Causal bandits

Result contract to fill

A separate record is required for each exact model and guarantee. Compare bounds only when assumptions, feedback, metrics and parameter regimes match.

Exact setting and model class

Pending source verification

Assumptions

Pending source verification

Feedback structure

Pending source verification

Algorithm / method

Pending source verification

Regret or other metric

Pending source verification

Expectation / high probability

Pending source verification

Horizon, dimension and other parameter dependencies

Pending source verification

Upper bound and conditions

Pending source verification

Lower bound and conditions

Pending source verification

Computation, oracle and relaxation requirements

Pending source verification

Paper / theorem / version / verification date

Pending source verification

Canonical Lean references and completion boundary

Pending source verification

Three separate evidence ledgers

• Literature results: Mapped source and repairs reviewed within the disclosed scope; remaining source audits are incomplete.
• Lean mapping: Mapped results independently reviewed with explicit differences; topic acceptance remains incomplete.
• Literature open problems: none asserted. Missing formalization is not an open mathematical problem.

Reviewed native heterogeneous causal sampling and actual expected regret

Native heterogeneous finite-node actual-law expected regret; complete frozen noisy diagnostics; repaired parallel allocation, actual graph-law/cost bridge, true-cost regret consumers and both boundary witnesses. Remaining source screening and ICLR evidence are open.

Mapped results independently reviewed with explicit scope differences; topic incomplete.

Causal Bandits: Learning Good Interventions via Causal Inference · F. Lattimore, T. Lattimore and M. D. Reid · 2016

• Each node has its own nonempty finite type; final performance theorem constructs the finite product codec internally. Binary readout extends the literal binary source reward node.
• Coverage is explicit; zero allocation weights are allowed only when Q covers every parent support.
• Unchanged source truncation tuning and fixed-order ties; coefficient and failure direction repairs disclosed.
• Uniform actual threshold uses m, with K substituted only in RHS. Optimal allocation minimizes cost, not a proved ordering of actual regrets.
• Frozen noisy test now proves concentrated coverage, cost 8/3, B=2 biases, conditional versus interventional means, an uncovered design, actual tuned T=1 recommendation and exact expected regret 1/5, plus concentrated all-horizon bound. The B=2 bias test is distinct from the actual one-round tuning.
• Parallel bridge and witnesses are reviewed with strict rarity, normalized observation mass 1-D, N>=2 and known q. Actual thresholds retain designCost. Remaining source audit and ICLR evidence remain required; no topic completion.
• Heterogeneous Fin3/Fin2 canary proves native factorization, means 5/12, 1/4, 3/4 and uniform all-horizon rate; no exact diagnostic cost or random recommendation distribution is claimed.
• Evidence erratum: three historical sampling file hashes are invalid. Fresh independent blind/source review binds five foundations and seven sampling modules plus canary to132a834; old records retained. Identity checks use explicit LF normalization and run in the harness. No Lean proof change or topic completion.

Shared graph · Shared reference registry

Related reading

Related routes await review.

Related links suggest starting points; they are not evidence for an unverified setting.

Back to Extended Chapters