Theorem 2 (p. 47) — for almost every game,
ProvedCalibratedCE.Generic.ae_limitSet_eq_CESetFix finite strategy sets , with and elements. A game is a pair of real payoff matrices, i.e. a point of ; "a set of games is of measure zero if the corresponding set of points in has Lebesgue measure zero" (p. 46). Let be the set of correlated equilibria of and the set of limit points of calibrated forecasts: the joint distributions for which there are stationary deterministic best-reply functions and history-dependent forecasting rules, calibrated against the opponent's plays, such that when each player plays of its forecast the empirical joint distribution of play converges to .
Theorem. For almost every game,
"In other words, for almost every game, the set of distributions which calibrated learning rules can converge to is identical to the set of correlated equilibriums." (p. 47)
The inclusion holds for every game (Theorem 1). The converse says that every correlated equilibrium is reached, as a genuine limit, by some calibrated forecasting rules with best replies; so calibration alone cannot single out any particular correlated equilibrium. The example of p. 48 shows that the converse fails for some non-generic games.
Formalization Note "Almost every" is with respect to Lebesgue measure on the pair of payoff matrices, . When or both sides are empty.
import Mathlib import Definitions.Def_CalibratedCE_Generic_Game import Definitions.Def_CalibratedCE_Shared_Calibration import Definitions.Def_CalibratedCE_Generic_LimitSet
namespace CalibratedCE.Generic
theorem ae_limitSet_eq_CESet (m n : ℕ) :
∀ᵐ G : (Fin m → Fin n → ℝ) × (Fin m → Fin n → ℝ) ∂MeasureTheory.volume,
LimitSet G.1 G.2 = CESet G.1 G.2 := by sorry
end CalibratedCE.Generic
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
For every pair of natural numbers and , the statement is about pairs where and are each a real array, that is, functions
with entries written and . The space of such pairs is , which can be identified with . It carries its standard volume measure: the product of the coordinatewise Lebesgue measures, i.e. Lebesgue measure on .
The statement asserts that for almost every pair with respect to this measure,
Put another way, the set of pairs where the two objects differ is contained in a set of Lebesgue measure zero. Because this is an "almost everywhere" claim, the exceptional set does not have to be measurable itself. It only has to lie inside a null set. There are no further hypotheses on , , or : no positivity, no genericity condition, no bound on the entries.
and are defined in this bundle's imported modules (a "Game", a "Calibration" and a "LimitSet" definitions file under the namespace CalibratedCE). Their definitions are not included in the code given here, so this read-back cannot unfold them. All the statement itself shows is that each one takes the pair of payoff arrays , in that order, and returns objects of the same type, which the statement claims are equal. What those objects actually are depends entirely on those definitions. Their names alone do not establish their meaning.
Degenerate cases:
- or : each array space has exactly one element, the empty array. The standard volume measure on that one-point space gives the point measure . The product space is then a single point of measure , so "almost every" leaves nothing out. In this case the statement asserts outright that for the unique pair of empty arrays.
- : each array is a single real number, and the claim is about Lebesgue-almost every point .
- All other cases: because only a null set may be excluded, the statement says nothing about any specific non-generic pair. For example, it makes no claim about the all-zero pair, which is a single point and so has measure zero whenever .
- Default values: the statement involves no division, natural-number subtraction, integrals or suprema. Whether such default values occur inside or cannot be determined from the code shown.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.