The isotropic constant of a log-concave probability measure

Description of constant

Let $\mu$ be a Borel probability measure on $\mathbb{R}^n$ with finite second moments. Its covariance matrix is

\[\mathrm{Cov}(\mu) \ :=\ \int_{\mathbb{R}^n} (x-m)(x-m)^{\mathsf T}\, d\mu(x), \qquad m:=\int_{\mathbb{R}^n} x\, d\mu(x).\]

Convex bodies

If $K\subset\mathbb{R}^n$ is a convex body, let $\lambda_{K}$ be the uniform probability measure on $K$ and abbreviate $\mathrm{Cov}(K):=\mathrm{Cov}(\lambda_{K})$. The isotropic constant of $K$ is

\(L_{K} \ :=\ \left(\frac{\det \mathrm{Cov}(K)}{\mathrm{Vol}_{n}(K)^2}\right)^{1/(2n)}.\) This quantity is invariant under invertible affine maps.

Define

\[L_{n}^{\mathrm{body}} \ :=\ \sup\{L_{K}:\ K\subset\mathbb{R}^n \text{ a convex body}\}.\]

Log-concave probability measures

If $\mu$ is absolutely continuous with density $f$, its (differential) entropy is

\[\mathrm{Ent}(\mu)\ :=\ -\int_{\mathbb{R}^n} f \log f.\]

For an absolutely-continuous log-concave probability measure $\mu$ on $\mathbb{R}^n$, define its isotropic constant by

\[L_\mu\ :=\ e^{-\mathrm{Ent}(\mu)/n}\cdot \det(\mathrm{Cov}(\mu))^{1/(2n)}.\]

(If $\mu$ is log-concave but supported on a proper affine subspace, define $L_{\mu}$ in that subspace.)

If $K$ is a convex body, then $\lambda_{K}$ has constant density $1/\mathrm{Vol}_{n}(K)$ on $K$, hence $\mathrm{Ent}(\lambda_{K})=\log\mathrm{Vol}_{n}(K)$, and therefore

\[L_{\lambda_{K}} := e^{-\log(\mathrm{Vol}_{n}(K))/n}\cdot \det(\mathrm{Cov}(K))^{1/(2n)} := \left(\frac{\det \mathrm{Cov}(K)}{\mathrm{Vol}_{n}(K)^2}\right)^{1/(2n)} := L_{K}.\]

Define

\[L_{n}^{\mathrm{lc}} \ :=\ \sup\{L_\mu:\ \mu \text{ log-concave on }\mathbb{R}^n\}, \qquad C_{20b} \ :=\ \sup_{n\ge 1} L_{n}^{\mathrm{lc}}.\]

The isotropic constant problem asked whether $C_{20b}<\infty$ (i.e. whether $L_\mu$ is bounded by a universal constant, independent of the dimension). This is now known to be true.

Known upper bounds

Below, bounds are stated for $L_{n}^{\mathrm{body}}$ (equivalently for $L_{n}^{\mathrm{lc}}$ up to universal factors; see comments).

Bound Reference Comments
$L_{n}^{\mathrm{body}} \le C\,n^{1/4}\log n$ [Bou1991], [Bou2002] Bourgain’s classical bound
$L_{n}^{\mathrm{body}} \le C\,n^{1/4}$ [K2006] First removal of the $\log n$ factor
$L_{n}^{\mathrm{body}} \le \exp\big(C\sqrt{\log n}\,\log\log n\big)$ [Che2021] First “subpolynomial” bound
$L_{n}^{\mathrm{body}} \le C\,(\log n)^4$ [KL2022] First polylogarithmic bound
$L_{n}^{\mathrm{body}} \le C\,(\log n)^{2.223\ldots}$ [JLV2022]  
$L_{n}^{\mathrm{body}} \le C\,(\log n)^{2.082\ldots}$ [K2023] Lehec (personal communication)
$L_{n}^{\mathrm{body}} \le C\,\sqrt{\log n}$ [K2023]  
$L_{n}^{\mathrm{body}} \le C\,\log\log n$ [Gua2024]  
$C_{20b}<\infty$ (dimension-free) [KL2024] Final dimension-free bound (slicing/hyperplane theorem)

Known lower bounds

These are lower bounds for the extremal constant $C_{20b}=\sup_{n} L_{n}^{\mathrm{lc}}$ (i.e. examples with large isotropic constant).

Bound Reference Comments
$C_{20b}\ge 1/e \approx 0.367879$ Classical Achieved asymptotically by simplices (their isotropic constants tend to $1/e$)
$C_{20b}\ge 1/\sqrt{12}\approx 0.288675$ Classical Achieved by the cube $[-\tfrac12,\tfrac12]^n$ (volume $1$)

(Separately: for every log-concave probability measure $\mu$, one has the universal lower bound $L_\mu \ge 1/\sqrt{2\pi e}$, with equality for Gaussian measures; see [KL2024].)

References

Acknowledgements

Prepared with ChatGPT 5.2 Pro.