Documentation

Complexitylib.Algebraic.LowerBound.Cutwidth.MultiOutput.TotallyRegular

Lower bounds for totally regular linear maps #

A matrix is totally regular when every square submatrix is nonsingular (TotallyRegular). Let a circuit over a finite field F, over any signature whose gates have at most two arguments (so any functions F × F → F, unary functions and constants are allowed), compute the linear map x ↦ M x for a totally regular N × N matrix M.

The finite bound (sub_one_le_of_totallyRegular). With the graph-ordering hypothesis for coefficient A, slack η and constant C, a circuit of size s has N - 1 ≤ (A + η) (s - N)⁺ + 3 log₂ (N + 3 s) + C.

The proof combines the two halves of the cut method. The ordering transfer (MultiOutput.exists_rank) ranks the wires so that every prefix is crossed by few signals, and the rank-cut bound (MultiOutput.blockRank_add_blockRank_le) shows that a prefix holding N of the 2 N inputs and outputs is crossed by at least rank M[Y_T, X_S] + rank M[Y_S, X_T] = |X_S| + |Y_S| = N signals, since every block of a totally regular matrix has full rank (TotallyRegular.min_card_le_blockRank). Inputs and outputs are counted separately: an output carried by an input wire counts twice, and as distinct outputs of a totally regular map lie on distinct wires the count grows by at most two per wire, so a prefix with N or N + 1 of them exists, which costs the - 1. All inputs and outputs lie in one component of the wire graph (its component is closed, hence uncrossed, so the rank-cut bound forces every input and output into it), and the prefix is charged to that component.

The asymptotic bound (eventually_lt_size_of_totallyRegular). With the edge-score ordering coefficient A = 2 κ_E, where κ_E = Gaussian.frontierCoefficient ≈ 0.14035, for every ε > 0 and all large N, every such circuit has more than (1 + 1/(2 κ_E) - ε) N gates, the coefficient being 1 + π/(3 arccos((1 + 2√2)/4)) ≈ 4.5625; as 2 κ_E ≤ 9/32, also more than (41/9 - ε) N gates (eventually_lt_size_of_totallyRegular_fortyOne_div_nine). The threshold depends only on ε, not on the field or the signature.

Rectangular totally regular maps (sub_one_le_of_totallyRegular_rect, eventually_lt_size_of_totallyRegular_rect). When M is an m × N totally regular matrix with N ≤ m (such as a systematic Reed–Solomon encoding matrix), distinct outputs still lie on distinct wires for N ≥ 2. Peeling the last gate of the program and at most one output wire seated on it m - N times reduces the circuit to one of size s - (m - N) computing an N × N totally regular row-submatrix of M, yielding the finite bound N - 1 ≤ (A + η) (s - m)⁺ + 3 log₂ (N + 3 s) + C and the asymptotic lower bound m + (1/(2 κ_E) - ε) N < s (hence m + (32/9 - ε) N < s, eventually_lt_size_of_totallyRegular_rect_thirtyTwo_div_nine).

An explicit family (totallyRegular_cauchy, eventually_lt_size_cauchyZMod). Cauchy matrices (1 / (x i - y j)) with distinct nodes are totally regular, as every square submatrix is again a nonsingular Cauchy matrix. Over ZMod q for a prime q ≥ 2 N, the nodes x i = i and y j = N + j give the explicit family cauchyZMod q N, and every fan-in-two circuit over ZMod q computing it has more than (41/9 - ε) N gates for large N (by Bertrand's postulate such a prime exists for every N ≥ 1).

Prior art. By a counting argument, the graph of a circuit computing a totally regular map is a superconcentrator, whatever its gate functions (Valiant), and size lower bounds for superconcentrators bound such circuits. The 5 N-edge lower bound of Lev and Valiant for superconcentrators is the edge-counting analogue of the cutwidth coefficient κ = 1/6, that is, of the ordering coefficient A = 1/3 (Multigraph.exists_orderingBound_one_third). Here gates are counted, and the ordering coefficient is 2 κ_E ≈ 0.2807.

Totally regular matrices #

theorem Algebraic.Cutwidth.MultiOutput.TotallyRegular.min_card_le_blockRank {F : Type u_1} [Field F] {m n : ℕ} {M : Matrix (Fin m) (Fin n) F} (hM : TotallyRegular M) (Y : Finset (Fin m)) (X : Finset (Fin n)) :

Blocks of a totally regular matrix have full rank: every block has rank at least the smaller of its dimensions.

theorem Algebraic.Cutwidth.MultiOutput.TotallyRegular.submatrix_rows {F : Type u_1} [Field F] {m m' n : ℕ} {M : Matrix (Fin m) (Fin n) F} (hM : TotallyRegular M) {r : Fin m' → Fin m} (hr : Function.Injective r) :

Row submatrices of a totally regular matrix are totally regular: any injective selection of rows of a totally regular matrix is totally regular.

theorem Algebraic.Cutwidth.MultiOutput.det_cauchy_ne_zero {F : Type u_1} [Field F] {k : ℕ} (x y : Fin k → F) (hx : Function.Injective x) (hy : Function.Injective y) (hxy : ∀ (i j : Fin k), x i ≠ y j) :
(cauchy x y).det ≠ 0

Square Cauchy matrices are nonsingular.

theorem Algebraic.Cutwidth.MultiOutput.totallyRegular_cauchy {F : Type u_1} [Field F] {m n : ℕ} (x : Fin m → F) (y : Fin n → F) (hx : Function.Injective x) (hy : Function.Injective y) (hxy : ∀ (i : Fin m) (j : Fin n), x i ≠ y j) :

Cauchy matrices are totally regular. With distinct nodes x i, distinct nodes y j, and x i ≠ y j for all i and j, every square submatrix of (1 / (x i - y j)) is nonsingular.

The explicit Cauchy family is totally regular over ZMod q for a prime q ≥ 2 N.

The finite bound #

theorem Algebraic.Cutwidth.MultiOutput.sub_one_le_of_totallyRegular {σ : Signature} {F : Type u_1} [Field F] [Fintype F] [DecidableEq F] {A η C : ℝ} (hAη : 0 ≤ A + η) (order : Multigraph.OrderingBound A η C) {N : ℕ} {M : Matrix (Fin N) (Fin N) F} (hM : TotallyRegular M) {I : Interpretation σ F} {c : Circuit σ N N} (hfan : c.FanInAtMost 2) (hc : c.Computes I fun (x : Fin N → F) => M.mulVec x) :
↑N - 1 ≤ (A + η) * max (↑c.size - ↑N) 0 + 3 * Real.logb 2 (↑N + 3 * ↑c.size) + C

The finite bound for totally regular maps. If a circuit over a finite field, over any signature with fan-in at most two, computes a totally regular linear map on N inputs, then under the graph-ordering hypothesis N - 1 ≤ (A + η) (s - N)⁺ + 3 log₂ (N + 3 s) + C for its size s.

theorem Algebraic.Cutwidth.MultiOutput.sub_one_le_of_totallyRegular_rect {σ : Signature} {F : Type u_1} [Field F] [Fintype F] [DecidableEq F] {A η C : ℝ} (hAη : 0 ≤ A + η) (order : Multigraph.OrderingBound A η C) {N m : ℕ} (hNm : N ≤ m) {M : Matrix (Fin m) (Fin N) F} (hM : TotallyRegular M) {I : Interpretation σ F} {c : Circuit σ N m} (hfan : c.FanInAtMost 2) (hc : c.Computes I fun (x : Fin N → F) => M.mulVec x) :
↑N - 1 ≤ (A + η) * max (↑c.size - ↑m) 0 + 3 * Real.logb 2 (↑N + 3 * ↑c.size) + C

The finite bound for rectangular totally regular maps. If a circuit over a finite field, over any signature with fan-in at most two, computes an m × N totally regular linear map with N ≤ m, then under the graph-ordering hypothesis N - 1 ≤ (A + η) (s - m)⁺ + 3 log₂ (N + 3 s) + C for its size s.

Asymptotic bounds #

theorem Algebraic.Cutwidth.MultiOutput.eventually_lt_size_of_totallyRegular_of_orderingBound {A : ℝ} (hA : 0 < A) (order : ∀ (η : ℝ), 0 < η → ∃ (C : ℝ), Multigraph.OrderingBound A η C) {ε : ℝ} (hε : 0 < ε) :
∀ᶠ (N : ℕ) in Filter.atTop, ∀ (F : Type u) [inst : Field F] [Fintype F] [DecidableEq F] (M : Matrix (Fin N) (Fin N) F), TotallyRegular M → ∀ (σ : Signature) (I : Interpretation σ F) (c : Circuit σ N N), c.FanInAtMost 2 → (c.Computes I fun (x : Fin N → F) => M.mulVec x) → (1 + 1 / A - ε) * ↑N < ↑c.size

The asymptotic bound with a general ordering coefficient. If the graph-ordering hypothesis holds with coefficient A > 0 for every positive slack, then for every ε > 0 and all large N, every fan-in-two circuit over a finite field computing a totally regular map on N inputs has more than (1 + 1/A - ε) N gates.

theorem Algebraic.Cutwidth.MultiOutput.eventually_lt_size_of_totallyRegular {ε : ℝ} (hε : 0 < ε) :
∀ᶠ (N : ℕ) in Filter.atTop, ∀ (F : Type u) [inst : Field F] [Fintype F] [DecidableEq F] (M : Matrix (Fin N) (Fin N) F), TotallyRegular M → ∀ (σ : Signature) (I : Interpretation σ F) (c : Circuit σ N N), c.FanInAtMost 2 → (c.Computes I fun (x : Fin N → F) => M.mulVec x) → (1 + 1 / (2 * Gaussian.frontierCoefficient) - ε) * ↑N < ↑c.size

Totally regular maps need (1 + 1/(2 κ_E) - ε) N gates. For every ε > 0 and all large N, every circuit over a finite field, over any signature with fan-in at most two, computing a totally regular linear map on N inputs has more than (1 + 1/(2 κ_E) - ε) N gates, where κ_E = Gaussian.frontierCoefficient; the coefficient is 1 + π/(3 arccos((1 + 2√2)/4)) ≈ 4.5625.

theorem Algebraic.Cutwidth.MultiOutput.eventually_lt_size_of_totallyRegular_fortyOne_div_nine {ε : ℝ} (hε : 0 < ε) :
∀ᶠ (N : ℕ) in Filter.atTop, ∀ (F : Type u) [inst : Field F] [Fintype F] [DecidableEq F] (M : Matrix (Fin N) (Fin N) F), TotallyRegular M → ∀ (σ : Signature) (I : Interpretation σ F) (c : Circuit σ N N), c.FanInAtMost 2 → (c.Computes I fun (x : Fin N → F) => M.mulVec x) → (41 / 9 - ε) * ↑N < ↑c.size

Totally regular maps need (41/9 - ε) N gates.

theorem Algebraic.Cutwidth.MultiOutput.eventually_lt_size_of_totallyRegular_rect_of_orderingBound {A : ℝ} (hA : 0 < A) (order : ∀ (η : ℝ), 0 < η → ∃ (C : ℝ), Multigraph.OrderingBound A η C) {ε : ℝ} (hε : 0 < ε) :
∀ᶠ (N : ℕ) in Filter.atTop, ∀ {m : ℕ}, N ≤ m → ∀ (F : Type u) [inst : Field F] [Fintype F] [DecidableEq F] (M : Matrix (Fin m) (Fin N) F), TotallyRegular M → ∀ (σ : Signature) (I : Interpretation σ F) (c : Circuit σ N m), c.FanInAtMost 2 → (c.Computes I fun (x : Fin N → F) => M.mulVec x) → ↑m + (1 / A - ε) * ↑N < ↑c.size

The asymptotic bound for rectangular totally regular maps with a general ordering coefficient. If the graph-ordering hypothesis holds with coefficient A > 0 for every positive slack, then for every ε > 0 and all large N, for every m ≥ N, every fan-in-two circuit over a finite field computing an m × N totally regular map has more than m + (1/A - ε) N gates.

theorem Algebraic.Cutwidth.MultiOutput.eventually_lt_size_of_totallyRegular_rect {ε : ℝ} (hε : 0 < ε) :
∀ᶠ (N : ℕ) in Filter.atTop, ∀ {m : ℕ}, N ≤ m → ∀ (F : Type u) [inst : Field F] [Fintype F] [DecidableEq F] (M : Matrix (Fin m) (Fin N) F), TotallyRegular M → ∀ (σ : Signature) (I : Interpretation σ F) (c : Circuit σ N m), c.FanInAtMost 2 → (c.Computes I fun (x : Fin N → F) => M.mulVec x) → ↑m + (1 / (2 * Gaussian.frontierCoefficient) - ε) * ↑N < ↑c.size

Rectangular totally regular maps need m + (1/(2 κ_E) - ε) N gates. For every ε > 0 and all large N, for every m ≥ N, every circuit over a finite field, over any signature with fan-in at most two, computing an m × N totally regular linear map has more than m + (1/(2 κ_E) - ε) N gates, where κ_E = Gaussian.frontierCoefficient.

theorem Algebraic.Cutwidth.MultiOutput.eventually_lt_size_of_totallyRegular_rect_thirtyTwo_div_nine {ε : ℝ} (hε : 0 < ε) :
∀ᶠ (N : ℕ) in Filter.atTop, ∀ {m : ℕ}, N ≤ m → ∀ (F : Type u) [inst : Field F] [Fintype F] [DecidableEq F] (M : Matrix (Fin m) (Fin N) F), TotallyRegular M → ∀ (σ : Signature) (I : Interpretation σ F) (c : Circuit σ N m), c.FanInAtMost 2 → (c.Computes I fun (x : Fin N → F) => M.mulVec x) → ↑m + (32 / 9 - ε) * ↑N < ↑c.size

Rectangular totally regular maps need m + (32/9 - ε) N gates.

theorem Algebraic.Cutwidth.MultiOutput.eventually_lt_size_cauchyZMod {ε : ℝ} (hε : 0 < ε) :
∀ᶠ (N : ℕ) in Filter.atTop, ∀ (q : ℕ) [Fact (Nat.Prime q)], 2 * N ≤ q → ∀ (σ : Signature) (I : Interpretation σ (ZMod q)) (c : Circuit σ N N), c.FanInAtMost 2 → (c.Computes I fun (x : Fin N → ZMod q) => (cauchyZMod q N).mulVec x) → (41 / 9 - ε) * ↑N < ↑c.size

The explicit family needs (41/9 - ε) N gates. For every ε > 0 and all large N, for every prime q ≥ 2 N, every circuit over ZMod q, over any signature with fan-in at most two, computing x ↦ cauchyZMod q N x has more than (41/9 - ε) N gates.