Winner-local gradient confidence on the complete terminal history
ProvedLG2021.Concentration.estimatorGood_full_gradientFix the same compact convex legal set, explicit configuration, complete history, normalized adaptive noise condition, candidate norm bounds, and paired-prefix confidence event as in the scalar confidence theorem. Assume additionally that the one-Lipschitz loss is globally convex and the confidence logarithm is nonnegative. Then, for every healthy actual trace state , every retained comparison point, and the candidate selected by the actual stopping rule, its directional-gradient deviation is bounded by the published radius
where are the configuration's radius, moment constant, and concentration constant, is its confidence logarithm, and is the actual sample count. The formal statement specifies the retained-point and winner conditions exactly. This transfers confidence from the paired observation prefix to the complete terminal history.
This is a formalization auxiliary for Algorithms 1–2 and their probability analysis in Tor Lattimore and András György, Improved Regret for Zeroth-Order Stochastic Convex Bandits, COLT 2021, PMLR 134, Lemma 4 and Sections 5–7. The implementation and event predicates are published definitions; the guarantee is proved in this theorem. Adaptive observation kernels and exact finite-history semantics are explicit formal interfaces. Statistical concentration and regret claims are not assumed by the implementation definitions.
import Mathlib import Definitions.Def_LG2021_Policy open LG2021 LG2021.Algorithm LG2021.Concentration BilateralProfitMHR MeasureTheory ProbabilityTheory Set Metric
theorem LG2021.Concentration.estimatorGood_full_gradient {d n u N : ℕ} (hd : 0 < d) (hn : 0 < n)
{D : ℝ} (hD : 0 ≤ D) {K : Set (Space d)} (hK : IsCompact K) (hcv : Convex ℝ K)
(hdiam : ∀ x ∈ K, ∀ y ∈ K, dist x y ≤ D) (R : ℝ) (anchor : Space d) (ha : anchor ∈ K)
(ν : ObsKernel (Space d)) (hνm : ν.IsMeasurable) (F : Space d → ℝ)
(hLip : LipschitzWith 1 F) (hconv : ConvexOn ℝ Set.univ F) (hF : ∀ x ∈ K, F x ∈ Set.Icc (0 : ℝ) 1)
(hν : CorollaryTransport.UnitNoiseCond K F ν) (hN : N ≤ n) (hpair : 2*N ≤ n)
(w : Seed d u n × (Fin n → ℝ)) (v : Fin (d+1) → Space d)
(hH : 0 ≤ confidenceLog (oracleConfig d D n) u)
(hseed : ∀ a k, ‖seedCandidates w.1 a k‖ ≤ (oracleConfig d D n).radius / (2*(d : ℝ)))
(hg : historyPrefix hpair w ∈ estimatorGood (oracleConfig d D n) K hK.isClosed.measurableSet R anchor ν hνm N v) :
∀ p < N, (interior (codedBody K (traceState (oracleConfig d D n) K R w.1 (terminalPairs w.2) p).1)).Nonempty →
(∀ i, v i ∈ codedBody K (traceState (oracleConfig d D n) K R w.1 (terminalPairs w.2) p).1) →
∀ k, traceWinner (oracleConfig d D n) K R w.1 (terminalPairs w.2) p = some k → ∀ i,
traceGradientDeviation (oracleConfig d D n) K R w.1 (terminalPairs w.2) F p k (v i) ≤
2*(oracleConfig d D n).radius*((oracleConfig d D n).momentConstant*Real.sqrt
(confidenceLog (oracleConfig d D n) u / ((oracleConfig d D n).concentrationConstant *
(traceSampleCount (oracleConfig d D n) K R w.1 (terminalPairs w.2) p : ℝ))) +
1/(8*(oracleConfig d D n).radius*(n : ℝ))) := by sorry