Penalized Langevin dynamics with vanishing penalty for smooth and log-concave targets

Authors: Avetik Karagulyan, Arnak Dalalyan

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical We study the problem of sampling from a probability distribution on Rp defined via a convex and smooth potential function. We first consider a continuous-time diffusion-type process, termed Penalized Langevin dynamics (PLD), the drift of which is the negative gradient of the potential plus a linear penalty that vanishes when time goes to infinity. An upper bound on the Wasserstein-2 distance between the distribution of the PLD at time t and the target is established. This upper bound highlights the influence of the speed of decay of the penalty on the accuracy of approximation. As a consequence, considering the low-temperature limit we infer a new nonasymptotic guarantee of convergence of the penalized gradient flow for the optimization problem. ... The main result of this work is an upper bound on W2(νPLD t , π) that is valid for every continuously differentiable and decreasing penalty factor α. Optimizing over α, we show that the choice α(t) 1/(2t), when t , leads to a simple upper bound of the order 1/√t. ... This paper is purely theoretical, thus we expect no direct or indirect ethical risks from it.
Researcher Affiliation Academia Avetik Karagulyan CREST, ENSAE, IP Paris avetik.karagulyan@ensae.fr Arnak S. Dalalyan CREST, ENSAE, IP Paris arnak.dalalyan@ensae.fr
Pseudocode No The paper does not contain any pseudocode or algorithm blocks. The processes are described using stochastic differential equations.
Open Source Code No The paper does not provide any information about open-source code for the described methodology.
Open Datasets No The paper is theoretical and does not use or refer to any specific publicly available datasets for training or evaluation. It discusses probability distributions in general terms.
Dataset Splits No The paper is theoretical and does not involve empirical experiments with dataset splits. Therefore, it does not provide train/validation/test split information.
Hardware Specification No The paper does not mention any specific hardware used for computations or simulations.
Software Dependencies No The paper does not provide specific software names with version numbers that would be necessary to replicate any computational aspects.
Experiment Setup No The paper is theoretical and does not include details on an experimental setup, such as hyperparameters or system-level training settings. The figures presented are illustrative of theoretical properties.