Sharp entropy bounds and the simplex isotropic constant

Source and verification

This article explains family 101 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

🏷️ Covariance, entropy, and affine shape

For a convex body $K\subset {\mathbb{R}}^n$, let $X$ be uniform on $K$ and ${\Sigma }_K=\mathrm{Cov}(X)$. The isotropic constant in this article is

$$L_K={\left(\frac{\det {\Sigma }_K}{|K{|}^2}\right)}^{1/(2n)}.$$

For an invertible affine change of variables, covariance determinant and squared volume acquire the same factor. Thus $L_K$ is affine invariant.

For a full-dimensional probability density $f$ on ${\mathbb{R}}^m$, differential entropy is

$$h(f)=-\int f\log f.$$

A log-concave density has the form $e^{-V}$ for a convex extended-valued potential $V$. It can have a proper convex support, so exponential and uniform densities are included.

🏷️ The sharp statements in the manuscript

Entropy statement

Every full-dimensional log-concave probability density on ${\mathbb{R}}^m$ satisfies

$$h(f)\ge m+\frac{1}{2}\log \det \mathrm{Cov}(f).$$

Equality holds exactly for invertible affine images of independent mean-one exponential variables. (OpenAI, 2026)

An exponential variable with density $e^{-t}{1}_{\{t\ge 0\}}$ has entropy one and variance one. Products therefore give equality in the stated normalization. Affine maps add $\log |\det A|$ to entropy and twice that amount to log covariance determinant, preserving the deficit.

The geometric conclusion is

$$L_K\le \frac{(n!{)}^{1/n}}{(n+1{)}^{(n+1)/(2n)}\sqrt{n+2}},$$

with equality exactly for simplices. (OpenAI, 2026) The constant on the right is the isotropic constant of a simplex. The entropy assertion is a sharp lower bound within log-concave densities; the familiar Gaussian maximum-entropy principle is a different inequality.

🏷️ The cone reduction, including the constant

Translate $K$ to have centroid zero and define

$${\mathcal{C}}_K=\{(tx,t):t\ge 0,x\in K\}.$$

The density

$$f_K(y,t)=\frac{e^{-t}}{n!|K|}{1}_{{\mathcal{C}}_K}(y,t)$$

is log-concave. At height $t$, the cross-section has volume $t^n|K|$, so its integral is one by ${\int }_0^{\infty }{t}^n{e}^{-t}dt=n!$.

Equivalently, it is the law of $(TX,T)$, where $T$ is independent of $X$ and has density $t^n{e}^{-t}/n!$. Integration by parts gives

$$\begin{array}{rl}\mathbb{E}T& =n+1,\\ \mathbb{E}{T}^2& =(n+1)(n+2),\\ \mathrm{Var}(T)& =n+1.\end{array}$$

Since $\mathbb{E}X=0$, the cross-covariance of $TX$ and $T$ vanishes. Therefore

$$\begin{array}{c}h(f_K)=n+1+\log (n!|K|),\\ \mathrm{Cov}(f_K)\\ =\left(\begin{array}{cc}(n+1)(n+2){\Sigma }_K& 0\\ 0& n+1\end{array}\right).\end{array}$$

Applying the entropy inequality in dimension $n+1$ and cancelling $n+1$ gives

$$\begin{array}{c}(n!{)}^2|K{|}^2\\ \ge (n+1{)}^{n+1}(n+2{)}^n\det {\Sigma }_K.\end{array}$$

Taking the appropriate $2n$th root yields exactly the displayed bound on $L_K$. This calculation also explains the shifted dimension and the factorial, rather than treating them as unexplained normalizations.

🏷️ Equality and simplex sections

Equality in the body inequality is equivalent to equality for $f_K$ in the entropy inequality. Its closed support is the cone ${\mathcal{C}}_K$. An affine image of a product exponential law has support an affine orthant.

For this support to equal ${\mathcal{C}}_K$, its apex must agree with the cone’s apex. Thus the cone is generated by $n+1$ independent rays. Intersecting it with the height-one hyperplane gives the convex hull of the corresponding normalized ray points: a simplex.

Conversely, a simplex generates such a cone. After a linear change of cone coordinates, its exponential density becomes a product of one-dimensional exponential densities. This recovers equality. The equality classification uses support geometry as well as the numerical entropy bound. (OpenAI, 2026)

As a low-dimensional check, an interval of length $\ell$ has variance ${\ell }^2/12$, so $L_K=1/\sqrt{12}$. The formula at $n=1$ is $1/(2\sqrt{3})$, the same number.

🏷️ The transport proof of the entropy inequality

The substantial analytic claim is the entropy inequality itself. The manuscript regularizes the density and uses a Brenier map from a Gaussian law. Such a map is the gradient of a convex function; its Jacobian $J$ is a positive matrix in the smooth approximation.

The scalar benchmark transports a standard normal variable to a mean-one exponential:

$$t(a)=-\log \Phi (-a),$$

where $\Phi$ is the normal distribution function. If $q=t^{\prime }$ and $q^{-1}$ is applied spectrally to $J$, the resulting matrix field provides coordinates adapted to exponential, rather than Gaussian, equality.

An affine normalization imposes a mean condition on this nonlinear matrix field. It is not merely covariance normalization. Transport energy estimates retain off-diagonal eigenvalue interactions, which would be lost by pretending that matrix fields at different points share eigenvectors. (OpenAI, 2026)

The field is decomposed into its first Gaussian-chaos component and higher chaoses, that is, orthogonal Hermite degrees. Ornstein–Uhlenbeck covariance identities and a log-determinant comparison express the entropy deficit in terms of these pieces. Scalar inequalities and commutator estimates control the higher-order remainder.

The source includes rational certificates on compact parameter intervals together with analytic tail bounds. They were not executed here. Passing from smooth densities to general convex supports also requires near-equality control: equality cannot be recovered just by saying all nonnegative deficits converge.

In the proposed rigidity argument, vanishing higher-chaos errors and compatibility of mixed Jacobian derivatives force commuting first-order matrices. This produces independent scalar exponential transports after an affine change of variables and supplies the equality statement used in the cone reduction.

🏷️ Key steps and proof outlines

Entropy under transport separates a Jacobian term

For a smooth Brenier map $F=\nabla \varphi$ transporting standard Gaussian measure to the target density, the change-of-variables formula gives

$$h(F(G))=h(G)+\mathbb{E}\log \det DF(G).$$

Thus the proposed entropy lower bound becomes a comparison between the mean log determinant of the positive Jacobian and the covariance determinant of $F(G)$. A direct Jensen inequality goes in the wrong direction for some of the required terms and would lose the exponential equality case. The source instead compares the transport to the one-dimensional Gaussian-to-exponential map. (OpenAI, 2026)

Let $t(a)=-\log \Phi (-a)$ and $q=t^{\prime }$. Gaussian tail identities show $q^{\prime }>0$, so $q$ is invertible on its range. Applying $q^{-1}$ spectrally to the Jacobian defines a nonlinear matrix coordinate $A$. This is designed so the product-exponential transport has a simple first-order form in $A$, even though its original Jacobian is nonlinear in the Gaussian variable.

The correct affine position normalizes a nonlinear matrix field

Changing affine coordinates in the target modifies both covariance and the transport Jacobian. The manuscript chooses a position by a variational condition on the nonlinear field $A$, rather than simply setting covariance to the identity. The transport estimates establish existence of this position and the integrability needed for the logarithmic determinant.

Gaussian integration by parts connects derivatives of the transport to its covariance. On the degree-$d$ Hermite component, differentiation contributes a factor $d$ in the energy. An inverse Ornstein–Uhlenbeck multiplier cancels this factor in the covariance representation. Summing orthogonal chaos components gives an exact expansion with a first-chaos part and a nonnegative higher-order contribution.

The expansion retains matrix order. Products such as a covariance factor and a Jacobian factor need not commute, so cyclicity of the trace permits only cyclic rearrangement, not arbitrary reordering. The source’s matrix surplus isolates the terms that would be lost by treating every eigenvalue as an independent scalar process.

Scalar remainders and commutators control the surplus

The first-chaos component is close to an affine matrix field; higher chaoses measure its deviation from that form. Two scalar inequalities bound the nonlinear functional calculus errors and absorb them using a second-moment budget. Commutator estimates control the additional error caused by changing eigenvectors.

The scalar estimates are verified on a compact interval by rational enclosures and outside it by separate positive- and negative-tail bounds. The tail arguments use the differential identities of the Gaussian hazard function $q$. This division is necessary because a finite numerical interval cannot establish a global inequality for all eigenvalues. The article attributes those certificates to the manuscript and does not treat an unexecuted certificate as an independently checked computation.

Combining the exact covariance expansion, the log-determinant comparison, and these remainder estimates yields the nonnegative entropy deficit. The sharp value is inherited from the scalar exponential benchmark, not from the Gaussian maximum-entropy inequality.

Equality uses compatibility of Jacobian derivatives

If the deficit vanishes, the higher-chaos errors disappear and $A$ becomes a centered linear matrix field. That alone does not imply a product structure: its coefficient matrices could fail to commute. Because $J=q(A)$ is a Jacobian, however, its mixed derivatives satisfy

$$({\partial }_{r}J)e_s=({\partial }_{s}J)e_r.$$

The rigidity section expands this identity using analyticity of $q$ and the centered linear field $A(x)={\sum }_r{x}_r{X}_r$, with ${\sum }_r{X}_r^2=I$. Comparing coefficients first gives $X_r{e}_s=X_s{e}_r$ for coordinate vectors $e_r,e_s$, and then $[X_r,X_s]=0$. These identities supply both compatibility and simultaneous diagonalization, yielding scalar transport directions after an affine change. Each scalar direction is the Gaussian-to-exponential map, so the target is an affine image of independent exponentials.

For nonsmooth densities, equality cannot be inferred merely from convergence of entropy and covariance. The approximation theorem controls the near-equality fields as well, extracts a compatible limiting transport, and applies the same rigidity mechanism. Convex supports with hard boundaries, including the exponential orthant, are therefore covered by a genuine limiting argument.

The cone application keeps equality and normalization together

The explicit cone computation earlier turns the entropy deficit of $f_K$ into the isotropic-constant deficit of $K$. The Gaussian transport proof operates in dimension $n+1$, while the cone covariance contributes $(n+1{)}^{n+1}(n+2{)}^n$. Equality identifies the cone support as a linear orthant; its height-one slice is a simplex. Thus the factorial and shifted-dimensional factors come from the elementary gamma-law calculation, while the sharp inequality and equality classification come from the transport and rigidity analysis.

See Also

Mahler volume products uses cone Laplace integrals for another simplex extremum. Subgaussian logarithmic Sobolev bounds relates log-concavity to entropy through gradient inequalities.

🏷️ Manuscripts and verification scope

• A sharp entropy bound and the simplex inequality for isotropic constants — pinned manuscript and companion source.

No family scope page was present at lean/docs/101.md in the pinned snapshot. This does not establish that all related formalization code is absent. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026. A sharp entropy bound and the simplex inequality for isotropic constants.