Embedding the Heisenberg group into a bounded-dimensional Euclidean space with optimal distortion (Q1998705)

Summary: Let \(H := \left(\begin{smallmatrix} 1 & \mathbb{R} & \mathbb{R} \\ 0 & 1 & \mathbb{R} \\ 0 & 0 & 1 \end{smallmatrix}\right)\) denote the Heisenberg group with the usual Carnot-Carathéodory metric \(d\). It is known (since the work of \textit{P. Pansu} [Ann. Math. (2) 129, No. 1, 1--60 (1989; Zbl 0678.53042)] and \textit{S. Semmes} [Rev. Mat. Iberoam. 12, No. 2, 337--410 (1996; Zbl 0858.46017)]) that the metric space \((H,d)\) cannot be embedded in a bilipschitz fashion into a Hilbert space; however, from a general theorem of \textit{P. Assouad} [Bull. Soc. Math. Fr. 111, 429--448 (1983; Zbl 0597.54015)], for any \(0 < \varepsilon \leq 1/2\), the snowflaked metric space \((H,d^{1-\varepsilon})\) embeds into an infinite-dimensional Hilbert space with distortion \(O( \varepsilon^{-1/2} )\). This distortion bound was shown by \textit{T. Austin} et al. [Groups Geom. Dyn. 7, No. 3, 497--522 (2013; Zbl 1284.46019)] to be sharp for the Heisenberg group \(H\). Assouad's argument allows \(\ell^2\) to be replaced by \(\mathbb{R}^{D(\varepsilon)}\) for some dimension \(D(\varepsilon)\) dependent on \(\varepsilon \). \textit{A. Naor} and \textit{O. Neiman} [Rev. Mat. Iberoam. 28, No. 4, 1123--1142 (2012; Zbl 1260.46016)] showed that \(D\) could be taken independent of \(\varepsilon \), at the cost of worsening the bound on the distortion to \(O( \varepsilon^{-1-c_D} )\), where \(c_D \to 0\) as \(D \to \infty \). In this paper we show that one can in fact retain the optimal distortion bound \(O( \varepsilon^{-1/2} )\) and still embed into a bounded-dimensional space \(\mathbb{R}^D\), answering a question of Naor and Neiman. As a corollary, the discrete ball of radius \(R \geq 2\) in \(\Gamma := \left(\begin{smallmatrix} 1 & \mathbb{Z} & \mathbb{Z} \\ 0 & 1 & \mathbb{Z} \\ 0 & 0 & 1 \end{smallmatrix}\right)\) can be embedded into a bounded-dimensional space \(\mathbb{R}^D\) with the optimal distortion bound of \(O(\log^{1/2} R)\). The construction is iterative, and is inspired by the Nash-Moser iteration scheme as used in the isometric embedding problem; this scheme is needed in order to counteract a certain ``loss of derivatives'' problem in the iteration.