Bizeul–Klartag–Lehec: cumulants and suspension - The KLS theorem and its methods

What to retain. The cumulants of an isotropic log-concave law satisfy $\abs{\kappa_m^\mu(u,\cdot,\dots,\cdot)}^2\le K^{m-1}((m-1)!)^2\abs u^2$ with $K$ independent of the dimension: up to a geometric factor, the factorial growth already forced by the exponential law on the line. Suspension encodes an arbitrary test function as an extra coordinate of a larger log-concave law, so this bound on linear observables becomes the exponential Appell coefficient bound for every function, which is KLS.

This chapter works through the proof of Bizeul, Klartag and Lehec (BKL) in their preprint of 4 October 2026 Bizeul et al., 2026, version 1: the analytic criterion, the cumulant induction and suspension. The criterion is the spectral criterion of the first version of Song–Zhang (SZ v1, Chapter Song–Zhang, first version: polynomial estimates and curvature) in exponential form. The second version of Song–Zhang (SZ v2, Chapter Song–Zhang, second version: repeated refinement with summable losses) closes the same criterion differently, and the proof of Balasubramanian and Kasiviswanathan (BK, Chapter Balasubramanian–Kasiviswanathan: compatible integration) does without it; the three are compared in Chapter The proofs of KLS compared.

A one-dimensional calibration¶

The exponential law forces factorial growth of cumulants, $(m-1)!$ at order $m$ (Section Bizeul–Klartag–Lehec: cumulants and suspension). In higher dimension the issue is to bound the full tensor with one argument fixed, summing the squares of all remaining entries without a dimension factor.

A tilted average is the expectation of a test function after reweighting the law by $e^{\langle z,x\rangle}$ . Its first derivative at zero is $\operatorname{Cov}(X,f)$ ; its higher derivatives retain the response to all small exponential tilts. We use the following conventions.

Definition 8.1 (Tilt averages, cumulants and regular measures — def:bkl-tilt-cumulants · Correction)

For a log-concave probability measure $\mu$ on $\mathbb R^n$ , set, in a neighborhood of the origin,

\Lambda_\mu(z)=\log\int e^{\langle z,x\rangle}\,d\mu(x),\qquad \kappa_m^\mu=\nabla^m\Lambda_\mu(0),\qquad F_f^\mu(z)=\frac{\int f(x)e^{\langle z,x\rangle}\,d\mu(x)} {\int e^{\langle z,x\rangle}\,d\mu(x)},\qquad \mathcal T_d^\mu f=\frac{\nabla^dF_f^\mu(0)}{d!}

(8.1)

for $f\in L^2(\mu)$ . Tensor norms and inner products are the Hilbert--Schmidt ones with all ordered indices summed. The positive semidefinite matrix $R_m^\mu$ is determined by

\langle R_m^\mu u,u\rangle=|\kappa_m^\mu(u,\cdot,\ldots,\cdot)|^2.

(8.2)

In this chapter, a BKL-regular measure has density $e^{-V}$ on $\mathbb R^n$ , where $V\in C^\infty$ , every derivative of $V$ has at most polynomial growth, and $aI\preceq\nabla^2V\preceq bI$ for some $0<a\le b<\infty$ depending on the measure. Write $\mathcal A$ for the smooth functions all of whose derivatives have at most polynomial growth, $\mathcal A_0=\{g\in\mathcal A:\int g\,d\mu=0\}$ , $L=\Delta-\langle\nabla V,\nabla\rangle$ , and $\nabla_0g=\nabla g-\int\nabla g\,d\mu$ .

From tilt averages to a spectral gap¶

The analytic part works first with smooth measures whose curvature is bounded above and below. The lower bound supplies a spectral gap at each fixed measure; it is not assumed uniform across measures. Approximation is used only after a bound independent of that curvature has been obtained.

Lemma 8.1 (Analytic foundations for the tilt criterion — Proved (from a preprint) · agent review (gpt-6-astra, 2026-10-06) · lem:bkl-analytic-foundations · Correction)

For every BKL-regular measure, $L$ maps $\mathcal A$ into $\mathcal A$ and

\int (Lu)v\,d\mu=-\int\langle\nabla u,\nabla v\rangle\,d\mu, \qquad \int(Lu)^2\,d\mu=\int|\nabla^2u|^2\,d\mu+ \int\langle\nabla^2V\nabla u,\nabla u\rangle\,d\mu

(8.3)

for $u,v\in\mathcal A$ . Its spectral gap $\lambda=\CP(\mu)^{-1}>0$ is attained by a function $f\in\mathcal A_0$ with $\|f\|_2=1$ and $Lf=-\lambda f$ . For every $g\in\mathcal A_0$ there is a unique $u\in\mathcal A_0$ solving $-Lu=g$ , and

\|u\|_2\le\lambda^{-1}\|g\|_2,\qquad \|\nabla u\|_2^2=\langle g,u\rangle_{L^2}\le\lambda^{-1}\|g\|_2^2, \qquad \|\nabla_0u\|_2\le\lambda^{-1/2}\|g\|_2.

(8.4)

These assertions hold componentwise for finite-dimensional tensor-valued functions. Every isotropic log-concave probability measure is the weak limit of isotropic BKL-regular measures, with convergence of every polynomial moment. A common Poincaré bound for these approximants passes to the limit.

The integration identities permit repeated inversion of the diffusion operator and centered differentiation of a first eigenfunction. The next estimate recovers a tensor from partial symmetrization. Its exponential cost is absorbed in the finite dyadic estimates used by the spectral argument.

This is the SZ v1 component of the BKL argument. One coefficient bound at all orders replaces successive curvature profiles with increasing constants.

The mechanism follows a first eigenfunction through centered gradients and inverse operators. A small spectral gap makes the resulting tensors large; the means removed by centering are expressed through tilted averages. Partial symmetrization compares these tensors with symmetric Taylor tensors. A single exponential bound controls the centering terms at every order. Summing the resulting estimates over finitely many dyadic orders up to an exit index forces a lower spectral gap. The analytic and tensor estimates above are separate inputs to this comparison.

The relation to the preceding chapter is exact, including the factorial: both constructions differentiate the same normalized exponential.

Proposition 8.1 (Duality between tilt derivatives and Appell coefficients — Proved (from a preprint) · agent review (gpt-6-astra, 2026-10-06) · prop:bkl-tilt-appell-duality · Correction)

For every full-dimensional log-concave probability measure $\mu$ , integer $d\ge1$ , symmetric order- $d$ tensor $T$ and $f\in L^2(\mu)$ , the Appell conventions of Theorem 7.1 give

\langle T,\nabla^dF_f^\mu(0)\rangle=\int f P_d^\mu[T]\,d\mu.

(8.7)

Consequently, with $K_d(\mu)=\sup_{T\text{ symmetric},\,|T|=1} \operatorname{Var}_\mu(P_d^\mu[T])$ and $c_d(\mu)=\sqrt{K_d(\mu)}/d!$ ,

\|\mathcal T_d^\mu\|_{L^2(\mu)\to\mathrm{HS}}=c_d(\mu).

(8.8)

The Appell generating function is $e^{\langle z,x\rangle-\Lambda_\mu(z)}$ . Differentiating its integral against $f$ gives the pairing. Positive-degree Appell polynomials have mean zero, so their $L^2$ norms equal the square roots of their variances. Taking the two Hilbert-space operator norms gives the coefficient identity. It identifies the common end point; it does not improve the older bound.

The cumulant induction¶

The stochastic part rescales its noise by the inverse square root of the current covariance. The deterministic covariance decay is then exponential. Higher cumulants are measured in the inverse covariance metric, keeping each index at its current variance scale.

Lemma 8.3 (Inverse-covariance localization and cumulant dynamics — Proved (from a preprint) · agent review (gpt-6-astra, 2026-10-06) · lem:bkl-cumulant-dynamics · Correction)

Let $\mu$ be compactly supported, isotropic and log-concave on $\mathbb R^n$ . There exists a global stochastic localization

d\mu_t(x)=\frac{e^{\langle c_t,x\rangle-\langle Q_tx,x\rangle/2}\,d\mu(x)} {\int e^{\langle c_t,y\rangle-\langle Q_ty,y\rangle/2}\,d\mu(y)},\qquad dc_t=A_t^{-1}a_t\,dt+A_t^{-1/2}\,dB_t,\qquad dQ_t=A_t^{-1}\,dt,

(8.9)

with $c_0=0$ , $Q_0=0$ , $a_t=\int x\,d\mu_t$ , and $A_t=\operatorname{Cov}(\mu_t)$ positive definite at every finite time. Put $\kappa_m(t)=\kappa_m^{\mu_t}$ and $H_i(t)=\kappa_3(t)(\cdot,\cdot,A_t^{-1/2}e_i)$ . Then

dA_t=\sum_iH_i(t)\,dB_t^i-A_t\,dt,\qquad \mathbb EA_t=e^{-t}I,\qquad \sum_iH_i(t)A_t^{-1}H_i(t)\preceq8A_t.

(8.10)

For $m\ge3$ ,

d\kappa_m(t)=\kappa_{m+1}(t)(\cdot,\ldots,\cdot,A_t^{-1/2}dB_t) -\bigl(m\kappa_m(t)+L_m(t)\bigr)dt,

(8.11)

where, for $h_J=(h_j)_{j\in J}$ ,

L_m(t)(h_1,\ldots,h_m)= \sum_{\substack{J\subset\{1,\ldots,m\},\,1\in J\\2\le |J|\le m-2}} \left\langle\kappa_{|J|+1}(t)(h_J,\cdot), A_t^{-1}\kappa_{m-|J|+1}(t)(h_{J^c},\cdot)\right\rangle.

(8.12)

For tensors of order $q$ define $\langle S,T\rangle_t= \langle S,(A_t^{-1})^{\otimes q}T\rangle$ and $|T|_t^2=\langle T,T\rangle_t$ . For fixed $u$ , write $\mathcal E_m(t)=|\kappa_m(t)(u,\cdot,\ldots,\cdot)|_t^2$ . Then $\mathcal E_2(t)=\langle A_tu,u\rangle$ and $\int_0^\infty\mathbb E\mathcal E_2(t)\,dt=|u|^2$ . Moreover $\mathcal E_3(t)\le8\langle A_tu,u\rangle$ . For fixed $n,m$ , the processes $\mathcal E_m(t)$ and $|L_m(t)(u)|_t^2$ have deterministic bounds depending on $n,m,u$ and the initial support radius. The Itô local martingale part of each $\mathcal E_m$ is a square-integrable martingale on every bounded time interval.

Differentiating the logarithmic Laplace transform along this localization produces a next-order cumulant in the noise and products of lower-order cumulants in the drift. The third-moment estimate controls covariance noise in its own metric. Positive definiteness at finite times and integrability after stopping justify use of that metric. These are mathematical inputs, not a convention for interpreting a singular inverse.

Lemma 8.4 (Differential and integrated cumulant energy estimates — Proved (from a preprint) · agent review (gpt-6-astra, 2026-10-06) · lem:bkl-cumulant-energy · Correction)

For the process and quantities of Lemma 8.3, a universal $C>0$ satisfies, for every $m\ge3$ and fixed $u$ ,

\operatorname{drift}\mathcal E_m(t)\ge\tfrac12\mathcal E_{m+1}(t) -Cm^2\mathcal E_m(t)-2\langle\kappa_m(t)(u),L_m(t)(u)\rangle_t.

(8.13)

If $I_m=\int_0^\infty\mathbb E\mathcal E_m(t)\,dt<\infty$ and $J_m=\int_0^\infty\mathbb E|L_m(t)(u)|_t^2\,dt<\infty$ , then

\mathcal E_m(0)+\tfrac12\int_0^\infty\mathbb E\mathcal E_{m+1}(t)\,dt \le Cm^2I_m+2\sqrt{I_mJ_m}.

(8.14)

The energy computation retains the variation of the inverse covariance and its cross variation with the cumulant. Noise supplies a positive next-order energy. The moving metric costs a quadratic factor in the order, and the remaining term is a product of lower-order cumulants. To integrate, finite total energy supplies terminal times at which the expected boundary term tends to zero. The induction must supply the stated integrability hypotheses.

The induction proves a static bound and an integrated next-order bound together. In each product of cumulants, the factor containing the fixed vector uses the integrated estimate; the other uses the static estimate for the whitened law. The number of index splittings is a binomial coefficient, canceled exactly by the two factorials. A polynomial cost in the order remains and is absorbed by one large exponential base $K$ . Conditioning on expanding balls and convergence of moments remove compact support. No KLS estimate enters this induction.

Suspension: putting a function into a coordinate¶

For an isotropic BKL-regular law, take a smooth test $f$ with bounded Hessian, unit $L^2$ norm and orthogonal to affine functions. Let $X_1,\ldots,X_N$ be independent copies from this law and put $F_N=N^{-1/2}\sum_i f(X_i)$ . Adjoin $S=(F_N+\eta_\beta)/\sqrt{1+2/\beta^2}$ , with an independent centered Laplace variable $\eta_\beta$ of rate $\beta$ . The joint law is isotropic. For large $N$ it is log-concave: the Hessian cost of each copy of $f$ is reduced by $N^{-1/2}$ and absorbed by the original positive curvature. The dimension has increased, which is precisely why the cumulant estimate must hold in every dimension with the same constant.

A cumulant with one slot in the new coordinate and the others in one copy of $X$ equals the tilt derivative divided by $\sqrt{N(1+2/\beta^2)}$ . The $N$ disjoint tensor blocks cancel the factor $N^{-1}$ in their squared norms. Increasing $\beta$ removes the added variance. Decomposing a test into constant, linear and affine-orthogonal parts, then using density, extends the estimate to $L^2$ . The suspension needs only its cumulant hypothesis.

Substitution of the all-order bound gives one exponential base. Approximation passes the Appell polynomial moment inequalities to arbitrary laws; linear contraction includes singular covariances.

Combining Theorem 8.3 with Theorem 8.1 gives the KLS conclusion Conjecture 0.1, first for regular measures and then by scalar Poincaré stability. This is the BKL argument.

Consequences for the earlier questions¶

The exponential estimate also supplies the initialization bound used in the first-version iteration of Chapter Song–Zhang, first version: polynomial estimates and curvature, without a curvature-profile hypothesis. Since that bound is already of KLS strength, feeding it back into the iteration gives no second proof.

Corollary 8.1 (Uniform conditional initialization from BKL — Proved · agent review (gpt-6-astra, 2026-10-06) · cor:bkl-uniform-conditional-initialization · Correction)

There exists a universal $G\ge1$ such that for every integer $r\ge2$ , every $\Gamma\ge G$ , every dimension, every centered log-concave probability measure $\mu$ with $\operatorname{Cov}(\mu)\preceq I$ , and every integer $d\ge1$ ,

c_d(\mu)\le\frac{((1+r^{-2})\Gamma)^d\ell_r(d)^d}{(d+1)^2}, \qquad \ell_0(x)=x,\quad\ell_{r+1}(x)=\log(e+\ell_r(x)).

(8.18)

Here the Appell coefficients use ambient tensors also for measures on a proper affine subspace. One may take $G=\max\{1,4\sqrt{2K}\}$ for $K$ as in Theorem 8.3. No curvature-profile assumption is needed.

Indeed $(d+1)^2\le4^d$ and $\ell_r(d)\ge1$ , so enlarging the exponential base absorbs the denominator uniformly in degree and depth. This does not estimate the accumulated centering losses of a different spectral comparison. What the three proofs contribute to each alternative mechanism is in Section What the three proofs contribute to each mechanism.

References¶

Bizeul, P., Klartag, B., & Lehec, J. (2026). Presenting a Proof of the Kannan–Lovasz–Simonovits Conjecture. https://arxiv.org/abs/2610.05474v1