Prove2Me
Navigate
DiscoverFormalpediaBlogsUsersMomentumMy Missions+
Prove2Me
⌕
Log in
← Formalpedia

(4.9), proof of Theorem 4.4, p. 307 — second term: η∇f(x_t)⊤(y_{t+1} − x_{t+1}) ≤ D_Φ(x_{t+1}, x_t) − (ρ/2)‖x_{t+1} − y_{t+1}‖² − (ρ/2)‖y_{t+1} − x_t‖²

Open
ConvexOptAlg.MirrorProx.eq_4_9

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

bregman-divergenceconvex-optimizationmirror-proxp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1strong-convexity

In the setting of Chapter 4, let Φ\PhiΦ be a mirror map on D\mathcal DD that is ρ\rhoρ-strongly convex on X∩D\mathcal X\cap\mathcal DX∩D with respect to ∥⋅∥\|\cdot\|∥⋅∥, and let (xt,yt,yt′,xt′)(x_t,y_t,y'_t,x'_t)(xt​,yt​,yt′​,xt′​) be a run of mirror prox with step size η\etaη. Then for every t≥1t\ge1t≥1:

  1. the bound through (4.9):
η ∇f(xt)⊤(yt+1−xt+1)≤DΦ(xt+1,xt)−DΦ(xt+1,yt+1)−DΦ(yt+1,xt);\eta\,\nabla f(x_t)^\top(y_{t+1}-x_{t+1})\le D_\Phi(x_{t+1},x_t)-D_\Phi(x_{t+1},y_{t+1})-D_\Phi(y_{t+1},x_t);η∇f(xt​)⊤(yt+1​−xt+1​)≤DΦ​(xt+1​,xt​)−DΦ​(xt+1​,yt+1​)−DΦ​(yt+1​,xt​);
  1. the final bound:
η ∇f(xt)⊤(yt+1−xt+1)≤DΦ(xt+1,xt)−ρ2∥xt+1−yt+1∥2−ρ2∥yt+1−xt∥2.\eta\,\nabla f(x_t)^\top(y_{t+1}-x_{t+1})\le D_\Phi(x_{t+1},x_t)-\frac\rho2\|x_{t+1}-y_{t+1}\|^2-\frac\rho2\|y_{t+1}-x_t\|^2.η∇f(xt​)⊤(yt+1​−xt+1​)≤DΦ​(xt+1​,xt​)−2ρ​∥xt+1​−yt+1​∥2−2ρ​∥yt+1​−xt​∥2.

This bounds the second of the three terms in the proof of Theorem 4.4; the negative quadratic terms are what absorbs the third term.

Formalization Note No sign condition on η\etaη or ρ\rhoρ is needed for this display. The dual vector ∇f(xt)\nabla f(x_t)∇f(xt​) is the value of the gradient map at xtx_txt​.

Preamble
import Mathlib
import Definitions.Def_ConvexOptAlg_MirrorProx_Defs
Formal statement
namespace ConvexOptAlg.MirrorProx

/-- The second term in the proof of Theorem 4.4 (Bubeck, arXiv:1405.4980v2, §4.5, p. 307, first
display, through (4.9)): for a run of mirror prox with step size `η`, a mirror map `Φ` that is
`ρ`-strongly convex on `X ∩ D`, and every `t ≥ 1`,
1. `η∇f(x_t)⊤(y_{t+1} − x_{t+1}) ≤ D_Φ(x_{t+1}, x_t) − D_Φ(x_{t+1}, y_{t+1}) − D_Φ(y_{t+1}, x_t)`
   (4.9);
2. `η∇f(x_t)⊤(y_{t+1} − x_{t+1})
     ≤ D_Φ(x_{t+1}, x_t) − (ρ/2)‖x_{t+1} − y_{t+1}‖² − (ρ/2)‖y_{t+1} − x_t‖²`. -/
theorem eq_4_9 {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E]
    (X D : Set E) (hXc : IsCompact X) (hXconv : Convex ℝ X) (hXD : X ⊆ closure D)
    (hXDne : (X ∩ D).Nonempty)
    (Φ : E → ℝ) (Φ' : E → E →L[ℝ] ℝ) (hΦ : IsMirrorMap D Φ Φ')
    (ρ : ℝ) (hsc : IsStronglyConvexWRT (X ∩ D) Φ Φ' ρ)
    (f' : E → E →L[ℝ] ℝ) (η : ℝ) (x y y' x' : ℕ → E)
    (hrun : IsMirrorProxRun X D Φ Φ' f' η x y y' x')
    (t : ℕ) (ht : 1 ≤ t) :
    η * f' (x t) (y (t + 1) - x (t + 1))
        ≤ bregman Φ Φ' (x (t + 1)) (x t) - bregman Φ Φ' (x (t + 1)) (y (t + 1))
            - bregman Φ Φ' (y (t + 1)) (x t) ∧
      η * f' (x t) (y (t + 1) - x (t + 1))
        ≤ bregman Φ Φ' (x (t + 1)) (x t) - ρ / 2 * ‖x (t + 1) - y (t + 1)‖ ^ 2
            - ρ / 2 * ‖y (t + 1) - x t‖ ^ 2 := by sorry

end ConvexOptAlg.MirrorProx
Source
Bubeck, arXiv:1405.4980v2, §4.5, proof of Theorem 4.4, p. 307, first display, Eq. (4.9)

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 paper
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me