[Paper Review] Cutoff for non-negatively curved Markov chains

Justin Salez|arXiv (Cornell University)|Feb 10, 2021

Markov Chains and Monte Carlo Methods4 citations

TL;DR

This paper establishes the cutoff phenomenon for non-negatively curved Markov chains under a refined product condition, proving that mixing times exhibit a sharp transition when the relaxation time grows slower than the square root of the mixing time. The key contribution is a quantitative entropic concentration principle that implies cutoff universally across this class of chains, including random walks on abelian Cayley expanders with bounded degree.

ABSTRACT

Discovered in the context of card shuffling by Aldous, Diaconis and Shahshahani, the cutoff phenomenon has since then been established in a variety of Markov chains. However, proving cutoff remains a delicate affair, which requires a detailed knowledge of the chain. Identifying the general mechanisms underlying this phase transition -- without having to pinpoint its precise location -- remains one of the most fundamental open problems in the area of mixing times. In the present paper, we make a step in this direction by establishing cutoff for Markov chains with non-negative curvature, under a suitably refined product condition. The result applies, in particular, to random walks on abelian Cayley expanders satisfying a mild degree condition, hence in particular to \emph{almost all} abelian Cayley graphs. Our proof relies on a quantitative \emph{entropic concentration principle}, which we believe to lie behind all cutoff phenomena.

Motivation & Objective

• To identify general conditions under which the cutoff phenomenon occurs in Markov chains without requiring precise knowledge of the cutoff location.
• To address the fundamental open problem of characterizing mechanisms behind cutoff beyond ad hoc analysis.
• To extend the applicability of cutoff results to non-reversible chains, particularly those with non-negative curvature.
• To establish a universal mechanism—entropic concentration—for cutoff in a broad class of chains, including random walks on abelian Cayley graphs.
• To provide a criterion based on order-of-magnitude comparisons between relaxation time and mixing time, avoiding exact spectral computations.

Proposed method

• Introduces a quantitative entropic concentration principle as the central mechanism driving cutoff, linking entropy decay to mixing behavior.
• Uses the heat-kernel formulation $\mathscr{P}_t(x,y) = e^{-t} \sum_{k=0}^\infty \frac{P^k(x,y)t^k}{k!}$ to model continuous-time Markov chains on finite state spaces.
• Applies a $\frac{3}{4}$-idle transition matrix $\widehat{P} = \frac{3}{4}\mathrm{Id} + \frac{1}{4}P$ to lower-bound transition probabilities over diameter steps.
• Derives a diameter bound using Chebyshev's inequality and Poincaré inequality, relating mixing time, relaxation time, and graph diameter.
• Combines entropy estimates with diameter and mixing time bounds to show $\mathscr{V}_{\textsc{kl}}^\star(t_{\textsc{mix}}(\varepsilon)) \lesssim \log^2 \Delta \cdot t_{\textsc{mix}}(\varepsilon)$, implying cutoff.
• Establishes that $\sqrt{t_{\textsc{rel}}} \ll t_{\textsc{mix}}(\varepsilon)$ implies cutoff via entropic concentration, under non-negative curvature.

Experimental results

Research questions

• RQ1 Can the cutoff phenomenon be established for non-negatively curved Markov chains without precise knowledge of the cutoff time?
• RQ2 Is there a universal mechanism—beyond spectral or path-based analysis—that explains cutoff across diverse Markov chains?
• RQ3 Does the product condition $t_{\textsc{rel}} \ll t_{\textsc{mix}}(\varepsilon)$ imply cutoff for non-reversible chains with non-negative curvature?
• RQ4 Can entropic concentration principles be used to derive cutoff in random walks on abelian Cayley graphs under mild degree assumptions?
• RQ5 What is the role of the diameter and relaxation time in determining the sharpness of mixing in non-negatively curved chains?

Key findings

• Cutoff occurs for all non-negatively curved Markov chains satisfying a refined product condition, specifically when $\sqrt{t_{\textsc{rel}}} \ll t_{\textsc{mix}}(\varepsilon)$.
• The entropic concentration principle is identified as the underlying mechanism driving cutoff, providing a general framework beyond specific models.
• For random walks on abelian Cayley expanders with bounded degree, cutoff holds under the same condition, extending known results to a broader class.
• A diameter bound is derived: $\mathrm{diam}(\mathscr{X}) \leq (2+o(1))t_{\textsc{mix}}(\varepsilon)$, linking geometric and mixing properties.
• The entropy variation $\mathscr{V}_{\textsc{kl}}^\star(t_{\textsc{mix}}(\varepsilon))$ is bounded by $18 t_{\textsc{mix}}(\varepsilon) (1 + \log \Delta)^2$, which implies cutoff when $\sqrt{t_{\textsc{rel}}} \ll t_{\textsc{mix}}(\varepsilon)$.
• The result resolves a long-standing open problem by showing that cutoff is implied by order-of-magnitude comparisons of $t_{\textsc{rel}}$ and $t_{\textsc{mix}}(\varepsilon)$ in non-negatively curved chains.