The Performance Of The Unadjusted Langevin Algorithm Without Smoothness Assumptions

Authors: Tim Johnston, Iosif Lytras, Nikolaos Makras, Sotirios Sabanis

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Examples and Numerical Experiments: 4.1.1 Sampling: We compare SGULA and MYULA on the task of sampling from a two dimensional mixture of Gaussians with a Laplacian prior, in order to assess their empirical behavior in a non-convex, non-smooth setting. ... Figure 1 compares the empirical densities obtained from pooled samples across all chains with the analytical ground truth densities. ... 4.4.1 Robust Regression: To compliment the theoretical analysis and illustrate the applicability of the Subgradient Unadjusted Langevin Algorithm (SGULA) to a practical optimization problem... Figure 2 displays the MRME boxplots for SCAD, LASSO, and the oracle.
Researcher Affiliation Academia Tim Johnston EMAIL Ceremade Université Paris Dauphine-PSL, France; Iosif Lytras EMAIL Archimedes Athena Research Center, Greece; Nikolaos Makras EMAIL School of Mathematics University of Edinburgh, United Kingdom; Sotirios Sabanis EMAIL School of Mathematics University of Edinburgh, United Kingdom School of Applied Mathematical and Physical Sciences National Technical University of Athens, Greece Archimedes Athena Research Center, Greece
Pseudocode Yes 3 Main Results: The Subgradient Unadjusted Langevin Algorithm (θλ n)n 0, is given by the Euler-Maruyama discretisation scheme of (1), in particular (SG-ULA): θλ n+1 = θλ n λh(θλ n) + p 2λβ 1ξn+1, θλ 0 = θ0, n N, (7)
Open Source Code No The paper does not contain an explicit statement about releasing source code, nor does it provide a link to a code repository.
Open Datasets No In this experiment we generate 100 datasets according to the following procedure. Let x Rd with Toeplitz covariance Σij = ρ|i j| and ρ = 0.5. For a fixed observations n = 60 and dimension d = 8, we sample X Rn d from the standard Gaussian. The response follows the model Y = XT β + ϵ, where β = (3, 1.5, 0, 0, 2, 0, 0, 0)T and the noise is drawn from a heavy tailed mixture, i.e. ϵ 0.9N(0, 1) + 0.1Cauchy(0, 1).
Dataset Splits Yes The tuning parameter γ > 0, is chosen independently for both objectives via 5-fold Cross-Validation.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes Both samplers were implemented with a fixed stepsize λ = 10 3 and inverse temperature parameter β = 1, and MYULA employed the same value for its smoothing parameter (γ = λ). For each method, we initialized 12 parallel chains from a broad uniform distribution on [min j µj 2 max j σ2 j , max j µj + 2 max j σ2 j ]2... Each chain was run for 52 103 iterations, discarding the first 12 103 as burn-in and retaining the rest for the assessment. ... The stepsize is fixed at λ = 10 3 and the tuning parameter γ > 0, is chosen independently for both objectives via 5-fold Cross-Validation. Each chain is run for 7.5 103 iterations, while each CV-fold is truncated at 1.25 103 iterations.