§3, after (13) — with δ ≥ maxₜ‖gₜ‖∞, Σₜ‖gₜ‖²_{ψ*_{t−1}} ≤ 2 Σᵢ‖g_{1:T,i}‖₂
ProvedAdaGrad.Diag.primal_dual_gradient_termadaptive-methodsonline-learningp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be any sequence, (with ), a horizon and with . With AdaGrad's proximal functions , whose squared dual norms are ,
This is the inequality (13) for the primal-dual subgradient update, whose regret bound (Proposition 2) measures round 's subgradient in the dual norm of the previous round's proximal function.
Formalization Note The page states the per-round comparison with and says the same reasoning gives (13); the statement here is that summed conclusion. Division uses Lean's .
Preamble
import Mathlib import Definitions.Def_AdaGrad_Diag_Setup
Formal statement
namespace AdaGrad.Diag
/-- Duchi, Hazan, Singer, JMLR 12 (2011), §3, paragraph after (13), p. 2131: if
`δ ≥ max_{t ≤ T} ‖g_t‖∞`, then the squared dual norms for the *previous* round's proximal
function `ψ_{t−1}(x) = ½⟨x, (δI + diag(s_{t−1})) x⟩` (with `s_0 = 0`) satisfy the inequality (13):
`∑_{t=1}^T ‖g_t‖²_{ψ_{t−1}*} = ∑_{t=1}^T ∑_i g_{t,i}² / (δ + s_{t−1,i}) ≤ 2 ∑_{i=1}^d ‖g_{1:T,i}‖₂`. -/
theorem primal_dual_gradient_term {d : ℕ} (g : ℕ → EuclideanSpace ℝ (Fin d)) (δ : ℝ)
(hδ : 0 ≤ δ) (T : ℕ) (hδg : ∀ t ∈ Finset.Icc 1 T, supNorm (g t) ≤ δ) :
∑ t ∈ Finset.Icc 1 T, dualNormSq (adaWeights δ g (t - 1)) (g t) ≤ 2 * ∑ i, s g T i := by sorry
end AdaGrad.Diag
Source
Duchi, Hazan, Singer, Adaptive Subgradient Methods for Online Learning and Stochastic Optimization, JMLR 12 (2011), p. 2131, §3, paragraph after (13)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.