When Gaussian tails in every direction control nonlinear fluctuations

← All digests

Family 093 · Convex and metric geometry
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-14

1. The question

Suppose a high-dimensional random vector looks Gaussian in one limited sense: every linear projection has a Gaussian-type tail. Does that force good concentration for nonlinear observations as well, such as distance from a fixed set?

For arbitrary distributions, linear information can miss complicated internal structure. The question here adds log-concavity, a convexity condition on the density that rules out separated peaks. The manuscript claims that this combination is enough for a dimension-free logarithmic Sobolev inequality.

Such an inequality is substantially stronger than a tail estimate for one coordinate. It controls an entropy functional for every smooth test function and implies Gaussian concentration for Lipschitz functions. “Dimension-free” means that the constant depends on the common one-dimensional tail scale, but does not deteriorate merely because the vector has more coordinates.

2. Background and essential definitions

A probability density $\rho$ on $\mathbb R^n$ is log-concave if its positive set is convex and $\log\rho$ is concave there. Equivalently, one can write $\rho=e^{-V}$ with a convex potential $V$, allowing $V=+\infty$ outside the support. Uniform measures on convex bodies fit this definition, despite their discontinuous boundary densities.

Let $X$ have law $\mu$, and assume it is centered: $\mathbb EX=0$. The source calls it linearly $a$-subgaussian when

$$\sup_{|\theta|=1}\mathbb E\exp\left(\frac{\langle X,\theta\rangle^2}{a^2}\right)\leq2.$$

Every unit direction has the same exponential-square bound. Markov’s inequality then gives a tail bound of the form $\mathbb P(|\langle X,\theta\rangle|\geq t)\leq2e^{-t^2/a^2}$. This controls more than variance: it restricts rare large deviations in every direction.

For a nonnegative integrable function $g$, define

$$\operatorname{Ent}_\mu(g) =\int g\log g\,d\mu -\left(\int g\,d\mu\right)\log\left(\int g\,d\mu\right),$$

with $0\log0=0$. This is the entropy functional of the test function relative to $\mu$, not simply the differential entropy of the random vector.

A logarithmic Sobolev inequality compares $\operatorname{Ent}_\mu(f^2)$ with the Dirichlet energy $\int|\nabla f|^2d\mu$. The gradient measures local variation. The inequality therefore says that a function cannot concentrate its squared mass too sharply without paying a gradient cost.

3. Development history and previous partial results

Gross’s Logarithmic Sobolev inequalities (1975) established the Gaussian framework and its connection to hypercontractivity. One of its characteristic features is a constant independent of the number of Gaussian coordinates.

The Bakry–Émery diffusion method gives such inequalities when a smooth potential has a uniformly positive Hessian: $D^2V\geq\kappa I$, $\kappa>0$. This is a local curvature condition. Ordinary log-concavity only requires nonnegative curvature and allows flat regions, including the interior of a uniform convex-body distribution.

Logarithmic Sobolev inequalities imply Gaussian concentration for all Lipschitz functions through Herbst’s argument. Under log-concavity, results of Milman connect strong concentration back to isoperimetry and logarithmic Sobolev inequalities. The unresolved step addressed here is whether it is enough to know concentration only for linear functions.

Bizeul formulated that sufficiency conjecture in work first posted in 2023, On the log-Sobolev constant of log-concave vectors. Earlier norm-based criteria gave an entropy coefficient of order $na^2$, because controlling all coordinates gives a subgaussian scale of order $\sqrt n\,a$ for the full Euclidean norm. Bizeul improved the coefficient to order $\sqrt n\,a^2$, and established a dimension-free conclusion under rotational invariance.

These coefficients must be compared in the same convention. Some authors define a log-Sobolev parameter through its square; a bound of order $n^{1/4}a$ for that parameter corresponds to a coefficient of order $\sqrt n\,a^2$ in the displayed entropy inequality.

The manuscript claims removal of the remaining dimension factor without rotational symmetry. Its information-theoretic tools also build on Guo, Shamai and Verdú’s Gaussian-channel identities, which relate information gained from noisy observations to mean-square prediction error.

4. What the new manuscripts claim

M1 states that a universal constant $C<\infty$ exists such that, in every dimension, every centered log-concave probability measure with a Lebesgue density and linear subgaussian parameter $a$ satisfies

$$\operatorname{Ent}_\mu(f^2) \leq Ca^2\int|\nabla f|^2\,d\mu \qquad(f\in C_c^\infty(\mathbb R^n)).$$

Here $C_c^\infty$ denotes smooth compactly supported real test functions. The density need not be smooth, strictly positive everywhere, or uniformly curved. Uniform measures on convex bodies are included whenever the stated marginal condition holds.

The manuscript also gives a quadratic transport–entropy consequence. For probability measures with finite second moments, the Wasserstein distance is

$$W_2(\nu,\mu)^2 =\inf_{\pi}\int|x-y|^2\,d\pi(x,y),$$

where $\pi$ ranges over joint distributions with marginals $\nu$ and $\mu$. It is the least mean-square cost of moving one distribution to the other.

If $\nu=f\mu$, its relative entropy is $H(\nu\mid\mu)=\int f\log f\,d\mu$; it is infinite when absolute continuity fails. By the established Otto–Villani implication and an approximation argument, M1 states

$$W_2(\nu,\mu)^2\leq Ca^2H(\nu\mid\mu).$$

The same universal coefficient convention is used in the two displayed statements. This consequence compares entire probability distributions, rather than only individual linear projections.

5. Why the advance matters

Importance: 5/5. The claimed theorem would give a sharp qualitative bridge from one-dimensional marginal tails to a strong global functional inequality throughout the log-concave class. It removes a dimension-dependent gap in a natural conjecture linking convexity, concentration and entropy.

The hypothesis is often easier to interpret than a logarithmic Sobolev inequality. It asks about scalar measurements $\langle X,\theta\rangle$. The conclusion controls arbitrary smooth nonlinear tests and, through standard consequences, Lipschitz fluctuations and transport costs.

Dimension-free concentration concerns fluctuations around a typical value, not the size of that value. The Euclidean norm of a high-dimensional Gaussian vector is typically of order $\sqrt n$, even though its deviations have a dimension-independent scale. There is no contradiction between the theorem and growth of the typical radius with dimension.

The result should also be distinguished from conjectures assuming only normalized covariance. A covariance bound controls second moments; the exponential-square condition controls much further into the tails. The manuscript does not assert a universal log-Sobolev constant for every isotropic log-concave measure without this stronger hypothesis.

6. The main idea

Normalize a hypothetical failure

The proof proceeds by contradiction. If no universal ratio between the log-Sobolev coefficient and $a^2$ existed, suitable truncation, smoothing and rescaling would produce measures whose log-Sobolev constants stay normalized while their linear subgaussian parameters tend to zero.

The argument then seeks incompatible descriptions of the fluctuations that could sustain such an extremal inequality. It treats separately the case where the entropy constant exceeds the Poincaré constant and the case where they coincide. A Poincaré inequality controls variance rather than entropy; its first nonconstant eigenfunction supplies the relevant object in the second case.

Observe an extremal field through Gaussian noise

From an extremizing function, the manuscript constructs a related probability law and a vector field resembling the gradient of a log-density ratio. Variational equations make that field behave approximately like a first diffusion eigenfunction.

The additional information comes from prediction. Observe a vector after adding independent Gaussian noise, and ask how much of the field can still be recovered by conditional expectation. Matching entropy contraction with a separate entropy-loss comparison forces a specific limiting prediction curve.

Repeated observations of independent copies control the cost of discarding conditional-mean directions. This information step is more restrictive than knowing a spectral estimate alone: it prescribes how predictability must deteriorate across a whole interval of noise levels.

Turn near-equality into too much information

Gaussian polynomial orthogonality converts the prediction curve into a hierarchy of nearly extremal derivative tensors. A tensor is simply a multi-index array, here recording repeated derivatives and their symmetries. Averaging matrix quantities built from these tensors produces a positive matrix field with normalized trace mass.

The symmetries force approximate commutation properties. The manuscript derives a scalar lognormal fluctuation law under a weighted trace, without assuming that all matrices in varying dimensions can be simultaneously diagonalized.

Finally, it considers Gaussian observations whose covariance depends on this matrix field. The log-Sobolev inequality bounds the information such observations can reveal. The derived fluctuation law forces a conflicting lower bound when diffusion time and a covariance cap are enlarged in the prescribed order.

Integrating over observation scales ensures that the contradiction retains matrix mass even when it is spread among many small eigenvalues. The contradiction rules out the normalized sequence of failures and yields existence of a universal constant. It is not a computation of the best constant.

7. Scope, limitations, and open questions

The measure is centered, log-concave and has a density with respect to full-dimensional Lebesgue measure. The theorem as stated is not a direct assertion about arbitrary singular measures or nonconvex mixtures.

The subgaussian bound must hold uniformly over every unit direction. Testing only the coordinate axes is weaker, and controlling covariance alone is weaker still. The parameter $a$ may depend on the distribution; dimension-free means no extra dependence on $n$ beyond that parameter.

No positive curvature lower bound or density smoothness is required in the final statement. The regularizations used in the proof are part of its argument, not hidden hypotheses imposed on the original measure.

Optimal constants, efficient sampling consequences for a particular access model, and related statements under weaker marginal assumptions are separate questions. The digest reports the manuscript’s theorem and contradiction strategy without verifying its information or tensor estimates.

8. Sources and further reading

M1 states the log-Sobolev and transport inequalities and carefully specifies its entropy conventions. Its introduction outlines the extremal-law construction, prediction identity, tensor hierarchy and information contradiction.

Bizeul’s linked paper is the primary source for the conjecture and earlier dimension-dependent bound. Guo–Shamai–Verdú supplies the Gaussian-channel background. Otto and Villani’s Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality (2000), cited in M1, is the source of the transport implication.

Source manuscripts

• M1. OpenAI, A dimension-free logarithmic Sobolev inequality for subgaussian log-concave measures (September 23, 2026).

Sources consulted on 14 October 2026 at repository revision 1581d4c. Download the manuscript citations.