Long-Run Accuracy of Variational Integrators in the Stochastic Context

1 Introduction

• GLA combines a variational integrator for the Hamiltonian flow with the exact Ornstein–Uhlenbeck flow in a Lie–Trotter splitting of inertial Langevin equations.The paper uses “Geometric Langevin Algorithm” for this generalization of variational integrators to inertial Langevin equations.Source: §1 · p. 2
• Long-time statistical accuracy of this splitting had few quantified results in the literature, despite its use in molecular dynamics and related settings,,.The paper focuses on molecular dynamics, where inertial Langevin integrators have often generalized the explicit, symmetric, second-order Störmer–Verlet method.Source: §1 · p. 2
• GLA is conformally symplectic, preserving the precise symplectic area change associated with inertial Langevin flow.This property goes beyond being quasi-symplectic as defined in RL1 and RL2 of.Source: §1 · p. 2
• Under the paper’s hypotheses, its main results establish finite-time pathwise convergence, geometric ergodicity, and long-run statistical accuracy governed by the variational integrator’s Hamiltonian-representation error.The paper states that if GLA is geometrically ergodic with a unique invariant measure, its invariant-measure sampling error is determined by that energy error.Source: §1 · p. 3
• For nonglobally Lipschitz drift with an explicit variational integrator, GLA may fail to be geometrically ergodic, so the paper’s results may not hold.The paper suggests Metropolizing GLA as a proposal to stochastically stabilize it; inertial Langevin requires a momentum flip for reversibility [4], with momentum flips studied in.Source: §1 · p. 3
• The analysis is set on flat space, although the authors state that GLA and its properties generalize to manifolds.They refer readers to Remark 2.1 and [2] for details.Source: §1 · p. 3

2 Main Results of Paper

• GLA applies a pth-order variational integrator to approximate Hamiltonian flow and composes it with the Ornstein-Uhlenbeck dynamics in a Lie-Trotter splitting.The Ornstein-Uhlenbeck component is defined by the nonconservative part of inertial Langevin.Source: §2 · pp. 5–6
• GLA extends to inertial Langevin equations on manifolds, including systems with holonomic constraints.For a compact configuration manifold and smooth potential energy, geometric ergodicity is typically satisfied for sufficiently small time steps.Source: §2 · p. 6
• Theorem 2.1 establishes finite-time pathwise convergence of GLA, consistent with first-order Lie-Trotter splitting for additive noise.The theorem quantifies trajectory accuracy over any fixed finite time interval.Source: §2 · pp. 6–7
• Under the stated assumptions, GLA is geometrically ergodic with respect to a discrete invariant measure for sufficiently small step sizes.The theorem applies to test functions bounded by a constant times the specified Lyapunov function; the assumptions are sufficient but not necessary.Source: §2 · p. 7
• The discrete invariant measure is O(h^p) away in total variation from the Boltzmann-Gibbs measure when the underlying variational integrator has global accuracy p.The result is established under the stated assumptions for sufficiently small step sizes.Source: §2 · p. 7
• The proof uses forward error analysis based on the existence of the discrete invariant measure and convergence from nonequilibrium states, rather than characterizing that measure directly.The paper notes that a backward error analysis would be substantially more involved.Source: §2 · pp. 7–8
• The equilibrium accuracy order can be tuned through the variational integrator, while each GLA step uses one n-dimensional stochastic integral rather than the multiple required by higher-order SDE discretizations.The paper contrasts this with the computational intensity of higher-order SDE discretizations.Source: §2 · p. 8

3 Preliminaries

• The potential conditions ensure almost-sure existence and pathwise uniqueness of the SDE solution (U1) and integrability of e^(-βH), making the Boltzmann-Gibbs measure a probability measure (U2).Source: §3 · pp. 8–9
• The variational integrator assumptions require a discrete Hamiltonian map from a hyperregular discrete Lagrangian and local (p + 1)th-order accuracy.The discrete-Lagrangian condition implies the map is symplectic and Lebesgue-measure preserving.Source: §3 · p. 9
• The structural assumptions include a Lyapunov function with quadratic growth bounds and a drift condition.Source: §3 · p. 9

4.1 Pathwise Convergence

• GLA is first-order mean-squared convergent to solutions on finite time intervals; its Lie-Trotter splitting and additive noise are noted in explaining this result.Source: §4.1 · pp. 10–11

4.2 Geometric Ergodicity

• Under the stated assumptions, GLA is geometrically ergodic with a discrete invariant measure, proved using an inherited Lyapunov function and a minorization condition.The minorization condition is established for GLA sampled every other step.Source: §4.2 · p. 12

4.3 Long-Run Accuracy

• If the variational integrator has no energy error, GLA preserves the invariant measure; the exact splitting is one such case.Source: §4.3 · p. 15
• Under the stated assumptions, GLA’s long-run equilibrium error reflects both its convergence to equilibrium and the local accuracy of its variational integrator in representing the Hamiltonian.The long-run accuracy analysis also requires volume preservation, exact preservation of the Boltzmann-Gibbs measure by the Ornstein-Uhlenbeck map, and geometric ergodicity.Source: §4.3 · pp. 16, 18

Validation

• The tested GLA variants compose the Ornstein-Uhlenbeck flow with explicit first-order symplectic Euler, second-order Störmer-Verlet, or fourth-order symplectic methods.The fourth-order method is included for validation.Source: §5 · pp. 19–20
• Although (36) and (37) are only first-order pathwise convergent, they approximate the specified ensemble averages to second- and fourth-order accuracy, respectively.Source: §5 · p. 20
• For the linear oscillator, the exact Hamiltonian splitting makes no error in the stationary correlation matrix because its Hamiltonian flow preserves volume and energy.Source: §5 · p. 22
• For the tested nonglobally Lipschitz nonlinear oscillator, theory does not apply, though the tested explicit schemes’ orbits appeared confined to a linearly stable compact region during the experiments.The authors describe explicit GLA as potentially transiently stochastically unstable for such forces.Source: §5 · p. 22
• Halving the time step decreases |µh(q2) − µ(q2)| linearly for (35), nearly quadratically for (36), and nearly quartically for (37).The table uses empirical time averages; subsequent rows halve the step size and double the number of steps while keeping the integration interval fixed.Source: Table 1 · p. 23

• The analysis is presented as a first step toward a deeper study of GLA for molecular systems.Source: §6 · p. 23
• Uniform Lipschitz continuity of the Hamiltonian vector field and coercivity of the Hamiltonian are assumed; these conditions suffice for GLA’s geometric ergodicity when the solution process is geometrically ergodic.The paper identifies uniform Lipschitz continuity as important for stochastic stability.Source: §6 · p. 24
• With nonglobally Lipschitz drift, an explicit variational integrator may fail to be geometrically ergodic; a non-globally-linearly-stable integrator can make GLA transient.The paper notes that MAGLA can instead use GLA as a proposal step within a Metropolis-Hastings algorithm.Source: §6 · p. 24
• The proof estimate uses geometric ergodicity, measure preservation by the variational integrator, preservation of µ by the Ornstein-Uhlenbeck flow, and (p + 1)th-order local energy error.These are described as ingredients of the estimate, while the paper stresses that its result holds under more general conditions.Source: §6 · p. 24
• If GLA is geometrically ergodic with a unique invariant measure, its invariant-measure sampling error is determined by the variational integrator’s energy error.Source: §6 · p. 24

7.1 Single-Step Error

• For sufficiently small h, the exact splitting has a bounded second moment, obtained using the uniform Lipschitz condition and Gronwall’s lemma.Source: §7.1 · p. 26
• The leading expected difference is O(h2), while the leading mean-squared difference is O(h3/2); bounding these terms uses second-moment bounds and Hessian control.Smoothness of U and the uniform Lipschitz condition on the potential force provide the Hessian control.Source: §7.1 · p. 27

7.2 Variational Integrators

• A variational integrator is defined by a discrete Lagrangian approximating the exact discrete Lagrangian, whose trajectory solves the Euler-Lagrange equations with fixed endpoints.Source: §7.2 · pp. 27–28
• In Hamiltonian form, the discrete Lagrangian defines an update that preserves the symplectic form and Lebesgue measure and can approximate Hamilton’s flow.Hyperregularity is expressed by a nonzero determinant of the mixed derivative of the discrete Lagrangian.Source: §7.2 · p. 28