empiricalJoint_col_deviation_le
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.empiricalJoint_col_deviation_le← empiricalJoint_col_deviation_le (TCSlib/LearningTheory/Minimax/CCE.lean:457): verbatim
- 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
Choose a tactic — a marked line — to see the goals Lean had before and after it, as in Lean’s infoview. Step with ← → or j k.
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]