The oracle mixture
DefinitionRobustKellyTV_Supplement_oracleMixtureWork in the finite-alphabet total-variation model: observations take values in , the null class is , and for a probability vector the pointwise Kelly oracle is the clipped likelihood ratio of eq. (12).
Let be a mixing measure on probability vectors and a set of alternatives. The oracle mixture is the process
which averages, over the prior , the wealth of the oracle bettors that each know one alternative . For the empty product gives , which equals when is a probability measure carried by .
This is the comparison process of Appendix B: a Bayesian mixture over pointwise oracles, whose regret against the oracle growth rate Theorem B.1 bounds by the prior mass of a neighbourhood of .
Formalization Note The observation () is coordinate of a sequence . The integral is the Bochner integral over ; under the assumptions of Appendix B ( compact) the integrand is continuous and bounded on , so the integral is the paper's. is the published RobustKellyTV.oracleE.
import Mathlib
import Definitions.Def_RobustKellyTV_oracleE
open MeasureTheory
namespace RobustKellyTV.Supplement
variable {d : ℕ}
/-- The **oracle mixture** `B^Π_n = ∫_A ∏_{t=1}^n E_U(X_t) Π(dU)` of A. Saha, A. Ramdas,
*Beyond least-favorable distribution pairs: asymptotically instance-optimal robust hypothesis
testing*, arXiv:2611.02130v1, Appendix B, eq. (53), p.35.
`prior` is the mixing measure `Π` on `Fin d → ℝ`, `A` the compact set of alternatives it is
supported on, and `E_U` the published pointwise Kelly oracle `RobustKellyTV.oracleE`.
**Formalization Note.** The paper's observation `X_t` (`t ≥ 1`) is coordinate `t − 1` of
`ω : ℕ → Fin d`, so the product runs over the coordinates `s ∈ {0, …, n − 1}`; the empty product
gives `B^Π_0 = Π(A)`, which is `1` for a probability measure carried by `A`. The integral is the
Bochner integral over `A` with respect to `Π`; under the assumptions of Appendix B
(`A ⊆ K_{m,δ}` compact) the integrand is continuous and bounded on `A` (Lemma A.1), so the integral
is the paper's. -/
noncomputable def oracleMixture (σ : Setting d) (prior : Measure (Fin d → ℝ))
(A : Set (Fin d → ℝ)) (n : ℕ) (ω : ℕ → Fin d) : ℝ :=
∫ U in A, ∏ s ∈ Finset.range n, oracleE σ U (ω s) ∂prior
end RobustKellyTV.Supplement