Sharp hypocoercive convergence estimates for underdamped Langevin dynamics via the modified $L^2$ method

Abstract: In this note, we consider the underdamped Langevin dynamics with invariant measure $μ(\mathrm{d}x\,\mathrm{d}v) \propto e^{-U(x)-|v|^2/2}\,\mathrm{d}x\,\mathrm{d}v$. Assume that the position marginal $μx(\mathrm{d}x)\propto e^{-U(x)}\,\mathrm{d}x$ satisfies a Poincaré inequality with constant $m>0$, and that $\nabla^2 U\ge -K\,\mathrm{Id}$ for some $K\ge 0$. We revisit the modified $L^2$ method of Dolbeault--Mouhot--Schmeiser, employing a gap-shifted corrector \begin{equation*} A_m=(m- L{\mathrm{o}})^{-1}(L_aΠ_v)^*, \end{equation*} where $L_{\mathrm{o}}=Δ_x-\nabla U\cdot\nabla_x$ is the overdamped generator, $L_a$ is the generator of the Hamiltonian flow, and $Π_v$ denotes averaging over the velocity variable. We establish an explicit hypocoercive $L^2$-convergence rate \begin{equation*} Λ=\frac{1}{6\Bigl(\sqrt{2+\frac{K}{2m}}+\sqrt{4+\frac{K}{2m}}\Bigr)}\sqrt{m}. \end{equation*} In particular, for convex $U$, this recovers the optimal $O(\sqrt{m})$ rate.

• The paper derives sharp, explicit L2 convergence rates for underdamped Langevin dynamics using a gap-shifted corrector in the DMS framework.
• It introduces a modified L2 Lyapunov functional that leverages a gap-shifted corrector to improve spectral estimates and achieve optimal O(√m) dependence.
• The results impact high-dimensional sampling and computational methods by providing actionable guidelines for tuning kinetic model parameters.

Sharp Hypocoercive Convergence Estimates in Underdamped Langevin Dynamics via a Modified $L^2$ Approach

Problem Setting and Motivation

The paper "Sharp hypocoercive convergence estimates for underdamped Langevin dynamics via the modified $L^2$ method" (2604.10068) investigates explicit quantitative hypocoercive $L^2$ convergence rates for underdamped Langevin dynamics. The study considers the kinetic Fokker–Planck (KFP) equation associated with a confining $C^2$ potential $U:\mathbb{R}^d\to\mathbb{R}$ whose Hessian is bounded below by $-K$ for $K\ge 0$. The central question is to derive sharp, explicit $L^2(\mu)$ contraction rates for the Langevin semigroup, under only a Poincaré inequality for the position marginal, utilizing the Dolbeault–Mouhot–Schmeiser (DMS) hypocoercivity framework, specifically via a modified $L^2$ Lyapunov functional.

The kinetic system is described by

$$\begin{aligned} \mathrm{d}X_t &= V_t\,\mathrm{d}t \ \mathrm{d}V_t &= -\nabla U(X_t)\,\mathrm{d}t - \gamma V_t\,\mathrm{d}t + \sqrt{2\gamma}\,\mathrm{d}W_t \end{aligned}$$

with the invariant measure $L^2$0

Hypocoercivity is essential here since dissipation (through stochastic forcing) acts only in $L^2$1, while the KFP operator couples $L^2$2 via the Hamiltonian vector field, breaking coercivity in $L^2$3. Precise convergence rates are required in high-dimensional sampling and statistical mechanics, making the development of explicit, dimension-independent rates under minimal assumptions critical.

Methodology: Modified $L^2$4 Functional and Gap-Shifted Corrector

The analysis employs the DMS method, which augments the $L^2$5 norm with an antisymmetric correction term, yielding a Lyapunov functional sensitive to hidden modes and enabling derivation of exponential decay even for non-coercive generators. Conventionally, the corrector is constructed as

$L^2$6

where $L^2$7 is the Hamiltonian part of the generator, $L^2$8 projects onto velocity-averaged functions, and $L^2$9 is the standard resolvent.

However, following [Cao et al., 2023], it is known that this choice does not yield the optimal $L^2$0 rate in $L^2$1 (the Poincaré constant of the marginal), but rather $L^2$2 as $L^2$3. The paper's critical insight, not previously recorded in the literature, is to incorporate a "gap-shifted" corrector,

$L^2$4

where $L^2$5 is the overdamped Langevin generator, and $L^2$6 is its spectral gap. The corresponding Lyapunov function is

$L^2$7

This construction, using the operator $L^2$8 in place of $L^2$9, modifies the contribution of the slowest $C^2$0-modes (arising from the slow macroscopic component). The advantage is that for any non-constant eigenfunction $C^2$1 of $C^2$2 with eigenvalue $C^2$3, the corrector provides a prefactor of $C^2$4, ensuring robust macroscopic coercivity. In contrast, fixed-shift correctors lead to weaker contributions for small $C^2$5, and thus suboptimal rates.

Main Results

The main theorem establishes:

• For the underdamped Langevin dynamics under the Poincaré inequality and Hessian lower bound,
• The optimal choice of friction $C^2$6,
• The modified $C^2$7 functional with gap-shifted corrector yields

$C^2$8

with the explicit rate

$C^2$9

and for convex $U:\mathbb{R}^d\to\mathbb{R}$0 ($U:\mathbb{R}^d\to\mathbb{R}$1), recovers the optimal $U:\mathbb{R}^d\to\mathbb{R}$2 dependence.

This rate matches the optimal bounds achieved in [Cao et al., 2023] via space-time Poincaré inequalities and subsequent results using dynamical lifting. The proof uses fine spectral calculus, block matrix estimation of the dissipation functional, and an explicit optimization over auxiliary parameters.

Key technical assertions in the paper include:

• The spectral lift connecting the Hamiltonian generator $U:\mathbb{R}^d\to\mathbb{R}$3 and $U:\mathbb{R}^d\to\mathbb{R}$4 (Eq. (13)),
• Precise operator bounds for $U:\mathbb{R}^d\to\mathbb{R}$5, $U:\mathbb{R}^d\to\mathbb{R}$6, and related combinations,
• Sharp matrix inequality for the dissipation estimate and explicit calculation of its eigenvalues to deliver the dependence of $U:\mathbb{R}^d\to\mathbb{R}$7 on $U:\mathbb{R}^d\to\mathbb{R}$8 and $U:\mathbb{R}^d\to\mathbb{R}$9,
• Keeping the entire analysis in a single time-slice, as opposed to integral or path-space inequalities used in other methods.

Notably, the result demonstrates that the DMS $-K$0 framework, with a minimal modification, is sufficient to recover optimal hypocoercive rates under the classical Poincaré assumption, eliminating the need for more complex space-time or lifting-based arguments.

Implications and Theoretical Significance

The implications of this work extend across mathematical statistical mechanics, computational chemistry, and Bayesian computation, where reversible and non-reversible Langevin samplers are used for high-dimensional inference. Explicit dependence of convergence rates on geometric and spectral quantities of the confining potential informs the tuning of algorithmic parameters (e.g., the friction $-K$1 and temperature scaling), especially in regimes of weak macroscopic coercivity.

Theoretically, the result refines the understanding of hypocoercivity:

• It clarifies that the gap-shifting principle is sufficient for DMS-based Lyapunov analyses to achieve sharp rates.
• The technique allows for clean, resolvent-based arguments and could be adapted to other kinetic models or to quantum generalizations [see also (Li et al., 18 May 2025)].

Practically, the "single slice" technique has computational advantages for error bounds and variance control, suggesting potential future improvements in algorithmic analysis for kinetic samplers. Additionally, the approach raises the prospect of analyzing complex multiscale kinetic systems with degenerate noise using analogous corrected Lyapunov functionals.

Future Directions

Possible developments include:

• Extension to non-quadratic kinetic or non-Gaussian noise, where the velocity marginal is not explicitly solvable,
• Adaptation to finite-volume schemes or mesh-based discretizations for numerical hypocoercivity,
• Treatment of microcanonical or constrained dynamics where the invariant measure lacks product structure,
• Exploring analogous gap-shifted correctors for non-linear or Mean Field kinetic equations.

Connections with recent advances in non-reversible Markov chain acceleration (lifting, skew-symmetric perturbations) and their impact on metastability, cutoff phenomena, or hydrodynamic limits present compelling avenues for new research.

Conclusion

This paper establishes that the modified $-K$2 method, when equipped with a gap-shifted corrector reflecting the Poincaré constant, is sufficient to recover optimal hypocoercive $-K$3 convergence estimates for underdamped Langevin dynamics. The result sharpens the DMS framework and provides a self-contained, operator-theoretic approach complementing recent path-space and dynamical lifting methodologies. The techniques and insights here are poised to generalize to a broad class of degenerate parabolic and hypoelliptic systems.