AFTD Auto-Formalizing Theoretical Domains

empiricalJoint_col_deviation_le

theoremverifiedimported Lean

Column regret bound gives column deviation bound. Let $G$ be a finite two-player game with column utility $u_c$ over $M$ row actions and $N$ column actions, and consider $T > 0$ rounds of play in which the row player uses the mixed strategy $p_t$ and the column player the mixed strategy $q_t$ at round $t$. Fix $\varepsilon \in \bbr$ and a pure column action $j'$, and suppose the column player's cumulative external regret against always playing $j'$ is at most $\varepsilon T$: \[ \sum_{t=1}^{T}\sum_{i}(p_t)_i\, u_c(i,j') \;-\; \sum_{t=1}^{T}\sum_{i,j}(p_t)_i (q_t)_j\, u_c(i,j) \;\le\; \varepsilon T. \] Then in the empirical joint distribution $\sigma$ of the play, deviating to $j'$ gains at most $\varepsilon$ over the expected utility: $\mathrm{colDevUtil}(\sigma, G, j') \le \mathrm{colExpUtil}(\sigma, G) + \varepsilon$. This is the column half of empiricalJoint_isApproxCCE, mirroring

Statement

set_option relaxedAutoImplicit false in
set_option autoImplicit false in
open Finset BigOperators in
lemma empiricalJoint_col_deviation_le {M N T : ℕ} (hT : 0 < T)
    (G : Game M N)
    (p : Fin T → MixedStrategy M) (q : Fin T → MixedStrategy N)
    (ε : ℝ) (j' : Fin N)
    (hCol : (∑ t : Fin T, ∑ i : Fin M, (p t).weights i * G.colUtility i j') -
        (∑ t : Fin T, ∑ i : Fin M, ∑ j : Fin N,
          (p t).weights i * (q t).weights j * G.colUtility i j) ≤ ε * T) :
    (empiricalJoint hT p q).colDeviationUtility G j' ≤
      (empiricalJoint hT p q).colExpectedUtility G + ε
Source
TCSlib, `empiricalJoint_col_deviation_le`
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
  • empiricalJointGiven $T > 0$ rounds of play in which the row player uses mixed strategy $p_t$ and the column player uses $q_t$, the empirical joint…
  • MixedStrategyA mixed strategy over $n$ actions is a record consisting of a weight function $w : \mathrm{Fin}\,n \to \mathbb{R}$ together with…
  • JointDistribution.colDeviationUtilityThe column player's expected utility from unilaterally deviating to the fixed pure action $j'$, while the row player keeps $\sigma$'s row…
  • empiricalJoint_colMarginalColumn marginal of the empirical joint distribution. Consider $T > 0$ rounds of play between a row player with $M$ actions and a column…
  • JointDistribution.colExpectedUtilityThe column player's expected utility under $\sigma$ in game $G$ is $\sum_{i,j} \sigma_{ij} \cdot u_c(i,j)$.
  • empiricalJoint_rowMarginalRow marginal of the empirical joint distribution. Fix $T \ge 1$ rounds of play in which the row player uses mixed strategies…
  • empiricalJoint_colExpectedUtilityColumn utility of the empirical joint distribution. Consider a finite two-player game with $M$ row actions and $N$ column actions, whose…
  • empiricalJoint_colDeviationUtilityColumn deviation utility of the empirical joint distribution. Let $G$ be a finite two-player game with $M$ row actions and $N$ column…
  • GameA finite two-player game with $M$ row actions and $N$ column actions, specified by a row utility function…
Used by 1
  • empiricalJoint_isApproxCCENo-regret play yields an approximate coarse correlated equilibrium. Let $G$ be a finite two-player game with row utility $u_r$ and column…

Lean source view module on GitHub

set_option relaxedAutoImplicit false inset_option autoImplicit false inopen Finset BigOperators in/-- Column half of `empiricalJoint_isApproxCCE`: if the column player's cumulative external regret against the fixed deviation `j'` is at most `ε · T`, then in the empirical joint distribution deviating to `j'` gains at most `ε`. The mirror image of `empiricalJoint_row_deviation_le`, with the same three steps. [Rou13-L17, Prop. 3.1]; [CBL06, §7.4]. **Proof sketch.** Step 1: rewrite both empirical utilities as time averages (`empiricalJoint_colDeviationUtility`, `empiricalJoint_colExpectedUtility`). Step 2: put the right-hand side over the common denominator `T`. Step 3: clear `T > 0` from both sides; the resulting inequality is the regret hypothesis. -/lemma empiricalJoint_col_deviation_le {M N T : ℕ} (hT : 0 < T)    (G : Game M N)    (p : Fin T → MixedStrategy M) (q : Fin T → MixedStrategy N)    (ε : ℝ) (j' : Fin N)    (hCol : (∑ t : Fin T, ∑ i : Fin M, (p t).weights i * G.colUtility i j') -        (∑ t : Fin T, ∑ i : Fin M, ∑ j : Fin N,          (p t).weights i * (q t).weights j * G.colUtility i j) ≤ ε * T) :    (empiricalJoint hT p q).colDeviationUtility G j' ≤      (empiricalJoint hT p q).colExpectedUtility G + ε := by  have hT_pos : (0 : ℝ) < (T : ℝ) := Nat.cast_pos.mpr hT  have hT_ne : (T : ℝ) ≠ 0 := ne_of_gt hT_pos  -- Step 1: rewrite both empirical utilities as time averages.  rw [empiricalJoint_colDeviationUtility, empiricalJoint_colExpectedUtility]  -- Step 2: put the right-hand side over the common denominator `T`.  have hRHS :      (∑ t : Fin T, ∑ i : Fin M, ∑ j : Fin N,          (p t).weights i * (q t).weights j * G.colUtility i j) / (T : ℝ) +        ε =        ((∑ t : Fin T, ∑ i : Fin M, ∑ j : Fin N,          (p t).weights i * (q t).weights j * G.colUtility i j) +          ε * (T : ℝ)) / (T : ℝ) := by field_simp  -- Step 3: clear `T` and conclude from the regret bound.  rw [hRHS, div_le_div_iff₀ hT_pos hT_pos]  nlinarith [hCol, hT_pos]
See it in the graphDiscuss this resultChallenge it