Reproduce “Optimal Scaling for Locally Balanced Proposals in Discrete Spaces” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haoran Sun, Hanjun Dai, Dale Schuurmans

Locally balanced proposals (LBP) tune the number of sites changed in each MCMC step using an acceptance-rate target. The authors analyze expected jump distance (EJD), the expected Hamming distance per Metropolis–Hastings iteration, and find an optimal LBP acceptance rate of 0.574 and scale O(N^(2/3)). Experiments show near-2/3 scaling across three model families; theory is proved for product Bernoulli targets with bounded marginals.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.