C7, the wall half.
ProvedForkPinning.kfactor_wallaether-catalogprobability
C7, the wall half. For N = p₁ ⋯ p_{k+1} with independent uniform Frobenius classes,
the class of N carries exactly zero information about any statistic of the first k
factors. No amount of factor structure leaks which prime is which.
theorem ForkPinning.kfactor_wall{k : ℕ} (F : (Fin k → G) → β) :
mutualInfo (fun v : Fin (k + 1) → G => vecProd v)
(fun v : Fin (k + 1) → G => F (fun i => v i.castSucc)) = 0 := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Probability/ForkPinningKFactors.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Probability/ForkPinningKFactors.lean
import Mathlib
import Definitions.Def_Probability_ForkPinningCore
import Definitions.Def_Probability_ForkPinningKFactors
/-
# The which-factor wall for a product of `k` primes
`ForkPinningSemiprimeGeneral` proves that for a semiprime `N = p q` the class of `N` in a finite
group `G` is *exactly* independent of every statistic of the first factor. This file closes the
"wall" half of conjecture **C7** of `FUTURE_DIRECTIONS.md`: the same is true for a product of
arbitrarily many primes, and in the strongest possible form.
The key structural fact is that the class map is a group homomorphism with *uniform fibres in the
last coordinate*: whatever the classes of the first `k` primes are, the class of the last one is
free, so the class of the product is uniform **conditionally on everything else**. We isolate
this as `ForkPinning.last_factor_wall`, which is stated for an arbitrary aggregate
`h : Ω → G` of the other factors, and then instantiate it at `Ω = Fin k → G`.
Main results:
* `ForkPinning.mutualInfo_congr_equiv` : mutual information is invariant under a relabelling of
the sample space (a reusable transport lemma).
* `ForkPinning.prb_lastMul_uniform` : the class of the product is exactly uniform.
* `ForkPinning.last_factor_wall` : `I( h(u)·v ; F(u) ) = 0` for every aggregate `h` and every
statistic `F` of the remaining data — the wall in its general form.
* `ForkPinning.kfactor_wall` : for `N = p₁ ⋯ p_{k+1}` with independent uniform classes, the class
of `N` carries **exactly zero** information about any statistic of `p₁, …, p_k`.
* `ForkPinning.kfactor_wall_which_splits` : in particular it says nothing about *which* of the
first `k` primes split, and `kfactor_class_uniform` : the class of `N` is uniform, so it says
nothing at all in isolation either.
-/
open ForkPinning
open Finset Real
/-! ## Transport of the information functionals along a relabelling of the sample space -/
variable {Ω Ω' : Type*} [Fintype Ω] [Nonempty Ω] [DecidableEq Ω]
variable [Fintype Ω'] [Nonempty Ω'] [DecidableEq Ω']
variable {κ β : Type*} [Fintype κ] [DecidableEq κ] [Fintype β] [DecidableEq β]
/-! ## The wall in its general form -/
variable {Ω : Type*} [Fintype Ω] [Nonempty Ω] [DecidableEq Ω]
variable {G : Type*} [Group G] [Fintype G] [DecidableEq G]
variable {β : Type*} [Fintype β] [DecidableEq β]
/-! ## `k`-fold products -/
variable {G : Type*} [Group G] [Fintype G] [Nonempty G] [DecidableEq G]
variable {β : Type*} [Fintype β] [DecidableEq β]Formal statement
theorem ForkPinning.kfactor_wall{k : ℕ} (F : (Fin k → G) → β) :
mutualInfo (fun v : Fin (k + 1) → G => vecProd v)
(fun v : Fin (k + 1) → G => F (fun i => v i.castSucc)) = 0 := by sorrySource