17 · Harmonic coordinates and weakly smooth forms

22 NAG Tags

17.1 Harmonic coordinates: the geometric obstruction to original smoothness

Numerical triviality and principal triviality differ

An original affine coordinate is built from logarithms of analytic units. Its residual line-bundle data is therefore principal in the appropriate local description. A numerical harmonic function only asks the associated residual class to be numerically trivial. A degree-zero line bundle on a residual curve need not be trivial or torsion. This difference can prevent a numerical harmonic function from being original smooth.

On a curve whose residue field is algebraic over a finite field, torsion properties of residual degree-zero classes improve the comparison. The relevant statement in GJR is a curve result. It should not be generalized to every dimension over a finite extension of Qp\mathbb Q_p merely because the residue field is finite.

What the enlarged sheaf is intended to repair

If harmonic functions are geometrically natural but unavailable as original smooth coordinates, differential constructions based only on analytic units may be too restrictive. GJR allows harmonic tropicalizations and pulls back real smooth superforms through them. The resulting weakly smooth sheaf contains the original one and supports a differential calculus and integration theory.

The adjective weakly smooth refers to the enlarged coordinate class; it has no connection with weak positivity. Its continuous dual consists of strong currents in the terminology of that paper; again, this says which tests are allowed, not which positive cone is used. The detailed constructions and the direction of the restriction map follow in the next section.

The curve cohomology calculation

Gubler–Jell–Rabinoff also construct a compatible differential theory on weighted metric graphs and use it to calculate weakly smooth Dolbeault cohomology of compact Berkovich curves. For a connected compact curve without boundary in their setting, the dimensions satisfy h0,0=h1,1=1,h1,0=h0,1=b1(Xan).h^{0,0}=h^{1,1}=1,\qquad h^{1,0}=h^{0,1}=b_1(X^{\mathrm{an}}). The integer on the right is the first Betti number of the underlying analytic topological space. It is not in general the algebraic genus. A smooth proper curve with good reduction can have positive algebraic genus and b1(Xan)=0b_1(X^{\mathrm{an}})=0.

This is an important warning for the proposed passage from tropical Hodge theory to Berkovich Hodge theory: one must specify which cohomology is being compared. Residual algebraic geometry is not automatically encoded in the first Betti number of a skeleton. A statement resembling a Hodge diamond may concern tropical Dolbeault groups rather than algebraic de Rham cohomology.

References. [GJR] (Sections 7, 10, 15–17); [GJRc] (Theorems 6.4.3 and 6.4.5).

17.2 Harmonic tropicalizations and weakly smooth forms

Throughout this section, KK is complete and nontrivially valued, and XX is a good strictly KK-analytic space. The version of Gubler–Jell–Rabinoff cited here is arXiv:2111.05741v3, dated 7 March 2025. The adjective weakly in weakly smooth concerns the allowed coordinate functions. It has no connection with the distinction between weakly, strongly, and ordinarily positive forms.

From analytic units to numerically trivial models

A smooth tropicalization is locally a map F=(−log⁡∣f1∣,…,−log⁡∣fr∣)F=(-\log|f_1|,\ldots,-\log|f_r|) defined by invertible analytic functions. To describe the enlargement of this class of maps, we first specify the model-theoretic meaning of harmonicity.

Definition 17.2.1. (Local model functions and harmonicity). Let x∈Xx\in X. An integral model function near xx has the form h=−log⁡∥1∥Lh=-\log\|1\|_{\mathfrak L} on an analytic neighbourhood VV, where V\mathfrak V is a formal K∘K^\circ-model of VV and L\mathfrak L is a model of OV\mathscr O_V, with its specified generic-fibre trivialization. An R\mathbb R-model function is locally a finite real linear combination of such functions.

An R\mathbb R-model function hh is semipositive at xx if it can be represented on a neighbourhood of xx by an R\mathbb R-model line bundle whose restriction to the special fibre is nef. It is harmonic at xx if both hh and −h-h are semipositive there. These conditions are imposed at every point to obtain semipositivity or harmonicity on XX.

Here nefness means nonnegative degree on every proper integral curve in the special fibre. An R\mathbb R-model line bundle means a finite real linear combination of integral model line bundles, with degrees extended linearly. Proposition 7.1 and Definition 7.2 of [GJR] identify this definition with the corresponding condition on the residue class Lh(x)L_h(x) on the reduction of the analytic germ (X,x)(X,x). In particular, h is harmonic at x⟺Lh(x) is numerically trivial.h\text{ is harmonic at }x \quad\Longleftrightarrow\quad L_h(x)\text{ is numerically trivial}. Numerical triviality is the condition deg⁡(Lh(x)∣C)=0\deg(L_h(x)|_C)=0 on the relevant proper curves. It does not assert that Lh(x)L_h(x) is trivial or torsion.

Proposition 17.2.2. (Elementary direction and functoriality). If ff is an invertible analytic function, then log⁡∣f∣\log|f| is harmonic. Harmonic functions are stable under pullback by analytic morphisms and under extension of the ground field. Every real harmonic function is locally a finite real linear combination of integral harmonic functions.

Open proof obligation (RC2). Prove the nontrivial harmonic and PL assertions currently assigned to GJR Propositions 7.6 and 7.5.

NAG-900304 · Proof status

Remark (The relevant topology). The PL functions form a sheaf for the analytic GG-topology. Harmonicity and semipositivity in Definition 17.2.1 are local in the ordinary analytic topology and do not form sheaves for the GG-topology [GJR] (Remark 7.3(3)). A covering by formal domains therefore cannot be substituted without argument for analytic neighbourhoods in this definition.

Definition 17.2.4. (Harmonic tropicalization). A harmonic tropicalization of a compact good strictly analytic space VV is a map h=(h1,…,hr):V→Rrh=(h_1,\ldots,h_r):V\to\mathbb R^r whose coordinates are integral harmonic functions.

This is [GJR] (Definition 8.1). A smooth tropicalization is harmonic, but a harmonic tropicalization need not arise from analytic units. The integrality of the coordinates is part of the definition. The last assertion of Proposition 17.2.2 explains how general real harmonic functions are nevertheless represented in such charts.

The sheaf and its differential operations

Definition 17.2.5. (Weakly smooth preforms and forms). Let U⊂XU\subset X be open with compact strictly analytic closure U‾\overline U. A preform of type (p,q)(p,q) on UU is a pair (h,α)(h,\alpha), where h:U‾→Rrh:\overline U\to\mathbb R^r is a harmonic tropicalization and α\alpha is a smooth Lagerberg (p,q)(p,q)-form on h(U‾)h(\overline U). Two pairs (h,α)(h,\alpha) and (k,β)(k,\beta) represent the same preform if pr⁡1∗α=pr⁡2∗βon (h,k)(U‾).\operatorname{pr}_1^*\alpha =\operatorname{pr}_2^*\beta \quad\text{on }(h,k)(\overline U). The sheafification of this presheaf is denoted Awkp,q\mathcal A_{\mathrm{wk}}^{p,q}. Its sections are weakly smooth forms.

This is [GJR] (Definitions 10.3 and 10.8). Passing to a product chart puts any finite collection of preforms on a common tropicalization. On that chart, wedge products and derivatives are those of Lagerberg forms: d′(h,α)=(h,d′α),d′′(h,α)=(h,d′′α).d'(h,\alpha)=(h,d'\alpha),\qquad d''(h,\alpha)=(h,d''\alpha). For types (p,q)(p,q) and (p′,q′)(p',q'), the commutativity sign is (−1)(p+q)(p′+q′)(-1)^{(p+q)(p'+q')}, and the Leibniz sign for either differential is (−1)p+q(-1)^{p+q}. These are total-degree signs.

Theorem 17.2.6. (Comparison of the three sheaves). There are injective morphisms of sheaves Asm∙,∙⟶Awk∙,∙⟶Aps∙,∙.\mathcal A_{\mathrm{sm}}^{\bullet,\bullet} \longrightarrow \mathcal A_{\mathrm{wk}}^{\bullet,\bullet} \longrightarrow \mathcal A_{\mathrm{ps}}^{\bullet,\bullet}. The first sheaf is the original CLD sheaf. The last consists of piecewise smooth forms and is endowed with the polyhedral differentials dP′,dP′′d'_P,d''_P. Both morphisms preserve wedge products and the indicated differentials. Weakly smooth forms admit functorial pullback by analytic morphisms.

Open proof obligation (RC2). Supply the exact inclusions and comparison of the smooth, weakly smooth and delta-form categories; GJR Proposition 10.14 and its supporting constructions.

NAG-900305 · Proof status

Proposition 17.2.7. (Harmonicity detected by the weak differential). A real-valued function ff on XX is harmonic if and only if f∈Awk0,0(X),d′d′′f=0.f\in\mathcal A_{\mathrm{wk}}^{0,0}(X), \qquad d'd''f=0.

Open proof obligation (RC2). Prove injectivity of pullback on the harmonic chart and the converse with the stated face hypotheses; GJR Proposition 10.9 supplies injectivity.

NAG-900306 · Proof status

Example 17.2.8. (A harmonic function which is not CLD-smooth). Assume that KK is algebraically closed and that its residue field is not algebraic over a finite field. Gubler–Jell–Rabinoff construct a smooth proper genus-three curve with a theta graph as skeleton and an elliptic residual component EE at one vertex vv. Let zz be the origin of EE and choose a nontorsion point x∈E(K~)x\in E(\widetilde K).

The local construction produces an integral harmonic function hh whose residue at vv has degree zero but represents [z]−[x]≠0inPic⁡0(E)⊗ZR.[z]-[x]\ne0\quad\text{in}\quad \operatorname{Pic}^0(E)\otimes_{\mathbb Z}\mathbb R. On a smooth analytic curve, the quotient of harmonic germs by CLD-smooth harmonic germs at such a point is identified with this residual Picard space. Thus hh is weakly smooth by Proposition 17.2.7 and is not CLD-smooth. The construction and the nonvanishing obstruction are given in [GJR] (§17.3 and Proposition 17.4); the sheaf comparison used here is explained in Remark 15.10.

The obstruction is algebraic: degree zero only detects numerical triviality, while the nonzero Picard class prevents a representation by the local logarithms that give smooth harmonic functions. Hence the first inclusion in Theorem 17.2.6 can be strict even in bidegree (0,0)(0,0).

Remark (Cases in which the first inclusion is equality). On a smooth analytic curve potentially locally isomorphic to P1,an\mathbb P^{1,\mathrm{an}}, or over a residue field algebraic over a finite field, every harmonic function is CLD-smooth and Asm=Awk\mathcal A_{\mathrm{sm}}=\mathcal A_{\mathrm{wk}} [GJR] (Remark 15.10). Strictness is therefore not a statement about every analytic curve.

Strong currents and restriction of test spaces

Now assume that XX is Hausdorff, boundaryless, and of pure dimension dd. Write Awk,cr,s(X)\mathcal A_{\mathrm{wk},c}^{r,s}(X) for compactly supported weakly smooth forms, endowed with the locally convex topology of [GJR] (§12.1). In charts, its defining convergence uses fixed compact support control and uniform convergence of every derivative of the representing Lagerberg forms.

Definition 17.2.10. (Strong currents, in the bidegree convention of these notes). Set Dwkp,q(X):=(Awk,cd−p,d−q(X))′,\mathcal D_{\mathrm{wk}}^{p,q}(X) :=\bigl(\mathcal A_{\mathrm{wk},c}^{d-p,d-q}(X)\bigr)', where the prime denotes the continuous real dual. Its elements are the strong currents of bidegree (p,q)(p,q).

In [GJR] (§12.2), the notation is Dr,s=(Acr,s)′\mathcal D_{r,s}=(\mathcal A_c^{r,s})': the indices record the degrees of the test forms. Thus their Dr,s\mathcal D_{r,s} is our Dwkd−r,d−s\mathcal D_{\mathrm{wk}}^{d-r,d-s}. The explicit dual formula is the reliable way to translate conventions.

NAG-900307 · Conditional on stated inputs

Proposition 17.2.11. (The direction of the comparison map). Restriction of functionals gives a canonical map ρ:Dwkp,q(X)⟶Dsmp,q(X).\rho:\mathcal D_{\mathrm{wk}}^{p,q}(X) \longrightarrow \mathcal D_{\mathrm{sm}}^{p,q}(X). Using the assertion in [GJR] (§12.1) that the original CLD test-space topology is the induced subspace topology, this map is surjective. Its kernel consists precisely of the functionals vanishing on all original smooth test forms. An extension of a smooth-test current is unique if those forms are dense in the weak test space.

Proof. Let E=Asm,cd−p,d−q(X)E=\mathcal A_{\mathrm{sm},c}^{d-p,d-q}(X) and F=Awk,cd−p,d−q(X)F=\mathcal A_{\mathrm{wk},c}^{d-p,d-q}(X). Continuity of E↪FE\hookrightarrow F gives restriction F′→E′F'\to E'. If T∈E′T\in E', the induced subspace topology supplies a continuous seminorm PP on FF such that ∣T(e)∣≤P(e)|T(e)|\le P(e) for every e∈Ee\in E. The real Hahn–Banach theorem extends TT to a linear functional on FF dominated by PP, hence to an element of F′F'. This proves surjectivity. The kernel description is immediate; a continuous functional vanishing on a dense subspace vanishes everywhere. ◻

NAG-900308 · Open proof obligation

Remark (What this proposition does not identify). The Hahn–Banach extension is not canonical and does not, by this argument, preserve positivity. Density is a separate question; it does not follow from the inclusion of sheaves. In particular, enlarging the test forms does not by itself give an inclusion of the corresponding spaces of currents, or an isomorphism of those spaces.

The differential of a strong current is defined by duality with the same total-degree signs as for smooth-test currents. Restriction commutes with d′d' and d′′d''. Proper direct images and multiplication by weakly smooth forms are available [GJR] (§12.2). These constructions do not give arbitrary currents a wedge product or a pullback under every analytic morphism.

Theorem 17.2.13. (Poincare–Lelong against weak tests). Let ff be a regular meromorphic function on XX, meaning locally a nonzero quotient of analytic functions that are non-zero-divisors on the local rings. Then d′d′′[log⁡∣f∣]=δdiv⁡(f)inDwk1,1(X).d'd''[\log|f|]=\delta_{\operatorname{div}(f)} \quad\text{in}\quad\mathcal D_{\mathrm{wk}}^{1,1}(X).

Open proof obligation (RC2). Prove the Poincare–Lelong equality on the full weakly smooth test class of the statement; GJR Theorem 12.7 is supplied.

NAG-900309 · Proof status

Example 17.2.14. (Polyhedral differentiation misses a corner mass). On X=GmanX=\mathbb G_m^{\mathrm{an}}, put t=−log⁡∣T∣t=-\log|T| and u=∣t∣u=|t|. This is a piecewise smooth function. On the two cells t≥0t\ge0 and t≤0t\le0 it is affine, so dP′dP′′u=0.d'_P d''_P u=0. Its current derivative is different. For a smooth compactly supported test function φ\varphi on the real skeleton, ∫R∣t∣φ′′(t) dt=2φ(0).\int_{\mathbb R}|t|\varphi''(t)\,dt=2\varphi(0). Consequently d′d′′[u]d'd''[u] is twice the Dirac measure at the radius-one Gauss point. This explicitly distinguishes polyhedral differentiation from differentiation of currents. On this genus-zero curve weakly smooth forms agree with CLD-smooth forms; uu is not weakly smooth at that point.

For weakly smooth forms α\alpha, Stokes gives d′[α]=[d′α]d'[\alpha]=[d'\alpha] and d′′[α]=[d′′α]d''[\alpha]=[d''\alpha]. For arbitrary piecewise smooth forms, these identities require the missing face terms; they cannot be inferred from the injection into currents in [GJR] (§12.4).

Positivity, regularization, and the scope of comparison

Theorem 17.2.15. (Model positivity implies original current positivity). Let XX be good, separated, and strictly analytic. If hh is a semipositive R\mathbb R-PL function, then the original CLD current d′d′′[h]d'd''[h] is positive on X∖∂XX\setminus\partial X. The same conclusion holds locally for uniform limits of semipositive PL metric weights.

Open proof obligation (RC2). Prove metric/current positivity and the converse with their exact hypotheses; GJR Theorem 7.14 and Remark 7.16 are supplied.

NAG-900310 · Proof status

Theorem 17.2.16. (Regularization on smooth analytic curves). Let XX be a rig-smooth boundaryless analytic curve over KK, and let ff be subharmonic in Thuillier’s sense. For every relatively compact open subset U⊂XU\subset X, there is a decreasing net of weakly smooth psh functions on UU converging pointwise to f∣Uf|_U.

Open proof obligation (RC2). Prove regularization on curves in the stated net and weakly smooth scope; GJR Proposition 15.11 is supplied.

NAG-900311 · Proof status

NAG-900312 · Open proof obligation

Remark 17.2.17 (Canonical metrics and an explicitly corrected argument). The canonical tropicalization in Raynaud uniformization is harmonic [GJR] (Proposition 16.2). It need not be CLD-smooth [GJR] (Corollary 17.5). Remark 16.4 explicitly explains that Gubler–Künnemann’s earlier argument in the 2017 paper, Examples 8.15 and 9.17, used the incorrect smoothness assertion to conclude that canonical metrics are δ\delta-metrics. GJR leaves the latter conclusion open in the original δ\delta-metric theory. This corrects the cited argument; it does not prove that the metrics are not δ\delta-metrics.

The version concordance is recorded in Section 18.3.4; the elementary formula corrections are collected in Remark A.1.12.

17.3 Ordinary forms, G-forms, and harmonic forms in CLD25

The change of psh definition in CLD25 does not change its ordinary smooth test forms. Three sheaves must nevertheless be distinguished. We retain the good Hausdorff hypotheses on XX; whenever currents enter the comparison, XX is purely nn-dimensional and boundaryless.

Definition 17.3.1. (The analytic affine and PL structures). An affine function in the sense of CLD25 is, locally for the ordinary topology, of the form alog⁡∣f∣+c,a∈Q,c∈R,f∈OX×.a\log|f|+c, \qquad a\in\mathbb Q,\quad c\in\mathbb R, \quad f\in\mathcal O_X^{\times}. A PL function is a continuous function which has this description GG-locally, on analytic domains. Finite sums are included in this description by clearing rational denominators and multiplying units. These are the structures of [CLD25] (Proposition 12.4.1 and §12.4.6). They use rational slopes and real constants; this notation does not silently mean arbitrary R\mathbb R-PL functions.

Here a GG-cover consists of analytic domains such that, at each point, finitely many members containing that point have union a neighbourhood. An ordinary open cover is a GG-cover, but a GG-cover need not provide a single member which is a neighbourhood of a point on an interface.

Definition 17.3.2. (The two smooth sheaves). For a polyhedral subset PP of a real affine space EE, let Ap,q(P)A^{p,q}(P) denote smooth ambient superforms restricted to PP, modulo forms whose restrictions to the tangent spaces of every cell in PP vanish. More precisely, this is the quotient sheaf of [CLD25] (§1.9.1–§1.9.3); it is independent of subdivision.

For an analytic domain VV, form the direct limit Pp,q(V)=lim→⁡(h:V⟶E, P)Ap,q(P),\mathcal P^{p,q}(V)= \varinjlim_{(h:V\longrightarrow E,\,P)} A^{p,q}(P), where hh has PL coordinates, PP contains h(V)h(V), and transition maps are affine maps between ambient spaces. Products of charts give common refinements. The associated sheaf on the analytic GG-site is denoted AG,25p,qA^{p,q}_{G,25} here. Its sections are the smooth GG-forms of [CLD25] (§3.3.3–Definition 3.3.4).

Write π:XG→X\pi:X_G\to X for the map to the ordinary topological site. The ordinary smooth sheaf A25p,q⊆π∗AG,25p,qA^{p,q}_{25}\subseteq\pi_*A^{p,q}_{G,25} consists of forms represented, on an ordinary neighbourhood of each point, by an affine chart and a smooth superform. This is [CLD25] (Definition 6.2.1).

Proposition 17.3.3. (The ordinary sheaf is unchanged). There is a canonical identification Asmp,q=A25p,q.A^{p,q}_{\mathrm{sm}}=A^{p,q}_{25}. It respects wedge product and the two differentials. The operators called JJ correspond by the sign conversion already specified above.

Open proof obligation (RC2). Prove both local moment-map presentation and matching nullity of representatives on the common product chart; CLD25 Proposition 15.1.1(b), CLD12 Lemma 3.2.2 and CLD25 Proposition 3.4.4.

NAG-900313 · Proof status

The distinction between the two sheaves is the covering condition, not the replacement of analytic units by arbitrary continuous coordinates. Indeed, [CLD25] (Proposition 15.1.1(a)) still presents every GG-form by analytic moment maps, but on a GG-cover. The ordinary smooth forms remain the test objects in [CLD25] (§7.1.1).

Definition 17.3.4. (The harmonic subsheaf). A PL function hh is harmonic in [CLD25] (Definition 7.8.2) if d′d′′[h]=0.d'd''[h]=0. More general harmonic functions are defined to be local finite real linear combinations of such PL harmonic functions. The sheaf Aharm,25p,qA^{p,q}_{\mathrm{harm},25} consists of GG-forms locally represented by h∗αh^*\alpha, where h=(h1,…,hr)h=(h_1,\ldots,h_r) has PL harmonic coordinates and α\alpha is a smooth superform on a neighbourhood of a polyhedron containing h(U)h(U); see [CLD25] (§7.8.7).

Affine functions are harmonic, so these definitions give inclusions Asmp,q=A25p,q ⊆ Aharm,25p,q ⊆ π∗AG,25p,q.A^{p,q}_{\mathrm{sm}}=A^{p,q}_{25} \ \subseteq\ A^{p,q}_{\mathrm{harm},25} \ \subseteq\ \pi_*A^{p,q}_{G,25}. All three are differential algebras, but integration into currents has different differential properties. Denote the operators on GG-forms by dG′,dG′′d'_G,d''_G. Every PL function satisfies dG′dG′′h=0,d'_Gd''_G h=0, because it is affine on a GG-cover. This does not imply d′d′′[h]=0d'd''[h]=0: distributional differentiation can detect the interfaces. Already on the real line, h(t)=∣t∣h(t)=|t| has zero piecewise Hessian, whereas its distributional Hessian is 2δ02\delta_0. The harmonic condition is therefore an additional condition on a PL function.

For a harmonic form ω\omega, the extra boundary terms vanish: d′[ω]=[dG′ω],d′′[ω]=[dG′′ω].d'[\omega]=[d'_G\omega],\qquad d''[\omega]=[d''_G\omega]. This is [CLD25] (Theorem 7.8.5 and Corollary 7.8.8). The resulting map to currents is injective in the present boundaryless setting by [CLD25] (Corollary 7.3.9 and Theorem 12.11.1). It does not redefine currents as duals of all GG-forms.

NAG-900314 · Open proof obligation

Remark 17.3.5 (Comparison with GJR requires a numerical theorem). CLD25 starts with a vanishing original curvature current for a PL function. GJR defines harmonicity through numerical triviality of the residual model class, with its own integral and real PL classes; see Definition 17.2.1. An identification of the two harmonic sheaves requires a PL curvature comparison with matching hypotheses, and then a comparison of their chart presentations. The degree-one vanishing assertion of [CLD25] (Theorem 18.8.4) addresses the numerical/current step for its integral PL metrics; its scope is discussed in Section 19.3. No unrestricted identification with the GJR weakly smooth sheaf is assumed here. Also, CLD25 does not define every continuous solution of d′d′′[h]=0d'd''[h]=0 to be harmonic: [CLD25] (Remark 7.8.3) explicitly distinguishes that question from the PL-generated definition.

17.4 Graph forms and harmonic cohomology

The first-order equations on a graph

Let Σ\Sigma be a finite connected metric graph without boundary, and choose orientations of its edges. On an open edge with coordinate tt, forms have the familiar expressions f(t)f(t), a(t)d′ta(t)d't, b(t)d′′tb(t)d''t, and c(t)d′t∧d′′tc(t)d't\wedge d''t. At vertices, the smoothness conditions are those of the harmonic-coordinate graph calculus. In particular, the first-order compatibility is tied to a weighted conservation law.

For constant coefficients on oriented edges, closed (1,0)(1,0)-forms correspond to flows satisfying Kirchhoff’s law: the weighted outgoing coefficients sum to zero at each vertex. In the unweighted case, the incidence matrix has rank ∣V∣−1|V|-1 for a connected graph. Its kernel therefore has dimension ∣E∣−∣V∣+1=b1(Σ).|E|-|V|+1=b_1(\Sigma). This linear algebra explains the occurrence of the graph genus in H1,0H^{1,0}. Subdividing an edge adds one vertex and one edge and leaves the dimension unchanged.

Why integration detects the top cohomology

For a compact graph without boundary, Stokes gives ∫Σd′′α=0\int_\Sigma d''\alpha=0 for a (1,0)(1,0)-form α\alpha. Thus integration descends to a functional on H1,1H^{1,1}. Conversely, a top-degree form of total integral zero can be integrated along edges to produce a primitive; constants of integration are chosen so that the vertex conditions hold. The obstruction is precisely the total integral.

This gives H1,1≃RH^{1,1}\simeq\mathbb R in the connected compact case. If the graph has boundary, flux may escape through boundary vertices, changing the obstruction and the cohomology. The source gives the exact boundary-dependent dimension formulas. The boundary must therefore be specified before applying a duality statement.

The circle as a complete example

For a circle of length ℓ\ell, take a periodic coordinate on its universal cover. A closed (1,0)(1,0)-form has a constant coefficient after passing to its cohomology class, and its period is the invariant. A (0,1)(0,1)-form b(t)d′′tb(t)d''t is exact precisely when its integral over one period is zero. A (1,1)(1,1)-form has a primitive of the required periodic type precisely when its total integral is zero. Hence the four cohomology groups in bidegrees (0,0),(1,0),(0,1),(1,1)(0,0),(1,0),(0,1),(1,1) all have dimension one.

The pairing of a,d′ta,d't with b,d′′tb,d''t is abℓab\ell in the simplest constant representatives. It is nondegenerate. This gives a concrete model for the Poincaré duality proved in the curve theory, while retaining the metric length in the pairing.

The good-reduction curve as a contrasting example

A smooth proper curve with good reduction can have a one-point skeleton and first Betti number zero. Its weakly smooth Dolbeault groups in bidegrees (1,0)(1,0) and (0,1)(0,1) are therefore zero in the cited theorem, while H1,1H^{1,1} remains one-dimensional by integration. If the algebraic genus is positive, algebraic differential forms still exist; they belong to a different cohomology theory.

This example prevents a misleading inference from the graph picture. The weakly smooth calculus is designed to have robust tropical analytic properties and duality. It is not claimed to reproduce every algebraic Hodge number. To compare it with algebraic de Rham theory, one would need an additional construction accounting for the residual geometry.

From graph forms to analytic forms

The passage to Berkovich curves uses skeleta adapted to the forms and harmonic maps between graph presentations. It is not a claim that every form factors through one minimal skeleton. Enlarging a skeleton can add trees and local coordinates while preserving the appropriate cohomological information. Compatibility under these changes is what makes the analytic sheaf and its cohomology intrinsic.

This gives a successful example of the general strategy discussed in Chapter 24: first understand a finite tropical complex, then prove compatibility under refinement, then identify the resulting analytic cohomology. The proof contains more than a topological deformation retraction, because the differential and vertex-smoothness conditions also have to be preserved.

References. [GJRc] (Sections 2–3 and Theorems 6.4.3, 6.4.5).

17.5 The differential complex on a compact graph

We give the graph calculation underlying the curve comparison. It is useful both as a complete example of a differential theory and as a warning about what a skeleton retains. Let Γ\Gamma be a finite connected metric graph with at least one edge. We first take all edge weights to be one and no designated boundary. Multiple edges and loops are allowed; a loop contributes two outgoing germs at its vertex.

Choose an orientation of every edge ee and a parameter te∈[0,ℓe]t_e\in[0,\ell_e]. A smooth function is continuous on Γ\Gamma, smooth up to each endpoint on each edge, and satisfies the following local conditions. At a vertex of valency one it is constant nearby. At valency two it is an ordinary smooth function across the vertex. At higher valency its outgoing first derivatives have sum zero.

A (1,0)(1,0)-form has edgewise expressions ae(te)d′tea_e(t_e)d't_e. Reversing orientation reverses the coefficient. At a leaf it vanishes nearby, at a two-valent vertex the one-form extends smoothly across, and at higher valency its outgoing coefficients have sum zero. The definition of (0,1)(0,1)-forms is identical with d′′ted''t_e. A (1,1)(1,1)-form has expressions ce(te)d′te∧d′′tec_e(t_e)d't_e\wedge d''t_e, unchanged by orientation reversal. It vanishes near leaves, is smooth across two-valent vertices, and has no extra matching condition at higher-valent vertices. These are the unweighted GJR graph conventions.

The edgewise formulas are d′′f=fe′ d′′te,d′′(ae d′te)=−ae′ d′te∧d′′te,∫Γc=∑e∫0ℓece(t) dt.d''f=f'_e\,d''t_e, \qquad d''(a_e\,d't_e)=-a'_e\,d't_e\wedge d''t_e, \qquad \int_\Gamma c=\sum_e\int_0^{\ell_e}c_e(t)\,dt. The vertex conditions ensure that these derivatives lie in the indicated form spaces. They also imply Stokes: the endpoint terms of ∫Γd′′a\int_\Gamma d''a sum to the outgoing coefficient sums, hence vanish.

The incidence matrix

Let V,EV,E denote the vertices and the chosen oriented edges. Define B:RE→RVB:\mathbb R^E\to\mathbb R^V by putting +1+1 at an edge’s tail and −1-1 at its head; a loop column is zero. Connectedness implies im⁡B={z∈RV:∑vzv=0},dim⁡ker⁡B=∣E∣−∣V∣+1.\operatorname{im}B=\{z\in\mathbb R^V:\textstyle\sum_vz_v=0\}, \qquad \dim\ker B=|E|-|V|+1. Indeed the columns have coordinate sum zero. Along a path, signed sums of columns produce the difference of the two endpoint basis vectors. Since any vertex can be joined to a fixed one, these differences span the zero-sum hyperplane. We write b=∣E∣−∣V∣+1=b1(Γ)b=|E|-|V|+1=b_1(\Gamma).

Proposition 17.5.1. The d′′d''-closed (1,0)(1,0)-forms are exactly the edgewise constant forms whose coefficient vector belongs to ker⁡B\ker B. Consequently H1,0(Γ)≃ker⁡BH^{1,0}(\Gamma)\simeq\ker B and h1,0=bh^{1,0}=b.

Proof. The displayed formula for d′′d'' forces ae′=0a'_e=0 on each edge. The vertex conditions then reduce to Ba=0Ba=0. At a leaf this equation forces the incident constant to be zero, so the required vanishing on a neighbourhood follows as well. At valency two it makes the constants match with the orientation signs. Conversely these conditions give a smooth closed form. There are no forms of type (1,−1)(1,-1), hence no boundaries in this group. ◻

Theorem 17.5.2. (Periods and exactness). Integration around cycles induces an isomorphism H0,1(Γ)≃Hom⁡(H1(Γ,R),R).H^{0,1}(\Gamma)\simeq \operatorname{Hom}(H_1(\Gamma,\mathbb R),\mathbb R).

Proof. For a (0,1)(0,1)-form β=be(t)d′′te\beta=b_e(t)d''t_e, set Ie=∫0ℓebe(t)dtI_e=\int_0^{\ell_e}b_e(t)dt. The period on a cycle z∈ker⁡Bz\in\ker B is ∑ezeIe\sum_e z_e I_e. If β=d′′f\beta=d''f, these integrals are endpoint differences and all periods vanish.

Conversely suppose all periods vanish. Fix a basepoint and define f(x)f(x) by integrating β\beta along a path from it to xx. Any two paths differ by cycles, so ff is well defined. It is continuous and has derivative beb_e on each edge. Near a leaf, be=0b_e=0, so ff is constant. At a two-valent vertex, smooth matching of the one-form gives smooth matching of its primitive. At a higher-valent vertex, the outgoing derivatives of ff are the outgoing coefficients of β\beta, whose sum is zero. Thus ff belongs to A0,0(Γ)A^{0,0}(\Gamma) and β=d′′f\beta=d''f.

To prove surjectivity, represent a functional on ker⁡B\ker B by some vector I∈REI\in\mathbb R^E. On each edge choose a smooth bump supported strictly inside the edge and of integral IeI_e. These coefficients define a (0,1)(0,1)-form because they vanish near every vertex. Its periods are the prescribed functional. ◻

This proof is stronger than a dimension count: it supplies an exactness criterion and a primitive. It also shows why smoothness at vertices cannot be ignored. Edgewise primitives always exist, but their values and derivatives must satisfy the graph’s matching conditions.

Theorem 17.5.3. (Top-degree exactness). The image of d′′:A1,0(Γ)→A1,1(Γ)d'':A^{1,0}(\Gamma)\to A^{1,1}(\Gamma) is precisely the kernel of integration. Hence H1,1(Γ)≃RH^{1,1}(\Gamma)\simeq\mathbb R.

Proof. Stokes gives one inclusion. For the other, let ce(t)d′te∧d′′tec_e(t)d't_e\wedge d''t_e have total integral zero and put Je=∫0ℓece(t)dtJ_e=\int_0^{\ell_e}c_e(t)dt. An edgewise solution is ae(t)=Ae−∫0tce(s)ds.a_e(t)=A_e-\int_0^t c_e(s)ds. At a tail the outgoing coefficient is AeA_e; at a head it is −Ae+Je-A_e+J_e. Therefore the vertex equations are (BA)v=−∑e: e+=vJe.(BA)_v=-\sum_{e:\,e^+=v}J_e. The right side has total sum −∑eJe=0-\sum_eJ_e=0. By the description of im⁡B\operatorname{im}B, constants AeA_e solving these equations exist. They give the value matching at two-valent vertices and the coefficient sums at higher-valent ones. Derivative matching at a two-valent vertex follows by differentiating ae′=−cea'_e=-c_e and using the smooth matching of cc. At a leaf, cc vanishes nearby and the vertex equation sets aa to zero there. Thus aa is an admitted (1,0)(1,0)-form and d′′a=cd''a=c.

An interior bump on any one edge gives a (1,1)(1,1)-form of nonzero integral, proving that integration is surjective on the quotient. ◻

Corollary 17.5.4. (Graph duality). For this compact connected graph, h0,0=h1,1=1,h1,0=h0,1=b.h^{0,0}=h^{1,1}=1,\qquad h^{1,0}=h^{0,1}=b. Integration of wedge products gives perfect pairings between complementary bidegrees.

Proof. A function with zero edge derivatives is constant by connectedness, giving h0,0=1h^{0,0}=1. The other dimensions were proved above. For (1,0)(1,0) and (0,1)(0,1), write the first class as a constant flow a∈ker⁡Ba\in\ker B and the second by its edge integrals II. Their pairing is ∑eaeIe\sum_ea_eI_e. It is the natural pairing of ker⁡B\ker B with its dual from Theorem 17.5.2, hence perfect. Constants pair perfectly with the integral coordinate of H1,1H^{1,1}. Stokes ensures independence of representatives. ◻

What changes when a boundary is designated

Let ∂Γ⊂V\partial\Gamma\subset V be nonempty and impose no vertex matching condition there, apart from continuity of functions. Write BintB_{\mathrm{int}} for the incidence matrix with only the rows indexed by V∖∂ΓV\setminus\partial\Gamma. These rows are independent. Indeed a linear relation, extended by zero at boundary vertices, gives a vector in ker⁡Bt\ker B^t, hence a constant vector; because it is zero at one boundary vertex, it is zero everywhere.

Proposition 17.5.5. For a connected graph with nonempty designated boundary, h0,0=1,h0,1=b,h1,0=b+∣∂Γ∣−1,h1,1=0.h^{0,0}=1,\quad h^{0,1}=b,\quad h^{1,0}=b+|\partial\Gamma|-1,\quad h^{1,1}=0.

Proof. Closed (1,0)(1,0)-forms are the constant flows in ker⁡Bint\ker B_{\mathrm{int}}, whose dimension is ∣E∣−∣V∣+∣∂Γ∣|E|-|V|+|\partial\Gamma|. The period proof for H0,1H^{0,1} is unchanged: path integration still gives a continuous primitive and there are fewer endpoint conditions to check. The proof of top-degree exactness now solves the vertex equations only at interior vertices. Since BintB_{\mathrm{int}} is onto, these equations always have a solution, with no total-integral constraint. Thus every (1,1)(1,1)-form is exact. ◻

For an interval with both endpoints designated as boundary, h1,0=1h^{1,0}=1 and h1,1=0h^{1,1}=0. For the same interval with no designated boundary, the leaf conditions give h1,0=0h^{1,0}=0 and h1,1=1h^{1,1}=1. The topological interval alone does not determine this differential complex: the boundary convention is part of its data.

Weights and the curve comparison

If an edge has positive weight wew_e, put t0,e=te/wet_{0,e}=t_e/w_e, so its new length is ℓe/we\ell_e/w_e. Pullback replaces a one-form coefficient aea_e by weaew_ea_e and a two-form coefficient cec_e by we2cew_e^2c_e. The weighted vertex conditions become the preceding unweighted ones. Since d/dt0,e=wed/dted/dt_{0,e}=w_e d/dt_e, this is an isomorphism of differential algebras. Its integration rule agrees with ∑ewe∫ce(te)dte\sum_e w_e\int c_e(t_e)dt_e.

The passage from these graph groups to weakly smooth forms on a Berkovich curve is a separate theorem. In the setting of GJR’s curve paper, a suitable strictly semistable model gives a pullback from graph forms inducing the cohomology isomorphism. Finite extension and Galois descent handle a curve without such a model over the initial field. We use that comparison, not a claim that every continuous retraction induces an isomorphism of differential complexes.

In particular, a skeleton consisting of one isolated point has no graph (1,1)(1,1)-forms under the definitions above. It cannot by itself compute the top-degree cohomology of a proper analytic curve. The comparison theorem uses an appropriate model; one may blow up to avoid smooth connected components of the special fibre and obtain a nontrivial skeleton. Attaching the resulting tree does not change b1b_1, but it supplies the graph differential data needed for top-degree integration.

17.6 Residual genus and topological genus

The discrepancy between algebraic genus and graph genus can be calculated without differential forms. Let C0C_0 be a connected proper nodal curve over an algebraically closed field. Let C~v\widetilde C_v be its normalized irreducible components and let Γ\Gamma be its dual graph. There is one vertex per component and one edge per node, including loop edges for self-nodes.

Proposition 17.6.1. The arithmetic genus satisfies pa(C0)=b1(Γ)+∑vg(C~v).p_a(C_0)=b_1(\Gamma)+\sum_v g(\widetilde C_v).

Proof. Normalization gives an exact sequence 0⟶OC0⟶⨁vν∗OC~v⟶⨁e∈Eke⟶0.0\longrightarrow\mathscr O_{C_0} \longrightarrow\bigoplus_v\nu_*\mathscr O_{\widetilde C_v} \longrightarrow\bigoplus_{e\in E} k_e\longrightarrow0. At a node, the last map takes the difference of the two branch values. Its exactness follows from the local description k[[x,y]]/(xy)k[[x,y]]/(xy): a pair of power series in xx and yy comes from this ring exactly when its constant terms agree. Taking Euler characteristics yields χ(OC0)=∑v(1−g(C~v))−∣E∣.\chi(\mathscr O_{C_0})=\sum_v(1-g(\widetilde C_v))-|E|. Since C0C_0 is connected and proper, h0(OC0)=1h^0(\mathscr O_{C_0})=1. Thus pa=1−χ=∣E∣−∣V∣+1+∑vg(C~v)p_a=1-\chi=|E|-|V|+1+\sum_vg(\widetilde C_v), as asserted. ◻

For a proper semistable model of a smooth curve, constancy of Euler characteristic in the flat family identifies this arithmetic genus with the genus of the generic fibre. The skeleton retains the graph contribution b1(Γ)b_1(\Gamma); the remaining summands are genera of residual components. Good reduction gives a one-vertex dual graph and all the genus is residual. Totally degenerate reduction has genus-zero residual components and all the genus is graph-theoretic.

Consequently the weakly smooth cohomology formula h1,0=b1h^{1,0}=b_1 cannot be identified with the dimension gg of algebraic differentials without an additional hypothesis. For a good-reduction elliptic curve these dimensions are respectively zero and one. This does not contradict either theory: their complexes retain different information. In a proposed tropical-to-Berkovich comparison, the relevant sheaf and the treatment of residual components must be specified before drawing a Hodge-theoretic conclusion.

References. Graph definitions and the unweighting are [GJRc] (Section 2.2); the graph calculation above is an explicit incidence-matrix proof of the cohomology and duality statements. The analytic comparison is [GJRc] (Theorems 6.4.2–6.4.5).

17.7 Residual divisors and harmonic coordinates

A harmonic function on a curve satisfies a conservation law for outgoing slopes. An original smooth harmonic function satisfies an additional local algebraic condition. The difference is visible on the residual curve of a type II point. This section calculates that difference and then revisits the canonical tropicalization example of Gubler–Jell–Rabinoff with consistent coordinates.

Work over a complete algebraically closed nontrivially valued field KK. At a type II point vv of a smooth analytic curve, the residual function field determines a smooth projective curve CvC_v over K~\widetilde K. Its closed points correspond to tangent directions. We use the slope formula: for an analytic meromorphic function whose normalized reduction at vv is f~\widetilde f, outgoing slopes of −log⁡∣f∣-\log|f| are the orders of f~\widetilde f, with the sign reversed for log⁡∣f∣\log|f|. The identification of tangent directions and this slope formula are geometric inputs; they are not consequences of the abstract graph calculation.

The sign can be checked in the open disc direction at the Gauss point: with outward distance t=−log⁡∣T∣t=-\log|T|, the reduction of TT has order one at 00, while log⁡∣T∣=−t\log|T|=-t has outgoing slope −1-1.

Proposition 17.7.1. Let hh be a harmonic germ represented by a PL function on a finite star at vv. If sps_p is its outgoing slope in the direction p∈Cvp\in C_v, then Dh=∑psp[p],deg⁡Dh=∑psp=0.D_h=\sum_p s_p[p],\qquad \deg D_h=\sum_p s_p=0. If hh is a real linear combination of logarithms of local analytic units, then DhD_h has zero class in Pic⁡0(Cv)⊗ZR\operatorname{Pic}^0(C_v)\otimes_{\mathbb Z}\mathbb R.

Proof. Only finitely many directions have a nonzero slope, so the divisor is defined. Harmonicity is precisely the vanishing of the outgoing slope sum; the residual field is algebraically closed, so all residual points have degree one. For h=log⁡∣f∣h=\log|f|, the slope formula identifies DhD_h with −div⁡(f~)-\operatorname{div}(\widetilde f), hence with a principal divisor. Multiplying by a real scalar and adding preserves the zero divisor class. Constants have all slopes zero. ◻

The converse is not asserted from this calculation alone: lifting a residual principal relation to analytic units is a local analytic theorem. Thuillier’s comparison identifies the relevant quotient of harmonic germs by smooth harmonic germs with Pic⁡0(Cv)⊗R\operatorname{Pic}^0(C_v)\otimes\mathbb R in the curve setting used below. GJR records this input in Remark 15.10 and the proof of Proposition 17.4. It also records that original smooth harmonic germs are the sheaf generated by logarithms of analytic units. We use these two comparison results explicitly.

Corollary 17.7.2. In this setting, a harmonic germ with nonzero class [Dh][D_h] cannot be an original smooth function. A degree-zero residual divisor is therefore not enough to guarantee original smoothness.

Proof. If the germ were original smooth, it would be a smooth harmonic germ, hence a real linear combination of logarithms of analytic units. Proposition 17.7.1 would make its residual class zero. ◻

Example 17.7.3. (Why the residue field matters). If a smooth projective curve is defined over an algebraic closure of a finite field, every point of its Jacobian is torsion. Indeed the variety and any chosen point descend to a finite field; the group of points over that finite field is finite. Thus Pic⁡0(Cv)⊗R=0\operatorname{Pic}^0(C_v)\otimes\mathbb R=0. Over other algebraically closed fields, an elliptic residual curve can have nontorsion points, and their degree-zero divisor classes survive tensoring with R\mathbb R. For an abelian group AA, the kernel of A→A⊗QA\to A\otimes\mathbb Q is its torsion subgroup, and extension from Q\mathbb Q to R\mathbb R is faithful. This justifies the last assertion. The vanishing of the residual obstruction is consistent with the equality of the two harmonic sheaves in the finite-residue-field case; the analytic equality itself uses the cited comparison theorem.

17.8 A consistent theta-graph calculation

Let a,b,ca,b,c be three edges from PP to QQ, all oriented in that direction, with lengths A,B,C>0A,B,C>0. Choose the cycles γ1=c−a,γ2=c−b.\gamma_1=c-a,\qquad \gamma_2=c-b. The notation means traverse cc from PP to QQ and return along the indicated other edge. Use the basis dual to (γ1,γ2)(\gamma_1,\gamma_2) on H1(Θ,R)H^1(\Theta,\mathbb R). The edge-length pairing has matrix G=(A+CCCB+C),det⁡G=AB+AC+BC>0.G=\begin{pmatrix}A+C&C\\ C&B+C\end{pmatrix},\qquad \det G=AB+AC+BC>0. Thus GZ2G\mathbb Z^2 is a full lattice in R2\mathbb R^2, and the graph Jacobian is R2/GZ2\mathbb R^2/G\mathbb Z^2. The positivity also follows directly from the positive edge-length pairing on the cycle space.

Proposition 17.8.1. A lift of the Abel–Jacobi map based at PP, on a star around PP, has the following formulas, where tt is distance from PP: a: (−t,0),b: (0,−t),c: (t,t).a:\ (-t,0),\qquad b:\ (0,-t),\qquad c:\ (t,t). Modulo GZ2G\mathbb Z^2, these formulas agree at QQ and define a continuous map on the whole graph. Each coordinate is harmonic near PP.

Proof. Integrating the two harmonic one-forms corresponding to γ1,γ2\gamma_1,\gamma_2 gives coefficient vectors (−1,0)(-1,0) on aa, (0,−1)(0,-1) on bb, and (1,1)(1,1) on cc. Their sums at PP vanish. At QQ, the three lifted values are (−A,0)(-A,0), (0,−B)(0,-B), and (C,C)(C,C). Their differences from the third are the negatives of the first and second columns of GG, so they represent the same point of the quotient. The period of each cycle is the corresponding column of GG, as required by the Abel–Jacobi definition. ◻

Attach an elliptic curve EE at PP, with the directions a,b,ca,b,c marked by distinct points x,y,zx,y,z, where zz is the identity and x,yx,y are nontorsion. Attach a projective line at QQ, with three distinct markings. The lifting theorem for metrized complexes, used in GJR Section 17.3, realizes this data by a smooth proper analytic curve after choosing the lengths in the value group. This realization and compatibility of analytic and tropical Abel–Jacobi maps are the two geometric inputs in the example.

Theorem 17.8.2. (The residual obstruction in the example). The two lifted coordinate functions are harmonic but not original smooth near PP. Their outgoing-slope divisors are respectively D1=[z]−[x],D2=[z]−[y].D_1=[z]-[x],\qquad D_2=[z]-[y]. Consequently the canonical tropicalization of the Jacobian of the realized curve need not be an original smooth tropicalization.

Proof. Proposition 17.8.1 gives the slope lists (−1,0,1)(-1,0,1) and (0,−1,1)(0,-1,1). Each has sum zero, giving harmonicity. Matching the directions with the three residual markings gives the displayed divisors. The standard identification E(K~)≃Pic⁡0(E)E(\widetilde K)\simeq\operatorname{Pic}^0(E) sends xx to [x]−[z][x]-[z]. Nontorsion of x,yx,y and Example 17.7.3 show that both classes remain nonzero after tensoring with R\mathbb R. Corollary 17.7.2 proves nonsmoothness. The analytic Abel–Jacobi map identifies these local functions with pullbacks of local coordinates of the canonical tropicalization of the Jacobian. If that tropicalization were original smooth, pullback of original smooth functions by the analytic map would make these functions original smooth, a contradiction. The realized curve has genus 2+1+0=32+1+0=3, using the graph and residual genus formula of Proposition 17.6.1; hence this produces the abelian threefold of GJR. ◻

Remark 17.8.3 (Two coordinate slips in the cited version). In GJR v3, Section 17.2 names the first cycle as c−bc-b and the second as c−ac-a, but its Gram matrix and Abel–Jacobi formulas use the opposite ordering. Taking (c−a,c−b)(c-a,c-b) as above makes the displayed formulas consistent. In the proof of Proposition 17.4, the printed first-coordinate slope list (1,−1,0)(1,-1,0) along (a,b,c)(a,b,c) is incompatible with both those formulas and the subsequently stated divisor [z]−[x][z]-[x]. The list (−1,0,1)(-1,0,1) gives the stated divisor. These are indexing and slope corrections; they preserve the non-smoothness example. In particular, one should not use the printed list to replace [z]−[x][z]-[x] by [x]−[y][x]-[y]: distinct nontorsion x,yx,y do not by themselves imply that x−yx-y is nontorsion.

17.9 Consequences for the choice of differential sheaf

A harmonic tropicalization is built from local numerical harmonic functions, including the functions above. Pullback of a real smooth function along such a map is weakly smooth by definition. Original smoothness, on the other hand, requires presentation through logarithms of analytic units. The obstruction is therefore a change of coordinate class, not a failure of the elementary real chain rule.

For a smooth real function FF and harmonic coordinates h1,…,hrh_1,\ldots,h_r, the weakly smooth calculus has d′d′′F(h)=∑i,jFij(h) d′hi∧d′′hj.\mathrm d'\mathrm d''F(h)=\sum_{i,j}F_{ij}(h)\,d'h_i\wedge d''h_j. The first-derivative terms vanish in that calculus because d′d′′hi=0\mathrm d'\mathrm d''h_i=0 there. This is not a proof that hih_i are original smooth functions, and it cannot be inserted into the original smooth Leibniz rule against an arbitrary original current without a comparison theorem. Chapter 14 isolates exactly that background-current kernel requirement.

The example also explains the limitation of an earlier argument about canonical metrics. GK17 used smoothness of canonical tropicalizations to claim that all canonical metrics on abelian varieties are original delta-metrics. GJR Remark 16.4 explicitly withdraws that argument: its conclusion requires a further proof, while weakly smooth coordinates are available. The non-smooth coordinate example does not by itself prove that a particular canonical metric fails to be an original delta-metric; a metric might admit another presentation. This distinction between a failed presentation argument and a false existence statement is essential when citing the correction.

References. [GJR] (Remark 15.10, Proposition 15.11, Proposition 16.2, Remark 16.4, Sections 17.1–17.4); the graph calculation above fixes all orientations before applying the cited geometric comparisons.