OctoInformBeta

OpenAI preprint, September 2026 · Appendix A. Electrical-flow localization

Appendix A (proof of Lemma 4.1): heat representation and entropy: The Gram projection has an integrable heat representation, with entropy dissipation on complete weighted graphs

Heat and entropy control electrical localization.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Electrical-flow localization concerns weighted sums of absolute matrix entries of a Euclidean range projection. Let EE and VV be finite edge and vertex sets, and put n=|V|n=|V|. All matrices and weights are real.

For a matrix A∈ℝE×VA\in\mathbb R^{E\times V}, choose an orthonormal eigenvector matrix UU of its Gram matrix G=A⊤AG=A^\top A, with eigenvalues λk\lambda_k. Put λk†=1/λk\lambda_k^\dagger=1/\lambda_k when λk≠0\lambda_k\ne0 and λk†=0\lambda_k^\dagger=0 otherwise. The Gram eigenbasis coordinate matrix CC of AA and its Gram projection QAQ_A, the orthogonal range projection, are defined by (1). C=AU,QA=Cdiag⁡(λk†)C⊤.C=AU,\qquad Q_A=C\operatorname{diag}(\lambda_k^\dagger)C^\top.(1)

Let s,r:E→Vs,r:E\to V be endpoint maps; the incidence matrix BB is the matrix in (2).✦ For endpoint maps s,r:E→Vs,r:E\to V and conductances c:E→ℝc:E\to\mathbb R, the weighted graph Laplacian LcL_c is the matrix defined by (2). Bei=𝟏{s(e)=i}−𝟏{r(e)=i},Lc=B⊤diag⁡(ce)B.B_{ei}=\mathbf 1_{\{s(e)=i\}}-\mathbf 1_{\{r(e)=i\}},\qquad L_c=B^\top\operatorname{diag}(c_e)B.(2)

Given a real matrix LL on VV and a function d:V→ℝd:V\to\mathbb R, put D=diag⁡(di)D=\operatorname{diag}(d_i). The symmetric heat generator Md(L)M_d(L) and the weighted heat kernel KtL,dK_t^{L,d} are the matrices defined by (3), for t∈ℝt\in\mathbb R. Md(L)=DLD,KtL,d=Dexp⁡(−tMd(L))D.M_d(L)=DLD,\qquad K_t^{L,d}=D\exp\bigl(-tM_d(L)\bigr)D.(3) For the graph kernel, write Kt=KtLc,dK_t=K_t^{L_c,d} and Kt(v)(i)=Kt(i,v)K_t^{(v)}(i)=K_t(i,v) for its column at vv.

Let w:E→ℝw:E\to\mathbb R be real edge weights, and put σ=∑ewe2\sigma=\sum_e w_e^2. The squared incident edge weight mwm_w and the normalized vertex mass μw\mu_w are the functions on VV defined by (4), with the normalized vertex mass set to zero when all edge weights vanish. mw(i)=∑e:s(e)=iwe2+∑e:r(e)=iwe2,μw(i)=mw(i)2σ.m_w(i)=\sum_{e:s(e)=i}w_e^2+\sum_{e:r(e)=i}w_e^2,\qquad \mu_w(i)=\frac{m_w(i)}{2\sigma}.(4)

For real edge weights c:E→ℝc:E\to\mathbb R and a real vertex function h:V→ℝh:V\to\mathbb R, the electrical energy ℰc(h)\mathcal E_c(h) is the scalar in (5). ℰc(h)=∑e∈Ece(h(s(e))−h(r(e)))(log⁡h(s(e))−log⁡h(r(e))).\mathcal E_c(h)=\sum_{e\in E}c_e\bigl(h(s(e))-h(r(e))\bigr) \bigl(\log h(s(e))-\log h(r(e))\bigr).(5)✦ All logarithms are natural.

The Claim

Theorem 1. For every real matrix A∈ℝE×VA\in\mathbb R^{E\times V}, each entry of t↦Aexp⁡(−tA⊤A)A⊤t\mapsto A\exp(-tA^\top A)A^\top is integrable on (0,∞)(0,\infty), with integral (6) for e,f∈Ee,f\in E. ∫0∞(Aexp⁡(−tA⊤A)A⊤)efdt=(QA)ef.\int_0^\infty\bigl(A\exp(-tA^\top A)A^\top\bigr)_{ef}\,dt=(Q_A)_{ef}.(6)✦

Suppose every pair of distinct vertices is joined by an edge in either direction, ce>0c_e>0, and μ\mu is a strictly positive probability mass on VV. Let di>0d_i>0 satisfy μidi2=1\mu_i d_i^2=1. For each v∈Vv\in V, the function t↦ℰc(Kt(v))t\mapsto\mathcal E_c(K_t^{(v)}) is integrable on (0,∞)(0,\infty) and satisfies (7). ∫0∞ℰc(Kt(v))dt=log⁡1μv.\int_0^\infty\mathcal E_c(K_t^{(v)})\,dt=\log\frac1{\mu_v}.(7)✦

For arbitrary endpoint maps, let ce≥0c_e\ge0, let wew_e be real with σ=∑ewe2>0\sigma=\sum_e w_e^2>0, and let hh be positive with ∑iμw(i)h(i)=1\sum_i\mu_w(i)h(i)=1. The electrical Cauchy-Schwarz inequality is (8). (∑e∈Ewece|h(s(e))−h(r(e))|)2≤σℰc(h).\left(\sum_{e\in E}w_e\sqrt{c_e}\, |h(s(e))-h(r(e))|\right)^2 \le \sigma\,\mathcal E_c(h).(8)✦

For any strictly positive probability mass μ\mu on VV, the entropy bound is (9). ∑v∈Vμvlog⁡1μv≤log⁡n.\sum_{v\in V}\mu_v\log\frac1{\mu_v}\le\log n.(9)✦

Let Q:E×E→ℝQ:E\times E\to\mathbb R, F:ℝ×E×E→ℝF:\mathbb R\times E\times E\to\mathbb R, I:ℝ×V→ℝI:\mathbb R\times V\to\mathbb R, μ:V→ℝ\mu:V\to\mathbb R, and S∈ℝS\in\mathbb R. Take we≥0w_e\ge0. Suppose every function F(⋅,e,f)F(\cdot,e,f) and I(⋅,v)I(\cdot,v) is integrable on (0,∞)(0,\infty) and the relations (10) hold for all e,f∈Ee,f\in E, v∈Vv\in V, and t>0t>0. Q(e,f)=2∫0∞F(t,e,f)dt,∫0∞I(t,v)dt=log⁡1μv,∑e,f∈Ewewf|F(t,e,f)|≤S∑v∈VμvI(t,v).\begin{aligned} Q(e,f)&=2\int_0^\infty F(t,e,f)\,dt,\\ \int_0^\infty I(t,v)\,dt&=\log\frac1{\mu_v},\\ \sum_{e,f\in E}w_e w_f|F(t,e,f)| &\le S\sum_{v\in V}\mu_v I(t,v). \end{aligned}(10) Then the localization bound is (11). ∑e,f∈Ewewf|Q(e,f)|≤2S∑v∈Vμvlog⁡1μv.\sum_{e,f\in E}w_e w_f|Q(e,f)| \le 2S\sum_{v\in V}\mu_v\log\frac1{\mu_v}.(11)✦

In the paper, the result reads: the Euclidean projection has the convergent heat-kernel integral QA=∫0∞Aexp⁡(−tA⊤A)A⊤dtQ_A=\int_0^\infty A\exp(-tA^\top A)A^\top\,dt, and, for the vertex mass μ\mu determined by positive edge weights, the heat-column entropy dissipation integrates to log⁡(1/μv)\log(1/\mu_v), while electrical Cauchy-Schwarz and ∑vμvlog⁡(1/μv)≤log⁡n\sum_v\mu_v\log(1/\mu_v)\le\log n yield the localization inequality.

What Is Checked

The heat representation ranges over arbitrary finite real matrices, and the graph assertions over the incidence constructions (2) with their stated positivity and normalization hypotheses. The localization estimate ranges over entrywise integrable kernels satisfying (10).

The Gram matrix A⊤AA^\top A of every real matrix AA is symmetric.✦