Convergence Rates for Non-Log-Concave Sampling and Log-Partition Estimation

Authors: David Holzmüller, Francis Bach

JMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental study nonetheless confirms practical differences between the convergence rates of some of the investigated efficient algorithms, although it is limited to a toy problem and simple algorithms.
Researcher Affiliation Academia David Holzmüller david dot holzmuller at inria.fr Francis Bach EMAIL INRIA Ecole Normale Supérieure PSL Research University
Pseudocode Yes Algorithm 1 Rejection sampling with proposal distribution Pg limited to n function evaluations. ... Algorithm 2 Bisection sampling algorithm using a log-partition algorithm L.
Open Source Code Yes Our plots can be reproduced using the code at github.com/dholzmueller/sampling_experiments
Open Datasets No To further investigate the convergence behavior of some simple algorithms, we study them numerically on functions of the form f : [0, 1]3 R, x 7 β(x1 + x2 + x3). While these functions are simple (and concave), they pose a challenge to some general algorithms as they have a large range in relation to their Lipschitz constant.
Dataset Splits No The paper uses a synthetic function for its experiments (functions of the form f : [0, 1]3 R, x 7 β(x1 + x2 + x3)) and does not mention any dataset splits for this synthetic data, nor for any external dataset.
Hardware Specification No No specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments are mentioned in the paper.
Software Dependencies No No specific ancillary software details (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4) are mentioned in the paper.
Experiment Setup Yes Figure 2: Convergence of the (median) error |Lf Lf| for different values of β {0.1, 40, 10000}. For the stochastic methods MC and PC+MC, the median is taken over 10001 independent runs. ... Figure 3: Convergence of different sampling methods in terms of the empirical energy distance, computed using N = 106 samples for each distribution, to the true distribution Pf for β = 15.