mme_laser_per_N_witness
DisprovedPer-(N, ) polynomial-bounded laser-method witness at a fixed probability distribution . For a -way mode-graded 3-tensor with cyclic-symmetric laser support , laser-aligned grading , and probability distribution supported on , let be the closed-form Wigderson--Zuiddam laser value at . Then and there exists a polynomial degree such that for every , frequently many admit with , a from into , and the H"older lower bound . This is the per- analytic-combinatorial core of the WZ §6 / CW 1990 §5--§7 construction at a single , absorbing the six Layer-4 leaves (3AP-free, Salem--Spencer, graded tensor power decomposition, independent direct-sum restriction, block-tensor MM identification, dimension count) plus the Stirling/H"older final optimization.
import Mathlib.Algebra.BigOperators.Group.Finset.Basic import Mathlib.Analysis.SpecialFunctions.Log.Basic import Mathlib.Analysis.SpecialFunctions.Pow.Real import Mathlib.LinearAlgebra.FiniteDimensional.Basic import Mathlib.Order.Filter.AtTopBot.Defs import Definitions.Def_mme_laser_pattern import Definitions.Def_mme_subrank_capacity_poly import Definitions.Def_mme_tensor_type_grading import Definitions.Def_mme_tensor_rank open MME BigOperators Filter universe u
/-- **Per-(N, ε) polynomial-bounded laser-method witness at a fixed
probability distribution π.**
For a `t`-way mode-graded 3-tensor `T : TensorObj K 3` with cyclic-symmetric
laser support `S ⊆ (Fin t)^3`, laser-aligned grading `G`, and a probability
distribution `π : (Fin t × Fin t × Fin t) → ℝ` supported on `S`: the
closed-form Wigderson–Zuiddam laser value at `π`,
`V(π) = exp(log 2 · ( H(π) + (1/3) · Σ_σ π(σ) · log₂(d₀(σ)·d₁(σ)·d₂(σ)) ))`
admits the polynomial-bounded witness construction defining
`subrankCapacityPoly T`. Concretely:
* `V(π) ≥ 1` (since `H(π) ≥ 0` and the type-dimensions are positive
integers, so the log-arg is `≥ 1`); and
* there is a polynomial degree `c` (depending on `π, G, S`) such that for
every `ε > 0` and frequently many `N`, one can exhibit `(k, a, b, c')`
with `k ≤ (N+1)^c`, a `Restrict` from `⨁ᵢ MMObj K (aᵢ bᵢ c'ᵢ)` into
`T^{⊗N}`, and the Hölder-controlled lower bound
`V(π)^N · (1 - ε) ≤ ∑ᵢ (aᵢ·bᵢ·c'ᵢ)^{1/3}`.
This is the **direct unwrapping** of the WZ §6 / CW 1990 §5–§7 construction
at a single `π`. Conceptually it bundles the six Layer-4 leaves
(`mme_3AP_free_no_collision`, `mme_salem_spencer_eps_form`,
`mme_graded_tensor_pow_block_decomp`,
`mme_independent_blocks_form_direct_sum_restrict_enum`,
`mme_laser_block_dimension_count_refined`,
`mme_block_tensor_is_matMul_kronPow_balanced`) plus the Stirling / Hölder
final optimization step.
**Status.** Open. Decomposes the value bound at one π into the per-N
construction, which is the deep analytic core.
**Reduction role.** Together with the paper-agnostic abstract bridge
`mme_subrankCapacityPoly_ge_of_witnesses`, this leaf proves the parent
`mme_laser_witness_at_prob_dist`. -/
theorem mme_laser_per_N_witness
{K : Type u} [Field K] {T : TensorObj K 3} {t : ℕ}
(G : T.TypeGrading t) (S : Finset (Fin t × Fin t × Fin t))
(_hSym : LaserSymmetric S)
(_hsupport : TensorObj.LaserAlignedSupport G S)
(π : (Fin t × Fin t × Fin t) → ℝ)
(_hπ_supp : ∀ σ, σ ∉ S → π σ = 0)
(_hπ_nn : ∀ σ, 0 ≤ π σ)
(_hπ_sum : (∑ σ ∈ S, π σ) = 1) :
let V : ℝ :=
Real.exp (Real.log 2 *
( (-(∑ σ ∈ S, π σ * (Real.log (π σ) / Real.log 2)))
+ (1 / 3) *
(∑ σ ∈ S, π σ *
(Real.log
(((Module.finrank K (G.classOf 0 σ.1) : ℕ) : ℝ) *
((Module.finrank K (G.classOf 1 σ.2.1) : ℕ) : ℝ) *
((Module.finrank K (G.classOf 2 σ.2.2) : ℕ) : ℝ))
/ Real.log 2)) ) )
(1 ≤ V) ∧ ∃ c : ℝ,
∀ ε > (0 : ℝ), ∃ᶠ (N : ℕ) in atTop,
∃ (k : ℕ) (a b c' : Fin k → ℕ),
(k : ℝ) ≤ ((N : ℝ) + 1) ^ c ∧
TensorObj.Restrict
(TensorObj.bigAdd (fun i => MMObj K (a i) (b i) (c' i)))
(T.kronPow N)
∧ V ^ N * (1 - ε) ≤ ∑ i, ((a i * b i * c' i : ℕ) : ℝ) ^ ((1 : ℝ) / 3) := by
sorry