Random vectors in the isotropic position (Q1294094)

Let \(y_1,\dots, y_M\) be independent copies of a random vector \(y\) in \(\mathbb{R}^n\) satisfying \(\mathbb{E} y\otimes y= \text{Id}\). The author studies the question of approximation of the identity map by sums \(y_j\otimes y_j\) and proves that \[ \mathbb{E} \Biggl\|{1\over M}\sum^M_{j= 1} y_j\otimes y_j- \text{Id}\Biggr\|\leq c{\sqrt{\log n}\over\sqrt M} (\mathbb{E}\| y\|^{\log M})^{1/\log M}, \] provided the last expression is smaller than 1, \(c\) being an absolute constant. The proof uses a noncommutative Khintchine inequality. The author applies the formula to improve a result of Bourgain about the number of points needed to bring a convex body into nearly isotropic position: namely, if \(y_1,\dots, y_M\) are independent random vectors uniformly distributed in a convex body \(K\) in isotropic position, not to big in \(\ell_2\)-norm, \(\mathbb{E}\left\|{1\over m}\sum^M_{j= 1} y_j\otimes y_j- \text{Id}\right\|\leq \varepsilon\) holds, where \(M\) is less than \(cn\log n/\varepsilon^2\). Another application is a better estimate on the number of contact points of approximate John-decompositions of the identity map, of order \(\leq cn\log n/\varepsilon^2\).