Rademacher type and Enflo type coincide (Q2206349)

The main goal of the present paper is to prove that Rademacher type \(p\) implies Enflo type \(p\) for Banach spaces, which settles a long-standing open problem. Enflo type and Rademacher type can be defined in the following way. Consider a vertex set of an \(n\)-dimensional cube identified with the set of \(n\)-tuples of \(\pm 1\) in \(\mathbb{R}^n\). For the cube, there are naturally defined notions of edges and (main) diagonals. We say that a Banach space \(X\) has Enflo type \(p\in [1,2]\) if there exists \(C\in(0,\infty)\) so that, for all \(n\ge 1\) and all maps of \(f:\{-1,1\}^n\to X\), \[ \mathbb{E} d^p(\varepsilon)\leq C^p \sum_{j=1}^n \mathbb{E} s_j^p(\varepsilon), \] where \(\varepsilon\) is uniformly distributed on the discrete cube \(\{-1,1\}^n\), \(d(\varepsilon)\) is the distance between the images of the ends of the diagonal which joins \(\varepsilon\in\{-1,1\}^n\) and \(-\varepsilon\), and \(s_j(\varepsilon)\) is the distance between the images of the ends of the edge which starts at \(\varepsilon\) and goes to the vertex for which only the \(j\)th coordinate differs. Rademacher type \(p\in[1,2]\) is defined in the same way, but now we require the inequality to hold only for mappings \(f:\{-1,1\}^n\to X\) which are restrictions of linear mappings \(\mathbb{R}^n\to X\). Enflo type was introduced in [\textit{P. Enflo}, Ark. Mat. 8, 103--105 (1970; Zbl 0196.14002)] where it was used as a tool for solving important open problems in non-linear Banach space theory. Rademacher type was explicitly or implicitly used in many works; the early period of this development was crowned by the foundational work of \textit{B. Maurey} and \textit{G. Pisier} [Stud. Math. 58, 45--90 (1976; Zbl 0344.47014)], and Rademacher type became a standard tool in Banach space theory. It is immediately clear from the definitions that Enflo type \(p\) implies Rademacher type \(p\). The main result of the paper is the converse implication. The proof is based on a dimension-free analogue of \textit{G. Pisier}'s inequality [Lect. Notes Math. 1206, 167--241 (1986; Zbl 0606.60008)]: \[ \mathbb{E}_\varepsilon\|f(\varepsilon)-\mathbb{E}_\varepsilon f(\varepsilon)\|^p \leq C^p\log^p n\, \mathbb{E}_{\varepsilon,\delta}\left\|\sum_{j=1}^n \delta_j s_j(\varepsilon)\right\|^p, \] where \(p\ge 1\), \(f:\{-1,1\}^n\to X\), and \(\varepsilon,\delta\) are independent random vectors, uniformly distributed on \(\{-1,1\}^n\). The words dimension-free mean the possibility of replacement of \(\log n\) by a constant which does not depend on \(n\). It is known [\textit{M. Talagrand}, Geom. Funct. Anal. 3, 295--314 (1993; Zbl 0806.46035)] to be impossible for the inequality just stated. For this reason, the authors replace the random vector \(\delta\) by a vector of what the authors call biased Rademacher variables, and prove a dimension-free analogue of Pisier's inequality for these variables (Theorem 1.4), the actual statement is a bit more technical. These biased Rademacher variables appear when the authors develop a ``correct'' analogue of Pisier's dimension-free analogue of the inequality above for Gauss space [Pisier, loc. cit., Theorem 2.2] for a discrete cube \(\{-1,1\}^n\). The authors use their Theorem 1.4 to get an answer to another important open problem, namely, they prove (Theorem 1.5) that a Banach space \(X\) satisfies the dimension-free Pisier inequality stated above if and only if \(X\) has finite cotype. The paper is nicely written, short, contains answers to some long-standing open problems, and develops potentially very useful tools.