Pointwise estimates for marginals of convex bodies (Q2426508)

In this paper the authors deal with a pointwise version of the multi-dimensional central limit theorem for convex bodies, earlier introduced and discussed by \textit{B. Klartag} [Invent. Math. 168, No.~1, 91--131 (2007; Zbl 1144.60021), J. Funct. Anal. 245, No.~1, 284--310 (2007; Zbl 1140.52004)]. Suppose that \(X\) is a random vector in \(\mathbb{R}^n\) that is distributed uniformly in some convex set \(K\subset\mathbb{R}^n\), and, for a subspace \(E\subset\mathbb{R}^n\), denote by \(\mathrm{Proj}_E\) the orthogonal projection operator onto \(E\) in \(\mathbb{R}^n\). From the above quoted papers by B. Klartag it derives that there exists a subspace \(E\subset\mathbb{R}^n\), with \(\mathrm{dim}(E)>n^c\), such that the random vector \(\mathrm{Proj}_E(X)\) is approximately Gaussian in the total variation sense. Here the authors get a stronger conclusion. Let \(X\) be an isotropic random vector in \(\mathbb{R}^n\) with a log-concave density, and let \(l\) be an integer, \(1\leq l\leq n^{c_1}\). Let \(G_{n,l}\) be the Grassman manifold of all \(l\)-dimensional subspaces of \(\mathbb{R}^n\), provided with its unique rotationally-invariant probability measure \(\mu_{n,l}\). Then there exists a subset \(\mathcal{E}\subseteq G_{n,l}\) with \(\mu_{n,l}(\mathcal{E})\geq 1-C\exp(-n^{c_2})\) such that, for any \(E\in\mathcal{E}\) the following holds. Denote by \(f_E\) the density of the random vector \(\mathrm{Proj}_E(X)\). Then \(\displaystyle{\left| \frac{f_E(x)}{\gamma(x)}-1\right| \leq\frac{C}{n^{c_3}}}\) for all \(x\in E\) with \(| x| \leq n^{c_4}\). Here \(\gamma(x)\) is the standard Gaussian density in \(E\), and \(C,c_1,c_2,c_3,c_4>0\) are universal constants. As it is pointed out in the paper, the basic idea of the proof of such an interesting result is based on the fact, proved by \textit{B. Klartag} [loc. cit.], that \(\mathrm{Proj}_E(X+Y)\) is pointwise approximately radial, where \(Y\) is an independent Gaussian random vector, and \(X+Y\) is concentrated in a thin spherical shell. By combining these properties it is proved th at the density of \(\mathrm{Proj}_E(X+Y)\) is not only radial, but also very close to the Gaussian density in \(E\). Then it is shown that the addition of \(Y\) is not required, so that the result follows