Upper bound
Known-delay DEXP3/DEW upper bound
Nonstochastic Multiarmed Bandits with Unrestricted Delays
Use the source's known-delay or skipping contracts and parameter choice.
Open primary sourceBanditRLwiki case · delayed-adversarial-bandit
Known-delay algorithms attain square-root dependence on rounds and total delay up to logs; local work currently compiles accounting and source-audit interfaces, not a regret endpoint.
Comparison judgment
The literature has strong delayed-feedback rates. Local declarations prove causal views and accounting identities but not the cited algorithm theorem.
Known gap. Logarithmic factors and the distinction between total-delay and fixed-delay contracts remain visible.
Local Lean boundary
Observed/outstanding partition identities, causal action-time views, active allocation, a nonnegative-domain D.11 core, Algorithm-5 line-10 eliminated-arm initialization, and several Delayed SAPO audit surfaces compile. The central generated delayed process and stochastic/adversarial regret endpoints do not.
Upper bound
Nonstochastic Multiarmed Bandits with Unrestricted Delays
Use the source's known-delay or skipping contracts and parameter choice.
Open primary sourceUpper bound
Delay and Cooperation in Nonstochastic Bandits
Fixed-delay feedback under the source's protocol.
Open primary sourceLower bound
Delay and Cooperation in Nonstochastic Bandits
Compare only to upper theorems with the same fixed-delay feedback contract.
Open primary sourceLocal Lean evidence
BanditRLProof.DelayedFeedback.card_observedBefore_add_card_outstandingAt BanditRLProof.DelayedFeedback.actionTimeViewAt BanditRLProof.DelayedFeedback.delayedSAPOProbability BanditRLProof.DelayedFeedback.two_mul_card_sourceStochasticLossGap_aboveTwiceAverage_le BanditRLProof.DelayedFeedback.DelayedSAPOEliminatedArmInitialization.initializeNewlyEliminated_spec_of_mem Not yet proved here
formalization frontier
The bookkeeping, causal-view, active-allocation, and conditional source-audit surfaces compile; no algorithm regret theorem is claimed.
Corrections should preserve the comparison signature and cite a primary theorem, theorem number, source edition, and exact gap being closed. Lean contributions should target one named missing leaf without weakening the mathematical contract.
Propose a sourced update