Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

BanditAlgorithm.mirror_descent_regret_bound

Proved

by Shuze Chen · 1 vote · Jul 29, 2026 · Mathlib c5ea003 (Lean v4.30.0)

mirror-descentonline-learningregret

(Mirror descent regret bound, Theorem 28.4) Let η>0\eta > 0η>0, FFF be Legendre with domain DDD, and A\mathcal{A}A a non-empty convex set with int(D)∩A≠∅\mathrm{int}(D) \cap \mathcal{A} \neq \emptysetint(D)∩A=∅. If a1,a2,…a_1, a_2, \ldotsa1​,a2​,… are the (well-defined) mirror-descent iterates with unprojected dual iterates a~t+1\tilde a_{t+1}a~t+1​, then for any comparator a0∈A∩Da_0 \in \mathcal{A} \cap Da0​∈A∩D, both

Rn(a0)≤F(a0)−F(a1)η+∑t<n⟨at−at+1,yt⟩−1η∑t<nDF(at+1,at),R_n(a_0) \le \frac{F(a_0)-F(a_1)}{\eta} + \sum_{t<n} \langle a_t - a_{t+1}, y_t\rangle - \frac{1}{\eta}\sum_{t<n} D_F(a_{t+1}, a_t),Rn​(a0​)≤ηF(a0​)−F(a1​)​+t<n∑​⟨at​−at+1​,yt​⟩−η1​t<n∑​DF​(at+1​,at​),

and (the Eq. (28.6)-conditioned second form, the condition being internalised in the iterate predicate as existence of the dual iterates)

Rn(a0)≤1η(F(a0)−F(a1)+∑t<nDF(at,a~t+1)).R_n(a_0) \le \frac{1}{\eta}\Big(F(a_0) - F(a_1) + \sum_{t<n} D_F(a_t, \tilde a_{t+1})\Big).Rn​(a0​)≤η1​(F(a0​)−F(a1​)+t<n∑​DF​(at​,a~t+1​)).
Preamble
import Definitions.Def_OnlineLinearOptimization


open RealInnerProductSpace
Formal statement
theorem BanditAlgorithm.mirror_descent_regret_bound
    {d : ℕ} {η : ℝ} (hη : 0 < η)
    {F : EuclideanSpace ℝ (Fin d) → ℝ} {D 𝒜 : Set (EuclideanSpace ℝ (Fin d))}
    (hF : IsLegendre F D) (h𝒜conv : Convex ℝ 𝒜) (h𝒜ne : 𝒜.Nonempty)
    (hmeet : (interior D ∩ 𝒜).Nonempty)
    (y a ã : ℕ → EuclideanSpace ℝ (Fin d)) (n : ℕ)
    (hiter : IsMirrorDescentIterates η F D 𝒜 y a ã) :
    ∀ a₀ ∈ 𝒜 ∩ D,
      (oloRegret a y n a₀ ≤
        (F a₀ - F (a 0)) / η +
          ∑ t ∈ Finset.range n,
            (⟪a t - a (t + 1), y t⟫ - (1 / η) * bregmanDiv F (a (t + 1)) (a t))) ∧
      (oloRegret a y n a₀ ≤
        (1 / η) * (F a₀ - F (a 0) +
          ∑ t ∈ Finset.range n, bregmanDiv F (a t) (ã (t + 1)))) := by
  sorry
Source
L&S Theorem 28.4, p.331
Read-back

What the Lean code literally says, in plain math · claude-fable-5

Work in E=RdE=\mathbb R^dE=Rd with the Euclidean inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle⟨⋅,⋅⟩. For F:E→RF:E\to\mathbb RF:E→R, ∇F\nabla F∇F is the gradient (junk value 000 where FFF is not differentiable) and DF(x,y)=F(x)−F(y)−⟨∇F(y),x−y⟩D_F(x,y)=F(x)-F(y)-\langle\nabla F(y),x-y\rangleDF​(x,y)=F(x)−F(y)−⟨∇F(y),x−y⟩ is the Bregman divergence. For iterates (at)(a_t)(at​), losses (yt)(y_t)(yt​) and comparator uuu, the regret is Regn(u)=∑t=0n−1⟨at−u, yt⟩\mathrm{Reg}_n(u)=\sum_{t=0}^{n-1}\langle a_t-u,\,y_t\rangleRegn​(u)=∑t=0n−1​⟨at​−u,yt​⟩.

The theorem asserts: for every comparator u∈A∩Du\in\mathcal A\cap Du∈A∩D, both of

Regn(u) ≤ F(u)−F(a0)η  +  ∑t=0n−1(⟨at−at+1, yt⟩−1η DF(at+1,at))\mathrm{Reg}_n(u)\ \le\ \frac{F(u)-F(a_0)}{\eta}\;+\;\sum_{t=0}^{n-1}\Big(\big\langle a_t-a_{t+1},\,y_t\big\rangle-\frac1\eta\,D_F\big(a_{t+1},a_t\big)\Big)Regn​(u) ≤ ηF(u)−F(a0​)​+t=0∑n−1​(⟨at​−at+1​,yt​⟩−η1​DF​(at+1​,at​))

and

Regn(u) ≤ 1η(F(u)−F(a0)  +  ∑t=0n−1DF(at, a~t+1)).\mathrm{Reg}_n(u)\ \le\ \frac1\eta\Big(F(u)-F(a_0)\;+\;\sum_{t=0}^{n-1}D_F\big(a_t,\ \tilde a_{t+1}\big)\Big).Regn​(u) ≤ η1​(F(u)−F(a0​)+t=0∑n−1​DF​(at​, a~t+1​)).

Hypotheses:

  • η>0\eta>0η>0.
  • (F,D)(F,D)(F,D) is a Legendre pair: D⊆RdD\subseteq\mathbb R^dD⊆Rd is convex with FFF convex on DDD; int⁡D≠∅\operatorname{int}D\ne\emptysetintD=∅; FFF is differentiable at every point of int⁡D\operatorname{int}DintD and strictly convex there; and along any sequence in int⁡D\operatorname{int}DintD converging to a boundary point of int⁡D\operatorname{int}DintD, ∥∇F∥→∞\|\nabla F\|\to\infty∥∇F∥→∞.
  • A⊆Rd\mathcal A\subseteq\mathbb R^dA⊆Rd is convex and nonempty, and int⁡D∩A≠∅\operatorname{int}D\cap\mathcal A\ne\emptysetintD∩A=∅.
  • (at)(a_t)(at​) and (a~t)(\tilde a_t)(a~t​) are mirror-descent iterates for the losses (yt)(y_t)(yt​) at rate η\etaη: a0∈A∩Da_0\in\mathcal A\cap Da0​∈A∩D minimizes FFF over A∩D\mathcal A\cap DA∩D; every at∈int⁡Da_t\in\operatorname{int}Dat​∈intD; each at+1∈A∩Da_{t+1}\in\mathcal A\cap Dat+1​∈A∩D minimizes b↦η⟨b,yt⟩+DF(b,at)b\mapsto\eta\langle b,y_t\rangle+D_F(b,a_t)b↦η⟨b,yt​⟩+DF​(b,at​) over A∩D\mathcal A\cap DA∩D; each a~t+1∈int⁡D\tilde a_{t+1}\in\operatorname{int}Da~t+1​∈intD satisfies ∇F(a~t+1)=∇F(at)−η yt\nabla F(\tilde a_{t+1})=\nabla F(a_t)-\eta\,y_t∇F(a~t+1​)=∇F(at​)−ηyt​; and each at+1a_{t+1}at+1​ minimizes b↦DF(b,a~t+1)b\mapsto D_F(b,\tilde a_{t+1})b↦DF​(b,a~t+1​) over A∩D\mathcal A\cap DA∩D ("minimizes" = value ≤\le≤ value at every point of the set; no uniqueness).

Edge cases and conventions:

  • The horizon nnn is an arbitrary natural number, including n=0n=0n=0, in which case both sums are empty and the regret is 000.
  • All Bregman terms take gradients at points (ata_tat​, a~t+1\tilde a_{t+1}a~t+1​) that the iterate predicate places in int⁡D\operatorname{int}DintD, where the Legendre hypothesis makes FFF genuinely differentiable; the Bregman divergence itself has no differentiability guard.
  • The two bounds are asserted jointly (a conjunction), for every comparator in A∩D\mathcal A\cap DA∩D — not merely for a best comparator.
  • The dimension ddd is arbitrary, including d=0d=0d=0.
  • The iterate predicate is an assumption; the theorem asserts nothing about existence of such iterates.
Human review
  • Endorsed by Community (Bot) · Jul 29, 2026

  • Endorsed by Shuze Chen · Jul 29, 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