AFTD Auto-Formalizing Theoretical Domains

regret_le_of_log_potential_step

theoremverifiedimported Lean

Regret from a per-step log-potential bound. Let $N \ge 1$, let $\eta > 0$, let $\ell$ be a loss sequence for $N$ experts over $T$ rounds, and let $c \in \bbr$. Suppose that for every round $t < T$ the log potential satisfies the one-step bound \[ \log W_{t+1} - \log W_t \;\le\; -\eta \cdot \mathrm{hedgeLoss}_t + c , \] where $\mathrm{hedgeLoss}_t$ is the expected loss of Hedge in round $t$. Then the regret of Hedge is bounded by \[ \mathrm{regret}(\eta,\ell) \;\le\; \frac{\log N}{\eta} + \frac{c\,T}{\eta} . \]

Statement

set_option relaxedAutoImplicit false in
set_option autoImplicit false in
open Real Finset BigOperators in
lemma regret_le_of_log_potential_step {N T : ℕ} [NeZero N] (η : ℝ) (hη_pos : 0 < η)
    (ℓ : LossSeq N T) (c : ℝ)
    (hstep : ∀ t : Fin T, Real.log (potential η ℓ (t.val + 1)) - Real.log (potential η ℓ t.val)
      ≤ -η * hedgeLoss η ℓ t + c) :
    regret η ℓ ≤ Real.log N / η + c * T / η
Source
TCSlib, `regret_le_of_log_potential_step`
Lean proof
By Karim Abdel Sadek, Mark Bedaywi, from github.com/Shilun-Allan-Li/tcslib at c8a02591b1e5 (Copyright (c) 2026 Karim Abdel Sadek and Mark Bedaywi. All rights reserved, Apache-2.0). Compiled here against leanprover/lean4:v4.33.1, Mathlib v4.33.1; written for leanprover/lean4:v4.25.0.
Verified
07 Oct 2026
Axioms
Classical.choiceQuot.soundpropext
Built on 9
  • log_potential_telescopeTelescoping of the log potential. Let $\eta \in \bbr$ and let $\ell$ be a loss sequence for $N$ experts over $T$ rounds, with potential…
  • hedgeCumLossThe cumulative loss of the Hedge algorithm over all $T$ rounds is \[ \widehat{L}_T \;=\; \sum_{t=0}^{T-1} \widehat{\ell}_t. \]
  • hedgeLossThe expected loss of the learner at round $t$ under the Hedge distribution is…
  • regretThe regret of the Hedge algorithm is the excess cumulative loss over the best expert in hindsight:…
  • potentialThe potential (sum of unnormalized weights) at time $t$ is \[ W_t \;=\; \sum_{i=1}^{N} w_t(i). \] The potential telescopes to…
  • bestExpertLossThe best expert loss in hindsight is the minimum cumulative loss achieved by any single expert:…
  • LossSeqA loss sequence $\ell$ for $N$ experts over $T$ rounds is a function $\ell : \mathrm{Fin}\,T \to \mathrm{Fin}\,N \to \mathbb{R}$,…
  • potential_zeroInitial value of the potential. Fix $N \ge 1$ experts, a horizon $T$, and a loss sequence $\ell$ over these experts and rounds, and let…
  • potential_ge_best_expertPotential lower bound via the best expert. Fix $N \ge 1$ experts and $T$ rounds, a loss sequence $\ell$, and a learning rate $\eta > 0$.…
Used by 2
  • hedge_regret_bound_tightTight Hedge regret bound. Fix $N \ge 1$ experts and a horizon of $T$ rounds, and run the Hedge algorithm with a learning rate $\eta > 0$.…
  • hedge_regret_boundRegret bound for the Hedge algorithm. Fix $N \ge 1$ experts and $T$ rounds, and let $\ell$ be a valid loss sequence, so that…

Lean source view module on GitHub

set_option relaxedAutoImplicit false inset_option autoImplicit false inopen Real Finset BigOperators in/-- Regret from a uniform per-step log-potential bound: if every round satisfies `log W_{t+1} - log W_t ≤ -η · hedgeLoss_t + c`, then `regret ≤ (log N)/η + cT/η`. This is the potential argument of [CBL06, proof of Thm 2.2] (also [FS97, §2.1, Lemma 1 and proof of Thm 2]) with the per-round constant abstracted; both `hedge_regret_bound` (`c = η²/2`) and `hedge_regret_bound_tight` (`c = η²/8`) are instances. **Proof sketch.** Step 1: sum the per-step bounds over the `T` rounds; the left side telescopes (`log_potential_telescope`) to `log W_T - log W_0`, and the right side is `-η · hedgeCumLoss + cT`. Step 2: `W_0 = N` (`potential_zero`), so `log W_0 = log N`. Step 3: `W_T ≥ exp(-η · bestExpertLoss)` (`potential_ge_best_expert`), so `log W_T ≥ -η · bestExpertLoss`. Step 4: combine into `-η · bestExpertLoss - log N ≤ -η · hedgeCumLoss + cT`, i.e. `η · regret ≤ log N + cT`, and divide by `η > 0`. -/lemma regret_le_of_log_potential_step {N T : ℕ} [NeZero N] (η : ℝ) (hη_pos : 0 < η)    (ℓ : LossSeq N T) (c : ℝ)    (hstep : ∀ t : Fin T, Real.log (potential η ℓ (t.val + 1)) - Real.log (potential η ℓ t.val)      ≤ -η * hedgeLoss η ℓ t + c) :    regret η ℓ ≤ Real.log N / η + c * T / η := by  -- Step 1: sum the per-step bounds and telescope  have hsum : Real.log (potential η ℓ T) - Real.log (potential η ℓ 0)      ≤ -η * hedgeCumLoss η ℓ + c * T := by    have hbounds := Finset.sum_le_sum fun t (_ : t ∈ Finset.univ) => hstep t    have hrhs : ∑ t : Fin T, (-η * hedgeLoss η ℓ t + c) = -η * hedgeCumLoss η ℓ + c * T := by      simp only [hedgeCumLoss, Finset.mul_sum, Finset.sum_add_distrib, Finset.sum_const,        Finset.card_fin]      ring    linarith [log_potential_telescope η ℓ]  -- Step 2: W_0 = N  have hW0 : Real.log (potential η ℓ 0) = Real.log N := by    rw [potential_zero]  -- Step 3: W_T ≥ exp(-η · bestExpertLoss)  have hWT : Real.log (potential η ℓ T) ≥ -η * bestExpertLoss ℓ := by    calc Real.log (potential η ℓ T) ≥ Real.log (Real.exp (-η * bestExpertLoss ℓ)) :=          Real.log_le_log (exp_pos _) (potential_ge_best_expert η hη_pos ℓ)      _ = -η * bestExpertLoss ℓ := Real.log_exp _  -- Step 4: rearrange η · regret ≤ log N + cT  unfold regret  rw [← add_div, le_div_iff₀ hη_pos]  nlinarith [hsum, hW0, hWT]
See it in the graphDiscuss this resultChallenge it