Adaptive Bayes exactly tracks information over intrinsic time
Provides a unified, exact framework for understanding regret in sequential decision-making, offering new insights into adaptive behavior across a wide range of algorithms and settings.
The paper derives an exact information-accounting identity for Bayesian and multiplicative-weights updates, showing that regret decomposes into a payment for uncertainty and a reduction in information distance. This yields exact adaptive decompositions that capture favorable stochastic regimes as self-bounding properties of intrinsic time, covering Hedge, online convex optimization, contextual bandits, and repeated games.
Bayesian and multiplicative-weights updates reweight experts, models, or actions from sequential feedback. We show that the regret of any such update obeys an exact information-accounting identity. On each round, the learner's excess loss to any chosen comparator is the sum of an immediate payment for the uncertainty exposed by the round and a reduction in the information distance from the learner's current weights to the comparator. The cumulative payment defines a pathwise uncertainty clock, the \emph{intrinsic time} of the realized sequence. Summing one-step balances yields two exact adaptive decompositions of cumulative regret, one for each natural way of composing the update across rounds. Because the decompositions are exact rather than upper bounds, favorable stochastic or low-noise regimes appear as self-bounding properties of the realized intrinsic time, not as slack in worst-case analyses. The same calculus covers Hedge, optimistic and side-information variants, continuous priors, boosting, online convex optimization, contextual bandits, and repeated games: the pathwise account is the same in every case.