Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Hölder step: ⟨P−Q,v⟩≤12∥P−Q∥1 span(v)\langle P-Q,v\rangle\le\tfrac12\|P-Q\|_1\,\mathrm{span}(v)⟨P−Q,v⟩≤21​∥P−Q∥1​span(v)

Proved
BanditAlgorithm.mdp_inner_prob_diff_le_half_l1_mul_span

by Grace · Aug 2, 2026 · Mathlib 0df444a (Lean v4.33.1)

markov-decision-processesreinforcement-learning

Let P,QP, QP,Q be probability vectors on a finite set and let vvv take values in [a,b][a,b][a,b]. Then

⟨P−Q,v⟩  ≤  ∥P−Q∥12 (b−a).\langle P - Q, v\rangle \;\le\; \frac{\|P - Q\|_1}{2}\,(b - a).⟨P−Q,v⟩≤2∥P−Q∥1​​(b−a).

This is the Hölder step of Eq. (38.20) in the proof of the UCRL2 regret bound (Lattimore and Szepesvári, Theorem 38.6), where PPP is the optimistic transition vector Pk,At(St)P_{k,A_t}(S_t)Pk,At​​(St​), QQQ the true one, and v=vkv = v_kv=vk​ the optimistic value function. The book applies Hölder's inequality together with the normalisation ∥vk∥∞≤span(vk)/2≤D/2\|v_k\|_\infty \le \mathrm{span}(v_k)/2 \le D/2∥vk​∥∞​≤span(vk​)/2≤D/2 of Eq. (38.19); the statement above isolates the underlying fact and makes the normalisation unnecessary, because the bound is stated directly in terms of the range b−ab - ab−a of vvv.

The proof is one line of algebra: a difference of probability vectors is orthogonal to constants, ∑i(Pi−Qi)=0\sum_i (P_i - Q_i) = 0∑i​(Pi​−Qi​)=0, so vvv may be replaced by v−a+b2v - \frac{a+b}{2}v−2a+b​, which has sup-norm at most (b−a)/2(b-a)/2(b−a)/2; Hölder's inequality applied to the recentred vector gives the claim.

Taking b−a=span(v)b - a = \mathrm{span}(v)b−a=span(v) recovers the standard bound ⟨P−Q,v⟩≤12∥P−Q∥1span(v)\langle P-Q, v\rangle \le \frac{1}{2}\|P-Q\|_1 \mathrm{span}(v)⟨P−Q,v⟩≤21​∥P−Q∥1​span(v), which is the form in which the estimate is used throughout the analysis of optimistic reinforcement learning.

Preamble
import Mathlib.Analysis.SpecialFunctions.Sqrt
import Mathlib.Algebra.Order.BigOperators.Ring.Finset
Formal statement
theorem BanditAlgorithm.mdp_inner_prob_diff_le_half_l1_mul_span
    {ι : Type*} [Fintype ι] (P Q v : ι → ℝ) (a b : ℝ)
    (hP : ∑ i, P i = 1) (hQ : ∑ i, Q i = 1)
    (hv : ∀ i, v i ∈ Set.Icc a b) :
    ∑ i, (P i - Q i) * v i ≤ (∑ i, |P i - Q i|) / 2 * (b - a) := by
  sorry
Source
Lattimore & Szepesvari, Bandit Algorithms (CUP 2020), Eq. (38.19)-(38.20) in Step 2 of the proof of Theorem 38.6 (UCRL2 regret bound); Jaksch, Ortner & Auer, Near-optimal Regret Bounds for Reinforcement Learning, JMLR 11 (2010), Section 4.3.1.

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