on a sharp quadratic-form inequality and the KLS constant

Overview

Let $X$ be a centered isotropic log-concave random vector in ${\mathbb{R}}^n$. Letwin proves the sharp estimate

$$\mathrm{Var}(X^{\mathsf{T}}MX)\le 8\mathrm{Tr}(M^2)$$

for every symmetric matrix $M$ (Letwin, 2026). Since $\nabla (x^{\mathsf{T}}Mx)=2Mx$, this is a Poincare inequality with constant $2$ for quadratic forms.

The proof represents $X$ by a moment map, obtains a matrix-valued estimate for the Hessian of its potential, and converts that estimate through a Stein kernel and an $H^{-1}$ inequality. Applied to the third-moment tensor, it gives a dimension-free parameter ${\kappa }_n\le 2\sqrt{2}$. Klartag’s stochastic-localization estimates then yield

$$C_{\mathrm{P}}(\mu )\lesssim \sqrt{\log (en)},{\psi }_n\lesssim {\log }^{1/4}(en).$$

The preprint leaves the last implication as “tracing Klartag’s inequalities.” The missing absorption argument is written out below.

🏷️ Log-Concave and Spectral Preliminaries

A probability measure $\mu$ on ${\mathbb{R}}^n$ is log-concave if

$$\mu (\lambda A+(1-\lambda )B)\ge \mu (A{)}^{\lambda }\mu (B{)}^{1-\lambda }$$

for compact $A,B$ and $\lambda \in [0,1]$. In the full-dimensional case it has a density $e^{-V}$ for a convex extended-valued function $V$. It is isotropic when $X\sim \mu$ satisfies

$$\mathbb{E}X=0,\mathbb{E}(X\otimes X)=I_n.$$

For a matrix $A$,

$$\Vert A{\Vert }_{\mathrm{HS}}^2=\mathrm{Tr}(A^{\mathsf{T}}A),\Vert A{\Vert }_{\mathrm{op}}=\underset{|u|=1}{\sup }|Au|.$$

If $A$ is symmetric, $\Vert A{\Vert }_{\mathrm{HS}}^2=\mathrm{Tr}(A^2)$. Define

$$|A|=(A^2{)}^{1/2},\mathrm{sgn}(A)=A|A{|}^{-1}$$

when $A$ is invertible. Then $A=\mathrm{sgn}(A)|A|$ and $\mathrm{sgn}(A{)}^2=I_n$.

The Poincare constant $C_{\mathrm{P}}(\mu )$ is the least $P$ such that

$${\mathrm{Var}}_{\mu }(f)\le P\int |\nabla f{|}^{2}d\mu .$$

Its reciprocal ${\lambda }_0=P^{-1}$ is the spectral gap. Define the outer boundary measure by

$${\mu }^{+}(S)=\underset{\epsilon \downarrow 0}{\mathrm{liminf}}\frac{\mu (S_{\epsilon })-\mu (S)}{\epsilon },S_{\epsilon }=\{x:\mathrm{dist}(x,S)\le \epsilon \}.$$

If

$$h(\mu )=\underset{0<\mu (S)\le 1/2}{\inf }\frac{{\mu }^{+}(S)}{\mu (S)},\psi (\mu )=h(\mu {)}^{-1},$$

then Cheeger and reverse-Cheeger inequalities for log-concave measures give

$$\psi (\mu {)}^2\asymp C_{\mathrm{P}}(\mu ).$$

The inverse KLS constant is

$${\psi }_n=\sup \{\psi (\mu ):\mu \text{ is isotropic and log-concave on }{\mathbb{R}}^n\}.$$

Convention for the KLS parameter

Some sources denote the lower expansion $h$ by $\psi$. This post follows Letwin and on Eldan’s stochastic localization and the KLS conjecture in writing $\psi =1/h$. Thus an upper bound on ${\psi }_n$ is an improvement.

🏷️ Moment-Map Preliminaries

For a convex $\phi$ with $e^{-\phi }$ integrable, its moment measure is

$$\mu =(\nabla \phi {)}_{\#}\nu ,d\nu (y)=e^{-\phi (y)}dy.$$

Thus $Y\sim \nu$ and $X=\nabla \phi (Y)$ give $X\sim \mu$. More precisely, every centered probability measure whose support is not contained in a proper linear subspace admits such a potential, unique up to translation of its argument (Cordero-Erausquin & Klartag, 2015).

The proof begins with

$$d\mu (x)=e^{-V(x)}{1}_K(x)dx,$$

where $K$ is bounded, open, and convex, while $V$ is smooth and convex near $\overline{K}$. The moment map is smooth, $H={\nabla }^2\phi$ is positive definite, and $\nabla \phi :{\mathbb{R}}^n\to K$ is a diffeomorphism. Change of variables gives the Monge-Ampere equation

$$\log \det H=-\phi +V(\nabla \phi ).$$

Isotropy becomes the mean-Hessian identity

$${\mathbb{E}}_{\nu }H={\mathbb{E}}_{\nu }[\nabla \phi \otimes \nabla \phi ]=I_n.$$

Write $({\phi }^{ij})=H^{-1}$. The weighted generator

$$\mathcal{L}f={\phi }^{ij}{\partial }_{ij}f-(V_i\circ \nabla \phi ){\partial }_{i}f$$

is symmetric in $L^2(\nu )$ and satisfies

$${\mathbb{E}}_{\nu }[(\mathcal{L}f)g]=-{\mathbb{E}}_{\nu }[{\phi }^{ij}({\partial }_{i}f)({\partial }_{j}g)].$$

In particular, justified cutoff arguments give ${\mathbb{E}}_{\nu }\mathcal{L}f=0$ for the functions below.

The Brascamp-Lieb inequality is

$${\mathrm{Var}}_{\nu }(f)\le {\mathbb{E}}_{\nu }⟨H^{-1}\nabla f,\nabla f⟩$$

(Brascamp & Lieb, 1976). It will be applied entrywise to a matrix-valued function.

Stein Kernels and the Negative Sobolev Norm

A matrix-valued, locally integrable function ${\tau }_{\lambda }$ is a Stein kernel for a centered law $\lambda$ if, for every $f\in C_c^{\infty }({\mathbb{R}}^n)$,

$$\mathbb{E}[W_{i}f(W)]=\mathbb{E}\sum _j({\tau }_{\lambda }(W){)}_{ij}{\partial }_{j}f(W).$$

The moment map supplies

$${\tau }_{\mu }(x)=H((\nabla \phi {)}^{-1}(x)),{\tau }_{\mu }(X)=H(Y)$$

under the moment-map coupling (Fathi, 2019).

For centered $g\in L^2(\lambda )$, define

$$\Vert g{\Vert }_{H^{-1}(\lambda )}=\sup \left\{\mathbb{E}[g(W)u(W)]:u\in L^2(\lambda )\text{ is locally Lipschitz},\mathbb{E}|\nabla u(W){|}^2\le 1\right\}.$$

The Barthe-Klartag inequality gives

$${\mathrm{Var}}_{\lambda }(f)\le \sum _i\Vert {\partial }_{i}f{\Vert }_{H^{-1}(\lambda )}^2$$

for a full-dimensional log-concave $\lambda$ whenever $f,{\partial }_{i}f\in L^2(\lambda )$ and $\mathbb{E}{\partial }_{i}f=0$ for every $i$ (Barthe & Klartag, 2020). The Stein identity and Cauchy-Schwarz also give

$$\Vert z\mapsto ⟨z,v⟩{\Vert }_{H^{-1}(\lambda )}^2\le \mathbb{E}|{\tau }_{\lambda }(W)v{|}^2.$$

🏷️ Main Quadratic-Form Theorem

Sharp quadratic-form Poincare inequality

If $X$ is centered, isotropic, and log-concave, then every symmetric $M$ satisfies

$$\mathrm{Var}(X^{\mathsf{T}}MX)\le 8\mathrm{Tr}(M^2)=2\mathbb{E}|\nabla (X^{\mathsf{T}}MX){|}^2.$$

The equality on the right follows from

$$\mathbb{E}|2MX{|}^2=4\mathrm{Tr}(M^2\mathbb{E}[X\otimes X])=4\mathrm{Tr}(M^2).$$

🏷️ Differentiated Monge-Ampere Equation

Differentiating

$$\log \det H=-\phi +V(\nabla \phi )$$

once gives

$${\phi }^{ab}{\phi }_{abi}=-{\phi }_i+V_a{\phi }_{ai}.$$

Using

$${\partial }_k{\phi }^{ij}=-{\phi }^{ia}{\phi }_{abk}{\phi }^{bj}$$

and differentiating again yields

$$\mathcal{L}({\phi }_{ij})+{\phi }_{ij}={\phi }_{ia}{V}_{ab}{\phi }_{bj}+{\phi }^{ac}{\phi }^{bd}{\phi }_{abi}{\phi }_{cdj}.$$

Equivalently,

$$\mathcal{L}H+H=H({\nabla }^{2}V)H+Q,$$

where

$$Q_{ij}={\phi }^{ac}{\phi }^{bd}{\phi }_{abi}{\phi }_{cdj}.$$

Both terms on the right are positive semidefinite. Convexity gives $H({\nabla }^{2}V)H⪰0$, while $Q$ is the Gram matrix of the matrices

$$H^{-1/2}({\partial }_{i}H)H^{-1/2}$$

in the Hilbert-Schmidt inner product.

🏷️ Matrix Energy of the Random Hessian

Random-Hessian estimate

For every symmetric $B$,

$${\mathbb{E}}_{\nu }\mathrm{Tr}(BHBH)\le 2\mathrm{Tr}(B^2).$$

Assume first that $B⪰0$ and set $q=\mathrm{Tr}(BHBH)$. The product rule for $\mathcal{L}$ and the Hessian equation give

$$\mathcal{L}q=-2q+2A+2D+2C,$$

where

$$A={\phi }^{ab}\mathrm{Tr}(B({\partial }_{a}H)B({\partial }_{b}H)),$$ $$D=\mathrm{Tr}((BHB)(H({\nabla }^{2}V)H)),$$

and

$$C=(BHB{)}_{ij}{\phi }^{ac}{\phi }^{bd}{\phi }_{abi}{\phi }_{cdj}.$$

These terms are nonnegative. Integrating gives

$${\mathbb{E}}_{\nu }q={\mathbb{E}}_{\nu }(A+D+C).$$

Pointwise Tensor Comparison

The delicate inequality is $C\ge A$. At a fixed point, the contractions are invariant under an invertible linear change of coordinates. Normalize $H=I_n$ and orthogonally diagonalize

$$B=\mathrm{diag}(b_1,\dots ,b_n),b_i\ge 0.$$

For $T_{ijk}={\phi }_{ijk}$, symmetry of the third derivative tensor gives

$$A=\sum _{i,j,k}{b}_i{b}_j{T}_{ijk}^2,C=\sum _{i,j,k}{b}_i^2{T}_{ijk}^2.$$

Therefore

$$C-A=\frac{1}{2}\sum _{i,j,k}(b_i-b_j{)}^2{T}_{ijk}^2\ge 0.$$

Point-dependent normalization

The coordinate change may depend on the point, but it is not differentiated. This is a pointwise algebraic inequality between complete tensor contractions, so no correction term appears.

Since $D\ge 0$,

$${\mathbb{E}}_{\nu }q\ge 2{\mathbb{E}}_{\nu }A.$$

Now set $F=B^{1/2}H{B}^{1/2}$. Since ${\mathbb{E}}_{\nu }F=B$, entrywise Brascamp-Lieb gives

$${\mathbb{E}}_{\nu }\Vert F-B{\Vert }_{\mathrm{HS}}^2={\mathbb{E}}_{\nu }q-\mathrm{Tr}(B^2)\le {\mathbb{E}}_{\nu }A\le \frac{1}{2}{\mathbb{E}}_{\nu }q.$$

Rearranging proves

$${\mathbb{E}}_{\nu }q\le 2\mathrm{Tr}(B^2).$$

For indefinite $B$, use an eigenbasis:

$$\mathrm{Tr}(BHBH)=\sum _{i,j}{b}_i{b}_j{H}_{ij}^2\le \sum _{i,j}|b_i||b_j|H_{ij}^2=\mathrm{Tr}(|B|H|B|H).$$

The positive-semidefinite result for $|B|$ completes the matrix estimate.

🏷️ From Hessian Energy to Quadratic Variance

Suppose that $M$ is invertible and put

$$S=\mathrm{sgn}(M),Z=|M{|}^{1/2}X.$$

Then $S$ is symmetric and orthogonal, while

$$X^{\mathsf{T}}MX=Z^{\mathsf{T}}SZ.$$

If $\eta$ is the law of $Z$, then

$${\tau }_{\eta }(Z)=|M{|}^{1/2}{\tau }_{\mu }(X)|M{|}^{1/2}=|M{|}^{1/2}H(Y)|M{|}^{1/2}$$

is its Stein kernel. The random-Hessian estimate gives

$$\mathbb{E}\Vert {\tau }_{\eta }(Z){\Vert }_{\mathrm{HS}}^2={\mathbb{E}}_{\nu }\mathrm{Tr}(|M|H|M|H)\le 2\mathrm{Tr}(M^2).$$

For

$$G(z)=z^{\mathsf{T}}Sz-\mathrm{Tr}(M),$$

we have ${\partial }_{i}G(z)=2⟨z,S{e}_i⟩$. Hence

$$\begin{array}{rl}\mathrm{Var}(Z^{\mathsf{T}}SZ)& \le 4\mathbb{E}\sum _i|{\tau }_{\eta }(Z)S{e}_i{|}^2\\ & =4\mathbb{E}\Vert {\tau }_{\eta }(Z)S{\Vert }_{\mathrm{HS}}^2\\ & =4\mathbb{E}\Vert {\tau }_{\eta }(Z){\Vert }_{\mathrm{HS}}^2\\ & \le 8\mathrm{Tr}(M^2).\end{array}$$

The change of variables is essential. Directly treating $X^{\mathsf{T}}MX$ would produce $\Vert {\tau }_{\mu }M{\Vert }_{\mathrm{HS}}^2$, which the Hessian estimate does not control because ${\tau }_{\mu }$ and $M$ need not commute.

If $M$ is singular, replace it by

$$M_{\epsilon }=M+\epsilon P_{\ker M}$$

and let $\epsilon \downarrow 0$.

🏷️ Third Moments and the KLS Parameter

Define

$${\kappa }_n=\underset{\mu ,\theta }{\sup }{\Vert \mathbb{E}[⟨X,\theta ⟩X\otimes X]\Vert }_{\mathrm{HS}},$$

where $\mu$ ranges over isotropic log-concave laws and $\theta \in S^{n-1}$. Fix $\mu ,\theta$ and set

$$M=\mathbb{E}[⟨X,\theta ⟩X\otimes X].$$

Then

$$\begin{array}{rl}\Vert M{\Vert }_{\mathrm{HS}}^2& =\mathrm{Tr}(M^2)\\ & =\mathbb{E}[⟨X,\theta ⟩X^{\mathsf{T}}MX]\\ & =\mathbb{E}[⟨X,\theta ⟩(X^{\mathsf{T}}MX-\mathrm{Tr}M)].\end{array}$$

Cauchy-Schwarz, isotropy, and the quadratic-form theorem imply

$$\Vert M{\Vert }_{\mathrm{HS}}^4\le \mathrm{Var}(⟨X,\theta ⟩)\mathrm{Var}(X^{\mathsf{T}}MX)\le 8\Vert M{\Vert }_{\mathrm{HS}}^2.$$

Consequently,

$${\kappa }_n\le 2\sqrt{2}.$$

🏷️ Explicit Stochastic-Localization Absorption

Klartag’s localization estimates have the form (Klartag, 2023)

$$P\le C(a+1)\frac{\mathbb{E}\sqrt{\Vert A_t{\Vert }_{\mathrm{op}}}}{\sqrt{t}},0<t<a{\lambda }_0,$$

where $P=C_{\mathrm{P}}(\mu )$, ${\lambda }_0=P^{-1}$, and $A_t$ is the localized covariance. They also give

$$\mathbb{E}\Vert A_t{\Vert }_{\mathrm{op}}\le C$$

when

$$t\le \frac{c}{1+{\kappa }_n^2\log (en)}.$$

Jensen’s inequality bounds $\mathbb{E}\sqrt{\Vert A_t{\Vert }_{\mathrm{op}}}$ universally.

Choose

$$t=\frac{c_0}{1+{\kappa }_n^2\log (en)}$$

with $c_0$ small, and set

$$a=2tP.$$

Since

$$a{\lambda }_0=2tP\cdot P^{-1}=2t,$$

the condition $t<a{\lambda }_0$ is automatic. Substitution gives

$$P\le C\frac{1+2tP}{\sqrt{t}}=\frac{C}{\sqrt{t}}+2C\sqrt{t}P.$$

Choose $c_0$ so that $2C\sqrt{t}\le 1/2$ and absorb:

$$P\lesssim \frac{1}{\sqrt{t}}\lesssim \sqrt{1+{\kappa }_n^2\log (en)}.$$

The bound ${\kappa }_n\le 2\sqrt{2}$ therefore yields

$$C_{\mathrm{P}}(\mu )\lesssim \sqrt{\log (en)}.$$

Taking the square root in $\psi (\mu {)}^2\asymp C_{\mathrm{P}}(\mu )$ gives

$${\psi }_n\lesssim {\log }^{1/4}(en).$$

Absence of circularity

The unknown $P$ enters $a=2tP$, but the admissibility condition simplifies to $2t>t$. Its occurrence on the right has coefficient $O(\sqrt{t})$ and is absorbed, so no prior estimate for $P$ is assumed.

🏷️ Sharpness, Approximation, and Status

Let $Y_i$ be independent mean-one exponential variables and set $X_i=Y_i-1$. Then

$$\mathbb{E}{X}_i^2=1,\mathbb{E}{X}_i^4=9,\mathrm{Var}(X_i^2)=8.$$

Taking $M=I_n$ gives

$$\mathrm{Var}(|X{|}^2)=8n=8\mathrm{Tr}(I_n^2),$$

so the constant is sharp. The scalar case $M=I_n$ is also proved independently in the contemporaneous Chen-Klartag thin-shell preprint (Chen & Klartag, 2026).

The argument first treats smooth densities on bounded convex supports. Smoothing, truncation to convex sublevel sets, and affine isotropic normalization reduce a general isotropic log-concave measure to this case. The sublevel thresholds may be chosen as regular values; equivalently, the convex sublevel sets may be approximated by smooth convex sets. Only moments through degree four are needed to pass the final inequality to the limit.

Status of the result

The paper was submitted on July 27, 2026, and arXiv currently lists only version 1 (Letwin, 2026). The matrix-energy argument and the factor $8$ are internally consistent. The KLS corollary needs the explicit absorption argument above, which the preprint omits. The paper also discloses AI assistance in discovering the Hessian estimate and preparing regularity arguments; the displayed derivations, rather than their provenance, are the relevant basis for mathematical assessment.

Links

• on Eldan’s stochastic localization and the KLS conjecture --- develops the covariance process and spectral-gap mechanism used in the final corollary.
• on an almost constant lower bound in the KLS conjecture --- presents Chen’s recursive covariance bootstrap and the preceding subpolynomial KLS bound.
• on KLS localization lemma --- explains the deterministic needle method that preceded stochastic localization.

📚 References

🐻  Barthe, F. & Klartag, B. 2020. Spectral Gaps, Symmetries and Log-Concave Perturbations. Bulletin of the Hellenic Mathematical Society 64, 1–31.

🐻  Brascamp, H.J. & Lieb, E.H. 1976. On Extensions of the Brunn–Minkowski and Prékopa–Leindler Theorems, Including Inequalities for Log-Concave Functions, and with an Application to the Diffusion Equation. Journal of Functional Analysis 22(4), 366–389.

🐻  Chen, Y. & Klartag, B. 2026. Digesting the Proof of the Sharp Thin-Shell Inequality.

🐻  Cordero-Erausquin, D. & Klartag, B. 2015. Moment Measures. Journal of Functional Analysis 268(12), 3834–3866.

🐻  Fathi, M. 2019. Stein Kernels and Moment Maps. The Annals of Probability 47(4), 2172–2185.

🐻  Klartag, B. 2023. Logarithmic Bounds for Isoperimetry and Slices of Convex Sets. Ars Inveniendi Analytica, Paper No. 4, 17 pp.

🐻  Letwin, B. 2026. The KLS Constant Is O(\log^1/4 n).