5 · Tropicalization, multiplicities and balancing

189 NAG Tags

5.1 Tropicalization of subvarieties of a torus

Throughout these foundations, KK is a complete non-Archimedean field with a nontrivial absolute value. We write v(a)=−log⁡∣a∣,v(0)=+∞,Γ=v(K×)⊂R.v(a)=-\log|a|,\qquad v(0)=+\infty,\qquad \Gamma=v(K^\times)\subset\mathbb R. Algebraic closure will be stated when needed; it is a standing hypothesis of GK17, not a consequence of completeness. The real polyhedral constructions below do not themselves involve a field. We retain the series normalization d′d′′=d′d′′\mathrm d'\mathrm d''=\mathrm d'\mathrm d'', without a factor 2π2\pi and without squaring a norm.

Cancellation and repeated minima

Suppose that a1+⋯+as=0a_1+\cdots+a_s=0, with each ai≠0a_i\ne0, in a non-Archimedean valued field. The smallest of the numbers v(ai)v(a_i) must occur at least twice. Indeed, if aja_j had strictly smaller valuation than all the others, then ∣aj∣>max⁡i≠j∣ai∣but∣aj∣=∣∑i≠jai∣≤max⁡i≠j∣ai∣,|a_j|>\max_{i\ne j}|a_i| \quad\text{but}\quad |a_j|=\Big|\sum_{i\ne j}a_i\Big| \leq \max_{i\ne j}|a_i|, a contradiction. This observation also applies to evaluation at a Berkovich point: all the evaluated terms lie in its completed residue field H(x)\mathscr H(x).

Let T=Gmr=Spec⁡K[z1±1,…,zr±1]T=\mathbb G_m^r=\operatorname{Spec}K[z_1^{\pm1},\ldots,z_r^{\pm1}]. For a nonzero Laurent polynomial f=∑u∈Zrauzu,zu=z1u1⋯zrur,f=\sum_{u\in\mathbb Z^r}a_u z^u,\qquad z^u=z_1^{u_1}\cdots z_r^{u_r}, where only finitely many aua_u are nonzero, define Ff(w)=min⁡au≠0(v(au)+u1w1+⋯+urwr).\begin{equation} F_f(w)=\min_{a_u\ne0}\bigl(v(a_u)+u_1w_1+\cdots+u_rw_r\bigr). \end{equation} This is a continuous, concave, piecewise affine function on Rr\mathbb R^r. Concavity follows because it is the minimum of affine functions. Its tropical hypersurface is the set where that minimum is attained for at least two distinct exponents. The word hypersurface denotes this set with suitable multiplicities, not the graph of FfF_f.

At an analytic point x∈Tanx\in T^\mathrm{an}, put trop⁡(x)=(−log⁡∣z1(x)∣,…,−log⁡∣zr(x)∣).\operatorname{trop}(x)=\bigl(-\log|z_1(x)|,\ldots,-\log|z_r(x)|\bigr). If f(x)=0f(x)=0, the preceding argument proves that trop⁡(x)\operatorname{trop}(x) belongs to the tropical hypersurface of ff. For arbitrary xx, the ultrametric inequality gives only −log⁡∣f(x)∣ ≥ Ff(trop⁡(x)).\begin{equation} -\log|f(x)|\ \geq\ F_f(\operatorname{trop}(x)). \end{equation} Equality can fail because terms of equal size can cancel. We will obtain equality at the particular Gauss points of Section 6.1. It must not be used at every analytic point.

The tropical line

Example 5.1.1. (The MIN tropical line). For f=1+z1+z2f=1+z_1+z_2, all three coefficients have valuation zero, so Ff(w1,w2)=min⁡(0,w1,w2).F_f(w_1,w_2)=\min(0,w_1,w_2). Two terms can be minimal in exactly the following three ways: minimal termsconditionsray0, w1w1=0, w2≥0R≥0(0,1)0, w2w2=0, w1≥0R≥0(1,0)w1, w2w1=w2≤0R≥0(−1,−1).\begin{array}{c|c|c} \text{minimal terms}&\text{conditions}&\text{ray}\\ \hline 0,\ w_1&w_1=0,\ w_2\geq0&\mathbb R_{\geq0}(0,1)\\ 0,\ w_2&w_2=0,\ w_1\geq0&\mathbb R_{\geq0}(1,0)\\ w_1,\ w_2&w_1=w_2\leq0&\mathbb R_{\geq0}(-1,-1). \end{array} Their union will be denoted by LL. Each ray has weight one; the meaning of this weight is explained in the next section.

Three rays of a tropical line in the MIN convention
The tropical line for the MIN convention. The open regions record the unique smallest term; the thick rays record ties. Replacing MIN by MAX would reverse the picture.

This calculation is worth doing before memorizing a familiar tropical-line diagram. With the convention v=−log⁡∣⋅∣v=-\log|\cdot|, the rays point right, up, and southwest. On the first quadrant the constant term is smallest. On the left-hand region the z1z_1-term is smallest, and on the lower region the z2z_2-term is smallest.

Definition 5.1.2. (Tropicalization of a closed subscheme). For a closed subscheme X⊂TX\subset T, its set-theoretic tropicalization is Trop⁡(X)=trop⁡(Xan)⊂Rr.\operatorname{Trop}(X)=\operatorname{trop}(X^\mathrm{an})\subset\mathbb R^r. The scheme structure will enter through the weights, even though this underlying set does not detect nilpotents.

Theorem 5.1.3. (Polyhedrality and the hypersurface theorem). For a closed subscheme X⊂TX\subset T, the set Trop⁡(X)\operatorname{Trop}(X) is a finite union of integral Γ\Gamma-affine polyhedra. If XX is pure of dimension dd, this union can be made into a pure dd-dimensional polyhedral complex. If X=V(f)X=V(f) for a nonzero Laurent polynomial, then Trop⁡(X)\operatorname{Trop}(X) is exactly the locus of repeated minima of FfF_f.

Here the definition uses all Berkovich points. It requires no assertion that a prescribed real valuation vector is realized by a KK-rational point. Over an algebraically closed nontrivially valued extension K′K', one can instead take the closure of the coordinatewise valuations of X(K′)X(K'). The converse implication in the hypersurface assertion is a theorem about lifting tropical data, not a consequence of the ultrametric inequality alone.

References. [Gub13] (Theorem 3.3, Proposition 3.8, Theorem 5.6 and Example 5.8). The guide treats arbitrary valued ground fields and performs analytification over their completions. The elementary argument above proves the necessary cancellation condition.

For an ideal with several generators, intersecting the tropical hypersurfaces of those generators gives a tropical prevariety. It may be larger than Trop⁡(X)\operatorname{Trop}(X). The fundamental theorem uses the whole ideal, or a specially chosen finite tropical basis. A generating set for the ordinary ideal need not be a tropical basis.

Example 5.1.4. (Tropical prevarieties and translation). Assume ∣2∣=1|2|=1. The polynomials f1=1+z1+z2f_1=1+z_1+z_2 and f2=2+z1+z2f_2=2+z_1+z_2 have the same tropical polynomial min⁡(0,w1,w2)\min(0,w_1,w_2). Their tropical hypersurfaces therefore agree. Nevertheless, f2−f1=1f_2-f_1=1, so (f1,f2)(f_1,f_2) is the unit ideal and V(f1,f2)V(f_1,f_2) is empty.

More generally, for abc≠0abc\ne0, the tropical hypersurface of a+bz1+cz2a+bz_1+cz_2 is a translate of the standard tropical line. Subtracting v(a)v(a) from the three term valuations gives min⁡(0,w1+v(b)−v(a),w2+v(c)−v(a)).\min\bigl(0,w_1+v(b)-v(a),w_2+v(c)-v(a)\bigr). Its vertex is (v(a)−v(b),v(a)−v(c))(v(a)-v(b),v(a)-v(c)). Translation changes the vertex but preserves the primitive directions and the weights, which are one.

5.2 Lattices, polyhedra, multiplicities, and balancing

Character lattices and integral affine structures

The expression zuz^u is intrinsic to the torus even though its coordinate notation is not. Let M=Hom⁡(T,Gm)≃Zr,N=Hom⁡(M,Z),MR=M⊗ZR,NR=N⊗ZR.\begin{gathered} M=\operatorname{Hom}(T,\mathbb G_m)\simeq\mathbb Z^r,\qquad N=\operatorname{Hom}(M,\mathbb Z),\\ M_{\mathbb R}=M\otimes_\mathbb Z\mathbb R,\qquad N_{\mathbb R}=N\otimes_\mathbb Z\mathbb R. \end{gathered} Thus MM is the character lattice and NN its dual, the cocharacter lattice. Write χm\chi^m for the character associated to m∈Mm\in M. Tropicalization is characterized without coordinates by ⟨m,trop⁡(x)⟩=−log⁡∣χm(x)∣.\langle m,\operatorname{trop}(x)\rangle=-\log|\chi^m(x)|. A torus homomorphism induces an integral linear map on cocharacter lattices and hence a real linear map on tropicalizations. Translation by a torus point adds a constant valuation vector. These are the origins of the integral affine changes of coordinates used throughout GK17.

Definition 5.2.1. (Integral affine data). An integral Γ\Gamma-affine function on NRN_{\mathbb R} is a function ⟨m,w⟩+c\langle m,w\rangle+c, with m∈Mm\in M and c∈Γc\in\Gamma. An integral Γ\Gamma-affine polyhedron is an intersection of finitely many half-spaces {w:⟨mi,w⟩≥ci},mi∈M, ci∈Γ.\{w:\langle m_i,w\rangle\geq c_i\},\qquad m_i\in M,\ c_i\in\Gamma. Allowing ci∈Rc_i\in\mathbb R gives integral R\mathbb R-affine polyhedra. A polytope is a bounded polyhedron. A face is the polyhedron itself or its intersection with a supporting hyperplane. The relative interior means interior in the smallest affine space containing the polyhedron.

The adjective integral controls slopes, not the coordinates of every vertex. For example, 2t≥λ2t\geq\lambda is an integral Γ\Gamma-affine inequality when λ∈Γ\lambda\in\Gamma; its endpoint λ/2\lambda/2 need not lie in Γ\Gamma. In general, vertices may lie in the divisible hull of Γ\Gamma. This distinction is relevant to strict analytic domains and to model metrics over a fixed ground field.

A finite polyhedral complex C\mathscr C contains every face of each of its cells, and the intersection of two cells is empty or a face of both. Its support is their union. It is pure dd-dimensional if every maximal cell has dimension dd. A subdivision cuts cells into smaller cells without changing the support. The underlying tropical cycle will not depend on such cuts.

For a cell σ\sigma, let LσL_\sigma be the vector space parallel to its affine span and put Nσ=N∩Lσ.N_\sigma=N\cap L_\sigma. For a facet τ\tau of σ\sigma, the quotient Nσ/NτN_\sigma/N_\tau is a free rank-one lattice. Its generator pointing from τ\tau into σ\sigma is a primitive normal class. Choose a lift nσ,τ∈Nσn_{\sigma,\tau}\in N_\sigma. Another lift differs by an element of NτN_\tau.

Primitive refers to the lattice: a vector cannot be a positive integer multiple of another lattice vector on the same ray. It does not mean Euclidean length one. For instance, (1,2)(1,2) is primitive, and the natural integral coordinate on its line is s↦s(1,2)s\mapsto s(1,2); its Euclidean length 5\sqrt5 plays no role in a tropical multiplicity.

References. [GK17] (Appendix A.1–A.6 and Definition 1.1); [Gub13] (Appendix A.1–A.4).

Tropical multiplicities

The reason for attaching weights is already visible in one variable. The scheme zd=az^d=a has length dd, while its tropicalization is the single point v(a)/dv(a)/d. That point must carry weight dd to retain the algebraic degree. This remains true if the scheme is nonreduced, for example when the characteristic divides dd.

In higher dimensions, the relevant algebra is an initial degeneration. Here is a construction that does not assume a chosen section of the valuation. For w∈NRw\in N_{\mathbb R}, pass to a valued extension K′/KK'/K and choose t∈T(K′)t\in T(K') with trop⁡(t)=w\operatorname{trop}(t)=w. Translate XK′X_{K'} by t−1t^{-1}, take its schematic closure in the split torus over (K′)∘(K')^\circ, and take the special fibre. The result is denoted in⁡w(X)\operatorname{in}_w(X). Different choices give equivalent schemes after residue-field extension and torus translation.

For a Laurent polynomial, this operation retains the terms attaining the minimum in (5.1.1), after rescaling them to have valuation zero, and reduces their coefficients. For an ideal, one takes the initial forms of all its elements. The fundamental theorem says w∈Trop⁡(X)⟺in⁡w(X)≠∅.w\in\operatorname{Trop}(X)\quad\Longleftrightarrow\quad \operatorname{in}_w(X)\ne\varnothing. In the Laurent coordinate ring, a monomial is a unit. Thus an initial ideal containing a monomial defines the empty scheme.

At a regular point of Trop⁡(X)\operatorname{Trop}(X), lying in the relative interior of a maximal cell, its tropical multiplicity is the sum of the multiplicities of the geometric irreducible components of in⁡w(X)\operatorname{in}_w(X). An irreducible-component multiplicity means the length of the local ring at its generic point. Passing to an algebraically closed residue-field extension ensures that geometric components are counted. This integer is locally constant on the regular locus and therefore gives a weight mσm_\sigma to each maximal cell.

Example 5.2.2. (A binomial weight). Let q=dq0∈Mq=dq_0\in M, with q0q_0 primitive and d>0d>0, and consider X=V(χq−a),a∈K×.X=V(\chi^q-a),\qquad a\in K^\times. Its tropicalization is the hyperplane ⟨q,w⟩=v(a)\langle q,w\rangle=v(a), with weight dd. Extend q0q_0 to a basis of MM; in the resulting coordinates the equation is z1d=az_1^d=a. Its initial degeneration is a product of a zero-dimensional scheme of length dd with a torus of dimension r−1r-1. Thus the weight is a scheme-theoretic multiplicity, including repeated roots, rather than a count of distinct points.

References. [Gub13] (Definitions 5.1 and 13.4, Remark 5.7, Theorems 5.6 and 13.8).

Computing an initial ideal: why all relations matter

The phrase “keep the lowest-weight terms” first defines in⁡w(f)\operatorname{in}_w(f) for a single polynomial. For X=V(I)X=V(I), the geometric object is instead determined by in⁡w(I)=(in⁡w(f):f∈I),\operatorname{in}_w(I)=(\operatorname{in}_w(f):f\in I), with the coefficient normalization made over a suitable valued extension as above. A generating set of II need not be a tropical basis or a Gröbner basis at ww.

Here is a calculation at w=(0,0)w=(0,0). Assume the residue characteristic is not 22, choose 0<∣π∣<10<|\pi|<1, and work in K[x±1,y±1]K[x^{\pm1},y^{\pm1}]. Let f1=x+y+1,f2=x+y+1+π(y−1),I=(f1,f2).f_1=x+y+1,\qquad f_2=x+y+1+\pi(y-1),\qquad I=(f_1,f_2). The initial forms of these two displayed generators both equal x+y+1x+y+1. They define a line in the residue torus. But f2−f1π=y−1∈I,f1−(y−1)=x+2∈I.\frac{f_2-f_1}{\pi}=y-1\in I,\qquad f_1-(y-1)=x+2\in I. Thus I=(y−1,x+2)I=(y-1,x+2), and its initial degeneration is the single residue-torus point (−2,1)(-2,1). Taking initials of only the two original generators loses a relation and even gives the wrong dimension.

The construction by translation and closure is useful precisely because it is intrinsic to the ideal. A coefficientwise computation is reliable when one knows a suitable basis, or proves directly that the proposed initial generators generate the entire initial ideal.

Generic weights and cell multiplicities

To attach the weight mσm_\sigma to a maximal cell σ\sigma, choose ww in its relative interior in a subdivision on which the initial-degeneration type is controlled. This is a regular point of the tropical support: locally that support is a piece of an affine space of the expected dimension. Moving ww inside that regular region does not change its multiplicity.

For balancing at a codimension-one face τ\tau, there are several neighbouring maximal cells. Each carries the weight computed on its own relative interior. A point of τ\tau is useful for studying the entire local star, but it is not substituted for all these interior points in the definition of their weights. The multiplicity function need not be described as one locally constant function across every singular face of the support.

The algebraic origin of balancing

For the tropical line, three primitive vectors sum to zero. For a hypersurface, the dual Newton subdivision often makes the balancing identity an elementary polygon calculation. Neither argument proves the theorem for an arbitrary higher-codimension ideal.

The general proof relates the star near a point to the tropicalization of an initial degeneration. After this reduction, the balancing relation is an intersection-theoretic degree identity. Roughly, a character defines a principal divisor on an appropriate toric compactification; its intersection degree with the relevant proper cycle is zero. The contributions along the toric boundary are the multiplicities times the primitive normal components. Their sum is therefore zero. Passing back through the initial degeneration recovers the balancing equation for the original tropicalization.

This is a proof mechanism, not a replacement for the compactification and multiplicity arguments in [Gub13] (§13, Theorem 13.11). It explains the content of the theorem: the algebraic multiplicities satisfy a compatibility that arbitrary weights on the same support would not satisfy. Once balancing is known, Stokes translates it into closedness of the integration current.

The balancing condition

Definition 5.2.3. (A balanced tropical cycle). Give each maximal cell σ\sigma of a pure dd-dimensional integral R\mathbb R-affine complex an integer weight mσm_\sigma. The complex is balanced if, for every (d−1)(d-1)-dimensional face τ\tau, ∑σ⊃τmσnσ,τ ∈ Lτ.\begin{equation} \sum_{\sigma\supset\tau}m_\sigma n_{\sigma,\tau} \ \in\ L_\tau. \end{equation} Equivalently, the sum is zero in NR/LτN_{\mathbb R}/L_\tau. With integer weights it lies in NτN_\tau. A tropical cycle identifies balanced weighted complexes that agree after subdivision; cells of zero weight do not contribute to its support.

For a curve, τ\tau is a vertex and Lτ=0L_\tau=0. Balancing says that the weighted sum of primitive outgoing direction vectors is zero. The tropical line satisfies (1,0)+(0,1)+(−1,−1)=0.(1,0)+(0,1)+(-1,-1)=0. An arbitrary graph drawn with straight edges need not satisfy this equation. Conversely, a balanced weighted graph is a valid abstract tropical cycle even when no algebraic realization has been specified.

Theorem 5.2.4. (Balancing of algebraic tropicalizations). If XX is a pure-dimensional closed subscheme of a split torus, its tropical multiplicities make Trop⁡(X)\operatorname{Trop}(X) a balanced tropical cycle of dimension dim⁡X\dim X. The resulting cycle is independent of the chosen polyhedral subdivision.

Open proof obligation (RC2). The sole supplied proof source for the stated algebraic tropical balancing/subdivision claim is GubGuide Theorem 13.11 and Remark 13.12.

NAG-900009 · Proof status

Later, integration over a constant-weight cycle will define a current δC\delta_C. Balancing is precisely the cancellation of facet terms that makes this current d′\mathrm d'- and d′′\mathrm d''-closed. This explains why the combinatorial condition belongs in a differential theory. GK17 also allows smoothly varying real weights, with a pointwise balancing equation. In that setting derivatives of the weights remain inside the cells: balancing alone does not imply that δC\delta_C is closed [GK17] (Proposition 2.16). We will retain constant weights until that distinction has been explained.

Example 5.2.5. (Primitive directions and lattice index). For the rays with primitive directions (1,0),(0,1),(−2,−3)(1,0),(0,1),(-2,-3), balancing is (m1−2m3,m2−3m3)=(0,0).(m_1-2m_3,m_2-3m_3)=(0,0). The smallest positive integral weight vector is (2,3,1)(2,3,1). Replacing the primitive directions by Euclidean unit vectors changes their normalization and hence the coefficients in the balancing equation.

The cone generated by (1,0)(1,0) and (1,2)(1,2) is simplicial, since the two vectors are linearly independent. Their determinant is two, so the lattice they generate has index two in Z2\mathbb Z^2. Simpliciality does not imply unimodularity.

Example 5.2.6. (A binomial of multiplicity two). Let a∈K×a\in K^\times. The tropical support of z16z24−az_1^6z_2^4-a is 6w1+4w2=v(a).6w_1+4w_2=v(a). The exponent difference is 2(3,2)2(3,2), with (3,2)(3,2) primitive. The weight is therefore two. Equivalently, use the primitive character s=z13z22s=z_1^3z_2^2; the transverse equation is s2=as^2=a and has length two. In residue characteristic two the initial transverse scheme can be nonreduced, but its length remains two. Counting distinct geometric roots would lose this multiplicity.

5.3 Initial degenerations and multiplicity calculations

Initial degenerations over valued extensions

Let M≃ZrM\simeq\mathbb Z^r, let T=Spec⁡K[M]T=\operatorname{Spec}K[M], and let Z⊂TZ\subset T have ideal II. Fix w∈NRw\in N_{\mathbb R}. The valuation of a Laurent term amχma_m\chi^m at weight ww is v(am)+⟨m,w⟩.v(a_m)+\langle m,w\rangle. If ww is realized by a torus point pp over a suitable valued extension, substitute χm=pmχnewm\chi^m=p^m\chi^m_{\mathrm{new}}. For each relation, divide by an element whose valuation is the minimum of its term valuations, then reduce. The terms attaining the minimum survive; the others vanish in the residue field. The ideal generated by all these reduced relations defines the initial degeneration.

This description explains why one sometimes extends the field before writing an initial ideal with ordinary residue coefficients. A weight in NRN_{\mathbb R} need not lie in the original value group. Graded reduction or the corresponding tilted algebra packages the same information without pretending that every real weight is realized by a point over the original field. The eventual tropical multiplicity must be invariant under the permitted extensions and choices.

A nonempty variety whose chosen equations give a larger prevariety

Consider the ideal in three Laurent variables I=(x+y+1, x+y+z).I=(x+y+1,\ x+y+z). The variety is nonempty: it is given by x+y=−1x+y=-1 and z=1z=1, inside the torus. At the weight w=(−1,−1,1)w=(-1,-1,1), the two displayed generators both have their minimum attained by the terms xx and yy. Thus this weight belongs to the intersection of their two tropical hypersurfaces.

But subtracting the equations gives z−1∈Iz-1\in I. At the same weight, its constant term has value 00 and its zz term has value 11. Its initial form is the nonzero constant −1-1. Therefore the full initial ideal is the unit ideal, and the weight is not in Trop⁡(Z)\operatorname{Trop}(Z).

This example isolates the issue without an empty algebraic variety: solving the minimum-tie condition for a chosen generating set does not necessarily impose the minimum-tie condition for every algebraic relation. The intersection of the chosen tropical hypersurfaces is a tropical prevariety; obtaining the actual tropical variety requires a tropical basis or another method controlling the full initial ideal.

A multiplicity calculation that keeps nilpotents

For the one-variable equation xm=ax^m=a, the tropical support is the point w=v(a)/mw=v(a)/m. After a suitable translation and residue extension, the initial equation is zm−b~z^m-\widetilde b with b~≠0\widetilde b\ne0. Its zero-dimensional length is mm. If the residual characteristic divides mm, the polynomial can be inseparable and some geometric roots can coincide. The sum of scheme-theoretic multiplicities is still mm.

Thus a tropical multiplicity is not generally the number of distinct geometric roots of an initial equation. It counts lengths at generic components. This is why taking a reduced initial degeneration before measuring multiplicity can give the wrong answer.

In higher dimension, at a regular point of a maximal tropical cell, the initial degeneration has a torus symmetry parallel to the cell. Quotienting by that tangent torus leaves a zero-dimensional transverse scheme. Its length is the conceptual source of the cell weight. The algebraic construction makes this description invariant under the choices needed to realize the weight.

For the hypersurface defined scheme-theoretically by (1+x+y)2(1+x+y)^2, the support of the tropicalization is unchanged but its generic multiplicity is two. This is a useful distinction between a reduced support and a cycle. If one first replaces the subscheme by its reduction, the cycle changes. The multiplicity is not inferred from the picture of the three rays.

Why the weight is constant only in the appropriate region

The initial degeneration can change when ww crosses a Gröbner wall. On a sufficiently fine polyhedral subdivision, the relevant initial data are constant in the required sense on relative interiors. The tropical multiplicity is locally constant on the regular locus of a maximal cell. At a lower-dimensional face, several maximal cells meet and there is no single weight to obtain by pretending that the face is a generic point of one chosen cell.

When defining the weight attached to σ\sigma, choose ww in its relative interior away from the lower-dimensional exceptional locus. When proving balancing at a face τ\tau, choose a point in the relative interior of τ\tau and examine the local star, with the weights already assigned to adjacent maximal cells. These are two different choices of point serving two different purposes.

A proof mechanism for balancing

Near a codimension-one face, quotient the local star by the tangent space of that face. The transverse geometry is one-dimensional. In the corresponding initial-degeneration picture, characters of the quotient torus give rational functions on a compactification of the relevant curves. A principal divisor has degree zero. Its orders at the boundary points are the pairings of the character with the primitive outgoing directions, weighted by the transverse multiplicities. Hence, for every quotient character mm, ∑σ⊃τmσ⟨m,uσ/τ⟩=0.\sum_{\sigma\supset\tau}m_\sigma\langle m,u_{\sigma/\tau}\rangle=0. Characters separate vectors in the quotient real vector space, so the weighted normal sum vanishes. This explains the algebraic source of balancing. Making the argument general requires the initial-degeneration, compactification, and multiplicity theorems; the vector identity is its conclusion, not an extra combinatorial assumption about an algebraic tropicalization.

Example. (Valuation of a coefficient). For v(π)=λ>0v(\pi)=\lambda>0, the tropical polynomial of 1+x+πy1+x+\pi y is min⁡(0,w1,λ+w2)\min(0,w_1,\lambda+w_2). Its repeated-minimum locus is the standard tropical line translated by (0,−λ)(0,-\lambda). Its vertex is (0,−λ)(0,-\lambda), and the outgoing primitive directions remain (1,0),(0,1),(−1,−1)(1,0),(0,1),(-1,-1), all with weight one. The coefficient valuation affects position, not the tangent lattices.

References. [Gub13] (Sections 5–6, 10–13); [MS15].

5.4 Corners and stable intersections

Corner divisors

On a weighted tropical curve, a continuous piecewise affine function φ\varphi has a divisor at its vertices: div⁡C(φ)=∑p(∑e∋pme ∂eφ(p))[p].\begin{equation} \operatorname{div}_C(\varphi)= \sum_p\left(\sum_{e\ni p}m_e\,\partial_e\varphi(p)\right)[p]. \end{equation} Here ∂e\partial_e is differentiation in the primitive lattice coordinate pointing away from pp, and subdivision inserts vertices at all breakpoints. This is the outgoing-slope convention used throughout the series. On the line, div⁡R(max⁡(0,t))=[0],div⁡R(min⁡(0,t))=−[0].\operatorname{div}_\mathbb R(\max(0,t))=[0],\qquad \operatorname{div}_\mathbb R(\min(0,t))=-[0]. The coefficient is the right derivative minus the left derivative. These are the previously studied atomic-curvature calculations in a tropical divisor notation.

In higher dimension, a normal vector is defined only modulo a tangent vector. A formula must compensate for this ambiguity.

Definition 5.4.1. (The corner divisor). Let CC be a constant-weight tropical cycle and let φ\varphi be continuous and affine on every cell of a subdivision. Denote the linear part of its restriction to σ\sigma by ℓσ\ell_\sigma. The weight of div⁡C(φ)=φ⋅C\operatorname{div}_C(\varphi)=\varphi\cdot C on a facet τ\tau is mτ(φ⋅C)=∑σ⊃τmσℓσ(nσ,τ)−ℓτ(∑σ⊃τmσnσ,τ).\begin{equation} m_\tau(\varphi\cdot C)= \sum_{\sigma\supset\tau}m_\sigma\ell_\sigma(n_{\sigma,\tau}) - \ell_\tau\left(\sum_{\sigma\supset\tau}m_\sigma n_{\sigma,\tau}\right). \end{equation} Integral slopes give integral divisor weights when the weights of CC are integral. Real slopes are also allowed, with real divisor weights.

Balancing puts the vector in the second term in LτL_\tau, where ℓτ\ell_\tau is defined. Replacing a lift nσ,τn_{\sigma,\tau} by nσ,τ+an_{\sigma,\tau}+a, with a∈Nτa\in N_\tau, changes the first term by mσℓσ(a)m_\sigma\ell_\sigma(a) and the second by mσℓτ(a)m_\sigma\ell_\tau(a). These agree because φ\varphi is continuous along τ\tau. Hence the formula is independent of the lifts. If φ\varphi is the restriction of one ambient affine function, the two terms cancel completely.

The resulting divisor is again balanced and is independent of subdivision. This is a substantive compatibility theorem; it makes it possible to iterate corner operations. The same construction in GK17 applies to piecewise smooth functions and smooth weights, using derivatives at each point of the face.

References. [GK17] (Definition 1.10, Remark 1.11 and Propositions 1.12–1.14); Section 10.3.

There is a useful analytic dictionary. For constant-weight CC, d′d′′[φ]=δφ⋅C\begin{equation} \mathrm d'\mathrm d''[\varphi]=\delta_{\varphi\cdot C} \end{equation} when φ\varphi is piecewise affine and the currents are taken on CC. For a piecewise smooth function there is also the facewise Hessian: d′d′′[φ]=[dP′dP′′φ]+δφ⋅C.\mathrm d'\mathrm d''[\varphi]=[\mathrm d'_P\mathrm d''_P\varphi]+\delta_{\varphi\cdot C}. Thus differentiating each affine piece and obtaining zero does not compute the derivative of the whole current. The missing terms live on the interfaces.

References. [GK17] (Corollaries 3.19–3.20).

Example 5.4.2. (The sign of a tropical polynomial). The tropical polynomial FfF_f in (5.1.1) is concave. Consequently the effective tropical hypersurface is the corner divisor of −Ff-F_f, not of FfF_f. Near a general facet where two exponents differ by dq0dq_0, with q0q_0 primitive, the slope jump of −Ff-F_f in a transverse integral coordinate is d>0d>0. This recovers the binomial multiplicity.

In particular, for the tropical line, u(w)=max⁡(0,−w1,−w2)=−Ff(w),div⁡R2(u)=L,d′d′′[u]=δL.u(w)=\max(0,-w_1,-w_2)=-F_f(w),\qquad \operatorname{div}_{\mathbb R^2}(u)=L,\qquad \mathrm d'\mathrm d''[u]=\delta_L. This statement is on the real tropical chart. It does not replace the analytic function −log⁡∣f∣-\log|f| by Ff∘trop⁡F_f\circ\operatorname{trop} at points where cancellation occurs.

Example 5.4.3. (The correction term in the corner formula). Subdivide R2\mathbb R^2 along x=0x=0, with both weights one, and use normal lifts n+=(1,a)n_+=(1,a) and n−=(−1,b)n_-=(-1,b). For φ=2x+3y\varphi=2x+3y, formula (5.4.2) gives (2+3a)+(−2+3b)−3(a+b)=0.(2+3a)+(-2+3b)-3(a+b)=0. For ψ=max⁡(0,x)+3y\psi=\max(0,x)+3y, it gives (1+3a)+3b−3(a+b)=1.(1+3a)+3b-3(a+b)=1. Thus the affine function has no corner, whereas the second function has divisor [x=0][x=0]. The correction along the common tangent space is exactly what removes the dependence on the chosen lifts.

Example. (Slopes on the tropical line). Give the three rays outgoing slopes a,b,ca,b,c, using their primitive parameters. The coefficient of the corner at the central vertex is a+b+ca+b+c. For the restriction of an ambient affine function ℓ(w)=Aw1+Bw2+C\ell(w)=Aw_1+Bw_2+C, these slopes are B,A,−A−BB,A,-A-B, so the sum is zero. For a general piecewise affine function they need not sum to zero. In particular, slopes 0,0,−10,0,-1 give negative curvature at the vertex although the function is affine, hence convex, on every ray.

This calculation is the simplest reason the continuous facewise-convexity criterion in CLD12 requires correction. A condition on each maximal cell does not control the terms on their interfaces. Appendix A gives the exact source locator and the corrected interpretation.

Stable intersection of tropical cycles

Two coincident lines have a one-dimensional set-theoretic intersection, whereas the intersection of two codimension-one cycles in a plane should have dimension zero. Tropical geometry resolves this discrepancy by a displacement procedure.

For cycles C,D⊂NRC,D\subset N_{\mathbb R}, translate DD by εv\varepsilon v, with vv generic and ε>0\varepsilon>0 small, intersect transverse faces, and take the limiting cycle as ε→0\varepsilon\to0. When cells σ,τ\sigma,\tau meet transversely in complementary dimensions, their contribution has multiplicity mσmτ [N:Nσ+Nτ].\begin{equation} m_\sigma m_\tau\,[N:N_\sigma+N_\tau]. \end{equation} The index counts how far the two integral tangent lattices are from generating the ambient lattice. It is finite under the stated transversality condition. For two primitive lines in Z2\mathbb Z^2, it is the absolute determinant of their direction vectors.

Theorem 5.4.5. (Stable intersection). For constant-weight integral R\mathbb R-affine tropical cycles in NRN_{\mathbb R}, the displacement construction defines a cycle C⋅DC\cdot D, independent of the sufficiently small generic displacement and of subdivision. It is bilinear, associative and commutative, its codimension is the sum of the codimensions, and its support is contained in ∣C∣∩∣D∣|C|\cap|D|. The construction extends to real weights and to the smooth weights used by GK17.

Independence of displacement uses balancing. It is not a general fact about arbitrary weighted polyhedral sets. The combinatorial intersection is defined before any attempt to multiply singular currents.

References. [GK17] (Sections 1.3–1.4, especially formula (1.4.1)).

Example 5.4.6. (A transverse intersection of multiplicity two). In R2\mathbb R^2, give each of the lines H={x=0},Q={x+2y=0}H=\{x=0\},\qquad Q=\{x+2y=0\} weight one. Their primitive tangent vectors are (0,1)(0,1) and (2,−1)(2,-1). They generate the sublattice of vectors with even first coordinate, of index two. Hence H⋅Q=2[0].H\cdot Q=2[0]. The corresponding algebraic example is z1=1,z1z22=1.z_1=1,\qquad z_1z_2^2=1. Its coordinate algebra is K[z2±1]/(z22−1)K[z_2^{\pm1}]/(z_2^2-1), of length two. In characteristic two the two roots merge into a nonreduced point; the intersection length and tropical index remain two. Both individual tropical lines have weight one because their defining exponent vectors (1,0)(1,0) and (1,2)(1,2) are primitive.

Two transversal lines and the fundamental parallelogram of their tangent vectors
Transversality does not force multiplicity one. The two primitive tangent vectors span a fundamental parallelogram of lattice area two; the quotient lattice has two elements.

The same index will appear in products of curvature currents. With u=max⁡(0,x)u=\max(0,x) and v=max⁡(0,x+2y)v=\max(0,x+2y), the corner cycles are HH and QQ. Their product in the delta-calculus is therefore 2δ02\delta_0. This formula is justified by that calculus; it is not a definition of the product of two arbitrary distributions.

Example 5.4.7. (Two different self-intersections). A straight line translated in a nonparallel direction becomes disjoint from itself, so its stable self-intersection is zero. The tropical line LL behaves differently. Translate its vertex by (ε,2ε)(\varepsilon,2\varepsilon). There is one transverse intersection, at (0,ε)(0,\varepsilon), between the original vertical ray and the translated southwest ray. Their primitive directions have determinant of absolute value one. Letting ε→0\varepsilon\to0 gives L⋅L=[0].L\cdot L=[0]. Although both set-theoretic self-intersections contain the entire original curve, the stable answers differ.

Example 5.4.8. (Determinants and intersection multiplicities). Let p=(a,b)p=(a,b) and q=(c,d)q=(c,d) be primitive integral covectors with ad−bc≠0ad-bc\ne0. The weight-one lines ⟨p,w⟩=0\langle p,w\rangle=0 and ⟨q,w⟩=0\langle q,w\rangle=0 have primitive tangent vectors (−b,a)(-b,a) and (−d,c)(-d,c). Their stable intersection is the origin with weight [Z2:Z(−b,a)+Z(−d,c)]=∣ad−bc∣.[\mathbb Z^2:\mathbb Z(-b,a)+\mathbb Z(-d,c)]=|ad-bc|. The torus homomorphism (z1,z2)↦(z1az2b,z1cz2d)(z_1,z_2)\mapsto(z_1^az_2^b,z_1^cz_2^d) has the same degree. Indeed, its character-lattice map has cokernel of order ∣ad−bc∣|ad-bc|, so its finite kernel has that scheme-theoretic length. The formula remains valid in inseparable characteristic, where the kernel need not be reduced.

Example. (A nontrivial lattice index). Let two weighted rational lines in R2\mathbb R^2 have primitive directions u=(1,0)u=(1,0) and v=(1,2)v=(1,2), with weights aa and bb. A generic small translate meets transversely, and the stable intersection multiplicity is ab[Z2:Zu+Zv]=ab∣det⁡(u,v)∣=2ab.ab[\mathbb Z^2:\mathbb Zu+\mathbb Zv]=ab|\det(u,v)|=2ab. The factor two is an index of integral tangent lattices. It is not the angle between the lines. If the two lines coincide, their set-theoretic intersection is one-dimensional, whereas the stable intersection is defined by a displacement construction and the expected codimension. This is why ordinary intersection of supports cannot be substituted for tropical intersection of cycles.

A displaced tropical line meeting the original line transversely

The set-theoretic self-intersection of a tropical line is the whole line. A generic displacement gives one transverse point of lattice multiplicity one, which tends to the vertex. The stable self-intersection is that weighted vertex.

5.5 Weighted Gauss points and the role of cancellation

The tropical image contains real vectors which need not be valuations of KK-rational points. Weighted Gauss seminorms explain this already for a torus, without changing the ground field.

Proposition 5.5.1. Let KK be a complete non-Archimedean field, M≃ZrM\simeq\mathbb Z^r, and w∈NRw\in N_{\mathbb R}. The rule ∣f∣w=max⁡m∣am∣e−⟨m,w⟩,f=∑mamχm∈K[M],|f|_w=\max_m|a_m|e^{-\langle m,w\rangle},\qquad f=\sum_m a_m\chi^m\in K[M], is a multiplicative norm extending that of KK. It determines a point s(w)s(w) of the analytified torus, and trop⁡(s(w))=w\operatorname{trop}(s(w))=w. The map ss is continuous.

Proof. Choose a basis of MM and put ri=e−wir_i=e^{-w_i}. The Laurent-polynomial case of Proposition 1.2.2 gives multiplicativity with these radii.

The norm is nonzero on every nonzero polynomial and has the specified values on characters. These facts give the analytic point and its tropicalization. Continuity follows from the topology of analytification: evaluation of each Laurent polynomial along ss is the maximum of finitely many continuous functions. ◻

At these points one has −log⁡∣f(s(w))∣=Ff(w)-\log|f(s(w))|=F_f(w). At a general analytic point with the same tropicalization only the inequality −log⁡∣f(x)∣≥Ff(w)-\log|f(x)|\geq F_f(w) holds. Thus ss is a section on the torus, but it is not a section on every subvariety: if ff vanishes on a subvariety, the Gauss norm of a nonzero ff remains nonzero. A point of the subvariety over ww must carry cancellation information absent from the torus Gauss point.

5.6 Primitive lattices and the index in a pushforward

Lemma 5.6.1. If L⊂NRL\subset N_{\mathbb R} is a rational linear subspace, then NL=N∩LN_L=N\cap L is saturated in NN. If Lτ⊂LσL_\tau\subset L_\sigma are rational subspaces with dimension difference one, Nσ/NτN_\sigma/N_\tau is free of rank one.

Proof. If kn∈NLkn\in N_L for a nonzero integer kk and n∈Nn\in N, then n=k−1(kn)∈Ln=k^{-1}(kn)\in L, so n∈NLn\in N_L. Thus the quotient has no torsion. The same argument inside NσN_\sigma proves that Nσ/NτN_\sigma/N_\tau is torsion free. Its rank is the dimension difference, since both subspaces are rational. A finitely generated torsion-free abelian group of rank one is infinite cyclic. ◻

Proposition 5.6.2. Let F:N→N′F:N\to N' be integral linear and injective on the span of a dd-dimensional rational polyhedron σ\sigma. Let σ′=F(σ)\sigma'=F(\sigma) and e=[Nσ′′:F(Nσ)].e=[N'_{\sigma'}:F(N_\sigma)]. For a compactly supported (d,d)(d,d)-superform α\alpha on the relative interior of σ′\sigma' one has ∫σF∗α=e∫σ′α.\int_\sigma F^*\alpha=e\int_{\sigma'}\alpha.

Proof. Choose bases of the two tangent lattices. The resulting square integral matrix AA has nonzero determinant and ∣det⁡A∣=e|\det A|=e. Pullback acts on both exterior factors, so the top-form coefficient is multiplied by (det⁡A)2(\det A)^2. The ordinary change of variables in the scalar integral contributes ∣det⁡A∣−1|\det A|^{-1}. Their product is ee. The sign convention for top superforms is the same on both sides and does not introduce an orientation sign. ◻

For finitely many maximal cells mapping onto one image cell, summing this formula produces the weight ∑σmσ[Nσ′′:F(Nσ)]\sum_\sigma m_\sigma[N'_{\sigma'}:F(N_\sigma)]. This explains the lattice factor in tropical pushforward and integration. Identifying that weighted pushforward with the tropicalization of an algebraic image, divided by the generic algebraic degree when appropriate, requires the Sturmfels–Tevelev multiplicity theorem. The change-of-variables calculation does not prove that algebraic theorem.

Example 5.6.3. For a nonzero integer dd and F(t)=dtF(t)=dt on the tropical line, F∗(a(y)d′y∧d′′y)=d2a(dt)d′t∧d′′tF^*(a(y)d'y\wedge d''y)=d^2a(dt)d't\wedge d''t. Integration gives ∣d∣∫a(y)dy|d|\int a(y)dy. The same factor is the degree of the torus isogeny z↦zdz\mapsto z^d. If the characteristic of the fibre field divides dd, distinct geometric points need not count that degree; scheme lengths do. For an initial fibre this is the residue characteristic. An isogeny over a characteristic-zero ground field remains separable even when its degree is divisible by the residue characteristic.

5.7 Monomial maps and scheme-theoretic degree

The lattice indices in tropical intersection are algebraic degrees. The following calculation proves this for a general isogeny of split tori, including inseparable characteristic. It also distinguishes the multiplicity of an initial degeneration from the number of geometric points in a generic fibre.

Let MM and M′M' be free abelian groups of rank rr, and let A:M′↪MA:M'\hookrightarrow M be injective with finite cokernel. Over any field kk, the inclusion of group algebras gives a morphism f:TM=Spec⁡k[M]⟶TM′=Spec⁡k[M′],f∗χm′=χAm′.f:T_M=\operatorname{Spec}k[M]\longrightarrow T_{M'}=\operatorname{Spec}k[M'], \qquad f^*\chi^{m'}=\chi^{A m'}.

Proposition 5.7.1. The map ff is finite faithfully flat of degree d=[M:A(M′)]d=[M:A(M')]. Every fibre algebra has vector-space dimension dd over the residue field of the target point.

Proof. Choose one representative bb from each coset in M/A(M′)M/A(M'). Every monomial has a unique expression χbχAm′\chi^{b}\chi^{A m'}, so k[M]=⨁b∈M/A(M′)k[M′]χbk[M]=\bigoplus_{b\in M/A(M')}k[M']\chi^b as a module over k[M′]k[M']. Independence follows because distinct monomials form a kk-basis of the group algebra. Hence this algebra is finite free of positive rank dd, and thus faithfully flat. Tensoring with the residue field of a point of TM′T_{M'} produces a dd-dimensional algebra over that field. It is finite, hence zero-dimensional, and its dimension is its scheme length with residue degrees counted. ◻

Proposition 5.7.2. (Smith normal form and inseparability). After integral changes of bases, ff is the map (z1,…,zr)⟼(z1d1,…,zrdr),di>0,d1∣d2∣⋯∣dr,(z_1,\ldots,z_r)\longmapsto(z_1^{d_1},\ldots,z_r^{d_r}), \quad d_i>0,\quad d_1\mid d_2\mid\cdots\mid d_r, and d=∏idi=∣det⁡A∣d=\prod_i d_i=|\det A|. Over an algebraically closed field of characteristic p>0p>0, write di=paieid_i=p^{a_i}e_i with p∤eip\nmid e_i. Each geometric fibre has ∏iei\prod_i e_i distinct points, each of length p∑iaip^{\sum_i a_i}. In characteristic zero every geometric fibre is reduced with dd points.

Proof. Smith normal form is the integer row-and-column reduction of a full-rank integral matrix to the displayed diagonal matrix. The basis changes are torus automorphisms, so they preserve the fibre schemes and degree. In one coordinate, after choosing a paip^{a_i}th root bib_i of the nonzero target coordinate, zidi−bipai=(ziei−bi)pai.z_i^{d_i}-b_i^{p^{a_i}} =(z_i^{e_i}-b_i)^{p^{a_i}}. The polynomial ziei−biz_i^{e_i}-b_i has eie_i distinct nonzero roots, since its derivative does not vanish at any root. The local quotient at each root is isomorphic to k[t]/(tpai)k[t]/(t^{p^{a_i}}) and has length paip^{a_i}. Tensoring these one-coordinate algebras proves the count and the length in rr coordinates. The same derivative argument with ai=0a_i=0 applies in characteristic zero. ◻

This statement uses the characteristic of the field over which the fibre is formed. If KK has characteristic zero and residue characteristic pp, the isogeny z↦zpz\mapsto z^p has reduced geometric fibres over KK, while its reduction can have a nonreduced fibre of length pp. The total degree is pp in both calculations; the source of multiplicity differs. Tropical multiplicity uses the appropriate initial degeneration over a residue field.

5.8 Binomial complete intersections and tropical weights

Return to a complete non-Archimedean field KK. Let m1,…,mrm_1,\ldots,m_r be linearly independent elements of M≃ZrM\simeq\mathbb Z^r, let ai∈K×a_i\in K^\times, and set Z=V(χm1−a1,…,χmr−ar)⊂TM.Z=V(\chi^{m_1}-a_1,\ldots,\chi^{m_r}-a_r)\subset T_M.

Theorem 5.8.1. The scheme ZZ is finite of length d=[M:∑iZmi]d=[M:\sum_i\mathbb Zm_i]. Its tropicalization is one point w0w_0, uniquely determined by ⟨mi,w0⟩=v(ai)(1≤i≤r).\langle m_i,w_0\rangle=v(a_i)\quad(1\leq i\leq r). This point has tropical multiplicity dd. The conclusion includes nonreduced schemes and weights outside the original value group.

Proof. The equations define the fibre of the isogeny associated with the sublattice generated by the mim_i, over the point (a1,…,ar)(a_1,\ldots,a_r). Proposition 5.7.1 proves the length and nonemptiness. Every analytic point of ZZ satisfies the displayed valuation equations. They have a unique real solution because the mim_i form a real basis. A finite nonzero KK-algebra has a maximal ideal with finite residue extension of KK; the extended absolute value gives an analytic point. Thus the tropicalization is exactly {w0}\{w_0\}.

Choose a valued extension containing a solution tt of the binomial system. Translation z=tyz=t y changes the equations to χmi(y)=1\chi^{m_i}(y)=1. The closure of this kernel in the torus over the valuation ring is finite free of rank dd: the same coset-basis proof works over a ring, and imposing the target coordinates equal to one preserves that free basis. Its generic fibre is the translated ZZ, and it has no valuation-ring torsion. Hence it is the schematic closure. Its special fibre has dimension dd as a vector space over the residue field. After algebraic closure of that residue field, the sum of local lengths is still dd. This is exactly the tropical multiplicity in dimension zero. ◻

Corollary 5.8.2. (Transverse hypersurfaces). The stable product of the rr tropical binomial hypersurfaces is d[w0]d[w_0], equal to the tropical cycle of ZZ.

Proof. Write mi=bimˉim_i=b_i\bar m_i with mˉi\bar m_i primitive. The iith hypersurface has weight bib_i. The primitive normal covectors mˉi\bar m_i are real-linearly independent. The transverse lattice-index rule gives multiplicity ∣det⁡(mˉ1,…,mˉr)∣|\det(\bar m_1,\ldots,\bar m_r)| for the weight-one hyperplanes; multiplying their weights gives ∣det⁡(m1,…,mr)∣=d|\det(m_1,\ldots,m_r)|=d. The support is their unique intersection point w0w_0. This uses the transverse part of the tropical intersection construction, not a general lifting theorem for intersections. ◻

Example 5.8.3. Consider in a three-dimensional torus the equations x2=a,xy3=b,yz5=c.x^2=a,\qquad xy^3=b,\qquad yz^5=c. The exponent matrix has determinant 3030, so the intersection has length 3030. Its tropical point satisfies wx=v(a)2,wy=v(b)−wx3,wz=v(c)−wy5.w_x=\frac{v(a)}2,\qquad w_y=\frac{v(b)-w_x}3,\qquad w_z=\frac{v(c)-w_y}5. The first hypersurface has weight two. The other two exponent vectors are primitive and give weight-one hypersurfaces. Their three primitive normal vectors have determinant 1515; multiplying by the weight two recovers 3030. Thus neither the individual weights nor the primitive intersection index alone gives the answer. The computation remains valid in characteristics two, three and five.

For non-binomial equations, a transverse intersection of tropical supports need not by itself identify an algebraic scheme intersection without the relevant lifting and proper-intersection hypotheses. The binomial proof succeeds because one finite free algebra controls both the generic fibre and the initial fibre. In the general theory, flatness, initial degenerations and multiplicity comparison replace that explicit algebra. This is the point at which the geometric results used by Gubler and collaborators enter the calculus.

References. [Gub13] (Sections 5 and 13); [GK17] (Section 1). The finite-free and initial-fibre calculations are proved here; the general tropical multiplicity and intersection conventions are those of the cited sources.

5.9 Real weights and valued extensions

A tropical point is a real vector; it is not required to be the valuation vector of a point over the original ground field. There is a direct construction of a complete valued extension realizing any prescribed vector. It explains why the definition by initial degenerations remains meaningful for arbitrary real weights.

Proposition 5.9.1. For every w=(w1,…,wr)∈Rrw=(w_1,\ldots,w_r)\in\mathbb R^r, there is a complete valued extension L/KL/K and elements ti∈L×t_i\in L^\times with v(ti)=wiv(t_i)=w_i. The absolute value on LL restricts to the given one on KK.

Proof. On the Laurent polynomial ring K[T1±1,…,Tr±1]K[T_1^{\pm 1},\ldots,T_r^{\pm 1}] take the weighted Gauss norm ∣∑mamTm∣w=max⁡m∣am∣e−⟨m,w⟩.\left|\sum_m a_m T^m\right|_w =\max_m |a_m|e^{-\langle m,w\rangle}. It is multiplicative. One proof uses the associated graded ring: the initial homogeneous part of a nonzero polynomial is nonzero, and the graded Laurent polynomial ring over the graded residue field is a domain, so initial parts of a product cannot multiply to zero. Equivalently, refine the real weight order by a monomial order to select a leading exponent; the leading coefficients multiply without cancellation. Extend the norm to the fraction field by ∣f/g∣w=∣f∣w/∣g∣w|f/g|_w=|f|_w/|g|_w, then complete. The images tit_i of TiT_i have the prescribed norms, while constants keep their original norms. ◻

This is a valued extension construction, not a finite-extension assertion. A finite extension cannot introduce an arbitrary real value over a discretely valued field: its value group remains a finite-index enlargement. Thus strictification by a general analytic extension and descent through a finite field extension are different operations.

5.10 The closure algebra and all initial relations

Fix L,tL,t as above and put R=L∘R=L^\circ, m=L∘∘\mathfrak m=L^{\circ\circ} and k=R/mk=R/\mathfrak m. Translate the given subscheme X⊂TX\subset T by t−1t^{-1}, so its ideal becomes It⊂L[z1±1,…,zr±1]I_t\subset L[z_1^{\pm 1},\ldots,z_r^{\pm 1}]. The closure algebra in the torus over RR is A=R[z1±1,…,zr±1]/J,J=It∩R[z1±1,…,zr±1].A=R[z_1^{\pm1},\ldots,z_r^{\pm1}]/J, \qquad J=I_t\cap R[z_1^{\pm1},\ldots,z_r^{\pm1}].

Proposition 5.10.1. The algebra AA is torsion-free over RR and hence flat over this valuation ring. Its special-fibre ideal is generated by the reductions of all normalized polynomials in ItI_t: divide each nonzero polynomial by one of its coefficients of maximal absolute value and then reduce.

Proof. The map A→L[z±1]/ItA\to L[z^{\pm 1}]/I_t is injective by the definition of JJ. Its target has no RR-torsion, because every nonzero element of RR is invertible in LL. Thus AA is torsion-free. We use the standard valuation-ring criterion that every torsion-free module is flat.

A normalized polynomial has coefficients in RR, with at least one unit coefficient, so it belongs to JJ and reduces to a nonzero initial polynomial. Conversely, take g∈Jg\in J. If all its coefficients belong to m\mathfrak m, its reduction is zero. Otherwise a coefficient of maximal absolute value is a unit of RR; dividing by that coefficient changes its reduction by a nonzero scalar in kk and gives one of the stated normalized relations. These two observations identify the ideals in k[z±1]k[z^{\pm 1}]. ◻

The proof shows why dividing by a small scalar is necessary even if it is not a unit in RR. The relation belongs to the generic-fibre ideal, where that scalar is invertible; taking the saturated closure then retains its normalized reduction. Reducing only an arbitrary list of original equations computes a potentially larger scheme. The explicit example with f2−f1=π(y−1)f_2-f_1=\pi(y-1) earlier in this chapter is exactly this saturation phenomenon.

If t′t' is another choice with the same valuation vector, the coordinate ratios ti′/tit'_i/t_i have absolute value one after passage to a common valued extension. Their reductions are units. They induce a torus translation of the special fibre, so geometric component lengths and the resulting tropical multiplicities are unchanged. This is the elementary part of independence from the translating point. General valued-base-change invariance of tropical multiplicities remains the comparison theorem used in their definition.

Where the lifting theorem enters

Suppose an analytic point of XX has tropicalization ww. Evaluation in its completed residue field sends each normalized relation to zero, so its reductions cannot generate the unit ideal in the appropriate residue torus. This gives the elementary implication from an analytic point to nonempty initial degeneration. The reverse implication requires lifting from a nonempty reduction to an analytic point of the prescribed tropical fibre. It is the geometric content of the fundamental theorem, not a consequence of the termwise minimum rule alone.

Likewise, flatness of a closure does not by itself identify the initial scheme obtained from any unsaturated presentation. The closure, its relevant analytic domain, and the reduction map must refer to the same normalized ideal. Keeping these objects aligned prevents a common circular argument in informal explanations of tropicalization.

5.11 Hypersurface weights from Newton edges

At a generic point of a maximal face of a tropical hypersurface, the active exponents form a segment in the Newton subdivision. Let its endpoints be m0m_0 and m0+dqm_0+dq, where q∈Mq\in M is primitive and d>0d>0. The initial polynomial has the form χm0P(χq),P(z)=c0+c1z+⋯+cdzd,c0cd≠0.\chi^{m_0}P(\chi^q),\qquad P(z)=c_0+c_1z+\cdots+c_dz^d, \quad c_0c_d\ne0.

Proposition 5.11.1. The tropical multiplicity of this maximal face is dd, the lattice length of the dual Newton edge. This remains valid in inseparable residue characteristic.

Proof. Extend qq to a basis of MM. Since χm0\chi^{m_0} is a unit, the initial scheme becomes Spec⁡k[z±1]/(P(z)) × (Gm)r−1.\operatorname{Spec}k[z^{\pm1}]/(P(z))\ \times\ (\mathbb G_m)^{r-1}. The first algebra has dimension dd over kk. Indeed PP has degree dd and nonzero constant term, so zz is already invertible in k[z]/(P)k[z]/(P) and localization does not change that quotient. After algebraic closure it decomposes into local Artinian algebras at its roots, whose lengths sum to dd. The product with the torus preserves these generic component lengths. No separability assumption is used. ◻

In a two-dimensional transverse slice at a codimension-two face, the dual Newton cell is a polygon. List its oriented edge vectors as diqid_iq_i around the boundary. Their sum is zero. After choosing dual oriented lattice bases, rotating each primitive qiq_i to the corresponding primitive outgoing normal uiu_i gives ∑idiui=0.\sum_i d_i u_i=0. By Proposition 5.11.1, the did_i are exactly the hypersurface weights. This proves balancing for the hypersurface in that local slice, and hence modulo the tangent lattice of the original face. The Newton-subdivision correspondence supplies the duality of the slice and polygon; the actual balancing equation is the closed-polygon identity. Higher-codimension ideals need the more general multiplicity theorem.

The sign of the tropical divisor

The tropical polynomial F(w)=min⁡m(v(am)+⟨m,w⟩)F(w)=\min_m(v(a_m)+\langle m,w\rangle) is concave. With the positive Hessian convention of this book, its effective hypersurface current is therefore δTrop⁡(V(f))=d′d′′[−F].\delta_{\operatorname{Trop}(V(f))}=\mathrm d'\mathrm d''[-F]. The corner coefficient across a maximal face is the jump between the two active slopes, and the primitive normal calculation gives the lattice length dd. This proves the formula from the corner identity. It is a formula on the real tropical space. It must not be replaced by an equality log⁡∣f∣=−F∘trop⁡\log|f|=-F\circ\operatorname{trop} on every analytic point: cancellation makes that equality fail away from the Gauss section, as seen at the beginning of the chapter.

5.12 Why the weight is locally constant on the regular locus

Proposition 5.12.1. Let CC be a balanced pure dd-dimensional tropical cycle with constant weights on a polyhedral complex. If an open part of its support is an open subset of a single dd-dimensional affine space, its weight is locally constant there.

Proof. At a facet introduced by the subdivision inside this affine space, precisely two local maximal sides occur. Their tangent lattices agree, and their outgoing primitive classes in the rank-one quotient are opposite. Balancing says m+n−m−n=0m_+n-m_-n=0, so m+=m−m_+=m_-. In a sufficiently small affine ball, two generic interior points can be joined by a polygonal path avoiding the codimension-two skeleton and crossing only facets. Repeated application of this equality identifies the weights along the path. Points on lower-dimensional subdivision faces inherit this common local weight. ◻

This is an alternative explanation of local constancy once the balancing theorem has been established. It does not say that the initial schemes are all isomorphic, nor that one weight can be extended unchanged across a singular tropical vertex. In computations, choose a point inside each maximal cell to obtain its multiplicity, and then use the separate weights in the balancing equation at their common face.

References. Initial degenerations and the lifting theorem are treated in [Gub13] (Sections 5–6 and 13); the closure and hypersurface calculations above make the normalization and multiplicities explicit. For the valuation-ring flatness criterion, see https://stacks.math.columbia.edu/tag/0539.

5.13 Initial ideals and analytic tropical fibres

Let KK be a complete non-Archimedean field, put v=−log⁡∣⋅∣v=-\log|\cdot|, and write M=ZnM=\mathbf Z^n. For w∈Rnw\in\mathbf R^n and f=∑mamχm∈K[M]f=\sum_m a_m\chi^m\in K[M], put ∥f∥w=max⁡m∣am∣e−⟨m,w⟩,μw(f)=min⁡am≠0(v(am)+⟨m,w⟩).\|f\|_w=\max_m |a_m|e^{-\langle m,w\rangle}, \qquad \mu_w(f)=\min_{a_m\ne0}\bigl(v(a_m)+\langle m,w\rangle\bigr). A term is active if it attains this maximum, or equivalently this minimum. Denote by Tw\mathcal T_w the completion of K[M]K[M] for ∥⋅∥w\|\cdot\|_w. The multilaurent presentation of Theorem 1.17.2 identifies it, as a normed algebra, with K{r1−1T1,r1S1,…,rn−1Tn,rnSn}/(TiSi−1)i,ri=e−wi.K\{r_1^{-1}T_1,r_1S_1,\ldots,r_n^{-1}T_n,r_nS_n\}/(T_iS_i-1)_i, \qquad r_i=e^{-w_i}.

Lemma 5.13.1. NAG-002080 (Coordinate bases for weighted subspaces). Let E=KqE=K^q have the norm ∥(ai)∥=max⁡i∣ai∣ri\|(a_i)\|=\max_i|a_i|r_i, where every ri>0r_i>0. For every ss-dimensional subspace W⊂EW\subset E there are distinct indices b1,…,bsb_1,\ldots,b_s and a basis g1,…,gsg_1,\ldots,g_s of WW such that (gi)bj=δij,∥gi∥=rbi,∥∑icigi∥=max⁡i∣ci∣rbi.(g_i)_{b_j}=\delta_{ij},\qquad \|g_i\|=r_{b_i},\qquad \Bigl\|\sum_i c_i g_i\Bigr\|=\max_i|c_i|r_{b_i}. The same assertions hold after every isometric field extension, with the same basis and weights. If all ri=1r_i=1, reduction of the unit ball of WW has dimension ss over K~\widetilde K, and its reduction commutes with residue-field extension.

Proof. Choose a nonzero vector in WW, choose a coordinate at which its weighted norm is attained, and divide by that coordinate. This gives g1g_1 with pivot b1b_1, pivot coefficient one, and norm rb1r_{b_1}. Apply induction to the kernel in WW of the b1b_1-coordinate. Its basis vectors have zero b1b_1-coordinate. Subtract their pivot multiples from g1g_1 to make all other pivot coordinates of g1g_1 zero. Each subtraction has norm at most rb1r_{b_1}: its scalar coefficient cc occurs in a pivot coordinate of weight rbr_b and satisfies ∣c∣rb≤rb1|c|r_b\le r_{b_1}. The b1b_1-coordinate stays one, so the norm stays rb1r_{b_1}. This completes the construction.

The ultrametric inequality gives the upper bound for a linear combination. Its pivot coordinates are exactly the cic_i, giving the reverse bound. These calculations remain valid over an isometric extension. For unit weights the unit ball is therefore ⨁iK∘gi\bigoplus_iK^\circ g_i, its reduction has basis g‾i\overline g_i, and the identical description over the extension proves the last assertion. ◻

Proposition 5.13.2. NAG-002081 (The tropical fibre algebra). Let I⊂K[M]I\subset K[M] be an ideal and X=Spec⁡(K[M]/I)X=\operatorname{Spec}(K[M]/I). For every real weight ww, the fibre Xw={x∈Xan:−log⁡∣χei(x)∣=wi}X_w=\{x\in X^{\mathrm{an}}:-\log|\chi^{e_i}(x)|=w_i\} is the affinoid space Xw=M(Tw/ITw).X_w=\mathcal M(\mathcal T_w/I\mathcal T_w). The ideal ITwI\mathcal T_w is closed and is the norm closure of II in Tw\mathcal T_w.

Proof. Theorem 1.18.3 makes every ideal of Tw\mathcal T_w closed. Write an element of ITwI\mathcal T_w as a finite sum ∑jhjfj\sum_j h_j f_j with fj∈If_j\in I and hj∈Twh_j\in\mathcal T_w. Approximating the finitely many hjh_j by Laurent polynomials approximates this sum by elements of II. Thus its closure is precisely ITwI\mathcal T_w.

An analytic point in the indicated fibre bounds every Laurent polynomial by ∥⋅∥w\|\cdot\|_w, so its seminorm extends to Tw\mathcal T_w and kills ITwI\mathcal T_w. Conversely, the bounds on Ti,SiT_i,S_i and the identities TiSi=1T_iS_i=1 force ∣Ti∣=ri|T_i|=r_i at every point of the displayed quotient. Restriction to K[M]/IK[M]/I gives the corresponding point of XanX^{\mathrm{an}}, by Proposition 2.10.1. The two constructions are inverse, continuous in the evaluation topologies, and identify the analytic structures by the rational-domain universal property of Theorem 2.31.2, applied to the finitely many coordinate and inverse-coordinate inequalities. ◻

Theorem 5.13.3. NAG-002082 (The tropical fundamental theorem over the original field). With the preceding notation, the following conditions on w∈Rnw\in\mathbf R^n are equivalent:

  1. w∈Trop⁡(X):=trop⁡(Xan)w\in\operatorname{Trop}(X):=\operatorname{trop}(X^{\mathrm{an}});

  2. Tw/ITw≠0\mathcal T_w/I\mathcal T_w\ne0;

  3. every nonzero f∈If\in I has at least two active terms at ww.

This holds with trivial valuation, arbitrary real radii, and nonreduced XX.

Proof. The equivalence of the first two conditions follows from Proposition 5.13.2 and Theorem 1.20.1. If ff has a unique active term amχma_m\chi^m, then f=amχm(1+h),∥h∥w<1.f=a_m\chi^m(1+h),\qquad \|h\|_w<1. Lemma 1.1.4 makes ff invertible in Tw\mathcal T_w, so the quotient is zero. Conversely, if that quotient is zero, then 1∈ITw1\in I\mathcal T_w. By the closure assertion in Proposition 5.13.2, there is g∈Ig\in I with ∥1−g∥w<1\|1-g\|_w<1. Its constant coefficient has absolute value one and every other term has weighted size less than one. Thus gg has a unique active term, contradicting the third condition. ◻

Proposition 5.13.4. NAG-002083 (Initial ideals and scalar extension). Put R=K∘R=K^\circ and k=K~k=\widetilde K. Let J=I∩R[M]J=I\cap R[M] and let J‾\overline J be its image in k[M]k[M]. For any complete isometric extension L/KL/K, with residue field ll, one has IL[M]∩L∘[M]‾=J‾ l[M].\overline{I L[M]\cap L^\circ[M]}=\overline J\,l[M]. Moreover, R[M]/JR[M]/J is flat over RR, its special fibre is k[M]/J‾k[M]/\overline J, and this special fibre is nonempty exactly when 0∈Trop⁡(X)0\in\operatorname{Trop}(X).

Proof. For a finite set F⊂MF\subset M, let VFV_F be its coordinate span and WF=I∩VFW_F=I\cap V_F. Every vector space over KK is torsion-free and therefore flat by [StacksAlg] (Tag 0539), applied to the valuation ring KK itself. Tensoring the kernel sequence WF→VF→K[M]/IW_F\to V_F\to K[M]/I thus gives (IL[M])∩(VF⊗KL)=WF⊗KL.(I L[M])\cap(V_F\otimes_KL)=W_F\otimes_KL. Apply Lemma 5.13.1 with unit weights. The reductions of the unit balls on the two sides therefore differ exactly by scalar extension from kk to ll. Taking the union over finite supports proves the ideal identity.

The closure algebra injects into K[M]/IK[M]/I, hence has no RR-torsion. It is flat by the valuation-ring criterion, [StacksAlg] (Tag 0539). Its special fibre has the stated presentation. If J‾\overline J contains a monomial, lifting the equality expressing this monomial as a finite sum of reduced relations gives an element of II whose reduction is that monomial. It has a unique active term. Conversely, normalize an element of II having a unique active term by its largest coefficient; its reduction is a unit monomial in J‾\overline J. Thus J‾\overline J is proper exactly when every relation has at least two active terms. Theorem 5.13.3 finishes the proof. ◻

Proposition 5.13.5. NAG-002084 (Completion retains the initial relations). Let I^=IT0\widehat I=I\mathcal T_0, let T0≤1\mathcal T_0^{\le 1} denote the Gauss unit ball, and reduce this unit ball coefficientwise to k[M]k[M]. Then red⁡(I^∩T0≤1)=J‾.\operatorname{red}(\widehat I\cap\mathcal T_0^{\le1})=\overline J.

Proof. One inclusion follows from I⊂I^I\subset\widehat I. Given h∈I^h\in\widehat I with ∥h∥0≤1\|h\|_0\le 1, approximate it by g∈Ig\in I with ∥g−h∥0<1\|g-h\|_0<1, using Proposition 5.13.2. Then ∥g∥0≤1\|g\|_0\le 1 and g,hg,h have identical reductions. This proves the reverse inclusion. Coefficientwise reduction is a Laurent polynomial because only finitely many coefficients of a convergent Laurent series can have absolute value one. ◻

Theorem 5.13.6. NAG-002085 (Initial degenerations at arbitrary real weights). Choose a complete isometric extension L/KL/K and t∈(L×)nt\in(L^\times)^n with v(t)=wv(t)=w. Such a choice exists for every real ww. Write ItI_t for the image of IL[M]I L[M] under χm↦tmχm\chi^m\mapsto t^m\chi^m, and set in⁡w,t,L(X)=Spec⁡(L~[M]/It∩L∘[M]‾).\operatorname{in}_{w,t,L}(X) =\operatorname{Spec}\left( \widetilde L[M]\big/\overline{I_t\cap L^\circ[M]}\right). After residue-field extension and translation by a point of the residue torus, this scheme is independent of L,tL,t. It is nonempty exactly when w∈Trop⁡(X)w\in\operatorname{Trop}(X). For every complete isometric extension E/KE/K, Trop⁡(XE)=Trop⁡(X).\operatorname{Trop}(X_E)=\operatorname{Trop}(X).

Proof. Extend the multiplicative weighted Gauss norm on K[M]K[M] to its fraction field and complete. The coordinate monomials give the required tt, with the original absolute value on KK. Multiplicativity is the Gauss calculation of Proposition 1.2.2. Theorem 2.35.5 places any two choices in a common complete valued extension. Proposition 5.13.4 compares the resulting initial ideals under that extension. For two translating points in the common field, the ratios ti′/tit'_i/t_i have absolute value one; their residues induce a torus translation between the two special fibres.

Theorem 2.33.13, applied to the fibre algebra of Proposition 5.13.2, shows that every tropical fibre is nonempty before a complete extension exactly when it is nonempty afterwards. This proves tropical base-change invariance. Translation by tt identifies the fibre at ww with the unit-radius fibre. Proposition 5.13.4 and Theorem 5.13.3 now give the assertion about nonemptiness. ◻

Proposition 5.13.7. NAG-002086 (Properness of toric tropicalization). The map trop⁡:Xan→Rn\operatorname{trop}:X^{\mathrm{an}}\to\mathbf R^n is continuous and topologically proper. Its image is closed.

Proof. Continuity follows from evaluation on the invertible characters. The inverse image of a closed box is the closed analytic subspace defined by II in the corresponding affinoid polyannulus; Proposition 2.24.1 gives its compactness. For a compact subset of Rn\mathbf R^n, choose a box containing it; its inverse image is closed in that compact polyannulus, hence compact. Closedness of the image also follows directly from Theorem 5.13.3: its complement is the union, over f∈If\in I, of the open loci where one term of ff is uniquely active. ◻

References. These arguments reconstruct the initial-degeneration and lifting mechanisms of [Gub13] (Sections 3–5). The analytic lifting step here is the closed-ideal approximation in the actual tropical-fibre algebra; the proof uses the previously established affinoid spectrum and base-change results.

5.14 A combinatorial calculus of corners

Let NN be a lattice and put V=N⊗ZRV=N\otimes_{\mathbf Z}\mathbf R. A rational polyhedron may have arbitrary real translation constants; its direction space is rational. All complexes in this section are finite. Write Nσ=N∩span⁡R(σ−σ)N_\sigma=N\cap\operatorname{span}_{\mathbf R}(\sigma-\sigma). A real weighted dd-cycle is a pure dd-dimensional rational polyhedral complex with constant real weights wσw_\sigma on its facets, satisfying ∑σ⊃τwσuσ/τ=0in V/Nτ,R.\sum_{\sigma\supset\tau}w_\sigma u_{\sigma/\tau}=0 \quad\text{in }V/N_{\tau,\mathbf R}. Here uσ/τu_{\sigma/\tau} is the primitive outward generator of the rank-one lattice Nσ/NτN_\sigma/N_\tau. Subdivision preserves a weight on each subdivided facet; cells of weight zero may be discarded. We identify cycles after a common weighted subdivision. A piecewise affine function on a cycle is continuous and affine on the cells of some finite rational subdivision. Its slopes may be real.

Lemma 5.14.1. NAG-002179 (Primitive vectors and saturated lattices). A primitive vector of Zn\mathbf Z^n extends to a lattice basis. A saturated subgroup A⊂ZnA\subset\mathbf Z^n is generated by part of a basis and has a free quotient. The group GLn(Z)\mathrm{GL}_n(\mathbf Z) is generated by row interchanges, sign changes, and adding an integer multiple of one row to another.

Proof. The Euclidean algorithm, applied successively to coordinate pairs, uses these invertible row operations to send a vector to (d,0,…,0)(d,0,\ldots,0), where dd is the greatest common divisor of its coordinates. For a primitive vector d=1d=1; applying the inverse row-operation matrix to the standard basis proves the first assertion. If A≠0A\ne 0 is saturated, divide a nonzero element by the greatest common divisor of its coordinates. Saturation puts the resulting primitive vector aa in AA. The image of AA in Zn/Za≃Zn−1\mathbf Z^n/\mathbf Za\simeq\mathbf Z^{n-1} is saturated: if a nonzero integer multiple of a class lies in that image, a representative has that multiple in AA, since Za⊂A\mathbf Za\subset A, and hence belongs to AA. Induction and lifting a quotient basis now give a basis whose first members generate AA.

For an invertible integral matrix, the greatest common divisor of its first column is one, since any common divisor would divide its determinant ±1\pm 1. Row operations reduce that column to the first standard vector. The lower-right block is again invertible over Z\mathbf Z; reduce it by induction, and then clear the remaining entries of the first row using the lower identity rows. This reduces the entire matrix to the identity and proves the generation assertion. ◻

Lemma 5.14.2. NAG-002180 (Refinement and products of balanced complexes). Finite rational polyhedral complexes admit common rational refinements. Balancing is invariant under refinement, and the Cartesian product of two balanced cycles, with product weights, is balanced. The star at a face and the quotient of that star by the direction space of the face are balanced.

Proof. Take all affine supporting hyperplanes of the finitely many faces, and cut every polyhedron by them. The resulting sign cells and their faces form a common refinement. At a new codimension-one face lying inside an old facet, there are two incident pieces with the same weight and opposite primitive normals. At a subdivision of an old codimension-one face, the old normal equation is unchanged. All other new faces have smaller dimension. These observations prove refinement invariance in both directions.

A codimension-one face of a product is either τ×ρ\tau\times\rho or σ×ν\sigma\times\nu, where exactly one factor has codimension one. Its normal equation is the normal equation in that factor multiplied by the weight of the other factor. For a star, the normals and weights are the original ones at the faces containing the chosen face. Its direction lattice is saturated, so passage to the quotient preserves the primitive normal quotient lattices and the balancing equations. ◻

Definition 5.14.3. (The corner operator). Let CC be a balanced dd-cycle and φ\varphi a piecewise affine function on it. On a subdivision on which φ\varphi is affine, let φσ\varphi_\sigma denote its linear part. At a codimension-one face τ\tau choose any real linear extension ℓτ:V→R\ell_\tau:V\to\mathbf R of φτ\varphi_\tau, and choose a lattice lift nσ/τn_{\sigma/\tau} of each primitive normal. The corner DφCD_\varphi C has weight wDφC(τ)=∑σ⊃τwσ(φσ−ℓτ)(nσ/τ).(C)w_{D_\varphi C}(\tau) =\sum_{\sigma\supset\tau}w_\sigma (\varphi_\sigma-\ell_\tau)(n_{\sigma/\tau}). \tag{C} For d=0d=0 the corner is zero.

Proposition 5.14.4. NAG-002183 (Well-defined corner weights). Formula ({C}) is independent of all normal lifts, linear extensions, and adapted subdivisions. It is linear in CC and in φ\varphi, and DℓC=0D_\ell C=0 for every globally affine function ℓ\ell. It commutes with integral affine lattice isomorphisms.

Proof. Changing a lift adds a vector in Nτ,RN_{\tau,\mathbf R}, on which φσ−ℓτ\varphi_\sigma-\ell_\tau vanishes by continuity. Changing ℓτ\ell_\tau by a linear form hh vanishing on that space changes the sum by −h(∑σwσnσ/τ)=0-h(\sum_\sigma w_\sigma n_{\sigma/\tau})=0, by balancing. At a new face inside an old affine piece, the two contributions cancel; in fact the same linear extension can be used on both sides. At a subdivided old face, the formula has the old value. Lemma 5.14.2 now compares any two adapted subdivisions. The formula gives both linearities. For a globally affine function choose its common linear part as every ℓτ\ell_\tau; all summands vanish. A lattice isomorphism transports the saturated lattices, their primitive generators, and this same formula, proving the last assertion. ◻

Theorem 5.14.5. NAG-002184 (Balancing and commutation of corners). For every balanced real cycle CC and continuous piecewise affine functions φ,ψ\varphi,\psi on it, DφCD_\varphi C is balanced and DψDφC=DφDψC.D_\psi D_\varphi C=D_\varphi D_\psi C.

Proof. Both assertions are local at an original face of codimension two. Take its star and quotient by its direction space. Subtract linear extensions of the two functions along that face. By Proposition 5.14.4, this changes neither corner. We have reduced to a two-dimensional balanced rational fan, with both functions homogeneous and linear on its cones. Cut cones with lineality by rational hyperplanes if necessary, so that the cones are pointed; their two-dimensional cones then have two extreme rays. Let the distinct primitive ray vectors be v1,…,vsv_1,\ldots,v_s. For a cone with extreme rays vi,vjv_i,v_j put dij=[N⟨vi,vj⟩:Zvi+Zvj],aij=aji=wij/dij;d_{ij}=[N_{\langle v_i,v_j\rangle}:\mathbf Zv_i+\mathbf Zv_j], \qquad a_{ij}=a_{ji}=w_{ij}/d_{ij}; put aij=0a_{ij}=0 if there is no such cone. Modulo Rvi\mathbf Rv_i, its primitive normal is the class of vj/dijv_j/d_{ij}. Indeed viv_i is primitive in the rank-two saturated lattice; Lemma 5.14.1 completes it to a basis, in which vj=bvi+dijev_j=bv_i+d_{ij}e with the appropriate sign of ee. Balancing at viv_i therefore says ∑jaijvj=λivifor some λi∈R.\sum_j a_{ij}v_j=\lambda_i v_i \quad\text{for some }\lambda_i\in\mathbf R. The weight of the first corner on the ray ii is consequently bi(φ)=∑jaijφ(vj)−λiφ(vi).(R)b_i(\varphi)=\sum_j a_{ij}\varphi(v_j)-\lambda_i\varphi(v_i). \tag{R} One may use vj/dijv_j/d_{ij} instead of a lattice lift in ({C}), since their difference is tangential. Symmetry of aija_{ij} and the displayed balancing equations give ∑ibi(φ)vi=∑jφ(vj)∑iaijvi−∑iλiφ(vi)vi=0.\sum_i b_i(\varphi)v_i =\sum_j\varphi(v_j)\sum_i a_{ij}v_i -\sum_i\lambda_i\varphi(v_i)v_i=0. This is balancing for the corner at the origin. The coefficient of the second corner at that origin is ∑ibi(φ)ψ(vi)=∑i,jaijφ(vj)ψ(vi)−∑iλiφ(vi)ψ(vi),\sum_i b_i(\varphi)\psi(v_i) =\sum_{i,j}a_{ij}\varphi(v_j)\psi(v_i) -\sum_i\lambda_i\varphi(v_i)\psi(v_i), which is symmetric in φ,ψ\varphi,\psi. Hence the two orders agree. For a cycle of dimension one, its corner is a zero-cycle, whose further corner is zero; in dimension zero both assertions are immediate. The local calculation and refinement invariance prove the theorem in every dimension. ◻

Lemma 5.14.6. NAG-002186 (Restriction, products, and quotient directions). The corner depends only on the restricted function on the support of its cycle. If pp is projection from a product, then Dφ∘p(C×E)=(DφC)×E.D_{\varphi\circ p}(C\times E)=(D_\varphi C)\times E. Taking a star commutes with the corner after replacing the function by its local piecewise linear part. If a fan has a rational lineality space LL, and the function is pulled back from its quotient by LL, the corner is the pullback of the corner on that quotient, with the same weights.

Proof. All terms of ({C}) are determined by the restricted affine functions. A codimension-one face of C×EC\times E belonging to the EE factor has zero weight for this corner, since the function has a common affine extension there. At a face belonging to the CC factor, the formula is the corner weight in CC multiplied by the weight in EE. Stars leave the normal quotient lattices unchanged. For the last assertion, N∩LN\cap L is saturated; Lemma 5.14.1 gives a lattice splitting of NN over N/(N∩L)N/(N\cap L). The fan is then a product of LL with its quotient, so the product calculation applies. ◻

References. The corner construction follows Allermann–Rau, First steps in tropical intersection theory, Section 3. The proof above gives the full rank-two calculation, including its tangential correction, before using any differential forms or toric Chow theory.

5.15 The standard-chart homotopy

Let KK be complete and nontrivially valued, with its original absolute value. Fix 0<∣a∣<10<|a|<1 and integers r,q≥0r,q\geq 0. Put A=K⟨X0,…,Xr,T1,…,Tq⟩/(X0⋯Xr−a),V=M(A).A=K\langle X_0,\ldots,X_r,T_1,\ldots,T_q\rangle /(X_0\cdots X_r-a),\qquad V=\mathcal M(A). Proposition 3.26.1 gives the continuous Gauss section s:Δa→Vs:\Delta_a\to V, where Δa={(w0,…,wr)∈R≥0r+1:∑iwi=v(a)}.\Delta_a=\{(w_0,\ldots,w_r)\in\mathbb R_{\geq0}^{r+1}: \sum_iw_i=v(a)\}. After eliminating X0=a/(X1⋯Xr)X_0=a/(X_1\cdots X_r), s(w)s(w) is the Gauss point of radii e−wie^{-w_i} in XiX_i, 1≤i≤r1\leq i\leq r, and radius 11 in every TjT_j. Let Σa=s(Δa)\Sigma_a=s(\Delta_a) and write val⁡:V→Δa\operatorname{val}:V\to\Delta_a for the values of X0,…,XrX_0,\ldots,X_r.

Lemma 5.15.1. NAG-002139 (Mixed Gauss displacement). For x∈Vx\in V, write E=H(x)E=\mathcal H(x), ξi=Xi(x)∈E×\xi_i=X_i(x)\in E^\times and bj=Tj(x)∈Eb_j=T_j(x)\in E. For 0<t≤10<t\leq 1, take the Gauss absolute value of radius tt in all variables S1,…,Sr,Z1,…,ZqS_1,\ldots,S_r,Z_1,\ldots,Z_q over EE, and extend it to the completed rational function field EtE_t. The substitutions Xi⟼ξi(1+Si)(1≤i≤r),X0⟼ξ0∏i=1r(1+Si)−1,Tj⟼bj+Zj(1≤j≤q).\begin{aligned} X_i&\longmapsto\xi_i(1+S_i) &&(1\leq i\leq r),\\ X_0&\longmapsto\xi_0\prod_{i=1}^r(1+S_i)^{-1},\\ T_j&\longmapsto b_j+Z_j &&(1\leq j\leq q). \end{aligned} define a bounded homomorphism A→EtA\to E_t and hence a point H(x,t)∈VH(x,t)\in V. Set H(x,0)=xH(x,0)=x. For every t∈[0,1]t\in[0,1], ∣Xi(H(x,t))∣=∣Xi(x)∣(0≤i≤r),∣Tj(H(x,t))∣=max⁡{∣Tj(x)∣,t}.|X_i(H(x,t))|=|X_i(x)|\quad(0\leq i\leq r),\qquad |T_j(H(x,t))|=\max\{|T_j(x)|,t\}.

Proof. The Gauss absolute value is multiplicative by Proposition 1.2.2. For t>0t>0 the polynomial 1+Si1+S_i has norm 11, so it is nonzero and invertible in the rational function field, also when t=1t=1. All displayed coordinate images have norm at most one, and their product in the first r+1r+1 coordinates is aa. Convergent restricted series can therefore be evaluated in the complete field EtE_t. This yields a homomorphism from the Tate algebra, contractive for its Gauss norm, which kills the defining relation. It descends contractively for the quotient norm on AA. Multiplicativity and the exact Gauss norms of 1+Si1+S_i and bj+Zjb_j+Z_j give the stated coordinate values. At t=0t=0 they are the values at xx, as required. ◻

Theorem 5.15.2. NAG-002140 (A jointly continuous strong deformation retraction). The map H:V×[0,1]→VH:V\times[0,1]\to V of Lemma 5.15.1 is jointly continuous and is a strong deformation retraction onto Σa\Sigma_a. More precisely, H(x,0)=x,H(x,1)=s(val⁡(x)),H(s(w),t)=s(w),val⁡(H(x,t))=val⁡(x).\begin{aligned} H(x,0)&=x,& H(x,1)&=s(\operatorname{val}(x)),\\ H(s(w),t)&=s(w),& \operatorname{val}(H(x,t))&=\operatorname{val}(x). \end{aligned} It commutes with complete isometric ground-field extension.

Proof. We first give a formula proving joint continuity, including at t=0,1t=0,1. Let P∈K[X1±1,…,Xr±1,T]P\in K[X_1^{\pm 1},\ldots,X_r^{\pm 1},T] and choose D∈NrD\in\mathbb N^r so that Q=XDPQ=X^DP is a polynomial. Expand the finite polynomial Q(X1(1+S1),…,Xr(1+Sr),T+Z)=∑β,γQβγ(X,T)SβZγ.Q\bigl(X_1(1+S_1),\ldots,X_r(1+S_r),T+Z\bigr) =\sum_{\beta,\gamma}Q_{\beta\gamma}(X,T)S^\beta Z^\gamma. Since the factors 1+Si1+S_i have norm one, the exact formula is ∣P(H(x,t))∣=max⁡β,γ∣Qβγ(x)∣ t∣β∣+∣γ∣∏i=1r∣Xi(x)∣Di,(H)|P(H(x,t))| =\frac{\displaystyle\max_{\beta,\gamma} |Q_{\beta\gamma}(x)|\,t^{|\beta|+|\gamma|}} {\displaystyle\prod_{i=1}^r|X_i(x)|^{D_i}}, \tag{H} where the term of degree zero uses 00=10^0=1. Its right side is continuous in (x,t)(x,t): it is a finite maximum divided by a continuous positive function. Every XiX_i is invertible on VV. The Laurent polynomial algebra is dense in AA by Proposition 3.26.1. All evaluations H(x,t)H(x,t) are contractive for one fixed quotient norm, so approximating an arbitrary f∈Af\in A by Laurent polynomials gives uniform approximation of the real functions ∣f(H(x,t))∣|f(H(x,t))| on V×[0,1]V\times[0,1]. Thus all such evaluation functions are continuous, which is precisely continuity into the Berkovich spectrum.

At t=1t=1, translation of any variable of the unit Gauss norm by a scalar of norm at most one is isometric. Indeed, the ultrametric inequality makes substitution contractive on polynomials, and the inverse translation gives the reverse bound; the assertion then extends to fractions. Apply this to Si↦1+SiS_i\mapsto 1+S_i and Zj↦bj+ZjZ_j\mapsto b_j+Z_j. For every Laurent polynomial P=∑cαμXαTμP=\sum c_{\alpha\mu}X^\alpha T^\mu, ∣P(H(x,1))∣=max⁡α,μ∣cαμ∣∏i=1r∣Xi(x)∣αi.|P(H(x,1))|=\max_{\alpha,\mu} |c_{\alpha\mu}|\prod_{i=1}^r|X_i(x)|^{\alpha_i}. This is the defining Gauss norm of s(val⁡(x))s(\operatorname{val}(x)), including radius one for every free coordinate TjT_j. Density gives the same equality on AA.

Let x=s(w)x=s(w). The norm of every monomial after displacement is at most its original Gauss norm, so ∣P(H(x,t))∣≤∣P(x)∣|P(H(x,t))|\leq|P(x)|. In ({H}), the constant coefficient is Q(x)Q(x) and has norm equal to the Gauss norm of QQ at s(w)s(w). Consequently the same formula gives the reverse inequality. Thus HH fixes every skeletal point. The identity at time zero is its definition, and preservation of val⁡\operatorname{val} is the coordinate formula in the lemma. These are all strong deformation retraction identities.

For a complete isometric extension L/KL/K and a point x′∈VLx'\in V_L over xx, the completed residue-field embedding H(x)↪H(x′)\mathcal H(x)\hookrightarrow\mathcal H(x') is isometric. Formula ({H}) therefore gives the same value on every original Laurent polynomial before and after projection. Density proves commutation on AA. At the endpoint the standard Gauss section over LL restricts to that over KK, with unchanged real radii. ◻

Proposition 5.15.3. NAG-002142 (The maximum-parameter composition law). For all x∈Vx\in V and s,t∈[0,1]s,t\in[0,1], H(H(x,s),t)=H(x,max⁡(s,t)).H(H(x,s),t)=H(x,\max(s,t)).

Proof. When one parameter is zero the assertion is immediate. For positive parameters, realize the completed residue field of H(x,s)H(x,s) inside the completed Gauss rational field used in its definition. The injection comes from its actual evaluation homomorphism, after quotienting its kernel, taking fractions, and completing. Gauss extension preserves the norm of every coefficient. Iterating the construction therefore uses independent variables (S,Z)(S,Z) of radius ss and (S′,Z′)(S',Z') of radius tt, and substitutes Xi↦ξi(1+Si)(1+Si′),Tj↦bj+Zj+Zj′.X_i\mapsto\xi_i(1+S_i)(1+S'_i),\qquad T_j\mapsto b_j+Z_j+Z'_j. The two sets of variables are symmetric, so assume s≤ts\leq t. Set Wi=Si+Si′+SiSi′,Vj=Zj+Zj′.W_i=S_i+S'_i+S_iS'_i,\qquad V_j=Z_j+Z'_j. The rational coordinate change from (S,Z,S′,Z′)(S,Z,S',Z') to (S,Z,W,V)(S,Z,W,V) is an isometry of Gauss fields with radii (s,s,t,t)(s,s,t,t). To see this directly, its inverse formulas are Si′=(Wi−Si)/(1+Si),Zj′=Vj−Zj.S'_i=(W_i-S_i)/(1+S_i),\qquad Z'_j=V_j-Z_j. The displayed images have their assigned norm bounds, because s≤ts\leq t and ∣1+Si∣=1|1+S_i|=1, including when s=t=1s=t=1. Hence substitution of an old polynomial has norm at most its original Gauss norm. Its image is a new polynomial divided by a product of powers of 1+Si1+S_i, of norm one. Applying the reverse substitution to that numerator gives the reverse bound. This proves equality for polynomials and then for their fractions. The same argument for V=Z+Z′V=Z+Z' is linear and has no denominator.

After this isometry the iterated substitution is simply Xi↦ξi(1+Wi)X_i\mapsto\xi_i(1+W_i) and Tj↦bj+VjT_j\mapsto b_j+V_j. It does not involve (S,Z)(S,Z); its Gauss norm is the single displacement norm of radius tt. This proves the identity on Laurent polynomials, and uniform approximation proves it on all of AA. ◻

Proposition 5.15.4. NAG-002143 (Monomial and bounded affine covariance). On a torus times a unit polydisc, the mixed displacement construction is unchanged by a torus coordinate change Yi=ciXmiY_i=c_iX^{m_i} whose exponent matrix lies in GLr(Z)\mathrm{GL}_r(\mathbb Z), together with a free-coordinate change U=BT+bU=BT+b with B,B−1B,B^{-1} having entries in K∘K^\circ and b∈(K∘)qb\in(K^\circ)^q. Whenever such a change identifies two standard chart domains, it intertwines their entire homotopies. In particular the construction is independent of which monomial coordinate is eliminated from X0⋯Xr=aX_0\cdots X_r=a.

Proof. For 0<t<10<t<1, the normalized relative displacement of YiY_i is Wi=∏j(1+Sj)mij−1.W_i=\prod_j(1+S_j)^{m_{ij}}-1. Integer binomial coefficients have norm at most one, including those for negative exponents. The binomial series therefore converge for radius tt, and ∣Wi∣≤t|W_i|\leq t. The inverse integral matrix gives the inverse substitutions with the same bounds. The two maps of radius-tt Tate algebras are contractive inverses, hence isometric. At t=1t=1 use instead the variables Ri=1+SiR_i=1+S_i. Their radius-one Gauss norm is unchanged by translation. A unimodular monomial change permutes the Laurent monomial basis, so preserves the Gauss norm and its extension to fractions. At time zero both constructions are evaluation at the original point.

For the free variables, the relative displacement of BT+bBT+b is BZBZ. Both BB and its inverse are contractive for the radius-tt coordinate maximum norm, so their induced Gauss substitutions are isometric. The constant bb merely changes the center. These isometries identify all polynomial evaluations defining the homotopies; density identifies the analytic evaluations. Finally, changing the eliminated coordinate expresses the new independent monomial variables in the old ones by an integral invertible exponent matrix and powers of aa, precisely the torus changes just treated. ◻

References. The construction uses explicit Gauss displacement on the standard semistable chart. At the endpoint, evaluation in a rational Gauss field retains the norm-one denominators and their multiplicative inverses.

5.16 Finite polyhedral initial data

Write S=K[x0,…,xn]S=K[x_0,\ldots,x_n] with its usual grading, and let H⊂SH\subset S be homogeneous. For ω∈Rn+1\omega\in\mathbf R^{n+1}, pass to a complete isometric extension realizing the coordinate weights, translate the variables, and reduce normalized homogeneous relations. This gives a homogeneous ideal in⁡ω(H)\operatorname{in}_\omega(H) in a polynomial ring over the residue field. The arguments of Proposition 5.13.4 apply degree by degree to polynomial rings as well. We compare ideals after residue-field extension and diagonal change of coordinates.

Lemma 5.16.1. NAG-002088 (The Hilbert function of an initial ideal). For every degree dd, the degree-dd component of in⁡ω(H)\operatorname{in}_\omega(H) is the reduction of the unit ball of the translated subspace HdH_d. Consequently dim⁡L~(L~[x]/in⁡ω(H))d=dim⁡K(S/H)d.\dim_{\widetilde L}\bigl(\widetilde L[x]/\operatorname{in}_\omega(H)\bigr)_d =\dim_K(S/H)_d.

Proof. The reduction of the unit ball of HdH_d is a vector subspace. Multiplying a normalized homogeneous relation of degree e≤de\le d by a monomial of degree d−ed-e gives a normalized relation in HdH_d. Conversely, every reduced vector in HdH_d is one of the homogeneous initial relations defining the ideal. Thus its degree-dd part is exactly this reduced subspace. After translating, Lemma 5.13.1 with unit weights gives its dimension dim⁡KHd\dim_KH_d. Subtract this from the number of degree-dd monomials. ◻

Lemma 5.16.2. NAG-002089 (Finitely many monomial ideals with a fixed Hilbert function). In a polynomial ring in finitely many variables, there are only finitely many monomial ideals with a prescribed Hilbert function. This finite set depends only on that function and the number of variables, not on the coefficient field.

Proof. Suppose there were infinitely many distinct such ideals. Enumerate the monomials by nondecreasing degree. Successively retain an infinite subfamily on which membership of the next monomial is constant. The limiting membership choices define a monomial ideal QQ: divisibility preserves membership because it does so in every ideal of every sufficiently late subfamily. In each degree its standard monomials have the prescribed cardinality. The polynomial ring is Noetherian by [StacksAlg] (Tag 00FN), so QQ has finitely many monomial generators, all of degree at most some DD.

An infinite subfamily agrees with QQ on all monomials of degree at most DD. Each member therefore contains QQ. Since its standard-monomial count equals that of QQ in every degree, the containment is equality. This contradicts the distinctness of the infinite subfamily. All choices involved only sets of exponent vectors. ◻

Proposition 5.16.3. NAG-002090 (A uniform degree bound for homogeneous initial ideals). There is an integer DD depending only on the Hilbert function of S/HS/H such that every in⁡ω(H)\operatorname{in}_\omega(H), over every residue extension, is generated in degrees at most DD. The same DD works for all further ordinary initial ideals of these ideals.

Proof. Fix a degree-compatible monomial order. For a homogeneous ideal JJ over a field, let LM⁡(J)\operatorname{LM}(J) be the monomial ideal generated by its leading monomials. Its standard monomials form a basis of the quotient in every degree. Indeed, finitely many leading monomials generate LM⁡(J)\operatorname{LM}(J) by Noetherianity. Choose homogeneous elements of JJ having these leading monomials. Polynomial division by them terminates because the monomial order is well ordered and reduces every homogeneous polynomial to a linear combination of standard monomials. Such a combination cannot belong to JJ unless it is zero, since otherwise its leading monomial would belong to LM⁡(J)\operatorname{LM}(J). The same division proves that the chosen elements generate JJ.

By Lemma 5.16.1, all ideals under consideration have the Hilbert function of HH. Their monomial leading ideals therefore belong to the finite list of Lemma 5.16.2. Take DD to be the largest degree among the minimal monomial generators in this list. Lifting these generators as in the preceding paragraph proves the bound. An ordinary weight initial ideal has the same Hilbert function, by the same degreewise argument over a trivially valued field, so the bound applies to it as well. ◻

Lemma 5.16.4. NAG-002091 (Plücker coordinates and weighted row reduction). Let W⊂KFW\subset K^F have dimension ss, where FF is a finite set of exponent vectors, and let CC be a full-rank row matrix for WW. For an ss-element column set BB with det⁡CB≠0\det C_B\ne 0, put pB(ω)=v(det⁡CB)+∑b∈B⟨b,ω⟩.p_B(\omega)=v(\det C_B)+\sum_{b\in B}\langle b,\omega\rangle. If BB minimizes pB(ω)p_B(\omega), the rows gbg_b of CB−1CC_B^{-1}C, indexed by b∈Bb\in B, satisfy (gb)b′=δbb′,v((gb)m)+⟨m,ω⟩≥⟨b,ω⟩.(g_b)_{b'}=\delta_{bb'},\qquad v((g_b)_m)+\langle m,\omega\rangle\ge\langle b,\omega\rangle. Their initial forms form a basis of the initial subspace. Which coefficients survive is determined by equalities among the finitely many functions pBp_B.

Proof. The pivot identity follows from the matrix inverse. Cramer’s rule expresses (gb)m(g_b)_m, up to sign, as the ratio of the minor obtained by replacing bb by mm to det⁡CB\det C_B. Minimality gives the displayed inequality. The resulting rows have distinct pivots and weighted norms e−⟨b,ω⟩e^{-\langle b,\omega\rangle}; the pivot-coordinate argument of Lemma 5.13.1 applies. After translating and dividing the bb-th row by the bb-th translating monomial, its pivot is one. Reduction therefore gives independent spanning rows. A nonpivot coefficient survives exactly when the corresponding minor has the same value of pp as BB. ◻

Theorem 5.16.5. NAG-002092 (Finite polyhedral strata for initial degenerations). There is a finite polyhedral complex covering Rn+1\mathbf R^{n+1}, defined by integer linear forms and constants in v(K×)v(K^\times), such that on the relative interior of each cell the schemes defined by in⁡ω(H)\operatorname{in}_\omega(H) are isomorphic after residue extension and diagonal torus translation. For an ideal I⊂K[M]I\subset K[M], there is consequently a finite integral v(K×)v(K^\times)-affine polyhedral complex on Rn\mathbf R^n for which Trop⁡(Spec⁡K[M]/I)\operatorname{Trop}(\operatorname{Spec}K[M]/I) is the support of a subcomplex.

Proof. Take DD from Proposition 5.16.3. For each d≤Dd\le D apply Lemma 5.16.4 to HdH_d. Cut weight space by all equalities between the finitely many affine functions pBp_B occurring in these degrees. Their normals are integral and their constant differences belong to v(K×)v(K^\times). The resulting finite hyperplane arrangement is the required polyhedral complex.

Fix weights ω,ω′\omega,\omega' in the same relative interior. In each degree choose one common minimizing pivot set BB; the same coefficients in its row basis survive at both weights. Realize both weights in a common valued extension, with translating points t,t′t,t'. For every surviving exponent difference m−bm-b, the unit ratio (t′/t)m−b(t'/t)^{m-b} has norm one. On the subgroup A⊂Zn+1A\subset\mathbf Z^{n+1} generated by all these differences, their residues define a character ρ:A→L~×\rho:A\to\widetilde L^\times. This character extends to the full exponent lattice after a finite algebraic residue extension. To see this directly, adjoin the standard basis vectors successively. When adjoining eie_i, if no positive multiple of eie_i belongs to the current subgroup, give eie_i any nonzero value. Otherwise let q>0q>0 be least with qeiqe_i in that subgroup, adjoin a root bb of Tq−ρ(qei)T^q-\rho(qe_i), and set ρ(a+mei)=ρ(a)bm\rho(a+me_i)=\rho(a)b^m. This is well defined and extends ρ\rho. A root is obtained by quotienting by an irreducible factor of that polynomial. After finitely many steps, translation by the resulting torus point sends every surviving coefficient in every row at ω\omega to its coefficient at ω′\omega'. The ideals are generated by these rows in degrees at most DD, so the translation identifies the ideals.

For the Laurent ideal, take the homogeneous kernel H=ker⁡(K[x0,…,xn]⟶(K[M]/I)[z,z−1]),x0↦z,xi↦zχei.H=\ker\bigl(K[x_0,\ldots,x_n]\longrightarrow (K[M]/I)[z,z^{-1}]\bigr), \qquad x_0\mapsto z,\quad x_i\mapsto z\chi^{e_i}. Localizing all the variables and putting x0=1x_0=1 recovers II. Initial relations commute with this operation: multiplication by a monomial shifts every term valuation by the same amount, and any Laurent relation can first be multiplied by a monomial and homogenized. Restrict the constructed complex to ω0=0\omega_0=0. Nonemptiness of the resulting initial scheme is constant on every relative interior. By Theorem 5.13.6 this is exactly membership in Trop⁡(X)\operatorname{Trop}(X). That set is closed by Proposition 5.13.7; hence it contains the closure of every selected relative interior. It is therefore a union of closed cells, together with all their faces. ◻

Theorem 5.16.6. NAG-002093 (Initial degenerations describe local stars). Let X⊂Spec⁡K[M]X\subset\operatorname{Spec}K[M] be a closed subscheme and w∈Trop⁡(X)w\in\operatorname{Trop}(X). Give the residue field of an initial degeneration Y=in⁡w(X)Y=\operatorname{in}_w(X) the trivial absolute value. Then Trop⁡(Y)=Star⁡wTrop⁡(X).\operatorname{Trop}(Y)=\operatorname{Star}_w\operatorname{Trop}(X). Here the right side consists of those uu for which w+εu∈Trop⁡(X)w+\varepsilon u\in\operatorname{Trop}(X) for all sufficiently small positive ε\varepsilon. More precisely, for every fixed uu, the schemes in⁡w+εu(X)\operatorname{in}_{w+\varepsilon u}(X) and in⁡u(Y)\operatorname{in}_u(Y) are equivalent after residue extension and translation for all sufficiently small positive ε\varepsilon.

Proof. Homogenize as in Theorem 5.16.5, and write ω=(0,w)\omega=(0,w) and υ=(0,u)\upsilon=(0,u). Choose the common degree bound DD of Proposition 5.16.3, also valid for the further initial ideals. In every degree at most DD, minimize the pair (pB(ω), ∑b∈B⟨b,υ⟩).\left(p_B(\omega),\ \sum_{b\in B}\langle b,\upsilon\rangle\right). There are finitely many pairs. Thus a lexicographically minimizing basis is minimizing for pB(ω+ευ)p_B(\omega+\varepsilon\upsilon) whenever ε>0\varepsilon>0 is sufficiently small. Cramer’s rule then says that its surviving row entries are exactly those which first survive at ω\omega and then have minimal υ\upsilon-weight. After the first reduction these rows are a basis of the degree piece of in⁡ω(H)\operatorname{in}_\omega(H); their second reduction is therefore its degree piece at υ\upsilon, by Lemma 5.16.4.

We spell out the coefficient comparison. Fix a complete field LL and tt of value ω\omega. Adjoin Gauss variables qiq_i of values ευi\varepsilon\upsilon_i to LL and complete their fraction field, obtaining EE. Use tqtq for the one-step translation. For a row entry cbmc_{bm} put a=m−ba=m-b. An entry survives both operations precisely when v(cbm)+⟨a,ω⟩=0,⟨a,υ⟩=0.v(c_{bm})+\langle a,\omega\rangle=0, \qquad \langle a,\upsilon\rangle=0. Its coefficient after the first reduction is cbmta‾∈L~\overline{c_{bm}t^a}\in\widetilde L. The ordinary weight initial form over the trivially valued field retains this coefficient. The one-step coefficient in E~\widetilde E is cbmta‾ qa‾\overline{c_{bm}t^a}\,\overline{q^a}. Thus the discrepancy is exactly the character a↦qa‾a\mapsto\overline{q^a} on the subgroup generated by the surviving exponent differences. Extend this character to the whole exponent lattice by the finite root-adjunction argument in the preceding theorem. A single torus translation then compares all the row generators.

For completeness, the ordinary initial ideal obtained by retaining the terms of least υ\upsilon-weight agrees, up to this same kind of translation, with the valued-extension definition of in⁡υ\operatorname{in}_{\upsilon} over a trivially valued field. Indeed, introduce Gauss variables sis_i of values υi\upsilon_i and divide each selected row by its pivot monomial. For a retained exponent difference aa one has ⟨a,υ⟩=0\langle a,\upsilon\rangle=0, so the only new coefficient factor is the character a↦sa‾a\mapsto\overline{s^a}. This also extends after algebraic residue extension. The common degree bound identifies the full homogeneous ideals from the compared rows. Localizing and dehomogenizing gives the claimed scheme equivalence. The nonemptiness criterion of Theorem 5.13.6, over both fields, gives the equality of supports. ◻

References. The classical finite Gröbner-complex and local-star theorems are discussed in [Gub13] (Sections 10 and 12–13). The proof here uses finite-dimensional weighted elimination, a Hilbert-function degree bound, and the analytic fundamental theorem already proved above.

5.17 Pushforwards and the projection formula

For an affine map, the same letter denotes its linear part when applied to direction vectors or direction lattices.

Lemma 5.17.1. NAG-002187 (The index along a facet). Let f:N→N′f:N\to N' be a lattice homomorphism, and let τ≺σ\tau\prec\sigma be rational polyhedra of dimensions d−1,dd-1,d. Suppose ff is injective on Nσ,RN_{\sigma,\mathbf R} and that τ′=f(τ)\tau'=f(\tau) is a facet of σ′=f(σ)\sigma'=f(\sigma). Put aσ=[Nσ′′:f(Nσ)],aτ=[Nτ′′:f(Nτ)].a_\sigma=[N'_{\sigma'}:f(N_\sigma)],\qquad a_\tau=[N'_{\tau'}:f(N_\tau)]. There is a positive integer mσ/τm_{\sigma/\tau} such that f(uσ/τ)=mσ/τuσ′/τ′ mod Nτ′′,aσ=aτmσ/τ.f(u_{\sigma/\tau})=m_{\sigma/\tau}u_{\sigma'/\tau'} \quad\bmod N'_{\tau'},\qquad a_\sigma=a_\tau m_{\sigma/\tau}.

Proof. Injectivity and saturation give f(Nσ)∩Nτ′′=f(Nτ)f(N_\sigma)\cap N'_{\tau'}=f(N_\tau): the inverse image of Nτ′,R′N'_{\tau',\mathbf R} inside Nσ,RN_{\sigma,\mathbf R} is Nτ,RN_{\tau,\mathbf R}. The image of the primitive outward generator in the rank-one quotient is a positive integer mσ/τm_{\sigma/\tau} times the target generator. The finite groups fit into the exact sequence 0⟶Nτ′′/f(Nτ)⟶Nσ′′/f(Nσ)⟶Nσ′′/(Nτ′′+f(Nσ))⟶0.0\longrightarrow N'_{\tau'}/f(N_\tau) \longrightarrow N'_{\sigma'}/f(N_\sigma) \longrightarrow N'_{\sigma'}/(N'_{\tau'}+f(N_\sigma)) \longrightarrow0. The final group has order mσ/τm_{\sigma/\tau}, so taking orders proves the equality. ◻

Theorem 5.17.2. NAG-002188 (Balanced pushforward). Let CC be a finite balanced real dd-cycle in NRN_{\mathbf R} and let f:NR→NR′f:N_{\mathbf R}\to N'_{\mathbf R} be an integral affine map. There is a well-defined balanced dd-cycle f∗Cf_*C. After subdivision so that images are cells, its weight on a dd-dimensional target cell σ′\sigma' is wf∗C(σ′)=∑dim⁡σ=d, f(σ)=σ′wC(σ)[Nσ′′:f(Nσ)].(P)w_{f_*C}(\sigma')= \sum_{\substack{\dim\sigma=d\,,\ f(\sigma)=\sigma'}} w_C(\sigma)[N'_{\sigma'}:f(N_\sigma)]. \tag{P} Facets on which the dimension drops contribute zero. Pushforward is linear and functorial under composition of integral affine maps.

Proof. Images of rational polyhedra are rational polyhedra: eliminate one coordinate at a time from their finite systems of affine inequalities, pairing a lower and an upper bound for that coordinate. Equalities are eliminated by substitution. This finite elimination also permits real constants. Choose a common target subdivision of all images and pull its defining hyperplanes back to the source. On every facet where ff is injective, source cells correspond bijectively to the target cells lying in its image. Formula ({P}) therefore defines a finite weighted complex. Adding more hyperplanes merely subdivides it with the same generic weights; two choices are compared by their combined hyperplanes. This also proves invariance under a source subdivision.

Fix a target face τ′\tau' of dimension d−1d-1 and sum over all source faces τ\tau of dimension d−1d-1 with f(τ)=τ′f(\tau)=\tau'. Multiply the balancing equation at τ\tau, after applying the linear part of ff, by aτ=[Nτ′′:f(Nτ)]a_\tau=[N'_{\tau'}:f(N_\tau)]. For an incident full-rank facet σ\sigma, Lemma 5.17.1 replaces its term by wC(σ)aσuf(σ)/τ′w_C(\sigma)a_\sigma u_{f(\sigma)/\tau'}. If ff has rank less than dd on σ\sigma, its normal maps into Nτ′,R′N'_{\tau',\mathbf R} and gives zero in the normal quotient. Each full-rank σ\sigma with f(σ)⊃τ′f(\sigma)\supset\tau' occurs exactly once: the inverse image of the target facet is its unique facet mapping to τ′\tau'. The sum is consequently the balancing equation for the weights ({P}). This proves balancedness, including faces bordering a zero-weight region.

Linearity follows from the sum. For a second map gg, a facet contributing to (g∘f)∗C(g\circ f)_*C has full rank at both stages, while a dimension drop cannot be reversed. On a contributing facet the two finite indices multiply, since gg is injective on the relevant direction space. After a common subdivision, summing these products proves (g∘f)∗=g∗f∗(g\circ f)_*=g_*f_*. ◻

Theorem 5.17.3. NAG-002190 (The corner projection formula). Let ff and CC be as in Theorem 5.17.2, and let φ\varphi be a continuous piecewise affine function on the target. Then f∗(Dφ∘fC)=Dφ(f∗C).f_*(D_{\varphi\circ f}C)=D_\varphi(f_*C).

Proof. Refine the complexes so that all images are cells and φ\varphi is affine on each image. For a target (d−1)(d-1)-cell τ′\tau', extend the linear part of φ∣τ′\varphi|_{\tau'} to a linear form ℓ\ell on the target. On any source face τ\tau mapping onto τ′\tau', use ℓ∘f\ell\circ f in the corner formula. For an incident facet σ\sigma of full image dimension, the difference φf(σ)−ℓ\varphi_{f(\sigma)}-\ell vanishes tangentially to τ′\tau'. Lemma 5.17.1 therefore gives aτ(φf(σ)−ℓ)(f(nσ/τ))=aσ(φf(σ)−ℓ)(nf(σ)/τ′).a_\tau\bigl(\varphi_{f(\sigma)}-\ell\bigr) (f(n_{\sigma/\tau})) =a_\sigma\bigl(\varphi_{f(\sigma)}-\ell\bigr) (n_{f(\sigma)/\tau'}). A facet collapsed to dimension d−1d-1 has f(Nσ,R)=Nτ′,R′f(N_{\sigma,\mathbf R})=N'_{\tau',\mathbf R}, so its term vanishes. A lower-rank facet cannot contain such a τ\tau. Multiply by wC(σ)w_C(\sigma) and sum over τ\tau and its incident facets. As in the pushforward proof, every full-rank facet over a target facet containing τ′\tau' occurs once. The resulting sum is exactly the corner weight of f∗Cf_*C at τ′\tau'. If no full-rank facet occurs, both sides have weight zero there. This proves equality of the cycles. ◻

Corollary 5.17.4. NAG-002191 (A constant weighted degree). Let CC be a finite balanced dd-cycle and f:NR→Rdf:N_{\mathbf R}\to\mathbf R^d an integral affine map. For a point bb outside the images of all cells of dimension at most d−1d-1 and all rank-deficient facets, the number deg⁡f(C;b)=∑x∈∣C∣ , f(x)=bwC(σx)[Zd:f(Nσx)]\deg_f(C;b)=\sum_{x\in |C|\,,\ f(x)=b} w_C(\sigma_x)[\mathbf Z^d:f(N_{\sigma_x})] is finite and independent of bb. It is the constant weight of f∗Cf_*C on Rd\mathbf R^d, allowing zero.

Proof. Each permitted preimage is in a full-rank facet, and each such facet contains at most one preimage. Formula  ({P}) gives the displayed sum. Extend a target subdivision to a subdivision of all of Rd\mathbf R^d, assigning weight zero to the omitted regions. Across every codimension-one face, the two full-dimensional normal vectors are opposite primitive vectors, so balancing equates their weights. Any two generic points can be joined by a polygonal path avoiding the codimension-two faces and crossing only finitely many codimension-one faces: choose its intermediate vertices outside the finitely many affine subspaces causing a nontransverse crossing. For d=1d=1 an interval suffices; for d=0d=0 there is only one target point. The weight is therefore constant throughout the target. ◻

References. The lattice-index method is the one used by Gathmann–Kerber–Markwig, Tropical fans and the moduli spaces of tropical curves, Proposition 2.25, and by Allermann–Rau, First steps in tropical intersection theory, Proposition 4.8. The index and cancellation arguments have been given here for finite affine complexes and real weights.

5.18 Displacement through a standard Jacobian chart

Retain the standard algebra AA and its integral coordinates from Lemma 5.15.1. Let F=(F1,…,Fm)F=(F_1,\ldots,F_m) be represented by restricted series in X0,…,Xr,T,Y1,…,YmX_0,\ldots,X_r,T,Y_1,\ldots,Y_m with coefficients in K∘K^\circ. Write J=∂F/∂YJ=\partial F/\partial Y and consider the nonzero affinoid algebra B=A⟨Y1,…,Ym,Q⟩/(F1,…,Fm,Qdet⁡J−1).(J)B=A\langle Y_1,\ldots,Y_m,Q\rangle/ (F_1,\ldots,F_m,Q\det J-1). \tag{J} We use its displayed quotient norm. In particular the coordinate images and J−1=Qadj⁡(J)J^{-1}=Q\operatorname{adj}(J) have norm at most one. All the statements below keep this presentation over the original valued field.

Lemma 5.18.1. NAG-002156 (Hensel lifting at a Banach-algebra center). Let CC be a complete non-Archimedean Banach algebra, let G∈C⟨V1,…,Vm⟩mG\in C\langle V_1,\ldots,V_m\rangle^m have norm at most one, and let v∈Cmv\in C^m have norm at most one. Suppose D=G′(v)D=G'(v) is invertible with ∥D−1∥≤1\|D^{-1}\|\leq 1 and ∥G(v)∥≤δ<1\|G(v)\|\leq\delta<1. There is a unique root v+hv+h with ∥h∥<1\|h\|<1, and it satisfies ∥h∥≤δ\|h\|\leq\delta. Its Jacobian inverse has norm at most one. Moreover, if two such systems differ in coefficient norm by at most ϵ\epsilon, and their roots both lie at distance less than one from the same center with an inverse Jacobian of norm at most one for the first system at that center, their roots differ by at most ϵ\epsilon.

Proof. Translation by the bounded tuple vv is a contractive substitution. Expanding gives G(v+H)=e+DH+N(H),e=G(v),G(v+H)=e+DH+N(H),\qquad e=G(v), where NN has terms of degree at least two and coefficient norms at most one. This assertion follows first for polynomials by the binomial formula, then for restricted series by uniform truncation. For ∥h∥,∥k∥≤ρ<1\|h\|,\|k\|\leq\rho<1, telescoping each monomial gives ∥N(h)∥≤∥h∥2,∥N(h)−N(k)∥≤ρ∥h−k∥.\|N(h)\|\leq\|h\|^2,\qquad \|N(h)-N(k)\|\leq\rho\|h-k\|. Consequently h↦−D−1e−D−1N(h)h\mapsto-D^{-1}e-D^{-1}N(h) maps the closed ball of radius δ\delta to itself and is a contraction of ratio at most δ\delta. Starting at zero, successive differences tend to zero geometrically; completeness gives a fixed point. Any other root at distance less than one lies with it in a ball of some radius ρ<1\rho<1, where the same estimate proves equality. The Jacobian changes by a matrix of norm at most δ\delta, so a matrix geometric series gives the asserted inverse bound.

For the perturbation assertion, let the two roots be z,wz,w, and write the first system as GG. Telescoping its translated monomials gives G(z)−G(w)=(D+E)(z−w),∥E∥≤max⁡(∥z−v∥,∥w−v∥)<1.G(z)-G(w)=(D+E)(z-w),\qquad \|E\|\leq\max(\|z-v\|,\|w-v\|)<1. The matrix D+ED+E has inverse norm at most one. One term on the left is zero; the other has norm at most ϵ\epsilon because ww is a root of the second system. This proves the bound. All constructions use Banach-algebra operations, so nilpotents are retained. ◻

Proposition 5.18.2. NAG-002157 (The bounded displacement expansion). Put U=(S1,…,Sr,Z1,…,Zq)U=(S_1,\ldots,S_r,Z_1,\ldots,Z_q). There are bounded KK-linear maps Dν:B→BD_\nu:B\to B, ν∈Nr+q\nu\in\mathbb N^{r+q}, with D0=idD_0=\mathrm{id} and ∥Dνb∥≤∥b∥\|D_\nu b\|\leq\|b\|, such that for every 0<δ<10<\delta<1 the series σδ(b)=∑νDν(b)Uν∈B{δ−1U}\sigma_\delta(b)=\sum_\nu D_\nu(b)U^\nu \in B\{\delta^{-1}U\} defines a contractive homomorphism, extending the base displacement Xi↦Xi(1+Si) (i≥1),X0↦X0∏i(1+Si)−1,Tj↦Tj+Zj.X_i\mapsto X_i(1+S_i)\ (i\geq1),\quad X_0\mapsto X_0\prod_i(1+S_i)^{-1},\quad T_j\mapsto T_j+Z_j. It is characterized by this base formula and ∥σδ(Y)−Y∥<1\|\sigma_\delta(Y)-Y\|<1. In fact ∥σδ(b)−b∥≤δ∥b∥,Dν(bc)=∑α+β=νDα(b)Dβ(c).\|\sigma_\delta(b)-b\|\leq\delta\|b\|, \qquad D_\nu(bc)=\sum_{\alpha+\beta=\nu}D_\alpha(b)D_\beta(c).

Proof. Work in Cδ=B{δ−1U}C_\delta=B\{\delta^{-1}U\}. The base formulas define a contractive homomorphism by geometric inversion and bounded substitution. On each integral coordinate the difference from its original image has norm at most δ\delta. Telescoping monomials, passing to restricted series, and taking quotient infima therefore give ∥σδ(a)−a∥≤δ∥a∥\|\sigma_\delta(a)-a\|\leq\delta\|a\| for a∈Aa\in A.

Apply the base displacement to the coefficients of FF. At the center Y∈BmY\in B^m the new error has norm at most δ\delta, since F(Y)=0F(Y)=0. Its Jacobian differs from J(Y)J(Y) by a matrix of norm at most δ\delta. As J(Y)−1=Qadj⁡(J(Y))J(Y)^{-1}=Q\operatorname{adj}(J(Y)) has norm at most one, the new Jacobian also has inverse norm at most one. Lemma 5.18.1 gives the unique root Y+hδY+h_\delta, with ∥hδ∥≤δ\|h_\delta\|\leq\delta. The determinant at that root has an inverse QδQ_\delta of norm at most one and ∥Qδ−Q∥≤δ\|Q_\delta-Q\|\leq\delta, again by the geometric inverse formula. These images kill all relations in ({J}) and give a contractive homomorphism B→CδB\to C_\delta. The same telescoping and quotient-infimum argument gives its displayed difference bound on every b∈Bb\in B.

If δ′≤δ<1\delta'\leq\delta<1, restriction from radius δ\delta to radius δ′\delta' carries this root to a root at distance less than one from YY. Uniqueness identifies it with the root constructed at radius δ′\delta'. Thus all coefficient expansions agree. Evaluation at U=0U=0 is the identity, by the same uniqueness. Write their coefficients as Dν(b)D_\nu(b). Contractivity at every radius gives ∥Dν(b)∥δ∣ν∣≤∥b∥.\|D_\nu(b)\|\delta^{|\nu|}\leq\|b\|. Letting δ\delta increase to one proves ∥Dν(b)∥≤∥b∥\|D_\nu(b)\|\leq\|b\|. The homomorphism identity at any positive radius, together with uniqueness of coefficients, gives the convolution identity and KK-linearity. ◻

Theorem 5.18.3. NAG-002158 (The strict-time analytic lift). For x∈M(B)x\in\mathcal M(B) and 0≤t<10\leq t<1, the formula ∣b(HB(x,t))∣=max⁡ν∣Dν(b)(x)∣t∣ν∣,00=1,(E)|b(H_B(x,t))| =\max_\nu |D_\nu(b)(x)|t^{|\nu|}, \qquad 0^0=1, \tag{E} defines a jointly continuous map HB:M(B)×[0,1)→M(B)H_B:\mathcal M(B)\times[0,1)\to\mathcal M(B) with HB(x,0)=xH_B(x,0)=x. For the structure map f:M(B)→Vf:\mathcal M(B)\to V, f(HB(x,t))=H(f(x),t).f(H_B(x,t))=H(f(x),t). If bb has quotient norm at most one, then ∣b(HB(x,t))∣<1|b(H_B(x,t))|<1 if and only if ∣b(x)∣<1|b(x)|<1. Thus every strict-time trajectory preserves its reduction prime in the displayed integral model.

Proof. For t>0t>0, evaluate the coefficients of σδ(b)\sigma_\delta(b) in H(x)\mathcal H(x) with t<δ<1t<\delta<1 and then use the multiplicative radius-tt Gauss norm. This gives a bounded homomorphism into a complete Gauss field and hence an actual analytic point. Its norm is ({E}). At t=0t=0 the constant coefficient gives xx. The base-coordinate formulas identify the composite with ff.

On M(B)×[0,δ]\mathcal M(B)\times[0,\delta], choose δ<δ′<1\delta<\delta'<1. The restricted series σδ′(b)\sigma_{\delta'}(b) has uniformly small tails on this product. Each finite truncation of ({E}) is a finite maximum of continuous functions, so its uniform limit is continuous. These intervals cover [0,1)[0,1) and prove joint continuity. For a norm-at-most-one element bb, its constant coefficient is b(x)b(x) and its nonconstant terms have norms at most tt. Thus a constant coefficient of norm one stays of norm one; if it is smaller than one, every term is bounded by max⁡(∣b(x)∣,t)<1\max(|b(x)|,t)<1. This proves the reduction assertion. ◻

Proposition 5.18.4. NAG-002160 (Functoriality and the composition law). A contractive homomorphism between two presentations ({J}) over the same standard base commutes with their coefficient operators DνD_\nu. The lifts therefore commute with such morphisms, with integral Laurent restrictions, and with complete isometric ground-field extension. They satisfy HB(HB(x,s),t)=HB(x,max⁡(s,t))(0≤s,t<1).H_B(H_B(x,s),t)=H_B(x,\max(s,t)) \qquad(0\leq s,t<1).

Proof. Let ψ:B→C\psi:B\to C be the homomorphism in question. Applying ψ\psi to the lifted YY-tuple gives a root of the base-displaced BB-equations in C{δ−1U}C\{\delta^{-1}U\}. Applying the CC-displacement to ψ(Y)\psi(Y) gives another such root. Both differ from ψ(Y)\psi(Y) by norm at most δ\delta, and the original Jacobian inverse has norm at most one. Lemma 5.18.1 identifies them. The inverse determinant and then every element of BB agree by bounded substitution. Comparing coefficients proves the assertion. The same proof treats a complete isometric scalar extension using the identical equations.

For an integral Laurent restriction adjoining g−1g^{-1}, both gg and g−1g^{-1} have norm at most one. The displacement of gg differs from gg by norm at most δ\delta; factoring out gg and using geometric inversion extends the map uniquely to the inverse, with the same bound. This proves the restriction assertion. It also shows that two bounded presentations identified over the fixed base give the same displacement.

For the composition law, take independent variables (S,Z)(S,Z) and (S′,Z′)(S',Z') of radii s,t<1s,t<1. The two successive lifts solve the base-displaced equations whose relative coordinates are Wi=Si+Si′+SiSi′,Vj=Zj+Zj′.W_i=S_i+S'_i+S_iS'_i,\qquad V_j=Z_j+Z'_j. The resulting YY-tuple differs from the original one by norm at most max⁡(s,t)<1\max(s,t)<1. The single lift evaluated at (W,V)(W,V) has the same property and solves the same equations. Hensel uniqueness identifies the two tuples. The isometric Gauss-coordinate change in Proposition 5.15.3 then identifies their norms with the single displacement of radius max⁡(s,t)\max(s,t). This gives the identity for all b∈Bb\in B, and the zero-parameter cases are immediate. ◻

Lemma 5.18.5. NAG-002161 (The pointwise endpoint). For every x∈M(B)x\in\mathcal M(B), the formula ∣b(HB(x,1))∣=sup⁡ν∣Dν(b)(x)∣=lim⁡t↑1∣b(HB(x,t))∣|b(H_B(x,1))|=\sup_\nu|D_\nu(b)(x)| =\lim_{t\uparrow1}|b(H_B(x,t))| defines a point of M(B)\mathcal M(B) and extends its trajectory continuously to [0,1][0,1]. Each endpoint evaluation is lower semicontinuous as a function of xx. The endpoint commutes with all strict-time maps and is idempotent: HB(HB(x,1),t)=HB(x,1)=HB(HB(x,t),1),0≤t≤1.H_B(H_B(x,1),t)=H_B(x,1)=H_B(H_B(x,t),1), \qquad 0\leq t\leq1.

Proof. For each bb, the strict-time value is nondecreasing and bounded by ∥b∥\|b\|. Its limit is the displayed coefficient supremum: each fixed coefficient is attained in the limit, and their supremum bounds every time value. Passing to limits in the multiplicative, ultrametric, and scalar identities proves that this is a bounded multiplicative seminorm extending the given absolute value. Evaluation topology gives convergence to its point, hence continuity of the trajectory at time one. A supremum of continuous coefficient evaluations is lower semicontinuous.

Fix t<1t<1. Continuity of the strict-time map in xx and Proposition 5.18.4 allow s↑1s\uparrow 1 in HB(HB(x,s),t)=HB(x,max⁡(s,t))H_B(H_B(x,s),t)=H_B(x,\max(s,t)). The resulting identity is HB(HB(x,1),t)=HB(x,1)H_B(H_B(x,1),t)=H_B(x,1). In the opposite order, pointwise convergence applied to the fixed point HB(x,t)H_B(x,t) gives HB(HB(x,t),1)=HB(x,1)H_B(H_B(x,t),1)=H_B(x,1). Finally let t↑1t\uparrow 1 in the first equality, again using the continuous trajectory of the fixed point HB(x,1)H_B(x,1), to get idempotence. ◻

References. The construction is an explicit Banach-algebra Hensel lift of mixed Gauss displacement. The norm estimates are uniform in the real parameter below one.

5.19 Pure dimension and geometric weights

The initial scheme is taken in the sense of Theorem 5.13.6. A scheme is pure of dimension dd if every irreducible component of its underlying space has dimension dd; embedded components are allowed.

Lemma 5.19.1. NAG-002105 (Finite filtrations and zero-dimensional algebras). Every finite module over a Noetherian ring has a finite filtration whose nonzero successive quotients are A/pA/\mathfrak p for prime ideals p\mathfrak p. A zero-dimensional algebra of finite type over a field kk is finite-dimensional over kk.

Proof. Finite free modules over a Noetherian ring are Noetherian by induction on their rank, using the first-coordinate ideal and the kernel of that coordinate; hence their quotients are Noetherian. In a nonzero finite module choose a nonzero element whose annihilator is maximal among annihilators of nonzero elements. Such a maximum exists by the ascending chain condition. Its annihilator p\mathfrak p is prime: if abab annihilates the element and bb does not, the annihilator of its nonzero bb-multiple is the same maximal annihilator, so a∈pa\in\mathfrak p. This yields a submodule A/pA/\mathfrak p. Repeat in the quotient. The resulting ascending chain of submodules terminates.

For the second assertion, finite-type kk-algebras are Noetherian by [StacksAlg] (Tag 00FN). Apply the filtration to the algebra itself. Each prime quotient is a zero-dimensional finite-type domain. Its fraction field has transcendence degree zero by [StacksAlg] (Tag 00P0). Its finitely many algebra generators are therefore algebraic over kk, so the quotient is a finite-dimensional field. A finite extension of these vector spaces is finite-dimensional. ◻

Proposition 5.19.2. NAG-002106 (Pure dimension of initial schemes). Let X⊂Spec⁡K[M]X\subset\operatorname{Spec}K[M] be pure of dimension dd. For every w∈Trop⁡(X)w\in\operatorname{Trop}(X), every initial scheme in⁡w,t,L(X)\operatorname{in}_{w,t,L}(X) is pure of dimension dd.

Proof. By [StacksAlg] (Tag 02NK), extension from KK to LL preserves pure dimension. Translate by tt and work at weight zero over LL. Write R=L∘R=L^\circ, k=L~k=\widetilde L and I⊂L[M]I\subset L[M] for the translated ideal. Its finitely many minimal primes are P1,…,PsP_1,\ldots,P_s. Let B=R[M]/(I∩R[M]),Bi=R[M]/(Pi∩R[M]).B=R[M]/(I\cap R[M]),\qquad B_i=R[M]/(P_i\cap R[M]). Each BiB_i is a domain of finite type containing RR, and its generic fibre is the dimension-dd domain L[M]/PiL[M]/P_i. The exact valuation-ring dimension theorem [StacksAlg] (Tag 00QK) says that, whenever Bi⊗RkB_i\otimes_R k is nonzero, its spectrum is equidimensional of dimension dd. Its hypotheses require finite type, not finite presentation or a discrete valuation.

The identity I∩R[M]=I∩R[M]=⋂i(Pi∩R[M])\sqrt{I\cap R[M]}=\sqrt I\cap R[M] =\bigcap_i(P_i\cap R[M]) shows that Spec⁡B\operatorname{Spec}B is the union of the closed subsets Spec⁡Bi\operatorname{Spec}B_i. Taking the inverse image of this finite union in the closed fibre shows that Spec⁡(B⊗Rk)\operatorname{Spec}(B\otimes_Rk) is the union of the nonempty pure dimension-dd special fibres just described. A finite union of pure dimension-dd closed subsets is pure of dimension dd: every irreducible component is a maximal irreducible component among them, and a proper containment of irreducible closed subsets strictly lowers their finite dimensions. The closed fibre is nonempty by Theorem 5.13.6. ◻

Proposition 5.19.3. NAG-002107 (The tangent torus of an initial stratum). Use the finite complex constructed in Theorem 5.16.5. Let ww lie in the relative interior of a cell τ⊂Trop⁡(X)\tau\subset\operatorname{Trop}(X), put Lτ=span⁡R(τ−τ)L_\tau=\operatorname{span}_{\mathbf R}(\tau-\tau) and M0=M∩Lτ⊥M_0=M\cap L_\tau^\perp. The ideal of Y=in⁡w(X)Y=\operatorname{in}_w(X) is generated by elements of k[M0]k[M_0], after multiplication of its generators by unit characters. Consequently it is extended from its intersection with k[M0]k[M_0], and Y≃TNτ×Zτ,Zτ=Spec⁡(k[M0]/(IY∩k[M0])),Nτ=N∩Lτ.Y\simeq T_{N_\tau}\times Z_\tau, \qquad Z_\tau=\operatorname{Spec}\bigl(k[M_0]/(I_Y\cap k[M_0])\bigr), \quad N_\tau=N\cap L_\tau. The product uses a lattice splitting, while the quotient scheme ZτZ_\tau is intrinsic to YY and NτN_\tau.

Proof. Use the finitely many homogeneous pivot rows in the proof of Theorem 5.16.5, and divide each row by its pivot character. A surviving exponent difference aa satisfies one of the affine Plücker equalities defining the cell. That equality holds throughout the relative interior of τ\tau, hence its linear part vanishes on LτL_\tau. Thus a∈M0a\in M_0 after dehomogenization. These rows generate the initial Laurent ideal, so it is extended from an ideal J0⊂k[M0]J_0\subset k[M_0].

The lattice NτN_\tau is saturated. Lemma 5.14.1 splits N=Nτ⊕QN=N_\tau\oplus Q, and dually k[M]k[M] is a Laurent polynomial ring over k[M0]k[M_0]. Comparing coefficients in these additional Laurent variables shows J0k[M]∩k[M0]=J0J_0 k[M]\cap k[M_0]=J_0. The quotient ring is therefore (k[M0]/J0)[Nτ∨](k[M_0]/J_0)[N_\tau^\vee], proving the assertion including its scheme structure. ◻

Lemma 5.19.4. NAG-002108 (A zero tropical support forces dimension zero). Let ZZ be a nonempty closed subscheme of a torus over a complete field kk. If Trop⁡(Z)⊆{0}\operatorname{Trop}(Z)\subseteq\{0\}, then ZZ is zero-dimensional and its coordinate algebra is finite over kk.

Proof. Suppose an integral component has positive dimension, with function field FF. By [StacksAlg] (Tag 00P0), F/kF/k has positive transcendence degree. At least one of its torus-coordinate generators, say z1z_1, is transcendental over kk. Extend z1z_1 to a maximal algebraically independent subset z1,…,zez_1,\ldots,z_e of a finite field-generating set. Then FF is finite over k(z1,…,ze)k(z_1,\ldots,z_e). Give the latter field its Gauss absolute value with ∣z1∣≠1|z_1|\ne 1 and arbitrary positive radii on the other variables, and complete it to EE. The Gauss norm is multiplicative by Proposition 1.2.2.

The finite nonzero EE-algebra D=F⊗k(z1,…,ze)ED=F\otimes_{k(z_1,\ldots,z_e)}E has a maximal proper ideal: choose one of maximal EE-dimension. Its quotient is a finite field extension E′/EE'/E, which has the extending complete absolute value by Lemma 1.12.1. The induced map F→E′F\to E' is injective, since it is a unital map from a field. Restricting its norm to the coordinate algebra gives an analytic point of ZZ by Proposition 2.10.1, with ∣z1∣≠1|z_1|\ne 1. This contradicts the tropical-support hypothesis. All components therefore have dimension zero. Apply Lemma 5.19.1 to obtain finiteness over kk. ◻

Theorem 5.19.5. NAG-002109 (Pure dimension of algebraic tropicalization). If X⊂Spec⁡K[M]X\subset\operatorname{Spec}K[M] is nonempty and pure of dimension dd, then Trop⁡(X)\operatorname{Trop}(X) is a nonempty finite pure dd-dimensional integral v(K×)v(K^\times)-affine polyhedral complex. At the relative interior of each maximal cell σ\sigma, the quotient ZσZ_\sigma of Proposition 5.19.3 is a nonempty finite scheme over the residue field.

Proof. The Gauss-field construction in Lemma 5.19.4, with arbitrary radii and allowing an empty transcendence basis, gives an analytic point on each nonempty integral component. Thus the tropicalization is nonempty. The finite complex is Theorem 5.16.5. Fix a maximal cell σ\sigma, a relative-interior point ww, and put s=dim⁡σs=\dim\sigma. The local star at ww is the vector space LσL_\sigma. By Theorem 5.16.6, Y=in⁡w(X)Y=\operatorname{in}_w(X) has tropical support LσL_\sigma. By Proposition 5.19.3, Y=TNσ×ZσY=T_{N_\sigma}\times Z_\sigma. Every analytic point of ZσZ_\sigma gives a point of this product by assigning value one to the first torus coordinates. It follows that Trop⁡(Zσ)⊆{0}\operatorname{Trop}(Z_\sigma)\subseteq\{0\}. The quotient is nonempty, since YY is nonempty. Lemma 5.19.4 makes ZσZ_\sigma finite.

Every component of TNσ×ZσT_{N_\sigma}\times Z_\sigma has dimension ss: a prime quotient of its finite coefficient algebra is a finite field extension, and its Laurent polynomial ring has transcendence degree ss, so [StacksAlg] (Tag 00P0) applies. Proposition 5.19.2 gives dimension dd for the same components. Thus s=ds=d. This holds for every maximal cell. ◻

Definition 5.19.6. NAG-002110 (The geometric tropical weight). At w∈relint⁡(σ)w\in\operatorname{relint}(\sigma) as in Theorem 5.19.5, define mX(σ)=dim⁡kΓ(Zσ,OZσ).m_X(\sigma)=\dim_k\Gamma(Z_\sigma,\mathcal O_{Z_\sigma}).

Proposition 5.19.7. NAG-002111 (Geometric multiplicity and its invariance). The weight is a positive integer. It is independent of the chosen weight-realizing field, translating point, and lattice splitting. It equals the sum of the generic lengths of the irreducible components of the initial scheme after a residue extension splitting its finite transverse algebra. In particular it equals the usual sum of geometric generic lengths over an algebraic closure.

Proof. Let A=Γ(Zσ,O)A=\Gamma(Z_\sigma,\mathcal O), a nonzero finite-dimensional kk-algebra. Finite-dimensional scalar extension preserves its dimension. A finite field extension k′/kk'/k splits the characteristic polynomials of a finite set of algebra generators of AA. Every residue field of A⊗kk′A\otimes_k k' is then k′k', because the images of those generators are roots in k′k'. Choose a composition series by successively taking a proper submodule of maximal vector-space dimension. Its factors are the one-dimensional residue fields. Localizing the series at each maximal ideal counts each factor exactly once, and gives dim⁡kA=dim⁡k′(A⊗kk′)=∑zlength⁡((A⊗kk′)z).\dim_kA=\dim_{k'}(A\otimes_k k') =\sum_z\operatorname{length}\bigl((A\otimes_k k')_z\bigr). Adjoining Laurent variables and localizing at a generic point tensors the local Artinian factor with the rational function field of the corresponding torus. Its composition series tensors to a composition series with the same number of simple factors. Thus the displayed summands are exactly the generic lengths of Yk′Y_{k'}. The intrinsic quotient description in Proposition 5.19.3 makes the lattice splitting irrelevant. Theorem 5.13.6 compares all other choices by residue extension and a torus translation; translation preserves the tangent subtorus and induces a translation on its quotient. The same finite dimension is therefore obtained from every choice. ◻

Proposition 5.19.8. NAG-002112 (Constancy, refinement, and complete base change). The integer mX(σ)m_X(\sigma) is constant on the relative interior of each maximal initial stratum. The resulting weighted polyhedral complex is independent of its choice of initial-stratum refinement, and every complete isometric extension of KK preserves its support and weights.

Proof. Theorem 5.16.5 compares the initial schemes at two relative-interior points by a single torus translation after residue extension. The cell has the same tangent lattice at both points, so the comparison descends to their transverse finite quotients. Proposition 5.19.7 equates the dimensions. A subdivision retains the tangent lattice on its dimension-dd pieces and therefore retains the weights. Any two initial-stratum complexes have a common refinement; at a point belonging to both relative interiors their weights are computed from the identical initial scheme and tangent lattice. Finally, the initial scalar-extension identity in Proposition 5.13.4, together with the common-extension comparison in Theorem 5.13.6, gives the same finite quotient after residue extension. Its dimension is unchanged. ◻

References. The classical pure-dimension and multiplicity statements are reconstructed here from the exact valuation-ring dimension theorem of the Stacks Project, the finite Plücker construction above, and the finite transverse algebra of an initial stratum.

5.20 Stable intersection from the diagonal

Retain the finite real cycle convention of Section 5.14. Put n=rank⁡Nn=\operatorname{rank}N. Choose a lattice basis of N∨N^\vee, and let xi,yix_i,y_i be the corresponding coordinates on NR×NRN_{\mathbf R}\times N_{\mathbf R}. Set hi(x,y)=max⁡{0,xi−yi},Dxy=Dh1⋯Dhn,Δ={(x,x):x∈NR}.\begin{gathered} h_i(x,y)=\max\{0,x_i-y_i\},\qquad \mathcal D_{xy}=D_{h_1}\cdots D_{h_n},\\ \Delta=\{(x,x):x\in N_{\mathbf R}\}. \end{gathered}

Lemma 5.20.1. NAG-002192 (A basis-independent diagonal operator). For every balanced cycle ZZ in NR×NRN_{\mathbf R}\times N_{\mathbf R}, the cycle DxyZ\mathcal D_{xy}Z is balanced, is supported in ∣Z∣∩Δ|Z|\cap\Delta, and is independent of the chosen basis of N∨N^\vee. Interchanging xx and yy preserves this operator.

Proof. By Theorem 5.14.5, successive corners are balanced and their order does not matter. A corner by max⁡(0,ℓ)\max(0,\ell) is supported where ℓ=0\ell=0, since off that hyperplane the function is affine. It follows inductively that the final support lies in Δ\Delta.

Every change of integral basis is a composition of row interchanges, sign changes, and addition of an integer multiple of one row to another. Row interchanges use commutation of corners. A sign change replaces max⁡(0,ℓ)\max(0,\ell) by max⁡(0,−ℓ)=max⁡(0,ℓ)−ℓ\max(0,-\ell)=\max(0,\ell)-\ell, which has the same corner operator. For a row operation ℓi↦ℓi+aℓj\ell_i\mapsto\ell_i+a\ell_j, perform the jj-th corner first. Its support is contained in ℓj=0\ell_j=0, where the old and new ii-th functions coincide. Restriction locality, Lemma 5.14.6, proves equality. The elementary row operations generate GLn(Z)\mathrm{GL}_n(\mathbf Z) by Lemma 5.14.1, so this proves basis independence. Interchanging x,yx,y changes each linear form to its negative and is already covered. ◻

Definition 5.20.2. (The stable product). For balanced cycles C,EC,E of dimensions p,qp,q in NRN_{\mathbf R}, define C⋅E=π∗(Dxy(C×E)),π(x,y)=x.C\mathbin{\cdot}E =\pi_*\bigl(\mathcal D_{xy}(C\times E)\bigr), \qquad \pi(x,y)=x. The projection on the support in Δ\Delta is a lattice isomorphism. The product has dimension p+q−np+q-n, and is zero when that number is negative.

Theorem 5.20.3. NAG-002194 (The ring of balanced cycles). The stable product is bilinear, commutative, associative, and compatible with weighted subdivision and integral affine lattice isomorphisms. Its unit is [NR][N_{\mathbf R}] with weight one, and ∣C⋅E∣⊆∣C∣∩∣E∣|C\cdot E|\subseteq |C|\cap|E|.

Proof. Bilinearity and refinement invariance follow from those of products, corners, and pushforwards. The preceding lemma gives the support assertion, commutativity, and invariance under changes of lattice basis. Simultaneous translation of x,yx,y leaves their differences unchanged, proving affine invariance.

For the unit, begin with NR×EN_{\mathbf R}\times E. Impose xi=yix_i=y_i one coordinate at a time by its corner. Before the ii-th step, xix_i is still a free lattice coordinate, so the two sides of xi−yi=0x_i-y_i=0 have opposite primitive normals and the slope jump of max⁡(0,xi−yi)\max(0,x_i-y_i) is exactly one. Each facet thus keeps the weight of its EE factor. At a face inherited from EE, the function has a common affine extension on either side away from that equality and contributes zero. The final cycle is the diagonal copy of EE with its original weights. Projection therefore gives [NR]⋅E=E[N_{\mathbf R}]\cdot E=E.

For associativity, work on C×E×FC\times E\times F with coordinates x,y,zx,y,z. By product locality, imposing the x=yx=y diagonal first gives the diagonal copy of (C⋅E)×F(C\cdot E)\times F. On this support the functions max⁡(0,xi−zi)\max(0,x_i-z_i) and max⁡(0,yi−zi)\max(0,y_i-z_i) agree. Hence the lift of (C⋅E)⋅F(C\cdot E)\cdot F to the small diagonal is DyzDxy(C×E×F).\mathcal D_{yz}\mathcal D_{xy}(C\times E\times F). Starting instead with E⋅FE\cdot F gives DxyDyz(C×E×F)\mathcal D_{xy}\mathcal D_{yz}(C\times E\times F). The corner operators commute, so the two cycles agree before projection to xx. If an intermediate dimension is negative, both expressions vanish by the same commuting-operator calculation. This proves associativity in all degrees. ◻

Lemma 5.20.4. NAG-002195 (Complementary fans and displacement). Let A,BA,B be finite balanced fans in NRN_{\mathbf R}, of dimensions p,qp,q with p+q=np+q=n. For a vector bb outside the finitely many lower-dimensional images of faces under (x,y)↦x−y(x,y)\mapsto x-y, the coefficient of their stable product at the origin is ∑σ∈A(p), τ∈B(q)b∈relint⁡(σ−τ)Nσ,R+Nτ,R=NRwA(σ)wB(τ)[N:Nσ+Nτ].(F)\sum_{\substack{\sigma\in A^{(p)},\ \tau\in B^{(q)}\\ b\in\operatorname{relint}(\sigma-\tau)\\ N_{\sigma,\mathbf R}+N_{\tau,\mathbf R}=N_{\mathbf R}}} w_A(\sigma)w_B(\tau)[N:N_\sigma+N_\tau]. \tag{F} In particular this sum is independent of the generic vector bb, including for signed real weights.

Proof. Put Z=A×BZ=A\times B and f(x,y)=x−yf(x,y)=x-y. Its dimension is nn. By Corollary 5.17.4, f∗Z=c[NR]f_*Z=c[N_{\mathbf R}], where cc is the generic weighted degree. On a full-rank product facet, ff is injective because the two direction spaces have dimensions adding to nn. Its lattice image is Nσ+NτN_\sigma+N_\tau, and a generic inverse image exists precisely when b∈relint⁡(σ−τ)b\in\operatorname{relint}(\sigma-\tau). Thus cc is the displayed sum.

On NRN_{\mathbf R} let gi(z)=max⁡(0,zi)g_i(z)=\max(0,z_i). Repeated use of Theorem 5.17.3 gives f∗(Dg1∘f⋯Dgn∘fZ)=Dg1⋯Dgn(c[NR])=c[0].f_*\bigl(D_{g_1\circ f}\cdots D_{g_n\circ f}Z\bigr) =D_{g_1}\cdots D_{g_n}\bigl(c[N_{\mathbf R}]\bigr) =c[0]. The final equality is the successive unit slope jumps of the coordinate hyperplanes. The source is a zero-dimensional fan supported on the diagonal, so it is a multiple of the origin. Pushforward does not change a zero-dimensional point weight. By the definition of the diagonal product this multiple is also the coefficient of A⋅BA\cdot B. It is therefore cc, as required. ◻

Theorem 5.20.5. NAG-002197 (The local fan displacement formula). Let C,EC,E be finite balanced real cycles in NRN_{\mathbf R}, of dimensions p,qp,q, and put r=p+q−n≥0r=p+q-n\geq 0. Choose a common polyhedral refinement of their intersection and let γ\gamma be an rr-dimensional face. At a∈relint⁡(γ)a\in\operatorname{relint}(\gamma), take the two local stars and quotient by L=Nγ,RL=N_{\gamma,\mathbf R}. Their dimensions are complementary in NR/LN_{\mathbf R}/L. The weight of C⋅EC\cdot E on γ\gamma is the sum ({F}) for these quotient fans. Equivalently it is the weighted small-displacement limit of the transverse intersections of CC and E+εbE+\varepsilon b, with lattice factors [N:Nσ+Nτ][N:N_\sigma+N_\tau].

Proof. Stars commute with corners and products. Both stars at aa have LL as a lineality space: the minimal face of either original complex containing aa contains the local directions of γ\gamma. Since NγN_\gamma is saturated, Lemma 5.14.1 allows us to choose N=Nγ⊕QN=N_\gamma\oplus Q. Write the two stars as L×AL\times A and L×BL\times B. The diagonal cuts in the LL coordinates give the diagonal copy of LL with weight one, by the unit calculation in Theorem 5.20.3. The remaining cuts give A⋅BA\cdot B. Their dimensions satisfy dim⁡A+dim⁡B=(p−r)+(q−r)=n−r=rank⁡Q.\dim A+\dim B=(p-r)+(q-r)=n-r=\operatorname{rank}Q. The preceding lemma computes the coefficient at their origin, which is the desired weight on γ\gamma. For every contributing pair of original facets, NγN_\gamma is contained in both direction lattices, so [N/Nγ:(Nσ+Nτ)/Nγ]=[N:Nσ+Nτ].[N/N_\gamma:(N_\sigma+N_\tau)/N_\gamma] =[N:N_\sigma+N_\tau]. Finally choose a small neighborhood of aa meeting only the faces that contain its minimal faces. In this neighborhood the original cells are translates of their local tangent cones. For a generic bb and sufficiently small ε>0\varepsilon>0, a transverse displaced pair approaching aa is therefore exactly a pair counted by the quotient-fan formula; the signs agree because x∈C∩(E+εb)x\in C\cap(E+\varepsilon b) means x−y=εbx-y=\varepsilon b. Only finitely many cells occur, so the incidence conditions are constant for all sufficiently small positive ε\varepsilon. Summing their weights gives the claimed weighted limit. There is no additional support in the relative interior of an intersection cell of dimension s>rs>r: quotienting its common ss-dimensional lineality gives a product of dimension r−s<0r-s<0, which is zero. Cells of dimension less than rr cannot carry an rr-cycle. Thus refining the rr-faces gives all coefficients of the cycle. ◻

Corollary 5.20.6. NAG-002198 (Pointwise smooth weights). Suppose the weights of C,EC,E are smooth functions on their cells, extending smoothly to their faces, and that their values satisfy balancing at every point of every face. Freezing the weights at each point and applying the local fan formula defines a smooth-weighted balanced product. It is independent of subdivision and generic displacement, and is bilinear, commutative, and associative.

Proof. At a fixed point, the frozen stars are constant-weight balanced fans. Theorem 5.20.5 makes their product coefficient well-defined. On any fixed prospective intersection face, this coefficient is a finite sum of products of restrictions of the original weights multiplied by constant lattice indices. It is therefore smooth. At a lower face, freezing again identifies the normal equation with balancing of the product of the frozen fans. Refinement and the two bracketings of a triple product are compared by the same frozen fan calculation, using Theorem 5.20.3. The equalities hold at every point, hence as equalities of smooth coefficient functions. ◻

References. The diagonal construction is due to Allermann–Rau, First steps in tropical intersection theory, Section 9. For the stable fan description see Jensen–Yu, Stable intersections of tropical varieties, Section 2. Here basis independence is proved by elementary row operations, and the displacement formula by the balanced pushforward of the difference map.

5.21 Continuity at the endpoint of a Jacobian chart

Lemma 5.21.1. NAG-002162 (Finite control of bounded coefficients). Let CC be a nonzero affinoid algebra, and let R⊆CR\subseteq C be a Noetherian subring all of whose elements have norm at most one. For a family cν∈Rc_\nu\in R indexed by Nd\mathbb N^d, put q(x,t)=max⁡ν∣cν(x)∣t∣ν∣(0≤t<1),q(x,1)=sup⁡ν∣cν(x)∣.q(x,t)=\max_\nu|c_\nu(x)|t^{|\nu|}\quad(0\leq t<1), \qquad q(x,1)=\sup_\nu|c_\nu(x)|. There is a finite set of indices EE for which q(x,1)=max⁡ν∈E∣cν(x)∣(x∈M(C)).q(x,1)=\max_{\nu\in E}|c_\nu(x)| \qquad(x\in\mathcal M(C)). If the family is nonzero and N=max⁡ν∈E∣ν∣N=\max_{\nu\in E}|\nu|, then tNq(x,1)≤q(x,t)≤q(x,1)t^Nq(x,1)\leq q(x,t)\leq q(x,1) for 0<t≤10<t\leq 1. Consequently qq is jointly continuous on M(C)×[0,1]\mathcal M(C)\times[0,1].

Proof. The ideal of RR generated by all cνc_\nu is finitely generated. Express a finite generating list in terms of the original family and take the union of its finite index sets; thus finitely many actual cνc_\nu, ν∈E\nu\in E, generate the same ideal. Every other coefficient is a finite sum of these generators multiplied by elements of RR. Since all those multipliers have norm at most one, evaluation gives the claimed finite maximum at every xx. If the ideal is zero, the conclusion is immediate. Otherwise retaining those same terms in q(x,t)q(x,t) gives the lower bound tNq(x,1)t^Nq(x,1); the upper bound is its definition. In particular 0≤q(x,1)−q(x,t)≤1−tN.0\leq q(x,1)-q(x,t)\leq1-t^N. At time one this is uniform convergence to a continuous finite maximum. For times bounded above by δ<1\delta<1, the coefficient bound makes the tails at most δn+1\delta^{n+1} after total degree nn, so finite maxima converge uniformly and give joint continuity there as well. ◻

Lemma 5.21.2. NAG-002163 (Polynomial implicit equations have finite endpoint control). Let CC be an affinoid algebra receiving the standard algebra AA contractively. Suppose the system G∈K∘[X0,…,Xr,T,Y]mG\in K^\circ[X_0,\ldots,X_r,T,Y]^m has a root y∈Cmy\in C^m of norm at most one and its determinant has an inverse q∈Cq\in C of norm at most one. Lift this root under the base displacement by Lemma 5.18.1. For every polynomial P∈K[X0,…,Xr,T,Y,Q]P\in K[X_0,\ldots,X_r,T,Y,Q], the norm of PP evaluated at the displaced base coordinates, the lifted root, and the inverse determinant extends jointly continuously to M(C)×[0,1]\mathcal M(C)\times[0,1] by taking its coefficient supremum at time one.

Proof. First assume the coefficients of PP belong to K∘K^\circ. Let R⊆CR\subseteq C be the ring generated over the image of Z\mathbb Z by the finitely many coefficients of G,PG,P, the base-coordinate images, the coordinates of yy, and qq. Every generator has norm at most one, hence so does every element of RR. The ring Z\mathbb Z is Noetherian, since each ideal is generated by its least positive element, or is zero. Its image is a quotient, and [StacksAlg] (Tag 00FN) therefore makes RR Noetherian.

We check that every formal displacement coefficient lies in this single ring. The base displacement has coefficients in RR: the inverse factors (1+Si)−1(1+S_i)^{-1} have integral coefficients. Write the lifted root as y+h(U)y+h(U), with h(0)=0h(0)=0. In each positive total degree, the coefficient of an as-yet-unknown term of hh in Gσ(y+h)G_\sigma(y+h) is the original matrix G′(y)G'(y) times that coefficient; all remaining terms use only already determined lower degrees. The inverse matrix qadj⁡(G′(y))q\operatorname{adj}(G'(y)) has its entries in RR. Induction on total degree therefore puts every coefficient of hh in RR. This formal recursion agrees with the analytic root at each radius less than one by uniqueness. Recursively inverting its determinant puts every coefficient of the new inverse determinant in RR as well. Polynomial substitution then does the same for PP.

Lemma 5.21.1 now applies to this coefficient family and proves the assertion. For arbitrary PP, multiply it by a nonzero scalar making its finitely many coefficients integral, apply the result, and divide the real-valued norm function by the absolute value of that scalar. Such a scalar exists because the ground-field valuation is nontrivial. ◻

Theorem 5.21.3. NAG-002164 (Joint continuity for restricted-series Jacobian charts). For every presentation ({J}), the pointwise endpoint of Lemma 5.18.5 makes HB:M(B)×[0,1]⟶M(B)H_B:\mathcal M(B)\times[0,1]\longrightarrow\mathcal M(B) jointly continuous. Its full homotopy commutes with the contractive morphisms, integral Laurent restrictions and complete isometric extensions of Proposition 5.18.4.

Proof. We first treat a polynomial representative b=P(X,T,Y,Q)∈Bb=P(X,T,Y,Q)\in B. Choose polynomial systems F(n)∈K∘[X,T,Y]mF^{(n)}\in K^\circ[X,T,Y]^m with ∥F(n)−F∥≤ϵn<1\|F^{(n)}-F\|\leq\epsilon_n<1 and ϵn→0\epsilon_n\to 0, using the fixed restricted-series representatives. Differentiation is contractive for the unit-radius Gauss norms, so the Jacobians are within the same bound. At the original tuple Y∈BmY\in B^m, the system F(n)F^{(n)} has error at most ϵn\epsilon_n and an inverse Jacobian of norm at most one. Thus Lemma 5.18.1, now in BB, gives an exact root YnY_n with ∥Yn−Y∥≤ϵn,Qn=det⁡((F(n))′(Yn))−1,∥Qn∥≤1,∥Qn−Q∥≤ϵn.\begin{gathered} \|Y_n-Y\|\leq\epsilon_n, \qquad Q_n=\det((F^{(n)})'(Y_n))^{-1},\\ \|Q_n\|\leq1,\qquad \|Q_n-Q\|\leq\epsilon_n. \end{gathered}

Displace the base and lift the two roots, obtaining YδY_\delta and Yn,δY_{n,\delta} in B{δ−1U}B\{\delta^{-1}U\} for δ<1\delta<1. Both roots lie at distance at most max⁡(δ,ϵn)<1\max(\delta,\epsilon_n)<1 from the original tuple YY. Their displaced systems differ by at most ϵn\epsilon_n, since base substitution is contractive. The perturbation part of Lemma 5.18.1 therefore gives, uniformly for every δ<1\delta<1, ∥Yn,δ−Yδ∥≤ϵn.\|Y_{n,\delta}-Y_\delta\|\leq\epsilon_n. The determinant difference has the same bound, by telescoping and the derivative bound. Both inverses have norm at most one, so the identity u−1−v−1=u−1(v−u)v−1u^{-1}-v^{-1}=u^{-1}(v-u)v^{-1} gives ∥Qn,δ−Qδ∥≤ϵn\|Q_{n,\delta}-Q_\delta\|\leq\epsilon_n. Let Pn(U)P_n(U) denote PP evaluated on the displaced base, Yn,δY_{n,\delta} and Qn,δQ_{n,\delta}. This definition is compatible with all radii. If CPC_P is the maximum absolute value of the coefficients of PP, telescoping its finitely many monomials yields ∥Pn(U)−σδ(b)∥δ≤CPϵn(0<δ<1).(A)\|P_n(U)-\sigma_\delta(b)\|_\delta \leq C_P\epsilon_n \qquad(0<\delta<1). \tag{A}

By Lemma 5.21.2, the norm functions defined by Pn(U)P_n(U), with coefficient suprema at time one, are jointly continuous on the full product. Inequality ({A}) bounds the differences of their real-valued norms from ∣b(HB(x,t))∣|b(H_B(x,t))| by CPϵnC_P\epsilon_n for all xx and t<1t<1. Passing to the already established pointwise limits gives the same bound at t=1t=1. Thus ∣b(HB(x,t))∣|b(H_B(x,t))| is a uniform limit of continuous functions on the full product and is continuous there.

Finally polynomial representatives are dense in the quotient Banach algebra BB. Contractivity of every strict-time evaluation and of the endpoint seminorm gives ∣∣b(HB(x,t))∣−∣b′(HB(x,t))∣∣≤∥b−b′∥(0≤t≤1).\bigl||b(H_B(x,t))|-|b'(H_B(x,t))|\bigr| \leq\|b-b'\| \quad(0\leq t\leq1). Uniform polynomial approximation proves continuity for every b∈Bb\in B, hence joint continuity into its spectrum. The previously proved strict-time functorial identities pass to the endpoint by pointwise convergence. ◻

Theorem 5.21.4. NAG-002166 (Retraction to the inverse-image skeleton). For the structure map f:M(B)→Vf:\mathcal M(B)\to V, put ΣB=f−1(Σa)\Sigma_B=f^{-1}(\Sigma_a). Then HBH_B is a strong deformation retraction onto ΣB\Sigma_B. It satisfies the maximum-parameter composition law for all s,t∈[0,1]s,t\in[0,1]. Every fibre of ΣB→Σa\Sigma_B\to\Sigma_a is finite.

Proof. The fibre over any z∈Vz\in V is the affinoid spectrum of B⊗^AH(z)B\widehat\otimes_A\mathcal H(z) by Theorem 2.33.11. The equations F=0F=0, Qdet⁡J−1=0Q\det J-1=0 in the m+1m+1 variables (Y,Q)(Y,Q) have the block Jacobian (J0Q ∂(det⁡J)/∂Ydet⁡J),det⁡=(det⁡J)2∈B×.\begin{pmatrix} J&0\\ Q\,\partial(\det J)/\partial Y&\det J \end{pmatrix}, \qquad \det=(\det J)^2\in B^\times. The matrix criterion Theorem 2.41.7 therefore gives a rank-zero quasi-smooth presentation at every fibre point. Lemma 2.50.1 makes every such point isolated. The fibre is compact, so this discrete space is finite.

If x∈ΣBx\in\Sigma_B, the projected trajectory is constant because the standard homotopy fixes Σa\Sigma_a. Thus the continuous trajectory HB(x,t)H_B(x,t) lies in a finite discrete fibre; connectedness of [0,1][0,1] makes it constant. Conversely every endpoint projects to s(val⁡(f(x)))∈Σas(\operatorname{val}(f(x)))\in\Sigma_a, so the endpoint image is precisely ΣB\Sigma_B. The time-zero identity, joint continuity and skeletal fixedness prove the strong deformation retraction assertion. The composition law was proved for strict times and, when a parameter is one, in Lemma 5.18.5. The preceding fibre calculation gives the final statement. ◻

References. The finite coefficient argument uses a finitely generated ring over the image of Z\mathbb Z; it imposes no Noetherian hypothesis on K∘K^\circ. Uniform Hensel perturbation then passes from polynomial to restricted-series Jacobian equations.

5.22 Initial multiplicities and balancing

Lemma 5.22.1. NAG-002114 (Lengths on a curve and on its normalization). Let BB be a pure one-dimensional finite-type algebra over a field kk. Let s∈Bs\in B be a nonzerodivisor such that B/sBB/sB is finite-dimensional over kk. For each minimal prime PiP_i, put νi=length⁡BPi(BPi),Bi=B/Pi,\nu_i=\operatorname{length}_{B_{P_i}}(B_{P_i}),\qquad B_i=B/P_i, and let B‾i\overline B_i be the normalization of BiB_i in its fraction field. Then dim⁡k(B/sB)=∑iνi∑q∈Spec⁡B‾i closeds(q)=0[κ(q):k]ord⁡q(s).(L)\dim_k(B/sB) =\sum_i\nu_i \sum_{\substack{q\in\operatorname{Spec}\overline B_i\ \mathrm{closed}\\s(q)=0}} [\kappa(q):k]\operatorname{ord}_q(s). \tag{L}

Proof. For every finite BB-module MM, the kernel and cokernel of multiplication by ss are finite over the finite-dimensional algebra B/sBB/sB. Define χs(M)=dim⁡k(M/sM)−dim⁡kker⁡(s:M→M).\chi_s(M)=\dim_k(M/sM)-\dim_k\ker(s:M\to M). The six-term kernel-cokernel sequence for a short exact sequence shows that χs\chi_s is additive. Apply the prime filtration of Lemma 5.19.1 to BB. A factor B/PB/P with nonminimal PP has dimension zero, hence is finite over kk, and its kernel and cokernel have equal dimensions. It contributes zero. Localizing the filtration at a minimal prime PiP_i counts exactly νi\nu_i copies of its residue field. The nonzerodivisor ss avoids each minimal prime: in the finite-length local ring BPiB_{P_i} every element of its maximal ideal is nilpotent, whereas multiplication by ss remains injective. Thus dim⁡k(B/sB)=∑iνidim⁡k(Bi/sBi).\dim_k(B/sB)=\sum_i\nu_i\dim_k(B_i/sB_i).

The normalization B‾i\overline B_i is a finite BiB_i-module by [StacksAlg] (Tag 0BXR). Its quotient Qi=B‾i/BiQ_i=\overline B_i/B_i is annihilated by a nonzero element of BiB_i, obtained by clearing denominators of finitely many module generators. Hence QiQ_i has zero-dimensional support and is finite over kk, by the same prime-filtration argument. Therefore χs(Qi)=0\chi_s(Q_i)=0. Multiplication by ss is injective on BiB_i and on B‾i\overline B_i, so additivity gives dim⁡k(Bi/sBi)=dim⁡k(B‾i/sB‾i).\dim_k(B_i/sB_i)=\dim_k(\overline B_i/s\overline B_i). Every closed local ring of the normal curve Spec⁡B‾i\operatorname{Spec}\overline B_i is a discrete valuation ring by [StacksAlg] (Tag 0B8Y). Its quotient by ss has length ord⁡q(s)\operatorname{ord}_q(s), as is seen from the powers of a uniformizer. A composition series of the finite-dimensional quotient, localized at its finitely many maximal ideals, gives the residue-degree-weighted sum in ({L}). ◻

Proposition 5.22.2. NAG-002116 (A ray multiplicity is a boundary-order sum). Let kk have the trivial absolute value and let C⊂Spec⁡k[M]C\subset\operatorname{Spec}k[M] be a pure one-dimensional closed subscheme. Write CiC_i for its integral components, with generic lengths νi\nu_i, and let C‾i\overline C_i be normal proper curves with function fields k(Ci)k(C_i). At a closed point p∈C‾ip\in\overline C_i, define ai,p∈Na_{i,p}\in N by ⟨m,ai,p⟩=ord⁡p(χm).\langle m,a_{i,p}\rangle=\operatorname{ord}_p(\chi^m). If u∈Nu\in N is primitive and R≥0u\mathbf R_{\ge 0}u is a ray of Trop⁡(C)\operatorname{Trop}(C), its geometric initial multiplicity is mC(u)=∑iνi∑p∈C‾i closedai,p=ei,pu, ei,p>0[κ(p):k]ei,p.(B)m_C(u)=\sum_i\nu_i \sum_{\substack{p\in\overline C_i\ \mathrm{closed}\\a_{i,p}=e_{i,p}u,\ e_{i,p}>0}} [\kappa(p):k]e_{i,p}. \tag{B}

Proof. By Lemma 5.14.1, choose torus coordinates x1,…,xnx_1,\ldots,x_n in which u=e1u=e_1. An integral lattice change transports weighted Laurent norms and their initial ideals, so it preserves the multiplicity definition. Let I⊂k[x1±1,…,xn±1]I\subset k[x_1^{\pm 1},\ldots,x_n^{\pm 1}] define CC, and put J=I∩k[x1,x2±1,…,xn±1],B=k[x1,x2±1,…,xn±1]/J.J=I\cap k[x_1,x_2^{\pm1},\ldots,x_n^{\pm1}],\qquad B=k[x_1,x_2^{\pm1},\ldots,x_n^{\pm1}]/J. This is the coordinate algebra of the schematic closure in A1×Gmn−1\mathbf A^1\times\mathbf G_m^{n-1}. Multiplication by x1x_1 is injective, and B[x1−1]B[x_1^{-1}] is the coordinate algebra of CC. Every Laurent relation, multiplied by a suitable power of x1x_1, has lowest x1x_1-exponent zero and belongs to JJ. Its ordinary initial form in direction e1e_1 is its reduction modulo x1x_1. Conversely, every nonzero reduction of an element of JJ is such an initial form. Thus there is an equality of ideals, and hence of schemes, in⁡e1(C)≃Gm×Spec⁡(B/x1B).\operatorname{in}_{e_1}(C) \simeq\mathbf G_m\times\operatorname{Spec}(B/x_1B). The ordinary initial ideal over a trivially valued field agrees with the initial-degeneration definition after extension and translation, by the explicit coefficient comparison in Theorem 5.16.6. The initial scheme is pure of dimension one by Proposition 5.19.2; consequently B/x1BB/x_1B is zero-dimensional, hence finite by Lemma 5.19.1. Its dimension is exactly mC(u)m_C(u). The closure Spec⁡B\operatorname{Spec}B is pure one-dimensional: each of its minimal components meets x1≠0x_1\ne 0, and its function field is that of a component of CC, of transcendence degree one. Its generic lengths are the same νi\nu_i, since localization at the corresponding minimal primes is unchanged by inverting x1x_1.

For precision about the completion, take the closure of each CiC_i in (P1)n(\mathbf P^1)^n and normalize it. Normalization is finite by [StacksAlg] (Tag 0BXR). The resulting curve is proper: projective space is proper by [StacksAlg] (Tag 01WC); finite products follow from proper base change and composition, and closed subspaces from proper closed immersions [StacksAlg] (Tags 01W3–01W5). A finite morphism is proper because it is of finite type, separated, and universally closed by [StacksAlg] (Tag 01WM). Here finiteness implies integrality by the determinant trick: multiplication by an algebra element on finitely many module generators satisfies the monic determinant relation. The normal proper model of a one-variable function field is unique up to its unique function-field-preserving isomorphism by [StacksAlg] (Tag 0BY1); hence this model may be used to compute the stated orders. Normalization commutes with restriction to an open set, since integral closure commutes with localization. Indeed, clearing the finitely many denominators in a monic relation shows that an element integral after localization becomes integral before localization after multiplying by one denominator.

The normalization of each component of Spec⁡B\operatorname{Spec}B is therefore the inverse image of A1×Gmn−1\mathbf A^1\times\mathbf G_m^{n-1} in C‾i\overline C_i. Its points above x1=0x_1=0 are exactly those with ord⁡p(x1)>0,ord⁡p(xj)=0(2≤j≤n),\operatorname{ord}_p(x_1)>0,\qquad \operatorname{ord}_p(x_j)=0\quad(2\le j\le n), equivalently ai,p=ei,pua_{i,p}=e_{i,p}u with ei,p>0e_{i,p}>0. Applying Lemma 5.22.1 with s=x1s=x_1 gives ({B}), including the residue degrees and generic component lengths. ◻

Theorem 5.22.3. NAG-002118 (Balancing for a trivially valued curve). For every pure one-dimensional closed torus subscheme CC over a trivially valued field, its tropicalization with the geometric initial multiplicities is a balanced one-dimensional fan. In a pointed subdivision, ∑ρmC(ρ)uρ=0in N,\sum_{\rho}m_C(\rho)u_\rho=0\quad\text{in }N, where uρu_\rho is the primitive generator of the ray ρ\rho.

Proof. Since v(k×)=0v(k^\times)=0, all hyperplanes in the finite construction of Theorem 5.16.5 pass through zero. Theorem 5.19.5 therefore makes the tropicalization a finite pure one-dimensional fan. For every component CiC_i, only finitely many points of C‾i\overline C_i have a nonzero vector ai,pa_{i,p}, because each of the finitely many coordinate rational functions has only finitely many zeros and poles. Every such vector lies in Trop⁡(Ci)\operatorname{Trop}(C_i): the absolute value f↦exp⁡(−ord⁡p(f))f\mapsto\exp(-\operatorname{ord}_p(f)) on its function field is trivial on kk and, after completion, gives an analytic point of CiC_i with that tropical vector. Its positive rescalings give the same ray.

Conversely, Proposition 5.22.2 expresses the positive weight on each tropical ray as a sum over these points, so none is missing. For every character m∈Mm\in M, the degree of its principal divisor on each proper integral curve is zero by [StacksAlg] (Tag 02RU): ∑p[κ(p):k]⟨m,ai,p⟩=0.\sum_p[\kappa(p):k]\langle m,a_{i,p}\rangle=0. Multiply by νi\nu_i, sum over ii, and group the nonzero vectors by their primitive rays. Formula ({B}) turns the result into ⟨m,∑ρmC(ρ)uρ⟩=0\langle m,\sum_\rho m_C(\rho)u_\rho\rangle=0. Since this holds for every mm, it is the required lattice-vector equality. ◻

Proposition 5.22.4. NAG-002119 (Weighted initial stars and tangent quotients). Let XX be pure of dimension dd, let τ\tau be a cell of its initial-stratum complex of dimension ss, and take w∈relint⁡(τ)w\in\operatorname{relint}(\tau). Write in⁡w(X)=TNτ×Zτ\operatorname{in}_w(X)=T_{N_\tau}\times Z_\tau as in Proposition 5.19.3. Give its residue field the trivial valuation. Then ZτZ_\tau is pure of dimension d−sd-s, and Trop⁡(Zτ)=(Star⁡wTrop⁡(X))/Lτ.\operatorname{Trop}(Z_\tau) =\bigl(\operatorname{Star}_w\operatorname{Trop}(X)\bigr)/L_\tau. The equality preserves the weights on maximal cells, using the quotient lattice N/NτN/N_\tau.

Proof. A prime quotient of the coordinate algebra of ZτZ_\tau, with ss independent Laurent variables adjoined, has dimension increased by ss, by [StacksAlg] (Tag 00P0). Proposition 5.19.2 therefore proves the pure-dimension assertion. For every analytic point of ZτZ_\tau, its completed residue field admits Gauss extensions with arbitrary values of the additional torus coordinates. Hence the tropical support of the product is exactly Lτ×Trop⁡(Zτ)L_\tau\times\operatorname{Trop}(Z_\tau). Apply Theorem 5.16.6 to obtain the displayed support equality.

Take a direction uu whose class in NR/LτN_{\mathbf R}/L_\tau lies in the relative interior of a maximal cell of a common refinement of the quotient star and the initial complex of ZτZ_\tau. For small ε>0\varepsilon>0, Theorem 5.16.6 identifies in⁡w+εu(X)\operatorname{in}_{w+\varepsilon u}(X) with in⁡u(in⁡w(X))\operatorname{in}_u(\operatorname{in}_w(X)) after residue extension and torus translation. In the split coordinate ring, all equations for ZτZ_\tau use quotient characters only. Taking their initial forms therefore gives in⁡u(in⁡w(X))≃TNτ×in⁡uˉ(Zτ).\operatorname{in}_u(\operatorname{in}_w(X)) \simeq T_{N_\tau}\times\operatorname{in}_{\bar u}(Z_\tau). The maximal tangent lattice on the left is the inverse image in NN of the maximal tangent lattice in N/NτN/N_\tau. These are saturated lattices. Quotienting the two schemes by that maximal tangent torus gives the same finite transverse algebra after residue extension. Its dimension is unchanged. Proposition 5.19.7 gives equality of the weights; Proposition 5.19.8 removes the chosen refinements. ◻

Theorem 5.22.5. NAG-002120 (Balancing with actual initial multiplicities). Let KK be any complete real-valued non-Archimedean field, and let X⊂Spec⁡K[M]X\subset\operatorname{Spec}K[M] be a pure dd-dimensional closed subscheme. Its tropicalization, endowed with the geometric initial multiplicities of Definition 5.19.6, is a balanced integral weighted polyhedral complex. Thus at every codimension-one cell τ\tau, ∑σ⊃τmX(σ)uσ/τ=0in N/Nτ,\sum_{\sigma\supset\tau}m_X(\sigma)u_{\sigma/\tau}=0 \quad\text{in }N/N_\tau, where uσ/τu_{\sigma/\tau} denotes the primitive generator towards σ\sigma. The assertion is stable under subdivision and complete isometric base extension.

Proof. The case d=0d=0 has no balancing condition. If d>0d>0, choose w∈relint⁡(τ)w\in\operatorname{relint}(\tau). Proposition 5.22.4 expresses the initial scheme as TNτ×CT_{N_\tau}\times C with CC a pure one-dimensional closed torus subscheme over a trivially valued residue field. Its tropical rays are precisely the images of the cells σ⊃τ\sigma\supset\tau in the quotient lattice. Their primitive vectors are uσ/τu_{\sigma/\tau}, and their actual geometric weights are mX(σ)m_X(\sigma) by that proposition. Theorem 5.22.3 gives exactly the displayed equation. Refinement and base-change compatibility are those of Proposition 5.19.8. ◻

References. This is the classical reduction of tropical balancing to the divisor of a character on a proper curve. The length calculation ({L}) and the exact identity in⁡e1(C)=Gm×Spec⁡(B/x1B)\operatorname{in}_{e_1}(C)=\mathbf G_m\times\operatorname{Spec}(B/x_1B) identify the boundary-order coefficients with the scheme-theoretic initial weights.

5.23 The homotopy of a strictly semistable formal model

Throughout this section KK is complete and nontrivially valued, O=K∘O=K^\circ, and π∈O\pi\in O satisfies 0<∣π∣<10<|\pi|<1. Formal algebras carry their π\pi-adic topology. A formally étalemap below has the unique lifting property for square-zero extensions of discrete algebras annihilated by a power of π\pi.

Lemma 5.23.1. NAG-002172 (A bounded presentation for a formal étalechart). Let A0→B0A_0\to B_0 be an adic, topologically finitely presented, formally étalehomomorphism of complete admissible OO-algebras. There are polynomials F1,…,Fm∈A0[Y1,…,Ym]F_1,\ldots,F_m\in A_0[Y_1,\ldots,Y_m] and an isomorphism B0≃A0[Y1,…,Ym,Q]/(F1,…,Fm,Qdet⁡(∂F/∂Y)−1)^.(FJ)B_0\simeq \widehat{A_0[Y_1,\ldots,Y_m,Q]/ (F_1,\ldots,F_m,Q\det(\partial F/\partial Y)-1)}. \tag{FJ} Here completion is π\pi-adic. If A0A_0 is a standard semistable algebra, the generic fibre has a presentation ({J}), with all coefficients integral and all coordinates bounded by one.

Proof. Put Aˉ=A0/πA0\bar A=A_0/\pi A_0 and Bˉ=B0/πB0\bar B=B_0/\pi B_0. Topological finite presentation makes Bˉ\bar B finitely presented over Aˉ\bar A. Its unique square-zero lifting property follows from that of B0B_0: a target Aˉ\bar A-algebra is annihilated by π\pi, and every lifted A0A_0-map therefore factors through Bˉ\bar B. By [StacksAlg] (Tag 00UR), Aˉ→Bˉ\bar A\to\bar B is étale. The global presentation [StacksAlg] (Tag 00U9) gives Bˉ=Aˉ[Y1,…,Ym]/(Fˉ1,…,Fˉm),det⁡(∂Fˉ/∂Y)∈Bˉ×.\bar B=\bar A[Y_1,\ldots,Y_m]/ (\bar F_1,\ldots,\bar F_m), \qquad \det(\partial\bar F/\partial Y)\in\bar B^\times. Lift the finitely many polynomial coefficients to A0A_0, and write CC for the uncompleted algebra on the right of ({FJ}). Then C/πC≃BˉC/\pi C\simeq\bar B. We check its algebraic unique lifting property directly. Across a square-zero ideal II in a target ring, choose any lifts yy of a prescribed solution modulo II. Its determinant is a unit: a lift of a unit modulo a nilpotent ideal is a unit, by a finite geometric inverse. The error F(y)F(y) lies in ImI^m. The correction h=−F′(y)−1F(y)∈Imh=-F'(y)^{-1}F(y)\in I^m gives F(y+h)=0F(y+h)=0, since all terms quadratic in hh vanish. The same linear identity proves uniqueness; QQ must be the inverse determinant at the corrected root. Thus CC has the required unique lifting property.

The identity modulo π\pi now lifts successively to mutually inverse maps C/πnC ≃ B0/πnB0(n≥1).C/\pi^n C\ \simeq\ B_0/\pi^n B_0 \qquad(n\geq1). Indeed the kernels of the maps from level n+1n+1 to level nn are square-zero. The lifting properties of CC and B0B_0 respectively lift the two directions, and uniqueness makes both composites the identity and makes the isomorphisms compatible in nn. Taking inverse limits, and using the assumed completeness of B0B_0, proves ({FJ}). Finally Proposition 3.14.7 and Lemma 3.14.8 identify the generic algebra with the stated restricted-series quotient. Coefficients in a standard A0A_0 have integral restricted-series lifts in its defining coordinates, so this is exactly a bounded presentation of type (J). ◻

Lemma 5.23.2. NAG-002174 (Intrinsic bounds on displacement coefficients). For a bounded presentation (J), every analytic unit v∈B×v\in B^\times satisfies ∣v(HB(x,t))∣=∣v(x)∣,∣Dν(v)(x)∣≤∣v(x)∣(x∈M(B), 0≤t≤1).|v(H_B(x,t))|=|v(x)|,\qquad |D_\nu(v)(x)|\leq|v(x)| \quad(x\in\mathcal M(B),\ 0\leq t\leq1). If b∈Bb\in B satisfies ∣b(y)∣≤1|b(y)|\leq 1 at every y∈M(B)y\in\mathcal M(B), then ∣Dν(b)(x)∣≤1|D_\nu(b)(x)|\leq 1 for all ν,x\nu,x. Consequently, for 0<δ<10<\delta<1 and E=H(x)E=\mathcal H(x), the evaluated displacement has ∥σδ,x(v)v(x)−1∥δ≤δ,∥σδ,x(b)−b(x)∥δ≤δ.\left\|\frac{\sigma_{\delta,x}(v)}{v(x)}-1\right\|_\delta \leq\delta, \qquad \|\sigma_{\delta,x}(b)-b(x)\|_\delta\leq\delta.

Proof. The constant coefficient in the displacement expansion gives ∣c(HB(x,t))∣≥∣c(x)∣|c(H_B(x,t))|\geq|c(x)| for every c∈Bc\in B. Apply this to vv and v−1v^{-1}. Multiplicativity makes the product of the two left sides and the product of the two right sides both equal to one, so both inequalities are equalities. For each coefficient, ∣Dν(v)(x)∣t∣ν∣≤∣v(x)∣(0<t<1).|D_\nu(v)(x)|t^{|\nu|}\leq |v(x)|\qquad(0<t<1). Letting tt increase to one proves the coefficient bound for a unit. For the second assertion, HB(x,t)H_B(x,t) is an actual point of M(B)\mathcal M(B), so ∣b(HB(x,t))∣≤1|b(H_B(x,t))|\leq 1; the same limiting argument gives the coefficient bound. Since D0D_0 is the identity, the remaining terms have positive total degree, and their Gauss norms at radius δ\delta are at most δ\delta. These arguments use only the values of the functions on the analytic space, not a comparison of quotient norms. ◻

Theorem 5.23.3. NAG-002175 (Independence of the semistable chart). Suppose the same affinoid algebra BB has two bounded presentations of type (J), possibly over different standard semistable algebras. Their homotopies on M(B)\mathcal M(B) are equal at every time in [0,1][0,1].

Proof. Use unprimed and primed notation for the two presentations. Fix xx, put E=H(x)E=\mathcal H(x), and fix 0<δ<10<\delta<1. In E{δ−1U}E\{\delta^{-1}U\} define Φi=σδ,x(Xi′)Xi′(x)−1(1≤i≤r′),Ψj=σδ,x(Tj′)−Tj′(x).\Phi_i=\frac{\sigma_{\delta,x}(X'_i)}{X'_i(x)}-1 \quad(1\leq i\leq r'),\qquad \Psi_j=\sigma_{\delta,x}(T'_j)-T'_j(x). The functions Xi′X'_i are global analytic units because their product is the nonzero scalar a′a'. The Tj′T'_j and the primed implicit coordinates have values at most one everywhere. Lemma 5.23.2 therefore gives ∥Φi∥,∥Ψj∥≤δ\|\Phi_i\|,\|\Psi_j\|\leq\delta. Substitution U′=(S′,Z′)↦(Φ,Ψ)U'=(S',Z')\mapsto(\Phi,\Psi) defines a contractive EE-homomorphism θ:E{δ−1U′}⟶E{δ−1U}.\theta:E\{\delta^{-1}U'\}\longrightarrow E\{\delta^{-1}U\}. The two maps σδ,x\sigma_{\delta,x} and θσδ,x′\theta\sigma'_{\delta,x} have identical images on the primed base coordinates: this is the definition for Xi′X'_i with i≥1i\geq 1 and for Tj′T'_j, and the product relation gives it for X0′X'_0. Their images of the primed implicit tuple solve the same displaced equations and both lie within δ\delta of Y′(x)Y'(x). At that center the new Jacobian differs by at most δ\delta from the original invertible Jacobian of inverse norm at most one. The geometric inverse formula and Lemma 5.18.1 therefore identify the two roots. Their inverse determinants coincide as well.

The primed presentation topologically generates BB. Both maps are bounded for its topology, so equality on its coordinates implies equality on every b∈Bb\in B. Contractivity of θ\theta now gives ∥σδ,x(b)∥δ≤∥σδ,x′(b)∥δ.\|\sigma_{\delta,x}(b)\|_\delta \leq\|\sigma'_{\delta,x}(b)\|_\delta. Reversing the two presentations gives the opposite inequality. These Gauss norms are precisely evaluation at the two time-δ\delta points, so the points agree. They agree at zero by definition and at one by the pointwise endpoint formula. ◻

Proposition 5.23.4. NAG-002176 (Restriction to formal open subsets). Let U=Spf⁡(B0)\mathfrak U=\operatorname{Spf}(B_0) admit a topologically finitely presented formal étalemap to a standard semistable formal chart. Its homotopy preserves the generic fibre of every formal open subset. On an admissible affine formal open V⊆U\mathfrak V\subseteq\mathfrak U, its restriction equals the homotopy obtained from the restricted chart, independently of its bounded presentation.

Proof. Choose a bounded presentation by Lemma 5.23.1. A principal formal open D(g)D(g), g∈B0g\in B_0, has generic fibre ∣g∣=1|g|=1 by Proposition 3.17.1. The integral function gg has values at most one. The constant coefficient of its displacement gives ∣g(HB(x,t))∣≥∣g(x)∣|g(H_B(x,t))|\geq|g(x)|, so a point where ∣g∣=1|g|=1 stays in that domain at every time. General formal opens are unions of principal formal opens, and their generic fibres are the corresponding unions by Proposition 3.17.2; hence they are preserved as well.

Write V=Spf⁡(C0)\mathfrak V=\operatorname{Spf}(C_0). An open immersion is formally étale: a square-zero thickening does not change underlying points, so a map whose reduction factors through that open factors through it uniquely. Composing unique lifting properties shows that the restricted map to the standard base is formally étale. We also verify relative finite presentation. By Theorem 3.14.4, choose finite presentations B0=O⟨W⟩/IB_0=O\langle W\rangle/I and C0=O⟨Z⟩/JC_0=O\langle Z\rangle/J. Lift the images of the finitely many WiW_i to integral restricted series hi(Z)h_i(Z). The two substitution maps give the presentation C0≃B0⟨Z⟩/(J(Z), Wi−hi(Z):i).C_0\simeq B_0\langle Z\rangle/ (J(Z),\,W_i-h_i(Z):i). The relations are finite; the same construction applies directly to the map from the standard base to C0C_0. Thus Lemma 5.23.1 applies over that original base. The integral map B0→C0B_0\to C_0 gives a contractive map on the displayed generic algebras: each integral presentation coordinate of BB maps into C0C_0, whose elements have norm at most one by Proposition 3.14.7; bounded substitution and quotient infima give contractivity. Thus Proposition 5.18.4 and Theorem 5.21.3 identify the restricted homotopy with that on Vη\mathfrak V_\eta. Theorem 5.23.3 removes the chosen presentation. ◻

Theorem 5.23.5. NAG-002177 (Global semistable deformation). Let X\mathfrak X be a quasi-compact admissible formal OO-scheme with Hausdorff generic fibre XX. Suppose it has a finite affine formal open cover Ui\mathfrak U_i with topologically finitely presented formal étalemaps fi:Ui⟶Spf⁡O⟨Xi0,…,Xiri,Ti1,…,Tiqi⟩/(Xi0⋯Xiri−ai),0<∣ai∣<1.\begin{gathered} f_i:\mathfrak U_i\longrightarrow \operatorname{Spf} O\langle X_{i0},\ldots,X_{ir_i},T_{i1},\ldots,T_{iq_i}\rangle/ (X_{i0}\cdots X_{ir_i}-a_i),\\ 0<|a_i|<1. \end{gathered} There is a canonical jointly continuous homotopy HX:X×[0,1]→XH_{\mathfrak X}:X\times[0,1]\to X whose restriction to each Ui=(Ui)ηU_i=(\mathfrak U_i)_\eta is the constructed chart homotopy. It is a strong deformation retraction onto the closed compact subset Σ(X)=HX(X,1),Σ(X)∩Ui=(fi)η−1(Σai).\Sigma(\mathfrak X)=H_{\mathfrak X}(X,1),\qquad \Sigma(\mathfrak X)\cap U_i=(f_i)_\eta^{-1}(\Sigma_{a_i}). It satisfies the maximum composition law, is independent of the affine cover and the semistable coordinate maps, and restricts to the corresponding homotopy on every quasi-compact formal open. Under a complete isometric extension L/KL/K, the projection XL→XX_L\to X commutes with these homotopies.

Proof. The preceding presentation lemma and Theorem 5.21.4 give a jointly continuous strong deformation HiH_i on each UiU_i. On an overlap of two formal charts, choose a principal formal open of the first chart contained in the second. It is admissible by Proposition 3.17.1. Its generic fibre is preserved by both homotopies, and Proposition 5.23.4 identifies their restrictions with the two homotopies of that affine algebra. Those homotopies coincide by Theorem 5.23.3. Such formal affine opens cover the overlap, and their generic fibres cover the analytic overlap by Proposition 3.17.2. Thus all HiH_i agree on overlaps.

Each UiU_i is compact and hence closed in the Hausdorff space XX, and the finite family covers XX. The product sets Ui×[0,1]U_i\times[0,1] consequently form a finite closed cover. The maps HiH_i glue to a continuous map: the inverse image of any closed subset of XX is the finite union of its closed inverse images in these closed product sets. All homotopy and composition identities hold on each preserved UiU_i, hence globally.

A point in UiU_i is fixed by the global endpoint exactly when it is fixed by the local endpoint, namely when it lies over Σai\Sigma_{a_i}. The endpoint is idempotent, so its image is precisely its fixed set. Since XX is a finite union of compact sets, it is compact; its endpoint image is compact and hence closed. This proves the displayed local description and strong deformation assertion. A second formal atlas has common affine formal refinements with the first one; the same overlap argument proves independence. It also proves restriction to a quasi-compact formal open. Finally, base change of a chart retains its integral Jacobian presentation. The local scalar-extension compatibility of Theorem 5.21.3 therefore gives the asserted projection identity on the cover, and hence globally. ◻

The faces of an étalesemistable chart

Let A0=O⟨X0,…,Xr,T1,…,Tq⟩/(X0⋯Xr−a)A_0=O\langle X_0,\ldots,X_r,T_1,\ldots,T_q\rangle/ (X_0\cdots X_r-a), where 0<∣a∣<10<|a|<1, and let A0→B0A_0\to B_0 be a topologically finitely presented formally étalemap of admissible algebras. Write f:X=M(B)→M(A)f:X=\mathcal M(B) \to\mathcal M(A) for the generic map, fˉ\bar f for the special-fibre map, and k=K~k=\widetilde K. For a nonempty subset I⊆{0,…,r}I\subseteq\{0,\ldots,r\} put ΔI={w∈Δa:wj=0 for j∉I},DI={Xi=0 (i∈I),Xj≠0 (j∉I)}⊆Spec⁡(A0⊗Ok),SI=fˉ−1(DI).\begin{aligned} \Delta_I&=\{w\in\Delta_a:w_j=0\text{ for }j\notin I\},\\ D_I&=\{X_i=0\ (i\in I),\quad X_j\ne0\ (j\notin I)\} \subseteq\operatorname{Spec}(A_0\otimes_O k),\\ S_I&=\bar f^{-1}(D_I). \end{aligned} The stratum DID_I is the integral smooth variety with coordinates Xj±1X_j^{\pm 1} for j∉Ij\notin I and T1,…,TqT_1,\ldots,T_q. Denote its generic point by ξI\xi_I and the finite set of generic points of SIS_I by QIQ_I. The empty inverse image has QI=∅Q_I=\varnothing.

Lemma 5.24.1. NAG-002211 (The residue field over a relative open face). Let w∈relint⁡(ΔI)w\in\operatorname{relint}(\Delta_I) and let zw∈Σaz_w\in\Sigma_a be its standard Gauss point. Specialization sends zwz_w to ξI\xi_I, and the induced extension FI:=κ(ξI)=k(Xj:j∉I,T1,…,Tq) ⟶ hw:=H(zw)~F_I:=\kappa(\xi_I)=k(X_j:j\notin I,T_1,\ldots,T_q) \ \longrightarrow\ h_w:=\widetilde{\mathcal H(z_w)} is purely transcendental. In particular, for every finite extension l/FIl/F_I, the tensor product l⊗FIhwl\otimes_{F_I}h_w is a field, of degree [l:FI][l:F_I] over hwh_w.

Proof. Choose i0∈Ii_0\in I and eliminate Xi0X_{i_0} using the product relation. The independent Gauss variables are the remaining XiX_i and the TjT_j. The variables XjX_j with j∉Ij\notin I and all the TjT_j have weight zero; all other original XiX_i have strictly positive weight. Their residues therefore give exactly the generic point ξI\xi_I: every nonzero polynomial in the zero-weight variables has Gauss norm one after normalizing an integral coefficient.

In the residue-lattice proof of Proposition 3.8.1, let LwL_w be the lattice of exponent vectors whose weight belongs to v(K×)v(K^\times). The subgroup generated by the zero-weight coordinate vectors is a direct summand of the full exponent lattice, hence is saturated in LwL_w. Lemma 5.14.1 extends its displayed coordinate basis to a basis of LwL_w. The Gauss residue formula with this basis expresses hwh_w as the rational function field over kk in the residues of the zero-weight variables and finitely many additional normalized Laurent monomials. It is therefore purely transcendental over the indicated subfield FIF_I.

Finally, write hw=FI(U1,…,Us)h_w=F_I(U_1,\ldots,U_s). The ring l⊗FIhwl\otimes_{F_I}h_w is the localization of l[U1,…,Us]l[U_1,\ldots,U_s] by the nonzero polynomials over FIF_I. It is an integral domain and is finite-dimensional over hwh_w, hence is a field: multiplication by a nonzero element is an injective endomorphism of a finite-dimensional vector space and is therefore surjective. Tensoring a basis of l/FIl/F_I gives its stated degree. ◻

Proposition 5.24.2. NAG-002212 (Actual skeletal fibres and residual strata). For w∈relint⁡(ΔI)w\in\operatorname{relint}(\Delta_I), specialization induces a bijection f−1(zw) → ∼  QI.f^{-1}(z_w)\ \xrightarrow{\ \sim\ }\ Q_I. For the point xη,wx_{\eta,w} labelled by η∈QI\eta\in Q_I, put Ew=H(zw)E_w=\mathcal H(z_w) and lη=κ(η)l_\eta=\kappa(\eta). Then H(xη,w)/Ew\mathcal H(x_{\eta,w})/E_w is the unramified finite extension lifting the finite separable residue extension lη⊗FIhw / hw.l_\eta\otimes_{F_I}h_w\,/\,h_w. Its degree is [lη:FI][l_\eta:F_I], and its value group is that of EwE_w.

Proof. Use Lemma 5.23.1 to write B0=C^B_0=\widehat C with C=A0[Y,Q]/(F,Qdet⁡J−1)C=A_0[Y,Q]/(F,Q\det J-1). The square-zero correction in that proof and [StacksAlg] (Tag 00UR) show that CC is an étaleA0A_0-algebra. Base change along the bounded evaluation A0→Ew∘A_0\to E_w^\circ gives an étale algebra CEC_E over Ew∘E_w^\circ by [StacksAlg] (Tag 00U2, part (3)). It is flat by part (6) of the same lemma, so multiplication by π\pi is injective. Its π\pi-adic completion D0D_0 is an admissible Ew∘E_w^\circ-algebra by Lemma 3.14.8 and Theorem 3.14.4. The explicit restricted presentation gives D=D0[1/π]=B⊗^AEw,D0/Ew∘∘D0=(B0⊗Ok)⊗A0⊗Okhw.\begin{aligned} D&=D_0[1/\pi] =B\widehat\otimes_A E_w,\\ D_0/E_w^{\circ\circ}D_0 &=(B_0\otimes_O k)\otimes_{A_0\otimes_O k}h_w. \end{aligned} For the second identity one simply reduces the finite list of equations in the restricted presentation: reduction of an integral restricted series is a polynomial, and reduction of the ideal is generated by the reductions of its displayed generators. The first identity is the completed-tensor fibre presentation of Theorem 2.33.11.

The special-fibre map is étale, and its fibre over ξI\xi_I is a finite disjoint union ∐η∈QISpec⁡(lη)\coprod_{\eta\in Q_I}\operatorname{Spec}(l_\eta) with each lη/FIl_\eta/F_I finite separable, by [StacksAlg] (Tag 02GL). Here the fibre points are exactly the generic points of SIS_I. Indeed SIS_I is smooth over kk and hence normal by [StacksAlg] (Tags 036U, 056T); flat going down, [StacksAlg] (Tag 00HS), makes every component dominate the integral DID_I, and the zero-dimensional generic fibre has no proper chains of primes. The preceding lemma now identifies the special fibre of D0D_0 with ∏η∈QI(lη⊗FIhw),\prod_{\eta\in Q_I} (l_\eta\otimes_{F_I}h_w), a finite product of fields. This is in particular reduced, so Lemma 4.12.1 identifies it with the canonical reduction of DD. Every prime of this reduction is minimal. Theorem 3.16.7 therefore gives one and only one actual analytic point over each factor, and these exhaust the fibre. If the product is empty, the integral quotient norm would have 11 of norm less than one unless D=0D=0; thus the analytic fibre is empty in exactly that case. This proves the bijection and its specialization labels.

For the valued-field assertion, fix a factor λ=lη⊗FIhw\lambda=l_\eta\otimes_{F_I}h_w and apply Lemma 4.15.2 over the complete field EwE_w. It supplies a finite complete field L/EwL/E_w of degree [λ:hw][\lambda:h_w], residue field λ\lambda, and unchanged value group. Lift the coordinates of the corresponding special-fibre point arbitrarily into L∘L^\circ. The finitely many equation errors have norms strictly below one and the Jacobian determinant has norm one. Lemma 5.18.1 corrects them to an exact bounded solution in LL. This gives a point of M(D)\mathcal M(D) with that residue label, hence the point already identified, and an isometric embedding H(xη,w)↪L.\mathcal H(x_{\eta,w})\hookrightarrow L. The embedding first holds for the fraction field of the evaluated affinoid algebra and then for its completion. Its residue field contains λ\lambda. Lifts of a basis of λ/hw\lambda/h_w are EwE_w-linearly independent: normalize any putative relation by a coefficient of largest absolute value and reduce. Hence the left field has degree at least [λ:hw][\lambda:h_w], while the embedding into LL gives the opposite inequality. It equals LL, proving the assertion. ◻

NAG-002213 · Open at full scope

Lemma 5.24.3. NAG-002213 (The unique stratum along a face). Let η∈QI\eta\in Q_I and let ∅≠J⊆I\varnothing\ne J\subseteq I. There is exactly one ηJ∈QJ\eta_J\in Q_J whose closure contains η\eta. For J′⊆JJ'\subseteq J, the generization of ηJ\eta_J labelled by J′J' is ηJ′\eta_{J'}.

Proof. The closed subscheme ZJZ_J of the standard special fibre cut out by Xj=0X_j=0 for j∈Jj\in J is an affine space: at least one factor of the product relation is zero, so the remaining coordinates are free. Its pullback YJY_J is étaleover ZJZ_J, hence smooth and normal by [StacksAlg] (Tags 036U, 056T). Its local ring at η\eta is therefore a domain. Flat going down for fˉ\bar f gives a prime below η\eta lying over the generic point ξJ\xi_J of ZJZ_J. It lies in SJS_J. Every such prime is a component generic point of YJY_J: going down forces minimal primes below it to lie over the generic point of ZJZ_J, and the étalegeneric fibre is a finite disjoint union of fields. The domain OYJ,η\mathcal O_{Y_J,\eta} has only one minimal prime, so there is precisely one such component through η\eta. This proves existence and uniqueness. If J′⊆JJ'\subseteq J, the successive generization lies over ξJ′\xi_{J'} and its closure contains η\eta; uniqueness identifies it with ηJ′\eta_{J'}. ◻

Nonempty face generizations. In the formally étalesemistable chart above, for each η∈QI\eta\in Q_I and ∅≠J⊆I\varnothing\ne J\subseteq I there is exactly one ηJ∈QJ\eta_J\in Q_J whose closure contains η\eta. For ∅≠J′⊆J\varnothing\ne J'\subseteq J, the J′J'-generization of ηJ\eta_J is ηJ′\eta_{J'}.

Here is the proof on this range. The formal chart is topologically finitely presented. Reduction modulo π\pi has finite presentation and inherits the unique square-zero lifting property from discrete tests killed by π\pi. It is étaleby [StacksAlg] (Tag 00UR). Base change to kk and to the actual coordinate face ZJZ_J preserves étaleness, and the resulting map YJ→ZJY_J\to Z_J is flat [StacksAlg] (Tag 00U2(3), (6)). As JJ is nonempty, ZJZ_J is an affine space over kk. In the local étalepresentation of YJY_J over that polynomial base, the square Jacobian is invertible [StacksAlg] (Tag 00U9). The same minor in the presentation over kk proves smoothness [StacksAlg] (Tag 00T7(1)). Thus the relevant local ring of YJY_J at η\eta is a normal domain [StacksAlg] (Tag 056T).

Flat going down gives a prime below η\eta over the generic point of ZJZ_J [StacksAlg] (Tag 00HS). Every such prime is minimal in YJY_J. Indeed any prime below it still lies over that generic point, whose étalefibre is a finite disjoint union of spectra of fields [StacksAlg] (Tag 02GL(2)). The local domain has one minimal prime, giving the required unique label. For nonempty J′J', the successive generization over the generic point of ZJ′Z_{J'} specializes to η\eta and hence is the unique ηJ′\eta_{J'}.

Theorem 5.24.4. NAG-002214 (A finite simplex realization of the chart skeleton). For each η∈QI\eta\in Q_I, the points xη,wx_{\eta,w} on relint⁡(ΔI)\operatorname{relint}(\Delta_I) extend uniquely to a continuous map sη:ΔI⟶X,sη(w)=xηJ,wif w∈relint⁡(ΔJ).s_\eta:\Delta_I\longrightarrow X, \qquad s_\eta(w)=x_{\eta_J,w} \quad\text{if }w\in\operatorname{relint}(\Delta_J). It is a homeomorphism onto its image. The finite collection of these closed simplices covers ΣB=f−1(Σa)\Sigma_B=f^{-1}(\Sigma_a), and the restriction to a face ΔJ\Delta_J is sηJs_{\eta_J}. More precisely, ΣB\Sigma_B is homeomorphic to the quotient of the finite disjoint union of the ΔI\Delta_I, one copy for each η∈QI\eta\in Q_I, by exactly these face identifications. In these charts all face maps are the coordinate face inclusions, integral affine for the lattices NI={(ni)i∈I∈ZI:∑ini=0}.N_I=\{(n_i)_{i\in I}\in\mathbb Z^I: \textstyle\sum_i n_i=0\}. The intersection with the generic fibre of a formal open subset is a union of closed faces in this finite realization.

Proof. First work over the relative open face. The primes in QIQ_I are distinct and incomparable. Choose an integral function g∈B0g\in B_0 whose reduction is outside η\eta and inside every other member of QIQ_I, by choosing one element for each other prime and multiplying. If there is no other prime, take g=1g=1. By the actual fibre description, the locus ∣g∣=1|g|=1 selects exactly xη,wx_{\eta,w} over each ww in this open face. For any compact subset W⊂relint⁡(ΔI)W\subset\operatorname{relint}(\Delta_I), the set {x∈f−1({zw:w∈W}):∣g(x)∣=1}\{x\in f^{-1}(\{z_w:w\in W\}):|g(x)|=1\} is closed in the compact affinoid space XX, and maps continuously and bijectively onto WW. It is therefore homeomorphic to WW. Every point of the relative open face has a compact relative neighborhood contained in that face, proving continuity of w↦xη,ww\mapsto x_{\eta,w} throughout the relative open face.

Consider the closure of its graph in the compact space ΔI×X\Delta_I\times X. Let (w,x)(w,x) be a point of that closure and let J={j:wj>0}J=\{j:w_j>0\}. Continuity of ff gives f(x)=zwf(x)=z_w, so the actual fibre theorem labels xx by some θ∈QJ\theta\in Q_J. If an integral function hh has reduction outside η\eta, then ∣h(xη,u)∣=1|h(x_{\eta,u})|=1 at every interior uu. Continuity makes ∣h(x)∣=1|h(x)|=1. Thus the reduction prime θ\theta is contained in η\eta. Since ∑iwi=v(a)>0\sum_i w_i=v(a)>0, JJ is nonempty. By nonempty face generization, θ=ηJ\theta=\eta_J. Every w∈ΔIw\in\Delta_I is a limit of interior points; compactness supplies a graph limit above it. We have proved that this limit is unique. The closed graph therefore projects continuously and bijectively onto the compact Hausdorff simplex ΔI\Delta_I, hence homeomorphically. Its second projection is the desired continuous extension. Its composition with the base coordinate map is the identity, so it is injective and is a homeomorphism onto its compact image.

The boundary formula and transitivity for nonempty faces prove the face restriction assertion. There are finitely many II and finitely many η∈QI\eta\in Q_I, and the fibre theorem shows that their relative interiors partition ΣB\Sigma_B. Two closed cells meet at a point of support JJ exactly when their JJ-generizations agree; hence their intersection is a union of the indicated common closed faces. The induced map from the finite face quotient to ΣB\Sigma_B is therefore a continuous bijection. Its source is compact and its target Hausdorff, so it is a homeomorphism. Coordinate inclusions preserve the displayed integer lattices, giving the final affine assertion.

On a relative open cell the specialization label is constant. If this label belongs to a given formal open, all its generizations belong to that open as well. Consequently inclusion of an open cell includes its whole closed simplex. The specialization description of formal-open generic fibres then gives the asserted union of closed faces. ◻

5.25 Integral-affine coordinates on the skeleton

Lemma 5.25.1. NAG-002215 (Units on an open polyannulus). Let LL be a complete non-Archimedean field and let Ω⊂Rd\Omega\subset\mathbf R^d be a nonempty convex open set. Let PΩ⊂(Gm,Ld)anP_\Omega\subset(\mathbf G_{m,L}^d)^{\mathrm{an}} be the analytic domain defined by v(Z)∈Ωv(Z)\in\Omega, and denote its Gauss section by w↦zww\mapsto z_w. For every analytic unit g∈O(PΩ)×g\in\mathcal O(P_\Omega)^\times there are c∈v(L×)c\in v(L^\times) and m∈Zdm\in\mathbf Z^d such that v(g(zw))=c+⟨m,w⟩(w∈Ω).v(g(z_w))=c+\langle m,w\rangle \qquad(w\in\Omega). The assertion includes d=0d=0, when PΩ=M(L)P_\Omega=\mathcal M(L).

Proof. Suppose first that d>0d>0. On the inverse image of a closed rectangular box W=∏j[αj,βj]W=\prod_j[\alpha_j,\beta_j] contained in Ω\Omega, the arbitrary-radius domain presentations of Theorem 2.31.2 give the product of closed annulus algebras; their actual analytic sections are these algebras by Theorem 2.39.4. Iterate Theorem 2.13.2, with the other coordinates as Banach coefficients. This gives a unique Laurent expansion ∑m∈ZdcmZm\sum_{m\in\mathbf Z^d}c_mZ^m, and the norm is max⁡m∣cm∣max⁡w∈Wexp⁡(−⟨m,w⟩).\max_m |c_m|\max_{w\in W}\exp(-\langle m,w\rangle). The weighted coefficients tend to zero outside finite subsets. To check the multivariable assertion directly, apply the one-coordinate normal-form operators successively to polynomial families. Their bounds and commuting actions on different exponent pairs extend them to the complete weighted coefficient space. Their composite kernel is the sum of the coordinate relations, and their common section is the Laurent monomial representative. The finite-tail condition is exactly the weighted c0c_0 condition of Lemma 2.13.1. Thus this iteration also proves uniqueness and all restriction formulas.

Use boxes with nonempty interior. On overlapping interiors, uniqueness makes their coefficients equal. Every two such boxes can be connected by a finite chain of overlapping smaller boxes inside the convex set Ω\Omega. Hence a function on PΩP_\Omega has a single Laurent coefficient family, converging on every compact box, and log⁡∣g(zw)∣=max⁡m{log⁡∣cm∣−⟨m,w⟩}.\log|g(z_w)|=\max_m\{\log|c_m|-\langle m,w\rangle\}. This is convex in ww. Apply the same argument to g−1g^{-1}. Multiplicativity of the Gauss seminorm says that its logarithm is the negative of the displayed function. A function both convex and concave on an open convex set is affine: equality in the two convexity inequalities gives its affine interpolation on every line segment. At an interior point w0w_0, a nonzero coefficient cmc_m attains the maximum. The difference between the affine function and this one affine term is nonnegative on Ω\Omega and zero at w0w_0; an affine function with these properties is zero, by varying a small distance in both directions along every vector. Thus v(g(zw))=v(cm)+⟨m,w⟩v(g(z_w))=v(c_m)+\langle m,w\rangle. If d=0d=0, the unit is an element of L×L^\times, which proves the same claim. ◻

Lemma 5.25.2. NAG-002216 (A rational-origin tube over a semistable chart). Let LL be complete and nontrivially valued, and let 0<∣b∣<10<|b|<1. Put D0=L∘⟨Z0,…,Zd⟩/(Z0⋯Zd−b).D_0=L^\circ\langle Z_0,\ldots,Z_d\rangle/ (Z_0\cdots Z_d-b). Let D0→E0D_0\to E_0 be a topologically finitely presented formally étalemap of admissible algebras, and let qq be a rational point of its special fibre lying over Z0=⋯=Zd=0Z_0=\cdots=Z_d=0. The generic map induces an analytic isomorphism ]q[ → ∼ Pb:={z∈M(D):∣Zi(z)∣<1 for every i}.]q[\ \xrightarrow{\ \sim\ } P_b:=\{z\in\mathcal M(D):|Z_i(z)|<1\text{ for every }i\}. After eliminating Z0Z_0, this is the open polyannulus with logarithmic domain wj>0w_j>0 and ∑j=1dwj<v(b)\sum_{j=1}^d w_j<v(b).

Proof. Choose the bounded presentation of Lemma 5.23.1, with coordinates Y,QY,Q, equations F(Y)=0F(Y)=0 and Qdet⁡F′(Y)=1Q\det F'(Y)=1. Lift the residue coordinates of qq to constants y0,q0y_0,q_0 in L∘L^\circ. Choose integral restricted-series lifts of all the coefficients in the variables Z0,…,ZdZ_0,\ldots,Z_d. Their evaluations at Z=0,Y=y0Z=0,Y=y_0 have FF-error of norm less than one, and the Jacobian determinant has norm one. These assertions follow by reduction at the specified rational point qq.

Let WW be a compact logarithmic box inside the displayed open simplex, with affinoid inverse image PWP_W. All ∣Zi∣|Z_i| are at most some ρ<1\rho<1 there. Every nonconstant monomial in their integral series has norm at most ρ\rho. Thus, in the Banach algebra O(PW)\mathcal O(P_W), the error F(y0)F(y_0) has norm at most some δ<1\delta<1, and F′(y0)F'(y_0) has inverse of norm at most one. Indeed its determinant differs by a term of norm less than one from its constant unit, so the geometric inverse applies. Lemma 5.18.1 produces the unique root yWy_W with ∣yW−y0∣≤δ|y_W-y_0|\leq\delta; set qW=(det⁡F′(yW))−1q_W=(\det F'(y_W))^{-1}. This gives a bounded analytic section PW→M(E)P_W\to\mathcal M(E), specializing to qq. Uniqueness makes the sections agree on overlaps, so they glue on all of PbP_b. When d=0d=0 use the same argument in LL, where Z0=bZ_0=b.

We check the inverse as an analytic map. Over PWP_W, cover the tube by affinoid subdomains on which ∣Y−y0∣≤λ<1|Y-y_0|\leq\lambda<1, increasing λ\lambda if necessary so that λ≥δ\lambda\geq\delta; one may impose the analogous bound on Q−q0Q-q_0. The original coordinate tuple and the pullback of yWy_W are roots of the same system in the Banach algebra of such a domain. Both lie within a radius less than one of y0y_0, where the same Jacobian inverse estimate holds. Hensel uniqueness makes them equal as algebra elements, and then the inverse-determinant equation identifies QQ. Consequently the section and the generic map are inverse on these subdomains. Every point of the tube lies in one of them, since its residue coordinates are those of qq. This proves the analytic isomorphism, including its compatibility for every compact box in the open simplex. ◻

Theorem 5.25.3. NAG-002217 (Units along an étaleskeletal simplex). Use the notation and hypotheses of Theorem 5.24.4. Put Γ=v(K×)\Gamma=v(K^\times). For every η∈QI\eta\in Q_I and every g∈B×g\in B^\times, the function w⟼v(g(xη,w))(w∈ΔI)w\longmapsto v(g(x_{\eta,w})) \qquad(w\in\Delta_I) is Γ\Gamma-integral-affine: after eliminating any one coordinate in II, it has an integral linear part and a constant term in Γ\Gamma.

Proof. Let H0H_0 be the completion of K(Xj:j∉I,T1,…,Tq)K(X_j:j\notin I,T_1,\ldots,T_q) in the unit Gauss norm. Its residue field is FIF_I, by Proposition 3.8.1, and its value group is ∣K×∣|K^\times|: a nonzero polynomial attains its Gauss norm at a coefficient, fractions have the same property for their values, and completion does not change the value group since a nonzero convergent limit has the same norm as every sufficiently close term.

Send the zero-weight variables to these fixed elements of H0∘H_0^\circ. The remaining formal base is H0∘⟨Xi:i∈I⟩/(∏i∈IXi−b),b=a(∏j∉IXj)−1,∣b∣=∣a∣.H_0^\circ\langle X_i:i\in I\rangle/ \left(\prod_{i\in I}X_i-b\right), \qquad b=a\left(\prod_{j\notin I}X_j\right)^{-1}, \qquad |b|=|a|. Base-change the algebraic square-Jacobian presentation of Lemma 5.23.1 and then complete. The result is an admissible formally étalealgebra over this base. The justification is the same explicit completion argument in Proposition 5.24.2: the algebraic presentation is flat, so it has no scalar torsion; Lemma 3.14.8 preserves this under completion, and Theorem 3.14.4 gives the admissible finite presentation.

Above the all-zero special point this model has the residual point with field lη=κ(η)l_\eta=\kappa(\eta), a finite separable extension of FIF_I. Lemma 4.15.2 supplies one finite extension L0/H0L_0/H_0 with residue field lηl_\eta and unchanged value group. Base-change the presentation once more to L0∘L_0^\circ. The multiplication map lη⊗FIlη→lηl_\eta\otimes_{F_I}l_\eta\to l_\eta chooses a rational residual point qq above the origin. Thus all fields, models, and the point qq have been fixed independently of ww.

Lemma 5.25.2 now gives an analytic map from the entire open simplex polyannulus over L0L_0 to XX, by its inverse section followed by the two base-change projections. Its Gauss point with logarithms ww maps to zwz_w on the original standard base. Indeed, after eliminating one XiX_i with i∈Ii\in I, the coefficient Gauss norm in the zero-weight variables followed by the Gauss norm in the remaining positive-weight variables is exactly the original Gauss norm. This identity first holds for Laurent polynomials and then for all functions by bounded continuity. The specialization of the image is η\eta, by the chosen multiplication map of residue fields. Proposition 5.24.2 therefore identifies this image with the actual point xη,wx_{\eta,w}.

The pullback of gg is an analytic unit on this single open polyannulus. Lemma 5.25.1 gives the required integral-affine formula on relint⁡(ΔI)\operatorname{relint}(\Delta_I), with constant term in v(L0×)=Γv(L_0^\times)=\Gamma. The closed-cell section is continuous by Theorem 5.24.4, and gg is nowhere zero. Hence both sides extend continuously and agree on all of ΔI\Delta_I. The proof also covers a zero-dimensional face. ◻

Theorem 5.25.4. NAG-002218 (The integral-affine skeleton of a formal model). Let X\mathfrak X satisfy Theorem 5.23.5. Its canonical fixed set Σ(X)\Sigma(\mathfrak X) is a finite Γ\Gamma-integral-affine face-glued cell complex of closed simplices. Distinct cells may have the same vertex set. Each affine formal semistable chart supplies exactly the cells of Theorem 5.24.4. The cell structure and its integral-affine transition maps are independent of the chosen formal chart atlas. The homotopy of Theorem 5.23.5 is a jointly continuous strong deformation retraction onto this finite complex.

Proof. Write sp⁡\operatorname{sp} for specialization to Xs\mathfrak X_s. In one chart, the relative interior of the cell labelled by η\eta is precisely Eη={x∈Σ(X):sp⁡(x)=η}.E_\eta=\{x\in\Sigma(\mathfrak X): \operatorname{sp}(x)=\eta\}. This is the actual-fibre classification in Proposition 5.24.2. There is no restriction to the chart on the right: the inverse image under specialization of its formal open set is its generic domain, by Proposition 3.17.2. Thus whenever a second chart contains η\eta, the same set EηE_\eta is one of its relative open cells. For clarity, choose x∈Eηx\in E_\eta. The canonical fixed-set identity makes its image in the second standard base a Gauss point. Its specialization is the generic point of the unique coordinate stratum whose vanishing coordinates are those belonging to η\eta. The actual-fibre classification therefore puts η\eta in that chart’s corresponding QIQ_I. The same classification identifies its whole relative open cell with EηE_\eta. This proves intrinsically which residual generic points label cells, without choosing a global ordering of the special-fibre branches. Its closure in the global Hausdorff generic fibre is the closed simplex supplied by either chart, since each generic affinoid chart is compact and closed. Consequently the cells themselves are independent of charts.

Compare two coordinate systems on such a cell. Choose a common principal formal neighbourhood of η\eta in their overlap. Its generic domain contains the entire closed cell: it contains EηE_\eta and is closed, or equivalently one may use the formal-open subcomplex assertion of Theorem 5.24.4. It is admissible by Proposition 3.17.1, and its map to either standard base has a bounded presentation by Proposition 5.23.4. Every monomial coordinate of the second chart is an analytic unit on this common domain, since the product of those coordinates is a nonzero scalar. Theorem 5.25.3 therefore makes each of its logarithmic coordinates a Γ\Gamma-integral-affine function of the first ones.

Reverse the two charts to obtain the same assertion for the inverse coordinate change. The linear parts are mutually inverse maps of the integer tangent lattices NI={(ni)∈ZI:∑ni=0}N_I=\{(n_i)\in\mathbf Z^I:\sum n_i=0\}. They are therefore unimodular isomorphisms. Indeed, the coordinate compositions are identities on a relative open simplex, so their derivatives are inverse real linear maps; both preserve the displayed lattices. The affine constants belong to Γ\Gamma, and all face maps are the already proved coordinate face inclusions.

There are finitely many cells because the given formal atlas and the cell lists in its charts are finite. Their intersections are unions of common faces. To see this for cells in different charts, first intersect the second cell with the first formal open set. The formal-open subcomplex assertion expresses this intersection as a finite union of its closed faces. Each such face now lies in the first chart, where the common-face property has already been proved. These faces are the same global cells by the intrinsic description EηE_\eta above. Finally the map from the finite face-gluing quotient of the compact simplices to Σ(X)\Sigma(\mathfrak X) is a continuous bijection from a compact space to a Hausdorff space, hence is a homeomorphism. This proves the asserted finite complex and its intrinsic integral-affine structure. The previously proved global homotopy has exactly this fixed set and remains the claimed strong deformation retraction. ◻

5.26 Faithful rational coordinates on a curve skeleton

Throughout this section, KK is algebraically closed, complete and nontrivially valued. Let XX be a smooth projective integral KK-curve and let X/K∘\mathcal X/K^\circ be a projective model satisfying Theorem 4.23.3. Write Σ=Σ(X)\Sigma=\Sigma(\mathcal X). It is a compact finite metric graph. Its finite edges have the logarithmic length defined by the nodal annuli.

Lemma 5.26.1. NAG-002199 (Rational approximation of a nonzero germ). Every point of Σ\Sigma has algebraic support the generic point of XX. If x∈Σx\in\Sigma and gg is an analytic germ with g(x)≠0g(x)\ne 0, there is f∈K(X)×f\in K(X)^\times such that ∣f∣=∣g∣|f|=|g| on an ordinary neighborhood of xx on which both are defined. If ∣g(x)∣=1|g(x)|=1, f(x)f(x) and g(x)g(x) can be required to have the same residue.

Proof. At a component vertex, Proposition 4.24.2 gives residual transcendence degree one. At an interior annular Gauss point, Proposition 3.8.1 gives either residual transcendence degree one or a nonzero relative value-group rank. A point supported at an algebraic closed point would have completed residue field KK, by algebraic closedness and Proposition 2.10.1. These alternatives exclude closed algebraic support; an integral algebraic curve has only generic and closed points.

By the same description of analytification, H(x)\mathcal H(x) is the completion of K(X)K(X) for the norm of xx. Choose f∈K(X)f\in K(X) with ∣f(x)−g(x)∣<∣g(x)∣|f(x)-g(x)|<|g(x)|. This makes f≠0f\ne 0. The rational function and its inverse are regular on a Zariski open containing the generic point, so their analytifications are defined near xx. Continuity preserves the displayed strict inequality on a smaller ordinary neighborhood. The ultrametric inequality gives ∣f∣=∣g∣|f|=|g| there. When ∣g(x)∣=1|g(x)|=1, the same inequality says that the residues agree. ◻

Lemma 5.26.2. NAG-002200 (Prescribing finitely many residual orders). Let CC be an irreducible component of the special fibre of X\mathcal X, and let p1,…,pmp_1,\ldots,p_m be distinct closed points of its normalization CνC^\nu. For each ii there is ui∈k(C)×u_i\in k(C)^\times with ord⁡pj(ui)=δij.\operatorname{ord}_{p_j}(u_i)=\delta_{ij}.

Proof. The field k=K~k=\widetilde K is algebraically closed: lift a monic polynomial to K∘K^\circ and factor it over KK; dominance of the leading term makes every root integral, and reduction gives the factorization over kk. Choose a projective embedding of CC. A hyperplane avoiding the finitely many images of the pjp_j exists: the forbidden coefficients form a finite union of proper linear subspaces over the infinite field kk. Its complement on CC is affine. The normalization is finite by [StacksAlg] (Tag 0BXR), so its inverse image is an affine normal curve Spec⁡B\operatorname{Spec}B containing all pjp_j. Its local rings at these points are discrete valuation rings by [StacksAlg] (Tag 0B8Y).

Write mj\mathfrak m_j for their maximal ideals. A local uniformizer at pip_i can be written ai/bia_i/b_i, with ai,bi∈Ba_i,b_i\in B and bi∉mib_i\notin\mathfrak m_i; hence aia_i has order one at pip_i. The ideals mi2\mathfrak m_i^2 and mj\mathfrak m_j, j≠ij\ne i, are pairwise comaximal. Indeed, powers of comaximal ideals remain comaximal by expanding a sufficiently high power of 1=a+b1=a+b. For two comaximal ideals I,JI,J, write a+b=1a+b=1 with a∈I,b∈Ja\in I,b\in J; then br+asbr+as has prescribed residues rr modulo II and ss modulo JJ. Induction gives the same assertion for finitely many such ideals. Choose ui∈Bu_i\in B congruent to aia_i modulo mi2\mathfrak m_i^2 and to 11 modulo each mj\mathfrak m_j, j≠ij\ne i. Its orders are exactly those claimed, so it is a nonzero rational function. ◻

Proposition 5.26.3. NAG-002201 (Rational functions are piecewise integral affine). For every f∈K(X)×f\in K(X)^\times, the function −log⁡∣f∣-\log|f| is finite and continuous on Σ\Sigma, and is integral affine on every edge of a finite subdivision.

Proof. All skeletal points have generic algebraic support, so ff and f−1f^{-1} are analytic near Σ\Sigma. At a vertex, the finitely many incident graph branches are among the residual directions of Theorem 4.25.3. Theorem 4.24.5 makes −log⁡∣f∣-\log|f| integral affine on a sufficiently short initial segment of every one of them.

For an open nodal edge choose an algebraic affine neighborhood of its residual node. Its formal generic fibre contains the entire nodal tube. Express f=P/Qf=P/Q in the fraction field of the generic coordinate algebra. The functions P,QP,Q are analytic on that tube. In its pointed annulus coordinate, their norms on every compact subinterval of radii are finite maxima of Laurent monomial norms, by Proposition 4.4.1 and its first, nonunit, assertion. The two nonzero functions have positive Gauss norm at every radius. Subtracting their logarithmic profiles proves finite piecewise integral affineness on each compact interior subinterval. The already affine endpoint segments leave one such compact subinterval. There are finitely many edges, so one finite common subdivision works. ◻

Proposition 5.26.4. NAG-002202 (Local rational coordinates of lattice slope one). Every point of Σ\Sigma has a graph neighborhood WW and finitely many rational functions whose negative logarithmic norms give an injective map W→RmW\to\mathbf R^m. On each sufficiently short edge of WW, one coordinate has slope 11 or −1-1. At a model vertex the incident branches can be sent to distinct positive coordinate axes, after subtracting the value at the vertex.

Proof. For a point interior to an edge, its nodal annulus coordinate zz is an analytic unit, and −log⁡∣z∣-\log|z| is its length coordinate. Lemma 5.26.1 gives a rational function with the same norm on a neighborhood of the point. Restricting to a small interval proves the assertion there, for every real radius.

At a vertex vCv_C, label its incident branches by the distinct normalization points p1,…,pmp_1,\ldots,p_m, counting the two ends of a loop separately. Choose the residual functions uiu_i of Lemma 5.26.2. By Proposition 4.24.2, each is the residue of a norm-one element of H(vC)\mathcal H(v_C). The density of K(X)K(X) in this field gives fi∈K(X)×f_i\in K(X)^\times of norm one with residue uiu_i. The slope formula Theorem 4.24.5 then gives, on the jj-th short branch with length parameter τ\tau, (−log⁡∣f1∣,…,−log⁡∣fm∣)(γj(τ))=τej.\bigl(-\log|f_1|,\ldots,-\log|f_m|\bigr)(\gamma_j(\tau)) =\tau e_j. There are only finitely many pairs (i,j)(i,j), so one shortening works for all of them. The resulting star is mapped injectively, with primitive edge direction vectors. An isolated vertex needs no local coordinates. ◻

Theorem 5.26.5. NAG-002203 (Faithful tropicalization of a compact model skeleton). There is a nonempty Zariski open U⊂XU\subset X containing Σ\Sigma in UanU^{\mathrm{an}} and a closed immersion φ:U↪Gmn\varphi:U\hookrightarrow\mathbf G_m^n such that φtrop∣Σ\varphi_{\mathrm{trop}}|_\Sigma is a homeomorphism onto its image. After a finite subdivision of Σ\Sigma, its restriction to each edge is integral affine with primitive derivative vector. Thus each edge has its original logarithmic length measured in the primitive lattice length of its tropical image.

Proof. Compactness gives finitely many neighborhoods from Proposition 5.26.4. Collect their coordinate functions into a map F:Σ→RaF:\Sigma\to\mathbf R^a. It is injective on each of these neighborhoods. Let VV be the union of their Cartesian squares in Σ×Σ\Sigma\times\Sigma. This is an open neighborhood of the diagonal. The compact set E=(Σ×Σ)∖VE=(\Sigma\times\Sigma)\setminus V contains no diagonal point.

Any two distinct points x,y∈Σx,y\in\Sigma are separated by the absolute value of a nonzero rational function. Indeed, choose an algebraic affine open of XX. It contains the algebraic support of every skeletal point, so its analytification contains x,yx,y. Their distinct seminorms on its coordinate ring are separated by some element of that ring. By generic support, its two values are nonzero. The inequality of its two norms persists on a neighborhood of (x,y)(x,y). Finitely many such neighborhoods cover EE; append their rational functions to FF. The resulting map GG is injective on all of Σ\Sigma. It is a homeomorphism onto its image because its source is compact and its target is Hausdorff.

Proposition 5.26.3 gives a finite common subdivision on which all coordinates of GG are integral affine. On the interior of each resulting edge, one of the original local coordinates has slope 11 or −1-1. Its derivative vector is therefore primitive. Refining further retains this property.

It remains to realize these coordinates in a closed torus immersion. Choose a nonempty affine open U0=Spec⁡AU_0=\operatorname{Spec}A of XX. Write the finitely many rational functions of GG as fractions of nonzero elements of AA, and invert all their numerators and denominators. This gives a finite-type algebra BB and an affine open on which these functions are units. Choose finitely many nonzero algebra generators of BB, and invert them as well, obtaining B′B'. The chosen generators and their inverses generate B′B', so they give a surjection from a Laurent polynomial algebra onto B′B'. Append the functions defining GG to this generating list. The resulting map U=Spec⁡B′→GmnU=\operatorname{Spec}B'\to\mathbf G_m^n is a closed immersion. Every inverted element is nonzero in K(X)K(X), so its zero locus contains no skeletal point; hence Σ⊂Uan\Sigma\subset U^{\mathrm{an}}. Adding coordinates retains injectivity and a derivative coordinate equal to 11 or −1-1. Apply Proposition 5.26.3 once more to the appended functions and take a finite common refinement. All coordinates of the closed immersion are then affine on every subdivided edge, while the primitive derivative coordinate is retained. The primitive derivative formula identifies the two edge lengths, as asserted. ◻

References. This is the rational-coordinate form of faithful tropicalization for compact curve skeleta. The proof constructs local lattice-slope-one coordinates first, then uses compact pair separation and an affine algebra presentation to obtain a closed torus immersion.

5.27 Faithful curve coordinates over the original field

Lemma 5.27.1. NAG-002135 (An affine neighborhood of finitely many projective points). Let ZZ be a projective scheme over any field kk and let q1,…,qmq_1,\ldots,q_m be distinct closed points. They lie in a common affine open of ZZ. If CC is a projective integral curve and p1,…,pmp_1,\ldots,p_m are distinct closed points of its normalization, there are functions ui∈k(C)×u_i\in k(C)^\times satisfying ord⁡pj(ui)=δij.\operatorname{ord}_{p_j}(u_i)=\delta_{ij}.

Proof. Embed ZZ in Pkn\mathbf P^n_k and let PiP_i be the homogeneous prime defining qiq_i. The PiP_i are pairwise incomparable. For each j≠ij\ne i, choose a homogeneous element of Pj∖PiP_j\setminus P_i, and let gig_i be their product; if m=1m=1, put g1=1g_1=1. Then gig_i vanishes at the other selected points and not at qiq_i. Choose a coordinate xai∉Pix_{a_i}\notin P_i and an integer D≥max⁡(1,max⁡ideg⁡gi)D\geq\max(1,\max_i\deg g_i). The homogeneous polynomial F=∑igixaiD−deg⁡giF=\sum_i g_i x_{a_i}^{D-\deg g_i} is outside every PiP_i: modulo PiP_i its only nonzero summand is the ii-th. The standard affine open D+(F)D_+(F) in projective space therefore gives the required affine open on ZZ. This construction works over finite fields as well.

Apply this to the finitely many distinct images of the pjp_j on CC. Its inverse image under normalization is affine, because normalization is finite by [StacksAlg] (Tag 0BXR). Write it as Spec⁡B\operatorname{Spec}B, and let mj\mathfrak m_j correspond to pjp_j. By [StacksAlg] (Tag 0B8Y), BmiB_{\mathfrak m_i} is a discrete valuation ring. A local uniformizer is ai/bia_i/b_i with ai,bi∈Ba_i,b_i\in B and bi∉mib_i\notin\mathfrak m_i, so aia_i has order one at pip_i. The ideals mi2\mathfrak m_i^2 and mj\mathfrak m_j, j≠ij\ne i, are pairwise comaximal. The elementary Chinese remainder construction in Lemma 5.26.2 uses only these comaximalities, not any assumption on kk: for I+J=BI+J=B and a+b=1a+b=1 with a∈I,b∈Ja\in I,b\in J, the element br+asbr+as has prescribed residues rr modulo II and ss modulo JJ, and induction treats finitely many ideals. Choose uiu_i congruent to aia_i modulo mi2\mathfrak m_i^2 and to 11 modulo every other mj\mathfrak m_j. This gives precisely the displayed orders, regardless of the residue degrees of the pjp_j. ◻

In the rest of this section, KK is complete and nontrivially valued, with its original absolute value. Let XX be a smooth projective integral KK-curve with a projective model X/K∘\mathcal X/K^\circ satisfying Theorem 4.23.3, or its characteristic-zero form Theorem 4.28.5. Thus all nodes used for the finite graph are split rational nodes with the stated pointed charts. Put Σ=Σ(X)\Sigma=\Sigma(\mathcal X).

Lemma 5.27.2. NAG-002136 (Generic support and rational approximation over KK). Every x∈Σx\in\Sigma has generic algebraic support. If gg is an analytic germ at xx with g(x)≠0g(x)\ne 0, some f∈K(X)×f\in K(X)^\times satisfies ∣f∣=∣g∣|f|=|g| on an ordinary neighborhood of xx. If ∣g(x)∣=1|g(x)|=1, the residues of f(x)f(x) and g(x)g(x) can also be made equal.

Proof. For a finite extension E/KE/K, Lemma 1.12.1 makes EE complete, and its proof shows that every value in v(E×)v(E^\times) has a positive integer multiple in v(K×)v(K^\times). Moreover E~/K~\widetilde E/\widetilde K is algebraic: for b∈Eb\in E of norm one, take a polynomial relation over KK, divide its coefficients by a coefficient of maximal norm, and reduce. The resulting polynomial over K~\widetilde K is nonzero and annihilates the residue of bb. Thus the residual transcendence degree and relative rational value-group rank of a finite extension are both zero.

At a component vertex, Proposition 4.24.2 gives residual transcendence degree one. At an annular Gauss point, Proposition 3.8.1 gives the sum of these two invariants equal to one. If a skeletal point had closed algebraic support, its residue field would be a finite extension of KK by [StacksAlg] (Tag 00P0), and Proposition 2.10.1 would identify its completed residue field with that same finite field. This contradicts the preceding invariants. An integral algebraic curve has only generic and closed points, so its support is generic.

Consequently H(x)\mathcal H(x) is the completion of K(X)K(X) for the norm of xx. Choose f∈K(X)f\in K(X) with ∣f(x)−g(x)∣<∣g(x)∣|f(x)-g(x)|<|g(x)|. The rational function and its inverse are regular on a Zariski open containing the generic support; their analytifications are defined near xx. Continuity preserves the strict inequality on an ordinary neighborhood, and the ultrametric inequality gives equality of the norms there. For norm-one values the error has norm less than one, which also identifies the residues. ◻

Proposition 5.27.3. NAG-002137 (Integral affine coordinates on the original skeleton). For every f∈K(X)×f\in K(X)^\times, the function −log⁡∣f∣-\log|f| is finite and continuous on Σ\Sigma and integral affine on the edges of a finite subdivision. Every point of Σ\Sigma has a graph neighborhood on which finitely many such functions give an injective map with a derivative coordinate equal to 11 or −1-1 on each short edge.

Proof. By Lemma 5.27.2, ff and its inverse are analytic near every skeletal point. At a vertex, each incident graph branch is a rational normalization point above a split node, by Lemma 4.24.1. Theorem 4.24.5, over the original field, makes the logarithmic norm affine with integral slope on a short initial segment. On the interior of a nodal edge, express f=P/Qf=P/Q on an algebraic affine neighborhood of the residual node. Both functions are analytic on the nodal tube. On each compact radius subinterval their nonzero Gauss norms are finite maxima of Laurent monomial norms by Proposition 4.4.1. Hence their logarithmic difference has finitely many integral affine pieces. The endpoint segments leave a compact interior interval; doing this for the finitely many edges proves the first assertion.

At an interior point, approximate its annular coordinate zz by Lemma 5.27.2. The rational norm agrees near that point with ∣z∣|z|, so its negative logarithm is an oriented length coordinate of slope one. At a vertex vCv_C, list the finitely many distinct normalization points pjp_j corresponding to incident edges, counting the two branches of a loop separately. Lemma 5.27.1 gives ui∈k(C)×u_i\in k(C)^\times with orders δij\delta_{ij}. Proposition 4.24.2 identifies these with norm-one residues in H(vC)\mathcal H(v_C). Approximate their lifts by fi∈K(X)f_i\in K(X) with error of norm less than one. Theorem 4.24.5 gives, after a simultaneous shortening of finitely many branches, (−log⁡∣f1∣,…,−log⁡∣fm∣)(γj(t))=tej.(-\log|f_1|,\ldots,-\log|f_m|)(\gamma_j(t))=t e_j. These coordinates embed the small graph star in distinct positive coordinate axes and have primitive derivative vectors. An isolated vertex requires no coordinates. ◻

Theorem 5.27.4. NAG-002138 (Original-field faithful tropicalization of a split model skeleton). There is a nonempty Zariski open U⊂XU\subset X with Σ⊂Uan\Sigma\subset U^{\mathrm{an}} and a closed immersion over the original field KK, φ:U↪Gm,Kn\varphi:U\hookrightarrow\mathbf G_{m,K}^n whose tropicalization restricts to a homeomorphism on Σ\Sigma. On a finite subdivision of its edges, it is integral affine with primitive derivative vector. Consequently it preserves every edge length with the original absolute value, when the image is measured by primitive lattice length.

Proof. Take finitely many graph neighborhoods from Proposition 5.27.3 covering the compact graph. Gather their rational coordinate functions into FF. The union VV of the squares of these neighborhoods is open in Σ×Σ\Sigma\times\Sigma and contains the diagonal. On every pair in VV the coordinates of FF separate distinct points. The complement EE is compact and disjoint from the diagonal.

Any two distinct x,y∈Σx,y\in\Sigma are distinguished by a nonzero rational function over KK. Indeed, any algebraic affine open of XX contains their generic algebraic supports, and the two distinct seminorms on its coordinate ring differ on some element. Both values are nonzero by generic support. The strict inequality of absolute values persists on an ordinary neighborhood of (x,y)(x,y). Finitely many such pair neighborhoods cover EE. Append their rational functions to FF, obtaining a globally injective continuous map GG. A continuous injection of this compact graph into a Hausdorff real vector space is a homeomorphism onto its image.

Choose a nonempty affine open Spec⁡A⊂X\operatorname{Spec}A\subset X. Invert the finitely many nonzero numerators and denominators of the coordinates of GG, obtaining BB, and then invert finitely many nonzero algebra generators of BB, obtaining B′B'. Those generators and their inverses generate B′B'. Together with the coordinates of GG they therefore give a surjection from a Laurent polynomial KK-algebra onto B′B', hence the required closed immersion for U=Spec⁡B′U=\operatorname{Spec}B'. Every function inverted is nonzero at every generic-support point, so Σ⊂Uan\Sigma\subset U^{\mathrm{an}}. All choices and functions are defined over KK.

Apply Proposition 5.27.3 to all final coordinates and take a finite common subdivision. On each resulting edge, one of the local coordinates has derivative 11 or −1-1 on a nonempty subinterval, hence on the entire edge because it is affine there. The full integer derivative vector is therefore primitive. Appending coordinates does not remove that coordinate or destroy injectivity. The primitive lattice parameter on the image is exactly the original annular length parameter, which proves the length assertion. ◻

References. The construction extends the compact-model-skeleton argument to the original field. Homogeneous finite-point separation replaces infinite-field hyperplane avoidance, and the residue/value invariants of finite extensions replace algebraic closedness in the support argument.

5.28 Geometric degrees and tropical lattice indices

Write R=K∘R=K^\circ, m=K∘∘\mathfrak m=K^{\circ\circ} and k=K~k=\widetilde K. In the first two lemmas, KK is complete and nontrivially valued, and π∈R\pi\in R satisfies 0<∣π∣<10<|\pi|<1. The formal algebras carry their quotient π\pi-adic topologies. For a lattice MM, the notation R⟨M⟩R\langle M\rangle denotes the Gauss unit ball in K⟨M⟩K\langle M\rangle.

Lemma 5.28.1. NAG-002241 (The quotient model of an initial degeneration). Let I⊂K[M]I\subset K[M] be an ideal and set P=K⟨M⟩,P0=R⟨M⟩,J=IP.P=K\langle M\rangle,\qquad P_0=R\langle M\rangle, \qquad J=IP. Then B0=P0/(J∩P0)B_0=P_0/(J\cap P_0) is an admissible topologically finitely presented RR-algebra, with generic algebra B=P/JB=P/J. Its special algebra is B0/mB0≃k[M]/I∩R[M]‾,B_0/\mathfrak mB_0 \simeq k[M]/\overline{I\cap R[M]}, where the bar denotes coefficientwise reduction in k[M]k[M].

Proof. The ideal JJ is closed by Theorem 1.18.3. The algebra B0B_0 embeds in P/JP/J, so it has no RR-torsion. It is a quotient of a restricted power-series algebra in finitely many variables: use generators for MM and their inverses. Consequently Theorem 3.14.4 gives a finite presentation, and Proposition 3.14.5 gives completeness and separatedness. Every restricted Laurent series becomes integral after multiplication by a power of π\pi. The same is true of every element of JJ, whence B0[1/π]=P/J.B_0[1/\pi]=P/J.

The elements of P0P_0 of Gauss norm less than one form mP0\mathfrak mP_0. To see this, choose a coefficient of maximal absolute value in a nonzero restricted series. Dividing by that coefficient gives an element of P0P_0; the coefficient belongs to m\mathfrak m precisely when the original norm is less than one. The reverse inclusion follows from the ultrametric inequality applied to a finite sum. Thus P0/mP0=k[M]P_0/\mathfrak mP_0=k[M]. Proposition 5.13.5 identifies the image of J∩P0J\cap P_0 in this ring with I∩R[M]‾\overline{I\cap R[M]}, proving the special-fibre formula. ◻

Lemma 5.28.2. NAG-002242 (Finite formal models at the Gauss point). Let A0=R⟨U1±1,…,Ud±1⟩,A=A0[1/π],A_0=R\langle U_1^{\pm1},\ldots,U_d^{\pm1}\rangle, \qquad A=A_0[1/\pi], and let A0→B0A_0\to B_0 be a formal homomorphism to an admissible topologically finitely presented RR-algebra. Suppose that B‾0=B0/mB0\overline B_0=B_0/\mathfrak mB_0 is finite over k[U1±1,…,Ud±1]k[U_1^{\pm1},\ldots,U_d^{\pm1}]. Then B0B_0 is finite over A0A_0, and the generic algebra B=B0[1/π]B=B_0[1/\pi] has its canonical finite-AA-module topology. If HH is the completed residue field of the Gauss point of M(A)\mathcal M(A), then dim⁡H(B⊗^AH)=dim⁡k(U)(B‾0⊗k[U±1]k(U)).\begin{equation} \dim_H(B\widehat\otimes_A H) =\dim_{k(U)}\left( \overline B_0\otimes_{k[U^{\pm1}]}k(U)\right). \end{equation}

Proof. Choose bounded formal generators y1,…,ysy_1,\ldots,y_s of B0B_0, giving a surjection A0⟨Y1,…,Ys⟩→B0A_0\langle Y_1,\ldots,Y_s\rangle\to B_0. Each y‾i\overline y_i satisfies a monic polynomial over the residual Laurent ring. Lift its coefficients to obtain Pi∈A0[Yi]P_i\in A_0[Y_i], monic of degree ri>0r_i>0, with all lower coefficients of Gauss norm at most one. Since Pi(yi)∈mB0P_i(y_i)\in\mathfrak mB_0, this element is a finite sum of scalars in m\mathfrak m times elements of B0B_0. Lift those elements through the formal presentation. This produces Gi∈A0⟨Y⟩G_i\in A_0\langle Y\rangle with Gi(y)=Pi(yi),∥Gi∥<1.G_i(y)=P_i(y_i),\qquad \|G_i\|<1. There are finitely many equations, so θ=max⁡i∥Gi∥<1\theta=\max_i\|G_i\|<1 is a common contraction bound.

Apply Lemma 3.16.4 over AA. Its convergent division formula expresses every bounded restricted series, modulo the Pi−GiP_i-G_i, as a linear combination of the monomials YνY^\nu with 0≤νi<ri0\leq\nu_i<r_i. The coefficient maps have norm at most one, so a series in A0⟨Y⟩A_0\langle Y\rangle has all its remainder coefficients in A0A_0. The images of these finitely many monomials therefore generate B0B_0 over A0A_0.

The same calculation compares the generic norms. Give BB the quotient norm from A⟨Y⟩A\langle Y\rangle and let ∥⋅∥fin\|\cdot\|_{\rm fin} be the quotient norm from the finite free AA-module on the remainder monomials. A finite remainder expression gives ∥b∥B≤∥b∥fin\|b\|_B\leq\|b\|_{\rm fin}. Conversely, apply the bounded division formula to each representative of bb and take the infimum of its norms. It gives ∥b∥fin≤∥b∥B\|b\|_{\rm fin}\leq\|b\|_B. Hence these quotient norms agree; the equivalent admissible norms give the same Banach topology.

Let S⊂A0S\subset A_0 consist of the elements of Gauss norm one, and put V=S−1A0V=S^{-1}A_0. The Gauss norm is multiplicative by Proposition 1.2.2. Every nonzero restricted Laurent series aa has a coefficient c∈K×c\in K^\times of maximal absolute value, so that a=csa=c s with s∈Ss\in S. Every element of Frac⁡(A)×\operatorname{Frac}(A)^\times is consequently of the form cs/tc s/t, where c∈K×c\in K^\times and s,t∈Ss,t\in S. It belongs to VV exactly when ∣c∣≤1|c|\leq1. Thus VV is the Gauss valuation ring in Frac⁡(A)\operatorname{Frac}(A), with residue field k(U)k(U): the reductions of SS are all the nonzero Laurent polynomials. Its completion is H∘H^\circ. Indeed, an element of norm at most one in the completion of Frac⁡(A)\operatorname{Frac}(A) can be approximated, with error less than one, by elements of norm at most one, which belong to VV.

The finite VV-module M=S−1B0M=S^{-1}B_0 is torsion-free. Every nonzero element of VV is a nonzero scalar from RR times a unit; multiplication by that scalar is injective on B0B_0 and remains injective after localization. A finite torsion-free module over a valuation ring is free. One can see this directly by choosing a generating set of minimal size: a nonzero relation has a coefficient dividing all the others, and torsion-freeness then expresses the corresponding generator in terms of the rest. This contradicts minimality.

The rank of MM is therefore its residue dimension over k(U)k(U), which is the right-hand side of (5.28.1). Tensoring its free basis with HH computes dim⁡H(B⊗AH)\dim_H(B\otimes_A H) with the same value. The norm comparison above and Proposition 2.33.19 identify this algebraic tensor product with B⊗^AHB\widehat\otimes_AH. ◻

Proposition 5.28.3. NAG-002244 (Initial multiplicity and an actual fibre degree). Let X⊂TNX\subset T_N be a pure dd-dimensional closed subscheme over a complete real-valued non-Archimedean field KK. Let ww belong to the relative interior of a maximal cell σ\sigma of an initial-stratum complex, and suppose that q:N→Zdq:N\to\mathbf Z^d restricts to an injective map on NσN_\sigma with finite index e=[Zd:q(Nσ)].e=[\mathbf Z^d:q(N_\sigma)]. After a complete isometric extension realizing ww, made nontrivially valued, translate ww to zero. The induced monomial map from the actual tropical fibre XwX_w to the dd-dimensional unit polycircle is finite and norm-admissible. Its fibre algebra over the completed Gauss residue field HH has dimension e mX(σ)e\,m_X(\sigma). Multiplying the target coordinates by constants of absolute value one preserves these conclusions.

Proof. Theorem 5.13.6 supplies the required complete extension and translation, and Proposition 5.19.8 preserves the weight. The quotient model of Lemma 5.28.1 has the full initial algebra as its special algebra. By Proposition 5.19.3 and Theorem 5.19.5, a lattice splitting identifies it with D[Nσ∨],dim⁡kD=mX(σ),D[N_\sigma^\vee],\qquad \dim_kD=m_X(\sigma), where DD is a finite kk-algebra.

In this presentation the target characters reduce to djχajd_j\chi^{a_j}, with dj∈D×d_j\in D^\times. The aja_j form a basis of a sublattice of index ee in Nσ∨N_\sigma^\vee. Choose representatives for its cosets. Every character has a unique expression as one of these representatives times a product of the χaj\chi^{a_j}. Substitution of χaj=dj−1Uj\chi^{a_j}=d_j^{-1}U_j consequently makes D[Nσ∨]D[N_\sigma^\vee] free of rank ee over D[U±1]D[U^{\pm1}]. The asserted independence follows by comparing character exponents, since the djd_j are units. A kk-basis of DD then gives a basis of size edim⁡kDe\dim_kD over k[U±1]k[U^{\pm1}].

The normalized monomials and their inverses belong to the quotient formal model, so this is the reduction of a formal homomorphism from the unit-polycircle model. Apply Lemma 5.28.2. It proves the finiteness, norm comparison and actual Gauss fibre dimension. A unit constant changes only one of the coefficients djd_j in the preceding calculation. The calculation uses a basis of the entire finite algebra DD, including its nilpotents. ◻

Lemma 5.28.4. NAG-002245 (A finite free locus). Let f:X→Yf:X\to Y be a dominant morphism of affine finite-type KK-schemes. Suppose that YY is integral and that the generic fibre is nonempty and zero-dimensional. Put F=K(Y)F=K(Y) and δ=dim⁡FΓ(XF,OXF).\delta=\dim_F\Gamma(X_F,\mathcal O_{X_F}). There is a nonzero h∈Γ(Y,OY)h\in\Gamma(Y,\mathcal O_Y) such that f−1(Yh)→Yhf^{-1}(Y_h)\to Y_h is finite free of rank δ\delta.

Proof. Write Y=Spec⁡(A)Y=\operatorname{Spec}(A) and X=Spec⁡(B)X=\operatorname{Spec}(B). Lemma 5.19.1 makes the generic algebra BFB_F finite-dimensional. Each of a finite set of AA-algebra generators of BB satisfies a monic polynomial over FF. Clear the finitely many coefficient denominators. Each resulting polynomial equality holds in BFB_F, so its error in the localized BB is killed by a further nonzero element of AA. After inverting the product of these finitely many elements, all the monic equalities hold. The bounded monomials in the generators then give a finite set of module generators for this localization of BB.

Choose c1,…,cδc_1,\ldots,c_\delta among those monomials whose images form an FF-basis of BFB_F. Express the other module generators in that basis. Clear their finitely many coefficient denominators and then kill their finitely many errors by another localization in A∖{0}A\setminus\{0\}. The cic_i now generate the localized module. They are independent: a relation maps to an FF-linear relation between their images, and its coefficients vanish because the localized AA injects into FF. The product of all elements inverted is the required hh, and the displayed basis has size δ\delta. ◻

Theorem 5.28.5. NAG-002246 (The algebraic tropical degree formula). Let φ:TN→TN′\varphi:T_N\to T_{N'} be a homomorphism of split tori over a complete real-valued non-Archimedean field KK, and write a:N→N′a:N\to N' for its lattice map. Let X⊂TNX\subset T_N be a pure dd-dimensional closed subscheme and Y⊂TN′Y\subset T_{N'} an integral dd-dimensional closed subscheme. Suppose that φ∣X:X→Y\varphi|_X:X\to Y is dominant with nonempty zero-dimensional generic fibre of scheme degree δ\delta.

There is a finite union EtropE_{\mathrm{trop}} of polyhedra of dimension less than dd in Trop⁡(Y)\operatorname{Trop}(Y) such that, for w′∉Etropw'\notin E_{\mathrm{trop}}, its tropical preimages form a finite set {wi}i\{w_i\}_i. Each wiw_i lies in the relative interior of a maximal initial cell σi\sigma_i, while w′w' lies in the relative interior of a maximal initial cell σ′\sigma', and δ mY(σ′)=∑imX(σi)[Nσ′′:a(Nσi)].\begin{equation} \delta\,m_Y(\sigma') =\sum_i m_X(\sigma_i) [N'_{\sigma'}:a(N_{\sigma_i})]. \end{equation} For the polyhedral pushforward of Theorem 5.17.2, this is the weighted-cycle identity a∗Trop⁡(X)=δTrop⁡(Y).a_*\operatorname{Trop}(X)=\delta\operatorname{Trop}(Y).

Proof. Step 1: choose the finite free locus and the regular weights. Choose hh by Lemma 5.28.4, and let E=Y∖YhE=Y\setminus Y_h have its reduced closed structure. Every irreducible component of EE has dimension less than dd: a chain of primes above a nonzero prime of the domain Γ(Y,OY)\Gamma(Y,\mathcal O_Y) becomes one step longer after adjoining its zero prime. Apply Theorem 5.19.5 to the finitely many components of EE. Their tropicalizations form a finite union of lower-dimensional polyhedra.

Also exclude the nonmaximal cells of Trop⁡(Y)\operatorname{Trop}(Y), the images of nonmaximal cells of Trop⁡(X)\operatorname{Trop}(X), and the images of maximal source cells on which aa has rank less than dd. Intersect these images with the target complex. The finite polyhedral elimination and refinement in Theorem 5.17.2 give a finite union EtropE_{\mathrm{trop}} of polyhedra of dimension less than dd. It cannot contain the relative interior of a maximal target cell.

Fix w′∉Etropw'\notin E_{\mathrm{trop}}. Each of its tropical preimages is in the relative interior of a maximal source cell on whose direction space aa is injective. Each of finitely many such cells contributes at most one point. Thus the preimage set is finite. Near each wiw_i, the image of its full-dimensional source cell is contained in the local tropical support at w′w', which is the direction space of σ′\sigma'. The dimensions are equal, so a(Nσi)R=Nσ′,R′.a(N_{\sigma_i})_{\mathbf R}=N'_{\sigma',\mathbf R}. The indices in (5.28.2) are therefore finite.

Step 2: decompose the actual inverse-image algebra. Use Theorem 5.13.6 to make one complete isometric extension L/KL/K realizing w′w' and all the wiw_i, with nontrivial valuation. The finite free algebra on YhY_h remains free of rank δ\delta after extension, and Proposition 5.19.8 preserves all the weights. In particular, the subsequent argument uses the same finite free module even if YLY_L is reducible or nonreduced.

Write Yw′=M(C)Y_{w'}=\mathcal M(C) for the actual target tropical fibre over LL. Since w′∉Trop⁡(E)w'\notin\operatorname{Trop}(E), hh has no zero on this fibre. It is therefore a unit in CC by Theorem 1.20.1, and Yw′⊂(Yh)anY_{w'}\subset(Y_h)^{\mathrm{an}}. The explicit tensor description in Theorem 2.33.26 shows that the inverse image is M(D)\mathcal M(D), where DD is a finite CC-algebra free of rank δ\delta.

Every point of this inverse image has tropical weight one of the wiw_i. Its subset of weight wiw_i is the actual affinoid domain Xwi=M(Bi)X_{w_i}=\mathcal M(B_i) of Proposition 5.13.2. These subsets are open and closed: they are fibres of a continuous map with values in a finite Hausdorff set. Applying the structure sheaf property of Theorem 2.39.4 to this finite disjoint cover gives D≃∏iBi.\begin{equation} D\simeq\prod_i B_i. \end{equation} The inverse image is nonempty. Indeed, C≠0C\ne0 and δ>0\delta>0, so D≠0D\ne0; the affinoid nonemptiness theorem 1.20.1 gives a point of M(D)\mathcal M(D). Consequently the set of wiw_i is nonempty.

Step 3: compare all degrees over one Gauss field. The saturated lattice Nσ′′N'_{\sigma'} is a direct summand of N′N' by Lemma 5.14.1. Choose q:N′→Zdq:N'\to\mathbf Z^d whose restriction to it is an isomorphism. Divide the corresponding target monomials by their values at a translating point of weight w′w'. This gives one unit-polycircle algebra A=L⟨U1±1,…,Ud±1⟩A=L\langle U_1^{\pm1},\ldots,U_d^{\pm1}\rangle mapping to CC and to every BiB_i. After the respective source translations, the induced monomials differ from those for q∘aq\circ a by constants of absolute value one, because a(wi)=w′a(w_i)=w'.

Let HH be the completed Gauss residue field of AA. Proposition 5.28.3 gives dim⁡H(C⊗^AH)=mY(σ′),dim⁡H(Bi⊗^AH)=mX(σi)[Nσ′′:a(Nσi)].\begin{aligned} \dim_H(C\widehat\otimes_A H)&=m_Y(\sigma'),\\ \dim_H(B_i\widehat\otimes_A H) &=m_X(\sigma_i)[N'_{\sigma'}:a(N_{\sigma_i})]. \end{aligned} All the maps are finite with the canonical finite-module norms, so Proposition 2.33.19 identifies these completed tensors with the algebraic ones. Since DD is free of rank δ\delta over CC, tensoring its free basis over AA with HH gives dim⁡H(D⊗AH)=δdim⁡H(C⊗AH).\dim_H(D\otimes_AH)=\delta\dim_H(C\otimes_AH). Taking dimensions in (5.28.3) proves (5.28.2).

Theorem 5.22.5 makes both weighted tropicalizations balanced cycles. Choose compatible source and target subdivisions as in the polyhedral pushforward theorem. The equality holds on a dense relative open subset of every maximal target cell. Both weights are constant on each such cell by Proposition 5.19.8 and the definition of polyhedral pushforward. This proves the weighted-cycle identity. ◻

Remark (Finite algebras and nonreduced targets). Every degree in the proof is the dimension of a finite algebra. For the map t↦tpt\mapsto t^p in characteristic pp, both this degree and the lattice index are pp, although each geometric fibre has one underlying point.

The same proof applies to a pure dd-dimensional target YY if there is an open subset meeting all its generic points on which the source algebra is finite free of one fixed rank δ\delta. The complement again has dimension less than dd, and Steps 2 and 3 use precisely this free module. For comparison, the closed immersion Td↪Td×Spec⁡K[ϵ]/(ϵ2),ϵ⟼0,T^d\hookrightarrow T^d\times\operatorname{Spec}K[\epsilon]/(\epsilon^2), \qquad \epsilon\longmapsto0, is finite and dominant on underlying spaces, but its generic module is not free. The source weight is one and the target weight is two. Thus its residue-field degree one does not give the displayed degree formula. Realize the target as a torus subscheme by writing ϵ=V−1\epsilon=V-1 and imposing (V−1)2=0(V-1)^2=0.

References. The weighted-cycle identity is the classical Sturmfels–Tevelev multiplicity formula; its trivial- and nontrivial-valuation forms are reviewed in [GRW16] (§8). The proof above derives it from the initial quotient model and the actual finite analytic fibres. The short valuation-module and generic-basis arguments agree with the algebraic facts in the Stacks Project, Tags 0539, 00NZ, 02NW and 051S.

5.29 Analytic coordinates on a skeletal chart

Let KK be complete and nontrivially valued, with its specified real-valued absolute value. Put O=K∘O=K^\circ, k=K~k=\widetilde K, v=−log⁡∣⋅∣v=-\log|\cdot|, and Γ=v(K×)\Gamma=v(K^\times). Consider a topologically finitely presented formally étalemap of admissible algebras A0=O⟨X0,…,Xr,T1,…,Tq⟩/(X0⋯Xr−a)⟶B0,0<∣a∣<1.A_0=O\langle X_0,\ldots,X_r,T_1,\ldots,T_q\rangle/ (X_0\cdots X_r-a)\longrightarrow B_0, \qquad 0<|a|<1. Write B=B0⊗OKB=B_0\otimes_O K, Y=M(B)Y=\mathcal M(B), and Σ⊂Y\Sigma\subset Y for the inverse image of the standard skeleton. The notation ΔI\Delta_I, QIQ_I, and xη,wx_{\eta,w} is that of Proposition 5.24.2 and Theorem 5.24.4. In particular, Σ\Sigma is a compact finite complex, and a point over the relative interior of ΔI\Delta_I is specified by ww and η∈QI\eta\in Q_I. A finite piecewise Γ\Gamma-integral-affine function has integral linear parts and constants in Γ\Gamma on a finite subdivision defined by integral linear inequalities with constants in Γ\Gamma.

Proposition 5.29.1. NAG-002249 (Finite piecewise-affine values). Let g∈Bg\in B be nonzero at every point of Σ\Sigma. Then v(g)∣Σv(g)|_\Sigma is finite piecewise Γ\Gamma-integral-affine. One finite hyperplane subdivision of the standard simplex induces a suitable subdivision on every skeletal cell.

Proof. Assume that Σ\Sigma is nonempty. Compactness gives α=min⁡Σ∣g∣>0\alpha=\min_\Sigma|g|>0. Use the bounded square-Jacobian presentation of Lemma 5.23.1, with root coordinates Y1,…,YmY_1,\ldots,Y_m, inverse determinant QQ, and integral polynomial equations F(Y)=0F(Y)=0 whose coefficients belong to A0A_0. Give its generic algebra the quotient Banach norm in which all these coordinates have norm at most one. Choose a polynomial P(X,T,Y,Q)P(X,T,Y,Q) such that ∥g−P(X,T,Y,Q)∥<ε<α.\|g-P(X,T,Y,Q)\|<\varepsilon<\alpha. Let M≥1M\geq1 bound the absolute values of its coefficients. The difference of two monomials, expanded one factor at a time, shows that PP is MM-Lipschitz on tuples of norm at most one.

Approximate the finitely many restricted coefficients of FF by integral polynomials in X,TX,T, with error δ<1\delta<1 and Mδ<εM\delta<\varepsilon, obtaining GG. At the original root, both G(Y)G(Y) and G′(Y)−F′(Y)G'(Y)-F'(Y) have norm at most δ\delta. The inverse of F′(Y)F'(Y) has norm at most one, since its adjugate is integral and QQ has norm at most one. The geometric inverse gives the same bound for G′(Y)−1G'(Y)^{-1}. Lemma 5.18.1 produces YG∈BmY_G\in B^m with G(YG)=0,∥YG−Y∥≤δ,∥YG∥≤1.G(Y_G)=0,\qquad \|Y_G-Y\|\leq\delta,\qquad \|Y_G\|\leq1. The determinant and reciprocal identities imply that QG=(det⁡G′(YG))−1Q_G=(\det G'(Y_G))^{-1} has norm at most one and ∥QG−Q∥≤δ\|Q_G-Q\|\leq\delta. Hence gG=P(X,T,YG,QG),∥gG−g∥<α,∣gG(x)∣=∣g(x)∣(x∈Σ).g_G=P(X,T,Y_G,Q_G),\qquad \|g_G-g\|<\alpha, \qquad |g_G(x)|=|g(x)|\quad(x\in\Sigma).

Put R=K[X0,…,Xr,T]/(X0⋯Xr−a),C=R[Y,Q]/(G,Qdet⁡G′−1),F0=Frac⁡(R).\begin{aligned} R&=K[X_0,\ldots,X_r,T]/(X_0\cdots X_r-a),\\ C&=R[Y,Q]/(G,Q\det G'-1),\qquad F_0=\operatorname{Frac}(R). \end{aligned} Eliminating X0X_0 identifies RR with a Laurent polynomial ring in X1,…,XrX_1,\ldots,X_r and a polynomial ring in TT, so it is a domain. The square-zero correction by the inverse Jacobian proves that C/RC/R is formally étale. It is finitely presented, and therefore étaleby [StacksAlg] (Tag 00UR). Its affine generic fibre is a finite product of finite separable extensions of F0F_0 by [StacksAlg] (Tag 02GL); finiteness here follows from quasi-compactness of this discrete fibre. Thus the image of PP satisfies a polynomial over F0F_0.

At a skeletal point, evaluation on RR is its standard Gauss norm and has zero kernel. The evaluation C→H(x)C\to\mathcal H(x) therefore extends to C⊗RF0C\otimes_R F_0. In particular this generic fibre is nonzero. Clear the denominators of an annihilating polynomial and remove its largest common power of the variable. Because gG(x)≠0g_G(x)\ne0, the resulting relation holds at every skeletal point and has nonzero constant and leading coefficients: ∑i=0sAigGi=0,Ai∈R.\sum_{i=0}^{s}A_i g_G^i=0,\qquad A_i\in R. There are at least two nonzero terms. Write λi(w)=v(Ai(zw))\lambda_i(w)=v(A_i(z_w)) for each nonzero AiA_i. A Laurent polynomial Gauss norm makes λi\lambda_i the minimum of finitely many Γ\Gamma-integral-affine functions. At least two terms of the displayed relation have the smallest valuation, whence v(g(xη,w))∈{λi(w)−λj(w)j−i:i<j, AiAj≠0}.v(g(x_{\eta,w}))\in \left\{\frac{\lambda_i(w)-\lambda_j(w)}{j-i}: i<j,\ A_iA_j\ne0\right\}. Subdivide the standard simplex by comparisons of coefficient monomials and then by comparisons of the resulting affine candidates. Multiplying the equations by positive integers leaves integral hyperplanes with constants in Γ\Gamma. On a connected relative open piece the distinct candidates have a fixed strict order. A continuous selection from them must select one throughout that piece. This gives finitely many rational-affine formulas, with a common subdivision on all the skeletal cells.

To establish integrality, fix a relative open cell. The proof of Theorem 5.25.3, before its use of the unit hypothesis, constructs a single coefficient field L0L_0 with v(L0×)=Γv(L_0^\times)=\Gamma and an open polyannulus parametrizing this cell’s Gauss points. Pull back gg to that polyannulus. The Laurent expansion constructed in Lemma 5.25.1 gives log⁡∣g(zw)∣=max⁡m{log⁡∣cm∣−⟨m,w⟩}.\log|g(z_w)|=\max_m\{\log|c_m|-\langle m,w\rangle\}. On a full-dimensional open piece where the left side is affine, choose a Laurent term attaining the maximum at an interior point. The difference between the left side and this term is affine, nonnegative, and zero at an interior point, so it is identically zero there. That piece consequently has an integral slope and a constant in Γ\Gamma. Continuity gives the same assertion on its faces. Each lower-dimensional piece is a face of a full-dimensional piece in its original cell. An original zero-dimensional cell is evaluated in L0L_0 itself. This proves the assertion on every cell. ◻

Lemma 5.29.2. NAG-002250 (Small scalar shifts). For h∈Bh\in B, at most finitely many c∈Kc\in K make h+ch+c vanish at a point of Σ\Sigma. Consequently a nonvanishing shift can be chosen with arbitrarily small positive ∣c∣|c|.

Proof. On each relative open cell use the fixed coefficient field and polyannulus from the preceding proof. If h+ch+c vanishes at one Gauss point, every Laurent coefficient vanishes. Its pullback is then the zero function on that polyannulus. Since the inclusion of KK is isometric, at most one scalar has this property on a given cell. The same argument for a zero-dimensional cell takes place in its coefficient field. There are finitely many cells. Powers of any π∈K\pi\in K with 0<∣π∣<10<|\pi|<1 give infinitely many distinct nonzero scalars converging to zero, so they avoid the finite exceptional set. ◻

Theorem 5.29.3. NAG-002251 (Faithful analytic coordinates). There are an ordinary open analytic neighbourhood VV of Σ\Sigma in YY and an analytic map φ:V→(Gm,KN)an\varphi:V\to(\mathbf G_{m,K}^{N})^{\mathrm{an}} whose tropicalization is a homeomorphism on Σ\Sigma onto its image. On every cell of a finite subdivision it is integral affine and induces an isomorphism onto the saturated lattice of its image.

Proof. Include the units X0,…,XrX_0,\ldots,X_r, whose product is aa; their valuations distinguish the base weights. For each II and η∈QI\eta\in Q_I, choose hˉη\bar h_\eta in the special-fibre algebra that belongs to every other prime in QIQ_I and does not belong to η\eta. Indeed these primes are incomparable, since they lie in the same zero-dimensional étalefibre; multiplying elements of η′∖η\eta'\setminus\eta for η′≠η\eta'\ne\eta gives the required element. Omit the construction when QIQ_I has fewer than two members. Lift hˉη\bar h_\eta to hη∈B0h_\eta\in B_0 and choose 0<∣cη∣<10<|c_\eta|<1 by Lemma 5.29.2 so that gη=hη+cηg_\eta=h_\eta+c_\eta is nonzero on Σ\Sigma. Specialization gives, for w∈relint⁡ΔIw\in\operatorname{relint}\Delta_I, ∣gη(xη,w)∣=1,0<∣gη(xη′,w)∣<1(η′≠η).|g_\eta(x_{\eta,w})|=1,\qquad 0<|g_\eta(x_{\eta',w})|<1\quad(\eta'\ne\eta). Thus this finite list separates points over the same weight.

Let VV be the complement of the zero loci of the gηg_\eta. It is an ordinary open neighbourhood of Σ\Sigma. Each gηg_\eta is an analytic unit on VV: near any point choose a positive lower bound for its norm and use the inverse on the corresponding rational domain, as in Theorem 2.31.2. These units and the XiX_i define φ\varphi. Its tropicalization is continuous and injective on Σ\Sigma, hence is a homeomorphism onto its image by compactness.

Apply Proposition 5.29.1 to the finite list. The union of its hyperplanes gives a common finite subdivision, compatible with all coordinate faces and their identifications. On a subcell the linear map is, after eliminating one standard coordinate, the graph n↦(n,Ln)n\mapsto(n,Ln) of an integral map. Projection to the first coordinates is an integral left inverse. If a point of its real image belongs to the ambient integer lattice, its first coordinates belong to the subcell’s saturated tangent lattice; the entire point is then the corresponding integral graph point. This proves the claimed lattice equality. ◻

5.30 Rational coordinates on a projective model

The passage to rational functions uses an algebraic model. The projective assumption also provides a common affine open on which the functions separating residual strata are bounded. The following elementary construction works over a finite residue field as well.

Lemma 5.30.1. NAG-002252 (Common affine neighbourhoods of residual points). Let OO be a local ring with residue field kk, and let P↪POnP\hookrightarrow\mathbf P^n_O be a closed immersion. Every finite subset Z⊂∣Pk∣Z\subset |P_k| is contained in an affine open of PP. If PkP_k is of finite type over kk, there is a finite family of affine opens Uj⊂PU_j\subset P such that each UjU_j contains ZZ and every two points of PkP_k lie together in some UjU_j.

Proof. Represent the points of ZZ by homogeneous primes of k[x0,…,xn]k[x_0,\ldots,x_n]. Discard any prime contained in another one in this finite list. Avoiding the remaining maximal primes also avoids all the discarded ones. Denote these pairwise incomparable primes by p1,…,ps\mathfrak p_1,\ldots, \mathfrak p_s. For i≠ji\ne j choose a homogeneous element aij∈pj∖pia_{ij}\in\mathfrak p_j\setminus\mathfrak p_i; one exists because both ideals are homogeneous. Put hi=∏j≠iaijh_i=\prod_{j\ne i}a_{ij}. Choose a coordinate xℓix_{\ell_i} outside pi\mathfrak p_i and an integer D≥1D\geq1 at least all the degrees of the hih_i. Then Hˉ=∑i=1shixℓiD−deg⁡hi\bar H=\sum_{i=1}^{s}h_i x_{\ell_i}^{D-\deg h_i} is homogeneous of positive degree and is outside every pi\mathfrak p_i: modulo pi\mathfrak p_i only its iith summand survives, and that summand is nonzero. For s=1s=1 the empty product is one, so the same construction applies. For Z=∅Z=\varnothing any coordinate suffices.

Lift the coefficients to a homogeneous polynomial HH over OO. The standard open D+(H)D_+(H) in projective space is affine by [StacksAlg] (Tag 01MB). Its intersection with PP is therefore an affine open containing ZZ.

Apply the first assertion to Z∪{p,q}Z\cup\{p,q\} for every p,q∈Pkp,q\in P_k. The open products Uk×kUkU_k\times_k U_k obtained in this way cover Pk×kPkP_k\times_k P_k: the two projections of a point of the product form such a pair. This product has a finite affine cover, obtained from the pairwise products of a finite affine cover of PkP_k. Each affine scheme is quasi-compact: refine an open cover by standard opens D(fα)D(f_\alpha). Their elements generate the unit ideal, since otherwise [StacksAlg] (Tag 00E0(2)) would give a maximal ideal outside the cover. An expression for 11 uses finitely many fαf_\alpha and gives a finite subcover. Choose a finite subcover. Finally, every underlying pair (p,q)(p,q) is the image of a point of the scheme product. Indeed κ(p)⊗kκ(q)\kappa(p)\otimes_k\kappa(q) is a nonzero ring: if 1⊗1=01\otimes1=0, this equality is witnessed by finitely many defining bilinear relations. Let E⊂κ(p)E\subset\kappa(p) and F⊂κ(q)F\subset\kappa(q) be the finite-dimensional kk-subspaces spanned by 11 and every entry in these relations. Choose linear functionals on EE and FF that take 11 to 11. Their product annihilates all the chosen relations but takes (1,1)(1,1) to 11, a contradiction. By [StacksAlg] (Tag 00E0(2)), the nonzero tensor ring has a maximal ideal, which is prime and provides the desired point. Thus the finite subcover has the stated pair property. ◻

Fix a projective, flat, finitely presented OO-scheme X\mathcal X with smooth integral generic fibre XX. Assume that Xs\mathcal X_s has a finite affine open cover Uλ=Spec⁡Cλ\mathcal U_\lambda=\operatorname{Spec}C_\lambda, where each Uλ\mathcal U_\lambda admits an étaleOO-morphism to Spec⁡O[X0,…,Xrλ,T1,…,Tqλ]/(X0⋯Xrλ−aλ),0<∣aλ∣<1.\operatorname{Spec} O[X_0,\ldots,X_{r_\lambda},T_1,\ldots,T_{q_\lambda}]/ (X_0\cdots X_{r_\lambda}-a_\lambda), \qquad 0<|a_\lambda|<1. Here an affine open cover of Xs\mathcal X_s means affine opens of X\mathcal X whose special fibres cover Xs\mathcal X_s. Complete along any π∈O\pi\in O with 0<∣π∣<10<|\pi|<1. Lemma 4.13.1 and Theorem 4.13.2 identify the completion with an admissible model of XanX^{\mathrm{an}}. The completed charts are formally étaleand form a finite formal-open cover: a square-zero lifting problem modulo πm\pi^m is the algebraic étalelifting problem, and its compatible solutions give the completed lifting property. Consequently Theorem 5.25.4 gives its compact finite integral-affine skeleton SS.

Proposition 5.30.2. NAG-002253 (Skeletal valuations of rational functions). Every point of SS is a norm on K(X)K(X). For every f∈K(X)×f\in K(X)^\times, the function v(f)∣Sv(f)|_S is continuous and finite piecewise Γ\Gamma-integral-affine. Moreover the set of specializations of points of SS is finite and is the set of generic residual-stratum points occurring in the completed charts.

Proof. Consider a chart meeting SS, and write CK=Cλ⊗OKC_K=C_\lambda\otimes_O K. It is a nonzero domain because Uλ,K\mathcal U_{\lambda,K} is a nonempty affine open of the integral scheme XX. The algebraic generic chart base R=K[X0,…,Xr,T]/(X0⋯Xr−a)R=K[X_0,\ldots,X_r,T]/(X_0\cdots X_r-a) is a domain. The map R→CKR\to C_K is étale, hence flat by [StacksAlg] (Tag 00U2(3),(6)). A nonzero flat algebra over a domain is torsion free, so this map is injective. For F0=Frac⁡(R)F_0=\operatorname{Frac}(R), the algebra CK⊗RF0C_K\otimes_R F_0 is a nonzero domain and an affine étalealgebra over F0F_0. By [StacksAlg] (Tag 02GL) it is a finite separable field extension. Since it contains CKC_K and is contained in Frac⁡(CK)\operatorname{Frac}(C_K), it equals Frac⁡(CK)=K(X)\operatorname{Frac}(C_K)=K(X).

For a point xx of the chart skeleton, evaluation on RR is the standard Gauss norm. It is nonzero on every nonzero element of RR. Thus the map CK→H(x)C_K\to\mathcal H(x) extends to CK⊗RF0=K(X)C_K\otimes_R F_0=K(X). This is a unital map from a field, hence injective. The absolute value on H(x)\mathcal H(x) gives the asserted norm on K(X)K(X). All points of SS belong to one of these charts.

On each chart write f=g/hf=g/h with nonzero g,h∈CKg,h\in C_K. Their images are analytic on its completed generic fibre and are nonzero at every skeletal point by the preceding argument. Proposition 5.29.1 applies to both. Their valuation difference is continuous and has finitely many integral-affine formulas there. The finite cover gives the assertion globally, with the formulas agreeing on common faces by evaluation of the same rational function.

By Proposition 5.24.2, a point of a chart skeleton over w∈relint⁡ΔIw\in\operatorname{relint}\Delta_I specializes to its label η∈QI\eta\in Q_I. Conversely every such label has its nonempty relative open simplex. Each QIQ_I is finite, there are finitely many II and charts, and the formal specializations agree on overlaps. This proves the last assertion. ◻

Theorem 5.30.3. NAG-002254 (Faithful rational tropicalization of a projective skeleton). In the stated projective algebraic model setting, there are a dense open U⊂XU\subset X with S⊂UanS\subset U^{\mathrm{an}} and a closed immersion φ:U↪Gm,KN\varphi:U\hookrightarrow\mathbf G_{m,K}^N such that trop⁡∘φan∣S\operatorname{trop}\circ\varphi^{\mathrm{an}}|_S is a homeomorphism onto its image. On every cell of a finite subdivision of SS, this map is Γ\Gamma-integral-affine and identifies the tangent lattice with the saturated lattice of its image. All coordinates belong to the original field K(X)K(X) and use the specified absolute value of KK.

Proof. Algebraic chart coordinates. For each chart include its multiplicative coordinates X0,…,XrX_0,\ldots,X_r as nonzero rational functions on XX. Proposition 5.30.2 makes all their norms finite and positive on all of SS. On each relative open cell of their own chart their valuations are the standard simplex coordinates. Thus they distinguish two points with the same residual label and provide an integral left inverse on each cell’s tangent lattice.

Separation of residual labels. Let Z⊂XsZ\subset\mathcal X_s be the finite set of skeletal specializations from the preceding proposition. Use Lemma 5.30.1 to choose one affine open W=Spec⁡D\mathcal W=\operatorname{Spec}D of X\mathcal X containing all of ZZ. Every point of SS belongs to the generic fibre of W^\widehat{\mathcal W}, because this generic fibre is the inverse image of Ws\mathcal W_s under specialization; see Lemma 4.13.1 and the formal-open restriction statement in Proposition 5.23.4. Elements of DD are therefore bounded by one at every point of SS.

View the members of ZZ as distinct prime ideals of D⊗OkD\otimes_O k. For each unordered distinct pair choose an orientation (p,q)(\mathfrak p,\mathfrak q) with p⊈q\mathfrak p\not\subseteq\mathfrak q, choose hˉ∈p∖q\bar h\in\mathfrak p\setminus\mathfrak q, and lift it to h∈Dh\in D. Flatness embeds DD in the coordinate ring of its generic open, hence in K(X)K(X). Since hˉ≠0\bar h\ne0, this is a nonzero rational function. At a skeletal point specializing to p\mathfrak p its norm is less than one; at one specializing to q\mathfrak q its norm is one. These values follow directly from the definition of specialization for a bounded formal function. Add all these finitely many lifts to the coordinate list.

The resulting tropical map is injective on SS. Indeed the new functions distinguish different specializations. If x,yx,y have the same specialization η\eta, choose an algebraic chart containing η\eta. Both points belong to its completed generic fibre and its skeleton, by formal-open restriction. In that chart they have the same residual label, and equality of the multiplicative coordinate values gives the same ww. The bijection in Proposition 5.24.2 then gives x=yx=y.

A closed immersion of a dense open. Let f1,…,fMf_1,\ldots,f_M denote the list just constructed. Choose a nonempty affine open V=Spec⁡AV=\operatorname{Spec}A of XX and nonzero algebra generators t1,…,tbt_1,\ldots,t_b for the finitely generated integral KK-algebra AA; zero generators can be omitted. Write fi=pi/qif_i=p_i/q_i with nonzero pi,qi∈Ap_i,q_i\in A, and put s=∏jtj∏ipiqi,U=Spec⁡A[1/s].s=\prod_j t_j\prod_i p_iq_i, \qquad U=\operatorname{Spec}A[1/s]. Append the functions tj,pi,qit_j,p_i,q_i to the list. They are units on UU, and the induced homomorphism from the Laurent polynomial coordinate ring of the torus onto A[1/s]A[1/s] is surjective: it contains the generators tjt_j of AA and the inverse of their displayed product ss. This gives a closed immersion. The norm statement in Proposition 5.30.2 says that every point of SS has generic scheme support. It therefore belongs to the analytification of every nonempty open of XX, in particular to UanU^{\mathrm{an}}.

Lattices and topology. All coordinates of the final list are finite piecewise Γ\Gamma-integral-affine on SS. Subdivide each original cell so that these finitely many coordinates are affine. Its original chart coordinates remain among them, and projection to those coordinates gives an integral left inverse. The graph argument in Theorem 5.29.3 shows that the resulting map on every such subcell is unimodular. Their images are finitely many Γ\Gamma-integral polyhedra. To see this explicitly, the coordinate projection identifies each image with its source polyhedron, and its other coordinates are the graphs of integral-affine formulas with constants in Γ\Gamma. Cut the finite family of image polyhedra by their supporting hyperplanes. As in Lemma 5.14.2, this gives a finite common polyhedral subdivision. Pulling it back under the injective map gives a subdivision compatible with every face identification in SS, and preserves the integral left inverses on tangent lattices. Finally SS is compact, and its continuous injective map to RN\mathbf R^N is a homeomorphism onto its image. ◻

References. The faithful tropicalization theorem is a reconstruction, for the stated projective unmarked algebraic models, of the method in Gubler–Rabinoff–Werner, Skeletons and tropicalizations, arXiv:1404.7044v3, Section 9. The affine construction above uses homogeneous prime avoidance; its projective standard-open input is [StacksAlg] (Tag 01MB).

5.31 Finite projections of analytic tropical fibres

Let KK be complete for its given real-valued non-Archimedean absolute value. The first two results allow the trivial valuation and generalized affinoid algebras. For an affinoid space XX, write ∂KX=X∖Int⁡(X/M(K)).\partial_KX=X\setminus\operatorname{Int}(X/\mathcal M(K)). For a tuple r=(r1,…,rd)r=(r_1,\ldots,r_d) of positive real numbers put Ar=K{r−1T,rS}/(TiSi−1)i=1d,Tr=M(Ar).A_r=K\{r^{-1}T,rS\}/(T_iS_i-1)_{i=1}^d, \qquad \mathbb T_r=\mathcal M(A_r). Thus Tr\mathbb T_r is the exact-radius polytorus ∣Ti∣=ri|T_i|=r_i. All quotient ideals below are actual algebraic ideals, and all tensor products with a hat are separated completed tensor products.

Theorem 5.31.1. NAG-002256 (Boundary avoidance gives an actual finite algebra). Let X=M(B)X=\mathcal M(B) be a generalized KK-affinoid space and let g1,…,gd∈B×g_1,\ldots,g_d\in B^\times. For ri>0r_i>0 set Fr={x∈X:∣gi(x)∣=ri for every i},Br=B{r−1U,rV}/(Ui−gi,Vi−gi−1)i=1d.\begin{gathered} F_r=\{x\in X:|g_i(x)|=r_i\text{ for every }i\},\\ B_r=B\{r^{-1}U,rV\}/(U_i-g_i,V_i-g_i^{-1})_{i=1}^d. \end{gathered} Suppose Fr∩∂KX=∅F_r\cap\partial_KX=\varnothing. Then BrB_r is finite over ArA_r under Ti↦UiT_i\mapsto U_i, Si↦ViS_i\mapsto V_i, and its given affinoid norm is equivalent to its canonical finite-module norm. The spectrum of this algebra is the actual affinoid domain FrF_r, with its induced analytic structure. The assertion includes the empty domain and uses the original field and absolute value.

Proof. The equations defining BrB_r are a joint Weierstrass presentation: they impose ∣gi∣≤ri|g_i|\leq r_i and ∣gi−1∣≤ri−1|g_i^{-1}|\leq r_i^{-1}. Theorems 2.31.1 and 2.31.2 identify its spectrum, algebra, and domain mapping property with FrF_r. They also cover the zero algebra. We may therefore assume B≠0B\ne0.

Choose positive real numbers ai,bia_i,b_i such that ai<ri<bi,∥gi∥<bi,∥gi−1∥<ai−1.a_i<r_i<b_i,\qquad \|g_i\|<b_i,\qquad \|g_i^{-1}\|<a_i^{-1}. The larger closed polyannulus has algebra C=K{b−1T,aS}/(TiSi−1)i=1d.C=K\{b^{-1}T,aS\}/(T_iS_i-1)_{i=1}^d. Substitution gives a bounded map C→BC\to B by Proposition 1.9.3. The exact-radius polytorus is an affinoid domain of M(C)\mathcal M(C). Its algebra is ArA_r: applying the rational universal property in both directions identifies the joint inequalities ∣Ti∣≤ri|T_i|\leq r_i and ∣Si∣≤ri−1|S_i|\leq r_i^{-1} with its displayed presentation.

For x∈Frx\in F_r, the hypothesis says that x∈Int⁡(M(B)/M(K))x\in\operatorname{Int}(\mathcal M(B)/\mathcal M(K)). Apply the affinoid composition formula 2.44.1 to M(B)⟶M(C)⟶M(K).\mathcal M(B)\longrightarrow\mathcal M(C) \longrightarrow\mathcal M(K). It gives x∈Int⁡(M(B)/M(C))x\in\operatorname{Int}(\mathcal M(B)/\mathcal M(C)). By the actual rational base-change presentation and Proposition 2.33.8, B⊗^CAr≃BrB\widehat\otimes_C A_r\simeq B_r boundedly, with bounded inverse. The bounded maps from BB and ArA_r to BrB_r agree on CC, so the pushout gives one map. In the pushout the images of gig_i and gi−1g_i^{-1} are the images of the two coordinates of ArA_r; their powers therefore have bounds by constant multiples of rimr_i^m and ri−mr_i^{-m}. Bounded substitution gives the reverse map from BrB_r. Both composites fix all coefficient and coordinate images, hence are identities by polynomial density. The base-change assertion of Proposition 2.33.21 now puts every point of M(Br)\mathcal M(B_r) in its relative interior over Tr\mathbb T_r. Theorem 2.43.8 proves both finite generation over ArA_r and equivalence with the canonical finite module norm. ◻

Proposition 5.31.2. NAG-002257 (The factors of a finite projected level). Let f1,…,fn∈B×f_1,\ldots,f_n\in B^\times define t:X⟶Rn,t(x)=(−log⁡∣f1(x)∣,…,−log⁡∣fn(x)∣).t:X\longrightarrow\mathbb R^n, \qquad t(x)=(-\log|f_1(x)|,\ldots,-\log|f_n(x)|). Let Q=(qji)∈Md×n(Z)Q=(q_{ji})\in M_{d\times n}(\mathbb Z) and w′∈Rdw'\in\mathbb R^d. Suppose w′∉Q(t(∂KX)),Q−1(w′)∩t(X)={w1,…,ws}\begin{gathered} w'\notin Q(t(\partial_KX)),\\ Q^{-1}(w')\cap t(X)=\{w_1,\ldots,w_s\} \end{gathered} is finite. Put gj=∏ifiqjig_j=\prod_i f_i^{q_{ji}} and rj=exp⁡(−wj′)r_j=\exp(-w'_j). Then the actual algebras of the fibres t−1(wi)t^{-1}(w_i) give a bounded product decomposition Br≃∏i=1sBwi.B_r\simeq\prod_{i=1}^s B_{w_i}. Each BwiB_{w_i} is finite over ArA_r with its canonical module norm. For every ξ∈Tr\xi\in\mathbb T_r, with H=H(ξ)H=\mathcal H(\xi), these are actual finite analytic fibre algebras and dim⁡H(Br⊗^ArH)=∑i=1sdim⁡H(Bwi⊗^ArH).\dim_H(B_r\widehat\otimes_{A_r}H) =\sum_{i=1}^s \dim_H(B_{w_i}\widehat\otimes_{A_r}H). The empty finite set has zero product algebra and degree zero.

Proof. The monomials gjg_j are analytic units and satisfy (−log⁡∣gj∣)j=Q∘t(-\log|g_j|)_j=Q\circ t. The boundary condition therefore gives the hypothesis of Theorem 5.31.1. Its finite algebra BrB_r represents (Q∘t)−1(w′)(Q\circ t)^{-1}(w'). The continuous map tt on this space has the finite image {w1,…,ws}\{w_1,\ldots,w_s\}. Each inverse image is consequently open and closed. It is also the actual rational domain defined by ∣fj∣=exp⁡(−(wi)j)(1≤j≤n).|f_j|=\exp(-(w_i)_j)\quad(1\leq j\leq n). The rational universal property identifies this domain with the one obtained by imposing the same equalities directly on XX. These domains are pairwise disjoint and form a finite affinoid cover of M(Br)\mathcal M(B_r). The norm-strict equalizer of Theorem 2.38.1 is thus precisely the asserted bounded product isomorphism, since all intersections have zero algebra. Each factor is a direct summand of the finite ArA_r-module BrB_r. Theorem 1.18.3 identifies its induced and canonical module norms.

By Proposition 2.33.19, tensoring these finite modules with HH is already complete. Algebraic tensor product preserves the finite direct sum underlying the product, so dimensions add as displayed. Theorem 2.33.11 identifies these tensor algebras with the actual analytic fibres over ξ\xi. ◻

5.32 Finite quotient models retaining moment coordinates

Now assume that KK is nontrivially valued, and put O=K∘O=K^\circ, m=K∘∘\mathfrak m=K^{\circ\circ}, k=O/mk=O/\mathfrak m. Choose π∈O\pi\in O with 0<∣π∣<10<|\pi|<1 and set A0=O⟨T1±1,…,Td±1⟩,A=A0[1/π].A_0=O\langle T_1^{\pm1},\ldots,T_d^{\pm1}\rangle, \qquad A=A_0[1/\pi]. The integral algebra A0A_0 is the unit ball for the Laurent Gauss norm, and A~=k[T1±1,…,Td±1]\widetilde A=k[T_1^{\pm1},\ldots,T_d^{\pm1}]. These identifications follow from the actual Laurent normal form 1.17.2 and the multiplicative Gauss norm 1.2.2.

Theorem 5.32.1. NAG-002259 (A finite formal model containing prescribed bounded elements). Let A→DA\to D be a bounded map of generalized KK-affinoid algebras. Assume that DD is finite over AA and its given norm is equivalent to its canonical finite-AA-module norm. Let h1,…,hm∈Dh_1,\ldots,h_m\in D be power-bounded elements. There is an admissible topologically finitely presented OO-algebra D0⊂DD_0\subset D, finite over A0A_0, such that D0[1/π]=D,h1,…,hm∈D0.D_0[1/\pi]=D, \qquad h_1,\ldots,h_m\in D_0. It can be chosen as the actual image of a finite restricted-series presentation A0⟨Y⟩→DA_0\langle Y\rangle\to D. Its generic-fibre Banach structure is the given one on DD.

Let ξ\xi be the Gauss point of M(A)\mathcal M(A) and H=H(ξ)H=\mathcal H(\xi). For this model put C0=(D0/mD0)⊗k[T±1]k(T1,…,Td).C_0=(D_0/\mathfrak mD_0) \otimes_{k[T^{\pm1}]} k(T_1,\ldots,T_d). Then C0C_0 is a finite-dimensional algebra, and dim⁡H(D⊗^AH)=dim⁡k(T)C0=∑n∈Max⁡C0length⁡(C0)n((C0)n)[κ(n):k(T)].\begin{equation} \begin{aligned} \dim_H(D\widehat\otimes_AH) &=\dim_{k(T)}C_0\\ &=\sum_{\mathfrak n\in\operatorname{Max}C_0} \operatorname{length}_{(C_0)_{\mathfrak n}} ((C_0)_{\mathfrak n}) [\kappa(\mathfrak n):k(T)]. \end{aligned} \end{equation} For fixed A→DA\to D, this total degree is independent of the chosen presentation and of the chosen admissible finite A0A_0-model with generic algebra DD.

Proof. For D=0D=0 use D0=0D_0=0. Otherwise choose finite AA-module generators b1,…,bℓb_1,\ldots,b_\ell. Multiplying each by a sufficiently small nonzero power of π\pi, obtain generators e1,…,eℓe_1,\ldots,e_\ell with ∥ej∥≤1\|e_j\|\leq1. Append the prescribed hih_i to this list, and denote the resulting power-bounded elements by z1,…,zNz_1,\ldots,z_N. By Proposition 1.9.3, evaluation defines a bounded surjection q:P=A⟨Y1,…,YN⟩⟶D,Yi⟼zi.q:P=A\langle Y_1,\ldots,Y_N\rangle\longrightarrow D, \qquad Y_i\longmapsto z_i. It is norm-admissible. Indeed the first ℓ\ell images still generate DD as an AA-module, and Theorem 1.18.3 gives uniformly controlled coefficients for a linear lift in these generators. That linear polynomial is a lift through qq. Consequently the quotient norm from PP is equivalent to the given norm on DD. This presentation also proves that DD is strict over KK.

Put P0=A0⟨Y1,…,YN⟩P_0=A_0\langle Y_1,\ldots,Y_N\rangle, J=ker⁡qJ=\ker q, and D0=P0/(J∩P0)⊂D.D_0=P_0/(J\cap P_0)\subset D. The map on the right is injective by construction. In particular D0D_0 has no OO-torsion. Since P0P_0 is topologically finite type over OO, Proposition 3.14.5, applied to the finite P0P_0-module P0/(J∩P0)P_0/(J\cap P_0), proves that its quotient π\pi-adic topology is complete and separated. Theorem  3.14.4 makes it topologically finitely presented. Also P0[1/π]=PP_0[1/\pi]=P, by bounding the coefficients of a restricted series, whence D0[1/π]=DD_0[1/\pi]=D. All the hih_i lie in D0D_0.

It remains to prove finiteness over A0A_0. The finite map A→DA\to D is a map of strict affinoids. By Theorem 3.16.5, D~\widetilde D is finite over A~\widetilde A. For each ziz_i choose a monic relation for its residue over k[T±1]k[T^{\pm1}], and lift its coefficients to a monic pi(Z)∈A0[Z]p_i(Z)\in A_0[Z]. Then wi=pi(zi)∈D∘∘.w_i=p_i(z_i)\in D^{\circ\circ}. In the quotient norm from qq, ∥win∥→0\|w_i^n\|\to0; topological nilpotence is unchanged by the equivalent norm just established. For each of the finitely many ii, choose ai≥1a_i\geq1 and an actual lift Gi∈PG_i\in P of wiaiw_i^{a_i} with ∥Gi∥<1\|G_i\|<1. Such a lift exists by the definition of the quotient infimum once ∥wiai∥<1\|w_i^{a_i}\|<1. Since there are finitely many lifts, θ=max⁡i∥Gi∥<1.\theta=\max_i\|G_i\|<1. In particular Gi∈P0G_i\in P_0 and pi(Yi)ai−Gi∈J∩P0p_i(Y_i)^{a_i}-G_i\in J\cap P_0. The strict Laurent coefficient norm of a nonzero GiG_i attains its maximum at a coefficient in KK. Choosing such a coefficient cic_i gives Gi=ciGi′G_i=c_iG'_i with Gi′∈P0G'_i\in P_0 and ci∈mc_i\in\mathfrak m. The zero lift has the same conclusion. Thus every GiG_i vanishes modulo mD0\mathfrak mD_0.

Reduction of P0P_0 is the polynomial algebra k[T±1,Y1,…,YN]k[T^{\pm1},Y_1,\ldots,Y_N], since restricted coefficients are eventually smaller than one. The images of the YiY_i in D0/mD0D_0/\mathfrak mD_0 satisfy the monic equations p‾i(Yi)ai=0\overline p_i(Y_i)^{a_i}=0. The bounded powers below their degrees span this quotient over k[T±1]k[T^{\pm1}]. Lemma 5.28.2 now applies to the actual admissible quotient A0→D0A_0\to D_0. It proves that D0D_0 is finite over A0A_0, compares its generic norm with the canonical finite module norm, and gives the first equality in (5.32.1).

For the second equality, decompose the finite-dimensional Artinian algebra C0C_0 into its local factors by [StacksAlg] (Tag 00JB). A composition series of the factor (C0)n(C_0)_{\mathfrak n} has length⁡(C0)n((C0)n)\operatorname{length}_{(C_0)_{\mathfrak n}}((C_0)_{\mathfrak n}) successive quotients equal to κ(n)\kappa(\mathfrak n). Taking vector-space dimensions proves the formula. In particular the local lengths are those of the actual quotient algebra. Finally any admissible finite A0A_0-model of the same DD has finite reduction, so Lemma 5.28.2 identifies its total degree with the same left side. This proves independence. ◻

For an application to the preceding projected fibres, suppose rj=∣cj∣r_j=|c_j| with cj∈K×c_j\in K^\times; the change of coordinate Tj↦cjTjT_j\mapsto c_jT_j identifies the target with the unit polytorus over KK. If the full moment weight wiw_i also has exp⁡(−(wi)a)=∣ba∣\exp(-(w_i)_a)=|b_a| with ba∈K×b_a\in K^\times, the functions fa/baf_a/b_a and ba/fab_a/f_a have absolute value one on t−1(wi)t^{-1}(w_i). They are power bounded by Theorem 1.20.2. Including both in the prescribed list places the normalized moment characters and their inverses in D0D_0. The resulting finite model and the degree in (5.32.1) use these original-field coordinates with the original absolute value.

5.33 Common models for finite moment projections

Fix a complete nontrivially valued field KK with its given real-valued absolute value, and write O=K∘,m=K∘∘,k=O/m.O=K^\circ,\qquad \mathfrak m=K^{\circ\circ},\qquad k=O/\mathfrak m. Choose π∈O\pi\in O with 0<∣π∣<10<|\pi|<1. An actual admissible model of a KK-affinoid algebra DD is an admissible topologically finitely presented OO-subalgebra D0⊂DD_0\subset D for which D0[1/π]=DD_0[1/\pi]=D and the generic-fibre Banach topology is the given topology of DD. The special algebra is always the actual quotient D‾0=D0/mD0\overline D_0=D_0/\mathfrak mD_0.

Proposition 5.33.1. NAG-002280 (Simultaneous finite models). For 1≤i≤r1\leq i\leq r, let Ai,0=O⟨Ui,1±1,…,Ui,d±1⟩,Ai=Ai,0[1/π],A_{i,0}=O\langle U_{i,1}^{\pm1},\ldots,U_{i,d}^{\pm1}\rangle, \qquad A_i=A_{i,0}[1/\pi], and let Ai→DA_i\to D be bounded homomorphisms. Assume that DD is finite over every AiA_i and that its given norm is equivalent to each canonical finite-module norm.

Every actual admissible model D0⊂DD_0\subset D containing the images of all the Ui,jU_{i,j} and Ui,j−1U_{i,j}^{-1} is finite over every Ai,0A_{i,0}. Such models exist and may contain any prescribed finite list of power-bounded elements of DD.

More generally, let E1,0,…,Es,0E_{1,0},\ldots,E_{s,0} be actual admissible models of DD, each containing the target coordinates and their inverses for at least one of the maps Ai→DA_i\to D. There is an actual admissible model C0⊂DC_0\subset D containing all the Ej,0E_{j,0} and all target coordinates and inverses. It is finite over every Ai,0A_{i,0} and over every Ej,0E_{j,0}.

Proof. An actual admissible model D0D_0 is a bounded open OO-submodule of DD. Indeed, take a formal presentation O⟨Z⟩↠D0O\langle Z\rangle\twoheadrightarrow D_0. Its generic presentation has the quotient Banach topology. The image of the Gauss unit ball is bounded and contains every element whose quotient norm is strictly less than one, by the definition of the quotient infimum. Equivalence with the given norm proves the assertion. In particular, every element of D0D_0 is power bounded: all its powers lie in the bounded set D0D_0. The same argument shows that πD0\pi D_0 is open in DD.

The images of Ui,jU_{i,j} and their inverses define a continuous formal map Ai,0→D0A_{i,0}\to D_0. Restricted-series substitution is valid in the complete π\pi-adic ring D0D_0, and agrees in DD with the given bounded homomorphism. Fix ii, and choose formal generators z1,…,zNz_1,\ldots,z_N of D0D_0 over OO. The generic algebra DD is strict, since it is the generic fibre of this model. Theorem 3.16.5 makes its canonical reduction D~\widetilde D finite over k[Ui±1]k[U_i^{\pm1}]. For each zaz_a, lift a monic equation of its canonical residue to a monic polynomial pa(Z)∈Ai,0[Z]p_a(Z)\in A_{i,0}[Z]. Its value pa(za)p_a(z_a) is topologically nilpotent in DD. Since πD0\pi D_0 is open, some positive power of this value belongs to πD0\pi D_0. Consequently the image z‾a∈D‾0\overline z_a\in\overline D_0 satisfies a monic equation p‾a(Z)na=0\overline p_a(Z)^{n_a}=0 over k[Ui±1]k[U_i^{\pm1}].

The algebra D‾0\overline D_0 is generated by the finitely many z‾a\overline z_a over this Laurent ring. The powers below the degrees of these monic equations span it as a module. Thus D‾0\overline D_0 is finite over k[Ui±1]k[U_i^{\pm1}]. Lemma 5.28.2 gives finiteness of D0D_0 over Ai,0A_{i,0}. This argument applies separately to every ii and does not require flatness over Ai,0A_{i,0}.

For existence, use Theorem 5.32.1 with A1→DA_1\to D. Append to its prescribed list all target coordinates and inverses, and all formal generators of all Ej,0E_{j,0}. The target elements are power bounded because the maps Ai→DA_i\to D are bounded. The model generators are power bounded by the first paragraph. The resulting actual model C0C_0 contains every Ej,0E_{j,0}: restricted series in its formal generators converge in C0C_0, and their values in DD are the original values. It is finite over all Ai,0A_{i,0} by the part already proved. If Ai,0A_{i,0} maps into Ej,0E_{j,0}, any finite list of Ai,0A_{i,0}-module generators of C0C_0 also generates it over Ej,0E_{j,0}. This proves the last assertion. For D=0D=0 all the models may be taken to be zero. ◻

Definition 5.33.2. (Top-dimensional special cycles). For a finite-type kk-algebra RR of dimension at most dd, put [Spec⁡R]d=∑p∈Min⁡Rdim⁡(R/p)=dlength⁡Rp(Rp)[Spec⁡(R/p)].[\operatorname{Spec}R]_d =\sum_{\substack{\mathfrak p\in\operatorname{Min}R\\ \dim(R/\mathfrak p)=d}} \operatorname{length}_{R_{\mathfrak p}}(R_{\mathfrak p}) [\operatorname{Spec}(R/\mathfrak p)]. For a finite morphism of finite-type kk-schemes, the pushforward of an integral component is its image multiplied by the full degree of the extension of function fields. Pushforward of cycles is extended additively. These are cycles before passage to rational equivalence.

Proposition 5.33.3. NAG-002282 (Component lengths under a finite model comparison). Let D0⊂C0⊂DD_0\subset C_0\subset D be actual admissible models of the same affinoid algebra, both finite over A0=O⟨U1±1,…,Ud±1⟩.A_0=O\langle U_1^{\pm1},\ldots,U_d^{\pm1}\rangle. The inclusion induces a finite morphism ν:Spec⁡C‾0→Spec⁡D‾0\nu:\operatorname{Spec}\overline C_0\to \operatorname{Spec}\overline D_0, and ν∗[Spec⁡C‾0]d=[Spec⁡D‾0]d.\nu_*[\operatorname{Spec}\overline C_0]_d =[\operatorname{Spec}\overline D_0]_d. More precisely, for every dd-dimensional minimal prime p\mathfrak p of D‾0\overline D_0, length⁡((D‾0)p)=∑q↦plength⁡((C‾0)q)[κ(q):κ(p)],\begin{equation} \operatorname{length}((\overline D_0)_{\mathfrak p}) =\sum_{\mathfrak q\mapsto\mathfrak p} \operatorname{length}((\overline C_0)_{\mathfrak q}) [\kappa(\mathfrak q):\kappa(\mathfrak p)], \end{equation} where the sum is over the dd-dimensional minimal primes of C‾0\overline C_0 above p\mathfrak p.

Proof. Finite A0A_0-module generators of C0C_0 also generate it over D0D_0, so ν\nu is finite. Both special algebras are finite over k[U±1]k[U^{\pm1}] and are Noetherian by [StacksAlg] (Tag 00FN). Their components have dimension at most dd. By [StacksAlg] (Tag 00P0), a component has dimension dd exactly when it dominates Spec⁡k[U±1]\operatorname{Spec}k[U^{\pm1}]: its function field is finite over the fraction field of the corresponding prime quotient of k[U±1]k[U^{\pm1}], and a nonzero prime imposes a nonzero polynomial relation on its dd Laurent generators.

Let S⊂A0S\subset A_0 be the elements of Laurent Gauss norm one. The Gauss-localization argument of Lemma 5.28.2 identifies V=S−1A0,F=Frac⁡(A0),κ(V)=k(U).V=S^{-1}A_0,\qquad F=\operatorname{Frac}(A_0),\qquad \kappa(V)=k(U). Here VV is a valuation ring, and every nonzero element of VV is a nonzero scalar from OO times a unit. The finite modules M=S−1D0M=S^{-1}D_0 and N=S−1C0N=S^{-1}C_0 are therefore torsion free over VV and free by the finite-generator argument in that lemma. They are subalgebras of the same generic algebra E=D⊗A0[1/π]F.E=D\otimes_{A_0[1/\pi]}F. Both span EE over FF. They consequently have the same rank nn. Their residue algebras are B=D‾0⊗k[U±1]k(U),C=C‾0⊗k[U±1]k(U).B=\overline D_0\otimes_{k[U^{\pm1}]}k(U),\qquad C=\overline C_0\otimes_{k[U^{\pm1}]}k(U). They retain the full generic local rings of the corresponding special fibres. If n=0n=0, there are no components to compare.

By [StacksAlg] (Tag 00JB), BB and CC are products of their Artinian local factors. Fix a maximal ideal p\mathfrak p of BB, and let b‾∈B\overline b\in B be the idempotent which is one on BpB_{\mathfrak p} and zero on the other factors. Choose any lift b∈Mb\in M. Multiplication by bb on the two free VV-modules MM and NN has the same characteristic polynomial: after scalar extension to FF, both operators are multiplication by the same element on EE. Equality over VV follows from V⊂FV\subset F.

Reducing this equality to k(U)k(U) gives tn−a(t−1)a=tn−a′(t−1)a′,a=dim⁡k(U)Bp,a′=dim⁡k(U)(b‾C).t^{n-a}(t-1)^a=t^{n-a'}(t-1)^{a'},\qquad a=\dim_{k(U)} B_{\mathfrak p},\quad a'=\dim_{k(U)}(\overline b C). The factors tt and t−1t-1 are relatively prime in every characteristic, so a=a′a=a'. The idempotent b‾\overline b selects exactly the local factors CqC_{\mathfrak q} whose maximal ideals contract to p\mathfrak p in BB. Taking a composition series of each Artinian local factor therefore gives length⁡(Bp)[κ(p):k(U)]=∑q↦plength⁡(Cq)[κ(q):k(U)].\operatorname{length}(B_{\mathfrak p}) [\kappa(\mathfrak p):k(U)] =\sum_{\mathfrak q\mapsto\mathfrak p} \operatorname{length}(C_{\mathfrak q}) [\kappa(\mathfrak q):k(U)]. Divide by the positive integer [κ(p):k(U)][\kappa(\mathfrak p):k(U)] and use the tower formula for field degrees. Localization identifies these local rings with those in (5.33.1), proving that formula and the cycle identity. No local factor has been replaced by its reduced quotient. ◻

5.34 The residual cycle of a finite analytic moment fibre

Let NN be a lattice, M=N∨M=N^\vee, and TN=Spec⁡k[M]\mathbb T_N=\operatorname{Spec}k[M]. A normalized moment tuple on DD is a homomorphism h:M⟶D×h:M\longrightarrow D^\times such that every character value and its inverse are power bounded. A basis and its inverses suffice for this condition. For Q:N→ZdQ:N\to\mathbf Z^d, put AQ=K⟨U1±1,…,Ud±1⟩,Uj⟼h(Q∗ej).A_Q=K\langle U_1^{\pm1},\ldots,U_d^{\pm1}\rangle, \qquad U_j\longmapsto h(Q^*e_j). We consider those QQ for which this map is finite and the given norm on DD is equivalent to the canonical finite-AQA_Q-module norm. Write HQH_Q for the completed residue field of the Gauss point of M(AQ)\mathcal M(A_Q), and put δQ=dim⁡HQ(D⊗^AQHQ).\delta_Q=\dim_{H_Q}(D\widehat\otimes_{A_Q}H_Q). An actual model is compatible with hh if it contains the values of a basis of MM and their inverses.

Theorem 5.34.1. NAG-002285 (A residual moment cycle independent of the projection). Suppose at least one such finite projection QQ exists. For any actual model D0D_0 compatible with hh, the residual moment morphism ϕ0:Spec⁡D‾0⟶TN\phi_0:\operatorname{Spec}\overline D_0\longrightarrow\mathbb T_N is finite, and the effective cycle Cd(D,h)=(ϕ0)∗[Spec⁡D‾0]d\mathcal C_d(D,h)=(\phi_0)_* [\operatorname{Spec}\overline D_0]_d is independent of the model and of the finite projection used to construct it. If Cd(D,h)=∑ZcZ[Z]\mathcal C_d(D,h)=\sum_Z c_Z[Z], then every finite projection QQ restricts to a finite dominant map qZ:Z→TZdq_Z:Z\to\mathbb T_{\mathbf Z^d}, and δQ=∑ZcZ[k(Z):k(U1,…,Ud)].\begin{equation} \delta_Q=\sum_Z c_Z [k(Z):k(U_1,\ldots,U_d)]. \end{equation}

The cycles also respect monomial changes of moment coordinates. Precisely, let a:N→N′a:N\to N' and h′=h∘a∗h'=h\circ a^*, and suppose that h′h' admits a finite projection to a dd-dimensional unit polytorus for the same actual algebra DD. The residual monomial map is finite on every component of Cd(D,h)\mathcal C_d(D,h) and Cd(D,h′)=a∗Cd(D,h).\mathcal C_d(D,h')=a_*\mathcal C_d(D,h).

Proof. Proposition 5.33.1 supplies compatible models and makes every such model finite over the formal target of every finite projection. Fix one projection QQ. The map k[U±1]→D‾0k[U^{\pm1}]\to\overline D_0 factors through the residual moment homomorphism k[M]→D‾0k[M]\to\overline D_0. Its finite module generators therefore also generate D‾0\overline D_0 over k[M]k[M]. This proves that ϕ0\phi_0 is finite.

For two compatible models, the same proposition gives a common model C0C_0 containing both and finite over both. All three models are finite over the target of the fixed QQ. Proposition 5.33.3 identifies the pushforwards of the top-dimensional cycle of Spec⁡C‾0\operatorname{Spec}\overline C_0 with those of both models. Their residual moment maps agree. Pushforwards of these finite cycles compose because degrees of finite extensions of function fields multiply. Pushing to TN\mathbb T_N proves the asserted independence, including comparison of models initially constructed using different projections.

For any other finite projection, use the same compatible model. Each dd-dimensional component of its special fibre dominates TZd\mathbb T_{\mathbf Z^d} by the dimension argument in Proposition 5.33.3. Its residual moment image ZZ is also dd-dimensional. The coordinate ring of ZZ lies between the image of k[U±1]k[U^{\pm1}] and the coordinate domain of the source component. The latter is finite over k[U±1]k[U^{\pm1}]. Every generator of the intermediate ring is consequently integral over k[U±1]k[U^{\pm1}], and finitely many generators satisfying monic equations make it finite. Thus qZq_Z is finite and dominant.

By Lemma 5.28.2 and the Artinian length formula in Theorem 5.32.1, δQ=∑dim⁡(D‾0/p)=dlength⁡((D‾0)p)[κ(p):k(U)].\delta_Q =\sum_{\dim(\overline D_0/\mathfrak p)=d} \operatorname{length}((\overline D_0)_{\mathfrak p}) [\kappa(\mathfrak p):k(U)]. Group the terms having the same residual moment image ZZ. The definition of cZc_Z and the field-degree tower formula give exactly (5.34.1).

For the last assertion, a model compatible with hh is also compatible with h′h'. It is finite over a target of a finite projection for h′h'. On every component, the coordinate ring of the residual hh-image is intermediate between the ring of the residual h′h'-image and a finite algebra over the latter. The same monic-equation argument proves finiteness of the induced map of image components. The two residual moment maps now compose through aa, and multiplication of the full field degrees proves the cycle identity. ◻

Theorem 5.34.2. NAG-002287 (The degree of a subtorus-invariant residual cycle). In the preceding theorem let L⊂NL\subset N be a saturated sublattice of rank dd. Suppose every reduced integral support ZZ of Cd(D,h)\mathcal C_d(D,h) is stable, as a closed subscheme, under translation by the split subtorus TL\mathbb T_L. Let ρ:TN→TN/L\rho:\mathbb T_N\to\mathbb T_{N/L} be the quotient map. There are distinct closed points s1,…,sts_1,\ldots,s_t of TN/L\mathbb T_{N/L} and positive integers c1,…,ctc_1,\ldots,c_t with Cd(D,h)=∑j=1tcj[ρ−1(sj)].\mathcal C_d(D,h) =\sum_{j=1}^t c_j[\rho^{-1}(s_j)]. The empty sum is allowed. For every finite projection QQ for which Q∣LQ|_L is injective, put eQ=[Zd:Q(L)].e_Q=[\mathbf Z^d:Q(L)]. Then δQ=eQμ,μ=∑j=1tcj[κ(sj):k]∈Z≥0.\begin{equation} \delta_Q=e_Q\mu, \qquad \mu=\sum_{j=1}^t c_j[\kappa(s_j):k]\in\mathbf Z_{\geq0}. \end{equation} In particular, δQ/eQ\delta_Q/e_Q is independent of all these projections and is positive exactly when the residual cycle is nonzero.

Proof. Lemma 5.14.1 gives a splitting N=L⊕N1N=L\oplus N_1. Put M0=(N/L)∨M_0=(N/L)^\vee and ML=L∨M_L=L^\vee. Then k[M]=k[M0][ML]k[M]=k[M_0][M_L], and translation by TL\mathbb T_L is the character grading by MLM_L.

Let I⊂k[M]I\subset k[M] be the prime ideal of an integral support ZZ. The scheme-theoretic invariance hypothesis says that the coaction sends II to I⊗kk[ML]I\otimes_k k[M_L]. If f=∑αfαχα∈If=\sum_\alpha f_\alpha\chi^\alpha\in I, with fα∈k[M0]f_\alpha\in k[M_0], independence of the characters in the second tensor factor shows that each fαχαf_\alpha\chi^\alpha belongs to II. Multiplication by the inverse character then puts fαf_\alpha in I∩k[M0]I\cap k[M_0]. Thus I=(I∩k[M0])k[M],k[Z]=(k[M0]/(I∩k[M0]))[ML].I=(I\cap k[M_0])k[M],\qquad k[Z]=(k[M_0]/(I\cap k[M_0]))[M_L]. The coefficient algebra is a finite-type domain. Since dim⁡Z=d\dim Z=d, the dimension–transcendence-degree equality [StacksAlg] (Tag 00P0) makes its fraction field algebraic over kk. Its finitely many algebra generators are therefore algebraic over kk. They span a finite-dimensional domain by their monic equations, and this domain is a field. It is κ(s)\kappa(s) for a closed point ss of TN/L\mathbb T_{N/L}. This proves that Z=ρ−1(s)Z=\rho^{-1}(s). Distinct integral supports give distinct points, establishing the first formula.

Fix one of these points, and write k′=κ(s)k'=\kappa(s). Under the splitting its fibre has coordinate ring k′[ML]k'[M_L]. The target characters for QQ have the form Ui⟼diχai,di∈(k′)×.U_i\longmapsto d_i\chi^{a_i}, \qquad d_i\in(k')^\times. The aia_i generate a sublattice of MLM_L of index eQe_Q; equivalently, this is the dual index of the injective map Q∣LQ|_L. Choose one representative for each coset. Every character has a unique expression as a representative times a product of the χai\chi^{a_i}, and the units did_i can be absorbed into its coefficient. These representatives are therefore a basis of k′[ML]k'[M_L] over k′[U±1]k'[U^{\pm1}]. A kk-basis of k′k' gives a basis of size eQ[k′:k]e_Q[k':k] over k[U±1]k[U^{\pm1}]. This is the same character-coset calculation as in Proposition 5.28.3. The generic field degree of ρ−1(s)→TZd\rho^{-1}(s)\to \mathbb T_{\mathbf Z^d} is consequently eQ[k′:k]e_Q[k':k]. Substitution in (5.34.1) of Theorem 5.34.1 gives (5.34.2). Full residue-field degrees, including inseparable degrees, enter this computation. The positive coefficients cjc_j already contain the unreduced generic lengths from the actual models. ◻

Theorem 5.34.3. NAG-002289 (Projection comparison at a flat analytic tropical germ). Let X=M(B)X=\mathcal M(B) be a generalized KK-affinoid space and let f1,…,fn∈B×f_1,\ldots,f_n\in B^\times define the moment map t:X⟶Rn,t(x)=(−log⁡∣f1(x)∣,…,−log⁡∣fn(x)∣).t:X\longrightarrow\mathbf R^n,\qquad t(x)=(-\log|f_1(x)|,\ldots,-\log|f_n(x)|). Let w∈Rnw\in\mathbf R^n, and choose bj∈K×b_j\in K^\times with ∣bj∣=exp⁡(−wj)|b_j|=\exp(-w_j). Put hj=fj/bjh_j=f_j/b_j, let D=BwD=B_w be the actual algebra of Fw=t−1(w)F_w=t^{-1}(w), and set N=ZnN=\mathbf Z^n. Suppose there are a saturated rank-dd sublattice L⊂NL\subset N and a neighborhood WW of ww such that t(X)∩W⊂w+LR,Fw∩∂KX=∅.t(X)\cap W\subset w+L_{\mathbf R},\qquad F_w\cap\partial_KX=\varnothing. Assume that the normalized moment tuple on DD admits a finite projection to a dd-dimensional unit polytorus with its canonical finite-module norm.

Every integral support of Cd(D,h)\mathcal C_d(D,h) is a fibre of TN→TN/L\mathbb T_N\to\mathbb T_{N/L} over a closed point. Consequently, for any finite projections Q0,Q1Q_0,Q_1 of this same actual algebra with Qi∣LQ_i|_L injective, dim⁡HQ0(D⊗^AQ0HQ0)[Zd:Q0(L)]=dim⁡HQ1(D⊗^AQ1HQ1)[Zd:Q1(L)]=∑scs[κ(s):k]∈Z≥0,\frac{\dim_{H_{Q_0}}(D\widehat\otimes_{A_{Q_0}}H_{Q_0})} {[\mathbf Z^d:Q_0(L)]} =\frac{\dim_{H_{Q_1}}(D\widehat\otimes_{A_{Q_1}}H_{Q_1})} {[\mathbf Z^d:Q_1(L)]} =\sum_s c_s[\kappa(s):k]\in\mathbf Z_{\geq0}, where Cd(D,h)=∑scs[ρ−1(s)]\mathcal C_d(D,h)=\sum_s c_s[\rho^{-1}(s)].

Proof. The functions hjh_j and hj−1h_j^{-1} have absolute value one on FwF_w, so they are power bounded in DD by Theorem 1.20.2. Theorem 5.34.1 therefore applies. If its cycle is zero, its support assertion is empty and the degree formula follows from Theorem 5.34.2.

Choose ε>0\varepsilon>0 so small that the closed box {v:∣vj−wj∣≤ε for all j}\{v:|v_j-w_j|\leq\varepsilon\text{ for all }j\} is contained in WW. Let Y⊂XY\subset X be its inverse image, the actual rational affinoid domain e−ε≤∣hj∣≤eε(1≤j≤n).e^{-\varepsilon}\leq|h_j|\leq e^{\varepsilon} \quad(1\leq j\leq n). Its presentation by the hjh_j and hj−1h_j^{-1} is the actual joint Laurent presentation of Theorem 2.31.2. At every x∈Fwx\in F_w, all the relative coordinate inequalities are strict, since 1<eε1<e^{\varepsilon}. Thus x∈Int⁡(Y/X)x\in\operatorname{Int}(Y/X). The boundary hypothesis and the composition inclusion in Proposition 2.33.21 give Fw⊂Int⁡(Y/K).F_w\subset\operatorname{Int}(Y/K). Imposing the equalities ∣hj∣=1|h_j|=1 on YY gives the same actual algebra DD as imposing them on XX, again by the rational universal property.

Put M0=M∩LR⊥M_0=M\cap L_{\mathbf R}^{\perp}. For u∈M0u\in M_0, the unit hu=∏jhjujh_u=\prod_j h_j^{u_j} has absolute value one at every point of YY, because t(Y)−w⊂LRt(Y)-w\subset L_{\mathbf R}. Theorem 1.20.2 and bounded substitution, Proposition 1.9.3, give a morphism hu:Y⟶M(K⟨T±1⟩).h_u:Y\longrightarrow\mathcal M(K\langle T^{\pm1}\rangle). For x∈Fwx\in F_w, apply the forward implication for absolute evaluation to the actual collar YY, using x∈Int⁡(Y/K)x\in\operatorname{Int}(Y/K). The image SxS_x of O(Y)~gr\widetilde{\mathcal O(Y)}^{\rm gr} in H(x)~gr\widetilde{\mathcal H(x)}^{\rm gr} is finite over K~gr\widetilde K^{\rm gr}. Choose finitely many homogeneous module generators for SxS_x. The component of K~gr\widetilde K^{\rm gr} that takes any given generator to degree one has dimension at most one over kk. Thus (Sx)1(S_x)_1 is a finite-dimensional kk-algebra. The identity ∣hu∣=1\lvert h_u\rvert=1 on YY gives spectral radius one, and ∣hu(x)∣=1\lvert h_u(x)\rvert=1 makes its degree-one initial image the nonzero residue of hu(x)h_u(x) in H(x)~\widetilde{\mathcal H(x)}. This residue lies in (Sx)1(S_x)_1 and is therefore algebraic over kk.

Choose an actual model D0D_0 compatible with hh and a dd-dimensional minimal prime p\mathfrak p of D‾0\overline D_0. The formal specialization theorem 3.16.15 supplies x∈M(D)=Fwx\in\mathcal M(D)=F_w specializing to p\mathfrak p. By the definition of specialization, evaluation induces an injection D‾0/p↪H(x)~,\overline D_0/\mathfrak p \hookrightarrow\widetilde{\mathcal H(x)}, and hence an injection of its fraction field. The preceding paragraph shows that the residue of every huh_u, u∈M0u\in M_0, is algebraic over kk in this fraction field.

Let ZZ be the residual moment image of this component. Its coordinate ring is the image of k[M]k[M] in that field, and dim⁡Z=d\dim Z=d by finiteness of the residual moment map. The image of k[M0]k[M_0] is a finitely generated algebraic domain over kk, hence a finite field k′k' by monic equations and finite-dimensional linear algebra. It corresponds to a closed point ss of TN/L\mathbb T_{N/L}. Split N=L⊕N1N=L\oplus N_1 using Lemma 5.14.1. The coordinate ring of ZZ is then an integral quotient of k′[L∨]k'[L^\vee]. A nonzero kernel would give a nonzero Laurent polynomial relation on its dd character generators and lower the transcendence degree below dd. This contradicts dim⁡Z=d\dim Z=d and [StacksAlg] (Tag 00P0). The quotient map is therefore an isomorphism. Thus Z=ρ−1(s)Z=\rho^{-1}(s) as a closed subscheme. This treats every support of the residual cycle, and Theorem 5.34.2 now gives the displayed degree comparison. ◻

5.35 Purity and degrees near a projected analytic level

Throughout this section KK is complete for its specified real-valued non-Archimedean absolute value. Generalized affinoid algebras and the trivial valuation are allowed. An affinoid algebra BB is pure of analytic dimension dd if dim⁡KM(B/p)=d\dim_K\mathcal M(B/\mathfrak p)=d for every minimal prime p\mathfrak p of BB. We use analytic dimension before strictification; it need not equal the Krull dimension of a generalized affinoid algebra.

Lemma 5.35.1. NAG-002291 (Components under a free-radius extension). Let BB be a generalized KK-affinoid algebra, let rr be a finite KK-free polyradius, and put L=KrL=K_r and BL=B⊗^KLB_L=B\widehat\otimes_KL. If BB is integral, then BLB_L is integral. For arbitrary nonzero BB, extension of ideals gives a bijection Min⁡(B)⟶Min⁡(BL),p⟼pBL.\operatorname{Min}(B)\longrightarrow\operatorname{Min}(B_L), \qquad \mathfrak p\longmapsto\mathfrak pB_L. Moreover BL/pBL≃(B/p)L,dim⁡LM(BL/pBL)=dim⁡KM(B/p).B_L/\mathfrak pB_L\simeq(B/\mathfrak p)_L, \qquad \dim_L\mathcal M(B_L/\mathfrak pB_L) =\dim_K\mathcal M(B/\mathfrak p). Consequently BB is pure of analytic dimension dd if and only if BLB_L is pure of analytic dimension dd.

Proof. Suppose first that BB is integral. By Lemma 2.52.1, there is x∈M(B)x\in\mathcal M(B) with kernel zero. In the coefficient description of Theorem 1.16.2, its Gauss lift is ∣∑νbνTν∣σr(x)=max⁡ν∣bν(x)∣rν\left|\sum_\nu b_\nu T^\nu\right|_{\sigma_r(x)} =\max_\nu |b_\nu(x)|r^\nu by Theorem 1.19.1. A nonzero coefficient family has a nonzero coefficient, whose value at xx is positive. Thus this multiplicative seminorm has kernel zero, and BLB_L is a domain. No freeness over H(x)\mathcal H(x) is used.

The actual quotient identity follows from Theorem 1.18.8. Applied to B/pB/\mathfrak p, the preceding argument shows that pBL\mathfrak pB_L is prime. Its contraction is p\mathfrak p, by the constant embedding in (B/p)L(B/\mathfrak p)_L. Every prime of BLB_L contracts to a prime of BB containing a minimal prime, by [StacksAlg] (Tag 00E0(4)). It therefore contains one of the primes pBL\mathfrak pB_L. These extended primes cannot contain one another, since their contractions are distinct minimal primes. They are exactly the minimal primes of BLB_L. Finally apply Theorem 3.21.8 to each quotient. This proves the dimension formula and both purity implications. ◻

Lemma 5.35.2. NAG-002292 (Dimension at maximal ideals of a strict affinoid). Let BB be a nonzero strict KK-affinoid algebra and let m\mathfrak m be a maximal ideal. Then dim⁡Bm=max⁡p∈Min⁡(B)p⊆mdim⁡(B/p).\dim B_{\mathfrak m} =\max_{\substack{\mathfrak p\in\operatorname{Min}(B)\\ \mathfrak p\subseteq\mathfrak m}} \dim(B/\mathfrak p). In particular, if BB is pure of dimension dd, then every maximal localization has dimension dd.

Proof. First let R=K⟨T1,…,Te⟩R=K\langle T_1,\ldots,T_e\rangle and let n\mathfrak n be maximal. Theorem 1.12.3 makes E=R/nE=R/\mathfrak n a finite extension of KK. Write aia_i for the image of TiT_i; boundedness of the residue point gives ∣ai∣≤1|a_i|\leq1. A finite KK-basis of EE, together with the norm comparison of Lemma 1.12.1, identifies E⟨T⟩E\langle T\rangle with the finite free RR-module R⊗KER\otimes_KE: expanding coefficients in that basis preserves convergence in both directions. This ring extension is integral by [StacksAlg] (Tag 00GK).

In E⟨T⟩E\langle T\rangle, the coordinate ideals (0)⊂(T1−a1)⊂⋯⊂(T1−a1,…,Te−ae)(0)\subset(T_1-a_1)\subset\cdots\subset (T_1-a_1,\ldots,T_e-a_e) form a strict chain of prime ideals. Translation by the aia_i is a bounded automorphism, and each quotient is a Tate algebra in the remaining variables. The last ideal contracts to n\mathfrak n. Incomparability for an integral extension, [StacksAlg] (Tag 00GT), makes all contractions in this chain distinct. Hence dim⁡Rn≥e\dim R_{\mathfrak n}\geq e. The reverse inequality follows from dim⁡R=e\dim R=e, proved in Proposition 3.21.1.

Now suppose that BB is integral, of dimension ee. Choose a finite injection R=K⟨T1,…,Te⟩↪BR=K\langle T_1,\ldots,T_e\rangle\hookrightarrow B by Theorem 1.12.2 and Proposition 3.21.1. The contraction n=m∩R\mathfrak n=\mathfrak m\cap R is maximal: its quotient is a KK-subalgebra of the finite extension B/mB/\mathfrak m, hence is a finite-dimensional domain and a field by [StacksAlg] (Tag 00GS(1)). The ring RR is normal by Theorem 1.13.1. Going down for an integral extension over a normal domain, [StacksAlg] (Tag 00H8), lifts a chain of length ee ending at n\mathfrak n to a chain ending at the specified m\mathfrak m. Thus dim⁡Bm=e\dim B_{\mathfrak m}=e.

For general BB, apply this result to B/pB/\mathfrak p for each minimal prime contained in m\mathfrak m. These are finitely many, by [StacksAlg] (Tag 00FR) and Theorem 1.11.5. Their localized spectra cover Spec⁡(Bm)\operatorname{Spec}(B_{\mathfrak m}): any prime in a chain contains a minimal prime of BB, by [StacksAlg] (Tag 00E0(4)). A chain is therefore a chain in one such quotient. Taking the maximal possible length gives the displayed equality. ◻

Theorem 5.35.3. NAG-002293 (Purity of affinoid domains). Let X=M(B)X=\mathcal M(B) be pure of analytic dimension dd. Every nonempty affinoid domain V=M(C)⊆XV=\mathcal M(C)\subseteq X is pure of analytic dimension dd.

Proof. First assume that KK is nontrivially valued and that BB and CC are strict. The restriction B→CB\to C is flat by Proposition 2.38.2. If n\mathfrak n is maximal in CC, then C/nC/\mathfrak n is finite over KK by Theorem 1.12.3. The image of BB in this field is a finite-dimensional domain, hence a field by [StacksAlg] (Tag 00GS(1)). Thus m=n∩B\mathfrak m=\mathfrak n\cap B is maximal. Flat going down, [StacksAlg] (Tag 00HS), and Lemma 5.35.2 give dim⁡Cn≥d.\dim C_{\mathfrak n}\geq d. On the other hand, Proposition 3.22.8 gives dim⁡C≤d\dim C\leq d. Analytic and Krull dimensions agree here because both algebras are strict.

Fix a minimal prime q\mathfrak q of CC. The minimal primes are finite by [StacksAlg] (Tag 00FR). Choose an element hh in all the other minimal primes and outside q\mathfrak q: for each other prime choose an element outside q\mathfrak q and take their product; take h=1h=1 if there is no other prime. By Lemma 2.52.1, the integral affinoid C/qC/\mathfrak q has a point where hh does not vanish. Choose c∈K×c\in K^\times with ∣c∣≤∣h∣|c|\leq|h| at that point. The strict Laurent domain M((C/q)⟨U⟩/(hU−c))\mathcal M\bigl((C/\mathfrak q)\langle U\rangle/ (hU-c)\bigr) is nonempty by Theorem 2.31.1. A maximal residue field of its algebra is finite over KK. Its map from CC has maximal kernel n\mathfrak n, contains q\mathfrak q, and does not contain hh. Hence qCn\mathfrak qC_{\mathfrak n} is the only minimal prime of CnC_{\mathfrak n}. The spectra of this local ring and its quotient by that prime have the same prime chains. It follows that d≤dim⁡Cn=dim⁡(C/q)n≤dim⁡(C/q)≤d.d\leq\dim C_{\mathfrak n} =\dim(C/\mathfrak q)_{\mathfrak n} \leq\dim(C/\mathfrak q)\leq d. Every minimal-prime quotient of CC consequently has dimension dd.

For arbitrary KK, BB, and CC, choose one free-radius extension L=KrL=K_r making both algebras strict, by Theorem 1.17.6. If KK is trivially valued, include a radius different from 11 among the additional radii, so that LL is nontrivially valued. Theorem 2.36.9 identifies M(CL)\mathcal M(C_L) with the corresponding affinoid domain in M(BL)\mathcal M(B_L). Lemma 5.35.1 makes BLB_L pure of Krull dimension dd and ensures CL≠0C_L\ne0. The strict case makes CLC_L pure of dimension dd. Apply the same lemma to CC to descend this purity, with analytic dimension, to the original field. ◻

For a nonempty closed rectangular box P=∏j=1d[αj,βj]⊂RdP=\prod_{j=1}^d[\alpha_j,\beta_j]\subset\mathbb R^d, allow αj=βj\alpha_j=\beta_j and put ℓj=e−βj,uj=e−αj,AP=K{u−1T,ℓS}/(TjSj−1)j.\ell_j=e^{-\beta_j},\qquad u_j=e^{-\alpha_j},\qquad A_P=K\{u^{-1}T,\ell S\}/(T_jS_j-1)_j. Its spectrum TP\mathbb T_P is the closed polyannulus ℓj≤∣Tj∣≤uj\ell_j\leq|T_j|\leq u_j. For s∈Ps\in P, let ηs\eta_s be its Gauss point, with ∣∑mamTm∣ηs=max⁡m∣am∣e−⟨m,s⟩|\sum_m a_mT^m|_{\eta_s}=\max_m|a_m|e^{-\langle m,s\rangle}, and put Hs=H(ηs)H_s=\mathcal H(\eta_s).

Proposition 5.35.4. NAG-002294 (Gauss degrees over a closed polyannulus). The algebra APA_P is an integral domain of analytic dimension dd, and every ηs\eta_s has kernel zero. Let DD be a finite APA_P-algebra with its canonical affinoid norm. Then dim⁡Hs(D⊗^APHs)=dim⁡Frac⁡(AP)(D⊗APFrac⁡(AP))\dim_{H_s}(D\widehat\otimes_{A_P}H_s) =\dim_{\operatorname{Frac}(A_P)} (D\otimes_{A_P}\operatorname{Frac}(A_P)) for every s∈Ps\in P. In particular this integer is independent of ss.

If D≠0D\ne0 is pure of analytic dimension dd, then every minimal prime of DD contracts to zero in APA_P, the displayed integer is positive, and M(D)→TP\mathcal M(D)\to\mathbb T_P is surjective. Every point above ηs\eta_s satisfies dK(x)=dd_K(x)=d.

Proof. Iterating Theorem 2.13.2 gives unique Laurent expansions in APA_P, with the finite-tail condition for the norm max⁡m∣am∣max⁡s∈Pe−⟨m,s⟩.\max_m |a_m|\max_{s\in P} e^{-\langle m,s\rangle}. The operators for different coordinate pairs commute on polynomials, and their bounds extend this assertion to the completed coefficient spaces. Thus their common section and actual relation kernel give the indicated expansion also when some intervals have length zero. The weighted Gauss argument of Proposition 1.2.2 is multiplicative at each ss. Every nonzero expansion has a nonzero coefficient, so its value at every ηs\eta_s is positive. This proves both assertions about integrality and kernels.

At any point of TP\mathbb T_P, Laurent polynomials are dense in the evaluated affinoid algebra. Approximating a nonzero denominator closely enough preserves its value and approximates its inverse. The completed residue field is therefore the completion of a field generated by at most dd elements over KK. Lemma 3.22.3 bounds its residual dimension by dd. At ηs\eta_s this field is the weighted Gauss completion of K(T1,…,Td)K(T_1,\ldots,T_d), whose residual dimension is dd by Lemma 3.22.4. Theorem 3.22.7 now gives dim⁡KTP=d\dim_K\mathbb T_P=d.

The kernel-zero map AP→HsA_P\to H_s extends to its fraction field. Since DD is finite, Proposition 1.18.7 identifies the completed tensor in the statement with the ordinary tensor. Extension of a finite-dimensional vector space from Frac⁡(AP)\operatorname{Frac}(A_P) to HsH_s preserves its dimension. This proves the displayed equality, without a flatness assumption on DD.

Suppose now that D≠0D\ne0 is pure of analytic dimension dd, and fix a minimal prime q\mathfrak q of DD. Set E=D/qE=D/\mathfrak q. Choose a free-radius extension L/KL/K making APA_P and EE strict. Then ELE_L is integral by Lemma 5.35.1, and has Krull dimension dd by Theorem 3.21.8. The same statements hold for (AP)L(A_P)_L. The finite-module comparison in Theorem 1.18.8 preserves the finite module presentation, so ELE_L is finite over (AP)L(A_P)_L. If the kernel JJ of this map were nonzero, any prime chain in (AP)L/J(A_P)_L/J could be extended by the zero prime of the domain (AP)L(A_P)_L. Hence dim⁡((AP)L/J)≤d−1\dim((A_P)_L/J)\leq d-1. The finite injection (AP)L/J↪EL(A_P)_L/J\hookrightarrow E_L is integral by [StacksAlg] (Tag 00GK), so [StacksAlg] (Tag 00OK) gives a contradiction. Thus J=0J=0. The isometric scalar embedding then shows that AP→EA_P\to E is injective as well. This holds for every minimal prime, and in particular AP→DA_P\to D is injective.

The generic algebra is consequently nonzero, proving positivity. For an arbitrary y∈TPy\in\mathbb T_P with kernel p\mathfrak p, lying over for the integral injection AP↪DA_P\hookrightarrow D, [StacksAlg] (Tag 00GQ), makes D⊗APκ(p)D\otimes_{A_P}\kappa(\mathfrak p) nonzero. Extending this vector space to H(y)\mathcal H(y) preserves nonzeroness. The finite tensor comparison and Theorem 1.20.1 therefore give a point of the actual analytic fibre, identified by Theorem 2.33.11. This proves surjectivity. Finally Lemma 3.22.6 and the Gauss residual-dimension calculation give dK(x)=dd_K(x)=d at every point above ηs\eta_s. ◻

Theorem 5.35.5. NAG-002295 (A finite polyannulus neighborhood of a level). Let X=M(B)X=\mathcal M(B), let g1,…,gd∈B×g_1,\ldots,g_d\in B^\times, and set q:X⟶Rd,q(x)=(−log⁡∣g1(x)∣,…,−log⁡∣gd(x)∣).q:X\longrightarrow\mathbb R^d, \qquad q(x)=(-\log|g_1(x)|,\ldots,-\log|g_d(x)|). Suppose s0∈q(X)∖q(∂KX)s_0\in q(X)\setminus q(\partial_KX). There is a closed rectangular neighborhood PP of s0s_0, with s0s_0 in its interior, such that Y=q−1(P)Y=q^{-1}(P) is an actual affinoid domain of XX, Y⊂Int⁡(X/K)Y\subset\operatorname{Int}(X/K), and g:Y⟶TPg:Y\longrightarrow\mathbb T_P is finite with the canonical finite-module norm on its algebra.

If XX is pure of analytic dimension dd, this map is surjective. Its Gauss-fibre dimension is one fixed positive integer on PP. For every s∈Ps\in P, its restriction to the exact-radius polytorus at ss is the actual algebra of the whole level q−1(s)q^{-1}(s).

Proof. The set ∂KX\partial_KX is closed, by the openness of affinoid interior in Proposition 2.33.21. It is compact by Proposition 2.24.1; its image under qq is compact and hence closed. Choose a closed box PP around s0s_0 disjoint from that image. The simultaneous inequalities on gjg_j and gj−1g_j^{-1} give the domain YY with its actual affinoid algebra, by Theorems 2.31.1 and 2.31.2.

Choose a larger closed polyannulus TP+\mathbb T_{P^+} whose coordinate bounds strictly exceed the norms of all gjg_j and gj−1g_j^{-1} and which contains TP\mathbb T_P. Bounded substitution, as in Theorem 5.31.1, gives X→TP+X\to\mathbb T_{P^+}. At every point of YY, the affinoid composition formula 2.44.1 makes this morphism inner, since the point is in Int⁡(X/K)\operatorname{Int}(X/K). The actual relative pushout of Proposition 2.33.8 identifies the algebra of YY with B⊗^AP+AP.B\widehat\otimes_{A_{P^+}}A_P. Indeed both presentations impose exactly the stated bounds on the images of gjg_j and gj−1g_j^{-1}; the rational universal maps in both directions fix these images and the coefficients. Interior is preserved by base change, by Proposition 2.33.21. Hence every point of YY is inner over TP\mathbb T_P. Theorem 2.43.8 proves finiteness and the norm assertion.

If XX is pure of analytic dimension dd, then YY is nonempty and pure of the same dimension by Theorem 5.35.3. Proposition 5.35.4 gives surjectivity and a constant positive Gauss degree. Base change from APA_P to the exact-radius polytorus algebra imposes the equalities ∣gj∣=e−sj|g_j|=e^{-s_j}, so the same actual rational universal property identifies it with the whole level algebra. This includes every real ss, without changing the original absolute value. ◻

Proposition 5.35.6. NAG-002296 (Constant degrees for separated groups of moment fibres). Let X=M(B)X=\mathcal M(B) be pure of analytic dimension dd, let f1,…,fn∈B×f_1,\ldots,f_n\in B^\times define t:X→Rnt:X\to\mathbb R^n, and let Q∈Md×n(Z)Q\in M_{d\times n}(\mathbb Z). Suppose s0∉Q(t(∂KX)),Q−1(s0)∩t(X)={w1,…,wm}≠∅s_0\notin Q(t(\partial_KX)),\qquad Q^{-1}(s_0)\cap t(X)=\{w_1,\ldots,w_m\}\ne\varnothing is finite. Choose pairwise disjoint neighborhoods NiN_i of the wiw_i. There are closed rectangular boxes Wi⊂NiW_i\subset N_i with wi∈Int⁡(Wi)w_i\in\operatorname{Int}(W_i) and a closed rectangular neighborhood PP of s0s_0 such that Y=(Q∘t)−1(P)=∐i=1mYi,Yi=Y∩t−1(Wi),t(Yi)⊂Int⁡(Wi).Y=(Q\circ t)^{-1}(P)=\coprod_{i=1}^m Y_i, \qquad Y_i=Y\cap t^{-1}(W_i), \qquad t(Y_i)\subset\operatorname{Int}(W_i). This is a finite disjoint affinoid cover and gives a bounded product decomposition D=∏iDiD=\prod_iD_i over APA_P. Each DiD_i is finite over APA_P, and there are positive integers δi\delta_i such that dim⁡Hs(Di⊗^APHs)=δi(s∈P).\dim_{H_s}(D_i\widehat\otimes_{A_P}H_s)=\delta_i \qquad(s\in P). At s=s0s=s_0, the algebra being measured is the actual finite analytic fibre of BwiB_{w_i} over the projected Gauss point. Thus each original moment fibre has positive projected Gauss degree, extending to the constant Gauss degree of its separated group. Whenever Q−1(s)∩t(X)Q^{-1}(s)\cap t(X) is finite, δi\delta_i is the sum of the actual projected Gauss degrees of its moment fibres whose values lie in WiW_i.

Proof. Take closed boxes WiW_i as stated, still pairwise disjoint. Put q=Q∘tq=Q\circ t; its coordinates are the logarithms of the units gj=∏kfkqjkg_j=\prod_k f_k^{q_{jk}}. The compact set X∖t−1 ⁣(⋃iInt⁡(Wi))X\setminus t^{-1}\!\left(\bigcup_i \operatorname{Int}(W_i)\right) has qq-image disjoint from s0s_0. Shrink the box supplied by Theorem 5.35.5 to avoid this compact image. Then every point of YY belongs to exactly one YiY_i, and its tt-image lies in the interior of the corresponding box. Each YiY_i is the actual Laurent domain obtained by imposing the coordinate bounds defining WiW_i, by Theorem 2.31.2. The sets are both open and closed in YY.

The norm-strict equalizer of Theorem 2.38.1 for this disjoint cover is the asserted product isomorphism: all pairwise intersections have zero algebra. Its factors are direct summands of the finite APA_P-module DD, so they are finite; their induced norms are canonical by Theorem 1.18.3. Each contains t−1(wi)t^{-1}(w_i) and is nonempty; Theorem 5.35.3 makes it pure of analytic dimension dd. Proposition 5.35.4 gives the constants δi>0\delta_i>0.

At s0s_0, the imposed equalities Q∘t=s0Q\circ t=s_0 leave only w1,…,wmw_1,\ldots,w_m as possible moment values. Within WiW_i the remaining domain is exactly t−1(wi)t^{-1}(w_i). The rational universal property identifies its algebra with BwiB_{w_i}, and identifies the base change of DiD_i to the exact-radius polytorus with that algebra. Laurent polynomial density identifies the completed residue fields of its Gauss point and ηs0\eta_{s_0}. The finite tensor comparison of Proposition 1.18.7 and Theorem 2.33.11 now identify the tensor in the statement with the actual fibre. If the joint moment image over another s∈Ps\in P is finite, apply Proposition 5.31.2 at ss and retain the factors belonging to WiW_i. Finite products commute with these tensors, so their dimensions add, proving the last assertion. ◻

5.36 Gauss degrees at arbitrary real radii

Let KK be complete for its specified real-valued non-Archimedean absolute value, which may be trivial. For r∈R>0dr\in\mathbb R_{>0}^d put AK,r=K{r−1T,rU}/(TiUi−1)i=1d.A_{K,r}=K\{r^{-1}T,rU\}/(T_iU_i-1)_{i=1}^d. The actual Laurent presentation of Theorem 1.17.2 identifies this algebra isometrically with the convergent Laurent series ∑m∈ZdamTm,∣am∣rm⟶0outside finite subsets,∥∑mamTm∥r=max⁡m∣am∣rm.\begin{gathered} \sum_{m\in\mathbb Z^d}a_mT^m, \qquad |a_m|r^m\longrightarrow0 \quad\text{outside finite subsets},\\ \left\|\sum_m a_mT^m\right\|_r=\max_m|a_m|r^m. \end{gathered} By Proposition 1.2.2 and density of finite Laurent sums, this norm is multiplicative. Write ξK,r\xi_{K,r} for its spectrum point and HK,rH_{K,r} for the completion of Frac⁡(AK,r)\operatorname{Frac}(A_{K,r}) for this norm. Thus HK,r=H(ξK,r)H_{K,r}=\mathcal H(\xi_{K,r}). Empty tuples give AK,r=HK,r=KA_{K,r}=H_{K,r}=K.

Proposition 5.36.1. NAG-002298 (Monomial covers of Gauss fields). Let C=(cji)∈Md(Z)C=(c_{ji})\in M_d(\mathbb Z) have nonzero determinant, put sj=∏iricjis_j=\prod_i r_i^{c_{ji}}, and set e=∣det⁡C∣e=|\det C|. The substitution Sj↦∏iTicjiS_j\mapsto\prod_iT_i^{c_{ji}} gives an isometric inclusion AK,s↪AK,rA_{K,s}\hookrightarrow A_{K,r}. If RR is a set of representatives for Zd/CtZd\mathbb Z^d/C^{\mathsf t}\mathbb Z^d, then AK,r=⨁u∈RAK,sTu,∥∑u∈RauTu∥r=max⁡u∈Rru∥au∥s.\begin{equation} A_{K,r}=\bigoplus_{u\in R}A_{K,s}T^u, \qquad \left\|\sum_{u\in R}a_uT^u\right\|_r =\max_{u\in R}r^u\|a_u\|_s. \end{equation} In particular the map is finite free of rank ee, with the canonical finite-module Banach structure. Its actual fibre over ξK,s\xi_{K,s} is the field AK,r⊗^AK,sHK,s=HK,r,[HK,r:HK,s]=e.A_{K,r}\widehat\otimes_{A_{K,s}}H_{K,s}=H_{K,r}, \qquad [H_{K,r}:H_{K,s}]=e. The Gauss point is the only point of this fibre.

Proof. The substitution preserves each Laurent coefficient and its weighted size, since sv=rCtvs^v=r^{C^{\mathsf t}v}. The exponent map is injective, so it preserves the coefficient norm and the finite-tail condition.

We first check the integer index. Integer row and column operations diagonalize a nonsingular integer matrix for this purpose. Move a nonzero entry to the upper left. If an entry in its row or column is not divisible by it, division with remainder followed by a swap strictly decreases the positive absolute value of the pivot. After finitely many such reductions the pivot divides its entire row and column. Clear both by integer elementary operations and repeat on the remaining nonsingular square block. These operations are unimodular. A diagonal matrix has quotient index the product of the absolute values of its diagonal entries, which is its absolute determinant. Applying the same operations in transposed order also proves [Zd:CZd]=[Zd:CtZd]=∣det⁡C∣.[\mathbb Z^d:C\mathbb Z^d] =[\mathbb Z^d:C^{\mathsf t}\mathbb Z^d]=|\det C|.

Every exponent has a unique expression u+Ctvu+C^{\mathsf t}v with u∈Ru\in R and v∈Zdv\in\mathbb Z^d. Grouping a convergent Laurent series by these finitely many cosets gives (5.36.1). Each grouped series converges: its weighted coefficients are those of the original series, divided by the fixed positive number rur^u. Conversely a finite sum of such grouped series has a vanishing weighted tail. Disjoint exponents prove uniqueness and the norm equality. Since RR is finite, its weighted maximum norm is equivalent to the usual maximum norm of a free AK,sA_{K,s}-module.

Put Fs=Frac⁡(AK,s)F_s=\operatorname{Frac}(A_{K,s}). The localization AK,r⊗AK,sFsA_{K,r}\otimes_{A_{K,s}}F_s is a finite-dimensional domain over FsF_s, hence a field: multiplication by a nonzero element is an injective linear endomorphism and is therefore surjective. It is thus Frac⁡(AK,r)\operatorname{Frac}(A_{K,r}), with the same coset basis. Clearing a common denominator in AK,sA_{K,s} proves the weighted norm identity for coefficients in FsF_s as well.

Its completion is consequently ⨁u∈RHK,sTu\bigoplus_{u\in R}H_{K,s}T^u with that weighted norm. Multiplication extends continuously, and its norm remains multiplicative by approximating two elements with elements of the dense valued field. It is a finite-dimensional domain over HK,sH_{K,s}, hence a field. By the definition of HK,rH_{K,r} this completed field is HK,rH_{K,r}, of degree ∣R∣=e|R|=e. Proposition 2.33.19 identifies the algebraic tensor with the stated completed tensor; Theorem 2.33.11 identifies its spectrum with the actual analytic fibre.

Finally, a bounded multiplicative seminorm on the field HK,rH_{K,r} is at most its given multiplicative norm, by applying its bound to successive powers. Applying the same inequality to the inverse of a nonzero element gives the reverse inequality. Its spectrum over HK,sH_{K,s} is therefore a single point, with completed residue field HK,rH_{K,r}. ◻

For a finite AK,rA_{K,r}-algebra DD whose given norm is equivalent to its canonical finite-module norm, define its Gauss fibre and degree by EK,r(D)=D⊗^AK,rHK,r,δK,r(D)=dim⁡HK,rEK,r(D).\begin{equation} E_{K,r}(D)=D\widehat\otimes_{A_{K,r}}H_{K,r}, \qquad \delta_{K,r}(D)=\dim_{H_{K,r}}E_{K,r}(D). \end{equation} These are an actual finite analytic fibre algebra and a nonnegative integer by Proposition 2.33.19 and Theorem 2.33.11.

Proposition 5.36.2. NAG-002301 (Scalar extension of a finite Gauss fibre). Let L/KL/K be any complete isometric extension, and put DL=D⊗^KLD_L=D\widehat\otimes_K L. There are natural bounded algebra isomorphisms (AK,r)L≃AL,r,EL,r(DL)≃EK,r(D)⊗HK,rHL,r,(A_{K,r})_L\simeq A_{L,r},\qquad E_{L,r}(D_L)\simeq E_{K,r}(D)\otimes_{H_{K,r}}H_{L,r}, and the coefficient inclusion induces an isometric embedding HK,r↪HL,rH_{K,r}\hookrightarrow H_{L,r}. In particular δL,r(DL)=δK,r(D).\delta_{L,r}(D_L)=\delta_{K,r}(D). For either of these finite fibre algebras EE over its indicated field HH, its degree is dim⁡HE=∑p∈Max⁡Elength⁡Ep(Ep)[κ(p):H].\dim_HE= \sum_{\mathfrak p\in\operatorname{Max}E} \operatorname{length}_{E_{\mathfrak p}}(E_{\mathfrak p}) [\kappa(\mathfrak p):H].

Proof. The actual relative pushout 2.33.8 extends the Laurent relation presentation coefficientwise. Its norm is the Laurent coefficient norm over LL, by Theorem 1.17.2. The coefficient inclusion preserves the Gauss norm, so it extends first to the fraction fields and then to an isometric embedding HK,r↪HL,rH_{K,r}\hookrightarrow H_{L,r}.

The two bounded pushouts of DD and LL identify DL≃D⊗^AK,rAL,r.D_L\simeq D\widehat\otimes_{A_{K,r}}A_{L,r}. By Proposition 2.33.19, this is finite over AL,rA_{L,r} with its canonical module norm. Applying Corollary 2.33.9 to the Gauss point, and then the same finite-base-change proposition, gives DL⊗^AL,rHL,r≃D⊗^AK,rHL,r≃(D⊗^AK,rHK,r)⊗HK,rHL,r.\begin{aligned} D_L\widehat\otimes_{A_{L,r}}H_{L,r} &\simeq D\widehat\otimes_{A_{K,r}}H_{L,r}\\ &\simeq \bigl(D\widehat\otimes_{A_{K,r}}H_{K,r}\bigr) \otimes_{H_{K,r}}H_{L,r}. \end{aligned} The tensors on the last line are already complete because the first factor is finite-dimensional over HK,rH_{K,r}. A basis over HK,rH_{K,r} becomes a basis after this field extension, proving the degree equality, including degree zero.

A finite-dimensional algebra over a field is Artinian. Its local product decomposition is [StacksAlg] (Tag 00JB). In a composition series of EpE_{\mathfrak p} every simple quotient is κ(p)\kappa(\mathfrak p): the annihilator of a generator of a simple module is maximal, and this local ring has only one maximal ideal. Taking dimensions gives the displayed sum. Apply this argument to each of the two actual fibre algebras. ◻

Theorem 5.36.3. NAG-002302 (Integral models computing an arbitrary-radius degree). Let DD be as above. Let L/KL/K be a complete nontrivially valued isometric extension, choose bj∈L×b_j\in L^\times with ∣bj∣=rj|b_j|=r_j, and write Sj=Tj/bjS_j=T_j/b_j. Put OL=L∘,kL=L~,AL,0=OL⟨S1±1,…,Sd±1⟩.O_L=L^\circ,\quad k_L=\widetilde L,\quad A_{L,0}=O_L\langle S_1^{\pm1},\ldots,S_d^{\pm1}\rangle. For any finite list of power-bounded elements of DLD_L there is an admissible topologically finitely presented algebra DL,0⊂DLD_{L,0}\subset D_L, finite over AL,0A_{L,0}, containing that list and having generic algebra DLD_L. For every such finite model put CL,0=(DL,0/L∘∘DL,0)⊗kL[S±1]kL(S1,…,Sd).C_{L,0}=(D_{L,0}/L^{\circ\circ}D_{L,0}) \otimes_{k_L[S^{\pm1}]}k_L(S_1,\ldots,S_d). Then δK,r(D)=∑p∈Max⁡CL,0length⁡(CL,0)p((CL,0)p)[κ(p):kL(S)].\begin{equation} \delta_{K,r}(D)= \sum_{\mathfrak p\in\operatorname{Max}C_{L,0}} \operatorname{length}_{(C_{L,0})_{\mathfrak p}} ((C_{L,0})_{\mathfrak p}) [\kappa(\mathfrak p):k_L(S)]. \end{equation} In particular this integer is independent of LL, of the scalars bjb_j, and of the finite model. Such an extension LL always exists.

If DD is the actual full fibre at a weight ww of analytic units f1,…,fnf_1,\ldots,f_n, and is finite with its canonical module norm over the polytorus of their monomials fqjf^{q_j}, one may choose LL realizing all exp⁡(−wi)\exp(-w_i). For ai∈L×a_i\in L^\times of these absolute values, the model may be chosen to contain all fi/aif_i/a_i and ai/fia_i/f_i. The target normalizers can be taken as bj=∏iaiqjib_j=\prod_i a_i^{q_{ji}}.

Proof. The coordinate change Sj=Tj/bjS_j=T_j/b_j is an isometric isomorphism between AL,rA_{L,r} and the unit-polytorus algebra over LL; it sends their Gauss points to each other. Proposition 5.36.2 gives DLD_L its canonical finite-module Banach structure. Apply Theorem 5.32.1 over LL. It constructs the required actual integral quotient model, with the prescribed elements, and computes its special-fibre length sum as δL,r(DL)\delta_{L,r}(D_L). The same theorem gives this equality for every admissible finite model with that generic algebra. Proposition  5.36.2 identifies this number with δK,r(D)\delta_{K,r}(D), proving (5.36.3) and all the independence assertions.

To obtain an extension realizing any finite list of radii, append one radius 0<ρ<10<\rho<1 and take the completion of the fraction field of its Laurent algebra for the multiplicative Gauss norm. Its coordinate elements realize the prescribed radii and ρ\rho, so this is a complete nontrivially valued extension of KK with the original absolute value on KK.

In the last assertion the full fibre is the actual simultaneous Laurent domain ∣fi∣=exp⁡(−wi)|f_i|=\exp(-w_i). The relative pushout 2.33.8 shows that its base change is the same actual Laurent domain over LL. On every point of this domain both fi/aif_i/a_i and ai/fia_i/f_i have absolute value one. Theorem 1.20.2 makes these elements power bounded, also with the canonical equivalent norm. They may therefore be included in the finite prescribed list. The displayed choice of bjb_j makes the normalized target coordinates exactly the indicated monomials in these elements. ◻

Proposition 5.36.4. NAG-002304 (Calibration under an isogeny of the target). Let AK,r→DA_{K,r}\to D be finite with its canonical Banach structure, and let CC and ss be as in Proposition 5.36.1. Regard DD also as an AK,sA_{K,s}-algebra by composition. It has the canonical finite-AK,sA_{K,s}-module Banach structure, and δK,s(D)=∣det⁡C∣ δK,r(D).\delta_{K,s}(D)=|\det C|\,\delta_{K,r}(D).

More generally let DD be an actual full moment-fibre algebra at w∈Rnw\in\mathbb R^n, and let Q0,Q1:Zn→ZdQ_0,Q_1:\mathbb Z^n\to\mathbb Z^d have rank dd and the same rational row space. Suppose the two monomial maps make DD finite with its canonical module norms over their polytori at QiwQ_iw. Let N⊂ZnN\subset\mathbb Z^n have rank dd and suppose Qi∣NQ_i|_N is injective. Writing δQi\delta_{Q_i} for the corresponding Gauss degree, one has δQ0[Zd:Q0(N)]=δQ1[Zd:Q1(N)].\frac{\delta_{Q_0}}{[\mathbb Z^d:Q_0(N)]} =\frac{\delta_{Q_1}}{[\mathbb Z^d:Q_1(N)]}.

Proof. Choose a finite free AK,rA_{K,r}-module surjecting onto DD and give DD its quotient norm. The finite weighted basis in (5.36.1) identifies that free module with a finite free AK,sA_{K,s}-module with an equivalent norm. The same quotient therefore gives the canonical finite-AK,sA_{K,s}-module topology on DD.

Proposition 5.36.1 and associativity of the finite tensors identify D⊗AK,sHK,s≃D⊗AK,r(AK,r⊗AK,sHK,s)≃D⊗AK,rHK,r.D\otimes_{A_{K,s}}H_{K,s} \simeq D\otimes_{A_{K,r}} \bigl(A_{K,r}\otimes_{A_{K,s}}H_{K,s}\bigr) \simeq D\otimes_{A_{K,r}}H_{K,r}. Their completed versions are the same by Proposition 2.33.19. The algebra on the right has dimension δK,r(D)\delta_{K,r}(D) over HK,rH_{K,r}, and [HK,r:HK,s]=∣det⁡C∣[H_{K,r}:H_{K,s}]=|\det C|, giving the first formula.

For the second assertion, equality of the rational row spaces gives an invertible rational matrix BB with Q1=BQ0Q_1=BQ_0. Choose a positive integer mm clearing its denominators and set C0=mBC_0=mB, C1=mIdC_1=mI_d. Then both CiC_i are nonsingular integer matrices and C0Q0=C1Q1C_0Q_0=C_1Q_1. These equal matrices define the same monomial map on the same actual algebra DD, with the same target radii. Applying the first assertion twice gives ∣det⁡C0∣δQ0=∣det⁡C1∣δQ1.|\det C_0|\delta_{Q_0} =|\det C_1|\delta_{Q_1}. For each ii, the finite chain CiQi(N)⊂CiZd⊂ZdC_iQ_i(N)\subset C_i\mathbb Z^d\subset\mathbb Z^d has index [Zd:CiQi(N)]=∣det⁡Ci∣[Zd:Qi(N)].[\mathbb Z^d:C_iQ_i(N)] =|\det C_i|[\mathbb Z^d:Q_i(N)]. Indeed representatives for the two successive quotients give representatives for the composite quotient, and CiC_i identifies the first quotient with Zd/Qi(N)\mathbb Z^d/Q_i(N). The left sides for i=0,1i=0,1 agree because C0Q0=C1Q1C_0Q_0=C_1Q_1. Dividing the two displayed identities proves the calibrated equality. ◻

5.37 Weights on interior flat tropical germs

Theorem 5.37.1. NAG-002306 (Positive intrinsic weights at arbitrary real flat germs). Let X=M(B)X=\mathcal M(B) be nonempty and pure of analytic dimension dd over the original complete field KK. Let f1,…,fn∈B×f_1,\ldots,f_n\in B^\times and put t(x)=(−log⁡∣fi(x)∣)it(x)=(-\log|f_i(x)|)_i. Fix w∈t(X)w\in t(X) and a saturated rank-dd sublattice N0⊂ZnN_0\subset\mathbb Z^n. Suppose that, for some open neighborhood WW of ww, t(X)∩W⊂w+(N0)R,Fw=t−1(w)⊂Int⁡(X/K).t(X)\cap W\subset w+(N_0)_{\mathbb R},\qquad F_w=t^{-1}(w)\subset\operatorname{Int}(X/K). Let DwD_w be the actual affinoid algebra of the whole fibre FwF_w. For every integral map Q:Zn→ZdQ:\mathbb Z^n\to\mathbb Z^d whose restriction to N0N_0 is injective, the monomials fqjf^{q_j} make DwD_w finite, with its canonical module norm, over the exact-radius polytorus algebra at QwQw. Its Gauss degree δQ(w)\delta_Q(w) is positive, and mX,f(w)=δQ(w)[Zd:Q(N0)]∈Z>0m_{X,f}(w)= \frac{\delta_Q(w)}{[\mathbb Z^d:Q(N_0)]} \in\mathbb Z_{>0} is independent of QQ. This assertion allows arbitrary real ww and the trivial valuation. It uses the actual fibre algebra, including its nilpotents.

More precisely, after any complete nontrivial isometric extension L/KL/K realizing the numbers e−wie^{-w_i}, choose ai∈L×a_i\in L^\times of these absolute values. Form the residual moment cycle of Dw⊗^KLD_w\widehat\otimes_KL with characters fi/aif_i/a_i. If ρ:TZn,L~→TZn/N0,L~\rho:\mathbb T_{\mathbb Z^n,\widetilde L} \to\mathbb T_{\mathbb Z^n/N_0,\widetilde L} is the quotient, that cycle has the form ∑scs[ρ−1(s)]\sum_s c_s[\rho^{-1}(s)], and mX,f(w)=∑scs[κ(s):L~].m_{X,f}(w)= \sum_s c_s[\kappa(s):\widetilde L]. This integer is independent of the extension, normalizing elements, and compatible admissible model.

Proof. Transverse integral maps QQ exist by the saturated-lattice splitting in Lemma 5.14.1. Choose a closed rectangular box PP containing ww in its interior and contained in WW. Its inverse image Y=t−1(P)Y=t^{-1}(P) is the actual simultaneous Laurent domain, by Theorem 2.31.2. At every point of FwF_w, the finitely many coordinate inequalities defining YY are strict. Its relative presentation therefore puts Fw⊂Int⁡(Y/X)F_w\subset\operatorname{Int}(Y/X). Proposition 2.33.21 and the hypothesis give Fw⊂Int⁡(Y/K)F_w\subset\operatorname{Int}(Y/K).

The linear map QQ is injective on (N0)R(N_0)_{\mathbb R}. Since t(Y)⊂w+(N0)Rt(Y)\subset w+(N_0)_{\mathbb R}, the entire level of Q∘t∣YQ\circ t|_Y at QwQw is FwF_w. Theorem 5.31.1, applied to YY and its units fqjf^{q_j}, makes its actual level algebra finite with the canonical module norm over the target polytorus. The level is the same affinoid domain as Fw⊂XF_w\subset X: the rational universal property gives bounded inverse identifications of their algebras. Thus the finite algebra is precisely DwD_w.

Theorem 5.35.3 makes the nonempty domain FwF_w pure of analytic dimension dd. Apply Proposition 5.35.4 to its finite map to the exact-radius polytorus, regarded as a polyannulus with degenerate intervals. Its Gauss degree is positive. This construction works for every indicated QQ over KK itself.

Choose a complete extension LL as in Theorem 5.36.3, realizing all e−wie^{-w_i} and one radius strictly between zero and one. The actual full fibre of XLX_L at ww is (Fw)L(F_w)_L, by Theorem 2.36.9. Every point of XLX_L restricts to a point of XX with the same character values, so t(XL)∩W⊂w+(N0)Rt(X_L)\cap W\subset w+(N_0)_{\mathbb R}. Every point of (Fw)L(F_w)_L is interior over LL by the base-change assertion of Proposition 2.33.21. Proposition 5.36.2 preserves each finite map, its canonical module norm, and its Gauss degree.

Normalize by the aia_i and apply Theorem 5.34.3 over LL. Its hypotheses have just been verified for the same actual fibre algebra and for every QQ. It identifies the residual supports with the indicated cosets and gives δQ,L(w)=[Zd:Q(N0)]∑scs[κ(s):L~].\delta_{Q,L}(w) =[\mathbb Z^d:Q(N_0)] \sum_s c_s[\kappa(s):\widetilde L]. Scalar extension preserves δQ(w)\delta_Q(w), and its positivity has already been proved over KK. This proves integrality, positivity, and independence of the projection. The left side divided by its fixed lattice index is defined over the original field and absolute value; it proves independence of LL and the aia_i. Independence of the compatible model also follows from Theorem 5.34.1. ◻

Theorem 5.37.2. NAG-002307 (Local constancy on an interior flat germ). Under the hypotheses of Theorem 5.37.1, there is an open neighborhood VV of ww such that t(X)∩V=(w+(N0)R)∩V.t(X)\cap V=(w+(N_0)_{\mathbb R})\cap V. Every vv in this set has its whole fibre in Int⁡(X/K)\operatorname{Int}(X/K), satisfies the same flat-germ hypotheses with lattice N0N_0, and has mX,f(v)=mX,f(w).m_{X,f}(v)=m_{X,f}(w). The value at a weight is unchanged by restriction to an affinoid domain containing its whole moment fibre, whenever the two weighted germs satisfy these hypotheses.

Proof. The boundary ∂KX\partial_KX is compact, by openness of interior and compactness of XX in Propositions 2.33.21 and 2.24.1. Its tropical image is compact and does not contain ww. Choose the box PP in the preceding proof also disjoint from that image. Then Y=t−1(P)Y=t^{-1}(P) lies in Int⁡(X/K)\operatorname{Int}(X/K) and is pure of analytic dimension dd, by Theorem 5.35.3.

Fix a transverse integral QQ; one exists by choosing a basis of N0N_0 extending to a basis of Zn\mathbb Z^n, using Lemma 5.14.1, and projecting onto the first dd coordinates. The entire level of Q∘t∣YQ\circ t|_Y at QwQw is Fw⊂Int⁡(Y/K)F_w\subset\operatorname{Int}(Y/K), as above. Theorem 5.35.5 therefore supplies a closed rectangular neighborhood P′P' of QwQw for which Y′=(Q∘t∣Y)−1(P′)⟶TP′Y'=(Q\circ t|_Y)^{-1}(P')\longrightarrow\mathbb T_{P'} is finite and surjective, with a fixed positive Gauss degree δ\delta at every s∈P′s\in P'.

The affine restriction Q:w+(N0)R→RdQ:w+(N_0)_{\mathbb R}\to\mathbb R^d is an isomorphism. Its inverse sends P′P' exactly to t(Y′)t(Y'): inclusion in one direction follows from t(Y)⊂w+(N0)Rt(Y)\subset w+(N_0)_{\mathbb R}, and the other follows from surjectivity of Y′→TP′Y'\to\mathbb T_{P'}. Take an open neighborhood V⊂Int⁡(P)∩Q−1(Int⁡(P′))V\subset\operatorname{Int}(P) \cap Q^{-1}(\operatorname{Int}(P')) of ww. The preceding equality proves t(X)∩V=(w+(N0)R)∩Vt(X)\cap V=(w+(N_0)_{\mathbb R})\cap V. For vv in this set, its whole fibre in XX belongs to Y′Y', since v∈Int⁡(P)v\in\operatorname{Int}(P) and Qv∈P′Qv\in P'; in particular it is interior over KK. The local flat-germ condition at vv follows by restricting to a sufficiently small neighborhood inside WW, since v+(N0)R=w+(N0)Rv+(N_0)_{\mathbb R}=w+(N_0)_{\mathbb R}.

Restriction of the finite polyannulus map to its exact level at s=Qvs=Qv has algebra DvD_v, by Theorem 5.35.5 and injectivity of QQ on the affine plane. Laurent polynomials are dense both in this target’s exact-radius algebra and in the polyannulus algebra evaluated at ηs\eta_s. Their completed Gauss residue fields are therefore the same completion of K(T1,…,Td)K(T_1,\ldots,T_d) with its weighted Gauss norm. Finite base change consequently identifies the Gauss fibre of DvD_v with that of Y′Y' over ηs\eta_s. Thus δQ(v)=δ\delta_Q(v)=\delta for every such vv, including ww. Theorem 5.37.1 gives mX,f(v)=δ[Zd:Q(N0)]=mX,f(w).m_{X,f}(v)=\frac{\delta}{[\mathbb Z^d:Q(N_0)]} =m_{X,f}(w).

Finally an affinoid domain containing the whole fibre has the same actual fibre domain, with the same analytic structure. This follows from domain restriction and the two universal factorizations in Theorem 2.36.9. The local image equality just proved, applied to both domains, and inclusion of their moment images identify their tangent planes: both have dimension dd. Their saturated tangent lattices are consequently equal. The fibre algebra and its monomial maps to each target polytorus are unchanged, so their Gauss degrees, and hence their calibrated weights, agree. ◻

Theorem 5.37.3. NAG-002308 (Geometric pushforward under a finite locally free map). Let p:X→Yp:X\to Y be a finite morphism of generalized KK-affinoid spaces, both pure of analytic dimension dd. Suppose its algebra is finite locally free of constant rank Δ>0\Delta>0 over the target algebra, with the canonical module Banach structure. Let unit moment tuples fXf_X and fYf_Y define tX:X→Rnt_X:X\to\mathbb R^n and tY:Y→Rn′t_Y:Y\to\mathbb R^{n'}, and suppose that tY∘p=a∘tXt_Y\circ p=a\circ t_X comes from the corresponding identity of characters for an integral map a:Zn→Zn′a:\mathbb Z^n\to\mathbb Z^{n'}. Let w′∈tY(Y)w'\in t_Y(Y) have its whole fibre in Int⁡(Y/K)\operatorname{Int}(Y/K), and suppose that the local tropical image at w′w' is contained in w′+(NY)Rw'+(N_Y)_{\mathbb R} for a saturated rank-dd lattice NY⊂Zn′N_Y\subset\mathbb Z^{n'}. Assume a−1(w′)∩tX(X)={w1,…,ws}a^{-1}(w')\cap t_X(X)=\{w_1,\ldots,w_s\} is finite and nonempty, and that the local tropical image at each wiw_i is contained in wi+(Ni)Rw_i+(N_i)_{\mathbb R} for a saturated rank-dd lattice Ni⊂ZnN_i\subset\mathbb Z^n.

Then a(Ni)⊂NYa(N_i)\subset N_Y has finite index for every ii, all the weights in Theorem 5.37.1 are defined, and Δ mY,fY(w′)=∑i=1smX,fX(wi)[NY:a(Ni)].\Delta\,m_{Y,f_Y}(w') =\sum_{i=1}^s m_{X,f_X}(w_i) [N_Y:a(N_i)]. The locally free rank, the fibre algebras, and their lengths are taken before reduction. No restriction on the real weights or on the characteristic or value group of KK is imposed.

Proof. Theorem 2.33.18 makes pp everywhere interior relative to YY. The composition inclusion in Proposition 2.33.21 therefore gives p−1(tY−1(w′))⊂Int⁡(X/K).p^{-1}(t_Y^{-1}(w'))\subset\operatorname{Int}(X/K). In particular the whole fibre at every wiw_i is interior. Theorem 5.37.1 defines its positive intrinsic weight and that at w′w'. Theorem 5.37.2 shows that the local images fill neighborhoods in their stated planes.

For vv sufficiently near wiw_i in wi+(Ni)Rw_i+(N_i)_{\mathbb R}, one has v∈tX(X)v\in t_X(X) and a(v)∈tY(Y)a(v)\in t_Y(Y) sufficiently near w′w'. It follows that a((Ni)R)⊂(NY)Ra((N_i)_{\mathbb R})\subset(N_Y)_{\mathbb R}. Saturation of NYN_Y implies a(Ni)⊂NYa(N_i)\subset N_Y. If the restriction of aa to (Ni)R(N_i)_{\mathbb R} had nonzero kernel, a small interval through wiw_i in that kernel would give infinitely many elements of a−1(w′)∩tX(X)a^{-1}(w')\cap t_X(X). This contradicts the finite-set hypothesis. Hence this restriction is injective, and a(Ni)a(N_i) has finite index in NYN_Y.

Choose an integral Q:Zn′→ZdQ:\mathbb Z^{n'}\to\mathbb Z^d transverse to NYN_Y. Take a small closed box in the target moment space, with w′w' in its interior and with its inverse image Y0⊂YY_0\subset Y contained in the specified flat-germ neighborhood. As in the proof of Theorem 5.37.1, the entire level of Q∘tY∣Y0Q\circ t_Y|_{Y_0} at Qw′Qw' is tY−1(w′)t_Y^{-1}(w') and is interior over KK in Y0Y_0. Put X0=p−1(Y0)X_0=p^{-1}(Y_0). Theorem 2.36.9 gives its actual affinoid algebra, and Proposition 2.33.19 preserves the finite canonical module structure. Ordinary base change of a locally free module preserves its rank: its local bases base change to local bases.

The entire level of Qa∘tX∣X0Q a\circ t_X|_{X_0} at Qw′Qw' is p−1(tY−1(w′))p^{-1}(t_Y^{-1}(w')). Indeed, the target moment image in Y0Y_0 lies in the plane on which QQ is injective. This level is interior in X0X_0 by finite relative interior and composition. Its moment values are exactly w1,…,wsw_1,\ldots,w_s. Proposition 5.31.2 therefore identifies its actual algebra with ∏i=1sDwi.\prod_{i=1}^s D_{w_i}. Each factor is the original whole moment-fibre algebra on XX, since that whole fibre already belongs to X0X_0.

Let Dw′′D'_{w'} be the actual whole target-fibre algebra, AA the exact-radius target polytorus at Qw′Qw', and HH its Gauss field. The algebra of the displayed source level is also the actual finite base change of the algebra of XX along the algebra map from YY to Dw′′D'_{w'}. It is consequently finite locally free of rank Δ\Delta over Dw′′D'_{w'}. After tensoring over AA with HH, it is finite locally free of that rank over the Artinian algebra E′=Dw′′⊗^AH.E'=D'_{w'}\widehat\otimes_A H. All these finite tensors are ordinary tensors with their complete canonical norms by Proposition 2.33.19.

By [StacksAlg] (Tag 00JB), E′E' is a finite product of Artinian local algebras. On each factor, a local basis of the finite locally free module at its unique maximal ideal is a global basis: a principal open neighborhood of that ideal is the entire spectrum, and its defining element is a unit. Its HH-dimension is therefore Δ\Delta times the HH-dimension of that factor. Summing the dimensions and using the actual product of source fibres gives Δ δQ(w′)=∑i=1sδQa(wi).\Delta\,\delta_Q(w') =\sum_{i=1}^s\delta_{Qa}(w_i). Apply Theorem 5.37.1 to every term. Since QQ is injective on NYN_Y, counting representatives in the subgroup chain Qa(Ni)⊂Q(NY)⊂ZdQa(N_i)\subset Q(N_Y)\subset\mathbb Z^d gives [Zd:Qa(Ni)]=[Zd:Q(NY)] [NY:a(Ni)].[\mathbb Z^d:Qa(N_i)] =[\mathbb Z^d:Q(N_Y)]\,[N_Y:a(N_i)]. Cancel the common positive first factor. This proves the stated geometric pushforward formula with all lengths and full field degrees already incorporated in the intrinsic weights. ◻

5.38 Local tropical geometry at interior fibres

Local polyhedrality of analytic moment images belongs to the theory developed by Berkovich and Ducros. For the compact-space and point-germ theorems, see A. Ducros, Espaces de Berkovich, polytopes, squelettes et théorie des modèles, Confluentes Math. 4 (2012), 1250007, Theorems 3.2 and 3.4, DOI: 10.1142/S1793744212500077. The proof here reconstructs the local star at a whole interior fibre from the graded germ reductions and finite affinoid projections established earlier in the book.

For a tuple of invertible analytic functions h=(h1,…,hn)h=(h_1,\ldots,h_n), write th(y)=(−log⁡∣h1(y)∣,…,−log⁡∣hn(y)∣),hm=∏i=1nhimi(m∈Zn).t_h(y)=(-\log|h_1(y)|,\ldots,-\log|h_n(y)|), \qquad h^m=\prod_{i=1}^n h_i^{m_i}\quad(m\in\mathbf Z^n). The tropicalization of an algebraic torus subscheme over a trivially valued field is its actual analytic moment image, with the convention of Theorem 5.13.6.

Lemma 5.38.1. NAG-002348 (Lifting strict sectors from a residue field). Let KK be a complete non-Archimedean field with its specified absolute value, let (X,x)(X,x) be a good KK-analytic germ, and suppose that x∈Int⁡(X/K)x\in\operatorname{Int}(X/K). Let h1,…,hnh_1,\ldots,h_n be invertible analytic functions near xx such that ∣hi(x)∣=1|h_i(x)|=1. Put k=K~k=\widetilde K, and let R=k[h~1±1,…,h~n±1]⊂H(x)~,Z=Spec⁡R⊂Gm,kn.R=k[\widetilde h_1^{\pm1},\ldots,\widetilde h_n^{\pm1}] \subset\widetilde{\mathcal H(x)}, \qquad Z=\operatorname{Spec}R\subset\mathbf G_{m,k}^n. If u∈Trop⁡(Z)u\in\operatorname{Trop}(Z) and m1,…,ms∈Znm_1,\ldots,m_s\in\mathbf Z^n satisfy ⟨mj,u⟩>0\langle m_j,u\rangle>0 for every jj, then every neighbourhood of xx contains a point yy for which ∣hmj(y)∣<1(1≤j≤s).|h^{m_j}(y)|<1\qquad(1\leq j\leq s). In particular the corresponding strict rational sector contains moment values arbitrarily close to 0=th(x)0=t_h(x).

Proof. Choose an analytic point of ZZ whose moment value is uu. Its kernel is a prime p⊂R\mathfrak p\subset R, and its absolute value induces a valuation ring VpV_{\mathfrak p} in κ(p)\kappa(\mathfrak p) containing the trivially valued field kk. Let B⊂RpB\subset R_{\mathfrak p} be the inverse image of VpV_{\mathfrak p} under the residue map. This is a local subring of Frac⁡(R)⊂H(x)~\operatorname{Frac}(R)\subset\widetilde{\mathcal H(x)}. Indeed, its maximal ideal is the inverse image of the maximal ideal of VpV_{\mathfrak p}, and an element outside that inverse image is a unit of RpR_{\mathfrak p} whose inverse again lies in BB. The character h~mj\widetilde h^{m_j} is a unit of RpR_{\mathfrak p}, and its image in κ(p)\kappa(\mathfrak p) has absolute value e−⟨mj,u⟩<1e^{-\langle m_j,u\rangle}<1. It therefore belongs to the maximal ideal of BB.

By [StacksAlg] (Tag 00IA), there is a valuation ring V⊂H(x)~V\subset\widetilde{\mathcal H(x)} dominating BB. It contains kk, and each h~mj\widetilde h^{m_j} belongs to its maximal ideal. Consequently h~−mj∉V\widetilde h^{-m_j}\notin V. Lemma 3.19.20 extends VV to a graded valuation ring W⊂Fx:=H(x)~grcontaining K~gr,W1=V.W\subset F_x:=\widetilde{\mathcal H(x)}^{\rm gr} \quad\hbox{containing }\widetilde K^{\rm gr}, \qquad W_1=V. No assertion about the rank of this valuation is needed.

Choose an ordinary affinoid neighbourhood of xx on which the hih_i are units. The ordinary local test preserves interiority over M(K)\mathcal M(K) on this neighbourhood. The absolute evaluation clause makes the image of its graded algebra in FxF_x finite over K~gr\widetilde K^{\rm gr}. Every graded valuation ring containing K~gr\widetilde K^{\rm gr} contains this image: apply the integral-closedness assertion of Lemma 3.19.1 to its homogeneous elements. Thus R(X,x)=PK~gr(Fx),\mathscr R(X,x) =\mathscr P_{\widetilde K^{\rm gr}}(F_x), using the actual germ reduction of Proposition 3.19.4. In particular W∈R(X,x)W\in\mathscr R(X,x).

Suppose that a neighbourhood of xx contains no point satisfying all the strict inequalities in the statement. Choose a smaller ordinary affinoid neighbourhood of xx inside it on which all hih_i are units. The finitely many Weierstrass domains ∣h−mj∣≤1(1≤j≤s)|h^{-m_j}|\leq1\qquad(1\leq j\leq s) cover this representative. They are a finite closed cover, hence admissible by Lemma 2.39.2, and all contain xx. Each inequality is active at xx. Proposition 3.19.6 says that their germ reductions cover R(X,x)\mathscr R(X,x). By Lemma 3.19.5, the jj-th reduction consists of the valuation rings containing h~−mj\widetilde h^{-m_j}. None contains WW, a contradiction. The case s=0s=0 is immediate. Finally continuity of tht_h allows the neighbourhood to be chosen inside the inverse image of any prescribed neighbourhood of 00. ◻

Lemma 5.38.2. NAG-002349 (A finite tropical basis over a trivially valued field). Let kk be a field with the trivial absolute value, and let J⊂k[T1±1,…,Tn±1]J\subset k[T_1^{\pm1},\ldots,T_n^{\pm1}] be an ideal. There are finitely many nonzero p1,…,pa∈Jp_1,\ldots,p_a\in J such that Trop⁡(Spec⁡k[T±1]/J)=⋂j=1aTrop⁡(V(pj)).\operatorname{Trop}(\operatorname{Spec}k[T^{\pm1}]/J) =\bigcap_{j=1}^a\operatorname{Trop}(V(p_j)). For J=0J=0 the empty list is allowed. The set on either side is the support of a finite rational polyhedral fan. At every weight outside this set, some pjp_j has a unique term of least weight.

Proof. Form the homogeneous ideal H⊂k[x0,…,xn]H\subset k[x_0,\ldots,x_n] used in Theorem 5.16.5; localization in all variables followed by x0=1x_0=1 recovers JJ. Use ordinary weight initials, taking terms of least weight. In the present trivial valuation, the hyperplane arrangement in that theorem is central. On each relatively open cone CC its proof gives one fixed ordinary homogeneous ideal HC=in⁡ω(H)H_C=\operatorname{in}_\omega(H) over kk: the minimizing pivot sets and the surviving entries in their row bases are fixed on CC, and these rows generate the ideal in the uniformly bounded degrees of Proposition 5.16.3. Restrict this finite arrangement to ω0=0\omega_0=0.

For any cone CC outside the tropicalization, localization of HCH_C in all variables is the unit ideal, by Theorem 5.13.6 and the localization argument in Theorem 5.16.5. Hence HCH_C contains a monomial xαCx^{\alpha_C}, of some degree eCe_C. In the finite-dimensional space HeCH_{e_C}, choose a fixed row matrix and consider every column set BB with a nonzero maximal minor. There are finitely many such sets. For every one include all members of its pivot row basis in a list GC\mathcal G_C.

Fix ω∈C\omega\in C and choose a minimizing pivot set BB for this degree. Lemma 5.16.4 and Lemma 5.16.1 show that the initial rows form a basis of (HC)eC(H_C)_{e_C} and have the identity matrix in the pivot columns. Since the single monomial vector xαCx^{\alpha_C} belongs to this space, its column must be a pivot: otherwise its pivot coordinates would all be zero, forcing the vector to be zero. Its pivot coordinates then show that the corresponding initial row is exactly xαCx^{\alpha_C}. Thus a member of the finite list GC\mathcal G_C has a monomial initial at every ω∈C\omega\in C.

Dehomogenize the lists for the finitely many excluded cones. These give nonzero Laurent polynomials in JJ. A homogeneous polynomial of fixed degree remains nonzero after x0=1x_0=1, since its remaining exponent vector determines the exponent of x0x_0. Every weight outside the tropicalization has a unique least term in one of the resulting Laurent polynomials. Such a weight is outside that polynomial’s hypersurface tropicalization by the strong triangle inequality. Conversely, an actual analytic point of the subscheme is an analytic point of every one of these hypersurfaces, with the same moment value. This proves the equality. The fan assertion is the central arrangement already constructed. If JJ is the unit ideal, the polynomial 11 gives the assertion directly. ◻

Lemma 5.38.3. NAG-002350 (The actual image of a finite affinoid map). Let A→DA\to D be a bounded homomorphism of generalized KK-affinoid algebras, finite with its canonical module norm, and put I=ker⁡(A→D)I=\ker(A\to D). Its analytic image is exactly M(A/I)\mathcal M(A/I), with the quotient affinoid structure, and dim⁡KM(D)=dim⁡KM(A/I)≤dim⁡KM(A)\dim_K\mathcal M(D)=\dim_K\mathcal M(A/I) \leq\dim_K\mathcal M(A) when D≠0D\ne0. If DD is pure of analytic dimension dd, so is A/IA/I.

Proof. The quotient is an actual affinoid algebra by Corollary 1.18.4. The injection A/I↪DA/I\hookrightarrow D is integral by [StacksAlg] (Tag 00GK). For z∈M(A/I)z\in\mathcal M(A/I) with kernel p\mathfrak p, lying over, [StacksAlg] (Tag 00GQ), makes D⊗A/Iκ(p)D\otimes_{A/I}\kappa(\mathfrak p) nonzero. It is a finite-dimensional vector space. Extending it to H(z)\mathcal H(z) keeps it nonzero. Proposition 1.18.7 identifies this algebra with the actual completed fibre algebra, which has an analytic point by Theorem 1.20.1. The fibre identification of Theorem 2.33.11 therefore gives a point above zz. This proves the image assertion.

At every point of M(D)\mathcal M(D), the completed residue field is finite over that of its image, by Lemma 3.22.6. Taking the maxima in Theorem 3.22.7 on the two spaces gives equality of their analytic dimensions. The dimension inequality follows from the same maximum formula, since the quotient space has the same completed residue fields as its points in M(A)\mathcal M(A).

Suppose DD is pure of dimension dd and let p\mathfrak p be a minimal prime of A/IA/I. A prime of DD above p\mathfrak p contains a minimal prime q\mathfrak q of DD, by [StacksAlg] (Tag 00E0(4)). Its contraction equals p\mathfrak p by minimality. Thus (A/I)/p↪D/q(A/I)/\mathfrak p\hookrightarrow D/\mathfrak q is a finite injection. These quotient algebras have the canonical finite module norms by Theorem 1.18.3. Apply the dimension equality just proved to this injection. The right-hand algebra has analytic dimension dd, and hence so does the left-hand algebra. This holds for every minimal prime. ◻

Proposition 5.38.4. NAG-002351 (A finite analytic image around a whole moment fibre). Let X=M(B)X=\mathcal M(B) be a generalized affinoid space over a complete field KK with its specified absolute value, let f1,…,fn∈B×f_1,\ldots,f_n\in B^\times, and put t=tft=t_f. If w∈t(X)w\in t(X) and Fw=t−1(w)⊂Int⁡(X/K)F_w=t^{-1}(w)\subset\operatorname{Int}(X/K), there is a closed box PP about ww, with ww in its interior, such that Y=t−1(P)Y=t^{-1}(P) lies in Int⁡(X/K)\operatorname{Int}(X/K) and its moment map to the closed polyannulus TP\mathbb T_P is finite. For AP=O(TP)A_P=\mathcal O(\mathbb T_P), DP=O(Y)D_P=\mathcal O(Y), and IP=ker⁡(AP→DP)I_P=\ker(A_P\to D_P), one has t(X)∩P=t(M(AP/IP)).t(X)\cap P =t\bigl(\mathcal M(A_P/I_P)\bigr). Every point of the right-hand analytic space has an actual lift to YY. If XX is pure of analytic dimension dd, the closed analytic image M(AP/IP)\mathcal M(A_P/I_P) is pure of dimension dd. For every v∈Pv\in P, restriction to the exact-radius polytorus at vv gives the actual algebra of the whole fibre FvF_v.

Proof. Apply the first assertion of Theorem 5.35.5 to the complete tuple of nn moment functions. That assertion places no equality between nn and the dimension of XX. It gives the stated domain, finite map, canonical module norm, and the exact-level identifications. Lemma 5.38.3 gives the actual image and the pointwise lifting assertion. In the pure case, Theorem 5.35.3 makes YY pure of dimension dd, and the last assertion of the same lemma gives the purity of its finite image. ◻

Proposition 5.38.5. NAG-002352 (A residual fan bounds and meets the actual germ). Let KK be complete and nontrivially valued, put O=K∘O=K^\circ and k=K~k=\widetilde K, and let X=M(B)X=\mathcal M(B) be a generalized affinoid space. Let h1,…,hn∈B×h_1,\ldots,h_n\in B^\times, and suppose the nonempty whole fibre F0=th−1(0)F_0=t_h^{-1}(0) belongs to Int⁡(X/K)\operatorname{Int}(X/K). Let DD be its actual affinoid algebra, and let D0⊂DD_0\subset D be an actual admissible model, finite over O⟨T1±1,…,Tn±1⟩,Ti⟼hi.O\langle T_1^{\pm1},\ldots,T_n^{\pm1}\rangle, \qquad T_i\longmapsto h_i. Set D‾0=D0/K∘∘D0,J=ker⁡(k[T±1]⟶D‾0),R=k[T±1]/J,Z=Spec⁡R.\begin{aligned} \overline D_0&=D_0/K^{\circ\circ}D_0, & J&=\ker(k[T^{\pm1}]\longrightarrow\overline D_0),\\ R&=k[T^{\pm1}]/J, & Z&=\operatorname{Spec}R. \end{aligned} Then C=Trop⁡(Z)C=\operatorname{Trop}(Z) is the nonempty support of a finite rational polyhedral fan of dimension at most dim⁡KX\dim_K X. There is a neighbourhood VV of 00 such that th(X)∩V⊂C.t_h(X)\cap V\subset C. For every u∈Cu\in C and every finite list mj∈Znm_j\in\mathbf Z^n with ⟨mj,u⟩>0\langle m_j,u\rangle>0, the strict sector {v:⟨mj,v⟩>0 for all j}\{v:\langle m_j,v\rangle>0\text{ for all }j\} contains points of th(X)t_h(X) arbitrarily close to 00.

Proof. The algebra D‾0\overline D_0 is finite over its subalgebra RR. Thus R↪D‾0R\hookrightarrow\overline D_0 is integral by [StacksAlg] (Tag 00GK), and lying over, [StacksAlg] (Tag 00GQ), applies to every prime of RR. Formal specialization 3.16.15 is surjective on Spec⁡D‾0\operatorname{Spec}\overline D_0. If q\mathfrak q lies above p⊂R\mathfrak p\subset R and x∈F0x\in F_0 specializes to q\mathfrak q, evaluation gives injections R/p↪D‾0/q↪H(x)~.R/\mathfrak p\hookrightarrow \overline D_0/\mathfrak q\hookrightarrow \widetilde{\mathcal H(x)}. Indeed a model element has zero residue at xx exactly when its class belongs to q\mathfrak q; every nonzero residue consequently has absolute value one before reduction. These injections extend to the fraction fields.

The finite-type algebra RR is Noetherian, by [StacksAlg] (Tag 00FO), and has finitely many minimal primes, by [StacksAlg] (Tag 00FR). For any such prime p\mathfrak p, the preceding injection and Theorem 3.22.7 give dim⁡(R/p)=trdeg⁡kFrac⁡(R/p)≤trdeg⁡kH(x)~≤dK(x)≤dim⁡KX,\dim(R/\mathfrak p) =\operatorname{trdeg}_k\operatorname{Frac}(R/\mathfrak p) \leq\operatorname{trdeg}_k\widetilde{\mathcal H(x)} \leq d_K(x)\leq\dim_KX, where the equality is [StacksAlg] (Tag 00P0). Apply Theorem 5.19.5 to each reduced irreducible component of ZZ and take a common refinement of their finite fans. The tropicalization of their finite union is their union, since every analytic seminorm has a prime kernel containing one minimal prime by [StacksAlg] (Tag 00E0(4)). This proves the fan and dimension assertions. Nonemptiness also follows from specialization of any point of the nonempty fibre.

Take a finite tropical basis p1,…,pap_1,\ldots,p_a for JJ from Lemma 5.38.2. Lift each nonzero coefficient to an element of OO of absolute value one, and let Pj(T)∈O[T±1]P_j(T)\in O[T^{\pm1}] be the resulting polynomial. Its value Pj(h)P_j(h) belongs to K∘∘D0K^{\circ\circ}D_0. By the definition of this ideal it is a finite sum ∑ℓcℓdℓ\sum_\ell c_\ell d_\ell, with ∣cℓ∣<1|c_\ell|<1 and dℓ∈D0d_\ell\in D_0. Model elements have absolute value at most one at every point of F0F_0, as in Theorem 3.16.15. It follows that max⁡x∈F0∣Pj(h)(x)∣<1.\max_{x\in F_0}|P_j(h)(x)|<1. Choose a real number cj<1c_j<1 strictly larger than this maximum. The closed subset of XX on which ∣Pj(h)∣≥cj|P_j(h)|\geq c_j is compact by Proposition 2.24.1 and has moment image avoiding 00. After shrinking a common neighbourhood VV of 00, one therefore has ∣Pj(h)(x)∣<cj|P_j(h)(x)|<c_j whenever th(x)∈Vt_h(x)\in V. Shrink VV again so that e−⟨m,v⟩>cje^{-\langle m,v\rangle}>c_j for every exponent mm occurring in pjp_j and every v∈Vv\in V. If v∉Cv\notin C, the last assertion of Lemma 5.38.2 supplies an index jj for which the least value of ⟨m,v⟩\langle m,v\rangle among these exponents occurs exactly once. At a point xx with th(x)=vt_h(x)=v, that term is the unique largest term of Pj(h)(x)P_j(h)(x), so ∣Pj(h)(x)∣=e−⟨m,v⟩>cj|P_j(h)(x)|=e^{-\langle m,v\rangle}>c_j. This contradiction proves the asserted inclusion in CC.

Finally let u∈Cu\in C and choose an actual analytic point of ZZ with moment value uu, with kernel p\mathfrak p. Choose q\mathfrak q above p\mathfrak p and x∈F0x\in F_0 specializing to q\mathfrak q, as in the first paragraph. The subalgebra generated by the residues of the hi±1h_i^{\pm1} in H(x)~\widetilde{\mathcal H(x)} is exactly R/pR/\mathfrak p. The chosen analytic point factors through this algebra. If the list of inequalities is empty, the assertion is immediate. Otherwise the chosen point induces a valuation ring in Frac⁡(R/p)\operatorname{Frac}(R/\mathfrak p). The residue injection already constructed identifies this ring with a local subring of H(x)~\widetilde{\mathcal H(x)}. By [StacksAlg] (Tag 00IA), a valuation ring of that field dominates this local subring. Each residue of hmjh^{m_j} lies in its maximal ideal, because its value at the chosen residual point is e−⟨mj,u⟩<1e^{-\langle m_j,u\rangle}<1. Lemma 3.19.20 extends the dominating valuation ring to a graded valuation ring WW containing K~gr\widetilde K^{\rm gr} and with that ring as its degree-one part. Put gj=h−mjg_j=h^{-m_j}, and let γj\gamma_j be its degree-one initial value at xx. Thus no γj\gamma_j belongs to WW.

For an arbitrary prescribed neighbourhood of xx, choose an actual smaller rational neighbourhood NN by Lemma 2.41.1, with all its defining inequalities strict at xx. The strict inequalities give relative interior and Proposition 2.33.21 composes it with the original analytic interior. Thus x∈Int⁡(N/K)x\in\operatorname{Int}(N/K). Take the bounded coordinate maps from the actual rational presentations. The residue field at xx is unchanged by Proposition 2.33.10. For an inner presentation of N/KN/K, the finite-generation part of Theorem Finite graded reduction for a fixed absolute presentation for the actual absolute admissible weighted Tate quotient makes the evaluated graded ring finite over K~gr\widetilde K^{\rm gr}: all new variables vanish at their assigned degrees. Every homogeneous element of that image therefore belongs to WW. Indeed it satisfies a homogeneous monic equation, and dividing by its leading power would otherwise put 11 in the maximal homogeneous ideal of WW.

Suppose the desired strict sector has no point in NN. Then for each y∈Ny\in N one has ∣gj(y)∣≤1|g_j(y)|\le1 for at least one jj. For each ii form the actual domain Vi={y∈N:∣gi/gj(y)∣≤1 for every j}.V_i=\{y\in N: |g_i/g_j(y)|\le1 \text{ for every }j\}. All ratios are units, so these are actual denominator-one radius-one domains containing xx. Total comparability in the graded valuation ring orders the finitely many γi\gamma_i by divisibility. Choose ii with γi/γj∈W\gamma_i/\gamma_j\in W for all jj. The graded reduction image of the actual ViV_i at xx is finite over the image generated by the reduction of NN and these degree-one ratios. For this use present NN by an absolute weighted Tate algebra, append the ratios by the actual coefficientwise quotient of Proposition 1.17.3, and invoke only the admissible Tate-presentation clause of Theorem Finite graded reduction for a fixed absolute presentation. Any surviving class of an original coordinate comes from its class in NN by spectral contraction, and any surviving appended class is the stated ratio. The residue identification along Vi→NV_i\to N is the rational Proposition 2.33.10. Graded integral closedness (Lemma 3.19.1) puts the image of ViV_i in WW. On ViV_i, however, ∣gi(y)∣≤∣gj(y)∣≤1|g_i(y)|\le |g_j(y)|\le1 for some jj at each point. Thus gig_i has spectral seminorm at most one on ViV_i, by the spectral maximum formula 1.20.3. Its class at xx is γi∈W\gamma_i\in W, a contradiction. This proves realization of the strict signs in every neighbourhood of xx. Continuity of tht_h proves the assertion about values approaching 00. The choice of xx used the whole fibre; it did not restrict to the fibre above one target Gauss point. ◻

Strict sectors at a normalized interior point. Let xx be an interior point of an affinoid over the complete ground field, and let analytic units hih_i have absolute value one at xx. Let R⊂H(x)~R\subset\widetilde{\mathcal H(x)} be the Laurent algebra generated over the residue field by their degree-one initial units. Suppose a residual analytic point of RR, over the trivially valued residue field, satisfies the prescribed finitely many signs ∣h~mj∣<1|\widetilde h^{m_j}|<1. Every ordinary neighborhood of xx then contains a point with ∣hmj∣<1|h^{m_j}|<1 for every jj.

To apply the preceding minimum-ratio argument at this fixed point, let q\mathfrak q be the kernel of the residual point. Its valuation ring lies in Frac⁡(R/q)\operatorname{Frac}(R/\mathfrak q). Take the inverse image of that ring under the residue map Rq→Frac⁡(R/q)R_{\mathfrak q}\to\operatorname{Frac}(R/\mathfrak q). The inverse image is local. An element whose residue is a unit of the valuation ring is a unit of RqR_{\mathfrak q} and its inverse also belongs to the inverse image. This gives an actual local subring of H(x)~\widetilde{\mathcal H(x)} with all the prescribed monomial residues in its maximal ideal. A dominating valuation ring exists by [StacksAlg] (Tag 00IA). By Lemma 3.19.20 it extends to a graded valuation ring WW containing the graded ground field. For gj=h−mjg_j=h^{-m_j}, each initial value γj\gamma_j is outside WW.

Take a strict rational neighborhood NN inside the prescribed neighborhood. Strict inequalities and forward interior composition make xx inner in NN over the ground field. In an inner weighted presentation all added coordinate classes evaluate to zero at the assigned degrees. The fixed absolute presentation result therefore puts the evaluated graded ring of NN inside WW by graded integral closedness. If the desired signs fail everywhere on NN, some ∣gj∣≤1|g_j|\leq1 at each point there. Choose ii with all γi/γj\gamma_i/\gamma_j in WW using total homogeneous divisibility. On the actual radius-one quotient domain where all ∣gi/gj∣≤1|g_i/g_j|\leq1, append those ratios to the same absolute weighted presentation. The surviving old coordinate classes come from NN by spectral contraction. The new ones evaluate to the stated ratios. The same fixed absolute result and graded integral closedness put the evaluated graded ring of the quotient inside WW. But on that domain ∣gi∣≤1|g_i|\leq1 everywhere, so the spectral maximum formula puts γi\gamma_i in WW, a contradiction.

Theorem 5.38.6. NAG-002353 (The local tropical star of a whole interior fibre). Let X=M(B)X=\mathcal M(B) be nonempty and pure of analytic dimension dd over a complete non-Archimedean field KK with its specified absolute value. Let f1,…,fn∈B×f_1,\ldots,f_n\in B^\times, put t=tft=t_f, and suppose w∈t(X),Fw=t−1(w)⊂Int⁡(X/K).w\in t(X),\qquad F_w=t^{-1}(w) \subset\operatorname{Int}(X/K). Choose a complete nontrivially valued isometric extension L/KL/K and elements ai∈L×a_i\in L^\times with ∣ai∣=e−wi|a_i|=e^{-w_i}. Let DwD_w be the actual algebra of FwF_w, put D=(Dw)LD=(D_w)_L, and set hi=fi/aih_i=f_i/a_i. There is an actual admissible model D0⊂DD_0\subset D, finite over L∘⟨T1±1,…,Tn±1⟩L^\circ\langle T_1^{\pm1},\ldots,T_n^{\pm1}\rangle by Ti↦hiT_i\mapsto h_i. For any such model set Z=Spec⁡im⁡(L~[T±1]⟶D0/L∘∘D0).Z=\operatorname{Spec}\operatorname{im}\left( \widetilde L[T^{\pm1}] \longrightarrow D_0/L^{\circ\circ}D_0\right). Then C=Trop⁡(Z)C=\operatorname{Trop}(Z) is the support of a finite pure dd-dimensional rational polyhedral fan, and there is an open neighbourhood VV of ww such that t(X)∩V=(w+C)∩V.t(X)\cap V=(w+C)\cap V. Its support CC is independent of the extension, normalizing elements, and compatible model. The theorem allows the trivial absolute value on KK, arbitrary real ww, and nonreduced BB.

Proof. Proposition 5.38.4, restricted to the exact level at ww, makes DwD_w finite with its canonical norm over the nn-dimensional exact-radius polytorus. Theorem 5.36.3 constructs an extension realizing all the radii, and gives the indicated actual finite model over every such LL. The actual domain base-change theorem 2.36.9 identifies M(D)\mathcal M(D) with the whole normalized zero fibre in XLX_L. It is contained in Int⁡(XL/L)\operatorname{Int}(X_L/L) by the base-change assertion of Proposition 2.33.21. Moreover th(XL)=t(X)−w,dim⁡LXL=dim⁡KX=d,t_h(X_L)=t(X)-w,\qquad \dim_LX_L=\dim_KX=d, by Theorems 2.33.13 and 3.21.8. No purity assertion after this scalar extension is used.

Apply Proposition 5.38.5 to XLX_L. It gives a nonempty finite rational fan CC of dimension at most dd, an inclusion of t(X)−wt(X)-w in CC near 00, and actual moment values approaching ww in every strict rational sector meeting CC. The last assertion means sectors containing a specified point of CC with all their defining inequalities strict. The boundary ∂KX\partial_KX is compact, by Propositions 2.33.21 and 2.24.1. Its moment image is compact and avoids ww. Choose ε>0\varepsilon>0 such that in the open Euclidean ball Bε(w)B_\varepsilon(w) one has t(X)∩Bε(w)⊂w+C,t(∂KX)∩Bε(w)=∅.t(X)\cap B_\varepsilon(w)\subset w+C, \qquad t(\partial_KX)\cap B_\varepsilon(w)=\varnothing.

Fix a maximal cone σ\sigma of this fan and put r=dim⁡σr=\dim\sigma. The set Eσ=t(X)∩(w+relint⁡σ)∩Bε(w)E_\sigma=t(X)\cap \bigl(w+\operatorname{relint}\sigma\bigr) \cap B_\varepsilon(w) is nonempty. If σ={0}\sigma=\{0\} this follows from w∈t(X)w\in t(X). Otherwise choose a nonzero rational vector u∈relint⁡σu\in\operatorname{relint}\sigma. There is a strict rational sector containing uu which meets CC only in relint⁡σ\operatorname{relint}\sigma. To see this, choose an integral linear form ℓ\ell with ℓ(u)>0\ell(u)>0 and use the slice ℓ(v)=1\ell(v)=1. In this slice a sufficiently small rational open box around u/ℓ(u)u/\ell(u) avoids the finitely many other cones and the proper faces of σ\sigma. Its defining inequalities, after multiplication by the positive ℓ(v)\ell(v) and clearing denominators, together with ℓ(v)>0\ell(v)>0, define the required strict sector. Sector lifting from the preceding proposition, followed by th(XL)=t(X)−wt_h(X_L)=t(X)-w, gives points of EσE_\sigma arbitrarily close to ww.

Take v∈Eσv\in E_\sigma and put Nσ=Zn∩span⁡RσN_\sigma=\mathbf Z^n\cap\operatorname{span}_{\mathbf R}\sigma. This is saturated of rank rr: a rational basis of the span becomes integral after clearing denominators. Since the fan is finite and v−wv-w belongs to the relative interior of a maximal cone, the moment image in a sufficiently small neighbourhood of vv lies in v+(Nσ)Rv+(N_\sigma)_{\mathbf R}. Choose a closed rectangular box PvP_v in that neighbourhood, with vv in its interior, and put Yv=t−1(Pv)Y_v=t^{-1}(P_v). Choose an integral projection Q:Zn→ZrQ:\mathbf Z^n\to\mathbf Z^r that is injective on NσN_\sigma, using Lemma 5.14.1. Its monomials are analytic units on YvY_v. The entire level of Q∘t∣YvQ\circ t|_{Y_v} at QvQv is exactly Fv=t−1(v)F_v=t^{-1}(v), since QQ is injective on this affine plane.

Every point of FvF_v is interior over KK, because the chosen ball avoids t(∂KX)t(\partial_KX). The defining inequalities of YvY_v are strict on FvF_v, so its actual rational presentation puts Fv⊂Int⁡(Yv/X)F_v\subset\operatorname{Int}(Y_v/X). Composition in Proposition 2.33.21 therefore gives Fv⊂Int⁡(Yv/K)F_v\subset\operatorname{Int}(Y_v/K). Theorem 5.31.1 now makes the actual algebra of FvF_v finite over the rr-dimensional exact-radius polytorus at QvQv. The rational universal property 2.31.2 identifies this level algebra with the original full fibre algebra in XX. By Theorem 5.35.3 its analytic dimension is dd. Lemma 5.38.3 and Proposition 5.35.4 give d≤r.d\leq r. The residual fan bound gives r≤dr\leq d. Thus every maximal cone has dimension dd.

At every point of EσE_\sigma, the hypotheses of Theorem 5.37.2 are now satisfied over the original field: XX is pure of analytic dimension dd, the whole fibre is interior, and the local image lies in the rank-dd saturated tangent plane. That theorem makes EσE_\sigma relatively open in Uσ=(w+relint⁡σ)∩Bε(w).U_\sigma=\bigl(w+\operatorname{relint}\sigma\bigr) \cap B_\varepsilon(w). It is relatively closed there because t(X)t(X) is compact. The set UσU_\sigma is convex and nonempty. Starting at a point of EσE_\sigma, join it to any point of UσU_\sigma by its line segment. The inverse image of EσE_\sigma on [0,1][0,1] is open and closed and contains 00. If it were not all of [0,1][0,1], the supremum of the initial interval it contains would contradict either its closedness or its openness. Thus Eσ=UσE_\sigma=U_\sigma. This applies also to the zero-dimensional case, where Uσ={w}U_\sigma=\{w\}. The relative interiors of the maximal cones are dense in their union. Closedness of t(X)t(X) therefore fills all their faces inside the ball. Together with the opposite inclusion already proved, this yields t(X)∩Bε(w)=(w+C)∩Bε(w).t(X)\cap B_\varepsilon(w) =(w+C)\cap B_\varepsilon(w).

Finally, two choices of normalizers, extension, or model give conical supports with the same germ at 00 by this equality. Multiplying any vector by a sufficiently small positive real number puts it in that common neighbourhood. Homogeneity of the two supports then makes them equal everywhere. ◻

Corollary 5.38.7. NAG-002354 (Local polyhedrality without a purity assumption). Let X=M(B)X=\mathcal M(B) be a generalized affinoid space over any complete non-Archimedean field with its specified absolute value, and let t=tft=t_f be the moment map of finitely many analytic units. If w∈t(X)w\in t(X) and the whole fibre FwF_w is interior over KK, there is a finite rational polyhedral fan CC and a neighbourhood VV of ww such that t(X)∩V=(w+∣C∣)∩V.t(X)\cap V=(w+|C|)\cap V. More precisely, ∣C∣|C| is the union of the pure fans supplied by Theorem 5.38.6 for the irreducible components of XX meeting FwF_w, using their analytic dimensions.

Proof. The generalized affinoid algebra BB is Noetherian by Theorem 1.18.3, applied to the ideals of the finite module BB over itself, and has finitely many minimal primes pi\mathfrak p_i, by [StacksAlg] (Tag 00FR). The closed analytic subspaces Xi=M(B/pi)X_i=\mathcal M(B/\mathfrak p_i) cover XX: the prime kernel of every analytic point contains a minimal prime by [StacksAlg] (Tag 00E0(4)), and the point factors boundedly through that quotient. The quotient affinoid structure is again given by Corollary 1.18.4. Each nonempty XiX_i is pure of its analytic dimension, since its algebra is a domain.

The closed immersion Xi→XX_i\to X is finite, hence everywhere inner by Theorem 2.33.18. Composition in Proposition 2.33.21 therefore makes every point of Xi∩FwX_i\cap F_w interior over KK. For those ii for which this intersection is nonempty, apply Theorem 5.38.6 to XiX_i. For every other ii, Proposition 2.24.1 makes t(Xi)t(X_i) compact. It avoids ww and can be removed by shrinking the neighbourhood of ww. Intersect the finitely many resulting neighbourhoods. The equality for XX follows by taking their union. Subdividing by the finitely many rational hyperplanes used to define all the cones gives a common rational polyhedral refinement, and hence CC with the stated support. ◻

5.39 Finite polyhedrality of actual analytic moment images

Throughout this section KK is complete for its specified real-valued non-Archimedean absolute value. It may be trivially valued, and affinoid presentation radii may be arbitrary positive real numbers. The dimension is analytic dimension. The polyhedral inequality descriptions have integral linear parts and arbitrary real constants.

Theorem 5.39.1. NAG-002356 (A finite Gauss-fibre cover of the analytic boundary). Let X=M(A)X=\mathcal M(A) be a nonempty generalized affinoid space of analytic dimension dd. There are finitely many elements aj∈Aa_j\in A, with positive spectral radii rj=ρA(aj)r_j=\rho_A(a_j), and bounded maps K{rj−1T}⟶A,T⟼aj,K\{r_j^{-1}T\}\longrightarrow A, \qquad T\longmapsto a_j, with the following property. Let ηj\eta_j be the radius-rjr_j Gauss point, let Hj=H(ηj)H_j=\mathcal H(\eta_j) with its specified absolute value, and put Xj=MHj(A⊗^K{rj−1T}Hj).X_j=\mathcal M_{H_j} \bigl(A\widehat\otimes_{K\{r_j^{-1}T\}}H_j\bigr). The natural maps identify these actual analytic fibres with compact subsets of XX, preserve completed residue fields, and satisfy ∂KX=⋃jXj,dim⁡HjXj≤d−1\partial_KX=\bigcup_j X_j, \qquad \dim_{H_j}X_j\leq d-1 for every nonempty XjX_j. The empty list is allowed. In particular an affinoid of analytic dimension zero has empty absolute boundary.

Proof. Choose homogeneous algebra generators aˉj\bar a_j of A~gr\widetilde A^{\rm gr} over K~gr\widetilde K^{\rm gr} by Theorem 3.18.7, and discard zero generators. If aˉj\bar a_j has degree rjr_j, choose a lift aja_j in the spectral filtration. Nonvanishing of its initial form means exactly that ρA(aj)=rj>0\rho_A(a_j)=r_j>0. Theorem 1.20.2, together with the spectral formula 1.20.3, gives a constant CjC_j with ∥ajm∥≤Cjrjm\|a_j^m\|\leq C_jr_j^m for every m≥0m\geq0. Proposition 1.9.3 therefore gives the displayed bounded substitution map. This uses the original admissible norm on AA; it does not require ∥aj∥≤rj\|a_j\|\leq r_j.

Fix x∈Xx\in X and let SxS_x be the image of A~gr\widetilde A^{\rm gr} in H(x)~gr\widetilde{\mathcal H(x)}^{\rm gr}. Write uju_j for the image of aˉj\bar a_j. The graded innerness criterion 3.18.10, applied to evaluation A→H(x)A\to\mathcal H(x) over KK, says that xx is interior precisely when SxS_x is finite over K~gr\widetilde K^{\rm gr}.

A homogeneous element uu of degree rr satisfying any nonzero polynomial relation over the graded base satisfies a monic homogeneous relation. Indeed, take a nonzero homogeneous component of that polynomial relation, for the grading with deg⁡T=r\deg T=r. Its highest nonzero coefficient is a nonzero homogeneous element of the graded field K~gr\widetilde K^{\rm gr}, and is therefore invertible. Divide by that coefficient. Thus every such uu is integral. If all the uju_j are integral, their finitely many monic relations express SxS_x as a finite module, generated by monomials with bounded exponents. Conversely, in a finite algebra every element is integral: expressing multiplication by it on a finite generating family and applying the adjugate to the resulting square matrix gives a monic annihilating polynomial, which annihilates 11 as well. Consequently x∈∂KX⟺some uj satisfies no nonzero homogeneouspolynomial relation over K~gr.\begin{aligned} x\in\partial_KX\quad\Longleftrightarrow\quad &\text{some }u_j\text{ satisfies no nonzero homogeneous}\\ &\text{polynomial relation over }\widetilde K^{\rm gr}. \end{aligned}

For a fixed jj, this last condition is equivalent to the image of xx in the radius-rjr_j disc being ηj\eta_j. To prove one direction, let P(T)=∑iciTiP(T)=\sum_i c_iT^i be a nonzero polynomial over KK, and put R=max⁡i∣ci∣rjiR=\max_i|c_i|r_j^i. The terms attaining this maximum give a nonzero homogeneous polynomial in the initial variable. If uju_j has no relation, its value is nonzero. Hence ∣P(aj(x))∣=R.|P(a_j(x))|=R. Weighted polynomials are dense in K{rj−1T}K\{r_j^{-1}T\}, so these equalities identify the bounded seminorm on the whole disc algebra with the Gauss seminorm. Conversely, at that Gauss seminorm one has ∣aj(x)∣=rj|a_j(x)|=r_j. Lift the finitely many coefficients of a proposed nonzero homogeneous relation for uju_j to elements of KK with their exact prescribed sizes. All its terms have one common weighted size. A relation of initial forms would make the evaluated polynomial have strictly smaller size, contrary to the Gauss equality. This proves the equivalence and the asserted equality of boundary sets.

Theorem 2.33.11 constructs XjX_j as the actual generalized HjH_j-affinoid fibre. It gives its homeomorphism onto the full inverse image of ηj\eta_j and the isometric identification of its completed residue field at a point zz with HA(x)\mathcal H_A(x) at the corresponding point x∈Xx\in X. The field HjH_j is the completion of K(T)K(T) for the radius-rjr_j Gauss absolute value: weighted polynomial density also gives density of their fractions. Lemma 3.22.4 gives d(Hj/K)=1d(H_j/K)=1. The tower formula of Lemma 3.22.2 therefore gives dHj(z)=d(HA(x)/Hj)=dK(x)−1≤d−1.d_{H_j}(z)=d(\mathcal H_A(x)/H_j)=d_K(x)-1\leq d-1. Taking the maximum on the nonempty affinoid XjX_j and using Theorem 3.22.7 proves the dimension bound. If d=0d=0, a nonempty fibre would have a point with negative relative invariant, which is impossible. Thus the boundary is empty in that case. Compactness of the fibres is Proposition 2.24.1. ◻

Theorem 5.39.2. NAG-002357 (Finite polyhedral upper bounds without an interior hypothesis). Let X=M(A)X=\mathcal M(A) be a nonempty generalized affinoid space of analytic dimension dd, and let analytic units f1,…,fnf_1,\ldots,f_n define t:X→Rnt:X\to\mathbb R^n. For every w∈t(X)w\in t(X) there are a finite rational polyhedral fan CC of dimension at most dd and a neighbourhood VV of ww such that t(X)∩V⊂(w+∣C∣)∩V.t(X)\cap V\subset(w+|C|)\cap V. Consequently t(X)t(X) is contained in the support of a finite integral affine polyhedral complex of dimension at most dd.

Proof. Choose a complete isometric extension L/KL/K containing elements with sizes equal to every radius in a fixed presentation of AA, and elements cic_i with ∣ci∣=e−wi|c_i|=e^{-w_i}; require also that LL be nontrivially valued. Such a field is explicit: take the completion of a rational function field with weighted Gauss norms assigning these finitely many positive radii to its variables, and add a variable of size strictly between zero and one. The Gauss construction is Proposition 1.2.2; its valued completion is the one in Lemma 3.22.4. The actual scalar presentation 2.33.5, followed by rescaling the finitely many presentation variables, makes XLX_L strict. Its dimension is dd by Theorem 3.21.8.

Put hi=fi/cih_i=f_i/c_i. Scalar spectral surjectivity 2.33.13 gives th(XL)=t(X)−w.t_h(X_L)=t(X)-w. It suffices to prove the upper bound at zero for this actual strict space. Its whole fibre F=th−1(0)F=t_h^{-1}(0) is a nonempty strict affinoid domain by Theorem 2.31.2. Let DD be its algebra. It has an admissible integral model D0D_0 by Proposition 3.14.7. The elements hi±1h_i^{\pm1} are power-bounded in DD, since their values on FF are one and Theorem 1.20.2 applies. Adjoin them to D0D_0 by Theorem 3.16.8. The resulting algebra is still an actual admissible model of the same DD, and contains all these elements. Use this enlarged model in what follows.

Write k=L~k=\widetilde L, m=L∘∘\mathfrak m=L^{\circ\circ}, and J=ker⁡(k[T1±1,…,Tn±1]⟶D0/mD0),R=k[T±1]/J.J=\ker\bigl(k[T_1^{\pm1},\ldots,T_n^{\pm1}] \longrightarrow D_0/\mathfrak mD_0\bigr), \qquad R=k[T^{\pm1}]/J. Thus RR injects into the actual special-fibre algebra. For every minimal prime p\mathfrak p of RR, localization preserves this injection, by [StacksAlg] (Tag 00HT). The algebra (D0/mD0)⊗RRp(D_0/\mathfrak mD_0)\otimes_RR_{\mathfrak p} is nonzero, so it has a prime by [StacksAlg] (Tag 00E0(1)). Its contraction is the unique prime of RpR_{\mathfrak p}. It therefore gives a prime q\mathfrak q of the special fibre lying above p\mathfrak p; no finiteness of the injection R→D0/mD0R\to D_0/\mathfrak mD_0 is required here. Actual specialization surjectivity, Theorem 3.16.15, supplies x∈Fx\in F specializing to q\mathfrak q. The induced injections give Frac⁡(R/p)↪H(x)~.\operatorname{Frac}(R/\mathfrak p) \hookrightarrow\widetilde{\mathcal H(x)}. By [StacksAlg] (Tag 00P0) and Theorem 3.22.7, dim⁡(R/p)=trdeg⁡kFrac⁡(R/p)≤dL(x)≤d.\dim(R/\mathfrak p) =\operatorname{trdeg}_k\operatorname{Frac}(R/\mathfrak p) \leq d_L(x)\leq d. The residue field here is the same as for x∈XLx\in X_L, by Corollary 2.33.15.

The finite-type kk-algebra RR is Noetherian by [StacksAlg] (Tag 00FO), and has finitely many minimal primes, by [StacksAlg] (Tag 00FR). Its tropical support is the union of the supports of the corresponding integral quotients: every analytic prime kernel contains a minimal prime by [StacksAlg] (Tag 00E0(4)). Theorem 5.19.5 bounds the dimension of each of these algebraic tropicalizations by dd. Lemma 5.38.2 gives a finite rational fan CC with support Trop⁡(Spec⁡R)\operatorname{Trop}(\operatorname{Spec}R) and finitely many nonzero pj∈Jp_j\in J cutting out that support tropically. Its dimension is therefore at most dd.

Lift every nonzero coefficient of each pjp_j to an element of L∘L^\circ of absolute value one, obtaining a Laurent polynomial PjP_j. Since pj∈Jp_j\in J, Pj(h)∈mD0P_j(h)\in\mathfrak mD_0. Express this membership as a finite sum ∑ℓbjℓdjℓ\sum_\ell b_{j\ell}d_{j\ell} with ∣bjℓ∣<1|b_{j\ell}|<1 and djℓ∈D0d_{j\ell}\in D_0. Every djℓd_{j\ell} has absolute value at most one on FF, by the actual model presentation. Hence there is a real cj<1c_j<1 with ∣Pj(h)∣<cj|P_j(h)|<c_j throughout FF. The compact subset of XLX_L where ∣Pj(h)∣≥cj|P_j(h)|\geq c_j has compact moment image missing zero. Avoiding these finitely many images gives a neighbourhood V0V_0 of zero where all these strict inequalities hold on the entire inverse image. This uses the compact spectrum of Proposition 2.24.1.

Shrink V0V_0 further so that, for every monomial of every PjP_j, its norm at a weight v∈V0v\in V_0 is greater than cjc_j. At zero all these finitely many norms equal one, so continuity permits the simultaneous choice. If an actual moment value v∈V0v\in V_0 lay outside ∣C∣|C|, some pjp_j would have a unique least-weight term there. The same term of Pj(h)P_j(h) would have strictly largest absolute value at any point over vv, and that value would exceed cjc_j. The non-Archimedean triangle equality would give ∣Pj(h)∣>cj|P_j(h)|>c_j, a contradiction. Thus th(XL)∩V0⊂∣C∣t_h(X_L)\cap V_0\subset |C|. Translating back proves the local assertion over the original field and norm.

Finally, at every w∈t(X)w\in t(X) choose a closed rectangular box centred at ww inside its upper-bound neighbourhood. Finitely many box interiors cover the compact set t(X)t(X). The union of the intersections of these boxes with the corresponding translated finite fans contains t(X)t(X). It is a finite union of integral affine polytopes, all of dimension at most dd. Cutting by their finitely many defining hyperplanes gives a finite common complex, as in Lemma 5.14.2. This proves the final assertion. ◻

Theorem 5.39.3. NAG-002358 (Global polyhedrality and the tropicalized boundary). Let XX be a nonempty generalized affinoid space of analytic dimension dd over KK, and let tt be the moment map of any finite tuple of analytic units. The actual image t(X)t(X) is the support of a finite integral affine polyhedral complex of dimension at most dd. The actual boundary image t(∂KX)t(\partial_KX) is likewise finite polyhedral, of dimension at most d−1d-1; it is empty when d=0d=0.

If XX is pure of analytic dimension dd, every point of t(X)∖t(∂KX)t(X)\setminus t(\partial_KX) has the pure dd-dimensional local fan of Theorem 5.38.6.

Proof. We induct on the analytic dimension, simultaneously for all complete specified ground fields, all positive real radii, and all unit tuples. Theorem 5.39.2 gives a finite polyhedral upper bound of dimension zero when d=0d=0, hence a finite set containing t(X)t(X). Any subset of a finite set is a finite zero-dimensional complex. Theorem 5.39.1 makes the boundary empty. This establishes the induction base.

Assume the assertion for every analytic dimension less than dd. First suppose XX is pure of dimension dd. Use Theorem 5.39.1 to cover its actual boundary by the finitely many actual affinoid fibres XjX_j over their specified fields HjH_j, each of dimension at most d−1d-1. The restrictions of all original moment functions remain analytic units there. The homeomorphisms and residue-field identifications in that theorem preserve their absolute values. By induction, B=t(∂KX)=⋃jt(Xj)B=t(\partial_KX)=\bigcup_j t(X_j) is a finite union of integral affine polytopes of dimension at most d−1d-1. In particular it admits a finite complex structure of that dimension.

Put S=t(X)S=t(X). By Theorem 5.39.2, choose a finite complex P\mathcal P of dimension at most dd whose support contains SS. The upper-bound construction uses bounded boxes, so all its cells may be taken to be polytopes. Since B⊂SB\subset S, cutting by the defining hyperplanes of both complexes gives a common refinement of P\mathcal P in which BB is a subcomplex.

For a dd-cell σ\sigma of this refinement, its relative interior misses BB. The set S∩relint⁡(σ)S\cap\operatorname{relint}(\sigma) is relatively closed because SS is compact. It is also relatively open. Indeed, near one of its points no other cell of the finite complex meets that point, so the actual moment image is contained in the affine span of σ\sigma. The whole moment fibre is interior, since its value does not belong to BB. Theorem 5.37.2, with the saturated rank-dd tangent lattice of σ\sigma, fills a neighbourhood in that affine plane. This proves relative openness. The relative interior of σ\sigma is convex and therefore connected, by the line-segment open-and-closed argument used in Theorem 5.38.6. Thus the intersection is either empty or all of relint⁡(σ)\operatorname{relint}(\sigma).

Select the cells for which it is nonempty and let TT be their union with all their faces. Closedness of SS shows T⊂ST\subset S. We claim S=B∪TS=B\cup T. If v∈S∖(B∪T)v\in S\setminus(B\cup T), choose a neighbourhood missing the closed sets BB and TT. The whole fibre at vv is interior, so Theorem 5.38.6 presents SS there by a pure dd-dimensional translated fan. A relatively open part of a dd-cone in this fan cannot lie in the finitely many affine spans of the lower-dimensional cells of P\mathcal P. To check this, choose a line direction in its span outside the finitely many proper direction subspaces; each remaining proper affine span meets a line in that direction at at most one point. Such a direction exists: in a basis of the dd-plane, the vectors (1,b,…,bd−1)(1,b,\ldots,b^{d-1}), for integers bb, avoid each fixed proper linear subspace except at finitely many bb, by applying a nonzero linear equation for that subspace. Consequently arbitrarily close to vv there is a point of SS in the relative interior of a dd-cell of P\mathcal P. That cell is selected, contradicting the chosen neighbourhood’s disjointness from TT. This proves the claim. The finite union B∪TB\cup T is polyhedral of dimension at most dd.

For a general X=M(A)X=\mathcal M(A) of dimension dd, the Noetherian algebra AA has finitely many minimal primes, by Theorem 1.18.3 and [StacksAlg] (Tag 00FR). The actual quotient affinoids of Corollary 1.18.4 M(A/p)\mathcal M(A/\mathfrak p) cover XX, because every analytic prime kernel contains a minimal prime, [StacksAlg] (Tag 00E0(4)). Each quotient is a domain, and is pure of its analytic dimension. This dimension is at most dd: its points have the same completed residue fields as their images in XX, so apply Theorem 3.22.7. The pure case just proved handles the components of dimension dd, and induction handles the others. Their finite union of moment images proves the assertion for t(X)t(X). The boundary assertion for this general XX follows directly from its own finite Gauss-fibre cover and the induction hypothesis, exactly as in the pure case. No purity after a coefficient extension is used.

Finally, the stated local purity off the boundary image is precisely Theorem 5.38.6, whose full-fibre hypothesis holds at every such value. ◻

5.40 Generic cycles in actual analytic Gauss fibres

Throughout, KK is complete for its specified real-valued non-Archimedean absolute value. The valuation may be trivial. All affinoid algebras may be generalized, and finite algebra maps carry their canonical finite-module Banach structures. For a positive real polyradius rr, write R=AK,rR=A_{K,r} and H=HK,rH=H_{K,r} for the actual Laurent algebra and its Gauss field in Section 5.36.

Proposition 5.40.1. NAG-002360 (A Gauss fibre sees the generic component lengths). Let AA be a generalized KK-affinoid algebra, and let A→DA\to D represent an affinoid domain of M(A)\mathcal M(A). Suppose that a bounded map R→DR\to D is finite, and put E=D⊗^RH.E=D\widehat\otimes_R H. Then E=D⊗RHE=D\otimes_RH is a finite-dimensional HH-algebra and is flat as an ordinary AA-module. Every prime of EE contracts to a minimal prime of AA.

For a finite AA-module MM and p∈Min⁡(A)\mathfrak p\in\operatorname{Min}(A), put βp(M)=length⁡Ap(Mp),λp(E)=dim⁡H(E/pE).\beta_{\mathfrak p}(M)= \operatorname{length}_{A_{\mathfrak p}}(M_{\mathfrak p}), \qquad \lambda_{\mathfrak p}(E)=\dim_H(E/\mathfrak pE). These are finite nonnegative integers, and dim⁡H(M⊗AE)=∑p∈Min⁡(A)βp(M)λp(E).\begin{equation} \dim_H(M\otimes_AE) =\sum_{\mathfrak p\in\operatorname{Min}(A)} \beta_{\mathfrak p}(M)\lambda_{\mathfrak p}(E). \end{equation} The term λp(E)\lambda_{\mathfrak p}(E) is positive exactly when a prime of EE contracts to p\mathfrak p.

If an ideal I⊂AI\subset A is contained in no minimal prime, then IE=EIE=E. If Mp=0M_{\mathfrak p}=0 at every minimal prime, then M⊗AE=0M\otimes_AE=0.

Proof. The Laurent Gauss norm is multiplicative and has zero kernel, by the actual presentation and norm calculation in Theorem 1.17.2 and Proposition 1.2.2. Hence RR is a domain and R→HR\to H factors through Frac⁡(R)\operatorname{Frac}(R). is flat by [StacksAlg] (Tag 00HT(1)). A module over a field is flat by the ideal criterion [StacksAlg] (Tag 00HD), since that field has only the zero and unit ideals. It follows that HH is flat over RR.

Proposition 2.33.19 identifies EE with the displayed ordinary tensor, with its complete canonical finite-dimensional norm. The homomorphism D→ED\to E is flat: tensoring a DD-module with EE is tensoring its underlying RR-module with HH. The domain map A→DA\to D is flat by Proposition 2.38.2. Composing these two exact tensor functors proves that A→EA\to E is flat.

algebra EE is Artinian and is a finite product of its local factors EjE_j, by [StacksAlg] (Tag 00JB). This includes the empty product when E=0E=0. Let qj\mathfrak q_j be the maximal ideal of a nonzero factor, and let pj\mathfrak p_j be its contraction in AA. A strictly smaller prime of AA would, by flat going down [StacksAlg] (Tag 00HS), lift to a strictly smaller prime under the corresponding prime of EE. Every prime of an Artinian ring is minimal. Thus pj\mathfrak p_j is minimal in AA.

The ring AA is Noetherian by Theorem 1.18.3, and has finitely many minimal primes by [StacksAlg] (Tag 00FR). For any such prime p\mathfrak p, the local ring ApA_{\mathfrak p} is Noetherian of dimension zero, by [StacksAlg] (Tag 00FN) and the minimality of p\mathfrak p. It is Artinian by [StacksAlg] (Tag 00KH). Therefore every finite module over it has finite length.

Fix jj and put p=pj\mathfrak p=\mathfrak p_j. Every element of A∖pA\setminus\mathfrak p maps to a unit of EjE_j, so A→EjA\to E_j factors through ApA_{\mathfrak p}. The factor EjE_j is flat over AA, being a direct summand of EE, and hence is flat over ApA_{\mathfrak p} by [StacksAlg] (Tag 00HT(2)). Take a composition series of MpM_{\mathfrak p}. Every simple quotient is κ(p)\kappa(\mathfrak p): a simple module over a commutative local ring is its residue field. Tensoring this series with EjE_j remains exact and gives dim⁡H(M⊗AEj)=βp(M) dim⁡H(Ej/pEj).\dim_H(M\otimes_AE_j) =\beta_{\mathfrak p}(M)\, \dim_H(E_j/\mathfrak pE_j). The quotient on the right is nonzero, since pEj⊂qj\mathfrak pE_j\subset\mathfrak q_j.

For a different minimal prime p′≠pj\mathfrak p'\ne\mathfrak p_j, choose an element of p′\mathfrak p' outside pj\mathfrak p_j. Its image is a unit of EjE_j, so Ej/p′Ej=0E_j/\mathfrak p'E_j=0. Consequently E/pEE/\mathfrak pE is exactly the product of Ej/pEjE_j/\mathfrak pE_j over the factors with contraction p\mathfrak p. Summing the preceding dimension identity proves (5.40.1) and the positivity assertion.

Finally, if II is contained in no minimal prime, then for each jj it contains an element outside pj\mathfrak p_j. Thus IEj=EjIE_j=E_j, and hence IE=EIE=E. The last assertion follows from the displayed length formula. ◻

Theorem 5.40.2. NAG-002362 (Weights of finite modules on an interior flat germ). Let Y=M(A)Y=\mathcal M(A) be nonempty and pure of analytic dimension dd. Let a unit moment tuple ff define t:Y→Rnt:Y\to\mathbb R^n, and let w∈t(Y)w\in t(Y). Suppose that the entire fibre Fw=t−1(w)F_w=t^{-1}(w) lies in Int⁡(Y/K)\operatorname{Int}(Y/K) and that, for some neighborhood WW of ww and a saturated rank-dd lattice N⊂ZnN\subset\mathbb Z^n, t(Y)∩W⊂w+NR.t(Y)\cap W\subset w+N_{\mathbb R}. For p∈Min⁡(A)\mathfrak p\in\operatorname{Min}(A), put Yp=M(A/p)Y_{\mathfrak p}=\mathcal M(A/\mathfrak p). Set mp(w)=0m_{\mathfrak p}(w)=0 if its moment fibre at ww is empty; otherwise this fibre satisfies the same flat-germ hypotheses, and let mp(w)m_{\mathfrak p}(w) be its positive intrinsic weight.

Let MM be a finite AA-module. For any integral Q:Zn→ZdQ:\mathbb Z^n\to\mathbb Z^d injective on NN, let DwD_w be the actual algebra of FwF_w, let RQR_Q be the polytorus algebra at QwQw, and let HQH_Q be its Gauss field. Put EQ=Dw⊗^RQHQ,eQ=[Zd:Q(N)],δQ(M,w)=dim⁡HQ(M⊗AEQ).E_Q=D_w\widehat\otimes_{R_Q}H_Q,\qquad e_Q=[\mathbb Z^d:Q(N)],\qquad \delta_Q(M,w)=\dim_{H_Q}(M\otimes_AE_Q). Then δQ(M,w)eQ=∑p∈Min⁡(A)βp(M)mp(w)∈Z≥0.\begin{equation} \frac{\delta_Q(M,w)}{e_Q} =\sum_{\mathfrak p\in\operatorname{Min}(A)} \beta_{\mathfrak p}(M)m_{\mathfrak p}(w) \in\mathbb Z_{\geq0}. \end{equation} In particular this integer, denoted mM,f(w)m_{M,f}(w), is independent of QQ and is additive in short exact sequences of finite AA-modules. For M=AM=A it is the intrinsic weight of YY: mY,f(w)=∑p∈Min⁡(A)length⁡(Ap)mp(w).m_{Y,f}(w)= \sum_{\mathfrak p\in\operatorname{Min}(A)} \operatorname{length}(A_{\mathfrak p})m_{\mathfrak p}(w). These module weights are unchanged by restriction to an affinoid domain containing the whole fibre, whenever the two underlying pure-dimensional germs satisfy the stated flat-germ hypotheses.

Proof. Theorem 5.37.1 gives every finite map RQ→DwR_Q\to D_w, with its canonical norm and actual Gauss fibre. Apply Proposition 5.40.1 to the domain Fw⊂YF_w\subset Y.

The quotient A→A/pA\to A/\mathfrak p is finite with its canonical norm. Its spectral map is everywhere interior relative to YY by Theorem 2.33.18. The composition inclusion in Proposition 2.33.21 therefore makes every point over FwF_w interior over KK. The affinoid YpY_{\mathfrak p} is pure of analytic dimension dd by the assumed purity of AA, and its moment image is contained in that of YY. If its fibre at ww is nonempty, Theorem 5.37.1 applies with the same lattice NN.

The actual domain base-change theorem 2.36.9 and finite tensor comparison 2.33.19 identify its full fibre algebra with Dw/pDwD_w/\mathfrak pD_w. This is finite over RQR_Q. Its actual Gauss algebra is (Dw/pDw)⊗^RQHQ=EQ/pEQ.(D_w/\mathfrak pD_w)\widehat\otimes_{R_Q}H_Q =E_Q/\mathfrak pE_Q. For a nonempty fibre its dimension is eQmp(w)e_Qm_{\mathfrak p}(w) by the flat-germ weight theorem. For an empty fibre the actual algebra is zero, by Theorem 1.20.1, and the same identity holds with mp(w)=0m_{\mathfrak p}(w)=0. Substitution in (5.40.1) proves (5.40.2).

All tensors in this argument also have their canonical analytic meaning for MM. The finite module M⊗ADwM\otimes_AD_w has its complete canonical norm by Proposition 1.18.7. A quotient of a finite sum of copies of the finite canonical RQR_Q-module DwD_w gives its canonical finite-RQR_Q-module topology: compose the two finite free quotient presentations. Applying the same tensor comparison over RQR_Q shows that M⊗AEQM\otimes_AE_Q is already complete. Thus no separated completion discards a submodule in the length calculation.

Flatness of A→EQA\to E_Q makes its tensor functor exact, proving additivity. For M=AM=A the numerator is the ordinary Gauss degree of DwD_w, so the formula gives exactly mY,f(w)m_{Y,f}(w).

Finally let A→AVA\to A_V represent a domain containing FwF_w, and replace MM by M⊗AAVM\otimes_AA_V. The actual full fibre algebra is still DwD_w, by Theorem 2.36.9. The saturated tangent lattice is unchanged by Theorem 5.37.2. For a common QQ, associativity gives (M⊗AAV)⊗AVEQ=M⊗AEQ.(M\otimes_AA_V)\otimes_{A_V}E_Q=M\otimes_AE_Q. Both the Gauss dimension and its lattice divisor are the same, which proves the restriction assertion. ◻

5.41 Finite analytic pushforward with generic cycle degrees

Theorem 5.41.1. NAG-002365 (Finite pushforward with full generic lengths). Let p:X=M(B)→Y=M(A)p:X=\mathcal M(B)\to Y=\mathcal M(A) be finite, with the canonical finite-AA-module norm on BB. Suppose that XX and YY are pure of analytic dimension dd. Let unit moment tuples define tX:X→Rnt_X:X\to\mathbb R^n and tY:Y→Rn′t_Y:Y\to\mathbb R^{n'}, and let a:Zn→Zn′a:\mathbb Z^n\to\mathbb Z^{n'} be integral. Assume the corresponding identity of characters, so that tY∘p=a∘tXt_Y\circ p=a\circ t_X.

Let w′∈tY(Y)w'\in t_Y(Y) have its whole fibre in Int⁡(Y/K)\operatorname{Int}(Y/K), and suppose that the local target image is contained in w′+(NY)Rw'+(N_Y)_{\mathbb R} for a saturated rank-dd lattice NYN_Y. Assume that a−1(w′)∩tX(X)={w1,…,ws}a^{-1}(w')\cap t_X(X)=\{w_1,\ldots,w_s\} is finite and nonempty and that the local image at each wiw_i is contained in wi+(Ni)Rw_i+(N_i)_{\mathbb R} for a saturated rank-dd lattice NiN_i.

For every ii, the lattice a(Ni)a(N_i) is contained in NYN_Y with finite index. Write mp(w′)m_{\mathfrak p}(w') for the component weights of YY in Theorem 5.40.2. Then ∑i=1smX,fX(wi)[NY:a(Ni)]=mB,fY(w′)=∑p∈Min⁡(A)βp(B)mp(w′).\begin{equation} \sum_{i=1}^s m_{X,f_X}(w_i)[N_Y:a(N_i)] =m_{B,f_Y}(w') =\sum_{\mathfrak p\in\operatorname{Min}(A)} \beta_{\mathfrak p}(B)m_{\mathfrak p}(w'). \end{equation} The coefficients are the full generic pushforward lengths: βp(B)=∑q∈Min⁡(B)q∩A=plength⁡Bq(Bq)[κ(q):κ(p)].\beta_{\mathfrak p}(B)= \sum_{\substack{\mathfrak q\in\operatorname{Min}(B)\\ \mathfrak q\cap A=\mathfrak p}} \operatorname{length}_{B_{\mathfrak q}}(B_{\mathfrak q}) [\kappa(\mathfrak q):\kappa(\mathfrak p)].

Consequently, if a positive rational number Δ\Delta satisfies βp(B)=Δ length⁡(Ap)\beta_{\mathfrak p}(B)= \Delta\,\operatorname{length}(A_{\mathfrak p}) for every component with nonempty moment fibre at w′w', then Δ mY,fY(w′)=∑i=1smX,fX(wi)[NY:a(Ni)].\Delta\,m_{Y,f_Y}(w') =\sum_{i=1}^s m_{X,f_X}(w_i)[N_Y:a(N_i)]. For a reduced integral target, Δ\Delta is the positive integer dim⁡Frac⁡(A)(B⊗AFrac⁡(A))\dim_{\operatorname{Frac}(A)} (B\otimes_A\operatorname{Frac}(A)). If BB is also integral, it is [Frac⁡(B):Frac⁡(A)][\operatorname{Frac}(B):\operatorname{Frac}(A)].

Proof. Finite relative interior, Theorem 2.33.18, and the composition inclusion 2.33.21 put every whole source fibre over w′w' in Int⁡(X/K)\operatorname{Int}(X/K). Theorem 5.37.1 therefore defines all source and target weights. Their local images fill relative neighborhoods in the specified planes by Theorem 5.37.2. The equality of moment maps then gives a((Ni)R)⊂(NY)Ra((N_i)_{\mathbb R})\subset(N_Y)_{\mathbb R}. Saturation gives a(Ni)⊂NYa(N_i)\subset N_Y. A nonzero vector in the kernel of a∣(Ni)Ra|_{(N_i)_{\mathbb R}} would produce a small interval of source moment weights above w′w', contrary to the stated finite set. Thus a∣Nia|_{N_i} is injective, and its image has finite index in NYN_Y.

Choose an integral Q:Zn′→ZdQ:\mathbf Z^{n'}\to\mathbf Z^d injective on NYN_Y by Lemma 5.14.1. Inside the target flat-germ neighbourhood choose a closed box P=∏j[αj,βj]P=\prod_j[\alpha_j,\beta_j] with w′w' in its interior. Write gj=fY,jg_j=f_{Y,j} and Y0=tY−1(P)Y_0=t_Y^{-1}(P). The inequalities ∣gj∣≤e−αj,∣gj−1∣≤eβj|g_j|\leq e^{-\alpha_j},\qquad |g_j^{-1}|\leq e^{\beta_j} give a single rational presentation over AA with denominator 11. Theorems 2.31.1 and 2.31.2 identify its quotient, with its quotient norm, as the affinoid domain algebra of Y0Y_0. At every yy in the full fibre F′=tY−1(w′)F'=t_Y^{-1}(w'), the two coordinates for each jj are strictly below their displayed presentation radii. Thus y∈Int⁡(Y0/Y)y\in\operatorname{Int}(Y_0/Y) by Definition 2.7.1. Since F′⊂Int⁡(Y/K)F'\subset\operatorname{Int}(Y/K) by hypothesis, Proposition 2.33.21 gives F′⊂Int⁡(Y0/K)F'\subset\operatorname{Int}(Y_0/K).

The image of Y0Y_0 lies in w′+(NY)Rw'+(N_Y)_{\mathbf R}, on which QQ is injective. Its entire projected level at Qw′Qw' is therefore F′F'. For every norm-admissible affinoid test over YY, mapping into F′F' is equivalent to mapping into Y0Y_0 and then into this projected level. The rational domain universal property gives inverse bounded maps between the two representing algebras. Denote this algebra by Dw′′D'_{w'} with its original characters.

By Theorem 2.36.9, X0=p−1(Y0)X_0=p^{-1}(Y_0) and p−1(F′)p^{-1}(F') are the actual pullbacks of the two target domains. Proposition 2.33.19 preserves the finite canonical module structure of X0→Y0X_0\to Y_0. Theorem 2.33.18 and the composition inclusion put every point above F′F' in Int⁡(X0/K)\operatorname{Int}(X_0/K). By the character identities, the whole QaQa-level on X0X_0 at Qw′Qw' is p−1(F′)p^{-1}(F'). Its source moment values are precisely the wiw_i: every original source point with tX=wit_X=w_i already maps into F′F'. The same universal property identifies the projected source level with the pulled-back domain over XX, and each exact wiw_i-domain in X0X_0 with the original whole domain in XX. Proposition 5.31.2 now splits the projected level algebra into the product of their actual algebras. Consequently there are bounded algebra isomorphisms ∏iDwi ≃ B⊗^ADw′′ ≃ B⊗ADw′′.\prod_iD_{w_i}\ \simeq\ B\widehat\otimes_A D'_{w'}\ \simeq\ B\otimes_A D'_{w'} . The last isomorphism uses the given finite canonical norm and Proposition 2.33.19. The maps respect the projected target characters, since they come from restriction and pullback of the specified character identity.

Let RQR_Q be the exact-radius target at Qw′Qw' and HQH_Q its Gauss field. All finite tensors are ordinary complete canonical tensors by Proposition 2.33.19. Tensoring the preceding product with HQH_Q and taking dimensions gives dim⁡HQ(B⊗AEQ)=∑iδQa(wi).\dim_{H_Q}(B\otimes_AE_Q) =\sum_i\delta_{Qa}(w_i). Theorem 5.40.2 evaluates the left side as [Zd:Q(NY)]mB,fY(w′)[\mathbb Z^d:Q(N_Y)]m_{B,f_Y}(w'). Theorem 5.37.1 evaluates each right-hand term as [Zd:Qa(Ni)]mX,fX(wi)[\mathbb Z^d:Qa(N_i)]m_{X,f_X}(w_i). Counting representatives in Qa(Ni)⊂Q(NY)⊂ZdQa(N_i)\subset Q(N_Y)\subset\mathbb Z^d gives [Zd:Qa(Ni)]=[Zd:Q(NY)] [NY:a(Ni)].[\mathbb Z^d:Qa(N_i)] =[\mathbb Z^d:Q(N_Y)]\,[N_Y:a(N_i)]. Cancellation proves (5.41.1).

For the coefficient formula, fix a minimal prime p\mathfrak p of AA. The ring ApA_{\mathfrak p} is Artinian by the first proposition. Its finite algebra Bp=B⊗AApB_{\mathfrak p}=B\otimes_AA_{\mathfrak p} has finite length as an ApA_{\mathfrak p}-module, and hence is Artinian as a ring. Decompose it into its local factors by [StacksAlg] (Tag 00JB). They are BqB_{\mathfrak q} for primes q\mathfrak q of BB over p\mathfrak p. Such primes are minimal in BB: a smaller prime would have the same contraction, contradicting incomparability [StacksAlg] (Tag 00GT) for the finite, hence integral, map [StacksAlg] (Tag 00GK). Every simple quotient in a composition series of BqB_{\mathfrak q} is κ(q)\kappa(\mathfrak q), whose length over ApA_{\mathfrak p} is its full finite dimension over κ(p)\kappa(\mathfrak p). Adding these lengths gives the asserted coefficient formula.

The target formula in Theorem 5.40.2 now proves the assertion with Δ\Delta. If AA is an integral domain, its unique generic local ring is Frac⁡(A)\operatorname{Frac}(A) and has length one. Thus β(0)(B)\beta_{(0)}(B) is the displayed vector-space dimension. It is positive because (5.41.1) has a positive left side. If BB is integral, this positivity first makes A→BA\to B injective: a nonzero element of its kernel would become invertible and annihilate the generic algebra. Localizing the domain BB at the nonzero elements of AA is therefore injective. The resulting finite-dimensional domain is a field and hence is precisely Frac⁡(B)\operatorname{Frac}(B). This proves the last assertion. ◻

Proposition 5.41.2. NAG-002367 (The free locus of a finite analytic module). Let AA be a generalized KK-affinoid algebra and let MM be a finite AA-module. The set UM={p∈Spec⁡(A):Mp is free over Ap}U_M=\{\mathfrak p\in\operatorname{Spec}(A): M_{\mathfrak p}\text{ is free over }A_{\mathfrak p}\} is open. It is a finite union of principal opens D(gj)D(g_j) on each of which MgjM_{g_j} is finite free. If MM is free at all minimal primes, then the ideal I=(g1,…,gt)I=(g_1,\ldots,g_t) is contained in no minimal prime. This condition holds for every finite MM when AA is reduced.

Let UMan={y∈M(A):ker⁡(y)∈UM}.U_M^{\rm an}=\{y\in\mathcal M(A):\ker(y)\in U_M\}. For every affinoid domain V⊂UManV\subset U_M^{\rm an}, its restriction M⊗AAVM\otimes_AA_V is finite locally free as an ordinary AVA_V-module. Its local bases give bounded isomorphisms with the canonical free-module norms on affinoid domains where the corresponding gjg_j is a unit.

Suppose MM is free at all minimal primes, and let EE be any actual Gauss algebra in Proposition 5.40.1. Then IE=EIE=E. On a local factor EjE_j whose residue prime contracts to p\mathfrak p, the module M⊗AEjM\otimes_AE_j is free of rank rank⁡ApMp\operatorname{rank}_{A_{\mathfrak p}}M_{\mathfrak p}.

Proof. The ring AA is Noetherian and every finite module is finitely presented, by Theorem 1.18.3. Suppose MpM_{\mathfrak p} is free of rank rr. A basis can be represented by elements of MM: multiply its finitely many members by their respective denominators, which are units in ApA_{\mathfrak p}. They give a map Ar→MA^r\to M whose localization at p\mathfrak p is an isomorphism. Its kernel and cokernel are finite and become zero at p\mathfrak p by exactness of localization, [StacksAlg] (Tag 00HT). For each member of finite generating lists, vanishing after localization supplies an annihilator outside p\mathfrak p. The product of these finitely many annihilators is an element g∉pg\notin\mathfrak p for which Mg≃AgrM_g\simeq A_g^r. This proves openness and gives a principal free neighborhood of every point of UMU_M.

Let II be the ideal generated by all these elements gg. It is generated by finitely many of them, since AA is Noetherian: express a finite generating set of II using finitely many of the original gg’s. Thus UM=D(I)=⋃jD(gj)U_M=D(I)=\bigcup_jD(g_j) with the asserted free bases. If every minimal prime belongs to UMU_M, none contains II. When AA is reduced, each ApA_{\mathfrak p} at a minimal prime is a reduced Artinian local ring, by Proposition 5.40.1. Its maximal ideal is nilpotent by [StacksAlg] (Tag 00JB) and hence zero; it is a field. The finite module MpM_{\mathfrak p} is therefore free.

Empty domains are immediate. Let V=M(AV)⊂UManV=\mathcal M(A_V)\subset U_M^{\rm an} be nonempty. There is no point of VV where all gjg_j vanish. The quotient AV/IAVA_V/IA_V is an actual affinoid algebra with its canonical norm, by Theorem 2.33.18, and has empty spectrum. Theorem 1.20.1 makes it zero. Hence IAV=AVIA_V=A_V and the principal opens D(gj)D(g_j) cover Spec⁡(AV)\operatorname{Spec}(A_V). The given algebraic bases base change on these opens and prove that M⊗AAVM\otimes_AA_V is finite locally free. If W⊂VW\subset V is an affinoid domain on which gjg_j is a unit, its corresponding basis gives an algebraic isomorphism M⊗AAW≃AWrj.M\otimes_AA_W\simeq A_W^{r_j}. Both maps are bounded for the canonical norms by Theorem 1.18.3. Such domains cover VV: compactness, Proposition 2.24.1, gives a positive lower bound for max⁡j∣gj(y)∣\max_j|g_j(y)|, and the resulting finitely many Laurent domains ∣gj∣≥ε|g_j|\geq\varepsilon cover. Their actual algebras are given by Theorem 2.31.2.

Finally, IE=EIE=E follows from Proposition 5.40.1. For a local factor over p\mathfrak p, the map A→EjA\to E_j factors through ApA_{\mathfrak p}. Tensor its prescribed free basis with EjE_j to obtain the stated rank. ◻

Theorem 5.41.3. NAG-002368 (Finite neighborhoods of a relative interior level). Let p:X=M(B)→Y=M(A)p:X=\mathcal M(B)\to Y=\mathcal M(A) be any affinoid morphism, and let a unit moment tuple on YY define tY:Y→Rn′t_Y:Y\to\mathbb R^{n'}. Fix w′∈tY(Y)w'\in t_Y(Y) and suppose p−1(tY−1(w′))⊂Int⁡(X/Y).p^{-1}(t_Y^{-1}(w'))\subset\operatorname{Int}(X/Y). There is a closed rectangular box PP about w′w', with w′w' in its interior, for which Y0=tY−1(P),X0=p−1(Y0),p0:X0⟶Y0Y_0=t_Y^{-1}(P),\qquad X_0=p^{-1}(Y_0),\qquad p_0:X_0\longrightarrow Y_0 are actual affinoid domains and a finite morphism with the canonical module Banach structure. The box may be chosen inside any prescribed neighborhood of w′w'.

Suppose in addition that XX and YY are pure of analytic dimension dd, that w′w' is a whole interior flat germ of tYt_Y with saturated lattice NYN_Y, and that unit moment coordinates on XX and an integral lattice map aa satisfy the character identity tY∘p=a∘tXt_Y\circ p=a\circ t_X. Suppose a−1(w′)∩tX(X)={w1,…,ws}a^{-1}(w')\cap t_X(X)=\{w_1,\ldots,w_s\} is finite and nonempty, and that the local source image at each wiw_i lies in wi+(Ni)Rw_i+(N_i)_{\mathbb R} for a saturated rank-dd lattice NiN_i. Write A0A_0 and B0B_0 for the actual algebras of Y0Y_0 and X0X_0. For a sufficiently small box as above, all the original source weights are defined and ∑imX,fX(wi)[NY:a(Ni)]=mB0,fY(w′)=∑p∈Min⁡(A0)length⁡(A0)p((B0)p)m(Y0)p,fY(w′).\begin{aligned} \sum_i m_{X,f_X}(w_i)[N_Y:a(N_i)] &=m_{B_0,f_Y}(w')\\ &=\sum_{\mathfrak p\in\operatorname{Min}(A_0)} \operatorname{length}_{(A_0)_{\mathfrak p}} ((B_0)_{\mathfrak p}) m_{(Y_0)_{\mathfrak p},f_Y}(w'). \end{aligned} The component weight is zero for an empty fibre. The resulting integer is independent of the chosen box. If the displayed generic module lengths are Δ\Delta times those of A0A_0 on all components meeting the fibre, then the middle term is ΔmY,fY(w′)\Delta m_{Y,f_Y}(w').

Proof. By Proposition 2.33.21, the relative boundary Z=X∖Int⁡(X/Y)Z=X\setminus\operatorname{Int}(X/Y) is closed. It is compact by Proposition 2.24.1. The compact set (tY∘p)(Z)(t_Y\circ p)(Z) does not contain w′w', so choose a closed box PP about w′w' avoiding it and inside the prescribed neighborhood. The simultaneous coordinate bounds give the actual domain Y0Y_0 by Theorem 2.31.2. Theorem 2.36.9 gives the actual domain X0X_0, with algebra B⊗^AA0B\widehat\otimes_AA_0.

Every point of X0X_0 comes from Int⁡(X/Y)\operatorname{Int}(X/Y). Base change of relative interior in Proposition 2.33.21 makes p0p_0 everywhere interior. Theorem 2.43.8 proves that B0B_0 is finite over A0A_0 and has its canonical module norm. This proves the first assertion over the original field and absolute value.

For the remaining assertions choose PP also inside the target flat-germ neighborhood. Purity of X0X_0 and Y0Y_0 follows from Theorem 5.35.3. The whole target fibre is contained in Y0Y_0; all its defining inequalities are strict at that fibre. The relative domain presentation and the composition inclusion 2.33.21 therefore make it interior over KK in Y0Y_0. Its weight is the original one by Theorem 5.37.2.

Each entire original source fibre at wiw_i is contained in X0X_0, because a(wi)=w′a(w_i)=w'. It is interior over KK in XX by the assumed relative interior and the target absolute interior. It is also interior in X0X_0 by the finite relative interior of p0p_0 and composition. The local flat containment persists under restriction. Theorem 5.37.1 defines these weights; Theorem 5.37.2 identifies their values before and after restriction. The complete set of source moment values over w′w' is unchanged. Theorem 5.41.1, applied to the actual finite map p0p_0, now gives the displayed formula and the assertion involving Δ\Delta.

For independence of the box, intersect any two sufficiently small boxes. Their intersection is again a closed box with w′w' in its interior. The associated finite maps restrict to the same actual map on this intersection, by Theorem 2.36.9 and Proposition 2.33.19. The finite module over either target restricts to this same module, and the intersection contains the whole target moment fibre. The restriction assertion in Theorem 5.40.2 proves equality of the two middle terms. This proves the claimed independence. ◻

5.42 Cycles of finite analytic modules

Theorem 5.42.1. NAG-002370 (Module cycles and finite geometric pushforward). Let Y=M(A)Y=\mathcal M(A) be pure of analytic dimension dd, with a unit moment tuple fYf_Y and moment map tYt_Y. Let Ω⊂Rn′\Omega\subset\mathbb R^{n'} be open and convex, with Ω∩tY(∂KY)=∅\Omega\cap t_Y(\partial_KY)=\varnothing. Suppose a finite pure dd-dimensional integral affine polyhedral complex CY\mathcal C_Y has tY(Y)∩Ω=∣CY∣∩Ω.t_Y(Y)\cap\Omega=|\mathcal C_Y|\cap\Omega. For every finite AA-module MM, the values mM,fY(v)m_{M,f_Y}(v) are constant on relint⁡(σ)∩Ω\operatorname{relint}(\sigma)\cap\Omega for each dd-cell σ\sigma. They define a nonnegative integral balanced dd-cycle, denoted Trop⁡fY,Ω(M)\operatorname{Trop}_{f_Y,\Omega}(M). These cycles are additive in short exact sequences and independent of the chosen polyhedral subdivision. For M=AM=A this is the analytic cycle of Theorem 5.43.2.

Let p:X=M(B)→Yp:X=\mathcal M(B)\to Y be finite, with XX pure of analytic dimension dd. Suppose unit moment coordinates on XX and an integral map a:Zn→Zn′a:\mathbb Z^n\to\mathbb Z^{n'} satisfy the corresponding character identity tY∘p=a∘tXt_Y\circ p=a\circ t_X. Suppose also that a finite pure dd-dimensional integral affine complex CX\mathcal C_X satisfies tX(X)∩a−1(Ω)=∣CX∣∩a−1(Ω).t_X(X)\cap a^{-1}(\Omega) =|\mathcal C_X|\cap a^{-1}(\Omega). Then lattice-index pushforward on these open sets gives a∗Trop⁡fX,a−1(Ω)(B)=Trop⁡fY,Ω(B).\begin{equation} a_*\operatorname{Trop}_{f_X,a^{-1}(\Omega)}(B) =\operatorname{Trop}_{f_Y,\Omega}(B). \end{equation} On the left BB is the structure module of XX; on the right it is its finite AA-module. If Δ∈Q≥0\Delta\in\mathbb Q_{\geq0} satisfies βp(B)=Δ length⁡(Ap)(p∈Min⁡(A)),\beta_{\mathfrak p}(B) =\Delta\,\operatorname{length}(A_{\mathfrak p}) \qquad(\mathfrak p\in\operatorname{Min}(A)), then the right side is Δ Trop⁡fY,Ω(A)\Delta\,\operatorname{Trop}_{f_Y,\Omega}(A). The same statement allows a positive rational Δ\Delta for a nonreduced target.

Proof. Form the finite AA-algebra A⊕M,(a,m)(a′,m′)=(aa′,am′+a′m),A\oplus M,\qquad (a,m)(a',m')=(aa',am'+a'm), with its canonical Banach structure from Theorem 2.33.18. The ideal 0⊕M0\oplus M has square zero. Every multiplicative seminorm annihilates it, so projection and the coefficient inclusion identify its spectrum with that of AA. Its minimal primes are p⊕M\mathfrak p\oplus M, and the corresponding integral quotients are A/pA/\mathfrak p. Consequently YM=M(A⊕M)Y_M=\mathcal M(A\oplus M) is pure of analytic dimension dd and has the same moment image under the units (fY,j,0)(f_{Y,j},0).

Every point of YMY_M above a fibre over Ω\Omega is interior over KK: the map YM→YY_M\to Y is finite and hence relatively interior, and all points of YY with moment value in Ω\Omega are interior by hypothesis. Apply the composition inclusion 2.33.21. Theorem 5.43.2 therefore gives positive balanced analytic cycles for both YY and YMY_M on the same complex over Ω\Omega.

At an interior point vv of a dd-cell, let EQE_Q be the actual target Gauss algebra for a transverse projection.

For the domain comparison at vv, use fi=fY,if_i=f_{Y,i} and ri=exp⁡(−vi)r_i=\exp(-v_i) in the following exact-level calculation.

Let AA be a generalized affinoid over the fixed complete non-Archimedean field KK, let MM be finite over AA, and give the idealization AM=A⊕MA_M=A\oplus M its canonical finite-algebra Banach structure from Theorem 2.33.18. For units fi∈A×f_i\in A^\times and positive radii rir_i, put D=A{r−1T,rS}/(Ti−fi,Si−fi−1)i,DM=AM{r−1T,rS}/(Ti−(fi,0),Si−(fi−1,0))i.D=A\{r^{-1}T,rS\}/(T_i-f_i,S_i-f_i^{-1})_i,\qquad D_M=A_M\{r^{-1}T,rS\}/(T_i-(f_i,0),S_i-(f_i^{-1},0))_i . These use actual ideals and the displayed quotient norms. The rational point and geometric factorization clauses of Theorems 2.31.1 and 2.31.2, with denominator 11 and the simultaneous upper bounds ri,ri−1r_i,r_i^{-1}, identify the algebras with the whole exact-radius domains below and above, respectively. The lifted unit has inverse (fi−1,0)(f_i^{-1},0). Thus the inequalities upstairs describe exactly the inverse image of the whole domain. The same descriptions include the empty domain.

There is a canonical bounded isomorphism, with bounded inverse, DM≃D⊗^AAM.D_M\simeq D\widehat\otimes_A A_M . Indeed, for a commutative Banach KK-algebra CC with a bounded unital map ϕ:AM→C\phi:A_M\to C, the bounded factorization clause of Theorem 2.31.2 says that extending ϕ\phi across DMD_M requires exactly the estimates ∥ϕ((fi,0))m∥≤ci+rim,∥ϕ((fi−1,0))m∥≤ci−ri−m(m≥0)\|\phi((f_i,0))^m\|\leq c_i^+r_i^m,\qquad \|\phi((f_i^{-1},0))^m\|\leq c_i^-r_i^{-m}\quad(m\geq0) for some constants ci±c_i^\pm. The factorization is unique. These are precisely the estimates for extending ϕ∣A\phi|_A across DD. Thus DMD_M represents the pairs of bounded maps from AMA_M and DD agreeing on AA. Their balanced product on finite sums ∑jdj⊗bj\sum_j d_j\otimes b_j has norm bounded by the product of the two operator bounds times max⁡j∥dj∥∥bj∥\max_j\|d_j\|\|b_j\|. Taking the projective infimum and passing through separation and completion gives the unique bounded map from D⊗^AAMD\widehat\otimes_A A_M. The two universal factorizations are inverse because their composites fix both structural maps. This comparison preserves the actual algebras over AMA_M, hence also the maps from the exact-radius target torus defined by the common projecting monomials in the units.

Proposition 2.33.19 gives the Gauss algebra of YMY_M as (A⊕M)⊗AEQ=EQ⊕(M⊗AEQ).(A\oplus M)\otimes_AE_Q =E_Q\oplus(M\otimes_AE_Q). Thus its dimension is the sum of the two dimensions. Theorem 5.40.2 and the common lattice divisor give mYM(v)=mY(v)+mM,fY(v).m_{Y_M}(v)=m_Y(v)+m_{M,f_Y}(v). Subtracting the two balanced cycles proves constancy and balancing for the module weights. Their nonnegativity and integrality are (5.40.2). Additivity follows from the exact-sequence assertion of that theorem, and common-refinement invariance follows from the same property of the two analytic cycles. This proves the first part.

For the finite map pp, every point of XX whose moment value lies in a−1(Ω)a^{-1}(\Omega) is interior over KK, again by finite relative interior and composition. Thus the stated source analytic cycle is defined. Exclude from the target the lower-dimensional cells of CY\mathcal C_Y, the images of all cells of CX\mathcal C_X of dimension less than dd, and the images of those dd-cells on whose tangent space aa has rank less than dd. Linear images of the finitely many polyhedra are polyhedra, by the finite elimination and refinement construction in Theorem 5.17.2. Each excluded set has dimension at most d−1d-1. For d=0d=0 the nonempty cells are points and every restriction has rank zero. The exclusions therefore remove no point, so every target point in the support is already in the complement. For d>0d>0 the complement is dense in the relative interior of every target dd-cell: a line in its affine span with direction outside the finitely many excluded proper direction subspaces meets those affine subspaces in at most finitely many points. Such a line direction exists by the finite-polynomial argument in Theorem 5.43.1.

Fix a target point vv in this complement and in the target support. Every source moment value above it lies in the relative interior of a dd-cell on which aa is injective. There is at most one such point per cell, so the full inverse image in tX(X)t_X(X) is finite. If it is nonempty, all the hypotheses of Theorem 5.41.1 hold: the target and source germs are flat at cell interiors, and the target whole fibre is interior. The theorem identifies the lattice-index sum of source weights with mB,fY(v)m_{B,f_Y}(v).

If this finite inverse image is empty, the actual inverse image of the entire target moment fibre is empty. Its algebra is B⊗ADvB\otimes_AD_v, by finite base change and Theorem 2.36.9. Theorem 1.20.1 makes this algebra zero. Hence B⊗AEQ=0B\otimes_AE_Q=0 and mB,fY(v)=0m_{B,f_Y}(v)=0, which is also the empty lattice-index sum.

The map aa on the actual source support over Ω\Omega is proper. For a compact subset C⊂ΩC\subset\Omega, its inverse image is the closed subset tX(X)∩a−1(C)t_X(X)\cap a^{-1}(C) of the compact set tX(X)t_X(X). Refine the target along the finitely many cell images and pull the defining hyperplanes back to the source, as in Theorem 5.17.2. The lattice-index sum then defines the local pushforward; facets with smaller-dimensional image contribute zero. On the generic part of every resulting target dd-cell, this sum equals the module weight just computed. Both are constant there, so the cycles agree. This proves (5.42.1). The last assertion follows on every such cell from Theorem 5.40.2 and its formula for the structure module AA. ◻

Theorem 5.42.2. NAG-002372 (Finite-module pushforward and composition). Keep the finite morphism p:X=M(B)→Y=M(A)p:X=\mathcal M(B)\to Y=\mathcal M(A), the moment tuples, the lattice map aa, and the given complexes over Ω\Omega in Theorem 5.42.1. For every finite BB-module MM, regarded also as a finite AA-module, one has a∗Trop⁡fX,a−1(Ω)(M)=Trop⁡fY,Ω(M).a_*\operatorname{Trop}_{f_X,a^{-1}(\Omega)}(M) =\operatorname{Trop}_{f_Y,\Omega}(M). Its target generic coefficients are length⁡Ap(Mp)=∑q∈Min⁡(B)q∩A=plength⁡Bq(Mq)[κ(q):κ(p)](p∈Min⁡(A)).\operatorname{length}_{A_{\mathfrak p}}(M_{\mathfrak p}) =\sum_{\substack{\mathfrak q\in\operatorname{Min}(B)\\ \mathfrak q\cap A=\mathfrak p}} \operatorname{length}_{B_{\mathfrak q}}(M_{\mathfrak q}) [\kappa(\mathfrak q):\kappa(\mathfrak p)] \qquad(\mathfrak p\in\operatorname{Min}(A)).

These pushforwards respect finite composition. More precisely, let Z=M(C)→XZ=\mathcal M(C)\to X also be finite with ZZ pure of analytic dimension dd, and let a unit tuple fZf_Z and an integral map bb satisfy the corresponding character identity. Assume a finite pure integral complex presents the actual fZf_Z-image on (ab)−1(Ω)(ab)^{-1}(\Omega). For every finite CC-module LL, (ab)∗Trop⁡fZ,(ab)−1(Ω)(L)=a∗b∗Trop⁡fZ,(ab)−1(Ω)(L)=Trop⁡fY,Ω(L).(ab)_*\operatorname{Trop}_{f_Z,(ab)^{-1}(\Omega)}(L) =a_*b_*\operatorname{Trop}_{f_Z,(ab)^{-1}(\Omega)}(L) =\operatorname{Trop}_{f_Y,\Omega}(L). Here abab means a∘ba\circ b, and the middle pushforwards use the indicated inverse open sets.

If PP is a finite locally free AA-module of constant rank rr, then the tensor projection formula is a∗Trop⁡fX,a−1(Ω)(M⊗AP)=Trop⁡fY,Ω(M⊗AP)=r Trop⁡fY,Ω(M).a_*\operatorname{Trop}_{f_X,a^{-1}(\Omega)}(M\otimes_AP) =\operatorname{Trop}_{f_Y,\Omega}(M\otimes_AP) =r\,\operatorname{Trop}_{f_Y,\Omega}(M).

Proof. The square-zero algebra B⊕MB\oplus M is finite over BB and over AA, with the canonical finite-module norm, by Theorem 2.33.18. It has the same analytic points, moment image, and pure analytic dimension as BB, as proved in Theorem 5.42.1. Apply that theorem to its finite map to YY. Its source cycle is the sum of the structure cycle of XX and the module cycle of MM. Its target module cycle is the sum of the cycles of BB and MM as AA-modules. The theorem for B→AB\to A and subtraction give the asserted equality. Subtraction is valid on a common finite refinement because each lattice-index coefficient is a finite linear sum of source coefficients. The same compact actual support gives properness for the square-zero source, including modules with zero generic weight.

For a minimal prime p\mathfrak p of AA, decompose the finite Artinian algebra BpB_{\mathfrak p} into its local factors BqB_{\mathfrak q}, as in the proof of Theorem 5.41.1. The finite module MpM_{\mathfrak p} decomposes into the corresponding MqM_{\mathfrak q}. A composition series of each factor as a BqB_{\mathfrak q}-module has residue field κ(q)\kappa(\mathfrak q) as every simple quotient. Its length over ApA_{\mathfrak p} is the full field degree [κ(q):κ(p)][\kappa(\mathfrak q):\kappa(\mathfrak p)]. Adding these lengths proves the displayed coefficient formula.

The composite of two finite canonical affinoid maps is finite and canonical: finite module generators compose, and composing their finite free quotient presentations gives the same Banach topology. Apply the first assertion to C→BC\to B, then to B→AB\to A with LL as a finite BB-module, and finally to C→AC\to A. All source support above the stated open sets is interior by finite relative interior and Proposition 2.33.21. The given complexes therefore meet the same hypotheses at each step. The three equalities give the composition formula.

Finally, tensor PP with an actual Artinian Gauss algebra EQE_Q used to compute any of the target module weights. On a local factor EjE_j, the map from AA factors through the local ring at the contracted prime. A local free basis of PP of rank rr therefore gives P⊗AEj≃EjrP\otimes_AE_j\simeq E_j^r. Consequently dim⁡HQ((M⊗AP)⊗AEQ)=r dim⁡HQ(M⊗AEQ).\dim_{H_Q}\bigl((M\otimes_AP)\otimes_AE_Q\bigr) =r\,\dim_{H_Q}(M\otimes_AE_Q). Divide by the common lattice index and apply Theorem 5.40.2. This proves the target tensor equality on every generic cell. The first assertion, applied to the finite BB-module M⊗APM\otimes_AP, gives its pushforward equality. All these tensors have their canonical complete finite-module norms by Proposition 1.18.7. ◻

Theorem 5.42.3. NAG-002373 (Pushforward on a relative interior weight domain). Let p:X=M(B)→Y=M(A)p:X=\mathcal M(B)\to Y=\mathcal M(A) be an affinoid morphism, with XX and YY pure of analytic dimension dd. Let a unit tuple fYf_Y define tY:Y→Rn′t_Y:Y\to\mathbb R^{n'}. Write ∂(X/Y)=X∖Int⁡(X/Y)\partial(X/Y)=X\setminus\operatorname{Int}(X/Y) and put Up=Rn′∖(tY(∂KY)∪(tY∘p)(∂(X/Y))).U_p=\mathbb R^{n'}\setminus \bigl(t_Y(\partial_KY)\cup (t_Y\circ p)(\partial(X/Y))\bigr). This is open. There is a canonical local nonnegative integral balanced dd-cycle Cp,fY\mathcal C_{p,f_Y} on UpU_p, with support (tY∘p)(X)∩Up(t_Y\circ p)(X)\cap U_p. It is the local analytic cycle Cp,fY=Trop⁡p∗fY(X)∣Up.\mathcal C_{p,f_Y} =\operatorname{Trop}_{p^*f_Y}(X)|_{U_p}.

Every w∈Up∩tY(Y)w\in U_p\cap t_Y(Y) has a closed target moment box PP about ww, with ww in its interior, such that Y0=tY−1(P),X0=p−1(Y0),p0:X0⟶Y0Y_0=t_Y^{-1}(P),\qquad X_0=p^{-1}(Y_0),\qquad p_0:X_0\longrightarrow Y_0 is an actual finite canonical affinoid map. Writing A0=O(Y0)A_0=\mathcal O(Y_0) and B0=O(X0)B_0=\mathcal O(X_0), there is an open neighbourhood W⊂int⁡(P)W\subset\operatorname{int}(P) of ww on which Cp,fY∣W=Trop⁡fY∣Y0(B0)∣W.\mathcal C_{p,f_Y}|_W =\operatorname{Trop}_{f_Y|Y_0}(B_0)|_W. These finite-module descriptions agree on overlaps and under further target-box restriction. At a flat target germ v∈W∩tY(Y)v\in W\cap t_Y(Y) their weight is mCp,fY(v)=∑p∈Min⁡(A0)length⁡(A0)p((B0)p)m(Y0)p,fY(v),m_{\mathcal C_{p,f_Y}}(v) =\sum_{\mathfrak p\in\operatorname{Min}(A_0)} \operatorname{length}_{(A_0)_{\mathfrak p}} ((B_0)_{\mathfrak p}) m_{(Y_0)_{\mathfrak p},f_Y}(v), where an empty component fibre has weight zero.

Suppose a unit tuple fXf_X and an integral map a:Zn→Zn′a:\mathbb Z^n\to\mathbb Z^{n'} satisfy the character identity p∗fY=fXa∗p^*f_Y=f_X^{a^*}. All source weights in a−1(Up)∩tX(X)a^{-1}(U_p)\cap t_X(X) have whole interior fibres, and a∗(Trop⁡fX(X)∣a−1(Up))=Cp,fY.a_*\bigl(\operatorname{Trop}_{f_X}(X)|_{a^{-1}(U_p)}\bigr) =\mathcal C_{p,f_Y}. The pushforward uses the entire actual source support and is proper over UpU_p.

Proof. The absolute boundary of YY and the relative boundary of XX are closed, by Proposition 2.33.21, and compact, by Proposition 2.24.1. Their moment images are compact, so UpU_p is open. If (tY∘p)(x)∈Up(t_Y\circ p)(x)\in U_p, then xx is interior relative to YY and p(x)p(x) is interior over KK. The composition inclusion in Proposition 2.33.21 makes xx interior over KK. Consequently Up∩(tY∘p)(∂KX)=∅.U_p\cap(t_Y\circ p)(\partial_KX) =\varnothing. Theorem 5.45.3, applied on XX to the pullback unit tuple p∗fYp^*f_Y, defines the asserted local cycle and its support.

Fix w∈Up∩tY(Y)w\in U_p\cap t_Y(Y). The entire inverse image of its target moment fibre is relatively interior by definition of UpU_p. Theorem 5.41.3 therefore supplies the stated actual finite map, and permits PP to be chosen inside UpU_p. Choose an open box W⊂int⁡(P)W\subset\operatorname{int}(P) about ww. The whole target fibre at every v∈Wv\in W satisfies the box inequalities strictly. It is interior in Y0Y_0 relative to YY, by the actual rational presentation, and interior in YY over KK, by P⊂UpP\subset U_p. Composition makes it interior in Y0Y_0 over KK. Thus WW belongs to the whole interior weight domain of Y0Y_0. Theorem 5.35.3 gives purity of Y0Y_0 and of every nonempty X0X_0.

The finite-map assertion of Theorem 5.45.3, with the same moment tuple on source and target and the identity lattice map, now gives Trop⁡p0∗(fY∣Y0)(X0)∣W=Trop⁡fY∣Y0(B0)∣W.\operatorname{Trop}_{p_0^*(f_Y|Y_0)}(X_0)|_W =\operatorname{Trop}_{f_Y|Y_0}(B_0)|_W. The whole pullback moment fibre at every v∈Wv\in W is the same in X0X_0 and XX, and it is absolute interior in both. For X0X_0 this follows from finite relative interior and the target interior just checked. The flat-germ restriction assertion of Theorem 5.37.2 identifies their weights on every local generic cell. Hence the source cycle in this display is Cp,fY∣W\mathcal C_{p,f_Y}|_W. If X0X_0 is empty, its algebra and module cycle are zero by Theorem 1.20.1, and the same identity holds.

For completeness, the finite-module descriptions also compare directly. Intersect two boxes at a weight in the overlap of their open interiors. The intersection has that weight in its interior and retains the whole target fibre and its entire inverse image. The actual domain and finite tensor comparisons, Theorem 2.36.9 and Proposition 2.33.19, identify both restricted finite algebras with the algebra of this same inverse-image domain. The whole-fibre restriction assertion of Theorem 5.40.2 gives equal module weights on all generic cells of a common local refinement. This proves the overlap and restriction statements. The generic-length expression is the formula of that theorem for the actual finite A0A_0-module B0B_0.

Finally, the equality tY∘p=a∘tXt_Y\circ p=a\circ t_X and the first paragraph show that every whole fXf_X-fibre above UpU_p is absolute interior. Apply the finite-map theorem 5.45.3 to the identity map of XX, with source tuple fXf_X, target tuple p∗fYp^*f_Y, and lattice map aa, and restrict its equality to UpU_p. Its target module is the structure module of XX, so the result is exactly the asserted pushforward. For a compact subset C⊂UpC\subset U_p, the complete inverse image tX(X)∩a−1(C)t_X(X)\cap a^{-1}(C) is closed in the compact set tX(X)t_X(X). This proves the stated properness on the actual support. ◻

Theorem 5.42.4. NAG-002374 (The exceptional image of a generically free module). Let Y=M(A)Y=\mathcal M(A) be nonempty and pure of analytic dimension dd, and let I⊂AI\subset A be an ideal contained in no minimal prime. Put Z=M(A/I)Z=\mathcal M(A/I), with its actual closed-subspace structure. If d=0d=0, then ZZ is empty. If ZZ is nonempty, then dim⁡KZ≤d−1.\dim_K Z\leq d-1.

For any unit tuple ff defining t:Y→Rnt:Y\to\mathbb R^n, the entire actual image t(Z)t(Z) is the support of a finite integral affine polyhedral complex of dimension at most d−1d-1; for d=0d=0 it is empty. Put UY=Rn∖t(∂KY)U_Y=\mathbb R^n\setminus t(\partial_KY). The intersection t(Z)∩UYt(Z)\cap U_Y is relatively closed in t(Y)∩UYt(Y)\cap U_Y, and its complement is relatively open and dense there.

In particular, let MM be a finite AA-module which is free at every minimal prime, and let IMI_M define its nonfree locus as in Proposition 5.41.2. For every w∈(t(Y)∩UY)∖t(M(A/IM)),w\in\bigl(t(Y)\cap U_Y\bigr) \setminus t(\mathcal M(A/I_M)), there is a closed moment box PP about ww such that the whole domain YP=t−1(P)Y_P=t^{-1}(P) is in the analytic free locus and M⊗AO(YP)M\otimes_A\mathcal O(Y_P) is finite locally free. If AA is integral and the generic rank of MM is rr, this local rank is constantly rr on Spec⁡O(YP)\operatorname{Spec}\mathcal O(Y_P). These conclusions apply to every finite module over reduced AA. For a finite algebra M=BM=B, the corresponding actual finite map over YPY_P is finite locally free, with the same constant rank in the integral-target case.

Proof. The finite quotient A/IA/I has its actual generalized affinoid structure by Theorem 2.33.18. Choose a free-radius extension L=KsL=K_s making AA and A/IA/I strict, by Theorem 1.17.6. The quotient comparison of Theorem 1.18.8 gives (A/I)L=AL/IAL.(A/I)_L=A_L/IA_L. Lemma 5.35.1 identifies the minimal primes of ALA_L with pAL\mathfrak pA_L for p∈Min⁡(A)\mathfrak p\in\operatorname{Min}(A), and their contractions are p\mathfrak p. Thus IALIA_L is contained in no minimal prime of ALA_L. The analytic-dimension theorem 3.21.8 gives dim⁡AL=d\dim A_L=d, since ALA_L is strict.

Every chain of primes of AL/IALA_L/IA_L lifts to a chain q0⊊⋯⊊qℓ\mathfrak q_0\subsetneq\cdots\subsetneq\mathfrak q_\ell of primes of ALA_L containing IALIA_L. By [StacksAlg] (Tag 00E0(4)) applied to the zero ideal in ALA_L and the prime q0\mathfrak q_0, there is a minimal prime pL⊆q0\mathfrak p_L\subseteq\mathfrak q_0 of ALA_L. The inclusion pL⊂q0\mathfrak p_L\subset\mathfrak q_0 is strict, because IALIA_L is not contained in pL\mathfrak p_L. Prepending pL\mathfrak p_L therefore gives a chain of length ℓ+1\ell+1 in ALA_L. Hence ℓ+1≤d\ell+1\leq d. Put R=AL/IALR=A_L/IA_L. If R≠0R\ne0, then [StacksAlg] (Tag 00E0(2)) gives a maximal ideal m\mathfrak m of RR. The quotient R/mR/\mathfrak m is a field, so m\mathfrak m is prime. There is thus a chain of length zero in this case, and the bound on all chains gives dim⁡R≤d−1\dim R\leq d-1. The quotient is zero when d=0d=0. The faithful quotient comparison detects zero before extension, and Theorem 3.21.8 identifies the dimension of every nonzero quotient with that of M(A/I)\mathcal M(A/I) over the original KK.

If ZZ is empty, its image has the empty complex. Otherwise apply Theorem 5.39.3 to ZZ and the restrictions of the original units. The dimension inequality just proved gives the asserted finite complex of dimension at most d−1d-1 for the entire actual image. Compactness of ZZ and its image, Proposition 2.24.1, proves relative closedness on UYU_Y.

At every point of t(Y)∩UYt(Y)\cap U_Y, Theorem 5.38.6 presents the ambient actual image by a finite pure dd-dimensional fan. The exceptional complex is contained in a finite union of affine subspaces of dimension at most d−1d-1. No such finite union contains a relatively open part of a dd-cell. For example, choose a line direction in its tangent plane outside the finitely many proper direction subspaces; the finite-polynomial construction in Theorem 5.43.1 supplies one. Every short line segment in that direction meets each excluded affine subspace at most once. Since a pure finite complex is the closure of the union of its maximal-cell interiors, this proves density of the complement at every point. Openness follows from the relative closedness already proved. For d=0d=0 the exceptional set is empty and the assertion is immediate.

For the module MM, Proposition 5.41.2 makes IMI_M an ideal avoiding every minimal prime, and identifies its analytic zero set with the nonfree locus. Its image under tt is compact. Choose a closed box PP about the given ww, with ww in its interior, contained in UYU_Y and disjoint from that compact image. The actual domain YPY_P is given by Theorem 2.31.2. It lies entirely in the analytic free locus, so the analytic restriction assertion of Proposition 5.41.2 makes the restricted module finite locally free.

If AA is integral, a free basis of MpM_{\mathfrak p} at any prime of the free locus becomes a basis over Frac⁡(A)\operatorname{Frac}(A) after further localization. Its rank is therefore the same integer rr at every such prime. The principal free bases in the proof of that proposition consequently all have rank rr. They base change to the indicated cover of Spec⁡O(YP)\operatorname{Spec}\mathcal O(Y_P), proving the constant-rank assertion. Reduced AA has the required generic freeness by the same proposition. For a finite algebra BB, finite base change and its canonical complete norm, Proposition 2.33.19, identify the restricted algebra with B⊗AO(YP)B\otimes_A\mathcal O(Y_P). The module assertion is exactly finite local freeness of this actual morphism. ◻

5.43 Balancing from finite analytic degrees

Fix a complete real-valued non-Archimedean field KK, with its specified absolute value. The valuation may be trivial. Let X=M(B)X=\mathcal M(B) be a generalized KK-affinoid space, pure of analytic dimension dd, and let f1,…,fn∈B×f_1,\ldots,f_n\in B^\times. Write t:X⟶NR,t(x)=(−log⁡∣f1(x)∣,…,−log⁡∣fn(x)∣),N=Zn.t:X\longrightarrow N_{\mathbb R},\qquad t(x)=(-\log|f_1(x)|,\ldots,-\log|f_n(x)|),\qquad N=\mathbb Z^n. At a whole interior flat dd-dimensional germ, use the positive integer mX,fm_{X,f} of Theorem 5.37.1. All tangent lattices are saturated in the ambient lattice.

Theorem 5.43.1. NAG-002376 (The normal equation at a polyhedral wall). Suppose d≥1d\geq1, w∈t(X)w\in t(X), and t−1(w)⊂Int⁡(X/K)t^{-1}(w)\subset\operatorname{Int}(X/K). Let Nτ⊂NN_\tau\subset N be saturated of rank d−1d-1. For 1≤i≤r1\leq i\leq r, let Ni⊂NN_i\subset N be saturated of rank dd and contain NτN_\tau, and let ui∈Niu_i\in N_i lift a primitive generator of Ni/NτN_i/N_\tau. Suppose the half-planes Ci=(Nτ)R+R≥0uiC_i=(N_\tau)_{\mathbb R}+\mathbb R_{\geq0}u_i are distinct, and that on an open neighborhood WW of ww the actual moment image is t(X)∩W=(w+⋃i=1rCi)∩W.t(X)\cap W=\left(w+\bigcup_{i=1}^r C_i\right)\cap W. For each ii, the weight mX,fm_{X,f} is a fixed positive integer mim_i on the relative interior of w+Ciw+C_i sufficiently near ww. These integers satisfy ∑i=1rmiui=0in N/Nτ.\sum_{i=1}^r m_i u_i=0 \quad\text{in }N/N_\tau.

Proof. The boundary of XX is compact, by Propositions 2.33.21 and 2.24.1. Its moment image is compact and misses ww. Shrink WW to avoid that image. Away from the common wall, distinct half-planes have disjoint relative interiors. Near such a point the image is an open part of its affine dd-plane, and its whole fibre is interior. Theorems 5.37.1 and 5.37.2 give a positive locally constant integer there. Intersecting the relative interior with a sufficiently small open box about ww gives a convex set. A locally constant function is constant on each segment in that set: otherwise the supremum of the initial interval of constancy contradicts local constancy. This gives the integers mim_i.

By Lemma 5.14.1, choose a homomorphism q:N→Zd−1q:N\to\mathbb Z^{d-1} whose restriction to NτN_\tau is an isomorphism. Subtracting from each uiu_i its unique element of NτN_\tau with the same qq-image gives representatives satisfying q(ui)=0q(u_i)=0, without changing the half-planes or the asserted normal equation. Let λ:N→Z\lambda:N\to\mathbb Z vanish on NτN_\tau and satisfy λ(ui)≠0\lambda(u_i)\ne0 for every ii. Put Q=(q,λ):N⟶Zd.Q=(q,\lambda):N\longrightarrow\mathbb Z^d. Since Ni/NτN_i/N_\tau is generated by the class of uiu_i, the chosen basis of NτN_\tau, followed by uiu_i, is a basis of NiN_i. In these bases Q∣NiQ|_{N_i} has matrix diag⁡(Id−1,λ(ui))\operatorname{diag}(I_{d-1},\lambda(u_i)). Consequently Q∣(Ni)R is injective,[Zd:Q(Ni)]=∣λ(ui)∣.\begin{equation} Q|_{(N_i)_{\mathbb R}}\text{ is injective},\qquad [\mathbb Z^d:Q(N_i)]=|\lambda(u_i)|. \end{equation}

Take a small closed rectangular box P⊂WP\subset W with ww in its interior, and let Y=t−1(P)Y=t^{-1}(P) be the actual affinoid domain given by Theorem 2.31.2. It is pure of analytic dimension dd by Theorem 5.35.3. At each point of t−1(w)t^{-1}(w) all inequalities defining YY are strict. Thus that fibre is interior in YY over XX, and composition in Proposition 2.33.21 makes it interior in YY over KK.

The whole level of Q∘t∣YQ\circ t|_Y at s0=Qws_0=Qw is t−1(w)t^{-1}(w). Indeed, for v−w=b+aui∈Civ-w=b+a u_i\in C_i, with b∈(Nτ)Rb\in(N_\tau)_{\mathbb R} and a≥0a\geq0, the equation Q(v−w)=0Q(v-w)=0 first gives aλ(ui)=0a\lambda(u_i)=0 and then q(b)=0q(b)=0. Thus a=b=0a=b=0. Theorem 5.35.5, applied to the monomial units with exponent matrix QQ on YY, gives a closed box P′P' about s0s_0 and an actual finite map Z=(Q∘t∣Y)−1(P′)⟶TP′.Z=(Q\circ t|_Y)^{-1}(P')\longrightarrow\mathbb T_{P'}. Its Gauss-fibre dimension is a fixed positive integer δ\delta on P′P'. Every point of ZZ is interior in YY over KK.

For a sufficiently small ε>0\varepsilon>0, both s±=s0±εeds_\pm=s_0\pm\varepsilon e_d lie in the interior of P′P'. The moment values in YY over s+s_+ are exactly vi+=w+ελ(ui)ui,λ(ui)>0,v_i^+=w+\frac{\varepsilon}{\lambda(u_i)}u_i, \qquad\lambda(u_i)>0, and those over s−s_- are exactly vi−=w−ελ(ui)ui,λ(ui)<0.v_i^-=w-\frac{\varepsilon}{\lambda(u_i)}u_i, \qquad\lambda(u_i)<0. To see this, the first d−1d-1 coordinate equations give b=0b=0, and the last gives the displayed coefficient of uiu_i. Choose ε\varepsilon so that all these finitely many points belong to the interior of PP; the local image hypothesis then also proves their existence. Distinct half-planes give distinct points on each side.

Every whole moment fibre at these values in XX belongs to YY. The projected levels avoid the absolute boundary of YY, since they belong to Z⊂Int⁡(Y/K)Z\subset\operatorname{Int}(Y/K). Proposition 5.31.2 therefore decomposes each actual level algebra into the product of its whole moment fibre algebras, before reduction. Theorem 5.35.5 identifies these level algebras with the exact-radius restrictions of the same finite algebra over TP′\mathbb T_{P'}. Their Gauss fields are the same weighted completions of the target Laurent rational function field. Consequently their Gauss dimensions are both δ\delta, and the product decompositions give δ=∑λ(ui)>0δQ(vi+),δ=∑λ(ui)<0δQ(vi−).\delta=\sum_{\lambda(u_i)>0}\delta_Q(v_i^+),\qquad \delta=\sum_{\lambda(u_i)<0}\delta_Q(v_i^-). By Theorem 5.37.1 and (5.43.1), these identities read ∑λ(ui)>0miλ(ui)=∑λ(ui)<0mi(−λ(ui)).\sum_{\lambda(u_i)>0}m_i\lambda(u_i) =\sum_{\lambda(u_i)<0}m_i(-\lambda(u_i)). Hence λ(∑imiui)=0\lambda(\sum_i m_i u_i)=0 for every such λ\lambda.

These integral covectors separate points of the free quotient N/NτN/N_\tau. Explicitly, in a basis of that quotient use covectors with coordinates (1,a,a2,…,ah−1)(1,a,a^2,\ldots,a^{h-1}), where h=rank⁡(N/Nτ)h=\operatorname{rank}(N/N_\tau) and a∈Za\in\mathbb Z. Their values on any fixed nonzero vector are a nonzero polynomial in aa, with finitely many roots. If ∑imiui\sum_i m_i u_i had nonzero class, choose an integer aa avoiding the roots for that class and for all the nonzero classes of the uiu_i. The resulting covector would contradict the preceding equality. Thus the normal equation holds in N/NτN/N_\tau. ◻

Theorem 5.43.2. NAG-002378 (The analytic weights on a given polyhedral image). Let Ω⊂NR\Omega\subset N_{\mathbb R} be open and convex, with Ω∩t(∂KX)=∅\Omega\cap t(\partial_KX)=\varnothing. Suppose a finite pure dd-dimensional integral affine polyhedral complex C\mathcal C satisfies t(X)∩Ω=∣C∣∩Ω.t(X)\cap\Omega=|\mathcal C|\cap\Omega. Discard cells not meeting Ω\Omega. Each dd-cell σ\sigma has a well-defined positive integer mσm_\sigma: it is mX,f(v)m_{X,f}(v) for any v∈relint⁡(σ)∩Ωv\in\operatorname{relint}(\sigma)\cap\Omega. For every (d−1)(d-1)-cell τ\tau meeting Ω\Omega these integers obey ∑σ⊃τmσuσ/τ=0in N/Nτ.\sum_{\sigma\supset\tau}m_\sigma u_{\sigma/\tau}=0 \quad\text{in }N/N_\tau. Thus they define a positive integral balanced dd-cycle on Ω\Omega. It is unchanged by a common polyhedral refinement or by an affinoid restriction containing all moment fibres over Ω\Omega and satisfying the same hypotheses.

Proof. At an interior point of a dd-cell, the finiteness and face intersection property of the complex give a neighborhood in which its support is an open part of that cell’s affine span. The whole fibre is interior by the condition on Ω\Omega. Theorems 5.37.1 and 5.37.2 therefore give a positive locally constant integer on relint⁡(σ)∩Ω\operatorname{relint}(\sigma)\cap\Omega. This intersection is convex, so the segment argument in the preceding proof makes the integer constant.

Choose w∈relint⁡(τ)∩Ωw\in\operatorname{relint}(\tau)\cap\Omega. Such a point exists whenever τ\tau meets the open set Ω\Omega: points of its relative interior approach every point of the polyhedron. Only cells containing τ\tau occur sufficiently near ww. In a finite inequality presentation of an incident dd-cell, inequalities strict at ww remain strict in a small neighborhood; the remaining inequality in its affine span cuts out the half-space pointing into that cell from τ\tau. Its direction is (Nτ)R+R≥0uσ/τ(N_\tau)_{\mathbb R}+\mathbb R_{\geq0}u_{\sigma/\tau}. Distinct incident cells give distinct half-spaces, since otherwise their relative interiors would intersect. Theorem 5.43.1 now gives the displayed equation. For d=0d=0 there is no normal equation to check.

On a subdivision, the generic interior of each new dd-cell has the same whole fibres and tangent lattice as its old cell, so its weight is the old weight. Lemma 5.14.2 identifies the resulting cycles under common refinement. For an affinoid restriction as in the statement, Theorem 5.37.2 identifies the actual fibre algebras, tangent lattices and weights. The same argument on a common refinement proves the asserted equality of cycles. ◻

5.44 Adding moment coordinates

Theorem 5.44.1. NAG-002380 (Pushforward after adding analytic unit coordinates). In the setting above let g1,…,gm∈B×g_1,\ldots,g_m\in B^\times and put t^=(t,(−log⁡∣gj∣)j):X⟶Rn+m,π:Zn+m⟶Zn.\widehat t=(t,(-\log|g_j|)_j):X\longrightarrow\mathbb R^{n+m}, \qquad \pi:\mathbb Z^{n+m}\longrightarrow\mathbb Z^n. Let Ω⊂Rn\Omega\subset\mathbb R^n be open and convex and avoid t(∂KX)t(\partial_KX). Suppose finite pure dd-dimensional integral affine complexes C\mathcal C and C^\widehat{\mathcal C} satisfy t(X)∩Ω=∣C∣∩Ω,t^(X)∩π−1(Ω)=∣C^∣∩π−1(Ω).\begin{aligned} t(X)\cap\Omega&=|\mathcal C|\cap\Omega,\\ \widehat t(X)\cap\pi^{-1}(\Omega) &=|\widehat{\mathcal C}|\cap\pi^{-1}(\Omega). \end{aligned} Equip both with their analytic weights from Theorem 5.43.2. Then on Ω\Omega π∗[C^]=[C].\pi_*[\widehat{\mathcal C}]=[\mathcal C]. More explicitly, outside a finite union of polyhedra of dimension at most d−1d-1, each v∈∣C∣∩Ωv\in|\mathcal C|\cap\Omega lies in the relative interior of a dd-cell σ\sigma, its full inverse image in t^(X)\widehat t(X) is finite, and mσ=∑u∈t^(X)π(u)=vmσ^(u)[Nσ:π(Nσ^(u))].\begin{equation} m_\sigma= \sum_{\substack{u\in\widehat t(X)\\\pi(u)=v}} m_{\widehat\sigma(u)} [N_\sigma:\pi(N_{\widehat\sigma(u)})]. \end{equation} Here uu lies in the relative interior of a unique dd-cell σ^(u)\widehat\sigma(u), and π\pi is injective on its tangent space. The equality uses all source moment fibres over vv.

Proof. The set π−1(Ω)\pi^{-1}(\Omega) is open and convex and misses t^(∂KX)\widehat t(\partial_KX). Thus the indicated weights exist on both complexes and are balanced by Theorem 5.43.2.

Remove from Ω\Omega the cells of C\mathcal C of dimension less than dd, the images of all cells of C^\widehat{\mathcal C} of dimension less than dd, and the images of those dd-cells on whose tangent space π\pi has rank less than dd. Each removed set is contained in a finite union of affine subspaces of dimension at most d−1d-1. The images are polyhedra: eliminating one real coordinate from a finite list of linear inequalities replaces the upper and lower bounds for that coordinate by their pairwise comparisons, and preserves the inequalities not involving it; iterating gives a finite inequality presentation of a linear image. Rational coefficients remain rational. This also supplies a finite polyhedral refinement adapted to all these images, by cutting along the finitely many resulting supporting hyperplanes.

The complement is dense in the relative interior of every dd-cell meeting Ω\Omega. To verify this directly, in its affine span choose a line direction outside the finitely many proper linear subspaces parallel to the excluded affine subspaces. A line in that direction meets each of them at at most one point. Arbitrarily short segments therefore contain points of the complement. Such a direction exists, for example by the finite-polynomial argument in the proof of Theorem 5.43.1.

Fix vv in this complement and in the target support. Every u∈t^(X)u\in\widehat t(X) above vv is in π−1(Ω)\pi^{-1}(\Omega), so it lies in the relative interior of a unique dd-cell of C^\widehat{\mathcal C}, and π\pi has full rank on that cell. The affine restriction is injective, so there is at most one such point for each of the finitely many cells. There is at least one, since t=π∘t^t=\pi\circ\widehat t. This proves finiteness and nonemptiness of the full inverse image.

Near vv the target image is contained in its affine dd-plane, and near every uu the joint image is contained in its affine dd-plane. The whole target fibre is interior because v∉t(∂KX)v\notin t(\partial_KX); the whole source fibres are contained in it. Apply Theorem 5.37.3 to the identity morphism X→XX\to X, with source moment tuple (f,g)(f,g), target moment tuple ff, and lattice map π\pi. Its algebra is finite free of rank one over itself, with its original Banach structure. All its full-fibre hypotheses have just been checked. The resulting formula is exactly (5.44.1).

The restriction of π\pi to the actual joint support over Ω\Omega is proper: above a compact subset of Ω\Omega its inverse image is a closed subset of the compact set t^(X)\widehat t(X). On the finite refinement constructed above, the weight defining π∗\pi_* at a generic point of a target dd-cell is the finite sum in (5.44.1); cells with smaller-dimensional image contribute no dd-weight. The formula agrees with the analytic target weight on a dense subset of every such cell. Both are constant there, so the weighted cycles are equal. ◻

5.45 Local analytic tropical cycles

Lemma 5.45.1. NAG-002383 (A finite presentation near a compact projected fibre). Let S⊂RnS\subset\mathbb R^n be compact, let a:Rn→Rma:\mathbb R^n\to\mathbb R^m be an integral linear map, and let w∈Rmw\in\mathbb R^m. Suppose that, at each u∈S∩a−1(w)u\in S\cap a^{-1}(w), there are a neighbourhood VuV_u and a finite pure dd-dimensional rational polyhedral fan CuC_u such that S∩Vu=(u+∣Cu∣)∩Vu.S\cap V_u=(u+|C_u|)\cap V_u. There are an open convex neighbourhood WW of ww and a finite pure dd-dimensional integral affine polyhedral complex D\mathcal D such that ∣D∣⊂S,S∩a−1(W)=∣D∣∩a−1(W).|\mathcal D|\subset S,\qquad S\cap a^{-1}(W)=|\mathcal D|\cap a^{-1}(W). If the fibre is empty, D\mathcal D may be empty. The neighbourhood WW may be required to lie in any prescribed neighbourhood of ww.

Proof. For each uu in the compact fibre, take a closed rectangular box PuP_u centred at uu, contained in VuV_u, and with uu in its interior. The interiors of finitely many such boxes Pu1,…,PusP_{u_1},\ldots,P_{u_s} cover the fibre. Put T=⋃j=1s((uj+∣Cuj∣)∩Puj)=⋃j=1s(S∩Puj).T=\bigcup_{j=1}^s\bigl((u_j+|C_{u_j}|)\cap P_{u_j}\bigr) =\bigcup_{j=1}^s(S\cap P_{u_j}). For every dd-cone σ\sigma of CujC_{u_j}, the polytope (uj+σ)∩Puj(u_j+\sigma)\cap P_{u_j} has dimension dd: the box contains a neighbourhood of its cone vertex. Every smaller-dimensional cone is a face of a dd-cone. Thus TT is a finite union of dd-dimensional integral affine polytopes and their faces. Cut these polytopes by all the finitely many hyperplanes in their inequality and affine-hull presentations. Their cells and faces form a finite common refinement D\mathcal D with support TT. Each cell is contained in the closure of a dd-cell. Indeed, in any original dd-polytope its relative interior is dense, and the finitely many proper cutting hyperplanes have empty relative interior. Hence this complex is pure of dimension dd. This is also the explicit refinement construction of Lemma 5.14.2.

The set E=S∖⋃j=1sint⁡(Puj)E=S\setminus\bigcup_{j=1}^s\operatorname{int}(P_{u_j}) is compact and has no point over ww. Its image a(E)a(E) is compact and misses ww. Choose a sufficiently small open convex neighbourhood WW of ww disjoint from that image. Every point of S∩a−1(W)S\cap a^{-1}(W) belongs to one of the box interiors, hence to TT. Conversely T⊂ST\subset S, proving both asserted inclusions. If the fibre is empty, apply the same compactness argument to E=SE=S and take the empty complex. Shrinking WW proves the last assertion. ◻

Definition 5.45.2. (A local weighted polyhedral cycle). A local integral dd-cycle on an open subset UU of a real lattice space is a collection of finite weighted integral affine polyhedral complexes on an open cover of UU, with integer weights on their dd-cells, which agree after common refinement on every overlap. It is balanced if each local complex satisfies the primitive-normal balancing equations. Complexes with all weights zero represent the zero cycle. Lattice-index pushforward is computed on such local presentations when the map on their support is proper and the required source presentations are finite near each target point.

Theorem 5.45.3. NAG-002385 (Analytic cycles on whole interior weight domains). Let Y=M(A)Y=\mathcal M(A) be a generalized affinoid space, pure of analytic dimension dd, over any complete non-Archimedean field with its specified absolute value. Let fYf_Y be a finite tuple of analytic units, let tY:Y→Rn′t_Y:Y\to\mathbb R^{n'} be its moment map, and put UY=Rn′∖tY(∂KY).U_Y=\mathbb R^{n'}\setminus t_Y(\partial_KY). There is a unique local positive integral balanced dd-cycle Trop⁡fY(Y)∣UY\operatorname{Trop}_{f_Y}(Y)|_{U_Y} whose support is tY(Y)∩UYt_Y(Y)\cap U_Y and whose weights at flat dd-dimensional germs are the actual Gauss-degree weights. For every finite AA-module MM there is a local nonnegative integral balanced cycle Trop⁡fY(M)∣UY\operatorname{Trop}_{f_Y}(M)|_{U_Y} with the module weights of Theorem 5.40.2. It is additive in short exact sequences and is unchanged by common subdivision. The structure module gives the preceding positive cycle.

Let p:X=M(B)→Yp:X=\mathcal M(B)\to Y be finite, with XX pure of analytic dimension dd. Let fXf_X be a unit tuple and let a:Zn→Zn′a:\mathbb Z^n\to\mathbb Z^{n'} satisfy the actual character identities p∗fY,j=∏ifX,iajip^*f_{Y,j}=\prod_i f_{X,i}^{a_{ji}}, so that tY∘p=a∘tXt_Y\circ p=a\circ t_X. The entire source support over UYU_Y lies in its whole interior weight domain, and a∗(Trop⁡fX(X)∣a−1(UY))=Trop⁡fY(B)∣UY.\begin{equation} a_*\bigl(\operatorname{Trop}_{f_X}(X)|_{a^{-1}(U_Y)}\bigr) =\operatorname{Trop}_{f_Y}(B)|_{U_Y}. \end{equation} In particular, adjoining any finite list of analytic unit coordinates to a moment tuple and projecting back to that tuple preserves its local analytic cycle.

Proof. The set UYU_Y is open: the absolute boundary is closed in the compact affinoid spectrum by Propositions 2.33.21 and 2.24.1. At a point of tY(Y)∩UYt_Y(Y)\cap U_Y, Theorem 5.38.6 gives a translated finite pure rational fan presenting the actual image on a sufficiently small open convex neighbourhood W⊂UYW\subset U_Y. Theorem 5.43.2 gives its positive integral balanced weights. At a point of UY∖tY(Y)U_Y\setminus t_Y(Y) take a neighbourhood disjoint from the compact image and use the zero cycle.

These local presentations agree on overlaps. Cut along the finitely many defining hyperplanes of two presentations. On the relative interior of every resulting dd-cell meeting the overlap, the actual image has the same flat germ and saturated tangent lattice. Both weights are therefore mY,fYm_{Y,f_Y} of that same whole fibre. The common refinement has equal weights, which is the required agreement. This also proves uniqueness. Theorem 5.42.1 on each of these neighbourhoods gives the module cycles. Their weights agree on overlaps by Theorem 5.40.2; its exact-sequence formula proves the stated additivity.

For the finite morphism, any point xx with a(tX(x))∈UYa(t_X(x))\in U_Y maps to an absolute interior point of YY. Finite morphisms are everywhere relatively inner by Theorem 2.33.18. Composition in Proposition 2.33.21 makes xx an absolute interior point of XX. Thus a−1(UY)∩tX(∂KX)=∅.a^{-1}(U_Y)\cap t_X(\partial_KX)=\varnothing. Every whole source moment fibre above this set is interior. Theorem 5.38.6 therefore applies at each of its weights.

Fix w∈UYw\in U_Y. Apply Lemma 5.45.1 to the compact set S=tX(X)S=t_X(X) and its complete inverse image over ww. It supplies a finite pure source complex on a−1(W)a^{-1}(W) for some open convex neighbourhood W⊂UYW\subset U_Y of ww. If w∈tY(Y)w\in t_Y(Y), shrink WW inside one target fan neighbourhood already constructed. The two finite complexes then satisfy exactly the hypotheses of Theorem 5.42.1. Its equality proves (5.45.1) over WW. If the source inverse image over ww is empty, the compactness part of the lemma allows WW to miss a(tX(X))a(t_X(X)); the empty-fibre argument in that same theorem makes the module cycle zero there. If w∉tY(Y)w\notin t_Y(Y), both sides are zero on a smaller neighbourhood disjoint from tY(Y)t_Y(Y).

The map on the actual source support is proper over UYU_Y: the inverse image of a compact subset of UYU_Y is closed in the compact set tX(X)t_X(X). These finite local presentations therefore define the lattice-index pushforward, and their equalities glue to the displayed identity. Finally take p=id⁡Yp=\operatorname{id}_Y, B=AB=A, and the coordinate projection from the enlarged moment tuple. Its finite module is the structure module, giving the last assertion. ◻

5.46 Integration on analytic moment cycles

Definition 5.46.1. NAG-002388 (Lattice normalization for real superforms). On an open subset of NRN_{\mathbb R}, a real superform of bidegree (p,q)(p,q) is a smooth function with values in ⋀pNR′∗⊗⋀qNR′′∗\bigwedge^pN_{\mathbb R}^{\prime *}\otimes \bigwedge^qN_{\mathbb R}^{\prime\prime *}. All one-form generators anticommute. The operators d′d' and d′′d'' differentiate coefficients and wedge the corresponding primed or double-primed differential on the left. The involution JJ exchanges the two copies and fixes functions. In particular d′′=Jd′Jd''=Jd'J, as is checked on a coefficient and on the constant one-form generators.

For an affine dd-plane with tangent lattice MM, choose an integral basis and its affine coordinates x1,…,xdx_1,\ldots,x_d. Use the positive top generator ωM=d′x1∧d′′x1∧⋯∧d′xd∧d′′xd.\omega_M=d'x_1\wedge d''x_1\wedge\cdots\wedge d'x_d\wedge d''x_d. The lattice-normalized integral of cωMc\omega_M on a polyhedron is the ordinary real integral of cc in these coordinates. A form is restricted to a polyhedron by restricting both cotangent factors to its affine span. Its support on a weighted complex is the closure, in that complex, of the points in the relative interiors of its maximal cells where this restriction is nonzero, after discarding cells of weight zero. Compact support below refers to this support.

Lemma 5.46.2. NAG-002389 (The real polyhedral integration calculation). Lattice-normalized integration of smooth top superforms with compact support is well-defined and additive under finite polyhedral subdivision. Let F:σ→τF:\sigma\to\tau be an affine isomorphism of dd-dimensional polyhedra whose linear part maps the tangent lattice NσN_\sigma into NτN_\tau. If that lattice map is injective, then ∫σF∗α=[Nτ:F(Nσ)]∫τα.\int_\sigma F^*\alpha =[N_\tau:F(N_\sigma)]\int_\tau\alpha. A linear map of rank less than dd has zero pullback in bidegree (d,d)(d,d).

Suppose d≥1d\geq1, and let β\beta have bidegree (d−1,d)(d-1,d) near a dd-polyhedron σ\sigma, with compact tangential support on σ\sigma. For each facet τ\tau choose a basis v1,…,vd−1v_1,\ldots,v_{d-1} of NτN_\tau and a primitive lift uσ/τ∈Nσu_{\sigma/\tau}\in N_\sigma pointing into σ\sigma. Write dμτd\mu_\tau for the real measure giving this basis covolume one. Then ∫σd′β=(−1)d(d+1)/2∑τ≺σ∫τβ(v1′,…,vd−1′;v1′′,…,vd−1′′,uσ/τ′′) dμτ.\begin{equation} \int_\sigma d'\beta =(-1)^{d(d+1)/2}\sum_{\tau\prec\sigma} \int_\tau \beta(v_1',\ldots,v_{d-1}'; v_1'',\ldots,v_{d-1}'',u_{\sigma/\tau}'') \,d\mu_\tau. \end{equation} The facet expression is independent of the chosen basis and of the lift modulo NτN_\tau.

Proof. We first give the elementary real integration facts used in the calculation. A closed bounded box in Rr\mathbb R^r is sequentially compact: repeatedly bisect its sides, retain a closed subbox containing infinitely many terms of a sequence, and choose terms with increasing indices in the nested boxes. Their diameters tend to zero, so completeness of the real coordinates gives a convergent subsequence in the original box. A continuous function on a compact subset is bounded and uniformly continuous. Otherwise a sequence witnessing failure of either assertion has a convergent subsequence; in the second case take pairs at distance tending to zero. Continuity at the common limit is a contradiction.

Let PP be a bounded polytope in Rr\mathbb R^r, and let cc be continuous on a neighbourhood of PP. In a fixed bounding box use a cubical grid of side length hh. The grid boxes meeting a given affine hyperplane have total volume at most ChC h. To see this, choose a nonzero coordinate coefficient in an equation for the hyperplane. For each box in the other r−1r-1 coordinates, the possible remaining coordinate ranges over an interval of length at most a constant times hh, so only boundedly many of the remaining grid intervals are met. There are at most C′h−(r−1)C' h^{-(r-1)} such boxes in the fixed bounding region. The boundary of PP is contained in finitely many such hyperplanes. Its grid boxes therefore have total volume tending to zero.

The upper and lower grid sums for cc on the boxes wholly inside PP differ by at most the bounding volume times the modulus of continuity of cc at distance rh\sqrt r h. The boundary boxes add an error bounded by Chsup⁡∣c∣C h\sup|c|. Both errors tend to zero. Refining two grids gives the same limit, which defines the integral. The same bounds allow signed functions and prove additivity under a finite polyhedral dissection: its additional faces lie in finitely many hyperplanes and contribute zero. They also prove that lower-dimensional polyhedra have zero rr-dimensional integral. For unbounded polyhedra, intersect with a box containing the compact support in its interior; a larger box gives the same value.

We will use iterated integration in these coordinates. The same construction proves it for the functions just considered. On each line parallel to a coordinate axis, the characteristic function of a polytope times a continuous function is piecewise continuous on finitely many intervals. The upper and lower step functions of a rectangular grid bound its one-dimensional integrals. Integrating these step functions in the remaining variables gives exactly their rectangular grid sums. Their difference tends to zero by the preceding estimates. Thus both orders of integration have the same limit. This argument also proves existence of the required outer integrals; it does not assume an interchange theorem.

One-dimensional translation and scaling follow directly by translating or scaling interval subdivisions. Interchanging coordinates follows from the equality of iterated integrals. A shear xi↦xi+cxjx_i\mapsto x_i+c x_j preserves the integral, because after fixing all other coordinates it is a translation in xix_i. Every invertible real matrix is a product of shears, coordinate interchanges and nonzero diagonal scalings, by Gaussian elimination. Hence ordinary real change of variables for an invertible affine map with linear matrix LL has factor ∣det⁡L∣|\det L|. All the intermediate regions are polyhedra, so the preceding integral constructions apply at every step.

In integral bases, pullback of a top superform contributes (det⁡L)2(\det L)^2, one determinant from each exterior factor. The scalar change of variables contributes ∣det⁡L∣−1|\det L|^{-1}. Their product is ∣det⁡L∣|\det L|. For a full-rank integral lattice map this is [Nτ:F(Nσ)][N_\tau:F(N_\sigma)], by the integer basis calculation in Lemma 5.14.1. A change of integral basis has determinant ±1\pm1, proving independence of the normalization. If the rank is less than dd, the pullback on ⋀d\bigwedge^d is zero. This proves the first assertions.

For completeness, the one-dimensional derivative formula also follows from the same integral. If gg is continuous, the difference quotient of G(t)=∫atg(s) dsG(t)=\int_a^t g(s)\,ds lies between the minimum and maximum of gg on the intervening interval; hence G′=gG'=g. A function with zero derivative is constant by the mean value argument: a continuous function on a closed interval attains its extrema by the compactness just proved, an interior extremum of a differentiable function has derivative zero, and subtraction of the line through the endpoint values gives Rolle’s argument. Applying this to a C1C^1 function minus the primitive of its derivative gives ∫abh′(s) ds=h(b)−h(a).\int_a^b h'(s)\,ds=h(b)-h(a).

Apply this identity on the coordinate lines through a polyhedron, followed by the established iterated integration. For a coordinate xjx_j, almost every such line meets a bounded convex polytope in an interval. Its endpoints lie on facets; the projections of intersections of distinct facets have dimension at most r−2r-2 and have zero (r−1)(r-1)-dimensional integral by the grid estimate. On a facet written xj=ℓ(x^j)x_j=\ell(\widehat x_j) the endpoint sign is the sign of the jj-th outward normal component. Give the facet its Euclidean measure in orthonormal affine coordinates; the proved linear change of variables computes this measure in any other affine parametrization. The graph’s Euclidean area factor is 1+∣∇ℓ∣2\sqrt{1+|\nabla\ell|^2}, obtained by taking the determinant of the Gram matrix I+∇ℓ ∇ℓtI+\nabla\ell\,\nabla\ell^{\mathsf t}. The absolute value of the corresponding unit-normal component is its reciprocal. Thus the endpoint integral is exactly the facet integral of that normal component times the coefficient. Facets parallel to these lines have zero normal component and give no term. Summing over jj proves the divergence formula on a polytope for a C1C^1 vector field. The argument applies to an unbounded polyhedron with compact support by cutting with a box outside the support; all the new facet terms vanish.

We now apply this coefficient formula to β\beta. Put sd=(−1)d(d−1)/2s_d=(-1)^{d(d-1)/2}. In coordinates for any integral basis of NσN_\sigma, write its tangential restriction uniquely as β=sd∑j=1d(−1)j−1bj d′x1∧⋯d′xj^⋯∧d′xd∧d′′x1∧⋯∧d′′xd.\beta=s_d\sum_{j=1}^d(-1)^{j-1}b_j\, d'x_1\wedge\cdots\widehat{d'x_j}\cdots\wedge d'x_d \wedge d''x_1\wedge\cdots\wedge d''x_d. Direct differentiation gives d′β=(∑j∂jbj)ωNσd'\beta=(\sum_j\partial_j b_j)\omega_{N_\sigma}. At a facet, the chosen basis of NτN_\tau followed by uσ/τu_{\sigma/\tau} is an integral basis of NσN_\sigma: the quotient is free of rank one and its indicated element is primitive. This is again Lemma 5.14.1. Use its affine coordinates (t1,…,td−1,s)(t_1,\ldots,t_{d-1},s), with the facet in s=0s=0 and the polyhedron locally in s≥0s\geq0. The divergence contribution of this facet is −bd(t,0) dt1⋯dtd−1-b_d(t,0)\,dt_1\cdots dt_{d-1}. On the other hand, β(v1′,…,vd−1′;v1′′,…,vd−1′′,uσ/τ′′)=sd(−1)d−1bd(t,0).\beta(v_1',\ldots,v_{d-1}'; v_1'',\ldots,v_{d-1}'',u_{\sigma/\tau}'') =s_d(-1)^{d-1}b_d(t,0). The ratio of these two expressions is (−1)dsd=(−1)d(d+1)/2(-1)^d s_d=(-1)^{d(d+1)/2}. A unimodular change of cell coordinates has absolute determinant one; the two exterior factors and the scalar change of variables above therefore preserve this facet calculation. This proves (5.46.1) for all facets with one common sign. A basis change in NτN_\tau contributes its determinant twice, and a change of the lift by a tangent vector gives a repeated vector in the double-primed alternating factor. These observations prove the last assertion. ◻

Theorem 5.46.3. NAG-002391 (Integration and closedness of a local balanced cycle). Let CC be a local weighted integral dd-cycle on an open subset U⊂NRU\subset N_{\mathbb R} in the sense of Definition 5.45.2. For smooth superforms with compact support on CC, the cell formula IC(α)=∑σmσ∫σαI_C(\alpha)=\sum_\sigma m_\sigma \int_\sigma\alpha defines a linear functional in bidegree (d,d)(d,d), independent of finite local presentations and subdivisions. It is nonnegative on forms whose tangential top coefficients are nonnegative whenever all weights are nonnegative. If CC is balanced, then IC(d′β)=0,IC(d′′γ)=0I_C(d'\beta)=0, \qquad I_C(d''\gamma)=0 for compactly supported forms of bidegrees (d−1,d)(d-1,d) and (d,d−1)(d,d-1), respectively. For d=0d=0 the functional is the weighted sum of values and both differential assertions are empty.

If an integral affine map aa is proper on the support of CC, has finite source presentations near the target support, and defines the lattice-index pushforward a∗Ca_*C, then IC(a∗α)=Ia∗C(α)I_C(a^*\alpha)=I_{a_*C}(\alpha) for every smooth α\alpha with compact ambient support in the target open set.

Proof. Let KK be the support of the form on CC. Around each point of this compact set choose a finite presentation and a smaller closed rectangular box contained in its open domain. Finitely many box interiors cover KK. Intersect the finitely many presented cells with these boxes and take a common hyperplane refinement. On overlaps their weights agree by the definition of a local cycle. Count each resulting cell once with that weight. This gives a finite polyhedral description on an open neighbourhood of KK in the support. Cells or artificial faces outside that neighbourhood meet no nonzero tangential coefficient of the form. The integral may therefore be computed on this finite description.

Two such descriptions have a common refinement near KK. The first assertion of Lemma 5.46.2 proves that the two sums coincide. Linearity and nonnegativity follow from the corresponding inequalities for real grid sums.

For β\beta, use such a finite description around its support. Any artificial exterior facet lies outside that support; all the adjacent tangential coefficients of β\beta vanish in a relative neighbourhood, so their boundary contractions vanish there as well. On a genuine (d−1)(d-1)-face τ\tau, formula (5.46.1) adds the normal lifts with their weights. Its contribution is (−1)d(d+1)/2∫τβ(v1′,…,vd−1′;v1′′,…,vd−1′′,(∑σ⊃τmσuσ/τ)′′) dμτ.(-1)^{d(d+1)/2}\int_\tau \beta(v_1',\ldots,v_{d-1}';v_1'',\ldots,v_{d-1}'', (\sum_{\sigma\supset\tau} m_\sigma u_{\sigma/\tau})'') \,d\mu_\tau. Balancing puts the vector in the last slot in (Nτ)R(N_\tau)_{\mathbb R}. The double-primed factor alternates, so this expression is zero. Summing the finitely many faces gives IC(d′β)=0I_C(d'\beta)=0. The identities d′′=Jd′Jd''=Jd'J and JωNσ=(−1)dωNσJ\omega_{N_\sigma}=(-1)^d \omega_{N_\sigma} give the assertion for γ\gamma by exchanging the two exterior factors. This verifies both closedness statements with the stated normalization.

For the projection formula, properness makes the support of the pullback on CC compact. Choose finite source and target presentations near these compact sets and refine along all cell images. On a source dd-cell of full rank, the affine map is injective on its affine span. The change-of-variables formula of Lemma 5.46.2 multiplies its contribution by [Nτ:a(Nσ)][N_\tau:a(N_\sigma)]. On a source cell of rank less than dd the pullback vanishes. Grouping the remaining finitely many contributions over each target cell gives exactly its lattice-index pushforward weight. This is the asserted identity. ◻

Theorem 5.46.4. NAG-002392 (Analytic moment integrals and finite change of charts). Let X=M(B)X=\mathcal M(B) be a generalized affinoid space, pure of analytic dimension dd, over a complete non-Archimedean field with its specified absolute value. For a unit tuple ff write tf:X→Rnt_f:X\to\mathbb R^n and Uf=Rn∖tf(∂KX)U_f=\mathbb R^n\setminus t_f(\partial_KX). For every finite BB-module MM, the formula IM,f(α)=ITrop⁡f(M)∣Uf(α)I_{M,f}(\alpha) =I_{\operatorname{Trop}_f(M)|_{U_f}}(\alpha) defines a linear functional on smooth top forms whose tangential support is compact in UfU_f. It is nonnegative when their tangential top coefficients are nonnegative. It is additive in short exact sequences of modules and satisfies IM,f(d′β)=0,IM,f(d′′γ)=0I_{M,f}(d'\beta)=0, \qquad I_{M,f}(d''\gamma)=0 for smooth compactly supported tests of the indicated complementary bidegrees in UfU_f.

Let p:X=M(B)→Y=M(A)p:X=\mathcal M(B)\to Y=\mathcal M(A) be finite, with YY also pure of analytic dimension dd. Suppose unit tuples fX,fYf_X,f_Y and an integral map aa satisfy tfY∘p=a∘tfXt_{f_Y}\circ p=a\circ t_{f_X}. For every finite BB-module MM, regarded also as a finite AA-module, and every smooth top form α\alpha with compact ambient support in UfYU_{f_Y}, IM,fX(a∗α)=IM,fY(α).\begin{equation} I_{M,f_X}(a^*\alpha)=I_{M,f_Y}(\alpha). \end{equation} The support of the left-hand test on its actual source cycle is compact and lies in its whole interior weight domain; there α\alpha is understood to be extended by zero outside UfYU_{f_Y}.

In particular, let f,gf,g be two unit tuples on the same XX, let πf,πg\pi_f,\pi_g be the projections from the joint tuple (f,g)(f,g), and let α,β\alpha,\beta be top forms with compact ambient support in Uf,UgU_f,U_g, respectively. Extend them by zero outside these open sets. If πf∗α=πg∗β\pi_f^*\alpha=\pi_g^*\beta on every tangential dd-plane of the actual joint local moment image over πf−1(Uf)∪πg−1(Ug)\pi_f^{-1}(U_f)\cup\pi_g^{-1}(U_g), then IB,f(α)=IB,g(β)I_{B,f}(\alpha)=I_{B,g}(\beta).

Adjoining unit coordinates is faithful on tangential superforms in every bidegree: on tf(X)∩Uft_f(X)\cap U_f, a smooth superform is zero on all tangential faces if and only if its pullback is zero on all tangential faces of the actual joint image over UfU_f.

Proof. Theorem 5.45.3 constructs the local balanced module cycles with nonnegative integral weights. Their definition uses the actual Gauss algebras and generic local lengths of Theorem 5.40.2. The functional, its positivity, and both differential identities are therefore Theorem 5.46.3. The exact-sequence equality of module weights gives the corresponding equality of these finite cell integrals.

First assume that the target units pull back to the actual monomials in the source units prescribed by aa. For this finite map, every point whose target weight belongs to UfYU_{f_Y} is interior in XX, by finite relative interior and composition, as proved in Theorem 5.45.3. The complete source moment support over a compact subset of UfYU_{f_Y} is compact. Near each target weight, Lemma 5.45.1 supplies a finite source presentation retaining the entire inverse support; the target has the local fan presentation of Theorem 5.38.6. Theorem 5.42.2 on these finite presentations gives a∗(Trop⁡fX(M)∣a−1(UfY))=Trop⁡fY(M)∣UfY.a_*\bigl(\operatorname{Trop}_{f_X}(M)|_{a^{-1}(U_{f_Y})}\bigr) =\operatorname{Trop}_{f_Y}(M)|_{U_{f_Y}}. The local identities agree after refinement because the module weights refer to the same actual fibres. The projection formula of Theorem 5.46.3 proves (5.46.2). Notice that its real factor is the lattice index, whereas the target coefficients are the full finite-module generic lengths. Both appear in the proved cycle formula.

For a general equality of moment maps as in the statement, use the joint source tuple (fX,p∗fY)(f_X,p^*f_Y). Its actual image is the graph of aa. The case just proved applies both to its projection onto fXf_X with the identity morphism and to its projection onto fYf_Y with the morphism pp. The first graph projection is an integral isomorphism, with inverse u↦(u,a(u))u\mapsto(u,a(u)). It identifies saturated tangent lattices and has index one on every cell. The second projection is aa after this identification. The corresponding finite lattice-index sums therefore give the same cycle formula for the stated moment-map equality, and its integral follows from the same real projection formula. All these comparisons are over UfYU_{f_Y}; finite relative interior and composition give the required absolute interior on the whole source fibres.

Apply this result to p=id⁡Xp=\operatorname{id}_X with source tuple (f,g)(f,g) and target tuple ff, and then with target tuple gg. The source structure module is BB in both cases. A compactly supported smooth form on UfU_f extends smoothly by zero to Rn\mathbb R^n: near the complement of UfU_f its support is absent. The support of its pullback on the joint actual image is a closed subset of the compact set t(f,g)(X)∩πf−1(supp⁡α)t_{(f,g)}(X)\cap\pi_f^{-1}(\operatorname{supp}\alpha). Every point of that set lies over a whole interior ff-fibre, so the joint cycle is defined there. The same holds for β\beta. Their tangential equality gives equal cell integrals on a common finite presentation. The two identity projection formulas now give the asserted equality of original integrals.

Finally consider a form η\eta of any bidegree on the target moment image over UfU_f. If its tangential restrictions vanish, refine the joint and target local presentations so that every source cell maps into one target cell. Pullback on tangent spaces then gives zero on the source. Conversely suppose the joint pullback vanishes. Near any target point take the finite presentations supplied above, and remove the lower-dimensional target cells and the images of lower-dimensional or rank-dropping source cells, exactly as in Theorem 5.44.1. On the dense remaining part of every target dd-cell, the positive target weight is the sum of the positive source weights times their lattice indices. Hence at least one full-rank source cell lies above every such point. Its tangent map is a real linear isomorphism onto the target tangent plane, so pullback is injective on both exterior factors. The tangential value of η\eta is zero there. Continuity of its smooth coefficients extends this vanishing over the entire target dd-cell. Every smaller face is contained in such a cell, and restriction to its tangent space preserves zero. This proves the converse, including the bidegrees above dd, where all these restrictions are already zero. ◻

5.47 Weights on partial affinoid domains

Throughout this section KK is complete for its specified real-valued non-Archimedean absolute value. Trivial valuation and arbitrary positive affinoid radii are allowed. Purity means analytic purity as in Section 5.35. All fibre algebras are the actual affinoid algebras, with their nilpotents.

Lemma 5.47.1. NAG-002395 (Finite affinoid domains and idempotents). Let A→BA\to B represent a generalized affinoid domain U⊆M(A)U\subseteq\mathcal M(A). Suppose that BB is finite over AA and that its given norm is equivalent to its canonical finite-module norm. There is a unique idempotent e∈Ae\in A such that the restriction map identifies BB with eA=A/(1−e)AeA=A/(1-e)A, boundedly in both directions. Moreover, U={x:∣e(x)∣=1},A≃eA×(1−e)A.U=\{x:|e(x)|=1\},\qquad A\simeq eA\times(1-e)A. The set UU is open and closed, and the displayed product is a bounded isomorphism with bounded inverse. These assertions include the empty domain, for which e=0e=0.

Proof. The self-intersection of the domain has actual algebra B⊗^AB≃BB\widehat\otimes_A B\simeq B, and the isomorphism is multiplication, by Theorem 2.36.9. Proposition 1.18.7 identifies this tensor with B⊗ABB\otimes_A B. Thus the two tensor arrows from BB agree, and A→BA\to B is an epimorphism of ordinary rings. The finite epimorphism theorem, [StacksAlg] (Tag 04VT), makes it surjective.

Put J=ker⁡(A→B)J=\ker(A\to B). It is finitely generated by Theorem 1.18.3. The domain restriction is flat by Proposition 2.38.2. Tensoring J↪AJ\hookrightarrow A with A/JA/J is therefore injective. Its source is J/J2J/J^2 and its map to A/JA/J is zero, so J=J2J=J^2. By [StacksAlg] (Tag 00EH), J=(1−e)AJ=(1-e)A for an idempotent ee. Multiplication by ee identifies A/JA/J with eAeA. If another idempotent e′e' has the same kernel, then e(1−e′)=e′(1−e)=0e(1-e')=e'(1-e)=0, so e=ee′=e′e=ee'=e'. This proves uniqueness.

Every point takes an idempotent to 00 or 11. The points killing (1−e)A(1-e)A are exactly those with ∣e∣=1|e|=1. The quotient norm is canonical, so they are precisely the points of BB. The two level sets of ∣e∣|e| are open and closed. The algebraic maps in the product decomposition and their inverses are maps between finite canonical AA-modules, hence bounded by Theorem 1.18.3. The factors are actual affinoid quotients by Corollary 1.18.4. Finally the empty-domain algebra is zero by Theorem 1.20.1, giving e=0e=0. ◻

Theorem 5.47.2. NAG-002396 (Inclusion–exclusion on interior fibre covers). Let X=M(A)X=\mathcal M(A) be pure of analytic dimension dd, let f1,…,fn∈A×f_1,\ldots,f_n\in A^\times, and put t=(−log⁡∣f1∣,…,−log⁡∣fn∣)t=(-\log|f_1|,\ldots,-\log|f_n|). Fix w∈t(X)w\in t(X) and a saturated rank-dd lattice N⊂ZnN\subset\mathbb Z^n. Suppose that Fw=t−1(w)⊂Int⁡(X/K),t(X)∩W⊂w+NRF_w=t^{-1}(w)\subset\operatorname{Int}(X/K),\qquad t(X)\cap W\subset w+N_{\mathbb R} for an open neighborhood WW of ww. Let V1,…,Vq⊆XV_1,\ldots,V_q\subseteq X be affinoid domains such that Fw⊆⋃i=1qVi,Fw∩Vi⊆Int⁡(Vi/K)(1≤i≤q).F_w\subseteq\bigcup_{i=1}^q V_i, \qquad F_w\cap V_i\subseteq\operatorname{Int}(V_i/K) \quad(1\leq i\leq q). For nonempty I⊆{1,…,q}I\subseteq\{1,\ldots,q\}, write VI=⋂i∈IViV_I=\bigcap_{i\in I}V_i, AI=O(VI)A_I=\mathcal O(V_I) and FI=Fw∩VIF_I=F_w\cap V_I. Let DD and DID_I be the actual algebras of FwF_w and FIF_I.

Each FIF_I is an open and closed affinoid domain of FwF_w. There are unique idempotents ei∈De_i\in D such that, with eI=∏i∈Ieie_I=\prod_{i\in I}e_i, restriction gives bounded isomorphisms DI≃eID,∏i=1q(1−ei)=0.\begin{equation} D_I\simeq e_ID, \qquad \prod_{i=1}^q(1-e_i)=0. \end{equation} For every integral Q:Zn→ZdQ:\mathbb Z^n\to\mathbb Z^d injective on NN, let RQR_Q be the exact-radius polytorus algebra at QwQw, let HQH_Q be its Gauss field, and put EQ=D⊗^RQHQ,EQ,I=DI⊗^RQHQ,cQ=[Zd:Q(N)].E_Q=D\widehat\otimes_{R_Q}H_Q, \qquad E_{Q,I}=D_I\widehat\otimes_{R_Q}H_Q, \qquad c_Q=[\mathbb Z^d:Q(N)]. Then EQ,I=eIEQE_{Q,I}=e_IE_Q. For every finite DD-module PP, dim⁡HQ(P⊗DEQ)=∑∅≠I⊆{1,…,q}(−1)∣I∣+1dim⁡HQ(P⊗DEQ,I).\begin{equation} \dim_{H_Q}(P\otimes_D E_Q) =\sum_{\varnothing\ne I\subseteq\{1,\ldots,q\}} (-1)^{|I|+1} \dim_{H_Q}(P\otimes_D E_{Q,I}). \end{equation} All these tensors also agree with their completed tensors for the canonical finite-module norms.

In particular, let MM be a finite AA-module and set MI=M⊗AAIM_I=M\otimes_A A_I. Write mMI,f(w)=0m_{M_I,f}(w)=0 if FIF_I is empty. For a nonempty FIF_I, its whole fibre is interior in VIV_I and its module weight is defined by Theorem 5.40.2. Then mM,f(w)=∑∅≠I⊆{1,…,q}(−1)∣I∣+1mMI,f(w).\begin{equation} m_{M,f}(w)= \sum_{\varnothing\ne I\subseteq\{1,\ldots,q\}} (-1)^{|I|+1}m_{M_I,f}(w). \end{equation} For M=AM=A this is the same identity for the intrinsic weights mX,fm_{X,f} and mVI,fm_{V_I,f}. If the selected parts Fw∩ViF_w\cap V_i are pairwise disjoint, it reduces to the sum of their weights.

There is an open neighborhood W′W' of ww on which the same domains cover every whole fibre. For every v∈t(X)∩W′=(w+NR)∩W′v\in t(X)\cap W'=(w+N_{\mathbb R})\cap W', all the preceding assertions apply at vv. The empty intersections remain empty, and each module weight in (5.47.3) is constant on this neighborhood in the indicated plane.

Proof. Finite intersections are actual affinoid domains by Corollary 2.36.10. Every nonempty one is pure of analytic dimension dd by Theorem 5.35.3. If x∈Fw∩Vix\in F_w\cap V_i, the composition equality 2.44.1 makes xx interior in Vi/XV_i/X. The domain interior theorem 2.44.3 says that it is an ordinary interior point of ViV_i in XX. Consequently, if x∈FIx\in F_I, the intersection of these finitely many neighborhoods is a neighborhood of xx contained in VIV_I. The same domain theorem, followed by the composition inclusion in Proposition 2.33.21, gives FI⊂Int⁡(VI/K).F_I\subset\operatorname{Int}(V_I/K). Thus no extra boundary condition on the intersections is required.

The whole flat-germ theorem 5.37.1 makes DD and every nonzero DID_I finite canonical modules over the same RQR_Q. It applies to VIV_I because its moment image is contained in that of XX, and hence in the same affine plane near ww. An empty FIF_I has zero algebra by Theorem 1.20.1. The domain base-change theorem 2.36.9 identifies D→DID\to D_I with the actual domain restriction to FI⊂FwF_I\subset F_w.

This restriction is also finite with its canonical DD-module norm. Indeed, any finite RQR_Q-generating family of DID_I also generates it over DD. Controlled lifts through its canonical RQR_Q-presentation, followed by the bounded map RQ→DR_Q\to D, bound the quotient norm from the corresponding DD-presentation by a constant times the given norm of DID_I. Conversely the bounded DD-action bounds its given norm by that quotient norm. These are exactly the two finite-generator norm comparisons of Theorem 1.18.3.

Lemma 5.47.1, first for each singleton II, supplies the unique eie_i. Domain intersections and their completed tensor presentations identify DI≃⨂^i∈I,DeiD≃eID.D_I\simeq\widehat\bigotimes_{i\in I,D}e_iD \simeq e_ID. Here completion changes nothing by Proposition 1.18.7; multiplication gives the last algebraic identification. If h=∏i(1−ei)h=\prod_i(1-e_i) were nonzero, the nonzero affinoid quotient hDhD would have a point by Theorem 1.20.1. Its image in FwF_w would belong to none of the ViV_i, a contradiction. This proves (5.47.1) with the actual algebras, including their nilpotents.

Finite base change, Proposition 2.33.19, gives EQ,I=eIEQE_{Q,I}=e_IE_Q. Put L=P⊗DEQL=P\otimes_D E_Q, a finite-dimensional HQH_Q-vector space. For each J⊆{1,…,q}J\subseteq\{1,\ldots,q\} set aJ=∏j∈Jej∏j∉J(1−ej).a_J=\prod_{j\in J}e_j\prod_{j\notin J}(1-e_j). These are pairwise orthogonal idempotents with sum 11, and a∅=0a_{\varnothing}=0. Therefore L=⨁J≠∅aJLL=\bigoplus_{J\ne\varnothing}a_JL. On the summand aJLa_JL, eIe_I acts as 11 precisely when I⊆JI\subseteq J, and otherwise as 00. Its coefficient in the right side of (5.47.2) is ∑∅≠I⊆J(−1)∣I∣+1=1,\sum_{\varnothing\ne I\subseteq J}(-1)^{|I|+1}=1, as follows by expanding ∏j∈J(1−1)=0\prod_{j\in J}(1-1)=0. This proves that formula.

The finite DD-module PP is finite over RQR_Q. Composing finite free quotient presentations gives equivalent canonical norms for this restriction of scalars. Proposition 1.18.7 consequently identifies every tensor just used with its complete canonical finite-dimensional tensor. In particular, if P=M⊗ADP=M\otimes_A D, associativity identifies its factor P⊗DEQ,IP\otimes_D E_{Q,I} with MI⊗AIEQ,IM_I\otimes_{A_I}E_{Q,I}. Dividing (5.47.2) by cQc_Q and applying Theorem 5.40.2 to XX and the nonempty VIV_I proves (5.47.3), independently of QQ. The structure-module case agrees with Theorem 5.37.1. Pairwise disjoint parts make every intersection of size at least two empty.

We finally choose the neighborhood uniformly. The ordinary interiors of the ViV_i cover FwF_w, as proved above. The complement of their union in compact XX is compact, and its moment image misses ww. Also the compact sets ∂KX\partial_KX and ∂KVi\partial_KV_i have moment images missing ww. Shrink WW to avoid all these images. Compactness and openness of interior are Propositions 2.24.1 and 2.33.21. The resulting neighborhood has whole fibres covered by the same ViV_i, and each selected fibre is interior in its own domain. The intersection argument above applies unchanged. Every VIV_I whose fibre at ww is empty has compact moment image missing ww; shrinking again keeps it empty throughout the neighborhood.

Theorem 5.37.2 makes the image of XX equal to the indicated plane nearby. For each fixed VIV_I, apply the finite generic-component formula in Theorem 5.40.2 to MIM_I. A component with empty fibre at ww remains absent nearby by compactness. Every other component has a whole interior flat fibre, as verified in that theorem’s proof, and its intrinsic weight is locally constant by Theorem 5.37.2. There are only finitely many components and intersections. One final shrinking gives their simultaneous local constancy and all the stated assertions. ◻

References. The algebraic inputs concerning finite epimorphisms and idempotent ideals are the cited Stacks lemmas. The contribution here is their application to the actual generalized analytic fibre algebras and the resulting boundary-sensitive localization of Gauss and tropical module weights. This is a foundational reconstruction; no claim of novelty in the literature is made.

5.48 Global additivity over affinoid covers

Use the global top-dimensional chains Tf(M)\mathcal T_f(M) and their lattice-normalized integrals IM,f\mathcal I_{M,f} from the finite-module chain and weights and the finite-cell functional. The original complete field and absolute value are fixed; trivial valuation and arbitrary positive presentation radii are allowed.

Theorem 5.48.1. NAG-002401 (Global additivity for a finite affinoid cover). Let X=M(A)X=\mathcal M(A) be nonempty and pure of analytic dimension dd, let MM be a finite AA-module, and let a finite tuple of analytic units ff define t:X→Rnt:X\to\mathbb R^n. Suppose that V1,…,VqV_1,\ldots,V_q is a finite affinoid-domain cover of XX. For nonempty I⊆{1,…,q}I\subseteq\{1,\ldots,q\} put VI=⋂i∈IVi,AI=O(VI),MI=M⊗AAI,fI=f∣VI.V_I=\bigcap_{i\in I}V_i,\qquad A_I=\mathcal O(V_I),\qquad M_I=M\otimes_A A_I,\qquad f_I=f|_{V_I}. An empty intersection contributes the zero chain and zero integral. Then, as integral weighted dd-chains in the same Rn\mathbb R^n, Tf(M)=∑∅≠I⊆{1,…,q}(−1)∣I∣+1TfI(MI).\begin{equation} \mathcal T_f(M)= \sum_{\varnothing\ne I\subseteq\{1,\ldots,q\}} (-1)^{|I|+1}\mathcal T_{f_I}(M_I). \end{equation} In particular, for every smooth real superform α\alpha of bidegree (d,d)(d,d) defined on a neighborhood of the actual compact image t(X)t(X), IM,f(α)=∑∅≠I⊆{1,…,q}(−1)∣I∣+1IMI,fI(α).\begin{equation} \mathcal I_{M,f}(\alpha)= \sum_{\varnothing\ne I\subseteq\{1,\ldots,q\}} (-1)^{|I|+1}\mathcal I_{M_I,f_I}(\alpha). \end{equation} The identities are unchanged by finite refinements of the cover or of the polyhedral complexes used to represent the chains.

Proof. Corollary 2.36.10 makes every VIV_I an actual affinoid domain of XX, with its compatible restriction algebra. Every nonempty VIV_I is pure of analytic dimension dd by Theorem 5.35.3. The finite module MIM_I has its canonical complete norm by Theorem 1.18.3. Thus all the nonzero terms in the claimed identity have the stated analytic meaning.

Theorem 5.39.3 makes the actual images t(X)t(X) and t(VI)t(V_I) finite integral affine polyhedral sets. It also makes E=t(∂KX) ∪ ⁣⋃∅≠I⊆{1,…,q}t(∂KVI)E=t(\partial_KX)\ \cup\! \bigcup_{\varnothing\ne I\subseteq\{1,\ldots,q\}} t(\partial_KV_I) a finite polyhedral subset of t(X)t(X) of dimension at most d−1d-1; for d=0d=0 it is empty. By Lemma 5.14.2, choose one finite common subdivision of t(X)t(X) in which every t(VI)t(V_I) and EE is a subcomplex, and which also refines the representatives of all the finitely many chains under consideration.

Fix a dd-cell σ\sigma of this subdivision and a point w∈relint⁡(σ)w\in\operatorname{relint}(\sigma). Its relative interior misses EE. Since the whole image has dimension at most dd, near ww no cell other than σ\sigma is present. Hence the local moment image is contained in the affine dd-plane w+(Nσ)Rw+(N_\sigma)_{\mathbb R}, where NσN_\sigma is its saturated tangent lattice. The whole fibre FwF_w is interior in XX, and each Fw∩ViF_w\cap V_i is interior in ViV_i, by the definition of EE. The domains cover FwF_w because they cover XX.

Apply Theorem 5.47.2 to this fibre and the finite module MM. It gives mM,f(w)=∑∅≠I⊆{1,…,q}(−1)∣I∣+1mMI,fI(w),m_{M,f}(w)= \sum_{\varnothing\ne I\subseteq\{1,\ldots,q\}} (-1)^{|I|+1}m_{M_I,f_I}(w), with zero for the empty selected fibres. The common subdivision ensures that relint⁡(σ)\operatorname{relint}(\sigma) is either contained in t(VI)t(V_I) or disjoint from it. By the finite-module chain and weights, the terms in this identity are exactly the coefficients of the corresponding global chains on σ\sigma. Thus these coefficients agree on every dd-cell. Lower-dimensional cells carry no coefficient in a dd-chain. This proves (5.48.1); it also covers the case dim⁡t(X)<d\dim t(X)<d, when there are no dd-cells.

All these chains have compact polyhedral support contained in t(X)t(X). Apply the finite-cell functional to the same smooth form on each term. Their common representatives have finitely many cells, and the integral is the sum of the cell integrals times their integer coefficients. The coefficient identity proves (5.48.2). Subdivision invariance is part of the finite-cell functional. Repeating the argument for any finite refinement of the affinoid cover gives the same global chain Tf(M)\mathcal T_f(M) and the same integral, which proves the final assertion. ◻

5.49 Global moment chains and finite analytic chart data

NAG-002405 · Open at full scope

Theorem 5.49.1. NAG-002405 (Global finite-module tropical chains). Let X=M(A)X=\mathcal M(A) be pure of analytic dimension dd over a complete non-Archimedean field with its specified absolute value, and let a unit tuple ff define tf:X→Rnt_f:X\to\mathbb R^n. For every finite AA-module MM there is a canonical finite nonnegative integral dd-dimensional polyhedral chain Tf(M).\mathcal T_f(M). Its support is contained in tf(X)t_f(X). On any common finite subdivision of the actual image and tf(∂KX)t_f(\partial_KX), the weight of a dd-cell is the module weight mM,f(v)m_{M,f}(v) at any point of its relative interior. Cells of dimension less than dd give no dd-chain term. The chain is additive in short exact sequences, independent of subdivision, and balanced off tf(∂KX)t_f(\partial_KX). For M=AM=A every occurring dd-cell has positive weight.

For a smooth (d,d)(d,d)-superform α\alpha on an open neighbourhood of the actual compact image, define IM,f(α)=∑σmσ∫σα.\mathcal I_{M,f}(\alpha) =\sum_\sigma m_\sigma\int_\sigma\alpha. This is independent of subdivision and is nonnegative when the tangential top coefficients of α\alpha are nonnegative. It agrees with IM,fI_{M,f} of Theorem 5.46.4 for tests supported in the whole interior weight domain.

Let p:M(B)→M(A)p:\mathcal M(B)\to\mathcal M(A) be finite, with both spaces pure of analytic dimension dd, and suppose moment tuples satisfy tfA∘p=a∘tfBt_{f_A}\circ p=a\circ t_{f_B} for an integral map aa. For every finite BB-module MM, a∗TfB(M)=TfA(M),IM,fB(a∗α)=IM,fA(α).a_*\mathcal T_{f_B}(M)=\mathcal T_{f_A}(M), \qquad \mathcal I_{M,f_B}(a^*\alpha) =\mathcal I_{M,f_A}(\alpha). On the right MM is its finite AA-module. In particular, adjoining analytic unit coordinates preserves these chains under projection and preserves the corresponding integrals.

Proof. Theorem 5.39.3 gives a finite polyhedral image of dimension at most dd and a finite boundary image of dimension at most d−1d-1. Choose one finite common subdivision with the boundary image as a subcomplex. At a point in the relative interior of a dd-cell, the whole moment fibre is therefore interior, and the local image is contained in that cell’s affine dd-plane. Theorem 5.40.2 defines the nonnegative integer weight there. It is locally constant, by Theorem 5.42.1 on small convex neighbourhoods with the actual local fan presentations of Theorem 5.38.6. The relative interior of the cell is convex, so the weight is constant there. For the structure module it is positive by Theorem 5.37.1.

On a common subdivision of two choices, the generic weights compute the same whole-fibre module invariant. Thus their chains agree. Additivity follows from the exact-sequence formula of Theorem 5.40.2. On the complement of the boundary image the chain is the local balanced module cycle of Theorem 5.45.3. All cells with nonzero weights are compact polytopes: the actual moment image is compact, and the complex is finite. Consequently the integral is a finite sum of the integrals constructed in Lemma 5.46.2. That lemma gives subdivision independence and positivity. Restriction to the whole interior weight domain gives exactly the previously defined local integral.

First suppose that the compatibility is an identity of analytic characters: each target unit pulls back to the monomial in the source units prescribed by aa. For this finite morphism, take finite global subdivisions of the two actual images and their boundary images. Refine along the images of all source cells, using the finite linear elimination and hyperplane construction in Theorem 5.17.2. Outside a finite union of polyhedra of dimension at most d−1d-1 in each target dd-cell, the target whole fibre is interior, and every source weight above it lies in a full-rank dd-cell. The entire inverse weight set is finite, with at most one point per such source cell. Finite relative interior and composition imply absolute interior for every point of those source fibres.

Near such a target weight, Lemma 5.45.1 retains the complete source inverse support in a finite pure presentation. Theorem 5.42.2 then identifies the finite lattice-index sum with the target weight of MM as an AA-module. Its empty-inverse- image case gives zero on both sides. Thus the generic weight of every target dd-cell is exactly its pushforward weight. The resulting equality of finite chains is a∗TfB(M)=TfA(M)a_*\mathcal T_{f_B}(M)=\mathcal T_{f_A}(M). This proves the chain formula under the character identity. In particular it applies to the identity morphism and adjoining any analytic unit coordinates.

Now suppose only the equality of moment maps in the statement. On the source use the joint tuple (fB,p∗fA)(f_B,p^*f_A). Its actual image is the graph of aa over tfB(M(B))t_{f_B}(\mathcal M(B)). Projection of that graph onto its first factor is an integral isomorphism: its inverse is u↦(u,a(u))u\mapsto(u,a(u)). On the saturated tangent lattices of every corresponding cell these two maps are inverse, so the index of this projection is one. The augmentation case just proved identifies the graph chain with TfB(M)\mathcal T_{f_B}(M) under this isomorphism. The second projection satisfies an actual character identity with the target tuple. Its pushforward is TfA(M)\mathcal T_{f_A}(M) by the already proved case, and on the graph it is aa after the first projection. The equality of the finite lattice-index sums therefore proves the chain formula under the stated moment-map equality as well.

The full actual source support is compact, so this pushforward is proper. On each full-rank source cell, Lemma 5.46.2 gives the factor [Nτ:a(Nσ)][N_\tau:a(N_\sigma)] in the integral. On a rank-dropping cell the top-degree pullback is zero. Grouping these finite cell integrals by their target cells proves the integral identity. No ambient support cutoff is needed here, because all the integration polytopes are compact. Taking pp to be the identity and aa a coordinate projection proves the last assertion. ◻

Theorem 5.49.2. NAG-002406 (Faithfulness after adjoining analytic units). Let XX be any generalized affinoid space, and let f,gf,g be finite unit tuples. Write S=tf(X),S^=t(f,g)(X),π:S^⟶S.S=t_f(X),\qquad \widehat S=t_{(f,g)}(X),\qquad \pi:\widehat S\longrightarrow S. Restrict a smooth ambient superform to every tangential face of these actual finite polyhedral sets. These restrictions are independent of subdivision. Such a form on SS has all tangential restrictions zero if and only if its pullback has all tangential restrictions zero on S^\widehat S.

Tangential equality is preserved by restriction from an affinoid space to any affinoid domain, using the restricted unit coordinates.

Proof. Both actual images have finite polyhedral structures by Theorem 5.39.3, and π\pi is surjective on their actual point sets. A finite subdivision does not change tangential vanishing: the relative interiors of the full-dimensional pieces in each original cell are dense there, and continuity of the coefficients extends vanishing from them to the cell and its faces. The converse follows by restricting an alternating covector to a smaller tangent space.

Refine the target by the images of all source cells and refine the source by the inverse images of the resulting faces. Finite inequality elimination and hyperplane subdivision, as in Theorem 5.17.2, make this a finite polyhedral refinement. Every source cell then maps into a target cell. If the form vanishes tangentially on the target, pullback through these tangent maps vanishes on the source.

For the converse, fix a target cell τ\tau of dimension rr. Its relative interior is covered by the images of the finitely many source cells mapping into it. Remove their image faces of dimension less than rr. The remaining subset of relint⁡(τ)\operatorname{relint}(\tau) is dense: in its affine span, a line direction outside finitely many proper linear subspaces meets the corresponding proper affine spans in at most finitely many points. Such a direction is obtained by the polynomial vectors (1,b,…,br−1)(1,b,\ldots,b^{r-1}) avoiding finitely many roots; for r=0r=0 no exclusion is needed. At a point vv of this dense set some source cell σ\sigma has an image of dimension rr containing vv in its relative interior.

There is a preimage uu of vv in relint⁡(σ)\operatorname{relint}(\sigma). To verify this elementary convex assertion, choose u0u_0 in that relative interior. A small ball in its affine span lies in σ\sigma; its linear image contains a small ball in the image span, since a surjective map of finite-dimensional real vector spaces has a linear right inverse. Thus π(u0)\pi(u_0) is in the relative interior of the image. Since vv is also there, for some ε>0\varepsilon>0 the point v′=v+ε(v−π(u0))v'=v+\varepsilon(v-\pi(u_0)) still belongs to that image. Choose u′∈σu'\in\sigma over v′v'. Then u=u′+εu01+εu=\frac{u'+\varepsilon u_0}{1+\varepsilon} maps to vv and is interior in σ\sigma, because the positive coefficient of u0u_0 brings a scaled open ball with it.

The tangent map from σ\sigma onto τ\tau is surjective. Its pullback on each exterior cotangent factor is injective, by a linear right inverse. Vanishing of the joint pullback at uu therefore forces the tangential value of the original form at vv to be zero. Density and continuity prove vanishing throughout τ\tau. This applies to every cell, including lower-dimensional maximal cells of the actual image, and proves the converse.

Finally the moment image of an affinoid domain is a finite polyhedral subset of the old image, again by Theorem 5.39.3. Refine both complexes by all their defining hyperplanes. Each new cell of the subset lies in a cell of the old complex, so its tangent space is a subspace of that cell’s tangent space. Restriction preserves tangential vanishing and hence tangential equality. ◻

Definition 5.49.3. (Finite affinoid moment-chart data). Let X=M(A)X=\mathcal M(A) be pure of analytic dimension dd. Finite moment-chart data of bidegree (p,q)(p,q) consist of a finite affinoid-domain cover X=⋃i=1sViX=\bigcup_{i=1}^s V_i, a finite analytic unit tuple fif_i on each ViV_i, and a smooth (p,q)(p,q)-superform αi\alpha_i on a neighbourhood of tfi(Vi)t_{f_i}(V_i). On every intersection Vi∩VjV_i\cap V_j, the two forms must have equal tangential pullbacks to the actual joint moment image for (fi,fj)(f_i,f_j). A refinement may restrict the domains, adjoin unit coordinates, and replace a superform by one with the same tangential pullback. Data are identified if they agree on a common finite affinoid refinement.

Theorem 5.49.4. NAG-002408 (An integral independent of finite affinoid chart data). Fix a finite AA-module MM. For finite moment-chart data of bidegree (d,d)(d,d) on XX, put VI=⋂i∈IViV_I=\bigcap_{i\in I}V_i, AI=O(VI)A_I=\mathcal O(V_I) and MI=M⊗AAIM_I=M\otimes_A A_I. Choose any i∈Ii\in I to represent the restricted form on VIV_I, and set IM(α)=∑∅≠I⊆{1,…,s}(−1)∣I∣+1IMI,fi∣VI(αi).\begin{equation} \mathcal I_M(\alpha) =\sum_{\varnothing\ne I\subseteq\{1,\ldots,s\}} (-1)^{|I|+1} \mathcal I_{M_I,f_i|_{V_I}}(\alpha_i). \end{equation} An empty intersection contributes zero. The expression is independent of each chosen ii, of the finite cover, of the unit tuples and of the representatives within the equivalence of Definition 5.49.3. It is linear in the form and additive in short exact sequences of finite AA-modules. It is nonnegative if all chart representatives have nonnegative tangential top coefficients. For a single global moment presentation it is exactly the global integral given by the finite-cell functional.

Proof. Finite intersections are actual affinoid domains by Corollary 2.36.10; their algebras and restriction maps are the actual domain ones. Every nonempty intersection is pure of analytic dimension dd by Theorem 5.35.3. Thus all the displayed integrals exist by the finite-cell functional. Compatibility of the chart data restricts to further intersections by Theorem 5.49.2. On VIV_I, the same-affinoid coordinate augmentation identifies the integrals of any two chosen representatives with the integral of their equal joint pullbacks. This proves independence of the choice of ii.

First suppose the data come from a single tuple on XX. The domain-cover identity of Theorem 5.48.1 gives exactly (5.49.1). The same identity holds after restricting a single chart presentation to any of its affinoid subdomains.

Now compare compatible data on two finite covers (Vi)i=1s(V_i)_{i=1}^s and (Wj)j=1t(W_j)_{j=1}^t. The module on every overlap is the restriction of the same MM, by the actual domain tensor identifications of Proposition 1.18.7 and Theorem 2.36.9. For each nonempty II, cover VIV_I by the VI∩WjV_I\cap W_j and apply the preceding single-chart identity to its chosen representative. Substitution in (5.49.1) gives ∑I≠∅∑J≠∅(−1)∣I∣+∣J∣IM∣VI∩WJ(α∣VI∩WJ).\sum_{I\ne\varnothing}\sum_{J\ne\varnothing} (-1)^{|I|+|J|} \mathcal I_{M|_{V_I\cap W_J}}(\alpha|_{V_I\cap W_J}). The two representatives on each of these intersections have equal integrals by their joint-coordinate comparison. Interchange the two finite sums and apply the domain-cover identity to each WJW_J. This recovers precisely the expression for the second cover. Thus the value is unchanged under compatible replacement or refinement. For data identified after a common refinement, apply this comparison to each of the two original covers and that common refinement. This proves all claimed independences.

Linearity is termwise. For a short exact sequence of finite AA-modules, domain flatness on a single actual domain keeps it exact on every VIV_I. The global module-chain additivity of the finite-module chain and weights gives additivity of each integral, and hence of the finite alternating sum. The assertion about a single global presentation was proved in the second paragraph.

To verify positivity, order the cover and group the alternating sum by the largest index jj in II. The jj-th group uses the single representative αj\alpha_j and is the integral of the chain Cj=∑J⊆{1,…,j−1}(−1)∣J∣Tfj∣Vj∩VJ(M∣Vj∩VJ),C_j=\sum_{J\subseteq\{1,\ldots,j-1\}}(-1)^{|J|} \mathcal T_{f_j|_{V_j\cap V_J}} (M|_{V_j\cap V_J}), with the empty intersection in this formula understood to be VjV_j. Each CjC_j is nonnegative. Indeed, choose a common finite subdivision of these actual moment images and their boundary images by the common finite subdivision lemma. The latter have dimension at most d−1d-1 by Theorem 5.39.3. At a generic point of a dd-cell, apply Theorem 5.47.2 on VjV_j to the family consisting of VjV_j itself and the domains Vj∩ViV_j\cap V_i for i<ji<j. It supplies idempotents eie_i in the actual whole-fibre algebra; the member VjV_j ensures the cover hypothesis. For a transverse Gauss algebra EQE_Q and its common lattice divisor cQc_Q, the weight of CjC_j is 1cQdim⁡HQ(∏i<j(1−ei))(M∣Vj⊗AVjEQ)≥0.\frac{1}{c_Q}\dim_{H_Q} \left(\prod_{i<j}(1-e_i)\right) \bigl(M|_{V_j}\otimes_{A_{V_j}}E_Q\bigr) \geq0. This is the elementary expansion of the product of the complementary idempotents, with all module structures retained. It is already an integer because it is the displayed alternating sum of integer module weights. Thus every generic coefficient of CjC_j is nonnegative. Its integral against αj\alpha_j is nonnegative by the nonnegative cell integral on each cell. Summing these finitely many groups proves positivity. ◻

5.50 Integration on compact good analytic spaces

The complete field KK and its real-valued non-Archimedean absolute value are fixed. Trivial valuation and arbitrary positive affinoid radii are allowed. Compact means quasi-compact and Hausdorff.

Lemma 5.50.1. NAG-002411 (Affinoid intersections in a separated good space). Let XX be a good Hausdorff KK-analytic space. Any finite intersection of affinoid domains of XX is an affinoid domain, with the empty space allowed. Two finite affinoid-domain covers therefore have a common finite affinoid refinement. If XX is compact, it has a finite affinoid cover whose ordinary interiors cover XX.

Proof. Lemma 2.53.3, applied to X→M(K)X\to\mathcal M(K), makes the analytic diagonal Δ:X→X×KX\Delta:X\to X\times_KX a closed immersion. Let U=M(A)U=\mathcal M(A) and V=M(B)V=\mathcal M(B) be affinoid domains of XX. Theorem 2.40.11 makes U×KV=M(A⊗^KB)U\times_KV=\mathcal M(A\widehat\otimes_KB) an actual affinoid domain of X×KXX\times_KX. The pullback of Δ\Delta to this domain is an affinoid closed subspace. Indeed, the coherent defining ideal restricts to a finite ideal by Theorem 2.39.6, and its quotient is the actual affinoid quotient of Corollary 1.18.4. The arbitrary-domain restriction argument in Lemma 2.53.2 identifies this quotient with the entire diagonal pullback, including its analytic structure.

This pullback is U×XVU\times_XV. By the domain universal property of Theorem 2.40.8, a map to it is exactly a map to XX whose image lies in both UU and VV. Thus it is the actual analytic domain U∩VU\cap V. There is no identification of an analytic product with an ordinary product of point sets in this argument. Iteration proves the finite-intersection assertion, using domain transitivity in Proposition 2.40.5. The pairwise intersections of two finite covers form the claimed common refinement. Affinoid domains are compact by Proposition 2.24.1, and hence closed in the Hausdorff space XX. Lemma 2.39.2 makes every finite cover by such domains admissible. Finally, goodness supplies an affinoid neighbourhood of each point. Compactness gives a finite subcover of their ordinary interiors. ◻

Definition 5.50.2. (Pure dimension on a good space). A good analytic space is pure of analytic dimension dd if every point has an ordinary affinoid neighbourhood pure of analytic dimension dd in the sense of Section 5.35.

Lemma 5.50.3. NAG-002413 (Purity is local on affinoid covers). Let V1,…,VsV_1,\ldots,V_s be a finite affinoid-domain cover of a nonempty affinoid XX. Then XX is pure of analytic dimension dd if and only if every nonempty ViV_i is so. Consequently every nonempty affinoid domain of a good Hausdorff space pure of analytic dimension dd is pure of analytic dimension dd.

Proof. The forward implication is Theorem 5.35.3. For the converse write X=M(A)X=\mathcal M(A) and Vi=M(Ci)V_i=\mathcal M(C_i), omitting empty members. Fix a minimal prime p\mathfrak p of AA. The actual quotient A/pA/\mathfrak p is affinoid by Corollary 1.18.4, and its spectrum is nonempty by Theorem 1.20.1. Theorem 2.36.9 gives an affinoid-domain cover of this spectrum with algebras Di=Ci⊗^A(A/p)=Ci/pCi.D_i=C_i\widehat\otimes_A(A/\mathfrak p) =C_i/\mathfrak pC_i. The equality with the ordinary quotient follows from Proposition 1.18.7, applied to the finite canonical AA-module A/pA/\mathfrak p.

Every nonzero DiD_i is pure of analytic dimension dd. To see this, let q\mathfrak q be a prime of CiC_i minimal over pCi\mathfrak pC_i. The domain map A/p→DiA/\mathfrak p\to D_i is flat by Proposition 2.38.2. Flat going down, [StacksAlg] (Tag 00HS), shows that the minimal prime q/pCi\mathfrak q/\mathfrak pC_i contracts to the zero prime of the domain A/pA/\mathfrak p: otherwise going down to zero would give a strictly smaller prime. Hence q∩A=p\mathfrak q\cap A=\mathfrak p. Choose a minimal prime q0⊆q\mathfrak q_0\subseteq\mathfrak q of CiC_i by [StacksAlg] (Tag 00E0(4)). Its contraction is a prime contained in the minimal prime p\mathfrak p, and so equals p\mathfrak p. Thus pCi⊆q0\mathfrak pC_i\subseteq\mathfrak q_0. Minimality over pCi\mathfrak pC_i gives q0=q\mathfrak q_0=\mathfrak q. The minimal-prime quotient of DiD_i in question is therefore the minimal-prime quotient Ci/qC_i/\mathfrak q, whose analytic dimension is dd by the hypothesis on CiC_i.

A nonempty affinoid whose minimal-prime quotients all have dimension dd itself has dimension dd. Indeed, each point kernel contains a minimal prime by [StacksAlg] (Tag 00E0(4)). Passage to that quotient does not change its evaluated fraction field or its completed residue field. The maximum formula of Theorem 3.22.7 consequently identifies the dimension with the maximum of those component dimensions. Applying the domain-cover formula 3.22.8 to the displayed cover now gives dim⁡KM(A/p)=d.\dim_K\mathcal M(A/\mathfrak p)=d. This holds for every minimal prime of AA, proving purity. All dimensions here are analytic; no generalized Krull dimension comparison is used.

For the last assertion let VV be a nonempty affinoid domain of the good Hausdorff space. Its compact underlying space is covered by finitely many pure affinoid neighbourhoods WjW_j of points of XX. By Lemma 5.50.1, the sets V∩WjV\cap W_j are affinoid domains of WjW_j. Their nonempty members are pure of dimension dd by Theorem 5.35.3. They cover VV, so the implication just proved applies to VV. ◻

Lemma 5.50.4. NAG-002414 (Finite module data on an analytic cover). Let (Vi)(V_i) be a finite affinoid-domain cover of a good Hausdorff analytic space XX. Put Ai=O(Vi)A_i=\mathcal O(V_i) and Aij=O(Vi∩Vj)A_{ij}=\mathcal O(V_i\cap V_j). Finite AiA_i-modules MiM_i, with isomorphisms Aij⊗AiMi≃Aij⊗AjMjA_{ij}\otimes_{A_i}M_i \simeq A_{ij}\otimes_{A_j}M_j satisfying the identity and cocycle conditions, determine a coherent OXG\mathcal O_{X_G}-module F\mathcal F, uniquely up to the isomorphism inducing the given identifications. Conversely every coherent module gives such data. The construction is compatible with morphisms, finite affinoid refinements and short exact sequences. For every affinoid domain W⊆ViW\subseteq V_i, Γ(W,F)=O(W)⊗AiMi\Gamma(W,\mathcal F) =\mathcal O(W)\otimes_{A_i}M_i with its canonical finite-module norm.

Proof. All finite intersections are affinoid by Lemma 5.50.1. Theorem 2.39.6 associates to MiM_i a coherent sheaf M~i\widetilde M_i on (Vi)G(V_i)_G and identifies its restrictions with the displayed finite tensors. It turns the given module isomorphisms into sheaf isomorphisms on the intersections, with the same cocycle condition.

For any analytic domain T⊆XT\subseteq X, define F(T)\mathcal F(T) to be the families of sections of M~i\widetilde M_i on T∩ViT\cap V_i whose restrictions agree under these isomorphisms. The domain and covering calculus of Proposition 2.40.5 shows that an admissible cover of TT restricts to one of each T∩ViT\cap V_i. Existence and uniqueness of glued sections for the sheaves M~i\widetilde M_i therefore prove the sheaf axiom for these families. On a domain contained in ViV_i, its ii-th section determines all the others by restriction and the transition isomorphisms. The cocycle condition makes them compatible. Thus F∣(Vi)G\mathcal F|_{(V_i)_G} is M~i\widetilde M_i; in particular it is coherent. Indeed, finite generation and finite generation of the kernels of maps from finite free sheaves are both checked on this admissible cover, where M~i\widetilde M_i is coherent. The same description proves uniqueness and glues compatible module morphisms.

Conversely, sections of a coherent sheaf on each ViV_i are finite modules by Theorem 2.39.6. That theorem gives both their restriction isomorphisms and the displayed formula on WW. Theorem 1.18.3 gives the canonical norms and boundedness of all these module maps, while Proposition 1.18.7 identifies the finite ordinary and completed tensors. The two constructions are inverse on every chart and hence inverse globally by the sheaf axiom. Refinements simply restrict these same sheaves. Finally exactness is local on the cover, and the affinoid equivalence in Theorem 2.39.6 is exact. This proves the assertion for short exact sequences. ◻

Definition 5.50.5. (Moment-chart data on a compact good space). Let XX be compact, good and pure of analytic dimension dd. Finite moment-chart data of bidegree (p,q)(p,q) consist of a finite affinoid-domain cover (Vi)(V_i), a finite tuple of analytic units fif_i on each ViV_i, and a smooth real (p,q)(p,q)-superform αi\alpha_i on a neighbourhood of tfi(Vi)t_{f_i}(V_i). On each intersection the two forms have equal tangential pullbacks to the actual joint moment image. Data are identified when they agree on a common finite affinoid refinement, allowing adjoining unit coordinates and equality of tangential pullbacks. These are the conventions of Definition 5.49.3, with XX no longer required to be affinoid. The same definition is used on a compact analytic domain QQ that is a finite union of affinoid domains in an ambient good Hausdorff space pure of dimension dd. Its presenting domains are affinoid domains of QQ; QQ itself may fail to be good.

Theorem 5.50.6. NAG-002416 (The integral on a compact good analytic space). Let XX be a compact good KK-analytic space pure of analytic dimension dd, let F\mathcal F be a coherent OXG\mathcal O_{X_G}-module, and let α\alpha be finite moment-chart data of bidegree (d,d)(d,d). For a presenting cover (Vi)i=1s(V_i)_{i=1}^s put VI=⋂i∈IVi,MI=Γ(VI,F).V_I=\bigcap_{i\in I}V_i, \qquad M_I=\Gamma(V_I,\mathcal F). For nonempty II, choose any i∈Ii\in I. Then IX,F(α)=∑∅≠I⊆{1,…,s}(−1)∣I∣+1IMI,fi∣VI(αi)\begin{equation} \mathcal I_{X,\mathcal F}(\alpha) =\sum_{\varnothing\ne I\subseteq\{1,\ldots,s\}} (-1)^{|I|+1} \mathcal I_{M_I,f_i|_{V_I}}(\alpha_i) \end{equation} is independent of these choices and of the representative of the finite chart data. Empty intersections contribute zero. The integral is linear in α\alpha, additive in short exact sequences of coherent modules, and nonnegative if the chart representatives have nonnegative tangential top coefficients. It may equivalently be constructed from compatible finite module data as in Lemma 5.50.4. For affinoid XX it agrees with Theorem 5.49.4.

More generally, let YY be good, Hausdorff and pure of analytic dimension dd, and let Q⊆YQ\subseteq Y be the union of finitely many affinoid domains of YY. The same construction and all these conclusions hold for coherent modules and finite moment-chart data on the compact analytic domain QQ, including when QQ is not good.

Proof. The intersections VIV_I are actual affinoid domains by Lemma 5.50.1, and their nonempty members are pure of dimension dd by Lemma 5.50.3. Lemma 5.50.4 makes MIM_I a finite module, identified with the restriction of any MiM_i for i∈Ii\in I. Thus every nonzero term is given by the finite-cell functional on the actual affinoid. Compatibility of the chart forms restricts to every further domain by Theorem 5.49.2. On VIV_I, adjoining two chosen tuples identifies their integrals with the integrals of their joint pullbacks, by the same-affinoid coordinate augmentation. The pullbacks have the same tangential top coefficients. Their cell integrals agree, proving independence of ii.

Consider two compatible presentations on finite covers (Vi)(V_i) and (Wj)(W_j). On each VIV_I the domains VI∩WjV_I\cap W_j form a finite affinoid cover. Apply Theorem 5.48.1 to the chosen single-chart representative on VIV_I and the module Γ(VI,F)\Gamma(V_I,\mathcal F). Inserting its formula in (5.50.1) gives the finite double sum ∑I≠∅∑J≠∅(−1)∣I∣+∣J∣IF∣VI∩WJ(α∣VI∩WJ).\sum_{I\ne\varnothing}\sum_{J\ne\varnothing} (-1)^{|I|+|J|} \mathcal I_{\mathcal F|_{V_I\cap W_J}} (\alpha|_{V_I\cap W_J}). The modules are the same restrictions of F\mathcal F, and the two local form representatives have equal integrals by the joint-coordinate argument just given. Interchange the finite sums and apply the cover formula on each WJW_J. This recovers the expression for the second presentation. If two data sets agree only after a common refinement, compare each with that refinement. Lemma 5.50.1 provides the needed finite intersections. Restriction preserves tangential equality by Theorem 5.49.2, and the same-affinoid coordinate augmentation preserves the integral. This proves all the asserted independences.

Linearity follows after passing to a common cover and adjoining the two tuples on each member, where addition is ordinary addition of their pulled-back superforms. For a short exact sequence of coherent modules, the affinoid exact equivalence 2.39.6 makes its section sequence on every VIV_I exact. Additivity of the affinoid global chains in the finite-module chain and weights therefore proves additivity of the displayed finite sum.

For positivity, order the cover and use the largest index jj of II as the representative in each term. The jj-th group is the integral against αj\alpha_j of the finite dd-chain Cj=∑J⊆{1,…,j−1}(−1)∣J∣Tfj∣Vj∩VJ(Γ(Vj∩VJ,F)),C_j=\sum_{J\subseteq\{1,\ldots,j-1\}}(-1)^{|J|} \mathcal T_{f_j|_{V_j\cap V_J}} \bigl(\Gamma(V_j\cap V_J,\mathcal F)\bigr), where Vj∩V∅V_j\cap V_\varnothing means VjV_j. All moment tuples in this formula are restrictions of the single tuple fjf_j on the affinoid VjV_j. By Theorem 5.39.3 and the common finite subdivision lemma, choose a finite common subdivision of these actual images and their boundary images. The boundary images have dimension at most d−1d-1, and are empty for d=0d=0.

At the relative interior of any dd-cell, its whole fibre in VjV_j and all its parts in Vj∩ViV_j\cap V_i are interior. Near that point the actual image lies in the cell’s saturated dd-plane. Apply Theorem 5.47.2 on VjV_j to the family consisting of VjV_j itself and the Vj∩ViV_j\cap V_i, i<ji<j. The member VjV_j ensures the cover hypothesis. Write DD for its whole-fibre algebra, ei∈De_i\in D for the overlap idempotents, EQE_Q for a transverse Gauss algebra, and cQc_Q for the common positive lattice index. The coefficient of CjC_j on this cell is 1cQdim⁡HQ(∏i<j(1−ei))(Mj⊗O(Vj)EQ)≥0.\frac{1}{c_Q}\dim_{H_Q} \left(\prod_{i<j}(1-e_i)\right) \bigl(M_j\otimes_{\mathcal O(V_j)}E_Q\bigr) \geq0. Indeed, expansion of the product gives exactly the alternating sum of the idempotent-restricted module dimensions in CjC_j. This retains all nilpotent lengths. The coefficient is already an integer, being an alternating sum of the integer module weights. Thus CjC_j is nonnegative on every dd-cell. If its images have dimension less than dd, the chain is zero and the assertion is immediate. The nonnegative cell integral on each cell makes its integral against αj\alpha_j nonnegative. Summing over jj proves positivity.

The coherent-module formulation and compatible-module formulation coincide by Lemma 5.50.4. When XX itself is affinoid, its defining expression is the one in Theorem 5.49.4, so the affinoid comparison follows.

Finally consider QQ in the second assertion. Each of its finitely many affinoid pieces is compact and closed in the Hausdorff space YY. Their finite cover is a quasi-net by Lemma 2.39.2, so the analytic-domain definition 2.40.4 gives QQ its actual analytic-domain structure. It is compact. By the transitivity in Proposition 2.40.5, every affinoid domain of QQ is also an affinoid domain of YY. Its finite intersections are therefore affinoid by Lemma 5.50.1 applied to YY, and every nonempty one is pure of dimension dd by Lemma 5.50.3 applied to YY.

A coherent sheaf on QGQ_G restricts to finite modules on these affinoids by Theorem 2.39.6. Conversely, compatible finite modules glue on QGQ_G by the same section-family construction as in Lemma 5.50.4: for each analytic T⊆QT\subseteq Q, take compatible sections on T∩ViT\cap V_i. The finite cover is admissible, so the sheaf axiom and coherence of its chart restrictions give the coherent glued sheaf, with the stated affinoid restriction tensors. Thus all modules, intersections and finite cover comparisons in the preceding proof are available on QQ. Each invocation of the affinoid cover formula occurs on a single affinoid intersection, and the complementary-idempotent positivity calculation occurs on a single affinoid VjV_j. Neither uses goodness of QQ. Those arguments prove precisely the same independence, exact-sequence additivity and positivity on QQ, as claimed. ◻

5.51 Finite pushforward for compact chart integrals

Theorem 5.51.1. NAG-002419 (Finite pushforward on compact analytic domains). Let p:Y→Xp:Y\to X be a norm-admissible finite morphism of good Hausdorff analytic spaces pure of the same analytic dimension dd, over a complete field with its specified real-valued absolute value. Let Q⊆XQ\subseteq X be a finite union of affinoid domains, and put R=p−1(Q)R=p^{-1}(Q). Let F\mathcal F be coherent on YGY_G. Then RR is a finite union of affinoid domains of YY, and both QQ and RR are compact analytic domains. Neither is required to be good.

Pulling back the unit tuples defines a pullback of finite moment-chart data on QQ to such data on RR. For every top-degree datum α\alpha, IR,F∣R(p∗α)=IQ,(p∗F)∣Q(α).\begin{equation} \mathcal I_{R,\mathcal F|_R}(p^*\alpha) =\mathcal I_{Q,(p_*\mathcal F)|_Q}(\alpha). \end{equation} Successive applications agree for composable finite morphisms between spaces of the same pure analytic dimension. If (p∗F)∣Q(p_*\mathcal F)|_Q is locally free of constant rank ρ\rho, then IR,F∣R(p∗α)=ρ IQ,OQ(α).\mathcal I_{R,\mathcal F|_R}(p^*\alpha) =\rho\,\mathcal I_{Q,\mathcal O_Q}(\alpha). All conclusions allow trivial valuation, arbitrary positive affinoid radii, and nonreduced spaces and modules.

Proof. Choose a finite affinoid-domain cover (Vi)(V_i) of QQ. Each ViV_i is an affinoid domain of XX by the domain transitivity of Proposition 2.40.5. Finiteness of pp makes Wi=p−1(Vi)W_i=p^{-1}(V_i) an actual affinoid domain of YY, whose algebra is a finite canonical module over O(Vi)\mathcal O(V_i). These domains cover RR. The finite-union assertion of Theorem 5.50.6 therefore gives the indicated compact analytic-domain structures. By Lemma 5.50.1, every finite intersection VIV_I is affinoid, and WI=⋂i∈IWi=p−1(VI)W_I=\bigcap_{i\in I}W_i=p^{-1}(V_I) is its full finite affinoid inverse image. Every nonempty such domain is pure of dimension dd by Lemma 5.50.3.

Suppose that (fi,αi)(f_i,\alpha_i) presents the given chart data on the ViV_i. Its proposed pullback has tuple p∗fip^*f_i and the same ambient superform αi\alpha_i on WiW_i. Units remain units. On Wi∩WjW_i\cap W_j the actual joint moment image is contained in the joint moment image on Vi∩VjV_i\cap V_j. Both are finite polyhedral sets by Theorem 5.39.3. Refine their complexes by their finitely many defining hyperplanes, using Lemma 5.14.2. Each cell of the source image lies in a cell of the target image, and its tangent space is a subspace of that cell’s tangent space. Tangential equality on the target consequently restricts to tangential equality on the source. This proves compatibility of the pulled-back data in every bidegree. The same argument applies to a common refinement of two representatives. Inverse images of its domains are actual finite affinoid domains, and pullback commutes with adjoining units. Thus the pullback depends only on the equivalence class of the data.

Proposition 2.35.15 makes p∗Fp_*\mathcal F coherent and identifies, for each nonempty II, Γ(VI,p∗F)=Γ(WI,F).\Gamma(V_I,p_*\mathcal F)=\Gamma(W_I,\mathcal F). Write AI=O(VI)A_I=\mathcal O(V_I), BI=O(WI)B_I=\mathcal O(W_I) and MI=Γ(WI,F)M_I=\Gamma(W_I,\mathcal F). The module MIM_I is finite over BIB_I by Theorem 2.39.6, and finite over AIA_I because BIB_I is finite over AIA_I. The cited direct-image theorem gives precisely this restriction of scalars, with its compatible canonical norms and domain restrictions.

Assume first that WIW_I is nonempty and put f=fi∣VIf=f_i|_{V_I}. The two tuples are ff and its literal pullback p∗fp^*f on the full inverse domain. The character identity therefore holds with the identity map on the moment lattice.

By Theorem 5.39.3, both actual moment images are finite polyhedral sets. The absolute boundary image of VIV_I has dimension at most d−1d-1 and is empty if d=0d=0. Use Lemma 5.14.2 to subdivide the two images and that boundary image together. At a relative-interior point ww of a target dd-cell, choose a small convex open neighbourhood in moment space disjoint from the boundary image and the other cells. The target image is presented there by the given cell. If the source image contains ww, it contains that cell and has the same local presentation. The whole target fibre is absolutely interior there. Finiteness gives relative interior everywhere on the actual source affinoid by Theorem 2.33.18, so Proposition 2.33.21 gives absolute interior of the whole source fibre there. With these pure local presentations, Theorem 5.42.2 compares the same finite module MIM_I over BIB_I and AIA_I. The index on the common saturated tangent lattice is one.

If the source image misses ww, its compactness lets us avoid it on a neighbourhood of ww. For the entire target simultaneous level algebra DwD_w, the inverse level algebra BI⊗AIDwB_I\otimes_{A_I}D_w is the actual finite pullback by Theorems 2.33.19 and 2.36.9. It is zero by Theorem 1.20.1, since its spectrum is empty. Thus MI⊗AIDw=0M_I\otimes_{A_I}D_w=0, and further tensoring with the transverse Gauss field gives zero. The target module weight is zero by Theorem 5.40.2, in agreement with the empty source coefficient.

These generic coefficients determine the closed finite-module chains on the common subdivision. Pairing the same ambient form with these equal chains by the finite-cell integral gives IMI,p∗fi∣WI(αi)=IMI,fi∣VI(αi),\mathcal I_{M_I,p^*f_i|_{W_I}}(\alpha_i) =\mathcal I_{M_I,f_i|_{V_I}}(\alpha_i), where the right side regards MIM_I as an AIA_I-module. An empty inverse image has zero module on both sides. Multiply by (−1)∣I∣+1(-1)^{|I|+1} and sum over the nonempty index sets. The two finite sums are exactly the compact chart integrals in Theorem 5.50.6, including its non-good-domain variant. This proves (5.51.1).

For a further finite map q:Z→Yq:Z\to Y, the inverse image of every WiW_i is again the full finite affinoid inverse image. Apply the same equality first to qq and then to pp. Pullback of unit functions is associative, and p∗(q∗G)=(p∘q)∗Gp_*(q_*\mathcal G)=(p\circ q)_*\mathcal G on sections of every analytic domain, since both sides take sections on its full inverse image. These identities prove the stated compatibility for successive applications.

Finally suppose the indicated direct image is locally free of constant rank ρ\rho on QGQ_G. Refine its local trivializing cover by affinoid domains and intersect them with the presenting chart domains. The quasi-net and compactness assertions of Lemma 2.39.2 give a finite such refinement. The compact integral is unchanged by this refinement. On every new VIV_I its finite module NIN_I is free of rank ρ\rho. At each minimal prime p\mathfrak p of AIA_I, length⁡((NI)p)=ρlength⁡((AI)p)\operatorname{length}((N_I)_{\mathfrak p}) =\rho\operatorname{length}((A_I)_{\mathfrak p}). The exact module-weight formula of Theorem 5.40.2 therefore multiplies every top-dimensional global-chain coefficient by ρ\rho. Lower-dimensional cells give no top-dimensional term. Applying the finite-cell integral to the scaled finite-module chain gives INI,fi(αi)=ρ IAI,fi(αi).\mathcal I_{N_I,f_i}(\alpha_i) =\rho\,\mathcal I_{A_I,f_i}(\alpha_i). The same finite cover sum proves the last formula. In particular the argument preserves all generic Artinian lengths before multiplying by ρ\rho. ◻

5.52 Smooth moment germs and compactly supported integration

The complete field and its specified absolute value remain fixed. We use the real superform conventions of Definition 5.46.1.

Lemma 5.52.1. NAG-002422 (Tangential equality on a finite affinoid cover). Let VV be a generalized affinoid space, let ff be a finite tuple of analytic units, and let V=⋃i=1sViV=\bigcup_{i=1}^s V_i be a finite affinoid-domain cover. A smooth ambient superform η\eta near tf(V)t_f(V) has zero tangential restriction on every face of tf(V)t_f(V) if and only if the same holds on each tf(Vi)t_f(V_i).

Tangential equality on a finite polyhedral set is preserved by d′d', d′′d'', JJ, and multiplication by another superform. These assertions hold in every bidegree and in every dimension.

Proof. All the actual images are finite polyhedral by Theorem 5.39.3, and tf(V)=⋃itf(Vi)t_f(V)=\bigcup_i t_f(V_i) as point sets. Cut their complexes by the finitely many supporting hyperplanes of all their cells, as in Lemma 5.14.2. This makes every subimage a subcomplex of the resulting complex on the union. Each cell of that union is contained in one subimage: a point of its relative interior belongs to a cell of one subcomplex, and the common face property then puts the entire cell in that subcomplex. Thus vanishing on the subimages implies vanishing on every cell of the common refinement. Independence of subdivision, proved in Theorem 5.49.2, gives vanishing on the original image. The converse is restriction to a polyhedral subset, as in that theorem.

Fix a cell and choose affine coordinates on its span. Restriction of both cotangent factors commutes with differentiation along that span. If a tangential coefficient vanishes on the relative interior, all its derivatives in tangent directions vanish there. Their continuity extends the assertion to its faces. This proves the claims for d′d' and d′′d''. The operations JJ and wedge product commute directly with restriction of alternating covectors. Applying these observations to the difference of two forms proves preservation of tangential equality. ◻

Definition 5.52.2. (Ordinary smooth moment germs). Let XX be a good Hausdorff analytic space and let x∈Xx\in X. A smooth moment representative of bidegree (p,q)(p,q) at xx is a triple (V,f,α)(V,f,\alpha), where VV is an ordinary affinoid neighbourhood of xx, ff is a finite tuple of analytic units on VV, and α\alpha is a smooth real (p,q)(p,q)-superform on an open neighbourhood of tf(V)t_f(V). Two representatives are equivalent if, on some ordinary affinoid neighbourhood W⊆V∩V′W\subseteq V\cap V' of xx, their pullbacks to the actual joint image for (f,f′)(f,f') have equal tangential restrictions on every face.

For an ordinary open set U⊆XU\subseteq X, write AXp,q(U)\mathcal A_X^{p,q}(U) for the families of such germs at the points of UU which are locally represented by one smooth moment representative. Here locally means on the ordinary interior of an affinoid neighbourhood contained in UU. The support of a section is the complement of the open set on which all its germs are zero. A top-degree section on a pure-dimensional space is positive if locally it has representatives with nonnegative tangential top coefficients.

On a finite union of affinoid domains, finite moment-chart data mean a finite affinoid cover with unit tuples and smooth ambient superforms whose joint tangential pullbacks agree on every overlap, identified under common finite refinement. This uses the same presentation convention as Definition 5.49.3 in arbitrary dimension; purity is imposed when integrating top-degree forms.

Theorem 5.52.3. NAG-002424 (The sheaf of ordinary smooth moment forms). The relation in Definition 5.52.2 is an equivalence relation. The spaces AXp,q\mathcal A_X^{p,q} form sheaves on the ordinary topology, with well-defined wedge product, JJ, d′d' and d′′d''. These operations satisfy the usual exterior-algebra rules, (d′)2=(d′′)2=0,d′d′′+d′′d′=0.(d')^2=(d'')^2=0,\qquad d'd''+d''d'=0. Every analytic morphism between good Hausdorff spaces over the specified field induces a functorial pullback on these sheaves, commuting with all these operations.

For any affinoid domain V⊆UV\subseteq U, a section on UU determines finite moment-chart data on VV, canonically up to common refinement and tangential equality. This assignment commutes with further affinoid restriction and all the displayed operations. If the section has zero germ at every point of VV, the induced finite chart data are zero. The same assertions hold on a finite union of affinoid domains contained in UU.

Proof. We first record the neighbourhood choices used below. In an affinoid spectrum its evaluation topology has a basis of finitely many strict bounds on absolute values. At a point satisfying such bounds, replace each upper bound by a slightly smaller positive real bound still strict at the point, and each positive lower bound by a slightly larger one still strict there. Their weak inequalities define an actual affinoid domain by Theorem 2.31.2 and Corollary 2.36.10. It contains an ordinary neighbourhood of the point and is contained in the prescribed open set. Goodness and domain transitivity 2.40.5 therefore give a basis of ordinary affinoid neighbourhoods in XX. Intersections of two such neighbourhoods are affinoid by Lemma 5.50.1 and remain ordinary neighbourhoods of the common point. No choice of a ground-field element with a prescribed real absolute value is needed for these inequalities.

Reflexivity and symmetry of the relation are immediate. For transitivity, restrict three representatives to a common smaller ordinary affinoid neighbourhood and adjoin all three tuples. The two given equalities pull back to that joint image and imply equality of the first and third forms there. Faithfulness under adding coordinates, Theorem 5.49.2, descends this equality to their two-tuple joint image. This proves transitivity. The same argument shows that one may shrink a representative or add unit coordinates without changing its germ.

Sums and wedge products are computed on a common smaller neighbourhood with the joint tuple. The preceding faithfulness and transitivity make the result independent of the representatives. Apply JJ, d′d' and d′′d'' to the ambient forms. Lemma 5.52.1 shows that these operations preserve all the germ relations, so they descend. Their algebraic identities can be checked on the coefficients and constant one-form generators of a representative. Equality of mixed partial derivatives follows from the real calculus proved in Lemma 5.46.2: the two iterated fundamental-theorem expressions for a rectangular second difference are equal; dividing by the two side lengths and letting them tend to zero identifies the continuous mixed derivatives. With the anticommuting one-form generators, the stated differential identities and the graded product rules follow.

The sheaf axiom is pointwise. Families agreeing on overlaps give one family of germs, and its local representatives are the given local representatives. This proves both existence and uniqueness of gluing. All the operations commute with restriction, as their representative constructions do.

Let h:Y→Xh:Y\to X be such a morphism and let y∈Yy\in Y. Choose a representative (V,f,α)(V,f,\alpha) at h(y)h(y) and an ordinary affinoid neighbourhood WW of yy mapping into VV. Continuity of hh and the affinoid basis give WW. The analytic-domain universal property 2.40.8 makes the restriction an actual morphism W→VW\to V. Use (W,h∗f,α)(W,h^*f,\alpha) as the pulled-back representative. Its actual moment image is contained in tf(V)t_f(V). Both sets are finite polyhedral by Theorem 5.39.3. Tangential vanishing restricts to any finite polyhedral subset: a common subdivision puts each of its cells inside a cell of the larger set, where restriction of alternating covectors preserves zero. Thus an equality on a smaller target germ gives an equality on a smaller source germ. This proves well-defined pullback. Composition is checked on the pulled-back unit functions, and the ambient form is unchanged in these coordinates. Pullback consequently commutes with the displayed operations and is functorial.

Now let V⊆UV\subseteq U be affinoid and let α\alpha be a section on UU. At every point of VV choose a representing neighbourhood WW and then an affinoid neighbourhood W′W' contained in the ordinary interior of WW. The preceding basis construction permits this choice. Finitely many of the ordinary interiors of these W′W' cover the compact set VV, by Proposition 2.24.1. The intersections V∩W′V\cap W' are affinoid domains by Lemma 5.50.1. Restrict the chosen tuples and ambient forms to them.

These representatives agree on every entire overlap. Indeed, each point of an overlap lies in the ordinary interiors of the two original representing neighbourhoods WW, so equality of its germs supplies a smaller ordinary affinoid neighbourhood on which the joint forms agree. Finitely many such neighbourhoods cover the compact overlap by their interiors. Intersect them with the overlap and apply Lemma 5.52.1 to the fixed joint tuple there. It upgrades the local equalities to tangential equality on the whole joint image of the overlap. Thus the restricted representatives are finite moment-chart data on VV.

The same argument applied to two choices shows that they agree on the finite common refinement by all pairwise intersections. It proves independence and compatibility with further restriction. The operations are computed on these same refinements, so commute with the assignment. If every germ on VV is zero, apply the overlap argument to each presenting chart and the zero representative. A finite cover suffices by compactness, and Lemma 5.52.1 makes each entire chart representative tangentially zero.

For a finite union of affinoid domains, perform the construction on its finitely many members. The same compact-overlap argument proves compatibility and independence on their union. This also proves the zero-germ assertion there. Its analytic-domain structure is supplied by the finite closed-cover quasi-net in Lemma 2.39.2 and the analytic-domain definition 2.40.4. ◻

Theorem 5.52.4. NAG-002425 (Integration of compactly supported smooth moment forms). Let XX be good, Hausdorff and pure of analytic dimension dd, and let F\mathcal F be a coherent OXG\mathcal O_{X_G}-module. There is a canonical linear functional ∫X,F:AX,cd,d(X)⟶R\int_{X,\mathcal F}:\mathcal A_{X,c}^{d,d}(X) \longrightarrow\mathbb R on sections of compact support. It is nonnegative on positive sections and additive in short exact sequences of coherent modules.

Choose finitely many affinoid domains whose ordinary interiors cover the support, and let QQ be their union. The functional is the finite-chart integral of the induced data on QQ and F∣QG\mathcal F|_{Q_G} from Theorem 5.50.6. It is independent of this compact neighbourhood and of all the presenting charts. It agrees with the earlier affinoid and compact-space integrals whenever those are applied to the induced data.

Proof. Put S=supp⁡(α)S=\operatorname{supp}(\alpha). The zero section has integral zero, so first suppose SS is nonempty. The affinoid-neighbourhood basis in the proof of Theorem 5.52.3 and compactness of SS give the required finite family. Its union QQ is a compact analytic domain by the finite-union assertion in Theorem 5.50.6. Its ordinary interior in XX contains SS. Theorem 5.52.3 supplies canonical finite chart data on QQ. The restriction of a coherent sheaf is coherent, as is seen from its finite-module presentations on affinoid domains and Theorem 2.39.6. The compact-domain assertion of Theorem 5.50.6 therefore defines a real number, even when QQ is not good.

We prove independence of QQ. Let Q1,Q2Q_1,Q_2 be two choices and put O=Int⁡top(Q1,X)∩Int⁡top(Q2,X),Z=X∖S.O=\operatorname{Int}_{\rm top}(Q_1,X) \cap\operatorname{Int}_{\rm top}(Q_2,X), \qquad Z=X\setminus S. Both sets are open, S⊆OS\subseteq O, and O∪Z=XO\cup Z=X. Cover the compact set Q1∪Q2Q_1\cup Q_2 by the ordinary interiors of finitely many affinoid domains WlW_l, each contained either in OO or in ZZ. Goodness and the affinoid basis justify this choice. Mark as central the members contained in OO, and mark the remaining members as zero members. Every central member belongs to both Q1Q_1 and Q2Q_2.

Fix finite affinoid presentations (Vi(a))(V_i^{(a)}) of QaQ_a, for a=1,2a=1,2. A finite affinoid cover of QaQ_a is given by all the central WlW_l, together with the domains Vi(a)∩WlV_i^{(a)}\cap W_l for the zero members. These are actual affinoid intersections by Lemma 5.50.1 applied in XX, and they cover QaQ_a since the WlW_l cover Q1∪Q2Q_1\cup Q_2. The section has zero germ at every point of a zero member. The zero-germ assertion of Theorem 5.52.3 makes its induced finite chart data zero on each zero piece and on every further intersection involving such a piece.

Use these covers to compute the two integrals by the finite inclusion-exclusion formula of Theorem 5.50.6. If a covering member needs several charts, first replace it by its finite presenting chart cover; the independence statement of that theorem allows this refinement. Every term involving a zero piece is zero. The surviving terms involve only the same central members, with the same restricted coherent module and the same section. Choose their finite presenting charts once for both covers. The two finite sums are then identical. This proves independence of the compact neighbourhood. The same argument with no central members proves zero when the support is empty.

Independence of the representatives on a fixed QQ is Theorem 5.50.6. For two compactly supported sections, choose one finite affinoid neighbourhood of the union of their supports. Its finite-chart integral is linear, which proves linearity of the functional. If α\alpha is positive, compactness permits the finite presenting charts on QQ to be chosen from positive local representatives and zero representatives outside its support. Positivity of that same theorem gives the claimed inequality. On every affinoid intersection an exact sequence of coherent modules has an exact sequence of finite section modules by Theorem 2.39.6. Additivity of the compact-domain integral therefore proves the asserted coherent-module additivity.

The comparisons with the earlier integrals follow by using the same finite affinoid cover and the same inclusion-exclusion expression. In the compact case one may compare any support neighbourhood with XX itself by the same central-and-zero cover argument. ◻

5.53 Smooth cutoffs on good analytic spaces

Lemma 5.53.1. NAG-002427 (Compact smooth cutoffs). Let XX be a good Hausdorff analytic space over the specified complete field. The sheaf AX0,0\mathcal A_X^{0,0} identifies with a subsheaf of continuous real-valued functions. For every point xx and ordinary open neighbourhood O⊆XO\subseteq X, there is a section χ∈AX0,0(X)\chi\in\mathcal A_X^{0,0}(X) such that 0≤χ≤1,χ=1 on an ordinary neighbourhood of x,supp⁡(χ) is compact in O.\begin{gathered} 0\leq\chi\leq1,\qquad \chi=1\ \hbox{on an ordinary neighbourhood of }x,\\ \operatorname{supp}(\chi)\ \hbox{is compact in }O. \end{gathered} The assertion allows a trivially valued ground field.

Proof. A degree-zero representative evaluates to the continuous function α∘tf\alpha\circ t_f. Two representatives have equal functions near a point exactly when their joint degree-zero form vanishes at every point of the actual joint image on a smaller ordinary neighbourhood. This is precisely the germ relation of Definition 5.52.2. Thus evaluation is injective on germs and identifies the indicated sheaf with a subsheaf of continuous functions. It is closed under composition with a smooth real function: in a representative, compose that function with the smooth ambient coefficient.

We use a smooth real profile ρ\rho with values in [0,1][0,1], equal to zero on (−∞,0](-\infty,0] and to one on [1,∞)[1,\infty). Here is an elementary construction. The power series eu=∑n≥0un/n!e^u=\sum_{n\geq0}u^n/n! and all its termwise derivatives converge uniformly on each bounded interval: after any fixed number of derivatives, their tails are bounded by a constant times a geometric series, once the index exceeds twice the interval bound. Applying the fundamental theorem from Lemma 5.46.2 to the partial sums proves termwise differentiation. The derivative of eue−ue^u e^{-u} is zero, so this product equals its value 11 at zero. Hence e−u=1/eue^{-u}=1/e^u; positivity follows for u≥0u\geq0 from the series and for u<0u<0 from this reciprocal identity. For t>0t>0 put E(t)=e−1/tE(t)=e^{-1/t}, and put E(t)=0E(t)=0 for t≤0t\leq0. Each right derivative is a polynomial in 1/t1/t times e−1/te^{-1/t}. Every such expression tends to zero at zero, since eu≥um+1/(m+1)!e^u\geq u^{m+1}/(m+1)! for u>0u>0. Induction with difference quotients therefore proves that EE is smooth with every derivative zero at zero. The function ρ(t)=E(t)E(t)+E(1−t)\rho(t)=\frac{E(t)}{E(t)+E(1-t)} has the required properties; its denominator is positive.

Choose an ordinary affinoid neighbourhood W=M(A)W=\mathcal M(A) of xx. In its evaluation topology, choose finitely many strict upper and positive lower bounds on absolute values of functions aj∈Aa_j\in A whose intersection contains xx and is contained in O∩Int⁡top(W,X)O\cap\operatorname{Int}_{\rm top}(W,X). The neighbourhood basis used in Theorem 5.52.3 permits this choice. For each upper condition ∣aj∣<bj|a_j|<b_j, choose ∣aj(x)∣<bj,0<bj,1<bj.|a_j(x)|<b_{j,0}<b_{j,1}<b_j. Use the factor 1−ρ(log⁡∣aj∣−log⁡bj,0log⁡bj,1−log⁡bj,0),1-\rho\left( \frac{\log|a_j|-\log b_{j,0}} {\log b_{j,1}-\log b_{j,0}}\right), with value one where aj=0a_j=0. For each lower condition ∣aj∣>cj>0|a_j|>c_j>0, choose cj<cj,0<cj,1<∣aj(x)∣c_j<c_{j,0}<c_{j,1}<|a_j(x)| and use the factor ρ(log⁡∣aj∣−log⁡cj,0log⁡cj,1−log⁡cj,0),\rho\left( \frac{\log|a_j|-\log c_{j,0}} {\log c_{j,1}-\log c_{j,0}}\right), with value zero where aj=0a_j=0.

Each factor is a smooth moment section on WW. At a zero of aja_j it is constant on an ordinary neighbourhood, hence has a constant representative. At a point where aja_j is nonzero, choose a smaller ordinary affinoid neighbourhood where ∣aj∣|a_j| is bounded below by a positive real constant. The unit criterion 1.20.1 makes its restriction an analytic unit. The displayed expression is then a smooth function of the moment coordinate −log⁡∣aj∣-\log|a_j|. These representatives agree as continuous functions, and hence as germs by the first paragraph. The sheaf theorem 5.52.3 glues them.

Multiply the finitely many factors. The resulting χW\chi_W lies between zero and one and equals one on an ordinary neighbourhood of xx. Its support is contained in the subset C⊆WC\subseteq W defined by the weak upper bounds ∣aj∣≤bj,1|a_j|\leq b_{j,1} and weak lower bounds ∣aj∣≥cj,0|a_j|\geq c_{j,0}. This subset is compact by Proposition 2.24.1, is closed in the Hausdorff space XX, and is contained in O∩Int⁡top(W,X)O\cap\operatorname{Int}_{\rm top}(W,X) by the strictly nested choices of bounds. On the open cover X=Int⁡top(W,X)∪(X∖C)X=\operatorname{Int}_{\rm top}(W,X) \cup(X\setminus C) glue χW\chi_W on the first member and zero on the second. They agree on the overlap. The resulting section has all the stated properties. All thresholds were positive real numbers; the construction never required them to be absolute values of ground-field constants. ◻

Theorem 5.53.2. NAG-002428 (Finite smooth partitions near a compact set). Let XX be good and Hausdorff, let S⊆XS\subseteq X be compact, and let O1,…,OmO_1,\ldots,O_m be ordinary open sets covering SS. There are compactly supported sections χi∈AX0,0(X)\chi_i\in\mathcal A_X^{0,0}(X) such that 0≤χi≤1,supp⁡(χi)⊆Oi,∑iχi=1on an ordinary neighbourhood of S.\begin{gathered} 0\leq\chi_i\leq1,\qquad \operatorname{supp}(\chi_i)\subseteq O_i,\\ \sum_i\chi_i=1 \quad\hbox{on an ordinary neighbourhood of }S. \end{gathered} If XX is pure of analytic dimension dd, F\mathcal F is coherent, and α∈AX,cd,d(X)\alpha\in\mathcal A_{X,c}^{d,d}(X) has support in SS, then ∫X,Fα=∑i∫Oi,F∣(Oi)G(χiα)∣Oi.\int_{X,\mathcal F}\alpha =\sum_i\int_{O_i,\mathcal F|_{(O_i)_G}} (\chi_i\alpha)|_{O_i}. Each term is a compactly supported integral on the indicated open analytic subspace.

Proof. The case S=∅S=\varnothing uses zero functions. Otherwise, apply Lemma 5.53.1 at each point of SS, with one OiO_i containing it. Compactness gives finitely many resulting functions bjb_j whose loci where they equal one cover SS by ordinary neighbourhoods. Group them by the chosen index and put bi′=∑j assigned to ibj,s=∑ibi′.b_i'=\sum_{j\text{ assigned to }i}b_j, \qquad s=\sum_i b_i'. Thus s≥1s\geq1 on SS and each bi′b_i' has compact support contained in OiO_i. Choose a smooth real function η\eta with 0≤η≤10\leq\eta\leq1, equal to zero for t≤1/4t\leq1/4 and to one for t≥1/2t\geq1/2, by the profile construction in that lemma. The function h(t)={η(t)/t,t>0,0,t≤0h(t)=\begin{cases}\eta(t)/t,&t>0,\\0,&t\leq0\end{cases} is smooth, since it is identically zero for t≤1/4t\leq1/4. Set χi=h(s)bi′\chi_i=h(s)b_i'. Smooth composition and the sheaf algebra of Theorem 5.52.3 make these global smooth moment sections. Since 0≤bi′≤s0\leq b_i'\leq s, they take values in [0,1][0,1], and their supports are contained in those of bi′b_i'. Their sum is η(s)\eta(s), which equals one on the open neighbourhood {s>1/2}\{s>1/2\} of SS. This proves the partition assertion.

The equality α=∑iχiα\alpha=\sum_i\chi_i\alpha holds as sections: it holds on a neighbourhood of its support, and every germ of α\alpha off that support is zero. The support of χiα\chi_i\alpha is compact and lies in OiO_i. By the affinoid-neighbourhood basis, one may choose its finite compact affinoid neighbourhood entirely inside OiO_i. The same finite-chart formula then computes its integral in XX and in OiO_i, by Theorem 5.52.4. Open subspaces are good by that same neighbourhood basis and Hausdorff by restriction; their nonempty affinoid charts have dimension dd by Lemma 5.50.3. Thus all the indicated integrals exist, and linearity gives the displayed formula. ◻

5.54 Finite projection for compactly supported smooth forms

Theorem 5.54.1. NAG-002430 (Finite projection for smooth moment forms). Let p:Y→Xp:Y\to X be a norm-admissible finite morphism of good Hausdorff analytic spaces pure of the same analytic dimension dd, over the specified complete field. Let F\mathcal F be coherent on YGY_G. For every α∈AX,cd,d(X)\alpha\in\mathcal A_{X,c}^{d,d}(X), the form p∗αp^*\alpha has compact support and ∫Y,Fp∗α=∫X,p∗Fα.\int_{Y,\mathcal F}p^*\alpha =\int_{X,p_*\mathcal F}\alpha. These identities are compatible with composition of finite morphisms. If p∗Fp_*\mathcal F is locally free of constant rank ρ\rho on the analytic G-site over an ordinary open neighbourhood of supp⁡(α)\operatorname{supp}(\alpha), then ∫Y,Fp∗α=ρ∫X,OXα.\int_{Y,\mathcal F}p^*\alpha =\rho\int_{X,\mathcal O_X}\alpha. The absolute value is unchanged, and the assertions allow trivial valuation and nonreduced spaces.

Proof. Put S=supp⁡(α)S=\operatorname{supp}(\alpha). The ordinary smooth-germ pullback is defined by Theorem 5.52.3. If y∉p−1(S)y\notin p^{-1}(S), a representative of α\alpha vanishes on an ordinary neighbourhood of p(y)p(y). Its pullback vanishes near yy, by continuity of pp and the definition of pullback on representatives. Thus supp⁡(p∗α)⊆p−1(S).\operatorname{supp}(p^*\alpha) \subseteq p^{-1}(S).

Choose finitely many affinoid domains Vi⊆XV_i\subseteq X whose ordinary interiors cover SS, using the affinoid neighbourhood basis in Theorem 5.52.3. Put Q=⋃iViQ=\bigcup_iV_i. Finiteness of pp makes Wi=p−1(Vi)W_i=p^{-1}(V_i) actual affinoid domains, and R=⋃iWi=p−1(Q)R=\bigcup_iW_i=p^{-1}(Q) is compact by Theorem 5.51.1. The set p−1(S)p^{-1}(S) is closed in RR: SS is compact and therefore closed in the Hausdorff space XX. It is consequently compact. The support of the pullback is closed and contained in this compact set, which proves the first assertion.

For every ii, the open set p−1(Int⁡top(Vi,X))p^{-1}(\operatorname{Int}_{\rm top}(V_i,X)) is contained in WiW_i, and therefore in Int⁡top(Wi,Y)\operatorname{Int}_{\rm top}(W_i,Y). These interiors cover p−1(S)p^{-1}(S) and hence the support of the pullback. Thus QQ and RR are valid compact neighbourhoods for the two integrals in Theorem 5.52.4; neither finite union needs to be good.

The coherent direct image p∗Fp_*\mathcal F exists by Proposition 2.35.15. The finite chart data induced on RR by the ordinary section p∗αp^*\alpha are the pullback of the data induced by α\alpha on QQ. To check this, take the finite presenting affinoid cover supplied by Theorem 5.52.3. On its full inverse images both constructions use the same pulled-back analytic units and the same ambient superform. Their relations agree on every overlap by that theorem’s restriction and pullback compatibility. If further affinoid restriction is needed, both sets of data are compared on that same finite common refinement.

Theorem 5.51.1 now gives IR,F∣R(p∗α)=IQ,(p∗F)∣Q(α).\mathcal I_{R,\mathcal F|_R}(p^*\alpha) =\mathcal I_{Q,(p_*\mathcal F)|_Q}(\alpha). By Theorem 5.52.4, these are precisely the two ordinary compact-support integrals in the statement.

For composable finite maps, functoriality of ordinary pullback is Theorem 5.52.3. Direct images compose on sections over full inverse images, as in Theorem 5.51.1. Applying the proved identity twice therefore gives the identity for the composite.

Finally, under the local rank hypothesis choose all ViV_i inside the indicated ordinary open neighbourhood of SS. The scalar-rank assertion of Theorem 5.51.1 applies on QQ. It multiplies the structure-module integral by ρ\rho, with the generic Artinian lengths retained. Independence of compact neighbourhoods in Theorem 5.52.4 identifies that structure-module integral with the one on XX. ◻

5.55 Graded independence and pointed moment images

Let KK be a complete non-Archimedean field with its specified real-valued absolute value, possibly trivial. For an analytic point xx, write Fx=H(x)~grF_x=\widetilde{\mathcal H(x)}^{\rm gr} and k=K~grk=\widetilde K^{\rm gr}. The initial form of a nonzero a∈H(x)a\in\mathcal H(x) is denoted by a~\widetilde a; it is a homogeneous unit of degree ∣a∣|a|. Algebraic independence of homogeneous elements means ordinary polynomial independence over the underlying graded ring kk. The kernel of polynomial evaluation is graded, so this is equivalent to the absence of nonzero homogeneous polynomial relations. For a graded field generated by finitely many such elements, trdeg⁡k\operatorname{trdeg}_k means the transcendence degree of the fraction fields of the underlying graded domains, or equivalently the size of a maximal algebraically independent subtuple of the homogeneous generators.

Lemma 5.55.1. NAG-002448 (Norm rigidity from a graded algebraic initial form). Let XX be a good KK-analytic space, let x∈Xx\in X, and let ff be an analytic unit on an ordinary neighbourhood of xx. If f(x)~∈Fx\widetilde{f(x)}\in F_x is algebraic over kk, there is an ordinary neighbourhood UU of xx on which ∣f(y)∣=∣f(x)∣(y∈U).|f(y)|=|f(x)|\qquad(y\in U). For a finite tuple of units the same assertion holds simultaneously when all its initial forms are algebraic over kk. No interior or purity hypothesis is required.

Proof. Put r=∣f(x)∣>0r=|f(x)|>0 and u=f(x)~u=\widetilde{f(x)}. Take a nonzero homogeneous relation ∑j∈Jcˉjuj=0\sum_{j\in J}\bar c_j u^j=0 with J⊂Z≥0J\subset\mathbf Z_{\ge0} finite, all cˉj\bar c_j nonzero and homogeneous, and all the terms of the same degree R>0R>0. Such a relation exists because the kernel of evaluation in the graded polynomial ring is graded. It has at least two distinct exponents, since each cˉjuj\bar c_j u^j is nonzero. Lift cˉj\bar c_j to cj∈Kc_j\in K with ∣cj∣rj=R|c_j|r^j=R, and put P(T)=∑j∈JcjTjP(T)=\sum_{j\in J}c_jT^j. Cancellation of the initial forms gives ∣P(f(x))∣<R|P(f(x))|<R.

Choose cc with ∣P(f(x))∣<c<R|P(f(x))|<c<R. By continuity, after shrinking to an ordinary neighbourhood UU of xx, ∣P(f(y))∣<c.|P(f(y))|<c. Write a=min⁡Ja=\min J and b=max⁡Jb=\max J; then a<ba<b. For a positive real qq near rr, the sizes of the terms are ∣cj∣qj=R(q/r)j.|c_j|q^j=R(q/r)^j. If q>rq>r, the unique largest size is the j=bj=b term; if q<rq<r, it is the j=aj=a term. Choose ε>0\varepsilon>0 so that the maximum of these finitely many sizes is strictly larger than cc whenever 0<∣log⁡(q/r)∣<ε0<|\log(q/r)|<\varepsilon. Continuity of ∣f∣|f| allows us to shrink UU further to ∣log⁡(∣f(y)∣/r)∣<ε|\log(|f(y)|/r)|<\varepsilon. If ∣f(y)∣≠r|f(y)|\ne r, the non-Archimedean strong triangle inequality gives ∣P(f(y))∣=max⁡j∈J∣cj∣∣f(y)∣j>c,|P(f(y))|=\max_{j\in J}|c_j||f(y)|^j>c, a contradiction. The finite-tuple assertion follows by intersecting the finitely many resulting neighbourhoods. ◻

Lemma 5.55.2. NAG-002449 (Independent initial forms survive Gauss normalization). Let V=M(A)V=\mathcal M(A) be a generalized KK-affinoid space, x∈Vx\in V, and let f1,…,fs∈A×f_1,\ldots,f_s\in A^\times have algebraically independent initial forms in FxF_x over kk. Put ri=∣fi(x)∣r_i=|f_i(x)|. Give K(S1,…,Ss)K(S_1,\ldots,S_s) the weighted Gauss absolute value with ∣Si∣=ri|S_i|=r_i and write LL for its completion. There exists an analytic point xL∈VLx_L\in V_L above xx with f1/S1~(xL),…,fs/Ss~(xL)\widetilde{f_1/S_1}(x_L),\ldots, \widetilde{f_s/S_s}(x_L) algebraically independent over L~\widetilde L. Moreover x∈Int⁡(V/K)x\in\operatorname{Int}(V/K) implies xL∈Int⁡(VL/L)x_L\in\operatorname{Int}(V_L/L).

Proof. Put H=H(x)H=\mathcal H(x) and let EE be the completion of H(S1,…,Ss)H(S_1,\ldots,S_s) for the same weighted Gauss norm. The given norm on K(S1,…,Ss)K(S_1,\ldots,S_s) is its restriction, so LL embeds isometrically in EE. Evaluation at xx gives a bounded map A→H⊂EA\to H\subset E. Together with L→EL\to E it induces a bounded map A⊗^KL→EA\widehat\otimes_K L\to E and hence a point xLx_L above xx.

Write ui=fi(x)~u_i=\widetilde{f_i(x)} and si=Si~s_i=\widetilde{S_i}. The defining Gauss norm says that the graded Laurent polynomial ring Fx[s1±1,…,ss±1]F_x[s_1^{\pm1},\ldots,s_s^{\pm1}] embeds in E~gr\widetilde E^{\rm gr}: a nonzero finite sum has a maximal Gauss term and its initial form does not vanish. Here polynomial independence holds even when two weighted monomials have the same numerical degree, because the Gauss norm does not cancel distinct monomials. The uiu_i are independent over kk by assumption, while the sis_i are independent over the entire FxF_x. Consequently the ui/siu_i/s_i are algebraically independent over the ordinary fraction field of k[s1±1,…,ss±1]k[s_1^{\pm1},\ldots,s_s^{\pm1}]: any polynomial relation, after clearing denominators and powers of the sis_i, would be a polynomial relation in the uiu_i with coefficients in k[s1±1,…,ss±1]k[s_1^{\pm1},\ldots,s_s^{\pm1}]; independence of the uiu_i, followed by independence of the sis_i, makes each coefficient zero. The degree-one residue field L~\widetilde L is a subfield of this fraction field, as follows directly by reducing fractions for the weighted Gauss norm. Thus the displayed normalized initial forms are independent over L~\widetilde L. The inclusion of the graded residue field of H(xL)\mathcal H(x_L) in that of EE preserves such independence.

Finally, base-change preservation of relative interior in Proposition 2.33.21, applied over the point M(K)\mathcal M(K), gives the last assertion. The argument did not change the originally specified absolute value on KK. ◻

Theorem 5.55.3. NAG-002450 (Local openness for graded-independent moments). Let V=M(A)V=\mathcal M(A) be a generalized KK-affinoid space, let x∈Int⁡(V/K)x\in\operatorname{Int}(V/K), and let f1,…,fs∈A×f_1,\ldots,f_s\in A^\times have algebraically independent homogeneous initial forms over kk in FxF_x. Put tf(y)=(−log⁡∣f1(y)∣,…,−log⁡∣fs(y)∣),w=tf(x).t_f(y)=(-\log|f_1(y)|,\ldots,-\log|f_s(y)|), \qquad w=t_f(x). For every ordinary neighbourhood UU of xx in VV, the set tf(U)t_f(U) contains an ordinary Euclidean neighbourhood of ww in Rs\mathbf R^s. In particular, for every ordinary affinoid neighbourhood WW of xx contained in VV, its actual finite polyhedral moment image has dimension ss. The assertion holds for the trivial valuation, arbitrary real radii and every analytic dimension of VV.

Proof. For s=0s=0 the image is the one-point space R0\mathbf R^0, so there is nothing to prove. Assume s>0s>0. Use Lemma 5.55.2 and write hi=fi/Sih_i=f_i/S_i on VLV_L. The point xLx_L is interior, every ∣hi(xL)∣=1|h_i(x_L)|=1, and the reductions of the hih_i are algebraically independent over L~\widetilde L. Thus the actual residue algebra L~[h~1±1,…,h~s±1]≃L~[T1±1,…,Ts±1]\widetilde L[\widetilde h_1^{\pm1},\ldots, \widetilde h_s^{\pm1}] \simeq\widetilde L[T_1^{\pm1},\ldots,T_s^{\pm1}] is a Laurent polynomial algebra. Its tropicalization over the trivially valued residue field is all of Rs\mathbf R^s: for any real vector vv, the weighted Gauss seminorm on its Laurent ring has moment vv.

Let CC be a nonempty open rational homogeneous polyhedral cone in Rs\mathbf R^s. Choose a vector v∈Cv\in C and integer vectors m1,…,mam_1,\ldots,m_a such that the strict inequalities ⟨mj,z⟩>0\langle m_j,z\rangle>0 define a nonempty smaller cone contained in CC and contain vv. The normalized sector argument applied to (VL,xL)(V_L,x_L) and (hi)(h_i) says that every ordinary neighbourhood of xLx_L contains a point where ⟨mj,th⟩>0\langle m_j,t_h\rangle>0 for all jj. For any δ>0\delta>0 apply it to the neighbourhood p−1(U)∩th−1(Bδ(0))p^{-1}(U)\cap t_h^{-1}(B_\delta(0)) of xLx_L, where p:VL→Vp:V_L\to V is projection. The projected point belongs to UU and its moment displacement tf(y)−w=th(yL)t_f(y)-w=t_h(y_L) belongs to C∩Bδ(0)C\cap B_\delta(0).

Choose an ordinary affinoid neighbourhood W⊆UW\subseteq U of xx by the affinoid-neighbourhood basis established in the proof of Theorem 5.52.3. Apply the previous paragraph with UU replaced by the ordinary interior of WW. Theorem 5.39.3 gives a finite integral affine polyhedral complex whose support is exactly tf(W)t_f(W). Because w∈tf(W)w\in t_f(W), the cells containing ww determine a finite union CWC_W of closed rational polyhedral tangent cones, and for some ϵ>0\epsilon>0 one has tf(W)∩Bϵ(w)=(w+CW)∩Bϵ(w).t_f(W)\cap B_\epsilon(w) =(w+C_W)\cap B_\epsilon(w). If CW≠RsC_W\ne\mathbf R^s, its closed conical support has a complementary nonempty open rational homogeneous cone. The preceding sector lifting then produces a moment displacement in this cone arbitrarily close to zero, contradicting the displayed equality. Hence CW=RsC_W=\mathbf R^s and the displayed equality puts Bϵ(w)B_\epsilon(w) inside tf(W)⊆tf(U)t_f(W)\subseteq t_f(U). The dimension assertion follows because the moment image is a subset of Rs\mathbf R^s containing a ball. ◻

Corollary 5.55.4. NAG-002451 (The two extreme ranks of pointed tropical germs). Let V,xV,x be as in Theorem 5.55.3, and let f=(f1,…,fn)f=(f_1,\ldots,f_n) be any finite unit tuple. Put s=trdeg⁡K~grK~gr(f1(x)~,…,fn(x)~).s=\operatorname{trdeg}_{\widetilde K^{\rm gr}} \widetilde K^{\rm gr} (\widetilde{f_1(x)},\ldots,\widetilde{f_n(x)}). If s=0s=0, there is an ordinary neighbourhood U0U_0 of xx on which tft_f is constant. If s=ns=n, for every ordinary neighbourhood UU of xx, the germ (tf(U),tf(x))(t_f(U),t_f(x)) is the full Euclidean germ (Rn,tf(x))(\mathbf R^n,t_f(x)). More generally, if ss of the initial forms are algebraically independent, the projection of tf(U)t_f(U) to those ss coordinates contains an ordinary Euclidean neighbourhood of its base point.

Proof. When s=0s=0, each initial form is algebraic over K~gr\widetilde K^{\rm gr}, so apply Lemma 5.55.1 to the finite tuple. When s=ns=n the tuple itself satisfies Theorem 5.55.3. The projection assertion follows by applying the same theorem to the indicated subtuple. ◻

Lemma 5.55.5. NAG-002452 (Finite tropical witnesses for homogeneous graded relations). Let kk be a graded field, let r1,…,rn>0r_1,\ldots,r_n>0, and let u1,…,unu_1,\ldots,u_n be nonzero homogeneous elements of degrees rir_i in a graded field extension F/kF/k. Put I=ker⁡(k[Z1±1,…,Zn±1]⟶F, Zi⟼ui),s=trdeg⁡kk(u1,…,un).I=\ker\bigl(k[Z_1^{\pm1},\ldots,Z_n^{\pm1}] \longrightarrow F,\ Z_i\longmapsto u_i\bigr), \qquad s=\operatorname{trdeg}_k k(u_1,\ldots,u_n). The Laurent polynomial ring is graded by deg⁡Zi=ri\deg Z_i=r_i, and II is homogeneous. Define C⊂RnC\subset\mathbf R^n by requiring that for every nonzero homogeneous relation P=∑mcmZm∈IP=\sum_m c_mZ^m\in I the minimum of ⟨m,q⟩\langle m,q\rangle over its nonzero terms is attained at least twice. Then CC is the support of a finite pure ss-dimensional rational polyhedral fan. There are finitely many nonzero homogeneous relations P1,…,PN∈IP_1,\ldots,P_N\in I whose tropical hypersurfaces already cut out CC. If I=0I=0, the list is empty and s=ns=n.

Proof. Adjoin algebraically independent homogeneous variables SiS_i of degrees rir_i and let ℓ\ell be the graded fraction field of k[S1±1,…,Sn±1]k[S_1^{\pm1},\ldots,S_n^{\pm1}]. The degree-one part ℓ1\ell_1 is an ordinary field. As the degrees rir_i lie in the support of ℓ\ell, replacing ZiZ_i by SiHiS_iH_i identifies the degree-one part of (k[Z±1]/I)⊗kℓ(k[Z^{\pm1}]/I)\otimes_k\ell with a finitely generated ℓ1\ell_1-algebra B=ℓ1[H1±1,…,Hn±1]/J.B=\ell_1[H_1^{\pm1},\ldots,H_n^{\pm1}]/J. The tensor algebra injects into the graded fraction field of F[S1±1,…,Sn±1]F[S_1^{\pm1},\ldots,S_n^{\pm1}]. Indeed polynomial independence of the SiS_i over FF preserves the original kernel II after scalar extension, and localization at nonzero homogeneous elements of k[Si±1]k[S_i^{\pm1}] remains injective. Thus BB is a domain.

Choose a maximal algebraically independent subtuple of the uiu_i over kk, with ss members. The ratios of those members to their corresponding SiS_i are algebraically independent over ℓ1\ell_1: a relation would, after clearing denominators and powers of the SiS_i, contradict joint independence of the chosen uiu_i and all the SiS_i. Every remaining uju_j is algebraic over the graded field generated by the chosen subtuple. Its nonzero homogeneous polynomial equation, after the substitutions ui=SiHiu_i=S_iH_i and division by a homogeneous unit of ℓ\ell, is an ordinary polynomial equation for HjH_j over ℓ1\ell_1 and the chosen HiH_i. Consequently dim⁡B=trdeg⁡ℓ1Frac⁡B=s,\dim B =\operatorname{trdeg}_{\ell_1} \operatorname{Frac}B=s, using the finite-type dimension theorem [StacksAlg] (Tag 00P0), already an authorized leaf in the frozen foundation.

The ordinary field ℓ1\ell_1 is given the trivial absolute value. Apply the finite tropical basis Lemma 5.38.2 to JJ. Write T=Trop⁡(Spec⁡B)T=\operatorname{Trop}(\operatorname{Spec}B). It is a finite rational polyhedral fan; the algebraic tropical pure-dimension theorem 5.19.5, applied to the integral Laurent variety Spec⁡B\operatorname{Spec}B, gives pure dimension ss. A nonzero homogeneous original relation becomes an ordinary Laurent relation after normalization. Therefore T⊆CT\subseteq C. The difficult reverse inclusion requires descending finite witnesses from JJ to II; we establish that inclusion next.

We now descend a finite tropical basis. Choose a finite basis Q1,…,Qa∈JQ_1,\ldots,Q_a\in J from the preceding lemma, and let S⊂ZnS\subset\mathbf Z^n be the finite union of their exponent supports. Put MS=⨁m∈SkZm,NS=I∩MS.M_S=\bigoplus_{m\in S}kZ^m, \qquad N_S=I\cap M_S. These are graded kk-modules, but no assertion about flatness of an infinite coefficient extension is needed. Write si=Sis_i=S_i for these auxiliary graded variables.

Fix q∉Tq\notin T. Some normalized QjQ_j has a unique least qq-weight monomial, say with exponent m0∈Sm_0\in S. Rewrite it in the variables Zi=SiHiZ_i=S_iH_i, obtaining Q=∑m∈Sam(s)Zm∈ℓ[Z±1],Q=\sum_{m\in S}a_m(s)Z^m \in\ell[Z^{\pm1}], where every nonzero am(s)a_m(s) is homogeneous and QQ has one common total graded degree. The coefficient at m0m_0 is nonzero, and all coefficients at other exponents of qq-weight at most ⟨m0,q⟩\langle m_0,q\rangle are zero. Choose a common nonzero homogeneous denominator d(s)∈k[s1±1,…,sn±1]d(s)\in k[s_1^{\pm1},\ldots,s_n^{\pm1}] of its finitely many coefficients. Then d(s)Q=∑a∈AsaPa(Z),Pa∈MS,d(s)Q=\sum_{a\in A}s^aP_a(Z), \qquad P_a\in M_S, for one finite set A⊂ZnA\subset\mathbf Z^n. Every nonzero PaP_a is homogeneous: in a fixed monomial saZms^aZ^m, the total degree of dQdQ determines the degree of its kk-coefficient, and therefore the total degree of the corresponding terms of PaP_a.

The normalized relation is zero after substituting Zi=uiZ_i=u_i in the graded polynomial ring F[si±1]F[s_i^{\pm1}]. The sis_i are algebraically independent over FF, so comparison of their Laurent coefficients gives Pa(u)=0P_a(u)=0 for every aa; hence Pa∈NSP_a\in N_S. The nonzero Laurent polynomial d(s)am0(s)d(s)a_{m_0}(s) has some nonzero coefficient, so one PaP_a has a nonzero m0m_0-coordinate. Since all other coordinates of qq-weight at most ⟨m0,q⟩\langle m_0,q\rangle vanish before multiplication by dd, they vanish in every PaP_a. This chosen homogeneous PaP_a has the single initial monomial cm0Zm0c_{m_0}Z^{m_0}. Thus q∉Cq\notin C, proving C⊆TC\subseteq T and hence C=TC=T.

It remains to replace the resulting weight-dependent choice by a finite list. A circuit of NSN_S is a nonzero homogeneous vector whose exponent support is minimal among supports of nonzero vectors of NSN_S. For each possible support there is at most one circuit up to multiplication by a nonzero homogeneous scalar: if two had that support, eliminate one coordinate by such a scalar, producing a nonzero vector of smaller support unless they were proportional. There are at most 2∣S∣−12^{|S|}-1 supports, so there are finitely many circuits up to proportionality.

Given a homogeneous P∈NSP\in N_S with unique least term m0m_0, choose a homogeneous vector with nonzero m0m_0-coordinate, support contained in supp⁡P\operatorname{supp}P, and support minimal under those two conditions. If some nonzero vector had support properly contained in that support, either it has a nonzero m0m_0-coordinate, contradicting minimality, or it can be scaled to cancel a different coordinate without changing the m0m_0-coordinate, again contradicting minimality. The chosen vector is thus a circuit. Its support lies in supp⁡P\operatorname{supp}P and contains m0m_0, so its initial at qq is still a single monomial. Taking one representative of each circuit gives a finite family of homogeneous original relations, and a vector is outside CC precisely when one of these has a single least term. This proves the assertion. ◻

Lemma 5.55.6. NAG-002453 (The actual residual algebra after free Gauss normalization). Let V=M(A)V=\mathcal M(A), x∈Vx\in V, and f1,…,fn∈A×f_1,\ldots,f_n\in A^\times be arbitrary analytic units. Put ri=∣fi(x)∣r_i=|f_i(x)|, Fx=H(x)~grF_x=\widetilde{\mathcal H(x)}^{\rm gr}, k=K~grk=\widetilde K^{\rm gr}, and ui=fi(x)~u_i=\widetilde{f_i(x)}. Let LL be the completion of K(S1,…,Sn)K(S_1,\ldots,S_n) for the weighted Gauss norm ∣Si∣=ri|S_i|=r_i, without assuming these radii are free over the value group of KK. There is a point xL∈VLx_L\in V_L above xx for which the residue algebra L~[f1/S1~(xL)±1,…,fn/Sn~(xL)±1]\widetilde L[\widetilde{f_1/S_1}(x_L)^{\pm1}, \ldots,\widetilde{f_n/S_n}(x_L)^{\pm1}] is precisely the ordinary degree-one algebra BB constructed in Lemma 5.55.5 from the graded relation ideal of the uiu_i. If x∈Int⁡(V/K)x\in\operatorname{Int}(V/K), then xL∈Int⁡(VL/L)x_L\in\operatorname{Int}(V_L/L).

Proof. Put H=H(x)H=\mathcal H(x) and let EE be the completion of H(S1,…,Sn)H(S_1,\ldots,S_n) for the same weighted Gauss norm. The formula on finite sums gives an isometric embedding L↪EL\hookrightarrow E. It also proves the graded identities L~gr=Frac⁡gr(k[s1±1,…,sn±1]),E~gr=Frac⁡gr(Fx[s1±1,…,sn±1]),\widetilde L^{\rm gr} =\operatorname{Frac}^{\rm gr}(k[s_1^{\pm1},\ldots, s_n^{\pm1}]), \quad \widetilde E^{\rm gr} =\operatorname{Frac}^{\rm gr}(F_x[s_1^{\pm1},\ldots, s_n^{\pm1}]), where si=Si~s_i=\widetilde{S_i} and the fraction fields invert nonzero homogeneous elements. Indeed the maximum Gauss term of a nonzero polynomial has nonzero initial polynomial, distinct Laurent monomials do not cancel, and the initial form of a rational fraction is the quotient of its numerator and denominator initial forms. Passing to the completion adds no homogeneous initial values: approximate every nonzero element by a rational fraction with smaller error.

The bounded evaluation A→HA\to H and the isometric map L→EL\to E produce a bounded map A⊗^KL→EA\widehat\otimes_KL\to E and hence xLx_L above xx. The completed residue field H(xL)\mathcal H(x_L) embeds isometrically into EE. The graded injection Fx⊗kL~gr⟶E~grF_x\otimes_k\widetilde L^{\rm gr} \longrightarrow\widetilde E^{\rm gr} comes from the algebraic independence of the sis_i over FxF_x, followed by homogeneous localization. Consequently the kernel of evaluation of the normalized initials ui/siu_i/s_i over L~=(L~gr)1\widetilde L=(\widetilde L^{\rm gr})_1 is exactly the degree-one normalization of the scalar extension of the original homogeneous relation ideal II. By the definition of BB in Lemma 5.55.5, this is the asserted ordinary residue algebra. The base-change interior inclusion in Proposition 2.33.21 gives the last assertion. ◻

Lemma 5.55.7. NAG-002454 (A graded residual fan bounds a pointed moment germ). Let V=M(A)V=\mathcal M(A) be a generalized KK-affinoid space, x∈Vx\in V, and let f=(f1,…,fn)f=(f_1,\ldots,f_n) be a tuple of analytic units. Put w=tf(x)w=t_f(x), set ui=fi(x)~∈Fxu_i=\widetilde{f_i(x)}\in F_x, and let CxC_x be the graded tropical fan of Lemma 5.55.5. There are an ordinary affinoid neighbourhood W0W_0 of xx and ε>0\varepsilon>0 such that tf(W0)∩Bε(w)⊆(w+Cx)∩Bε(w).t_f(W_0)\cap B_\varepsilon(w) \subseteq(w+C_x)\cap B_\varepsilon(w). This upper bound has pure fan dimension s=trdeg⁡kk(u1,…,un)s=\operatorname{trdeg}_k k(u_1,\ldots,u_n) and requires no interior or purity hypothesis on VV.

Proof. Choose the finite homogeneous relations Pj=∑mcˉjmZmP_j=\sum_m \bar c_{jm}Z^m from Lemma 5.55.5, and lift each nonzero cˉjm\bar c_{jm} to cjm∈Kc_{jm}\in K with its exact absolute value. Because PjP_j is homogeneous, there is an Rj>0R_j>0 such that ∣cjm∣∏i∣fi(x)∣mi=Rj|c_{jm}|\prod_i|f_i(x)|^{m_i}=R_j for every monomial of PjP_j. Put Fj=∑mcjmfm∈AF_j=\sum_mc_{jm}f^m\in A. The graded relation gives ∣Fj(x)∣<Rj|F_j(x)|<R_j. Choose real bjb_j with ∣Fj(x)∣<bj<Rj|F_j(x)|<b_j<R_j. By continuity, all the inequalities ∣Fj(y)∣<bj|F_j(y)|<b_j hold on one ordinary affinoid neighbourhood W0W_0 of xx. Shrink ε>0\varepsilon>0 so that for every v∈Bε(w)v\in B_\varepsilon(w), every monomial of the finite list has size ∣cjm∣exp⁡(−⟨m,v⟩)>bj.|c_{jm}|\exp(-\langle m,v\rangle)>b_j. If v=tf(y)v=t_f(y) does not belong to w+Cxw+C_x, then q=v−w∉Cxq=v-w\notin C_x and some PjP_j has one unique least qq-weight exponent. All terms of Fj(y)F_j(y) have sizes Rjexp⁡(−⟨m,q⟩)R_j\exp(-\langle m,q\rangle); the unique least exponent is therefore the unique largest term, and the strong triangle inequality gives ∣Fj(y)∣>bj|F_j(y)|>b_j. This contradicts the choice of W0W_0. The fan and dimension assertions are precisely those of Lemma 5.55.5. ◻

Theorem 5.55.8. NAG-002455 (Stabilization and purity of pointed tropical germs). Let V=M(A)V=\mathcal M(A) be a generalized KK-affinoid space, x∈Int⁡(V/K)x\in\operatorname{Int}(V/K), and f=(f1,…,fn)f=(f_1,\ldots,f_n) a tuple of analytic units. Put w=tf(x)w=t_f(x) and s=trdeg⁡K~grK~gr(f1(x)~,…,fn(x)~).s=\operatorname{trdeg}_{\widetilde K^{\rm gr}} \widetilde K^{\rm gr} (\widetilde{f_1(x)},\ldots,\widetilde{f_n(x)}). There is a finite pure ss-dimensional rational polyhedral fan Cx⊆RnC_x\subseteq\mathbf R^n and an ordinary affinoid neighbourhood W0W_0 of xx such that for every ordinary neighbourhood UU of xx contained in W0W_0, there is εU>0\varepsilon_U>0 with tf(U)∩BεU(w)=(w+Cx)∩BεU(w).t_f(U)\cap B_{\varepsilon_U}(w) =(w+C_x)\cap B_{\varepsilon_U}(w). Thus the pointed moment-image germ is independent of the choice of sufficiently small ordinary neighbourhood. Its dimension is exactly ss, for every complete real-valued ground field, including the trivial valuation, with the original absolute value fixed. The statement allows nonreduced or nonpure AA.

Proof. Let CxC_x and W0W_0 be supplied by Lemmas 5.55.5 and 5.55.7. The finite upper relation applies to every smaller neighbourhood U⊆W0U\subseteq W_0.

For the opposite inclusion, use Lemma 5.55.6 to make the simultaneous weighted Gauss extension with all nn radii and its specified point xLx_L above xx. That lemma identifies the ordinary residual Laurent algebra generated by the normalized initial forms fi/Si~(xL)\widetilde{f_i/S_i}(x_L) with the algebra BB of Lemma 5.55.5. Its tropicalization is therefore exactly CxC_x, and xLx_L is interior over LL.

Fix a nonzero vector q∈Cxq\in C_x and a strict rational open homogeneous sector SS containing qq. The normalized sector argument, applied to the normalized tuple hi=fi/Sih_i=f_i/S_i at xLx_L, supplies points with tht_h in SS in every ordinary neighbourhood of xLx_L. To require arbitrarily small displacements, intersect this neighbourhood with th−1(Bδ(0))t_h^{-1}(B_\delta(0)) for any δ>0\delta>0. For any prescribed ordinary neighbourhood UU of xx the inverse image of UU under VL→VV_L\to V is an ordinary neighbourhood of xLx_L. Projection then gives points y∈Uy\in U with tf(y)−w∈St_f(y)-w\in S and ∥tf(y)−w∥<δ\|t_f(y)-w\|<\delta.

Choose an ordinary affinoid neighbourhood WW of xx contained in U∩W0U\cap W_0. The actual image tf(W)t_f(W) is a finite integral affine polyhedral set by Theorem 5.39.3. In a sufficiently small Euclidean ball about ww, it equals w+DWw+D_W for the finite closed rational polyhedral tangent fan DWD_W formed by the cells containing ww. By the upper inclusion, DW⊆CxD_W\subseteq C_x. If some nonzero q∈Cxq\in C_x were absent from DWD_W, the finite closed conical set DWD_W would have a strict rational open sector about qq disjoint from it. The preceding residue-sector lifting into WW would contradict the local equality with w+DWw+D_W. Since 0∈DW0\in D_W, this proves DW=CxD_W=C_x. Hence tf(W)t_f(W) agrees near ww with w+Cxw+C_x. It is contained in tf(U)t_f(U), while the upper inclusion for U⊆W0U\subseteq W_0 gives the reverse containment near ww. Shrinking the Euclidean ball once more proves the displayed equality. The finite pure-dimensional fan calculation in Lemma 5.55.5 gives the dimension assertion. Every field extension was isometric and every projected moment used the original norm on KK. ◻

5.56 Proper morphisms and top-dimensional moment pushforward

Throughout this section, KK is a complete non-Archimedean field with its specified real-valued absolute value. The valuation may be trivial. All analytic spaces and completed residue fields use this absolute value. A finite morphism is norm-admissible finite. Compactness and properness have the meanings of Definition 2.53.1. For a unit tuple f=(f1,…,fn)f=(f_1,\ldots,f_n) write tf=(−log⁡∣f1∣,…,−log⁡∣fn∣)t_f=(-\log|f_1|,\ldots,-\log|f_n|). The exact polytorus at w∈Rnw\in\mathbb R^n is the affinoid domain of the analytic torus on which ∣Tj∣=e−wj|T_j|=e^{-w_j} for every jj. Empty spaces have zero moment chain and zero integral.

Lemma 5.56.1. NAG-002457 (A finite neighbourhood of the entire analytic fibre). Let p:Z→Yp:Z\to Y be a compact morphism of good Hausdorff analytic spaces. Suppose that y∈Yy\in Y, that the underlying set of the analytic fibre ZyZ_y is finite, and that every point above yy belongs to Int⁡(Z/Y)\operatorname{Int}(Z/Y). There is an ordinary open neighbourhood NN of yy such that the full restriction p−1(N)⟶Np^{-1}(N)\longrightarrow N is finite. One may require NN to lie in any prescribed ordinary neighbourhood of yy. Empty fibres are allowed.

Consequently, if pp is proper, the locus of target points with finite analytic fibres is ordinary open, and the full restriction of pp to that locus is finite.

Proof. Choose an affinoid neighbourhood VV of yy inside the prescribed target neighbourhood. It is compact by Proposition 2.24.1. Its inverse image ZVZ_V is compact Hausdorff by compactness of pp. The topological identification of Theorem 2.33.11, applied on charts, identifies the points of ZyZ_y with the points of p−1(y)p^{-1}(y).

Suppose first that this fibre consists of z1,…,zrz_1,\ldots,z_r. A finite subset of a Hausdorff space has disjoint ordinary open neighbourhoods of its points: choose separating neighbourhoods for every pair and take the finitely many intersections. Theorem 2.49.1 therefore supplies pairwise disjoint ordinary open neighbourhoods WiW_i of ziz_i, ordinary open neighbourhoods NiN_i of yy, and finite maps Wi→NiW_i\to N_i. The same theorem permits the WiW_i to be contained in the chosen disjoint neighbourhoods and in p−1(Int⁡top(V,Y))p^{-1}(\operatorname{Int}_{\rm top}(V,Y)).

The set C=ZV∖⋃i=1rWiC=Z_V\setminus\bigcup_{i=1}^r W_i is compact. Its image is closed in the Hausdorff affinoid space VV and does not contain yy. Choose NN inside Int⁡top(V,Y)∩⋂iNi\operatorname{Int}_{\rm top}(V,Y)\cap\bigcap_iN_i and disjoint from p(C)p(C). Then p−1(N)=∐i=1r(Wi∩p−1(N)).p^{-1}(N)=\coprod_{i=1}^r \bigl(W_i\cap p^{-1}(N)\bigr). Every map in this finite disjoint union is the base change of Wi→NiW_i\to N_i, hence is finite by Proposition 2.33.19 and the gluing of Theorem 2.40.8. A finite disjoint union of finite morphisms is finite: over an affinoid target its algebra is the finite product of the finite canonical algebras of the summands. The product norm is the canonical finite-module norm, and its coordinate idempotents identify its spectrum and sheaf with the indicated disjoint union. This proves finiteness of the full restriction.

If the fibre is empty, use C=ZVC=Z_V. The same compact-image argument gives p−1(N)=∅p^{-1}(N)=\varnothing, which is the finite space represented by the zero algebra.

For a proper morphism the interior condition holds everywhere. The neighbourhood just constructed consists of finite-fibre points, since finite analytic fibres are finite discrete spaces by Lemma 2.35.1. Thus the finite-fibre locus is open. On this open locus the restrictions are finite on an ordinary open cover of the target; the target-local definition of a finite analytic morphism glues them to the full finite restriction. ◻

Theorem 5.56.2. NAG-002458 (The finite locus at maximal-dimensional analytic points). Let p:X→Yp:X\to Y be a proper morphism of good Hausdorff analytic spaces, with dim⁡KX≤d\dim_K X\leq d. If y∈Yy\in Y satisfies dK(y)=dd_K(y)=d, then its entire analytic fibre is finite, and pp is finite over an ordinary neighbourhood of yy. If YY is pure of analytic dimension dd, the ordinary finite-fibre locus is open and dense in YY.

Proof. We first verify the zero-dimensional fact needed for the actual fibre. A nonempty generalized affinoid space of analytic dimension zero has a finite underlying set. Indeed choose a strictifying complete extension by Theorem 1.17.6. Theorem 3.21.8 makes the strict algebra after this extension zero-dimensional. Finite normalization, Theorem 1.12.2, and Proposition 3.21.1 make it finite over the new ground field itself. Its actual strict norm is the canonical finite-module norm by Proposition 1.14.4. Its spectrum is finite by Lemma 2.35.1. The scalar projection onto the original spectrum is surjective by Theorem 2.33.13, so the original spectrum is finite too. This argument concerns analytic dimension, with Krull dimension used only after strictification.

Put H=H(y)H=\mathcal H(y). To see that the actual analytic fibre XyX_y is good, choose at each zz above yy ordinary affinoid neighborhoods of yy and zz, the latter mapping into the former. The completed-residue fibre of the resulting affinoid map is affinoid by Theorem 2.33.11. It contains the intersection with the ordinary neighborhood of zz and thus is an ordinary affinoid neighborhood in the whole fibre. Its underlying topology is that of p−1(y)p^{-1}(y), and the completed residue field at a point zz is the original HX(z)\mathcal H_X(z) with its structural isometric inclusion H↪HX(z)H\hookrightarrow\mathcal H_X(z). These assertions follow on each affinoid chart from Theorem 2.33.11, and agree under the domain restrictions of Theorem 2.40.11. The fibre is compact Hausdorff: it is closed in the compact inverse image of any affinoid neighbourhood of yy.

For every point zz in this fibre, the valued-field tower formula, Lemma 3.22.2, gives dH(z)=dK(z)−dK(y)≤dim⁡KX−d≤0.d_H(z)=d_K(z)-d_K(y)\leq\dim_K X-d\leq0. Here the point-dimension bound is Theorem 3.22.9, and dH(z)d_H(z) is nonnegative. It is consequently zero. The maximum formula of Theorem 3.22.7 makes every nonempty affinoid chart of XyX_y zero-dimensional over HH. Compactness supplies finitely many such charts, each with a finite underlying set by the first paragraph. Thus the entire fibre is finite. Properness supplies relative interiority of all its points, so Lemma 5.56.1 gives the asserted ordinary finite neighbourhood, including the empty-fibre case.

If YY is pure of dimension dd, every nonempty ordinary open subset contains a nonempty affinoid neighbourhood VV of dimension dd, by goodness and Lemma 5.50.3. The Abhyankar maximum formula supplies a point of VV with dK(y)=dd_K(y)=d. Corollary 2.33.15 leaves its completed residue field unchanged under domain restriction, so it is a point of the finite locus just proved. That locus therefore meets every nonempty ordinary open subset. Its openness is Lemma 5.56.1, proving density. ◻

Theorem 5.56.3. NAG-002459 (Finite moment collars on compact analytic spaces). Let ZZ be a compact good Hausdorff analytic space of analytic dimension at most dd, and let ff be a finite tuple of global analytic units on ZZ. Choose a finite affinoid-domain cover (Ui)i=1r(U_i)_{i=1}^r whose ordinary interiors cover ZZ, and put E=⋃i=1rtf(∂KUi).E=\bigcup_{i=1}^r t_f(\partial_K U_i). Then EE is a compact finite integral-affine polyhedral set of dimension at most d−1d-1; it is empty if d=0d=0. For every w∉Ew\notin E there is a closed rectangular box PP about ww, with ww in its interior, for which ZP=tf−1(P)⟶TPZ_P=t_f^{-1}(P)\longrightarrow\mathbb T_P is finite. In particular ZPZ_P is an actual affinoid domain of ZZ. The box can be chosen in any prescribed neighbourhood of ww. The inverse image here is the entire inverse image, and restriction to any exact polytorus in PP gives the entire original moment fibre with its analytic structure.

Proof. The cover exists by Lemma 5.50.1. Each UiU_i has dimension at most dd by Theorem 3.22.9. The boundary-image assertion of Theorem 5.39.3 proves the claims about EE. Each boundary is closed by Proposition 2.33.21 and compact by Proposition 2.24.1, so its image and their finite union are compact.

Let q:Z→Gmn,anq:Z\to\mathbb G_m^{n,\mathrm{an}} be the actual analytic map defined by ff, and let Tw\mathbb T_w be the exact polytorus at ww. The map qq is compact: the inverse image of a compact analytic domain in the Hausdorff torus is a closed subset of the compact Hausdorff space ZZ.

For each ii with w∈tf(Ui)w\in t_f(U_i), the entire fibre in UiU_i avoids ∂KUi\partial_K U_i. Proposition 5.38.4 supplies a finite map from its full moment neighbourhood to a closed target polyannulus about ww. If w∉tf(Ui)w\notin t_f(U_i), compactness of this image supplies a box about ww with empty inverse image in UiU_i. Intersect these finitely many boxes and shrink slightly. We obtain a box QQ about ww such that every restriction Ui∩tf−1(Q)⟶TQU_i\cap t_f^{-1}(Q)\longrightarrow\mathbb T_Q is finite, including the empty restrictions.

Fix a point a∈Twa\in\mathbb T_w. Its entire fibre in ZZ is a finite union of fibres of these finite maps, and is therefore finite. At a point zz of this fibre, choose ii for which z∈Int⁡top(Ui,Z)z\in\operatorname{Int}_{\rm top}(U_i,Z). Since ww satisfies all inequalities of QQ strictly, the affinoid Ui∩tf−1(Q)U_i\cap t_f^{-1}(Q) is an ordinary neighbourhood of zz, and TQ\mathbb T_Q is an ordinary neighbourhood of aa. Finite morphisms are everywhere interior by Theorem 2.33.18. Ordinary local testing, Proposition 2.44.6, gives z∈Int⁡(Z/Gmn,an)z\in\operatorname{Int}(Z/\mathbb G_m^{n,\mathrm{an}}). Lemma 5.56.1 now gives an ordinary neighbourhood NaN_a of aa over which the full map qq is finite. The union N=⋃a∈TwNaN=\bigcup_{a\in\mathbb T_w}N_a is an ordinary open neighbourhood of the whole exact polytorus, and q−1(N)→Nq^{-1}(N)\to N is finite by target locality.

Choose a closed box Q′Q' about ww inside the prescribed real neighbourhood. The complement TQ′∖N\mathbb T_{Q'}\setminus N is compact and its tropical image does not contain ww. A smaller closed box PP about ww can therefore be chosen so that TP⊂N\mathbb T_P\subset N. The full restriction q−1(TP)→TPq^{-1}(\mathbb T_P)\to\mathbb T_P is finite, proving the assertion. The analytic-domain inverse-image construction 2.40.11 identifies this source with tf−1(P)t_f^{-1}(P). Finite base change gives the actual exact-level algebras. No point or factor of the original inverse image has been removed. ◻

Corollary 5.56.4. NAG-002460 (Proper morphisms are finite near generic moment levels). Let p:X→Yp:X\to Y be a proper morphism of good Hausdorff analytic spaces, with dim⁡KX≤d\dim_K X\le d. Let V⊆YV\subseteq Y be an affinoid domain and let ff be a unit tuple on VV. There is a compact finite integral-affine polyhedral set E⊆RnE\subseteq\mathbb R^n of dimension at most d−1d-1 such that, for every w∉Ew\notin E, one can choose a closed moment box PP about ww for which VP=tf−1(P),XP=p−1(VP),pP:XP⟶VPV_P=t_f^{-1}(P),\qquad X_P=p^{-1}(V_P),\qquad p_P:X_P\longrightarrow V_P are actual affinoid domains and a finite morphism. Empty domains are allowed. The box can be made arbitrarily small.

In particular, if Fin⁡(pV)\operatorname{Fin}(p_V) is the ordinary finite-fibre locus of pV:p−1(V)→Vp_V:p^{-1}(V)\to V, then tf(V∖Fin⁡(pV))⊆E.t_f\bigl(V\setminus\operatorname{Fin}(p_V)\bigr) \subseteq E. The same inclusion holds for the image of every positive-dimensional analytic fibre. When VV is pure of dimension dd, the complement of EE is dense on every dd-dimensional cell of its actual moment image.

Proof. Properness makes Z=p−1(V)Z=p^{-1}(V) compact Hausdorff. It is good by the actual product charts. At a point of ZZ, choose an ordinary affinoid neighborhood UU of its image in YY and ordinary affinoid neighborhoods X0X_0 and V0V_0 of its projection points in XX and VV, both mapping into UU. The affinoid product X0×UV0X_0\times_U V_0 identifies with the intersection of the two inverse images in X×YVX\times_Y V by Theorem 2.40.11. It contains the intersection of the inverse images of their ordinary interiors, an ordinary neighborhood of the chosen point. Its dimension is at most dd, since it is an analytic domain of XX. Apply Theorem 5.56.3 to ZZ and the unit tuple p∗fp^*f. It gives the set EE and a finite map XP→TPX_P\to\mathbb T_P. The intermediate domain VPV_P is affinoid by the actual simultaneous Laurent inequalities of Theorem 2.31.2. Write their algebras as RP=O(TP)⟶AP=O(VP)⟶BP=O(XP).R_P=\mathcal O(\mathbb T_P)\longrightarrow A_P=\mathcal O(V_P)\longrightarrow B_P=\mathcal O(X_P). The algebra BPB_P is finite over RPR_P, and the same finite generating family makes it finite over APA_P.

Here its given norm is the canonical finite-APA_P-module norm. Indeed, using Theorem 1.18.3, choose RPR_P-module generators b1,…,bmb_1,\ldots,b_m and their controlled lifts for the canonical RPR_P-module norm. They are also APA_P-module generators. Boundedness of RP→APR_P\to A_P makes the quotient APA_P-module norm bounded above by a constant times the given norm of BPB_P. Conversely, the bounded analytic map AP→BPA_P\to B_P and the finitely many bib_i bound the given norm by a constant times that quotient norm. The two norms are therefore equivalent. This proves norm-admissible finiteness over VPV_P.

Every point of VV of weight w∉Ew\notin E lies in the ordinary interior of VPV_P, because the defining moment inequalities are strict there. Its full fibre is consequently finite. Lemma 5.56.1 identifies the resulting ordinary finite locus. A positive-dimensional analytic fibre cannot be the finite fibre of a finite morphism: its affinoid pieces would be finite algebras over the completed target residue field and hence have analytic dimension zero, by finite residue extensions and the Abhyankar maximum formula. Finally EE has dimension at most d−1d-1, so it cannot contain a relatively open subset of a dd-cell. This proves the density assertion exactly on those cells. ◻

Theorem 5.56.5. NAG-002461 (Moment chains on compact analytic domains). Let ZZ be a compact good Hausdorff analytic space pure of dimension dd, let F\mathcal F be coherent on ZGZ_G, and let ff be a global unit tuple. There is a canonical finite nonnegative integral dd-chain TZ,f(F)\mathcal T_{Z,f}(\mathcal F), with support contained in the actual compact image tf(Z)t_f(Z). For any finite affinoid-domain cover (Ui)i=1r(U_i)_{i=1}^r, it is given by TZ,f(F)=∑∅≠I⊆{1,…,r}(−1)∣I∣+1Tf∣UI(Γ(UI,F)),UI=⋂i∈IUi.\mathcal T_{Z,f}(\mathcal F)= \sum_{\varnothing\ne I\subseteq\{1,\ldots,r\}} (-1)^{|I|+1} \mathcal T_{f|_{U_I}}\bigl(\Gamma(U_I,\mathcal F)\bigr), \qquad U_I=\bigcap_{i\in I}U_i. Empty intersections contribute zero. This chain is additive in short exact sequences of coherent modules. For every smooth (d,d)(d,d)-superform α\alpha defined near tf(Z)t_f(Z), ∫TZ,f(F)α=IZ,F(tf∗α).\int_{\mathcal T_{Z,f}(\mathcal F)}\alpha =\mathcal I_{Z,\mathcal F}(t_f^*\alpha). Here the right side is the compact finite-chart integral.

The same construction applies to a finite union of affinoid domains in a good Hausdorff pure dd-dimensional ambient space, even if the union is not good. It is additive by finite inclusion–exclusion for covers by such compact analytic domains. The analogous inclusion–exclusion identity holds for compatible finite moment-chart integrals, without requiring one global tuple.

If ZP=tf−1(P)Z_P=t_f^{-1}(P) is an affinoid moment domain, then TZ,f(F)∣int⁡(P)=Tf∣ZP(Γ(ZP,F))∣int⁡(P).\mathcal T_{Z,f}(\mathcal F)|_{\operatorname{int}(P)} =\mathcal T_{f|_{Z_P}}\bigl(\Gamma(Z_P,\mathcal F)\bigr) |_{\operatorname{int}(P)}. If global unit tuples g,fg,f on ZZ satisfy tf=a∘tgt_f=a\circ t_g for an integral lattice map aa, then a∗TZ,g(F)=TZ,f(F).a_*\mathcal T_{Z,g}(\mathcal F) =\mathcal T_{Z,f}(\mathcal F). This pushforward is proper on the actual support. Cells on which aa drops rank contribute zero.

Proof. All finite intersections are actual affinoid domains by Lemma 5.50.1, and every nonempty one is pure of dimension dd by Lemma 5.50.3. Its sections of F\mathcal F form a finite module by Theorem 2.39.6. Thus the displayed finite sum is a well-defined integral dd-chain. By Theorem 5.39.3, the actual image of ZZ is the finite union of the polyhedral images of the UiU_i, so it is compact and finite polyhedral of dimension at most dd.

To compare covers (Ui)(U_i) and (Vj)(V_j), apply Theorem 5.48.1 on each UIU_I to its affinoid cover (UI∩Vj)j(U_I\cap V_j)_j. The first display becomes ∑I≠∅∑J≠∅(−1)∣I∣+∣J∣Tf∣UI∩VJ(Γ(UI∩VJ,F)).\sum_{I\ne\varnothing}\sum_{J\ne\varnothing} (-1)^{|I|+|J|} \mathcal T_{f|_{U_I\cap V_J}} \bigl(\Gamma(U_I\cap V_J,\mathcal F)\bigr). Interchanging the finite sums and applying the same theorem on each VJV_J gives the expression for the second cover. This also proves independence under further refinement. Exact sequences remain exact on every affinoid intersection by the coherent equivalence, and the affinoid chains are additive by the additivity of finite-module chains; hence so is this chain.

For nonnegativity, order the cover and group each term by its largest index jj. Use the supporting-hyperplane construction of Lemma 5.14.2 to subdivide all images and boundary images together. Consider the relative interior of a dd-cell. All occurring full fibres and their overlaps are interior. The complementary-idempotent calculation in Theorem 5.50.6, using Theorem 5.47.2, identifies the coefficient of the jj-th group with 1eQdim⁡HQ(∏i<j(1−ei))(Mj⊗O(Uj)EQ).\frac{1}{e_Q}\dim_{H_Q} \left(\prod_{i<j}(1-e_i)\right) \bigl(M_j\otimes_{\mathcal O(U_j)}E_Q\bigr). Here Mj=Γ(Uj,F)M_j=\Gamma(U_j,\mathcal F), EQE_Q is the actual transverse Artinian Gauss algebra, eQe_Q is its lattice index, and the eie_i are the idempotents of the overlap domains. Expanding the product gives exactly that group’s alternating sum of module dimensions. The displayed coefficient is nonnegative, and the chain coefficient is an integer because it was defined as an alternating sum of integral weights. This proves positivity, including all module nilpotents. The integral identity is precisely the same finite sum as the defining formula of Theorem 5.50.6.

For a finite affinoid union in a good Hausdorff ambient space, the domain, intersection, purity and coherence statements are the ones verified in the last paragraph of that theorem. The preceding double-cover argument therefore applies verbatim to the chains and to compatible finite chart data. We give the finite-union bookkeeping explicitly. In the commutative ring generated by symbols [U][U] for the finitely many affinoid domains involved, impose [U][V]=[U∩V][U][V]=[U\cap V]; all these symbols are idempotent. A finite covering family is represented by 1−∏i(1−[Ui]).1-\prod_i(1-[U_i]). Expanding gives its inclusion–exclusion expression. Products represent intersection covers: the Boolean identity (1−∏i(1−xi))(1−∏j(1−yj))=1−∏i,j(1−xiyj)(1-\prod_i(1-x_i))(1-\prod_j(1-y_j)) =1-\prod_{i,j}(1-x_i y_j) holds for commuting idempotents. To verify it, expand the identity decomposition into the products which choose either each variable or its complement. On every such product the two sides are the same zero or one. A union of compact domains represented by such covering expressions is consequently represented by 1−∏l(1−Pl)1-\prod_l(1-P_l). Expansion first by the PlP_l, or first by their affinoid constituents, gives the same finite sum of intersection symbols. Cover independence just proved permits either presentation of any intersection. Applying its chain or compatible-integral value proves the asserted finite inclusion–exclusion formula.

For the restriction assertion, intersect the chosen cover with ZPZ_P. On the interior of PP, the whole moment fibre in every UIU_I is unchanged. Refine by the finitely many boundary images of both UIU_I and UI∩ZPU_I\cap Z_P. At every remaining generic dd-cell, the whole-fibre module restriction assertion of Theorem 5.40.2 identifies their weights. Equality on these generic cells is equality of dd-chains. The restricted inclusion–exclusion sum is the affinoid cover expression on ZPZ_P, proving the claim.

Finally apply the same-affinoid chain projection to the identity map of each UIU_I, with its two restricted tuples. It gives the claimed lattice pushforward on each term. Finite linearity of the lattice-index formula gives the equality for their sum. The actual source support is compact, so the pushforward is proper. The rank-dropping convention is the one in that affinoid theorem and is unchanged by finite summation. ◻

Theorem 5.56.6. NAG-002462 (Proper moment pushforward with generic lengths). Let p:X→Yp:X\to Y be proper, where XX and YY are good Hausdorff and pure of the same analytic dimension dd. Let F\mathcal F be coherent on XGX_G. For an affinoid domain V⊆YV\subseteq Y and a unit tuple ff on VV, put Z=p−1(V)Z=p^{-1}(V) and Pp,F,f=TZ,p∗f(F∣Z).\mathcal P_{p,\mathcal F,f} =\mathcal T_{Z,p^*f}(\mathcal F|_Z). This is a finite nonnegative integral dd-chain. There is a compact finite polyhedral set EE of dimension at most d−1d-1 such that at every generic dd-cell point w∉Ew\notin E it has the following description. There is an actual finite restriction pP:XP=p−1(VP)⟶VP=tf−1(P),w∈int⁡(P),p_P:X_P=p^{-1}(V_P)\longrightarrow V_P=t_f^{-1}(P), \qquad w\in\operatorname{int}(P), with affinoid algebras AP,BPA_P,B_P. For the finite BPB_P-module MP=Γ(XP,F)M_P=\Gamma(X_P,\mathcal F) its coefficient is mP(w)=∑p∈Min⁡(AP)length⁡(AP)p((MP)p)mM(AP/p),f(w),length⁡(AP)p((MP)p)=∑q∈Min⁡(BP)q∩AP=plength⁡(BP)q((MP)q)[κ(q):κ(p)].\begin{align*} m_{\mathcal P}(w) &=\sum_{\mathfrak p\in\operatorname{Min}(A_P)} \operatorname{length}_{(A_P)_{\mathfrak p}} ((M_P)_{\mathfrak p}) m_{\mathcal M(A_P/\mathfrak p),f}(w),\\ \operatorname{length}_{(A_P)_{\mathfrak p}} ((M_P)_{\mathfrak p}) &=\sum_{\substack{\mathfrak q\in\operatorname{Min}(B_P)\\ \mathfrak q\cap A_P=\mathfrak p}} \operatorname{length}_{(B_P)_{\mathfrak q}} ((M_P)_{\mathfrak q}) [\kappa(\mathfrak q):\kappa(\mathfrak p)]. \end{align*} An empty component fibre contributes zero. Residue-field degrees in this expression are full degrees, including their inseparable factors. The coefficient and the chain do not depend on the finite collar or the affinoid cover of ZZ.

If a global unit tuple gg on ZZ and an integral map aa satisfy tf∘p=a∘tgt_f\circ p=a\circ t_g, then a∗TZ,g(F∣Z)=Pp,F,f.a_*\mathcal T_{Z,g}(\mathcal F|_Z) =\mathcal P_{p,\mathcal F,f}. In particular its generic coefficient is the full lattice-index sum over all full-rank source cells above that target cell; contracted cells contribute zero. For every smooth (d,d)(d,d)-superform α\alpha defined near tf(V)t_f(V), IZ,F(p∗tf∗α)=∫Pp,F,fα.\mathcal I_{Z,\mathcal F}(p^*t_f^*\alpha) =\int_{\mathcal P_{p,\mathcal F,f}}\alpha.

Proof. The space ZZ is compact Hausdorff and good as in Corollary 5.56.4. Its affinoid domains are domains of XX and are pure of dimension dd by Lemma 5.50.3. Thus ZZ is pure of dimension dd, and Theorem 5.56.5 defines the displayed chain with its stated positivity and finiteness.

Choose the exceptional set in Corollary 5.56.4, and enlarge it by tf(∂KV)t_f(\partial_KV). On a finite common subdivision of the actual image, the source-chain image and this exceptional set, obtained by Lemma 5.14.2, a generic point ww of a dd-cell avoids the exceptional set. Near it, the target image lies in the affine span of that cell, with its saturated rank-dd tangent lattice. The proper finite-collar theorem gives pPp_P; choose the box small enough that ww satisfies its inequalities strictly. Every point of the entire target fibre is absolute interior in VV and interior in the domain VPV_P. Domain interior and composition, Propositions 2.44.5 and 2.33.21, make it absolute interior in VPV_P. Every point of its full inverse image is absolute interior in XPX_P, because pPp_P is finite. The nonempty domains VPV_P and XPX_P are pure of dimension dd, and MPM_P is finite over both BPB_P and APA_P by Proposition 2.35.15.

The restriction statement in Theorem 5.56.5 identifies the source chain near ww with that of XPX_P. Apply the one-affinoid chain calculation in finite pushforward to pP:XP→VPp_P:X_P\to V_P and the single domain VPV_P. The finite collar gives norm-admissibility, and the coefficient module is MPM_P on the full inverse image and the same module regarded over APA_P on the target. The source tuple is the literal pullback of f∣VPf|_{V_P}; the map of moment lattices is the identity. The calculation identifies the two finite chains, with index one on each common saturated top tangent lattice and zero target module weight where the source fibre is empty. The generic-length formula of Theorem 5.40.2 is the first equality in the statement. The Artinian local factor calculation of Theorem 5.42.2 is the second. It uses the actual local module lengths, and its simple factors have dimension the full field degree over the target residue field. Thus neither nilpotents nor inseparability have been removed.

Two collars compare on a smaller moment box retaining the whole target fibre and its entire inverse image. Finite base change identifies their restricted algebras and modules with the algebra and module of that same actual inverse image. The whole-fibre module restriction assertion of Theorem 5.40.2 gives the same coefficient. Cover independence also follows from the already canonical compact chain.

The lattice pushforward is the last assertion of Theorem 5.56.5 on ZZ. It requires only the displayed equality of moment maps; the actual joint-character graph argument is included in its affinoid prerequisite. A common finite refinement of the compact source image gives the finite full-rank lattice-index sum on each generic target cell. Its support is proper, and rank-dropping source cells have zero top-degree contribution. Finally the compact-chain integral identity gives the last display directly. Every construction used the complete inverse image ZZ, and every finite collar retained its entire inverse image over the target box. ◻

NAG-002463 · Open at full scope

Theorem 5.56.7. NAG-002463 (Proper projection from generic coefficient agreement). Let p:X→Yp:X\to Y be proper between good Hausdorff analytic spaces pure of the same dimension dd. Let F\mathcal F and G\mathcal G be coherent on XGX_G and YGY_G, respectively, and fix Δ∈Q≥0\Delta\in\mathbb Q_{\geq0}. Let D⊆YD\subseteq Y be a closed analytic subspace whose defining ideal, on every affinoid domain of YY, is contained in no minimal prime. The empty subspace is permitted.

Suppose that whenever W=M(A)W=\mathcal M(A) is an affinoid domain of Y∖DY\setminus D for which p−1(W)=M(B)→Wp^{-1}(W)=\mathcal M(B)\to W is finite, the modules M=Γ(p−1(W),F),L=Γ(W,G)M=\Gamma(p^{-1}(W),\mathcal F),\qquad L=\Gamma(W,\mathcal G) satisfy, for every p∈Min⁡(A)\mathfrak p\in\operatorname{Min}(A), length⁡Ap(Mp)=Δlength⁡Ap(Lp).\begin{equation} \operatorname{length}_{A_{\mathfrak p}}(M_{\mathfrak p}) =\Delta\operatorname{length}_{A_{\mathfrak p}} (L_{\mathfrak p}). \end{equation} Then for every affinoid domain V⊆YV\subseteq Y and unit tuple ff on VV, Pp,F,f=ΔTf(Γ(V,G)).\mathcal P_{p,\mathcal F,f} =\Delta\mathcal T_f\bigl(\Gamma(V,\mathcal G)\bigr). For every ordinary smooth moment form α∈AY,cd,d(Y)\alpha\in\mathcal A^{d,d}_{Y,c}(Y), the pullback has compact support and ∫X,Fp∗α=Δ∫Y,Gα.\int_{X,\mathcal F}p^*\alpha =\Delta\int_{Y,\mathcal G}\alpha. These scalar formulas multiply under composition when the corresponding coefficient conditions hold for the two maps.

For F=OX\mathcal F=\mathcal O_X and G=OY\mathcal G=\mathcal O_Y, condition (5.56.1) uses the ratio of the full generic local lengths. If AA is reduced, its denominator is one, so a nonzero such ratio is a positive integer. If an actual finite restriction has integral rings A,BA,B and an injective structural map, its numerator is [Frac⁡(B):Frac⁡(A)][\operatorname{Frac}(B):\operatorname{Frac}(A)], with the full inseparable degree retained.

Proof. Fix VV and ff. The actual closed subspace D∩VD\cap V has moment image of dimension at most d−1d-1 by Theorem 5.42.4. This image is compact and finite polyhedral. Add it to the exceptional polyhedral set of Theorem 5.56.6 and take a common finite subdivision by Lemma 5.14.2. At a generic point ww of every remaining dd-cell, choose the proper finite collar small enough that its entire target domain VPV_P misses D∩VD\cap V. This is possible because the moment image of D∩VD\cap V is compact and does not contain ww.

Apply (5.56.1) to W=VPW=V_P. The generic-length formula in Theorem 5.56.6 and the module-weight formula in Theorem 5.40.2 give mP(w)=Δ mΓ(VP,G),f(w)=Δ mΓ(V,G),f(w).m_{\mathcal P}(w) =\Delta\,m_{\Gamma(V_P,\mathcal G),f}(w) =\Delta\,m_{\Gamma(V,\mathcal G),f}(w). The second equality is the whole-fibre restriction assertion of that theorem: the full fibre lies in VPV_P, and the restriction of the coherent module is its actual domain tensor. Equality on the generic part of every dd-cell is equality of the finite dd-chains. The excluded sets have dimension at most d−1d-1 and contain no dd-cell. This proves the first assertion, including components of the source whose moment image has dimension smaller than dd.

The chain equality and Theorem 5.56.5 give the corresponding integral equality for a single smooth moment representative on any affinoid VV. We pass carefully to ordinary forms. Let S=supp⁡(α)S=\operatorname{supp}(\alpha). Pullback of representatives in Theorem 5.52.3 gives supp⁡(p∗α)⊆p−1(S).\operatorname{supp}(p^*\alpha)\subseteq p^{-1}(S). Choose a finite union QQ of affinoid domains of YY whose ordinary interiors cover SS. Every full inverse image of one of these domains is compact by properness and good by Corollary 2.46.4. Choose a finite affinoid cover of each such inverse image. Thus R=p−1(Q)R=p^{-1}(Q) is a finite union of affinoid domains in XX. It is compact Hausdorff, and p−1(S)p^{-1}(S) is closed in it. Consequently the pullback support is compact. The ordinary interior of RR contains it, since it contains p−1(Int⁡top(Q,Y))p^{-1}(\operatorname{Int}_{\rm top}(Q,Y)).

Represent the induced finite chart data of α\alpha on QQ by a finite affinoid cover (Vi)(V_i), with one moment tuple and one smooth representative on each member, as in Theorem 5.52.3. Put Zi=p−1(Vi)Z_i=p^{-1}(V_i). Each ZiZ_i is compact good and admits a finite affinoid cover; the same holds for ZI=p−1(VI)Z_I=p^{-1}(V_I), where VI=⋂i∈IViV_I=\bigcap_{i\in I}V_i. Their pulled-back finite chart representatives agree with the data induced by p∗αp^*\alpha, because both use the same analytic unit pullbacks on every affinoid refinement. The compact-domain inclusion–exclusion assertion of Theorem 5.56.5 computes the integral on RR as the alternating sum of the integrals on the ZIZ_I. On each ZIZ_I, use a representative from any i∈Ii\in I. The single-chart equality just proved identifies this term with Δ\Delta times the corresponding integral on VIV_I. The alternating sum is therefore Δ\Delta times the finite-chart integral on QQ. Theorem 5.52.4 identifies these two compact-neighbourhood integrals with the stated ordinary compact-support integrals.

For two composable maps satisfying the respective conditions, the pullback of a compactly supported form under the second map has compact support by the assertion just proved. Apply the two integral formulas successively and use the functoriality of ordinary smooth pullback in Theorem 5.52.3. The factors multiply.

Finally, for reduced AA, every ApA_{\mathfrak p} at a minimal prime is a field, as checked in Proposition 5.41.2, so its own length is one and the length of a finite module is an integer. For a finite injection of integral rings A→BA\to B, localizing BB at the nonzero elements of AA gives a finite-dimensional domain over Frac⁡(A)\operatorname{Frac}(A), hence a field. It contains BB and is contained in Frac⁡(B)\operatorname{Frac}(B); conversely every nonzero element of BB is invertible in that field, so it is Frac⁡(B)\operatorname{Frac}(B). Its vector-space dimension is the claimed full degree for this actual finite restriction. ◻

NAG-002465 · Open at full scope

Corollary 5.56.8. NAG-002465 (Proper modifications and generic finite free degree). Keep p:X→Yp:X\to Y and D⊆YD\subseteq Y as in Theorem 5.56.7. If pp is an isomorphism over Y∖DY\setminus D and coherent modules F,G\mathcal F,\mathcal G are identified there by F≃p∗G\mathcal F\simeq p^*\mathcal G, then ∫X,Fp∗α=∫Y,Gα\int_{X,\mathcal F}p^*\alpha =\int_{Y,\mathcal G}\alpha for every compactly supported ordinary smooth top form. In particular a proper modification which is an isomorphism away from such a closed subspace preserves the structure-module integral. Additional source components mapped into DD are allowed, provided XX remains pure of dimension dd.

More generally, suppose that over Y∖DY\setminus D the map is finite and its coherent direct image of F\mathcal F is locally free of constant rank rr on the analytic G-site. Then ∫X,Fp∗α=r∫Y,OYα.\int_{X,\mathcal F}p^*\alpha =r\int_{Y,\mathcal O_Y}\alpha. The rank may be the full degree of a finite inseparable map; separability is not required.

Proof. In the isomorphism case the two finite modules on every affinoid restriction disjoint from DD are identified. Their generic local lengths are equal, so Theorem 5.56.7 applies with Δ=1\Delta=1.

For the rank assertion, fix a target affinoid moment chart and exclude the actual moment image of its intersection with DD, as in the proof of that theorem. On a sufficiently small proper finite collar about any remaining generic dd-cell point, the whole target domain is disjoint from DD. The direct image is therefore analytically locally free of rank rr on that domain. The rank formula in Theorem 5.51.1, proved there by an actual finite affinoid trivializing refinement, identifies its finite-module moment chain with rr times the structure chain. Equivalently, on that refinement each actual Artinian Gauss factor carries rr copies of the original module; inclusion–exclusion gives the same equality on the whole domain. Thus the compact source chain equals rr times the target structure chain on every generic dd-cell. The ordinary-form argument in Theorem 5.56.7 now gives the displayed equality without replacing analytic local freeness by an unproved algebraic-local assertion.

If a source component maps into DD, its image under every pulled-back target tuple lies in the finite polyhedral image of DD of dimension at most d−1d-1. Its pulled-back top chain and top-form integral are zero. The preceding proof retains it in every full inverse image; its zero contribution is a dimension calculation. All remaining factors are the full finite-module factors, so the field degrees, including inseparable degrees, retain their original normalization. ◻

5.57 Rational coordinates and initial models along a skeleton

Let KK be complete for its specified nontrivial real-valued absolute value. Write O=K∘O=K^\circ, k=K~k=\widetilde K, and v=−log⁡∣⋅∣v=-\log|\cdot|, and fix π∈O\pi\in O with 0<∣π∣<10<|\pi|<1. We use the algebraic model setting of Theorem 5.30.3: X\mathcal X is a given projective, flat, finitely presented OO-scheme, its generic fibre XX is smooth and integral, and a finite collection of affine opens covering Xs\mathcal X_s has étalemaps to Spec⁡Rλ,Rλ=O[X0,…,Xrλ,T1,…,Tqλ]/(X0⋯Xrλ−aλ),0<∣aλ∣<1.\operatorname{Spec}R_\lambda,\qquad R_\lambda= O[X_0,\ldots,X_{r_\lambda},T_1,\ldots,T_{q_\lambda}]/ (X_0\cdots X_{r_\lambda}-a_\lambda), \quad 0<|a_\lambda|<1. Let S⊂XanS\subset X^{\mathrm{an}} be the compact skeleton of the completed model. All rational functions below belong to the original field K(X)K(X).

For such a chart V=Spec⁡C\mathcal V=\operatorname{Spec}C and a nonempty I⊂{0,…,r}I\subset\{0,\ldots,r\}, write SIS_I for its residual stratum over DI={Xi=0 (i∈I), Xj≠0 (j∉I)}D_I=\{X_i=0\ (i\in I),\ X_j\ne0\ (j\notin I)\} and QIQ_I for the finite set of generic points of SIS_I. These are the strata and labels of Proposition 5.24.2. If s=xη,ws=x_{\eta,w}, then I={i:wi>0}I=\{i:w_i>0\} and η∈QI\eta\in Q_I. An irreducible residual component in this section is a component of this locally closed stratum.

Lemma 5.57.1. NAG-002467 (Algebraic coordinates on a common affine open). There is an affine open W=Spec⁡D\mathcal W=\operatorname{Spec}D of X\mathcal X containing every specialization of SS. There is a finite collection of affine étalesemistable charts Vλ=Spec⁡Cλ\mathcal V_\lambda=\operatorname{Spec}C_\lambda contained in W\mathcal W whose completed generic fibres cover SS. For each pair (I,η)(I,\eta) in a chart one may choose hλ,I,η∈Cλh_{\lambda,I,\eta}\in C_\lambda such that Eλ,I,η:=(SI)h‾λ,I,ηE_{\lambda,I,\eta}:=(S_I)_{\overline h_{\lambda,I,\eta}} is a nonempty smooth integral open of the component labelled by η\eta and meets no other component of SIS_I.

Fix a faithful rational tuple F0F_0 as in Theorem 5.30.3. There are a finite list P⊂D∖{0}P\subset D\setminus\{0\}, a dense open U=Spec⁡DK[1/∏p∈Pp]⊂X,U=\operatorname{Spec}D_K\left[1/\prod_{p\in P}p\right] \subset X, and a finite tuple F⊂K(X)×F\subset K(X)^\times with the following properties. The open UU is contained in every Vλ,K\mathcal V_{\lambda,K} and contains SS analytically. The tuple FF contains PP, F0F_0, finite nonzero OO-algebra generating sets of DD and every CλC_\lambda, the chart coordinates, and every hλ,I,ηh_{\lambda,I,\eta}. Each member of FF is a ratio of members of PP, or itself a member of PP. All members of FF are units on UU, and they define a closed immersion U↪Gm,K∣F∣U\hookrightarrow\mathbf G_{m,K}^{|F|}. Its restriction to SS is faithful and unimodular on a finite subdivision.

Proof. Proposition 5.30.2 makes the set ZZ of skeletal specializations finite. Apply Lemma 5.30.1 to obtain W\mathcal W. For each point of ZZ in an original chart, choose a principal open neighbourhood in that chart contained in W\mathcal W. There are only finitely many choices. A principal open is étaleover its chart, and the restricted map to the standard model is étaleby [StacksAlg] (Tag 00U2). At every skeletal point specializing to the chosen point, the defining function of this principal open has norm one. Lemma 4.13.1 and Proposition 5.23.4 therefore show that these completed opens cover SS and have the restricted skeletons claimed.

The algebras DD and CλC_\lambda are finitely presented over OO: apply [StacksAlg] (Tag 01TQ) to the restrictions of the finitely presented structural morphism to these affine opens. They are OO-flat. Thus they inject into their generic coordinate rings, which in turn inject into K(X)K(X).

For fixed (λ,I)(\lambda,I), the distinct primes in QIQ_I are incomparable. For each η′≠η\eta'\ne\eta, choose an element of the prime of η′\eta' outside the prime of η\eta, and multiply these finitely many elements. They may be chosen in Cλ⊗OkC_\lambda\otimes_O k, by taking the inverse images of the stratum primes in this ring. Lift the product to hλ,I,η∈Cλh_{\lambda,I,\eta}\in C_\lambda. If QIQ_I is a singleton, take hλ,I,η=1h_{\lambda,I,\eta}=1. The stratum is smooth over kk, by étalebase change and smooth composition [StacksAlg] (Tags 00U2, 036U); it is reduced by [StacksAlg] (Tag 056T). The principal open thus has exactly the chosen minimal prime and is integral. It is nonempty, smooth, and contains η\eta. In particular each chosen hh is a nonzero element of K(X)K(X).

Let AA be the finite list consisting of F0F_0, nonzero algebra generators for the charts, their standard coordinates, and the chosen hh’s. For every f∈Af\in A write f=pf/qff=p_f/q_f with nonzero pf,qf∈Dp_f,q_f\in D. This is possible because Frac⁡(D)=K(X)\operatorname{Frac}(D)=K(X); denominators from KK can be cleared by a power of π\pi. For each λ\lambda, choose bλ∈D∖{0}b_\lambda\in D\setminus\{0\} whose principal generic open is contained in the nonempty open Vλ,K\mathcal V_{\lambda,K} of WK\mathcal W_K. Such a principal open exists by the affine open basis; again clear a scalar denominator to put bλb_\lambda in DD. Let PP contain these bλb_\lambda, all pf,qfp_f,q_f, and a finite nonzero generating set for DD over OO, and put F=A∪PF=A\cup P. Zero generators can be omitted.

Every point of SS is a norm on K(X)K(X) by Proposition 5.30.2. It consequently has generic scheme support and belongs to UanU^{\mathrm{an}}. The elements bλb_\lambda give U⊂Vλ,KU\subset\mathcal V_{\lambda,K}. The Laurent coordinate map associated with FF surjects onto DK[1/∏P]D_K[1/\prod P]: its image contains the generators of DKD_K and the inverses of every member of PP. This proves the closed immersion assertion. Finally all added rational coordinates have finite piecewise v(K×)v(K^\times)-integral-affine values on SS by the same proposition. Retaining F0F_0 retains injectivity and an integral left inverse on every tangent lattice. A common finite subdivision, followed by the integral graph argument in Theorem 5.30.3, gives the last assertion. ◻

Proposition 5.57.2. NAG-002468 (A weighted chart with one residual component). Fix a chart Spec⁡C\operatorname{Spec}C, a label η∈QI\eta\in Q_I, and its separator hh, and put E=(SI)hˉE=(S_I)_{\bar h}. Let w∈relint⁡ΔIw\in\operatorname{relint}\Delta_I and s=xη,ws=x_{\eta,w}. Let L/KL/K be a complete isometric extension in whose value group all wiw_i occur. Choose ci∈L×c_i\in L^\times with v(ci)=wiv(c_i)=w_i and ∏ici=a\prod_i c_i=a, taking cj=1c_j=1 for j∉Ij\notin I. Write l=L~l=\widetilde L.

There is an admissible L∘L^\circ-model B0B_0 of the actual domain in the completed chart over LL given by ∣Xi∣=∣ci∣(0≤i≤r),∣h∣=1.|X_i|=|c_i|\quad(0\leq i\leq r), \qquad |h|=1. Its special fibre is Spec⁡(B0/mLB0)≃El×Gm,l∣I∣−1.\operatorname{Spec}(B_0/\mathfrak m_LB_0) \simeq E_l\times\mathbf G_{m,l}^{|I|-1}. This fibre is smooth and geometrically reduced, and B0B_0 is the spectral unit ball of B=B0[1/π]B=B_0[1/\pi]. Every Shilov point of ML(B)\mathcal M_L(B) restricts to ss. Consequently, for every g∈Cg\in C nonzero as a rational function on XX, ρB(g)=∣g(s)∣>0,∣g(ξ)∣=∣g(s)∣(ξ∈ΓB).\rho_B(g)=|g(s)|>0, \qquad |g(\xi)|=|g(s)| \quad(\xi\in\Gamma_B).

Proof. Use the separator supplied by Lemma 5.57.1. Choose i0∈Ii_0\in I. Choose the cic_i for i≠i0i\ne i_0 first and then put ci0=a/∏i≠i0cic_{i_0}=a/\prod_{i\ne i_0}c_i. The relation ∑iwi=v(a)\sum_iw_i=v(a) gives its required norm. Consider the algebraic base change Aw=L∘[U0±1,…,Ur±1,T1,…,Tq]/(∏iUi−1),Cw=(C⊗RAw)[1/h],B0=Cw^,\begin{aligned} A_w&=L^\circ[U_0^{\pm1},\ldots,U_r^{\pm1},T_1,\ldots,T_q] /(\prod_iU_i-1),\\ C_w&=(C\otimes_R A_w)[1/h],\\ B_0&=\widehat{C_w}, \end{aligned} where Xi↦ciUiX_i\mapsto c_iU_i and completion is π\pi-adic. Eliminating Ui0U_{i_0} makes AwA_w a Laurent polynomial algebra with polynomial variables over L∘L^\circ; it is flat by the coefficientwise ideal criterion [StacksAlg] (Tag 00HD). The map Aw→CwA_w\to C_w is étaleby [StacksAlg] (Tag 00U2). Thus CwC_w is finitely presented and flat over L∘L^\circ, and Lemma 4.13.1 applies.

Its explicit bounded-coordinate presentation imposes the original chart bounds, ∣Ui∣=1|U_i|=1, and ∣h−1∣≤1|h^{-1}|\leq1. As hh was already integral, these are exactly the displayed domain conditions. This proves the assertion about the actual analytic domain, with its analytic structure. On reduction, cic_i vanishes for i∈Ii\in I and equals one for i∉Ii\notin I. The coordinates UjU_j for j∉Ij\notin I are therefore the invertible stratum coordinates XjX_j. After eliminating Ui0U_{i_0}, the remaining UiU_i, i∈I∖{i0}i\in I\setminus\{i_0\}, are independent invertible variables. Reducing the displayed finite presentation gives precisely El×Gm,l∣I∣−1E_l\times\mathbf G_{m,l}^{|I|-1}, also for nondiscrete valuations: integral restricted series reduce to polynomials. Smoothness and geometric reducedness follow from [StacksAlg] (Tags 00U2, 036U, 056T). Lemma 4.12.1 identifies B0B_0 with B∘B^\circ and its special fibre with B~\widetilde B.

Let ξ∈ΓB\xi\in\Gamma_B. Its reduction is a component generic point by Theorem 3.16.7. The reduced algebra is étale, hence flat, over l[Ui±1:i≠i0,T]l[U_i^{\pm1}:i\ne i_0,T] after the above identification. Flat going down, [StacksAlg] (Tag 00HS), makes its minimal prime contract to the zero prime of this Laurent polynomial algebra. The image of ξ\xi in its analytic unit polytorus times polydisc therefore reduces to the generic point. It is the Gauss point: divide any nonzero restricted series by a coefficient of maximum norm; its nonzero polynomial reduction cannot vanish at that generic point. Polynomial approximation, or Proposition 1.2.2, gives the Gauss value for every series.

Restriction to KK gives the standard point zwz_w. Moreover the reduction of ξ\xi contracts to η\eta in the original stratum. Indeed the field extension k→lk\to l is flat by [StacksAlg] (Tag 00HD), and going down applied to El→EE_l\to E contracts a minimal prime to the unique minimal prime of EE; adjoining the independent Laurent variables has the same property. Compatibility of algebraic completion, scalar extension and specialization is the explicit presentation comparison of Lemma 4.28.1. Proposition 5.24.2 now says that the restricted point is exactly xη,w=sx_{\eta,w}=s. Taking maxima on the finite Shilov boundary gives the last formula. Its value is positive by Proposition 5.30.2. ◻

Lemma 5.57.3. NAG-002469 (Localization at a common Shilov value). Let LL be complete and nontrivially valued, and let B0B_0 be an admissible L∘L^\circ-algebra with reduced special fibre, and put B=B0[1/π]B=B_0[1/\pi]. Suppose g1,…,gt∈Bg_1,\ldots,g_t\in B and b1,…,bt∈L×b_1,\ldots,b_t\in L^\times satisfy ∣gj(ξ)∣=∣bj∣(ξ∈ΓB,1≤j≤t).|g_j(\xi)|=|b_j| \quad(\xi\in\Gamma_B, 1\leq j\leq t). Put uj=gj/bju_j=g_j/b_j. Then uj∈B0u_j\in B_0, and B0′=B0⟨V1,…,Vt⟩/(ujVj−1:1≤j≤t)B'_0=B_0\langle V_1,\ldots,V_t\rangle/ (u_jV_j-1:1\leq j\leq t) is an admissible model of the domain ∣gj∣=∣bj∣|g_j|=|b_j| for all jj. Its reduction is B~[1/uˉ1⋯uˉt]\widetilde B[1/\bar u_1\cdots\bar u_t], and this open meets every irreducible component, including every geometric component. Its Shilov points are the original Shilov points, regarded in the restricted domain. In particular a smooth special fibre remains smooth after this localization.

Proof. If B=0B=0, then B0=0B_0=0 by π\pi-torsion-freeness, and every assertion is immediate. Assume B≠0B\ne0. The Shilov maximum formula gives ρB(uj)=1\rho_B(u_j)=1, and Lemma 4.12.1 gives uj∈B0u_j\in B_0. At each component generic point its residue is nonzero by Theorem 3.16.7 and the equality ∣uj(ξ)∣=1|u_j(\xi)|=1. Apply Proposition 3.17.1 successively. It gives the displayed admissible algebra, the exact domain ∣uj∣=1|u_j|=1, and the principal localization on reduction. Its special fibre is still reduced, so the reduced-model lemma and the generic-point theorem apply again. All original generic points survive. Their original analytic lifts belong to the restricted domain and are the unique lifts of those generic points. This identifies the Shilov boundaries.

After any field extension of the residue field, each minimal prime contracts to an original minimal prime by [StacksAlg] (Tags 00HD, 00HS). None can therefore contain any uˉj\bar u_j. Thus the localization retains every geometric component as well. Finally, restriction to an open preserves smoothness, as in [StacksAlg] (Tag 036U). ◻

Theorem 5.57.4. NAG-002470 (Actual initial models along a projective skeleton). In the stated model setting, the tuple FF of Lemma 5.57.1 can be chosen so that the resulting closed immersion φ:U↪Gm,KN\varphi:U\hookrightarrow\mathbf G_{m,K}^N has the following additional property. Let s∈Ss\in S, put ω=trop⁡φ(s)\omega=\operatorname{trop}_\varphi(s), and choose a refined chart containing its specialization, with label (I,η)(I,\eta). For every complete isometric L/KL/K realizing the coordinates of ω\omega in its value group, the actual initial scheme of ULU_L at ω\omega is isomorphic to a principal open of Eλ,I,η,L~×Gm,L~∣I∣−1E_{\lambda,I,\eta,\widetilde L} \times\mathbf G_{m,\widetilde L}^{|I|-1} which meets every geometric irreducible component. In particular this initial scheme is smooth and geometrically reduced. The quotient norm of the normalized tropical-fibre presentation is its spectral norm, and its integral quotient model is the spectral unit ball.

Proof. Use precisely the finite lists constructed in Lemma 5.57.1. Fix ss and its chart, and write h=hλ,I,ηh=h_{\lambda,I,\eta}. The normalization field LL exists by Theorem 5.13.6; the argument below works for every such LL. Choose the standard-coordinate normalizers with product aλa_\lambda as in Proposition 5.57.2. Let B0B_0 be that proposition’s selected weighted model.

We first locate the entire tropical fibre. Every point of the actual fibre Yω=trop⁡φ−1(ω)⊂UanY_\omega=\operatorname{trop}_\varphi^{-1}(\omega) \subset U^{\mathrm{an}} has, on every algebra generator of the selected chart, the same absolute value as ss, which is at most one. Algebraically UU is contained in this chart by construction. Hence the actual bounded-coordinate description of Lemma 4.13.1 puts the entire fibre inside the completed chart. The coordinates XiX_i impose its weights wiw_i, and the coordinate hh imposes ∣h∣=∣h(s)∣=1|h|=|h(s)|=1. The same argument over LL places the entire base-changed fibre in ML(B0[1/π])\mathcal M_L(B_0[1/\pi]). This also shows explicitly how points outside the selected formal chart are excluded: their chart-generator norms cannot satisfy all these equalities.

For each p∈Pp\in P, choose bp∈L×b_p\in L^\times with ∣bp∣=∣p(s)∣|b_p|=|p(s)|. The function pp is regular on this chart because Vλ⊂W\mathcal V_\lambda\subset\mathcal W. Proposition 5.57.2 gives this same norm at every Shilov point of B0[1/π]B_0[1/\pi]. Lemma 5.57.3 therefore gives B0′=B0⟨Vp:p∈P⟩/((p/bp)Vp−1:p∈P).B'_0=B_0\langle V_p:p\in P\rangle/ ((p/b_p)V_p-1:p\in P). Its reduction is a principal open of the stated smooth product and meets every geometric component.

Its generic fibre is exactly Yω,LY_{\omega,L}. In one direction this follows from the preceding containment and the equalities ∣p∣=∣p(s)∣|p|=|p(s)|. In the other direction those equalities make all members of PP invertible and thus put a point in ULanU_L^{\mathrm{an}}. Every remaining coordinate of FF is pf/qfp_f/q_f with pf,qf∈Pp_f,q_f\in P, so it has the required absolute value as well. The presentations of rational domains identify the analytic structures, not only the sets of points; equivalently, use Proposition 2.33.12 and the universal property of these displayed equations.

It remains to identify the integral model with the actual initial quotient. Choose normalizers bfb_f for all coordinates of FF. Changing them by norm-one constants only changes torus coordinates. Put T0=L∘⟨Zf±1:f∈F⟩,T=T0[1/π],Zf⟼f/bf.T_0=L^\circ\langle Z_f^{\pm1}:f\in F\rangle, \qquad T=T_0[1/\pi], \qquad Z_f\longmapsto f/b_f. All these images and their inverses belong to B0′B'_0. For f=pf/qff=p_f/q_f this follows by using the normalized pf,qfp_f,q_f and a norm-one scalar. The map T0→B0′T_0\to B'_0 is surjective, as can be checked on its actual restricted presentation. Every chart algebra generator uu is the image of buZub_u Z_u, with bu∈L∘b_u\in L^\circ because ∣u(s)∣≤1|u(s)|\leq1. Every Ui±1U_i^{\pm1} is the image of ZXi±1Z_{X_i}^{\pm1} multiplied by a norm-one scalar. The generator h−1h^{-1} is the image of a norm-one scalar times Zh−1Z_h^{-1}. Finally the generators (p/bp)−1(p/b_p)^{-1} are the images of Zp−1Z_p^{-1}. These are all the generators in the displayed construction of B0′B'_0. Their lifts are integral Laurent polynomials of Gauss norm at most one. Substitution lifts every integral restricted series in them to T0T_0, by Proposition 1.2.3. Thus the map is surjective, rather than merely having dense image.

Let a\mathfrak a be the algebraic ideal of the closed immersion UL↪Gm,LNU_L\hookrightarrow\mathbf G_{m,L}^N, in these normalized coordinates. Proposition 5.13.2 identifies the generic map with T/aT ≃ B0′[1/π].T/\mathfrak aT\ \simeq\ B'_0[1/\pi]. Since B0′B'_0 is L∘L^\circ-flat and is embedded in its generic algebra, the kernel of T0→B0′T_0\to B'_0 is precisely aT∩T0\mathfrak aT\cap T_0. Consequently T0/(aT∩T0) ≃ B0′.T_0/(\mathfrak aT\cap T_0)\ \simeq\ B'_0. Proposition 5.13.5 identifies the reduction of this quotient with the actual initial scheme of the algebraic immersion. Its special fibre was computed above. Lemma 4.12.1 now identifies the quotient norm with the spectral norm and the quotient model with the spectral unit ball. The arbitrary-real-weight interpretation and independence of normalization are those of Theorem 5.13.6; the calculation has retained every component throughout. ◻

Corollary 5.57.5. NAG-002471 (Shilov fibres and geometric generic lengths). For the immersion of Theorem 5.57.4, the original-KK tropical fibre at trop⁡φ(s)\operatorname{trop}_\varphi(s) has Shilov boundary {s}\{s\} for every s∈Ss\in S. The inverse of the faithful map on SS is therefore a continuous section over trop⁡φ(S)\operatorname{trop}_\varphi(S) selecting the unique Shilov point.

For s=xη,ws=x_{\eta,w}, let gηg_\eta be the number of irreducible components of Eλ,I,η,kˉE_{\lambda,I,\eta,\bar k}, where kˉ\bar k is an algebraic closure of kk. After any normalization and algebraically closed residue extension, the sum of the generic component lengths of the actual initial scheme at trop⁡φ(s)\operatorname{trop}_\varphi(s) equals gηg_\eta. In particular this sum is one if the corresponding residual component is geometrically integral. If every residual component in these charts is geometrically integral, this conclusion holds at every point of SS. At weights lying in relative interiors of maximal tropical cells, the sum is the tropical multiplicity of Proposition 5.19.7.

Proof. Normalize the weights over a complete isometric extension L/KL/K. Theorem 5.57.4 and Lemma 5.57.3 identify the Shilov points of the base-changed fibre with those of the selected weighted chart. Every one restricts to ss by Proposition 5.57.2. For any analytic function gg on the original tropical fibre, its base change consequently satisfies max⁡Yω,L∣g∣=∣g(s)∣.\max_{Y_{\omega,L}}|g|=|g(s)|. The map Yω,L→YωY_{\omega,L}\to Y_\omega is surjective by Theorem 2.33.13, with scalar compatibility supplied by Proposition 5.13.2. The same maximum formula therefore holds on YωY_\omega. The singleton {s}\{s\} is a boundary, and the existence and minimality of the Shilov boundary in Corollary 3.12.5 give the assertion. The faithful map on the compact SS is a homeomorphism onto its image, so its inverse is continuous and has the stated values.

Over an algebraically closed residue field the initial scheme is smooth and hence reduced. Each of its generic local rings is consequently a field and has length one. The principal localization in Theorem 5.57.4 removes no geometric component. Taking the product with a split torus changes no component count: the Laurent polynomial ring of an integral component is a domain, and the finite minimal-prime decomposition gives all components this way.

For completeness, this count is unchanged by extending the algebraically closed residue field. Every irreducible component over an algebraically closed field is geometrically irreducible by [StacksAlg] (Tag 020J). Thus its base change remains irreducible. The components stay distinct and incomparable: contraction of a prime under a field extension preserves their original defining prime, by faithful flatness of a nonzero extension of fields and [StacksAlg] (Tags 00HD, 00HP, 00HQ). Their finite union still covers the base change. To compare kk with the residue field ll of a chosen normalization, take an algebraic closure of ll and inside it the algebraic closure of kk consisting of elements algebraic over kk. Applying the preceding argument gives exactly gηg_\eta. The asserted length sum follows. For a maximal-cell weight, Proposition 5.19.7 identifies this sum with its registered tropical weight. ◻

References. This construction reconstructs the rational-coordinate and initial-fibre methods of Gubler–Rabinoff–Werner, Skeletons and tropicalizations, arXiv:1404.7044v3, Sections 9–10, in the stated projective algebraic model scope. The non-Archimedean inputs are the internally proved results cited above. The scheme-theoretic initial quotient is identified by its integral presentation, and the arithmetic component count is retained; no independent-priority claim is made.