Cross-population wall stability (corrected form).
ProvedWhichFactorWall.imbalance_dist_leaether-catalogalgebra
Cross-population wall stability (corrected form). Two imbalances in the
guarded range [0, 1/2 - η] are controlled by the gap of their walls, with the
explicit constant log ((1/2+η)/(1/2-η)), which is the slope of binEntropy at
the guard point — not at δ.
theorem WhichFactorWall.imbalance_dist_le{η p q : ℝ} (hη : 0 < η) (hη2 : η < 2⁻¹)
(hp : p ∈ Icc (0 : ℝ) (2⁻¹ - η)) (hq : q ∈ Icc (0 : ℝ) (2⁻¹ - η)) :
(log (2⁻¹ + η) - log (2⁻¹ - η)) * |p - q| ≤ |binEntropy p - binEntropy q| := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Algebra/WhichFactorWallInvariant.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Algebra/WhichFactorWallInvariant.lean
import Mathlib
import Definitions.Def_Algebra_WhichFactorWallInvariant
/-
# The which-factor wall as a cross-population invariant: how much does it really pin down?
This file continues `Speculative.AutoResearch.TraceBatteryWall`
(`TraceBattery.binary_wall_inversion`), which shows that a *binary* capacity
("wall") value determines the class imbalance of a two-valued statistic
**uniquely** on the balanced side `[0, 1/2]`. Uniqueness, however, is a
qualitative statement. The research question of this cycle is quantitative:
> if two independent populations report walls that agree to within `ε`,
> how close are their class imbalances?
The mission proposal was:
`|binEntropy p - binEntropy q| ≥ c(δ) |p - q|` on `[δ, 1/2]`,
with `c(δ) = log ((1-δ)/δ)`.
**This is false**, and `binEntropy_conjectured_lower_bound_false` refutes it with
an exact counterexample (`δ = q = 1/4`, `p = 1/2`), where the failure reduces to
`log 16 ≤ log 27`. The reason is structural: `c(δ)` is the *supremum* of
`|binEntropy'|` on `[δ, 1/2]`, not its infimum, so it controls the Lipschitz
(upper) bound, while the true inverse bound must be governed by the derivative
at the endpoint *closest to* `1/2`.
What survives, and is proved here with zero sorries:
* `binEntropy_sub_ge` — the sharp mean-value lower bound
`(q - p) * (log (1-q) - log q) ≤ binEntropy q - binEntropy p` for
`0 ≤ p ≤ q ≤ 1/2` (including the boundary case `p = 0`).
* `binEntropy_lipschitz` — the true version of the proposed inequality, with the
inequality reversed: `|binEntropy p - binEntropy q| ≤ c(δ) |p - q|` on
`[δ, 1-δ]`.
* `imbalance_dist_le` / `imbalance_dist_le_div` — the corrected **cross-population
stability theorem**: imbalances in `[0, 1/2 - η]` whose walls agree within `ε`
agree within `ε / log ((1/2+η)/(1/2-η))`.
* `binary_wall_stability` — the same statement at the level of two binary
statistics on two different finite populations, via the empirical entropy `H`.
* `log_two_sub_binEntropy_le_sq` — `log 2 - binEntropy (1/2 - t) ≤ 4 t²`, and
`no_uniform_inversion_constant`: **no** constant inverts the wall near `1/2`.
So the guard `η > 0` is not an artefact: the wall genuinely loses all
resolution at balance, at a quadratic rate.
* `wall_imbalance_bracket` — the reported wall `0.4677` bits is a falsifiable
claim about the split: the unique minority fraction in `[0, 1/2]` realising it
lies strictly between `1/12` and `1/9` (i.e. between 8.34% and 11.11%),
consistent with the reported 9.96% and inconsistent with, say, a 5% or a 15%
split.
Because the catalog module `Combinatorics.TraceBatteryEntropy` carrying the
empirical-entropy definitions is not present in this snapshot, the small
population layer (`img`, `cnt`, `H`, `H_two_values`) is restated here in the
open `WhichFactorWall`,
self-contained and compiles on its own.
-/
open WhichFactorWall
open Real Set
/-! ## 1. Mean-value machinery for `Real.binEntropy`
`binEntropy` is differentiable away from `{0,1}` with derivative
`log (1-x) - log x` (Mathlib's `Real.deriv_binEntropy`). We turn one-sided
bounds on that derivative into slope bounds by monotonicity of an auxiliary
function; this is the mean value theorem in the form we need. -/
/-! ## 2. Refutation of the proposed inverse bound -/
/-! ## 3. Corrected cross-population stability -/Formal statement
theorem WhichFactorWall.imbalance_dist_le{η p q : ℝ} (hη : 0 < η) (hη2 : η < 2⁻¹)
(hp : p ∈ Icc (0 : ℝ) (2⁻¹ - η)) (hq : q ∈ Icc (0 : ℝ) (2⁻¹ - η)) :
(log (2⁻¹ + η) - log (2⁻¹ - η)) * |p - q| ≤ |binEntropy p - binEntropy q| := by sorrySource