Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Proposition 2 — regret of primal-dual subgradient (dual averaging) with time-varying proximal functions, (10)

Proved
AdaGrad.Diag.proposition_2

by mikedeng1 · Oct 5, 2026 · Mathlib 0df444a (Lean v4.33.1)

dual-averagingonline-learningp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1regret

Let η>0\eta>0η>0, let X⊆Rd\mathcal X\subseteq\mathbb R^dX⊆Rd be closed and convex with 0∈X0\in\mathcal X0∈X, and let φ\varphiφ and every ftf_tft​ be convex functions Rd→R\mathbb R^d\to\mathbb RRd→R, with φ\varphiφ minimized over X\mathcal XX at 000. For t=0,1,2,…t=0,1,2,\dotst=0,1,2,… let ht∈Rdh_t\in\mathbb R^dht​∈Rd have positive entries, non-decreasing in ttt (ht,i≤ht+1,ih_{t,i}\le h_{t+1,i}ht,i​≤ht+1,i​), and let ψt(y)=12∑iht,iyi2\psi_t(y)=\frac12\sum_ih_{t,i}y_i^2ψt​(y)=21​∑i​ht,i​yi2​, so that ψt+1≥ψt\psi_{t+1}\ge\psi_tψt+1​≥ψt​, with squared dual norm ∥g∥ψt∗2=∑igi2/ht,i\|g\|^2_{\psi_t^*}=\sum_ig_i^2/h_{t,i}∥g∥ψt∗​2​=∑i​gi2​/ht,i​. Let x1=0x_1=0x1​=0 and let x1,x2,…x_1,x_2,\dotsx1​,x2​,… be a run of the primal-dual subgradient update (3): gtg_tgt​ is a subgradient of ftf_tft​ at xtx_txt​ and xt+1∈Xx_{t+1}\in\mathcal Xxt+1​∈X minimizes

η⟨1t∑τ=1tgτ,y⟩+ηφ(y)+1tψt(y)\eta\Big\langle\frac1t\sum_{\tau=1}^tg_\tau,y\Big\rangle+\eta\varphi(y)+\frac1t\psi_t(y)η⟨t1​τ=1∑t​gτ​,y⟩+ηφ(y)+t1​ψt​(y)

over X\mathcal XX, for every t≥1t\ge1t≥1. Then for every TTT and every x∗∈Xx^*\in\mathcal Xx∗∈X,

∑t=1T[ft(xt)+φ(xt)−ft(x∗)−φ(x∗)]≤1ηψT(x∗)+η2∑t=1T∥gt∥ψt−1∗2.\sum_{t=1}^T\big[f_t(x_t)+\varphi(x_t)-f_t(x^*)-\varphi(x^*)\big]\le\frac1\eta\psi_T(x^*)+\frac\eta2\sum_{t=1}^T\|g_t\|^2_{\psi_{t-1}^*}.t=1∑T​[ft​(xt​)+φ(xt​)−ft​(x∗)−φ(x∗)]≤η1​ψT​(x∗)+2η​t=1∑T​∥gt​∥ψt−1∗​2​.

This is the regret bound of dual averaging when the proximal function may grow from round to round; note that round ttt's subgradient is measured in the dual norm of the previous round's proximal function.

Formalization Note Specialised to diagonal quadratic proximal functions; the standing assumptions of §2 (p. 2129), ψt+1≥ψt\psi_{t+1}\ge\psi_tψt+1​≥ψt​ and 1-strong convexity of ψt\psi_tψt​ with respect to ∥⋅∥ψt\|\cdot\|_{\psi_t}∥⋅∥ψt​​, are the monotonicity of the weights and the quadratic form. The weights are taken strictly positive, so every dual norm is finite. The proof's "recalling that x1=argmin⁡x∈Xφ(x)x_1=\operatorname{argmin}_{x\in\mathcal X}\varphi(x)x1​=argminx∈X​φ(x)" (p. 2153) and the identity ∇ψ0∗(0)=x1\nabla\psi_0^*(0)=x_1∇ψ0∗​(0)=x1​ it uses are made explicit, with Figure 1's initialization, as x1=0∈Xx_1=0\in\mathcal Xx1​=0∈X and φ(0)≤φ(y)\varphi(0)\le\varphi(y)φ(0)≤φ(y) for y∈Xy\in\mathcal Xy∈X.

Preamble
import Mathlib
import Definitions.Def_AdaGrad_Diag_Setup
Formal statement
namespace AdaGrad.Diag

/-- Duchi, Hazan, Singer, JMLR 12 (2011), Proposition 2, (10), p. 2129, specialised to diagonal
quadratic proximal functions `ψ_t(y) = ½ ∑_i h_{t,i} y_i²` (`t = 0, 1, 2, …`) with positive weights
that are non-decreasing in `t` (so `ψ_{t+1} ≥ ψ_t`, the standing assumption of §2). Let the `f_t`
and `ϕ` be convex, `X` closed and convex with `0 ∈ X`, let `x_1 = 0` minimize `ϕ` over `X`, and let
`x_1, x_2, …` be a run of the primal-dual subgradient update (3) with subgradients `g_t` of `f_t` at
`x_t`. Then for every `T` and every `x* ∈ X`,
`∑_{t=1}^T [f_t(x_t) + ϕ(x_t) − f_t(x*) − ϕ(x*)] ≤ (1/η) ψ_T(x*) + (η/2) ∑_{t=1}^T ‖g_t‖²_{ψ_{t−1}*}`,
with `‖g‖²_{ψ*} = ∑_i g_i² / h_i`. -/
theorem proposition_2 {d : ℕ} (η : ℝ) (hη : 0 < η) (X : Set (EuclideanSpace ℝ (Fin d)))
    (hXc : Convex ℝ X) (hXcl : IsClosed X) (hX0 : (0 : EuclideanSpace ℝ (Fin d)) ∈ X)
    (ϕ : EuclideanSpace ℝ (Fin d) → ℝ) (hϕ : ConvexOn ℝ Set.univ ϕ)
    (hϕ0 : ∀ y ∈ X, ϕ 0 ≤ ϕ y)
    (f : ℕ → EuclideanSpace ℝ (Fin d) → ℝ) (hf : ∀ t, ConvexOn ℝ Set.univ (f t))
    (h : ℕ → Fin d → ℝ) (hpos : ∀ t i, 0 < h t i) (hmono : ∀ t i, h t i ≤ h (t + 1) i)
    (x g : ℕ → EuclideanSpace ℝ (Fin d)) (hx1 : x 1 = 0)
    (hrun : IsDualAveragingRun η X ϕ f h x g)
    (T : ℕ) (xstar : EuclideanSpace ℝ (Fin d)) (hxstar : xstar ∈ X) :
    regret f ϕ x xstar T
      ≤ 1 / η * prox (h T) xstar
        + η / 2 * ∑ t ∈ Finset.Icc 1 T, dualNormSq (h (t - 1)) (g t) := by sorry

end AdaGrad.Diag
Source
Duchi, Hazan, Singer, Adaptive Subgradient Methods for Online Learning and Stochastic Optimization, JMLR 12 (2011), p. 2129, Proposition 2, (10), standing assumptions of §2; proof pp. 2152–2153
Human review
  • Endorsed by Shuze Chen · Oct 6, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 6, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me