Log-concave sampling: Metropolis-Hastings algorithms are fast

Authors: Raaz Dwivedi, Yuansi Chen, Martin J. Wainwright, Bin Yu

JMLR 2019 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We study the problem of sampling from a strongly log-concave density supported on Rd, and prove a non-asymptotic upper bound on the mixing time of the Metropolis-adjusted Langevin algorithm (MALA)... We provide numerical examples that support our theoretical findings, and demonstrate the benefits of Metropolis-Hastings adjustment for Langevin-type sampling algorithms.
Researcher Affiliation Academia Raaz Dwivedi , EMAIL Yuansi Chen , EMAIL Martin J. Wainwright , , EMAIL Bin Yu , EMAIL Department of Statistics Department of Electrical Engineering and Computer Sciences University of California, Berkeley Voleon Group , Berkeley
Pseudocode Yes Algorithm 1: Metropolis adjusted Langevin algorithm (MALA)
Open Source Code No The paper does not contain any explicit statements about releasing code or links to source code repositories.
Open Datasets No The paper uses simulated data or describes methods for generating data based on theoretical models (e.g., "sampling a multivariate Gaussian", "sampling a Gaussian mixture", "Bayesian logistic regression" where data is "randomly draw[n] i.i.d. samples"). It does not provide concrete access information to any publicly available or open datasets.
Dataset Splits No The experiments involve simulating Markov chains for a number of samples or generating i.i.d. data points for logistic regression. No training/test/validation splits are mentioned for pre-existing datasets, as the datasets are either custom-generated or theoretical distributions.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments.
Software Dependencies No The paper does not provide specific version numbers for any software components or libraries used.
Experiment Setup Yes The step-sizes are chosen according to Table 3. For ULA, the error-tolerance δ is chosen to be 0.2. The step-size is error-dependent, precisely chosen to be 10 times of δ. For weakly log-concave density, m is chosen to be δ/(d L). The initial distribution is chosen as µ = N(0, L-1Id) and the step-sizes are chosen according to Table 1, where for ULA, we set three different choices of δ = 0.2 (ULA), δ = 0.1 (small-step ULA) and δ = 1.0 (large-step ULA).