Adaptive covariance identity.
ProvedFactoringLab.adaptive_cov_eq_cov_bandMeanaether-catalogprobability
Adaptive covariance identity. The covariance of an adaptive N-only
strategy with the target equals its covariance with the band means.
theorem FactoringLab.adaptive_cov_eq_cov_bandMean[DecidableEq κ]
(Ω : Finset ι) (n : ι → κ) (Y : ι → ℝ)
(t : DTree ι) (ht : t.BandOnly Ω n) :
cov Ω t.eval Y = cov Ω t.eval (bandMean Ω n Y) := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Probability/AdaptiveBarrier.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Probability/AdaptiveBarrier.lean
import Mathlib
import Definitions.Def_Probability_AdaptiveBarrier
import Definitions.Def_Probability_StructuralOrthogonality
/-
# The Adaptive Barrier (Factoring Lab, Phase A v19c — cycle 2)
Closing **Conjecture 5** of `FUTURE_DIRECTIONS.md`: *the structural-orthogonality
barrier is closed under adaptivity.*
The previous cycle proved that no invariant `g ∘ n` computable from the band
label alone predicts the target better than the band mean
(`FactoringLab.bandMean_is_best_predictor`), and that arbitrary *fixed*
nonlinear aggregation `Φ(w₁, …, w_m)` of finitely many free witnesses stays
`N`-only (`FactoringLab.aggregation_no_better_than_bandMean`).
An obvious escape route is *adaptivity*: run an `N`-only test, and depending on
its outcome run a different test, and so on, finally emitting a candidate
factor. This file formalizes such a strategy as a finite **decision tree**
whose internal tests and whose leaf outputs are band-measurable, and proves:
* `FactoringLab.DTree.bandMeasurable_eval` — the value computed by such a tree
is itself band-measurable (structural induction on the tree);
* `FactoringLab.factors_through_band` — every band-measurable function factors
as `g ∘ n` for some `g : κ → ℝ`;
* `FactoringLab.adaptive_structural_orthogonality` — the output of any such
tree is orthogonal to the residual `Y − E[Y | n]`;
* `FactoringLab.adaptive_barrier` — no such tree beats the band mean, *for any
depth and any branching*; the excess error is again exactly the squared
deviation from the band mean (`adaptive_sq_error_decomposition`);
* `FactoringLab.adaptive_cov_eq_cov_bandMean` and
`FactoringLab.adaptive_nearEqualN_test` — the near-equal-`N` test applies
verbatim to adaptive strategies: constant band means force covariance `0`.
Boundary (Stage 4, adversarial review). The band-measurability hypothesis is
not decorative: `FactoringLab.adaptive_barrier_fails_without_bandMeasurable`
exhibits a two-point population and a depth-`0` tree which attains error `0`
while the band mean has error `1/2`.
Finally `FactoringLab.witnessTree_bandMeasurable` shows the hypothesis is met
by the strategies the framework is about: any tree whose tests are threshold
comparisons of free witnesses `w : κ → ℝ` and whose leaves are `N`-only
invariants is band-measurable, so the barrier covers the whole adaptive
free-witness family.
-/
open Finset
open FactoringLab
variable {ι κ : Type*}
/-! ## 1. Band-measurable functions -/
/-! ## 2. Adaptive strategies as decision trees -/
open DTree
/-! ## 3. The adaptive barrier -/Formal statement
theorem FactoringLab.adaptive_cov_eq_cov_bandMean[DecidableEq κ]
(Ω : Finset ι) (n : ι → κ) (Y : ι → ℝ)
(t : DTree ι) (ht : t.BandOnly Ω n) :
cov Ω t.eval Y = cov Ω t.eval (bandMean Ω n Y) := by sorrySource