6 · Charts, skeleta and faithful tropicalization

11 NAG Tags

Chapter 4 constructed the annular and semistable curve skeleta. We use them here to compare tropical maps, integral lattice indices and canonical analytic lifts. Section 3.8 supplies the residue-dimension calculation for the higher-dimensional local model.

6.1 Skeleta and tropical images

The Gauss section of a torus

For w∈NRw\in N_{\mathbb R}, define a weighted Gauss norm on the Laurent algebra by ∣∑mamχm∣σ(w)=max⁡m∣am∣e−⟨m,w⟩.\begin{equation} \left|\sum_m a_m\chi^m\right|_{\sigma(w)} =\max_m |a_m|e^{-\langle m,w\rangle}. \end{equation} It is a multiplicative norm and therefore a point of TanT^\mathrm{an}. The minus sign in the exponential is forced by v=−log⁡∣⋅∣v=-\log|\cdot|. In particular, −log⁡∣χm∣σ(w)=⟨m,w⟩,trop⁡∘σ=id⁡NR.-\log|\chi^m|_{\sigma(w)}=\langle m,w\rangle, \qquad \operatorname{trop}\circ\sigma=\operatorname{id}_{N_{\mathbb R}}. The set S(T)=σ(NR)S(T)=\sigma(N_{\mathbb R}) is the torus skeleton. The map σ\sigma is continuous and identifies NRN_{\mathbb R} with this subset of TanT^\mathrm{an}; a separate deformation-retraction theorem makes σ∘trop⁡\sigma\circ\operatorname{trop} a strong deformation retraction. Having a section alone would not prove that theorem.

Equation (6.1.1) also gives the exact identity −log⁡∣f∣σ(w)=Ff(w).-\log|f|_{\sigma(w)}=F_f(w). The Gauss norm retains the largest size among the terms, even when several sizes agree. It does not perform the cancellation that can occur at other points in the same tropical fibre.

References. [GRW16] (Sections 3–4); Proposition 5.5.1.

Example 6.1.1. (The ambient section misses the hypersurface). Let X=V(1+z1+z2)⊂Gm2X=V(1+z_1+z_2)\subset\mathbb G_m^2. For every w∈R2w\in\mathbb R^2, ∣1+z1+z2∣σ(w)=max⁡(1,e−w1,e−w2)>0.|1+z_1+z_2|_{\sigma(w)} =\max(1,e^{-w_1},e^{-w_2})>0. Thus σ(w)∉Xan\sigma(w)\notin X^\mathrm{an}, including when w∈Trop⁡(X)=Lw\in\operatorname{Trop}(X)=L. A point of XanX^\mathrm{an} must annihilate its defining equation. The ambient Gauss point does not.

Consequently, restricting σ\sigma to LL does not produce a section L→XanL\to X^\mathrm{an}. A section for a subvariety, when a theorem supplies one, must choose different analytic points. Likewise, the ambient torus deformation retraction need not preserve XanX^\mathrm{an}.

This elementary example is a useful safeguard. The tropicalization is a subset of a real vector space; a skeleton is a subset of an analytic space. A map can identify suitable pieces of them, but the identification is additional structure or an additional theorem.

Annuli, lengths, and integral slopes

The annular skeleton and its retraction were constructed in Section 4.3 and illustrated in Figure 4.1. Its coordinate t=−log⁡∣T∣t=-\log|T| has integral slope one. We now compare the logarithmic metric with lattice indices and model multiplicities.

For the local semistable equation UV=aUV=a, with 0<∣a∣<10<|a|<1, the annular skeleton has coordinate t=v(U)t=v(U) and interval 0≤t≤v(a),v(V)=v(a)−t.0\leq t\leq v(a),\qquad v(V)=v(a)-t. If a=πba=\pi^b over a discretely valued field, its length is bλb\lambda, where λ=−log⁡∣π∣\lambda=-\log|\pi|. It becomes bb only after the explicit normalization λ=1\lambda=1.

For a curve with a chosen semistable model, annular pieces assemble into a metric graph. Over a non-algebraically-closed residue field, residual degrees can contribute weights to its slope formulas. The model, the graph, its metric, and these weights must all be specified. A smallest skeleton need not contain enough cells to represent every differential form: a curve with good reduction can have a one-point minimal skeleton, although its local differential calculus is not zero.

There are also several kinds of boundary. The endpoints of a closed annulus belong to its analytic boundary, whereas an open annulus is boundaryless. A facet in a chosen polyhedral subdivision is a combinatorial notion. A component removed from an algebraic compactification is yet another notion. Balancing away from an analytic boundary cannot be silently applied at the endpoint of a restricted annulus.

References. [GRW16] (Sections 3–4 and 9–10).

For a discretely valued field with uniformizer π\pi, write λ=−log⁡∣π∣\lambda=-\log|\pi|. The annulus ∣π∣m≤∣T∣≤1|\pi|^m\leq|T|\leq1 has length mλm\lambda. Its integral slope for the function −log⁡∣T∣-\log|T| is one, while the change in function value is mλm\lambda. The integer mm, the real length, and the slope are different quantities. If an algebraic intersection length is mm, the associated logarithmic height has the factor λ\lambda in this normalization. By contrast, the total mass of the Chern measure of a degree-one model line is one.

Example 6.1.2. (Power maps on annuli). Let 0<r<R0<r<R and d≥1d\geq1. The map T↦TdT\mapsto T^d takes A[r,R]A[r,R] to A[rd,Rd]A[r^d,R^d]. If t=−log⁡∣T∣t=-\log|T|, the target skeletal coordinate is dtdt. Thus the target interval has dd times the source length, and the image of the source integral tangent lattice has index dd in the target lattice. The map is finite of degree dd: the equation Td=ST^d=S is monic and the source algebra is free of rank dd over the target algebra. This includes the inseparable case. The integral slope and algebraic degree retain the multiplicity that a count of distinct fibres can miss.

A semistable local model and its simplex

Over a discretely valued field, a basic strictly semistable model has an eˊ\acute{e}tale local chart Spec⁡(K∘[x0,…,xd]/(x0⋯xr−π)),0≤r≤d.\operatorname{Spec}\bigl(K^\circ[x_0,\ldots,x_d]/(x_0\cdots x_r-\pi)\bigr), \qquad 0\leq r\leq d. Write λ=v(π)>0\lambda=v(\pi)>0. Its skeleton is locally described by ti=v(xi)≥0,∑i=0rti=λ.t_i=v(x_i)\geq0,\qquad \sum_{i=0}^r t_i=\lambda. This is an rr-simplex. Its dimension is rr, while the generic fibre has dimension dd. The remaining coordinates describe directions not recorded by this simplex.

On the special fibre, the equation becomes x0⋯xr=0x_0\cdots x_r=0: the components are the coordinate hyperplanes. Their intersections are normal-crossing strata. For r=1r=1 the transversal picture is a node; for larger rr, several components meet normally. These semistable equations control the singularities of the special fibre. Over a general valuation ring we do not additionally assert that the underlying formal model is a noetherian regular scheme. A stratum where more components meet corresponds to a simplex of higher dimension. Thus a skeleton records incidence and valuation data, not every residual algebraic direction.

Faithful tropicalization and multiplicity-one sections

In this subsection, let KK be complete, algebraically closed, and nontrivially valued. Let (X,H)(\mathfrak X,H) be a strictly semistable pair in the proper algebraic setting of Gubler–Rabinoff–Werner. Its skeleton S(X,H)S(\mathfrak X,H) lies in the analytic generic fibre, with unbounded directions when the horizontal divisor HH is present. A tropicalization is faithful on this skeleton if it is injective there and unimodular on its polyhedral pieces: it preserves the integral tangent lattices, not just the topology.

Theorem 6.1.3. (GRW, Theorem 9.5). In the strictly semistable-pair setting of [GRW16], there exists a finite collection of nonzero rational functions on the generic fibre whose tropicalization is faithful on S(X,H)S(\mathfrak X,H).

The rational functions are regular units on a suitable dense open containing the relevant skeletal points. The assertion does not say that every preassigned moment map is faithful. Nor does it say that the entire analytic space is homeomorphic to the skeleton. The added rational functions can have divisors outside HH; the resulting tropicalization need not factor through the skeletal retraction [GRW16] (Remark 9.6).

A different theorem starts with an embedding of a very affine integral variety into a torus. On a subset ZZ of its tropicalization where the tropical multiplicity is one, each tropical fibre has a unique Shilov boundary point. These points give a continuous partial section Z→XanZ\to X^{\mathrm{an}}. If ZZ is contained in the closure of its interior in the whole tropicalization, this is the unique continuous section on ZZ [GRW16] (Theorem 10.6). This statement concerns multiplicity-one fibres, and is not obtained by restricting the ambient torus Gauss section.

On a curve with good reduction, a natural compact skeleton may consist of one point. If all differential forms factored through that point, every positive-degree form would vanish. Local annular and punctured-disc domains still admit nonconstant tropical coordinates and nonzero forms. Therefore global factorization through a chosen small skeleton is a property of special data, not a general consequence of the existence of a skeleton.

Payne’s inverse limit: what is recovered

For a quasiprojective variety over a complete algebraically closed nontrivially valued field, Payne identifies its analytification with the inverse limit of extended tropicalizations over the appropriate system of toric embeddings: Xan≃lim←⁡ιTrop⁡ι(X).X^{\mathrm{an}}\simeq\varprojlim_{\iota}\operatorname{Trop}_{\iota}(X). This is a homeomorphism of topological spaces [Pay09] (Theorem 4.2). The transition maps remember how a larger collection of characters projects to a smaller one.

The reason a whole system is needed can already be seen in one tropical fibre. A fixed tuple of values ∣fi(x)∣|f_i(x)| may be shared by many analytic points. Adding more functions can distinguish them. The inverse limit requires compatible values for all such choices and recovers the seminorm, whereas one finite tropicalization records only part of it.

There are three different questions here:

Question Additional structure needed
Recover the underlying topology The inverse system of extended tropicalizations.
Recover analytic functions and their norms Analytic or formal models and compatible sheaves; a bare topological homeomorphism is insufficient.
Recover forms, currents, and their operations Test spaces, quotient relations, refinements, and continuity of the operations.

A smooth form has locally finite tropical presentations by definition. This fact is not a consequence that every current factors through one finite tropicalization. Likewise, a Hodge theorem on selected tropical spaces does not automatically pass through this inverse limit: the relevant complexes, transition maps, cohomology, and exactness must first be compared.

6.2 Fans and extended tropicalization: keeping the boundary

Cones specify which limits are admitted

A torus excludes the zero and pole loci of its characters. This is convenient locally, but a projective variety cannot in general be embedded in one torus. To describe a compactification, tropical geometry must allow some character valuations to tend to +∞+\infty.

A rational polyhedral cone σ⊂NR\sigma\subset N_{\mathbb R} is generated by finitely many lattice vectors with nonnegative real coefficients. It is strongly convex if it contains no line. Its dual cone and character monoid are σ∨={m∈MR:⟨m,w⟩≥0 for every w∈σ},Sσ=σ∨∩M.\sigma^\vee=\{m\in M_{\mathbb R}:\langle m,w\rangle\geq0\text{ for every }w\in\sigma\}, \qquad S_\sigma=\sigma^\vee\cap M. The associated affine toric variety is Uσ=Spec⁡K[Sσ].U_\sigma=\operatorname{Spec}K[S_\sigma]. The monoid consists precisely of the characters permitted to extend as regular functions on this affine chart. For σ=R≥0⊂R\sigma=\mathbb R_{\geq0}\subset\mathbb R, it is N\mathbb N, and Uσ=A1U_\sigma=\mathbb A^1; for the zero cone it is all of MM, and the chart is the torus.

A fan Σ\Sigma is a finite collection of strongly convex rational polyhedral cones, closed under faces, with pairwise intersections common faces. The affine varieties UσU_\sigma glue to the normal toric variety XΣX_\Sigma. A cone of dimension kk corresponds to a torus orbit of dimension r−kr-k. The fan is complete if its support is all of NRN_{\mathbb R}; this is equivalent to properness of XΣX_\Sigma.

Two adjectives must be separated. A cone is simplicial when its ray generators are linearly independent over R\mathbb R. It is smooth or unimodular when its primitive ray generators are part of a Z\mathbb Z-basis of NN. The cone spanned by (1,0),(1,2)(1,0),(1,2) is simplicial but not smooth. This distinction matters when a theorem assumes a simplicial fan: it has not thereby assumed a smooth toric variety.

References. [Gub13] (Section 6.2–6.3 and Appendix A); [GK17] (Appendix A.7).

The topology is part of the construction

Set R∞=R∪{+∞}\mathbb R_\infty=\mathbb R\cup\{+\infty\}, with addition extended by a+∞=∞a+\infty=\infty. Neighbourhoods of +∞+\infty contain a tail (c,+∞](c,+\infty]. The map t↦e−tt\mapsto e^{-t}, with e−∞=0e^{-\infty}=0, identifies this space with R≥0\mathbb R_{\geq0}.

Definition 6.2.1. (Extended tropicalization of a toric chart). Define Uσtrop=Hom⁡mon(Sσ,R∞),U_\sigma^{\mathrm{trop}}=\operatorname{Hom}_{\mathrm{mon}}(S_\sigma,\mathbb R_\infty), where a monoid homomorphism is additive and sends 00 to 00. Give this set the topology of pointwise convergence. A finite set of monoid generators realizes it as a subspace of a finite product of copies of R∞\mathbb R_\infty.

For a global fan these extended charts assemble into the space NΣ=∐σ∈ΣN(σ),N(σ)=NR/span⁡R(σ).\begin{equation} N_\Sigma=\coprod_{\sigma\in\Sigma}N(\sigma), \qquad N(\sigma)=N_{\mathbb R}/\operatorname{span}_\mathbb R(\sigma). \end{equation} The extended charts glue along the opens associated to common faces, giving the topology on this disjoint union. It is not the disjoint-union topology. The stratum N(0)=NRN(0)=N_{\mathbb R} is dense, and a sequence can converge from it to a lower-dimensional stratum.

For x∈Uσanx\in U_\sigma^\mathrm{an}, the rule m⟼−log⁡∣χm(x)∣,−log⁡0=+∞,m\longmapsto-\log|\chi^m(x)|,\qquad -\log0=+\infty, is a monoid homomorphism and gives the extended tropicalization map. These maps glue to XΣan→NΣX_\Sigma^\mathrm{an}\to N_\Sigma. The orbit corresponding to σ\sigma maps to N(σ)N(\sigma). For a closed subvariety of XΣX_\Sigma, its extended tropicalization is the image under this map.

References. [Pay09] (Section 3, especially Remark 3.1); [BGK25] (Section 1.13).

Example 6.2.2. (The line, affine line and projective line). For Gm\mathbb G_m, the target is R\mathbb R. For A1\mathbb A^1 with coordinate TT, it is R∪{+∞}\mathbb R\cup\{+\infty\}, and T=0T=0 maps to +∞+\infty. For P1\mathbb P^1, the two rays of its complete fan add two endpoints: NΣ=R∪{−∞,+∞}.N_\Sigma=\mathbb R\cup\{-\infty,+\infty\}. The point T=0T=0 corresponds to +∞+\infty; the point T=∞T=\infty, with local coordinate T−1T^{-1}, corresponds to −∞-\infty. These are toric boundary points. The proper analytic space P1,an\mathbb P^{1,\mathrm{an}} itself has empty analytic boundary.

For P2\mathbb P^2, use the dense-torus coordinates z1=X1/X0z_1=X_1/X_0, z2=X2/X0z_2=X_2/X_0, with x=v(z1)x=v(z_1), y=v(z2)y=v(z_2). Its fan has rays (1,0),(0,1),(−1,−1)(1,0),(0,1),(-1,-1), and its maximal cones are the sectors between consecutive rays. The extended tropicalization has one dense R2\mathbb R^2, three one-dimensional strata R\mathbb R, and three zero-dimensional strata. A useful topological model is the closed triangle (x,y)⟼(1,e−x,e−y)1+e−x+e−y.(x,y)\longmapsto \frac{(1,e^{-x},e^{-y})}{1+e^{-x}+e^{-y}}. The interior is R2\mathbb R^2, open sides are the one-dimensional strata, and vertices are the zero-dimensional strata. This model is a homeomorphism, not an identification of integral affine structures with the usual affine structure of a bounded triangle.

The fan of the projective plane and its strata mapped onto a triangle
A cone of dimension kk produces a boundary stratum of dimension 2−k2-k. Thus the two-dimensional cone σ12\sigma_{12} gives the vertex [1:0:0][1:0:0], while the ray e1e_1 gives the open side X1=0X_1=0. The triangle records topology and strata.

Boundary compatibility and Green functions

A function or form defined on the dense NRN_{\mathbb R} need not extend across these strata. The smooth tropical calculus imposes compatibility with projection to the lower-dimensional stratum: near the boundary, the form must come from that stratum in the directions that disappear. On R∞\mathbb R_\infty, an ordinary smooth function in this sense is constant for all sufficiently large tt, locally near +∞+\infty; a positive-degree form vanishes there. A function such as tt instead has the affine growth expected of a Green potential for a boundary divisor.

This is why extended tropicalization contributes more than a compact picture. It records where a divisor lies and what growth a Green function must have. BGK25 develops the compatible delta-calculus, Green functions and star-products needed to compute local heights on a suitable tropical toric compactification. Its subdivision theorems and boundary conditions are additional results; the affine stable-intersection construction alone does not supply them.

There is a complementary global lesson. Over a complete, algebraically closed, nontrivially valued field, Payne’s theorem identifies the analytification of a quasiprojective variety with the inverse limit of its extended tropicalizations over all quasiprojective toric embeddings. One tropicalization forgets information; the compatible system recovers it. This topological reconstruction does not say that every current, metric or potential is represented on a single finite tropicalization.

References. [Pay09] (Theorem 4.2); [BGK25] (Sections 1.7–1.11 and 3). For a systematic further introduction, see [MS15].

Example 6.2.3. (Limits in the extended tropical plane). In the triangle coordinates above, the sequences (n,n)(n,n), (n,0)(n,0), and (−n,−n)(-n,-n) have respective limits (1,0,0),(1/2,0,1/2),(0,1/2,1/2).(1,0,0),\qquad (1/2,0,1/2),\qquad (0,1/2,1/2). The first sequence escapes through the interior of the cone generated by e1,e2e_1,e_2 and reaches its zero-dimensional stratum. The second escapes along the ray e1e_1 and reaches the one-dimensional stratum X1=0X_1=0. The third escapes along −e1−e2-e_1-e_2 and reaches the one-dimensional stratum X0=0X_0=0. A divergent sequence need not approach a toric fixed point.

6.3 Skeleta, slope formulas, and tropical degree

Skeleta of semistable pairs

For a strictly semistable pair (X,H)(\mathfrak X,H), the skeleton is assembled from the standard simplices of local semistable equations, with unbounded directions contributed by the horizontal divisor HH. Its faces correspond to strata. The resulting integral-affine structure remembers the orders of vanishing of functions along model components and horizontal divisors.

In dimension one this is a metrized graph. In higher dimension, the normal direction along a codimension-one face carries data involving the neighbouring strata and intersection numbers. The higher-dimensional slope formula is therefore not obtained by summing arbitrary Euclidean normal derivatives. The model geometry specifies the correct weights and normalizations.

Slope formulas and principal divisors

If a rational function has divisor supported in the allowed boundary, its logarithmic absolute value restricts to an integral-affine function on each canonical cell of the skeleton. The slope formula expresses the balancing of its divisor data. On a curve it reduces to the familiar sum of outgoing slopes. In higher dimension it uses the incidence and intersection data of the semistable model.

This result is a bridge between algebraic principal divisors and tropical principal divisors. It tells us that the polyhedral data obtained from an analytic function obey the conservation law needed for integration. Conversely, an arbitrary PL function on a chosen skeleton need not be the logarithm of one analytic unit on a neighbourhood; lifting prescribed slopes is a separate problem.

The Sturmfels–Tevelev mechanism

For a generically finite map between suitable varieties in tori, the pushforward of the tropical cycle carries lattice-index weights. In a region where the map is combinatorially well behaved, the target weight is obtained by summing mσ[Nτ:F(Nσ)]m_\sigma [N_\tau:F(N_\sigma)] over source cells mapping onto the target cell, with the generic degree of the algebraic map accounted for in the cycle pushforward formula. This is the tropical version of counting sheets of a finite map, including ramification and multiplicity.

GRW develops related degree formulas for skeleta and tropicalizations. The lesson for CLD is direct: if a form is presented by a larger moment map and then projected, the calibrated integral must transform by precisely these degrees. The lesson for GK is equally direct: a tropical intersection product has to descend through the same refinement maps.

Existence and scope of faithful tropicalization

The faithful tropicalization theorem says that suitable rational functions can be chosen so that a specified skeleton is mapped injectively and unimodularly on its cells. It does not say that every tropicalization is faithful. Nor does it say that an arbitrary analytic datum factors through the retraction to that skeleton.

The distinction can already be seen on an annulus: the coordinate TmT^m has slope mm and fails unimodularity when ∣m∣>1|m|>1 as an integer lattice index, while TT supplies the primitive parameter. Additional functions can separate cells that a first tropicalization identifies. In higher-dimensional proofs, constructing enough such functions is a geometric theorem involving models and rational functions.

The rational functions supplied for faithfulness may have divisors outside the original boundary divisor. Therefore the theorem should not be strengthened to a global factorization through the original skeleton without checking the actual functions and their domains.

Multiplicity one and the tropical skeleton

For a tropical fibre of multiplicity one, the unique Shilov boundary point gives a canonical candidate for a section of tropicalization. The continuity theorem in GRW applies on the multiplicity-one locus in its stated setting. The stronger uniqueness assertion for a continuous section includes an additional closure-of-interior hypothesis. These are different assertions.

The later paper Tropical Skeletons studies the collection of Shilov boundary points of tropical fibres and its topology and functorial behaviour. This is a more intrinsic construction relative to a tropicalization than simply choosing a semistable skeleton in advance. It helps explain when combinatorial points have canonical analytic lifts.

For the differential calculus, however, even a canonical section of one tropicalization does not eliminate all other charts. Equality of currents still requires every actual test. Chapter 24 explains why a topological inverse-limit theorem or a faithful section does not by itself transfer a Hodge theory.

References. [GRW16] (Theorems 6.12, 8.4, 9.5 and 10.6); [GRW17]; [Gub13].

Tropical contractions and SYZ fibrations.

Yamamoto’s 2026 theorem gives a further geometric use of tropicalization. For the Batyrev–Borisov degenerations over k((t))k((t)) specified in [Yam26] (Section 3), with kk algebraically closed of characteristic zero, a good unimodular subdivision and irreducible generic fibre, tropicalization identifies the essential skeleton piecewise integrally affinely with the dual intersection complex. Composing with a tropical contraction gives an affinoid torus fibration off its discriminant, with the prescribed integral affine structure [Yam26] (Theorem 1.1). This is a structured degeneration theorem; it supplies no general descent theorem for CLD currents or cohomology.

6.4 Recovering an affine analytification from finite coordinate data

Payne’s theorem has a useful elementary core: every seminorm identity involves only finitely many functions. We prove a precise affine version using actual tropical images. This also identifies the point at which a topological reconstruction stops short of a differential or cohomological reconstruction.

Let AA be a finitely generated reduced KK-algebra and X=Spec⁡AX=\operatorname{Spec}A. Fix a finite set F0F_0 of algebra generators. Let F\mathscr F be the directed set of finite subsets F⊂AF\subset A containing F0F_0. The functions in FF give a closed embedding ιF:X↪AF\iota_F:X\hookrightarrow\mathbb A^F. Use absolute-value coordinates and set TF={(∣f(x)∣)f∈F:x∈Xan}⊂R≥0F.T_F=\{(|f(x)|)_{f\in F}:x\in X^{\mathrm{an}}\}\subset\mathbb R_{\geq0}^F. For F⊂GF\subset G, coordinate projection maps TGT_G onto TFT_F. With v=−log⁡∣⋅∣v=-\log|\cdot|, the same construction takes values in (R∪{+∞})F(\mathbb R\cup\{+\infty\})^F. The value +∞+\infty records a zero and is essential for embeddings into affine space.

Theorem 6.4.1. (Affine coordinate reconstruction). The natural map Φ:Xan⟶lim←⁡F∈FTF\Phi:X^{\mathrm{an}}\longrightarrow\varprojlim_{F\in\mathscr F}T_F is a homeomorphism. Here the analytic space on the left is described by multiplicative seminorms on AA extending the absolute value of KK, and every TFT_F is the actual image defined above.

Proof. A point in the inverse limit gives compatible numbers sf≥0s_f\geq 0 for every f∈Af\in A: choose a finite FF containing ff and read its coordinate. Directedness and compatibility show that the answer does not depend on the choice. For a,b∈Aa,b\in A, choose one finite set containing a,b,a+b,ab,0,1a,b,a+b,ab,0,1. Its coordinate vector belongs to TFT_F, so it is realized by an analytic point. Consequently s0=0,s1=1,sab=sasb,sa+b≤max⁡(sa,sb).s_0=0,\quad s_1=1,\quad s_{ab}=s_as_b,\quad s_{a+b}\leq\max(s_a,s_b). Adding any c∈Kc\in K to the finite set gives sc=∣c∣s_c=|c|. Hence f↦sff\mapsto s_f is a multiplicative seminorm extending KK, and determines a point of XanX^{\mathrm{an}}. Its image is the original compatible family. This proves surjectivity. If two analytic points have the same family, they have the same value on every f∈Af\in A, so they are the same seminorm. This proves injectivity.

Every coordinate of Φ\Phi is an evaluation map, so Φ\Phi is continuous. Conversely, the topology of XanX^{\mathrm{an}} is generated by all evaluations x↦∣f(x)∣x\mapsto|f(x)|, and each of these is a coordinate projection from some TFT_F. The inverse map is therefore continuous. No compactness assumption on XanX^{\mathrm{an}} is needed for this final step. ◻

The proof uses actual images, not an unproved identification with a combinatorial tropical variety. The theorem equating analytic images with algebraically defined tropicalizations is a separate input. Likewise, gluing affine descriptions into the toric-embedding formulation for a quasiprojective variety is the content of the global theorem cited in Section 6.1.5. Keeping these ingredients separate makes the reconstruction argument reusable.

Proposition 6.4.2. (Finite-coordinate neighbourhoods). Every ordinary neighbourhood of a point in XanX^{\mathrm{an}} contains a neighbourhood defined by finitely many conditions on values ∣fi∣|f_i|. On the open subset where all fif_i are nonzero, these conditions can be written using a finite tropicalization by analytic units.

Proof. The evaluation topology has a subbasis consisting of inverse images of open subsets of R≥0\mathbb R_{\geq 0} under single evaluations. Finite intersections form a basis, proving the first assertion. On the common nonvanishing locus, each regular fif_i is an analytic unit and the change of coordinate s↦−log⁡ss\mapsto-\log s is a homeomorphism from R>0\mathbb R_{>0} to R\mathbb R. This gives the second assertion. If some fif_i vanishes at the point, extended coordinates or a different local construction are required; deleting that coordinate would discard information. ◻

6.5 What a refinement must preserve

The preceding theorem recovers points and topology because their defining relations are finite. Differential forms involve an additional quotient: two presentations may agree after restriction to tangent spaces of the tropical support. Integration also involves weights, and current convergence involves a test topology. These three structures cannot be read from the underlying inverse-limit topology alone.

Example 6.5.1. (A homeomorphism that loses the lattice). The map F:R→RF:\mathbb R\to\mathbb R, t↦2tt\mapsto 2t, is a homeomorphism and is integral affine, but its tangent-lattice map Z→Z\mathbb Z\to\mathbb Z has index two. For a top superform α=a(y)d′y∧d′′y\alpha=a(y)d'y\wedge d''y, ∫RF∗α=2∫Rα.\int_{\mathbb R}F^*\alpha=2\int_{\mathbb R}\alpha. Thus topological injectivity does not give unimodularity, nor does it preserve calibrated integration with weight one. This is why faithful tropicalization includes an integral-affine requirement.

Example 6.5.2. (One tropical coordinate does not determine all tests). On the analytic torus with coordinate zz, the map x↦−log⁡∣z(x)∣x\mapsto-\log|z(x)| forgets the values of z−az-a for a∈K×a\in K^\times. On the open locus z≠az\ne a, the latter function is a unit and produces the additional smooth coordinate −log⁡∣z−a∣-\log|z-a|. Points with the same ∣z∣|z| can have different ∣z−a∣|z-a|. A test depending on this additional coordinate therefore need not be a pullback from the first tropical line. An identity checked only against functions of −log⁡∣z∣-\log|z| has not yet been checked against all original smooth tests.

To pass a calculation through chart refinements, one consequently needs: pullback preserving the quotient of forms; compatibility of d′d' and d′′d'' with that pullback; the weighted pushforward formula for integration; and continuity on the chosen spaces of compactly supported tests. Chapters 7–9 establish or explicitly import these ingredients in their stated categories. An inverse limit of spaces supplies none of them merely by being a homeomorphism.

The same distinction applies to cohomology. A direct limit of complexes requires actual transition maps of complexes. An identification with a complex on XanX^{\mathrm{an}} then requires exactness or a comparison argument, and a statement about cohomology requires a further passage through kernels and images. This is why a tropical Hodge theorem cannot be transferred to Berkovich cohomology solely by citing Payne’s topological theorem. It does not make such a transfer impossible; it specifies the additional mathematical work.

References. [Pay09] (Theorem 4.2); [GRW16] (Sections 9–10). The affine reconstruction above is proved directly for the explicitly defined system of actual images.

6.6 Tropical fibres and their Shilov points

Section 3.4 established the criterion for a singleton Shilov boundary. We now apply it to tropical fibres, using the formal-model comparison rather than identifying two different reductions. Now let UU be an integral closed subvariety of a split torus over a complete, algebraically closed, nontrivially valued field. For a tropical point ww with coordinates in the value group, the fibre UwU_w is strictly affinoid. Its algebra AwA_w is a quotient of a Laurent affinoid algebra. Distinguish its quotient Gauss norm from its spectral seminorm:

∥f∥quot=inf⁡g↦f∥g∥w,∥f∥sp=sup⁡x∈Uw∣f(x)∣.\|f\|_{\mathrm{quot}}=\inf_{g\mapsto f}\|g\|_w,\qquad \|f\|_{\mathrm{sp}}=\sup_{x\in U_w}|f(x)|. A quotient norm is submultiplicative and need not itself be an analytic point. The polyhedral formal model has the initial degeneration as special fibre; the canonical formal model has the canonical reduction as special fibre. The latter is reduced, and these two special fibres need not agree.

GRW Sections 10.2–10.5 supply a finite comparison of the two formal models and the associated special-fibre projection formula. Under equidimensionality, tropical multiplicity one forces a single component of the canonical reduction, hence a singleton Shilov boundary. The unique point is the spectral seminorm. Identifying it with the quotient norm requires the stronger reducedness condition on the initial degeneration. For arbitrary real ww, the valued-extension and boundary-descent results are additional inputs. This discussion separates the geometric theorem from the elementary norm argument above.

6.7 Adding coordinates without changing the lattice

Lemma 6.7.1. Let NN be a lattice. Suppose A:N→ZrA:N\to\mathbb Z^r is injective with saturated image. For every integral map B:N→ZsB:N\to\mathbb Z^s, the map (A,B)(A,B) has saturated image.

Proof. The cokernel of AA is finitely generated and torsion-free, hence free. Therefore AA admits an integral left inverse RR. Composing RR with projection onto the first factor gives an integral left inverse of (A,B)(A,B):

(R∘pr⁡1)∘(A,B)=id⁡N.(R\circ\operatorname{pr}_1)\circ(A,B)=\operatorname{id}_N. The image of (A,B)(A,B) is thus a direct summand, in particular saturated. ◻

For tropicalizations, this lemma applies to each tangent lattice after taking a common integral-affine subdivision. It proves that adding coordinates preserves unimodularity wherever it was already established. The existence of a finite subdivision on a semistable skeleton is a separate assertion about the restrictions of rational functions.

Example 6.7.2. On an annular skeleton, the map t↦mtt\mapsto mt is injective for every nonzero integer mm, but its lattice index is ∣m∣|m|. The map t↦(mt,t)t\mapsto(mt,t) is unimodular because projection to the second coordinate is an integral left inverse. More generally, t↦(at,bt)t\mapsto(at,bt) is unimodular exactly when gcd⁡(a,b)=1\gcd(a,b)=1. Thus injectivity and preservation of the integral lattice are distinct requirements.

6.8 Finite separation on a compact skeleton

Proposition 6.8.1. Let SS be a compact polyhedral space, and let F\mathcal F be a family of continuous real functions separating its points. Suppose a finite tuple F0F_0 drawn from F\mathcal F is locally injective on SS and unimodular on a finite subdivision. Assume every further function is integral affine on a finite subdivision. Then a finite tuple extending F0F_0 gives a faithful embedding of SS into a real vector space.

Proof. Choose finitely many open sets ViV_i covering SS such that F0F_0 is injective on each ViV_i. Set

O=⋃i(Vi×Vi),C=(S×S)∖O.O=\bigcup_i(V_i\times V_i),\qquad C=(S\times S)\setminus O. The set OO is an open neighborhood of the diagonal. Consequently CC is compact and contains no diagonal point. For every pair (x,y)(x,y) in CC, some function ff in F\mathcal F separates xx and yy. The inequality f(x)≠f(y)f(x)\ne f(y) persists on an open neighborhood of the pair. Compactness gives finitely many such functions which separate every pair in CC.

Together with F0F_0 these functions separate all points of SS: a pair in OO is separated by F0F_0, and a pair outside OO by one of the added functions. Take a common finite subdivision. Lemma 6.7.1 preserves unimodularity on its cells. Finally, a continuous injection from a compact space into a Hausdorff real vector space is a homeomorphism onto its image. ◻

Remark 6.8.2. Local injectivity on SS is stronger than injectivity on each closed cell separately. Different cells meeting at a vertex may still be folded onto one another. Proposition 6.8.1 does not remove this hypothesis. Nor does its compactness argument apply directly to a noncompact skeleton with horizontal boundary.

In the GRW construction, semistable coordinates first supply integral-affine embeddings of individual canonical polyhedra. The global separation of different polyhedra uses additional rational functions and model geometry. Proposition 6.8.1 explains one possible finite assembly step when its stronger local hypothesis holds; it is not a replacement for the complete global theorem.

6.9 From faithful coordinates to differential calculus

Faithfulness concerns a specified subset of the analytic space. Its conclusions are injectivity and unimodularity on that subset. It does not say that every analytic function, smooth form, or current factors through a retraction onto the subset. A current is still tested on the full admitted test space, and a representative still requires compatibility under refinements.

The distinction matters already for a curve with good reduction: its smallest proper skeleton may be one point, while annular domains carry nonzero differential forms. Enlarging the system of charts recovers these directions. A finite faithful tropicalization of one skeleton therefore cannot by itself transfer a global Hodge or Green theory.

References. [GRW16] (Section 9 and Sections 10.1–10.6); [GRW17] (Sections 2 and 8).

6.10 Pure dimension away from the tropicalized boundary

A moment map on a dd-dimensional analytic space can have a lower-dimensional tropical image. Affinoidness and avoidance of the analytic boundary restore the expected dimension. The proof is useful beyond the dimension theorem: it turns a tropical projection into a finite morphism of affinoid spaces and then reads dimension at a Shilov boundary point.

Lemma 6.10.1. (A simultaneous integral projection). Let P⊂RnP\subset\mathbb R^n be a finite polyhedral complex of dimension at most dd, and let B⊂PB\subset P be contained in finitely many polyhedra of dimension at most d−1d-1. For ω∈P∖B\omega\in P\setminus B, there is an integral linear map Q:Rn→RdQ:\mathbb R^n\to\mathbb R^d which is injective on every face of PP and satisfies Qω∉Q(B)Q\omega\notin Q(B). Assume that the stated finite polyhedral cover of BB consists of closed polyhedra, none containing ω\omega.

Proof. For each face σ\sigma of PP, require QQ to be injective on its tangent space LσL_\sigma. For each polyhedron τ\tau in the chosen cover of BB, require injectivity on the linear span of τ−ω\tau-\omega. The first spaces have dimension at most dd, and the second have dimension at most dim⁡τ+1≤d\dim\tau+1\leq d. For each of these finitely many subspaces, injectivity is a nonempty open condition on the space of real d×nd\times n matrices; its complement is given by vanishing of all relevant maximal minors and has empty interior. The intersection of the conditions is therefore nonempty and open. It contains a rational matrix. Clearing denominators gives an integral matrix without changing injectivity.

If b∈Bb\in B had Qb=QωQb=Q\omega, choose a covering polyhedron τ\tau containing bb. Then b−ω≠0b-\omega\ne 0 would lie in the kernel of QQ on span⁡(τ−ω)\operatorname{span}(\tau-\omega), a contradiction. ◻

In the analytic application, BB is compact and polyhedral, so a finite subdivision provides a closed polyhedral cover avoiding ω\omega. The last condition in the lemma is material. Merely knowing that ω\omega is not in a subset of a polyhedron would not imply that projection separates it from the whole covering polyhedron.

Theorem 6.10.2. (Affinoid tropical purity). Let KK be complete and nontrivially valued, let XX be a strict affinoid space of pure dimension dd, and let φ:X→(Gmn)an\varphi:X\to(\mathbb G_m^n)^{\mathrm{an}} be a moment map. Then φtrop(X)\varphi_{\rm trop}(X) has pure dimension dd at every point of φtrop(X)∖φtrop(∂X).\varphi_{\rm trop}(X)\setminus\varphi_{\rm trop}(\partial X).

Open proof obligation (RC2). Prove the local germ and finite-projection step in the stated dimension, support and base-change scope; GRPL Theorem 6.4 and the proof of 6.6.

NAG-900030 · Proof status

Remark 6.10.3. A constant moment map from a positive-dimensional proper space has a point as tropical image, although the source is boundaryless. This does not contradict the theorem, because the source is not affinoid. Conversely, a constant moment map from a positive-dimensional affinoid space sends its boundary to the same point, so the asserted open set is empty. Both hypotheses have visible roles.

This result concerns smooth tropicalization by analytic units. It does not prove purity of the image of a subvariety under the canonical tropicalization of a general abelian variety. That map comes from Raynaud uniformization and may forget an abelian part with good reduction. The distinction becomes important in Chapter 21, where canonical measures can live on cells of several dimensions.

References. [GRPL] (Sections 6.1–6.5 and Theorem 6.6). The proof above expands the finite-projection and Shilov-boundary argument; analytic dimension, properness, reduction and base-change results are explicit geometric inputs.

6.11 Matroids and the local models of tropical manifolds

Balancing is enough for Stokes, but not for all cohomological dualities. The matroidal local models used in tropical Poincaré duality satisfy a stronger incidence structure. This section constructs their fans from finite combinatorial data and proves balancing directly.

Definition 6.11.1. A matroid on a finite set EE is specified by a rank function r:2E→Z≥0r:2^E\to\mathbb Z_{\geq 0} satisfying 0≤r(A)≤∣A∣0\leq r(A)\leq|A|, monotonicity, and submodularity r(A∪B)+r(A∩B)≤r(A)+r(B).r(A\cup B)+r(A\cap B)\leq r(A)+r(B). Its closure is cl⁡(A)={e:r(A∪{e})=r(A)}\operatorname{cl}(A)=\{e:r(A\cup\{e\})=r(A)\}, with AA included. A flat is a set equal to its closure. A circuit is a minimal dependent set, where dependence means r(A)<∣A∣r(A)<|A|. We assume the matroid is loopless, meaning r({e})=1r(\{e\})=1 for every ee. Its rank is r(E)r(E).

For vectors (ve)e∈E(v_e)_{e\in E} over a field, r(A)=dim⁡span⁡{ve:e∈A}r(A)=\dim\operatorname{span}\{v_e:e\in A\} gives the basic example. The axioms do not require such a representation. Thus a matroidal tropical model need not arise from a linear subspace over the particular ground field under consideration.

Lemma 6.11.2. For e∉Ae\notin A, one has e∈cl⁡(A)e\in\operatorname{cl}(A) if and only if some circuit CC satisfies e∈C⊆A∪{e}e\in C\subseteq A\cup\{e\}.

Proof. Choose a maximal independent subset BB of AA. Its cardinality is r(A)r(A), and adding ee is dependent exactly when r(A∪{e})=r(A)r(A\cup\{e\})=r(A). In that case a minimal dependent subset of B∪{e}B\cup\{e\} is a circuit and must contain ee, since BB is independent. Conversely, in a circuit CC, deleting any element leaves an independent set of rank ∣C∣−1=r(C)|C|-1=r(C). Thus e∈cl⁡(C∖{e})e\in\operatorname{cl}(C\setminus\{e\}), and closure is monotone, giving the assertion for AA. The facts about maximal independent sets and monotone closure follow from the rank axioms by adding one element at a time and applying submodularity. ◻

The circuit and flag descriptions

Let N=ZE/Z1N=\mathbb Z^E/\mathbb Z\mathbf 1, where 1=∑e∈Eee\mathbf 1=\sum_{e\in E}e_e. Define the Bergman support by the minimum convention ∣B(M)∣={w∈RE/R1:#argmin⁡e∈Cwe≥2 extforeverycircuitC}.|B(M)|=\bigl\{w\in\mathbb R^E/\mathbb R\mathbf1: \#\operatorname{argmin}_{e\in C}w_e\geq2\ ext{for every circuit }C\bigr\}. Adding a constant to all coordinates does not change this condition. For a nonempty proper flat FF, let eFe_F be the class of its indicator vector. A flag of proper flats F=(F1⊊⋯⊊Fk)\mathscr F=(F_1\subsetneq\cdots\subsetneq F_k) gives the cone σF=R≥0eF1+⋯+R≥0eFk.\sigma_{\mathscr F}=\mathbb R_{\geq0}e_{F_1}+\cdots+\mathbb R_{\geq0}e_{F_k}.

Proposition 6.11.3. The cones σF\sigma_{\mathscr F} form a fan with support ∣B(M)∣|B(M)|. If r=r(E)r=r(E), this fan is pure of dimension r−1r-1. Each cone is unimodular in the lattice of its linear span.

Proof. A weight ww satisfies the circuit condition exactly when each nonempty proper upper level set Fc={e:we>c}F_c=\{e:w_e>c\} is a flat. If such a level set is not flat, choose e∉Fce\notin F_c in its closure. Lemma 6.11.2 gives a circuit contained in Fc∪{e}F_c\cup\{e\}, with ee its unique minimum. Conversely, a unique minimum ee in a circuit is separated from the other circuit weights by a level cc, making ee lie in the closure of FcF_c but not in FcF_c.

The distinct upper level sets form a nested flag. Subtracting the minimum weight expresses ww uniquely as a positive linear combination of their indicator vectors. Allowing some coefficients to vanish gives the faces of the cones. Thus intersections of cones are their common flag faces, proving the fan assertion. Strict inclusions of flats increase rank. Every flag can be filled to one containing a flat of each rank 1,…,r−11,\ldots,r-1, by adjoining an element outside the current flat and taking closure. Hence maximal cones have r−1r-1 generators.

To check the lattice, put F0=∅F_0=\varnothing, Fk+1=EF_{k+1}=E, and choose an element bi∈Fi∖Fi−1b_i\in F_i\setminus F_{i-1} for 1≤i≤k+11\leq i\leq k+1. For 1≤i≤k1\leq i\leq k, the integral coordinate difference w↦wbi−wbi+1w\mapsto w_{b_i}-w_{b_{i+1}} descends to NN and takes the value δij\delta_{ij} on eFje_{F_j}. These differences give an integral left inverse to the generator map Zk→N\mathbb Z^k\to N. Its image is saturated, proving both linear independence and unimodularity. ◻

Theorem 6.11.4. (Weight-one balancing). Assign weight one to each maximal cone of the flag fan. The resulting weighted fan is balanced.

Proof. A codimension-one flag omits exactly one rank between two successive retained flats A⊂BA\subset B, with r(B)=r(A)+2r(B)=r(A)+2; allow A=∅A=\varnothing and B=EB=E. The maximal cones containing it correspond to intermediate flats GG of rank r(A)+1r(A)+1. In the rank-two minor (M∣B)/A(M|B)/A, the sets G∖AG\setminus A are its rank-one flats. They partition B∖AB\setminus A: each element has a unique rank-one closure, and two such closures either agree or are disjoint. If there are ss of them, their indicator vectors satisfy ∑GeG=eB+(s−1)eA.\sum_G e_G=e_B+(s-1)e_A. Modulo the span of the codimension-one cone, the right side vanishes; when B=EB=E, its vector already vanishes in NN. By the preceding unimodularity argument, the image of each eGe_G is the primitive outgoing generator for its adjacent cone. Thus the displayed identity is exactly the balancing equation with all weights one. Rank one has no balancing faces to check. ◻

The proof uses only the rank-two intervals in the lattice of flats. This explains why matroidal fans have particularly controlled local stars. It also separates two uses of the word smooth: a unimodular cone is a lattice property, whereas being a tropical manifold requires every local model to have the specified matroidal form. A fan all of whose individual cones are unimodular need not have that local form.

References. [AK]; the use of these local models in tropical Poincaré duality is [JSS] (Definition 4.15 and Theorem 4.33).

6.12 Accessible skeleta and the limits of functoriality

A semistable skeleton is constructed from a model and comes with its associated retraction. The tropical skeleton of a fixed torus embedding is defined through Shilov boundary points of tropical fibres. Neither description by itself supplies a class stable under all the operations needed in analytic geometry. Ducros–Thuillier introduce accessible skeleta to address this functorial problem. We retain the coefficient data cc and the Γ\Gamma-strict category of their paper. Locally in the GG-topology, an accessible skeleton is covered by compact pieces that are images of elementary classical skeleta under analytic morphisms into compact analytic domains. The image is not required to consist only of points with maximal possible Abhyankar rank in the target. This is one reason the accessible category is larger than the classical category.

Theorem 6.12.1. (Ducros–Thuillier, fixed version). In the category of [DT26], accessible skeleta form the smallest admissible class containing the standard torus skeleta. This class contains classical skeleta, is stable under finite intersections and the specified piecewise linear subspaces, and has the following functorial properties.

  1. If f:Y→Xf:Y\to X and T⊂YT\subset Y is accessible, with f∣Tf|_T topologically proper, then f(T)f(T) is accessible.

  2. If Σ⊂X\Sigma\subset X is accessible and every fibre over Σ\Sigma is empty or zero-dimensional, then f−1(Σ)f^{-1}(\Sigma) is accessible.

  3. The function x↦dK(x)x\mapsto d_K(x) is locally constant on an accessible skeleton. Its locus where dK(x)=dim⁡xXd_K(x)=\dim_xX is open and closed and is classical.

Open proof obligation (RC2). Prove the accessible-skeleton geometric image assertion; this block explicitly offers DT26 and APW26 as imported proof.

NAG-900034 · Proof status

An accessible skeleton is not, by its definition, a deformation retract of the ambient space. Nor does the image theorem assign the tropical multiplicities needed to push differential forms. To use it in an integration formula one still needs weights, lattice normalizations and compatibility with calibration. To use it in a Green-form construction one additionally needs descent in the chosen sheaf of actual forms. These are separate inputs. There is a further distinction from faithfulness. Faithful tropicalization concerns the behaviour of a chosen map on a chosen skeleton. In the strong faithfulness used in [APW26], injectivity on the relevant dual complex is supplemented by weight one on its top-dimensional tropical faces. The additional weight condition allows multiplicity formulas to control other branches above generic tropical points. A homeomorphism on the chosen skeleton alone supplies no such control.

References. [DT26] (Theorems 9.1.12 and 9.2.3); [APW26] (Section 12.4).