11 · Psh functions and generated classes

149 NAG Tags

All continuity and local-uniform approximation in this chapter use the ordinary Berkovich topology, not the path metric on a skeleton. Example 4.2.1 shows why these topologies cannot be interchanged. A neighborhood on which an approximating family is defined is fixed before taking its limit; restricting to an analytic domain is not a substitute for proving such a neighborhood exists.

11.1 Continuous potentials, smooth approximation, and curvature

Assume for this section that XX has no analytic boundary. A function is locally integrable in the current formalism if it is locally integrable against the measures represented by smooth top-degree test forms. There is no unspecified ambient Lebesgue measure on XX in this definition. For a continuous real function uu, the current ⟨[u],α⟩=∫Xuα\langle[u],\alpha\rangle=\int_Xu\alpha is well-defined and has bidegree (0,0)(0,0).

Definition 11.1.1. (The continuous psh condition in CLD 2012). A continuous real function uu is psh in the sense of [CLD12] (Definition 5.5.1) if d′d′′[u]d'd''[u] is a positive current. It is locally smoothly psh-approximable if each point has a neighborhood on which uu is a uniform limit of original smooth psh functions.

The second condition is Definition 5.6.3 of the same source. It has an existential approximation requirement that does not occur in the first condition.

Proposition 11.1.2. (Stability under locally uniform limits). Let uju_j be continuous current-psh functions converging locally uniformly to a finite continuous function uu. Then uu is current-psh. In particular, local smooth psh approximation implies the continuous psh condition of Definition 11.1.1.

Proof. Let α\alpha be a positive smooth test form of bidegree (n−1,n−1)(n-1,n-1), and choose a compact neighborhood CC of its support. The smooth top-degree form d′d′′αd'd''\alpha represents a signed measure μα\mu_\alpha of finite total variation on CC. Hence ∣⟨d′d′′[uj−u],α⟩∣=∣∫C(uj−u) dμα∣≤∥uj−u∥C ∣μα∣(C)⟶0.\left|\langle d'd''[u_j-u],\alpha\rangle\right| =\left|\int_C(u_j-u)\,d\mu_\alpha\right| \leq\|u_j-u\|_C\,|\mu_\alpha|(C)\longrightarrow0. Each ⟨d′d′′[uj],α⟩\langle d'd''[u_j],\alpha\rangle is nonnegative, so its limit is nonnegative. Positivity is local and the test form was arbitrary. ◻

Remark (What the 2012 construction does not establish). The converse to Proposition 11.1.2 is not proved in CLD 2012; [CLD12] (Remark 5.6.4) explicitly records this difficulty. Consequently, a theorem whose hypotheses require local smooth psh approximation cannot be invoked merely because a continuous function has positive d′d′′d'd''-current.

The sentence following [CLD12] (Definition 5.5.1) also asserts functoriality for all continuous current-psh functions. Later work distinguishes current-psh from functorial psh and does not supply that general assertion as an automatic consequence of the current definition. It must therefore be cited with a separately verified hypothesis, rather than proved by an undefined pullback of currents. Smooth potentials and local uniform limits of smooth psh potentials are legitimate settings for the corresponding pullback argument; the smooth comparison is also covered by [GK19] (Corollary 8.9).

Example 11.1.4. (A corner contributes to the current). On an interval around 00, set u(t)=−∣t∣u(t)=-|t|. Both restrictions to t<0t<0 and t>0t>0 are affine, hence convex. Nevertheless, for a smooth compactly supported test function χ\chi, ⟨d′d′′[u],χ⟩=∫R−∣t∣ χ′′(t) dt=−2χ(0).\langle d'd''[u],\chi\rangle =\int_{\mathbb R}-|t|\,\chi''(t)\,dt=-2\chi(0). Thus d′d′′[u]=−2δ0d'd''[u]=-2\delta_0 and uu is not psh. The identical calculation applies in a tropical coordinate on an annulus. The facewise Hessians vanish; the negative contribution lies at the corner.

For a tropical curve with three outgoing primitive directions (1,0)(1,0), (0,1)(0,1), (−1,−1)(-1,-1), a continuous piecewise affine potential with outgoing slopes a,b,ca,b,c has vertex mass a+b+ca+b+c. Choosing a+b+c<0a+b+c<0 gives the counterexample in [GK19] (Example 3.3) to the continuous version of [CLD12] (Lemma 5.5.3). The smooth restriction of that criterion is the one retained in [GK19] (Remark 3.2).

Theorem 11.1.5. (The mixed product constructed in CLD 2012). Let u1,…,umu_1,\ldots,u_m be locally smoothly psh-approximable continuous functions, where 0≤m≤n0\leq m\leq n. There is a positive current of bidegree (m,m)(m,m) denoted by d′d′′u1∧⋯∧d′d′′um.d'd''u_1\wedge\cdots\wedge d'd''u_m. On an open subset where ui,ju_{i,j} are smooth psh functions converging uniformly to uiu_i, its value on a test form α\alpha is lim⁡j→∞∫Xd′d′′u1,j∧⋯∧d′d′′um,j∧α.\lim_{j\to\infty}\int_X d'd''u_{1,j}\wedge\cdots\wedge d'd''u_{m,j}\wedge\alpha. The limit is independent of the approximating sequences and is compatible with restriction. If m=nm=n, it is a positive measure.

Open proof obligation (RC2). Prove the native mixed-current assertion under its local uniform approximation hypothesis. CLD12 Corollary 5.6.5 is supplied here; the described internal route also has open multiplier convergence.

NAG-900104 · Proof status

Definition 11.1.6. (Metric weights and curvature). Let LL be a line bundle with a finite continuous metric. For a local nonvanishing frame ee, write φe=−log⁡∥e∥,c1(L,∥⋅∥)∣U=d′d′′[φe].\varphi_e=-\log\|e\|, \qquad c_1(L,\|\cdot\|)|_U=d'd''[\varphi_e]. The metric is current-psh if its local weights satisfy Definition 11.1.1.

Proposition 11.1.7. (Frame independence and the Poincaré–Lelong sign). The local curvature currents in Definition 11.1.6 agree on overlaps. If ZZ is an algebraic variety and ss a nonzero meromorphic section of LL, then c1(L,∥⋅∥)=d′d′′[−log⁡∥s∥]+δdiv⁡(s).c_1(L,\|\cdot\|) =d'd''[-\log\|s\|]+\delta_{\operatorname{div}(s)}.

Open proof obligation (RC2). Prove the scalar divisor/current identity with the stated sign and scope; CLD12 Theorem 4.6.5 or the cited Gubler alternative is the premise.

NAG-900105 · Proof status

A local smooth approximation of each weight, a global smooth approximation of a metric on a fixed line bundle, and an approximation by relatively nef algebraic models are three different requirements. The first two are already separated in [CLD12] (§6.3.1); the third belongs to model positivity. The weak convergence of curvature currents proved in Proposition 11.1.2 does not identify these approximation categories. Likewise, adding a constant to every local weight leaves the curvature unchanged but changes the metric. Curvature comparison alone does not identify metrized line bundles.

11.2 The change of psh definition from 2012 to 2025

We work on a good Hausdorff boundaryless analytic space of pure dimension nn and use the original smooth-form sheaf. An original smooth function is called psh when the smooth (1,1)(1,1)-form d′d′′ud'd''u is positive. The smooth-form/current comparison identifies this with positivity of d′d′′[u]d'd''[u] in this regularity class; see [CLD25] (Corollary 7.3.10 and Theorem 12.11.1). The issue in this section is the passage to nonsmooth functions.

The classes that must remain separate

Definition 11.2.1. (Local approximation classes). For an open set UU, let U(U)\mathcal U(U) be the continuous real functions uu such that every x∈Ux\in U has an open neighbourhood V⊂UV\subset U on which uu is the locally uniform limit of a sequence of original smooth psh functions defined on that same VV. Let M(U)\mathcal M(U) be the finite continuous functions which, locally on one fixed open neighbourhood, are decreasing limits of sequences of original smooth psh functions. Write Psh⁡12cont(U)\operatorname{Psh}_{12}^{\rm cont}(U) for the continuous current-psh functions of Definition 11.1.1.

The neighbourhood in the definition is fixed for the whole approximating sequence. Neither definition is satisfied merely by finding, for every accuracy parameter, a function which is psh on a different shrinking neighbourhood of the point.

Definition 11.2.2. (The two generated sheaves in CLD25). The sheaf P25\mathcal P_{25} is the smallest subsheaf of functions with values in R∪{−∞}\mathbb R\cup\{-\infty\} which contains the original smooth psh functions and is stable under decreasing limits of sequences on every open set. The sheaf C25\mathcal C_{25} is the smallest subsheaf of finite continuous functions which contains the same smooth functions and is stable under uniform limits. These are [CLD25] (Definitions 8.1.1 and 8.1.7).

The ordinal stages of the first construction are defined by P25,0={original smooth psh functions},\mathcal P_{25,0}=\{\text{original smooth psh functions}\}, with P25,α+1\mathcal P_{25,\alpha+1} consisting of local decreasing limits of sequences from P25,α\mathcal P_{25,\alpha}; at a limit ordinal one takes the sheaf of functions which locally belong to an earlier stage. In particular, the first stage P25,1\mathcal P_{25,1} has an explicit local decreasing smooth approximation. The generated sheaf includes the later stages as well.

The word “smallest” refers to intersection among subsheaves satisfying the two defining properties. It is an exact definition and does not assert that every resulting section is a one-step limit of smooth functions. The transfinite description in [CLD25] (Lemma 8.1.4) makes this distinction visible.

Proposition 11.2.3. (The continuous first stage). As sheaves of finite continuous functions, M=U=C0∩P25,1.\begin{equation} \mathcal M=\mathcal U =C^0\cap\mathcal P_{25,1}. \end{equation}

Proof. If uj↓uu_j\downarrow u on a fixed neighbourhood and uj,uu_j,u are continuous, Dini’s theorem gives uniform convergence on every compact subset. This proves M⊂U\mathcal M\subset\mathcal U.

Conversely, fix a smaller open neighbourhood VV with compact closure inside a common approximation neighbourhood. Choose smooth psh functions vjv_j such that ∥vj−u∥V‾≤4−j.\|v_j-u\|_{\overline V}\leq4^{-j}. Put wj=vj+2⋅4−jw_j=v_j+2\cdot4^{-j}. Addition of a constant preserves smooth psh. On VV, u+4−j≤wj≤u+3⋅4−j,wj+1≤u+34 4−j<wj.u+4^{-j}\leq w_j\leq u+3\cdot4^{-j},\qquad w_{j+1}\leq u+\tfrac34\,4^{-j}<w_j. Hence wj↓uw_j\downarrow u on VV. This proves the reverse inclusion. The last equality is the definition of the first generated stage, with the continuous endpoint required explicitly. ◻

Proposition 11.2.4. (Implications obtained without identifying the higher stages). In the setting above one has U ⊆ C25 ⊆ C0∩P25 ⊆ Psh⁡12cont.\begin{equation} \mathcal U\ \subseteq\ \mathcal C_{25}\ \subseteq\ C^0\cap\mathcal P_{25}\ \subseteq\ \operatorname{Psh}_{12}^{\rm cont}. \end{equation} More generally, a locally integrable section uu of P25\mathcal P_{25} has positive original current d′d′′[u]d'd''[u]. None of the reverse inclusions for general analytic spaces is inferred here from the definitions.

Proof. The first inclusion follows by restricting to relatively compact neighbourhoods and applying the uniform-limit closure property of C25\mathcal C_{25}. The generated psh sheaf is a cone, is stable under addition of constants, and is local. A locally uniform limit of finite members can therefore be replaced, on a smaller relatively compact neighbourhood, by a decreasing subsequence after adding positive constants, exactly as in Proposition 11.2.3. Thus C0∩P25C^0\cap\mathcal P_{25} contains smooth psh functions and is stable under uniform limits; the minimality of C25\mathcal C_{25} proves the second inclusion.

For the last assertion, use transfinite induction on the decreasing-limit construction. Suppose uj↓uu_j\downarrow u, where uu is locally integrable and each uju_j belongs to an earlier stage. All generated functions are upper semicontinuous, so uju_j is bounded above on compact sets. The inequalities u≤uj≤u1u\leq u_j\leq u_1 imply local integrability of uju_j; Proposition 11.6.7 records this inheritance at every ordinal. If α\alpha is a positive top-degree test form, monotone convergence applied after subtracting a local upper bound gives ∫ujα⟶∫uα.\int u_j\alpha\longrightarrow\int u\alpha. A general top-degree test is a difference of positive tests. Hence [uj]→[u][u_j]\to[u], and d′d′′[uj]⟶d′d′′[u].d'd''[u_j]\longrightarrow d'd''[u]. Positivity is preserved under this convergence because each positive test has a nonnegative value at every stage. Locality treats sheafification and limit ordinals. The smooth initial stage follows from integration of positive smooth forms. This supplies the usable argument for [CLD25] (Proposition 8.3.1(a)) without invoking Proposition 8.1.11 or the false weak-positive order-zero statement. ◻

Remark 11.2.5 (The equality printed in Proposition 8.1.11). CLD25 states the stronger equality C25=C0∩P25\mathcal C_{25}=C^0\cap\mathcal P_{25} and an equality at every ordinal stage. Its printed successor-stage argument starts with approximants in the continuous stage, whereas the decreasing-limit definition provides approximants in the possibly discontinuous stage. The missing replacement is discussed in Remark A.1.4. We use (11.2.2) and the proved first-stage equality (11.2.1); no argument in this book requires the stronger printed equality. This is a limitation of the available argument, not a counterexample to that equality. Sections 11.6.1–11.6.3 prove the compact-domain barrier theorem, the indexed forward inclusion and the locally integrable comparison, and isolate the additional domain requirement. A projective route, using Wu’s comparison and envelope inputs, is given in Section 11.7. Theorem 11.7.5 proves that bounded first-stage decreasing stability is sufficient for the full continuous-stage equality.

Remark 11.2.6 (Translation of the word approchable). In [CLD12] (Definition 5.6.3), “locally psh-approachable” is the positive approximation condition represented here by U\mathcal U, whereas “locally approachable” denotes a difference of two such functions. In [CLD25] (Definition 8.1.13), a psh function is “locally approachable” when it belongs to P25,1\mathcal P_{25,1}. Thus the latter expression is on the positive side of the theory. For finite continuous endpoints, Proposition 11.2.3 compares the positive classes; it does not identify the old difference class with the new positive class.

Pulling back functions and pulling back currents

Proposition 11.2.7. (Pullback for generated and approximable psh functions). Let f:Y→Xf:Y\to X be an analytic morphism in the preceding category. Pullback of functions preserves P25\mathcal P_{25} and U\mathcal U. In the generated class it preserves each ordinal stage.

Proof. Pullback of an original smooth function is original smooth, and pullback of its positive (1,1)(1,1)-form is positive. It commutes with composition and with pointwise decreasing limits. Transfinite induction, followed by locality, proves the assertion for P25\mathcal P_{25}, as in [CLD25] (Proposition 8.1.5). For U\mathcal U, pull back the smooth approximants on the inverse image of a common neighbourhood. A compact subset of that inverse image has compact image, so local uniform convergence is preserved. ◻

The proof constructs u∘fu\circ f and, whenever this function is locally integrable, its curvature current. Local integrability holds for continuous uu; a singular psh pullback may instead be identically −∞-\infty. The proof never defines a pullback of an arbitrary current. In particular, the notation d′d′′[u∘f]d'd''[u\circ f] does not by itself justify an operation f∗(d′d′′[u])f^*(d'd''[u]) on the whole current space. The general pullback assertion following [CLD12] (Definition 5.5.1) cannot be used as a proof of this operation or as a comparison with all current-psh functions; see Section A.1 and the functorial conditions in the delta-form section.

A complete comparison on real convex domains

Proposition 11.2.8. (Continuous distributional convexity). Let Ω⊂Rr\Omega\subset\mathbb R^r be open and convex, and let u:Ω→Ru:\Omega\to\mathbb R be continuous. The following are equivalent:

  1. uu is convex;

  2. d′d′′[u]d'd''[u] is a positive supercurrent;

  3. uu is locally a uniform limit of smooth convex functions;

  4. uu is locally a decreasing limit of smooth convex functions.

Proof. Convolution on a slightly larger relatively compact domain preserves convexity and gives smooth functions converging uniformly on smaller compact sets. Thus (1) implies (3). The shift construction in Proposition 11.2.3 proves (3) implies (4), and local uniform convergence of the corresponding distributions gives (3) implies (2).

If (2) holds, convolve the distributional Hessian with a nonnegative smooth kernel. The resulting matrix is positive semidefinite, so each regularization uεu_\varepsilon is convex by Proposition 7.2.4. Continuity gives uε→uu_\varepsilon\to u uniformly on compact subsets. Passing to the limit in the convexity inequality proves local convexity. Restricting to a line segment in Ω\Omega then gives global convexity on the convex domain. Finally, (4) implies (3) by Dini’s theorem. ◻

Remark 11.2.9 (Why continuity cannot be omitted from a pointwise statement). On (−1,1)(-1,1) define u(0)=1u(0)=1 and u(t)=0u(t)=0 for t≠0t\ne0. Then uu is locally integrable and even upper semicontinuous, and [u]=[0][u]=[0]. Thus d′d′′[u]=0d'd''[u]=0, but uu is not convex: u(0)=1>12u(−a)+12u(a)=0(0<a<1).u(0)=1>\tfrac12u(-a)+\tfrac12u(a)=0 \qquad(0<a<1). The implication from the locally integrable current condition to pointwise convexity in [CLD25] (Proposition 8.3.2) is therefore false literally. A distribution determines an almost-everywhere class; a convex representative must be specified, or continuity must be assumed as in Proposition 11.2.8. Upper semicontinuity alone does not repair this example.

11.3 Positive distributional Hessians and convex representatives

A current defined by a locally integrable function only remembers its almost-everywhere class. To repair a pointwise convexity assertion, one must first construct the canonical representative of that class. The continuous comparison in Section 11.2 is a special case of the result below; the proof also explains precisely why changing one point can invalidate the unrepaired statement.

Let Ω⊂Rn\Omega\subset\mathbb R^n be a nonempty convex open set. For u∈Lloc1(Ω)u\in L^1_{\rm loc}(\Omega), say that its distributional Hessian is positive if ∑i,jξiξj ∂i∂ju≥0as a distribution for every ξ∈Rn.\sum_{i,j}\xi_i\xi_j\,\partial_i\partial_j u\geq0 \quad\text{as a distribution for every }\xi\in\mathbb R^n. This is equivalent to ordinary positivity of DD[u]DD[u] in the real superform convention. It is a condition on a distribution, so it is unchanged by modifications on Lebesgue-null sets.

Lemma 11.3.1. (An interior bound from an L1L^1 bound). Let vv be a smooth convex function on B(0,4r)B(0,4r) and put M=∫B(0,4r)∣v∣ dxM=\int_{B(0,4r)}|v|\,dx. With Vr=vol⁡(B(0,r))V_r=\operatorname{vol}(B(0,r)), one has −3M/Vr≤v(x)≤M/Vr(x∈B(0,r)).-3M/V_r\leq v(x)\leq M/V_r\quad(x\in B(0,r)). The functions with a common MM bound are uniformly Lipschitz on B(0,r/2)B(0,r/2), with a constant depending only on M,r,nM,r,n.

Proof. For z∈B(0,3r)z\in B(0,3r), the average of points in B(z,r)B(z,r) is zz. Jensen’s inequality gives v(z)≤M/Vr=:Av(z)\leq M/V_r=:A. For x∈B(0,r)x\in B(0,r) and y∈B(x,r)y\in B(x,r), the point 2y−x2y-x lies in B(0,3r)B(0,3r); convexity gives v(y)≤12v(x)+12v(2y−x)≤12(v(x)+A).v(y)\leq\tfrac12v(x)+\tfrac12v(2y-x)\leq\tfrac12(v(x)+A). If v(x)≥−Av(x)\geq-A there is nothing to prove. Otherwise the right side is negative throughout that ball, and integration of ∣v∣|v| yields M≥Vr(−v(x)−A)/2M\geq V_r(-v(x)-A)/2. Thus v(x)≥−3Av(x)\geq-3A. For x∈B(0,r/2)x\in B(0,r/2) and a unit vector ee, the two points x±(r/2)ex\pm(r/2)e lie in B(0,r)B(0,r) after an arbitrarily small inward adjustment. The tangent-line inequality for a convex function bounds ∣∂ev(x)∣|\partial_ev(x)| by 8A/r8A/r. This gives a uniform gradient bound and the Lipschitz assertion. ◻

Theorem 11.3.2. (The convex representative theorem). A function u∈Lloc1(Ω)u\in L^1_{\rm loc}(\Omega) has positive distributional Hessian if and only if its almost-everywhere class contains a finite convex function vv. Such a representative is unique. Every standard compactly supported nonnegative smooth approximate identity ρε\rho_\varepsilon gives u∗ρε⟶vlocally uniformly on Ω.u*\rho_\varepsilon\longrightarrow v \quad\text{locally uniformly on }\Omega.

Proof. Assume first that the Hessian is positive. On the inner domain where convolution is defined, uε=u∗ρεu_\varepsilon=u*\rho_\varepsilon is smooth and its Hessian is positive semidefinite: for each vector ξ\xi, its quadratic Hessian coefficient is the convolution of the nonnegative distribution ∂ξ2u\partial_\xi^2u with a nonnegative kernel. Thus uεu_\varepsilon is convex on every convex inner subdomain.

The convolutions converge to uu in L1L^1 on every relatively compact open subset. Their L1L^1 norms are therefore uniformly bounded there for sufficiently small ε\varepsilon. Cover a given compact subset by finitely many balls whose fourfold enlargements remain in Ω\Omega. Lemma 11.3.1 gives uniform bounds and equicontinuity on the compact subset. Arzelà–Ascoli and a diagonal argument over an exhaustion give a subsequence converging locally uniformly to a finite continuous function vv. The convex inequality passes to the limit on each compact line segment, so vv is convex. Its L1L^1 limit is also uu, hence v=uv=u almost everywhere.

Two finite convex representatives are continuous. If they differed at one point, they would differ on a nonempty open set, which has positive Lebesgue measure. Hence they coincide. This uniqueness and the same local compactness argument show convergence of the whole family uεu_\varepsilon, not just the selected subsequence: any sequence violating uniform convergence on a compact set would have a locally uniformly convergent subsequence with the same unique limit.

Conversely, if u=vu=v almost everywhere for a finite convex vv, its convolutions are smooth convex functions and converge locally uniformly to vv. Their positive Hessian distributions converge to the Hessian distribution of uu. Positivity is closed under convergence on tests, proving the converse. ◻

Which pointwise representative is admissible

Corollary 11.3.3. The representative in Theorem 11.3.2 is characterized by v(x)=lim⁡r↓01vol⁡B(x,r)∫B(x,r)u(y) dy=lim⁡r↓0ess sup⁡B(x,r)u.v(x)=\lim_{r\downarrow0}\frac1{\operatorname{vol}B(x,r)}\int_{B(x,r)}u(y)\,dy =\lim_{r\downarrow0}\operatorname*{ess\,sup}_{B(x,r)}u. If the chosen pointwise function uu is strongly upper semicontinuous, in the sense that u(x)=lim sup⁡y→x, y∉Eu(y)u(x)=\limsup_{y\to x,\,y\notin E}u(y) for every Lebesgue-null set EE, then u=vu=v everywhere. Ordinary upper semicontinuity alone does not suffice.

Proof. The integral and essential supremum are unchanged by replacing uu with vv. Continuity of vv makes both limits equal to v(x)v(x). For the last assertion take E={u≠v}E=\{u\ne v\}, which is null. Its complement is dense in every open ball. Strong upper semicontinuity and continuity of vv therefore give u(x)=v(x)u(x)=v(x). To see the failure for ordinary upper semicontinuity, put u(0)=1u(0)=1 and u=0u=0 elsewhere on a ball containing zero. This function is upper semicontinuous and defines the zero distribution, but is not convex. ◻

This gives two precise replacements for the pointwise claim in CLD25 Proposition 8.3.2: either work with a continuous function from the outset, or work with an L1L^1 class and its unique convex representative. Strong upper semicontinuity selects that representative. The approximation converges uniformly to the representative, not to an arbitrarily modified pointwise version.

11.4 The Hessian kernel and potential independence on a real domain

Proposition 11.4.1. If a distribution ww on a connected open subset of Rn\mathbb R^n satisfies ∂i∂jw=0\partial_i\partial_jw=0 for all i,ji,j, then ww is the distribution of an affine function. Consequently two finite convex functions on a convex open set with the same Hessian current differ by an affine function.

Proof. On any ball compactly contained in the domain, small convolutions wεw_\varepsilon have zero Hessian, hence are affine on a smaller concentric ball. The finite-dimensional space of affine distributions on that smaller ball is closed in the distribution topology. To see this directly, choose n+1n+1 test functions whose integrals against 1,x1,…,xn1,x_1,\ldots,x_n form an invertible matrix; convergence on these tests forces convergence of all affine coefficients. Since wε→ww_\varepsilon\to w as distributions, ww is affine on the smaller ball. On overlapping balls the affine polynomials agree as distributions and hence as functions on an open set, so their coefficients agree. Connectedness glues them to one affine function. Apply this to the difference of the two functions and then use their continuity. ◻

Corollary 11.4.2. Let SS be an ordinary positive current closed under both differentials on a convex real domain. If finite convex functions u,vu,v have DD[u]=DD[v]DD[u]=DD[v], then DD(uS)=DD(vS).DD(uS)=DD(vS).

Proof. Both functions are continuous, so multiplication is defined by the order-zero theorem. Their difference a=u−va=u-v is affine by Proposition 11.4.1. The smooth Leibniz formula and closedness give DD(aS)=(DDa)∧S=0DD(aS)=(DDa)\wedge S=0. Subtracting the two products proves the result. ◻

On a singular tropical support, the kernel of the current Hessian can be larger than the affine functions inherited from the ambient space. The cross example in Chapter 14 exploits exactly that enlargement. Thus the Euclidean argument proves potential independence on its own domain and simultaneously identifies the step that cannot simply be carried over to an arbitrary quotient of charts.

References. This section supplies the representative-sensitive repair of [CLD25] (Proposition 8.3.2). The distinction between strong upper semicontinuity and a mere distributional class is also used in [BGJK25] (Section 3.1).

11.5 A convex potential for a positive closed real current

On a convex real domain, every positive closed (1,1)(1,1) supercurrent has a convex potential. The assertion concerns an ambient Euclidean domain. It is a useful comparison theorem for the local analytic problem, but its proof uses linear coordinates and cannot be transferred to a singular tropical quotient without checking those operations. We include the distributional step so that regularizing the current does not conceal a missing convergence argument for its potentials.

Lemma 11.5.1. (Positive Laplacian gives local integrability). Let uu be a distribution on an open subset of Rn\mathbb R^n. If Δu\Delta u is a positive distribution, then uu is represented by a locally integrable function.

Proof. A positive distribution is a positive Radon measure; write μ=Δu\mu=\Delta u. Choose a compactly supported smooth cutoff χ\chi equal to one on a smaller open set. A fundamental solution EnE_n for Δ\Delta is E1(x)=12∣x∣,E2(x)=12πlog⁡∣x∣,En(x)=−∣x∣2−n(n−2) area⁡(Sn−1)(n≥3).\begin{gathered}E_1(x)=\tfrac12|x|,\qquad E_2(x)=\tfrac1{2\pi}\log|x|,\\ E_n(x)=-\frac{|x|^{2-n}}{(n-2)\,\operatorname{area}(S^{n-1})}\quad(n\geq3).\end{gathered} Each kernel is locally integrable. The convolution v=En∗(χμ)v=E_n*(\chi\mu) is locally integrable by Fubini: on a compact set, the integral of ∣En(x−y)∣|E_n(x-y)| in xx is uniformly bounded for yy in the compact support of χμ\chi\mu. In distributions, Δv=χμ\Delta v=\chi\mu. Thus u−vu-v is a harmonic distribution on the smaller open set.

For completeness, a harmonic distribution is smooth. Its small convolutions are smooth harmonic functions. Fix a smooth radial averaging kernel supported in a ball small enough for the harmonic mean-value identity. Every such convolution equals its convolution with this fixed kernel on a smaller ball. Passing to distributions shows that the original distribution equals its convolution with the fixed kernel there, which is smooth. This is the distributional Weyl lemma. Consequently u=v+(u−v)u=v+(u-v) is locally integrable. ◻

Corollary 11.5.2. A distribution whose Hessian is positive semidefinite in the distributional sense is represented by a unique finite convex function on a convex domain.

Proof. Its Laplacian is a sum of positive distributions, hence positive. Lemma 11.5.1 gives a locally integrable representative, and Theorem 11.3.2 gives the unique finite convex representative. ◻

Theorem 11.5.3. (The real DDDD potential theorem). Let Ω⊂Rn\Omega\subset\mathbb R^n be convex and open. A symmetric ordinary positive (1,1)(1,1) supercurrent TT with d′T=d′′T=0d'T=d''T=0 has a finite convex potential uu satisfying DD[u]=TDD[u]=T. The potential is unique up to addition of an affine function.

Proof. Write T=∑i,jTij d′xi∧d′′xjT=\sum_{i,j}T_{ij}\,d'x_i\wedge d''x_j. Symmetry gives Tij=TjiT_{ij}=T_{ji}, and closedness gives ∂kTij=∂iTkj.\partial_kT_{ij}=\partial_iT_{kj}. The trace μ=∑iTii\mu=\sum_iT_{ii} is a positive Radon measure. Commuting distributional derivatives and using symmetry gives ΔTij=∑k∂k∂iTkj=∂i∂j∑kTkk=∂i∂jμ.\Delta T_{ij} =\sum_k\partial_k\partial_iT_{kj} =\partial_i\partial_j\sum_kT_{kk} =\partial_i\partial_j\mu.

First work on a relatively compact convex subdomain U⋐ΩU\Subset\Omega. Choose χ=1\chi=1 near its closure and form v=En∗(χμ)v=E_n*(\chi\mu). The preceding lemma’s argument makes vv locally integrable. On UU, the distributions Sij=Tij−∂i∂jvS_{ij}=T_{ij}-\partial_i\partial_jv are harmonic, hence smooth. They remain symmetric and satisfy ∂kSij=∂iSkj\partial_kS_{ij}=\partial_iS_{kj}. Translate an interior base point of UU to zero and define aj(x)=∫01∑ixiSij(tx) dt.a_j(x)=\int_0^1\sum_i x_iS_{ij}(tx)\,dt. Differentiating and using the closed-column identity gives ∂kaj(x)=∫01(Skj(tx)+t∑ixi∂iSkj(tx))dt=Skj(x).\partial_ka_j(x) =\int_0^1\left(S_{kj}(tx)+t\sum_i x_i\partial_iS_{kj}(tx)\right)dt =S_{kj}(x). Symmetry now says that the smooth one-form ∑jajdxj\sum_ja_jdx_j is closed. A second radial integration, b(x)=∫01∑jxjaj(tx) dt,b(x)=\int_0^1\sum_jx_ja_j(tx)\,dt, gives ∂jb=aj\partial_jb=a_j and D2b=SD^2b=S. Hence the locally integrable function v+bv+b has Hessian TT on UU. Positivity and Theorem 11.3.2 replace it by a finite convex representative uUu_U.

Exhaust Ω\Omega by nested relatively compact convex open subdomains UjU_j. Such an exhaustion is obtained by intersecting large balls with inner parallel convex domains. On UjU_j, the difference uUj+1−uUju_{U_{j+1}}-u_{U_j} has zero Hessian, hence is affine by Proposition 11.4.1. Subtract this affine function from uUj+1u_{U_{j+1}} and proceed inductively. Affine subtraction preserves convexity, and the adjusted potentials agree on overlaps. They glue to the required finite convex function on Ω\Omega. The same Hessian-kernel proposition gives uniqueness modulo affine functions. ◻

This proof makes two uses of the ambient real structure explicit. Coordinate derivatives commute as distributions, allowing the Laplacian calculation; and straight segments remain in the convex domain, allowing the two radial integrations. The equation DD[u]=TDD[u]=T on a tropical support is an equality modulo tangential relations. It need not furnish an ambient matrix TijT_{ij} satisfying these unrestricted coordinate identities. This is the precise obstruction to treating the real potential theorem as a local-potential theorem on every Berkovich chart.

Corollary 11.5.4. For a positive closed (1,1)(1,1) current on all of Rn\mathbb R^n, every connected component of the complement of its support is convex.

Proof. Let uu be its global convex potential and let VV be such a component. The potential agrees there with one affine function aa. At any point of VV, that affine function is a supporting plane for uu, so u≥au\geq a globally. The zero set of the convex nonnegative function u−au-a is convex. It contains VV, hence its convex hull. Since VV is open, its convex hull is open as well; uu is affine on that open hull, so the current vanishes there. Connectedness and maximality of the component then imply conv⁡(V)=V\operatorname{conv}(V)=V. ◻

References. The potential theorem and the support-complement consequence occur in [Lag12] (Propositions 1.13–1.14). The proof here supplies the distributional integrability and gluing steps explicitly, using the representative theorem of Section 11.3.

11.6 Which approximation statement is available

The first-stage theorem and the generated-sheaf comparison serve different purposes. The former constructs a single smooth sequence on a fixed neighbourhood and is the input for Theorem 13.6.3. The latter allows repeated local limiting operations. A section obtained after several such operations must not be used as a first-stage section without an additional argument.

Lemma 11.6.1. (The upper-semicontinuous form of Dini’s argument). Let KK be compact, let uju_j be finite upper semicontinuous functions decreasing to a continuous finite function uu. Then uj→uu_j\to u uniformly on KK.

Proof. For ϵ>0\epsilon>0, the sets Oj={uj−u<ϵ}O_j=\{u_j-u<\epsilon\} are open, increase with jj, and cover KK. Compactness gives K=ONK=O_N for some NN. Thus 0≤uj−u<ϵ0\leq u_j-u<\epsilon on KK for all j≥Nj\geq N. ◻

This observation locates the difficulty in CLD25 Proposition 8.1.11 more precisely. The obstacle is not the uniformity of a decreasing limit with continuous endpoint: upper semicontinuity already suffices for that. It is the replacement of the possibly discontinuous predecessors in P25,α\mathcal P_{25,\alpha} by members of C25,α\mathcal C_{25,\alpha}. Uniform convergence of discontinuous predecessors does not make those predecessors continuous. Proposition 11.2.3 avoids this issue because its predecessors are smooth.

Assertion Usable justification
U=M=C0∩P25,1\mathcal U=\mathcal M=C^0\cap\mathcal P_{25,1} Proposition 11.2.3, with fixed local approximation domain.
U⊆C25⊆C0∩P25\mathcal U\subseteq\mathcal C_{25}\subseteq C^0\cap\mathcal P_{25} Proposition 11.2.4; no reverse inclusion is used.
C25=C0∩P25\mathcal C_{25}=C^0\cap\mathcal P_{25} Stronger source assertion; the predecessor replacement remains unproved here.
RX(u1,…,uk;T)R_X(u_1,\ldots,u_k;T) Theorem 13.6.3 for ui∈UXu_i\in\mathcal U_X and the stated ordinary positive closed TT.

The kernel question is independent of this existence statement. The cross calculation in Proposition 14.2.2 refutes its unrestricted abstract version. A positive result on an analytic space must specify the harmonic kernel, the test quotient and the permitted background currents.

References. CLD25 Definitions 8.1.1, 8.1.7 and Proposition 8.1.11; the compactness argument above is elementary.

A continuous upper barrier on a fixed compact set

To understand the missing step in Proposition 8.1.11, we first remove one difficulty from the problem. Repeated decreasing limits on a fixed compact set can be handled without making every predecessor continuous. A continuous upper barrier allows us to select one predecessor at a time. The distinction between this construction and a sheaf construction will then be explicit.

Definition 11.6.2. (The decreasing hierarchy on a fixed compact set). Let KK be a compact Hausdorff space and let S⊂C(K,R)\mathscr S\subset C(K,\mathbf R) be nonempty. Define classes of actual functions K→R∪{−∞}K\to\mathbf R\cup\{-\infty\} by D0(S)=S,Dα+1(S)={f:there are fj∈Dα(S) with fj↓f on K},Dλ(S)=⋃β<λDβ(S)(λ a limit ordinal).\begin{split} D_0(\mathscr S)&=\mathscr S,\\ D_{\alpha+1}(\mathscr S)&= \{f:\text{there are }f_j\in D_\alpha(\mathscr S) \text{ with }f_j\downarrow f\text{ on }K\},\\ D_\lambda(\mathscr S)&=\bigcup_{\beta<\lambda}D_\beta(\mathscr S) \quad(\lambda\text{ a limit ordinal}). \end{split} Write D(S)D(\mathscr S) for the class obtained after stabilization. All functions at all stages have domain KK. There is no sheafification or passage to a smaller domain in this definition.

The classes increase because constant sequences are allowed. They stabilize since they are subsets of the fixed set of all functions on KK. Every member is upper semicontinuous: this is preserved by a decreasing limit and is unchanged by taking a union of classes. These observations alone do not give first-stage stabilization for discontinuous functions. Corollary 11.24.6 below proves it by a separate compact diagonal argument.

Lemma 11.6.3. (Selection below a continuous barrier). If f∈D(S)f\in D(\mathscr S) and b∈C(K,R)b\in C(K,\mathbf R) satisfy f<bf<b everywhere on KK, then there is s∈Ss\in\mathscr S such that f≤s<bon K.\begin{equation} f\leq s<b\quad\text{on }K. \end{equation}

Proof. We prove the assertion by induction on the stage containing ff. At stage zero, take s=fs=f. At a successor stage, write fj↓ff_j\downarrow f with all fjf_j in the preceding stage. The sets Oj={z∈K:fj(z)<b(z)}O_j=\{z\in K:f_j(z)<b(z)\} are open, increase with jj, and cover KK. Openness uses upper semicontinuity of fj−bf_j-b, and the covering property uses f<bf<b. By compactness, ON=KO_N=K for some NN. Apply the induction hypothesis to fNf_N and the same barrier bb. It gives f≤fN≤s<bf\leq f_N\leq s<b. At a limit ordinal, ff already belongs to one earlier stage by definition, so the induction hypothesis applies directly. ◻

Only one predecessor is retained at each successor step. The selected predecessor may be discontinuous; its upper semicontinuity is enough for the next selection. All selections use the same compact set and the same barrier. This is the point that a direct appeal to continuity of the predecessors would miss.

Theorem 11.6.4. (Continuous endpoints of the compact hierarchy). Suppose S\mathscr S is stable under addition of real constants. Then C(K,R)∩D(S)=S‾∥⋅∥∞=C(K,R)∩D1(S).\begin{equation} C(K,\mathbf R)\cap D(\mathscr S) =\overline{\mathscr S}^{\|\cdot\|_\infty} =C(K,\mathbf R)\cap D_1(\mathscr S). \end{equation} Consequently, the continuous parts of Dα(S)D_\alpha(\mathscr S) agree for every α≥1\alpha\geq 1.

Proof. Let uu be a finite continuous member of D(S)D(\mathscr S). Apply Lemma 11.6.3 with b=u+ϵb=u+\epsilon. For every ϵ>0\epsilon>0 it gives sϵ∈Ss_\epsilon\in\mathscr S with 0≤sϵ−u<ϵ0\leq s_\epsilon-u<\epsilon on all of KK. Thus uu is in the uniform closure of S\mathscr S.

Conversely, let uu be in that uniform closure. Choose sj∈Ss_j\in\mathscr S with ∥sj−u∥∞≤δj\|s_j-u\|_\infty\leq\delta_j, where δj=4−j\delta_j=4^{-j}, and set tj=sj+2δjt_j=s_j+2\delta_j. Then tj∈St_j\in\mathscr S and u+δj≤tj≤u+3δj,tj+1≤u+34δj<tj.u+\delta_j\leq t_j\leq u+3\delta_j, \qquad t_{j+1}\leq u+\tfrac34\delta_j<t_j. Hence tj↓ut_j\downarrow u uniformly on KK, so u∈D1(S)u\in D_1(\mathscr S). The remaining inclusion follows from D1⊂DD_1\subset D. The assertion for every positive ordinal follows by inserting D1⊂Dα⊂DD_1\subset D_\alpha\subset D into (11.6.2). ◻

This is an elementary statement about monotone closure on a fixed compact domain. It does not use positivity, tropical charts or a regularization theorem. Its application to the generated psh sheaf requires an additional representation property, stated next.

The remaining domain requirement

Corollary 11.6.5. (A compact representation gives first-stage approximation). Let KK be a compact neighbourhood of xx in a Hausdorff locally compact analytic space XX. Let SK\mathscr S_K consist of the restrictions to KK of original smooth psh functions defined on open neighbourhoods of KK. These neighbourhoods may depend on the function. Let uu be finite and continuous near KK. If u∣K∈D(SK),\begin{equation} u|_K\in D(\mathscr S_K), \end{equation} then uu has a decreasing original smooth psh approximation on the single open neighbourhood Int⁡X(K)\operatorname{Int}_X(K) of xx, uniformly on compact subsets there. In particular, uu is locally in UX\mathcal U_X near xx.

Proof. Theorem 11.6.4 gives a decreasing sequence tj∈SKt_j\in\mathscr S_K converging uniformly to u∣Ku|_K. Choose smooth psh representatives near KK. Every representative is defined on Int⁡X(K)\operatorname{Int}_X(K), even though their larger neighbourhoods may differ. Their restrictions to this fixed interior therefore give the required sequence. The inequalities and the uniform bounds on KK remain valid there. ◻

Condition (11.6.3) is the substantive hypothesis. Membership in P25,α\mathcal P_{25,\alpha} allows local representations at every successor stage and sheafification at limit stages. Restricting one section to KK does not by itself give a representation in Dα(SK)D_\alpha(\mathscr S_K): the lower-stage sections may only have representations on a cover of KK, and their smooth predecessors need not extend across that cover. The compact hierarchy has no such local operation.

In particular, approximation on a neighbourhood depending on the error is insufficient. On the real tropical line, the function u(t)=−t2u(t)=-t^2 is not psh on any interval, since u′′=−2u''=-2. Nevertheless, for every ϵ>0\epsilon>0, the smooth psh constant zero satisfies ∣0−u(t)∣<ϵfor ∣t∣<ϵ.|0-u(t)|<\epsilon\quad\text{for }|t|<\sqrt\epsilon. Indeed, every continuous function admits such shrinking-domain constant approximations at a point. Thus this condition carries no psh information. In Corollary 11.6.5, the compact neighbourhood is fixed before the error tends to zero.

The conclusion of the corollary is stronger than the unindexed equality asserted in CLD25 Proposition 8.1.11: it gives first-stage approximation. Consequently, constructing compact representations for every continuous generated psh function would settle more than that equality. Failure to construct them does not refute the weaker equality. For that weaker target, continuous replacements already in C25\mathcal C_{25} on one fixed domain would suffice; their ordinal need not match the ordinal of the original predecessor. The same-ordinal assertion in the source imposes this further constraint.

What survives at every ordinal

We record the valid direction of the continuous comparison with its ordinal intact. Define C25,0=P25,0\mathcal C_{25,0}=\mathcal P_{25,0}, let C25,α+1\mathcal C_{25,\alpha+1} be the sheaf of local uniform limits of sections of C25,α\mathcal C_{25,\alpha}, and at a limit ordinal take the sheaf of functions locally in an earlier continuous stage. All its sections are finite and continuous.

Proposition 11.6.6. (The indexed inclusion). For every ordinal α\alpha, C25,α⊂C0∩P25,α.\mathcal C_{25,\alpha} \subset C^0\cap\mathcal P_{25,\alpha}. Equality holds for α=0,1\alpha=0,1. No reverse inclusion at later stages is asserted here.

Proof. First, addition of a constant preserves every C25,α\mathcal C_{25,\alpha}: prove this by induction, using the assertion for smooth psh functions, compatibility of translation with uniform limits, and locality at limit ordinals. Now argue by induction for the displayed inclusion. The zero stage is immediate. At a successor stage, work on an open set where uj∈C25,αu_j\in\mathcal C_{25,\alpha} converges uniformly to uu. After passing to a subsequence, assume ∥uj−u∥≤4−j\|u_j-u\|\leq 4^{-j}. The functions uj+2⋅4−ju_j+2\cdot 4^{-j} remain in C25,α\mathcal C_{25,\alpha} and decrease to uu, by the estimate in Proposition 11.2.3. By induction they lie in P25,α\mathcal P_{25,\alpha}; hence u∈P25,α+1u\in\mathcal P_{25,\alpha+1}. If the convergence is only locally uniform, first restrict to a relatively compact open neighbourhood. Sheafification and limit ordinals are treated by restriction to the corresponding local representations. Equality at stage one is Proposition 11.2.3. ◻

Local integrability behaves differently under the order relation defining a decreasing limit. If uu is locally integrable, vv is upper semicontinuous, and u≤vu\leq v, then on a compact set where v≤Mv\leq M one has ∣v∣≤∣u∣+max⁡(M,0)almost everywhere.\begin{equation} |v|\leq |u|+\max(M,0) \quad\text{almost everywhere}. \end{equation} Thus vv is integrable for each locally finite measure used in the definition. In CLD25 paragraph 5.7.2 this can be checked on skeletal cells with their Lebesgue measures, or against the specified locally bounded volume forms. It does not mean integrability against every arbitrary background current.

Proposition 11.6.7. (The locally integrable hierarchy). Let I25,α\mathcal I_{25,\alpha} be the hierarchy obtained from the original smooth psh functions by the same local decreasing-limit and limit-ordinal operations as P25,α\mathcal P_{25,\alpha}, but retaining only locally integrable functions and requiring locally integrable endpoints. Then, for every ordinal α\alpha, I25,α=Lloc1∩P25,α.\begin{equation} \mathcal I_{25,\alpha} =L^1_{\mathrm{loc}}\cap\mathcal P_{25,\alpha}. \end{equation} These are classes of actual functions; the order and limits are pointwise, not operations on equivalence classes modulo null sets. Moreover, if uj↓uu_j\downarrow u are generated psh functions and uu is locally integrable, then every uju_j is locally integrable and uj→uu_j\to u in local L1L^1 for each of the measures in that definition.

Proof. This is the argument of CLD25 Proposition 8.1.12, with the inherited integrability made explicit. The inclusion from left to right in (11.6.5) follows from the construction. For the reverse inclusion, use induction on α\alpha. At a successor stage, write a locally integrable uu locally as uj↓uu_j\downarrow u with uj∈P25,αu_j\in\mathcal P_{25,\alpha}. The functions uju_j are upper semicontinuous and bounded above on compact sets. Since u≤uju\leq u_j, estimate (11.6.4) shows that each uju_j is locally integrable. The induction hypothesis puts them in I25,α\mathcal I_{25,\alpha}, so their locally integrable endpoint belongs to I25,α+1\mathcal I_{25,\alpha+1}. The zero and limit stages follow directly from the definitions and locality.

For the convergence assertion, work on a compact set for one of the defining measures. The nonnegative integrable function u1−uu_1-u dominates uj−uu_j-u, which decreases to zero almost everywhere. Dominated convergence gives ∫∣uj−u∣ dμ⟶0.\int|u_j-u|\,d\mu\longrightarrow0. Apply this on the local cells or to each required volume measure. Thus the currents [uj][u_j] converge to [u][u] against the original smooth tests wherever these currents are defined, as used in Proposition 11.2.4. No continuous predecessor replacement is needed. ◻

Remark 11.6.8 (Consequences for the approximation audit). The compact barrier theorem resolves the transfinite selection problem when one common compact representation is supplied. It does not prove that sheaf-generated sections admit such a representation. The indexed inclusion and the locally integrable equality above are valid without that representation. Consequently, the gap identified in Proposition 8.1.11 does not invalidate Proposition 8.1.12 or the locally integrable current-positivity argument. Neither local L1L^1 convergence nor membership in a generated class alone defines products of unbounded potentials with arbitrary currents. Such products require their own integrability and continuity statements.

References. CLD25 v2, paragraph 5.7.2, Definitions 8.1.1 and 8.1.7, Lemmas 8.1.4, 8.1.8 and 8.1.10, Propositions 8.1.11–8.1.12. The compact hierarchy and its barrier proof are elementary deductions stated separately from the source’s sheaf hierarchy. The indexed forward inclusion and the L1L^1 equality expand valid portions of the source argument; no counterexample to its continuous equality is asserted.

11.7 Envelope methods and bounded localization

The compact argument of Section 11.6.1 leaves a geometric question: how can approximants on different local domains be compared on one neighbourhood? A projective ambient space provides another route. One first extends bounded local potentials to a common global background, uses an envelope comparison there, and then restricts back. This section gives the order estimates needed for that route and identifies the additional theorems it uses.

Three classes in the recent local theory

We fix the preprint [Wu26], arXiv version 1 of 23 September 2026. Its local theory is formulated on opens in algebraic analytifications over a complete discretely or trivially valued field. Write PW\mathcal P_W for its class of decreasing limits of tropically convex PL functions, and RW\mathcal R_W for its smooth-approximable class. These are first-stage local approximation conditions. They must be distinguished from the sheaf P25\mathcal P_{25} obtained by iterating decreasing limits and localization.

For the comparisons below we use locally bounded functions. They are finite, so the exclusion of functions identically −∞-\infty on a component causes no ambiguity. The smooth generators are the original CLD generators; replacing them by weakly smooth or harmonic-chart generators would change the comparison.

Class Local requirement Relevant source
PW\mathcal P_W One decreasing PL sequence on one open neighbourhood [Wu26], Definition 3.8
RW\mathcal R_W One decreasing original smooth convex sequence [Wu26], Definition A.1
P25\mathcal P_{25} The generated sheaf, including further decreasing limits [CLD25], Definition 8.1.1

Theorem A.3 of [Wu26] asserts PW=RW\mathcal P_W=\mathcal R_W. Corollary 4.18(1) supplies decreasing-limit stability of PW\mathcal P_W when the projective ambient variety has the envelope property. Remark A.2 explicitly retains this hypothesis in the comparison with the generated CLD definition. These are substantial source inputs, not consequences of the terminology “psh”. Their implication for continuous generated functions is proved in Section 11.7.4.

In the convex terminology here, the supporting affine function touches the function at the point. Example A.8.1 records a defect in the source’s displayed definition and its repair. For the explicit rational-generator class of Section 11.9.1, Theorem 11.9.5 supplies one direction of approximation independently of Theorem A.3, over a nontrivially valued field.

Continuous obstacles on a compact space

An envelope argument needs continuous majorants of a local potential. Compactness alone does not make every upper semicontinuous function a decreasing limit of a sequence of continuous functions on an arbitrary Hausdorff space. The local approximation hypothesis supplies precisely the missing information here.

Lemma 11.7.1. (Continuous majorants from local sequences). Let XX be compact Hausdorff and let u:X→Ru:X\to\mathbf R. Suppose every point has an open neighbourhood on which uu is a decreasing limit of finite continuous functions. Then there are Fj∈C(X,R)F_j\in C(X,\mathbf R) with Fj↓uF_j\downarrow u pointwise on XX. No metrizability assumption is needed.

Proof. Choose a finite open cover U1,…,UrU_1,\ldots,U_r with continuous sequences ui,j↓u∣Uiu_{i,j}\downarrow u|_{U_i}. By normality of a compact Hausdorff space, choose a continuous partition of unity ρi\rho_i with supp⁡ρi⊂Ui\operatorname{supp}\rho_i\subset U_i. Extend ρiui,j\rho_i u_{i,j} by zero outside UiU_i and put Fj=∑i=1rρiui,j.F_j=\sum_{i=1}^r\rho_i u_{i,j}. Each summand is continuous: its support stays inside its defining open set. Nonnegative coefficients give Fj+1≤FjF_{j+1}\leq F_j. At every point only finitely many terms occur, and their limit is ∑iρiu=u\sum_i\rho_i u=u. ◻

The functions FjF_j are continuous obstacles; the proof does not assert that they are psh. Applying a partition of unity directly to psh potentials would introduce uncontrolled derivatives of the coefficients. Positivity is recovered by an envelope, under the following explicit hypotheses.

Proposition 11.7.2. (The bounded envelope comparison). Fix a projective ambient space and an ample semipositive model background ω\omega. Let Gω\mathscr G_\omega and Lω\mathscr L_\omega be respectively its global and local classes of bounded ω\omega-psh potentials, with Gω⊂Lω\mathscr G_\omega\subset\mathscr L_\omega. Assume the following inputs:

  1. Every u∈Lωu\in\mathscr L_\omega is locally a decreasing limit of continuous functions.

  2. For F∈C(Xan,R)F\in C(X^{\rm an},\mathbf R), the envelope PωFP_\omega F is a continuous member of Gω\mathscr G_\omega, satisfies PωF≤FP_\omega F\leq F, and its Monge–Ampère measure is supported on {PωF=F}\{P_\omega F=F\}.

  3. Domination holds: for u∈Lωu\in\mathscr L_\omega and v∈Gωv\in\mathscr G_\omega, the inequality u≤vu\leq v almost everywhere for MA⁡ω(v)\operatorname{MA}_\omega(v) implies u≤vu\leq v everywhere.

  4. Gω\mathscr G_\omega is closed under decreasing limits with bounded endpoints.

Then Lω=Gω\mathscr L_\omega=\mathscr G_\omega.

Proof. Take u∈Lωu\in\mathscr L_\omega. Lemma 11.7.1 gives Fj↓uF_j\downarrow u with FjF_j continuous. Put pj=PωFjp_j=P_\omega F_j. The definition of an envelope is order preserving, so pj+1≤pjp_{j+1}\leq p_j. On the support of MA⁡ω(pj)\operatorname{MA}_\omega(p_j) one has u≤Fj=pju\leq F_j=p_j. Domination gives u≤pju\leq p_j on the whole space. Thus u≤pj≤Fj,pj↓u.u\leq p_j\leq F_j,\qquad p_j\downarrow u. The endpoint is bounded, so the last hypothesis gives u∈Gωu\in\mathscr G_\omega. ◻

This is the bounded part of the mechanism in [Wu26], Theorem 4.17. Envelope continuity, orthogonality, and domination remain imported inputs; none follows from Lemma 11.7.1. The lemma makes the continuous-obstacle construction independent of a countable dense subfield. It does not remove countability hypotheses from other theorems in that paper.

An order-preserving extension on a fixed neighbourhood

The next construction explains why the global comparison is relevant to a local decreasing sequence. The extensions must preserve order; choosing an unrelated global extension for every term would not suffice. We use a local potential class closed under finite maxima, positive scaling and addition of constants. Model backgrounds in this subsection are curvature classes of metrics, not automatically original smooth CLD forms.

Proposition 11.7.3. (A fixed bounded extension formula). Let XX be projective. Let ω,θ\omega,\theta be semipositive model backgrounds on the same Q\mathbf Q-line bundle, admitting local psh potentials in the chosen class, and choose a global continuous PL function bb with θ=ω+d′d′′b\theta=\omega+d'd''b. On an open Y⊂XanY\subset X^{\rm an} let vv be a continuous psh potential of θ\theta. Assume Q={v≤0}⋐Y∩U,P={v≤−ϵ},x∈Int⁡(P),ϵ>0,Q=\{v\leq0\}\Subset Y\cap U,\qquad P=\{v\leq-\epsilon\},\qquad x\in\operatorname{Int}(P),\qquad \epsilon>0, where UU is an open set on which the local potentials will be given. Fix M>0M>0 and a positive rational number AA with Aϵ>2M+1A\epsilon>2M+1, and set h=Av+M+1h=Av+M+1. For any psh uu on UU with ∣u∣≤M|u|\leq M on QQ, define Ψ(u)={max⁡{u−h,0},on Q,0,on Xan∖Q,Φ(u)=A−1Ψ(u)+b.\begin{equation} \Psi(u)= \begin{cases} \max\{u-h,0\},&\text{on }Q,\\ 0,&\text{on }X^{\rm an}\setminus Q, \end{cases} \qquad \Phi(u)=A^{-1}\Psi(u)+b. \end{equation} Then Φ(u)\Phi(u) is a bounded local ω\omega-psh potential. On the same open neighbourhood V=Int⁡(P)V=\operatorname{Int}(P) of xx, for all such uu, Φ(u)+g=A−1u,g=v−b+A−1(M+1).\begin{equation} \Phi(u)+g=A^{-1}u, \qquad g=v-b+A^{-1}(M+1). \end{equation} The map Φ\Phi preserves order and decreasing pointwise limits within this bounded family. Moreover, ∥Φ(u)−Φ(w)∥Xan≤A−1∥u−w∥Q.\begin{equation} \|\Phi(u)-\Phi(w)\|_{X^{\rm an}} \leq A^{-1}\|u-w\|_Q. \end{equation}

Proof. On PP one has h<−M≤uh<-M\leq u. On the compact collar C={−1/(2A)≤v≤0}C=\{-1/(2A)\leq v\leq 0\} one has h≥M+1/2>uh\geq M+1/2>u. Consequently Ψ(u)\Psi(u) vanishes on a neighbourhood of the boundary of QQ, and the piecewise formula is well defined without a boundary jump. On Int⁡(Q)\operatorname{Int}(Q), Ψ(u)+h=max⁡{u,h}\Psi(u)+h=\max\{u,h\} is psh. Outside a smaller compact sublevel set, Ψ(u)=0\Psi(u)=0. Since θ\theta is semipositive with admitted local potentials, these statements show that Ψ(u)\Psi(u) is locally AθA\theta-psh. After scaling and changing the background, ω+d′d′′Φ(u)=θ+A−1d′d′′Ψ(u)≥0\omega+d'd''\Phi(u)=\theta+A^{-1}d'd''\Psi(u)\geq0 in the local-potential sense. Boundedness follows from the bounds on uu, continuity of hh on QQ, and continuity of bb on the compact space. On PP the maximum selects u−hu-h, which proves (11.7.2). Finally, the scalar map t↦max⁡(t−h,0)t\mapsto\max(t-h,0) is increasing, continuous, and 11-Lipschitz. This proves the order, limit, and norm assertions using the same A,h,bA,h,b throughout. ◻

Lemma 11.7.4. (One collar for a whole decreasing tail). In the preceding setting, suppose uju_j are finite upper semicontinuous functions on UU with uj↓uu_j\downarrow u and ∣u∣≤M|u|\leq M on QQ. Then, for all sufficiently large jj, uj<hu_j<h on the fixed collar CC. If the uju_j are continuous, their extensions by (11.7.1) are continuous for this entire tail. On PP the identity (11.7.2) holds for every jj.

Proof. The open sets {uj<M+1/4}\{u_j<M+1/4\} increase and cover the compact set CC, since u≤Mu\leq M. One index therefore works on all of CC, and every later index works as well. There uj<M+1/4<hu_j<M+1/4<h. On PP, the inequalities uj≥u≥−M>hu_j\geq u\geq-M>h give the claimed identity. ◻

These formulas reconstruct the bounded localization mechanism of [Wu26], Proposition 3.28, while recording the background change and the common collar. The existence of the projective barrier data is a separate geometric input, constructed for integral projective varieties in Theorem 11.8.9. In a weighted Fubini–Study presentation its local shape can be v=max⁡{−a,ℓ1,…,ℓr},a>ϵ>0,ℓi(x)<−ϵ.\begin{equation} v=\max\{-a,\ell_1,\ldots,\ell_r\}, \qquad a>\epsilon>0,\quad \ell_i(x)<-\epsilon. \end{equation} Here the ℓi\ell_i are normalized logarithmic section ratios with their weights, and the outer sublevel set must be chosen compactly inside UU. The negative constant branch is essential. A formula max⁡{0,ℓ1,…,ℓr}\max\{0,\ell_1,\ldots,\ell_r\} is nonnegative and cannot satisfy v(x)<0v(x)<0. Also, equality of underlying line bundles does not identify their model backgrounds; the term bb in (11.7.1) performs the required conversion.

What bounded decreasing stability would settle

We now isolate a sufficient input that is weaker than constructing the raw compact representation of Corollary 11.6.5. Let R=P25,1\mathcal R=\mathcal P_{25,1} denote the original first-stage sheaf, allowing discontinuous functions.

Theorem 11.7.5. (Bounded stability implies collapse of continuous stages). Assume that on every open set a decreasing sequence of locally bounded sections of R\mathcal R, whose endpoint is locally bounded, again has its endpoint in R\mathcal R. Then P25∩Lloc∞=R∩Lloc∞.\begin{equation} \mathcal P_{25}\cap L^\infty_{\rm loc} =\mathcal R\cap L^\infty_{\rm loc}. \end{equation} For every ordinal α≥1\alpha\geq 1, one moreover has C25,α=C0∩P25,α=U.\begin{equation} \mathcal C_{25,\alpha} =C^0\cap\mathcal P_{25,\alpha} =\mathcal U. \end{equation} All boundedness conditions here concern actual pointwise functions.

Proof. We prove by induction on α\alpha that each locally bounded section of P25,α\mathcal P_{25,\alpha} belongs to R\mathcal R. This holds at stage zero. At a successor stage, work on an open set where uj↓uu_j\downarrow u, with uj∈P25,αu_j\in\mathcal P_{25,\alpha} and uu locally bounded. Since uj≥uu_j\geq u and uju_j is upper semicontinuous with no value +∞+\infty, every uju_j is locally bounded. The induction hypothesis puts every predecessor in R\mathcal R. The assumed bounded decreasing stability now applies on this open set. At a limit ordinal use the local earlier-stage representations and the sheaf property of R\mathcal R. This proves (11.7.5).

A finite continuous u∈P25,αu\in\mathcal P_{25,\alpha} is locally bounded, hence belongs to R\mathcal R. Dini’s argument on a smaller relatively compact neighbourhood turns its decreasing smooth approximation into local uniform approximation. Thus u∈Uu\in\mathcal U. Conversely, U=C25,1\mathcal U=\mathcal C_{25,1} by Proposition 11.2.3, and the continuous hierarchy is increasing. Combine these inclusions with Proposition 11.6.6 to obtain (11.7.6). ◻

The hypothesis is a genuine stability statement on a common open set. It does not follow merely by choosing a separate regularization neighbourhood for each predecessor. Once it is known, no additional ordinal estimate is needed for continuous functions: all positive stages already coincide. This strengthens the earlier repair plan, which treated ordinal control as a further possible task.

The projective scope and its source dependencies

Corollary 11.7.6. (Application of the envelope comparison inputs). Let KK be complete and discretely or trivially valued, let XX be projective over KK, and suppose XX has the envelope property over KK in the sense of [Wu26]. Import Theorem A.3 and Corollary 4.18(1) of that source, with its comparison of analytic and algebraic smooth generators. Then on opens of XanX^{\rm an}, the conclusions (11.7.5) and (11.7.6) hold.

Proof. For locally bounded functions, Theorem A.3 identifies R\mathcal R with PW\mathcal P_W. A locally bounded decreasing endpoint is generically finite, so Corollary 4.18(1) gives the bounded stability assumed in Theorem 11.7.5. Apply that theorem. ◻

Wu’s Theorem 7.1 supplies the envelope property for projective geometrically unibranch varieties over discretely or trivially valued fields of equicharacteristic zero; normal projective varieties are included. Accordingly, this is a source-dependent route to the continuous-stage comparison in that category. The paper’s Fornæss–Narasimhan step, Theorem 6.5, includes a countable-dense-subfield hypothesis; the proof of Theorem 7.1 contains a further descent argument to remove it. Neither hypothesis nor descent step should be suppressed when reconstructing that proof.

Remark 11.7.7 (What has been proved here and what is imported). The continuous-majorant lemma, the fixed extension estimates, and the transfinite bounded-stability implication have complete proofs above. The projective corollary uses the indicated preprint theorems. It is not an independent reconstruction of the full proof of Theorem A.3, the domination principle, or the envelope theorem. In particular, Appendix A of [Wu26] invokes two Amini–Piquerez–Wu preprints and passes from approximation of generators to approximation classes. Section 11.8 matches the localization and PL-approximation locators, corrects a false whole-domain inequality, and proves fixed-domain estimates. The general generator-to-class passage and the separate mixed-calculus input remain to be verified.

The general analytic question from Proposition 8.1.11 remains separate. The cited envelope theorem does not cover arbitrary complete fields or arbitrary Berkovich spaces; in particular, its equicharacteristic-zero statement is not a theorem for every finite extension of Qp\mathbf Q_p. Nor does a comparison of functions establish an arbitrary-background-current Bedford–Taylor calculus. The next task is to verify the source comparison chain in the projective category, while retaining bounded first-stage stability as the precise remaining input outside it.

References. Wu, arXiv:2609.27610v1, Definitions 3.8 and A.1, Proposition 3.28, Theorem 4.17, Corollary 4.18, Theorems 6.5 and 7.1, Remark A.2 and Theorem A.3. The present bounded extension is a corrected reconstruction of the localization mechanism; the continuous-majorant and transfinite arguments are elementary deductions. The external comparison theorems remain attributed inputs.

11.8 Common domains in local approximation

A comparison of generators becomes a comparison of first-stage classes only when the approximations can be made on one neighbourhood. We first give the required diagonal argument. We then reconstruct two localization procedures that produce such neighbourhoods: one on a tropical compactification and one on a projective analytic space. The tropical argument also identifies an incorrect extra inequality in the printed source. The local conclusion survives after a precise restriction of its domain.

A diagonal sequence on a prescribed domain

Lemma 11.8.1. Let VV be a fixed set, let uj:V→Ru_j:V\to\mathbb R decrease pointwise to u:V→R∪{−∞}u:V\to\mathbb R\cup\{-\infty\}, and let D(V)\mathcal D(V) be a class stable under addition of rational constants. Suppose that, for every j≥1j\geq 1, there is aj∈D(V)a_j\in\mathcal D(V) with ∥aj−uj∥V≤ϵj,ϵj=4−j.\|a_j-u_j\|_V\leq\epsilon_j,\qquad\epsilon_j=4^{-j}. Then bj=aj+2ϵjb_j=a_j+2\epsilon_j belongs to D(V)\mathcal D(V) and decreases pointwise to uu on this same VV.

Proof. The errors give uj+ϵj≤bj≤uj+3ϵj,bj+1≤uj+1+3ϵj+1≤uj+ϵj≤bj.u_j+\epsilon_j\leq b_j\leq u_j+3\epsilon_j, \qquad b_{j+1}\leq u_{j+1}+3\epsilon_{j+1} \leq u_j+\epsilon_j\leq b_j. The first pair of inequalities gives the pointwise limit, including at points where u=−∞u=-\infty. No boundedness of uu and no continuity of the endpoint are required. ◻

Proposition 11.8.2. (A sufficient criterion for transferring first-stage classes). Let A\mathcal A and B\mathcal B be classes of finite continuous functions on ordinary open sets of a locally compact Hausdorff space, stable under restriction, and suppose B\mathcal B is stable under addition of rational constants. Assume that whenever V⋐UV\Subset U are open and a∈A(U)a\in\mathcal A(U), then for every δ>0\delta>0 there exists b∈B(V)b\in\mathcal B(V) with ∥a−b∥V<δ\|a-b\|_V<\delta. Every function locally given by a decreasing sequence of A\mathcal A-functions is then locally given by a decreasing sequence of B\mathcal B-functions. Any separate endpoint condition, such as generic finiteness, is unchanged.

Proof. At a point xx, choose a representation uj↓uu_j\downarrow u on one open UU. Choose x∈V⋐Ux\in V\Subset U once. The hypothesis gives an approximation to uj∣Vu_j|_V with error at most 4−j4^{-j} for every jj, without changing VV. Lemma 11.8.1 applies. ◻

The hypothesis is stronger than approximation near each point of each individual generator. From statements with domains VjV_j depending on jj, one cannot simply form ⋂jVj\bigcap_jV_j and call it a neighbourhood. For example, the neighbourhoods (−1/j,1/j)(-1/j,1/j) of zero have intersection {0}\{0\}. This logical distinction identifies a missing justification; it is not by itself a counterexample to equality of the two function classes.

Corollary 11.8.3. (Decreasing finite logarithmic maxima). Let VV be an ordinary open set of a good analytic space in the original CLD category. Suppose uju_j decreases pointwise on VV and each term has a finite presentation uj=max⁡{cj, aj1log⁡∣fj1∣+dj1,…,ajrjlog⁡∣fjrj∣+djrj},aji>0,u_j=\max\{c_j,\ a_{j1}\log|f_{j1}|+d_{j1},\ldots, a_{jr_j}\log|f_{jr_j}|+d_{jr_j}\}, \quad a_{ji}>0, where fjif_{ji} are analytic on all of VV and cj,djic_j,d_{ji} are real constants. There is one decreasing sequence of original smooth psh functions on VV with the same pointwise limit. Different terms may use different analytic functions; these functions may have zeros.

Proof. For each jj, regularize its finite maximum by the compactly supported regularized maximum used in Theorem 20.8.1. The constant branch makes uju_j finite and continuous. At a zero of one of the fjif_{ji}, its logarithmic branch tends to −∞-\infty and is strictly below the constant branch on a neighbourhood. Exact deletion of inactive coordinates makes the regularization smooth there. At other points the surviving functions are analytic units, and convexity of the regularized maximum gives a positive Hessian in a moment chart. The local formulas agree by exact deletion. Thus they define a smooth psh function on all of VV, with a uniform error bounded by the regularization parameter. Choose that parameter at most 4−j4^{-j} and apply Lemma 11.8.1. ◻

This is a common-domain result for the displayed presentations. It does not assert that an arbitrary locally tropically convex PL function has one such presentation on an arbitrarily prescribed open set. Log-sum-exp alone would not justify smoothness at zeros; the exact-deletion property is the reason for using this regularization. The unit-only formula in Theorem 19.18.15 has no such zero issue.

The local affine support calculation

Lemma 11.8.4. Let Π\Pi be a locally finite polyhedral complex in a finite-dimensional real vector space EE, let t∈∣Π∣t\in|\Pi|, and let WW be an open neighbourhood of tt. There is r>0r>0, depending only on Π,W,t\Pi,W,t, such that every cell meeting B(t,r)B(t,r) contains tt and B(t,r)⊂WB(t,r)\subset W. If F∈C1(W)F\in C^1(W) is convex on σ∩B(t,r)\sigma\cap B(t,r) for every cell σ\sigma, then F(z)≥ℓt(z),ℓt(z)=F(t)+dFt(z−t),z∈B(t,r)∩∣Π∣.\begin{equation} F(z)\geq\ell_t(z),\qquad \ell_t(z)=F(t)+dF_t(z-t),\qquad z\in B(t,r)\cap|\Pi|. \end{equation} In particular, the same radius works for any family satisfying these hypotheses for the same complex and open set.

Proof. Local finiteness gives a neighbourhood meeting only finitely many cells. Each of those not containing tt is closed and has positive distance from tt. A sufficiently small ball avoids all of them and lies in WW. If zz is in the remaining ball and in a cell σ\sigma, then t,z∈σt,z\in\sigma and their segment lies in σ∩B(t,r)\sigma\cap B(t,r). Apply the supporting-line inequality to the convex function s↦F(t+s(z−t))s\mapsto F(t+s(z-t)) on [0,1][0,1]. Its right derivative at zero is dFt(z−t)dF_t(z-t). ◻

The affine expression must be centred at tt. Interpreting dFtdF_t as a linear functional on EE, the expression dFt(z)+F(t)dF_t(z)+F(t) generally fails to take the value F(t)F(t) at z=tz=t. Also, the cells here contain the tropical point t=Trop⁡(x)t=\operatorname{Trop}(x), not the analytic point xx. These observations correct the notation in the supporting-affine-function step of Wu’s Appendix A. They prove local support for a fixed polyhedral presentation, not ambient convexity on all of WW.

A false global inequality and the local replacement

We fix [APW26], arXiv:2610.25602v1. Its Proposition 5.19 contains both a local extension assertion and the additional inequality φ+g≥Cf\varphi+g\geq Cf on the whole original open set UU. Only the latter is contradicted by the following example.

Example 11.8.5. (Failure of majorization on all of UU). For every n≥1n\geq 1, take the projective tropical toric variety X=(TP1)nX=(\mathbb{TP}^1)^n, with dense stratum X∘=RnX^\circ=\mathbb R^n and product fan. Set g(t)=∑i=1n∣ti∣,U=(−1,1)n,f(t)=11−t1+∑i=2nti2.g(t)=\sum_{i=1}^n|t_i|,\qquad U=(-1,1)^n,\qquad f(t)=\frac1{1-t_1}+\sum_{i=2}^nt_i^2. The function gg satisfies the positivity and conewise-linearity assumptions in that proposition, and ff is smooth and strictly convex on UU. Nevertheless, for ω=∂∂ˉg\omega=\partial\bar\partial g, there are no C>0C>0 and φ∈CPSH⁡ω(X)\varphi\in\operatorname{CPSH}_\omega(X) satisfying φ+g≥Cf\varphi+g\geq Cf throughout UU.

Proof. The function gg is convex, homogeneous, linear on each orthant, and positive away from zero. The Hessian of ff is diagonal with entries 2/(1−t1)3,2,…,22/(1-t_1)^3,2,\ldots,2, hence positive definite. By Definition 4.3 of the source, every member of CPSH⁡ω(X)\operatorname{CPSH}_\omega(X) is bounded. Since g≤ng\leq n on UU, the left-hand side of the proposed inequality is bounded above there. The right-hand side tends to +∞+\infty along (t1,0,…,0)(t_1,0,\ldots,0) as t1↑1t_1\uparrow 1. This is impossible for C>0C>0. ◻

The printed proof first bounds ff on an inner compact neighbourhood U′U' but then uses that bound to control the set where a global barrier lies below ff on all of UU. The example shows why that enlargement of the domain is invalid. The construction can instead be carried out on the inner neighbourhood alone. The following formulation makes the domain and all constants uniform for a bounded family.

Proposition 11.8.6. (Uniform tropical localization on an inner domain). Use the projective tropical toric setting and the bounded-growth correspondence of [APW26] (Example 4.9). Write Y=X∘⊂E=RNY=X^\circ\subset E=\mathbb R^N, let gg be convex homogeneous PL with g>0g>0 away from zero, put h=g∣Yh=g|_Y, and let ω=∂∂ˉh\omega=\partial\bar\partial h. Fix x∈Yx\in Y, an open U⊂YU\subset Y, and r>0r>0 such that Q=Y∩B‾(x,r)⋐U.Q=Y\cap\overline B(x,r)\Subset U. Fix M>0M>0. There are one neighbourhood VV of xx in QQ and one λ>0\lambda>0 such that every f∈LConv⁡(U)f\in\operatorname{LConv}(U) with ∣f∣≤M|f|\leq M on QQ admits Φf∈CPSH⁡ω(X)\Phi_f\in\operatorname{CPSH}_\omega(X) satisfying Φf+h=λ−1f on V,Φf+h≥λ−1f on Q.\begin{equation} \Phi_f+h=\lambda^{-1}f\ \text{on }V,\qquad \Phi_f+h\geq\lambda^{-1}f\ \text{on }Q. \end{equation} The assignment preserves order and is λ−1\lambda^{-1}-Lipschitz for the uniform norms on QQ and XX. If ff is PL, the extension is PL in the source’s category. No inequality is asserted on U∖QU\setminus Q.

Proof. Continuity on the unit sphere gives c>0c>0 with g(z)≥c∥z∥g(z)\geq c\|z\|. Choose λ\lambda with λcr>2M+2\lambda cr>2M+2 and put gx(y)=g(y−x),H=λgx−M−1,Ff(y)={max⁡{f(y),H(y)},y∈Q,H(y),y∈Y∖Q.\begin{gathered} g_x(y)=g(y-x),\qquad H=\lambda g_x-M-1,\\ F_f(y)=\begin{cases}\max\{f(y),H(y)\},&y\in Q,\\H(y),&y\in Y\setminus Q. \end{cases} \end{gathered} On a fixed inner collar of ∂Q\partial Q, the inequality H>MH>M holds, so the two formulas agree on a neighbourhood of that boundary. Thus FfF_f is continuous and locally convex on YY: use the maximum of locally convex functions inside, and HH outside. On the fixed neighbourhood V=Y∩{gx<1/(2λ)}⊂Int⁡YQV=Y\cap\{g_x<1/(2\lambda)\}\subset\operatorname{Int}_Y Q one has H<−M−1/2<fH<-M-1/2<f, so Ff=fF_f=f. Set Φf=λ−1Ff−h\Phi_f=\lambda^{-1}F_f-h on YY. Outside QQ this is gx−g−(M+1)/λg_x-g-(M+1)/\lambda. A finite convex homogeneous PL function is globally Lipschitz, hence ∣gx−g∣≤L∥x∥|g_x-g|\leq L\|x\|. On QQ boundedness follows from the hypotheses. Example 4.9 of the source therefore gives the bounded ω\omega-psh extension to XX. In the PL case, the translated background extends as a PL potential difference as in the proof of its Proposition 5.19; the modification has compact support in YY and is locally a maximum of PL functions. This preserves PL.

The two identities follow directly from the maximum. On QQ, this maximum is increasing and 11-Lipschitz as a function of ff; outside QQ the extension is independent of ff. The norm bound follows on YY, hence on XX by continuity and density. These observations prove order preservation as well. If a decreasing family stays within the stated bounded class, the same formulas also commute with its pointwise limit. ◻

This proves the uniformity needed to localize a sequence already uniformly bounded on a prescribed compact tropical domain. Combined with the source’s uniform PL density theorem, it repairs the localization input in Proposition 5.21. That proof uses the local identity, not majorization on all of UU. The correction consequently does not refute its local PL approximation conclusion. Neither the tropical extension nor uniform PL density, however, supplies the analytic pullback comparison invoked separately in Wu’s Appendix A.

Projective Fubini–Study barriers

We now supply the geometric data assumed in Proposition 11.7.3, first as an explicit construction of weighted metrics. Let KK be any complete non-Archimedean field, let XX be an integral projective KK-variety, and let LL be ample. A Q\mathbf Q-line bundle is handled by first taking an actual tensor power. All topological neighbourhoods below are ordinary neighbourhoods in XanX^{\rm an}.

Lemma 11.8.7. (An affine compact buffer). For x∈U⊂Xanx\in U\subset X^{\rm an} with UU open, there are d>0d>0, a nonzero s∈H0(X,Ld)s\in H^0(X,L^d), the affine open Y=D(s)Y=D(s) containing xx, finitely many fi∈K[Y]f_i\in K[Y], and rational constants cic_i, such that K0={y∈Yan:∣fi(y)∣≤e−ci for all i}⋐U∩Yan,∣fi(x)∣<e−cifor all i.\begin{equation} \begin{gathered} K_0=\{y\in Y^{\rm an}:|f_i(y)|\leq e^{-c_i}\text{ for all }i\} \Subset U\cap Y^{\rm an},\\ |f_i(x)|<e^{-c_i}\qquad\text{for all }i. \end{gathered} \end{equation}

Proof. A globally generated power of LL has a section s0s_0 nonvanishing at xx; its principal open Y0=D(s0)Y_0=D(s_0) is affine. The topology of its analytification is generated by absolute values of regular functions, as in Section 2.10. Thus a finite intersection of strict inequalities ∣pi∣<Ri|p_i|<R_i and ∣qj∣>rj>0|q_j|>r_j>0 contains xx and is contained in U∩Y0anU\cap Y_0^{\rm an}. On Y=Y0∖⋃jZ(qj)Y=Y_0\setminus\bigcup_jZ(q_j) the lower inequalities become upper inequalities for qj−1q_j^{-1}. This is again a principal affine open of the required type: clear the denominators of the finite product of the qjq_j using a power of s0s_0, then multiply its extending section by s0s_0 to ensure nonvanishing only inside Y0Y_0. The resulting nonzero section of a power of LL defines exactly YY.

Replace each strict upper bound by a closed bound strictly between the value at xx and that upper bound. These bounds may be chosen in exp⁡(Q)\exp(\mathbf Q), which is dense in R>0\mathbb R_{>0}. Add bounds, again strict at xx, for a finite set of algebra generators of K[Y]K[Y]. Those latter bounds cut out the intersection of a closed polydisc with a closed affine embedding of YY, hence a compact affinoid domain; arbitrary positive radii are allowed. Intersecting with the other closed inequalities gives a compact subset K0K_0 of it. All original strict inequalities hold on K0K_0, so K0⊂UK_0\subset U. Their strictness at xx supplies its ordinary interior neighbourhood. This argument applies also to a trivially valued field and to non-strict affinoid radii. ◻

Lemma 11.8.8. (One denominator for finitely many functions). Let ss be a nonzero section of LdL^d on an integral projective variety, and let f1,…,fmf_1,\ldots,f_m be regular on Y=D(s)Y=D(s). For every sufficiently large NN, each fisNf_is^N extends to a section of LdNL^{dN} on XX.

Proof. Cover XX by finitely many affine opens trivializing LdL^d. On such an open Spec⁡A\operatorname{Spec}A, the section ss is a nonzero element a∈Aa\in A, and fi∣D(a)∈Aaf_i|_{D(a)}\in A_a. Clearing the finitely many denominators gives a common exponent for every ii and every open in the cover. The resulting local sections agree on the dense open D(s)D(s) and hence on overlaps, since XX is integral. They glue. A larger exponent is obtained by multiplication by a further power of ss. This also justifies the denominator-clearing step in Lemma 11.8.7. ◻

Theorem 11.8.9. (A negative Fubini–Study barrier). In the projective setting above, given x∈Ux\in U there exist a rational-weighted finite Fubini–Study metric θ\theta on LL, an affine open Y=D(s)Y=D(s) containing xx, and a continuous local weight vv for θ\theta on YanY^{\rm an} such that, for some ϵ>0\epsilon>0, Q={v≤0}⋐Yan∩U,x∈Int⁡Xan{v≤−ϵ}.\begin{equation} Q=\{v\leq0\}\Subset Y^{\rm an}\cap U, \qquad x\in\operatorname{Int}_{X^{\rm an}}\{v\leq-\epsilon\}. \end{equation} The weight is locally tropically convex and is locally uniformly approximable by original smooth psh functions. These assertions concern a metric and its local weights. Identifying its curvature with a model background in a narrower source category requires that category’s Fubini–Study/model comparison.

Proof. Take Y,s,fi,ci,K0Y,s,f_i,c_i,K_0 from Lemma 11.8.7. By Lemma 11.8.8, choose one NN such that all ti=fisNt_i=f_is^N extend to sections of LdNL^{dN} and that line bundle is globally generated. Choose a finite globally generating list r1,…,rar_1,\ldots,r_a of its sections. Every rk/sNr_k/s^N is regular on YY. Choose rational bkb_k so negative that log⁡∣rk/sN∣(x)+bk<0\log|r_k/s^N|(x)+b_k<0, with log⁡0=−∞\log 0=-\infty. Choose also a rational constant c0<0c_0<0. On YanY^{\rm an} put v=1dNmax⁡{c0, (log⁡∣fi∣+ci)i,(log⁡∣rk/sN∣+bk)k}.\begin{equation} v=\frac1{dN}\max\left\{c_0,\ (\log|f_i|+c_i)_i, (\log|r_k/s^N|+b_k)_k\right\}. \end{equation} This is the weight, relative to the nonvanishing section sNs^N of LdNL^{dN}, of the metric θ=1dNmax⁡{log⁡∣sN∣+c0, (log⁡∣ti∣+ci)i,(log⁡∣rk∣+bk)k}.\theta=\frac1{dN}\max\left\{ \log|s^N|+c_0,\ (\log|t_i|+c_i)_i, (\log|r_k|+b_k)_k\right\}. The last group of sections has no common zero, so this formula defines a continuous metric globally. Every branch of (11.8.5) is strictly negative at xx, and the maximum is finite because of c0c_0. Hence v(x)<0v(x)<0 and vv is continuous. Choose 0<ϵ<−v(x)0<\epsilon<-v(x); a neighbourhood of xx then lies in {v<−ϵ}\{v< -\epsilon\}. Moreover, v≤0v\leq 0 implies all inequalities defining K0K_0, so Q⊂K0Q\subset K_0. The set QQ is closed in YanY^{\rm an} and therefore closed in the compact set K0K_0. This proves the required compact containment.

Locally, branches vanishing at a point can be deleted because the constant branch dominates them nearby. The remaining branches are logarithms of analytic units with constants; their maximum is a convex PL function of moment coordinates. This proves local tropical convexity. Theorem 20.8.1, or its local exact-deletion argument, supplies the original smooth psh approximations. The construction uses rational weights but does not require those weights to belong to the value group, nor does it change the field. ◻

In the normalized discrete/trivial framework of [Wu26], Definition 2.3 admits these rational-weighted metrics; Proposition 2.4 is the cited model identification. Thus the theorem reconstructs the barrier geometry for integral projective varieties in that framework. If ω\omega is the initial model background on LL, let bb be the difference of the constructed metric and the initial metric. It is a global continuous PL function there, and θ=ω+d′d′′b\theta=\omega+d'd''b. Inserting this bb and (11.8.4) into Proposition 11.7.3 gives its extension formula on a fixed neighbourhood, with the required background correction. The geometric construction is proved here; the model identification is an explicitly attributed input. Reducible varieties and more general analytic spaces have not been added to this statement.

The remaining comparison input

The preceding results settle the quantitative diagonal step, bounded-family tropical localization and projective barrier geometry. Section 11.9 now assembles these with model approximation to obtain the PL-to-smooth direction on one ordinary neighbourhood, for its explicitly stated nontrivially valued rational-generator category. Proposition 11.22.14 supplies the smooth-to-PL direction for one original smooth psh input in this projective setting. The common-domain passage for arbitrary sequences, identification of the full source generator class and extension to trivial valuation remain separate.

There is also a source dependency that cannot be replaced by the elementary constructions above. Wu invokes Theorem 6.13 of Amini–Piquerez–Wu, Mixed differential calculus and Monge–Ampère equations on tropical varieties. Both inspected bibliographies list it as a 2026 preprint without an identifier. A targeted search of the title and the authors’ publication pages did not locate its primary full text in this revision. Accordingly, the analytic comparison assigned to that theorem is recorded as an unverified imported input. This is an access limitation, not evidence that the paper or theorem does not exist. Corollary 11.7.6 retains its explicit source-dependent status.

References. Wu26, Definitions 2.3 and A.1, Propositions 2.4, 2.11 and 3.28, and Appendix A, pp. 52–54; APW26, Definition 4.3, Example 4.9 and Propositions 5.19 and 5.21, fixed October v1. The counterexample and the common-domain estimates are proved here. Tropical bounded-growth extension and the indicated model comparison remain cited framework inputs.

11.9 From PL sequences to smooth sequences on one neighbourhood

The projective barrier of Section 11.8.4 can be chosen before the functions to be approximated. The bound of an individual function then affects only a scalar in the construction, not the final neighbourhood. Combining this observation with the known approximation of nef model metrics gives one direction of the local comparison without an envelope hypothesis. We state the coefficient and model conventions explicitly: arbitrary real PL weights cannot be identified with rational model metrics without a further argument.

The model-compatible class of generators

Throughout this section KK is complete and nontrivially valued, its absolute value is fixed, and XX is an integral projective KK-variety of dimension at least one. Put Γ=−log⁡∣K×∣\Gamma=-\log|K^\times| and ΓQ=QΓ\Gamma_{\mathbf Q}=\mathbf Q\Gamma. For an ordinary open U⊂XanU\subset X^{\rm an}, write G(U)\mathcal G(U) for the finite continuous functions which locally have the form u=max⁡1≤a≤r{ca+∑i=1mqailog⁡∣fi∣},ca∈ΓQ,qai∈Q,\begin{equation} u=\max_{1\leq a\leq r}\left\{c_a+\sum_{i=1}^m q_{ai}\log|f_i|\right\}, \qquad c_a\in\Gamma_{\mathbf Q},\quad q_{ai}\in\mathbf Q, \end{equation} where the fif_i are analytic units on the neighbourhood in question. Presentations and neighbourhoods may vary with the point. This is the local convex rational PL class used below. It is stable under finite maxima, positive rational scaling, and addition of constants in ΓQ\Gamma_{\mathbf Q}. Products of the finite lists of units give common moment charts for these operations.

Every such local presentation is uniformly approximable by original smooth psh functions on that neighbourhood: apply the regularized maximum to its affine branches. This assertion concerns a single presentation. It does not yet give a common approximation domain for a sequence of functions, each with its own covering by presentations. A finite convex PL function on an ordinary real open set has the displayed form after shrinking to a convex neighbourhood. At a boundary point of an extended tropical space, the reduction to finite coordinates is supplied for ambient convex PL functions in Section 11.10.1. The present theorem assumes (11.9.1); the relative-image class is discussed separately in Remark 11.10.10.

Lemma 11.9.1. (Positivity at a continuous endpoint). On an ordinary open set, suppose uj∈G(U)u_j\in\mathcal G(U) decreases to a finite continuous function uu. Then d′d′′[u]d'd''[u] is positive against the original CLD compactly supported smooth tests.

Proof. Local regularization of (11.9.1) gives positivity of d′d′′[uj]d'd''[u_j]. For a fixed test η\eta, Dini’s lemma gives uniform convergence on its compact support. The smooth top-degree form d′d′′ηd'd''\eta defines a signed measure of finite total variation there, so ⟨d′d′′[uj],η⟩=⟨[uj],d′d′′η⟩⟶⟨[u],d′d′′η⟩.\langle d'd''[u_j],\eta\rangle =\langle[u_j],d'd''\eta\rangle \longrightarrow \langle[u],d'd''\eta\rangle. Nonnegativity on positive tests passes to the limit. No smooth approximating sequence on a common neighbourhood has been used. ◻

We use two model-theoretic inputs. First, on a proper algebraic analytification, integral PL metrics are formal/model metrics; rational versions follow by taking a common tensor power on a finite cover. The model/formal/PL comparison is [GM] (Theorem 1.1). Constants in ΓQ\Gamma_{\mathbf Q} cause no extension of KK: after a tensor power they are logarithms of absolute values of elements of KK. Second, a model metric whose curvature current is positive on the whole projective analytification is nef. This is the model criterion in [CLD25] (Corollary 18.8.8), with the residual mechanism recorded in Theorem 19.18.12 and Section 19.18.13. These are named geometric inputs. For the specific maximum weights below, Proposition 11.10.7 also supplies a route through Gubler–Künnemann functorial positivity and Gubler–Martin descent; its proof uses the displayed generators and does not depend on the present approximation theorem.

A fixed barrier with variable scales

Lemma 11.9.2. For x∈Ux\in U, the data L,Y,v,Q,ϵL,Y,v,Q,\epsilon of Theorem 11.8.9 can be chosen so that all weights of the Fubini–Study metric θ\theta lie in ΓQ\Gamma_{\mathbf Q}. In particular θ\theta is a rational model metric and v∈G(Yan)v\in\mathcal G(Y^{\rm an}).

Open proof obligation (RC2). Prove the value-group barrier in the stated scope; BE21 Theorem 5.14 is invoked.

NAG-900140 · Proof status

Proposition 11.9.3. (Localization with a domain independent of the bound). Fix the preceding barrier, put P={v≤−ϵ}P=\{v\leq-\epsilon\} and V=Int⁡XanPV=\operatorname{Int}_{X^{\rm an}}P, and let u∈G(U)u\in\mathcal G(U). Choose M,τ∈ΓQM,\tau\in\Gamma_{\mathbf Q} with M>0M>0, τ>0\tau>0, ∣u∣≤M|u|\leq M on QQ, and a rational A>0A>0 with Aϵ>2M+τA\epsilon>2M+\tau. Then there is a nef rational model metric mum_u on the fixed line bundle LL whose weight wuw_u relative to the same power-section frame as vv satisfies Awu+M+τ=uon V.\begin{equation} A w_u+M+\tau=u\quad\hbox{on }V. \end{equation} The neighbourhood VV depends only on x,U,Lx,U,L and the chosen barrier. The numbers M,AM,A may depend on uu.

Proof. Define a global function by ϕu={A−1max⁡{u−Av−M−τ,0},on Q,0,on Xan∖Q.mu=θ+ϕu.\begin{equation} \phi_u=\begin{cases} A^{-1}\max\{u-Av-M-\tau,0\},&\text{on }Q,\\ 0,&\text{on }X^{\rm an}\setminus Q. \end{cases} \qquad m_u=\theta+\phi_u. \end{equation} Every boundary point of QQ lies in Yan∩UY^{\rm an}\cap U and has v=0v=0. There u−Av−M−τ≤−τu-Av-M-\tau\leq-\tau. Continuity gives an open collar on which the maximum vanishes, so the formulas glue. The collar may depend on u,Au,A, which is harmless: it is used only to construct this one global metric. Since u,vu,v are locally rational PL with constants in ΓQ\Gamma_{\mathbf Q}, the glued function is rational PL. The model/PL comparison stated above makes mum_u a rational model metric.

On the interior of QQ, its local weight is wu=v+ϕu=max⁡{A−1(u−M−τ),v}.w_u=v+\phi_u =\max\{A^{-1}(u-M-\tau),v\}. Both branches, and their maximum on a common moment chart, are in G\mathcal G. Outside the modification the metric is θ\theta. Hence its curvature current is positive everywhere. The quoted current/model criterion gives nefness; alternatively apply Proposition 11.10.7 directly to these weights. Finally, on PP, Av+M+τ≤−Aϵ+M+τ<−M≤u,Av+M+\tau\leq-A\epsilon+M+\tau<-M\leq u, so the first branch is active and (11.9.2) follows. ◻

Unlike Proposition 11.7.3, this proposition does not claim that the global extensions preserve the order of an entire family. Different scales usually destroy that order. Order is restored only after restricting the approximants to the fixed VV. This is why a uniform bound for the whole family is unnecessary.

Global approximation and the fixed-domain diagonal

Proposition 11.9.4. (A known global approximation input). A nef rational model metric mm on an ample line bundle over a projective KK-scheme is a uniform limit of pure Fubini–Study metrics. Consequently, for every δ>0\delta>0 it admits a global original smooth semipositive metric m~\widetilde m with ∥m~−m∥Xan<δ\|\widetilde m-m\|_{X^{\rm an}}<\delta. Here KK is complete and nontrivially valued.

Open proof obligation (RC2). Prove the required FS density for the nef classes and its exact hypotheses; BE21 Theorem 7.8 is invoked.

NAG-900142 · Proof status

This approximation theorem is already in the literature. Its role here is to supply functions on all of YanY^{\rm an}, so that the local domains no longer depend on the generator being approximated. No envelope continuity theorem is used in this step.

Theorem 11.9.5. (One neighbourhood for a decreasing PL sequence). Let uj∈G(U)u_j\in\mathcal G(U) decrease pointwise to u:U→R∪{−∞}u:U\to\mathbf R\cup\{-\infty\}. For every x∈Ux\in U there is one ordinary open neighbourhood VV of xx and original smooth psh functions bjb_j on VV such that bj↓u∣Vb_j\downarrow u|_V. Neither a uniform bound on the family (uj)(u_j) nor local boundedness or continuity of uu is required. Any generic-finiteness condition imposed in the definition of the endpoint class is preserved.

Proof. Choose the barrier and VV once by Lemma 11.9.2. Fix τ>0\tau>0 in ΓQ\Gamma_{\mathbf Q}. Each uju_j is continuous on the compact set QQ, so choose its own Mj∈ΓQM_j\in\Gamma_{\mathbf Q} and rational Aj>0A_j>0 as in Proposition 11.9.3. This gives nef model metrics mjm_j on the same ample LL, with weights wjw_j satisfying uj=Ajwj+Mj+τon V.u_j=A_jw_j+M_j+\tau\quad\text{on }V. Put ϵj=4−j\epsilon_j=4^{-j}. Apply Proposition 11.9.4 to mjm_j with error less than ϵj/Aj\epsilon_j/A_j. Let w~j\widetilde w_j be the resulting smooth weight in the same power-section frame on YanY^{\rm an}. The function aj=Ajw~j+Mj+τsatisfies∥aj−uj∥V<ϵj.a_j=A_j\widetilde w_j+M_j+\tau \quad\text{satisfies}\quad \|a_j-u_j\|_V<\epsilon_j. It is original smooth and psh because Aj>0A_j>0. A power-section frame defines a normalized weight by dividing by its tensor exponent; it does not require a root of that section over KK. Finally set bj=aj+2ϵjb_j=a_j+2\epsilon_j. Lemma 11.8.1 proves bj↓ub_j\downarrow u on this same VV. ◻

Corollary 11.9.6. Under the hypotheses of Theorem 11.9.5, if uu is finite continuous, then uu is locally uniformly approximable by original smooth psh functions. Thus the continuous endpoints of this PL first-stage class belong to U\mathcal U.

Proof. Shrink to a relatively compact neighbourhood inside VV and apply Dini’s lemma to the decreasing continuous functions bjb_j and their continuous limit. ◻

Remark 11.9.7. This combines Wu’s localization mechanism, a model semipositivity criterion, Gubler–Martin’s PL/model comparison and Boucksom–Eriksson’s approximation theorem. The explicit choices are a fixed neighbourhood and errors ϵj/Aj\epsilon_j/A_j; no priority claim is made. Wu’s convention Z⊆ΓK\mathbf Z\subseteq\Gamma_K places his rational constants in ΓQ\Gamma_{\mathbf Q}. The ambient analytic generators are identified in Theorem 11.10.3; the relative finite-selection presentation is supplied in Theorem 11.12.5, with its explicit convention. Arbitrary real PL weights and trivial valuation are not covered.

First-stage comparison and further stability

Let PG\mathcal P_{\mathcal G} be the local decreasing class generated by G\mathcal G, with the same endpoint convention as the original smooth first-stage class R\mathcal R. Theorem 11.9.5 proves PG⊆R\mathcal P_{\mathcal G}\subseteq\mathcal R. Corollary 11.24.4 now supplies the reverse inclusion in the same nontrivially valued integral-projective category, by rational replacement on a fixed affinoid. Thus PG=R\mathcal P_{\mathcal G}=\mathcal R in this precisely specified first-stage sense.

The following sufficient criterion is retained for its stability implication. Generator membership and first-stage equality are no longer the missing steps in this setting. Closure under decreasing limits of arbitrary sections of the resulting sheaf remains additional: their row domains may vary. Lemma 11.24.5 handles common-domain rows, but does not provide a common domain from termwise germ membership.

Proposition 11.9.8. (A sufficient stability criterion). Suppose every original smooth psh function locally belongs to PG\mathcal P_{\mathcal G}, and PG\mathcal P_{\mathcal G} is closed under decreasing limits of its sections on ordinary open sets whenever the endpoint is allowed. Then R=PG\mathcal R=\mathcal P_{\mathcal G}. In particular, its bounded decreasing stability implies the continuous-stage equalities of Theorem 11.7.5.

Proof. On a neighbourhood representing a section of R\mathcal R, each smooth predecessor is a section of the sheaf PG\mathcal P_{\mathcal G} by the first hypothesis. The second hypothesis places their decreasing limit there on that same open set. This proves the reverse inclusion; combine it with Theorem 11.9.5. Apply Theorem 11.7.5 for the last assertion. ◻

11.10 Convex PL generators at the extended tropical boundary

The logarithms in the presentation of G\mathcal G are logarithms of units. An analytic tropicalization, however, may be defined by functions with zeros, so its image meets the boundary of Tk\mathbb T^k, where T=R∪{+∞}\mathbb T=\mathbf R\cup\{+\infty\}. We explain why this distinction disappears locally for finite ambient convex PL functions. The argument expands [Wu26] (Remark 3.2); it also makes precise the absolute-value estimate used in Remark 3.9 of that paper. The word ambient is essential: the defining function is given on an open subset of Tk\mathbb T^k, not merely on a tropical image.

Finite affine selections and boundary factorization

We use the finite-valued convention of [Wu26] (Definition 3.1(1)–(2)). On an ambient open Ω⊂Tk\Omega\subset\mathbb T^k, an affine function is a finite continuous function ℓ(t)=c+∑iaiti\ell(t)=c+\sum_i a_i t_i for which every summand is finite; in particular ai=0a_i=0 if Ω\Omega contains a point with ti=+∞t_i=+\infty. A finite PL function is locally a continuous selection from a finite list of such affine functions. Locally convex means that at every point there is a touching affine lower support on a neighbourhood. Thus the contact condition is f(z)=ℓ(z)f(z)=\ell(z) and f≥ℓf\geq\ell there, with the correction explained in Example A.8.1.

Lemma 11.10.1. (Eliminating infinite coordinates). Let z∈Ts×Rdz\in\mathbb T^s\times\mathbf R^d have its first ss coordinates infinite and its remaining coordinates equal to yy. Let ff be finite PL on an ambient open neighbourhood of zz. There are R∈RR\in\mathbf R, an open ball B⊂RdB\subset\mathbf R^d about yy, and a finite continuous PL function FF on BB such that N=(R,+∞]s×B,f∣N=F∘p,\begin{equation} N=(R,+\infty]^s\times B,\qquad f|_N=F\circ p, \end{equation} where pp is the projection onto the last dd coordinates. The affine pieces of FF are obtained by deleting the first ss coordinates from one local affine list for ff. If d=0d=0, the conclusion is that ff is constant on NN.

Proof. Choose one open patch containing zz on which ff selects from finitely many affine functions ℓ1,…,ℓr\ell_1,\ldots,\ell_r, and choose the product NN inside that patch. Finiteness of each summand at zz forces all coefficients of the first ss coordinates to vanish. Write the remaining affine functions as LaL_a on BB. For fixed b∈Bb\in B, the continuous function ff on the connected fibre (R,+∞]s×{b}(R,+\infty]^s\times\{b\} takes values in the finite set {La(b):1≤a≤r}\{L_a(b):1\leq a\leq r\}. It is therefore constant on that fibre. Evaluation along any fixed section of pp defines a continuous FF, and F(b)F(b) is one of the La(b)L_a(b). This proves the factorization and the PL assertion. ◻

Independence of the infinite coordinates is not yet a maximum presentation. The following elementary real argument supplies that remaining step, and keeps the original affine coefficients.

Lemma 11.10.2. (From affine selection to a convex maximum). Let B⊂RdB\subset\mathbf R^d be an open ball and let FF be continuous, with F(b)∈{L1(b),…,Lr(b)}F(b)\in\{L_1(b),\ldots,L_r(b)\} for affine La(b)=ca+⟨aa,b⟩L_a(b)=c_a+\langle a_a,b\rangle. Then FF is Lipschitz on BB, with constant max⁡a∥aa∥\max_a\|a_a\| for the Euclidean norm. If FF has a local touching affine lower support at every point, it is convex and there is a nonempty subset I⊆{1,…,r}I\subseteq\{1,\ldots,r\} such that F=max⁡a∈ILaon B.\begin{equation} F=\max_{a\in I} L_a\quad\text{on }B. \end{equation}

Proof. On a line segment in BB, restrict the finite list of affine functions and discard identical restrictions. Distinct restrictions intersect at at most one point. Between these finitely many intersection points, continuity forces FF to follow one affine restriction. Its slope on each interval is bounded by max⁡a∥aa∥\max_a\|a_a\| times the speed of the segment. Summing the interval estimates proves the Lipschitz bound.

At an interior breakpoint of the segment, a touching affine lower support has slope mm satisfying m−≤m≤m+m_-\leq m\leq m_+, where m−m_- and m+m_+ are the left and right slopes of FF. Hence the slopes never decrease. The restriction to every segment is convex, so FF is convex on BB.

Discard duplicate affine functions on BB. The hyperplanes La=LbL_a=L_b cut a dense open subset of BB into finitely many chambers. On each chamber the selected affine function is fixed. Let II consist of the indices selected on some chamber. If a∈Ia\in I, choose bb in that chamber. For any c∈Bc\in B and sufficiently small t>0t>0, the point b+t(c−b)b+t(c-b) remains in the chamber. Convexity gives (1−t)La(b)+tLa(c)=F(b+t(c−b))≤(1−t)F(b)+tF(c),(1-t)L_a(b)+tL_a(c)=F(b+t(c-b)) \leq(1-t)F(b)+tF(c), so La(c)≤F(c)L_a(c)\leq F(c). Equality with one of these supporting pieces holds on every chamber and hence, by continuity, throughout BB. This proves (11.10.2). For d=0d=0 both assertions are immediate. ◻

If the function ff of Lemma 11.10.1 is locally convex, then FF has the lower supports required by Lemma 11.10.2: fix finite values of the first ss coordinates and restrict each local support to that slice. Thus a finite ambient convex PL function near a boundary point is a maximum of finitely many affine functions of the finite coordinates alone. No convex-extension theorem for a tropical variety enters this argument.

Analytic pullbacks and the value group

Fix the absolute value of KK and put Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|, ΓQ=QΓ\Gamma_{\mathbf Q}=\mathbf Q\Gamma. In this subsection a compatible PL function has rational slopes and constants in ΓQ\Gamma_{\mathbf Q} in its local affine lists. This is a coefficient convention, not a normalization of the field.

Theorem 11.10.3. (Ambient analytic generators and G\mathcal G). Let WW be an ordinary open in a KK-analytic space, let g1,…,gkg_1,\ldots,g_k be analytic functions on WW, and write trop⁡g=(−log⁡∣g1∣,…,−log⁡∣gk∣)\operatorname{trop}_g=(-\log|g_1|,\ldots,-\log|g_k|), with −log⁡0=+∞-\log 0=+\infty. Suppose ff is a finite compatible locally convex PL function on an ambient open Ω⊂Tk\Omega\subset\mathbb T^k containing trop⁡g(W)\operatorname{trop}_g(W). Then u=f∘trop⁡gu=f\circ\operatorname{trop}_g locally has the unit-logarithm maximum presentation (11.9.1).

Conversely, if KK is nontrivially valued, every germ with that presentation can be written as H∘trop⁡hH\circ\operatorname{trop}_h for analytic units h1,…,hrh_1,\ldots,h_r and a finite compatible locally convex PL function HH on all of Tr\mathbb T^r. Consequently these global-ambient analytic generators and G\mathcal G define the same sheaf of germs.

Proof. Fix x∈Wx\in W. Reorder the coordinates so that the functions vanishing at xx are precisely the first ss functions. Apply Lemmas 11.10.1 and 11.10.2 at trop⁡g(x)\operatorname{trop}_g(x). On a product neighbourhood NN of that point, f(t)=max⁡a∈I(ca+∑i>saaiti),aai∈Q,ca∈ΓQ.f(t)=\max_{a\in I}\left(c_a+\sum_{i>s} a_{ai}t_i\right), \qquad a_{ai}\in\mathbf Q,\quad c_a\in\Gamma_{\mathbf Q}. On the inverse image of NN, shrink about xx so that the remaining gig_i are analytic units. Substitution gives (11.9.1), with qai=−aaiq_{ai}=-a_{ai}. The functions that vanish at xx have been deleted exactly, not assigned logarithmic values by a limiting convention. If all gi(x)g_i(x) vanish, the pullback is constant near xx.

For the converse write u=max⁡a(ca+∑iqailog⁡∣fi∣)u=\max_a(c_a+\sum_i q_{ai}\log|f_i|) on a neighbourhood of xx. Choose a positive integer mm with mqai∈Zmq_{ai}\in\mathbf Z and put ha=∏ifimqaih_a=\prod_i f_i^{mq_{ai}}. Negative exponents are allowed because the fif_i are units. Nontrivial valuation makes ΓQ\Gamma_{\mathbf Q} dense in R\mathbf R. Choose c0∈ΓQc_0\in\Gamma_{\mathbf Q} with c0<u(x)c_0<u(x) and shrink so that c0<uc_0<u. Define H(t1,…,tr)=max⁡{c0,c1−t1/m,…,cr−tr/m},\begin{equation} H(t_1,\ldots,t_r)=\max\{c_0,c_1-t_1/m,\ldots,c_r-t_r/m\}, \end{equation} where a branch with ta=+∞t_a=+\infty contributes −∞-\infty to this maximum. At such a coordinate, the inequality ta>m(ca−c0)t_a>m(c_a-c_0) deletes that branch on a neighbourhood. The remaining branches are finite affine functions. More explicitly, cover each coordinate by R\mathbf R and (m(ca−c0),+∞](m(c_a-c_0),+\infty]; their finitely many product sets give finite affine presentations after the inactive branches are deleted. Thus HH is finite continuous PL on all of Tr\mathbb T^r, with the stated coefficients. On each such neighbourhood it is a maximum of affine functions and hence has touching lower supports. Finally H∘trop⁡h=max⁡(c0,u)=uH\circ\operatorname{trop}_h=\max(c_0,u)=u on the chosen neighbourhood. No extension of KK or change of its absolute value has occurred. ◻

Remark 11.10.4. Wu assumes Z⊆ΓK\mathbf Z\subseteq\Gamma_K for a nontrivially valued field in [Wu26] (Section 2.1, p. 5). Under that convention, the rational constants in Definition 3.1 belong to ΓQ\Gamma_{\mathbf Q}, so the preceding theorem applies to the ambient analytic generators in Remark 3.9, after correcting the touching-support formula. For a fixed absolute value without this convention, rational constants need not lie in ΓQ\Gamma_{\mathbf Q}; our theorem retains the actual value-group condition. The boundary factorization itself allows real coefficients, but the rational model argument of Section 11.9 does not thereby acquire real coefficients.

Approximation through zeros

At a zero of gg, uniform control of log⁡∣g∣\log|g| is unavailable. The useful coordinates for perturbing the analytic functions are their absolute values. Boundary factorization provides precisely the needed regularity in these coordinates.

Proposition 11.10.5. (A local Lipschitz estimate in absolute values). Let ff be any finite PL function with real affine pieces on an ambient open Ω⊂Tk\Omega\subset\mathbb T^k. Define D=(−log⁡)−1(Ω)⊂[0,+∞)kD=(-\log)^{-1}(\Omega)\subset[0,+\infty)^k, with −log⁡0=+∞-\log 0=+\infty. Then h=f∘(−log⁡)h=f\circ(-\log) is locally Lipschitz on DD for the Euclidean distance. Its restriction to every compact subset C⊂DC\subset D is Lipschitz. Convexity is not required.

Proof. Fix r∈Dr\in D, with its first ss coordinates zero and the rest positive. Lemma 11.10.1 makes hh independent of the first ss coordinates on a neighbourhood of rr. For the remaining coordinates choose a neighbourhood on which ri≥a>0r_i\geq a>0 and the logarithmic image lies in the ball BB of that lemma. The first part of Lemma 11.10.2 gives a Lipschitz constant LL for FF on BB. Since ∣log⁡ri−log⁡ri′∣≤a−1∣ri−ri′∣,|\log r_i-\log r_i'|\leq a^{-1}|r_i-r_i'|, the resulting local constant for hh is at most L/aL/a. If there are no positive coordinates, hh is locally constant.

To pass to a compact CC, take a finite cover by relative neighbourhoods on which these pairwise Lipschitz estimates hold. A Lebesgue number δ>0\delta>0 ensures that two points of CC at distance less than δ\delta lie in one member of the cover. Use the largest of the finitely many local constants for these pairs. For the other pairs, use the bound osc⁡C(h)/δ\operatorname{osc}_C(h)/\delta. The maximum of the two constants works on all of CC. ◻

Corollary 11.10.6. Let CC be compact. Suppose that continuous maps rj,rr_j,r from CC to [0,+∞)k[0,+\infty)^k satisfy rj→rr_j\to r uniformly and r(C)⊂Dr(C)\subset D. For all sufficiently large jj, the compositions are defined and ∥f∘(−log⁡)∘rj−f∘(−log⁡)∘r∥C≤LC∥rj−r∥C⟶0\begin{equation} \|f\circ(-\log)\circ r_j-f\circ(-\log)\circ r\|_C \leq L_C\|r_j-r\|_C\longrightarrow0 \end{equation} for a constant LCL_C independent of jj. In particular, this applies to uniformly approximating tuples of absolute values of analytic functions, including at their zeros.

Proof. The compact image r(C)r(C) has a compact neighbourhood inside the relatively open set DD. Uniform convergence places all sufficiently large images rj(C)r_j(C) in that neighbourhood. Apply Proposition 11.10.5 there. ◻

This proves the change-of-coordinates estimate used in [Wu26] (Remark 3.9). It does not construct the algebraic approximants to the analytic functions: that is a separate approximation assertion on the specified rational domain. Once absolute-value approximants on a fixed compact domain are supplied, the displayed estimate transfers their error without taking logarithms at zeros.

Model semipositivity through functorial positivity

For these generators the model-nefness step also follows from the earlier work of Gubler–Künnemann and Gubler–Martin. The reason is stronger than positivity of a single current: the same maximum presentation survives pullback. This gives another route for the metrics constructed in Proposition 11.9.3.

Proposition 11.10.7. (A criterion for the constructed model metrics). Let KK be complete and nontrivially valued, let XX be a proper KK-scheme, and let mm be a rational model metric on a line bundle LL. Suppose its local weights have the maximum presentation (11.9.1). Then mm is a nef model metric.

Open proof obligation (RC2). Establish the model-nef criterion and its functoriality in the stated category; the proof uses GK19 and GM.

NAG-900151 · Proof status

This proof uses the published model comparison and descent theorems as inputs; it does not re-prove their lifting arguments. The passage through CKC_K descends a numerical property of a model, rather than an equality of currents or representatives. The local regularizations and the final sequence of Theorem 11.9.5 are still constructed over the original KK. The criterion is specific to weights with the displayed presentations and does not assert that arbitrary positive currents are functorially positive.

The resulting regularization statement

Corollary 11.10.8. (Decreasing ambient PL generators). Let KK be complete and nontrivially valued, let XX be an integral projective KK-variety of positive dimension, and let U⊂XanU\subset X^{\rm an} be ordinary open. Suppose the finite continuous functions uju_j decrease pointwise to uu, and that each uju_j locally admits an ambient analytic presentation as in Theorem 11.10.3, with rational slopes and constants in ΓQ\Gamma_{\mathbf Q}. Then near each x∈Ux\in U there is one ordinary open VV and original smooth psh functions bjb_j on VV with bj↓u∣Vb_j\downarrow u|_V. The analytic functions in the presentations may have zeros; their charts and neighbourhoods may vary with both jj and the point. If uu is finite continuous, the convergence is uniform on compact subsets of VV.

Proof. Theorem 11.10.3 gives uj∈G(U)u_j\in\mathcal G(U) for every jj as a local membership assertion on the original UU. Apply Theorem 11.9.5. Its projective barrier construction, not an intersection of countably many presentation neighbourhoods, provides the common VV. The last assertion is Dini’s lemma on each compact subset. ◻

A relative touching-support condition on a polyhedral set must be distinguished from ambient convexity even before passing to analytic spaces. The following example explains why a same-chart extension cannot be taken for granted.

Example 11.10.9. (Relative supports do not give a same-chart convex extension). In R2\mathbf R^2, let Σ\Sigma be the one-dimensional fan with primitive ray vectors v1=(1,0),v2=(0,1),v3=(1,1),v4=(−1,−1),(w1,w2,w3,w4)=(1,1,1,2).\begin{gathered} v_1=(1,0),\quad v_2=(0,1),\quad v_3=(1,1),\quad v_4=(-1,-1),\\ (w_1,w_2,w_3,w_4)=(1,1,1,2). \end{gathered} It is positively weighted and balanced. Define a continuous rational PL function on its support by F(tv3)=tF(tv_3)=t for t≥0t\geq 0, and F(tvi)=0F(tv_i)=0 for i=1,2,4i=1,2,4. At every point of ∣Σ∣|\Sigma|, FF has an affine lower support touching on a relative neighbourhood. Nevertheless its germ at the origin is not the restriction of an ambient convex function on any open convex neighbourhood of the origin.

Proof. Balancing is the identity v1+v2+v3+2v4=0v_1+v_2+v_3+2v_4=0. The function is continuous at the origin and affine on each ray; it selects from 00 and (x+y)/2(x+y)/2. At the origin, the constant zero function is a touching lower support. At a nonzero point a sufficiently small relative neighbourhood meets only its ray, where either 00 or (x+y)/2(x+y)/2 agrees with FF. Suppose that a convex ambient extension F~\widetilde F existed near the origin. For sufficiently small t>0t>0, its midpoint inequality would give t=F~(t,t)≤12F~(2t,0)+12F~(0,2t)=0,t=\widetilde F(t,t) \leq\tfrac12\widetilde F(2t,0)+\tfrac12\widetilde F(0,2t)=0, a contradiction. The obstruction depends on the presentation. Indeed, take the balanced fan Σ~\widetilde\Sigma in R3\mathbf R^3 with rays generated by e1,e2,e3,−e1−e2−e3e_1,e_2,e_3,-e_1-e_2-e_3, all of weight one. The integral linear map A(x,y,z)=(x+z,y+z)A(x,y,z)=(x+z,y+z) maps these generators to v1,v2,v3,2v4v_1,v_2,v_3,2v_4, respectively. It maps the supports homeomorphically, and its ray lattice indices give the weights (1,1,1,2)(1,1,1,2) on Σ\Sigma. On the lifted support one has F∘A=max⁡{z,0}.F\circ A=\max\{z,0\}. This is an ambient convex PL presentation after changing the fan. This calculation is polyhedral; Proposition 11.16.1 realizes it by analytic units over fields with ∣2∣=1|2|=1. ◻

Remark 11.10.10 (The remaining source-class interface). The ambient analytic generators of [Wu26] (Remark 3.9) are now covered, with the stated coefficient convention and corrected convexity. Definitions 3.6–3.7 there instead use PL and convex functions on Trop⁡(W)\operatorname{Trop}(W), and the two properties may initially be presented in different charts. Such an image need not be ambient open. A general identification therefore needs a presentation theorem with any necessary changes of chart, not a presumed same-chart convex extension. Example 11.10.9 refutes that elementary extension shortcut; Section 11.16 supplies its analytic realization and a smooth psh variant, but no obstruction after changing tropicalization. The unit-chart PL identification and the relative-support positivity step are proved in Section 11.11. Section 11.12 completes the extended-image PL identification for its explicit relative finite-selection convention. The prescribed-affinoid reverse replacement and the comparison of four explicit first-stage classes are proved in Section 11.24.3. Source-category identification and sheaf-level decreasing closure remain separate before identifying the full classes in Theorem A.3.

A transfer for locally semipositive rational PL functions

There is another way to use the projective barrier without demanding an ambient maximum presentation for every generator. Assume directly that the rational PL generators are locally semipositive in the model sense. The maximum theorem of Gubler–Rabinoff then supplies the positivity of the patched metric. We record the precise synthesis because it isolates a potentially weaker interface for comparison with another local psh theory.

Write SQ(U)\mathcal S_{\mathbf Q}(U) for the finite continuous Q\mathbf Q-PL functions on an ordinary open UU whose residues are nef at every point, in the sense of [GJR] (Proposition 7.1 and Definition 7.2). Equivalently, locally a tensor power of the metric ∥1∥=e−h\|1\|=e^{-h} is given on a strictly analytic neighbourhood by a nef formal model. The equivalence is the cited proposition; it is not deduced merely from positivity of d′d′′[h]d'd''[h]. Rational PL here uses logarithms of analytic units on a strict analytic domain cover, so constants in ΓQ\Gamma_{\mathbf Q} are admitted.

Theorem 11.10.11. (Fixed-domain transfer from locally nef PL generators). Let KK be complete and nontrivially valued, let XX be an integral projective KK-variety of positive dimension, and let U⊂XanU\subset X^{\rm an} be ordinary open. If uj∈SQ(U)u_j\in\mathcal S_{\mathbf Q}(U) decreases pointwise to u:U→R∪{−∞}u:U\to\mathbf R\cup\{-\infty\}, then every x∈Ux\in U has one ordinary open neighbourhood VV carrying original smooth psh functions bj↓u∣Vb_j\downarrow u|_V. No uniform bound on the family and no local boundedness of the endpoint are required. For a finite continuous endpoint the convergence is uniform on compact subsets of VV.

Open proof obligation (RC2). Prove the transfer between nef, PL and approximated metrics in the original hypotheses; the proof uses the listed GM, GRPL and GJR inputs.

NAG-900153 · Proof status

Every member of G\mathcal G is locally in SQ\mathcal S_{\mathbf Q}: its affine unit-logarithm branches are rationally harmonic, and repeated application of the same maximum theorem preserves semipositivity. The new theorem therefore gives a second proof of the earlier transfer. More generally, it suffices to verify model compatibility and local semipositivity of a proposed source generator; an ambient convex extension is unnecessary for this route. Theorem 11.11.6 completes the semipositivity step whenever the function is already model-compatible PL; Proposition 11.11.5 verifies that membership for unit-chart relative presentations. Theorem 11.12.5 now supplies extended relative-image membership under its common-branch convention. The rational coefficient restriction also remains substantive: [GRPL] (the paragraph following Corollary 5.5) warns that maxima need not remain Λ\Lambda-PL in the strict domain category for larger groups Λ\Lambda.

For a single h∈SQ(U)h\in\mathcal S_{\mathbf Q}(U), apply Theorem 11.10.11 to the constant sequence uj=hu_j=h. This gives decreasing original smooth psh approximants near each point. The related [GRPL] (Theorem 1.5) applies to semipositive real PL functions on a boundaryless strictly analytic space and gives uniform approximation by locally Fubini–Study functions. The present argument keeps rational coefficients and a projective ambient space, while handling a whole decreasing family on one neighbourhood and choosing original smooth approximants. The locally Fubini–Study conclusion and the original smooth conclusion should not be identified without the smoothing step. Likewise, [GRPL] (Theorem 1.4) gives semipositivity of a pointwise limit only when that limit is still real PL; it does not supply the decreasing closure of an arbitrary generated psh class needed in Section 11.9.4.

11.11 Relative supports and the curvewise route to nefness

The obstruction in Example 11.10.9 concerns extension in a fixed tropical chart. It does not obstruct a different route: pull a touching affine support back to an analytic curve, use harmonic comparison there, and then apply the curve criterion for semipositive PL functions. The two steps to distinguish are positivity and membership in the model-compatible PL class. This section proves the positivity step even at extended tropical boundary points, and proves the PL identification when the presentation uses analytic units.

A lower-contact criterion on curves

We first isolate a comparison argument that requires no polyhedral subdivision and no differentiability. Harmonicity in this subsection is Thuillier harmonicity on smooth strictly analytic curves over a nontrivially valued field. We use its vector-space property, its local maximum principle, and the comparison definition of subharmonicity; precise modern references are [GRPSH] (Sections 3.2–3.3 and Proposition 3.14).

Lemma 11.11.1. (Harmonic lower contacts). Let CC be a smooth strictly analytic curve over a complete nontrivially valued field, and let u:C→Ru:C\to\mathbf R be continuous. Suppose that every z∈Cz\in C has an open neighbourhood NzN_z and a harmonic function hzh_z on NzN_z such that hz≤uon Nz,hz(z)=u(z).\begin{equation} h_z\leq u\quad\hbox{on }N_z,\qquad h_z(z)=u(z). \end{equation} Then uu is subharmonic.

Open proof obligation (RC2). Establish the comparison for the pulled-back function on the analytic curve; GRPSH 3.1(3) used.

NAG-900154 · Proof status

The contact inequality is used on a whole neighbourhood. Convexity on individual cells would not give it at a branching point. The proof also tests every interior point of DD; no restriction to skeleton vertices or to a chosen collection of outgoing directions has been made.

Relative tropical supports, including zeros

Here is the exact relative interpretation used below. For a subset S⊂TrS\subset\mathbb T^r, a finite continuous F:S→RF:S\to\mathbf R has relative affine lower supports if for every t∈St\in S there is a relative open neighbourhood Ot⊂SO_t\subset S and an expression ℓt(s)=c+∑iaisi\ell_t(s)=c+\sum_i a_i s_i, with real coefficients, such that every summand is finite on OtO_t, F≥ℓtF\geq\ell_t there, and F(t)=ℓt(t)F(t)=\ell_t(t). This is the corrected touching convention of Example A.8.1, now explicitly on the relative image. We do not silently extend a definition stated only for ambient opens to arbitrary subsets.

Theorem 11.11.2. (Relative supports give classical psh functions). Let KK be any complete non-Archimedean field and WW a KK-analytic space. Suppose locally on ordinary opens that u=F∘τ,τ=(−log⁡∣g1∣,…,−log⁡∣gr∣),gi∈O(W),u=F\circ\tau,\qquad \tau=(-\log|g_1|,\ldots,-\log|g_r|),\qquad g_i\in\mathcal O(W), where FF is finite continuous on τ(W)\tau(W) and has relative affine lower supports in the convention just specified. Then uu is classically psh in the sense of [GRPSH] (Definition 4.1). The functions gig_i may vanish; no PL assumption on FF is needed.

Open proof obligation (RC2). Prove the curvewise test in the specified relative setting; the proof uses GRPSH Proposition 3.14.

NAG-900155 · Proof status

Historical proof-local display withheld.\begin{equation} \text{Historical proof-local display withheld.} \end{equation}

Corollary 11.11.3. (Decreasing limits in the curvewise class). A pointwise decreasing sequence of functions covered by Theorem 11.11.2, all defined on the same ordinary open UU, has a classically psh limit u:U→R∪{−∞}u:U\to\mathbf R\cup\{-\infty\}. The presentations may vary with the sequence index.

Open proof obligation (RC2). Prove the relevant limit property inside the curvewise psh class; the proof applies GRPSH Proposition 4.2(4).

NAG-900156 · Proof status

Identifying PL functions on a unit chart

We now impose the nontrivial valuation needed for the strict model theory. Put Γ=−log⁡∣K×∣\Gamma=-\log|K^\times| and ΓQ=QΓ\Gamma_{\mathbf Q}=\mathbf Q\Gamma. A rational affine expression has rational slopes and its constant in ΓQ\Gamma_{\mathbf Q}. This coefficient condition is essential for the strict analytic domains in the proof.

Lemma 11.11.4. (Selection on a finite polyhedral image). Let S⊂RrS\subset\mathbf R^r be the support of a finite rational ΓQ\Gamma_{\mathbf Q}-polytopal complex. Let F:S→RF:S\to\mathbf R be continuous and satisfy F(t)∈{L1(t),…,Lm(t)}F(t)\in\{L_1(t),\ldots,L_m(t)\} for finitely many rational affine expressions LaL_a. There is a finite rational ΓQ\Gamma_{\mathbf Q}-polytopal subdivision of SS on each of whose closed cells FF equals one LaL_a.

Proof. Subdivide every cell by all hyperplanes La=LbL_a=L_b. On the relative interior σ∘\sigma^\circ of any resulting cell, two branches either agree identically or are everywhere distinct, and their order is fixed. At a point of σ∘\sigma^\circ, distinct values have a positive separation, so continuity of FF makes the chosen equivalence class of branches locally constant. The set σ∘\sigma^\circ is connected; hence one branch is selected throughout it. Continuity extends the identity to σ\sigma. Finiteness and rationality follow from the finite hyperplane arrangement. This argument uses connected polyhedral cells, not connected fibres of the tropicalization. ◻

Proposition 11.11.5. (Relative unit-chart PL functions are model-compatible). Let XX be a good strictly KK-analytic space of pure dimension, with KK complete and nontrivially valued. Suppose that, locally on ordinary opens WW, a finite continuous function uu has a presentation u=F∘τu=F\circ\tau, where τ=(−log⁡∣g1∣,…,−log⁡∣gr∣)\tau=(-\log|g_1|,\ldots,-\log|g_r|) for analytic units gig_i on WW, and FF is locally a continuous finite selection of rational affine expressions on the relative image τ(W)\tau(W). Then uu is Q-PL in the sense of [GJR] (Definition 5.1). No convexity is assumed.

Open proof obligation (RC2). Prove the polyhedral image and analytic sheaf gluing required by the local presentation; GJR section 3.2 and 5.2(4).

NAG-900158 · Proof status

Nef residues and the fixed-domain consequence

Theorem 11.11.6. (The relative-support model criterion). Let XX be good and strictly analytic over a complete nontrivially valued KK. Let uu be an R-PL function in the GJR sense. If uu has the relative-support presentations of Theorem 11.11.2, then its residue line bundle Lu(x)L_u(x) is nef at every x∈Xx\in X. In particular, on a pure-dimensional XX, it suffices that uu have both a unit-chart rational PL presentation as in Proposition 11.11.5 and a relative-support presentation. These two presentations need not use the same tropicalization.

Open proof obligation (RC2). Prove the model-nef comparison and the residual criterion; GRPSH Theorem 4.3 and GRPL Proposition 5.1/Corollary 5.4.

NAG-900159 · Proof status

There is a version-numbering discrepancy worth recording for citation. The proof of [GRPSH] (Theorem 4.3) in arXiv:2506.13548v2 refers to Proposition 3.1 and Corollary 3.5 of its companion paper. In the fixed arXiv:2503.05513v1 consulted here, the corresponding statements are Proposition 5.1 and Corollary 5.4. We use the latter locators for that file; the discrepancy is bibliographic and does not invalidate the comparison theorem.

Corollary 11.11.7. (Relative generators on one final neighbourhood). Let XX be integral projective of positive dimension over a complete nontrivially valued KK, and let U⊂XanU\subset X^{\mathrm{an}} be ordinary open. Suppose uj:U→Ru_j:U\to\mathbf R decrease pointwise to uu, each uju_j has the relative-support presentations above, and each uju_j is Q-PL in the GJR sense. For every x∈Ux\in U there are one ordinary open neighbourhood V⊂UV\subset U and original smooth psh functions bjb_j on VV decreasing to u∣Vu|_V. It is enough to verify the Q-PL hypothesis by the local unit-chart presentations of Proposition 11.11.5. If uu is finite continuous, convergence is uniform on compact subsets of VV.

Proof. Theorem 11.11.6 gives nef residues for every uju_j. Apply Theorem 11.10.11. In that theorem the barrier and final neighbourhood are fixed before the sequence terms are extended; their bounds and scaling factors may depend on jj. Consequently no intersection of countably many chart neighbourhoods is taken here. The last assertion is Dini’s theorem on each compact subset. ◻

Example 11.11.8. (Cellwise affine does not suffice). Choose a∈Ka\in K with 0<∣a∣<10<|a|<1, and on the open annulus ∣a∣<∣T∣<∣a∣−1|a|<|T|<|a|^{-1} put u=−∣log⁡∣T∣∣u=-|\log|T||. This function is Q-PL and affine on each of the two skeletal intervals separated by ∣T∣=1|T|=1, but it is not subharmonic. Indeed, choose r∈∣K×∣Qr\in|K^\times|^{\mathbf Q} with ∣a∣<r<1|a|<r<1. On the strict closed annulus r≤∣T∣≤r−1r\leq|T|\leq r^{-1}, the constant harmonic function H=log⁡r<0H=\log r<0 agrees with uu at both boundary points, whereas u=0>Hu=0>H at the unit-radius Gauss point. Thus harmonic comparison fails. Equivalently, a touching affine lower support at the central tropical point would require a slope simultaneously at least 11 and at most −1-1.

Example 11.11.9. (A monomial inequality at a common zero). In the closed unit bidisc B=M(K⟨X,Y⟩)B=\mathcal M(K\langle X,Y\rangle) over a complete nontrivially valued field, put D={∣X∣≤∣Y∣}D=\{|X|\leq|Y|\}. There is no rational domain RR of BB with 0∈R⊂D0\in R\subset D. In particular DD is not a finite union of rational domains of BB. This explains why the Laurent-domain step in Proposition 11.11.5 cannot simply be repeated with vanishing coordinates.

Proof. Write a rational domain containing 00 in the form ∣fi∣≤ri∣g∣|f_i|\leq r_i|g|, where ri>0r_i>0 and g,f1,…,fmg,f_1,\ldots,f_m generate the unit ideal. Necessarily g(0)≠0g(0)\ne 0, since otherwise all fi(0)f_i(0) would vanish. An analytic function nonzero at the KK-rational point 00 has constant absolute value on some ordinary neighbourhood of 00: use ∣f−f(0)∣<∣f(0)∣|f-f(0)|<|f(0)|. For the finitely many fif_i vanishing at 00, continuity makes ∣fi∣<ri∣g(0)∣|f_i|<r_i|g(0)| after shrinking. Hence every such rational domain contains an ordinary neighbourhood of 00. But, for any a∈Ka\in K with 0<∣a∣<10<|a|<1, the points (an,0)(a^n,0) converge to 00 and lie outside DD. This proves the claim. ◻

Two local proof repairs in tropical regularization

The reverse approximation direction leads back to [APW26] (Proposition 5.21). We record two elementary repairs in its surrounding convexity and localization arguments. They concern the stated auxiliary claim and a displayed zero set, not counterexamples to the intended regularization theorem.

Proposition 11.11.10. (The missing hypothesis in the interval claim). A continuous g:[0,1]→Rg:[0,1]\to\mathbf R with a touching affine lower support at every interior point is convex. If an affine ℓ\ell also touches gg from below on [0,ϵ)[0,\epsilon) at 00, then g≥ℓg\geq\ell on all of [0,1][0,1]. The contact condition at 00 alone does not imply this conclusion.

Proof. On any interval [a,b]⊂[0,1][a,b]\subset[0,1], let HH be the affine interpolant of the endpoint values. If g−Hg-H had a positive maximum MM, its maximum set would be a nonempty compact subset of (a,b)(a,b). At a maximum point, subtracting HH from a touching affine support gives an affine function with an interior local maximum, hence a constant. As in Lemma 11.11.1, the maximum set is therefore open as well as closed in [a,b][a,b], a contradiction. Thus g≤Hg\leq H, which proves convexity. If g(t)<ℓ(t)g(t)<\ell(t) for some t>0t>0, convexity gives, for 0<s<min⁡(t,ϵ)0<s<\min(t,\epsilon), g(s)≤(1−s/t)g(0)+(s/t)g(t)<(1−s/t)ℓ(0)+(s/t)ℓ(t)=ℓ(s),g(s)\leq (1-s/t)g(0)+(s/t)g(t) <(1-s/t)\ell(0)+(s/t)\ell(t)=\ell(s), contradicting the lower support near 00. Without the interior hypothesis, g(t)=t−2t2g(t)=t-2t^2 and ℓ=0\ell=0 give a counterexample: g≥0g\geq 0 near 00, but g(1)=−1g(1)=-1. ◻

The auxiliary interval claim printed in the proof of [APW26] (Proposition 3.3), p. 23, states only the condition at 00. Its subsequent proof uses a support at an interior maximum point. The proposition’s actual local-convexity hypothesis supplies exactly that missing condition, so the corrected claim above repairs this inference.

Lemma 11.11.11. (The strict-support zero set). Let QQ be a compact Hausdorff space, x∈Qx\in Q, and let A,BA,B be continuous nonnegative functions with A−1(0)={x}A^{-1}(0)=\{x\} and B(x)=0B(x)=0. For positive decreasing sequences Cj,δj→0C_j,\delta_j\to 0, put Vj={A<CjB+δj}⊂Q.V_j=\{A<C_jB+\delta_j\}\subset Q. Then x∈Vjx\in V_j, the closures are nested, and ⋂jVj‾=A−1(0)={x}.\begin{equation} \bigcap_j\overline{V_j}=A^{-1}(0)=\{x\}. \end{equation} For every neighbourhood OO of xx, eventually Vj‾⊂O\overline{V_j}\subset O.

Proof. Positivity of δj\delta_j gives x∈Vjx\in V_j, and nonnegativity of BB makes VjV_j decreasing. Continuity and the uniform bound give Vj‾⊂{A≤CjB+δj},∥CjB+δj∥Q≤Cj∥B∥Q+δj⟶0.\overline{V_j}\subset\{A\leq C_jB+\delta_j\},\qquad \|C_jB+\delta_j\|_Q\leq C_j\|B\|_Q+\delta_j\longrightarrow0. Thus their intersection lies in A−1(0)A^{-1}(0) and contains xx. If the final assertion failed, the nonempty nested compact sets Vj‾∩(Q∖O)\overline{V_j}\cap(Q\setminus O) would have a common point, contradicting (11.11.3). ◻

In [APW26] (Proposition 5.10), put g=φ+hg=\varphi+h, let ℓg\ell_g be its strict local affine support at xx, and let ℓf\ell_f support ff. Shrink to a compact neighbourhood on which both support inequalities hold, and take A=g−ℓg,B=f−ℓf.A=g-\ell_g,\qquad B=f-\ell_f. The strict support gives A−1(0)={x}A^{-1}(0)=\{x\}. Thus the displayed intersection in the proof on p. 37 must use {g−ℓg≤0}\{g-\ell_g\leq 0\}, not {f−ℓf≤0}\{f-\ell_f\leq 0\}. The latter can be the whole neighbourhood when f=ℓff=\ell_f. The shrinking argument is valid with the corrected zero set. In addition, the maximum modification has the exact pointwise increment ψ−φ=max⁡{0,CB+δ−A},0≤ψ−φ≤C∥B∥Q+δ.\begin{equation} \psi-\varphi=\max\{0,CB+\delta-A\},\qquad 0\leq\psi-\varphi\leq C\|B\|_Q+\delta. \end{equation} Taking the right side at most ϵ\epsilon and the modification set compactly inside the support neighbourhood makes the function equal to φ\varphi near its boundary. Hence it glues to φ\varphi outside. On V={A<CB+δ}V=\{A<CB+\delta\} one obtains ψ+h=Cf+ℓ\psi+h=Cf+\ell, where ℓ=δ+ℓg−Cℓf\ell=\delta+\ell_g-C\ell_f. Preservation of the tropical psh class under this maximum and gluing is the separate cited input [APW26] (Lemma 4.13). We have repaired its localization calculation, not re-proved the entire tropical regularization theorem.

Remark 11.11.12 (The coefficient issue in local PL density). In [APW26] (Sections 2.1.1, 2.1.4–2.1.5), rationality of a polyhedron constrains the normals of its defining half-spaces, while their constants may be real; the affine functions used for PL functions are real affine. Proposition 5.21 is stated for this PL class. It does not state that its approximants have rational slopes and constants in the value group of an independently chosen analytic field. This extra conclusion cannot be inferred from the word “rational” attached to the polyhedral structure.

For example, the real convex PL function F(t)=max⁡(1−πt,0)F(t)=\max(1-\pi t,0) is allowed in that polyhedral category. Over a field whose fixed additive value group is Q\mathbf Q, its pullback max⁡(1+πlog⁡∣T∣,0)\max(1+\pi\log|T|,0) on Gman\mathbf G_m^{\mathrm{an}} is not R-PL in the strict GJR category: the bend at t=1/πt=1/\pi cannot occur in a finite strict polyhedral presentation. This is the example of [GJR] (Example 5.3), applied to the present comparison. It does not disprove the existence of better approximants. It shows that a coefficient-preserving approximation theorem, or an explicit replacement argument, is required before using APW local density as a Q-PL/model input.

There is nevertheless a simple replacement on a bounded unit chart when an ambient convex maximum is already available. This distinguishes a coefficient issue from the separate problem of extending a relatively convex function off its tropical image.

Proposition 11.11.13. (Coefficient-controlled approximation on a bounded chart). Let KK be complete and nontrivially valued, with Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|. Let S⊂RrS\subset\mathbf R^r be bounded and F(t)=max⁡1≤a≤m{ca+⟨va,t⟩},ca∈R,va∈Rr.F(t)=\max_{1\leq a\leq m}\{c_a+\langle v_a,t\rangle\}, \qquad c_a\in\mathbf R,\quad v_a\in\mathbf R^r. There are convex maxima HjH_j of affine functions with slopes in Qr\mathbf Q^r and constants in ΓQ\Gamma_{\mathbf Q} such that Hj∣SH_j|_S decreases uniformly to F∣SF|_S. Consequently, on a good strict analytic unit chart with bounded tropical image, the pullback of FF has a decreasing uniform approximation by GJR Q-PL functions having relative affine lower supports. No change of the absolute value or extension of KK is required.

Open proof obligation (RC2). Prove the analytic Q-PL assertion for the finite maxima. The coefficient approximation is in the withheld proof; the non-Stacks use is GJR 5.2(7).

NAG-900163 · Proof status

The monotonicity is asserted on SS, not on all of Rr\mathbf R^r. Every point of a unit chart has a neighbourhood with bounded tropical image. Thus the proposition resolves the local coefficient problem for finite ambient convex maxima, including the irrational-bend example above. It does not supply such a maximum for an arbitrary relatively convex tropical function; Example 11.10.9 prevents that shortcut on a fixed chart.

What remains in the source comparison

Under the stated relative interpretation, Wu’s convexity condition gives classical psh, and known Q-PL membership then gives nef residues. The unit-chart membership is proved above. With vanishing coordinates, the ambient connected-fibre proof of Lemma 11.10.1 does not apply directly. Section 11.12 supplies the missing strict G-local unit-log presentation using finite polyhedral strata, with explicit common-branch and coefficient hypotheses.

Proposition 11.22.14 proves the reverse original-smooth-to-PL comparison for a single input in the stated projective setting. The arbitrary-family comparison and first-stage decreasing closure remain open; Corollary 11.11.3 proves only classical closure. Full Wu Theorem A.3, CLD25 Proposition 8.1.11 and arbitrary-current mixed products are not established here. The primary text of mixed-calculus Theorem 6.13 remains unavailable and is not used.

11.12 Finite selections at the extended tropical boundary

The remaining PL identification in Section 11.11.6 can be settled for a precise relative finite-selection class. The missing ingredient is finiteness of the polyhedral strata of an analytic tropical image. It permits a function to forget the coordinates tending to infinity, even though the fibres of the projection need not be connected. We first prove that assertion as a statement about polyhedra, and then identify its analytic hypotheses.

A recession argument

A closed polyhedron is an intersection of finitely many closed affine half-spaces in a finite-dimensional real vector space; the empty set is allowed. For P={v:Av≤b}≠∅P=\{v:Av\leq b\}\ne\varnothing, its recession cone is rec⁡(P)={w:Aw≤0}\operatorname{rec}(P)=\{w:Aw\leq 0\}. Thus v+λw∈Pv+\lambda w\in P for every v∈Pv\in P, w∈rec⁡(P)w\in\operatorname{rec}(P) and λ≥0\lambda\geq 0.

Lemma 11.12.1. (Simultaneous escape). Let P⊂Rs×RdP\subset\mathbf R^s\times\mathbf R^d be a nonempty closed polyhedron, where s>0s>0, and suppose that its projection to the second factor is bounded. Exactly one of the following alternatives holds:

  1. There is (w,0)∈rec⁡(P)(w,0)\in\operatorname{rec}(P) with wi>0w_i>0 for every ii.

  2. There is M∈RM\in\mathbf R such that min⁡iti≤M\min_i t_i\leq M for every (t,y)∈P(t,y)\in P.

Proof. We use the elementary finite-dimensional Farkas alternative in the following form: if a linear functional cc is nonpositive on {w:Aw≤0}\{w:Aw\leq 0\}, then c=λtAc=\lambda^tA for some column vector λ≥0\lambda\geq 0. Here is a brief justification. The cone generated by the rows of AA is closed: in each nonnegative representation, a linear dependence can eliminate one positive coefficient, until its support is linearly independent. There are finitely many such supports; on a fixed independent support convergence of the represented vectors forces convergence of the coefficients. If cc were outside this closed convex cone, separation would give ww with Aw≤0<c(w)Aw\leq 0<c(w), a contradiction.

It follows that a linear functional unbounded above on a nonempty polyhedron Q={v:Av≤b}Q=\{v:Av\leq b\} is positive on some recession vector: otherwise c=λtAc=\lambda^tA gives c(v)≤λtbc(v)\leq\lambda^tb on QQ. Apply this observation to Q={(t,y,z):(t,y)∈P, z≤ti for all i}.Q=\{(t,y,z):(t,y)\in P,\ z\leq t_i\text{ for all }i\}. If alternative (2) fails, zz is unbounded above on QQ. Hence QQ has a recession vector (w,v,a)(w,v,a) with a>0a>0. The inequalities give wi≥aw_i\geq a. Boundedness of the yy-projection forces v=0v=0, proving (1). Conversely, a ray with every wi>0w_i>0 makes min⁡iti\min_i t_i unbounded. ◻

Theorem 11.12.2. (Factorization on a polyhedral tail). Let B⊂RdB\subset\mathbf R^d be bounded and let S⊂[R0,+∞]s×BS\subset[R_0,+\infty]^s\times B be compact, with s>0s>0. For each I⊂{1,…,s}I\subset\{1,\ldots,s\}, assume that the stratum SI={(t,y)∈S:ti=+∞ ⟺ i∈I}S_I=\{(t,y)\in S:t_i=+\infty\ \Longleftrightarrow\ i\in I\} is, after deleting the infinite coordinates, a finite union of closed polyhedra. Let F:S→RF:S\to\mathbf R be continuous and suppose that there are finitely many functions L1,…,Lm:B→RL_1,\ldots,L_m:B\to\mathbf R such that F(t,y)∈{L1(y),…,Lm(y)}F(t,y)\in\{L_1(y),\ldots,L_m(y)\} everywhere. Then there exists RR such that, for every (t,y)∈S(t,y)\in S with all ti≥Rt_i\geq R, (+∞,…,+∞,y)∈S,F(t,y)=F(+∞,…,+∞,y).\begin{equation} (+\infty,\ldots,+\infty,y)\in S, \qquad F(t,y)=F(+\infty,\ldots,+\infty,y). \end{equation} The threshold depends on SS and its finite polyhedral presentations, not on FF or on the functions LaL_a.

Proof. For a polyhedron PP in a stratum with at least one finite tt-coordinate, apply Lemma 11.12.1 to those coordinates. If its second alternative holds, choose MPM_P as there. Finitely many strata and polyhedra occur, so choose RR greater than all these bounds and at least R0R_0; if there are no such bounds, any R≥R0R\geq R_0 suffices.

A point (t,y)(t,y) in the stated tail either already has all ti=+∞t_i=+\infty, or lies in a polyhedron of the first type. In the latter case choose its recession vector (w,0)(w,0). The ray obtained by increasing the finite tt-coordinates by λw\lambda w, and leaving the others infinite, lies in SS for every λ≥0\lambda\geq 0. Along this connected ray, FF takes values in the fixed finite subset {La(y)}a\{L_a(y)\}_a of R\mathbf R. Continuity therefore makes FF constant on the ray. Its limit in the ambient extended space is (+∞,…,+∞,y)(+\infty,\ldots,+\infty,y); compactness of SS places that point in SS, and continuity gives (11.12.1). ◻

No connectedness of a whole projection fibre was used. The proof instead connects each point in a sufficiently deep tail to the same boundary point in its fibre. This distinction is important for relative images.

Example 11.12.3. (Why finite polyhedral structure is needed). In [1,+∞]×[0,1][1,+\infty]\times[0,1], let S={(1/y,y):0<y≤1}∪{(2/y,y):0<y≤1}∪{(+∞,0)}.S=\{(1/y,y):0<y\leq1\}\cup\{(2/y,y):0<y\leq1\}\cup\{(+\infty,0)\}. Set F=yF=y on the first curve, F=−yF=-y on the second, and F(+∞,0)=0F(+\infty,0)=0. Both compactified curves are continuous images of [0,1][0,1], so SS is compact. They meet only at their common endpoint, where both values tend to zero; hence FF is continuous and selects from the two affine functions y,−yy,-y. It also has relative affine lower supports: use the active branch away from the endpoint, and −y-y at the endpoint. For every RR, small positive yy gives two points in the tail with the same yy and opposite values. Thus FF does not factor through yy on any tail. The finite-coordinate stratum is not a finite union of polyhedra. This is an abstract counterexample to dropping that hypothesis, not an asserted analytic tropical image.

The analytic image input and zero strata

We use the following named input from Martin: for a compact strictly KK-analytic space over a complete nontrivially valued field, the positive-coordinate image of finitely many analytic functions is a finite polyhedral set. In logarithmic coordinates its normals are rational and its constants lie in ΓQ\Gamma_{\mathbf Q}, where Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|. See [Mar13] (Theorem 4.3.1, pp. 172–173; conventions in Section 4.1.1). The published strict version is [Mar15]. We checked the thesis proof; the published article was available only through its abstract. No broader non-strict or trivially valued formulation is used.

Lemma 11.12.4. (Stratified polyhedral analytic images). Let VV be strictly KK-affinoid, with KK complete and nontrivially valued. Let g1,…,gsg_1,\ldots,g_s be analytic functions and h1,…,hdh_1,\ldots,h_d analytic units on VV. The image S={((−log⁡∣gi(z)∣)i=1s,(−log⁡∣hj(z)∣)j=1d):z∈V}S=\{(( -\log|g_i(z)|)_{i=1}^s,(-\log|h_j(z)|)_{j=1}^d):z\in V\} satisfies the compactness, boundedness and finite-stratum hypotheses of Theorem 11.12.2, for some R0R_0 and bounded box BB.

Proof. Compactness follows from continuity and compactness of VV. The functions gig_i are bounded, so their negative logarithms have a common lower bound. Both hjh_j and hj−1h_j^{-1} are bounded, so the last coordinates lie in a bounded box. Fix I⊂{1,…,s}I\subset\{1,\ldots,s\} and take the closed analytic subspace ZI⊂VZ_I\subset V defined by gi=0g_i=0 for i∈Ii\in I. It is strictly affinoid. Apply Martin’s theorem on ZIZ_I to the remaining gig_i, together with all hjh_j, and take its positive-coordinate image. Positivity excludes exactly the additional zeros gi=0g_i=0 for i∉Ii\notin I. After taking negative logarithms this is precisely SIS_I with its infinite coordinates deleted, not its closure across other strata. Martin’s finite union of closed polyhedra gives the assertion; empty strata cause no difficulty. If no coordinates remain, the stratum is empty or a single point, and the conclusion is immediate. ◻

From relative finite selections to model PL functions

Specify the relative convention before using it. On a relative neighbourhood OO of a point in a tropical image, a rational finite selection is a continuous real-valued function selecting from one finite list ℓa(t)=ca+∑iqaiti,ca∈ΓQ,qai∈Q,a=1,…,m.\begin{equation} \ell_a(t)=c_a+\sum_i q_{ai}t_i, \qquad c_a\in\Gamma_{\mathbf Q},\quad q_{ai}\in\mathbf Q, \qquad a=1,\ldots,m. \end{equation} Every individual summand of every branch must be finite everywhere on OO, with 0⋅(+∞)=00\cdot(+\infty)=0. In particular, a branch cannot be defined only on the subset where it is selected. This is a substantive hypothesis.

Theorem 11.12.5. (Relative rational PL identification through zeros). Let XX be a pure-dimensional good strictly KK-analytic space, where KK is complete and nontrivially valued. Suppose a function uu is locally, on ordinary opens, the pullback under the negative logarithms of finitely many analytic functions of a relative rational finite selection in the convention (11.12.2). Then uu is Q-PL in the strict GJR sense. More precisely, near each point it has such a finite-selection presentation using only analytic units. No convexity assumption is required.

Proof. Fix xx. Shrink to an ordinary neighbourhood on whose tropical image one finite list works. Reorder the coordinates as g1,…,gsg_1,\ldots,g_s, vanishing at xx, and h1,…,hdh_1,\ldots,h_d, nonvanishing at xx. Shrink again so that the latter are units. If s=0s=0, Proposition 11.11.5 applies directly. Otherwise finiteness of each summand at the image of xx forces the coefficient of each gig_i to vanish in every branch. Thus all branches have the form La(y)L_a(y), where y=(−log⁡∣hj∣)jy=(-\log|h_j|)_j.

Choose a compact strict affinoid neighbourhood VV of xx inside this open, and put S=τ(V)S=\tau(V). Lemma 11.12.4 and Theorem 11.12.2 give a threshold RR. Enlarge it to an element of ΓQ\Gamma_{\mathbf Q}; this is possible since Γ\Gamma contains a positive element. Then VR={z∈V:∣gi(z)∣≤e−R for all i}V_R=\{z\in V:|g_i(z)|\leq e^{-R}\text{ for all }i\} is a strict affinoid domain containing an ordinary neighbourhood of xx. Indeed, choose N>0N>0 and a∈K×a\in K^\times with NR=−log⁡∣a∣NR=-\log|a|; its inequalities are ∣giN/a∣≤1|g_i^N/a|\leq 1. At xx they are strict, and xx lies in the ordinary interior of VV.

Define the boundary slice and its restricted function by S∞={y:(+∞,…,+∞,y)∈S},F∞(y)=F(+∞,…,+∞,y).\begin{gathered} S_\infty=\{y:(+\infty,\ldots,+\infty,y)\in S\},\\ F_\infty(y)=F(+\infty,\ldots,+\infty,y). \end{gathered} The function F∞F_\infty is continuous on the compact set S∞S_\infty. The tail theorem gives y(VR)⊂S∞y(V_R)\subset S_\infty and u∣VR=F∞∘yu|_{V_R}=F_\infty\circ y. It still selects from the same rational branches LaL_a. Restrict to the ordinary interior of VRV_R and apply Proposition 11.11.5. Varying xx proves the Q-PL assertion by sheafiness. ◻

This proof avoids subdividing inequalities between vanishing monomials into rational domains at their common zero, the invalid shortcut excluded by Example 11.11.9. Only inequalities against a nonzero constant are used to select VRV_R; the subsequent Laurent subdivision involves units. The argument combines Martin’s image theorem with an elementary recession calculation and the earlier unit-chart identification; no priority claim is made for that combination.

Corollary 11.12.6. (Relative generators in the fixed-domain transfer). Let XX be an integral projective variety of positive dimension over a complete nontrivially valued field KK. On one ordinary open U⊂XanU\subset X^{\mathrm{an}}, let uju_j be a decreasing sequence of functions satisfying Theorem 11.12.5, and assume each has local relative affine lower supports as in Theorem 11.11.2. The presentations and their charts may depend on jj. For every x∈Ux\in U, there is one ordinary neighbourhood V⊂UV\subset U and a decreasing sequence of original CLD smooth psh functions bjb_j on VV with bj↓u=inf⁡jujb_j\downarrow u=\inf_j u_j. If uu is finite continuous, this convergence is uniform on compact subsets of VV.

Proof. Theorem 11.12.5 gives Q-PL membership for each term on UU. Theorem 11.11.6 gives nef residues, using the classical psh criterion and its cited comparison theorem. Now apply Theorem 11.10.11. The varying neighbourhoods used to identify individual terms are not intersected: they establish a sheaf property on UU before the fixed projective barrier is chosen. Its termwise scaling constants remain free to vary. Dini’s theorem gives the final assertion. ◻

Remark 11.12.7 (One tube for a family of finite selections). The unit-reduction part of the proof does not require rational coefficients. On a fixed compact image SS as above, every continuous finite selection from branches depending only on yy factors on the same tail, since the threshold in Theorem 11.12.2 is independent of those branches. The finite lists may differ for different members of a family. The resulting boundary restriction preserves pointwise order, maxima and minima. This gives a uniform domain when the image and the vanishing-coordinate set are already fixed. It gives no uniform domain for an arbitrary family of different analytic charts, and real coefficients do not by themselves give strict GJR PL membership.

Example 11.12.8. (Rational constants and a fixed valuation). Equip K=QpK=\mathbf Q_p with the absolute value ∣a∣=exp⁡(−2 vp(a))|a|=\exp(-\sqrt 2\,v_p(a)). This complete nontrivially valued field has ΓQ=2 Q\Gamma_{\mathbf Q}=\sqrt 2\,\mathbf Q. The constant function 11 on the analytic point M(K)\mathcal M(K) is not Q-PL in the strict unit-log sense: such a presentation would give N=log⁡∣b∣∈2 ZN=\log|b|\in\sqrt 2\,\mathbf Z for some positive integer NN and b∈K×b\in K^\times, which is impossible. The same obstruction holds at a KK-rational point of an analytic space, by evaluating the unit-log presentation there. Nevertheless 11 is an original smooth pluriharmonic function. Thus this example separates coefficient categories; it is not a counterexample to smooth approximation or to psh comparison. No absolute value is changed during either argument.

Coefficient control for ambient convex functions

The remaining relative approximation problem should be distinguished from a coefficient problem that can be solved directly. Proposition 11.11.13 treats finite ambient maxima. Supporting hyperplanes extend that argument to every finite convex function on an ambient convex open set.

Proposition 11.12.9. (Decreasing rational approximation of ambient convex functions). Let Ω⊂Rd\Omega\subset\mathbf R^d be convex and open, let F:Ω→RF:\Omega\to\mathbf R be finite continuous and convex, and let S⊂ΩS\subset\Omega be nonempty compact. For every nonzero additive subgroup Γ⊂R\Gamma\subset\mathbf R, there are finite maxima HjH_j of affine functions with rational slopes and constants in ΓQ\Gamma_{\mathbf Q} whose restrictions to SS decrease uniformly to F∣SF|_S. Each HjH_j is defined and convex on all of Rd\mathbf R^d, but monotonicity is asserted only on SS.

Proof. At each x∈Ωx\in\Omega a supporting affine function ℓx\ell_x satisfies ℓx≤F\ell_x\leq F on Ω\Omega and ℓx(x)=F(x)\ell_x(x)=F(x). To recall the input, separate the open convex epigraph {(t,r):t∈Ω, r>F(t)}\{(t,r):t\in\Omega,\ r>F(t)\} from its boundary point (x,F(x))(x,F(x)). The separating functional has nonnegative last coefficient; that coefficient cannot be zero because xx is interior to Ω\Omega. Division by its positive last coefficient gives ℓx\ell_x.

Choose ρ∈Γ\rho\in\Gamma positive and ϵj=ρ8−j\epsilon_j=\rho 8^{-j}. For fixed jj, continuity and compactness give finitely many supports with maximum MjM_j satisfying F−ϵj≤Mj≤FF-\epsilon_j\leq M_j\leq F on SS: the relative open sets {F−ℓx<ϵj}∩S\{F-\ell_x<\epsilon_j\}\cap S, for x∈Sx\in S, cover SS. Since SS is bounded, approximate each selected slope by a rational vector and each constant by an element of ρQ\rho\mathbf Q so that the resulting affine function differs by at most ϵj\epsilon_j on SS. Let GjG_j be the maximum of these rational affine functions. Then F−2ϵj≤Gj≤F+ϵjon S.F-2\epsilon_j\leq G_j\leq F+\epsilon_j\quad\text{on }S. Set Hj=Gj+3ϵjH_j=G_j+3\epsilon_j. The shift is still in ΓQ\Gamma_{\mathbf Q}, and F+ϵj≤Hj≤F+4ϵj,Hj+1≤F+12ϵj≤Hjon S.F+\epsilon_j\leq H_j\leq F+4\epsilon_j, \qquad H_{j+1}\leq F+\tfrac12\epsilon_j\leq H_j\quad\text{on }S. This proves both assertions. It is not necessary to choose nested sets of supporting hyperplanes: the positive shifts enforce monotonicity on the prescribed compact set. ◻

Corollary 11.12.10. (A coefficient-compatible local analytic class). On a pure-dimensional good strictly analytic space over a complete nontrivially valued KK, suppose locally that u=F∘τu=F\circ\tau, where τ\tau is given by analytic units and FF is finite continuous convex on a convex ambient open containing the image. Near every point, uu is a uniform decreasing limit of finite rational maxima of unit logarithms, with constants in ΓQ\Gamma_{\mathbf Q}. These approximants are Q-PL and have relative affine lower supports. For an integral projective XX of positive dimension, Corollary 11.12.6 consequently gives a decreasing original smooth psh approximation on one neighbourhood.

Open proof obligation (RC2). Prove the Q-PL analytic presentation and stability under the maximum operation; the proof takes GJR 5.2(7) as input.

NAG-900170 · Proof status

The supporting hyperplanes in this proof lie below FF on the entire ambient convex open. Relative local supports on a polyhedral union need not have that property. Thus the proposition cannot be used to fill the remaining general relative-density step by replacing the ambient set with a tropical image.

A maximum statement that must retain PL membership

The coefficient distinction also corrects a literal closure assertion in the consulted GJR v3, also present in the published 2026 version, pp. 191–192. Remark 7.11 suggests that the maximum of two semipositive R-PL functions is again semipositive R-PL, and its proposed argument treats maxima of R-PL LFS functions as R-PL LFS. Example 5.3 of that same version prevents this without an additional membership hypothesis. The later [GRPL] (Corollary 5.5) explicitly restricts the unconditional statement to Λ=Z\Lambda=\mathbf Z or Q\mathbf Q.

Proposition 11.12.11. (Maximum with the correct coefficient scope). Let XX be a good strictly analytic space over a complete nontrivially valued field. If h1,h2h_1,h_2 are semipositive R-PL functions and h=max⁡(h1,h2)h=\max(h_1,h_2) is R-PL, then hh is semipositive. For Q-PL or integral PL inputs, membership of the maximum is automatic and the conclusion holds in the same coefficient class. For arbitrary R-PL inputs, that membership can fail even when both inputs are harmonic.

Open proof obligation (RC2). Prove the applicable positivity and closure implications for the specified maximum; the proof uses GRPSH, GJR and GRPL.

NAG-900171 · Proof status

The obstruction also occurs in the proper ambient setting of GJR Remark 7.11, which asks for semipositivity at a point. On P1,an\mathbf P^{1,\mathrm{an}} extend q=min⁡{0,max⁡{−1,log⁡∣T∣}}q=\min\{0,\max\{-1,\log|T|\}\} by its constant values near 00 and ∞\infty. This is integral PL when Γ=Q\Gamma=\mathbf Q. Then 1+πq1+\pi q and 00 are global R-PL functions, harmonic near the point of the standard skeleton with log⁡∣T∣=−1/π\log|T|=-1/\pi. Their maximum has the same irrational bend there and is not R-PL near that point. Thus properness of the ambient curve does not repair the omitted local membership hypothesis.

The other expectation in GJR Remark 7.11, concerning pointwise limits, has a different status. It is now supplied by [GRPL] (Theorem 5.6): a pointwise limit of semipositive R-PL functions is semipositive when the limit is itself R-PL. That final membership hypothesis must also be retained. It neither asserts that all such limits are PL nor proves first-stage original-smooth decreasing closure.

Gaussian regularization at a boundary stratum

The projective tropical localization used earlier rests on a Gaussian regularization of a finite convex conewise-linear function. The calculation at infinity in [APW26] (Proposition 5.7, pp. 34–35) can be made independent of the chosen approach sequence. We give the elementary calculation, including the growth hypothesis and the direction of monotonicity.

Proposition 11.12.12. (Gaussian boundary limit for a finite maximum). Let EE be a finite-dimensional Euclidean space, let v1,…,vsv_1,\ldots,v_s be linearly independent with s≥1s\geq 1, and put δ=∑iR≥0vi\delta=\sum_i\mathbf R_{\geq 0}v_i, T=span⁡(δ)T=\operatorname{span}(\delta) and Q=T⊥Q=T^\perp. Let p:E→Qp:E\to Q be orthogonal projection. Suppose h(x)=max⁡1≤a≤m⟨ma,x⟩≥0,h∣δ=0.h(x)=\max_{1\leq a\leq m}\langle m_a,x\rangle\geq0, \qquad h|_\delta=0. Set Aδ={a:⟨ma,vi⟩=0 for every i}A_\delta=\{a:\langle m_a,v_i\rangle=0\text{ for every }i\}, and hδ(z)=max⁡a∈Aδ⟨ma,z⟩h^\delta(z)=\max_{a\in A_\delta}\langle m_a,z\rangle for z∈Qz\in Q. This set of active indices is nonempty. If ρE\rho_E and ρQ\rho_Q are the standard normalized Gaussian densities and H=h∗ρEH=h*\rho_E, then lim⁡min⁡iti→+∞H(∑itivi+z)=∫Qhδ(z−b)ρQ(b) db.\begin{equation} \lim_{\min_i t_i\to+\infty} H\left(\sum_i t_i v_i+z\right) =\int_Q h^\delta(z-b)\rho_Q(b)\,db. \end{equation} The convergence is uniform for zz in compact subsets of QQ. No monotonicity of the approaching parameters is assumed. Moreover H(x+v)≤H(x)H(x+v)\leq H(x) for v∈δv\in\delta. If h(−v)>0h(-v)>0 for a nonzero v∈δv\in\delta, the inequality is strict.

Proof. Every ⟨ma,vi⟩\langle m_a,v_i\rangle is nonpositive. Since h(∑ivi)=0h(\sum_i v_i)=0, a maximizing index there has all these coefficients zero, proving that AδA_\delta is nonempty. For y=∑itiviy=\sum_i t_i v_i, every inactive linear form in h(y+z−w)h(y+z-w) tends to −∞-\infty as min⁡iti→∞\min_i t_i\to\infty; active forms annihilate TT. Consequently, for each fixed z,wz,w, h(y+z−w)⟶hδ(z−p(w)).h(y+z-w)\longrightarrow h^\delta(z-p(w)). Write L=max⁡a∥ma∥L=\max_a\|m_a\|. The function hh is LL-Lipschitz and subadditive. For ti≥0t_i\geq 0, its nonnegativity and h(y)=0h(y)=0 give the integrable bound 0≤h(y+z−w)≤h(z−w)≤L(∥z∥+∥w∥).0\leq h(y+z-w)\leq h(z-w)\leq L(\|z\|+\|w\|). Dominated convergence now gives the limit of the integral. Orthogonal decomposition E=T⊕QE=T\oplus Q factors the normalized Gaussian into the product of the two Gaussian densities; integration in TT gives exactly (11.12.3). Both sides, as functions of zz, have Lipschitz constant at most LL. A finite net on a compact subset of QQ upgrades pointwise convergence to uniform convergence. Equivalently, any failure of uniform convergence would give a sequence of parameters and a convergent sequence of points contradicting that bound and the already proved pointwise limit.

Finally, h(x+v)≤h(x)+h(v)=h(x)h(x+v)\leq h(x)+h(v)=h(x), and integration gives monotonicity. If h(−v)>0h(-v)>0, the difference h(x−w)−h(x+v−w)h(x-w)-h(x+v-w) is a nonnegative continuous function of ww and is positive at w=x+vw=x+v: its value there is h(−v)−h(0)>0h(-v)-h(0)>0. It is therefore positive on a nonempty open set of positive Gaussian measure. Its integral is strictly positive. This proves strictness without a separate strict-convexity argument. ◻

Corollary 11.12.13. (Strict affine contact across all faces). In Proposition 11.12.12, suppose in addition that h−1(0)=δh^{-1}(0)=\delta. The Gaussian convolution extends continuously to [0,+∞]s×Q[0,+\infty]^s\times Q by its quotient formulas on all infinity strata. At each point (+∞,…,+∞,z0)(+\infty,\ldots,+\infty,z_0) it admits a strict affine lower support depending only on zz. Here strictness means equality only at that point, including comparison with points on partial infinity strata.

Proof. For I⊂{1,…,s}I\subset\{1,\ldots,s\} put TI=span⁡{vi:i∈I}T_I=\operatorname{span}\{v_i:i\in I\} and QI=TI⊥Q_I=T_I^\perp, with orthogonal projection pIp_I. Retain only the linear forms annihilating TIT_I, and let their maximum on QIQ_I be hIh^I. The preceding proposition, applied to the face generated by the viv_i for i∈Ii\in I, gives the boundary value HI=hI∗ρQIH_I=h^I*\rho_{Q_I}, evaluated at pI(∑i∉Itivi+z)p_I(\sum_{i\notin I}t_i v_i+z). For I=∅I=\varnothing this is the original convolution.

We check the zero cone, since it controls strictness at partially infinite points. For I≠∅I\ne\varnothing, put vI=∑i∈Iviv_I=\sum_{i\in I}v_i. If hI(q)=0h^I(q)=0, all active linear forms at vIv_I have value at most zero on qq. Every inactive form is strictly negative on vIv_I, so for sufficiently large aa all forms have value at most zero on q+avIq+av_I. Since h≥0h\geq 0 and h−1(0)=δh^{-1}(0)=\delta, this gives q+avI∈δq+av_I\in\delta. Conversely, if q∈QI∩(δ+TI)q\in Q_I\cap(\delta+T_I), every active form is nonpositive at qq, and hI(q)=0h^I(q)=0 by nonnegativity. Thus (hI)−1(0)=pI(δ)=∑i∉IR≥0pI(vi).(h^I)^{-1}(0)=p_I(\delta) =\sum_{i\notin I}\mathbf R_{\geq0}p_I(v_i). The projected generators are linearly independent, because the original generators are. In particular this cone is pointed. The formula is immediate also for I=∅I=\varnothing. If I⊂JI\subset J, repeating the active-form construction on QIQ_I gives precisely hJh^J; orthogonal Gaussian pushforwards compose because pJpI=pJp_Jp_I=p_J. Proposition 11.12.12, with its locally uniform assertion, therefore proves continuity on approaching any higher stratum. There are only finitely many strata, so this proves continuity of the assembled extension.

Let II contain all indices, so QI=QQ_I=Q and (hI)−1(0)={0}(h^I)^{-1}(0)=\{0\}. Gaussian convolution of a Lipschitz function is smooth, by differentiation of the Gaussian under the integral. It is strictly convex here. Indeed, for each nonzero direction dd, the restriction hI(td)h^I(td) has different one-sided slopes: their difference is hI(d)+hI(−d)>0h^I(d)+h^I(-d)>0. A translate of any nontrivial segment in direction dd can cross this bend. The convexity gap is then positive on an open set of translates, and integration against the positive Gaussian makes the gap strictly positive. Hence the tangent affine function ℓ\ell to HIH_I at z0z_0 lies strictly below HI(z)H_I(z) for z≠z0z\ne z_0. If Q=0Q=0, take its unique constant affine function instead.

On a proper infinity stratum indexed by JJ, move in the nonzero direction ∑i∉JpJ(vi)\sum_{i\notin J}p_J(v_i). Its negative is outside the pointed zero cone, so Proposition 11.12.12 gives strict decrease after any positive step. Continuing toward the full infinity stratum yields HJ>HI(z)≥ℓ(z)H_J>H_I(z)\geq\ell(z). On the full infinity stratum equality occurs only at z=z0z=z_0. This proves the stated strict contact, including all partial strata. ◻

Remark 11.12.14 (What the corrected calculation supplies). In the application of APW Proposition 5.7, h=g−ℓh=g-\ell is a finite maximum of linear forms with zero set the strongly convex cone δ\delta. Thus h(−v)>0h(-v)>0 for every nonzero v∈δv\in\delta, and finite points lie strictly above the boundary limit: first move by a positive multiple of ∑ivi\sum_i v_i, then pass to infinity using monotonicity. The normalized quotient Gaussian in (11.12.3) is the pushforward of the ambient Gaussian; an arbitrary nonorthogonal coordinate splitting must instead retain that pushforward density.

On p. 35 the source initially allows an arbitrary sequence tending to infinity and then calls the translated functions decreasing. Arbitrary approach sequences need not be monotone; the dominated-convergence proof above removes that restriction. The integrand before taking the limit must be h(yi+z−w)h(y_i+z-w), with hδ(z−p(w))h^\delta(z-p(w)) appearing after passage to the limit. Later on the same page, for yi+1≥yiy_{i+1}\geq y_i, the printed strict inequality has the wrong direction: for a nonzero increment it is H(yi+1+z)<H(yi+z)H(y_{i+1}+z)<H(y_i+z). If two terms coincide, only equality is available. These corrections preserve the intended finite-point-to-boundary inequality. Corollary 11.12.13 supplies the same calculation on every partial infinity stratum, with its pointed zero cone verified explicitly. This proposition supplies the analytic limit calculation, not a new proof here of every global tropical-geometric input of APW Proposition 5.7.

Example 11.12.15. (Why Gaussian convolution needs a growth hypothesis). The finite-valued formulation of [APW26] (Lemma 5.8) needs an integrability qualification for arbitrary convex functions. On R\mathbf R, g(t)=exp⁡(t4)g(t)=\exp(t^4) is convex, since g′′(t)=(12t2+16t6)exp⁡(t4)≥0g''(t)=(12t^2+16t^6)\exp(t^4)\geq 0, but (g∗ρ)(x)=+∞(g*\rho)(x)=+\infty for every xx: the quartic term (x−w)4(x-w)^4 dominates w2/2w^2/2 in the exponent as ∣w∣→∞|w|\to\infty. Thus arbitrary finite convex input does not give finite Gaussian convolution. For the globally Lipschitz finite PL input actually used in Proposition 5.7, the bound in the preceding proof gives finiteness, and convolution preserves convexity by integrating its defining inequality. The local proof repair does not contradict that intended application.

Rational approximation on a fixed rational fan

There is a relative case beyond ambient convex extension in which coefficient control can also be proved. The essential restrictions are that the input is already piecewise linear and that its fixed fan is rational. The argument is a finite-dimensional rational-polyhedral calculation, not a general regularization theorem for continuous relatively convex functions.

Proposition 11.12.16. (Coefficient density for relatively convex fan functions). Let Σ\Sigma be a finite rational polyhedral fan in Rn\mathbf R^n, let S⊂∣Σ∣S\subset|\Sigma| be nonempty compact, and let F:∣Σ∣→RF:|\Sigma|\to\mathbf R be continuous, homogeneous and linear on each cone. Suppose that FF has a relative affine lower support near every point of ∣Σ∣|\Sigma|. For any nonzero additive subgroup Γ⊂R\Gamma\subset\mathbf R, there exist functions HjH_j on ∣Σ∣|\Sigma|, affine with rational slopes and constants in ΓQ\Gamma_{\mathbf Q} on the same cones, which have relative affine lower supports everywhere and decrease uniformly to FF on SS. The monotonicity is asserted only on SS. No ambient convex extension of FF is assumed.

Proof. Choose ambient covectors qσq_\sigma with F∣σ=qσ∣σF|_\sigma=q_\sigma|_\sigma. For every face τ\tau choose a point in its relative interior. A local affine lower support at that point has a slope aτa_\tau satisfying qσ∣σ∩σ′=qσ′∣σ∩σ′,aτ∣τ=qτ∣τ,(qσ−aτ)∣σ≥0(σ⊇τ).\begin{equation} \begin{aligned} q_\sigma|_{\sigma\cap\sigma'}&=q_{\sigma'}|_{\sigma\cap\sigma'},\\ a_\tau|_\tau&=q_\tau|_\tau,\\ (q_\sigma-a_\tau)|_\sigma&\geq0\qquad(\sigma\supseteq\tau). \end{aligned} \end{equation} Indeed, comparison in both tangent directions inside τ\tau identifies the slope on span⁡(τ)\operatorname{span}(\tau), and the support’s constant is then zero by homogeneity. For x∈relint⁡(τ)x\in\operatorname{relint}(\tau) and v∈σ⊇τv\in\sigma\supseteq\tau, the point x+tvx+tv lies in σ\sigma for t≥0t\geq 0. Applying the local support inequality for small tt gives the last line. Conversely, these constraints imply a relative lower support at every point of relint⁡(τ)\operatorname{relint}(\tau): cones not containing that point can be excluded by shrinking, and on every remaining cone the covector aτa_\tau is a lower support with contact.

Because the fan is rational and finite, the displayed conditions are finitely many homogeneous rational linear equalities and inequalities in the coordinates of all qσ,aτq_\sigma,a_\tau. For the inequalities it suffices to test a finite rational generating set of each cone, including generators of its lineality space with both signs if necessary. Thus the tuples form a rational polyhedral cone CC in a finite-dimensional vector space. Rational points are dense in CC. To see this at a particular tuple cc, impose as equalities the rational inequalities active at cc, together with the original equalities. Their nullspace has a rational basis. Approximate the coordinates of cc in that basis by rational numbers. The inactive inequalities remain strict for sufficiently close approximations, and the active equalities are preserved exactly.

Choose ρ∈Γ\rho\in\Gamma positive and set ϵj=ρ8−j\epsilon_j=\rho 8^{-j}. Approximate the tuple of FF inside CC so closely that its rational covectors define a continuous homogeneous function GjG_j with ∣Gj−F∣≤ϵj|G_j-F|\leq\epsilon_j on SS. This is possible because SS is bounded and there are finitely many cones. The constraints give relative affine lower supports for GjG_j. Set Hj=Gj+2ϵjH_j=G_j+2\epsilon_j. Then F+ϵj≤Hj≤F+3ϵj,Hj+1≤F+38ϵj≤Hjon S.F+\epsilon_j\leq H_j\leq F+3\epsilon_j, \qquad H_{j+1}\leq F+\tfrac38\epsilon_j\leq H_j\quad\text{on }S. The constants 2ϵj2\epsilon_j belong to ΓQ\Gamma_{\mathbf Q}, and shifting the supports by the same constants preserves contact. This proves the assertion. ◻

This argument combines a finite support criterion with rational-point density in a rational polyhedral cone; no priority claim is attached to the combination. On a unit chart whose image is contained in ∣Σ∣|\Sigma|, take SS to be the image of a compact strict affinoid neighbourhood. The pullbacks of HjH_j are Q-PL and have relative lower supports. Thus, in the projective setting of Corollary 11.12.6, they yield a decreasing original smooth psh approximation to F∘τF\circ\tau near the point. This does not require a same-chart ambient convex extension.

For a nonhomogeneous complex with vertices involving arbitrary elements of ΓQ\Gamma_{\mathbf Q}, the displayed system need not have rational coefficients, and its affine constants require separate control. The argument above does not treat that case. Nor does it show that a general continuous relatively convex function is approximable by functions linear on one fixed rational fan. These are distinct remaining questions.

Example 11.12.17. (A fixed irrational-length subdivision can obstruct density). Let a=2a=\sqrt 2, and let SS be the boundary of the rectangle with vertices A=(0,0)A=(0,0), B=(1,0)B=(1,0), C=(1,a)C=(1,a) and D=(0,a)D=(0,a), with its four-edge polyhedral complex. Take Γ=Z+2Z\Gamma=\mathbf Z+\sqrt 2\mathbf Z. All directions are rational and all vertex coordinates belong to ΓQ\Gamma_{\mathbf Q}. Define FF by the vertex values F(A)=0,F(B)=F(C)=F(D)=a,F(A)=0,\qquad F(B)=F(C)=F(D)=a, and extend linearly on each edge. This is a continuous relatively convex real PL function, but any continuous function GG affine with rational slopes on these same four edges satisfies ∥G−F∥C0(S)≥24.\begin{equation} \|G-F\|_{C^0(S)}\geq\frac{\sqrt2}{4}. \end{equation} The obstruction persists if the affine constants of GG are allowed to be arbitrary real numbers.

Proof. At an interior point of an edge, its affine expression is a relative support after shrinking. At a vertex the two adjacent edge directions are linearly independent, so there is an ambient affine function agreeing with FF on both incident edges. This supplies a relative support near that vertex. Thus FF has the stated relative convexity; no ambient convex extension is being asserted.

For GG let pb,pt∈Qp_b,p_t\in\mathbf Q be the derivatives in the positive horizontal direction on the bottom and top edges, and ql,qr∈Qq_l,q_r\in\mathbf Q the derivatives in the positive vertical direction on the left and right edges. Continuity around the rectangle gives pb+aqr−pt−aql=0.p_b+a q_r-p_t-a q_l=0. Since 11 and aa are linearly independent over Q\mathbf Q, necessarily pb=ptp_b=p_t and qr=qlq_r=q_l. In particular G(A)+G(C)−G(B)−G(D)=0.G(A)+G(C)-G(B)-G(D)=0. For FF the same expression equals −a-a. Comparing the four vertex values proves (11.12.5). This calculation uses only the slopes and continuity, so there is no restriction on the constants. ◻

In this example refinement does remove the obstruction. Choose rational r→ar\to a and rational q1,q2→1q_1,q_2\to 1 with q1<r/a<q2q_1<r/a<q_2. Put G=rxG=rx on the bottom edge, G=rG=r on the top and right edges, and on the left edge set G(0,y)=max⁡{q1y, q2y+r−q2a},0≤y≤a.G(0,y)=\max\{q_1y,\ q_2y+r-q_2a\},\qquad 0\leq y\leq a. The values at the four vertices match. The new breakpoint is (q2a−r)/(q2−q1)∈ΓQ(q_2a-r)/(q_2-q_1)\in\Gamma_{\mathbf Q}, and all slopes are rational, with constants in ΓQ\Gamma_{\mathbf Q}. The left-edge slopes increase from q1q_1 to q2q_2, so relative convexity holds there; at the vertices affine supports again exist by independence of the incident directions. The uniform error is at most max⁡{∣r−a∣,a∣q1−1∣,a∣q2−1∣}\max\{|r-a|,a|q_1-1|,a|q_2-1|\} and tends to zero. Choosing these errors geometrically small and adding the positive shifts used in Proposition 11.12.16 also makes the approximation decreasing. Thus the example isolates the fixed-subdivision obstruction rather than an obstruction to approximation itself.

Taking S×[0,1]d−1S\times[0,1]^{d-1} and pulling back FF gives the same obstruction in every positive dimension dd, for the product complex: restrict any proposed approximant to S×{0}S\times\{0\}. Relative supports pull back by projection. This is a counterexample in the category of finite polyhedral complexes with rational directions. The rectangle boundary is not asserted to be a balanced tropical variety or an analytic tropical image. The example does not rule out approximation after subdivision, changing charts, or using other analytic presentations, and does not refute the unrestricted regularization conclusion of APW. It explains why a proof that rationalizes coefficients while freezing an arbitrary ΓQ\Gamma_{\mathbf Q}-subdivision is insufficient. The distinction is sharpened in Section 11.19.4: the obstruction survives each fixed value-group refinement, but Example 11.19.9 gives an explicit decreasing approximation with a moving breakpoint.

Scope of the completed direction

The relative PL interface is now established with the stated common-branch and coefficient conventions. Wu’s Definition 3.1(1) uses rational constants, and his standing convention in Section 2.1 is Z⊆ΓK\mathbf Z\subseteq\Gamma_K when the valuation is nontrivial. Under that convention, these constants do lie in ΓQ\Gamma_{\mathbf Q}. Our formulation retains arbitrary fixed nontrivial absolute values by stating the coefficient condition explicitly; Example 11.12.8 explains why one must not omit it when leaving Wu’s normalization. It does not refute a claim under his standing convention. Definition 3.6 also uses images that need not be ambient open sets; our relative convention is explicit, rather than an attribution of an unstated convention to the source.

The direction obtained is from these relative rational convex generators to original smooth approximation on a fixed analytic neighbourhood. The reverse direction, coefficient-compatible approximation for general relatively convex functions, and the first-stage decreasing-closure problem remain separate. Neither full Wu Theorem A.3 nor CLD25 Proposition 8.1.11 follows from this section. No enlargement to an unrestricted positive-current Bedford–Taylor calculus, change of actual tests or support, or field-descent theorem is asserted.

11.13 Rational approximation on a fixed residue tube

The coefficient problem has a second, geometric form. A function already known to be R-PL has a finite presentation by integral PL functions. Their residue line bundles can be placed on one projective model. This permits rational perturbation after adding a small ample class, without rationalizing the vertices of a tropical subdivision. We carry out this argument with one neighbourhood for the whole finite-dimensional family.

Throughout this section KK is complete and nontrivially valued, Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|, and ΓQ=Γ⊗ZQ⊂R\Gamma_{\mathbf Q}=\Gamma\otimes_{\mathbf Z}\mathbf Q\subset\mathbf R. PL, Q-PL and R-PL have the strict analytic meaning of [GJR] (Section 5). In particular an arbitrary relative real PL presentation is not assumed to be an R-PL presentation in this sense. The geometric input below is the residue-model construction in [GJR] (Lemma 6.14 and proof of Proposition 7.10); the coefficient estimate and its ordered-family application are written out here.

A finite numerical estimate

Lemma 11.13.1. (Rational perturbation with an ample margin). Let VV be a projective variety over a field, AA an ample line bundle, and L1,…,LmL_1,\ldots,L_m line bundles. Choose positive integers NiN_i such that NiA+LiN_iA+L_i and NiA−LiN_iA-L_i are ample, using additive notation in Pic⁡(V)R\operatorname{Pic}(V)_{\mathbf R}. If L(a)=∑iaiLiL(a)=\sum_i a_iL_i is nef, ε>0\varepsilon>0, and ri∈Rr_i\in\mathbf R satisfy D:=∑i=1mNi∣ri−ai∣<ε,\begin{equation} D:=\sum_{i=1}^m N_i|r_i-a_i|<\varepsilon, \end{equation} then ∑iriLi+εA\sum_i r_iL_i+\varepsilon A is ample. In particular rir_i and ε\varepsilon may be chosen rational, with rr arbitrarily close to aa as ε\varepsilon tends to zero. No rational polyhedrality of the nef cone is required.

Open proof obligation (RC2). Prove the ample margin used in the construction; the proof uses the GJR Proposition 7.10 argument.

NAG-900175 · Proof status

Realizing a family before choosing its coefficients

Lemma 11.13.2. (One residue tube for a finite PL presentation). Let XX be good and strictly KK-analytic, let x∈Int⁡(X)x\in\operatorname{Int}(X), and let h1,…,hmh_1,\ldots,h_m be integral PL functions on a neighbourhood of xx. After shrinking once, there are a compact strictly analytic neighbourhood WW of xx, an ordinary open neighbourhood U⊂WU\subset W of xx, an integral PL function bb on WW, a projective variety V/K~V/\widetilde K, line bundles LiL_i on VV, and an ample line bundle AA on VV with the following properties.

  1. LiL_i and AA model the residues Lhi(x)L_{h_i}(x) and Lb(x)L_b(x). For every a∈Rma\in\mathbf R^m, ∑iaihi\sum_i a_ih_i is semipositive at xx if and only if ∑iaiLi\sum_i a_iL_i is nef on VV.

  2. If ∑iriLi+tA\sum_i r_iL_i+tA is semiample, then ∑irihi+tb\sum_i r_ih_i+tb is locally Fubini–Study on UU. If that class is nef, the function is semipositive on UU. Here ri,tr_i,t are real and locally Fubini–Study means semiampleness of the residue at every point.

All these data depend on the finite list (hi)(h_i) and xx, not on a,r,ta,r,t.

Open proof obligation (RC2). Prove the simultaneous neighbourhood/tube with the required residual class control; the proof applies the listed GJR inputs.

NAG-900176 · Proof status

A decreasing Q-PL approximation

Theorem 11.13.3. (Rational approximation of a semipositive R-PL germ). Let XX be good and strictly KK-analytic, x∈Int⁡(X)x\in\operatorname{Int}(X), and let hh be R-PL and semipositive at xx. There are one ordinary open neighbourhood UU of xx and Q-PL locally Fubini–Study functions qjq_j on UU which decrease uniformly to h∣Uh|_U. More precisely, for any fixed ρ∈Γ>0\rho\in\Gamma_{>0}, the construction can be made to satisfy h+ηj≤qj≤h+3ηjon U,ηj=ρ 8−j(j≥1).\begin{equation} h+\eta_j\leq q_j\leq h+3\eta_j\quad\hbox{on }U, \qquad \eta_j=\rho\,8^{-j}\quad(j\geq1). \end{equation} The approximants need not be harmonic when hh is harmonic. The claim is an explicit rational and ordered refinement of the construction of [GJR] (Proposition 7.10), not an assertion that that proposition is false.

Proof. Write h=∑iaihih=\sum_i a_ih_i on one neighbourhood, with hih_i integral PL, and apply Lemma 11.13.2. All hih_i and bb are continuous on the compact outer neighbourhood WW, so put Mi=∥hi∥WM_i=\|h_i\|_W and Mb=∥b∥WM_b=\|b\|_W. Choose the integers NiN_i in Lemma 11.13.1. For each jj choose εj∈Q>0\varepsilon_j\in\mathbf Q_{>0} and rji∈Qr_{ji}\in\mathbf Q such that εjMb<ηj2,∑iMi∣rji−ai∣<ηj2,∑iNi∣rji−ai∣<εj2.\begin{equation} \varepsilon_jM_b<\frac{\eta_j}{2},\qquad \sum_i M_i|r_{ji}-a_i|<\frac{\eta_j}{2},\qquad \sum_i N_i|r_{ji}-a_i|<\frac{\varepsilon_j}{2}. \end{equation} First choose the positive rational εj\varepsilon_j; then choose the rational vector sufficiently close to aa. Zero values of some MiM_i or MbM_b impose no restriction. Since L(a)L(a) is nef, the class ∑irjiLi+εjA\sum_i r_{ji}L_i+\varepsilon_jA is ample. Consequently pj=∑irjihi+εjbp_j=\sum_i r_{ji}h_i+\varepsilon_jb is Q-PL on WW and locally Fubini–Study on UU, while ∥pj−h∥W<ηj\|p_j-h\|_W<\eta_j. Set qj=pj+2ηjq_j=p_j+2\eta_j. The added constant is in ΓQ\Gamma_{\mathbf Q}, hence is Q-PL with zero residue: after clearing denominators it is a logarithm of an element of K×K^\times. Thus qjq_j is still locally Fubini–Study on UU. Its error band is (11.13.2), and qj+1≤h+3ηj+1=h+38ηj≤h+ηj≤qj.q_{j+1}\leq h+3\eta_{j+1} =h+\tfrac38\eta_j\leq h+\eta_j\leq q_j. The bound 3ηj→03\eta_j\to 0 proves uniform convergence on all of UU. No intersection of a sequence of shrinking neighbourhoods occurs. ◻

Corollary 11.13.4. (Ordered families in a fixed finite presentation). In the setting of Lemma 11.13.2, suppose uj=∑iajihiu_j=\sum_i a_{ji}h_i are semipositive at xx and decrease pointwise on their common original neighbourhood. Then on one ordinary open neighbourhood UU of xx there are Q-PL locally Fubini–Study functions qjq_j such that uj+ηj≤qj≤uj+3ηj,qj+1≤qj,inf⁡jqj=inf⁡juj.\begin{equation} u_j+\eta_j\leq q_j\leq u_j+3\eta_j, \qquad q_{j+1}\leq q_j, \qquad \inf_jq_j=\inf_ju_j. \end{equation} No bound on the coefficient vectors aja_j or on the family (uj)(u_j) is required. The common finite presentation is required.

Proof. Choose the data W,U,b,V,A,Li,NiW,U,b,V,A,L_i,N_i once for (hi)(h_i). In (11.13.3) replace aia_i by ajia_{ji} for the jjth term. The constants Mi,Mb,NiM_i,M_b,N_i remain fixed; the rational choices may vary. The same estimate gives the first band, and qj+1≤uj+1+3ηj+1≤uj+38ηj≤uj+ηj≤qj.q_{j+1}\leq u_{j+1}+3\eta_{j+1} \leq u_j+\tfrac38\eta_j\leq u_j+\eta_j\leq q_j. Since qj−uj→0q_j-u_j\to 0 uniformly, their decreasing limits agree, also when the limit takes the value −∞-\infty. ◻

Consequences for original smooth approximation

Corollary 11.13.5. (Semipositive R-PL functions in the projective case). Let XX be an integral projective variety of positive dimension over KK. If hh is R-PL near x∈Xanx\in X^{\mathrm{an}} and semipositive at xx, then on one ordinary open neighbourhood of xx there is a decreasing sequence of original CLD smooth psh functions converging to hh, uniformly on compact subsets. The same assertion holds for u=inf⁡juju=\inf_ju_j from Corollary 11.13.4, except that uniform convergence on compact subsets is asserted only when uu is finite continuous.

Proof. The preceding results give decreasing Q-PL functions on one ordinary open set, with nef residues at all its points. Apply Theorem 11.10.11 to that sequence. Its projective barrier and termwise rescaling supply the common smooth approximation domain. Dini’s theorem gives the asserted compact uniform convergence for a finite continuous endpoint. The residue rationalization itself used neither a scalar extension nor a change of the chosen absolute value. ◻

Theorem 11.13.6. (A comparison confined to strict R-PL germs). Under the projective hypotheses of Corollary 11.13.5, let hh be an R-PL germ at xx. The following three conditions are equivalent: hh is semipositive at xx; hh is classically psh on a neighbourhood of xx in the curvewise sense of [GRPSH] (Definition 4.1); on some fixed neighbourhood hh is a uniform decreasing limit of Q-PL LFS functions. They imply that, on a fixed neighbourhood, hh is a decreasing limit of original CLD smooth psh functions, uniformly on compact subsets, and hence that d′d′′[h]d'd''[h] is positive on the original CLD tests. Each neighbourhood here is ordinary open.

Open proof obligation (RC2). Prove both comparisons in the R-PL scope; the proof applies GRPL Theorem 5.6 and GRPSH Theorem 4.3.

NAG-900180 · Proof status

The last current-positivity conclusion is not used to reverse the argument. In GRPSH, “classically psh” requires subharmonic pullback to curves after every analytic field extension. Positivity on tests on XX alone is a different condition. Neither a pullback of an arbitrary current nor an identification of these two definitions is hidden in this proof.

Remark 11.13.7 (What this coefficient argument does not identify). For an already R-PL function, classical psh is equivalent to semipositivity by [GRPSH] (Theorem 4.3), so that condition may replace semipositivity on a neighbourhood in the preceding corollary. An arbitrary decreasing sequence of semipositive R-PL functions need not come with one finite presentation valid for all terms. Applying Theorem 11.13.3 separately then gives neighbourhoods depending on the term; Corollary 11.13.4 does not remove that problem without its additional hypothesis. Nor does the construction identify arbitrary relative real PL functions with GJR R-PL functions, or approximate every original smooth psh function by relative rational generators. Those are different interfaces.

The projective corollary concerns the original CLD smooth class and consequently its original tests; it does not replace those tests by weakly smooth or tropical tests. It introduces no current pushforward, altered support, or unverified descent. Full Wu Theorem A.3, CLD25 Proposition 8.1.11, and an unrestricted positive-current Bedford–Taylor calculus do not follow. The improvement over the preceding section is a proved coefficient conversion for already semipositive strict R-PL germs, including ordered families with a fixed finite presentation.

11.14 Ample extensions: a fixed-model obstruction and the cofinal repair

The approximation argument in Section 11.9 needs an ample model after passing to a suitable higher model. It does not need an ample extension on every prescribed projective model. These assertions are different. The published [GM] (Lemma 4.12, p. 353) states the latter assertion; the example below disproves that literal statement. We then recover exactly the cofinal formal-model input needed for the metric approximation. This does not disprove [GM] (Proposition 4.11) or [BE21] (Theorem 7.8).

The missing generic-fibre factor

In that lemma, f:Y→X1f:\mathscr Y\to\mathscr X_1 is a projective morphism of models inducing the identity on XX, and an ample L1\mathscr L_1 extends Lm1L^{m_1}. The proof chooses an arbitrary ff-ample bundle EE from a projective embedding and forms f∗L1m2⊗Ef^*\mathscr L_1^{m_2}\otimes E. Its generic restriction is (f∗L1m2⊗E)∣X≃Lm1m2⊗E∣X.\bigl(f^*\mathscr L_1^{m_2}\otimes E\bigr)|_X \simeq L^{m_1m_2}\otimes E|_X. The assumption f∣X=idXf|_X=\mathrm{id}_X does not trivialize E∣XE|_X: every line bundle is relatively ample for an identity morphism. The final identification in the published proof drops this factor. A missing justification alone would not refute the lemma. The following construction supplies the additional obstruction.

Example 11.14.1. (No ample extension on the prescribed model). Let R=k[[t]]R=k[[t]], K=k((t))K=k((t)), with ∣t∣=e−1|t|=e^{-1}, where kk is any field. There are projective algebraic RR-models Y,X1\mathscr Y,\mathscr X_1 of X=PK1×PK1X=\mathbf P^1_K\times\mathbf P^1_K and a morphism f:Y→X1f:\mathscr Y\to\mathscr X_1 inducing idX\mathrm{id}_X such that L=OX(1,1)L=\mathcal O_X(1,1) extends to an ample line bundle on X1\mathscr X_1, but no LmL^m, m>0m>0, extends to an ample line bundle on Y\mathscr Y. The same obstruction exists in every generic dimension n≥2n\geq 2.

Proof. On PR1\mathbf P^1_R, take the rank-two vector bundle FF defined by the extension class tξt\xi, where ξ\xi generates Ext⁡1(O(2),O)=H1(PR1,O(−2))≃R,0⟶O⟶F⟶O(2)⟶0.\operatorname{Ext}^1(\mathcal O(2),\mathcal O) =H^1(\mathbf P^1_R,\mathcal O(-2))\simeq R, \qquad 0\longrightarrow\mathcal O\longrightarrow F \longrightarrow\mathcal O(2)\longrightarrow0. The middle term is locally free since the quotient is locally free. On the generic fibre the extension is nonzero, hence, by rescaling the one-dimensional extension class and using the Euler sequence, FK≃O(1)⊕O(1)F_K\simeq\mathcal O(1)\oplus\mathcal O(1). On the special fibre the extension splits: Fk≃O⊕O(2)F_k\simeq\mathcal O\oplus\mathcal O(2). These are the classical deformation data relating the Hirzebruch surfaces F0\mathbb F_0 and F2\mathbb F_2; the argument uses them as an explicit model obstruction, not as a new deformation construction.

Use the quotient convention for Y=P(F)\mathscr Y=\mathbf P(F) and put M=OP(F)(1)M=\mathcal O_{\mathbf P(F)}(1). This is a smooth projective RR-scheme, hence regular and integral, with irreducible special fibre. Its generic fibre is XX, and MK=LM_K=L. The special quotient Fk↠OF_k\twoheadrightarrow\mathcal O defines a section C≃Pk1C\simeq\mathbf P^1_k on which MM has degree zero.

The vanishing H1(PR1,O)=0H^1(\mathbf P^1_R,\mathcal O)=0 shows that all three standard global sections of O(2)\mathcal O(2) lift to FF. Together with the section of the subbundle O\mathcal O, they generate FF: locally a section can first be matched in the quotient and then in the kernel. Thus MM is generated by four sections and defines g:Y→PR3g:\mathscr Y\to\mathbf P^3_R. Its generic restriction is the Segre embedding associated to O(1,1)\mathcal O(1,1). Let X1\mathscr X_1 be the schematic image. It is projective and flat over RR: its structure sheaf embeds into g∗OYg_*\mathcal O_{\mathscr Y}, which has no tt-torsion, and torsion-free modules over a DVR are flat. Its generic fibre is the Segre copy of XX. The induced f:Y→X1f:\mathscr Y\to\mathscr X_1 is projective (use the closed graph and a projective embedding of Y\mathscr Y over RR). With this identification of the generic fibres it induces the identity. The bundle L1=OPR3(1)∣X1\mathscr L_1=\mathcal O_{\mathbf P^3_R}(1)|_{\mathscr X_1} is ample and extends LL, and f∗L1=Mf^*\mathscr L_1=M.

Suppose that a line bundle NN on Y\mathscr Y extends LmL^m. Then N⊗M−mN\otimes M^{-m} is generically trivial. A generic trivialization gives a rational section whose divisor is vertical. Since Y\mathscr Y is regular and its special fibre is irreducible and reduced, this divisor is aYka\mathscr Y_k for an integer aa. But Yk=div⁡(t)\mathscr Y_k=\operatorname{div}(t) is principal. Thus N≃MmN\simeq M^m, and deg⁡(N∣C)=0\deg(N|_C)=0. An ample bundle has positive degree on every projective integral curve, so NN is not ample. This proves the claimed failure for every m>0m>0.

For n>2n>2, take the products of both models with PRn−2\mathbf P^{n-2}_R and replace L,M,L1L,M,\mathscr L_1 by their external tensor products with O(1)\mathcal O(1). Regularity and irreducibility of the special fibre persist. The same vertical-divisor argument forces any extension of LmL^m to be the indicated tensor power, whose degree on C×{p}C\times\{p\} is zero for a kk-rational point pp. The ample extension on the target is the external tensor product of two ample bundles. ◻

Proposition 11.14.2. (The obstruction over any nontrivially valued field). The conclusion of Example 11.14.1 holds over every complete nontrivially valued field KK, with its fixed absolute value, and in every dimension at least two. No discrete valuation, characteristic assumption or field extension is needed.

Proof. Put R=K∘R=K^\circ and choose 0≠a∈K∘∘0\ne a\in K^{\circ\circ}. On PR1\mathbf P^1_R, with homogeneous coordinates [x:y][x:y], define FF by 0⟶O(−1)2→ B O4⟶F⟶0,B=(−y0x−y0x0ay).0\longrightarrow\mathcal O(-1)^2 \xrightarrow{\ B\ }\mathcal O^4\longrightarrow F\longrightarrow0, \qquad B=\begin{pmatrix}-y&0\\x&-y\\0&x\\0&ay\end{pmatrix}. The minors include x2x^2 and y2y^2, so BB is locally a split injection and FF is locally free of rank two. On the special fibre its cokernel is O(2)⊕O\mathcal O(2)\oplus\mathcal O, as seen from the quotient map (x2,xy,y2,0)(x^2,xy,y^2,0) and the fourth basis vector. On the generic fibre, adding a−1a^{-1} times the fourth row to the second row makes the matrix block diagonal, with columns (−y,x)(-y,x) and (x,ay)(x,ay); each block has cokernel O(1)\mathcal O(1). Thus FK=O(1)2F_K=\mathcal O(1)^2. This is an explicit form of the preceding classical deformation; see also [Addington21] (Worksheet 2, part 3).

Set Y=P(F)\mathscr Y=\mathbf P(F) and M=OP(F)(1)M=\mathcal O_{\mathbf P(F)}(1). Its four generating sections give a morphism to the quadric X1={z0z2−z12+az1z3=0}⊂PR3.\mathscr X_1=\{z_0z_2-z_1^2+a z_1z_3=0\}\subset\mathbf P^3_R. Indeed the incidence relations are xz1−yz0=0xz_1-yz_0=0 and xz2−yz1+ayz3=0xz_2-yz_1+ayz_3=0, whose determinant is the displayed equation. The homogeneous coordinate ring is free of rank two over R[z0,z2,z3]R[z_0,z_2,z_3], because the equation is monic in z1z_1. Hence X1\mathscr X_1 is flat and of finite presentation over RR. The generic quadric is split and smooth: replacing z3z_3 by az3−z1az_3-z_1 gives the usual Segre equation. The generic morphism is the complete O(1,1)\mathcal O(1,1) embedding, so, after identifying generic fibres, f:Y→X1f:\mathscr Y\to\mathscr X_1 is a projective model morphism inducing the identity. The target has the ample extension O(1)\mathcal O(1).

To replace the DVR divisor argument, use the projective-bundle Picard formula. Over a connected base SS, this formula says that a line bundle on P(E)\mathbf P(E) is uniquely π∗N⊗O(d)\pi^*N\otimes\mathcal O(d). Here is the needed general-base justification. At any s∈Ss\in S, twist by the degree dsd_s on the projective-space fibre. Its fibre cohomology is then one-dimensional in degree zero and zero otherwise. Cohomology and base change, in the form [StacksPB] (Tag 0B9S), gives locally a line bundle N=π∗L(−ds)N=\pi_*\mathcal L(-d_s); the evaluation map is an isomorphism on every fibre and hence an isomorphism after shrinking. Thus dsd_s is locally constant, and connectedness makes it constant. The local NN glue by the evaluation isomorphisms, and uniqueness follows from π∗O=OS\pi_*\mathcal O=\mathcal O_S. This argument applies without noetherian assumptions.

Since RR is local, Pic⁡(R)=0\operatorname{Pic}(R)=0. Applying the formula twice gives Pic⁡(Y)=Z[M]⊕Z[π∗O(1)]\operatorname{Pic}(\mathscr Y)=\mathbf Z[M]\oplus\mathbf Z[\pi^*\mathcal O(1)]. On X=PK1×PK1X=\mathbf P^1_K\times\mathbf P^1_K these generators restrict to (1,1)(1,1) and (1,0)(1,0), a basis of Pic⁡(X)\operatorname{Pic}(X). Therefore restriction is injective, and any extension of LmL^m must be MmM^m. Its degree on the section defined by FK~↠OF_{\widetilde K}\twoheadrightarrow\mathcal O is zero, so it cannot be ample. Products with projective space give every higher dimension, with injectivity of Picard restriction again supplied by the same formula. No renormalization of ∣⋅∣|\cdot| has been made. ◻

In this example the failure has a precise geometric cause. Every generically trivial line bundle on Y\mathscr Y is trivial, whereas an ff-ample bundle must have positive degree on the contracted curve CC. Hence no ff-ample bundle can be generically trivial. Even though f∗L1=Mf^*\mathscr L_1=M is semiample, no perturbation within the same generic line bundle on this fixed model can make it ample. Passing to a higher model in the approximation theorem is therefore essential.

The sufficient vertical hypothesis

Lemma 11.14.3. (Ample extension along a vertical polarization). Let f:Y→X1f:\mathscr Y\to\mathscr X_1 be as above. If there is an ff-ample line bundle EE with E∣X≃OXE|_X\simeq\mathcal O_X, then some positive power of LL extends to an ample bundle on Y\mathscr Y. This hypothesis holds when ff is the blowup of a coherent vertical ideal.

Proof. For sufficiently large qq, f∗L1q⊗Ef^*\mathscr L_1^q\otimes E is ample by the relative ample twisting theorem, and its generic restriction is now Lm1qL^{m_1q}. For a vertical blowup use the relatively ample tautological bundle E=O(1)E=\mathcal O(1) of the Rees Proj. The ideal becomes the unit ideal on XX, so this bundle is generically trivial. This is the original proof with its missing hypothesis supplied. ◻

Proposition 11.14.4. (Cofinal ample models for metric approximation). Let KK be complete and nontrivially valued, X/KX/K projective, LL ample, and X0\mathfrak X_0 a formal model of XanX^{\mathrm{an}}. There is a projective algebraic K∘K^\circ-model Y\mathscr Y whose completion dominates X0\mathfrak X_0 and which carries an ample extension of some LmL^m, m>0m>0. No prescribed algebraic model is required to remain fixed.

Open proof obligation (RC2). Prove the cofinal construction and its ample-model assertion; the proof applies the construction in GM Lemma 2.4.

NAG-900183 · Proof status

Corollary 11.14.5. (Finitely many ample generic polarizations). In the setting of Proposition 11.14.4, let L1,…,LrL_1,\ldots,L_r be ample line bundles on XX. One can choose a single projective algebraic model whose completion dominates X0\mathfrak X_0 and on which every LiL_i has an ample Q-model.

Open proof obligation (RC2). Prove the relative ampleness step for the simultaneous models; the proof uses the GM Proposition 4.13 argument.

NAG-900184 · Proof status

Remark 11.14.6 (Effect on the approximation argument). For a nef model metric of an ample LL, pull its formal model bundle to the completion in Proposition 11.14.4; nefness persists under pullback. Proper formal GAGA algebraizes the pulled-back formal line bundle on Y\mathscr Y, preserving its generic identification. Divide the ample extension by its exponent to obtain an ample Q-model of the same LL. The rational convex combinations of these two model bundles are ample and their metrics converge uniformly to the original metric. Sufficient powers are generated by global sections, giving the pure Fubini–Study approximation used in Section 11.9. Thus the fixed-model counterexample removes an unnecessary stronger assertion without removing this metric approximation route. Corollary 11.14.5 also supplies the finite-polarization geometric input used in [GM] (Proposition 4.14) and [BGM22] (arXiv v1, Lemma 1.1). The generic Picard relation and the small model-function bound require a further argument, supplied in Proposition 11.14.8 below with explicit formal-domination scope. The reduced-special-fibre input of BGM Lemma 1.1 is separate. No scalar extension or change of valuation is involved.

Preserving the generic Picard relation

There is one further distinction in the proof of [GM] (Proposition 4.14). A numerical equality on the generic fibre does not itself trivialize a line bundle. To obtain a model function from the difference of two model bundles, one needs equality in Pic⁡(X)Q\operatorname{Pic}(X)_{\mathbf Q}. The ample generic class permits the following choice, which supplies that stronger relation before rational approximation.

Lemma 11.14.7. (Exact ample decomposition and rational relations). Let XX be projective over a field and L1,…,LmL_1,\ldots,L_m line bundles. Suppose Θ=∑iλi[Li]∈Pic⁡(X)R\Theta=\sum_i\lambda_i[L_i]\in\operatorname{Pic}(X)_{\mathbf R} has ample numerical class. There are ample line bundles H1,…,HrH_1,\ldots,H_r and real numbers μj>0\mu_j>0 such that ∑iλi[Li]=∑jμj[Hj]in Pic⁡(X)R.\sum_i\lambda_i[L_i]=\sum_j\mu_j[H_j] \quad\text{in }\operatorname{Pic}(X)_{\mathbf R}. Moreover (λ,μ)(\lambda,\mu) can be approximated arbitrarily closely by rational vectors (λ′,μ′)(\lambda',\mu') with all μj′>0\mu'_j>0 and the same exact relation in Pic⁡(X)Q\operatorname{Pic}(X)_{\mathbf Q}.

Proof. The inverse image of the numerical ample cone under Rm→N1(X)R\mathbf R^m\to N^1(X)_{\mathbf R} is open and contains λ\lambda. Choose a rational box containing λ\lambda in its interior and contained in this inverse image. Every vertex vv gives an ample Q-line bundle ∑ivi[Li]\sum_i v_i[L_i]; multiplying by a positive integer dvd_v gives an ample integral line bundle HvH_v. Express λ\lambda as the convex combination of the vertices with its positive multilinear barycentric coefficients tvt_v. Then the required equality in the Picard group, not merely in the numerical quotient, holds with μv=tv/dv\mu_v=t_v/d_v. This is an explicit use of openness of the ample cone.

For the second assertion, map Qm+r\mathbf Q^{m+r} to the finite-dimensional Q-subspace of Pic⁡(X)Q\operatorname{Pic}(X)_{\mathbf Q} spanned by these bundles, sending (a,b)(a,b) to ∑iai[Li]−∑jbj[Hj]\sum_i a_i[L_i]-\sum_j b_j[H_j]. Its kernel has a rational basis, and its realification contains (λ,μ)(\lambda,\mu). Rational linear combinations of that basis are dense. Choose one sufficiently close to preserve all strict inequalities μj′>0\mu'_j>0. Equality in Pic⁡(X)Q\operatorname{Pic}(X)_{\mathbf Q} means that, after clearing denominators and killing a possible torsion class, the corresponding integral bundles are isomorphic. This is exactly what is needed for a generic trivialization. ◻

Proposition 11.14.8. (Ample correction of a real model class). Let KK be complete and nontrivially valued and X/KX/K projective. Let θ\theta be a model (1,1)(1,1)-class in the sense of [GM] (4.4–4.6), represented by a finite real combination of model line bundles, with ample generic numerical class. Given a formal model carrying those data, there is a projective algebraic model whose completion dominates it and a rational model function φ\varphi such that θ+ddcφ\theta+dd^c\varphi is represented by an ample real class on that model. If θ\theta is semipositive, then for every ϵ>0\epsilon>0 one can arrange −ϵ≤φ≤0-\epsilon\leq\varphi\leq 0 on XanX^{\mathrm{an}}. This statement concerns GM model classes; it is not a new local theorem about arbitrary CLD currents.

Open proof obligation (RC2). Prove the model-class correction and uniform comparison with the stated positivity prerequisites; the proof uses GM 4.5 and Proposition 4.14.

NAG-900186 · Proof status

Section 11.15.1 now supplies the additional morphism-algebraization theorem and strengthens these statements to actual algebraic domination. Section 11.15.2 verifies the separate reduced-fibre input in the published BGM Lemma 1.1.

11.15 Algebraic domination and reduced models

The cofinal constructions of Section 11.14 were formulated for formal completions. To use them in algebraic intersection theory, one must recover actual morphisms of algebraic models. This passage is available over a complete rank-one valuation ring, but it requires the non-Noetherian form of formal GAGA. We give the hypotheses explicitly, then apply the result to the simultaneous ample-model construction. A second, independent step produces a reduced special fibre over an algebraically closed field. Its role is to turn an analytic supnorm into an algebraic lattice norm.

Algebraizing the morphism, not merely the line bundle

Throughout this subsection, KK is complete and nontrivially valued, with its given real-valued absolute value, and R=K∘R=K^\circ. Fix π∈R\pi\in R with 0<∣π∣<10<|\pi|<1. Completion always means π\pi-adic completion. We work with proper finitely presented RR-schemes. No discrete-valuation assumption is made.

Proposition 11.15.1. (Formal GAGA for model morphisms). For proper finitely presented RR-schemes Y,Z\mathscr Y,\mathscr Z, completion induces a bijection Hom⁡R(Y,Z)→ ∼ Hom⁡Spf⁡R(Y^,Z^).\operatorname{Hom}_R(\mathscr Y,\mathscr Z) \xrightarrow{\ \sim\ } \operatorname{Hom}_{\operatorname{Spf}R} (\widehat{\mathscr Y},\widehat{\mathscr Z}). In particular, if they are models of the same proper KK-scheme XX, a formal morphism inducing idXan\mathrm{id}_{X^{\mathrm{an}}} algebraizes uniquely to a morphism inducing idX\mathrm{id}_X.

Open proof obligation (RC2). Prove formal existence and faithfulness of proper analytification in the stated formal model category; the proof uses the cited FK17 corollaries/proposition.

NAG-900187 · Proof status

The distinction between sheaves and morphisms matters here. Formal GAGA for coherent sheaves, [FK17] (I, Theorem 10.1.2), algebraizes a formal line bundle on an already algebraized proper model. Proposition 11.15.1 also algebraizes the map to the prescribed model. In the source, the latter result is obtained by algebraizing the ideal of the graph. This is the additional step needed for domination in the algebraic directed system. All FK locators in this section refer to the fixed arXiv v5 of 28 February 2017.

Corollary 11.15.2. (Simultaneous algebraic domination). Let X/KX/K be projective, let L1,…,LrL_1,\ldots,L_r be ample, and prescribe finitely many proper finitely presented algebraic RR-models Xa\mathscr X_a of XX. There is a projective model Y\mathscr Y with morphisms Y→Xa\mathscr Y\to\mathscr X_a extending idX\mathrm{id}_X, on which every LiL_i has an ample Q-model. All prescribed commutative diagrams among the Xa\mathscr X_a remain commutative after pullback.

Proof. Choose a common formal refinement of the finitely many completions. Corollary 11.14.5 supplies Y\mathscr Y and the ample Q-models, with a formal map from Y^\widehat{\mathscr Y} to each X^a\widehat{\mathscr X}_a. Proposition 11.15.1 algebraizes each map and identifies its generic restriction. Equality of the two composites in any diagram follows from faithfulness of completion. Pullback of model line bundles along these algebraic maps has the prescribed formal pullback, so the metric identifications used in Section 11.14 are unchanged. ◻

Corollary 11.15.3. (The corrected algebraic model-class perturbation). Proposition 11.14.8 holds with algebraic domination of any prescribed proper finitely presented model. In particular, its rational model function can be represented on one algebraic model carrying the pulled-back initial data. If the initial class is semipositive, the bound −ϵ≤φ≤0-\epsilon\leq\varphi\leq 0 holds on XanX^{\mathrm{an}} with the original absolute value.

Proof. Include the prescribed model and all finitely many representatives of the initial class in Corollary 11.15.2. The exact generic Picard relation and the rational perturbation are those of Lemma 11.14.7 and Proposition 11.14.8. The new algebraic maps realize exactly the formal pullbacks used there. Hence their numerical classes, model functions and uniform bounds agree. No additional approximation or limiting argument is required. ◻

These corollaries reconstruct the conclusions needed from GM Propositions 4.11 and 4.14, using a different cofinal model. They do not restore the false fixed-model assertion of GM Lemma 4.12. The counterexample in Section 11.14.1 continues to apply.

The reduced-special-fibre step

For the next theorem assume additionally that KK is algebraically closed and that XX is reduced. The closure used below is the integral closure of a model in its given generic fibre; it is not a replacement of XX by its normalization. This distinction preserves the model’s generic identification even when XX is singular.

Theorem 11.15.4. (Reduced cofinal models with finitely many polarizations). Let KK be complete, algebraically closed and nontrivially valued. Let X/KX/K be reduced and projective, and L1,…,LrL_1,\ldots,L_r ample. Given finitely many algebraic models as in Corollary 11.15.2, there is a projective model Z\mathscr Z dominating all of them, with reduced special fibre and ample Q-models of all LiL_i.

Open proof obligation (RC2). Prove finiteness of the normalization and the reduced-special-fibre model step from the stated hypotheses; the argument invokes BE21 and BGM22.

NAG-900190 · Proof status

Remark 11.15.5 (The field hypothesis in the reduced fibre theorem). BE Theorem 4.20 also gives finiteness of the integral closure when KK is discretely valued. It does not assert a reduced special fibre in that case. For an arbitrary densely valued field, even finiteness can fail; [BE21] (Remark 4.21) explains the obstruction from values of spectral norms outside ∣K×∣|K^\times|. Thus algebraic domination, proved above over every complete nontrivially valued KK, and reduced cofinality, asserted here over algebraically closed KK, have different scopes.

Exact norms and the error that disappears in volume

Let L\mathscr L be a line bundle on a projective flat RR-model X\mathscr X, with generic fibre LL, and put Λ=H0(X,L)⊂H0(X,L)\Lambda=H^0(\mathscr X,\mathscr L)\subset H^0(X,L). The lattice is finitely generated and spans the generic space. Its norm and the supnorm of the induced metric are ∥s∥Λ=inf⁡{∣a∣:a∈K×, s∈aΛ},∥s∥sup⁡=sup⁡x∈Xan∣s(x)∣L.\|s\|_\Lambda=\inf\{|a|:a\in K^\times,\ s\in a\Lambda\},\qquad \|s\|_{\sup}=\sup_{x\in X^{\mathrm{an}}}|s(x)|_{\mathscr L}. By [BE21] (Lemma 6.3), ∥s∥sup⁡≤∥s∥Λ\|s\|_{\sup}\leq\|s\|_\Lambda, with equality when the special fibre is reduced. The equality is also the input of [BGM22] (Lemma 1.3). The next calculation shows why replacing the hypothesis by normality of the model would not justify the same conclusion.

Example 11.15.6. (A normal ramified model with unequal norms). Let KK be complete discretely valued, π\pi a uniformizer, and let e≥2e\geq 2 be prime to the residue characteristic. Put E=K(α)E=K(\alpha), where αe=π\alpha^e=\pi. The polynomial is Eisenstein, E/KE/K is separable and totally ramified, and RE=R[α]R_E=R[\alpha] is the valuation ring of EE. For any n≥0n\geq 0 take X=PREn,X=PEnas an R-model and a K-scheme,L=O(1).\mathscr X=\mathbf P^n_{R_E},\quad X=\mathbf P^n_E \quad\hbox{as an \(R\)-model and a \(K\)-scheme},\qquad \mathscr L=\mathcal O(1). Then X\mathscr X is normal, projective and flat over RR, and XX is smooth over KK. Its special fibre over RR is not reduced. For m≥1m\geq 1 and s=∑∣β∣=m∑i=0e−1biβαiTβ,biβ∈K,s=\sum_{|\beta|=m}\sum_{i=0}^{e-1}b_{i\beta}\alpha^iT^\beta, \qquad b_{i\beta}\in K, the two norms are ∥s∥Λm=max⁡i,β∣biβ∣,∥s∥mϕ=max⁡i,β∣biβ∣ ∣π∣i/e,Λm=H0(X,O(m)).\begin{equation} \|s\|_{\Lambda_m}=\max_{i,\beta}|b_{i\beta}|, \qquad \|s\|_{m\phi}=\max_{i,\beta}|b_{i\beta}|\,|\pi|^{i/e}, \qquad \Lambda_m=H^0(\mathscr X,\mathcal O(m)). \end{equation} Consequently the norms have the same unit ball but are unequal, and ∣π∣(e−1)/e∥s∥Λm≤∥s∥mϕ≤∥s∥Λm.|\pi|^{(e-1)/e}\|s\|_{\Lambda_m} \leq\|s\|_{m\phi}\leq\|s\|_{\Lambda_m}.

Proof. The special fibre is Pk[α]/(αe)n\mathbf P^n_{k[\alpha]/(\alpha^e)}, whereas X\mathscr X is a projective space over the DVR RER_E and hence is normal. The monomials αiTβ\alpha^iT^\beta form an RR-basis of Λm\Lambda_m, proving the first formula. Write a coefficient as cβ=∑ibiβαic_\beta=\sum_i b_{i\beta}\alpha^i. Distinct nonzero summands have valuations in different cosets of Z\mathbf Z in e−1Ze^{-1}\mathbf Z, when v(π)=1v(\pi)=1. Thus there is a unique term of least valuation and no cancellation at that valuation: ∣cβ∣=max⁡i∣biβ∣ ∣π∣i/e.|c_\beta|=\max_i |b_{i\beta}|\,|\pi|^{i/e}. For the standard projective model metric, the supnorm of a homogeneous polynomial equals the largest absolute value of its coefficients. The ultrametric inequality gives the upper bound, and the Gauss point in the chart T0≠0T_0\ne 0 gives equality. This proves the second formula. Finally, v(biβ)+i/e≥0v(b_{i\beta})+i/e\geq 0 holds for every i,βi,\beta if and only if every integer v(biβ)v(b_{i\beta}) is nonnegative. Hence the unit balls agree. Taking s=αT0ms=\alpha T_0^m gives supnorm ∣π∣1/e|\pi|^{1/e} and lattice norm 11, so the norms differ in every dimension. ◻

The discrepancy can be computed on the determinant line as well. Put Mm=(m+nn)M_m=\binom{m+n}{n} and λ=−log⁡∣π∣\lambda=-\log|\pi|. The eMmeM_m basis elements in (11.15.1) are orthogonal for both norms. The determinant norm of an orthogonal basis is the product of its norms: the ultrametric determinant inequality gives the lower bound for any decomposition of the same determinant vector, and the chosen basis attains it. Consequently, log⁡det⁡∥⋅∥Λmdet⁡∥⋅∥mϕ=Mmλ∑i=0e−1ie=e−12Mmλ.\log\frac{\det\|\cdot\|_{\Lambda_m}}{\det\|\cdot\|_{m\phi}} =M_m\lambda\sum_{i=0}^{e-1}\frac{i}{e} =\frac{e-1}{2}M_m\lambda. This grows like mnm^n, but multiplication by n!/mn+1n!/m^{n+1} makes it asymptotic to (e−1)λ/(2m)(e-1)\lambda/(2m). Thus a persistent, nonzero norm discrepancy can disappear in the volume normalization. The pointwise comparison constant ∣π∣−(e−1)/e|\pi|^{-(e-1)/e} is optimal, as the section αe−1T0m\alpha^{e-1}T_0^m shows. These exact calculations illustrate the two different scales in Proposition 11.15.8.

The absolute value on EE in this example is the extension of the one on KK: ∣α∣=∣π∣1/e|\alpha|=|\pi|^{1/e}. Renormalizing the valuation of EE so that α\alpha has value 11 would multiply its restriction to KK by ee and would change the problem. Also, equality of unit balls determines a norm when ∣K×∣|K^\times| is dense, by arbitrarily close scalar rescaling; it need not do so over a discretely valued field. This is why a lattice description by itself does not establish an isometry.

This failure of isometry does not refute the effectivity criterion for vertical Cartier divisors on normal models. GM Remark 4.10 gives that criterion without a reduced-special-fibre assumption. In the explicit model above, a vertical divisor k div⁡(α)k\,\operatorname{div}(\alpha) is effective exactly when k≥0k\geq 0, and its model function is −klog⁡∣α∣-k\log|\alpha|, with the same sign. Effectivity detects the unit-ball inequality; an isometry determines every radius.

Example 11.15.7. (Strict norm inequality after wild scalar extension). Let KK be a finite extension of Qp\mathbf Q_p containing a primitive ppth root of unity ζ\zeta, let π\pi be a uniformizer, and put E=K(α)E=K(\alpha) with αp=π\alpha^p=\pi. Then E/KE/K is cyclic of degree pp. For X=Spec⁡EX=\operatorname{Spec}E and the trivial metric, the original supnorm is ∣⋅∣E|\cdot|_E. On E⊗KEE\otimes_KE, the scalar extension of this norm is strictly larger than the supnorm of the pulled-back metric. More precisely, q=1p∑i=0p−1α−i⊗αisatisfies∥q∥tensor=∣p∣−1,∥q∥sup=1.\begin{equation} q=\frac1p\sum_{i=0}^{p-1}\alpha^{-i}\otimes\alpha^i \quad\text{satisfies}\quad \|q\|_{\mathrm{tensor}}=|p|^{-1},\qquad \|q\|_{\mathrm{sup}}=1. \end{equation} The same discrepancy occurs for sections of O(m)\mathcal O(m) on PEn\mathbf P^n_E regarded as a KK-scheme, in every dimension nn and every m≥1m\geq 1. This is an explicit instance of the obstruction in [BE21] (Lemma 6.2), not a counterexample to that lemma.

Proof. Eisenstein’s criterion gives [E:K]=p[E:K]=p, and the automorphisms are σj(α)=ζjα\sigma_j(\alpha)=\zeta^j\alpha, 0≤j<p0\leq j<p. The distinct valuation classes of 1,α,…,αp−11,\alpha,\ldots,\alpha^{p-1} make this an orthogonal KK-basis of the normed space EE. Its scalar-extended norm, with the left tensor factor as coefficient field, is therefore ∥∑ici⊗αi∥tensor=max⁡i∣ci∣ ∣α∣i.\left\|\sum_i c_i\otimes\alpha^i\right\|_{\mathrm{tensor}} =\max_i |c_i|\,|\alpha|^i. For qq all displayed terms have norm ∣p∣−1|p|^{-1}. On the other hand, the splitting isomorphism E⊗KE⟶∏j=0p−1E,a⊗b⟼(aσj(b))jE\otimes_KE\longrightarrow\prod_{j=0}^{p-1}E, \qquad a\otimes b\longmapsto(a\sigma_j(b))_j sends qq to (1,0,…,0)(1,0,\ldots,0), by the finite geometric sum p−1∑iζijp^{-1}\sum_i\zeta^{ij}. The pulled-back metric has supnorm equal to the maximum of the component norms, so its value on qq is 11. For projective space, take the section qT0mqT_0^m. The monomial basis together with the orthogonal basis of E/KE/K gives the same tensor norm, and the Gauss point on the first split component gives the same supnorm. All norms use the unique extension of the original absolute value. ◻

Proposition 11.15.8. (Uniform norm comparison and relative volume). Let XX be geometrically reduced and projective over complete nontrivially valued KK, let LL be ample, and let L1,L2\mathscr L_1,\mathscr L_2 be integral models of LL inducing metrics ϕ1,ϕ2\phi_1,\phi_2. Put Nm=h0(X,Lm)N_m=h^0(X,L^m), si,m=∥⋅∥mϕis_{i,m}=\|\cdot\|_{m\phi_i} and ℓi,m=∥⋅∥H0(mLi)\ell_{i,m}=\|\cdot\|_{H^0(m\mathscr L_i)}. There are Ci≥1C_i\geq 1, independent of mm, such that si,m≤ℓi,m≤Cisi,m.s_{i,m}\leq\ell_{i,m}\leq C_i s_{i,m}. With vol⁡(u,v)=log⁡(det⁡v/det⁡u)\operatorname{vol}(u,v)=\log(\det v/\det u) on the determinant line, one has ∣vol⁡(s1,m,s2,m)−vol⁡(ℓ1,m,ℓ2,m)∣≤Nmmax⁡{log⁡C1,log⁡C2}.\begin{equation} \left|\operatorname{vol}(s_{1,m},s_{2,m}) -\operatorname{vol}(\ell_{1,m},\ell_{2,m})\right| \leq N_m\max\{\log C_1,\log C_2\}. \end{equation} If d=dim⁡Xd=\dim X, the difference multiplied by d!/md+1d!/m^{d+1} tends to zero. Thus replacing supnorms by these lattice norms preserves relative volume, although it need not preserve any individual norm.

Open proof obligation (RC2). Prove the uniform bound and determinant-to-volume limit; BE21 Theorem 6.4 and Lemma 9.17 are used.

NAG-900191 · Proof status

There is a further distinction in the scalar-extension step. For a complete extension F/KF/K, the pulled-back metric has the same supnorm on sections originally defined over KK. On all of H0(XF,LF)H^0(X_F,L_F), however, its supnorm is only bounded above by the scalar extension of the original norm in general; see [BE21] (Lemmas 6.1 and 6.2). Equality on the original sections is sufficient in the comparison proof. One must not replace it by an unproved isometry on the enlarged vector space. None of these model and norm results identifies an arbitrary positive CLD current with a nef metric, or establishes an unrestricted Bedford–Taylor product.

Geometric reducedness also has a concrete role in the extension step. Over K=Fp((π))K=\mathbf F_p((\pi)), put E=K(α)E=K(\alpha) with αp=π\alpha^p=\pi. The scheme Spec⁡E\operatorname{Spec}E is reduced over KK, but E⊗KE≃E[ε]/(εp)E\otimes_KE\simeq E[\varepsilon]/(\varepsilon^p). The nonzero element 1⊗α−α⊗11\otimes\alpha-\alpha\otimes 1 has zero analytic sup-seminorm after extension, whereas its scalar-extended vector-space norm is positive. Thus the proof cannot replace geometric reducedness by reducedness when it requires a norm on every extended section space. This calculation is not a counterexample to BE Theorem 6.4 under its stated hypotheses.

11.16 An analytic realization of the fixed-chart obstruction

Example 11.10.9 separated relative convexity from ambient convex extension on an abstract fan. We now realize that fan by analytic units. The same chart also supports an original CLD smooth psh function with no ambient convex extension, even locally at its central vertex. This gives a concrete test for a proposed smooth-to-PL argument: it must allow relative convexity or a change of chart. It does not contradict such a comparison when these choices are allowed.

Four punctures and two tropical coordinates

Let KK be complete and nontrivially valued, with ∣2∣=1|2|=1. On C=PK1∖{0,1,−1,∞}C=\mathbf P^1_K\setminus\{0,1,-1,\infty\} use the analytic units f=t(t−1)f=t(t-1) and g=t(t+1)g=t(t+1), and write x=−log⁡∣f∣,y=−log⁡∣g∣,s=−log⁡∣t∣,τ=(x,y):Can⟶R2.\begin{gathered} x=-\log|f|,\qquad y=-\log|g|,\qquad s=-\log|t|,\\ \tau=(x,y):C^{\mathrm{an}}\longrightarrow\mathbf R^2. \end{gathered} The unit-disc Gauss point ξ\xi belongs to CanC^{\mathrm{an}} and maps to the origin. No extension of KK is made.

Proposition 11.16.1. (The fan and its analytic refinement). The image of τ\tau is the support of the fan in Example 11.10.9: the four rays generated by (1,0),(0,1),(1,1),(−1,−1)(1,0),(0,1),(1,1),(-1,-1). The refinement by the three units t−1,t+1,tt-1,t+1,t has image the four rays generated by e1,e2,e3,−e1−e2−e3e_1,e_2,e_3,-e_1-e_2-e_3. Under (z1,z2,z3)=(−log⁡∣t−1∣,−log⁡∣t+1∣,−log⁡∣t∣),A(z1,z2,z3)=(z1+z3,z2+z3).\begin{gathered} (z_1,z_2,z_3)=(-\log|t-1|,-\log|t+1|,-\log|t|),\\ A(z_1,z_2,z_3)=(z_1+z_3,z_2+z_3). \end{gathered} one has τ=A∘τ~\tau=A\circ\widetilde\tau. The relative PL function of that example pulls back to max⁡{s,0}\max\{s,0\}. Thus its previously abstract convex presentation after refinement is realized by analytic units.

Proof. If ∣t∣<1|t|<1, the ultrametric inequality gives ∣t−1∣=∣t+1∣=1|t-1|=|t+1|=1, so (x,y)=(s,s)(x,y)=(s,s) with s>0s>0. If ∣t∣>1|t|>1, both ∣t±1∣|t\pm 1| equal ∣t∣|t|, so (x,y)=(2s,2s)(x,y)=(2s,2s) with s<0s<0. If ∣t∣=1|t|=1, both ∣t±1∣≤1|t\pm 1|\leq 1. They cannot both be strictly smaller than 11, because their difference is 22 and ∣2∣=1|2|=1. Thus the image lies on a nonnegative coordinate ray or at the origin. These arguments apply to the multiplicative seminorm of every Berkovich point.

Conversely, the Gauss points of discs centered at 0,1,−10,1,-1 with radius r<1r<1, and of discs centered at 00 with radius r>1r>1, realize every point on the four rays. Arbitrary positive real radii are allowed for Berkovich points; no value-group enlargement is needed. The same case distinction gives the three-coordinate image and the identity involving AA. On the ray over 00 the PL function is ss, and it is zero on the other rays, hence equals max⁡{s,0}\max\{s,0\}. The orders of f,gf,g at the four punctures are (1,1),(1,0),(0,1),(−2,−2)(1,1),(1,0),(0,1),(-2,-2), explaining the primitive ray weights 1,1,1,21,1,1,2 and their balancing identity. ◻

Lemma 11.16.2. (A smooth barrier does not remove the first-order obstruction). Let F0F_0 denote the relative PL function of Example 11.10.9, so F0(h,h)=hF_0(h,h)=h and F0(2h,0)=F0(0,2h)=0F_0(2h,0)=F_0(0,2h)=0. For every ambient C2C^2 function BB near the origin, F0+B∣ΣF_0+B|_\Sigma still has no ambient convex extension near the origin. If ∥D2B∥op≤M\|D^2B\|_{\mathrm{op}}\leq M on the relevant segments, its midpoint convexity defect is at least h−Mh2h-Mh^2.

Proof. Put a=(h,h)a=(h,h) and v=(h,−h)v=(h,-h). The affine part of the Taylor expansion cancels in B(a)−12B(a+v)−12B(a−v)B(a)-\frac 12B(a+v)-\frac 12B(a-v). The remainder is bounded below by −M∥v∥2/2=−Mh2-M\|v\|^2/2=-Mh^2. Adding the exact defect hh of F0F_0 gives the assertion, which is positive for all sufficiently small h>0h>0. Thus even an arbitrarily large fixed quadratic barrier cannot correct this first-order failure. ◻

For a convex smooth BB, the pullback max⁡{s,0}+B(x,y)\max\{s,0\}+B(x,y) is nevertheless psh: the first summand is a locally uniform limit of smooth convex functions of the pluriharmonic unit logarithm ss, and the second is an ambient convex pullback. The lemma therefore concerns a false extension shortcut even inside a positive class. It does not obstruct relative PL approximation or a change of analytic presentation.

A smooth psh function that is not ambient convex in this chart

Example 11.16.3. (A fixed-chart obstruction already for smooth psh functions). The function u=xyu=xy on CanC^{\mathrm{an}} is original CLD smooth and psh on all original tests. Nevertheless, the function F(a,b)=abF(a,b)=ab restricted to τ(Can)\tau(C^{\mathrm{an}}) has no ambient convex extension on any convex open neighbourhood of the origin. Moreover, on any subset containing (h,h),(2h,0),(0,2h)(h,h),(2h,0),(0,2h) with h>0h>0, every restriction GG of an ambient convex function satisfies max⁡z∈{(h,h),(2h,0),(0,2h)}∣G(z)−F(z)∣≥h22.\begin{equation} \max_{z\in\{(h,h),(2h,0),(0,2h)\}}|G(z)-F(z)|\geq\frac{h^2}{2}. \end{equation} Every ordinary open neighbourhood of ξ\xi has a tropical image containing such a triple for some h>0h>0. Consequently a fixed-chart uniform approximation by ambient convex functions fails on every such neighbourhood.

Proof. Smoothness follows because x,yx,y are logarithms of analytic units and multiplication is a smooth operation. The case distinction in Proposition 11.16.1 gives the additional identity u=H(s),H(v)=(max⁡{v,0})2+4(max⁡{−v,0})2.u=H(s),\qquad H(v)=(\max\{v,0\})^2+4(\max\{-v,0\})^2. The real function HH is convex: its derivative is 8v8v for v<0v<0 and 2v2v for v>0v>0, with common value zero at the origin. Choose smooth convex regularizations HϵH_\epsilon converging uniformly on compact real intervals to HH. The function s=−log⁡∣t∣s=-\log|t| is original smooth and pluriharmonic, since tt is a unit. Therefore d′d′′Hϵ(s)=Hϵ′′(s) d′s∧d′′s≥0.d'd''H_\epsilon(s)=H_\epsilon''(s)\,d's\wedge d''s\geq0. For every compactly supported original CLD test, ss is bounded on its support, so Hϵ(s)→uH_\epsilon(s)\to u uniformly there. Passing to curvature currents preserves positivity on every original positive test. This proves psh without asserting that the ambient polynomial abab is convex. The relative function has touching affine lower supports: zero at the vertex and a tangent affine function along a ray away from the vertex.

Suppose the maximum error on the three points is ϵ\epsilon. Since (h,h)(h,h) is the midpoint of (2h,0)(2h,0) and (0,2h)(0,2h), convexity gives h2−ϵ≤G(h,h)≤12G(2h,0)+12G(0,2h)≤ϵ.h^2-\epsilon\leq G(h,h) \leq\tfrac12G(2h,0)+\tfrac12G(0,2h)\leq\epsilon. This proves (11.16.1) and rules out an exact extension. The three Gauss-point paths toward 0,1,−10,1,-1 are continuous and have common limit ξ\xi as their radii tend to 11. An ordinary neighbourhood of ξ\xi therefore contains initial segments of all three paths, giving the asserted triple for small hh. The argument also rules out uniform convergence on a fixed compact set containing the corresponding three analytic points. ◻

Corollary 11.16.4. (The obstruction in every positive dimension). For every d≥1d\geq 1, the preceding smooth psh example occurs on an ordinary open subset of the integral projective variety (PK1)d(\mathbf P^1_K)^d. In the product moment chart the same fixed-chart obstruction persists, while a different analytic presentation makes the function ambient convex.

Proof. Use C×(Gm)d−1C\times(\mathbf G_m)^{d-1}, pull uu back from its first factor, and take the chart consisting of f,gf,g and the torus coordinates. The smooth convex approximations Hϵ(−log⁡∣t∣)H_\epsilon(-\log|t|) still have positive original curvature, so their locally uniform limit is psh. Fix the remaining tropical coordinates to zero. Restricting any proposed ambient convex extension to that affine slice contradicts (11.16.1). On the other hand, the unit tt alone presents the function as the pullback of the convex real function HH. The coefficient-controlled approximation of Section 11.12.4 can be applied to this latter presentation on bounded tropical images. ◻

The example closes the analytic-realization question left by the earlier fan calculation, but not the general chart-selection problem. It proves failure of one fixed-chart shortcut, together with an explicit successful change of analytic presentation in the example itself. In particular, it is not a counterexample to full Wu Theorem A.3, and it does not identify the general original smooth psh class with a prescribed tropical generator class.

For uniform approximation, the fixed-domain qualification is essential: the error bound concerns ambient convex functions defined at the three fixed points, or locally convex functions on one fixed ambient neighbourhood containing a convex ball around the origin. It does not rule out approximants whose ambient neighbourhoods shrink with the index, nor approximants defined only by relative supports on the tropical image.

Removing the residue-characteristic restriction

The four rational punctures above require ∣2∣=1|2|=1 so that their residue directions remain distinct. That restriction can be removed by replacing one rational puncture with the zero locus of a polynomial whose reduction avoids both 00 and 11. The resulting function and its convex alternative presentation are still defined over the original field.

Theorem 11.16.5. (A fixed-chart smooth psh obstruction over every nontrivial field). Let KK be any complete nontrivially valued non-Archimedean field, with its fixed absolute value. Set P(t)=t2−t+1,C=PK1∖({0,1,∞}∪V(P)),f=t(t−1),g=tP(t),x=−log⁡∣f∣,y=−log⁡∣g∣,s=−log⁡∣t∣.\begin{gathered} P(t)=t^2-t+1,\\ C=\mathbf P^1_K\setminus\bigl(\{0,1,\infty\}\cup V(P)\bigr),\\ f=t(t-1),\qquad g=tP(t),\\ x=-\log|f|,\quad y=-\log|g|,\quad s=-\log|t|. \end{gathered} Then u=xyu=xy is original smooth psh on CanC^{\mathrm{an}}, but it has no ambient convex presentation in the fixed chart (x,y)(x,y) near the unit-disc Gauss point. The quantitative bound (11.16.1) still holds. The example extends to an ordinary open subset of (PK1)d(\mathbf P^1_K)^d for every d≥1d\geq 1. In particular, residue characteristics 22 and 33 are included.

Proof. Both ff and gg are units on CC. If ∣t∣<1|t|<1, then ∣t−1∣=∣P(t)∣=1|t-1|=|P(t)|=1, giving (x,y)=(s,s)(x,y)=(s,s). If ∣t∣>1|t|>1, then ∣t−1∣=∣t∣|t-1|=|t| and ∣P(t)∣=∣t∣2|P(t)|=|t|^2, giving (x,y)=(2s,3s)(x,y)=(2s,3s). If ∣t∣=1|t|=1, put f=t(t−1)f=t(t-1) and note the exact identity P=f+1P=f+1. The numbers ∣f∣,∣P∣|f|,|P| are at most 11 and cannot both be smaller than 11. Since ∣g∣=∣P∣|g|=|P| in this case, at least one of x,yx,y is zero. Therefore u=H(s),H(v)=(max⁡{v,0})2+6(max⁡{−v,0})2.u=H(s),\qquad H(v)=(\max\{v,0\})^2+6(\max\{-v,0\})^2. This function HH is convex, with derivatives 12v12v for v<0v<0 and 2v2v for v>0v>0. The same convex regularization argument as above proves positivity against all original compactly supported tests. The formula u=xyu=xy proves original smoothness.

The diagonal and horizontal ray segments at the origin are realized by Gauss points of small discs around 00 and 11. To exhibit the vertical segment, choose a root aa of PP in a finite extension K′/KK'/K, endowed with the unique extended absolute value. Both roots have absolute value 11, and P(1)=1P(1)=1 implies ∣a−1∣=1|a-1|=1. Gauss points of K′K'-discs centered at aa with radii r<1r<1 and close to 11 therefore have ∣t∣=∣t−1∣=1|t|=|t-1|=1, while ∣P(t)∣<1|P(t)|<1 approaches 11 as r↑1r\uparrow 1. More precisely, factor PP over a finite splitting extension. On such a disc, ∣P∣a,r=∏P(b)=0max⁡{r,∣a−b∣},|P|_{a,r}=\prod_{P(b)=0}\max\{r,|a-b|\}, where roots are counted with multiplicity. For rr sufficiently close to 11, this is rqr^q with q∈{1,2}q\in\{1,2\}, so it realizes a whole vertical interval. Restricting these multiplicative seminorms to K[t]K[t] gives actual points of CanC^{\mathrm{an}} tending to its unit-disc Gauss point. This use of an auxiliary field constructs points over KK; the function, the positivity proof and the approximation question have not been base-changed, and no current-descent assertion is used.

Every ordinary neighbourhood of that Gauss point consequently contains analytic points with tropical coordinates (h,h),(2h,0),(0,2h)(h,h),(2h,0),(0,2h) for small h>0h>0. Their function values are h2,0,0h^2,0,0, giving the same midpoint obstruction and error bound. The product construction of Corollary 11.16.4 gives every positive dimension. Repeated roots of PP cause no difficulty: their multiplicities are already included in the displayed Gauss-norm formula. ◻

11.17 Relative approximation in a chart without convex extension

The obstruction in Section 11.16 concerns ambient convex functions. It leaves open whether relative convex PL functions suffice on the very same chart. We answer this question for the explicit family constructed there, including a perturbation with a uniform positive tangential Hessian. The proof keeps the value group visible: all affine slopes of the approximants are rational, and all constants belong to ΓQ\Gamma_{\mathbf Q}. It gives a worked comparison in every positive dimension, rather than the general converse in Wu’s Theorem A.3.

The source distinction behind the construction is already important for smooth functions. Gubler–Künnemann retain the smooth facewise-convexity criterion in [GK19] (Remark 3.2), while their Example 3.3 disproves its extension to arbitrary continuous functions; see Example 11.1.4. For a smooth representative on a locally finite complex, Lemma 11.8.4 supplies a centred affine lower support at each point. Neither statement supplies an ambient convex extension, nor a lower bound on the contact radius independent of the point. The following calculation separates all three assertions.

Tangential positivity and the size of a contact neighbourhood

Fix a complete nontrivially valued field KK with its given absolute value. Use the all-field construction of Theorem 11.16.5: P(t)=t2−t+1,C=PK1∖({0,1,∞}∪V(P)),Y=C×(Gm)d−1,d≥1,x=−log⁡∣t(t−1)∣,y=−log⁡∣tP(t)∣,zi=−log⁡∣wi∣,τ=(x,y,z1,…,zd−1).\begin{gathered} P(t)=t^2-t+1,\qquad C=\mathbf P^1_K\setminus\bigl(\{0,1,\infty\}\cup V(P)\bigr),\\ Y=C\times(\mathbf G_m)^{d-1},\qquad d\geq1,\\ x=-\log|t(t-1)|,\quad y=-\log|tP(t)|,\quad z_i=-\log|w_i|,\\ \tau=(x,y,z_1,\ldots,z_{d-1}). \end{gathered} Here wiw_i are the torus coordinates. The image is S×Rd−1S\times\mathbf R^{d-1}, where SS is the union of the four rays generated by (1,1),(1,0),(0,1),(−2,−3)(1,1),(1,0),(0,1),(-2,-3). The unit-disc Gauss point in the first factor, together with wi=1w_i=1 in the other factors, gives a point ξd\xi_d with tropical image zero.

Proposition 11.17.1. (Positive tangential Hessian without ambient convex extension). Let 0<ε<1/20<\varepsilon<1/2, and put Fε(x,y,z)=xy+ε(x2+y2+∑i=1d−1zi2),uε=Fε∘τ.\begin{equation} F_\varepsilon(x,y,z)=xy+\varepsilon\left(x^2+y^2+\sum_{i=1}^{d-1}z_i^2\right), \qquad u_\varepsilon=F_\varepsilon\circ\tau. \end{equation} Then uεu_\varepsilon is original CLD smooth and psh. For each maximal face σ\sigma of S×Rd−1S\times\mathbf R^{d-1} and each vector vv tangent to σ\sigma, D2Fε(v,v)≥2ε∥v∥2.\begin{equation} D^2F_\varepsilon(v,v)\geq2\varepsilon\|v\|^2. \end{equation} Nevertheless, Fε∣S×Rd−1F_\varepsilon|_{S\times\mathbf R^{d-1}} has no ambient convex extension near zero. If an ambient convex function GG is defined on a convex set containing ph=(h,h,0),qh=(2h,0,0),rh=(0,2h,0),h>0,p_h=(h,h,0),\qquad q_h=(2h,0,0),\qquad r_h=(0,2h,0),\quad h>0, then max⁡a∈{ph,qh,rh}∣G(a)−Fε(a)∣≥1−2ε2h2.\begin{equation} \max_{a\in\{p_h,q_h,r_h\}}|G(a)-F_\varepsilon(a)| \geq\frac{1-2\varepsilon}{2}h^2. \end{equation} Every ordinary neighbourhood of ξd\xi_d contains analytic points realizing such triples for all sufficiently small h>0h>0. Thus ambient convex approximation in this chart cannot converge uniformly on any fixed such neighbourhood.

Proof. Theorem 11.16.5 proves that xyxy is original smooth psh, by expressing it as the convex function s+2+6s−2s_+^2+6s_-^2 of s=−log⁡∣t∣s=-\log|t| and using smooth regularization. The added quadratic is an ambient smooth convex function of unit logarithms. Its pullback is smooth psh, so their sum is psh on all original CLD tests. No pullback operation on arbitrary currents is involved.

Write a tangent vector as v=(aα,aβ,vz)v=(a\alpha,a\beta,v_z), where (α,β)(\alpha,\beta) is a generator of one of the four rays. In each case αβ≥0\alpha\beta\geq 0, and D2Fε(v,v)=2a2αβ+2ε∥v∥2,D^2F_\varepsilon(v,v)=2a^2\alpha\beta+2\varepsilon\|v\|^2, which proves (11.17.2). This is the precise tangential positivity assertion used here; it is not a claim about every possible convention for strict psh.

Since ph=(qh+rh)/2p_h=(q_h+r_h)/2, the values of FεF_\varepsilon give the midpoint defect Fε(ph)−Fε(qh)+Fε(rh)2=(1−2ε)h2.F_\varepsilon(p_h)-\frac{F_\varepsilon(q_h)+F_\varepsilon(r_h)}2 =(1-2\varepsilon)h^2. If the error at all three points is at most ee, convexity of GG bounds this defect above by 2e2e. This proves (11.17.3). The Gauss-point paths in Theorem 11.16.5, with wi=1w_i=1, converge to ξd\xi_d and realize the three rays in question. Fixing one sufficiently small hh in a proposed approximation domain gives a positive error lower bound independent of the approximant. ◻

The threshold in this quadratic family is exact: the ambient Hessian has eigenvalues 2ε+12\varepsilon+1, 2ε−12\varepsilon-1, and 2ε2\varepsilon in the remaining directions. Thus FεF_\varepsilon itself is ambient convex when ε≥1/2\varepsilon\geq 1/2. A positive lower bound on its restrictions to face tangent spaces is therefore strictly weaker than ambient convexity.

Corollary 11.17.2. (Monomial refinements and nonuniform contact radii). Replacing or augmenting the coordinates of τ\tau only by constants times monomials in its original units does not remove the ambient convex obstruction. Also, any affine lower support ℓ\ell for Fε∣S×Rd−1F_\varepsilon|_{S\times\mathbf R^{d-1}} at php_h, valid on a relative ball B(ph,R)∩(S×Rd−1)B(p_h,R)\cap(S\times\mathbf R^{d-1}), must satisfy R≤2hR\leq\sqrt 2h. In particular, the contact radii cannot have a positive uniform lower bound as ph→0p_h\to 0.

Proof. A monomial change has tropical form b↦Bb+cb\mapsto Bb+c, with an integral matrix BB and c∈Γmc\in\Gamma^m. It preserves the midpoint relation of the three points. The composition of an ambient convex function with this affine map is convex on the inverse image of its convex domain, so the preceding defect argument applies unchanged. Injectivity of BB is unnecessary. For the second assertion, if R>2hR>\sqrt 2h, both qh,rhq_h,r_h belong to the ball. An affine lower support with equality at php_h would give Fε(ph)=ℓ(ph)=ℓ(qh)+ℓ(rh)2≤Fε(qh)+Fε(rh)2,F_\varepsilon(p_h)=\ell(p_h)=\frac{\ell(q_h)+\ell(r_h)}2 \leq\frac{F_\varepsilon(q_h)+F_\varepsilon(r_h)}2, contrary to the positive defect. ◻

There is no conflict with Lemma 11.8.4: that lemma fixes the contact point, and its radius is allowed to depend on the point. Nor does this corollary require a chart change for relative PL approximation. The construction below takes place in the original chart. New analytic units can change the ambient problem: the unit tt used above is not a monomial in t(t−1)t(t-1) and tP(t)tP(t), even after allowing a nonzero integral power of tt. Indeed, in tm=c[t(t−1)]a[tP(t)]bt^m=c[t(t-1)]^a[tP(t)]^b, orders at 11 force a=0a=0, orders at a zero of PP force b=0b=0, and the order at 00 then forces m=0m=0.

An explicit decreasing approximation with controlled coefficients

The general ambient density argument in Section 11.12.4 can be made particularly explicit for the positive-part square. We give the formula because it allows composition with relative PL functions without introducing constants outside the value group.

Lemma 11.17.3. (Rational PL approximation of the positive-part square). Let Γ⊂R\Gamma\subset\mathbf R be a nonzero additive subgroup, fix λ∈Γ>0\lambda\in\Gamma_{>0} and M>0M>0, and put ΓQ=Γ⊗ZQ⊂R\Gamma_{\mathbf Q}=\Gamma\otimes_{\mathbf Z}\mathbf Q\subset\mathbf R. There are nonnegative, nondecreasing, convex PL functions qj:R→Rq_j:\mathbf R\to\mathbf R, with rational slopes and constants in ΓQ\Gamma_{\mathbf Q}, such that on [−M,M][-M,M] qj+1≤qj,0≤qj(t)−t+2≤83λ4−j,j≥0.\begin{equation} q_{j+1}\leq q_j,\qquad 0\leq q_j(t)-t_+^2\leq\frac83\lambda4^{-j},\qquad j\geq0. \end{equation} The decrease in jj is asserted on this fixed interval, not on all of R\mathbf R.

Proof. Choose an integer N0≥1N_0\geq 1 with (4N02)−1≤λ(4N_0^2)^{-1}\leq\lambda, and set Nj=N02jN_j=N_0 2^j and γj=λ4−j\gamma_j=\lambda 4^{-j}. For 0≤k≤⌈2MNj⌉0\leq k\leq\lceil 2MN_j\rceil, put pj,k=kNj,cj,k=γj⌊−pj,k24γj⌋,aj(t)=max⁡k{pj,kt+cj,k}.p_{j,k}=\frac{k}{N_j},\qquad c_{j,k}=\gamma_j\left\lfloor-\frac{p_{j,k}^2}{4\gamma_j}\right\rfloor, \qquad a_j(t)=\max_k\{p_{j,k}t+c_{j,k}\}. The slopes are rational and nonnegative. The rounded constants are in ΓQ\Gamma_{\mathbf Q}; in particular we have not assumed that squares of value-group elements remain in that group. The k=0k=0 branch is zero. Every other branch lies below the tangent line pt−p2/4pt-p^2/4 of t2t^2 at p/2p/2, and hence below t+2t_+^2 on R\mathbf R. For t∈[0,M]t\in[0,M], select p=k/Njp=k/N_j with ∣2t−p∣≤Nj−1|2t-p|\leq N_j^{-1}. The two errors satisfy t2−(pt−p2/4)=(t−p/2)2≤14Nj2≤γj,0≤−p2/4−cj,k<γj.\begin{split} t^2-(pt-p^2/4)&=(t-p/2)^2\leq\frac1{4N_j^2}\leq\gamma_j,\\ 0\leq-p^2/4-c_{j,k}&<\gamma_j. \end{split} For t<0t<0, aj(t)=0=t+2a_j(t)=0=t_+^2. Thus 0≤t+2−aj(t)≤2γj0\leq t_+^2-a_j(t)\leq 2\gamma_j on [−M,M][-M,M]. To obtain decrease without requiring the supporting lines to be nested, define δj=∑k=j∞2γk=83γj,qj=aj+δj.\delta_j=\sum_{k=j}^{\infty}2\gamma_k=\frac83\gamma_j, \qquad q_j=a_j+\delta_j. Both δj\delta_j and every affine constant of qjq_j belong to ΓQ\Gamma_{\mathbf Q}. On the fixed interval, qj−qj+1≥−2γj+δj−δj+1=0,23γj≤qj−t+2≤83γj.q_j-q_{j+1}\geq-2\gamma_j+\delta_j-\delta_{j+1}=0, \qquad \frac23\gamma_j\leq q_j-t_+^2\leq\frac83\gamma_j. These inequalities prove the claim. Nonnegativity and monotonicity as functions of tt follow from the zero branch, the positive shift, and the nonnegative slopes. ◻

Proposition 11.17.4. (Composition with relative PL supports). Let TT be a subset of a real vector space. Suppose that the continuous functions ψ1,…,ψm:T→[0,M]\psi_1,\ldots,\psi_m:T\to[0,M] have local affine lower supports and finite-selection PL presentations with rational slopes and constants in ΓQ\Gamma_{\mathbf Q}. For bi∈Q≥0b_i\in\mathbf Q_{\geq 0}, the functions vj=∑i=1mbiqj(ψi)v_j=\sum_{i=1}^m b_i q_j(\psi_i) have the same two properties. On all of TT, they decrease uniformly to ∑ibiψi2\sum_i b_i\psi_i^2, with error at most (8/3)λ4−j∑ibi(8/3)\lambda 4^{-j}\sum_i b_i. If TT is the image of a chart of analytic units over KK and Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|, their pullbacks are actual Q\mathbf Q-PL functions by Proposition 11.11.5, on its good strictly analytic pure-dimensional domain. They are classically psh by Theorem 11.11.2, and have nef residues by Theorem 11.11.6.

Proof. Fix a∈Ta\in T and an affine lower support ℓ\ell for ψi\psi_i at aa. Choose an active affine branch pt+cpt+c of qjq_j at ψi(a)\psi_i(a). Since p≥0p\geq 0, near aa one has qj(ψi(b))≥pψi(b)+c≥pℓ(b)+c,q_j(\psi_i(b))\geq p\psi_i(b)+c\geq p\ell(b)+c, with equality at b=ab=a. Summing the resulting affine lower supports proves the support assertion. Finite selections remain finite after composition with a finite maximum and after a finite sum. Rational coefficients remain rational, and ΓQ\Gamma_{\mathbf Q} is closed under rational multiplication and addition. The uniform estimate and the decrease follow term by term from Lemma 11.17.3. The analytic conclusions use the three cited results with exactly these coefficient and support hypotheses; no ambient convex extension is required. ◻

Theorem 11.17.5. (Same-chart relative approximation of the obstructed family). In (11.17.1), assume also ε∈Q\varepsilon\in\mathbf Q. For every M>0M>0, on the single ordinary open set V=τ−1((−M,M)d+1)⊂Yan,V=\tau^{-1}\bigl((-M,M)^{d+1}\bigr)\subset Y^{\mathrm{an}}, there is a decreasing sequence of relative convex rational PL functions in the original chart, with constants in ΓQ\Gamma_{\mathbf Q}, whose pullbacks converge uniformly to uεu_\varepsilon. These pullbacks are actual Q\mathbf Q-PL functions with nef residues. No extension of the base field and no change of the analytic chart are needed.

Proof. On SS, define A(x,y)=max⁡{0,min⁡(x,y)},B(x,y)=max⁡{0,−x/2,−y/3}.A(x,y)=\max\{0,\min(x,y)\},\qquad B(x,y)=\max\{0,-x/2,-y/3\}. The function AA has local affine lower supports: at the vertex, zero is a support; away from it a sufficiently small relative neighbourhood lies on one ray, where AA is affine. The same argument applies on S×Rd−1S\times\mathbf R^{d-1}, since AA is independent of zz. The function BB is ambient convex. Both have rational finite-selection PL presentations with zero constants. Evaluating on the four rays gives xy=A2+6B2on S.\begin{equation} xy=A^2+6B^2\quad\text{on }S. \end{equation} Indeed, the values of (A,B)(A,B) are (s,0)(s,0) on (x,y)=(s,s)(x,y)=(s,s) with s≥0s\geq 0, (0,−s)(0,-s) on (x,y)=(2s,3s)(x,y)=(2s,3s) with s≤0s\leq 0, and (0,0)(0,0) on both nonnegative coordinate rays.

Choose λ∈Γ>0\lambda\in\Gamma_{>0}, which is possible because the valuation is nontrivial, and use the functions qjq_j of Lemma 11.17.3. On the fixed tropical image of VV, put vj=qj(A)+6qj(B)+ε(qj(∣x∣)+qj(∣y∣)+∑i=1d−1qj(∣zi∣)).\begin{equation} v_j=q_j(A)+6q_j(B)+\varepsilon\left(q_j(|x|)+q_j(|y|)+\sum_{i=1}^{d-1}q_j(|z_i|)\right). \end{equation} Every argument of qjq_j lies in [0,M][0,M]. The functions ∣x∣,∣y∣,∣zi∣|x|,|y|,|z_i| are ambient convex rational PL. Proposition 11.17.4 therefore applies to all summands, and (11.17.5) identifies the limit as FεF_\varepsilon. More precisely, throughout τ(V)\tau(V), 0≤vj−Fε≤(7+ε(d+1))83λ4−j,vj+1≤vj.\begin{equation} 0\leq v_j-F_\varepsilon\leq \bigl(7+\varepsilon(d+1)\bigr)\frac83\lambda4^{-j}, \qquad v_{j+1}\leq v_j. \end{equation} The set VV is an ordinary open subset of the smooth analytification YanY^{\mathrm{an}}, hence good, strictly analytic, and pure-dimensional. The coefficient, classical psh and nef-residue conclusions of Proposition 11.17.4 apply on VV. All constants were constructed inside the original ΓQ\Gamma_{\mathbf Q}; the use of its divisible hull means rational model data, not a change of the absolute value or an implicit field extension. ◻

This example gives both the failed ambient approximation and a successful relative approximation on the same fixed neighbourhood. The argument extends beyond squares. The extra point in the next statement is to preserve nonnegative slopes when rationalizing supporting hyperplanes: that sign is what makes composition preserve relative lower supports.

Proposition 11.17.6. (Monotone convex composition in any dimension). Let ψ=(ψ1,…,ψm):T→[0,M]m\psi=(\psi_1,\ldots,\psi_m):T\to[0,M]^m satisfy the hypotheses of Proposition 11.17.4. Let HH be a finite convex function on a convex open neighbourhood of [0,M]m[0,M]^m, nondecreasing in each coordinate there. For every λ∈Γ>0\lambda\in\Gamma_{>0} there are finite maxima QjQ_j of affine functions with slopes in Q≥0m\mathbf Q_{\geq 0}^m and constants in ΓQ\Gamma_{\mathbf Q} such that Qj+1≤Qj,0≤Qj−H≤4λ4−jon [0,M]m.\begin{equation} Q_{j+1}\leq Q_j,\qquad 0\leq Q_j-H\leq4\lambda4^{-j} \quad\text{on }[0,M]^m. \end{equation} Consequently Qj∘ψQ_j\circ\psi are coefficient-compatible relative PL functions decreasing uniformly to H∘ψH\circ\psi on the same set TT. The analytic PL, classical psh and nef-residue conclusions of Proposition 11.17.4 apply to their pullbacks.

Proof. Put D=[0,M]mD=[0,M]^m and γj=λ4−j\gamma_j=\lambda 4^{-j}. As in Proposition 11.12.9, the supporting-hyperplane theorem supplies an affine minorant ℓa(t)=pa⋅t+ba\ell_a(t)=p_a\cdot t+b_a touching HH at each a∈Da\in D. All coordinates of pap_a are nonnegative: if (pa)i<0(p_a)_i<0, the support inequality at a−heia-he_i for small h>0h>0 would contradict the assumed monotonicity of HH. This uses the open neighbourhood of DD, including at its boundary.

The sets where H−ℓa<γjH-\ell_a<\gamma_j cover DD. By continuity and compactness, finitely many of these supports suffice. For each selected pair (p,b)(p,b), choose r∈Q≥0mr\in\mathbf Q_{\geq 0}^m with ri≤pir_i\leq p_i and ∥p−r∥1≤γj/(M+1)\|p-r\|_1\leq\gamma_j/(M+1), and set c=γj⌊b/γj⌋∈ΓQc=\gamma_j\lfloor b/\gamma_j\rfloor\in\Gamma_{\mathbf Q}. On DD, 0≤(p⋅t+b)−(r⋅t+c)≤M∥p−r∥1+(b−c)<2γj.0\leq (p\cdot t+b)-(r\cdot t+c) \leq M\|p-r\|_1+(b-c)<2\gamma_j. Let RjR_j be the maximum of the finitely many resulting affine functions. Then 0≤H−Rj≤3γj0\leq H-R_j\leq 3\gamma_j on DD. Set Qj=Rj+4γjQ_j=R_j+4\gamma_j. Since 4γj−4γj+1=3γj4\gamma_j-4\gamma_{j+1}=3\gamma_j, Qj−Qj+1≥−3γj+4γj−4γj+1=0,γj≤Qj−H≤4γj.Q_j-Q_{j+1}\geq-3\gamma_j+4\gamma_j-4\gamma_{j+1}=0, \qquad \gamma_j\leq Q_j-H\leq4\gamma_j. This proves (11.17.8). At a point of TT, choose local affine lower supports LiL_i for the ψi\psi_i and an active affine branch ∑iriti+c\sum_i r_i t_i+c of QjQ_j at ψ\psi. The function ∑iriLi+c\sum_i r_iL_i+c is a local affine lower support for Qj∘ψQ_j\circ\psi, because every ri≥0r_i\geq 0. Finite selections and the coefficient conditions are preserved just as in Proposition 11.17.4. ◻

The same construction is compatible with a decreasing tuple. More precisely, suppose each ψi,j:T→[0,M]\psi_{i,j}:T\to[0,M] has the stated coefficient-compatible PL and support presentations and ψi,j+1≤ψi,j→ψi\psi_{i,j+1}\leq\psi_{i,j}\to\psi_i. The functions QjQ_j depend only on H,D,ΓH,D,\Gamma, not on these presentations. Hence Qj+1(ψ1,j+1,…,ψm,j+1)≤Qj(ψ1,j+1,…,ψm,j+1)≤Qj(ψ1,j,…,ψm,j).\begin{equation} Q_{j+1}(\psi_{1,j+1},\ldots,\psi_{m,j+1}) \leq Q_j(\psi_{1,j+1},\ldots,\psi_{m,j+1}) \leq Q_j(\psi_{1,j},\ldots,\psi_{m,j}). \end{equation} The error estimate and continuity of HH identify the pointwise limit as H(ψ1,…,ψm)H(\psi_1,\ldots,\psi_m). Each term is still a relative rational PL function on the same TT. Uniform convergence of the tuple gives uniform convergence of these composites by uniform continuity on DD; pointwise convergence alone is not upgraded to uniform convergence. This is a finite-tuple closure statement for the displayed generator class, not closure under arbitrary successive approximation stages.

Corollary 11.17.7. (Original smooth approximation of these composites). Let XX be integral projective of positive dimension over a complete nontrivially valued KK. Suppose an ordinary open W⊂XanW\subset X^{\mathrm{an}} carries a unit chart and a bounded tuple ψ\psi as in Proposition 11.17.6. Then, near each point of WW, the pullback of H∘ψH\circ\psi is a decreasing limit of original smooth psh functions on one ordinary neighbourhood, uniformly on compact subsets.

Proof. The proposition supplies a decreasing sequence of actual Q-PL functions with relative supports on all of WW. Apply Corollary 11.11.7. Its neighbourhood is fixed before the individual extension and smoothing steps. The endpoint is continuous, so the compact-uniform conclusion follows from the last assertion of that corollary. ◻

These results combine the supporting-hyperplane argument of Section 11.12.4, sign-preserving rational approximation and the previously established relative-support bridge. They are recorded as reconstructions and consequences of those inputs, not as a claim of a new general regularization theorem. In particular, we have not shown that every smooth facewise-convex function admits such a finite monotone-convex composition presentation. Section 11.22 supplies coefficient-controlled density on any fixed compact image for continuous relatively supported inputs, without requiring such a composition presentation. The remaining reverse-comparison step concerns arbitrary families with different charts and their common analytic domain. The missing full text of the mixed-calculus Theorem 6.13 is not an input to any argument in this section.

11.18 Relative coefficient approximation on a finite polytopal complex

Section 11.12.7 already proves coefficient-compatible decreasing approximation for homogeneous relative convex functions on a fixed rational fan. Its proof approximates the function and all supporting covectors simultaneously inside a rational system. The obstruction following it shows why an arbitrary nonhomogeneous complex cannot be substituted without checking its affine constants. Here we extend that argument to finite polytopal complexes whose vertices share one value-group scale.

The source convention matters. In [APW26] (Section 2.1), a rational halfspace has an integral normal but an arbitrary real intercept; Section 2.1.6 assumes rational simplicial polyhedral structures in that sense. This does not impose the value-group condition on the vertices used below. The two qualifications must therefore be checked separately when importing PL approximants.

The affine support system

Theorem 11.18.1. (Rational approximation with one geometric scale). Let Γ⊂R\Gamma\subset\mathbf R be a nonzero additive subgroup, and fix λ∈Γ>0\lambda\in\Gamma_{>0}. Let Π\Pi be a finite polytopal complex in Rr\mathbf R^r such that every vertex belongs to λQr\lambda\mathbf Q^r. Let F:∣Π∣→RF:|\Pi|\to\mathbf R be continuous, affine on each cell, and admit a relative affine lower support near every point. Then there are continuous functions FjF_j on ∣Π∣|\Pi|, affine on the same cells with rational slopes and constants in λQ⊂ΓQ\lambda\mathbf Q\subset\Gamma_{\mathbf Q}, admitting relative affine lower supports and satisfying Fj+1≤Fj,0≤Fj−F≤3λ8−jon ∣Π∣.\begin{equation} F_{j+1}\leq F_j,\qquad 0\leq F_j-F\leq3\lambda8^{-j}\quad\text{on }|\Pi|. \end{equation} Here polytopal means that the cells are compact polytopes. In particular, the hypothesis on the vertices holds for a finite (Z,Γ)(\mathbf Z,\Gamma)-polytopal complex when Γ=λZ\Gamma=\lambda\mathbf Z. The conclusion does not require Γ\Gamma itself to be discrete if the displayed one-scale hypothesis holds.

Proof. First write t=λzt=\lambda z and F^(z)=F(λz)/λ\widehat F(z)=F(\lambda z)/\lambda. This is a change of real coordinates and function units for the calculation; the absolute value on any analytic base field remains unchanged. The scaled complex has rational vertices. An affine function with rational coefficients in zz pulls back, after multiplication by λ\lambda, to a function with rational slopes and λQ\lambda\mathbf Q-constants in tt. It therefore suffices to approximate the function and its supports with rational coefficients in the scaled coordinates.

Write the affine expression on a cell σ\sigma as pσ(z)+cσp_\sigma(z)+c_\sigma. For every face ρ\rho, introduce an affine support variable aρ(z)+bρa_\rho(z)+b_\rho. The conditions are pσ(v)+cσ=pτ(v)+cτ(v a vertex of σ∩τ),aρ(v)+bρ=pρ(v)+cρ(v a vertex of ρ),aρ(v)+bρ≤pσ(v)+cσ(σ⊇ρ, v a vertex of σ).\begin{equation} \begin{aligned} p_\sigma(v)+c_\sigma&=p_\tau(v)+c_\tau &&(v\text{ a vertex of }\sigma\cap\tau),\\ a_\rho(v)+b_\rho&=p_\rho(v)+c_\rho &&(v\text{ a vertex of }\rho),\\ a_\rho(v)+b_\rho&\leq p_\sigma(v)+c_\sigma &&(\sigma\supseteq\rho,\ v\text{ a vertex of }\sigma). \end{aligned} \end{equation} The first line is continuity. To justify the support conditions, choose x∈relint⁡(ρ)x\in\operatorname{relint}(\rho). A touching affine lower support agrees with F^\widehat F on the affine span of ρ\rho, since their affine difference has a local minimum of zero there. For y∈σ⊇ρy\in\sigma\supseteq\rho, apply the support inequality at x+s(y−x)x+s(y-x) for small s>0s>0. Linearity on σ\sigma and equality at xx extend the inequality to all of σ\sigma. Testing it on vertices is sufficient. Conversely, the last two lines give a lower support on every cell containing ρ\rho. A small neighbourhood of xx avoids the other cells: the complex is finite, and a cell containing a point in relint⁡(ρ)\operatorname{relint}(\rho) contains ρ\rho. Thus the system is equivalent to the required local relative supports.

All vertices are rational in the scaled coordinates, so (11.18.2) is a finite homogeneous rational linear system in the combined tuple (pσ,cσ,aρ,bρ)(p_\sigma,c_\sigma,a_\rho,b_\rho). Use the rational-density argument already given in Section 11.12.7: impose the active inequalities as equalities, choose a rational basis of the resulting nullspace, and approximate the tuple within that subspace. The inactive inequalities remain strict for sufficiently close approximations. This preserves both continuity and the existence of supports, rather than merely approximating the cellwise formulas.

Return to the tt-coordinates and choose the approximation GjG_j so that ∣Gj−F∣≤ϵj:=λ8−j|G_j-F|\leq\epsilon_j:=\lambda 8^{-j} on ∣Π∣|\Pi|. This is possible by controlling the error at the finitely many vertices: on each cell the difference is affine, and its absolute value is bounded by the largest vertex error. Put Fj=Gj+2ϵjF_j=G_j+2\epsilon_j. Constants still belong to λQ\lambda\mathbf Q, and F+ϵj≤Fj≤F+3ϵj,Fj+1≤F+38ϵj≤Fj.F+\epsilon_j\leq F_j\leq F+3\epsilon_j, \qquad F_{j+1}\leq F+\tfrac38\epsilon_j\leq F_j. Shifting the supports by 2ϵj2\epsilon_j preserves contact. This proves (11.18.1). For the discrete value-group case, a vertex of a (Z,λZ)(\mathbf Z,\lambda\mathbf Z)-polytope is the solution of a full-rank rational linear system with right-hand side in λZ\lambda\mathbf Z, so its coordinates lie in λQ\lambda\mathbf Q. ◻

On an ordinary unit chart whose image is contained in ∣Π∣|\Pi|, the pullbacks of FjF_j are actual Q-PL functions by Proposition 11.11.5, provided the analytic space is good, strictly analytic and pure-dimensional over nontrivially valued KK, and Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|. Their supports give nef residues by Theorem 11.11.6. In the integral projective setting, Corollary 11.11.7 gives an original smooth decreasing approximation near every point on one ordinary neighbourhood. These are consequences of the stated analytic inputs, not a new current-positivity criterion.

The compactness in the theorem serves to control the error by finitely many vertex values. It need not restrict the original tropical image to a compact polytope. A single truncation gives the following local version; the translation also clarifies which affine constants are allowed.

Corollary 11.18.2. (Translation and one fixed truncation). Let Π\Pi be a finite polyhedral complex in Rr\mathbf R^r. Suppose that there are a∈ΓQra\in\Gamma_{\mathbf Q}^r and λ∈Γ>0\lambda\in\Gamma_{>0} such that every cell is given by rational affine inequalities in the coordinates z=(t−a)/λz=(t-a)/\lambda; both normals and intercepts are rational in these coordinates. Let FF be continuous, affine on each cell, and have local relative affine lower supports. For every compact S⊂∣Π∣S\subset|\Pi|, there is an integer N>0N>0 such that S⊂U:=∣Π∣∩(a+(−λN,λN)r),S\subset U:=|\Pi|\cap\bigl(a+(-\lambda N,\lambda N)^r\bigr), and F∣UF|_U admits a decreasing uniform approximation with the bound in (11.18.1), by continuous functions affine on each σ∩U\sigma\cap U, with rational slopes, ΓQ\Gamma_{\mathbf Q}-constants and local relative affine lower supports. The set UU is fixed for the whole sequence.

Proof. Choose NN so that SS lies in the interior of the displayed cube. Intersect the scaled complex with [−N,N]r[-N,N]^r and include all faces. This is a finite rational polytopal complex. Apply Theorem 11.18.1 to F^(z)=F(a+λz)/λ\widehat F(z)=F(a+\lambda z)/\lambda, with subgroup Z\mathbf Z and scale 11. Its hypotheses concern real polyhedra, so this use of Z\mathbf Z makes no change to an analytic value group. A rational affine branch p⋅z+cp\cdot z+c gives, after multiplying by λ\lambda and returning to tt, p⋅t+λc−p⋅a.p\cdot t+\lambda c-p\cdot a. Its slope is rational and its constant belongs to ΓQ\Gamma_{\mathbf Q}. The error is multiplied by λ\lambda; order and relative supports are preserved. Restrict to UU. Every original cell meets the closed cube in a cell of the truncated complex, so the approximants remain affine on σ∩U\sigma\cap U. No extension beyond this fixed truncation is asserted. ◻

In particular, this applies to finite (Z,ΓQ)(\mathbf Z,\Gamma_{\mathbf Q})-polyhedral complexes over a discrete nontrivial value group Γ=λZ\Gamma=\lambda\mathbf Z, even when the cells are unbounded or have lineality. On a unit chart the inverse image of the open cube is an ordinary open set, so the truncation does not force the analytic domain to vary with the approximation index.

Two distinct restrictions on a fixed subdivision

Example 11.18.3. (Irrational directions). In R2\mathbf R^2, put L(x,y)=x+2yL(x,y)=x+\sqrt 2y, and subdivide the plane into L≥0L\geq 0, L≤0L\leq 0 and their common face. The convex function F=max⁡{0,L}F=\max\{0,L\} is linear on these cones. Every continuous function GG which is affine with rational slopes on these same two cones is globally affine. Thus max⁡a∈{(−1,0),(0,0),(1,0)}∣G(a)−F(a)∣≥14.\begin{equation} \max_{a\in\{(-1,0),(0,0),(1,0)\}}|G(a)-F(a)|\geq\frac14. \end{equation}

Proof. Continuity makes the branch difference aLaL for some real aa. Both coordinates of the slope difference are rational, so aa and a2a\sqrt 2 are rational; hence a=0a=0. The midpoint defect of FF at the three displayed points has absolute value 1/21/2, whereas an affine function has zero defect. Three errors bounded by ee change this defect by at most 2e2e, giving e≥1/4e\geq 1/4. ◻

This subdivision is not rational in the APW convention, so the example does not contradict a statement under that convention. The next example has rational directions and isolates the separate intercept condition.

Example 11.18.4. (An intercept outside the value-group hull). Suppose ΓQ≠R\Gamma_{\mathbf Q}\ne\mathbf R, and choose b∉ΓQb\notin\Gamma_{\mathbf Q}. On Rr\mathbf R^r, put F(t)=max⁡{0,t1−b}F(t)=\max\{0,t_1-b\}. Every continuous function that is affine with rational slopes and ΓQ\Gamma_{\mathbf Q}-constants on each of the same halfspaces t1≤bt_1\leq b and t1≥bt_1\geq b is globally affine. For h>0h>0, its uniform error from FF at (b−h,0),(b,0),(b+h,0)(b-h,0),(b,0),(b+h,0) is at least h/4h/4.

Proof. The branch difference vanishes on t1=bt_1=b, so it is a(t1−b)a(t_1-b) with a∈Qa\in\mathbf Q. If a≠0a\ne 0, its constant −ab∈ΓQ-ab\in\Gamma_{\mathbf Q} would force b∈ΓQb\in\Gamma_{\mathbf Q}. Thus the branches agree. The midpoint defect of FF has absolute value h/2h/2, giving the error bound as before. ◻

Both examples allow rational approximation when the subdivision varies: approximate 2\sqrt 2 by rational numbers in the first, and bb by elements of ΓQ\Gamma_{\mathbf Q} in the second. On bounded sets the errors tend uniformly to zero; the error shifts of Proposition 11.12.9 enforce decrease. Neither example refutes unrestricted rational approximation. They exclude only freezing the displayed cells while rounding their coefficients. The earlier nonhomogeneous-complex obstruction in Section 11.12.7 addresses a further issue: vertices may belong to a higher-rank ΓQ\Gamma_{\mathbf Q} without sharing the one scale required in Theorem 11.18.1.

The remaining source-comparison task is therefore precise. The real PL output must be replaced by coefficient-compatible relative PL functions on one fixed domain, while allowing controlled changes of subdivision if its vertices or normals fail these conditions. Theorems for an already rational fan or for a one-scale complex do not establish that replacement for arbitrary smooth inputs. No full Wu Theorem A.3, CLD25 Proposition 8.1.11 or mixed-calculus comparison is inferred here.

11.19 Coefficient closure and relative supports

The one-scale hypothesis of Section 11.18 makes the entire support system rational. With several independent value-group scales, continuity alone may already prevent rational approximation. We first describe that obstruction exactly. We then separate it from the additional requirement of relative convexity, and give a finite strict-support condition under which no further obstruction occurs. All approximations use the same compact support; each statement specifies whether its subdivision is fixed or allowed to vary.

Continuity and the coefficient space

Fix a nonzero additive subgroup Γ⊂R\Gamma\subset\mathbf R and put ΓQ=Γ⊗ZQ⊂R\Gamma_{\mathbf Q}=\Gamma\otimes_{\mathbf Z}\mathbf Q\subset\mathbf R. Let Π\Pi be a nonempty finite complex of compact polytopes in Rr\mathbf R^r, with support S=∣Π∣S=|\Pi|. Write V(Π)V(\Pi) for the real vector space of continuous functions on SS that are affine on each cell. Its uniform norm is the maximum of the absolute values at the finitely many vertices. Define the rational vector space AΓ(Π)={F∈V(Π):F∣σ=pσ⋅t+cσ, pσ∈Qr, cσ∈ΓQ}.A_\Gamma(\Pi)=\{F\in V(\Pi):F|_\sigma=p_\sigma\cdot t+c_\sigma, \ p_\sigma\in\mathbf Q^r,\ c_\sigma\in\Gamma_{\mathbf Q}\}. The formula is required for every cell, but its ambient extension need not be unique. Let VΓ(Π)V_\Gamma(\Pi) denote the uniform closure of AΓ(Π)A_\Gamma(\Pi) in V(Π)V(\Pi). This is a real vector subspace: rational scalar multiplication extends to real scalar multiplication by taking limits. The following description makes this closure explicit when the vertices have value-group coordinates.

Theorem 11.19.1. (Continuity closure). Suppose that SS is connected and every vertex of Π\Pi lies in ΓQr\Gamma_{\mathbf Q}^r. There is a rational linear subspace L⊂Rr#ΠL\subset\mathbf R^{r\#\Pi} and a real linear map T:L×R→V(Π)T:L\times\mathbf R\to V(\Pi) such that AΓ(Π)=T((L∩Qr#Π)×ΓQ),VΓ(Π)=T(L×R).\begin{equation} A_\Gamma(\Pi)=T\bigl((L\cap\mathbf Q^{r\#\Pi})\times\Gamma_{\mathbf Q}\bigr), \qquad V_\Gamma(\Pi)=T(L\times\mathbf R). \end{equation} The construction below uses only the vertex coordinates and continuity equations. For disconnected SS, apply it to each connected component with a separate constant.

Proof. Let EE be the rational span of all vertex coordinates, enlarged by one nonzero element of Γ\Gamma if necessary. Choose a rational basis γ1,…,γs\gamma_1,\ldots,\gamma_s of EE. Introduce one ambient slope variable pσ∈Rrp_\sigma\in\mathbf R^r for each nonempty cell. The graph joining two cells with nonempty intersection is connected: otherwise the corresponding finite compact unions would disconnect SS. Choose a root cell σ0\sigma_0, a spanning tree and, for each edge στ\sigma\tau, a vertex vστv_{\sigma\tau} of the intersection. Along the rooted tree define dσ0(p)=0,dτ(p)=dσ(p)+(pσ−pτ)⋅vστ.d_{\sigma_0}(p)=0,\qquad d_\tau(p)=d_\sigma(p)+(p_\sigma-p_\tau)\cdot v_{\sigma\tau}. Continuity at the remaining intersection vertices requires (pσ−pτ)⋅v+dσ(p)−dτ(p)=∑a=1sγaAσ,τ,v,a(p)=0.\begin{equation} (p_\sigma-p_\tau)\cdot v+d_\sigma(p)-d_\tau(p) =\sum_{a=1}^s\gamma_a A_{\sigma,\tau,v,a}(p)=0. \end{equation} Each Aσ,τ,v,aA_{\sigma,\tau,v,a} is a rational linear form. Let LL be their common kernel, and set T(p,c)∣σ(t)=pσ⋅t+dσ(p)+c.T(p,c)|_\sigma(t)=p_\sigma\cdot t+d_\sigma(p)+c. For p∈Lp\in L, the expressions agree at every intersection vertex, hence on every intersection; thus TT is well-defined. If pp is rational and c∈ΓQc\in\Gamma_{\mathbf Q}, all dσ(p)d_\sigma(p) belong to E⊂ΓQE\subset\Gamma_{\mathbf Q}. This proves one inclusion in the first equality of (11.19.1). Conversely, choose rational ambient slopes and value-group constants for F∈AΓ(Π)F\in A_\Gamma(\Pi). Continuity along the tree forces cσ=cσ0+dσ(p)c_\sigma=c_{\sigma_0}+d_\sigma(p). In (11.19.2), all Aσ,τ,v,a(p)A_{\sigma,\tau,v,a}(p) are rational. The rational independence of the γa\gamma_a therefore forces each to vanish, so p∈Lp\in L.

The rational points of LL are dense in LL, and ΓQ\Gamma_{\mathbf Q} is dense in R\mathbf R, since it contains ρQ\rho\mathbf Q for some ρ>0\rho>0. The image of their product is therefore dense in T(L×R)T(L\times\mathbf R). This image space is finite-dimensional and hence closed in V(Π)V(\Pi). These facts prove the second equality. For disconnected SS, there are only finitely many components, and their function spaces form a finite product. ◻

Remark 11.19.2. The criterion for a target FF is the existence of a representation F=T(p,c)F=T(p,c) with p∈Lp\in L. On a lower-dimensional cell it is not legitimate to test an arbitrary ambient extension and declare failure. Nor does the construction claim an effective procedure for identifying rational relations among arbitrary unspecified real numbers. When all vertices lie in λQr\lambda\mathbf Q^r, one can take s=1s=1, and the split equations are exactly the original continuity equations. Then VΓ(Π)=V(Π)V_\Gamma(\Pi)=V(\Pi), consistently with Theorem 11.18.1. With several scales the equality can fail, as the rectangle calculation in Section 11.12.7 shows.

The relative-support cone

Let C(Π)⊂V(Π)C(\Pi)\subset V(\Pi) consist of functions admitting an affine lower support in the relative topology of SS near every point, with equality at that point. The support is allowed to depend on the point. This is the local support condition of Section 11.18; it is not ambient convexity on a neighbourhood of SS.

Lemma 11.19.3. (Finite support inequalities). The set C(Π)C(\Pi) is a closed convex polyhedral cone in V(Π)V(\Pi), and contains all constants with either sign. Its defining real coefficients need not be rational. More precisely, F∈C(Π)F\in C(\Pi) if and only if for every nonempty face ρ\rho there is an affine function aρa_\rho satisfying aρ(v)=F(v)(v a vertex of ρ),aρ(v)≤F(v)(v a vertex of σ, σ⊇ρ).\begin{equation} \begin{aligned} a_\rho(v)&=F(v)&& (v\text{ a vertex of }\rho),\\ a_\rho(v)&\leq F(v)&& (v\text{ a vertex of }\sigma,\ \sigma\supseteq\rho). \end{aligned} \end{equation}

Proof. Choose x∈relint⁡(ρ)x\in\operatorname{relint}(\rho). A lower support at xx agrees with FF on aff⁡(ρ)\operatorname{aff}(\rho), since their affine difference has a local minimum there. For any y∈σ⊇ρy\in\sigma\supseteq\rho, apply the local inequality at x+t(y−x)x+t(y-x) and divide by t>0t>0; this extends it to yy. Conversely, (11.19.3) gives a support on each coface of ρ\rho, and a sufficiently small neighbourhood of xx meets no other cells. This proves the equivalence.

The system is finite and homogeneous linear in the vertex values and support coefficients. Eliminating the latter preserves polyhedrality. For completeness, to eliminate a variable yy, write each inequality as aiy+bi(z)≥0a_i y+b_i(z)\geq 0. Keep the inequalities with ai=0a_i=0, and require each lower bound from ai>0a_i>0 to be no larger than each upper bound from ai<0a_i<0. If one type is absent there is no additional restriction. Equalities may be written as pairs of inequalities. Repeating this elimination gives finitely many weak linear inequalities in the vertex values alone, proving closedness as well. Addition, nonnegative scaling and constant shifts preserve the support condition directly. ◻

Closedness alone does not justify intersecting with a dense rational class. For instance, the ray {(x,2x):x≥0}\{(x,\sqrt 2 x):x\geq 0\} is a closed real polyhedral cone whose intersection with Q2\mathbf Q^2 is only the origin. This is a warning about a linear-algebra inference, not a tropical or analytic counterexample. The next statement gives the exact replacement.

Theorem 11.19.4. (Closure of admissible rational functions). For any finite compact polytopal complex, set WΓ(Π)=span⁡R(AΓ(Π)∩C(Π)).W_\Gamma(\Pi)=\operatorname{span}_{\mathbf R} \bigl(A_\Gamma(\Pi)\cap C(\Pi)\bigr). Then AΓ(Π)∩C(Π)‾ ∥⋅∥∞=WΓ(Π)∩C(Π).\begin{equation} \overline{A_\Gamma(\Pi)\cap C(\Pi)}^{\,\|\cdot\|_\infty} =W_\Gamma(\Pi)\cap C(\Pi). \end{equation} If FF belongs to this set and ρ∈Γ>0\rho\in\Gamma_{>0}, there are Fj∈AΓ(Π)∩C(Π)F_j\in A_\Gamma(\Pi)\cap C(\Pi) such that Fj+1≤Fj,0≤Fj−F≤3ρ8−jon S.\begin{equation} F_{j+1}\leq F_j,\qquad 0\leq F_j-F\leq3\rho8^{-j}\quad\text{on }S. \end{equation}

Proof. Both WΓ(Π)W_\Gamma(\Pi) and C(Π)C(\Pi) are closed, so the left side of (11.19.4) is contained in the right side. To prove density, choose a real basis g1,…,gmg_1,\ldots,g_m of WΓ(Π)W_\Gamma(\Pi) from AΓ(Π)∩C(Π)A_\Gamma(\Pi)\cap C(\Pi). Its positive cone lies in C(Π)C(\Pi) and has nonempty interior in WΓ(Π)W_\Gamma(\Pi). In particular, H=∑igiH=\sum_i g_i has a ball BW(H,a)B_W(H,a) contained in C(Π)C(\Pi). For F∈WΓ(Π)∩C(Π)F\in W_\Gamma(\Pi)\cap C(\Pi) and ϵ>0\epsilon>0, BW(F+ϵH,ϵa)=F+ϵBW(H,a)⊂C(Π).B_W(F+\epsilon H,\epsilon a) =F+\epsilon B_W(H,a)\subset C(\Pi). Approximate the coordinates of F+ϵHF+\epsilon H in the basis gig_i by rational numbers so closely that the resulting function lies in this ball. It belongs to AΓ(Π)A_\Gamma(\Pi), since that is a rational vector space, and it belongs to C(Π)C(\Pi) by the ball inclusion. Letting ϵ\epsilon and the coordinate errors tend to zero proves density. The argument does not require the rational coordinates to be nonnegative.

Choose GjG_j in this intersection with ∥Gj−F∥∞≤ηj:=ρ8−j\|G_j-F\|_\infty\leq\eta_j:=\rho 8^{-j} and put Fj=Gj+2ηjF_j=G_j+2\eta_j. The added constant belongs to ΓQ\Gamma_{\mathbf Q}, and shifting every support by it preserves contact. Thus F+ηj≤Fj≤F+3ηjF+\eta_j\leq F_j\leq F+3\eta_j and Fj+1≤F+3ηj/8≤FjF_{j+1}\leq F+3\eta_j/8\leq F_j. This proves (11.19.5). ◻

This is an exact closure formula, but not by itself an effective description of WΓ(Π)W_\Gamma(\Pi). A useful finite certificate is a list of admissible rational functions gig_i whose real span contains the target FF. The finite support system checks admissibility, and the theorem then gives decreasing approximation. In general we have only WΓ(Π)⊆VΓ(Π)W_\Gamma(\Pi)\subseteq V_\Gamma(\Pi); equality requires an argument. The proof is an elementary combination of finite-dimensional convexity and rational approximation, and no priority claim is made for that combination.

A strict-support certificate

We now give a geometric sufficient condition for this equality. Say that H∈V(Π)H\in V(\Pi) has strict face supports if, for every nonempty face ρ\rho, there is an affine aρa_\rho agreeing with HH on ρ\rho and satisfying H(v)>aρ(v)(σ⊇ρ, v a vertex of σ, v∉ρ).\begin{equation} H(v)>a_\rho(v)\quad (\sigma\supseteq\rho,\ v\text{ a vertex of }\sigma,\ v\notin\rho). \end{equation} This is a finite condition on a specified complex. It is not a definition of strict CLD plurisubharmonicity. For a maximal face the inequalities are vacuous.

Proposition 11.19.5. (Strict supports remove the positivity obstruction). If H∈AΓ(Π)H\in A_\Gamma(\Pi) has strict face supports, then HH is an interior point of C(Π)C(\Pi) in V(Π)V(\Pi), and WΓ(Π)=VΓ(Π),AΓ(Π)∩C(Π)‾=VΓ(Π)∩C(Π).W_\Gamma(\Pi)=V_\Gamma(\Pi),\qquad \overline{A_\Gamma(\Pi)\cap C(\Pi)}=V_\Gamma(\Pi)\cap C(\Pi). Consequently every target in the latter intersection has the decreasing approximants of (11.19.5).

Proof. For each ρ\rho, choose a linear extension map EρE_\rho from affine functions on aff⁡(ρ)\operatorname{aff}(\rho) to affine functions on Rr\mathbf R^r. Such a map exists by extending affine coordinates, and all the relevant spaces are finite-dimensional. If δ∈V(Π)\delta\in V(\Pi), replace aρa_\rho by aρ+Eρ(δ∣ρ)a_\rho+E_\rho(\delta|_\rho). It agrees with H+δH+\delta on ρ\rho. At the finitely many other incident vertices the change in the support gap is δ(v)−Eρ(δ∣ρ)(v).\delta(v)-E_\rho(\delta|_\rho)(v). The absolute values of these expressions are bounded by M∥δ∥∞M\|\delta\|_\infty for a fixed finite MM. The finitely many strict gaps in (11.19.6) have a positive minimum. For sufficiently small ∥δ∥∞\|\delta\|_\infty, all remain positive, so Lemma 11.19.3 gives H+δ∈C(Π)H+\delta\in C(\Pi). If there are no strict gaps to test, the same conclusion holds for every δ\delta. This proves interiority.

For F∈C(Π)∩VΓ(Π)F\in C(\Pi)\cap V_\Gamma(\Pi), the points F+ϵHF+\epsilon H are interior to C(Π)C(\Pi) relative to VΓ(Π)V_\Gamma(\Pi). The density of AΓ(Π)A_\Gamma(\Pi) in VΓ(Π)V_\Gamma(\Pi) gives admissible rational approximants tending to FF, exactly as in Theorem 11.19.4. Moreover, a sufficiently small ball about HH in VΓ(Π)V_\Gamma(\Pi) is contained in the closure of the admissible class, hence in WΓ(Π)W_\Gamma(\Pi). A linear subspace containing such a ball is all of VΓ(Π)V_\Gamma(\Pi). This proves both equalities and the decreasing conclusion. ◻

The hypothesis has an explicit source of examples, valid in arbitrary dimension and without a one-scale restriction. The subdivision must be the one specified below; adding arbitrary faces can destroy the strict certificate.

Proposition 11.19.6. (Subdivisions of a finite rational maximum). Let Q⊂RrQ\subset\mathbf R^r be a full-dimensional compact polytope given by hb≥0h_b\geq 0, where the hbh_b have rational slopes and constants in ΓQ\Gamma_{\mathbf Q}. Let H=max⁡1≤i≤Nℓi,ℓi(t)=pi⋅t+ci,pi∈Qr, ci∈ΓQ.H=\max_{1\leq i\leq N}\ell_i, \qquad \ell_i(t)=p_i\cdot t+c_i,\quad p_i\in\mathbf Q^r,\ c_i\in\Gamma_{\mathbf Q}. Form the complex of all nonempty regions Ri={t∈Q:H(t)=ℓi(t)}R_i=\{t\in Q:H(t)=\ell_i(t)\} and all their faces. On any nonempty subcomplex Π\Pi of this complex, H∣∣Π∣H|_{|\Pi|} belongs to AΓ(Π)A_\Gamma(\Pi) and has strict face supports. Every vertex of Π\Pi lies in ΓQr\Gamma_{\mathbf Q}^r. Thus the continuity criterion of Theorem 11.19.1, component by component, is necessary and sufficient for a relatively supported F∈V(Π)F\in V(\Pi) to admit the approximants (11.19.5).

Proof. The regions form a polyhedral complex: Ri∩RjR_i\cap R_j is the equality face for ℓi−ℓj≥0\ell_i-\ell_j\geq 0 on RiR_i, and similarly on RjR_j. All defining normals are rational and constants belong to ΓQ\Gamma_{\mathbf Q}. At a vertex, a full-rank set of active equations has a rational inverse matrix, so the coordinates belong to ΓQ\Gamma_{\mathbf Q}. This argument also applies to lower-dimensional regions.

Fix a face ρ\rho, choose x∈relint⁡(ρ)x\in\operatorname{relint}(\rho), and set I={i:ℓi(x)=H(x)}I=\{i:\ell_i(x)=H(x)\} and B={b:hb(x)=0}B=\{b:h_b(x)=0\}. Define aρ=1∣I∣∑i∈Iℓi−∑b∈Bhb.\begin{equation} a_\rho=\frac1{|I|}\sum_{i\in I}\ell_i-\sum_{b\in B}h_b. \end{equation} On QQ, both H−∣I∣−1∑i∈IℓiH-|I|^{-1}\sum_{i\in I}\ell_i and every hbh_b in this sum are nonnegative. Their sum vanishes precisely where all the indicated maximum branches and boundary constraints are active. This set is exactly ρ\rho: inside any Ri0R_{i_0} with i0∈Ii_0\in I, these are all the constraints active at xx, hence they define the minimal face containing xx. Consequently aρ=Ha_\rho=H on ρ\rho and aρ<Ha_\rho<H at every point of Q∖ρQ\setminus\rho. Formula (11.19.7) supplies the required strict face supports, even after restricting to a subcomplex. The function HH is rational affine on each cell. Proposition 11.19.5 and Theorem 11.19.1 finish the proof. ◻

One fixed refinement

A complex need not itself have strict face supports. Refinement gives a precise remedy when its cells have rational normals and value-group constants. The continuity obstruction, however, survives every fixed refinement of the kind considered here.

Lemma 11.19.7. (Continuity closure under a fixed refinement). Let Π′\Pi' be a finite compact polytopal refinement of Π\Pi with the same support: each refined cell lies in a coarse cell, and each coarse cell is a union of refined cells. Suppose that every cell of Π\Pi has rational tangent space, and all vertices of both complexes belong to ΓQr\Gamma_{\mathbf Q}^r. Under the natural inclusion V(Π)⊆V(Π′)V(\Pi)\subseteq V(\Pi'), VΓ(Π′)∩V(Π)=VΓ(Π).\begin{equation} V_\Gamma(\Pi')\cap V(\Pi)=V_\Gamma(\Pi). \end{equation}

Proof. Restriction to the refined cells gives AΓ(Π)⊆AΓ(Π′)A_\Gamma(\Pi)\subseteq A_\Gamma(\Pi'), hence one inclusion. For the reverse inclusion, use Theorem 11.19.1 on Π′\Pi', component by component, to represent F∈VΓ(Π′)∩V(Π)F\in V_\Gamma(\Pi')\cap V(\Pi) by slope parameters p∈Lp\in L and root constants. For each coarse cell σ\sigma, take all refined cells τ⊆σ\tau\subseteq\sigma of dimension dim⁡σ\dim\sigma. They cover σ\sigma: their finite closed union contains the relative interior densely and therefore contains its closure. Choose a rational basis uσ,ku_{\sigma,k} for the tangent space of σ\sigma. Being affine on σ\sigma implies (pτ−pτ0)⋅uσ,k=0(τ,τ0⊆σ, dim⁡τ=dim⁡τ0=dim⁡σ).\begin{equation} (p_\tau-p_{\tau_0})\cdot u_{\sigma,k}=0 \quad(\tau,\tau_0\subseteq\sigma,\ \dim\tau=\dim\tau_0=\dim\sigma). \end{equation} Conversely, these equations, together with continuity, force affinity on σ\sigma. Subtract an ambient affine function with slope pτ0p_{\tau_0}; the remainder is constant on every full-dimensional refined cell. The intersection graph of these cells is connected because they cover the connected compact polytope σ\sigma, so their constants agree.

Equations (11.19.9) are rational linear equations in the slope parameters. Their intersection with LL is therefore a rational linear subspace containing the chosen representative of FF. Approximate within that subspace by rational parameters, and approximate each root constant in ΓQ\Gamma_{\mathbf Q}. The resulting functions converge uniformly to FF, belong to AΓ(Π′)A_\Gamma(\Pi'), and are affine on every coarse cell. On a coarse cell their expression agrees with that of any full-dimensional refined cell, so it has rational ambient slope and constant in ΓQ\Gamma_{\mathbf Q}. They thus belong to AΓ(Π)A_\Gamma(\Pi), proving (11.19.8). Zero-dimensional coarse cells require no slope equation. ◻

Theorem 11.19.8. (Relative approximation after one fixed refinement). Suppose each cell of the finite compact complex Π\Pi is given by finitely many affine weak inequalities with rational slopes and constants in ΓQ\Gamma_{\mathbf Q}. There is a finite compact refinement Π∗\Pi^*, depending only on Π\Pi, such that for every F∈C(Π)F\in C(\Pi) the following are equivalent:

  1. F∈VΓ(Π)F\in V_\Gamma(\Pi).

  2. FF is a uniform limit of functions in AΓ(Π∗)∩C(Π∗)A_\Gamma(\Pi^*)\cap C(\Pi^*).

  3. FF has approximants in that class satisfying (11.19.5).

Proof. List all affine functions h1,…,hNh_1,\ldots,h_N occurring in inequalities for all cells, writing equalities as two inequalities. Choose a cube Q=[−Mρ,Mρ]rQ=[-M\rho,M\rho]^r containing SS in its interior, with ρ∈Γ>0\rho\in\Gamma_{>0} and MM a sufficiently large integer. On this cube set H=∑l=1N∣hl∣=max⁡ϵ∈{−1,1}N∑l=1Nϵlhl.\begin{equation} H=\sum_{l=1}^N|h_l| =\max_{\epsilon\in\{-1,1\}^N}\sum_{l=1}^N\epsilon_l h_l. \end{equation} The regions of this maximum are exactly the nonempty sign regions ϵlhl≥0\epsilon_lh_l\geq 0 in QQ: equality in the maximum requires equality in each nonnegative term ∣hl∣−ϵlhl|h_l|-\epsilon_lh_l. Include their faces. Each original cell is a union of these arrangement cells, since its defining inequalities occur in the list. Take the subcomplex consisting of arrangement cells contained in SS; call it Π∗\Pi^*. It refines every original cell and has the same support. All its vertices lie in ΓQr\Gamma_{\mathbf Q}^r, and Proposition 11.19.6 supplies strict face supports for HH.

The relative-support condition depends on the function and the support SS, not on a chosen subdivision. Thus F∈C(Π)F\in C(\Pi) also belongs to C(Π∗)C(\Pi^*). If (1) holds, restriction gives F∈VΓ(Π∗)F\in V_\Gamma(\Pi^*), and Proposition 11.19.5 proves (3). Clearly (3) implies (2). Finally (2) gives F∈VΓ(Π∗)∩V(Π)F\in V_\Gamma(\Pi^*)\cap V(\Pi), which is VΓ(Π)V_\Gamma(\Pi) by Lemma 11.19.7. Its rational-tangent hypothesis holds because the original defining normals are rational. ◻

In particular, a fixed rational value-group refinement can remove the extra support obstruction, but cannot remove the continuity obstruction for a target affine on the old cells. This does not exclude progressively changing subdivisions: their arithmetic continuity equations also change. The following construction shows that the distinction is substantive, even when the target is already PL.

Example 11.19.9. (Moving subdivisions resolve the rectangle test). Use the rectangle and function FF of Example 11.12.17, with a=2a=\sqrt 2 and Γ=Z+aZ\Gamma=\mathbf Z+a\mathbf Z. Thus F(x,0)=axF(x,0)=ax, F(0,y)=yF(0,y)=y, and F=aF=a on the top and right edges. Put δj=8−j,qj=δj⌊a/δj⌋,qj+=qj+δj,bj=qj+−aδj∈(0,1)∩ΓQ.\delta_j=8^{-j},\qquad q_j=\delta_j\lfloor a/\delta_j\rfloor, \qquad q_j^+=q_j+\delta_j, \qquad b_j=\frac{q_j^+-a}{\delta_j}\in(0,1)\cap\Gamma_{\mathbf Q}. Define Gj=FG_j=F on the left, top and right edges, and on the bottom edge set Gj(x,0)=max⁡{qjx, a−qj+(1−x)},Fj=Gj+2δj.\begin{equation} G_j(x,0)=\max\{q_jx,\ a-q_j^+(1-x)\},\qquad F_j=G_j+2\delta_j. \end{equation} The two branches meet at x=bjx=b_j. Both have rational slopes and constants in ΓQ\Gamma_{\mathbf Q}. Their maximum equals 00 at x=0x=0 and aa at x=1x=1, so GjG_j is continuous around the rectangle. On the bottom edge it is convex. Near each corner its affine branch and the branch on the adjacent edge extend to one ambient affine function, because the edge directions are independent. It therefore has relative affine lower supports everywhere. The same holds for FjF_j.

Since qj<a<qj+q_j<a<q_j^+, both branches lie below axax, and 0≤ax−Gj(x,0)=min⁡{(a−qj)x,(qj+−a)(1−x)}≤δj.0\leq ax-G_j(x,0) =\min\{(a-q_j)x,(q_j^+-a)(1-x)\}\leq\delta_j. Consequently F+δj≤Fj≤F+2δjF+\delta_j\leq F_j\leq F+2\delta_j on all of SS, and Fj+1≤F+δj/4≤FjF_{j+1}\leq F+\delta_j/4\leq F_j. This is a coefficient-compatible decreasing uniform approximation, obtained by adding just one moving breakpoint. It cannot be carried by any single finite refinement with vertices in ΓQ2\Gamma_{\mathbf Q}^2: Lemma 11.19.7 would then imply F∈VΓ(Π)F\in V_\Gamma(\Pi), contrary to (11.12.5). Pullback to S×[0,1]d−1S\times[0,1]^{d-1} gives the same construction in every positive dimension. These remain abstract polyhedral examples; no balancing or analytic realization is asserted. The changing arithmetic equations can be seen directly. Write N=8jN=8^j and m=⌊Na⌋m=\lfloor Na\rfloor, so bj=m+1−Nab_j=m+1-Na. For rational slopes p1,p2p_1,p_2 on the two bottom segments, ptp_t on the top and ql,qrq_l,q_r on the vertical edges, continuity splits as (m+1)p1−mp2−pt=0,N(p2−p1)+qr−ql=0.(m+1)p_1-mp_2-p_t=0,\qquad N(p_2-p_1)+q_r-q_l=0. The approximant has (p1,p2,pt,ql,qr)=(m/N,(m+1)/N,0,1,0)(p_1,p_2,p_t,q_l,q_r)=(m/N,(m+1)/N,0,1,0) and satisfies both equations. The target would have (a,a,0,1,0)(a,a,0,1,0) and satisfies neither. There is no contradiction with convergence: the equation coefficients change with NN. This also explains why a failure of fixed-complex density cannot decide moving-subdivision density.

Propagation of a common scale

There is another useful route back to the one-scale theorem. Rational simplices of dimension at least two have a stronger arithmetic property than general polytopes with rational facet normals. The property propagates through a shared edge.

Proposition 11.19.10. (Scale propagation through edges). Let Π\Pi be a finite compact simplicial complex in Rr\mathbf R^r. Suppose every maximal simplex has dimension at least two, every edge has rational direction, and the graph joining maximal simplices that share an edge is connected. If all vertices belong to ΓQr\Gamma_{\mathbf Q}^r, then for some a∈ΓQra\in\Gamma_{\mathbf Q}^r and λ∈Γ>0\lambda\in\Gamma_{>0} every vertex lies in a+λQra+\lambda\mathbf Q^r. Consequently every F∈C(Π)F\in C(\Pi) has decreasing approximants in AΓ(Π)∩C(Π)A_\Gamma(\Pi)\cap C(\Pi), uniformly on ∣Π∣|\Pi|.

Proof. First take one simplex with vertices v0,…,vdv_0,\ldots,v_d, d≥2d\geq 2. Write vi−v0=λieiv_i-v_0=\lambda_i e_i, with nonzero ei∈Qre_i\in\mathbf Q^r. The eie_i are linearly independent. The edge vj−vi=λjej−λieiv_j-v_i=\lambda_j e_j-\lambda_i e_i is a real multiple of a rational vector. A rational left inverse to the matrix with columns eie_i shows that the ratio λi/λj\lambda_i/\lambda_j is rational: its two nonzero coordinates are real multiples of two nonzero rational numbers. Thus all vertex differences of this simplex belong to λσQr\lambda_\sigma\mathbf Q^r for one positive real λσ\lambda_\sigma.

If two simplices share an edge, its nonzero vector belongs to both λσQr\lambda_\sigma\mathbf Q^r and λτQr\lambda_\tau\mathbf Q^r, forcing λσ/λτ∈Q\lambda_\sigma/\lambda_\tau\in\mathbf Q. Propagate this relation along the connected graph. Starting from one vertex aa, shared vertices then show that every vertex lies in a+λ0Qra+\lambda_0\mathbf Q^r for a common positive λ0\lambda_0. A nonzero coordinate of an edge shows λ0∈ΓQ\lambda_0\in\Gamma_{\mathbf Q} after the initial choice of scale; multiplying by a positive integer makes it an element λ\lambda of Γ\Gamma without changing λ0Qr\lambda_0\mathbf Q^r. Translate by aa and apply Theorem 11.18.1. Translation preserves rational slopes and changes a rational affine constant by p⋅a∈ΓQp\cdot a\in\Gamma_{\mathbf Q}, so the approximants retain the required coefficients. ◻

Example 11.19.11. (Rational normals do not ensure rational triangulations). The rectangle Q=[0,1]×[0,2]Q=[0,1]\times[0,\sqrt 2] has integral facet normals and vertices in Γ2\Gamma^2 for Γ=Z+2Z\Gamma=\mathbf Z+\sqrt 2\mathbf Z. Nevertheless it has no finite triangulation all of whose edges have rational direction, even allowing extra vertices. Indeed, the triangles in any finite triangulation of a rectangle are connected through shared edges. To see this, join interior points of two triangles by a polygonal path in the interior, chosen to avoid the finitely many vertices and to cross edges transversely. The proof of Proposition 11.19.10, up to the value-group assertion, puts all vertices of such a rational-direction triangulation in a+λQ2a+\lambda\mathbf Q^2 for some real λ>0\lambda>0. The two boundary vectors (1,0)(1,0) and (0,2)(0,\sqrt 2) would then force both 1/λ1/\lambda and 2/λ\sqrt 2/\lambda to be rational, a contradiction.

This example concerns a compact polytope, not a balanced tropical variety and not a counterexample to the conditional theory in APW26. Section 2.1.6 of that source explicitly assumes the existence of a rational simplicial structure. The example explains why this assumption cannot be replaced silently by rationality of facet normals. Nor may Proposition 11.19.10 be applied to simplices meeting only at vertices or to one-dimensional complexes: the shared nonzero edge is the mechanism that compares their scales.

Convex extension on individual cells

The moving-subdivision problem has a further elementary part that can be completed without a global comparison theorem: matching convex boundary data on one polytope. Recording it separately identifies the remaining issue at joins between cells.

Lemma 11.19.12. (Coefficient-compatible convex boundary extension). Let QQ be a positive-dimensional compact polytope defined by rational-slope, ΓQ\Gamma_{\mathbf Q}-constant affine inequalities. Suppose gg is continuous on its relative boundary, and on each facet is the maximum of finitely many ambient affine functions with those coefficients. Then gg extends to a function GG on QQ which is itself a finite maximum of such affine functions. If an affine real function AA satisfies 0≤A−g≤e0\leq A-g\leq e on the boundary, then for every ϵ>0\epsilon>0 one can arrange G∣∂Q=g,0≤A−G≤e+ϵon Q.\begin{equation} G|_{\partial Q}=g,\qquad 0\leq A-G\leq e+\epsilon\quad\text{on }Q. \end{equation}

Proof. For each facet ρ\rho, choose a rational-slope, value-group-constant affine function hρ≥0h_\rho\geq 0 on QQ whose zero set on QQ is exactly ρ\rho. Take one affine branch ℓ\ell of g∣ρg|_\rho. We claim that ℓ−Chρ≤g\ell-C h_\rho\leq g on the whole boundary for some nonnegative rational CC. Refine the boundary into finitely many compact polytopes on which gg is affine. At a vertex vv with hρ(v)=0h_\rho(v)=0, we already have ℓ(v)≤g(v)\ell(v)\leq g(v). At all other vertices impose C≥ℓ(v)−g(v)hρ(v).C\geq\frac{\ell(v)-g(v)}{h_\rho(v)}. There are only finitely many bounds, so a nonnegative rational CC exists. Vertex inequalities imply the inequalities on every boundary cell. The modified branch still restricts to ℓ\ell on ρ\rho and retains rational slopes and constants in ΓQ\Gamma_{\mathbf Q}. Take the maximum of all modified branches, over all facets and all their branches. It agrees with gg on each facet and is convex on QQ.

If g≤Ag\leq A, every modified affine branch is at most AA at all vertices of QQ, hence at most AA throughout QQ. To obtain the lower bound, choose a rational-slope, ΓQ\Gamma_{\mathbf Q}-constant affine function ℓ0\ell_0 with A−e−ϵ≤ℓ0≤A−eon Q.A-e-\epsilon\leq\ell_0\leq A-e\quad\text{on }Q. Such a choice follows by approximating the slope and constant of A−e−ϵ/2A-e-\epsilon/2, using compactness and the density of ΓQ\Gamma_{\mathbf Q}. On the boundary, ℓ0≤g\ell_0\leq g. Adding ℓ0\ell_0 to the finite maximum therefore preserves the boundary values and gives (11.19.12). ◻

Proposition 11.19.13. (Decreasing cellwise-convex approximation). Let Π\Pi have the coefficient-compatible defining inequalities of Theorem 11.19.8, and let F∈V(Π)F\in V(\Pi). There are continuous rational-slope, ΓQ\Gamma_{\mathbf Q}-constant PL functions FjF_j on SS, carried by subdivisions allowed to depend on jj, such that each FjF_j is convex on every original cell and Fj+1≤Fj,0≤Fj−F≤2ρ8−j(ρ∈Γ>0 fixed).F_{j+1}\leq F_j,\qquad 0\leq F_j-F\leq2\rho8^{-j} \quad(\rho\in\Gamma_{>0}\text{ fixed}). No relative affine support across different original cells is asserted.

Proof. Put d=dim⁡Πd=\dim\Pi and fix ϵ>0\epsilon>0. At each vertex choose a value in ΓQ\Gamma_{\mathbf Q} between F(v)−ϵF(v)-\epsilon and F(v)F(v). Induct on cell dimension. Once a continuous function has been defined on the (k−1)(k-1)-skeleton, its restriction to the boundary of each kk-cell is a finite maximum of coefficient-compatible affine functions on each facet, by the preceding induction steps. Apply Lemma 11.19.12 with A=F∣σA=F|_\sigma and additional error ϵ\epsilon. The extensions agree on common faces because those values have already been fixed. After dimension dd, the resulting continuous function GG satisfies 0≤F−G≤(d+1)ϵ0\leq F-G\leq(d+1)\epsilon on SS and is convex on every original cell. A common refinement by the finitely many branch comparisons gives a finite coefficient-compatible PL presentation.

For the jjth construction choose ϵ=ρ8−j/(d+1)\epsilon=\rho 8^{-j}/(d+1), write ηj=ρ8−j\eta_j=\rho 8^{-j}, and set Fj=Gj+2ηjF_j=G_j+2\eta_j. Then F+ηj≤Fj≤F+2ηjF+\eta_j\leq F_j\leq F+2\eta_j, and Fj+1≤F+ηj/4≤FjF_{j+1}\leq F+\eta_j/4\leq F_j. The shift belongs to ΓQ\Gamma_{\mathbf Q} and preserves convexity on each cell. ◻

The qualification is necessary. On [−1,1][-1,1] with cells [−1,0][-1,0] and [0,1][0,1], the functions ϵ(1−∣x∣)\epsilon(1-|x|) are affine on each cell and converge uniformly to the relatively convex function 00, but have no affine lower support at 00: its slope would have to be both at least ϵ\epsilon and at most −ϵ-\epsilon. Small uniform errors therefore do not control the support condition at a join. In the boundary-extension proof, the constants CC provide domination but no bound on the normal slopes. A relative-convex version needs such control, or another argument preserving supports across the old faces. Section 11.20 supplies that slope control for compatible polyhedral geometry. The value-only cellwise result by itself is not used to infer nef residues or original smooth psh approximation.

Analytic use and remaining limits

To use these results on a fixed ordinary unit chart, retain the hypotheses of Proposition 11.11.5 and Theorem 11.11.6: the ground field is nontrivially valued, Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|, and the analytic space is good, strictly analytic and pure-dimensional. Suppose the chart image is contained in SS, and F∈C(Π)∩WΓ(Π)F\in C(\Pi)\cap W_\Gamma(\Pi). Theorem 11.19.4 gives a decreasing sequence on the whole same SS, with uniform error. Pullback therefore gives actual Q-PL functions on the whole same ordinary chart domain, with nef residues by the cited inputs. If the integral-projective hypotheses of Corollary 11.11.7 also hold, that corollary supplies an original smooth decreasing sequence near each point on one fixed ordinary neighbourhood.

The same analytic conclusion follows from Theorem 11.19.8 when the original complex has rational normals and value-group constants and F∈C(Π)∩VΓ(Π)F\in C(\Pi)\cap V_\Gamma(\Pi). Its refinement changes neither the tropical support nor the ordinary inverse-image domain. In contrast, Example 11.19.9 has not been realized as an analytic tropical chart, so no analytic conclusion is inferred from that example alone.

This application uses the existing model and residue comparisons; a formal real PL expression is not substituted for actual Q-PL membership. In particular, [GRPSH] (Theorem 4.3) assumes an actual R-PL function and nontrivial valuation before identifying classical psh with semipositivity. The real coordinate calculations above do not rescale the absolute value, extend the ground field, replace original CLD tests, or enlarge their actual supports.

The distinction from [APW26] (Proposition 5.21) remains essential. That proposition produces locally uniform real PL approximants on one neighbourhood; its proof passes through a projective compactification and functions (φi+h)/C(\varphi_i+h)/C. Neither a common finite subdivision for all approximants nor the coefficient and strict-support conditions of this section are supplied by that statement. Our criteria can be applied to an individual output only after those data are checked, or to a sequence after the domains and error choices are made compatible. We have not proved that every relevant analytic tropical image admits the certificate of Proposition 11.19.6. Thus the general relative coefficient replacement, Wu Theorem A.3 and the stronger CLD25 closure comparison remain open here. The primary full text of the separate mixed-calculus Theorem 6.13 has still not been obtained; it is not an input to any proof in this section.

References. Finite-dimensional reconstruction in this manuscript, extending the continuity-only note of v0.21. Source scope checked against APW26 Sections 2.1, 5.2 and Proposition 5.21 (v1), and GRPSH Theorem 4.3 (v2). The strict-support certificate and closure proofs are given above; no general source theorem is declared corrected or disproved by them.

11.20 Controlled slopes and moving relative approximation

The value-only extension in Section 11.19.6 leaves the normal slopes uncontrolled. Here the fixed geometry of the original cells supplies the missing estimate. We first strengthen the extension lemma, then show that small cellwise Lipschitz errors preserve strict relative supports. A fixed strict barrier reduces a general relatively supported PL target to that case. The resulting approximants may change subdivision, but never change the compact support.

A boundary extension with controlled slopes

Use the standard Euclidean norm in the fixed ambient space. For a function on a compact convex cell σ\sigma, write Lip⁡σ\operatorname{Lip}_\sigma for its Lipschitz seminorm, and set this seminorm to zero on a point. All constants below depend on the original finite polytopes, not on subdivisions of their boundary data.

Lemma 11.20.1. (Controlled convex extension).

Let QQ be as in Lemma 11.19.12. There is MQ≥1M_Q\geq 1 with the following property. Let AA be real affine, let ϵ>0\epsilon>0, and let gg be continuous on ∂Q\partial Q, a finite maximum of rational-slope, ΓQ\Gamma_{\mathbf Q}-constant affine functions on each facet. Suppose 0≤A−g≤ϵon ∂Q,Lip⁡τ(g−A)≤ϵfor every facet τ.0\leq A-g\leq\epsilon\quad\text{on }\partial Q, \qquad \operatorname{Lip}_\tau(g-A)\leq\epsilon \quad\text{for every facet }\tau. Then gg has a finite-max extension GG with the same coefficient conditions and G∣∂Q=g,0≤A−G≤MQϵ,Lip⁡Q(G−A)≤MQϵ.\begin{equation} G|_{\partial Q}=g,\qquad 0\leq A-G\leq M_Q\epsilon,\qquad \operatorname{Lip}_Q(G-A)\leq M_Q\epsilon. \end{equation}

Proof. Fix a facet ρ\rho and a compatible affine barrier hρ≥0h_\rho\geq 0 vanishing exactly on ρ\rho. In a finite-max representation of g∣ρg|_\rho, discard branches never active on a relatively open subset. This does not change the function, including at its boundary, by continuity. For a remaining branch ℓ\ell, the tangential part of ∇ℓ−∇A\nabla\ell-\nabla A on ρ\rho has norm at most ϵ\epsilon. The tangent space of ρ\rho is rational. Its orthogonal complement therefore has dense rational points. Add to ℓ\ell a function q⋅(t−vρ)q\cdot(t-v_\rho), where vρ∈ΓQrv_\rho\in\Gamma_{\mathbf Q}^r is a vertex and qq is rational, orthogonal to the tangent space, chosen so that the new branch ℓ~\widetilde\ell satisfies ℓ~∣ρ=ℓ∣ρ,∥∇ℓ~−∇A∥≤2ϵ.\widetilde\ell|_\rho=\ell|_\rho, \qquad \|\nabla\widetilde\ell-\nabla A\|\leq2\epsilon. The added constant is in ΓQ\Gamma_{\mathbf Q}. For a zero-dimensional facet the same construction uses its single boundary value and no tangential constraint. At a point where the original branch is active, ∣ℓ~−A∣≤ϵ|\widetilde\ell-A|\leq\epsilon. If D=diam⁡(Q)D=\operatorname{diam}(Q), it follows that ∣ℓ~−A∣≤(1+2D)ϵ|\widetilde\ell-A|\leq(1+2D)\epsilon on QQ.

We next bound the penalty needed to dominate every other facet. If τ∩ρ≠∅\tau\cap\rho\ne\varnothing, then for every y∈τy\in\tau one can choose z∈τ∩ρz\in\tau\cap\rho such that ∥y−z∥≤Kρτhρ(y),\begin{equation} \|y-z\|\leq K_{\rho\tau}h_\rho(y), \end{equation} with a fixed finite constant. Indeed, write yy as a convex combination of the vertices of τ\tau, keep the vertices in ρ\rho, and replace all other vertices by one fixed vertex z0∈τ∩ρz_0\in\tau\cap\rho. One may take the largest ratio ∥v−z0∥/hρ(v)\|v-z_0\|/h_\rho(v) over vertices outside ρ\rho. Since ℓ~(z)≤g(z)\widetilde\ell(z)\leq g(z) and the function ℓ~−g\widetilde\ell-g is 3ϵ3\epsilon-Lipschitz on τ\tau, ℓ~(y)−g(y)≤3Kρτϵhρ(y).\widetilde\ell(y)-g(y)\leq3K_{\rho\tau}\epsilon h_\rho(y). If the two facets are disjoint, hρh_\rho has a positive minimum on τ\tau, and the uniform bounds above give the same inequality with another fixed constant. Thus, independently of the branch and the boundary subdivision, there is Kρ≥0K_\rho\geq 0 such that every modified branch is dominated on the boundary after subtracting CρhρC_\rho h_\rho, where CρC_\rho is any rational number in [Kρϵ,(Kρ+1)ϵ][K_\rho\epsilon,(K_\rho+1)\epsilon].

Take the maximum of these penalized branches over all facets. It extends gg exactly and retains the coefficient conditions. Every branch has gradient within M1ϵM_1\epsilon of ∇A\nabla A, where M1M_1 depends only on the chosen facet barriers. Hence G−AG-A, a maximum of these affine differences, is M1ϵM_1\epsilon-Lipschitz on QQ. Each branch is at most AA at the vertices, so G≤AG\leq A throughout QQ. Comparing with any boundary point gives A−G≤(1+M1D)ϵA-G\leq(1+M_1D)\epsilon. Taking MQ=max⁡{1,M1,1+M1D}M_Q=\max\{1,M_1,1+M_1D\} proves the lemma. ◻

Corollary 11.20.2. (Cellwise Lipschitz approximation). For a compatible finite compact complex Π\Pi, every F∈V(Π)F\in V(\Pi) and every δ>0\delta>0 admit a continuous coefficient-compatible PL function GG, convex on each original cell, such that 0≤F−G≤δ,max⁡σ∈ΠLip⁡σ(G−F)≤δ.\begin{equation} 0\leq F-G\leq\delta,\qquad \max_{\sigma\in\Pi}\operatorname{Lip}_\sigma(G-F)\leq\delta. \end{equation} Its subdivision may depend on δ\delta.

Proof. Repeat the skeleton induction of Proposition 11.19.13, now using Lemma 11.20.1. Let M≥1M\geq 1 bound its constants for all positive-dimensional cells and let d=dim⁡Πd=\dim\Pi. Choose the vertex errors at most δ/Md\delta/M^d. At step kk, both the uniform error and the facetwise Lipschitz error are at most Mk−1δ/MdM^{k-1}\delta/M^d, and the lemma bounds both errors on the new cell by Mkδ/MdM^k\delta/M^d. The functions agree on shared faces by construction. The finite-max presentations give rational slopes and value-group constants on a finite refinement. If d=0d=0, the vertex choices alone suffice. ◻

Preserving the relative supports

Lemma 11.20.3. (Stability near a strict PL function). Let F0∈V(Π)F_0\in V(\Pi) have strict face supports. There is δ0>0\delta_0>0 such that every continuous GG, a finite maximum of real affine functions on each cell, has local relative affine lower supports whenever max⁡σ∈ΠLip⁡σ(G−F0)≤δ0\max_{\sigma\in\Pi}\operatorname{Lip}_\sigma(G-F_0)\leq\delta_0 holds. More quantitatively, if F0=F+tHF_0=F+tH with F∈C(Π)F\in C(\Pi) and fixed HH having strict face supports, one may take δ0=t/(4KH)\delta_0=t/(4K_H) for a constant KH≥1K_H\geq 1 independent of FF and t>0t>0.

Proof. Choose the strict support aρa_\rho for a face ρ\rho. On any coface σ\sigma, put Dρσ=F0−aρD_{\rho\sigma}=F_0-a_\rho. It is affine and nonnegative, with zero set exactly ρ\rho. The vertex replacement argument used in (11.20.2) gives, for every y∈σy\in\sigma, a point z∈ρz\in\rho with ∥y−z∥≤KDρσ(y)\|y-z\|\leq K D_{\rho\sigma}(y), for one finite K≥1K\geq 1 over all proper incidences.

Take x∈relint⁡(ρ)x\in\operatorname{relint}(\rho) and an active affine branch supporting G∣ρG|_\rho at xx. Its tangential gradient differs from that of F0∣ρF_0|_\rho by a vector ee of norm at most δ:=max⁡σLip⁡σ(G−F0)\delta:=\max_\sigma\operatorname{Lip}_\sigma(G-F_0). This follows by testing the supporting inequality in both tangent directions near xx. Regard ee as a tangent vector in the ambient Euclidean space and set b(y)=aρ(y)+e⋅(y−x)+G(x)−F0(x).b(y)=a_\rho(y)+e\cdot(y-x)+G(x)-F_0(x). Then b≤Gb\leq G on ρ\rho and b(x)=G(x)b(x)=G(x). For y∈σ⊇ρy\in\sigma\supseteq\rho, use the point zz above and the Lipschitz estimate within σ\sigma to obtain G(y)−b(y)=(G−F0)(y)−(G−F0)(z)+G(z)−b(z)+Dρσ(y)−e⋅(y−z)≥Dρσ(y)−2δ∥y−z∥≥(1−2δK)Dρσ(y).\begin{align*} G(y)-b(y) &= (G-F_0)(y)-(G-F_0)(z)+G(z)-b(z)\\ &\hspace{1em}+D_{\rho\sigma}(y)-e\cdot(y-z)\\ &\geq D_{\rho\sigma}(y)-2\delta\|y-z\|\\ &\geq (1-2\delta K)D_{\rho\sigma}(y). \end{align*} Thus δ≤1/(4K)\delta\leq 1/(4K) suffices. A small neighbourhood of xx meets only cofaces of ρ\rho, so bb is the required relative support. If ρ\rho is maximal, the active branch alone suffices.

For the quantitative statement, add a support of FF to tt times a strict support of HH. The gap on each coface is at least tt times the corresponding gap for HH. The same vertex replacement construction therefore gives the distance bound with constant KH/tK_H/t, which may itself be less than one. This proves the stated bound without using a uniform bound on the values of G−F0G-F_0. ◻

Theorem 11.20.4. (Relative rational approximation with moving subdivisions). Let Γ≠0\Gamma\ne 0 and let Π\Pi be a finite compact polytopal complex whose cells are defined by rational-slope, ΓQ\Gamma_{\mathbf Q}-constant affine inequalities. Every continuous cellwise-affine FF with local relative affine lower supports has continuous coefficient-compatible PL approximants FjF_j on the same support S=∣Π∣S=|\Pi|, themselves admitting such supports, with Fj+1≤Fj,0≤Fj−F≤3ρ8−j(ρ∈Γ>0 fixed).\begin{equation} F_{j+1}\leq F_j,\qquad 0\leq F_j-F\leq3\rho8^{-j} \quad(\rho\in\Gamma_{>0}\text{ fixed}). \end{equation} The subdivisions of the approximants may vary with jj. No continuity-closure condition F∈VΓ(Π)F\in V_\Gamma(\Pi) is required.

Proof. Use the fixed arrangement refinement Π∗\Pi^* and strict function HH from Theorem 11.19.8. The function FF is still affine on each refined cell and has relative supports. Set B=max⁡{1,∥H∥∞}B=\max\{1,\|H\|_\infty\} and ηj=ρ8−j\eta_j=\rho 8^{-j}. Choose tj=ηj2B,0<δj≤min⁡{ηj2,tj4KH}.t_j=\frac{\eta_j}{2B},\qquad 0<\delta_j\leq\min\left\{\frac{\eta_j}{2},\frac{t_j}{4K_H}\right\}. Apply Corollary 11.20.2 to the real PL target F+tjHF+t_jH with error δj\delta_j, obtaining GjG_j. Lemma 11.20.3 gives relative affine supports for GjG_j, and ∥Gj−F∥∞≤δj+tj∥H∥∞≤ηj.\|G_j-F\|_\infty\leq\delta_j+t_j\|H\|_\infty\leq\eta_j. Put Fj=Gj+2ηjF_j=G_j+2\eta_j. This is an allowed value-group shift, preserving the supports and coefficients. The inequalities F+ηj≤Fj≤F+3ηjF+\eta_j\leq F_j\leq F+3\eta_j imply (11.20.4). Every construction takes place on the same SS. The arrangement refinement is fixed once; only the further finite-max subdivisions used to build GjG_j vary. ◻

The same proof applies termwise to a decreasing sequence of such real PL targets uju_j on the same compact support, even if their compatible complexes differ. Choose supported rational GjG_j with ∣Gj−uj∣≤ηj|G_j-u_j|\leq\eta_j and add 2ηj2\eta_j. Then Gj+1+2ηj+1≤uj+1+3ηj+1≤uj+ηj≤Gj+2ηjG_{j+1}+2\eta_{j+1}\leq u_{j+1}+3\eta_{j+1}\leq u_j+\eta_j\leq G_j+2\eta_j. The limit is the original pointwise limit of uju_j; uniform convergence is claimed only when the targets themselves converge uniformly.

The source interface still to be checked

On an actual ordinary unit chart, the analytic consequences in Section 11.19.7 now apply to every relatively supported real PL target whose finite compact complex has the stated coefficient-compatible geometry. Actual Q-PL identification, nontrivial valuation, nef-residue comparison and integral-projective original-smooth transfer remain separate cited inputs. The arithmetic normalization and original tests are unchanged.

This does not yet cover the arbitrary real intercepts allowed by APW26, nor an arbitrary smooth target without a suitable PL approximation on a common domain. The argument also does not extend the current calculus to arbitrary positive backgrounds or prove first-stage closure. These distinctions are part of the theorem: the new result removes the fixed-subdivision obstruction by controlled moving subdivisions within the compatible-polyhedral class. It is an elementary reconstruction assembled from the explicit extension and support estimates above, not a claim that an unverified source theorem has been proved in its full generality.

Section 11.21 goes further for rational-direction PL germs and for families with an explicit common transport domain. Section 11.21.4 isolates an earlier issue in the literal APW26 PL convention; the gap is not resolved merely by replacing real intercepts with value-group constants.

11.21 Real intercepts, moving germs and common domains

A real intercept can be an obstruction to keeping a subdivision fixed without being an obstruction to approximating the function. There is a local way to remove it: move the centre of a PL star inside the stratum of the original support. The support remains fixed, whereas the break loci move. We prove this germ statement first. We then give a separate transport criterion that works on a prescribed compact set, and explain why the germ statement alone cannot be applied termwise to a source approximation on one neighbourhood.

Section 11.22 gives a subsequent replacement for the fixed-domain approximation question: continuous relative supports suffice, without rational directions or a one-scale input subdivision. The present constructions retain their explicit geometric interpretation and illustrate why termwise germ arguments alone do not control a common domain.

Moving a centre within its stratum

Throughout this section Γ⊂R\Gamma\subset\mathbf R is a nonzero additive subgroup. A compatible polyhedron is one defined by rational-slope affine inequalities with constants in ΓQ\Gamma_{\mathbf Q}. A rational polyhedron, without the word compatible, may have arbitrary real constants. All complexes used locally are finite after restriction to a sufficiently small neighbourhood. Relative supports always mean ambient affine lower supports in the relative topology of the specified support.

Lemma 11.21.1. (Compatible points in a face). If τ\tau is a nonempty compatible polyhedron, then relint⁡(τ)∩ΓQr\operatorname{relint}(\tau)\cap\Gamma_{\mathbf Q}^r is dense in relint⁡(τ)\operatorname{relint}(\tau). Its tangent space L=TτL=T\tau is rational.

Proof. The affine hull is specified by those rational equations that hold identically on τ\tau. Gaussian elimination gives a particular solution a∈ΓQra\in\Gamma_{\mathbf Q}^r and a basis e1,…,ed∈Qre_1,\ldots,e_d\in\mathbf Q^r for the homogeneous solutions. Thus aff⁡(τ)=a+∑kRek\operatorname{aff}(\tau)=a+\sum_k\mathbf R e_k. Since ΓQ\Gamma_{\mathbf Q} contains ρQ\rho\mathbf Q for some ρ>0\rho>0, points a+∑kbkeka+\sum_k b_ke_k with bk∈ΓQb_k\in\Gamma_{\mathbf Q} are dense in this affine hull and have compatible coordinates. Intersect with the relatively open set relint⁡(τ)\operatorname{relint}(\tau). The case d=0d=0 is included. ◻

Lemma 11.21.2. (A fixed support near a stratum). Let SS be the support of a locally finite compatible complex, and let x∈relint⁡(τ)x\in\operatorname{relint}(\tau). There are R>0R>0 and a finite union CC of rational polyhedral cones such that S∩B(x,R)=(x+C)∩B(x,R),C+L=C,L=Tτ.S\cap B(x,R)=(x+C)\cap B(x,R),\qquad C+L=C, \quad L=T\tau. In particular, b+C=x+Cb+C=x+C for every b∈aff⁡(τ)b\in\operatorname{aff}(\tau). If FF is continuous and affine on a finite rational polyhedral subdivision near xx, then, after decreasing RR, F(x+v)=F(x)+f(v),v∈C,∥v∥<R,\begin{equation} F(x+v)=F(x)+f(v),\qquad v\in C,\quad\|v\|<R, \end{equation} where ff is continuous and homogeneous of degree one, affine on a finite rational conical subdivision of CC. If FF has relative supports near xx, then ff has relative supports on all of CC.

Proof. Only cells containing xx meet a sufficiently small ball. In each such cell, inequalities that are strict at xx remain strict in a smaller ball; the active inequalities become homogeneous after translating xx to the origin. Their normals are rational, and vanish on LL, because they vanish on the face τ\tau. Their tangent cones form the asserted union CC, invariant under both signs of every vector in LL. This also proves the equality after changing the centre to bb.

Apply the same active-constraint construction to the finite rational subdivision of FF, together with the original complex. On a cell containing xx, its affine expression becomes F(x)+pσ⋅vF(x)+p_\sigma\cdot v. The expressions agree on overlaps, first near zero and then everywhere by homogeneity. They define ff on CC. To transfer a relative support to a point v∈Cv\in C, choose s>0s>0 so small that svsv and a neighbourhood of it lie in the ball where (11.21.1) holds. A support aa for ff near svsv gives the support s−1a(s ⋅)s^{-1}a(s\,\cdot) near vv. At zero use the given support directly. ◻

Theorem 11.21.3. (Rational approximation of a real-intercept PL germ). Let SS be locally the support of a compatible complex and let FF be continuous, affine on a locally finite rational polyhedral subdivision of SS, with local relative affine lower supports. For every x∈Sx\in S there is a compact neighbourhood KK of xx in SS and continuous coefficient-compatible PL functions FjF_j on KK, themselves having relative supports, such that F+ηj≤Fj≤F+3ηj,Fj+1≤Fj,ηj=ρ8−j,ρ∈Γ>0.\begin{equation} F+\eta_j\leq F_j\leq F+3\eta_j,\qquad F_{j+1}\leq F_j,\qquad \eta_j=\rho8^{-j},\quad\rho\in\Gamma_{>0}. \end{equation} The intercepts of the subdivision of FF need not belong to ΓQ\Gamma_{\mathbf Q}. The neighbourhood KK depends on the germ FF; the theorem makes no uniform-domain assertion for a family of different germs.

Proof. Use Lemma 11.21.2. Choose a closed box BB with coordinate endpoints in ΓQ\Gamma_{\mathbf Q} such that x∈int⁡(B)x\in\operatorname{int}(B) and B⊂B(x,R/4)B\subset B(x,R/4), and put K=S∩B=(x+C)∩BK=S\cap B=(x+C)\cap B. For b∈relint⁡(τ)∩ΓQrb\in\operatorname{relint}(\tau)\cap\Gamma_{\mathbf Q}^r close to xx, define Hb(t)=F(x)+f(t−b),t∈K.H_b(t)=F(x)+f(t-b),\qquad t\in K. The equality b+C=x+Cb+C=x+C makes this a function on exactly the same KK. It has relative supports, and its cells are intersections of BB with translates by bb of rational cones. These cells are compatible. Thus HbH_b satisfies the hypotheses of Theorem 11.20.4, although its function coefficients can still be real.

There is a constant M<∞M<\infty with ∣Hb−F∣≤M∥b−x∥|H_b-F|\leq M\|b-x\| on KK. Indeed, the segment from t−xt-x to t−bt-b lies in CC, since its direction belongs to LL. Partition that segment at the finitely many conical cells of ff. On each piece its slope is bounded by the norm of a chosen ambient slope of that cell; take MM to be the largest of these norms. Lemma 11.21.1 now gives bjb_j so close to xx that this error is at most ηj/2\eta_j/2.

Theorem 11.20.4 supplies a supported coefficient-compatible GjG_j on KK with ∥Gj−Hbj∥∞≤ηj/2\|G_j-H_{b_j}\|_\infty\leq\eta_j/2. Hence ∥Gj−F∥∞≤ηj\|G_j-F\|_\infty\leq\eta_j. Set Fj=Gj+2ηjF_j=G_j+2\eta_j. The shift is in ΓQ\Gamma_{\mathbf Q} and gives (11.21.2), including decrease because 3ηj+1<ηj3\eta_{j+1}<\eta_j. No affine map of an analytic space has been introduced: only the real function used in a fixed set of coordinates has changed. ◻

Transport on a prescribed compact set

For a single germ the preceding construction is enough. For a sequence, a useful alternative is to move an entire one-scale presentation on a domain fixed in advance. The scale below is a real geometric parameter; it is not the normalization of a field valuation.

Theorem 11.21.4. (One-scale transport). Let CC be a finite union of rational cones, let L⊂RrL\subset\mathbf R^r be a rational subspace with C+L=CC+L=C, and let a∈ΓQra\in\Gamma_{\mathbf Q}^r. Let UU be relatively open in S=a+CS=a+C. Let BB be a closed compatible box with nonempty K=S∩B⊂UK=S\cap B\subset U. Suppose FF is continuous, has relative supports on UU, and is affine on a finite polyhedral structure in a neighbourhood of KK, whose cells are defined by inequalities q⋅(t−a−w)≥λc,q∈Qr,c∈Q,w∈L,λ>0.\begin{equation} q\cdot(t-a-w)\geq\lambda c, \qquad q\in\mathbf Q^r,\quad c\in\mathbf Q, \quad w\in L,\quad\lambda>0. \end{equation} Here the same ww and λ\lambda are used for all cells, and equalities are allowed. Then F∣KF|_K has decreasing coefficient-compatible relative PL approximants satisfying (11.21.2) on this prescribed KK. Neither ww nor λ\lambda is required to have value-group coordinates.

Proof. Choose wj∈L∩ΓQrw_j\in L\cap\Gamma_{\mathbf Q}^r and λj∈ΓQ,>0\lambda_j\in\Gamma_{\mathbf Q,>0} tending to ww and λ\lambda, respectively, and define Tj(t)=a+wj+λjλ(t−a−w).T_j(t)=a+w_j+\frac{\lambda_j}{\lambda}(t-a-w). The maps TjT_j preserve SS bijectively and converge to the identity, as do their inverses, uniformly on compact sets. Choose a relatively open neighbourhood NN of KK with compact closure in UU, contained in the neighbourhood where the stated finite presentation is valid. Sufficiently close choices ensure Tj−1(K)⊂NT_j^{-1}(K)\subset N. On KK put Hj=F∘Tj−1H_j=F\circ T_j^{-1}. Supports are transported by this invertible affine map in the real coordinate space. The transformed cell inequalities are q⋅(t−a−wj)≥λjc.q\cdot(t-a-w_j)\geq\lambda_j c. They are compatible, since only rational numbers multiply elements of ΓQ\Gamma_{\mathbf Q}. Their intersections with BB form a compatible compact complex with support exactly KK. Uniform continuity of FF on N‾\overline N allows the choices to satisfy ∥Hj−F∥K≤ηj/2\|H_j-F\|_K\leq\eta_j/2. Apply Theorem 11.20.4 with error ηj/2\eta_j/2, and add 2ηj2\eta_j as in the previous proof. ◻

Corollary 11.21.5. (An edge-connected simplicial certificate). In Theorem 11.21.4, the one-scale presentation condition follows if FF is affine on a finite compact simplicial complex covering a neighbourhood of KK in SS, every maximal simplex has dimension at least two, every edge has rational direction, the maximal simplices are connected by shared edges, and one vertex pp belongs to a+La+L. The vertices need not have value-group coordinates.

Proof. Use the geometric part of the proof of Proposition 11.19.10, before its value-group step. It gives one positive real λ\lambda such that all vertices lie in p+λQrp+\lambda\mathbf Q^r. Indeed, within a simplex the rational directions of its edges force the edge scales to have rational ratios; a shared nonzero edge propagates the ratio between simplices. Each simplex is therefore p+λp+\lambda times a polytope with rational vertices, allowing lower-dimensional affine-hull equations. Its defining inequalities have the form (11.21.3), with w=p−a∈Lw=p-a\in L. Apply Theorem 11.21.4. This is a consequence of the earlier scale-propagation proof and the new transport argument, rather than a separate density theorem. ◻

This certificate is deliberately checkable. It does not apply merely because a source uses the word simplicial: unbounded polyhedra, pieces joined only at vertices, one-dimensional cells, and a missing lineality anchor require additional arguments. In an ambient vector space one may take L=RrL=\mathbf R^r, so the anchor condition is automatic; on a branching support it is a real restriction.

Corollary 11.21.6. (Families with a common transport domain). Let uju_j decrease pointwise on the same U⊂SU\subset S. Suppose each uju_j satisfies the hypotheses of Theorem 11.21.4 near the same KK, with its own finite structure, scale and lineality translation. Then there are supported coefficient-compatible PL functions vjv_j on KK with uj+ηj≤vj≤uj+3ηj,vj+1≤vj.u_j+\eta_j\leq v_j\leq u_j+3\eta_j,\qquad v_{j+1}\leq v_j. The pointwise limit is unchanged. If instead uju_j converges uniformly to a finite function uu on KK, without monotonicity, a subsequence and the same error-shift construction give decreasing supported coefficient-compatible PL approximants to uu on KK.

Proof. For the first assertion choose GjG_j with ∥Gj−uj∥K≤ηj\|G_j-u_j\|_K\leq\eta_j and set vj=Gj+2ηjv_j=G_j+2\eta_j. Then vj+1≤uj+1+3ηj+1≤uj+ηj≤vjv_{j+1}\leq u_{j+1}+3\eta_{j+1}\leq u_j+\eta_j\leq v_j. For the second, select ukju_{k_j} with ∥ukj−u∥K≤ηj/2\|u_{k_j}-u\|_K\leq\eta_j/2 and approximate that term within ηj/2\eta_j/2. The resulting GjG_j lies within ηj\eta_j of uu, so the same shift proves the claim. Parameters may be chosen separately for each target; no uniform slope bound or common subdivision is used. ◻

Why the germ quantifiers matter

Example 11.21.7. (Shrinking stars despite uniform control). Take Γ=Z\Gamma=\mathbf Z and let pjp_j be the secant interpolation of t2t^2 at the dyadic nodes of [−1,1][-1,1], with mesh hj=2−jh_j=2^{-j}. On [khj,(k+1)hj][kh_j,(k+1)h_j] one has pj(t)−t2=(t−khj)((k+1)hj−t),0≤pj(t)−t2≤hj2/4.p_j(t)-t^2=(t-kh_j)((k+1)h_j-t), \qquad 0\leq p_j(t)-t^2\leq h_j^2/4. The slopes (2k+1)hj(2k+1)h_j increase with kk, so pjp_j is convex. Refining each interval inserts a point of the convex graph below its old chord, hence pj+1≤pjp_{j+1}\leq p_j. Also Lip⁡(pj)≤2\operatorname{Lip}(p_j)\leq 2. Nevertheless, the maximal centred interval on which pj(t)=pj(0)+fj(t)p_j(t)=p_j(0)+f_j(t) for a degree-one homogeneous PL function fjf_j is [−hj,hj][-h_j,h_j]: there pj=hj∣t∣p_j=h_j|t|, and the slope changes at ±hj\pm h_j. Thus no positive-radius star works for all jj. Pullback by the first-coordinate projection gives the same phenomenon in every positive dimension.

This is not a failure of approximation: the functions pjp_j are already coefficient-compatible, convex and decreasing uniformly to a smooth convex target. It disproves only the inference from termwise PL stars, uniform convergence and a uniform Lipschitz bound to a common star radius. The domain in Theorem 11.21.3 therefore cannot simply be intersected over the approximants furnished by a source theorem.

Corollary 11.21.8. (Original smooth approximation of a rational-direction PL germ). Let KK be complete and nontrivially valued, let XX be an integral projective KK-variety of positive dimension, and let W⊂XanW\subset X^{\rm an} be ordinary open. On WW fix analytic units g1,…,grg_1,\ldots,g_r and the map τ=(−log⁡∣g1∣,…,−log⁡∣gr∣)\tau=(-\log|g_1|,\ldots,-\log|g_r|). Let x∈Wx\in W. Suppose near τ(x)\tau(x) that the actual image is contained in a support SS of a locally finite ΓQ\Gamma_{\mathbf Q}-compatible complex, where Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|, and u=F∘τ,u=F\circ\tau, where FF is finite continuous, affine on a locally finite rational subdivision of SS, and has local relative affine lower supports. Then some ordinary open neighbourhood V⊂WV\subset W of xx carries original smooth psh functions bj↓u∣Vb_j\downarrow u|_V. Convergence is uniform on compact subsets of VV.

Open proof obligation (RC2). Prove the analytic germ bridge using the required curvewise-to-model implication; GRPSH Theorem 4.3.

NAG-900226 · Proof status

For a family on a prescribed compatible box, Corollary 11.21.6 can replace the germ theorem in this proof, after its geometric presentation conditions are checked term by term. Neither route changes the chosen valuation, extends the ground field, enlarges the support of an actual analytic test, or replaces original CLD tests. The auxiliary real support SS must be specified in the hypothesis; no unsupported extension of FF from the actual image is being asserted. These are consequences of the displayed real constructions and the named model and smoothing inputs, not a full comparison with the generated source classes.

The PL convention in the source regularization

There is a logically earlier issue than the value-group constants. In the fixed version [APW26] (Section 1.6; Sections 2.1, 2.1.4–2.1.6), affine functions have real coefficients, rational half-spaces have integral normals and arbitrary real intercepts, and polyhedral structures are declared rational and simplicial unless specified otherwise. The last step of Proposition 5.12, printed page 38, invokes a continuous finite-selection argument to conclude PL membership. The following two examples explain why that step needs an explicit convention or an additional geometric argument. They do not disprove the existence statements of Propositions 5.12 or 5.21.

The examples below concern finite or locally finite subdivisions, as explicitly stated. The displayed definition in the source Section 2.1.6 does not itself list local finiteness. This convention also needs clarification when that definition is used: we do not infer a statement about arbitrary non-locally-finite decompositions from the finite-boundary argument below.

Example 11.21.9. (A maximum with an irrational break direction). On R2\mathbf R^2, the affine functions ℓ1(x,y)=x\ell_1(x,y)=x and ℓ2(x,y)=2 y\ell_2(x,y)=\sqrt 2\,y are affine on every rational triangulation. Their maximum m(x,y)=max⁡{x,2 y}m(x,y)=\max\{x,\sqrt2\,y\} is not affine on the cells of any locally finite rational polyhedral structure near the origin. Indeed, each full-dimensional affine cell must lie on one side of the break line D={x=2 y}D=\{x=\sqrt 2\,y\}. Hence an open segment of DD would lie in the finite union of rational supporting lines of the cells near the origin. One such line would have to contain an open segment of DD, and therefore equal DD. But DD has no nonzero rational tangent vector. This is a contradiction. The function mm is nevertheless an ordinary real-polyhedral PL function and is ambient convex. Products with Rr−2\mathbf R^{r-2} give the same distinction in every dimension r≥2r\geq 2.

Proposition 11.21.10. (Uniform density does not give Lipschitz-small errors). Let m(x,y)=max⁡{x,2 y}m(x,y)=\max\{x,\sqrt 2\,y\} and let BB be a closed ball of positive radius centred at the origin. If GG is continuous and affine on a locally finite rational polyhedral subdivision of a neighbourhood of BB, then Lip⁡B(G−m)≥32.\operatorname{Lip}_{B}(G-m)\geq\frac{\sqrt3}{2}. This holds even when the affine coefficients of GG are real. Nevertheless, rational numbers qj→2q_j\to\sqrt 2 give coefficient-compatible convex functions max⁡{x,qjy}\max\{x,q_jy\} converging uniformly to mm on BB.

Proof. A nontrivial segment of the irrational line D={x=2 y}D=\{x=\sqrt 2\,y\} inside BB cannot be covered by the finitely many rational boundary lines of the subdivision. Choose a point zz on that segment in the interior of a full-dimensional cell. Near zz, the function GG is one affine function with gradient pp. On the two open sides of DD, the gradients of G−mG-m are respectively p−(1,0)p-(1,0) and p−(0,2)p-(0,\sqrt 2). Its Lipschitz seminorm on BB bounds both norms, by evaluating differences in small line segments lying in the corresponding open side. The triangle inequality gives 3=∥(1,0)−(0,2)∥≤∥p−(1,0)∥+∥p−(0,2)∥≤2Lip⁡B(G−m).\sqrt3=\|(1,0)-(0,\sqrt2)\| \leq\|p-(1,0)\|+\|p-(0,\sqrt2)\| \leq2\operatorname{Lip}_{B}(G-m). For uniform approximation, the maximum operation is 11-Lipschitz in the sup norm of its two entries, so the error is at most ∣qj−2∣max⁡B∣y∣|q_j-\sqrt 2|\max_B|y|. The same obstruction holds in every ambient dimension at least two by restricting to the rational plane where the other coordinates vanish. The lower bound is sharp among rational-subdivision PL functions with real affine coefficients: the single affine function G=(x+2 y)/2G=(x+\sqrt 2\,y)/2 has G−m=−∣x−2 y∣/2G-m=-|x-\sqrt 2\,y|/2 and Lipschitz error 3/2\sqrt 3/2. Among coefficient-compatible functions the infimum is the same, by approximating this affine gradient by rational vectors; an allowed constant can be chosen independently because the Lipschitz seminorm ignores constants. ◻

Thus the cellwise Lipschitz estimates used in Section 11.20 cannot simply be demanded for arbitrary real-polyhedral targets on a preassigned rational complex. A rational subdivision of the input was a substantive hypothesis of the germ construction. This obstruction concerns the stronger error norm only; it neither obstructs uniform rational approximation nor disproves the source regularization statement.

The same proof applies to the maximum of two affine functions with gradients p1,p2p_1,p_2 whenever their break hyperplane cannot be a facet in the permitted subdivision class; the bound is ∥p1−p2∥/2\|p_1-p_2\|/2. This also distinguishes value-group intercepts from rational directions. For example, with Γ=Z\Gamma=\mathbf Z, on [−2,2][-2,2] every coefficient-compatible PL approximation GG to ∣t−2∣|t-\sqrt 2| has Lip⁡(G−∣t−2∣)≥1\operatorname{Lip}(G-|t-\sqrt 2|)\geq 1: its finitely many rational breakpoints miss 2\sqrt 2, so GG is affine there. Yet ∣t−aj∣|t-a_j| converges uniformly when aj∈Qa_j\in\mathbf Q tends to 2\sqrt 2. Thus moving the centre in Theorem 11.21.3 is compatible with uniform error control, without implying Lipschitz convergence. Products give this intercept phenomenon in every positive dimension.

Example 11.21.11. (Compatible coefficients without rational triangulability). Let Γ=Z+2Z\Gamma=\mathbf Z+\sqrt 2\mathbf Z. Each of the four functions ∣x∣|x|, ∣x−1∣|x-1|, ∣y∣|y| and ∣y−2∣|y-\sqrt 2| admits a rational simplicial conical subdivision of R2\mathbf R^2, translated to a point on its break line. Their sum P(x,y)=∣x∣+∣x−1∣+∣y∣+∣y−2∣P(x,y)=|x|+|x-1|+|y|+|y-\sqrt2| has rational slopes and ΓQ\Gamma_{\mathbf Q} constants on a compatible polyhedral complex. But it admits no locally finite rational simplicial subdivision on a neighbourhood of [0,1]×[0,2][0,1]\times[0,\sqrt 2]. Its gradient changes across each of the four bounding lines. No full-dimensional cell of an affine presentation can cross any of those lines. Thus the cells inside the rectangle would give a finite rational-direction triangulation of it, contradicting Example 11.19.11. All the cells contained in the bounded rectangle are compact simplices, irrespective of how simplicial unbounded polyhedra are defined outside it.

The first example concerns rational directions, independently of any valuation. The second retains rational directions and compatible coefficients, but concerns the extra simplicial requirement. Neither example is a counterexample to approximation by supported rational functions: the first is ambient convex and admits such approximations by Section 11.12.4, while the second is already coefficient-compatible. They identify the category in which a finite-selection proof is valid.

Lemma 11.21.12. (Coefficient-compatible finite selections). On the support of a finite compatible complex, a continuous selection from finitely many affine functions qi⋅t+ciq_i\cdot t+c_i, with qi∈Qrq_i\in\mathbf Q^r and ci∈ΓQc_i\in\Gamma_{\mathbf Q}, is coefficient-compatible PL on a finite compatible refinement. The same conclusion holds for a continuous selection from finitely many coefficient-compatible PL functions. No simplicial refinement is asserted.

Proof. Cut each cell by all equations (qi−qk)⋅t=ck−ci(q_i-q_k)\cdot t=c_k-c_i. These are compatible hyperplanes. On the relative interior of a resulting cell, two branches are either identical or everywhere distinct, with a fixed strict order. A continuous selection must choose one of these distinct branches on each connected relative interior: locally it cannot change to a different branch across a positive value gap. Continuity extends that affine expression to the closed cell. For PL inputs first take their finite common compatible refinement and apply the same argument to their affine branches there. ◻

With arbitrary real affine coefficients the same proof produces a real-polyhedral refinement, whose face directions may be irrational. Thus there are two honest ways to use the last step of the source Proposition 5.12: either specify the unrestricted real-polyhedral PL category and separately prove the later rational/model interface, or establish coefficient and subdivision control during the construction. Merely citing a finite-selection theorem does not supply that control. The germ theorem above assumes a rational subdivision of the input and therefore does not silently cover the first alternative. Section 11.22 supplies the second route for continuous relative-support inputs on a fixed compact set. The companion mixed-calculus Theorem 6.13 remains unverified here because its primary full text has not been obtained.

The primary-file check also confirms the localization correction already proved in Lemma 11.11.11. On the compact neighbourhood used in APW Proposition 5.10, the intersection of the closures of its shrinking sets is {g−ℓg=0}={x}\{g-\ell_g=0\}=\{x\}, not the printed zero set of f−ℓff-\ell_f. This is a recheck of that earlier repair, not an additional result of the present section.

11.22 Rational affine bumps on a fixed compact support

The coefficient problem can be treated before the PL subdivision is produced. The local modification and finite-cover argument of [APW26] (Propositions 5.10 and 5.12) suggest the following replacement: approximate the raised affine supports by coefficient-compatible affine functions, then perform the gluing. A small quadratic term supplies the strict separation needed to absorb that perturbation. The proof below is given in full, including the common domain and the last finite-selection step. It is an elementary reconstruction of that argument with coefficient control, not a claim of a new regularization principle.

Local modifications with controlled coefficients

Fix a nonzero additive subgroup Γ⊂R\Gamma\subset\mathbf R, and set ΓQ=Γ⊗ZQ⊂R\Gamma_{\mathbf Q}=\Gamma\otimes_{\mathbf Z}\mathbf Q\subset\mathbf R. An allowed affine function is a(t)=q⋅t+ca(t)=q\cdot t+c with q∈Qrq\in\mathbf Q^r and c∈ΓQc\in\Gamma_{\mathbf Q}. Let S⊂RrS\subset\mathbf R^r be nonempty and compact. A continuous F:S→RF:S\to\mathbf R has relative affine lower supports if, for every x∈Sx\in S, some ambient affine ℓx\ell_x satisfies ℓx(x)=F(x)\ell_x(x)=F(x) and ℓx≤F\ell_x\leq F on a relative neighbourhood of xx. No polyhedrality of SS is needed until the polyhedral conclusion is stated.

Lemma 11.22.1. (A rational affine local bump). Suppose that g:S→Rg:S\to\mathbf R is continuous and has relative affine lower supports. Fix x∈Sx\in S, r,t>0r,t>0 and an ambient affine ℓ\ell such that ℓ(x)=g(x),g(y)−ℓ(y)≥t∥y−x∥2(y∈S∩B‾(x,2r)).\ell(x)=g(x),\qquad g(y)-\ell(y)\geq t\|y-x\|^2 \quad (y\in S\cap\overline B(x,2r)). For every 0<δ≤tr2/160<\delta\leq tr^2/16 there is an allowed affine function aa and a continuous function b:S→Rb:S\to\mathbf R with relative affine lower supports, satisfying g≤b≤g+5δ/4,b(x)−g(x)≥3δ/4,b=gon S∖B(x,r/2),b(y)>g(y)⟹b(y)=a(y).\begin{gather} g\leq b\leq g+5\delta/4,\qquad b(x)-g(x)\geq3\delta/4,\\ b=g\quad\text{on }S\setminus B(x,r/2),\qquad b(y)>g(y)\Longrightarrow b(y)=a(y).\nonumber \end{gather}

Proof. The groups Q\mathbf Q and ΓQ\Gamma_{\mathbf Q} are dense in R\mathbf R. Approximate the coefficients of ℓ+δ\ell+\delta so closely that ∣a−(ℓ+δ)∣≤δ/4on B‾(x,2r).|a-(\ell+\delta)|\leq\delta/4 \quad\text{on }\overline B(x,2r). This is possible by coefficientwise approximation on a bounded ball; the centre xx and the radius rr need not be compatible. At xx one has a−g≥3δ/4a-g\geq 3\delta/4. On S∩B‾(x,2r)S\cap\overline B(x,2r) one has a−g≤5δ/4a-g\leq 5\delta/4. If also ∥y−x∥≥r/2\|y-x\|\geq r/2, then a(y)−g(y)≤5δ/4−tr2/4<0.a(y)-g(y)\leq 5\delta/4-tr^2/4<0. Define b=max⁡{g,a}b=\max\{g,a\} on S∩B(x,r)S\cap B(x,r) and b=gb=g elsewhere. On an annulus around the patching boundary both formulas equal gg. Hence the definition is continuous and gives (11.22.1). Locally it is either gg or a maximum of gg and an affine function. A support of an active term supports the maximum at that point; this proves the support assertion, including the tie locus. No smooth cutoff is used. ◻

Lemma 11.22.2. (A finite active cover removes the background). Let g,b1,…,bNg,b_1,\ldots,b_N be continuous functions on compact SS, with bi≥gb_i\geq g. Suppose that {bi>g}\{b_i>g\} cover SS, and that bi=aib_i=a_i on {bi>g}\{b_i>g\} for allowed affine functions aia_i. Then P=max⁡ibiP=\max_i b_i is a continuous selection from a1,…,aNa_1,\ldots,a_N. Moreover, near each point of SS, it is the restriction of a maximum of a fixed subset of these affine functions. In particular, it has relative affine lower supports.

Proof. At y∈Sy\in S, one has P(y)>g(y)P(y)>g(y). Every maximizing index therefore has bi(y)=ai(y)b_i(y)=a_i(y). Let Iy={i:bi(y)=P(y)}I_y=\{i:b_i(y)=P(y)\}. For these indices, continuity keeps bi>gb_i>g near yy, so bi=aib_i=a_i there. For the other indices, the strict inequalities bi(y)<P(y)b_i(y)<P(y) persist on a sufficiently small common neighbourhood, because the list is finite. Consequently P=max⁡i∈IyaiP=\max_{i\in I_y}a_i near yy. The covering also gives min⁡S(P−g)>0\min_S(P-g)>0, although no uniform lower bound independent of the approximation is needed. ◻

The local and global assertions in this lemma are different. An affine branch aia_i may exceed PP outside the patch on which it was used; the construction does not give P=max⁡iaiP=\max_i a_i on all of SS. Near a point, however, every active modification is strictly above the background and therefore equals its affine branch. This is the precise reason why the local maximum conclusion survives the gluing.

Coefficient-compatible density and fixed-domain sequences

Theorem 11.22.3. (Density by rational finite selections). Let S⊂RrS\subset\mathbf R^r be nonempty and compact and let F:S→RF:S\to\mathbf R be continuous with relative affine lower supports. For every ε>0\varepsilon>0 there is a continuous function P:S→RP:S\to\mathbf R, locally a finite maximum of allowed affine functions and globally a selection from one finite list of them, such that F≤P≤F+εon the same S.\begin{equation} F\leq P\leq F+\varepsilon\quad\text{on the same }S. \end{equation} If SS is the support of a finite coefficient-compatible polytopal complex, then PP is coefficient-compatible PL on a finite refinement. No rational subdivision, differentiability or PL hypothesis is imposed on FF. The refinement is not asserted to be simplicial.

Proof. Set Q(y)=∥y∥2Q(y)=\|y\|^2 and choose t>0t>0 so that tmax⁡SQ≤ε/4t\max_S Q\leq\varepsilon/4. The function g=F+tQg=F+tQ is continuous and has relative supports. More precisely, if ℓxF\ell_x^F supports FF near xx, then ℓx(y)=ℓxF(y)+t(Q(x)+2x⋅(y−x))\ell_x(y)=\ell_x^F(y)+t\bigl(Q(x)+2x\cdot(y-x)\bigr) satisfies g(y)−ℓx(y)≥t∥y−x∥2g(y)-\ell_x(y)\geq t\|y-x\|^2 in that neighbourhood. Choose rx>0r_x>0 so that this holds on S∩B‾(x,2rx)S\cap\overline B(x,2r_x), and choose 0<δx≤min⁡{ε/4,trx2/16}.0<\delta_x\leq\min\{\varepsilon/4,tr_x^2/16\}. Lemma 11.22.1 gives bxb_x, and the relatively open sets {bx>g}\{b_x>g\} cover SS. Take a finite subcover and set PP equal to the maximum of its modifications. Then F≤g≤P≤g+5ε/16≤F+9ε/16<F+ε.F\leq g\leq P\leq g+5\varepsilon/16 \leq F+9\varepsilon/16<F+\varepsilon. Lemma 11.22.2 gives the local maxima and the global finite selection. If SS is coefficient-compatible polyhedral, apply Lemma 11.21.12: cut by the equations ai=aka_i=a_k, whose slopes are rational and constants lie in ΓQ\Gamma_{\mathbf Q}. The original balls and the possibly nonpolyhedral sets {ai>g}\{a_i>g\} disappear from this final presentation. It is therefore unnecessary to make their radii or boundaries compatible. ◻

Corollary 11.22.4. (Decreasing approximation on the same support). Under the hypotheses of Theorem 11.22.3, fix ρ∈Γ>0\rho\in\Gamma_{>0} and put ηj=ρ8−j\eta_j=\rho 8^{-j}. There are functions PjP_j of the type constructed there such that F+ηj≤Pj≤F+2ηj,Pj+1≤Pj.\begin{equation} F+\eta_j\leq P_j\leq F+2\eta_j,\qquad P_{j+1}\leq P_j. \end{equation} Thus Pj↓FP_j\downarrow F uniformly on SS. The local maximum neighbourhoods may vary with jj; their domains of definition are all SS.

Proof. Approximate FF from above within ηj\eta_j and add the allowed constant ηj\eta_j. This gives the displayed bounds, and 2ηj+1<ηj2\eta_{j+1}<\eta_j gives decrease. Adding a constant preserves every support and every local maximum presentation. This is the geometric-error shift used earlier, now applied after the rational bump construction. ◻

Corollary 11.22.5. (Continuous supported families). Let FjF_j be continuous functions with relative affine lower supports on the same compact SS. If Fj↓FF_j\downarrow F pointwise, then there are rational finite selections PjP_j, each locally a maximum of allowed affine functions, with Fj+ηj≤Pj≤Fj+2ηj,Pj↓F.F_j+\eta_j\leq P_j\leq F_j+2\eta_j,\qquad P_j\downarrow F. If instead Fj→FF_j\to F uniformly with no monotonicity, then a subsequence and suitable allowed shifts give such a decreasing approximation of FF. In the polyhedral case all these functions are coefficient-compatible PL. No common subdivision or common local maximum radius is needed.

Proof. For decreasing targets, apply Theorem 11.22.3 to the jjth target with error ηj\eta_j, then add ηj\eta_j. The bounds imply Pj+1≤Fj+1+2ηj+1≤Fj+ηj≤PjP_{j+1}\leq F_{j+1}+2\eta_{j+1}\leq F_j+\eta_j\leq P_j. For uniformly convergent targets, choose FkjF_{k_j} within ηj/4\eta_j/4 of FF, approximate it from above within ηj/4\eta_j/4, and add ηj\eta_j. Then F+3ηj/4≤Pj≤F+3ηj/2F+3\eta_j/4\leq P_j\leq F+3\eta_j/2; the same geometric ratio gives decrease. The pointwise statement does not assert uniform convergence to a discontinuous or infinite limit. ◻

Lemma 11.22.6. (Smooth cellwise convex inputs have relative supports). Let SS be the support of a locally finite polyhedral complex in Rr\mathbf R^r, and let ff be C1C^1 on an ambient neighbourhood of SS. Assume that its restriction to every cell is locally convex. Then f∣Sf|_S has relative affine lower supports. It is enough, for C2C^2 functions, that the Hessian be positive semidefinite on the tangent space of each cell at its relative interior points.

Proof. Fix x∈Sx\in S. By local finiteness and closedness, a sufficiently small ball meets only cells containing xx. On each of the finitely many incident cells, local convexity at xx gives convexity on its intersection with a sufficiently small ball. Take a common such ball; every segment from xx to a point in that cell stays in the convex intersection. The one-variable tangent inequality gives f(y)≥f(x)+dfx(y−x).f(y)\geq f(x)+df_x(y-x). The right side is one ambient affine function, independent of the cell, and it equals ff at xx. This proves the assertion. In the C2C^2 case the tangential Hessian condition extends to cell boundaries by continuity, and its restriction to each segment gives the required convexity. In affine coordinates the support is y↦f(x)+dfx(y−x)y\mapsto f(x)+df_x(y-x), not y↦f(x)+dfx(y)y\mapsto f(x)+df_x(y) unless the coordinates have first been centred at xx. ◻

This is the local tangent-support argument used in [Wu26] (Appendix A, p. 53), with its affine translation written explicitly and local finiteness included. Combined with Theorem 11.22.3, it gives coefficient-compatible approximation on any fixed compact compatible support inside the ambient domain. It does not assume, without a separate comparison theorem, that positivity of an arbitrary CLD current supplies this cellwise Hessian hypothesis.

Corollary 11.22.7. (Finite ordered families). Let {Fα}α∈A\{F_\alpha\}_{\alpha\in A} be a finite family satisfying Theorem 11.22.3, indexed by a finite partially ordered set, with Fα≤FβF_\alpha\leq F_\beta whenever α≤β\alpha\leq\beta. The decreasing rational approximants may be chosen to preserve all these inequalities simultaneously.

Proof. Choose approximants Qα,jQ_{\alpha,j} from Corollary 11.22.4, with the same ηj\eta_j for all α\alpha. Set Pα,j=max⁡γ≤αQγ,jP_{\alpha,j}=\max_{\gamma\leq\alpha}Q_{\gamma,j}. The index sets are nested when α≤β\alpha\leq\beta, so the order is preserved. Also Fα+ηj≤Pα,j≤Fα+2ηjF_\alpha+\eta_j\leq P_{\alpha,j}\leq F_\alpha+2\eta_j, because the term γ=α\gamma=\alpha is present and Fγ≤FαF_\gamma\leq F_\alpha. Finite maxima preserve the local affine-maximum presentations and decrease in jj. ◻

The actual analytic consequence

Theorem 11.22.8. (Continuous relative-support potentials on unit charts). Let KK be complete and nontrivially valued, let XX be an integral projective KK-variety of positive dimension, and let W⊂XanW\subset X^{\rm an} be ordinary open. Fix analytic units g1,…,grg_1,\ldots,g_r on WW and τ=(−log⁡∣g1∣,…,−log⁡∣gr∣)\tau=(-\log|g_1|,\ldots,-\log|g_r|). Suppose that u=F∘τu=F\circ\tau, where F:τ(W)→RF:\tau(W)\to\mathbf R is continuous and has ambient affine lower supports locally in the relative topology of the actual image. Then, for every x∈Wx\in W, there is one ordinary neighbourhood V⊂WV\subset W carrying original smooth psh functions bj↓u∣Vb_j\downarrow u|_V. Convergence is uniform on compact subsets of VV.

Open proof obligation (RC2). Establish the analytic chart replacement and positivity inputs for the regularization; the proof applies GJR and GRPL.

NAG-900236 · Proof status

Corollary 11.22.9. (Ordered families in one actual chart). In Theorem 11.22.8, replace FF by continuous relatively supported functions FjF_j on the same actual image, and assume Fj↓FF_j\downarrow F pointwise. Then F∘τF\circ\tau has original smooth psh approximation on one ordinary neighbourhood of each point. The limit may be −∞-\infty; compact-uniform convergence is asserted only when the analytic limit is finite continuous. The same conclusion holds for locally uniform limits of such FjF_j, after restriction to a compact image.

Proof. Use the same affinoid AA and the same S=τ(A)S=\tau(A) for every term, apply Corollary 11.22.5, and repeat the local unit-logarithm maximum argument and the fixed-domain transfer steps in the preceding proof. ◻

There is also a consistent longer route: Proposition 11.11.5 gives Q-PL membership from finite selections, and Theorem 11.11.6 gives nef residues from relative supports. The latter uses [GRPSH] (Theorem 4.3), whose hypotheses are a good strictly analytic space over a nontrivially valued field and an already R-PL function. The direct maximum argument above is sufficient here because the new approximants have the stronger local maximum presentation. Neither route infers PL membership from current positivity alone.

For a finite continuous endpoint uu, the conclusion also gives d′d′′[u]≥0d'd''[u]\geq 0 on every original CLD positive test. Indeed, for such a compactly supported test η\eta, compact-uniform convergence gives [bj](d′d′′η)→[u](d′d′′η)[b_j](d'd''\eta)\to[u](d'd''\eta), by the finite signed measure defined by the top-degree form d′d′′ηd'd''\eta. This is the passage in Lemma 11.9.1. The endpoint is not thereby asserted to be PL, nor is an arbitrary-background BT product being defined.

This proves a local forward implication for the stated continuous relative-support presentations. It does not identify every original positive current with such a presentation, treat arbitrary positive-current backgrounds, remove integral projectivity from the transfer input, or settle the discontinuous predecessor and first-stage closure issues in CLD25 Proposition 8.1.11. The complete Wu Theorem A.3 is not deduced. The ground field and valuation normalization are unchanged, and the conclusion concerns original CLD smooth functions and tests.

What the construction does not preserve

Example 11.22.10. (A balanced branching support without an ambient convex extension). In R2\mathbf R^2, take the rays with primitive directions e1,e2,e1+e2,−e1−e2e_1,e_2,e_1+e_2,-e_1-e_2 and weights 1,1,1,21,1,1,2. They form a balanced one-dimensional fan. Define F(s(e1+e2))=sF(s(e_1+e_2))=s for s≥0s\geq 0 and F=0F=0 on the other three rays. This continuous coefficient-compatible PL function has relative affine lower supports: zero supports it at the origin, and away from the origin it is affine on a single local ray. But it is not the restriction of an ambient convex function on any neighbourhood of the origin. Indeed, for small s>0s>0, s(e1+e2)=12(2se1)+12(2se2),F(s(e1+e2))=s>0=12F(2se1)+12F(2se2).\begin{gather*} s(e_1+e_2)=\tfrac12(2se_1)+\tfrac12(2se_2),\\ F(s(e_1+e_2))=s>0=\tfrac12F(2se_1)+\tfrac12F(2se_2). \end{gather*} On a compact compatible truncation, Theorem 11.22.3 nevertheless applies. Its local ambient-convex neighbourhoods must be allowed to shrink with the approximation. Taking a product with Rd−1\mathbf R^{d-1} gives the same example on a balanced support in every positive dimension dd.

Example 11.22.11. (Relative supports need not survive a uniform limit at a boundary). On [0,1][0,1], set fϵ(t)=ϵ−t+ϵf_\epsilon(t)=\sqrt\epsilon-\sqrt{t+\epsilon} for ϵ>0\epsilon>0. Each function is smooth on a neighbourhood of the interval and convex there, hence has affine lower supports at every point of the interval. As ϵ↓0\epsilon\downarrow 0, the functions decrease uniformly to f(t)=−tf(t)=-\sqrt t, with error between zero and ϵ\sqrt\epsilon. Yet ff has no affine lower support at zero: such a support would be atat, and at≤−tat\leq-\sqrt t for all sufficiently small positive tt is impossible. Products give the same boundary phenomenon in every dimension.

This is a statement about relative supports on a compact polyhedral boundary, not a counterexample to an analytic regularization theorem. It explains why Corollary 11.22.5 treats limits directly instead of silently assuming that the limit still satisfies the input support definition.

The local bump proof controls values, not Lipschitz differences. Proposition 11.21.10 remains valid, including its irrational-intercept variant. The construction also uses Γ≠0\Gamma\ne 0 essentially for coefficient approximation: on S={0}S=\{0\} with Γ=0\Gamma=0, every allowed affine function has value zero, so even the constant function one cannot be approximated in this coefficient class. This does not assert failure of original smooth regularization over a trivially valued field.

For the local analytic purpose, Theorem 11.22.8 supersedes the rational-direction and one-scale restrictions in Section 11.21, once the continuous relative-support input is verified. The earlier proofs remain useful for their explicit geometry and, in Section 11.20, their stronger error estimates. The present proof does not repair the global twisted, simplicial or sedentary formulations of the source by redefining their terminology. It supplies a coefficient-controlled local replacement and leaves those comparisons separate.

For comparison of conventions, [BBS22] (Definition 2.1) explicitly requires a locally finite closed polyhedral cover; its real Euclidean framework should not be read as imposing value-group constants. The finite-selection characterization in [FL00] (p. 74, equation (1.1)) likewise concerns real affine data. Our rational conclusion comes from choosing the branches before gluing, not from attributing a rationality assertion to either source.

Continuous supports at vanishing coordinates

The Euclidean quadratic in the first proof cannot simply be used at an infinite tropical coordinate. There is, however, a bounded strictifying function on each compact extended image. We now give the argument separately, with the precise boundary support convention. This treats continuous inputs by constructing new locally finite-max approximants; it does not identify every pre-existing relative PL presentation at the boundary.

Put T=R∪{+∞}\mathbb T=\mathbf R\cup\{+\infty\} with the topology for which t↦e−tt\mapsto e^{-t}, with e−∞=0e^{-\infty}=0, is a homeomorphism onto [0,∞)[0,\infty). At x∈Trx\in\mathbb T^r, an admissible affine germ is a function c+∑i∈Ixqitic+\sum_{i\in I_x}q_it_i on a neighbourhood of xx, where Ix={i:xi<∞}I_x=\{i:x_i<\infty\}. For a real affine germ the coefficients are real; for an allowed germ qi∈Qq_i\in\mathbf Q and c∈ΓQc\in\Gamma_{\mathbf Q}. Thus an affine germ used at a boundary point is independent of every coordinate that is infinite at that point. Relative lower supports below must be admissible real affine germs, with equality at the point.

Theorem 11.22.12. (Rational bumps on compact extended images). Let Γ≠0\Gamma\ne 0, let S⊂TrS\subset\mathbb T^r be nonempty compact, and let F:S→RF:S\to\mathbf R be continuous with relative admissible affine lower supports. For every ε>0\varepsilon>0, there is continuous P:S→RP:S\to\mathbf R with F≤P≤F+εF\leq P\leq F+\varepsilon, such that SS has a finite relative open cover on each member of which PP is a finite maximum of allowed admissible affine functions. The decreasing, decreasing-family and uniformly-convergent-family conclusions of Corollaries 11.22.4 and 11.22.5 hold on this same SS, with this local presentation.

Proof. Write z(t)=(e−t1,…,e−tr)z(t)=(e^{-t_1},\ldots,e^{-t_r}) and Q(t)=∑ie−tiQ(t)=\sum_i e^{-t_i}. The function QQ is continuous and bounded on SS. At x∈Sx\in S put Lx(t)=∑i∈Ixe−xi(1−(ti−xi)).L_x(t)=\sum_{i\in I_x}e^{-x_i}\bigl(1-(t_i-x_i)\bigr). On a neighbourhood where all coordinates in IxI_x stay finite, this is an admissible affine function, Lx(x)=Q(x)L_x(x)=Q(x), and Q(t)−Lx(t)=∑i∈Ixe−xi(e−(ti−xi)−1+(ti−xi))+∑i∉Ixe−ti≥0.\begin{align*} Q(t)-L_x(t) &=\sum_{i\in I_x}e^{-x_i} \bigl(e^{-(t_i-x_i)}-1+(t_i-x_i)\bigr) +\sum_{i\notin I_x}e^{-t_i}\geq0. \end{align*} Equality holds exactly at t=xt=x: the first summands vanish exactly when ti=xit_i=x_i, and the last ones exactly when ti=+∞t_i=+\infty. Choose a>0a>0 with amax⁡SQ≤ε/4a\max_SQ\leq\varepsilon/4 and set g=F+aQg=F+aQ. If ℓxF\ell_x^F supports FF relatively near xx, then ℓx=ℓxF+aLx\ell_x=\ell_x^F+aL_x supports gg, with strictly positive gap off xx in that neighbourhood.

Use the metric d(s,t)=∥z(s)−z(t)∥d(s,t)=\|z(s)-z(t)\|. Choose rx>0r_x>0 so that S∩B‾d(x,2rx)S\cap\overline B_d(x,2r_x) lies in the support neighbourhood and all its coordinates in IxI_x stay in bounded real intervals. On the compact annulus Ax={t∈S:rx/2≤d(x,t)≤2rx}A_x=\{t\in S:r_x/2\leq d(x,t)\leq2r_x\} the continuous gap g−ℓxg-\ell_x has a positive minimum mxm_x, if the annulus is nonempty. If it is empty, choose any mx>0m_x>0. Take 0<δx≤min⁡{ε/4,mx/2}0<\delta_x\leq\min\{\varepsilon/4,m_x/2\}. Approximate ℓx+δx\ell_x+\delta_x within δx/4\delta_x/4 on the closed ball by an allowed affine axa_x using only coordinates in IxI_x. Their boundedness makes coefficient approximation uniform; no infinite-coordinate term is introduced.

Define bx=max⁡{g,ax}b_x=\max\{g,a_x\} on Bd(x,rx)∩SB_d(x,r_x)\cap S and bx=gb_x=g outside. On AxA_x, ax−g≤5δx/4−mx<0a_x-g\leq 5\delta_x/4-m_x<0. Thus the patching is continuous, equals gg outside Bd(x,rx/2)B_d(x,r_x/2), raises the value at xx by at least 3δx/43\delta_x/4, and changes gg by at most 5ε/165\varepsilon/16. As before, a finite positive cover yields P=max⁡bxP=\max b_x with F≤P≤F+9ε/16F\leq P\leq F+9\varepsilon/16.

At a point yy, every active modification is strictly above gg, so its affine branch is defined on a neighbourhood of yy. Each coordinate on which that branch depends is finite there. The active-index argument of Lemma 11.22.2 therefore gives a local finite maximum of admissible affine functions. Compactness gives a finite open cover by such presentations. The affine branches have specified local domains; no evaluation of an inactive branch at a point where it is undefined is needed. Finally the geometric shifts and subsequence arguments of Corollaries 11.22.4 and 11.22.5 apply without change. ◻

Corollary 11.22.13. (The analytic transfer through zeros). In Theorem 11.22.8, allow g1,…,grg_1,\ldots,g_r to be analytic functions, possibly vanishing, set −log⁡0=+∞-\log 0=+\infty, and impose the admissible relative-support hypothesis of Theorem 11.22.12 on the actual image. The same original smooth decreasing approximation exists on one ordinary neighbourhood of every point. The corresponding decreasing and locally uniformly convergent continuous families on the same extended chart also satisfy Corollary 11.22.9.

Proof. Take one compact strict affinoid neighbourhood A⊂WA\subset W, put S=τ(A)⊂TrS=\tau(A)\subset\mathbb T^r, and apply Theorem 11.22.12. This image is compact by continuity; polyhedrality is not needed. For each local maximum presentation of an approximant near τ(y)\tau(y), all coordinates used in its active branches are finite. The corresponding gig_i are nonvanishing near yy, and hence analytic units after restriction to an ordinary neighbourhood. The denominator and unit-logarithm calculation in Theorem 11.22.8 applies there. Thus each approximant is actual Q-PL with nef residues on the same Int⁡(A)\operatorname{Int}(A). Apply the fixed-domain transfer, and use the same compact image for families. The neighbourhoods used to verify individual PL germs are not intersected over the sequence. ◻

The support condition at the boundary is essential to this statement. It is a condition on a real-valued function on the actual extended image, expressed by affine germs independent of the infinite coordinates. It is not obtained merely by checking separate finite strata, taking their closures, or allowing cancellation of infinite affine terms. The construction does not alter the original field or its absolute value. General source-class identification, the first-stage closure problem and the full comparison theorems remain separate.

Single smooth inputs and sequences

Proposition 11.22.14. (Local rational approximation of one original smooth psh function). In the integral-projective, nontrivially valued setting of Theorem 11.22.8, let hh be an original CLD smooth function on WW with d′d′′[h]≥0d'd''[h]\geq 0. For each x∈Wx\in W there is one ordinary neighbourhood UU and a sequence of actual Q-PL functions pjp_j with nef residues on UU, decreasing to h∣Uh|_U uniformly. Each pjp_j is locally a finite maximum of rational unit logarithms. The neighbourhood and the tropical presentation in this assertion may depend on hh.

Open proof obligation (RC2). Supply the local boundary-separation and approximation steps used to pass from one original smooth input; GRPL Theorem 7.7 argument and section 7.2.

NAG-900240 · Proof status

For a sequence of different smooth inputs hih_i, this proposition supplies neighbourhoods UiU_i and presentations that may depend on ii. Their intersection need not be a neighbourhood. Corollaries 11.22.9 and 11.22.13 apply when the continuous supported inputs are already presented on one common actual chart; they do not manufacture such a chart for an arbitrary smooth sequence. Thus the single-input comparison above does not close the full family comparison in Wu Theorem A.3 or the first-stage closure issue in CLD25 Proposition 8.1.11.

For finitely many h1,…,hmh_1,\ldots,h_m, intersect their presentation neighbourhoods, take the product unit chart and add boundary-separating coordinates once. Each ambient HiH_i remains smooth; the same integration argument gives relative supports on the common image away from the boundary. One compact affinoid A′A' therefore works for all inputs. Corollary 11.22.7 also preserves any prescribed finite order in the rational approximants. This uses finite intersections and products; it supplies no countable common chart.

The common-domain replacement for arbitrary original smooth families in this projective setting is completed in Theorem 11.24.3. It uses finite extended presentations on each prescribed affinoid, rather than a countable product of unit charts. The remaining sheaf-level decreasing closure is distinguished in Section 11.24.4.

11.23 Varying unit charts on one affinoid neighbourhood

A fixed finite chart is stronger than the approximation argument requires. It is enough that the charts of the individual terms be defined on one affinoid neighbourhood: boundary-separating coordinates can then be chosen once, while the other coordinates vary. We prove this criterion first. We then give a decreasing smooth psh sequence with a finite continuous uniform limit for which the criterion’s presentation hypothesis fails on every neighbourhood of a rational point. Thus the criterion advances the common-domain argument but does not remove its general gluing problem.

Separating the boundary before choosing the terms

Throughout this section, KK is complete and nontrivially valued, XX is an integral projective KK-variety of positive dimension, and W⊂XanW\subset X^{\rm an} is ordinary open. Interiors are taken in XanX^{\rm an}; ∂B\partial B denotes the analytic boundary of an affinoid domain BB over KK. A global unit presentation on AA means an equality h∣A=H∘τh|_A=H\circ\tau, where τ=(−log⁡∣g1∣,…,−log⁡∣gr∣)\tau=(-\log|g_1|,\ldots,-\log|g_r|) for finitely many gi∈O(A)×g_i\in\mathcal O(A)^\times, and HH is C∞C^\infty on an ambient open neighbourhood of τ(A)\tau(A). This is a hypothesis on the whole of AA, stronger than original smoothness at each point.

Lemma 11.23.1. (A boundary separator independent of the input). Let A⊂WA\subset W be a compact strictly affinoid neighbourhood of xx. There are compact strictly affinoid neighbourhoods B⊂AB\subset A and A′⊂Int⁡(B)A'\subset\operatorname{Int}(B) of xx, and a unit tropicalization β:B→Rs\beta:B\to\mathbf R^s, such that, for every finite unit tropicalization τ\tau on AA, the product γ=(τ∣B,β)\gamma=(\tau|_B,\beta) satisfies γ(A′)⊂γ(B)∖γ(∂B).\begin{equation} \gamma(A')\subset \gamma(B)\setminus\gamma(\partial B). \end{equation} The choices of BB, A′A' and β\beta depend only on (A,x)(A,x), not on τ\tau or on the number of charts subsequently considered.

Open proof obligation (RC2). Supply boundary separation by analytic functions in the given setup; the proof applies GRPL’s Theorem 7.7 construction.

NAG-900241 · Proof status

The lemma fixes the analytic domain, not the tropical image. With infinitely many input charts the sets γj(A′)\gamma_j(A') may have different dimensions and subdivisions. They need not embed into one finite-dimensional common chart. The estimates below are imposed after pullback to the same A′A', where they can be compared.

Theorem 11.23.2. (Simultaneous rational replacement on a common affinoid). Let hjh_j be original CLD smooth psh functions on WW, and suppose that every hjh_j has a global unit presentation on the same compact strictly affinoid neighbourhood AA of xx. The finite tuples of units and the ambient smooth functions may depend on jj, with no uniform bound on their number or derivatives. There is one compact strictly affinoid neighbourhood A′A' of xx such that, for every sequence εj>0\varepsilon_j>0, there are continuous functions qjq_j on A′A' satisfying hj≤qj≤hj+εjon A′,\begin{equation} h_j\leq q_j\leq h_j+\varepsilon_j\quad\hbox{on }A', \end{equation} and each qj∣Int⁡(A′)q_j|_{\operatorname{Int}(A')} is actual Q-PL with nef residue classes, locally a finite maximum of rational unit logarithms. The same A′A' works before the errors εj\varepsilon_j are prescribed.

Open proof obligation (RC2). Supply the uniform chart replacement with the asserted error and positivity in the stated scope; GRPL (7.7.1).

NAG-900242 · Proof status

Decreasing families and continuous uniform limits

Corollary 11.23.3. (A family transfer without a common finite chart). Under the hypotheses of Theorem 11.23.2, assume either that hj↓hh_j\downarrow h pointwise on WW, with values of hh in R∪{−∞}\mathbf R\cup\{-\infty\}, or that hj→hh_j\to h locally uniformly on WW with hh finite and continuous. Then one ordinary neighbourhood UU of xx carries decreasing actual Q-PL functions with nef residues and pointwise limit h∣Uh|_U. On one further neighbourhood V⊂UV\subset U, there are original smooth psh functions bj↓h∣Vb_j\downarrow h|_V. In the second case, and whenever the endpoint in the first case is finite and continuous, convergence is uniform on compact subsets of VV.

Proof. Fix 0<ρ∈Γ0<\rho\in\Gamma and ηj=ρ8−j∈ΓQ\eta_j=\rho 8^{-j}\in\Gamma_{\mathbf Q}. In the decreasing case apply Theorem 11.23.2 with error ηj\eta_j and put pj=qj+ηjp_j=q_j+\eta_j. On A′A' one has hj+ηj≤pj≤hj+2ηj,pj+1≤hj+2ηj+1≤pj.h_j+\eta_j\leq p_j\leq h_j+2\eta_j, \qquad p_{j+1}\leq h_j+2\eta_{j+1}\leq p_j. Thus pj↓hp_j\downarrow h on U=Int⁡(A′)U=\operatorname{Int}(A'). Adding ηj\eta_j preserves Q-PL membership and nef residues because ηj∈ΓQ\eta_j\in\Gamma_{\mathbf Q}.

In the locally uniform case, choose increasing indices njn_j with sup⁡A′∣hnj−h∣≤ηj/4\sup_{A'}|h_{n_j}-h|\leq\eta_j/4. Replace hnjh_{n_j} with error ηj/4\eta_j/4 and again add ηj\eta_j. The resulting functions satisfy h+34ηj≤pj≤h+32ηj,32ηj+1<34ηj,h+\tfrac34\eta_j\leq p_j\leq h+\tfrac32\eta_j, \qquad \tfrac32\eta_{j+1}<\tfrac34\eta_j, so they decrease to hh on the same UU. In both cases Theorem 11.10.11 supplies the required original smooth sequence on one final V⊂UV\subset U. For a finite continuous endpoint, Dini’s theorem on each compact subset gives the stated convergence. No change of ground field, normalization or original test class occurs. ◻

For an already decreasing original smooth sequence, the last original-smooth conclusion by itself is tautological. The substantive assertion is its replacement by decreasing actual Q-PL functions with nef residues on one fixed domain. For a locally uniform, nonmonotone sequence, the argument additionally produces a decreasing sequence. These are consequences of the stated presentation hypothesis, not an identification of all sheaf-generated classes.

Why a common affinoid presentation is a genuine condition

Lemma 11.23.4. (Unit logarithms on a closed polydisc). Let Dsn={∣Ti∣≤s (1≤i≤n)}⊂AKn,anD_s^n=\{|T_i|\leq s\ (1\leq i\leq n)\}\subset\mathbf A_K^{n,\rm an}, with s∈∣K×∣s\in|K^\times|. Every g∈O(Dsn)×g\in\mathcal O(D_s^n)^\times has constant absolute value on DsnD_s^n. Consequently, if an affinoid neighbourhood AA of 00 contains DsnD_s^n, every function given by any finite unit presentation on AA is constant on this same DsnD_s^n. The statement holds for any number of such functions, even when their tuples of units differ.

Open proof obligation (RC2). Prove the analytic Tate-algebra unit criterion used in the argument; Bosch14 Section 2.1, Corollary 4.

NAG-900244 · Proof status

Example 11.23.5. (Uniform decreasing approximation without common unit presentations). For every n≥1n\geq 1, on W={∣Ti∣<1 (1≤i≤n)}⊂PKn,anW=\{|T_i|<1\ (1\leq i\leq n)\}\subset\mathbf P_K^{n,\rm an} there is a decreasing sequence of original smooth psh functions converging uniformly to the finite continuous function u=∣T1∣2u=|T_1|^2, but no affinoid neighbourhood of 00 carries global unit presentations for every term of the sequence.

Proof. Fix 0<δ<10<\delta<1, and let κ\kappa be a nonnegative even C∞C^\infty function supported in (−δ,δ)(-\delta,\delta) with integral one. Define, for z≥0z\geq 0, ψ(z)=∫−δδmax⁡{z,1+a}κ(a) da.\psi(z)=\int_{-\delta}^{\delta}\max\{z,1+a\}\kappa(a)\,da. This is smooth, convex and nondecreasing, equals 11 for z≤1−δz\leq 1-\delta, and equals zz for z≥1+δz\geq 1+\delta. Moreover 0≤ψ(z)−z≤1+δ,ψ(z)−zψ′(z)=∫a>z−1(1+a)κ(a) da≥0.\begin{gather*} 0\leq\psi(z)-z\leq1+\delta,\\ \psi(z)-z\psi'(z) =\int_{a>z-1}(1+a)\kappa(a)\,da\geq0. \end{gather*} Choose any rj↓0r_j\downarrow 0 and put hj(T)=rjψ(∣T1∣2/rj).h_j(T)=r_j\psi\bigl(|T_1|^2/r_j\bigr). Each hjh_j equals rjr_j on a neighbourhood of 00. On {T1≠0}\{T_1\ne 0\}, it is an ambient smooth function of the unit logarithm t=log⁡∣T1∣t=\log|T_1|. With z=e2t/rjz=e^{2t}/r_j, its second derivative is d2dt2(rjψ(e2t/rj))=4rj(zψ′(z)+z2ψ′′(z))≥0.\frac{d^2}{dt^2}\bigl(r_j\psi(e^{2t}/r_j)\bigr) =4r_j\bigl(z\psi'(z)+z^2\psi''(z)\bigr)\geq0. Thus hjh_j is original smooth and psh on WW, using a constant presentation near {T1=0}\{T_1=0\} and the unit chart elsewhere; the displayed nonnegative Hessian pulls back to a positive (1,1)(1,1)-form. For fixed v≥0v\geq 0, differentiation in r>0r>0 gives ddr(rψ(v/r))=ψ(v/r)−(v/r)ψ′(v/r)≥0.\frac{d}{dr}\bigl(r\psi(v/r)\bigr) =\psi(v/r)-(v/r)\psi'(v/r)\geq0. Consequently hjh_j decreases, and 0≤hj−u≤(1+δ)rj0\leq h_j-u\leq(1+\delta)r_j on all of WW. This proves uniform convergence, including at 00.

Suppose a common affinoid neighbourhood AA supplied the claimed unit presentations. It contains a polydisc DsnD_s^n with s∈∣K×∣s\in|K^\times| and s>0s>0. Lemma 11.23.4 would make every hjh_j constant on this same disc. For jj large enough that (1+δ)rj<s2(1+\delta)r_j<s^2, however, hj(0)=rjh_j(0)=r_j whereas at the Gauss point ζs\zeta_s with all radii equal to ss one has hj(ζs)=s2h_j(\zeta_s)=s^2. This is a contradiction. ◻

This is a counterexample only to automatic simultaneous global unit presentations. It is not a counterexample to regularization: the functions hjh_j themselves are the required regularization. They share the extended coordinate −log⁡∣T1∣-\log|T_1|, with an admissible constant germ at T1=0T_1=0, as permitted in Corollary 11.22.13. It also shows that local uniform convergence does not, by itself, give a uniform radius for the original smooth presentations. The same obstruction occurs in every positive dimension, although it is already detected by the first coordinate.

A common-domain argument must therefore do more than ask for global units on a smaller affinoid. The next section allows global analytic functions with zeros outside their local patches. It proves that these extended presentations suffice to handle arbitrary original smooth psh families on a fixed affinoid in the stated setting. The counterexample above remains valid for global unit presentations.

11.24 Gluing presentations on a fixed affinoid

The obstruction in Example 11.23.5 concerns global units. It disappears if a global analytic function is allowed to vanish outside the patch on which its logarithm is used. The following argument exploits this distinction. On a Laurent domain, a sufficiently accurate rational approximation to a unit has exactly the same absolute value. Finitely many such replacements, together with the functions defining the domains, place all local presentations in one extended tropical image. Relative affine supports then pass to that image. This gives rational approximation on a prescribed compact affinoid, before a sequence of inputs is chosen.

Exact absolute values from rational approximation

Lemma 11.24.1. (Replacing local units by global fractions). Let A=M(A)A=\mathcal M(\mathscr A) be strictly affinoid, and let C={y∈A:∣ai(y)∣≤1,∣bk(y)∣≥1}C=\{y\in A:|a_i(y)|\leq1, |b_k(y)|\geq1\} be a Laurent domain, where ai,bk∈Aa_i,b_k\in\mathscr A. For every g∈O(C)×g\in\mathcal O(C)^\times there are P,Q∈AP,Q\in\mathscr A, both nonvanishing on CC, such that ∣g(y)∣=∣P(y)/Q(y)∣(y∈C).\begin{equation} |g(y)|=|P(y)/Q(y)|\qquad(y\in C). \end{equation} One can take QQ to be a product of nonnegative integral powers of the bkb_k. No equality of analytic functions with gg is asserted.

Proof. If CC is empty, take P=Q=1P=Q=1. Otherwise the Laurent-domain algebra is O(C)=A⟨Ui,Vk⟩/(Ui−ai,bkVk−1).\mathcal O(C)= \mathscr A\langle U_i,V_k\rangle/(U_i-a_i, b_kV_k-1). This is the affinoid localization of Section 2.1; see also [Bosch14] (Section 3.3, Proposition 11). Truncation in the Tate algebra shows that the image of A[b1−1,…,bs−1]\mathscr A[b_1^{-1},\ldots,b_s^{-1}] is dense in O(C)\mathcal O(C). In particular its elements approximate gg in the supremum seminorm on CC. Since gg is a unit, m=min⁡C∣g∣>0m=\min_C|g|>0. Choose a fraction P/QP/Q in this dense subalgebra with sup⁡C∣g−P/Q∣<m\sup_C|g-P/Q|<m. The ultrametric equality proves (11.24.1) at every Berkovich point. The denominator is a product of the bkb_k, hence is nonvanishing on CC; the equality makes PP nonvanishing there as well. ◻

We also need the ordinary topology, rather than just an affinoid-domain cover. For any point of AA and any relative ordinary neighbourhood, choose finitely many strict inequalities ∣f∣<c|f|<c and ∣g∣>d|g|>d defining a smaller neighbourhood. Nontriviality makes ∣K×∣Q|K^\times|^{\mathbf Q} dense in R>0\mathbf R_{>0}. The thresholds can therefore be chosen in this group, with strict inequalities at the point and with their closed versions still inside the prescribed neighbourhood. Raising to positive integral powers and dividing by constants in K×K^\times writes the resulting compact domain in the form CC above. Its strict locus V={y∈A:∣ai(y)∣<1,∣bk(y)∣>1}V=\{y\in A:|a_i(y)|<1, |b_k(y)|>1\} is a relative ordinary neighbourhood of the point. Both CC and VV are determined by absolute values of global analytic functions on AA. This elementary construction does not say that every member of an arbitrary affinoid-domain cover contains an ordinary neighbourhood.

A finite extended presentation with relative supports

Call a finite continuous function hh locally unit-supported if, near each point in the ordinary topology, it has a unit tropical presentation h=H∘σh=H\circ\sigma, where HH is continuous on the actual image and has relative affine lower supports. The ambient affine slopes in these supports may be real. This definition includes local finite maxima of rational unit logarithms. In our integral-projective, nontrivially valued setting, it also includes every original smooth psh function: the boundary-adapted positivity argument in the proof of Proposition 11.22.14 supplies precisely these local supports.

Proposition 11.24.2. (Globalizing a finite local presentation). Let KK be complete and nontrivially valued, let AA be compact strictly affinoid, and let h:A→Rh:A\to\mathbf R be continuous and locally unit-supported in its relative ordinary topology. Then finitely many global analytic functions f1,…,fN∈O(A)f_1,\ldots,f_N\in\mathcal O(A) give an extended tropicalization τ:A⟶TN,τ=(−log⁡∣f1∣,…,−log⁡∣fN∣),S=τ(A),\tau:A\longrightarrow\mathbb T^N,\qquad \tau=(-\log|f_1|,\ldots,-\log|f_N|), \quad S=\tau(A), such that h=F∘τh=F\circ\tau for a continuous F:S→RF:S\to\mathbf R having relative admissible affine lower supports in the sense of Theorem 11.22.12. In particular, at a point where a coordinate is infinite, its supporting affine germ does not use that coordinate. No global-unit assertion is made for the fif_i.

Proof. At each point choose a local supported presentation, then a Laurent domain CiC_i inside its ordinary presentation domain whose strict locus ViV_i contains the point, as constructed above. Compactness yields finitely many ViV_i covering AA. Write the presentation on CiC_i as h=Hi∘σi,σi=(−log⁡∣gi1∣,…,−log⁡∣gimi∣).h=H_i\circ\sigma_i,\qquad \sigma_i=(-\log|g_{i1}|,\ldots,-\log|g_{im_i}|). Apply Lemma 11.24.1 to each unit gikg_{ik} on CiC_i, obtaining Pik,Qik∈O(A)P_{ik},Q_{ik}\in\mathcal O(A) with the same absolute-value ratio. Include all these numerators and denominators, and every global function defining the inequalities of the CiC_i, in the finite list defining τ\tau. Identically zero constraint functions are harmless; unused coordinates can be omitted.

Because the strict inequalities defining ViV_i are recorded among the coordinates, there is an open Oi⊂TNO_i\subset\mathbb T^N such that Vi=τ−1(Oi)V_i=\tau^{-1}(O_i). These OiO_i cover SS. On CiC_i, the coordinates of σi\sigma_i are differences of the finite coordinates corresponding to PikP_{ik} and QikQ_{ik}. Denote this linear map on that finite-coordinate locus by LiL_i. If two points of AA have the same τ\tau-image, choose ViV_i containing one; saturation puts both in ViV_i, and h=Hi∘Li∘τh=H_i\circ L_i\circ\tau gives equal values. Thus hh defines a function FF on SS. The map τ:A→S\tau:A\to S is a continuous surjection from a compact space to a Hausdorff space, hence a quotient map; consequently FF is continuous.

Finally fix w∈S∩Oiw\in S\cap O_i. Every PikP_{ik} and QikQ_{ik} is nonvanishing on CiC_i, so all coordinates used by LiL_i are finite near ww. A relative affine lower support ℓ\ell of HiH_i at Li(w)L_i(w), restricted to σi(Ci)\sigma_i(C_i), pulls back to ℓ∘Li\ell\circ L_i. On a sufficiently small relative neighbourhood of ww in S∩OiS\cap O_i, this affine function is at most FF and equals it at ww. It only uses coordinates finite at ww, so is admissible. This proves the asserted support property at every point, including the extended boundary. ◻

Only the presentations are glued. The fractions need not agree as analytic functions on overlaps, because equality of their absolute values with the prescribed local units is enough. The functions Hi∘LiH_i\circ L_i agree on the actual image by their common pullback hh; no ambient extension across all patches is asserted. Polyhedrality of the global extended image is also unnecessary for Theorem 11.22.12.

Fixed-domain replacement and first decreasing stages

Theorem 11.24.3. (Rational approximation on a prescribed affinoid). Let K,X,WK,X,W have the hypotheses of Section 11.23, and fix a compact strictly affinoid domain A⊂WA\subset W with nonempty ordinary interior UU. If hh is finite continuous and locally unit-supported on WW, then for every ε>0\varepsilon>0 there is a continuous q:A→Rq:A\to\mathbf R with h≤q≤h+εon A,h\leq q\leq h+\varepsilon\quad\hbox{on }A, and q∣Uq|_U is locally a finite maximum of rational unit logarithms, hence actual Q-PL with nef residues. This applies in particular to every original smooth psh hh on WW. The affinoid AA is fixed independently of hh and ε\varepsilon.

Proof. Restrict the local supported presentations of hh to AA and apply Proposition 11.24.2. Apply Theorem 11.22.12 to the resulting compact image SS, with Γ=−log⁡∣K×∣\Gamma=-\log|K^\times|. Pull back its approximant to AA. At every point of UU, each active affine branch uses only coordinates finite there. Their global analytic functions are therefore units on an ordinary neighbourhood of that point. The denominator calculation and finite-maximum semipositivity argument of Theorem 11.22.8 give the asserted rational PL and nef properties. The domain was never reduced in this construction. For original smooth psh inputs, local supported presentations come from the already verified boundary-adapted argument, as stated above. ◻

Corollary 11.24.4. (First-stage comparison in the stated projective setting). For an ordinary open O⊂XanO\subset X^{\rm an}, let H(O)\mathcal H(O) be its finite original smooth psh functions, let C(O)\mathcal C(O) be its finite continuous locally unit-supported functions, let G(O)\mathcal G(O) be its functions locally expressible as finite maxima of rational unit logarithms with constants in ΓQ\Gamma_{\mathbf Q}, and let SQ(O)\mathcal S_{\mathbf Q}(O) have the meaning of Section 11.10.6. For one of these classes E\mathcal E, write Dx(E)D_x(\mathcal E) for germs at xx representable as ej↓ue_j\downarrow u on one ordinary neighbourhood of xx, with all eje_j in that class on that neighbourhood. Allow endpoints in R∪{−∞}\mathbf R\cup\{-\infty\}. Then Dx(H)=Dx(C)=Dx(G)=Dx(SQ).\begin{equation} D_x(\mathcal H)=D_x(\mathcal C) =D_x(\mathcal G)=D_x(\mathcal S_{\mathbf Q}). \end{equation} If a sequence in C(W)\mathcal C(W) converges locally uniformly to a finite continuous uu, then its germ also belongs to these equal classes. The resulting decreasing original smooth approximation converges uniformly on compact subsets of one neighbourhood.

Proof. The local supported-presentation argument gives H⊂C\mathcal H\subset\mathcal C. Let hj∈C(W)h_j\in\mathcal C(W) decrease to uu, and fix a compact strictly affinoid neighbourhood A⊂WA\subset W of xx once. Apply Theorem 11.24.3 to every hjh_j, with error ηj=ρ8−j∈ΓQ\eta_j=\rho 8^{-j}\in\Gamma_{\mathbf Q}, and add ηj\eta_j. The inequalities in the proof of Corollary 11.23.3 give a decreasing sequence pj∈G(Int⁡(A))p_j\in\mathcal G(\operatorname{Int}(A)) with limit uu. Thus Dx(H)⊂Dx(C)⊂Dx(G)D_x(\mathcal H)\subset D_x(\mathcal C)\subset D_x(\mathcal G). The local unit-log residue and maximum statements give G⊂SQ\mathcal G\subset\mathcal S_{\mathbf Q}. Finally Theorem 11.10.11 gives Dx(SQ)⊂Dx(H)D_x(\mathcal S_{\mathbf Q})\subset D_x(\mathcal H), closing the chain. For a locally uniform sequence use the subsequence and separated errors from Corollary 11.23.3, now with Theorem 11.24.3 in place of the common-unit-presentation hypothesis. Dini’s theorem gives compact-uniform convergence at a finite continuous endpoint. ◻

Equation (11.24.2) is an equality of explicitly defined first-stage germ classes over the fixed nontrivially valued field and integral projective variety. It does not assert that this class is already closed under every further locally defined decreasing operation, or under the full transfinite procedure in CLD25. Identifying Wu’s precise analytic/algebraic generator conventions and all of its field hypotheses remains separate. No arbitrary-positive-current Bedford–Taylor product, new field descent, or verification of the unavailable mixed-calculus Theorem 6.13 is used or obtained. The proof is a synthesis of affinoid localization, the local positivity input, compact extended-image density, and the existing projective transfer; no independent priority claim is made.

A second diagonal when all rows have one domain

The first-stage equality leaves a separate question: can a decreasing limit of first-stage germs again be represented on one neighbourhood? The following elementary diagonal argument settles the case in which all rows already have a common domain. It does not manufacture that domain. Unlike the uniform-error diagonal, it permits discontinuous and −∞-\infty row limits.

Lemma 11.24.5. (Compact diagonal for ordered row limits). Let QQ be compact, and let fj,k:Q→Rf_{j,k}:Q\to\mathbf R be continuous. Suppose fj,k↓ujf_{j,k}\downarrow u_j as k→∞k\to\infty, and uj↓uu_j\downarrow u as j→∞j\to\infty, allowing limits in R∪{−∞}\mathbf R\cup\{-\infty\}. Fix any positive strictly decreasing ηj→0\eta_j\to 0. There are integers mj≥jm_j\geq j such that aj=fj,mj+ηj↓uon Q.a_j=f_{j,m_j}+\eta_j\downarrow u\quad\hbox{on }Q. No metrizability, countable dense subset, continuity of uju_j, or uniform convergence of the rows is assumed.

Proof. Choose mjm_j inductively so that, in addition to mj≥jm_j\geq j, fj,mj+ηj≤aj−1(j>1),fj,mj≤fk,ℓ+ηj(1≤k,ℓ≤j).\begin{align} f_{j,m_j}+\eta_j&\leq a_{j-1} &&(j>1),\\ f_{j,m_j}&\leq f_{k,\ell}+\eta_j &&(1\leq k,\ell\leq j). \end{align} For every pair in the second line, uj≤uk≤fk,ℓ<fk,ℓ+ηju_j\leq u_k\leq f_{k,\ell}<f_{k,\ell}+\eta_j. For the first line the inductive lower bound aj−1≥uj−1+ηj−1a_{j-1}\geq u_{j-1}+\eta_{j-1} gives uj<aj−1−ηju_j<a_{j-1}-\eta_j wherever uj−1u_{j-1} is finite. Where it equals −∞-\infty, also uj=−∞u_j=-\infty, so the same strict inequality holds. For any one of these finitely many continuous upper bounds bb, the open sets {fj,m<b}\{f_{j,m}<b\} increase with mm and cover QQ. Compactness makes one of them equal to QQ. Taking the largest of finitely many resulting indices imposes all the inequalities simultaneously. This constructs mjm_j and proves decrease of aja_j.

Since aj≥uj≥ua_j\geq u_j\geq u, its limit aa satisfies a≥ua\geq u. For each fixed (k,ℓ)(k,\ell), inequality (11.24.4) holds for all sufficiently large jj, so a≤fk,ℓa\leq f_{k,\ell}. Taking the infimum first in ℓ\ell, then in kk, gives a≤ua\leq u. Thus a=ua=u. The finite upper bounds were used only as constraints in the selection; their minimum is not used as an approximating function. ◻

For original smooth psh rows on one open WW, restrict to a compact affinoid neighbourhood Q⊂WQ\subset W and then to its ordinary interior. Each selected aja_j remains original smooth psh. For rows in G\mathcal G or SQ\mathcal S_{\mathbf Q}, choose the shifts in ΓQ\Gamma_{\mathbf Q} and obtain the same conclusion in that generator class. This is a common-domain second-stage closure statement. Merely knowing, for each jj, that the germ of uju_j belongs to Dx(H)D_x(\mathcal H) provides possibly different row domains; it does not satisfy this lemma’s hypothesis. That ordinary-neighbourhood/sheafification issue, together with the precise transfinite predecessor requirement, is the next comparison task.

Corollary 11.24.6. (Full collapse of the raw compact hierarchy). Let QQ be compact Hausdorff and let S⊂C(Q,R)\mathscr S\subset C(Q,\mathbf R) be nonempty and stable under addition of real constants. For the raw hierarchy of Definition 11.6.2, Dα(S)=D1(S)(α≥1).D_\alpha(\mathscr S)=D_1(\mathscr S)\qquad(\alpha\geq1). This equality includes discontinuous and −∞-\infty-valued endpoints. It concerns actual functions on the same compact domain, without sheafification.

Proof. If uj∈D1(S)u_j\in D_1(\mathscr S) decreases to uu, choose a row fj,k∈Sf_{j,k}\in\mathscr S decreasing to each uju_j on QQ. Lemma 11.24.5 gives fj,mj+ηj↓uf_{j,m_j}+\eta_j\downarrow u. Stability under constants places every selected term in S\mathscr S, so D1(S)D_1(\mathscr S) is closed under decreasing sequences. Since constant sequences give S⊂D1(S)\mathscr S\subset D_1(\mathscr S), transfinite induction in the stated recursion proves the equality: successor steps use this closure, and limit steps take unions. ◻

Theorem 11.6.4 remains useful for its uniform-closure characterization at continuous endpoints. The present corollary strengthens only its raw, fixed-domain conclusion. A local sheaf section may be produced on changing neighbourhoods before any compact representation is available; that operation is absent from the recursion used here. Thus this corollary does not supply the missing representation property in Corollary 11.6.5.