From Poincaré inequalities to nonlinear matrix concentration (Q2040088): Difference between revisions

Let \(\mathbb{H}_d\) be the real linear space of all \(d\times d\) self-adjoint complex matrices, equipped with the \(\ell_2\to\ell_2\) operator norm \(\|\cdot\|\). Suppose that \(\mu\) is a probability measure (on some Polish space \(\Omega\)), which is a stationary measure of a reversible, homogeneous Markov process with \textit{carré du champ} operator \(\Gamma\) and Dirichlet form \(\mathcal{E}\). The goal of the paper under review is to deduce exponential matrix concentration from the Poincaré inequality, first the authors prove the equivalence of the scalar, trace, and matrix Poincaré inequalities (cf. [\textit{A. Garg} et al., ``Scalar Poincaré implies matrix Poincaré'', Preprint, \url{arXiv:2006.09567}]). One says that the (scalar) Poincaré inequality holds if for all functions \(f\colon \Omega\to\mathbb{R}\), \[ \operatorname{Var}_\mu(f) \leq \alpha \mathcal{E}(f).\tag{1} \] The variance and the quantities \(\Gamma(f)\), \(\mathcal{E}(f)\) can also be defined for matrix-valued functions (e.g., if \(f\colon \Omega\to\mathbb{H}_d\), then \(\operatorname{Var}_\mu(f) = ( f - \mathbb{E}_\mu f)^2 = \mathbb{E}_\mu f^2 - (\mathbb{E}_\mu f)^2\) is a positive semidefinite matrix). One says that the trace (resp. matrix) Poincaré inequality holds if for all \(d\in\mathbb{N}\) and all functions \(f\colon \Omega\to\mathbb{H}_d\) the inequality (1) is satisfied after taking the trace of both sides (resp. without the trace and with the inequality sign denoting the semidefinite order). It turns out that all these formulations are in fact equivalent (with the same constant \(\alpha\)). Under the assumption that the trace (equivalently: scalar or matrix) Poincaré inequality holds, the authors prove that if \(f\colon \Omega\to\mathbb{H}_d\) is sufficiently nice, then \[ \mathbb{P}_\mu(\| f- \mathbb{E}_\mu f\| \geq \sqrt{\alpha \nu_f}\, t) \leq 6d\, e^{-t}, \qquad t\geq 0, \] where \(\nu_f = \bigl\| \| \Gamma(f) \| \bigr\|_{L^\infty(\mu)}\) is the variance proxy. Under the assumption of the matrix Poincaré inequality such a result was earlier obtained in [\textit{R. Aoun} et al., Adv. Math. 371, Article ID 107251, 32 p. (2020; Zbl 1452.60017)]. The proof in the paper under review is rather short and avoids difficulties associated with direct extensions of the classical, scalar argument. It is based on the subadditivity of the (trace) Poincaré inequality, a chain rule inequality (in trace form), and an exponential moment bound (an estimate of \(\mathbb{E}_\mu \operatorname{tr} \cosh(\theta f) \)). The interested reader should also consult the companion paper [\textit{De Huang} and \textit{J. A. Tropp}, Electron. J. Probab. 26, Paper No. 8, 31 p. (2021; Zbl 1470.60013)] about local Poincaré inequalities.