Theorem 3 — a randomized forecast calibrated against every learning rule of the opponent
ProvedCalibratedCE.Forecast.exists_calibrated_randomized_forecastLet player 2 have pure strategies. There exists a randomized forecaster for player 1, whose forecasts are always probability vectors over player 2's strategies, such that no matter what learning rule player 2 uses (deterministic or randomized, depending on the whole past, but not on player 1's current forecast), player 1's calibration score
converges to zero in probability: for every ,
where is the law of the first rounds when forecast and play are drawn independently in each round given the past.
The forecaster is chosen first and works against every opponent; the convergence rate may depend on the opponent. By the impossibility theorem of Oakes (1985), no deterministic forecaster has this property, which is why randomization is needed.
Formalization Note Forecaster and opponent are maps from histories (lists of (forecast, play) pairs) to Mathlib PMFs; the probability of the event is the outer measure of the history law, a value in , and the limit is taken there. is required: with no strategies there is no probability vector.
import Mathlib import Definitions.Def_CalibratedCE_Forecast_IsDist import Definitions.Def_CalibratedCE_Forecast_CalibScoreH import Definitions.Def_CalibratedCE_Forecast_HistLaw
namespace CalibratedCE.Forecast
theorem exists_calibrated_randomized_forecast (n : ℕ) (hn : 0 < n) :
∃ F : List ((Fin n → ℝ) × Fin n) → PMF (Fin n → ℝ),
(∀ h, ∀ p ∈ (F h).support, IsDist p) ∧
∀ A : List ((Fin n → ℝ) × Fin n) → PMF (Fin n),
∀ ε : ℝ, 0 < ε →
Filter.Tendsto
(fun t : ℕ => (histLaw F A t).toOuterMeasure {h | calibScoreH h < ε})
Filter.atTop (nhds 1) := by sorry
end CalibratedCE.Forecast
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
The statement is about a fixed positive integer . It uses three definitions from its own bundle, IsDist, calibScoreH and histLaw, whose code is not shown here. Below they are described only by the types their use forces on them. What they compute is not visible, and the claim's content depends entirely on them.
Objects involved.
- A forecast is a vector , meaning a function . Nothing in the types requires it to be nonnegative or to sum to .
- An outcome is an element .
- A history is a finite list of pairs (forecast, outcome), . The empty list is allowed.
- is a property of a single vector from the bundle. Its definition is not shown.
- is a real number attached to each history by the bundle's
calibScoreH. Its formula is not shown. - takes a forecasting rule , an outcome rule and a natural number . It returns a probability distribution on histories. It is used with the "outer measure" of a probability mass function, so it is presumably a discrete probability distribution. How it is built from , and is not shown. In particular, the code does not show whether the histories it produces have length exactly , or how the forecast and outcome at each step are drawn.
The claim. Suppose . Then there exists a map with the following properties:
- sends every history to a discrete probability distribution on . It is a probability mass function, so it puts all its mass on at most countably many vectors.
- Validity of forecasts: for every history whatever, including histories that can never actually occur, every vector with satisfies .
- Calibration: for every map and every real ,
Here sends every history to a probability distribution on the outcomes .
The probability above is the total mass the distribution gives to histories whose score is strictly below . Measurability is not an issue because the distribution is discrete. The limit is taken along the natural numbers .
Order of quantifiers. A single is chosen first, and it must work simultaneously for every and every . may not depend on or on .
Each of and sees only a history, meaning the list of past (forecast, outcome) pairs. Nothing else is passed to them. How much they "see" in practice depends on how chains them together, and that is not shown. For example, the code does not show whether sees the forecast realized at the current step.
The only convergence required is convergence of this probability to . Nothing is said about any rate, about almost-sure behaviour along a single infinite play, or about the size of beyond the strict threshold .
Degenerate cases.
- : excluded by the hypothesis .
- : forecasts are single real numbers, and there is exactly one outcome. The only possible always returns the point mass on that outcome. The statement then reduces to the existence of an whose supported values satisfy and for which the displayed limit holds for that single .
- Finitely many : the claim says nothing about any particular finite , including and whatever is (possibly the point mass on the empty history). Only the behaviour as matters.
- If is below every on all histories (for example, if it is always , or on every history can produce), the calibration part holds for any . The statement then reduces to the existence of an whose supported forecasts satisfy . Whether this happens depends on the unshown definition.
- If holds for no vector at all, then no distribution can meet the validity condition. A probability mass function always has at least one supported point, so the statement would be false.
- Junk values inside the definitions: any division by zero or other default value that produces on some histories (for example on the empty history, or where a count is zero) enters the event exactly as computed. The code shown does not reveal whether this happens.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.