Cov comm
ProvedLogic.PhaseRoute.cov_commaether-cataloglogic
Formal statement of Logic.PhaseRoute.cov_comm from the Aether Catalog (Logic). The mathematical content is given by the Lean statement below; a human-readable write-up is pending.
theorem Logic.PhaseRoute.cov_comm(y h : ι → ℝ) : cov y h = cov h y := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Logic/PhaseRouteAlignment.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Logic/PhaseRouteAlignment.lean
import Mathlib
import Definitions.Def_Logic_PhaseRouteAlignment
import Definitions.Def_Logic_PhaseRouteLeastSquares
/-
# The phase route is closed, but the interaction route is open — exactly
This file proves a *separation theorem* for feature encodings, in the exact
finite-sample least-squares calculus of `Logic.PhaseRouteLeastSquares`.
Setting. The sample space is a product `α × β` (think: the residue of a root
position modulo one prime, paired with the residue modulo a second prime, or
paired with itself), and the target is an **alignment indicator**
`graphInd σ (a, b) = if b = σ a then 1 else 0` (`σ : α ≃ β`).
*Singleton* (linear-phase) encodings are the additive predictors
`additive u v (a, b) = u a + v b`, where `u`, `v` are **arbitrary** real
functions of a single coordinate — this is strictly more general than one-hot
dummies, sines/cosines of phases, or any other per-coordinate featurisation, and
it also covers arbitrarily many such features used simultaneously, since the
additive family is closed under sums.
Results.
* `cov_graphInd_additive_eq_zero` : every additive predictor has covariance
*exactly* `0` with the alignment target. There is no small residual signal to
be found: the linear phase route is closed identically, not approximately.
* `graphInd_additive_no_gain` / `graphInd_additive_strict_loss` : consequently no
additive predictor beats the intercept-only baseline, and every *nonconstant*
one is strictly worse; the excess error equals the predictor's own variance.
* `Rsq_additive_nonpos` : the same statement in `R²` units — the attainable gain
is `≤ 0`, never the pre-stated `+0.05`, and never even `+0.0215`.
* `nuisance_no_gain` : adding arbitrarily many features from an *independent*
block of coordinates (other primes / other windows) does not change this.
* `graphInd_eq_sum_interactions` : the alignment target is *exactly* a sum of
`card α` degree-2 products of one-hot singleton features, and
`Rsq_interaction_eq_one` : that degree-2 encoding attains `R² = 1`.
So the missing variance is not merely hard to reach with singleton phases; it
lives, provably and entirely, in the degree-2 (joint alignment) layer.
-/
open Logic.PhaseRoute
open Finset
/-! ### Bilinearity tools for the empirical covariance -/
variable {ι : Type*} [Fintype ι] [Nonempty ι]
omit [Nonempty ι] inFormal statement
theorem Logic.PhaseRoute.cov_comm(y h : ι → ℝ) : cov y h = cov h y := by sorry
Source