Theorem 3, pp. 21β22 β πβ(π«, f_k) β₯ M max{(1/(8k*p_k))(β(p_k(1βp_k)/(16n)) β§ Β½(1βp_k) β§ p_k), β Ξ²_k^{1/k}c_k(Ο)(1/(4n) β§ p_k β§ (1β(1βΞ²_k)^{1βk*}p_k))^{1/k*}}
ProvedUniformDRO.LowerBound.theorem_3Let , , let be arbitrary but fixed, and define
Fix , let be the set of all distributions on , and let be the minimax risk (16) of estimating the CressieβRead robust risk , , from i.i.d. observations , over all estimators and all . Then
The two branches scale as and , so no estimator of the robust risk can converge faster than uniformly over distributions on two points. This matches the rate of the concentration upper bound for the plug-in robust risk (Theorem 2 of the paper), showing that the slow rate for small is unavoidable.
Formalization Note. (1) The class , the estimators and the expectation are encoded on the finite sample space , and takes values in ; the robust risk is the likelihood-ratio form (3) of . (2) Correction of the print: the page has inside the square root of the first branch. The paper's argument (p. 46) bounds and applies Pinsker in the form , which yields , not the that the constant requires. With the same argument gives , , and exactly the printed constant. The second branch is as printed. (3) The hypothesis is implicit on the page (the bound contains ). (4) The minimax risk agrees, for these laws, with the platform's general HighDimStat.Minimax.minimaxRisk.
import Mathlib import Definitions.Def_UniformDRO_LowerBound_Setting
namespace UniformDRO.LowerBound
theorem theorem_3 (k Ο M : β) (hk : 1 < k) (hΟ : 0 < Ο) (hM : 0 < M) (n : β) (hn : 0 < n) :
ENNReal.ofReal (M * max
(1 / (8 * UniformDRO.Concentration.kstar k * pk k Ο) *
min (min (Real.sqrt (pk k Ο * (1 - pk k Ο) / (16 * n))) ((1 - pk k Ο) / 2)) (pk k Ο))
(1 / 8 * Ξ²k k Ο ^ (1 / k) * UniformDRO.Concentration.ck k Ο *
(min (min (1 / (4 * (n : β))) (pk k Ο))
(1 - (1 - Ξ²k k Ο) ^ (1 - UniformDRO.Concentration.kstar k) * pk k Ο)) ^ (1 / UniformDRO.Concentration.kstar k))) β€
minimaxRisk k Ο M n := by sorry
end UniformDRO.LowerBound
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.