Products of Moore determinants of independent subspaces #
Let F be a finite field with 2 ^ n elements in which every proper subfield has at most
2 ^ s elements. Given 𝔽₂-subspaces Yᵢ of F of dimensions dᵢ > s and integers
2 ≤ kᵢ < dᵢ with ∑ᵢ (dᵢ - kᵢ) ≥ n + 1 - maxᵢ dᵢ, every nonzero 𝔽₂-linear functional
z ↦ Tr (μ z) is nonzero on some product ∏ᵢ Δ(uᵢ) of Moore determinants of families
uᵢ : Fin kᵢ → Yᵢ. This is the algebraic heart of the disperser for independent affine sources
of Ben-Sasson and Kopparty [BenSassonKopparty2012, Theorem 6.1], with the hypothesis that no
Yᵢ lies in a multiple of a proper subfield replaced by dᵢ > s; for a field of prime degree
s = 1.
The proof follows [BSK12]. Parametrize Yᵢ by its image-subspace polynomial
Qᵢ; then w ↦ Tr (μ ∏ᵢ Δ(Qᵢ ∘ wᵢ)) is a reduced polynomial function whose coefficient at
∏ᵢ ∏ₗ wᵢₗ ^ 2 ^ τᵢₗ is ∑_ρ μ ^ 2 ^ ρ ∏ᵢ per(Mᵢ[τᵢ - ρ]) ^ 2 ^ ρ, where
Mᵢ[τ] = (aᵢ (τ l - m) ^ 2 ^ m). Using the zero pattern of the coefficients aᵢ
(Cslib.CoeffPattern) we choose τ so that every Mᵢ[τᵢ] is triangular with
nonzero diagonal while for every rotation ρ ≠ 0 some Mᵢ[τᵢ - ρ] has a zero row
[BenSassonKopparty2012, Claims 6.2 and 6.3]. That coefficient is then μ ∏ᵢ per(Mᵢ[τᵢ]) ≠ 0,
and a reduced polynomial with a nonzero coefficient is a nonzero function
(FiniteField.eq_zero_of_forall_sum_mul_prod_pow_eq_zero).
Main results #
Cslib.AffineDisperser.exists_trace_mul_prod_det_mooreMatrix_ne_zero: the theorem above.
References #
Choosing the monomial #
Moore products of independent sources (after [BenSassonKopparty2012, Theorem 6.1]): in a
field with 2 ^ n elements whose proper subfields have at most 2 ^ s elements, let Yᵢ be
subspaces of dimensions dᵢ > s, and 2 ≤ kᵢ < dᵢ with n < d_{i₀} + ∑ᵢ (dᵢ - kᵢ) for some
i₀. Then for every μ ≠ 0 there are families uᵢ : Fin kᵢ → Yᵢ whose Moore determinants
satisfy Tr (μ ∏ᵢ Δ(uᵢ)) ≠ 0.
By sum_cube_pow_eq_det_mooreMatrix, Δ(uᵢ) is the kᵢ-th finite difference of
x ↦ x ^ (2 ^ kᵢ - 1) in the directions uᵢ. So the conclusion says that
(xᵢ) ↦ Tr (μ ∏ᵢ xᵢ ^ (2 ^ kᵢ - 1)) has a nonzero mixed finite difference along the Yᵢ, the
form in which the paper's theorem is used. The threshold n < maxᵢ dᵢ + ∑ᵢ (dᵢ - kᵢ) is one
better than the paper's ∑ᵢ (dᵢ - kᵢ) > n - maxᵢ dᵢ + 1.