Improved Analysis for Hessian-free High-resolution Monte Carlo Sampling

Wujun Lv; Xiaoyu Wang; Yingli Wang; Lingjiong Zhu

Source-linked AI summary

Improved Analysis for Hessian-free High-resolution Monte Carlo Sampling

Wujun Lv, Xiaoyu Wang, Yingli Wang, Lingjiong Zhu

arXiv:2608.25052v1Aug 25, 2026stat.MLcs.LGmath.APmath.PR

TL;DR

Sampling methods need quantitative guarantees for potentially nonconvex machine-learning targets, including behavior at the ULD endpoint. The paper analyzes HFHR dynamics with an adapted hypocoercive framework and HFHRMC through a path-space Girsanov argument. It obtains explicit continuous- and discrete-time convergence results across α ≥ 0, with finite-accuracy benefits from positive position diffusion and improved iteration complexity over existing HFHR analyses.

Problem
The paper addresses convergence analysis for HFHR sampling when the potential need not be convex, including the α = 0 ULD endpoint and discrete-time guarantees.
Method
It combines an HFHR-adapted time-augmented Poincaré inequality with L2 energy arguments, then uses a path-space Girsanov analysis for HFHRMC.
Results
The paper proves an explicit L2 decay rate and non-asymptotic total-variation convergence with iteration complexity for every α ≥ 0 and γ > 0, improving existing HFHR bounds.
Takeaways & Limitations
Positive position diffusion can improve constants at finite accuracy, while the optimized method’s leading high-accuracy order agrees with the optimized ULD endpoint.
Takeaways & Limitations
The explicit rate depends on a Poincaré constant that can be exponentially small for genuinely nonconvex multi-well targets, so polynomial dimension dependence is not guaranteed.

Abstract

from arXiv · show

Hessian-free high-resolution (HFHR) dynamics augments underdamped Langevin dynamics (ULD) with reversible position diffusion for sampling problems that arise in machine learning. We establish an explicit quantitative contraction rate for HFHR dynamics under a position Poincaré inequality, weighted Hessian and Laplacian bounds, and a compact Sobolev embedding, where the potential function is not necessarily convex. An adapted time-augmented Poincaré inequality yields an explicit rate that improves upon the contraction rate of the underdamped Langevin dynamics. We also give a weak-solution construction and a self-contained spectral proof of the divergence lemma underlying the argument. For HFHR Monte Carlo (HFHRMC) algorithm, which is based on a discretization scheme of HFHR dynamics, we use a path-space Girsanov argument to obtain a non-asymptotic convergence bound and an explicit iteration complexity in total variation distance. The bounds hold for every $α\geq0$ and $γ>0$ and remain regular at the ULD endpoint. Optimizing the iteration complexity bound yields a positive, accuracy-dependent position-diffusion parameter at finite accuracy, while its leading high-accuracy order coincides with that of the optimized ULD endpoint. Our iteration complexity bound improves upon the existing work on HFHR algorithms. Numerical experiments including Bayesian learning problems on real data are provided to illustrate the effect of positive $α$ and its benefit.

1 Introduction

The paper develops HFHR dynamics and HFHRMC for sampling potentially nonconvex targets, extending convergence analysis beyond earlier parameter and convexity restrictions. It establishes explicit continuous- and discrete-time guarantees across α ≥ 0, including the ULD endpoint.

Sampling target distributions is central to Bayesian statistics, inverse problems, and large-scale machine learning.
HFHR augments ULD with reversible position diffusion while retaining a drift depending only on ∇U, making it Hessian-free.
The continuous-time analysis targets explicit L2 decay without strong convexity or a uniform upper bound on ∇2U.
HFHRMC achieves non-asymptotic total-variation convergence and explicit iteration complexity for every α ≥ 0 and γ > 0, with a Girsanov factor α + γ−1 finite at α = 0.
The HFHR rate combines the ULD contribution with an additive position-diffusion contribution, recovering the ULD rate at α = 0 and yielding HFHR scaling αm + m/γ in the convex high-friction regime.
Optimizing iteration complexity gives a positive accuracy-dependent α at finite accuracy, while its high-accuracy order matches optimized ULD and improves on existing HFHR bounds.

2 Preliminaries and Assumptions

The preliminaries define the kinetic framework, Sobolev notation, and assumptions used for the HFHR analysis. The assumptions combine position Poincaré control, weighted derivative bounds, and compactness, while allowing important nonconvex and non-globally-smooth settings.

Notation: The analysis uses probability measures μ, κ, and π = μ ⊗ κ, together with time-space measures and Gaussian velocity projections.
Notation: Hk(ν) denotes the usual weighted Sobolev space, with Bochner spaces formed over the time-position variables.
Operators: The kinetic graph operator is paired with a time-position operator, while the Hamiltonian operator is antisymmetric and the dissipative operators are self-adjoint and nonpositive.
Assumptions: The assumptions require a position Poincaré inequality, weighted derivative bounds for U ∈ C2(Rd), and compact embedding H1(μ) ↪ L2(μ).
Assumptions: Position Poincaré inequalities can follow from strong convexity, log-concavity, bounded perturbations, or suitable dissipativity conditions, including many nonconvex confining potentials.
Assumptions: Weighted derivative bounds are weaker than global gradient Lipschitzness: quartic confining potentials may satisfy them even when their gradients are not globally Lipschitz.
Assumptions: Compact embedding is a technical condition used in the spectral resolution of the position operator rather than a curvature condition.
Rate quantities: The framework introduces a curvature-contribution quantity used in the explicit rate and distinguishes weak convexity through a lower Hessian bound.

3 Continuous-Time Analysis

The continuous-time analysis establishes an explicit L2 decay rate for HFHR dynamics under position-Poincaré and weighted regularity assumptions, including nonconvex potentials and the ULD endpoint. It combines hypocoercive and direct position-diffusion mechanisms, yielding convergence consequences and parameter comparisons.

Analytic framework: The analysis combines direct position-diffusion coercivity with an HFHR-adapted time-augmented Poincaré inequality for the hypocoercive contribution.The position diffusion is incorporated into the kinetic graph operator and controlled using integration by parts against the position derivative of the divergence test.
Main rate: ν0,γ ≥ c mγ/(√m + R + γ)^2 at α = 0, retaining a positive hypocoercive contribution at the ULD endpoint.This endpoint rate is uniform in the position-diffusion parameterization and supports the ULD comparison.
Comparisons and scope: Unlike the strongly convex comparison, the rate applies over the full parameter space α ≥0, γ > 0, while the Poincaré constant may be exponentially small for nonconvex multi-well targets.In the high-friction regime, the result matches the γ + αm scaling up to universal constants and can improve optimization when m is small.
Convergence consequences: The continuous-time bound yields χ2, KL, and total-variation convergence for initial densities in L2(π), with an additional 2-Wasserstein consequence under a log-Sobolev inequality.The L2 result itself does not automatically identify a Wasserstein contraction rate.
Analytic framework: The semigroup framework constructs conservative Markov evolutions for α > 0 and α = 0 and provides weak forward solutions at the kinetic endpoint.The constructions use coercive forms and identify the associated forward equation.

4 Discrete-Time Analysis

The discrete-time analysis identifies HFHRMC with a frozen-force interpolation, establishes non-asymptotic TV and W2 guarantees through path-space Girsanov and entropy–moment control, and optimizes its iteration complexity. The resulting bounds remain regular at α=0, while positive α can improve finite-accuracy certificates without changing leading high-accuracy order.

Algorithm: HFHRMC uses a frozen-force interpolation whose grid-point chain is exactly the implementable algorithm, requiring one new evaluation of ∇U per iteration.The interpolation keeps the momentum in the position equation unfrozen and uses independent centered joint Gaussian increments.
Endpoint behavior: For α>0, the leading path-space KL error has Th scaling, whereas at α=0 the Th term vanishes and the bound improves to Th^2 in KL.The corresponding total-variation error scales as √(Th) for α>0 and as Th for α=0.
Convergence analysis: HFHRMC admits explicit non-asymptotic convergence bounds in total variation and Wasserstein distance, together with an explicit iteration-complexity estimate.The TV result is stated in Corollary 4.6, the iteration complexity in Corollary 4.7, and a corresponding W2 guarantee is also provided.
Parameter optimization: The optimized positive position-diffusion parameter improves the finite-accuracy certificate, but HFHR and optimized ULD share the same leading high-accuracy order.The optimizer is determined by the crossover of decreasing and increasing terms in the complexity proxy, while the stability term is lower order at the optimized scale.
Comparison: Compared with the common HFHRMC discretization analyzed previously, the present certified bound improves dependence on accuracy and dimension, and on m when m≤1.This is explicitly a comparison of proved upper bounds, not a claim of pointwise algorithmic dominance.

5 Numerical Experiments

The experiments evaluate HFHRMC on an anisotropic Gaussian target and three Bayesian learning problems, using exact Gaussian calculations or ensemble diagnostics. Across the reported real-data comparisons, positive position diffusion is associated with faster or lower-error convergence than KLMC.

Experimental setup: The experiments include an anisotropic Gaussian target, Bayesian linear regression, Bayesian logistic regression, and a Bayesian neural network.The Gaussian experiments use exact calculations, while the final two real-data experiments compare ensemble diagnostics with numerical posterior references.
5.1 Anisotropic Gaussian target: At t = 18, increasing α from 0 to 0.7 lowers exact χ2 divergence from 2.1391 × 10−4 to 4.3128 × 10−6 at γ = 0.5, and from 2.7566×10−6 to 2.4157 × 10−9 at γ = 1.The toy experiment has no time-discretization or Monte Carlo error.
5.2 Bayesian linear regression on real data: At ε = 10−4 and γ = 2A, HFHRMC requires 548652 iterations, compared with 1318015 for KLMC.The iteration count is based on the analytically computed χ2-divergence upper bound on joint total variation, with positive α selected separately by accuracy.
5.2 Bayesian linear regression on real data: At ε = 10−4 and γ = A, HFHRMC requires 1034914 iterations, compared with 4159467 for KLMC.The grid selects α = 0.03 at every displayed accuracy in this panel.
5.3 Bayesian logistic regression on real data: HFHRMC reaches sliced-W2 and posterior-mean-probability-RMSE thresholds with fewer gradient evaluations than both KLMC configurations.Each iteration uses one full data gradient evaluation; the threshold crossings are K = 53, 77, 164 for sliced W2 and K = 30, 53, 93 for probability RMSE.
5.4 Bayesian neural network with bounded weights and biases on real data: Over 260 iterations, HFHRMC has lower empirical areas under both reported error curves and lower late-window variability than KLMC for the Bayesian neural network.The sliced-W2 areas are 7.3413 versus 9.8377, and the probability-RMSE areas are 6.1325 versus 8.9156, respectively.

6 Conclusion

The paper establishes explicit continuous- and discrete-time convergence results for HFHR dynamics and HFHRMC across the full parameter range, including the ULD endpoint. Positive position diffusion improves finite-accuracy iteration complexity, while preserving the leading high-accuracy order of optimized ULD.

Continuous-time analysis: HFHR dynamics has an explicit L2(π) convergence rate under a position Poincaré inequality and weighted regularity assumptions, without requiring convexity.The rate combines hypocoercive ULD and direct position-diffusion contributions and is uniform at α = 0.
Discrete-time analysis: The HFHRMC analysis gives non-asymptotic convergence and iteration-complexity bounds using a path-space KL estimate without a separate time-uniform moment-stability assumption.The results hold for α ≥0 and γ > 0 and remain regular at the ULD endpoint.
Conclusion: Optimized positive position diffusion improves finite-accuracy complexity, while the leading high-accuracy order agrees with that of the separately optimized ULD endpoint.This allows parameter tuning throughout the stated parameter space.

A Weighted Elliptic Estimates

The appendix develops weighted elliptic estimates for the time–position operators used in the divergence construction. It establishes coercivity, operator-domain regularity, and mixed elliptic regularity under the paper’s weighted assumptions.

Coercivity: Tensoring the position Poincaré inequality with the time Neumann Poincaré inequality provides coercivity for the mixed time–position operator.The construction uses the product measure on the time interval and position space.
Weighted multiplier control: A weighted multiplier estimate controls multiplication by the unbounded coefficient ∇U for H1 functions.The estimate follows from the paper’s weighted regularity assumption, integration by parts, and Young’s inequality.
Position operator: The position operator Aq is identified as the Friedrichs operator, with compactly supported smooth functions forming an operator core.The proof also establishes the relevant H2 domain regularity and the L2 realization of Aq.
Mixed elliptic regularity: For zero-mean h, the mixed Neumann problem has a unique mean-zero weak solution with the prescribed zero time-boundary derivatives.The solution satisfies the stated H1 and H2 regularity estimates.

B Spectral Construction of the Divergence Test

The appendix constructs the divergence test spectrally by decomposing the datum into harmonic and orthogonal components. Modewise estimates, compactness, and Sobolev control justify the infinite spectral assembly and its endpoint traces.

Spectral decomposition: Compactness of the position resolvent yields an orthonormal eigenbasis for L2(µ), enabling modewise construction of the divergence test.The basis separates the constant position mode from nonconstant modes.
Component decomposition: The datum is decomposed into harmonic and orthogonal components, and the orthogonal component is handled through the mean-zero mixed elliptic problem.Integration by parts with harmonic tests determines the endpoint behavior of the solution.
Modewise estimates: For each nonconstant spectral mode, exponential solutions provide explicit control of the Gram matrix and its smallest eigenvalue.These bounds are used to control the mode coefficients across the full time interval.
Assembly and regularity: Summing the modewise estimates gives the required Sobolev bounds, while spectral-tail control justifies passage from finite sums to the full divergence test.Continuity of trace maps preserves the zero endpoint traces in the limit.

C Weighted Kinetic Density, Traces, and Green’s Formula

This appendix establishes density, kinetic traces, and Green’s formula for weak kinetic solutions in weighted whole-space settings, including unbounded forces. The results also yield a kinetic energy chain rule for endpoint semigroup construction.

Weighted Kinetic Density: The graph-density lemma approximates kinetic-graph functions by compactly supported smooth functions despite unbounded Gaussian velocity multipliers and whole-space cutoffs.Cutoffs in position and velocity, followed by local smoothing, converge in the kinetic graph norm.
Weighted Kinetic Density: Gaussian integration by parts and multiplier bounds control multiplication by each velocity coordinate in the weighted kinetic spaces.The argument specifically addresses the unbounded Gaussian velocity multiplier.
Traces and Green’s Formula: The density result transfers classical smooth integration by parts to weak kinetic solutions, producing strong interior time traces and Green’s formula.The Hamiltonian vector field is divergence-free and preserves the Gibbs weight, supporting the weighted identity.
Traces and Green’s Formula: The trace representatives are continuous in time, with strong convergence at an endpoint under the stated endpoint hypothesis.Green’s formula consequently extends to the endpoint, including finite left or right endpoints.
Kinetic Chain Rule: Applying Green’s formula to identical arguments yields the kinetic energy chain rule used to construct the endpoint semigroup.The endpoint version follows when the endpoint hypothesis of the trace proposition holds.