Two observations regarding embedding subsets of Euclidean spaces in normed spaces (Q817572)

The first part of the paper deals with the dependence of constants on parameters in the Dvoretzky theorem. Milman's version of the theorem says that there exists a function \(c(\varepsilon)>0\) such that for all \(n\leq c(\varepsilon)\log N\), \(\ell_2^n\) can be \((1+\varepsilon)\)-embedded into any normed space \(Y\) of dimension~\(N\) [\textit{V.\,D.\,Milman} [Funct.\ Anal.\ Appl.\ 5, 28--37 (1971; Zbl 0239.46018); see also \textit{V.\,D.\,Milman} and \textit{G.\,Schechtman}, ``Asymptotic theory of finite dimensional normed spaces. With an appendix by M. Gromov: Isoperimetric inequalities in Riemannian manifolds'' (Lecture Notes in Mathematics 1200, Springer-Verlag, Berlin et al.) (1986; Zbl 0606.46013)]. Milman's original proof gives \(c(\varepsilon)\sim \varepsilon^2/\log (2/\varepsilon)\). Later, \textit{Y.\,Gordon} [Isr.\ J.\ Math.\ 50, No.\,4, 265--289 (1985; Zbl 0663.60034)] proved that \(c(\varepsilon) \sim \varepsilon^2\) is sufficient. Both proofs also provide randomness, i.\,e., if one identifies \(Y\) with \((\mathbb{R}^N, \| \cdot\| _Y)\), then most subspaces (with respect to the normalized Haar measure on the Grassmann manifold) of \(Y\) are \((1+\varepsilon)\)-Euclidean. In the present paper, the author significantly improves the estimate. He shows that \(c(\varepsilon) \sim \varepsilon/\left(\log(2/\varepsilon)\right)^2\) is enough. However, the approach is not random. The best possible dependence on \(\varepsilon\) is still unknown for both random and nonrandom cases. Very recently, \textit{G.\,Schechtman} [``The random Version of Dvoretzky's theorem in \(\ell^n_\infty\)'', in: Geometric aspects of functional analysis. Papers from the Israel seminar (GAFA) 2004--2005 (Lecture Notes in Mathematics 1910, Springer-Verlag, Berlin et al.), 265--270 (2007; ISBN 978-3-540-72052-2)] has shown that for the random Dvoretzky theorem, \(c(\varepsilon)\) cannot be better than \(c \varepsilon\) for some absolute positive constant \(c\). In the nonrandom case, no lower bound except the trivial one \(c(\varepsilon)\geq 1/\log (2/\varepsilon)\) is known. In the second part of the present paper, the author generalizes the Johnson-Lindenstrauss lemma [Contemp.\ Math.\ 26, 189--206 (1984; Zbl 0539.46017)] as well as its recent generalization by \textit{B.\,Klartag} and \textit{S.\,Mendelson} [J.~Funct.\ Anal.\ 225, No.~1, 229--245 (2005; Zbl 1079.60033)]. The Johnson-Lindenstrauss lemma says that every \(k\)-point subset of an Euclidean space can be \((1+\varepsilon)\)-Lipschitz embedded into \(\ell _2^n\) with \(n\leq C \log k / \varepsilon ^2\), where \(C\) is an absolute positive constant. The author proves that there exists an absolute positive constant \(c\) such that for every normed space \(X\), for every subset \(T\) of the \((m-1)\)-dimensional Euclidean sphere, and for every \(\varepsilon >0 \), if \(E_T^* \leq c \varepsilon E(X)\), then there exists a linear operator \(A: \mathbb{R}^m \to X\) with \(1-\varepsilon\leq \| A t\| \leq 1+\varepsilon\) for all \(t\in T\). Here, \[ E_T^* = \mathbb{E} \sup \left\{ \left| \sum_{i=1}^m t_i g_i \right| \mid t=(t_1, t_2, \dots, t_m) \in T \right\} \] and \[ E(X) = \sup \left\{\mathbb{E} \left\| \sum_{i=1}^n g_i u(e_i) \right\| _X \mid n\in \mathbb{N}, u : \ell_2^n \to X, \| u\| =1\right\}, \] \(g_1, g_2,\dots\) are independent copies of a standard Gaussian random variable. Very recently, an alternative proof was observed by \textit{Y.\,Gordon} [``A note on an observation of G.\,Schechtman'', in: Geometric aspects of functional analysis. Papers from the Israel seminar (GAFA) 2004--2005 (Lecture Notes in Mathematics 1910, Springer-Verlag, Berlin et al.), 127--132 (2007; Zbl 1113.46301)].