20 · Chern measures and intersection degrees

22 NAG Tags

20.1 Residual bundles and the recovery of model measures

What a formal metric contributes

Let X\mathfrak X be a formal model of a proper analytic space and let L\mathfrak L be a line bundle on it. A local formal frame has norm one on the generic fibre. This rule determines the model metric on L=L∣XL=\mathfrak L|_X. A change of formal frame is multiplication by a formal unit, whose norm is compatible with the gluing. Tensor powers allow rational model metrics.

The resulting weights are piecewise linear in the appropriate analytic sense. Their curvature is generally singular. A basic example is the standard metric on OP1(1)\mathcal O_{\mathbb P^1}(1), whose curvature is concentrated at the Gauss point. Calling this metric formal does not mean its original CLD weight is a smooth function.

Why residual geometry enters positivity

At an analytic point, a PL metric determines residual line-bundle data on a suitable reduction of the germ. Positivity of that data is tested by degrees on residual curves. This can see geometry invisible in one real coordinate picture. If the residual variety contains a curve on which the line bundle has negative degree, a numerical semipositivity assertion fails even when a chosen local polyhedral presentation looks harmless.

CLD12 develops the residual construction and uses residual positivity to obtain approximation results and identify model Monge–Ampère measures. It does not thereby prove every converse between continuous current positivity and model semipositivity. Chapter 19 develops the later model and curve tests; Chapter 19 states the stronger residual comparison claimed in CLD25 with its precise regularity scope.

The special-fibre formula

For orientation, assume a normal proper flat model over a discrete valuation ring, with special fibre ∑YmYY\sum_Y m_Y Y, and integral model line bundles L1,…,Ln\mathfrak L_1,\ldots,\mathfrak L_n. In the usual model-measure normalization, c1(L‾1)⋯c1(L‾n)=∑YmYdeg⁡Y(L1⋯Ln) δξY.c_1(\overline L_1)\cdots c_1(\overline L_n) =\sum_Y m_Y\deg_Y(\mathfrak L_1\cdots\mathfrak L_n)\,\delta_{\xi_Y}. Here ξY\xi_Y is the divisorial point associated with the component YY. Rational model metrics introduce the corresponding tensor-power denominators. The conventions for ξY\xi_Y and the component multiplicities must be retained together; omitting mYm_Y while keeping the same normalized valuation changes the formula.

The mass is the generic-fibre intersection degree. In particular, it does not acquire an extra factor −log⁡∣π∣-\log|\pi| merely because the underlying absolute value has that scale. That logarithmic scale enters the values of vertical model functions and local heights; the Chern measure itself is normalized to have algebraic degree as mass. A detailed blow-up calculation in Section 20.6 checks both facts simultaneously.

How the analytic and model constructions meet

The comparison is a theorem, with two constructions to compare. One side comes from approximating local metric weights and taking limits of differential products. The other side is a finite sum of algebraic intersection degrees. The proof identifies the analytic mass near a divisorial point with the corresponding residual intersection number and shows that no other mass remains. Locality, Poincaré–Lelong, and tropical degree calculations are all used in this passage.

Once this comparison is established for suitable model metrics and the appropriate continuity theorem is available, uniform limits extend the agreement to a larger class of semipositive metrics. It remains necessary to say which approximation is being used: local smooth approximation, global approximation on one line bundle, or approximation by semipositive models. Their equivalence is not part of the notation c1(L‾)c_1(\overline L).

References. [CLD12] (§§6.1–6.9); [CL06].

20.2 Metrics and residual curvature in CLD25 Chapters 17–18

Vector bundles first, curvature products for line bundles

CLD25 Chapter 17 develops metrics on vector bundles, their PL forms, and residual bundles. The statements about curvature products in CLD25 Chapter 18 principally use line bundles and first Chern currents. The presence of vector-bundle metrics in the foundations should not be read as a complete higher-Chern-form or Bott–Chern theory for every such metric.

For a line bundle, a metric is locally a weight, but the transition rule is constrained by analytic units. Residual reduction attaches geometric data to PL weights and makes numerical positivity available. Formal metrics and strictly integral PL metrics are compared in the appropriate nontrivially valued strict analytic category. Extending a conclusion from strictly integral PL to integral or real PL data requires the relevant reduction, tensor, or descent argument.

Relative Monge–Ampère and residual intersections

The model-measure formula of Chapter 20 is a top-degree case. CLD25 develops curvature products of PL metrics as structured currents and calculates them using residual intersection data. Relative formulas allow products to be evaluated along fibres. This is the mechanism that feeds the complete-intersection and abelian applications.

The central distinction is between a numerical residual class and an actual metric. Residual numerical triviality can force a curvature current to vanish in a specified theorem, but it does not make the metric canonically trivial. Constants and algebraically nontrivial degree-zero data remain relevant to heights and pairings.

The canonical current of a curve is not a probability measure

Section 18.2 treats a different geometric metric: the semi-metric on the canonical bundle induced from Temkin’s Kähler seminorm. Over an algebraically closed nontrivially valued field, the basic example on the affine line is ∥dT∥(ηa,r)=r.\|dT\|(\eta_{a,r})=r. At a rigid point the seminorm vanishes. Hence this is not a finite continuous metric of the sort used in the elementary BT recursion. Its logarithm is nevertheless locally integrable in the source construction, and its restriction to suitable skeletal pieces is PL.

Theorem 18.2.16 identifies its curvature on a smooth analytic curve with a canonical skeletal current, evaluated using a skeleton adapted to the test. The adapted skeleton may change when the test changes. On a compact curve, the degree of the canonical bundle is 2g−22g-2, so this object is not by definition a positive probability measure. In particular, it should not be confused with the normalized admissible measure used in Arakelov theory on a metrized graph or with the canonical probability measure of a polarized dynamical system.

This example is important for the learning programme: a natural geometric norm can be singular and fail to be continuous at a dense set, while still admitting a meaningful curvature current. The required integrability and support arguments are part of the theorem. They cannot be replaced by an informal appeal to the continuous theory.

References. [CLD25] (Chapters 10–18, especially §§12.10–12.11, §14.7, Proposition 17.7.5, and Theorem 18.2.16).

20.3 Metrics, curvature, and the first substantial applications

A metric is locally a function, but globally more than curvature

Let LL be a line bundle with a continuous metric. For a nonvanishing local frame ee, set he=−log⁡∥e∥.h_e=-\log\|e\|. If e′=fee'=fe with ff invertible, then he′=he+v(f),v(f)=−log⁡∣f∣.\begin{equation} h_{e'}=h_e+v(f),\qquad v(f)=-\log|f|. \end{equation} The unit logarithm v(f)v(f) is original smooth, being a moment-map coordinate, and has d′d′′[v(f)]=0\mathrm d'\mathrm d''[v(f)]=0. Hence the local currents d′d′′[he]\mathrm d'\mathrm d''[h_e] glue to the first Chern current c1(L,∥⋅∥)c_1(L,\|\cdot\|). This gluing uses a smooth harmonic change of weight and does not require a kernel theorem for arbitrary continuous harmonic multipliers. The trivialization rule is why a local Hessian becomes a global curvature.

For a nonzero rational or meromorphic section ss, the Poincaré–Lelong formula in our normalization is d′d′′[−log⁡∥s∥]+δdiv⁡(s)=c1(L,∥⋅∥),\begin{equation} \mathrm d'\mathrm d''[-\log\|s\|]+\delta_{\operatorname{div}(s)}=c_1(L,\|\cdot\|), \end{equation} interpreted where the locally integrable current is defined. Equivalently, d′d′′[log⁡∣f∣]=δdiv⁡(f)\mathrm d'\mathrm d''[\log|f|]=\delta_{\operatorname{div}(f)} for a meromorphic function in the trivial-metric case. This is consistent with the concavity of v(f)v(f) on a torus Gauss section: the positive divisor appears with a minus sign in d′d′′[v(f)]\mathrm d'\mathrm d''[v(f)].

Changing the metric to e−c∥⋅∥e^{-c}\|\cdot\|, with c∈Rc\in\mathbb R, adds cc to every local weight and leaves curvature unchanged. It does change the metrized line bundle and can change a local height. Thus curvature is an invariant of metric data, not a complete replacement for those data.

The standard metric on the projective line

Example 20.3.1. On PK1\mathbb P^1_K with homogeneous coordinates [Z0:Z1][Z_0:Z_1], the standard model metric on O(1)\mathcal O(1) satisfies, on the chart z=Z1/Z0z=Z_1/Z_0, ∥Z0∥(x)=1max⁡(1,∣z(x)∣).\|Z_0\|(x)=\frac1{\max(1,|z(x)|)}. On the torus Gauss skeleton, with t=v(z)t=v(z), its weight is h(t)=log⁡max⁡(1,e−t)=max⁡(0,−t).h(t)=\log\max(1,e^{-t})=\max(0,-t). Its left slope is −1-1 and its right slope is 00. The one-dimensional corner formula gives one unit of curvature at t=0t=0. Analytically this is c1(O(1),∥⋅∥std)=δξG,\begin{equation} c_1\bigl(\mathcal O(1),\|\cdot\|_{\mathrm{std}}\bigr)=\delta_{\xi_G}, \end{equation} where ξG\xi_G is the Gauss point of the closed unit disc. The identity on all of P1,an\mathbb P^{1,\mathrm{an}} also follows from the model formula in Section 20.6; the skeleton computation identifies the slope and its sign.

The section Z0Z_0 vanishes at infinity. Therefore d′d′′[−log⁡∥Z0∥]+δ∞=δξG.\mathrm d'\mathrm d''[-\log\|Z_0\|]+\delta_\infty=\delta_{\xi_G}. Both sides have total mass one after the divisor term is included. Integrating the displayed Hessian alone gives zero, as Stokes requires on the proper curve. The pole of the Green function at infinity is part of the global calculation and must not be lost when using the real line chart.

What an original GK delta-metric requires

A piecewise smooth metric has local weights that can be represented by piecewise smooth functions on tropical charts. The original GK definition of a delta-metric imposes an additional condition: the piecewise Hessian on the tropical image must come from a piecewise smooth ambient (1,1)(1,1)-superform in the prescribed way. This is [GK17] (Definition 9.9). Its purpose is to ensure that the complete curvature, including the corner part, is a delta-form with the required pullback behavior.

The characterization in [GK17] (Proposition 9.11) is functorial: a first Chern delta-form must realize the Chern delta-current after every allowed pullback, with the equality taking place in GK’s delta-current space, not only after restriction to original smooth tests. The resulting curvature is d′\mathrm d'- and d′′\mathrm d''-closed. Piecewise linear metrics satisfy the original condition because their interior Hessian is zero. Formal/model metrics therefore provide the fundamental examples. It is inaccurate to replace this statement by “every piecewise smooth metric is an original GK delta-metric.” The later Mihatsch category incorporates piecewise smooth functions more systematically, as explained in the previous section.

Positivity has a geometric meaning

For a convex smooth weight on a real chart, positivity is simply positivity of its Hessian. For a piecewise smooth weight, the facewise Hessians and the corner contributions both enter. On a line, convexity means nonnegative second derivatives on the pieces and nonnegative slope jumps. This local calculation is the prototype of the metric positivity criteria.

For piecewise linear metrics in the hypotheses of [GK19] (Theorem 6.10), model semipositivity is equivalent to the stated functorial positivity conditions, to positivity of the Chern delta-form, and to the relevant curve tests. The adjective functorial is essential in that theorem. Its list of equivalences should not be shortened to arbitrary positivity of a current on one fixed space. The later work of Gubler–Rabinoff and CLD develops additional analytic and residual criteria; these have their own scope, discussed in Section 20.6.

Comparison with model intersection numbers

On a proper nn-dimensional variety with formally metrized line bundles L‾1,…,L‾n\overline L_1,\ldots,\overline L_n, GK identifies the wedge of the Chern delta-forms with the Chambert–Loir measure obtained from intersection numbers on a model: c1(L‾1)∧⋯∧c1(L‾n)=MA⁡(L‾1,…,L‾n).\begin{equation} c_1(\overline L_1)\wedge\cdots\wedge c_1(\overline L_n) =\operatorname{MA}(\overline L_1,\ldots,\overline L_n). \end{equation} This is [GK17] (Theorem 10.5). It does not require the model line bundles to be nef for the equality to make sense: in general the measure is signed. Semipositivity is what makes the coefficients nonnegative.

This comparison is a major reason to learn the calculus. A wedge product defined by tropical local data computes a measure whose atoms and masses come from algebraic intersection theory on the special fibre. Section 20.6 explains the model formula; Section 22.2 uses it in local heights.

Remark 20.3.2 (Constant rescaling). Let L‾′=(L,e−c∥⋅∥)\overline L'=(L,e^{-c}\|\cdot\|) for a real constant cc. The weight changes from φe\varphi_e to φe+c\varphi_e+c, and d′d′′[c]=0\mathrm d'\mathrm d''[c]=0, so c1(L‾′)=c1(L‾)c_1(\overline L')=c_1(\overline L). Whenever the local-height recursion and degree formula apply on a proper nn-fold, rescaling one metric by e−ce^{-c} changes the local height by c deg⁡X(L1⋯Ln),c\,\deg_X(L_1\cdots L_n), where the displayed line bundles are the other nn entries. This follows by taking ρ=c\rho=c in (22.2.4); it does not follow from the curvature alone.

Fubini–Study weights and mixed Chern measures

Sections produce a metric

Let LL be a line bundle and let s0,…,sNs_0,\ldots,s_N be sections of L⊗mL^{\otimes m} with no common zero. Choose real weights aia_i. On an open set with frame ee of LL, write si=fie⊗ms_i=f_i e^{\otimes m} and put ϕe=1mmax⁡i(log⁡∣fi∣+ai),∥e∥=e−ϕe.\begin{equation} \phi_e=\frac1m\max_i\bigl(\log|f_i|+a_i\bigr),\qquad \|e\|=e^{-\phi_e}. \end{equation} Use log⁡0=−∞\log0=-\infty. The maximum is finite because the sections have no common zero. Under e′=gee'=ge, the coefficients become fi′=fig−mf_i'=f_i g^{-m}, so ϕe′=ϕe−log⁡∣g∣\phi_{e'}=\phi_e-\log|g|. Hence the local formulas define a metric.

Near a point, any section vanishing at that point is dominated in the maximum by a nonvanishing section after shrinking. The relevant nonvanishing coefficients are analytic units. Thus (20.4.1) is locally a maximum of affine functions of unit logarithms. This explains its piecewise affine character and positive corner curvature. Arbitrary real aia_i give real weighted FS data; to regard the metric as a formal or rational model metric over the original field, one must impose the corresponding value-group and denominator conditions.

Smoothing the maximum without losing the chart conditions

On a real chart, max⁡(ℓ0,…,ℓN)\max(\ell_0,\ldots,\ell_N) can be approximated by smooth convex regularized maxima. For instance, log-sum-exp gives a convenient local estimate max⁡iℓi≤εlog⁡∑ieℓi/ε≤max⁡iℓi+εlog⁡(N+1).\max_i\ell_i\leq\varepsilon\log\sum_i e^{\ell_i/\varepsilon} \leq\max_i\ell_i+\varepsilon\log(N+1). This formula illustrates uniform approximation and convexity on the chart. Near a zero of a section, it should not be used to pretend that log⁡∣fi∣\log|f_i| is a smooth finite coordinate. A compatible regularized-maximum construction discards dominated terms and agrees with the surviving affine term where it is separated. CLD’s FS regularization provides the analytic construction under its stated hypotheses.

On P1\mathbb P^1, the standard weight max⁡(0,−t)\max(0,-t) is particularly instructive. Smooth it only in a small interval around t=0t=0, leaving it equal to its affine branches outside. Its second derivative is a nonnegative measure of total mass one, since the endpoint slopes remain −1-1 and 00. As the interval shrinks, these measures converge weakly to δ0\delta_0. In analytic terms, the curvature tends to the Gauss-point mass. The need to keep the affine branches is also what permits smoothness in the original sense near the rigid endpoints.

Three legitimate meanings of a mixed Chern expression

The notation μ=c1(L1,∥⋅∥1)∧⋯∧c1(Ln,∥⋅∥n)\mu=c_1(L_1,\|\cdot\|_1)\wedge\cdots\wedge c_1(L_n,\|\cdot\|_n) does not define itself. There are several constructions, with comparisons on their common domains.

For smooth semipositive metrics, the local positive (1,1)(1,1)-forms d′d′′ϕi,e\mathrm d'\mathrm d''\phi_{i,e} glue to c1(L‾i)c_1(\overline L_i). Their ordinary wedge is a top-degree form and hence a measure.

For continuous metrics whose local weights lie in C25\mathcal C_{25}, use the recursion of Proposition 13.1.4 with background [1]X[1]_X. Smooth unit-logarithm frame changes have zero Hessian, so the local products glue. In top degree the resulting positive current is a Radon measure. Metrics given as controlled differences of admissible positive metrics can be treated by the corresponding multilinear signed theory once its independence statements have been established.

For formal/model metrics, the special-fibre formula gives an atomic measure, while GK constructs it as a product of Chern delta-forms. The equality of these constructions is a comparison theorem. The model definition does not rely on pretending the original curvature current was smooth.

What uniform approximation does to the measure

Suppose globally defined smooth semipositive metrics ∥⋅∥i,j\|\cdot\|_{i,j} on the fixed LiL_i converge uniformly in their weights to ∥⋅∥i\|\cdot\|_i. On every trivializing neighbourhood their weights satisfy the hypotheses of the continuous multiplier and BT theorems. Consequently ⋀i=1nc1(Li,∥⋅∥i,j)⇀⋀i=1nc1(Li,∥⋅∥i).\bigwedge_{i=1}^n c_1(L_i,\|\cdot\|_{i,j}) \rightharpoonup\bigwedge_{i=1}^n c_1(L_i,\|\cdot\|_i). Independence of approximating sequences follows from the intrinsic recursive construction; it is not inferred just from the existence of some weakly convergent subsequence.

For uniform limits of semipositive FS or model metrics, the model-theoretic measure is likewise defined using its continuity theorem. To compare it with the CLD product, compare first on the admissible common class and then pass to the uniform limit using continuity on both sides. Thus the proof has two inputs: equality at the approximating level and convergence of each construction. A raw current inequality alone does not supply either input.

Why the total mass is the intersection degree

Let Y/KY/K be smooth projective of dimension nn. In the smooth intersection comparison, one has ∫Yanc1(L‾1)∧⋯∧c1(L‾n)=deg⁡K(L1⋯Ln).\int_{Y^{\mathrm{an}}}c_1(\overline L_1)\wedge\cdots\wedge c_1(\overline L_n) =\deg_K(L_1\cdots L_n). For a globally smoothly approximable semipositive tuple as above, every approximating measure has this mass. Properness makes YanY^{\mathrm{an}} compact, so the constant function 11 is an admissible test for the top-degree current. Weak convergence therefore preserves the mass.

This explains the statement studied in the earlier discussion: global approximation plus the smooth degree formula gives the total mass formula. The word global cannot be dropped in this short argument. On a noncompact open space, vague convergence can lose mass at infinity; for example δj→0\delta_j\to0 on compact tests in R\mathbb R, although each measure has mass one.

References. [CLD12] (§§6.3–6.9); [CLD25] (§§8.7, 17.8, 18.8); [GK17] (Theorem 10.5).

20.5 Total variation and the curve estimate

The norm of a signed measure

For a finite signed Radon measure ν\nu, write its Jordan decomposition as ν=ν+−ν−\nu=\nu^+-\nu^-, with mutually singular positive measures. Its total variation measure is ∣ν∣=ν++ν−|\nu|=\nu^++\nu^-, and ∥ν∥TV=∣ν∣(X)=ν+(X)+ν−(X).\|\nu\|_{\mathrm{TV}}=|\nu|(X)=\nu^+(X)+\nu^-(X). Equivalently, it is the supremum of ∑i∣ν(Ei)∣\sum_i|\nu(E_i)| over finite measurable partitions, or the supremum of ∣∫f dν∣|\int f\,d\nu| over bounded measurable ff with ∣f∣≤1|f|\leq1. For ν=∑aiδxi\nu=\sum a_i\delta_{x_i} at distinct points, the norm is ∑∣ai∣\sum|a_i|. Its signed mass, by contrast, is ∑ai\sum a_i.

If μ0,μ1\mu_0,\mu_1 are positive finite measures, then ∣μ1−μ0∣≤μ1+μ0,∥μ1−μ0∥TV≤μ1(X)+μ0(X).|\mu_1-\mu_0|\leq\mu_1+\mu_0,\qquad \|\mu_1-\mu_0\|_{\mathrm{TV}}\leq\mu_1(X)+\mu_0(X). Cancellation can make the inequality strict. For example μ0=μ1≠0\mu_0=\mu_1\ne0 gives zero on the left.

The bound 2deg⁡L2\deg L and its hypotheses

On a smooth projective analytic curve, let two admissible semipositive metrics on the same line bundle LL have curvature measures μ0,μ1\mu_0,\mu_1. Their local weight difference g=ϕ1−ϕ0g=\phi_1-\phi_0 is a global continuous function, and d′d′′[g]=μ1−μ0.\mathrm d'\mathrm d''[g]=\mu_1-\mu_0. Each positive measure has mass deg⁡KL\deg_KL. Therefore ∥d′d′′[g]∥TV≤2deg⁡KL.\begin{equation} \|\mathrm d'\mathrm d''[g]\|_{\mathrm{TV}}\leq2\deg_KL. \end{equation} The signed mass of d′d′′[g]\mathrm d'\mathrm d''[g] is zero, but its total variation need not be zero. Inequality (20.5.1) measures how much positive and negative curvature can move while preserving the degree.

The bound does not hold for an arbitrary continuous function gg without the hypothesis that its Hessian is the difference of these two positive measures of controlled mass. It is a measure estimate after the curvature construction and degree comparison have been established, not a separate general regularization theorem.

What the estimate buys in an integral

For a continuous test coefficient aa, ∣∫a d′d′′[g]∣≤∥a∥∞∥d′d′′[g]∥TV≤2deg⁡KL ∥a∥∞.\left|\int a\,\mathrm d'\mathrm d''[g]\right| \leq\|a\|_\infty\|\mathrm d'\mathrm d''[g]\|_{\mathrm{TV}} \leq2\deg_KL\,\|a\|_\infty. Thus a uniform error in a potential or coefficient produces a controlled error against this curvature difference. For smooth functions on a proper curve, integration by parts gives ∫a d′d′′g=∫g d′d′′a.\int a\,\mathrm d'\mathrm d''g=\int g\,\mathrm d'\mathrm d''a. Extensions to nonsmooth functions require the admissible approximation or multiplier theorem; the formal symmetry of the display alone does not justify them. This is the analytic use of total variation in curve energy and approximation arguments.

20.6 Model measures with the original absolute value

The special-fibre formula

Let KK now be discretely valued, R=K∘R=K^\circ, and λ=−log⁡∣π∣>0\lambda=-\log|\pi|>0 for a uniformizer. Let X\mathcal X be a normal projective flat model of a projective nn-fold XX, with special fibre ∑YmYY\sum_Ym_YY. Each component determines a divisorial point ξY\xi_Y, whose valuation extends the original valuation of KK, so vξY(π)=λv_{\xi_Y}(\pi)=\lambda.

If Li\mathcal L_i extends Li⊗eiL_i^{\otimes e_i}, the model measure is μ=1e1⋯en∑YmYdeg⁡K~(c1(L1)⋯c1(Ln)∩[Y])δξY.\begin{equation} \mu=\frac1{e_1\cdots e_n} \sum_Ym_Y\deg_{\widetilde K} \bigl(c_1(\mathcal L_1)\cdots c_1(\mathcal L_n)\cap[Y]\bigr)\delta_{\xi_Y}. \end{equation} The residual degree includes residue-field degrees. There is no extra factor λ\lambda multiplying this Chern measure in the normalization used here: its mass is the algebraic intersection degree. The scale λ\lambda does appear in values of logarithmic model functions and in lengths of skeleton edges. Keeping both is what makes the slope formula agree with (20.6.1).

Relative nefness makes the coefficients nonnegative. Without nefness, the same formula defines a signed measure. The equality with the analytic CLD/GK construction on the common class is the content of the measure comparison, and compatibility with base change is needed when using the algebraically closed version of the GK theorem over a smaller ground field.

A blow-up with the scale still visible

Blow up PR1\mathbb P^1_R at a rational smooth point of the special fibre. The resulting special fibre is F+EF+E, where EE is exceptional, and F⋅E=1,F2=E2=−1.F\cdot E=1,\qquad F^2=E^2=-1. The generic fibre is still PK1\mathbb P^1_K. The model line O(E)\mathcal O(E), with its canonical generic trivialization, defines a function g=−log⁡∥1∥g=-\log\|1\| with g(ξF)=0,g(ξE)=λ.g(\xi_F)=0,\qquad g(\xi_E)=\lambda. The intervening skeletal edge has length λ\lambda. Its slope from FF to EE is therefore +1+1, not +λ+\lambda. The outgoing slope at the other end is −1-1, and d′d′′[g]=δξF−δξE.\mathrm d'\mathrm d''[g]=\delta_{\xi_F}-\delta_{\xi_E}. This agrees with E⋅F=1E\cdot F=1 and E2=−1E^2=-1. Normalizing λ=1\lambda=1 hides the distinction between value and slope, but is unnecessary.

More generally, D=aF+bED=aF+bE gives values aλ,bλa\lambda,b\lambda and curvature (b−a)δξF+(a−b)δξE.(b-a)\delta_{\xi_F}+(a-b)\delta_{\xi_E}. Its mass is zero. If the curvature is positive, then a=ba=b, and D=adiv⁡(π)D=a\operatorname{div}(\pi) gives a constant model function. Thus even this simple example separates a metric constant from its vanishing curvature.

References. [CL06] (§2); [GK17] (Theorem 10.5 and Remark 12.7).

20.7 Hessians, corner measures, and mixed mass

The smooth calculation fixes the factorial

On Rn\mathbb R^n, write d′d′′u=∑i,juij d′xi∧d′′xj,Ω=⋀i=1n(d′xi∧d′′xi).\mathrm d'\mathrm d''u=\sum_{i,j}u_{ij}\,d'x_i\wedge d''x_j, \qquad \Omega=\bigwedge_{i=1}^n(d'x_i\wedge d''x_i). Expanding the nnth power and using the exterior signs gives (d′d′′u)n=n!det⁡(D2u) Ω.(\mathrm d'\mathrm d''u)^n=n!\det(D^2u)\,\Omega. There are n!n! ways to choose the row indices, and the signed sum over column indices is the determinant. In this paired-order normalization, Ω\Omega is positive. If all d′d' factors are moved to the left, the factor (−1)n(n−1)/2(-1)^{n(n-1)/2} must be inserted.

For u(x)=12xTAxu(x)=\tfrac12 x^{\mathsf T}Ax with AA positive semidefinite, the measure has constant density n!det⁡An!\det A. If AA has rank less than nn, the top-degree measure vanishes although d′d′′u\mathrm d'\mathrm d''u can be nonzero. This returns to the warning in Chapter 7: lower-degree curvature is not determined by its self-top-degree mass.

The PL analogue uses subgradients

For a finite convex function uu, its subgradient at xx consists of covectors aa such that u(y)≥u(x)+⟨a,y−x⟩for all yu(y)\geq u(x)+\langle a,y-x\rangle\quad\text{for all }y in the relevant convex domain. For a maximum of affine functions, the subgradient at a point is the convex hull of the slopes of the active affine branches. The real Alexandrov Monge–Ampère measure measures the volume swept out by these subgradients. In the superform normalization above, the corresponding top-degree measure has the additional factor n!n!.

For u(x,y)=max⁡(0,x,y)u(x,y)=\max(0,x,y), the subgradient at the origin is the triangle with vertices (0,0),(1,0),(0,1)(0,0),(1,0),(0,1), of area 1/21/2. Away from the origin, the subgradient is a segment or a point and has area zero. Consequently (d′d′′u)2=δ0.(\mathrm d'\mathrm d''u)^2=\delta_0. The corner set is one-dimensional, but the top-degree product is concentrated where two independent slope changes meet. This is the convex-geometric prototype of the complete-intersection mass in the logarithmic maximum formula.

Mixed products are polarized intersections

Let u=max⁡(0,x)u=\max(0,x) and v=max⁡(0,y)v=\max(0,y). The subgradient of su+tvsu+tv at the origin is the rectangle [0,s]×[0,t][0,s]\times[0,t], so its normalized top-degree mass is 2st2st. On the other hand, (d′d′′(su+tv))2=s2(d′d′′u)2+2st d′d′′u∧d′d′′v+t2(d′d′′v)2.(\mathrm d'\mathrm d''(su+tv))^2=s^2(\mathrm d'\mathrm d''u)^2+2st\,\mathrm d'\mathrm d''u\wedge\mathrm d'\mathrm d''v+t^2(\mathrm d'\mathrm d''v)^2. Each self-square vanishes because only one real direction is involved. Comparing coefficients gives d′d′′u∧d′d′′v=δ0\mathrm d'\mathrm d''u\wedge\mathrm d'\mathrm d''v=\delta_0. Replacing the two slopes by primitive vectors produces the lattice determinant from tropical stable intersection.

This calculation explains why the tropical intersection and continuous convex-potential constructions agree on their common elementary examples. A theorem identifying the constructions globally must still handle charts, multiplicities, and limits. Agreement of one formula is a normalization check, not the full comparison theorem.

Why smoothing is local and why truncation is different

On a convex real domain, convolution with a nonnegative normalized kernel preserves convexity, but it is defined only on a smaller domain unless the function has been extended. This is the origin of the fixed-neighbourhood requirement in local approximation. A sequence whose domains shrink to the point does not give a uniform approximation theorem on an open neighbourhood.

Truncating a singular psh function by max⁡(u,−M)\max(u,-M) is a different operation. It produces bounded functions, but their curvature measures can carry mass on the truncation level. As M→∞M\to\infty, deciding what mass remains and what mass is excluded from a non-pluripolar product is part of the definition and theorem. It cannot be settled by treating the truncations as uniformly convergent on a neighbourhood of the singular set.

The analytic logarithmic maximum

For u=log⁡max⁡i∣fi∣u=\log\max_i|f_i|, the real maximum calculation explains the local convex structure on the complement of the common zero set. The analytic theorem must also establish local integrability, extension across the zero set, and the multiplicity of the resulting cycle. A regular sequence supplies the expected codimension, and the calibrated fibre comparison supplies the intersection multiplicity. The real determinant and subgradient calculations explain the shape of the answer; the analytic geometry proves that it is the correct answer for the given equations.

References. [CLD12] (§5.7); [CLD25] (§§1.7, 8.9 and 18.7); [BGJK25] (Section 4).

20.8 A global smooth approximation of a fixed Fubini–Study metric

The local regularized maximum of Chapter 12 gives a global construction when the local expressions come from one finite system of sections. The transition rule for a line-bundle frame is exactly the translation rule required by the regularization. This is a useful case in which local smooth approximation can be upgraded to global approximation without an additional gluing theorem.

Let XX be a good analytic space in the original smooth-form category, let LL be an analytic line bundle, and let s0,…,sNs_0,\ldots,s_N be global sections of L⊗mL^{\otimes m} without a common zero, with m≥1m\geq 1. Fix real constants aia_i. In a frame ee, write si=fie⊗ms_i=f_ie^{\otimes m} and put ϕe=1mmax⁡0≤i≤N(log⁡∣fi∣+ai).\phi_e=\frac1m\max_{0\leq i\leq N}\bigl(\log|f_i|+a_i\bigr). Use the normalized regularized maximum Mε,kM_{\varepsilon,k} constructed in Chapter 12, with the same even one-variable kernel in every coordinate. We will use its four properties: convexity and coordinatewise monotonicity; translation covariance by the diagonal vector; the bound max⁡≤M≤max⁡+ε\max\leq M\leq\max+\varepsilon; and exact deletion of a coordinate lying more than 2ε2\varepsilon below another coordinate.

Theorem 20.8.1. For every ε>0\varepsilon>0, the Fubini–Study metric above admits a global original smooth semipositive metric with local weights ϕe,ε\phi_{e,\varepsilon} satisfying 0≤ϕe,ε−ϕe≤εmon every trivializing open set.0\leq\phi_{e,\varepsilon}-\phi_e\leq\frac{\varepsilon}{m} \qquad\text{on every trivializing open set.} The difference is independent of the frame. The metrics therefore converge uniformly on the whole space in the metric-ratio distance, even when XX is not compact.

Proof. Fix x∈Xx\in X and a frame near xx. Let Ix={i:fi(x)≠0}I_x=\{i:f_i(x)\ne 0\}. This set is nonempty. Shrink the neighbourhood so that every fif_i, i∈Ixi\in I_x, is a unit. Choose i0∈Ixi_0\in I_x. Since ∣fj∣→0|f_j|\to 0 at xx for j∉Ixj\notin I_x, a further shrinking gives log⁡∣fj∣+aj<log⁡∣fi0∣+ai0−2ε(j∉Ix),\log|f_j|+a_j<\log|f_{i_0}|+a_{i_0}-2\varepsilon \qquad(j\notin I_x), with the usual interpretation log⁡0=−∞\log 0=-\infty. On this neighbourhood define ϕe,ε=1mMε,∣Ix∣((log⁡∣fi∣+ai)i∈Ix).\phi_{e,\varepsilon} =\frac1m M_{\varepsilon,|I_x|} \bigl((\log|f_i|+a_i)_{i\in I_x}\bigr). This is original smooth: the finite logarithms are coordinates of a moment map by units, and MM is a smooth real function. If two such neighbourhoods overlap, enlarge the coordinate list to the union of their surviving indices near a point of overlap. Every extra coordinate is either finite and inactive, or vanishes and remains inactive after shrinking. Exact deletion shows that both formulas give the same value. Thus the local constructions agree, including at zeros of sections.

If e′=gee'=ge, then each finite input changes by −mlog⁡∣g∣-m\log|g|. Diagonal translation covariance gives ϕe′,ε=ϕe,ε−log⁡∣g∣\phi_{e',\varepsilon}=\phi_{e,\varepsilon}-\log|g|. Hence these weights define a metric on LL. Each input has zero d′d′′d'd'', and the Hessian of MM is positive semidefinite. The convex-composition formula of Chapter 12 proves d′d′′ϕe,ε≥0d'd''\phi_{e,\varepsilon}\geq 0. This proves semipositivity. Finally, the maximum bound for MM gives the stated estimate on each neighbourhood. Since the transition term is the same for the smoothed and original weights, their difference is global, and the estimate is uniform on all of XX. ◻

This proof uses a common finite section system. It does not imply that every metric with locally smoothly approximable weights has a globally approximating sequence of metrics. General local-to-global approximation must also preserve the line bundle’s transition data. Here that preservation is supplied by the diagonal translation identity.

Proposition 20.8.2. Pullback preserves finite real-weighted Fubini–Study metrics. Tensor product also preserves this class, after passing to a common tensor exponent. The corresponding smooth approximations have positive curvature, and the uniform errors add under tensor product.

Proof. Pullback takes the section system to its pullback, which still has no common zero: the original evaluation map from a trivial bundle onto L⊗mL^{\otimes m} is surjective, and pullback preserves this surjection. The local weight formula is simply composed with the analytic map.

For tensor product, first make both exponents equal by taking powers of the individual sections and multiplying their constants accordingly. If the resulting local lists are Ai=log⁡∣fi∣+aiA_i=\log|f_i|+a_i and Bj=log⁡∣gj∣+bjB_j=\log|g_j|+b_j, the product sections have list Ai+BjA_i+B_j. The identity max⁡i,j(Ai+Bj)=max⁡iAi+max⁡jBj\max_{i,j}(A_i+B_j)=\max_i A_i+\max_j B_j identifies the metric with the tensor product. Curvatures of smooth metric weights add, so semipositivity is preserved. The triangle inequality gives addition of uniform errors. The tensor product of the chosen smooth approximations need not equal the particular regularized maximum of the product section system; both are legitimate approximations of the same limiting metric. ◻

20.9 Metric variation, closedness and conservation of total mass

We next isolate the formal calculation used in degree comparisons. Work on a proper boundaryless pure nn-dimensional analytic space where smooth integration and Stokes’ theorem are available. All metrics in the next proposition are original smooth.

Proposition 20.9.1. If two smooth metrics on LL have curvature forms θ\theta and θ′\theta', then θ′−θ=d′d′′u\theta'-\theta=d'd''u for a global smooth function uu. For every smooth d′d'- and d′′d''-closed form SS of bidegree (n−1,n−1)(n-1,n-1), ∫X(θ′−θ)∧S=0.\int_X(\theta'-\theta)\wedge S=0. Consequently the total mixed Chern integral of a fixed tuple of line bundles is independent of the smooth metrics.

Proof. The differences of local weights agree on overlaps, because the two frame-change terms cancel. They define uu, and differentiating gives θ′−θ=d′d′′u\theta'-\theta=d'd''u. Closedness of SS implies d′d′′(uS)=d′d′′u∧Sd'd''(uS)=d'd''u\wedge S. Stokes’ theorem gives zero after integration. To compare two tuples, replace the metrics one at a time. The wedge of all unchanged Chern forms is closed under both differentials, so the preceding identity applies at every replacement. ◻

This argument proves metric independence, not the value of the integral. To identify that value with deg⁡K(L1⋯Ln)\deg_K(L_1\cdots L_n) still requires an intersection comparison or a Poincaré–Lelong induction with admissible integrability. These are supplied by the source theorems cited earlier. Once that normalization is known, every smooth semipositive metric tuple has the same algebraic mass.

Lemma 20.9.2. (From smooth tests to continuous tests). Let XX be compact Hausdorff, let S⊂C(X,R)\mathscr S\subset C(X,\mathbb R) be uniformly dense, and let μj,μ\mu_j,\mu be finite signed Radon measures with sup⁡j∥μj∥TV+∥μ∥TV<∞.\sup_j\|\mu_j\|_{\mathrm{TV}}+\|\mu\|_{\mathrm{TV}}<\infty. If ∫a dμj→∫a dμ\int a\,d\mu_j\to\int a\,d\mu for every a∈Sa\in\mathscr S, the convergence holds for every a∈C(X,R)a\in C(X,\mathbb R).

Proof. Choose b∈Sb\in\mathscr S with ∥a−b∥∞<η\|a-b\|_\infty<\eta. If the displayed bound is MM, then ∣∫a d(μj−μ)∣≤Mη+∣∫b d(μj−μ)∣.\left|\int a\,d(\mu_j-\mu)\right| \leq M\eta+\left|\int b\,d(\mu_j-\mu)\right|. Take the upper limit in jj and then let η\eta tend to zero. ◻

In the algebraic CLD setting of Chapter 8, smooth functions are uniformly dense on compact spaces by the cutoff and partition argument there. Positive measures of a fixed degree have uniformly bounded total variation because total variation equals mass. Thus weak convergence as top-degree currents upgrades to weak convergence of Radon measures. This is often used implicitly when a current formula is stated as a measure formula.

Corollary 20.9.3. Assume the proper algebraic CLD setting and the mixed-product construction for local uniform smooth psh limits established in Chapter 13. For a tuple of finite Fubini–Study metrics, their smoothed mixed Chern measures converge weakly to the intrinsic mixed measure of the original metrics. The limit is independent of the chosen regularizing kernels. Its mass is deg⁡K(L1⋯Ln)\deg_K(L_1\cdots L_n) whenever the smooth degree comparison is normalized as above.

Proof. Theorem 20.8.1 gives uniformly convergent smooth semipositive metrics on the fixed line bundles. The local mixed-product theorem identifies their current limit and proves independence of the approximating sequences. The local limits glue because frame changes are smooth unit logarithms with zero Hessian. Positivity and the degree formula bound all masses. Lemma 20.9.2 upgrades convergence to continuous tests. Finally, test against the constant function one to recover the mass. ◻

There is no assertion of total-variation convergence in this result. Even on R\mathbb R, positive smooth probability densities concentrating at the origin converge weakly to δ0\delta_0 while their difference from δ0\delta_0 has total variation two. Likewise, smoothing a model metric can replace an atom by a diffuse measure without making the total-variation distance small. Weak convergence is the appropriate conclusion.

References. [CLD12] (Sections 6.3–6.9); [CLD25] (Sections 17.8 and 18.8); [GK17] (Theorem 10.5). The global approximation and the measure-theoretic passage are proved above; the algebraic degree comparison remains an explicitly cited input.

20.10 Integration by parts for metric differences

A global difference of metric weights is more flexible than a global psh function: its curvature is measured relative to a fixed line bundle. This leads to energy identities which connect the local differential calculus with arithmetic metric variation. We prove the identities in an explicit approximation domain. No general regularization theorem is hidden in the choice of that domain.

Let XX be proper, pure nn-dimensional and boundaryless, in the algebraic CLD setting where the preceding integration, positive-current and continuous-product theorems hold. Fix an original smooth semipositive metric on a line bundle LL, with curvature θ\theta. Write θu=θ+d′d′′u,V=∫Xθn.\theta_u=\theta+\mathrm d'\mathrm d''u,\qquad V=\int_X\theta^n. An admitted function in this section is a continuous global function uu which is a uniform limit of global original smooth functions uju_j with θ+d′d′′uj≥0\theta+\mathrm d'\mathrm d''u_j\geq 0. Thus the metric with weights φe+u\varphi_e+u is a uniform limit of smooth semipositive metrics on the same LL. Locally its weights belong to UX\mathcal U_X. This global requirement is stronger than merely having local smooth approximants, and stronger than positivity of a curvature current.

Lemma 20.10.1. The admitted domain contains constants and is convex. All mixed products of its curvature currents are well-defined, positive and closed, and ∫Xθu1∧⋯∧θun=V.\int_X\theta_{u_1}\wedge\cdots\wedge\theta_{u_n}=V.

Proof. Constants do not change curvature. If uj,vju_j,v_j are admissible smooth approximants, (1−t)uj+tvj(1-t)u_j+tv_j has curvature (1−t)θuj+tθvj≥0(1-t)\theta_{u_j}+t\theta_{v_j}\geq 0 for 0≤t≤10\leq t\leq 1, and approximates (1−t)u+tv(1-t)u+tv uniformly. Local continuous products give existence and compatibility with uniform limits. Changes of frame are smooth pluriharmonic functions, so these local products glue. For smooth approximants, Proposition 20.9.1 gives total mass VV. Weak convergence of the positive measures and testing against one give the same formula in the limit. ◻

Proposition 20.10.2. (Integration by parts). Let u,v,a1,…,an−1u,v,a_1,\ldots,a_{n-1} be admitted, and put S=θa1∧⋯∧θan−1S=\theta_{a_1}\wedge\cdots\wedge\theta_{a_{n-1}}. Then ∫Xu (θv−θ)∧S=∫Xv (θu−θ)∧S.\begin{equation} \int_Xu\,(\theta_v-\theta)\wedge S =\int_Xv\,(\theta_u-\theta)\wedge S. \end{equation} The equality also holds for differences of admitted functions in the two scalar positions, interpreting the curvature factors multilinearly.

Proof. First assume that all functions are smooth. Closedness of SS and Stokes give ∫Xu d′d′′v∧S=−∫Xd′u∧d′′v∧S=∫Xv d′d′′u∧S.\int_Xu\,\mathrm d'\mathrm d''v\wedge S =-\int_Xd'u\wedge d''v\wedge S =\int_Xv\,\mathrm d'\mathrm d''u\wedge S. The second equality follows by the other integration by parts, or by applying JJ to the middle integrand and using symmetry of SS. There are no boundary terms because XX is proper and boundaryless.

Choose simultaneous global smooth approximants for the finitely many functions. Every expression such as θvj∧Sj\theta_{v_j}\wedge S_j is a positive mixed measure of mass VV and converges weakly to the corresponding continuous mixed measure. Uniform convergence of uju_j permits multiplication: the error from uj−uu_j-u is bounded by V∥uj−u∥∞V\|u_j-u\|_\infty, and the remaining error tends to zero by weak convergence against the continuous function uu. Treat the term with θ\theta in the same way. Passing to the limit proves the displayed equality. Subtracting equalities proves its bilinear extension. ◻

This proof does not define a wedge of distributional first derivatives of arbitrary continuous functions. The smooth middle expression is used to establish an identity whose limiting terms are already defined mixed measures. That distinction is useful whenever a formal Dirichlet expression occurs in a non-Archimedean argument.

20.11 Energy and the variation of Monge–Ampère measures

Definition 20.11.1. For an admitted uu, define the unnormalized energy Eθ(u)=1n+1∑j=0n∫Xu θuj∧θn−j.E_\theta(u)=\frac1{n+1}\sum_{j=0}^n \int_Xu\,\theta_u^j\wedge\theta^{n-j}. In particular Eθ(0)=0E_\theta(0)=0. We do not divide by VV, so the definition is meaningful even when V=0V=0.

Theorem 20.11.2. (First variation and the difference formula). For admitted u,vu,v put h=u−vh=u-v and ut=v+thu_t=v+th, 0≤t≤10\leq t\leq 1. Then ddtEθ(ut)=∫Xh θutn,Eθ(u)−Eθ(v)=1n+1∑j=0n∫Xh θuj∧θvn−j.\begin{align} \frac{d}{dt}E_\theta(u_t)&=\int_Xh\,\theta_{u_t}^{n},\\ E_\theta(u)-E_\theta(v)&=\frac1{n+1}\sum_{j=0}^n \int_Xh\,\theta_u^j\wedge\theta_v^{n-j}. \end{align}

Proof. All expressions along this segment are polynomials in tt, by multilinearity of the continuous mixed products. Differentiation is therefore algebraic. Put Θ=θut\Theta=\theta_{u_t}. Differentiating the jjth summand of the energy gives ∫Xh Θjθn−j+j∫Xut d′d′′h∧Θj−1θn−j.\int_Xh\,\Theta^j\theta^{n-j} +j\int_Xu_t\,\mathrm d'\mathrm d''h\wedge\Theta^{j-1}\theta^{n-j}. By Proposition 20.10.2, the second integral equals j∫Xh(Θ−θ)Θj−1θn−jj\int_Xh(\Theta-\theta)\Theta^{j-1}\theta^{n-j}. Thus the summand becomes (j+1)∫Xh Θjθn−j−j∫Xh Θj−1θn−j+1.(j+1)\int_Xh\,\Theta^j\theta^{n-j} -j\int_Xh\,\Theta^{j-1}\theta^{n-j+1}. Summing over jj cancels consecutive terms and leaves (n+1)∫XhΘn(n+1)\int_Xh\Theta^n. This proves the derivative formula. Integrate from zero to one and expand θut=(1−t)θv+tθu\theta_{u_t}=(1-t)\theta_v+t\theta_u. The coefficient of each mixed term is (nj)∫01tj(1−t)n−j dt=1n+1.\binom nj\int_0^1 t^j(1-t)^{n-j}\,dt=\frac1{n+1}. The integral identity follows by jj integrations by parts, or by the factorial formula j!(n−j)!/(n+1)!j!(n-j)!/(n+1)!. This proves the difference formula. ◻

Corollary 20.11.3. Energy is increasing on the admitted domain and satisfies ∣Eθ(u)−Eθ(v)∣≤V∥u−v∥∞,Eθ(u+c)=Eθ(u)+cV.|E_\theta(u)-E_\theta(v)|\leq V\|u-v\|_\infty, \qquad E_\theta(u+c)=E_\theta(u)+cV.

Proof. All n+1n+1 mixed measures in (20.11.2) are positive of mass VV. If u≥vu\geq v, each summand is nonnegative. Bounding the absolute value of hh gives the Lipschitz estimate. For a constant difference cc, every curvature equals θu\theta_u and every mass equals VV, giving the translation formula. ◻

Proposition 20.11.4. (Concavity). The energy is concave on the admitted convex domain. If all functions along the segment are smooth, its second derivative is d2dt2Eθ(ut)=−n∫Xd′h∧d′′h∧θutn−1≤0.\frac{d^2}{dt^2}E_\theta(u_t) =-n\int_Xd'h\wedge d''h\wedge\theta_{u_t}^{n-1}\leq0.

Proof. Differentiate (20.11.1) in the smooth case to get n∫hd′d′′h∧θutn−1n\int h\mathrm d'\mathrm d''h\wedge\theta_{u_t}^{n-1}. Integration by parts gives the displayed expression. The form d′h∧d′′hd'h\wedge d''h is strongly positive of bidegree (1,1)(1,1), so its product with the semipositive curvature power has nonnegative integral. For continuous admitted endpoints, choose smooth approximants for both endpoints and use their affine segments. The Lipschitz estimate gives convergence of their energies at each of the three arguments in the concavity inequality. Passing to the limit proves concavity without assigning a separate meaning to the nonsmooth first-derivative integral. ◻

Combining concavity with the first variation gives the useful pair of bounds ∫X(u−v)θun≤Eθ(u)−Eθ(v)≤∫X(u−v)θvn.\int_X(u-v)\theta_u^n \leq E_\theta(u)-E_\theta(v) \leq\int_X(u-v)\theta_v^n. Indeed, the derivative of a differentiable concave function on [0,1][0,1] is nonincreasing, so its integral lies between its two endpoint values. These inequalities concern energy, not pointwise comparison of the two measures.

Remark 20.11.5 (Relation to local heights and variational theories). The sum of mixed integrals in (20.11.2) is the differential-calculus expression that occurs when all factors of an intersection pairing are changed from one metric to another. Identifying it with an arithmetic or Deligne pairing requires that pairing’s normalization and comparison theorem. The identities proved here use neither an adelic summation nor a claim that every semipositive model metric has a global original smooth approximation.

Global variational theories often define energy first on model or Fubini–Study metrics and extend it by order or completion. Their domains can be larger and their topology can be different. The present formulas provide a common calculation on the overlap of domains; they do not identify those completions. Likewise, energy concavity alone supplies neither a comparison principle for arbitrary bounded psh functions nor a solution of a Monge–Ampère equation.

References. [CLD12] (Sections 6.3–6.9); [CLD25] (Sections 8.7 and 17.8); [GK17] (Sections 10–12). The proofs combine continuous mixed-product convergence, smooth Stokes and elementary polynomial identities. Chapter 21 discusses the larger variational frameworks separately.

20.12 The graph Laplacian as an intersection matrix

The slope formula, model measure, and negative intersection pairing become the same calculation on a semistable curve. Keeping the length scale visible also resolves several normalization ambiguities in local heights. We first work on a finite connected compact metric graph Γ\Gamma. Subdivide loops if necessary so that every edge has two distinct endpoints. Denote its edge lengths by ℓe>0\ell_e>0.

For a continuous function which is C2C^2 on edges and has one-sided derivatives at vertices, set Δf=−f′′ dt−∑v(∑e∋v∂e,vf)δv.\Delta f=-f''\,dt-\sum_v\left(\sum_{e\ni v}\partial_{e,v}f\right)\delta_v. This is the nonnegative-operator convention: it is the negative of the curvature convention d′d′′f\mathrm d'\mathrm d''f used for graph potentials in this book. The graph has no designated boundary here; leaf vertices contribute to the displayed sum. For a graph serving as an analytic domain with boundary, one must instead specify which tests vanish at the boundary.

Proposition 20.12.1. For continuous edgewise C2C^2 functions f,gf,g, with integrable second derivatives and endpoint derivative limits, ∫Γf dΔg=∑e∫ef′g′ dt=∫Γg dΔf.\int_\Gamma f\,d\Delta g =\sum_e\int_e f'g'\,dt =\int_\Gamma g\,d\Delta f. In particular Δf\Delta f has total mass zero. If Δf=0\Delta f=0, then ff is constant.

Proof. On an oriented edge [0,ℓ][0,\ell], integration by parts gives −∫fg′′=∫f′g′−[fg′]0ℓ-\int fg''=\int f'g'-[fg']_0^\ell. At the initial endpoint the outgoing derivative is g′(0)g'(0), and at the final endpoint it is −g′(ℓ)-g'(\ell). Thus the two atomic contributions in Δg\Delta g cancel precisely the bracket term. Sum over all edges. Setting f=1f=1 proves zero total mass; setting f=gf=g when Δf=0\Delta f=0 proves f′=0f'=0 on each edge. Continuity and connectedness make the edgewise constants equal. ◻

For a function affine on every edge, its vertex values h=(hv)h=(h_v) determine it. The coefficient vector of Δh\Delta h is LhLh, where (Lh)v=∑e=vwhv−hwℓe,hTLh=∑e=vw(hv−hw)2ℓe.(Lh)_v=\sum_{e=vw}\frac{h_v-h_w}{\ell_e},\qquad h^{\mathsf T}Lh=\sum_{e=vw}\frac{(h_v-h_w)^2}{\ell_e}. Parallel edges are counted separately. In terms of an oriented incidence matrix BB and C=diag⁡(ℓe−1)C=\operatorname{diag}(\ell_e^{-1}), one has L=BCBTL=BCB^{\mathsf T}. Reversing an edge changes one column of BB by a sign and leaves LL unchanged. The kernel is the space of constant vectors, and the image is the hyperplane of vectors with coordinate sum zero.

Proposition 20.12.2. (The vertical divisor calculation). Let X\mathcal X be a regular proper split strictly semistable curve over a complete DVR RR, with reduced special fibre ∑vCv\sum_v C_v. Assume distinct branches at each node belong to distinct components, after subdivision if needed. Put λ=−log⁡∣π∣\lambda=-\log|\pi| for a uniformizer. Every node gives an edge of length λ\lambda. For vertical divisors D=∑vavCvD=\sum_v a_vC_v and E=∑vbvCvE=\sum_v b_vC_v, their model functions have values hD(v)=λavh_D(v)=\lambda a_v and hE(v)=λbvh_E(v)=\lambda b_v, and d′d′′hD=∑vdeg⁡(O(D)∣Cv) δv,∫ΓhD d′d′′hE=λ(D⋅E)=−∑e=vw(hD(v)−hD(w))(hE(v)−hE(w))ℓe.\begin{align*} \mathrm d'\mathrm d''h_D&=\sum_v\deg(\mathcal O(D)|_{C_v})\,\delta_v,\\ \int_\Gamma h_D\,\mathrm d'\mathrm d''h_E &=\lambda(D\cdot E) =-\sum_{e=vw}\frac{(h_D(v)-h_D(w))(h_E(v)-h_E(w))}{\ell_e}. \end{align*}

Proof. The off-diagonal intersection number Cv⋅CwC_v\cdot C_w is the number of their nodes, since the model is regular and split semistable. The whole special fibre is the principal divisor of π\pi, so its intersection with each component is zero. Hence Cv2C_v^2 is minus the number of edges incident to vv. It follows that deg⁡(O(D)∣Cv)=∑e=vw(aw−av).\deg(\mathcal O(D)|_{C_v})=\sum_{e=vw}(a_w-a_v). The model-weight formula of Lemma 19.9.1 gives hD(v)=λavh_D(v)=\lambda a_v. The slope along e=vwe=vw away from vv is (hD(w)−hD(v))/λ=aw−av(h_D(w)-h_D(v))/\lambda=a_w-a_v. Summing gives the first formula. Multiplying its version for EE by hDh_D and summing over vertices gives λ∑vav(E⋅Cv)=λ(D⋅E)\lambda\sum_v a_v(E\cdot C_v)=\lambda(D\cdot E). Grouping the two endpoint contributions of each edge gives the last expression. ◻

Corollary 20.12.3. One has D2≤0D^2\leq 0, with equality exactly when all coefficients ava_v are equal. Thus the vertical intersection pairing is negative definite modulo the whole special fibre.

This is the graph form of the vertical Hodge index statement. It is proved here under the split regular reduced-fibre hypotheses. With component multiplicities or nonsplit residue fields, the intersection matrix and graph weights must be adjusted; the unweighted formula should not be transplanted unchanged. In a model with a node of thickness mm, the corresponding skeleton edge has length mλm\lambda; resolving that node replaces it by a chain of length-λ\lambda edges. Linear interpolation on the chain gives the same energy.

20.13 Solving the graph Green equation

Theorem 20.13.1. For a finite signed atomic measure ν=∑vdvδv\nu=\sum_v d_v\delta_v with ∑vdv=0\sum_vd_v=0, there exists a continuous PL function gνg_\nu such that Δgν=ν,equivalentlyd′d′′gν=−ν.\Delta g_\nu=\nu,\qquad\text{equivalently}\qquad \mathrm d'\mathrm d''g_\nu=-\nu. It is unique modulo constants. Given a probability measure μ\mu, the normalization ∫gν dμ=0\int g_\nu\,d\mu=0 determines it uniquely.

Proof. Choose a vertex set containing the support of ν\nu. The symmetric matrix LL has kernel the constants, so its image is their orthogonal complement. The sum-zero condition gives d∈im⁡Ld\in\operatorname{im}L. Solve Lh=dLh=d and interpolate affinely on every edge. Its Laplacian is exactly ν\nu. Any two solutions differ by a harmonic function, hence by a constant by Proposition 20.12.1. Subtracting the μ\mu-integral imposes the normalization, since μ\mu has mass one. ◻

The Green pairing of two degree-zero atomic measures is ⟨ν,η⟩Γ=∫Γgν dη.\langle\nu,\eta\rangle_\Gamma=\int_\Gamma g_\nu\,d\eta. It is independent of the additive constant and of the normalizing probability measure, since η\eta has mass zero. Integration by parts gives symmetry and positivity: ⟨ν,η⟩Γ=∑e∫egν′gη′ dt,⟨ν,ν⟩Γ≥0,\langle\nu,\eta\rangle_\Gamma =\sum_e\int_e g_\nu'g_\eta'\,dt, \qquad\langle\nu,\nu\rangle_\Gamma\geq0, with equality in the second expression only for ν=0\nu=0. This positive Green pairing corresponds to the negative vertical intersection form with its sign reversed.

Example 20.13.2. (Two terminals and effective resistance). For vertices p≠qp\ne q, put ν=δp−δq\nu=\delta_p-\delta_q and choose gν(q)=0g_\nu(q)=0. The number r(p,q)=gν(p)−gν(q)=⟨ν,ν⟩Γ>0r(p,q)=g_\nu(p)-g_\nu(q)=\langle\nu,\nu\rangle_\Gamma>0 is the effective resistance when an edge of length ℓ\ell is a resistor of resistance ℓ\ell. On a segment it is the distance. For two parallel paths of lengths a,ba,b, a potential difference VV produces total current V/a+V/bV/a+V/b. Setting that current to one gives r=ab/(a+b)r=ab/(a+b). These statements follow directly from Lh=dLh=d; no probabilistic interpretation is required.

Proposition 20.13.3. (Subdivision does not change the solution). Inserting a vertex with no charge into an edge leaves the PL solution, Green pairing and effective resistances unchanged.

Proof. If the two new lengths are a,ba,b, the zero-charge equation at the inserted vertex is (h−h0)/a+(h−h1)/b=0(h-h_0)/a+(h-h_1)/b=0. Hence h=(bh0+ah1)/(a+b)h=(bh_0+ah_1)/(a+b), exactly linear interpolation. The endpoint currents are (h0−h1)/(a+b)(h_0-h_1)/(a+b) and its negative, as before. The edge energy is (h−h0)2a+(h1−h)2b=(h1−h0)2a+b.\frac{(h-h_0)^2}{a}+\frac{(h_1-h)^2}{b} =\frac{(h_1-h_0)^2}{a+b}. Uniqueness modulo constants now proves the assertion for the full graph. ◻

20.14 Canonical graph measures and residual genus

Let ReR_e be the effective resistance between the endpoints of ee after deleting its interior. If deletion disconnects those endpoints, set Re=+∞R_e=+\infty and (ℓe+Re)−1=0(\ell_e+R_e)^{-1}=0. The resistance in the full graph is the parallel combination re=ℓeReℓe+Re,r_e=\frac{\ell_eR_e}{\ell_e+R_e}, interpreted as re=ℓer_e=\ell_e for a bridge. Write b=∣E∣−∣V∣+1b=|E|-|V|+1.

Lemma 20.14.1. (Foster’s identity). ∑ereℓe=∣V∣−1,∑eℓeℓe+Re=b.\sum_e\frac{r_e}{\ell_e}=|V|-1, \qquad \sum_e\frac{\ell_e}{\ell_e+R_e}=b.

Proof. Let L+L^+ be the inverse of LL on the sum-zero hyperplane, extended by zero on constants. If beb_e is the incidence column for ee, the defining Green equation gives re=beTL+ber_e=b_e^{\mathsf T}L^+b_e. Therefore ∑eℓe−1beTL+be=tr⁡(L+BCBT)=tr⁡(L+L)=∣V∣−1.\sum_e\ell_e^{-1}b_e^{\mathsf T}L^+b_e =\operatorname{tr}(L^+BCB^{\mathsf T})=\operatorname{tr}(L^+L)=|V|-1. The last operator is the identity on a space of dimension ∣V∣−1|V|-1 and zero on constants. Since re/ℓe+ℓe/(ℓe+Re)=1r_e/\ell_e+\ell_e/(\ell_e+R_e)=1, subtracting the first equality from ∣E∣|E| gives the second. ◻

The canonical measure of a metric graph is the signed measure μcan=∑v(1−val⁡(v)2)δv+∑edtℓe+Re.\mu_{\rm can}=\sum_v\left(1-\frac{\operatorname{val}(v)}2\right)\delta_v +\sum_e\frac{dt}{\ell_e+R_e}. This explicit formula and its resistance-kernel characterization are classical; see Baker–Rumely, Section 14. Its mass is one: the vertex part has mass ∣V∣−∣E∣=1−b|V|-|E|=1-b, and Foster’s identity makes the edge part have mass bb. It is generally not positive. A trivalent vertex has coefficient −1/2-1/2.

A semistable curve carries more data than the unweighted graph: attach the residual genus q(v)≥0q(v)\geq 0 to each vertex. Put g=b+∑vq(v)g=b+\sum_vq(v) and assume g≥1g\geq 1. The polarized canonical divisor and admissible probability measure are KΓ=∑v(2q(v)+val⁡(v)−2)[v],μad=δKΓ+2μcan2g=1g(∑vq(v)δv+∑edtℓe+Re).\begin{align*} K_\Gamma&=\sum_v(2q(v)+\operatorname{val}(v)-2)[v],\\ \mu_{\rm ad}&=\frac{\delta_{K_\Gamma}+2\mu_{\rm can}}{2g} =\frac1g\left(\sum_vq(v)\delta_v+ \sum_e\frac{dt}{\ell_e+R_e}\right). \end{align*} The identification of this measure with Zhang’s admissible measure is the graph-theoretic input from admissible pairing theory. Its positivity and normalization are immediate from the formula, and can be checked without that identification: all displayed coefficients are nonnegative and the total mass is (∑q(v)+b)/g=1(\sum q(v)+b)/g=1. Also deg⁡KΓ=2g−2\deg K_\Gamma=2g-2.

Example 20.14.2. For good reduction, the graph is a point with q(v)=gq(v)=g, so μad=δv\mu_{\rm ad}=\delta_v. For a totally degenerate genus-one curve, the graph is a circle of length ℓ\ell and all q(v)q(v) vanish. The formula gives μad=dt/ℓ\mu_{\rm ad}=dt/\ell. For a tree with positive residual genera, every edge is a bridge, and the measure is g−1∑vq(v)δvg^{-1}\sum_vq(v)\delta_v. Thus neither total graph length nor the number of components alone determines the measure.

Three measures must now be distinguished: the generally signed μcan\mu_{\rm can} of the bare graph, the positive probability measure μad\mu_{\rm ad} of the polarized graph, and the Chern measure of a specified metrized line bundle. An admissibly metrized dualizing line has Chern measure (2g−2)μad(2g-2)\mu_{\rm ad}, not a probability measure unless that degree is one. A general semipositive model line of degree dd instead gives its own positive atomic measure of mass dd. Equality with an admissible measure is a theorem about the chosen metric, not a consequence of the word canonical.

References. [Zh93] (Appendix, especially Lemma a.6 and Theorem a.7); [BRg] (Sections 1, 4 and 14). The finite-dimensional Green, energy, subdivision and Foster calculations above are included to make all signs and scale factors explicit. The full analytic admissible metric construction is cited, not reconstructed from a skeleton alone.

20.15 Green kernels for finite measures on a compact graph

The finite matrix calculation extends to measures which need not be atomic. This extension is useful for admissible metrics: the target curvature may have a density along the edges, whereas a model metric has atomic curvature. We give the functional-analytic construction on a finite connected compact metric graph Γ\Gamma with positive edge lengths. It does not involve probability theory.

Fix z∈Γz\in\Gamma, inserting it as a vertex if necessary. Let HzH_z be the real vector space of continuous functions uu on Γ\Gamma whose restriction to each edge is in H1H^1 and for which u(z)=0u(z)=0. Put E(u,v)=∑e∫eu′v′ dt,∥u∥H2=E(u,u).\mathcal E(u,v)=\sum_e\int_e u'v'\,dt, \qquad \|u\|_H^2=\mathcal E(u,u). The derivatives here are weak derivatives on real intervals. Continuity at vertices is imposed separately; the construction does not impose a Kirchhoff condition on an arbitrary member of HzH_z.

Lemma 20.15.1. The space HzH_z is a Hilbert space. With dd the path metric and D=diam⁡ΓD=\operatorname{diam}\Gamma, ∣u(x)−u(y)∣2≤d(x,y)E(u,u),∥u∥∞≤D ∥u∥H.|u(x)-u(y)|^2\leq d(x,y)\mathcal E(u,u), \qquad \|u\|_\infty\leq\sqrt D\,\|u\|_H.

Proof. Integrate u′u' along a shortest path from yy to xx. Cauchy–Schwarz bounds the square of the integral by the path length times its derivative energy, which is at most the total energy. Taking y=zy=z gives the second estimate. A Cauchy sequence in the energy norm is therefore uniformly Cauchy, and its edge derivatives converge in L2L^2. On each edge, the fundamental theorem for H1H^1 functions identifies the uniform limit with the function obtained by integrating the limiting derivative and the limiting endpoint value. The limit has matching vertex values and vanishes at zz. It belongs to HzH_z and the convergence is in energy norm. Definiteness follows because zero energy makes the function constant on a connected graph, and its value at zz is zero. ◻

Theorem 20.15.2. For every finite signed Radon measure ν\nu on Γ\Gamma with ν(Γ)=0\nu(\Gamma)=0, there is a unique gν∈Hzg_\nu\in H_z such that E(gν,v)=∫Γv dν(v∈Hz).\mathcal E(g_\nu,v)=\int_\Gamma v\,d\nu \quad(v\in H_z). Equivalently, Δgν=ν\Delta g_\nu=\nu in the weak graph sense. The same equation holds against all continuous edgewise H1H^1 tests, without a condition at zz. Moreover, ∥gν∥H≤D ∥ν∥TV,∥gν∥∞≤D ∥ν∥TV.\|g_\nu\|_H\leq\sqrt D\,\|\nu\|_{\rm TV}, \qquad \|g_\nu\|_\infty\leq D\,\|\nu\|_{\rm TV}.

Proof. Lemma 20.15.1 bounds the linear functional v↦∫v dνv\mapsto\int v\,d\nu by D∥ν∥TV∥v∥H\sqrt D\|\nu\|_{\rm TV}\|v\|_H. The Hilbert-space representation theorem gives its unique representing vector and the first bound; the second follows from the sup-norm estimate. Subtracting v(z)v(z) from a general test leaves its derivative unchanged and does not change its integral against ν\nu, because the latter has mass zero. This proves the assertion about all tests. For a smooth edgewise function, integration by parts recovers precisely the convention Δ=−d′d′′\Delta=-\mathrm d'\mathrm d'' of Section 20.12. ◻

The necessary mass condition can also be read directly from the equation: the constant test has zero derivative, so ∫1 dν=0\int 1\,d\nu=0. On a single-point graph the condition forces ν=0\nu=0 and the normalized solution is zero. The nontrivial case is the graph with edges considered below.

Construction and symmetry of the kernel

For y∈Γy\in\Gamma, apply the theorem to δy−δz\delta_y-\delta_z and denote the solution by kyk_y. Define jz(x,y)=ky(x).j_z(x,y)=k_y(x).

Proposition 20.15.3. The function jzj_z is continuous on Γ×Γ\Gamma\times\Gamma, is symmetric, vanishes when either argument is zz, and satisfies Δxjz(x,y)=δy−δz,∣jz(x,y)∣≤d(x,z)d(y,z).\Delta_xj_z(x,y)=\delta_y-\delta_z, \qquad |j_z(x,y)|\leq\sqrt{d(x,z)d(y,z)}.

Proof. The representing-vector identity gives ky(x)=E(kx,ky)=kx(y)k_y(x)=\mathcal E(k_x,k_y)=k_x(y). The evaluation functional at xx has norm at most d(x,z)\sqrt{d(x,z)}, so ∥kx∥H≤d(x,z)\|k_x\|_H\leq\sqrt{d(x,z)} and Cauchy–Schwarz gives the stated bound. Likewise, the functional v↦v(x)−v(x′)v\mapsto v(x)-v(x') has norm at most d(x,x′)\sqrt{d(x,x')}; hence ∥kx−kx′∥H≤d(x,x′).\|k_x-k_{x'}\|_H\leq\sqrt{d(x,x')}. This proves continuity of the representing vectors in HzH_z and therefore joint continuity of their inner products. The differential equation and normalization are the defining properties of kyk_y. ◻

Corollary 20.15.4. The normalized potential and energy pairing are gν(x)=∫Γjz(x,y) dν(y),E(gν,gω)=∬Γ2jz(x,y) dν(x) dω(y)\begin{align*} g_\nu(x)&=\int_\Gamma j_z(x,y)\,d\nu(y),\\ \mathcal E(g_\nu,g_\omega) &=\iint_{\Gamma^2}j_z(x,y)\,d\nu(x)\,d\omega(y) \end{align*} for finite signed measures of mass zero. The second expression is independent of zz and positive definite when ω=ν\omega=\nu.

Proof. The continuous bounded map y↦ky∈Hzy\mapsto k_y\in H_z can be integrated against a finite signed measure. Pairing that integral with v∈Hzv\in H_z gives ∫v dν\int v\,d\nu, so uniqueness identifies it with gνg_\nu. Evaluation is continuous on HzH_z, giving the first formula. The second follows by pairing and Fubini; the kernel is bounded and both measures have finite total variation. Changing the normalization point adds only a constant to each solution, which leaves energy unchanged. Finally, zero energy implies gν=0g_\nu=0, and hence ν\nu vanishes on all continuous edgewise H1H^1 functions. These include the PL functions on arbitrary subdivisions, which are uniformly dense in C(Γ)C(\Gamma). Thus ν=0\nu=0. ◻

Normalization by a probability measure

Let μ\mu be any positive Radon measure of mass one. The word probability specifies only this normalization; no stochastic construction is used. Set a(x)=∫jz(x,t) dμ(t),c=∬jz(s,t) dμ(s)dμ(t),gμ(x,y)=jz(x,y)−a(x)−a(y)+c.\begin{align*} a(x)&=\int j_z(x,t)\,d\mu(t),\qquad c=\iint j_z(s,t)\,d\mu(s)d\mu(t),\\ g_\mu(x,y)&=j_z(x,y)-a(x)-a(y)+c. \end{align*}

Theorem 20.15.5. This is the unique continuous symmetric Green kernel with energy-class sections satisfying Δxgμ(x,y)=δy−μ,∫gμ(x,y) dμ(x)=0.\Delta_xg_\mu(x,y)=\delta_y-\mu, \qquad \int g_\mu(x,y)\,d\mu(x)=0. It is independent of the auxiliary point zz. Its sections are in general neither PL nor smooth.

Proof. Integration of the defining equation for jzj_z gives Δa=μ−δz\Delta a=\mu-\delta_z. Subtract to obtain the equation for gμg_\mu. Symmetry is inherited from jzj_z, and integration in xx cancels all four terms because ∫a dμ=c\int a\,d\mu=c. If another normalized kernel has the same equation, the difference of its sections has zero energy and is constant. The normalization makes this constant zero. This also proves independence of zz. A measure with a nonzero density on an edge requires a nonzero distributional second derivative there, so its potential need not be PL. ◻

Proposition 20.15.6. (Weak convergence of measures and uniform convergence of potentials). Suppose νj\nu_j and ν\nu have mass zero, νj⇀ν\nu_j\rightharpoonup\nu against C(Γ)C(\Gamma), and their total variations are bounded uniformly. Then gνj→gνg_{\nu_j}\to g_\nu uniformly, and their self-energies converge.

Proof. Pointwise convergence follows by testing against jz(x,⋅)j_z(x,\cdot). Uniform continuity of jzj_z on the compact product and the total-variation bound make the potentials equicontinuous. A finite net in Γ\Gamma upgrades pointwise convergence to uniform convergence: control the finitely many values on the net, then the oscillation on each neighboring ball. Finally split ∫gνj dνj−∫gν dν=∫(gνj−gν) dνj+∫gν d(νj−ν).\int g_{\nu_j}\,d\nu_j-\int g_\nu\,d\nu =\int(g_{\nu_j}-g_\nu)\,d\nu_j +\int g_\nu\,d(\nu_j-\nu). The first term tends to zero by uniform convergence and bounded total variation, and the second by weak convergence. ◻

The scope of this theorem is a fixed compact metric graph. A classical point of a Berkovich curve can lie at infinite distance along a ray outside such a graph, and its logarithmic Green function need not be continuous. To pass from the graph construction to an admissible metric on the analytic curve, one uses the curve-theoretic comparison between the Laplacian, the retraction, and the Chern measure. That comparison, and the treatment of divisor singularities, are additional analytic theorems. The Hilbert-space proof does not establish a higher-dimensional non-Archimedean Green theorem.

References. [BRg] [Zh93]. The argument above combines the elementary graph energy estimate with Hilbert-space representation and bounded-kernel integration; it supplies the measure extension of the preceding finite-matrix construction.