Spectral Ergodic Control with Provable Regret Guarantees
Unreviewed draft. This paper has not been human-reviewed.
Mathematical claims may be unverified. Use with appropriate caution.
Abstract
Ergodic control and online regret analysis are usually developed with different tools and assumptions. We present a spectral ergodic control framework where long-run average control objectives and finite-time regret can be analyzed jointly. Our main results provide policy classes with explicit stability and regret bounds derived from mode-wise contraction and uncertainty propagation.
Length
1,775 words
Claims
3 theorems
Status
Draft
Novelty
Unifying ergodic average-cost control and finite-time regret analysis through a shared spectral (OU mode) decomposition, so both guarantees fall out of the same representation.