Programming

Walking the proximal sampler around a corner: implementing Liu & Chewi's composite log-concave RGO algorithm

A two-week-old arXiv preprint promises a sampler that handles `π ∝ exp(-f-g)` with a non-smooth `g` in `Õ(κ√d log⁴(1/ε))` gradient calls, beating older `Õ(d)` methods. I implement it from the algorithm boxes, run it against two ground-truth problems, watch the cost-vs-dimension slope land at 0.38 (within shouting distance of the predicted 0.5), and surface a small sign typo in Appendix C along the way.