Live demo: https://chris-nemeth.github.io/sampling-gallery/

An interactive gallery of Markov-chain Monte Carlo (and related) sampling algorithms — animated, tunable, and explained in plain language. Static, buildless HTML: open it in a browser, or serve the folder.

Pages

• index.html — landing atlas: all algorithms grouped by family, with animated thumbnails and plain-language descriptions. Links into the sampler.
• notes.html — field notes: a companion note for every sampler (what it does, what to watch while it runs, which knobs matter, references), written as plain markdown in notes/*.md and also available in-app via the "Notes" button in the sampler.
• sampler.html — the interactive app, driven by the project's real MCMC engine (lib/, main/, algorithms/): every sampler in the rail runs its genuine algorithm. Nine target distributions (banana / donut / Gaussian / mixture / funnel / squiggle / flower / swiss roll / correlated ridge), per-sampler tunable parameters, live diagnostics (steps / acceptance / mean / ESS), a 2D view with proposal arrows, gradients and trajectories rendered live, and a rotating 3D density-surface view with the chain and step geometry lifted onto it.
• classic.html — the original interface (app.html), kept for continuity: a full-window 2D visualizer with a lil-gui control panel.
• styles.css — Broadsheet design tokens + component classes.
• notes/*.md — the field notes source; docs/ — thumbnails and social card, captured from the live engine.

Samplers

Rejection sampling, importance sampling (with SIR and PSIS), quasi-Monte Carlo, slice sampling, elliptical slice sampling, Random Walk MH, Adaptive MH, HMC (optional dual-averaging step-size adaptation), NUTS (efficient dual-averaged by default; naive tree and fixed step size as advanced options), Riemannian-manifold HMC (SoftAbs metric), the apogee-to-apogee path sampler (AAPS), MALA, ULA, SGLD (optional control variates), Barker Proposal, tuning-free ULA (FUSE), H2MC, Gibbs, ensemble MCMC (stretch and differential-evolution moves), parallel tempering, tempered SMC (with an annealed importance sampling mode), Zig-Zag, Bouncy Particle Sampler, stochastic-gradient PDMPs (SG Zig-Zag and SG-BPS), SVGD, Coin SVGD (learning-rate free), Wasserstein particle descent (Blob, GFSD and GFSF modes), SPOS, and nested sampling. The particle methods share a live kernel Stein discrepancy diagnostic, an attraction/repulsion force decomposition, and selectable particle initializations.

Run locally

Open index.html in any modern browser, or from the folder root:

python3 -m http.server

then visit http://localhost:8000/

Notes

• No build step; all libraries are vendored in lib/ (fonts load from Google Fonts when online; the pages remain functional offline).
• Every sampler runs the real, tested implementation from algorithms/ — there are no placeholder fall-backs.
• Redesigned in the Broadsheet system; based on the original demo by Chi Feng: https://github.com/chi-feng/mcmc-demo

Using this in teaching

Please do — that is what it is for. Link to the live site or to a specific sampler/target (e.g. sampler.html?algorithm=RiemannianHMC&target=funnel); the field notes at notes.html are written as companion reading. If you use the gallery in a course or talk, a link back here is appreciated.

Acknowledgements & licensing

Built from Chi Feng's wonderful mcmc-demo (MIT). Nested sampling adapts Johannes Buchner's ultranest-js (algorithms/NSRadFriends.js, AGPL-3.0 — the one non-MIT file; see LICENSE). Markdown rendering by marked (MIT). Algorithm sources are cited in the gallery's reference list and in each field note.

Community contributions: the Barker proposal sampler — implementation, visualization and field note — was contributed by Rui-Yang Zhang. Further contributions are very welcome; see the field notes for the house conventions.