Lemma 2 —
ProvedForecastAggLearn.lemma_2forecast-aggregationonline-learningp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1probability
Let be a probability space carrying a state and, for each expert , an information -algebra and a forecast that is -measurable and a version of , i.e. for every . This is an arbitrary information structure (not necessarily a partial evidence one): in the paper , is generated by and . For let
be the events that some expert's forecast is extreme low, resp. extreme high.
Lemma 2. For every ,
i.e. whenever has positive probability. Conditioned on some expert forecasting at most , the state is with probability at most ; this is what lets the aggregator forecast safely on periods with an extreme forecast, without learning.
Differences from the source:
- generalization: information structure I ∈ Δ(Ω × S_1 × … × S_n) with F_i = I(ω = 1, s_i)/(I(ω = 1, s_i) + I(ω = 0, s_i)) → any probability space with forecasts F_i that are 𝓖_i-measurable versions of P(ω = 1 | 𝓖_i) for sub-σ-algebras 𝓖_i; reason: the paper's structure is the case 𝓖_i = σ(s_i)
- encoding: I(ω = 1 | ζ_0(α)) ≤ nα → ν({ω = 1} ∩ ζ_0(α)) ≤ nα · ν(ζ_0(α)); reason: the conditional probability is undefined when ν(ζ_0(α)) = 0; the product form is the same inequality otherwise
Preamble
import Mathlib
Formal statement
namespace ForecastAggLearn
open MeasureTheory
/-- **Lemma 2: an extreme low forecast caps the probability of ω = 1.**
Babichenko, Garber, arXiv:1802.07107v1, Lemma 2, p.20: for an arbitrary information structure
`I ∈ Δ(Ω × S_1 × … × S_n)` and `ζ_0(α)`, the event that at least one expert's forecast is in
`[0, α]`, "For every α > 0 and every information structure I, I(ω = 1 | ζ_0(α)) ≤ nα."
**Formalization Note.** The information structure is a probability space `(X, ν)` with the state
`ω : X → Bool` and, for each expert `i`, the information `𝓖 i` (a sub-σ-algebra) and the forecast
`F i`, a `𝓖 i`-measurable version of `P(ω = 1 | 𝓖 i)`, i.e. `ν({ω = 1} ∩ E) = ∫_E F_i dν` for every
`E ∈ 𝓖 i`. The paper's `Ω × S_1 × … × S_n` with `F_i = I(ω = 1, s_i)/(I(ω = 1, s_i) + I(ω = 0, s_i))`
is the case `X = Ω × Π_i S_i`, `𝓖 i` generated by `s_i`. The conditional probability, undefined
when `ν(ζ_0(α)) = 0`, is stated in product form `ν({ω = 1} ∩ ζ_0) ≤ nα · ν(ζ_0)`. -/
theorem lemma_2 {X : Type*} [mX : MeasurableSpace X] (ν : Measure X) [IsProbabilityMeasure ν]
(ω : X → Bool) (hω : Measurable ω) {n : ℕ} (𝓖 : Fin n → MeasurableSpace X)
(h𝓖 : ∀ i, 𝓖 i ≤ mX) (F : Fin n → X → ℝ) (hFm : ∀ i, Measurable[𝓖 i] (F i))
(hF : ∀ i, ∀ E, MeasurableSet[𝓖 i] E →
ν ({x | ω x = true} ∩ E) = ∫⁻ x in E, ENNReal.ofReal (F i x) ∂ν)
(α : ℝ) (hα : 0 < α) :
ν ({x | ω x = true} ∩ {x | ∃ i, F i x ≤ α}) ≤
ENNReal.ofReal (n * α) * ν {x | ∃ i, F i x ≤ α} := by sorry
end ForecastAggLearn
Source
Babichenko, Garber, Learning of Optimal Forecast Aggregation in Partial Evidence Environments, arXiv:1802.07107v1, Lemma 2, p.20; proof Appendix A, pp.27–28
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.