Optimal non-reversible linear drift for the convergence to equilibrium of a diffusion

Tony Lelièvre; Francis Nier; Grigorios A. Pavliotis

Source-linked AI summary

Optimal non-reversible linear drift for the convergence to equilibrium of a diffusion

Tony Lelièvre, Francis Nier, Grigorios A. Pavliotis

arXiv:1212.0876v1Dec 4, 2012math.NA

TL;DR

The paper studies whether non-reversible perturbations can accelerate convergence to equilibrium without changing the invariant distribution. For linear drifts, it proves that optimal perturbations exist, provides an algorithm to construct them, and analyzes the prefactor in the resulting exponential estimate.

Problem
The paper asks whether one can choose a non-reversible perturbation that maximizes convergence to equilibrium while preserving the invariant distribution.
Method
For linear drifts, the paper reduces perturbation design to an algebraic problem involving antisymmetric matrices and constructs optimal matrices through an implementable algorithm.
Results
Optimal matrices Jopt exist and attain the maximum convergence rate for the linear-drift problem.
Takeaways & Limitations
Non-reversible linear perturbations provide a systematic way to seek faster equilibrium convergence without changing the target distribution.
Takeaways & Limitations
The analysis is developed for linear drift, and controlling the prefactor for general non-self-adjoint semigroups is difficult.

Abstract

from arXiv · show

We consider non-reversible perturbations of reversible diffusions that do not alter the invariant distribution and we ask whether there exists an optimal perturbation such that the rate of convergence to equilibrium is maximized. We solve this problem for the case of linear drift by proving the existence of such optimal perturbations and by providing an easily implementable algorithm for constructing them. We discuss in particular the role of the prefactor in the exponential convergence estimate. Our rigorous results are illustrated by numerical experiments.

1 Introduction

The paper asks whether non-reversible perturbations can accelerate convergence to equilibrium without changing the invariant distribution, and solves this optimization problem for linear drifts. It constructs such perturbations through antisymmetric matrices while analyzing both the exponential rate and its prefactor.

Motivation: The reversible rate is governed by the spectral gap, while metastability can make this rate small and motivate non-reversible acceleration.The spectral gap is tied to the smallest nonzero eigenvalue of the Fokker–Planck operator.
Non-reversible diffusion: For linear vector fields, invariant-measure preservation is equivalent to using b = −JSx with J antisymmetric, reducing the design problem to choosing J.This gives the matrix parameterization used to construct non-reversible dynamics.
Non-reversible diffusion: The central problem is whether a perturbation can maximize convergence to equilibrium, which the paper solves for linear drift.The goal is to choose the perturbation optimally while retaining the invariant measure.
Non-reversible diffusion: Non-reversible perturbations preserve the invariant distribution and cannot worsen the exponential convergence rate relative to reversible dynamics.The paper uses perturbations that are divergence-free with respect to the invariant distribution, and establishes the same upper bound as in the reversible case.
Non-reversible diffusion: The analysis separates maximizing the exponential rate from controlling the prefactor C(V, b), which must remain sufficiently moderate.The paper first optimizes the rate and then studies the behavior of the prefactor for the optimal perturbation.

1.4 Outline of the paper

The paper studies optimal non-reversible perturbations for linear-drift diffusions, reducing spectral optimization to linear algebra and analyzing convergence rates and prefactors. It proves existence and constructibility of optimal matrices, extends convergence results to general initial data, and develops operator-theoretic estimates and possible extensions.

Linear algebraic analysis: The linear-drift setting is tractable because the convergence problem reduces to linear algebra involving the eigenvalues of BJ.The spectral gap is determined by eigenvalues of the drift matrix, while controlling the prefactor requires Wick calculus.
Main results: The paper proves that an optimal antisymmetric matrix Jopt exists and provides an algorithm for constructing it to achieve the best convergence rate λopt.The optimization concerns the spectrum of BJ=(I+J)S, which determines the spectral gap.
Quantitative convergence bounds: The analysis estimates the prefactor C(S,J), including its dependence on the condition number of S and the dimension N.The paper emphasizes this prefactor because it affects quantitative convergence bounds, especially for systems with multiple timescales.
Extensions and scope: The paper identifies extensions to Langevin dynamics and time-dependent antisymmetric perturbations while noting that some sharper bounds require additional regularity.It presents the linear-drift study as a first step toward analyzing nonlinear drift terms.
PDE and operator analysis: The PDE analysis establishes that the drift-diffusion operator generates a contraction semigroup with compact resolvent and supports exponential convergence estimates.The associated operator is studied in weighted L2 spaces, with regularity obtained through global ellipticity and semigroup analysis.
General initial data: The convergence result extends from Gaussian to general initial probability laws through conditioning on the initial data.The paper relates backward and forward Kolmogorov formulations through conjugation when treating probability densities.

2 A useful rescaling

The paper rescales the linear diffusion to simplify both the operator and invariant Gaussian, converting the problem to standard coordinates with identity covariance. This transformation preserves the relevant dynamics while enabling spectral analysis of the rescaled operator.

2 A useful rescaling: For quadratic potentials, the rescaled generator remains a drift-diffusion operator with drift matrix expressed through S and the antisymmetric perturbation.The transformed operator acts in ordinary L2(R^N,dx;C), where standard spectral techniques can be used.
2 A useful rescaling: The coordinate transformation is x=S^-1/2y with ∇x=S^1/2∇y, so the new process is Yt=S^1/2Xt.In these coordinates, the kernel becomes the standard Gaussian distribution.
2 A useful rescaling: The rescaled formulation preserves equivalence between the original and transformed evolutions, allowing estimates for one representation to transfer to the other.The initial data and semigroups are related by the unitary transformation induced by the Gaussian weight and coordinate change.

3 The linear algebra problem

The linear-drift optimization reduces convergence to a matrix problem for S+J̃: maximize the minimum real part of its spectrum while controlling the exponential prefactor. The authors characterize optimal perturbations and give a constructive algorithm, with a scope limitation for non-diagonalizable or repeated-eigenvalue cases.

3 The linear algebra problem: The spectrum of the linear drift matrix determines the generator’s spectral gap, reducing optimization of convergence to a finite-dimensional linear algebra problem.The transformed drift matrix is B̃J=S+J̃, with J̃ antisymmetric, and its spectrum agrees with that of the original drift matrix.
3 The linear algebra problem: The authors characterize optimal pairs (J̃,Q) through a positive-definite matrix Q and an associated antisymmetric matrix J̃ defined from an orthonormal eigenbasis.For simple eigenvalues of Q, the antisymmetric matrix is completely determined by Q and the basis conditions.
3 The linear algebra problem: The construction algorithm first builds an orthonormal basis satisfying the required condition, then chooses distinct positive eigenvalues and defines J̃.A Gram-Schmidt-style procedure constructs the basis and is used in the numerical experiments.
3 The linear algebra problem: The set Popt(S) does not include every admissible perturbation because the characterization assumes diagonalizability and simple eigenvalues of Q.Limits of optimal pairs can produce additional admissible matrices, including cases where the drift matrix has Jordan blocks.
3 The linear algebra problem: The optimal perturbation can strictly improve the exponential convergence rate if and only if S is not proportional to the identity.When S has two distinct eigenvalues, the relevant spectral inequality is strict; otherwise the reversible rate cannot be improved.
3 The linear algebra problem: Optimizing the exponential rate does not by itself optimize the prefactor, whose behavior depends on the conditioning of the associated matrices.The paper separately studies minimizing the constant multiplying exp(-Tr(S)t/2), including asymptotic constructions for the prefactor.

4 Convergence to equilibrium for Gaussian laws and applications

For Gaussian initial data, the optimized non-reversible dynamics yields exponential decay bounds for the mean, covariance, and density. The improved rate comes with a larger prefactor, while broader initial-data results require additional regularity or sufficiently long times.

4 Convergence to equilibrium for Gaussian laws and applications: The optimized perturbation gives a better exponential semigroup rate than the reversible dynamics, at the cost of worsening the prefactor.In the reversible case the prefactor is one, whereas the non-reversible estimates introduce a larger constant.
4 Convergence to equilibrium for Gaussian laws and applications: The mean of the Gaussian process converges exponentially fast to zero under a suitably chosen perturbation.The estimate is derived using the Q^-1 norm and the matrix identity associated with the optimal pair.
4 Convergence to equilibrium for Gaussian laws and applications: The prefactor can remain controlled with dimension while the perturbation norm grows linearly in N for a suitable eigenvalue choice.Choosing λk=N+k gives κ(Q)=2, while the norm of J̃ remains linear in N relative to S.
4 Convergence to equilibrium for Gaussian laws and applications: The covariance also admits an exponential convergence estimate for every t≥0 in the Euclidean matricial norm.This follows from the semigroup estimate for the drift and the corresponding transpose bound.
4 Convergence to equilibrium for Gaussian laws and applications: The Gaussian density converges to equilibrium through combined control of the mean and covariance.The result is obtained as a corollary of the separate mean and covariance estimates.
4 Convergence to equilibrium for Gaussian laws and applications: For general initial laws, the analysis extends Gaussian estimates by conditioning on the initial data, but some bounds require sufficient regularity or only hold after a positive time.Proposition 7 does not require the same L2 regularity as the uniform-in-time corollary, while data outside that space require t≥t0.

5 Convergence to equilibrium for initial data in L2(RN, ψ∞dx)

The section analyzes the non-self-adjoint semigroup associated with the perturbed linear diffusion, combining spectral information with resolvent and weighted-norm estimates to characterize convergence to equilibrium.

The generator is non-self-adjoint for nonzero perturbations, so spectral information must be supplemented by resolvent-norm estimates for exponential decay.The analysis uses a weighted L2-norm associated with the matrices Q and J.
The semigroup preserves the Hermite-degree subspaces, yielding a block-diagonal decomposition that reduces convergence to equilibrium to decay on the orthogonal complement of the ground-state space.
The spectral result gives the optimal asymptotic decay rate, but obtaining a uniform exponential bound with a controlled prefactor requires additional analysis and is not straightforward for general non-self-adjoint semigroups.
For optimal pairs (J~, Q), the semigroup is a contraction on the weighted space H1_Q.

6 Numerical Experiments

The numerical experiments apply the construction algorithm to low- and high-dimensional linear systems, showing substantially faster convergence for optimally perturbed dynamics and accelerated second-moment relaxation.

The experiments construct the perturbation by computing an orthonormal basis, selecting Q eigenvalues, forming J, and comparing the norms of e^−tB and e^−tBJ.
The optimally perturbed 3 × 3 system has a substantially larger spectral gap than the symmetric matrix, whose gap is 0.01.
For a 100 × 100 random diagonal matrix, the spectral gap increases from 0.0012 to 0.4762 under the optimal perturbation.
For the 100-dimensional discrete-Laplacian example, the improvement in convergence rate exceeds three orders of magnitude.
In the two-dimensional potential example, δ = 10 speeds convergence of the second moment but produces an oscillatory transient.

A Wick calculus

The appendix develops Wick calculus through creation and annihilation operators, Wick quantization, and positivity properties used to analyze the quadratic operators arising in the diffusion problem.

The canonical commutation relations provide the algebraic rules used to manipulate these operators and compose Wick-quantized expressions.
Wick quantization maps polynomials in complex variables to operators built from creation and annihilation operators on the Hermite-function decomposition.
The Wick quantization of a nonnegative polynomial is nonnegative on the polynomial domain, supplying the positivity property used in operator estimates.
For the diffusion analysis, the generator and auxiliary operator are Wick quantizations of quadratic polynomials, with the auxiliary operator globally elliptic, self-adjoint, and equipped with compact resolvent.