[Paper Review] Intertwinings and Generalized Brascamp-Lieb Inequalities

Marc Arnaudon, Michel Bonnefont+1 more|arXiv (Cornell University)|Feb 11, 2016

Point processes and geometric inequalities29 references3 citations

TL;DR

This paper introduces a novel intertwining framework for multi-dimensional diffusions using weighted gradients to derive generalized Brascamp-Lieb inequalities beyond the classical log-concave setting. By constructing a matrix-weighted gradient and analyzing its intertwinement with the generator and a Feynman-Kac semigroup, the authors establish new spectral gap estimates and functional inequalities without requiring convexity of the potential, extending results from one-dimensional settings to higher dimensions with explicit lower bounds on the spectral gap.

ABSTRACT

We continue our investigation of the intertwining relations for Markov semigroups and extend the results of [9] to multi-dimensional diffusions. In particular these formulae entail new functional inequalities of Brascamp-Lieb type for log-concave distributions and beyond. Our results are illustrated by some classical and less classical examples.

Motivation & Objective

• To extend intertwining relations from one-dimensional diffusions to multi-dimensional diffusions on R^d for d ≥ 2.
• To derive new functional inequalities of Brascamp-Lieb type for non-log-concave and non-convex potentials.
• To establish lower bounds on the spectral gap of the diffusion generator using a weighted gradient approach.
• To generalize the Chen-Wang variational formula to higher dimensions via intertwining techniques.
• To illustrate the method on classical and non-classical examples, including a double-well potential with degenerate Hessian.

Proposed method

• Introduce a weighted gradient defined by an invertible matrix A, transforming the standard gradient into a distorted differential operator.
• Establish an intertwining identity between the Markov semigroup acting on functions and a Feynman-Kac-type semigroup acting on weighted gradients.
• Use the matrix A to diagonalize the Hessian of the potential V in a transformed space, enabling spectral analysis via the matrix M_A.
• Apply the Bakry-Émery Γ2 theory in the weighted setting to derive sub-commutation relations and functional inequalities.
• Construct a weight function a = e^{-W} such that the resulting matrix A^{-1}M_A A has uniformly positive lower spectral bound.
• Use variational estimates and Young’s inequality to control the lower bound of the smallest eigenvalue of the transformed Hessian matrix.

Experimental results

Research questions

• RQ1 Can intertwining relations between gradients and Markov semigroups be generalized from one-dimensional to multi-dimensional diffusions on R^d for d ≥ 2?
• RQ2 What functional inequalities of Brascamp-Lieb type can be derived when the potential V is not strictly convex or log-concave?
• RQ3 How can the spectral gap of a multi-dimensional diffusion generator be estimated without relying on convexity of the potential?
• RQ4 Can the Chen-Wang variational formula for the spectral gap be extended to higher-dimensional diffusions using intertwining techniques?
• RQ5 What is the role of matrix-weighted gradients in enabling spectral gap estimates for degenerate or non-convex potentials?

Key findings

• A generalized Brascamp-Lieb inequality is derived for the variance of functions under a non-log-concave measure, with a modified energy term involving a weighted gradient.
• An asymmetric Brascamp-Lieb inequality is established, bounding the covariance of two functions via the product of their weighted L^p norms.
• A new lower bound on the spectral gap λ₁(−L,μ) is derived as √(3/2 − β²) − 1 − β for a two-dimensional double-well potential with parameter β.
• The bound is valid for sufficiently small β > 0, ensuring positivity of the spectral gap estimate.
• The method successfully handles degenerate Hessian matrices by introducing a non-identity weight matrix A with different diagonal entries.
• The approach generalizes the one-dimensional Chen-Wang formula to higher dimensions through matrix-weighted gradient constructions.