Catching a Moving Subspace: Low-Rank Bandits Beyond Stationarity

In three linesTheoretical work on piecewise-stationary low-rank linear contextual bandits with drifting subspaces. Introduces SPSC algorithm combining isotropic probes with windowed projected ridge-UCB, achieving dynamic regret Õ(r√T) instead of Õ(d√T). Characterizes identification boundary for moving subspace recovery and validates on 11 benchmarks (synthetic, MovieLens, clinical, ZOZOTOWN production logs).

Summary generated by Claude — human-verified