Reproduce “Involutive MCMC: a Unifying Framework” 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
Kirill Neklyudov, Max Welling, Evgenii Egorov, Dmitry Vetrov
iMCMC provides a framework for constructing MCMC samplers from auxiliary-variable draws and involutions. The authors compose reversible steps to retain direction after acceptance and reverse it after rejection, preserving target invariance without guaranteeing ergodicity. On a two-dimensional Gaussian mixture, Irr-NICE-MC reports higher ESS and ESS per second than A-NICE-MC using the same learned proposal; results are not uniformly better.
Key insights
- Involutions provide a common structural explanation for algorithms as diverse as Metropolis-Hastings, leapfrog-based HMC, reversible-jump MCMC, and flow-based samplers.
- The basic iMCMC update is reversible, whereas irreversibility arises from ordered kernel compositions or from retaining auxiliary direction variables across iterations; stationarity does not require detailed balance.
- Persistent direction can help traverse separated modes, but it can also lower acceptance and sampling efficiency on unimodal targets where moving through low-density regions is less useful.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.