§4.2.1 — for injective , is invertible and
ProvedForecastAggLearn.hstar_aggregatesforecast-aggregationonline-learningp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1probability
Let be a real matrix such that is injective. Then is invertible, and with ,
Adding to both sides gives the paper's display with : the Bayes-optimal aggregate is a fixed linear function of what the prior-aware aggregator observes. This is why the aggregation can be learned when is injective.
Differences from the source:
- generalization: A the 0/1 evidence matrix → any real n × m matrix with injective mulVec; reason: the evidence matrix is the 0/1 case
- encoding: chain 1_m·ỹ + µ̃ = 1_m^⊤A_l^{−1}Aỹ + µ̃ = (A_l^{−1⊤}1_m)·z̃ + µ̃ = h*·z̃ + µ̃ → its two ends without the common term µ̃: h* ⬝ᵥ (Aỹ) = Σ_j ỹ_j; reason: adding µ̃ to both sides is the same identity; the middle terms are the definition of h*
Preamble
import Mathlib import Definitions.Def_ForecastAggLearn_hstar
Formal statement
namespace ForecastAggLearn
/-- **The optimal aggregation is linear in `z̃ = Aỹ`.**
Babichenko, Garber, arXiv:1802.07107v1, §4.2.1, display, p.17: "In case A is injective it holds
that A^⊤A is invertible, and hence, denoting the left inverse of A by A_l^{−1} := (A^⊤A)^{−1}A^⊤,
the optimal aggregation has the form
1_m · ỹ + µ̃ = 1_m^⊤ A_l^{−1} Aỹ + µ̃ = (A_l^{−1⊤} 1_m) · z̃ + µ̃ = h* · z̃ + µ̃."
If `A` is injective then `A^⊤A` is invertible and `h* · (Aỹ) = 1_m · ỹ` for every `ỹ ∈ ℝ^m`.
**Formalization Note.** Stated for every real `n × m` matrix with injective `mulVec`; the
evidence matrix is the 0/1 case. The `µ̃` on both sides cancels and is omitted. -/
theorem hstar_aggregates {n m : ℕ} (A : Matrix (Fin n) (Fin m) ℝ)
(hA : Function.Injective A.mulVec) :
IsUnit (A.transpose * A).det ∧ ∀ y : Fin m → ℝ, hstar A ⬝ᵥ A.mulVec y = ∑ j, y j := by sorry
end ForecastAggLearn
Source
Babichenko, Garber, Learning of Optimal Forecast Aggregation in Partial Evidence Environments, arXiv:1802.07107v1, §4.2.1, display, p.17
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.