On nearly radial marginals of high-dimensional probability measures (Q973862)

Let \(\mu\) be a Borel probability measure on \(\mathbb R^n\) and \(\varepsilon>0\). \(\mu\) is called decently high-dimensional with parameter \(\varepsilon\), or \(\varepsilon\)-decent for short, if for any linear subspace \(E\subset\mathbb R^n\), \(\mu(E)\leq \varepsilon\dim(E)\). In particular, \(\mu\) is decent if it is \(\varepsilon\)-decent for \(\varepsilon=1/n\), the minimal possible value of \(\varepsilon\). Clearly, all absolutely continuous probability measures on \(\mathbb R^n\) are decent, as are many discrete measures. Note that a decent measure \(\mu\) necessarily satisfies \(\mu(\{0\})=0\), however, this feature should not be taken too seriously. A measure \(\mu\) is weakly \(\varepsilon\)-decent if \(\mu(E)\leq \varepsilon\dim(E)\) holds for all subspaces \(E\subset\mathbb R^n\) except \(E=\{0\}\). For a measure \(\mu\) on a measurable space \(\Omega\) and a measurable map \(T:\Omega \to \Omega'\), denote by \(T_*(\mu)\) the push-forward of \(\mu\) under \(T\), i.e., \(T_*(\mu)(A)=\mu(T^{-1}(A))\) for all measurable sets \(A\subset \Omega'\). When \(\mu\) is a probability measure on \(\mathbb R^n\) and \(T:\mathbb R^n \to\mathbb R^l\) is a linear map with \(l< n\), \(T_*(\mu)\) is a marginal of \(\mu\), or a measure projection of \(\mu\). The classical Dvoretzky theorem asserts that appropriate geometric projections of any high-dimensional convex body are approximately Euclidean balls, see \textit{V. D. Milman}, [Geom. Funct. Anal., 2, No.~4, 455--479 (1992; Zbl 0787.46016)] and references therein. The analogous statement for probability measures should perhaps be the following [see \textit{M. Gromov}, Geometric aspects of functional analysis, Isr. Semin. 1986--87, Lect. Notes Math. 1317, 132--184 (1988; Zbl 0664.41019)]: Appropriate measure projections of any decent high-dimensional probability measure are appropriately spherically-symmetric. When we can say that a probability measure \(\mu\) on \(\mathbb R^d\) is approximately radially-symmetric? Let \(\mu\) be a finite measure on a measurable space \(\Omega\). For a subset \(E\subset \Omega\) with \(\mu(A)>0\) we write \(\mu|_A\) for the conditioning of \(\mu\) on \(A\), i.e., \(\mu|_A(B)=\mu(A\cap B)/\mu(A)\) for any measurable set \(B\subset \Omega\). Let \(S^{d-1}\) be the unit sphere centered at the origin in \(\mathbb R^d\). The uniform probability measure on \(S^{d-1}\) is denoted by \(\sigma_{d-1}\). For two probability measures \(\mu\) and \(\nu\) on the sphere \(S^{d-1}\) and \(1\leq p<\infty\), we write \(W_p(\mu, \nu)\) for the \(L^p\) Monge-Kantorovich transportation distance between \(\mu\) and \(\nu\) in the sphere \(S^{d-1}\) endowed with the geodesic distance. The metrics \(W_p\) are all equivalent (we have \(W_1\leq W_p\leq \pi W_1^{1/p}\)) and they metrize the weak spherical shell \(S(J)=\{x\in \mathbb R^d: |x|\in J\}\), where \(|\cdot|\) is the standard Euclidean norm in \(\mathbb R^d\). The radial projection in \(\mathbb R^d\) is the map \({\mathcal R}(x) = x/|x|\). Let \(\mu\) be a Borel probability measure on \(\mathbb R^d\) and let \(\varepsilon>0\). We say that \(\mu\) is \(\varepsilon\)-radial if for any interval \(J\subset (0, \infty)\) with \(\mu(S(J))\geq \varepsilon\), we have \(W_1({\mathcal R}_*(\mu|_{S(J)}), \sigma_{d-1})\leq \varepsilon\). An \(\varepsilon\)-radial measure \(\mu\) is said to be proper if \(\mu(\{0\})=0\). The main result of the paper is the following Theorem. There exists a universal constant \(C>0\) for which the following holds: Let \(0< s < 1\) and let \(d\), \(n\) be positive integers. Suppose that \(n\geq (C/\varepsilon)^{C d}\). Then, for any decent probability measure \(\mu\) on \(\mathbb R^n\), there exists a linear map \(T:\mathbb R^n\to\mathbb R^d\) such that \(T_*(\mu)\) is \(\varepsilon\)-radial proper. Furthermore, let \(\eta>0\) be such that \(\eta^{-1}\geq (C/\varepsilon)^{C d}\). Then, for any \(\eta\)-decent probability measure \(\mu\) on \(\mathbb R^n\), there exists a linear map \(T:\mathbb R^n\to\mathbb R^d\) such that \(T_*(\mu)\) is \(\varepsilon\)-radial.