Tightness of the randomized barrier.
ProvedFactoringLab.randomized_barrier_eq_iffaether-catalogprobability
Tightness of the randomized barrier. Equality holds exactly for the degenerate mixtures: those all of whose positively weighted strategies reproduce the band mean on the whole population.
theorem FactoringLab.randomized_barrier_eq_iff[DecidableEq κ] {m : ℕ} (Ω : Finset ι) (n : ι → κ)
(Y : ι → ℝ) (w : Fin m → ℝ) (T : Fin m → DTree ι) (hw : ∑ j, w j = 1)
(hw0 : ∀ j, 0 ≤ w j) (hT : ∀ j, (T j).BandOnly Ω n) :
mixRisk Ω w T Y = ∑ i ∈ Ω, (bandMean Ω n Y i - Y i) ^ 2 ↔
∀ j, 0 < w j → ∀ i ∈ Ω, (T j).eval i = bandMean Ω n Y i := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Probability/RandomizedBarrier.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Probability/RandomizedBarrier.lean
import Mathlib
import Definitions.Def_Probability_AdaptiveBarrier
import Definitions.Def_Probability_RandomizedBarrier
import Definitions.Def_Probability_StructuralOrthogonality
/-
# The randomized barrier (Factoring Lab, Phase A v19c — cycle 3)
This file closes the first (deterministic-reduction) half of **Conjecture D** of
`FUTURE_DIRECTIONS.md`: randomization does not break the structural barrier.
`Catalog/Probability/AdaptiveBarrier.lean` proves that a *single* adaptive
strategy whose tests and outputs are band-measurable cannot predict the hidden
factor better than the band mean. A natural escape route is to randomize: run
a finite mixture `μ = (w_1, …, w_m)` over strategies `t_1, …, t_m` and hope that
the mixture beats every one of its members. It cannot, and the reason is
sharper than convexity alone:
* `FactoringLab.mix_sq_error_decomposition` — an exact *bias–variance identity*
for mixtures: the expected squared error of the randomized strategy equals
the squared error of its **mean predictor** `m(i) = Σ_j w_j t_j(i)` plus the
randomization variance `Σ_i Σ_j w_j (t_j(i) − m(i))²`. Randomizing therefore
*strictly increases* the error unless all strategies with positive weight
agree pointwise on the population (`FactoringLab.randomization_never_helps`,
`FactoringLab.randomization_strictly_hurts`).
* `FactoringLab.mixEval_bandMeasurable` — the mean predictor of a mixture of
band-measurable strategies is itself band-measurable, so it factors through
the band label (`FactoringLab.randomized_is_N_only`) and is dominated by the
band mean.
* `FactoringLab.randomized_barrier` — consequently the expected squared error of
*any* finite mixture of `N`-only adaptive strategies is at least the
irreducible band-conditional error, uniformly in the number of strategies,
their sizes and the mixing weights.
* `FactoringLab.randomized_barrier_eq_iff` — the barrier is *tight exactly* on
the degenerate mixtures: equality holds iff every strategy carrying positive
weight reproduces the band mean on the whole population. This is the
equality clause conjectured in Conjecture D.
Together with `FactoringLab.adaptive_barrier`, this says: neither adaptivity nor
randomness — nor any combination of the two — extracts information about the
hidden factor beyond what the band label already carries.
-/
open Finset
open FactoringLab
variable {ι κ : Type*}
/-! ### Mixtures of strategies -/
/-! ### The bias–variance identity for mixtures -/Formal statement
theorem FactoringLab.randomized_barrier_eq_iff[DecidableEq κ] {m : ℕ} (Ω : Finset ι) (n : ι → κ)
(Y : ι → ℝ) (w : Fin m → ℝ) (T : Fin m → DTree ι) (hw : ∑ j, w j = 1)
(hw0 : ∀ j, 0 ≤ w j) (hT : ∀ j, (T j).BandOnly Ω n) :
mixRisk Ω w T Y = ∑ i ∈ Ω, (bandMean Ω n Y i - Y i) ^ 2 ↔
∀ j, 0 < w j → ∀ i ∈ Ω, (T j).eval i = bandMean Ω n Y i := by sorrySource