Local positive moment of the compensated normalized variable on the alive window
ProvedG26G.Sharp.win_posSetting. is a compact action set spanning , is a compact generator (), is the direction enumeration, and is a GIA-FTRL run of rounds (IsGIAFTRLRun) with cumulative loss estimate , run filtration , learning rate (gbeta) and magnitude constant (gB). The filtration used below is for and (Fx). The standing hypotheses Sharp hold: lies in the unit ball, , , is the unique minimizer of on , and the instance is sharp with modulus (H.sig): on . The corruption is bounded by : a.s. (IsCorruptedG). The projective body is and the cone body is ; , , and is the index of among the reference atoms (js).
Assume is surjective and . Let , , , , , and with ,
Then, with , , as in runN_SI and ,
The starting value at is controlled by the absolute envelope of the normalized field and the projective extension constant . The drift is at most , and the scalar positive-moment estimate propagates this bound to time . This is the positive part of (G.568).
Formalization Note Rounds are 0-based.
import Mathlib
import Definitions.Def_IAFTRL_LinEnv
import Definitions.Def_IAFTRL_GIAFTRL
import Definitions.Def_IAFTRL_G26_body
import Definitions.Def_IAFTRL_G26_run
import Definitions.Def_IAFTRL_G26_window
open MeasureTheory ProbabilityTheory Set IAFTRL Matrix J11
open G26G
open G26 G26R
variable {d : ℕ} {𝒜 ℬ : Set (Fin d → ℝ)} {e : ℕ → Fin d → ℚ} {θ astar : Fin d → ℝ}
variable {Ω : Type*} [mΩ : MeasurableSpace Ω] {P : Measure Ω} [IsProbabilityMeasure P]
{T : ℕ} {R : GRun d Ω}
open G26G.Sharp
variable (H : Sharp 𝒜 ℬ θ astar) (he : Function.Surjective e)
include H heomit he in
theorem G26G.Sharp.win_pos (he : Function.Surjective e) (hR : IsGIAFTRLRun P 𝒜 ℬ e T R) {C : ℝ}
(hC : 0 ≤ C) (hcor : IsCorruptedG P 𝒜 θ C T R) {u : Fin d → ℝ} (hu : u ∈ UT ℬ θ astar)
{z : ℝ} (hz : 0 ≤ z) {m t : ℕ} (hmt : m ≤ t) (htT : t < T)
(hmc : cTw d H.sig⁻¹ H.kb ≤ root8 ((m : ℝ) + 1)) (hlC : H.sig⁻¹ * C ≤ (m : ℝ) + 1)
(hzt : z ≤ ((t : ℝ) + 1) * root8 (((m : ℝ) + 1) / ((t : ℝ) + 1)) ^ 7)
(hξ : (2 * (Lam1 ℬ θ astar * max 1 (cK d) + 4) + 16) * root8 (((m : ℝ) + 1) / ((t : ℝ) + 1)) ≤ 1) :
kpos P (YN R θ H.sig⁻¹ C (tauB z) u) (IN 𝒜 ℬ e (js ℬ e astar) R θ astar H.sig⁻¹ C (tauB z) m)
(4 * kk d) t ≤ (8 * ((t : ℝ) + 1)) ^ (4 * kk d) := by sorry