Reversible Stratonovich perturbations for accelerating convergence to equilibrium of Langevin dynamics

1 Introduction

• Small spectral gaps can require long sampling intervals, while multiscale potentials can impose severe time-step restrictions on standard methods such as ULA.The convergence rate of overdamped Langevin dynamics is governed by the generator’s spectral gap.Source: §1 · p. 2
• Existing approaches have trade-offs: nonreversible perturbations can worsen transient convergence, while optimal preconditioners can be difficult to compute and expensive to simulate.A large constant in the nonreversible convergence bound can worsen the transient phase; preconditioned dynamics may cost more than overdamped Langevin or nonreversible dynamics.Source: §1 · p. 3

Our contributions

• For quadratic potentials, the paper constructs a Stratonovich perturbation that improves the spectral gap and proves an optimal perturbation can be built algorithmically.The theoretical analysis assumes a symmetric positive definite matrix S; in two dimensions, the induced rate is bounded below by γ2.Source: Our contributions · p. 4
• Numerical experiments report superior performance over other Langevin-based samplers and suggest acceleration also holds for non-Gaussian targets.The paper compares the proposed sampler with existing Langevin-based methods.Source: Our contributions · p. 4

2.1 Nonreversible perturbation to improve the spectral gap

• The nonreversible construction [22] pairs small and large modes to build an optimal perturbation, but its eigenvalue choice is unclear and repeated Gram–Schmidt is costly.The construction equilibrates the action of S on an orthonormal basis.Source: §2.1 · p. 6
• Nonreversible perturbations can improve convergence but also induce oscillatory and transient behavior.Complex eigenvalues can produce oscillations.Source: §2.1 · p. 5, §2.2 · p. 7

2.2 Reversible perturbation to improve the spectral gap

• A one-dimensional Stratonovich noise replaces direct drift perturbation; when g is divergence-free under π, the resulting dynamics are reversible.For g = J∇V, the generator is self-adjoint with real spectrum, so convergence remains non-oscillatory.Source: §2.2 · p. 7
• The earlier reversible result [1] showed that this perturbation form does not worsen convergence, but did not quantify improvement or identify an optimal rate or matrix J.Source: §2.2 · p. 7

3.1 Optimal scaling in dimension two

• For λ > 0, Theorem 2 states that the perturbed stochastic differential equation (17), with δ = λ−1/2, gives perturbed operator R spectral gap at least γ.The optimal scaling is the smallest α > 0 such that the convergence rate of (10) with skew-symmetric matrix δJ2, where δ = λ−α, remains bounded away from zero as λ → 0+.Source: §3 · p. 7, §3.1 · p. 8

3.2 Pairwise matching in a general dimension

• For even-dimensional S, pairwise matching targets γ_min = min_i(s_i + s_{d+1−i}), with the bound attained asymptotically.The construction assumes positive ordered eigenvalues and the theorem’s stated spectral conditions.Source: §3.2 · p. 8
• Pairwise matching can be weak when S has a few large and many small eigenvalues; for diag(1, λ, …, λ), the gap changes only from λ to 2λ.When eigenvalues are approximately symmetric around their mean, γ_min/2 is approximately the nonreversible optimal rate Tr(S)/d.Source: §3.2 · p. 9
• The general-dimensional theorem permits scaling O(s_d^-1), unlike the two-dimensional O(s_d^-1/2); experiments report significant improvement with δ̂ = 1.Source: §3.2 · p. 9
• Algorithm 1 diagonalizes the positive-definite input A, builds the perturbation in its eigenbasis, then transforms it back to the original coordinates.The resulting skew-symmetric matrix is J(A).Source: §3.2 · p. 9
• For nonquadratic potentials, experiments report improved sampler performance with a Stratonovich perturbation built from J(A), while theoretical justification is left to future work.The section also proposes constructing a perturbation from the Fisher information matrix F.Source: §3.2 · p. 9

3.3 Splitting eigenvalues in a general dimension

• For spectra with a few large and many small eigenvalues, the alternative pairs small eigenvalues with the largest using multiple Stratonovich noises.The authors interpret this construction as splitting the largest eigenvalue’s strength across the smaller ones; its second moment converges at a stated rate.Source: §3.3 · p. 10

4 Proof in dimension two: moment estimates

• The moment proof is needed because the perturbed dynamics is non-Gaussian even from a Gaussian initial condition, so controlling only the first two moments is insufficient.The original dynamics with J = 0 remains Gaussian under the stated condition.Source: §4 · p. 10

4.1 Convergence of the expectation

• For the expectation, the coordinate rates are 1 + δ^2λ and λ + δ^2λ; choosing δ = λ^-1/2 makes the smallest eigenvalue of Bδ equal to 1 + λ.Increasing δ can make the expectation’s convergence speed arbitrarily large.Source: §4.1 · p. 11

4.2 Convergence of the covariance

• For δ = λ^-1/2, µ− remains bounded away from zero as λ → 0, establishing this choice’s stated eigenvalue behavior without asserting that it is a critical scaling.Source: §4.2 · p. 12

4.3 Convergence of higher moments

• For δ = λ^-1/2, the even-moment convergence rate stays bounded away from zero as λ → 0 for every order, and the second moment is the slowest.Moments containing odd powers converge at a rate at least 1 + λ.Source: §4.3 · pp. 13, 15, 16

4.4 Extension to general observables

• The polynomial convergence bound extends to general L2(π) observables for the reversible, self-adjoint, negative-semidefinite generator R.The argument targets a spectral-gap bound γ for R.Source: §4.4 · pp. 17–18
• Continuity of e^{tR} and density of polynomials in L2(π) extend the bound to general observables, and Lemma 5 yields λ ≥ γ, completing Theorem 2.Source: §4.4 · p. 18

5 Proof in general dimension: Rayleigh estimation

• Estimating the Rayleigh quotient separately on A_k = ker(A) ∩ H_k and its orthogonal complement B_k yields a lower bound for the reversible generator's spectral gap.Source: §5 · p. 19

5.1 Reduction to finite dimensional space

• The Wiener chaoses H_k provide finite-dimensional blocks because the generator A preserves each H_k, enabling blockwise Rayleigh estimates before density extends the result to the whole space.Source: §5 · p. 19, §5.1 · p. 20

5.2 Characterization of the kernel Ak

• Kernel elements are characterized as polynomials in Re(z_i conjugate z_j) and Im(z_i conjugate z_j), and the odd-degree kernels vanish.Source: §5.2 · pp. 21–22, §5.3 · p. 22

5.3 Spectral gap on Ak

• For even-order A_2k, the Rayleigh quotient is bounded below by γ_min = min_i(s_i + s_{d+1-i}), with equality at k = 1; the bound extends to the whole kernel of A.Source: §5.3 · pp. 22–24

5.4 Spectral gap on Bk

• When B is diagonalizable, Proposition 8 characterizes the polynomial-space spectrum of T from B's eigenvalues; JS is diagonalizable over C with purely imaginary eigenvalues.Source: §5.4 · pp. 25–27
• On Bk, T has eigenvalues i(k − 2r) for r = 0,…,k, and the resulting Rayleigh quotient is bounded below by 1 + k s_d.The proof excludes a contribution from a zero-eigenvalue eigenvector on Bk.Source: §5.4 · pp. 27–28

5.5 Spectral gap on Pk

• Choosing δ = O(s_d^-1) yields a non-asymptotic lower bound on the Rayleigh quotient, with a finite-exception qualification across chaos orders.The proof gives the bound using a decomposition into ker(A) ∩ Hk and its orthogonal complement.Source: §5.5 · p. 29
• Odd moments converge at rate at least 1 + s_d; for δ = 1, controlling cross terms yields a γ_min/9 lower bound on the second Wiener chaos but a negative bound for large k.The authors suggest the true spectral gap lies in the second Wiener chaos and believe δ = 1 gives significant improvement, consistent with their numerical experiments.Source: §5.5 · p. 29

6.1 Two dimensional experiments

• With δ = λ^-1/4, the Stratonovich perturbation’s convergence rate collapses to the original process for small λ, unlike the optimal-scaling comparison.Figure 2 compares second-moment convergence for δ = λ^-1/2 and δ = λ^-1/4.Source: §6.1 · p. 32, Figure 2 · p. 33
• For a two-mode Gaussian mixture, the reversible perturbation appears to help overcome the energetic barrier.The experiment compares process-based approximations of the invariant mixture.Source: §6.1 · p. 33, Figure 3 · p. 34

6.2 Multidimensional Gaussian experiments

• In the d = 100 discrete-Laplacian experiment, the random perturbation is unstable and does not significantly improve convergence.The comparison uses the Algorithm 1 construction, a random skew-symmetric matrix, and a nonreversible perturbation.Source: §6.2 · p. 34, Figure 4 · p. 35
• In a pathological d = 4 Gaussian example, the Section 3.2 construction may be ineffective, while the Section 3.3 construction maintains satisfactory convergence.Figure 5 plots error over time for this example.Source: §6.2 · p. 35, Figure 5 · p. 35

6.3 Multidimensional non-Gaussian experiments

• Under a fixed computational budget, the Stratonovich and nonreversible perturbations offer a significant computational advantage in the non-Gaussian setting.Figure 7 compares relative MSE across step sizes; the perturbed methods use 16 force evaluations per step, versus 8 for the original and preconditioned processes.Source: §6.3 · pp. 36–37, Figure 7 · p. 37
• For reaching a fixed error, the Stratonovich perturbation behaves similarly to the nonreversible one and leads the other process.Figure 8 reports the number of steps required to reach that error in dimension d = 20.Source: §6.3 · p. 37, Figure 8 · p. 37

7 Conclusions

• For Gaussian targets, an optimal Stratonovich perturbation maximizing the spectral gap exists and can be constructed algorithmically.Numerical experiments also demonstrate superior performance relative to other Langevin-based sampling schemes.Source: §7 · p. 38
• Rigorous convergence analysis of the discretized dynamics, high-dimensional implementation, and rigorous analysis for non-Gaussian targets remain open.The paper notes that the non-Gaussian analysis would require tools different from those used in this article.Source: §7 · p. 38

• For n ≥ 11, the minimum is attained at a boundary, k = 1 or k = n.Source: A Complementary results for the 2… · p. 41
• For 2 ≤ n ≤ 8, the polynomial decreases on [1, n], so its minimum is at k = n; this endpoint evaluation is smallest at n = 2.Source: A Complementary results for the 2… · p. 43
• For n = 9, 10, and 11, the vertex lies inside [1, n], and its evaluated values are 23/11, 133/55, and 147/55, respectively.Source: A Complementary results for the 2… · p. 43
• The global minimum of the lower bound is reported at n = 2, k = 2, with value 9/11.Source: A Complementary results for the 2… · p. 43