Introduction
In [21 ] Hrubeš discovered a family of reductions showing that sufficiently strong monotone lower bounds imply unconditional lower bounds for general, non-monotone computation: on arithmetic circuit size, on Boolean circuit size, and — for a proof system whose proofs are monotone circuits and which polynomially simulates Resolution — on proof size. All of them ask for ϵ \epsilon ϵ -sensitive lower bounds: one bounds the monotone complexity not of a single hard object but of a perturbation of an easy one, the perturbation governed by a parameter ϵ > 0 \epsilon>0 ϵ > 0 tending to 0 0 0 . In the two instances this paper is concerned with, the arithmetic and the proof-complexity one, the perturbation is a difference
F − ϵ ⋅ G F-\epsilon\cdot G F − ϵ ⋅ G
with F F F alone monotonically easy. They differ in where the hardness resides and in the quantification over ϵ \epsilon ϵ : in the arithmetic instance F F F is fixed and easy while G G G is arbitrary and carries the hardness, and it suffices to have monotone lower bounds along some sequence of ϵ \epsilon ϵ tending to 0 0 0 ; in the proof complexity instance both F F F and G G G are monotone-easy, and the lower bound is required for every ϵ > 0 \epsilon> 0 ϵ > 0 , since a proof may use any single ϵ \epsilon ϵ . We briefly describe the two instances.
General arithmetic circuits. The arithmetic instance of the reduction is [21 Theorem 2.1] : if a polynomial f ∈ R [ x 1 , … , x n ] f \in\mathbb{R}[x_1,\ldots,x_n] f ∈ R [ x 1 , … , x n ] of degree d d d has an arithmetic circuit of size s s s , then for every sufficiently small ϵ > 0 \epsilon> 0 ϵ > 0 the polynomial
g ϵ : = ( x 1 + ⋯ + x n + 1 ) d + ϵ ⋅ f g_\epsilon:= (x_1+\cdots+x_n+1)^d+\epsilon\cdot f g ϵ := ( x 1 + ⋯ + x n + 1 ) d + ϵ ⋅ f
has a monotone arithmetic circuit of size O ( s d 2 + n log n ) O(sd^2+n\log n) O ( s d 2 + n log n ) . This is F − ϵ G F-\epsilon G F − ϵ G with F F F the universal polynomial U : = ( x 1 + ⋯ + x n + 1 ) d U := (x_1+\cdots+x_n+1)^d U := ( x 1 + ⋯ + x n + 1 ) d and G = − f G=-f G = − f an arbitrary, possibly non-monotone polynomial; since U U U contains every monomial of degree at most d d d with coefficient at least 1 1 1 , once ϵ \epsilon ϵ is small enough for these coefficients to absorb those of ϵ f \epsilon f ϵ f , every coefficient of g ϵ g_\epsilon g ϵ is indeed nonnegative, so g ϵ g_\epsilon g ϵ is a legitimate monotone target. Up to the factor d 2 d^2 d 2 the reduction is an exact equivalence: the identity f = ( g ϵ − g 0 ) / ϵ f=(g_\epsilon-g_0)/\epsilon f = ( g ϵ − g 0 ) / ϵ turns a monotone circuit for g ϵ g_\epsilon g ϵ at any single ϵ > 0 \epsilon>0 ϵ > 0 into a general circuit for f f f of essentially the same size, so the monotone size of g ϵ g_\epsilon g ϵ and the general circuit size of f f f coincide for all sufficiently small ϵ \epsilon ϵ , up to the factor d 2 d^2 d 2 . In particular, monotone lower bounds on g ϵ i g_{\epsilon_i} g ϵ i along any sequence ϵ i → 0 \epsilon_i\to0 ϵ i → 0 yield an unconditional lower bound for f f f , one of the ϵ i \epsilon_i ϵ i eventually falling below the relevant threshold; a bound at a single fixed ϵ \epsilon ϵ does not suffice, since the easiness of g ϵ g_\epsilon g ϵ for easy f f f is guaranteed only below an ϵ 0 \epsilon_0 ϵ 0 depending on the coefficients of f f f , and ϵ 0 \epsilon_0 ϵ 0 is not known in advance. The identity also shows that the required lower bounds are genuine: if f f f is hard, then g ϵ g_\epsilon g ϵ is monotone-hard at every ϵ > 0 \epsilon>0 ϵ > 0 . An equivalent formulation, via the standard observation that every circuit of size s s s can be simulated with a single subtraction as f = f + − f − f=f^+-f^- f = f + − f − with f ± f^\pm f ± monotone of monotone circuit size O ( s ) O(s) O ( s ) : the negative part f − f^- f − can always be taken to be a scalar multiple of the fixed universal polynomial U U U .
The measure ϵ ↦ size + ( g ϵ ) \epsilon\mapsto\operatorname{size}_+(g_\epsilon) ϵ ↦ size + ( g ϵ ) is thus highly discontinuous at ϵ = 0 \epsilon=0 ϵ = 0 : g 0 = U g_0=U g 0 = U has monotone circuits of size O ( n + log d ) O(n+\log d) O ( n + log d ) , yet whenever f f f is hard the monotone size of g ϵ g_\epsilon g ϵ jumps, for every ϵ > 0 \epsilon>0 ϵ > 0 , to essentially the general circuit size of f f f —an arbitrarily small perturbation of the easy polynomial U U U blows up its monotone complexity, and the hardness is created entirely by the ϵ \epsilon ϵ -perturbation. This discontinuity is both the power and the price of the approach. The power: it converts a lower bound in the monotone model — a restricted setting in which strong unconditional lower bounds are known — into a lower bound for general arithmetic circuits, where no such techniques exist. The price: most classical monotone lower bound methods are inapplicable here, as they depend only on the set of monomials that occur — the exponential monotone lower bound for the permanent of Jerrum and Snir [23 ] , for instance, is of this kind — but supp ( g ϵ ) = supp ( U ) \operatorname{supp}(g_\epsilon)=\operatorname{supp}(U) supp ( g ϵ ) = supp ( U ) is the full set of monomials of degree at most d d d , the support of an easy polynomial. An ϵ \epsilon ϵ -sensitive lower bound must therefore be genuinely coefficient-sensitive and, moreover, discontinuous in the coefficients.
Previous ϵ \epsilon ϵ -sensitive lower bounds use communication complexity for monotone arithmetic circuits [7 , 6 ] and extend to separations among monotone circuits, branching programs, formulas and bounded-depth formulas [8 ] . Here we study the tautology encodings arising from AMC.
The proof complexity setting: Algebraic Monotone Calculus. Hrubeš [21 ] applied the same idea to proof complexity , defining the Algebraic Monotone Calculus (AMC). The lines of AMC are formal inequalities f ⪯ g f\preceq g f ⪯ g between monotone polynomials, meaning that every monomial occurring in f f f occurs also in g g g , that is, the support containment supp ( f ) ⊆ supp ( g ) \operatorname{supp}(f)\subseteq\operatorname{supp}(g) supp ( f ) ⊆ supp ( g ) ; accordingly the axioms are f ⪯ f f \preceq f f ⪯ f and a ⪯ b a \preceq b a ⪯ b for reals a ≥ 0 a \ge0 a ≥ 0 , b > 0 b > 0 b > 0 , and the rules are transitivity together with compatibility with the two operations, from f 1 ⪯ g 1 f_1 \preceq g_1 f 1 ⪯ g 1 and f 2 ⪯ g 2 f_2 \preceq g_2 f 2 ⪯ g 2 infer f 1 ∘ f 2 ⪯ g 1 ∘ g 2 f_1 \circ f_2 \preceq g_1 \circ g_2 f 1 ∘ f 2 ⪯ g 1 ∘ g 2 for ∘ ∈ { + , × } \circ\in\{+, \times\} ∘ ∈ { + , × } . The size of a proof f 1 ⪯ g 1 , … , f m ⪯ g m f_1 \preceq g_1,\ldots,f_m \preceq g_m f 1 ⪯ g 1 , … , f m ⪯ g m is the size of a smallest monotone arithmetic circuit computing all the polynomials f 1 , g 1 , … , f m , g m f_1,g_1,\ldots,f_m,g_m f 1 , g 1 , … , f m , g m simultaneously. Nothing so far mentions ϵ \epsilon ϵ : the lines assert only support containment. The parameter enters through the semantics, since supp ( f ) ⊆ supp ( g ) \operatorname{supp}(f) \subseteq\operatorname{supp}(g) supp ( f ) ⊆ supp ( g ) holds iff g − ϵ ⋅ f g-\epsilon\cdot f g − ϵ ⋅ f is monotone for some ϵ > 0 \epsilon> 0 ϵ > 0 iff f ⪯ g f \preceq g f ⪯ g is provable. Quantitatively, the AMC proof size of f ⪯ g f \preceq g f ⪯ g equals, up to a constant factor, the minimum size of a monotone circuit computing f f f , g g g and g − ϵ ⋅ f g-\epsilon\cdot f g − ϵ ⋅ f simultaneously, over all ϵ > 0 \epsilon> 0 ϵ > 0 [21 Proposition 2.6, Theorem 2.7] .
For a DNF A A A , the coefficient of the set-multilinear monomial x σ = ∏ i x i σ ( i ) x^\sigma=\prod_i x_i^{\sigma(i)} x σ = ∏ i x i σ ( i ) in its characteristic polynomial χ ( A ) \chi(A) χ ( A ) counts the terms of A A A satisfied by σ \sigma σ . Here ( x i 0 , x i 1 ) (x_i^0,x_i^1) ( x i 0 , x i 1 ) is the variable pair for the Boolean variable x i x_i x i , while χ ( 1 ) = ∏ i ( x i 0 + x i 1 ) \chi(1)=\prod_i(x_i^0+x_i^1) χ ( 1 ) = ∏ i ( x i 0 + x i 1 ) collects all such monomials, one per assignment; thus A A A is a tautology precisely when supp ( χ ( 1 ) ) ⊆ supp ( χ ( A ) ) \operatorname{supp}(\chi(1)) \subseteq\operatorname{supp}(\chi(A)) supp ( χ ( 1 )) ⊆ supp ( χ ( A )) , and a proof of A A A is an AMC proof of χ ( 1 ) ⪯ χ ( A ) \chi(1) \preceq\chi(A) χ ( 1 ) ⪯ χ ( A ) . A proof of size s s s gives a monotone circuit of size O ( s ) O(s) O ( s ) for χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) for some ϵ > 0 \epsilon> 0 ϵ > 0 . Conversely, a monotone circuit of size s s s for this difference gives a proof of size O ( s + m n ) O(s+mn) O ( s + mn ) , since χ ( A ) \chi(A) χ ( A ) and χ ( 1 ) \chi(1) χ ( 1 ) have monotone circuits of size O ( m n ) O(mn) O ( mn ) when A A A has m m m terms: both are monotone-easy, and all the difficulty is in the perturbation. Thus a lower bound on the AMC complexity of A A A reduces to a monotone circuit lower bound for χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) for all ϵ > 0 \epsilon> 0 ϵ > 0 . Since χ ( A ) \chi(A) χ ( A ) and χ ( 1 ) \chi(1) χ ( 1 ) , and hence their difference, have small general arithmetic circuits, such a bound exhibits polynomials χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) in VP with no polynomial-size monotone circuit. Separations of monotone VP from VP as such are classical [27 ] ; what is demanded here is that the bound hold for every ϵ > 0 \epsilon> 0 ϵ > 0 , for targets whose support, for every sufficiently small positive ϵ \epsilon ϵ , is that of the easy polynomial χ ( 1 ) \chi(1) χ ( 1 ) .
On the proof side such a bound has unconditional consequences, because AMC is a strong system: it polynomially simulates Resolution — a Resolution refutation with k k k lines yields an AMC proof of size O ( ( m + k ) n ) O((m+k)n) O (( m + k ) n ) [21 Proposition 5.8] — so an AMC lower bound, and hence a monotone lower bound on χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) for every ϵ > 0 \epsilon> 0 ϵ > 0 , implies a Resolution size lower bound. As we show in § 3 , considerably more is true already for the weakest fragment considered here: it simulates Sherali–Adams, and therefore Resolution, when size counts displayed conjunctions. A distinctive feature of the proof complexity setting is that here the “positive part” F = χ ( A ) F=\chi(A) F = χ ( A ) varies with the instance while G = χ ( 1 ) G=\chi(1) G = χ ( 1 ) is fixed, whereas in the circuit-complexity applications of the reduction it is the other way around: F F F is a fixed, simple polynomial (such as a power ( 1 + ∑ i x i ) d (1+\sum_i x_i)^d ( 1 + ∑ i x i ) d ) and the varying, hardness-carrying part is G G G . The interest of AMC as a proof system is that it may be much stronger than Resolution, and here the relation to the monotone calculus MLK of Atserias, Galesi and Pudlák [1 ] is the relevant benchmark. MLK proves implications A → B A \to B A → B between monotone Boolean formulas, and an LK proof of size s s s of a monotone sequent can be turned into one containing only monotone formulas, of size s O ( log s ) s^{O(\log s)} s O ( l o g s ) and with polynomially many lines [1 ] ; on monotone sequents MLK therefore quasipolynomially simulates Frege. Replacing ∧ , ∨ \land,\lor ∧ , ∨ in a monotone circuit A A A by × , + \times,+ × , + yields a monotone polynomial F A F_A F A , and supp ( F A ) ⊆ supp ( F B ) \operatorname{supp}(F_A) \subseteq\operatorname{supp}(F_B) supp ( F A ) ⊆ supp ( F B ) implies that A → B A \to B A → B is a tautology; the converse fails — x ∧ y → y x \land y \to y x ∧ y → y is a tautology while supp ( F x ∧ y ) = { x y } \operatorname{supp}(F_{x\land y})=\{xy\} supp ( F x ∧ y ) = { x y } and supp ( F y ) = { y } \operatorname{supp}(F_y)=\{y\} supp ( F y ) = { y } — so AMC is, over such translations, a genuine weakening of MLK: it is blind to the Boolean identities x 2 = x x^2=x x 2 = x and x ≤ 1 x \le1 x ≤ 1 . Hrubeš [21 ] indicates — as a remark, without proof — how Frege could be reached in principle by restoring exactly these: AMC augmented with the Boolean axioms f ⪯ 1 f \preceq1 f ⪯ 1 and f ⪯ f 2 f \preceq f^2 f ⪯ f 2 simulates MLK. Whether the Boolean axioms actually help, and whether AMC itself polynomially simulates Frege, are open — both under this translation and under the encoding used in the rest of the paper, where a proof of a DNF tautology A A A is a proof of χ ( 1 ) ⪯ χ ( A ) \chi(1) \preceq\chi(A) χ ( 1 ) ⪯ χ ( A ) .
Nonnegative rank and extension complexity. The nonnegative rank of a matrix M ∈ R ≥ 0 m × n M \in\mathbb{R}_{\ge0}^{m \times n} M ∈ R ≥ 0 m × n with nonnegative entries is the smallest r r r such that M M M can be written as a sum of r r r nonnegative rank-one matrices, equivalently M = U V M = UV M = U V for nonnegative matrices U ∈ R ≥ 0 m × r U \in\mathbb{R}_{\geq0}^{m \times r} U ∈ R ≥ 0 m × r and V ∈ R ≥ 0 r × n V \in\mathbb{R}_{\geq0}^{r \times n} V ∈ R ≥ 0 r × n . The relevance of this measure to optimization was established in the seminal work of Yannakakis [29 ] , who showed that the smallest number of facets of any linear program projecting onto a polytope P P P — the extension complexity of P P P — equals the nonnegative rank of its slack matrix. Lower bounds on extension complexity therefore rule out small linear programming relaxations, and proving them amounts to bounding the nonnegative rank of an explicit matrix from below. After a long hiatus this program produced the first superpolynomial bounds, for the correlation, cut and travelling salesman polytopes [15 ] , and culminated in the breakthrough of Rothvoß [26 ] , who proved that the perfect matching polytope has exponential extension complexity; Braun, Fiorini, Pokutta and Steurer [3 ] and Braun and Pokutta [4 ] extended such bounds to approximate extended formulations.
The same measure is the common currency of several lower bound programs. In circuit complexity, Hrubeš [21 ] attaches to every Boolean function f f f an explicit nonnegative matrix M ( f ) M(f) M ( f ) , with rows indexed by the 1-inputs of f f f and columns by its 0-inputs, and shows that R ( f ) : = min ϵ > 0 rk + ( M ( f ) − ϵ J ) R(f) := \min_{\epsilon> 0} \operatorname{rk}_{+}(M(f) - \epsilon J) R ( f ) := min ϵ > 0 rk + ( M ( f ) − ϵ J ) , with J J J the all-ones matrix, satisfies R ( f ) = O ( C ( f ) + n ) R(f) = O(C(f) + n) R ( f ) = O ( C ( f ) + n ) for the Boolean circuit complexity C ( f ) C(f) C ( f ) [21 Thm. 2.3] : a superpolynomial lower bound on R ( f ) R(f) R ( f ) is a superpolynomial circuit lower bound. The same quantity measures the cost of separating f − 1 ( 1 ) f^{-1}(1) f − 1 ( 1 ) from f − 1 ( 0 ) f^{-1}(0) f − 1 ( 0 ) by a linear program, to within an additive term linear in n n n : writing sep ( f ) \operatorname{sep}(f) sep ( f ) for the least extension complexity of a separating polyhedron, the linear separation complexity , R ( f ) − 2 n − 1 ≤ sep ( f ) ≤ R ( f ) R(f) - 2n - 1 \leq\operatorname{sep}(f) \leq R(f) R ( f ) − 2 n − 1 ≤ sep ( f ) ≤ R ( f ) [21 Thm. 2.5] . For monotone f f f , the corresponding monotone separation complexity sep + ( f ) \operatorname{sep}_{+}(f) sep + ( f ) uses monotone separating polyhedra and the matrix M + ( f ) y , x = ∑ i x i ( 1 − y i ) M_{+}(f)_{y,x} = \sum_i x_i(1-y_i) M + ( f ) y , x = ∑ i x i ( 1 − y i ) , with x ∈ f − 1 ( 1 ) x \in f^{-1}(1) x ∈ f − 1 ( 1 ) and y ∈ f − 1 ( 0 ) y \in f^{-1}(0) y ∈ f − 1 ( 0 ) [21 Thm. 3.16] . Up to polynomial factors, this monotone quantity is the weak monotone linear programming gate complexity of de Oliveira Oliveira and Pudlák [11 ] . An explicit monotone function with superpolynomial sep + \operatorname{sep}_{+} sep + is sought in [21 Open Problem 3] and [11 Open Problem 1] ; nonconstructive lower bounds for separation complexity are known [21 , 22 ] . In formula complexity, splitting the variable blocks of a monotone set-multilinear polynomial into two groups and reading its coefficients as a matrix sends every product term of a monotone Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ formula to a nonnegative rank-one matrix, so the nonnegative rank of that matrix bounds the number of terms from below. Moreover ⌈ log 2 rk + ( M ) ⌉ \lceil\log_2 \operatorname{rk}_{+}(M) \rceil ⌈ log 2 rk + ( M )⌉ is exactly the randomized communication complexity of computing the entries of the nonnegative matrix M M M in expectation [13 ] . Information-theoretic estimates enter more directly: a nonnegative factorization of M M M with r r r terms, read as a mixture of r r r product distributions on the normalised matrix, bounds log r \log r log r from below by Wyner’s common information [28 ] , the route taken by Braun and Pokutta [4 ] . In proof complexity the measure is carried by Sherali–Adams: by Fleming, Göös, Grosser and Robere [16 ] a refutation of an unsatisfiable CNF C C C normalises to a single conical junta viol C − ϵ = ∑ j η j Q j \operatorname{viol}_C - \epsilon= \sum_j \eta_j Q_j viol C − ϵ = ∑ j η j Q j , with η j > 0 \eta_j > 0 η j > 0 and the Q j Q_j Q j conjunctions, where viol C ( σ ) \operatorname{viol}_C(\sigma) viol C ( σ ) is the number of clauses of C C C falsified by σ \sigma σ — so that degree becomes width, size becomes a number of subcubes, which is again a nonnegative rank of viol C − ϵ \operatorname{viol}_C - \epsilon viol C − ϵ with its rank-one terms constrained to that shape, and ϵ > 0 \epsilon> 0 ϵ > 0 is quantified existentially; and [11 ] use extension complexity in the same spirit to prove monotone feasible interpolation for Lovász–Schrijver and its mixed variant.
Our contributions
In this paper we study the strength of bounded-depth formula fragments of AMC using nonnegative rank measures. We obtain the following results.
Depth three: Sigma Pi Sigma-AMC, Sherali–Adams and size bounds
At depth three the minimum circuit and formula sizes differ by at most a factor O ( n ) O(n) O ( n ) , so we take Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC proofs to be formulas and measure their size by the number of product terms. Let $f be monotone, homogeneous and set-multilinear over blocks X 1 , … , X k X_1,\ldots,X_k X 1 , … , X k , and put n : = ∑ i ∣ X i ∣ n := \sum_i |X_i| n := ∑ i ∣ X i ∣ . A monotone Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ formula computing f f f is then set-multilinear itself — monotonicity is what forces this, a formula with cancellation being free to compute a set-multilinear polynomial through gates that are not set-multilinear — so it is a nonnegative tensor decomposition f = ∑ i ≤ r ℓ 1 , i ( X 1 ) ⊗ ⋯ ⊗ ℓ k , i ( X k ) f = \sum_{i\leq r} \ell_{1,i}(X_1)\otimes\cdots\otimes\ell_{k,i}(X_k) f = ∑ i ≤ r ℓ 1 , i ( X 1 ) ⊗ ⋯ ⊗ ℓ k , i ( X k ) with each ℓ j , i \ell_{j,i} ℓ j , i a nonnegative linear form on X j X_j X j ; the least such r r r is the monotone tensor rank rk + ⊗ ( f ) \operatorname{rk}_{+}^{\otimes}(f) rk + ⊗ ( f ) . Flattening f f f along a bipartition of the blocks into Y Y Y and Z Z Z sends each rank-one tensor to a nonnegative rank-one matrix, so rk + ⊗ ( f ) ≥ rk + ( M Y ∣ Z ( f ) ) \operatorname{rk}_{+}^{\otimes}(f)\geq\operatorname{rk}_{+}(M_{Y|Z}(f)) rk + ⊗ ( f ) ≥ rk + ( M Y ∣ Z ( f )) : a nonnegative-rank bound for a single flattening bounds the whole formula.
In the encoding of a DNF A A A the blocks are the pairs { x i 0 , x i 1 } \{x_i^0,x_i^1\} { x i 0 , x i 1 } , one per variable of $A, and a rank-one term is η ∏ i ( p i x i 1 + ( 1 − p i ) x i 0 ) \eta\prod_i(p_i x_i^1+(1-p_i)x_i^0) η ∏ i ( p i x i 1 + ( 1 − p i ) x i 0 ) , a nonnegative multiple of a product distribution with biases p i ∈ [ 0 , 1 ] p_i\in[0,1] p i ∈ [ 0 , 1 ] ; a Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC proof of size r r r is a representation viol C − ϵ = ∑ i ≤ r η i μ i \operatorname{viol}_C-\epsilon=\sum_{i\leq r}\eta_i\mu_i viol C − ϵ = ∑ i ≤ r η i μ i , where viol C ( σ ) \operatorname{viol}_C(\sigma) viol C ( σ ) counts the clauses of the m m m -clause CNF C = ¬ A C=\neg A C = ¬ A falsified by σ \sigma σ . In the bounded-bias fragment Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b the biases lie in { 0 , 1 2 , 1 } \{0,\frac{1}{2},1\} { 0 , 2 1 , 1 } , each term is a subcube indicator, and a proof is a conical junta representation of viol C − ϵ \operatorname{viol}_C-\epsilon viol C − ϵ — the Sherali–Adams normal form of [16 ] .
Theorem 1.1 (Informal; Theorem 3.2 and Corollary 3.4). Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b polynomially simulates Sherali–Adams : a degree-d d d refutation of C C C , displayed as an explicit list of conjunction multipliers of total length s s s , normalises to a single conical junta viol C − ϵ = K \operatorname{viol}_C-\epsilon=K viol C − ϵ = K with at most d s + m ds+m d s + m conjunctions, which is an Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof of A A A with that many terms. Conversely Sherali–Adams over the twin variables x i , x ˉ i x_i,\bar{x}_i x i , x ˉ i , where a conjunction counts as a single monomial, polynomially simulates Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b : an r r r -term Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof is a Sherali–Adams refutation with r + m r+m r + m conjunction occurrences. Since d ≤ n d\leq n d ≤ n , the two systems are polynomially equivalent.
The first direction partitions the complement of an r r r -literal conjunction into r r r first-mismatch cells rather than 2 r − 1 2^r-1 2 r − 1 complete ones, which is what removes the factor 2 d 2^d 2 d of [16 ] . The second is polynomial only in the twin variables: after substituting x ˉ i = 1 − x i \bar{x}_i=1-x_i x ˉ i = 1 − x i a single conjunction expands into as many as 2 d 2^d 2 d ordinary monomials, and in that measure the equivalence fails.
Two further facts place the fragment. Hrubeš’ simulation of Resolution [21 Prop. 5.8] can be carried out inside it: a Resolution refutation with k k k lines gives an Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof with at most 2 n k + m 2nk+m 2 nk + m terms (Corollary 3.5). And an r r r -term Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof is a sums-of-squares refutation with r + m r+m r + m monomials in the twin-variable measure of Atserias and Hakoniemi, so their bounds [2 ] give 2 Ω ( n ) 2^{\Omega(n)} 2 Ω ( n ) for Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b on Tseitin and on random 3-SAT, at every ϵ > 0 \epsilon>0 ϵ > 0 at once (Corollary 3.7). In the other direction P H P n + 1 n \mathrm{PHP}_{n+1}^{n} PHP n + 1 n has an Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof with O ( n 4 ) O(n^4) O ( n 4 ) terms (Theorem 3.8).
None of this reaches arbitrary bias, and that we prove directly, by hyperplane separation. Write T S G \mathrm{TS}_G TS G for the Tseitin contradiction of a graph G G G of odd total charge: the DNF, in the edge variables of G G G , whose terms are the local assignments violating one of its parity constraints (Section 3.3.1). One exhibits a signed test tensor, a witness , scoring high against the target χ ( T S G ) − ϵ ⋅ χ ( 1 ) \chi(\mathrm{TS}_G)-\epsilon\cdot\chi(1) χ ( TS G ) − ϵ ⋅ χ ( 1 ) but low against every rank-one term lying entrywise below it, and the ratio of the two scores bounds the number of terms from below. This is Rothvoß’ method [26 ] , which we adapt to monotone tensor rank and to the regime ϵ < 1 \epsilon<1 ϵ < 1 . Our witness is calibrated so that no subcube term scores much against it, and removing the bias restriction is the substantial step. A general product term is not a subcube, but each of its coordinates is a mixture of two simpler alternatives, pin it to its more likely value or leave it unbiased:
p x 1 + ( 1 − p ) x 0 = ( 2 p − 1 ) ⋅ x 1 + ( 2 − 2 p ) ⋅ 1 2 ( x 0 + x 1 ) ( p ≥ 1 2 ) , p x^1+(1-p)x^0=(2p-1)\cdot x^1+(2-2p)\cdot\frac{1}{2}(x^0+x^1)\qquad\left(p\geq\frac{1}{2}\right), p x 1 + ( 1 − p ) x 0 = ( 2 p − 1 ) ⋅ x 1 + ( 2 − 2 p ) ⋅ 2 1 ( x 0 + x 1 ) ( p ≥ 2 1 ) ,
a convex combination in which the pinned branch carries the weight 2 p − 1 2p-1 2 p − 1 . Tensoring these exhibits the whole term as a mixture of subcubes, which the witness scores by averaging. The subcubes that could spoil the estimate are held down by the expansion of G G G , and the average survives because the term lies entrywise below the target; Section 3.3.7 carries this out.
Theorem 1.2 (Informal; Theorem 3.24). Let G G G be connected, bridgeless and d d d -regular, d ≥ 3 d \ge3 d ≥ 3 , on n n n vertices, with edge expansion h h h , girth exceeding 4 / h 4/h 4/ h and odd total charge. For ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] with ϵ n ≥ 50 e 4 \epsilon n \ge50e^{4} ϵ n ≥ 50 e 4 ,
rk + ⊗ ( χ ( T S G ) − ϵ ⋅ χ ( 1 ) ) ≥ ϵ 2 ( n − ϵ ) 2 ϵ n min { h , 1 } / ( 50 e 4 ) . \operatorname{rk}_{+}^{\otimes}(\chi(\mathrm{TSG})-\epsilon\cdot\chi(1))\ge\frac{\epsilon}{2(n-\epsilon)}2^{\epsilon n\min\{h,1\}/(50e^{4})}. rk + ⊗ ( χ ( TSG ) − ϵ ⋅ χ ( 1 )) ≥ 2 ( n − ϵ ) ϵ 2 ϵ n m i n { h , 1 } / ( 50 e 4 ) . The same lower bound holds for every depth-three certificate at that shift.
On constant-degree expanders this is 2 Ω ( n ) 2^{\Omega(n)} 2 Ω ( n ) at constant ϵ \epsilon ϵ and superpolynomial throughout ϵ = ω ( log n / n ) \epsilon=\omega(\log n/n) ϵ = ω ( log n / n ) ((Corollary 3.26)). The proof gives an ϵ \epsilon ϵ -sensitive tensor-rank bound directly for the Tseitin target. It cannot be pushed to all ϵ > 0 \epsilon>0 ϵ > 0 , and the obstruction is an upper bound: telescoping the inconsistency of T S G \mathrm{TSG} TSG along an ordering of the vertices, and decomposing a weakly biased parity into few nonnegative products by finite differences, gives a polynomial-size proof once ϵ \epsilon ϵ is exponentially small.
Theorem 1.3 (Informal; Theorem 3.29 and Corollary 3.30). For fixed degree and every ϵ ≤ 2 − Θ ( n log n ) \epsilon\le2^{-\Theta(n\log n)} ϵ ≤ 2 − Θ ( n l o g n ) ,
rk + ⊗ ( χ ( T S G ) − ϵ ⋅ χ ( 1 ) ) = O ( n 3 ) . \operatorname{rk}_{+}^{\otimes}(\chi(\mathrm{TSG})-\epsilon\cdot\chi(1))=O(n^{3}). rk + ⊗ ( χ ( TSG ) − ϵ ⋅ χ ( 1 )) = O ( n 3 ) . With Corollary 3.7 this separates Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_{b} ΣΠΣ - AMC b from Σ Π Σ - A M C \Sigma\Pi\Sigma\text{-}\mathrm{AMC} ΣΠΣ - AMC exponentially at depth three.
Depth three is what the whole of this point is about, and the restriction is not cosmetic: at every ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] the same target has a monotone formula of size O ( n 3 ) O(n^{3}) O ( n 3 ) and depth O ( log n ) O(\log n) O ( log n ) , so the exponential bounds above are bounds on tensor rank and not on monotone proof size as such (Theorem 3.33).
So no ϵ \epsilon ϵ -uniform superpolynomial bound for Σ Π Σ - A M C \Sigma\Pi\Sigma\text{-}\mathrm{AMC} ΣΠΣ - AMC can be proved on Tseitin by any method, and the quantifier over ϵ \epsilon ϵ must be settled on a different formula. A constant-size index gadget supplies one, and in doing so converts a Sherali–Adams degree bound into a depth-three size bound: giving each base variable a private table addressed by selector bits plants a copy of the base instance on every address tuple at once, so the copy to be tested is chosen only after the purported decomposition.
Theorem 1.4 (Informal; Theorem 3.36 and Corollary 3.37). Let I 0 I_0 I 0 be an unsatisfiable CSP on m m m variables with L L L constraints and Sherali–Adams degree exceeding d d d , and let A A A be the DNF encoding the composition I 0 ∘ I N D m q I_0\circ\mathrm{IND}_{m}^{q} I 0 ∘ IND m q , with one copy of the index gadget I N D q ( x , y ) = y x \mathrm{IND}_{q}(x,y)=y_x IND q ( x , y ) = y x per base variable. Then
rk + ⊗ ( χ ( A ) − ϵ ⋅ χ ( 1 ) ) ≥ ϵ q d + 1 L ∑ i < d ( m i ) . \operatorname{rk}_{+}^{\otimes}(\chi(A)-\epsilon\cdot\chi(1))\ge\frac{\epsilon q^{d+1}}{L\sum_{i<d}\binom{m}{i}}. rk + ⊗ ( χ ( A ) − ϵ ⋅ χ ( 1 )) ≥ L ∑ i < d ( i m ) ϵ q d + 1 . Taking T S G \mathrm{TSG} TSG as the base, whose Sherali–Adams degree is linear [18 ] , gives an explicit family of tautological DNFs with O ( N ) O(N) O ( N ) terms of constant width and size 2 Ω ( N ) 2^{\Omega(N)} 2 Ω ( N ) for every ϵ ≥ 2 − Ω ( N ) \epsilon\ge2^{-\Omega(N)} ϵ ≥ 2 − Ω ( N ) at once.
What the gadget buys is the range: for T S G \mathrm{TSG} TSG itself the direct argument stops at ϵ = Ω ( log n / n ) \epsilon=\Omega(\log n/n) ϵ = Ω ( log n / n ) , while the lifted family is hard for every ϵ ≥ 2 − Ω ( N ) \epsilon\ge2^{-\Omega(N)} ϵ ≥ 2 − Ω ( N ) . The bound is still not uniform in ϵ \epsilon ϵ , and being a tensor-rank bound it says nothing about rk + \operatorname{rk}_{+} rk + .
Higher mixed-product depth: the flattening bound, and what it yields
At depth three the inequality rk + ⊗ ( f ) ≥ rk + ( M Y ∣ Z ( f ) ) \operatorname{rk}_{+}^{\otimes}(f)\ge\operatorname{rk}_{+}(M_{Y|Z}(f)) rk + ⊗ ( f ) ≥ rk + ( M Y ∣ Z ( f )) is immediate: a proof is a nonnegative tensor decomposition, and each of its terms flattens to a rank-one matrix. A deeper formula admits no such reading, and its flattening rank can exceed the number of its leaves; the bound must therefore say by how much. Depth itself is not the governing parameter. Multiplying by a factor living on one side of the bipartition only permutes and duplicates rows or columns, so what must be counted is the number of doubly mixed product gates on a root-to-leaf path — those both of whose children involve blocks from both sides. We call this the mixed-product depth t t t . Gates are binary here, so a depth-three term, whose product has unbounded fan-in, is bracketed first; multiplying each side separately and only then combining the two one-sided products makes no gate doubly mixed (Figure 1 ).
Figure 1. Two bracketings of one depth-three term g 1 g 2 h 1 h 2 g_1g_2h_1h_2 g 1 g 2 h 1 h 2 , the g i g_i g i depending on blocks from X \mathcal{X} X and the h i h_i h i on blocks from Y \mathcal{Y} Y . Shaded gates are mixed, the doubly drawn gate is doubly mixed. On the left no gate has two mixed children; on the right the root has two, and with 2 k 2k 2 k factors the interleaved bracketing reaches μ = ⌈ log 2 k ⌉ \mu= \lceil\log_2 k\rceil μ = ⌈ log 2 k ⌉ . The mixed-product depth of an unbounded fan-in expression is accordingly read as the minimum over bracketings.
Theorem 1.5 (Informal; Theorem 2.11 ). A monotone formula with L L L leaves and mixed-product depth t t t computing a set-multilinear f f f has rk + ( Flat ( f ) ) ≤ L 2 t \operatorname{rk}_{+}(\operatorname{Flat}(f)) \le L^{2^t} rk + ( Flat ( f )) ≤ L 2 t , so a submatrix of Flat ( f ) \operatorname{Flat}(f) Flat ( f ) of nonnegative rank R R R forces L ≥ R 1 / 2 t L \ge R^{1/2^t} L ≥ R 1/ 2 t . A depth-three proof has mixed-product depth zero.
Mixed-product depth zero is strictly more permissive than depth three. For every ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] and every bipartition, χ ( T S G ) − ϵ ⋅ χ ( 1 ) \chi(\mathrm{TSG})-\epsilon\cdot\chi(1) χ ( TSG ) − ϵ ⋅ χ ( 1 ) has a monotone formula of mixed-product depth zero with O ( n 3 ) O(n^3) O ( n 3 ) leaves (Theorem 3.33 ), while on a constant-degree expander its tensor rank, the least size of a depth-three proof, is 2 Ω ( n ) 2^{\Omega(n)} 2 Ω ( n ) at constant ϵ \epsilon ϵ . Every bipartite flattening of the target therefore has nonnegative rank O ( n 3 ) O(n^3) O ( n 3 ) , so the flattening bound is empty for T S G \mathrm{TSG} TSG at every t t t (Corollary 3.34 ) and a DNF to feed it has to be found elsewhere (§ 4 ).
The restriction to formulas cannot be dropped: an explicit O ( n ) O(n) O ( n ) -size set-multilinear monotone circuit with no doubly mixed product gate computes a polynomial of flattening rank 2 n 2^n 2 n (Proposition 2.13 ).
Feeding it needs a DNF one of whose flattenings has, after the shift, a submatrix of large nonnegative rank. Both of our sources of such DNFs use one construction, Proposition 2.14 . Its input is a matrix M ≥ J M \ge J M ≥ J whose rows and columns are indexed by Boolean encodings x x x and y y y , and whose entry M ( x , y ) M(x,y) M ( x , y ) counts how many of some fixed conjunctions W 1 , … , W q W_1,\ldots,W_q W 1 , … , W q (the witnesses) the pair ( x , y ) (x,y) ( x , y ) satisfies; its output is an explicit tautological DNF whose flattening contains M M M as the submatrix on the rows and columns that carry valid encodings. We call this submatrix the valid slice . After the shift it becomes M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J , so if M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J has large nonnegative rank, the DNF is hard, with the leaf bound of Corollary 2.15 . The hypothesis M ≥ J M \ge J M ≥ J is also what makes the rank hard to certify: it forces M − ϵ ⋅ J ≥ ( 1 − ϵ ) J > 0 M-\epsilon\cdot J \ge(1-\epsilon)J > 0 M − ϵ ⋅ J ≥ ( 1 − ϵ ) J > 0 , so for every ϵ < 1 \epsilon< 1 ϵ < 1 the slice has no zero entry, its rectangle covering number is one and its fooling sets are singletons. The entire support-based toolkit is thus unavailable below ϵ = 1 \epsilon= 1 ϵ = 1 , and the rank must come from a method tolerating a strictly positive matrix: hyperplane separation, or the information-theoretic bounds we quote. Even hyperplane separation survives only for signed test matrices — against a nonnegative one the floor ( 1 − ϵ ) J (1-\epsilon)J ( 1 − ϵ ) J is itself an admissible rank-one term and the bound collapses, as § 2.4 computes. The test tensor of § 3.3.3 and the functional behind § 4.4 both carry negative entries for this reason.
We have two recipes for a count matrix M ≥ J M \ge J M ≥ J as above whose perturbation M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J has large nonnegative rank. Recipe I (Proposition 4.1 ) takes a 0 / 1 0/1 0/1 polytope P = conv { χ H : H ∈ H } P=\operatorname{conv}\{\chi_H:H\in\mathcal{H}\} P = conv { χ H : H ∈ H } — say, the perfect matching polytope of K 2 n K_{2n} K 2 n — together with a family F \mathcal{F} F of covering inequalities ⟨ a f , x ⟩ ≥ b \langle a_f, x\rangle\ge b ⟨ a f , x ⟩ ≥ b with a f ∈ { 0 , 1 } E a_f \in\{0,1\}^{\mathcal{E}} a f ∈ { 0 , 1 } E valid on it, and sets M ( f , H ) : = ∣ supp ( a f ) ∩ H ∣ M(f,H) := |\operatorname{supp}(a_f) \cap H| M ( f , H ) := ∣ supp ( a f ) ∩ H ∣ , which is the slack matrix of F \mathcal{F} F at H \mathcal{H} H shifted by b b b . The nonnegative rank of M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J is then at least the least extension complexity of a polyhedron between P P P and the relaxation that lowers every right-hand side from b b b to ϵ \epsilon ϵ , and exceeds it by at most one (see § 4 ); only the lower bound is used. That least extension complexity is an approximate extension complexity . Recipe II (Proposition 4.5) takes a Max-k k k -CSP instance with a degree-d d d Sherali–Adams integrality gap and plants many copies of it in a larger variable set, the rank coming from that gap; its rows are the planted copies rather than valid inequalities, no polytope occurs, and the admissible shift is limited only by the quality of the planting estimate. Beyond the encoding lemma the two have nothing in common.
Theorem 1.6 (Informal; Theorem 4.2 and Corollary 4.3). Recipe I applied to the odd cut inequalities of K 2 n K_{2n} K 2 n gives DNFs A P M , n A_{\mathrm{PM},n} A PM , n and A O F , n A_{\mathrm{OF},n} A OF , n — every odd cut meets every perfect matching, respectively every odd factor — for which every monotone formula of mixed-product depth t t t computing the target has at least rk + ( M P M − ϵ ⋅ J ) 1 / 2 t \operatorname{rk}_{+}(M_{\mathrm{PM}}-\epsilon\cdot J)^{1/2^t} rk + ( M PM − ϵ ⋅ J ) 1/ 2 t leaves. By [4 ] this is 2 Ω ( n / 2 t ) 2^{\Omega(n/2^t)} 2 Ω ( n / 2 t ) for constant ϵ \epsilon ϵ , and still 2 n Ω ( 1 ) / 2 t 2^{n^{\Omega(1)}/2^t} 2 n Ω ( 1 ) / 2 t for ϵ ≥ c 0 log log n / log n \epsilon\ge c_0\log\log n/\log n ϵ ≥ c 0 log log n / log n with c 0 > 8 c_0>8 c 0 > 8 fixed.
Here rk + ( M P M ) = O ( n 2 ) \operatorname{rk}_{+}(M_{\mathrm{PM}})=O(n^2) rk + ( M PM ) = O ( n 2 ) while the perturbation alone creates the exponential bound — the ϵ \epsilon ϵ -sensitivity Hrubeš’ reductions call for, in the form supplied by Rothvoß [26 ] and Braun and Pokutta [4 ] . Appendix § 7 makes the constants explicit and proves the sufficient range ϵ ≥ c 0 log log n / log n \epsilon\ge c_0\log\log n/\log n ϵ ≥ c 0 log log n / log n for every fixed c 0 > 8 c_0>8 c 0 > 8 . Recipe II reaches further down.
Theorem 1.7 (Informal; Theorem 4.6). For every fixed 0 < γ < 1 / 4 0<\gamma<1/4 0 < γ < 1/4 , there is an explicit family of tautological DNFs A K A_K A K on K K K variables with K O ( 1 ) K^{O(1)} K O ( 1 ) terms such that, for all sufficiently large members and every K − γ ≤ ϵ < 1 K^{-\gamma}\le\epsilon<1 K − γ ≤ ϵ < 1 , the flattening of χ ( A K ) − ϵ ⋅ χ ( 1 ) \chi(A_K)-\epsilon\cdot\chi(1) χ ( A K ) − ϵ ⋅ χ ( 1 ) has nonnegative rank K Ω ( log K ) K^{\Omega(\log K)} K Ω ( l o g K ) .
It combines a quantitative form of the planted-CSP argument of Chan, Lee, Raghavendra and Steurer [5 ] , in which we track the dependence of the admissible shift on all parameters, with the linear-degree Sherali–Adams integrality gap for Tseitin on explicit constant-degree expanders [18 ] , also used in Section 3.5.
This is the smallest shift at which we can prove superpolynomial hardness for any explicit polynomial-size DNF at mixed-product depth above zero, and its mechanism — a degree-d d d Sherali–Adams integrality gap at the base instance, planted so as to defeat not that one relaxation but every linear program of the given size at once — needs neither the extension complexity of a combinatorial polytope nor hyperplane separation. It is a statement about a new DNF family, and does not reach the Tseitin range of § 3.3 , which goes down to log K / K \log K/K log K / K but only at depth three.
Tseitin is not the only possible base: the explicit sparse 3-XOR instances of Hopkins and Lin [20 ] , which have opt ( I 0 ) ≤ 1 − μ 1 \operatorname{opt}(I_0)\le1-\mu_1 opt ( I 0 ) ≤ 1 − μ 1 and SoS μ 2 m ( I 0 ) = 1 \operatorname{SoS}_{\mu_2m}(I_0)=1 SoS μ 2 m ( I 0 ) = 1 for constants μ 1 , μ 2 > 0 \mu_1,\mu_2>0 μ 1 , μ 2 > 0 with L = O ( m ) L=O(m) L = O ( m ) constraints, satisfy the hypotheses of Proposition 4.5 equally well and give the same bound, with a constant gap in place of the 1 / n 1/n 1/ n available here.
AI disclosure. The initial version of this paper was written by the author, with minor assistance from Claude Opus 4.6. It was much shorter and contained the pigeonhole principle upper bound, the depth-three Tseitin lower bounds, and the lower bounds for the perfect matching and odd factor DNFs. When an early draft was ready, at the end of summer 2026, the author put a few questions about it to ChatGPT Sol 5.6 for the first time. Several items were retrieved from the literature that the author had not been aware of: the Sherali–Adams normal form of Fleming, Göös, Grosser and Robere [16 ] , the planted-instance argument of Chan, Lee, Raghavendra and Steurer [5 ] , and the moment-matching mixtures of product distributions of Chen and Moitra [9 ] , on which our decomposition of a weakly biased parity into few nonnegative products is modelled. It then became clear that several further important results follow and should be included. The mathematics of those new results is essentially due to ChatGPT Sol 5.6. The writing of the final paper was done in close collaboration with Claude Opus 4.6–5. The author has revised everything that was generated by AI and is fully responsible for the content.
Acknowledgements. The author is grateful to Pavel Hrubeš for introducing him to ε \varepsilon ε -sensitive complexity and for discussions.
Preliminaries
AMC proof system
The algebraic monotone calculus (AMC) was introduced by Hrubeš [21 ] . It is a Hilbert-style proof system whose lines are inequalities f ⪯ g f \preceq g f ⪯ g , where f , g f,g f , g are monotone polynomials over R \mathbb{R} R (i.e. all coefficients nonnegative) in some variable set X X X . The intended semantics is
f ⪯ g ⟺ supp ( f ) ⊆ supp ( g ) , f \preceq g \quad\Longleftrightarrow\quad\operatorname{supp}(f) \subseteq\operatorname{supp}(g), f ⪯ g ⟺ supp ( f ) ⊆ supp ( g ) ,
where supp ( p ) \operatorname{supp}(p) supp ( p ) is the set of monomials that occur in p p p with nonzero coefficient (with the conventions supp ( 0 ) = ∅ \operatorname{supp}(0)=\varnothing supp ( 0 ) = ∅ and supp ( a ) = { 1 } \operatorname{supp}(a)=\{1\} supp ( a ) = { 1 } for a ∈ R ∖ { 0 } a \in\mathbb{R}\setminus\{0\} a ∈ R ∖ { 0 } ).
Definition 2.1 (AMC, [21 Sec. 2.3] ). The axioms of AMC are
(A1) f ⪯ f f \preceq f f ⪯ f for any monotone polynomial f f f ;
(A2) a ⪯ b a \preceq b a ⪯ b for any a , b ∈ R a,b \in\mathbb{R} a , b ∈ R with a ≥ 0 a \ge0 a ≥ 0 , b > 0 b>0 b > 0 .
The inference rules are
(R1) (cut/transitivity) from f 1 ⪯ g f_1 \preceq g f 1 ⪯ g and g ⪯ f 2 g \preceq f_2 g ⪯ f 2 derive f 1 ⪯ f 2 f_1 \preceq f_2 f 1 ⪯ f 2 ;
(R2) (compatibility with ∘ ∈ { + , × } \circ\in\{+,\times\} ∘ ∈ { + , × } ) from f 1 ⪯ g 1 f_1 \preceq g_1 f 1 ⪯ g 1 and f 2 ⪯ g 2 f_2 \preceq g_2 f 2 ⪯ g 2 derive f 1 ∘ f 2 ⪯ g 1 ∘ g 2 f_1 \circ f_2 \preceq g_1 \circ g_2 f 1 ∘ f 2 ⪯ g 1 ∘ g 2 .
An AMC proof of f ⪯ g f \preceq g f ⪯ g is a sequence of lines f 1 ⪯ g 1 , … , f m ⪯ g m f_1 \preceq g_1,\ldots,f_m \preceq g_m f 1 ⪯ g 1 , … , f m ⪯ g m with f m = f f_m=f f m = f , g m = g g_m=g g m = g in which every line is an axiom or follows from earlier lines by (R1) or (R2). The size of the proof is the smallest s ≥ m s \ge m s ≥ m such that the 2 m 2m 2 m polynomials f 1 , g 1 , … , f m , g m f_1,g_1,\ldots,f_m,g_m f 1 , g 1 , … , f m , g m can be simultaneously computed by a single monotone arithmetic circuit of size s s s .
Proposition 2.2 (Soundness and completeness; [21 Prop. 2.6] ). For monotone polynomials f , g f,g f , g the following are equivalent:
supp ( f ) ⊆ supp ( g ) \operatorname{supp}(f) \subseteq\operatorname{supp}(g) supp ( f ) ⊆ supp ( g ) ;
there exists ϵ > 0 \epsilon>0 ϵ > 0 such that g − ϵ ⋅ f g-\epsilon\cdot f g − ϵ ⋅ f is monotone;
there exists an AMC proof of f ⪯ g f \preceq g f ⪯ g .
Theorem 2.3 (Size characterization; [21 Thm. 2.7] ). Let f , g f,g f , g be monotone polynomials.
(a) If f ⪯ g f \preceq g f ⪯ g has an AMC proof of size s s s , then there exists ϵ > 0 \epsilon>0 ϵ > 0 such that f f f , g g g , and g − ϵ ⋅ f g-\epsilon\cdot f g − ϵ ⋅ f are simultaneously computable by a single monotone arithmetic circuit of size O ( s ) O(s) O ( s ) .
(b) Conversely, if ϵ > 0 \epsilon>0 ϵ > 0 and f f f , g g g , g − ϵ ⋅ f g-\epsilon\cdot f g − ϵ ⋅ f are simultaneously computable by a single monotone arithmetic circuit of size s s s , then f ⪯ g f \preceq g f ⪯ g has an AMC proof of size O ( s ) O(s) O ( s ) .
We use the certificate presentation supplied by Theorem 2.3: a certificate for f ⪯ g f \preceq g f ⪯ g is a monotone circuit for g − ϵ ⋅ f g-\epsilon\cdot f g − ϵ ⋅ f for some ϵ > 0 \epsilon>0 ϵ > 0 . The conversion back to a derivation also includes the circuit costs of f f f and g g g . For the DNF encodings below these costs are O ( m n ) O(mn) O ( mn ) , so the two presentations are polynomially equivalent. Depth and formula restrictions in this paper are imposed directly on the certificate. In particular, a depth-3 certificate is a Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ expression g − ϵ ⋅ f = ∑ i = 1 m ∏ j L i , j g-\epsilon\cdot f=\sum_{i=1}^{m}\prod_j L_{i,j} g − ϵ ⋅ f = ∑ i = 1 m ∏ j L i , j , with nonnegative-coefficient affine forms L i , j L_{i,j} L i , j . We use the term “proof” for these certificates as well.
Definition 2.4. Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC is the certificate proof system whose proofs are depth-3 certificates, taken as Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ formulas and measured by the number r r r of product terms. As noted in § 1.1 , at depth three the minimum circuit and formula sizes differ by at most a factor O ( n ) O(n) O ( n ) , so nothing below turns on this choice.
AMC as a propositional proof system
Following [21 ] , the statement “the DNF A A A is a tautology” is encoded as an inequality χ ( 1 ) ⪯ χ ( A ) \chi(1) \preceq\chi(A) χ ( 1 ) ⪯ χ ( A ) as follows. For a DNF A = A 1 ∨ ⋯ ∨ A m A = A_1 \vee\cdots\vee A_m A = A 1 ∨ ⋯ ∨ A m in n n n Boolean variables one works in 2 n 2n 2 n variables x 1 0 , x 1 1 , … , x n 0 , x n 1 x_1^0,x_1^1,\ldots,x_n^0,x_n^1 x 1 0 , x 1 1 , … , x n 0 , x n 1 , one per literal, grouped into blocks X i = { x i 0 , x i 1 } X_i = \{x_i^0,x_i^1\} X i = { x i 0 , x i 1 } ; to an assignment σ ∈ { 0 , 1 } n \sigma\in\{0,1\}^n σ ∈ { 0 , 1 } n corresponds the set-multilinear monomial x σ = ∏ i = 1 n x i σ i x^\sigma= \prod_{i=1}^n x_i^{\sigma_i} x σ = ∏ i = 1 n x i σ i , and
χ ( A ) : = ∑ σ ∈ { 0 , 1 } n c A ( σ ) x σ , c A ( σ ) : = ∣ { j : σ ⊨ A j } ∣ . \chi(A) := \sum_{\sigma\in\{0,1\}^n} c_A(\sigma)x^\sigma,\qquad c_A(\sigma):=\left|\{j:\sigma\models A_j\}\right|. χ ( A ) := σ ∈ { 0 , 1 } n ∑ c A ( σ ) x σ , c A ( σ ) := ∣ { j : σ ⊨ A j } ∣ .
Equivalently χ \chi χ is additive over the terms of A A A , and the characteristic polynomial of a single term A j = ⋀ i ∈ S j x i ∧ ⋀ i ∈ T j ¬ x i A_j = \bigwedge_{i\in S_j}x_i \wedge\bigwedge_{i\in T_j}\neg x_i A j = ⋀ i ∈ S j x i ∧ ⋀ i ∈ T j ¬ x i (with S j , T j ⊆ [ n ] S_j,T_j\subseteq[n] S j , T j ⊆ [ n ] disjoint) is
χ ( A j ) = ∏ i ∈ S j x i 1 ∏ i ∈ T j x i 0 ∏ i ∉ S j ∪ T j ( x i 0 + x i 1 ) . \chi(A_j)=\prod_{i\in S_j}x_i^1\prod_{i\in T_j}x_i^0\prod_{i\notin S_j\cup T_j}(x_i^0+x_i^1). χ ( A j ) = i ∈ S j ∏ x i 1 i ∈ T j ∏ x i 0 i ∈ / S j ∪ T j ∏ ( x i 0 + x i 1 ) .
Each χ ( A j ) \chi(A_j) χ ( A j ) is thus a single shallow monotone product, which is why χ ( A ) \chi(A) χ ( A ) and χ ( 1 ) = ∏ i ( x i 0 + x i 1 ) \chi(1)=\prod_i(x_i^0+x_i^1) χ ( 1 ) = ∏ i ( x i 0 + x i 1 ) admit small, shallow monotone circuits. Since χ ( 1 ) \chi(1) χ ( 1 ) collects all set-multilinear monomials with coefficient 1 1 1 , the DNF A A A is a tautology precisely when supp ( χ ( 1 ) ) ⊆ supp ( χ ( A ) ) \operatorname{supp}(\chi(1))\subseteq\operatorname{supp}(\chi(A)) supp ( χ ( 1 )) ⊆ supp ( χ ( A )) .
We write AMC ϵ \operatorname{AMC}_{\epsilon} AMC ϵ , Σ Π Σ - AMC ϵ \Sigma\Pi\Sigma\text{-}\operatorname{AMC}_{\epsilon} ΣΠΣ - AMC ϵ , Σ Π Σ - AMC b ϵ \Sigma\Pi\Sigma\text{-}\operatorname{AMC}_{b}^{\epsilon} ΣΠΣ - AMC b ϵ for the system in which the parameter is required to be at least ϵ \epsilon ϵ , and drop the superscript when it is existentially quantified as in Proposition 2.2. Demanding a lower bound rather than equality costs nothing: since χ ( 1 ) \chi(1) χ ( 1 ) is itself a single rank-one term, adding ( ϵ ′ − ϵ ) χ ( 1 ) (\epsilon'-\epsilon)\chi(1) ( ϵ ′ − ϵ ) χ ( 1 ) turns a proof with parameter ϵ ′ \epsilon' ϵ ′ into one with parameter ϵ < ϵ ′ \epsilon<\epsilon' ϵ < ϵ ′ at the cost of one further term.
Keeping track of both the parameter and the fragment is essential. With the exception of Corollary 3.7, our lower bounds apply over specified ranges of ϵ \epsilon ϵ to depth-three certificates or formulas of bounded mixed-product depth. A lower bound for an existentially quantified fragment must hold uniformly for every ϵ > 0 \epsilon>0 ϵ > 0 ; a lower bound for unrestricted AMC must additionally allow arbitrary monotone circuits.
Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC, Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b and conical juntas
It is convenient to record the form the two fragments take as identities of functions on the Boolean cube. For a DNF A = A 1 ∨ ⋯ ∨ A m A=A_1\vee\cdots\vee A_m A = A 1 ∨ ⋯ ∨ A m in variables x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n the coefficient c A ( σ ) c_A(\sigma) c A ( σ ) defined above is also the number of clauses of the CNF C : = ¬ A C:=\neg A C := ¬ A falsified by σ \sigma σ , the quantity written viol C ( σ ) \operatorname{viol}_C(\sigma) viol C ( σ ) in the semialgebraic proof complexity literature. Since [ x σ ] χ ( 1 ) = 1 [x^\sigma]\chi(1)=1 [ x σ ] χ ( 1 ) = 1 for every σ \sigma σ , we have [ x σ ] ( χ ( A ) − ϵ ⋅ χ ( 1 ) ) = c A ( σ ) − ϵ [x^\sigma](\chi(A)-\epsilon\cdot\chi(1))=c_A(\sigma)-\epsilon [ x σ ] ( χ ( A ) − ϵ ⋅ χ ( 1 )) = c A ( σ ) − ϵ . Sending a set-multilinear polynomial f f f to its coefficient function σ ↦ [ x σ ] f \sigma\mapsto[x^\sigma]f σ ↦ [ x σ ] f is a linear bijection onto the real functions on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n , the set-multilinear monomials being indexed by σ \sigma σ ; under it the target χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) becomes c A − ϵ = viol C − ϵ c_A-\epsilon=\operatorname{viol}_C-\epsilon c A − ϵ = viol C − ϵ . An AMC proof is therefore an identity of functions on the cube with viol C − ϵ \operatorname{viol}_C-\epsilon viol C − ϵ on the left; the two fragments differ in what may stand on the right, and that is what the rest of this subsection records.
Definition 2.5 (Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b ; the bounded-bias fragment). An Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof of a DNF tautology A A A with parameter ϵ > 0 \epsilon>0 ϵ > 0 is a family { ( η j , S j , T j ) } j = 1 s \{(\eta_j,S_j,T_j)\}_{j=1}^s {( η j , S j , T j ) } j = 1 s with η j ∈ R > 0 \eta_j\in\mathbb{R}_{>0} η j ∈ R > 0 and S j , T j ⊆ [ n ] S_j,T_j\subseteq[n] S j , T j ⊆ [ n ] disjoint
such that
χ ( A ) − ϵ ⋅ χ ( 1 ) = ∑ j = 1 s η j ∏ i ∈ S j x i 1 ∏ i ∈ T j x i 0 ∏ i ∉ S j ⊔ T j ( x i 0 + x i 1 ) . \chi(A)-\epsilon\cdot\chi(1)=\sum_{j=1}^{s}\eta_j\prod_{i\in S_j}x_i^1\prod_{i\in T_j}x_i^0\prod_{i\notin S_j\sqcup T_j}(x_i^0+x_i^1). χ ( A ) − ϵ ⋅ χ ( 1 ) = j = 1 ∑ s η j i ∈ S j ∏ x i 1 i ∈ T j ∏ x i 0 i ∈ / S j ⊔ T j ∏ ( x i 0 + x i 1 ) .
Its size is s s s and its width is max j ∣ S j ⊔ T j ∣ \max_j|S_j\sqcup T_j| max j ∣ S j ⊔ T j ∣ , the largest number of coordinates a term fixes. Polynomial degree is not a useful measure here: every term has degree exactly n n n .
Reading off coefficients, the j j j -th summand contributes η j \eta_j η j to exactly those σ \sigma σ lying in the subcube C j : = { σ : σ i = 1 ( i ∈ S j ) , σ i = 0 ( i ∈ T j ) } C_j:=\{\sigma:\sigma_i=1\ (i\in S_j),\ \sigma_i=0\ (i\in T_j)\} C j := { σ : σ i = 1 ( i ∈ S j ) , σ i = 0 ( i ∈ T j )} , so Definition 2.5 says precisely that
viol C ( σ ) − ϵ = ∑ j = 1 s η j 1 C j ( σ ) for all σ ∈ { 0 , 1 } n , (1) \operatorname{viol}_C(\sigma)-\epsilon=\sum_{j=1}^{s}\eta_j\mathbf{1}_{C_j}(\sigma)\qquad\text{for all }\sigma\in\{0,1\}^{n}, \tag*{(1)} viol C ( σ ) − ϵ = j = 1 ∑ s η j 1 C j ( σ ) for all σ ∈ { 0 , 1 } n , (1)
that is, that viol C − ϵ \operatorname{viol}_C-\epsilon viol C − ϵ is a conical junta of size s s s and width max j ∣ S j ⊔ T j ∣ \max_j|S_j\sqcup T_j| max j ∣ S j ⊔ T j ∣ .
The fragment ΣΠΣ-AMC relaxes (1) in one step. A rank-1 monotone set-multilinear tensor is
η ⋅ ∏ i = 1 n ( p i x i 1 + ( 1 − p i ) x i 0 ) , η ≥ 0 , p i ∈ [ 0 , 1 ] , \eta\cdot\prod_{i=1}^{n}(p_i x_i^1+(1-p_i)x_i^0),\qquad\eta\geq0,\ p_i\in[0,1], η ⋅ i = 1 ∏ n ( p i x i 1 + ( 1 − p i ) x i 0 ) , η ≥ 0 , p i ∈ [ 0 , 1 ] ,
whose coefficient at x σ x^\sigma x σ is η ⋅ ∏ i p i σ i ( 1 − p i ) 1 − σ i \eta\cdot\prod_i p_i^{\sigma_i}(1-p_i)^{1-\sigma_i} η ⋅ ∏ i p i σ i ( 1 − p i ) 1 − σ i , i.e. η \eta η times the value at σ \sigma σ of a product distribution on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n . Hence a ΣΠΣ-AMC proof of size r r r is exactly a representation
viol C ( σ ) − ϵ = ∑ i = 1 r η i μ i ( σ ) , η i ≥ 0 , μ i a product distribution , (2) \operatorname{viol}_C(\sigma)-\epsilon=\sum_{i=1}^{r}\eta_i\mu_i(\sigma),\qquad\eta_i\geq0,\ \mu_i\text{ a product distribution}, \tag*{(2)} viol C ( σ ) − ϵ = i = 1 ∑ r η i μ i ( σ ) , η i ≥ 0 , μ i a product distribution , (2)
and ΣΠΣ-AMCb _b b is the special case p i ∈ { 0 , 1 2 , 1 } p_i\in\{0,\frac{1}{2},1\} p i ∈ { 0 , 2 1 , 1 } . We call the numbers p i p_i p i the biases of the term and accordingly refer to ΣΠΣ-AMC as the arbitrary-bias and to ΣΠΣ-AMCb _b b as the bounded-bias fragment. A single term of (2) can fail to be a conical junta of subexponential size—not, however, merely because its biases lie in the open interval, since p i = 1 2 p_i=\frac{1}{2} p i = 2 1 for every i i i is one subcube term; what is needed is that the biases be generic. Taking odds ratios p i / ( 1 − p i ) = e 2 i − 1 p_i/(1-p_i)=e^{2^{i-1}} p i / ( 1 − p i ) = e 2 i − 1 makes the coefficient at x σ x^\sigma x σ proportional to e m ( σ ) e^{m(\sigma)} e m ( σ ) with m ( σ ) = ∑ i σ i 2 i − 1 m(\sigma)=\sum_i\sigma_i2^{i-1} m ( σ ) = ∑ i σ i 2 i − 1 , so the 2 n 2^n 2 n coefficients of the term are proportional to 1 , e , e 2 , … , e 2 n − 1 1,e,e^2,\ldots,e^{2^n-1} 1 , e , e 2 , … , e 2 n − 1 . Let χ C 1 , … , χ C r \chi_{C_1},\ldots,\chi_{C_r} χ C 1 , … , χ C r be subcube indicators with r < 2 n r<2^n r < 2 n and consider the homogeneous system ⟨ χ C j , φ ⟩ = 0 \langle\chi_{C_j},\varphi\rangle=0 ⟨ χ C j , φ ⟩ = 0 , j ≤ r j\leq r j ≤ r , in the unknown φ \varphi φ : it has 0/1 coefficients and fewer equations than unknowns, so it has a nonzero rational solution φ \varphi φ , which by linearity annihilates the whole real span of the χ C j \chi_{C_j} χ C j . Pairing φ \varphi φ with the coefficient vector of the term would then give a nontrivial rational linear relation among 1 , e , … , e 2 n − 1 1,e,\ldots,e^{2^n-1} 1 , e , … , e 2 n − 1 , which transcendence of e e e forbids. So the term lies outside the real span of any r < 2 n r<2^n r < 2 n subcube indicators. Hence (2) is a genuine relaxation of (1) at the level of representation. Whether it is one at the level of proof size is a different question, since a proof target must have the special form viol C − ϵ \operatorname{viol}_C-\epsilon viol C − ϵ with viol C \operatorname{viol}_C viol C integer-valued; Corollary 3.30 answers it at exponentially small shift, where the two fragments separate exponentially on Tseitin. We write rk + ⊗ ( ⋅ ) \operatorname{rk}^{\otimes}_{+}(\cdot) rk + ⊗ ( ⋅ ) for the least r r r in (2) and rk + , b ⊗ ( ⋅ ) \operatorname{rk}^{\otimes}_{+,b}(\cdot) rk + , b ⊗ ( ⋅ ) for the least s s s in (1) , so that rk + ⊗ ≤ rk + , b ⊗ \operatorname{rk}^{\otimes}_{+}\leq\operatorname{rk}^{\otimes}_{+,b} rk + ⊗ ≤ rk + , b ⊗ always.
The two subsections on Tseitin work in ± 1 \pm1 ± 1 coordinates, and we record once how the terms of (2) look there.
Lemma 2.6 (Rank-one terms in ± 1 \pm1 ± 1 coordinates). For σ ∈ { 0 , 1 } n \sigma\in\{0,1\}^n σ ∈ { 0 , 1 } n , put z i : = ( − 1 ) σ i z_i:=(-1)^{\sigma_i} z i := ( − 1 ) σ i , and read a set-multilinear tensor as the function σ ↦ \sigma\mapsto σ ↦ its coefficient at x σ x^\sigma x σ . Call ∏ i g i ( z i ) \prod_i g_i(z_i) ∏ i g i ( z i ) a nonnegative product. Since z i = − 1 z_i=-1 z i = − 1 exactly when σ i = 1 \sigma_i=1 σ i = 1 , each factor p i σ i ( 1 − p i ) 1 − σ i p_i^{\sigma_i}(1-p_i)^{1-\sigma_i} p i σ i ( 1 − p i ) 1 − σ i equals
( 1 + β i z i ) / 2 (1+\beta_i z_i)/2 ( 1 + β i z i ) /2 for β i : = 1 − 2 p i \beta_i:=1-2p_i β i := 1 − 2 p i , so the rank-one term of (2) with weight η ≥ 0 \eta\ge0 η ≥ 0 and biases p i ∈ [ 0 , 1 ] p_i\in[0,1] p i ∈ [ 0 , 1 ] reads, in these coordinates,
η ∏ i = 1 n p i σ i ( 1 − p i ) 1 − σ i = η 2 n ∏ i = 1 n ( 1 + β i z i ) , β i : = 1 − 2 p i ∈ [ − 1 , 1 ] , (3) \eta\prod_{i=1}^{n}p_i^{\sigma_i}(1-p_i)^{1-\sigma_i}=\frac{\eta}{2^n}\prod_{i=1}^{n}(1+\beta_i z_i),\qquad\beta_i:=1-2p_i\in[-1,1], \tag*{(3)} η i = 1 ∏ n p i σ i ( 1 − p i ) 1 − σ i = 2 n η i = 1 ∏ n ( 1 + β i z i ) , β i := 1 − 2 p i ∈ [ − 1 , 1 ] , (3)
and conversely η ′ ∏ i ( 1 + β i z i ) \eta'\prod_i(1+\beta_i z_i) η ′ ∏ i ( 1 + β i z i ) with η ′ ≥ 0 \eta'\ge0 η ′ ≥ 0 and β i ∈ [ − 1 , 1 ] \beta_i\in[-1,1] β i ∈ [ − 1 , 1 ] is the rank-one term with η = 2 n η ′ \eta=2^n\eta' η = 2 n η ′ and p i = ( 1 − β i ) / 2 p_i=(1-\beta_i)/2 p i = ( 1 − β i ) /2 . The rank-one terms are therefore exactly the nonnegative products: a nonnegative function on { − 1 , 1 } \{-1,1\} { − 1 , 1 } is g ˉ ( 1 + β z ) \bar g(1+\beta z) g ˉ ( 1 + β z ) with g ˉ : = 1 2 ( g ( 1 ) + g ( − 1 ) ) \bar g:=\frac{1}{2}(g(1)+g(-1)) g ˉ := 2 1 ( g ( 1 ) + g ( − 1 )) , so one with no g i g_i g i identically zero equals the right hand side of (3) for
η ′ = ∏ i 1 2 ( g i ( 1 ) + g i ( − 1 ) ) , β i = g i ( 1 ) − g i ( − 1 ) g i ( 1 ) + g i ( − 1 ) ∈ [ − 1 , 1 ] , \eta'=\prod_i\frac{1}{2}(g_i(1)+g_i(-1)),\qquad\beta_i=\frac{g_i(1)-g_i(-1)}{g_i(1)+g_i(-1)}\in[-1,1], η ′ = i ∏ 2 1 ( g i ( 1 ) + g i ( − 1 )) , β i = g i ( 1 ) + g i ( − 1 ) g i ( 1 ) − g i ( − 1 ) ∈ [ − 1 , 1 ] ,
and one with some g i g_i g i identically zero, that is with g ˉ i = 0 \bar g_i=0 g ˉ i = 0 , is the zero tensor. In particular rk + ⊗ ( f ) \operatorname{rk}^{\otimes}_{+}(f) rk + ⊗ ( f ) is the least number of nonnegative products summing to f f f , and Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b is the case β i ∈ { − 1 , 0 , 1 } \beta_i\in\{-1,0,1\} β i ∈ { − 1 , 0 , 1 } .
The bounds below are proved by induction on a formula, gate by gate, and what they need is set-multilinearity as a condition on the formula rather than on the polynomial it computes. For monotone formulas the two agree after zero gates are pruned, as we record below. Gates have fan-in two throughout; an expression whose sums and products have larger fan-in, such as the depth-three expressions of Definition 2.4 , is read as a formula by bracketing them, which changes neither the polynomial computed nor the number of leaves. It can change the mixed-product depth of Definition 2.10 , and the convention there says how we read it.
Definition 2.7 (Syntactic set-multilinearity). The block support S v S_v S v of a gate v v v is the set of blocks occurring below it, so S v S_v S v is a single block at a variable leaf and empty at a constant leaf. A formula Φ \Phi Φ is syntactically set-multilinear if, at every internal gate v v v with children u , w u,w u , w ,
S u = S w = S v at a sum gate , S u ∩ S w = ∅ and S v = S u ⊔ S w at a product gate . S_u=S_w=S_v\quad\text{at a sum gate},\qquad S_u\cap S_w=\varnothing\ \text{and}\ S_v=S_u\sqcup S_w\quad\text{at a product gate}. S u = S w = S v at a sum gate , S u ∩ S w = ∅ and S v = S u ⊔ S w at a product gate . An induction up the formula then gives that F v F_v F v is set-multilinear with block support S v S_v S v at every gate, which is the form in which the condition is used.
For a general formula this is a genuine restriction: cancellation lets a formula compute a set-multilinear polynomial through gates that are not set-multilinear. In the monotone case it costs nothing.
Lemma 2.8 (Automatic syntactic set-multilinearity). Let Φ \Phi Φ be a monotone formula, none of whose gates computes the zero polynomial, whose output F F F is set-multilinear with block support S S S . Then Φ \Phi Φ is syntactically set-multilinear, and S v S_v S v is the common block support of the monomials of F v F_v F v at every gate v v v .
Proof. Since Φ \Phi Φ is a tree, each gate v v v reaches the output through a single multiplicative context: writing M v M_v M v for the product of the polynomials computed at the siblings of the gates on the path from v v v to the root that are children of a product gate, and R R R for the remaining contributions, F = M v ⋅ F v + R F=M_v\cdot F_v+R F = M v ⋅ F v + R . All coefficients are nonnegative, so nothing cancels and every monomial of M v F v M_vF_v M v F v is a monomial of F F F .
Fix a monomial μ \mu μ of M v M_v M v , which exists as M v ≠ 0 M_v\ne0 M v = 0 , and let m 1 , m 2 m_1,m_2 m 1 , m 2 be monomials of F v F_v F v . Then μ m 1 \mu m_1 μ m 1 and μ m 2 \mu m_2 μ m 2 are monomials of F F F , hence set-multilinear over S S S , so blocks ( μ ) ⊔ blocks ( m 1 ) = S = blocks ( μ ) ⊔ blocks ( m 2 ) \operatorname{blocks}(\mu)\sqcup\operatorname{blocks}(m_1)=S=\operatorname{blocks}(\mu)\sqcup\operatorname{blocks}(m_2) blocks ( μ ) ⊔ blocks ( m 1 ) = S = blocks ( μ ) ⊔ blocks ( m 2 ) and blocks ( m 1 ) = blocks ( m 2 ) \operatorname{blocks}(m_1)=\operatorname{blocks}(m_2) blocks ( m 1 ) = blocks ( m 2 ) . Every monomial of F v F_v F v therefore has the same block support, which we call S v S_v S v , and uses each of its blocks once; that is, F v F_v F v is set-multilinear with block support S v S_v S v , and S v S_v S v is the set of blocks occurring below v v v .
The two clauses follow. At a sum gate, no cancellation makes every monomial of F u F_u F u and of F w F_w F w a monomial of F v F_v F v , so S u = S w = S v S_u=S_w=S_v S u = S w = S v . At a product gate, a monomial of F v F_v F v is m u m w m_um_w m u m w with blocks ( m u m w ) = S u ∪ S w \operatorname{blocks}(m_um_w)=S_u\cup S_w blocks ( m u m w ) = S u ∪ S w , and it meets each of its blocks once, so S u ∩ S w = ∅ S_u\cap S_w=\varnothing S u ∩ S w = ∅ and S v = S u ⊔ S w S_v=S_u\sqcup S_w S v = S u ⊔ S w .
The targets χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) of Section 2.2 are set-multilinear. A formula computing a nonzero target can first be pruned of gates computing zero, without increasing its leaf count, ordinary depth or mixed-product depth. The resulting formula is syntactically set-multilinear by the lemma, so we omit this restriction in the lower bounds below. A zero target has the constant-zero certificate and requires no such reduction.
This is what the next corollary is for. By Definition 2.4 a Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC proof is a depth-three monotone expression ∑ j ∏ j L i , j \sum_j\prod_j L_{i,j} ∑ j ∏ j L i , j in which each L i , j L_{i,j} L i , j is an arbitrary affine form with nonnegative coefficients in all 2 n 2n 2 n variables at once, with no set-multilinearity imposed. The corollary below, which is Lemma 2.8 read on such an expression, says that the freedom buys nothing — the claim made informally in Section 1.1.
Corollary 2.9 (Automatic set-multilinearity at depth three). Let F F F be a nonzero monotone polynomial in the variables { x i 0 , x i 1 } i ∈ [ n ] \{x_i^0,x_i^1\}_{i\in[n]} { x i 0 , x i 1 } i ∈ [ n ] that is homogeneous of degree n n n and set-multilinear with respect to the blocks X i = { x i 0 , x i 1 } X_i=\{x_i^0,x_i^1\} X i = { x i 0 , x i 1 } , and suppose F = ∑ i = 1 r ∏ j = 1 k i L i , j F=\sum_{i=1}^{r}\prod_{j=1}^{k_i}L_{i,j} F = ∑ i = 1 r ∏ j = 1 k i L i , j with every L i , j L_{i,j} L i , j a nonzero affine form with nonnegative coefficients. Then each product term equals η i ∏ b = 1 n ℓ i , b \eta_i\prod_{b=1}^{n}\ell_{i,b} η i ∏ b = 1 n ℓ i , b with η i > 0 \eta_i>0 η i > 0 and ℓ i , b \ell_{i,b} ℓ i , b a linear form supported in the block X b X_b X b . In particular every Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC proof of a DNF tautology A A A is a nonnegative tensor decomposition of χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) with at most as many rank-one terms as the proof has size.
Proof. Read the expression as a monotone formula: bracket its sums and products into gates of fan-in two, each L i , j L_{i,j} L i , j becoming a sum over its constant leaf and its variable leaves, and discard the gates computing zero, which changes nothing as F ≠ 0 F\ne0 F = 0 . The output is set-multilinear with block support [ n ] [n] [ n ] , so by Lemma 2.8 the monomials of every gate share a block support. At the gate L i , j L_{i,j} L i , j those monomials are the constant term, of support ∅ \varnothing ∅ , and the variables occurring in it, each of support its own block; so L i , j L_{i,j} L i , j is either a positive constant, which we absorb into η i \eta_i η i , or a linear form ℓ i , b \ell_{i,b} ℓ i , b inside a single block X b X_b X b . Within a term the product clause makes these blocks disjoint and the sum clause makes their union [ n ] [n] [ n ] , so its linear factors are exactly one per block. □
Flattenings and the flattening bound
Let A A A be a DNF in Boolean variables x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n , encoded as in Section 2.2 via the blocks X i : = { x i 0 , x i 1 } X_i:=\{x_i^0,x_i^1\} X i := { x i 0 , x i 1 } . For a subset I ⊆ [ n ] I\subseteq[n] I ⊆ [ n ] write X I : = ⋃ i ∈ I X i X_I:=\bigcup_{i\in I}X_i X I := ⋃ i ∈ I X i . Let f f f be a monotone polynomial that is set-multilinear with block support S ⊆ [ n ] S\subseteq[n] S ⊆ [ n ] , and let I ⊔ I ‾ = [ n ] I\sqcup\overline{I}=[n] I ⊔ I = [ n ] be a bipartition of the blocks. The flattening of f f f across it is the nonnegative matrix M X I ∣ X I ‾ ( f ) M_{X_I\mid X_{\overline{I}}}(f) M X I ∣ X I ( f ) whose rows are indexed by the set-multilinear monomials in X I X_I X I , whose columns are indexed by those in X I ‾ X_{\overline{I}} X I , and whose entry at a pair of monomials is the coefficient in f f f of their product; the monomials in X ∅ X_{\varnothing} X ∅ are the single empty one. The targets χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) have S = [ n ] S=[n] S = [ n ] , where this is the full coefficient matrix of f f f ; the gates of a formula compute polynomials of smaller support, where the index sets shrink accordingly, a variable giving a single column or a single row and a constant a 1 × 1 1\times1 1 × 1 matrix. When the bipartition is clear from the context we abbreviate this to Flat ( f ) \operatorname{Flat}(f) Flat ( f ) .
Fix such a bipartition. Since χ ( 1 ) = ∏ i = 1 n ( x i 0 + x i 1 ) \chi(1)=\prod_{i=1}^{n}(x_i^0+x_i^1) χ ( 1 ) = ∏ i = 1 n ( x i 0 + x i 1 ) flattens to the all-ones matrix, which we denote J X I ∣ X I ‾ J_{X_I\mid X_{\overline{I}}} J X I ∣ X I ,
M X I ∣ X I ‾ ( χ ( A ) − ϵ ⋅ χ ( 1 ) ) = M X I ∣ X I ‾ ( χ ( A ) ) − ϵ ⋅ J X I ∣ X I ‾ . (4) M_{X_I\mid X_{\overline{I}}}\bigl(\chi(A)-\epsilon\cdot\chi(1)\bigr)=M_{X_I\mid X_{\overline{I}}}\bigl(\chi(A)\bigr)-\epsilon\cdot J_{X_I\mid X_{\overline{I}}}. \tag*{(4)} M X I ∣ X I ( χ ( A ) − ϵ ⋅ χ ( 1 ) ) = M X I ∣ X I ( χ ( A ) ) − ϵ ⋅ J X I ∣ X I . (4)
Because each rank-one tensor of (2) flattens to a nonnegative rank-one matrix, a Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC derivation of A A A with parameter ϵ \epsilon ϵ has at least
rk + ( M X I ∣ X I ‾ ( χ ( A ) ) − ϵ ⋅ J X I ∣ X I ‾ ) (5) \operatorname{rk}_{+}\left(M_{X_I\mid X_{\overline{I}}}(\chi(A))-\epsilon\cdot J_{X_I\mid X_{\overline{I}}}\right) \tag*{(5)} rk + ( M X I ∣ X I ( χ ( A )) − ϵ ⋅ J X I ∣ X I ) (5)
terms, and, rk + \operatorname{rk}_{+} rk + being monotone under taking submatrices, also at least rk + ( N ) \operatorname{rk}_{+}(N) rk + ( N ) for every submatrix N N N of the matrix in (4) . Section 4 proceeds throughout by exhibiting such a submatrix. A derivation in Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCϵ ^{\epsilon} ϵ carries some parameter ϵ ′ ≥ ϵ \epsilon' \ge\epsilon ϵ ′ ≥ ϵ , so what bounds its size is the minimum of (5) over ϵ ′ ≥ ϵ \epsilon' \ge\epsilon ϵ ′ ≥ ϵ ; the lower bounds below hold throughout a range of shifts and so bound that minimum as well.
The rest of this subsection is not standard material: it states this bound for monotone formulas of arbitrary ordinary depth, the parameter being instead the mixed-product depth t t t of Definition 2.10 . That is the form used throughout the paper, and the display above is its case t = 0 t=0 t = 0 read at depth three.
The governing parameter is not depth. Multiplying by a factor that lives on one side of the bipartition permutes and duplicates rows, or columns, and cannot raise the rank, so what has to be counted is only the product gates that genuinely mix the two sides.
Definition 2.10 (Mixed-product depth). Fix a partition of the variable blocks into X ⊔ Y \mathcal{X}\sqcup\mathcal{Y} X ⊔ Y and let Φ \Phi Φ be a syntactically set-multilinear monotone arithmetic formula. A gate is mixed if its block support meets both X \mathcal{X} X and Y \mathcal{Y} Y . A product gate is doubly mixed if both of its children are mixed. For a gate v v v let μ ( v ) \mu(v) μ ( v ) be the maximum number of doubly mixed product gates on a path from v v v to a leaf and L ( v ) L(v) L ( v ) the number of leaves below v v v . The mixed-product depth μ ( Φ ) \mu(\Phi) μ ( Φ ) of Φ \Phi Φ is μ \mu μ of its output gate. A sum or product of larger fan-in has first to be bracketed into gates of fan-in two, and μ \mu μ depends on the choice; since Theorem 2.11 applies to every bracketing, we read μ \mu μ of such an expression as the minimum over bracketings.
Theorem 2.11 (Flattening bound at bounded mixed-product depth). Let Φ \Phi Φ be a syntactically set-multilinear monotone arithmetic formula with respect to the block partition X ⊔ Y \mathcal{X}\sqcup\mathcal{Y} X ⊔ Y . For every gate v v v , with F v F_v F v the polynomial computed at v v v and Flat ( F v ) \operatorname{Flat}(F_v) Flat ( F v ) its flattening across the induced X ∣ Y \mathcal{X}\mid\mathcal{Y} X ∣ Y bipartition,
rk + ( Flat ( F v ) ) ≤ L ( v ) 2 μ ( v ) . \operatorname{rk}_{+}(\operatorname{Flat}(F_v)) \le L(v)^{2^{\mu(v)}}. rk + ( Flat ( F v )) ≤ L ( v ) 2 μ ( v ) . Consequently, if Φ \Phi Φ has L L L leaves and mixed-product depth at most t t t , then
L ≥ rk + ( N ) 1 / 2 t for every submatrix N of Flat ( F Φ ) . (6) L \ge\operatorname{rk}_{+}(N)^{1/2^t}
\quad\text{for every submatrix }N\text{ of }\operatorname{Flat}(F_{\Phi}). \tag*{(6)} L ≥ rk + ( N ) 1/ 2 t for every submatrix N of Flat ( F Φ ) . (6)
Proof. Write r ( v ) : = rk + ( Flat ( F v ) ) r(v):=\operatorname{rk}_{+}(\operatorname{Flat}(F_v)) r ( v ) := rk + ( Flat ( F v )) and induct. A leaf computes a variable, of block support a single block X i X_i X i , or a constant, of empty block support, so by the definition above its flattening is ∣ X i ∣ × 1 |X_i|\times1 ∣ X i ∣ × 1 or 1 × ∣ X i ∣ 1\times|X_i| 1 × ∣ X i ∣ according to the side of X i X_i X i , or 1 × 1 1\times1 1 × 1 ; in each case it has one nonzero entry and r ≤ 1 r\le1 r ≤ 1 .
At a sum gate v v v with children u , w u,w u , w , Definition 2.7 gives S u = S w = S v S_u=S_w=S_v S u = S w = S v . The rows of the flattening of a set-multilinear polynomial with block support S S S are indexed by the set-multilinear monomials in the blocks of S ∩ X S\cap\mathcal{X} S ∩ X and its columns by those in S ∩ Y S\cap\mathcal{Y} S ∩ Y , index sets determined by S S S alone; so Flat ( F u ) \operatorname{Flat}(F_u) Flat ( F u ) and Flat ( F w ) \operatorname{Flat}(F_w) Flat ( F w ) have the same rows and the same columns, and flattening being a re-indexing of the coefficient vector, Flat ( F v ) = Flat ( F u ) + Flat ( F w ) \operatorname{Flat}(F_v)=\operatorname{Flat}(F_u)+\operatorname{Flat}(F_w) Flat ( F v ) = Flat ( F u ) + Flat ( F w ) entrywise. All three matrices are nonnegative because Φ \Phi Φ is monotone, so concatenating nonnegative rank-one decompositions of the two summands gives one of the sum: r ( v ) ≤ r ( u ) + r ( w ) r(v)\le r(u)+r(w) r ( v ) ≤ r ( u ) + r ( w ) . Put q : = 2 μ ( v ) ≥ 1 q:=2^{\mu(v)}\ge1 q := 2 μ ( v ) ≥ 1 ; since μ ( u ) , μ ( w ) ≤ μ ( v ) \mu(u),\mu(w)\le\mu(v) μ ( u ) , μ ( w ) ≤ μ ( v ) , induction and a q + b q ≤ ( a + b ) q a^q+b^q\le(a+b)^q a q + b q ≤ ( a + b ) q for q ≥ 1 q\ge1 q ≥ 1 , a , b ≥ 0 a,b\ge0 a , b ≥ 0 give r ( v ) ≤ L ( u ) q + L ( w ) q ≤ L ( v ) q r(v)\le L(u)^q+L(w)^q\le L(v)^q r ( v ) ≤ L ( u ) q + L ( w ) q ≤ L ( v ) q . At a product gate S u ∩ S w = ∅ S_u \cap S_w = \varnothing S u ∩ S w = ∅ by Definition 2.7, so up to permutations Flat ( F v ) = Flat ( F u ) ⊗ Flat ( F w ) \operatorname{Flat}(F_v) = \operatorname{Flat}(F_u) \otimes\operatorname{Flat}(F_w) Flat ( F v ) = Flat ( F u ) ⊗ Flat ( F w ) and r ( v ) ≤ r ( u ) r ( w ) r(v) \le r(u)r(w) r ( v ) ≤ r ( u ) r ( w ) . If v v v is not doubly mixed, one child is one-sided, its flattening has a single row or column, and r ( v ) ≤ r ( u ) ≤ L ( v ) 2 μ ( v ) r(v) \le r(u) \le L(v)^{2^{\mu(v)}} r ( v ) ≤ r ( u ) ≤ L ( v ) 2 μ ( v ) . If v v v is doubly mixed then μ ( v ) = 1 + max { μ ( u ) , μ ( w ) } \mu(v) = 1 + \max\{\mu(u),\mu(w)\} μ ( v ) = 1 + max { μ ( u ) , μ ( w )} , so with q : = 2 μ ( v ) q := 2^{\mu(v)} q := 2 μ ( v ) both child exponents are at most q / 2 q/2 q /2 and
r ( v ) ≤ L ( u ) q / 2 L ( w ) q / 2 ≤ ( L ( u ) + L ( w ) ) q = L ( v ) q . r(v) \le L(u)^{q/2}L(w)^{q/2} \le(L(u)+L(w))^q = L(v)^q. r ( v ) ≤ L ( u ) q /2 L ( w ) q /2 ≤ ( L ( u ) + L ( w ) ) q = L ( v ) q .
Finally, nonnegative rank is monotone under taking submatrices. □
Remark 2.12 (Depth three gives mixed-product depth zero). A depth-three monotone Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ expression for a set-multilinear polynomial is, by Corollary 2.9 , a sum of rank-one set-multilinear tensors. Its product gates have unbounded fan-in while μ \mu μ counts binary gates, so a bracketing has to be chosen, and the choice matters: interleaving the two sides makes a gate doubly mixed at every level. Bracketing each term by multiplying all X \mathcal{X} X -factors, then all Y \mathcal{Y} Y -factors, and finally the two one-sided products leaves no product gate with two mixed children (Figure 1 ), so the mixed-product depth is zero and (6) reads
rk + ⊗ ( f ) ≥ rk + ( Flat ( f ) ) ≥ rk + ( N ) , \operatorname{rk}_{+}^{\otimes}(f) \ge\operatorname{rk}_{+}(\operatorname{Flat}(f)) \ge\operatorname{rk}_{+}(N), rk + ⊗ ( f ) ≥ rk + ( Flat ( f )) ≥ rk + ( N ) , the bound on nonnegative tensor rank quoted in § 2.4 . Sharing the bottom linear forms in a depth-three circuit also does not help: its flattening rank is at most the number of top product gates. The converse implication fails, and fails exponentially: Remark 3.35 exhibits formulas of mixed-product depth zero, of polynomial size and logarithmic depth, for targets of exponential tensor rank. Mixed-product depth is thus not a measure of ordinary depth.
The restriction to formulas is not cosmetic. Reuse of an already mixed partial polynomial can compress a flattening of exponential rank into a linear-size circuit, even one with no doubly mixed product gate at all.
Proposition 2.13 (Sharing defeats the bound for circuits). For every n n n there is a set-multilinear monotone polynomial P n P_n P n whose flattening has nonnegative rank 2 n 2^n 2 n , but which has a syntactically set-multilinear monotone circuit of size O ( n ) O(n) O ( n ) containing no doubly mixed product gate.
Proof. Use the 2 n 2n 2 n blocks X i = { x i 0 , x i 1 } X_i = \{x_i^0,x_i^1\} X i = { x i 0 , x i 1 } and Y i = { y i 0 , y i 1 } Y_i = \{y_i^0,y_i^1\} Y i = { y i 0 , y i 1 } , partitioned into all X i X_i X i versus all Y i Y_i Y i , and set P n = ∏ i = 1 n ( x i 0 y i 0 + x i 1 y i 1 ) P_n = \prod_{i=1}^{n}(x_i^0y_i^0 + x_i^1y_i^1) P n = ∏ i = 1 n ( x i 0 y i 0 + x i 1 y i 1 ) . The coefficient of ∏ i x i a i ∏ i y i b i \prod_i x_i^{a_i}\prod_i y_i^{b_i} ∏ i x i a i ∏ i y i b i is 1 if a = b a=b a = b and 0 otherwise, so the flattening is the identity matrix of order 2 n 2^n 2 n , of nonnegative rank 2 n 2^n 2 n . On the other hand P i = x i 0 ( y i 0 P i − 1 ) + x i 1 ( y i 1 P i − 1 ) P_i = x_i^0(y_i^0P_{i-1}) + x_i^1(y_i^1P_{i-1}) P i = x i 0 ( y i 0 P i − 1 ) + x i 1 ( y i 1 P i − 1 ) , reusing one gate for P i − 1 P_{i-1} P i − 1 in both summands, adds O ( 1 ) O(1) O ( 1 ) gates per i i i ; every gate is syntactically set-multilinear, and at each product gate one child is a single variable, hence one-sided. □
The polynomial itself is not hard for formulas: the displayed product is one, with 4 n 4n 4 n leaves. But its product gates are doubly mixed, and bracketed in a balanced tree it has mixed-product depth ⌈ log 2 n ⌉ \lceil\log_2 n\rceil ⌈ log 2 n ⌉ , where (6) asks only for L ≥ 2 L \ge2 L ≥ 2 . At mixed-product depth zero the same bound forces 2 n 2^n 2 n leaves, and that is what the circuit above evades by reusing P i − 1 P_{i-1} P i − 1 .
So the method allows arbitrary ordinary depth and arbitrary one-sided multiplication, but a bounded number of genuinely two-sided multiplications per path and no reuse of a mixed partial result. Writing R = rk + ( N ) R = \operatorname{rk}_{+}(N) R = rk + ( N ) , the bound is log L ≥ ( log R ) / 2 t \log L \ge(\log R)/2^t log L ≥ ( log R ) / 2 t . It diverges when 2 t = o ( log R ) 2^t = o(\log R) 2 t = o ( log R ) , and is superpolynomial in an input parameter n n n when ( log R ) / 2 t = ω ( log n ) (\log R)/2^t = \omega(\log n) ( log R ) / 2 t = ω ( log n ) .
The following elementary lemma is the standard way of producing a DNF whose flattening contains a prescribed nonnegative integer matrix. We use it in § 4.4 .
Proposition 2.14 (Local-witness DNF encoding). Let R ⊆ { 0 , 1 } X R \subseteq\{0,1\}^{X} R ⊆ { 0 , 1 } X and C ⊆ { 0 , 1 } Y C \subseteq\{0,1\}^{Y} C ⊆ { 0 , 1 } Y be sets of valid row and column encodings over disjoint variable sets X X X and Y Y Y . Suppose invalid encodings are recognised by DNFs I R ( X ) I_R(X) I R ( X ) and I C ( Y ) I_C(Y) I C ( Y ) , that is, I R ( x ) = 1 I_R(x)=1 I R ( x ) = 1 iff x ∉ R x \notin R x ∈ / R and I C ( y ) = 1 I_C(y)=1 I C ( y ) = 1 iff y ∉ C y \notin C y ∈ / C , and suppose W 1 , … , W q W_1,\ldots,W_q W 1 , … , W q are conjunctions of literals such that for every ( x , y ) ∈ R × C (x,y)\in R\times C ( x , y ) ∈ R × C
M ( x , y ) : = ∣ { i : W i ( x , y ) = 1 } ∣ ≥ 1. M(x,y) := \lvert\{i: W_i(x,y)=1\}\rvert\ge1. M ( x , y ) := ∣{ i : W i ( x , y ) = 1 }∣ ≥ 1. Then
A ( X , Y ) : = I R ( X ) ∨ I C ( Y ) ∨ ⋁ i = 1 q W i ( X , Y ) A(X,Y) := I_R(X) \lor I_C(Y) \lor\bigvee_{i=1}^{q} W_i(X,Y) A ( X , Y ) := I R ( X ) ∨ I C ( Y ) ∨ i = 1 ⋁ q W i ( X , Y ) is a tautological DNF with ∣ I R ∣ + ∣ I C ∣ + q \lvert I_R\rvert+\lvert I_C\rvert+q ∣ I R ∣ + ∣ I C ∣ + q terms whose valid slice — the submatrix of the X ∣ Y X \mid Y X ∣ Y flattening of χ ( A ) \chi(A) χ ( A ) indexed by R × C R\times C R × C , that is, the rows and columns carrying valid encodings — is exactly M M M . Since χ ( 1 ) \chi(1) χ ( 1 ) flattens to the all-ones matrix, the submatrix of the flattening of χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) indexed by R × C R\times C R × C is then M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J ; we call it the perturbed valid slice.
Corollary 2.15 (Leaf bound from the perturbed valid slice). In the setting of Proposition 2.14 , every monotone formula of mixed-product depth at most t t t computing χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) has at least rk + ( M − ϵ ⋅ J ) 1 / 2 t \operatorname{rk}_{+}(M-\epsilon\cdot J)^{1/2^{t}} rk + ( M − ϵ ⋅ J ) 1/ 2 t leaves.
Proof. By Proposition 2.14 the X ∣ Y X\mid Y X ∣ Y flattening of χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) contains M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J as the submatrix indexed by the valid rows and columns. The target is set-multilinear, so the formula is syntactically set-multilinear by Lemma 2.8 and (6) applies. □ \square □
The hypothesis M ≥ J M\ge J M ≥ J is exactly what makes A A A a tautology, and its cost, announced in Section 1.1, is this. Since M − ϵ ⋅ J ≥ ( 1 − ϵ ) J > 0 M-\epsilon\cdot J\ge(1-\epsilon)J>0 M − ϵ ⋅ J ≥ ( 1 − ϵ ) J > 0 for ϵ < 1 \epsilon<1 ϵ < 1 , the matrix has no zero entry, so its rectangle covering number is one and its largest fooling set is a single position. The hyperplane separation bound survives, but only for signed test matrices. Writing α W : = max { ⟨ W , R ⟩ : R nonnegative of rank one , R ≤ M − ϵ ⋅ J } \alpha_W:=\max\{\langle W,R\rangle:R\text{ nonnegative of rank one}, R\le M-\epsilon\cdot J\} α W := max {⟨ W , R ⟩ : R nonnegative of rank one , R ≤ M − ϵ ⋅ J } , the r r r summands of a nonnegative factorization each lie entrywise below M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J and so score at most α W \alpha_W α W apiece, giving rk + ( M − ϵ ⋅ J ) ≥ ⟨ W , M − ϵ ⋅ J ⟩ / α W \operatorname{rk}_{+}(M-\epsilon\cdot J)\ge\langle W,M-\epsilon\cdot J\rangle/\alpha_W rk + ( M − ϵ ⋅ J ) ≥ ⟨ W , M − ϵ ⋅ J ⟩ / α W whenever α W > 0 \alpha_W>0 α W > 0 . For W ≥ 0 W\ge0 W ≥ 0 this is vacuous: the floor ( 1 − ϵ ) J (1-\epsilon)J ( 1 − ϵ ) J is itself an admissible R R R , so α W ≥ ( 1 − ϵ ) ∥ W ∥ 1 \alpha_W\ge(1-\epsilon)\lVert W\rVert_1 α W ≥ ( 1 − ϵ ) ∥ W ∥ 1 , while ⟨ W , M − ϵ ⋅ J ⟩ ≤ ⟨ W , M ⟩ ≤ ∥ W ∥ 1 max x , y M ( x , y ) \langle W,M-\epsilon\cdot J\rangle\le\langle W,M\rangle\le\lVert W\rVert_1\max_{x,y}M(x,y) ⟨ W , M − ϵ ⋅ J ⟩ ≤ ⟨ W , M ⟩ ≤ ∥ W ∥ 1 max x , y M ( x , y ) , and the quotient never exceeds max x , y M ( x , y ) / ( 1 − ϵ ) \max_{x,y}M(x,y)/(1-\epsilon) max x , y M ( x , y ) / ( 1 − ϵ ) .
A witness must therefore have negative entries, and the achievable bound is governed by how small its total mass is relative to its correlation with M M M .
The depth-three fragment
Representation (1) is exactly the normal form for Sherali–Adams refutations of Fleming, Göös, Grosser and Robere [16 Thm. 5] : an unsatisfiable CNF C C C with a Sherali–Adams refutation of degree d d d and size s s s , measured in expanded ordinary monomials, has an ϵ > 0 \epsilon>0 ϵ > 0 and a conical junta J J J of degree d d d and size s 2 d s2^d s 2 d with viol C − ϵ = J \operatorname{viol}_C-\epsilon=J viol C − ϵ = J , and conversely; read as widths, the two sides agree once each is maximised with the clause width w ( C ) w(C) w ( C ) [16 Lem. 10] . In particular the existential quantifier ∃ ϵ > 0 \exists\epsilon>0 ∃ ϵ > 0 built into AMC by Proposition 2.2 is precisely the one appearing on the Sherali–Adams side, and Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b and Sherali–Adams coincide as width measures, each read as its maximum with the clause width w ( C ) w(C) w ( C ) :
max { SA ( C ) , w ( C ) } = max { min ϵ > 0 width ( Σ Π Σ - A M C b ϵ -proof of A ) , w ( C ) } , C : = ¬ A . (7) \max\{\operatorname{SA}(C),w(C)\}=\max\left\{\min_{\epsilon>0}\operatorname{width}(\Sigma\Pi\Sigma\text{-}\mathrm{AMC}^{\epsilon}_b\text{-proof of }A),w(C)\right\},\qquad C:=\neg A. \tag*{(7)} max { SA ( C ) , w ( C )} = max { ϵ > 0 min width ( ΣΠΣ - AMC b ϵ -proof of A ) , w ( C ) } , C := ¬ A . (7)
That maximum cannot be dropped on either side.
In size the two directions are not symmetric. In a Sherali–Adams refutation − ∑ i B i J i + J 0 = − 1 -\sum_i B_iJ_i+J_0=-1 − ∑ i B i J i + J 0 = − 1 the multipliers J i J_i J i may be arbitrary conical juntas, whereas (1) rearranges to
− ∑ i = 1 m B i ϵ + J ϵ = − 1 , (8) -\sum_{i=1}^{m}\frac{B_i}{\epsilon}+\frac{J}{\epsilon}=-1, \tag*{(8)} − i = 1 ∑ m ϵ B i + ϵ J = − 1 , (8)
the special case of constant multipliers. An r r r -term Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b proof is therefore a Sherali–Adams refutation of size r + m r+m r + m : the r r r conjunctions of J / ϵ J/\epsilon J / ϵ , and the one constant multiplier 1 / ϵ 1/\epsilon 1/ ϵ attached to each of the m m m clauses. The converse is [16 Thm. 5] , which reduces arbitrary multipliers to constants at a cost 2 d 2^d 2 d — incurred by writing the complement of an r r r -literal conjunction as the sum of all 2 r − 1 2^r-1 2 r − 1 complete cells. We replace those by r r r cells, which brings the cost down to d d d .
Simulation of Sherali–Adams and the transfer of size lower bounds
We work in the twin variables x i x_i x i and x ˉ i \bar{x}_i x ˉ i , taken as syntactically distinct; they are the variables x i 1 x_i^1 x i 1 and x i 0 x_i^0 x i 0 of Section 2.2 under the names standard in the Sherali–Adams literature. They live in the quotient ring
B n : = R [ x 1 , x ˉ 1 , … , x n , x ˉ n ] / I n , I n : = ⟨ x i 2 − x i , x ˉ i 2 − x ˉ i , x i + x ˉ i − 1 : i ∈ [ n ] ⟩ . \mathcal{B}_n:=\mathbb{R}[x_1,\bar{x}_1,\ldots,x_n,\bar{x}_n]/\mathcal{I}_n,\qquad\mathcal{I}_n:=\langle x_i^2-x_i,\ \bar{x}_i^2-\bar{x}_i,\ x_i+\bar{x}_i-1:i\in[n]\rangle. B n := R [ x 1 , x ˉ 1 , … , x n , x ˉ n ] / I n , I n := ⟨ x i 2 − x i , x ˉ i 2 − x ˉ i , x i + x ˉ i − 1 : i ∈ [ n ]⟩ .
Evaluation on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n identifies B n \mathcal{B}_n B n with the ring of real functions on the cube, so an identity of Boolean functions and a congruence modulo I n \mathcal{I}_n I n are the same statement; we phrase all manipulations symbolically, in B n \mathcal{B}_n B n . A conjunction is a product of consistent literals and a conical junta a nonnegative combination of conjunctions. For each clause D i = ℓ i 1 ∨ ⋯ ∨ ℓ i w i D_i=\ell_{i1}\vee\cdots\vee\ell_{iw_i} D i = ℓ i 1 ∨ ⋯ ∨ ℓ i w i of C = D 1 ∧ ⋯ ∧ D m C=D_1\wedge\cdots\wedge D_m C = D 1 ∧ ⋯ ∧ D m let B i : = ℓ ˉ i 1 ⋯ ℓ ˉ i w i B_i:=\bar{\ell}_{i1}\cdots\bar{\ell}_{iw_i} B i := ℓ ˉ i 1 ⋯ ℓ ˉ i w i be its falsifying conjunction. Under the identification just made, the count viol C \operatorname{viol}_C viol C of Section 2.3 is the element ∑ i B i \sum_i B_i ∑ i B i of B n \mathcal{B}_n B n . A Sherali–Adams refutation of C C C is an identity
− ∑ i = 1 m B i J i + J 0 = − 1 , J i = ∑ a = 1 t i λ i a Q i a , J 0 = ∑ b = 1 t 0 μ b R b , (9) -\sum_{i=1}^{m}B_iJ_i+J_0=-1,\qquad J_i=\sum_{a=1}^{t_i}\lambda_{ia}Q_{ia},\qquad J_0=\sum_{b=1}^{t_0}\mu_bR_b, \tag*{(9)} − i = 1 ∑ m B i J i + J 0 = − 1 , J i = a = 1 ∑ t i λ ia Q ia , J 0 = b = 1 ∑ t 0 μ b R b , (9)
where the Q i a Q_{ia} Q ia and R b R_b R b are conjunctions and the coefficients λ i a , μ b \lambda_{ia},\mu_b λ ia , μ b are positive. Its degree d d d is the largest degree of a nonzero B i Q i a B_iQ_{ia} B i Q ia or R b R_b R b before cancellation, and its size is the number of conjunctions displayed,
s : = t 0 + ∑ i = 1 m t i , (10) s:=t_0+\sum_{i=1}^{m}t_i, \tag*{(10)} s := t 0 + i = 1 ∑ m t i , (10)
repetitions counted, coefficients free. In the twin variables a conjunction is a single monomial, so this is monomial size. [16 ] work instead in the single-variable multilinear quotient R [ x ] / ⟨ x i 2 − x i ⟩ \mathbb{R}[x]/\langle x_i^2-x_i\rangle R [ x ] / ⟨ x i 2 − x i ⟩ , where a conjunction is ∏ i ∈ S x i ∏ j ∉ T ( 1 − x j ) \prod_{i\in S}x_i\prod_{j\notin T}(1-x_j) ∏ i ∈ S x i ∏ j ∈ / T ( 1 − x j ) and there is no x ˉ j \bar{x}_j x ˉ j to begin with, and count the distinct monomials of the expansion; one conjunction is charged up to 2 d 2^d 2 d times.
The normal form (1) is the case of constant multipliers. To reach it, an occurrence − λ B i Q -\lambda B_iQ − λ B i Q in (9) is rewritten as − λ B i + λ B i ( 1 − Q ) -\lambda B_i+\lambda B_i(1-Q) − λ B i + λ B i ( 1 − Q ) , which leaves the clause B i B_i B i with a constant multiplier and moves the remainder to the positive side — provided the complement 1 − Q 1-Q 1 − Q is itself a conical junta. Writing it as the sum of all 2 r − 1 2^r-1 2 r − 1 complete cells on the variables of Q Q Q does that; the following partition uses r r r of them.
Lemma 3.1 (First mismatch). Let Q = ℓ 1 ℓ 2 ⋯ ℓ r Q=\ell_1\ell_2\cdots\ell_r Q = ℓ 1 ℓ 2 ⋯ ℓ r be a conjunction of consistent literals and write ℓ ˉ j \bar{\ell}_j ℓ ˉ j for the complementary literal on the same variable. Then
1 − Q = ∑ j = 1 r ( ∏ h < j ℓ h ) ℓ ˉ j (11) 1-Q=\sum_{j=1}^{r}\left(\prod_{h<j}\ell_h\right)\bar{\ell}_j \tag*{(11)} 1 − Q = j = 1 ∑ r h < j ∏ ℓ h ℓ ˉ j (11)
is an identity of polynomials in the twin variables, a consequence of the linear relations ℓ j + ℓ ˉ j = 1 \ell_j+\bar{\ell}_j=1 ℓ j + ℓ ˉ j = 1 alone; the Boolean relations x i 2 = x i x_i^2=x_i x i 2 = x i are not used. The r r r conjunctions on the right-hand side are moreover pairwise disjoint: the product of any two of them is divisible by some ℓ j ℓ ˉ j \ell_j\bar{\ell}_j ℓ j ℓ ˉ j , hence vanishes modulo x i 2 − x i x_i^2-x_i x i 2 − x i .
Proof. Substituting ℓ ˉ j = 1 − ℓ j \bar{\ell}_j=1-\ell_j ℓ ˉ j = 1 − ℓ j , the j j j -th summand of the right-hand side is ∏ h < j ℓ h − ∏ h ≤ j ℓ h \prod_{h<j}\ell_h-\prod_{h\leq j}\ell_h ∏ h < j ℓ h − ∏ h ≤ j ℓ h , so the sum telescopes to 1 − ∏ h ≤ r ℓ h = 1 − Q 1-\prod_{h\leq r}\ell_h=1-Q 1 − ∏ h ≤ r ℓ h = 1 − Q . For the disjointness let j < k j<k j < k : the k k k -th summand contains ℓ j \ell_j ℓ j in its prefix, while the j j j -th contains ℓ ˉ j \bar{\ell}_j ℓ ˉ j .
The identity coarsens the 2 r − 1 2^r-1 2 r − 1 complete cells by the location of the first mismatch, and is exact. It also separates two factors of 2 d 2^d 2 d that are easily conflated. The one in [16 Thm. 5] is a cost of replacing the multipliers by constants: their argument replaces a multiplier monomial ℓ 1 ⋯ ℓ r \ell_1\cdots\ell_r ℓ 1 ⋯ ℓ r by the complement junta obtained from ∑ S ⊆ [ r ] ∏ i ∈ S ℓ i ∏ j ∉ S ℓ ˉ j = 1 \sum_{S\subseteq[r]}\prod_{i\in S}\ell_i\prod_{j\notin S}\bar{\ell}_j=1 ∑ S ⊆ [ r ] ∏ i ∈ S ℓ i ∏ j ∈ / S ℓ ˉ j = 1 , that is, by a sum of 2 r − 1 2^r-1 2 r − 1 complete cells; this is the factor the lemma removes, replacing those cells by r r r . The second is a cost of the size convention alone, independent of how the multipliers are handled: under expanded ordinary-monomial counting a width-d d d conjunction is worth up to 2 d 2^d 2 d monomials however it was produced. Our accounting avoids the first outright; the second it avoids only in the twin-variable measure.
Theorem 3.2 (Sherali–Adams to conical junta normal form). Let C C C be an unsatisfiable CNF with m m m clauses and maximum clause width w ( C ) w(C) w ( C ) , and let (9) be a Sherali–Adams refutation of C C C of degree d ≥ 1 d\geq1 d ≥ 1 and size s s s , the size (10) being counted in the twin variables, where a conjunction is a single monomial. Put N : = ∑ i = 1 m t i N:=\sum_{i=1}^{m}t_i N := ∑ i = 1 m t i . Then there are an ϵ > 0 \epsilon>0 ϵ > 0 and a conical junta K K K with
viol C − ϵ = K , (12) \operatorname{viol}_C-\epsilon=K, \tag*{(12)} viol C − ϵ = K , (12) where the displayed representation of K K K uses at most
t 0 + d N + m ≤ d s + m t_0+dN+m\leq ds+m t 0 + d N + m ≤ d s + m conjunctions and has width at most max { d , w ( C ) } \max\{d,w(C)\} max { d , w ( C )} .
Proof. Expand the chosen conical-junta representations of the multipliers, but do not expand literal conjunctions into ordinary monomials. The negative part of (9) becomes − ∑ i = 1 m ∑ a = 1 τ i λ i a B i Q i a -\sum_{i=1}^{m}\sum_{a=1}^{\tau_i}\lambda_{ia}B_iQ_{ia} − ∑ i = 1 m ∑ a = 1 τ i λ ia B i Q ia . If B i Q i a B_iQ_{ia} B i Q ia contains two complementary literals it is divisible by some x i x ˉ i x_i\bar{x}_i x i x ˉ i , hence zero in B n \mathcal{B}_n B n , and may be deleted; we do so once and for all.
Consider one surviving occurrence and abbreviate λ = λ i a \lambda=\lambda_{ia} λ = λ ia , B = B i B=B_i B = B i and Q = Q i a = ℓ 1 ⋯ ℓ r Q=Q_{ia}=\ell_1\cdots\ell_r Q = Q ia = ℓ 1 ⋯ ℓ r . Lemma 3.1 gives
− λ B Q = − λ B + λ B ( 1 − Q ) = − λ B + λ ∑ j = 1 r B ( ∏ h < j ℓ h ) ℓ ˉ j . (13) -\lambda BQ=-\lambda B+\lambda B(1-Q)=-\lambda B+\lambda\sum_{j=1}^{r}B\left(\prod_{h<j}\ell_h\right)\bar{\ell}_j. \tag*{(13)} − λ B Q = − λ B + λ B ( 1 − Q ) = − λ B + λ j = 1 ∑ r B h < j ∏ ℓ h ℓ ˉ j . (13) Since r ≤ deg ( B Q ) ≤ d r\leq\deg(BQ)\leq d r ≤ deg ( B Q ) ≤ d , replacing this occurrence creates at most d d d positive conjunctions, each of degree at most deg ( B Q ) \deg(BQ) deg ( B Q ) , and so does not increase the degree. Apply (13) to every surviving occurrence and gather all the positive terms, together with J 0 J_0 J 0 , into a conical junta J ′ J' J ′ . This yields
− ∑ i = 1 m a i B i + J ′ = − 1 , (14) -\sum_{i=1}^{m}a_iB_i+J'=-1, \tag*{(14)} − i = 1 ∑ m a i B i + J ′ = − 1 , (14)
where a i ≥ 0 a_i \ge0 a i ≥ 0 is the sum of the coefficients λ i a \lambda_{ia} λ ia of the surviving multiplier occurrences attached to clause i i i . The junta J ′ J' J ′ has at most t 0 + d N t_0 + dN t 0 + d N displayed conjunctions and width at most d d d .
Not all a i a_i a i vanish: otherwise (14) would assert that the nonnegative function J ′ J' J ′ equals − 1 -1 − 1 . Let a : = max i a i > 0 a := \max_i a_i > 0 a := max i a i > 0 and equalise the clause coefficients by setting
J ′ ′ : = J ′ + ∑ i = 1 m ( a − a i ) B i . J'' := J' + \sum_{i=1}^{m}(a-a_i)B_i. J ′′ := J ′ + i = 1 ∑ m ( a − a i ) B i .
This is again a conical junta; it uses at most m m m additional conjunctions and has width at most max { d , w ( C ) } \max\{d,w(C)\} max { d , w ( C )} , the second alternative arising only when a clause with a i = 0 a_i=0 a i = 0 must be added and contributes its full width. Now (14) reads − a ∑ i B i + J ′ ′ = − 1 -a\sum_i B_i+J''=-1 − a ∑ i B i + J ′′ = − 1 . Dividing by a a a and setting ϵ : = 1 / a \epsilon:= 1/a ϵ := 1/ a , and using ∑ i B i = viol C \sum_i B_i=\operatorname{viol}_C ∑ i B i = viol C , gives viol C − ϵ = J ′ ′ / a \operatorname{viol}_C-\epsilon=J''/a viol C − ϵ = J ′′ / a , which is (12) with the claimed term count and degree. Finally t 0 + d N + m ≤ d ( t 0 + N ) + m = d s + m t_0+dN+m\le d(t_0+N)+m=ds+m t 0 + d N + m ≤ d ( t 0 + N ) + m = d s + m because d ≥ 1 d\ge1 d ≥ 1 .
It remains to translate (12) into the AMC encoding of § 2.2 . For f ∈ B n f\in B_n f ∈ B n , written as the function it represents, define its coefficient lift
Lift ( f ) : = ∑ σ ∈ { 0 , 1 } n f ( σ ) ∏ i = 1 n x i σ i . \operatorname{Lift}(f):=\sum_{\sigma\in\{0,1\}^n}f(\sigma)\prod_{i=1}^{n}x_i^{\sigma_i}. Lift ( f ) := σ ∈ { 0 , 1 } n ∑ f ( σ ) i = 1 ∏ n x i σ i .
a linear bijection onto the set-multilinear polynomials in the paired variables x i 0 , x i 1 x_i^0,x_i^1 x i 0 , x i 1 .
Lemma 3.3. If Q = ∏ i ∈ S x i ∏ j ∈ T x ˉ j Q=\prod_{i\in S}x_i\prod_{j\in T}\bar{x}_j Q = ∏ i ∈ S x i ∏ j ∈ T x ˉ j is a conjunction, then
Lift ( Q ) = ∏ i ∈ S x i 1 ∏ j ∈ T x j 0 ∏ k ∉ S ∪ T ( x k 0 + x k 1 ) . (15) \operatorname{Lift}(Q)=\prod_{i\in S}x_i^1\prod_{j\in T}x_j^0\prod_{k\notin S\cup T}(x_k^0+x_k^1). \tag*{(15)} Lift ( Q ) = i ∈ S ∏ x i 1 j ∈ T ∏ x j 0 k ∈ / S ∪ T ∏ ( x k 0 + x k 1 ) . (15) Moreover χ ( A ) = Lift ( viol C ) \chi(A)=\operatorname{Lift}(\operatorname{viol}_C) χ ( A ) = Lift ( viol C ) and χ ( 1 ) = Lift ( 1 ) \chi(1)=\operatorname{Lift}(1) χ ( 1 ) = Lift ( 1 ) for A = ¬ C A=\neg C A = ¬ C : both sides of (15) are the sum of the monomials x σ x^\sigma x σ over the subcube Q = 1 Q=1 Q = 1 , and the last two identities are those of § 2.2 read through c A = viol C c_A=\operatorname{viol}_C c A = viol C .
A conjunction thus becomes exactly one Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b subcube term, with no expansion over its free coordinates, and applying Lift to (12) termwise gives
χ ( A ) − ϵ ⋅ χ ( 1 ) = ∑ u = 1 T η u ∏ i ∈ S u x i 1 ∏ j ∈ T u x j 0 ∏ k ∉ S u ∪ T u ( x k 0 + x k 1 ) , η u > 0 , (16) \chi(A)-\epsilon\cdot\chi(1)=\sum_{u=1}^{T}\eta_u\prod_{i\in S_u}x_i^1\prod_{j\in T_u}x_j^0\prod_{k\notin S_u\cup T_u}(x_k^0+x_k^1),\qquad\eta_u>0, \tag*{(16)} χ ( A ) − ϵ ⋅ χ ( 1 ) = u = 1 ∑ T η u i ∈ S u ∏ x i 1 j ∈ T u ∏ x j 0 k ∈ / S u ∪ T u ∏ ( x k 0 + x k 1 ) , η u > 0 , (16)
with T ≤ t 0 + d N + m T\le t_0+dN+m T ≤ t 0 + d N + m . By Definition 2.5 this is precisely an Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof of A A A .
Corollary 3.4 (Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b and Sherali–Adams are polynomially equivalent). Let C C C be an unsatisfiable CNF with m m m clauses, let A = ¬ C A=\neg C A = ¬ C , and measure Sherali–Adams by the size (10) .
A degree-d d d , size-s s s Sherali–Adams refutation of C C C yields a depth-three Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof of A A A , for some ϵ > 0 \epsilon>0 ϵ > 0 , with O ( d s + m ) O(ds+m) O ( d s + m ) product terms and width at most max { d , w ( C ) } \max\{d,w(C)\} max { d , w ( C )} . As a monotone Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ circuit it has top fan-in O ( d s + m ) O(ds+m) O ( d s + m ) , product gates of fan-in n n n and bottom linear forms of fan-in at most 2, hence O ( n ( d s + m ) ) O(n(ds+m)) O ( n ( d s + m )) wires.
Conversely, an Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof of A A A with r r r terms is, by (8) , a Sherali–Adams refutation of C C C with s ≤ r + m s\le r+m s ≤ r + m : one constant multiplier for each clause, and the r r r terms of the junta J / ϵ J/\epsilon J / ϵ .
Since d ≤ n d \le n d ≤ n , the two systems polynomially simulate one another as existential-ϵ \epsilon ϵ proof systems; by (7) the minimum Sherali–Adams degree and the minimum Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b width agree above the clause width w ( C ) w(C) w ( C ) , and either may fall below it.
Proof. In (1), Theorem 3.2 displays T ≤ d s + m T \le ds+m T ≤ d s + m conjunctions of width at most max { d , w ( C ) } \max\{d,w(C)\} max { d , w ( C )} , each of which lifts by Lemma 3.3 to one product term of (16). Its three layers are an output addition of fan-in T T T , product gates of fan-in n n n , and bottom linear forms x i 1 x_i^1 x i 1 , x i 0 x_i^0 x i 0 or x k 0 + x k 1 x_k^0+x_k^1 x k 0 + x k 1 , so there are at most 2 n T + T 2nT+T 2 n T + T wires. Part (2) is (8) read from right to left.
The simulation also stays inside the bounded-bias fragment: its output (16) uses only subcube terms, so the arbitrary biases available in general AMC are never needed.
Three qualifications belong with the statement. It is an equivalence of existential-ϵ \epsilon ϵ systems: neither direction normalises a proof to a prescribed ϵ \epsilon ϵ , the value produced by Theorem 3.2 being 1 / max i a i 1/\max_i a_i 1/ max i a i . It counts product terms and displayed conjunction occurrences, so a wire count carries the additional factor O ( n ) O(n) O ( n ) of part (1), and real coefficients have unit cost on both sides, so it is not a bit-complexity statement. Above all it is an equivalence in the size measure of (10) and not in the expanded ordinary-monomial measure of [16 ] .
Corollary 3.5 (Resolution). Let C C C be an unsatisfiable CNF with m m m clauses in n n n variables and A = ¬ C A=\neg C A = ¬ C . A Resolution refutation of C C C with k k k lines yields, for some ϵ > 0 \epsilon>0 ϵ > 0 , an Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b proof of A A A with at most 2 n k + m 2nk+m 2 nk + m terms. In particular Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b polynomially simulates Resolution.
Proof. Work in B n B_n B n and write B D B_D B D for the falsifying conjunction of a clause D D D , so that B ∅ = 1 B_{\varnothing}=1 B ∅ = 1 , and B D = 0 B_D=0 B D = 0 when D D D contains complementary literals. Let a line D ∨ D ′ D\lor D' D ∨ D ′ be derived by resolving D ∨ x D\lor x D ∨ x with D ′ ∨ ¬ x D'\lor\neg x D ′ ∨ ¬ x . Then B D ∨ x = B D x ˉ B_{D\lor x}=B_D\bar{x} B D ∨ x = B D x ˉ , B D ′ ∨ ¬ x = B D ′ x B_{D'\lor\neg x}=B_{D'}x B D ′ ∨ ¬ x = B D ′ x and B D ∨ D ′ = B D B D ′ B_{D\lor D'}=B_DB_{D'} B D ∨ D ′ = B D B D ′ , and splitting B D B D ′ = B D B D ′ x ˉ + B D B D ′ x B_DB_{D'}=B_DB_{D'}\bar{x}+B_DB_{D'}x B D B D ′ = B D B D ′ x ˉ + B D B D ′ x gives
B D ∨ x + B D ′ ∨ ¬ x = B D ∨ D ′ + B D x ˉ ( 1 − B D ′ ) + B D ′ x ( 1 − B D ) . B_{D\lor x}+B_{D'\lor\neg x}=B_{D\lor D'}+B_D\bar{x}(1-B_{D'})+B_{D'}x(1-B_D). B D ∨ x + B D ′ ∨ ¬ x = B D ∨ D ′ + B D x ˉ ( 1 − B D ′ ) + B D ′ x ( 1 − B D ) . By Lemma 3.1 the last two terms form a conical junta J ℓ J_\ell J ℓ with at most ∣ D ′ ∣ + ∣ D ∣ ≤ 2 n |D'|+|D|\le2n ∣ D ′ ∣ + ∣ D ∣ ≤ 2 n conjunctions, products containing a complementary pair being zero and dropped.
Weight the lines from the last to the first: the final line, the empty clause, gets weight 1 1 1 , and each derived line passes its weight w ℓ w_\ell w ℓ to both of its premises, a line used several times collecting the sum. Multiply the identity of each derived line ℓ \ell ℓ by w ℓ w_\ell w ℓ and add. A derived line other than the last occurs once as a conclusion, with coefficient − w ℓ -w_\ell − w ℓ , and as a premise with total coefficient w ℓ w_\ell w ℓ , so it cancels, and what remains is
∑ i = 1 m W i B i = 1 + ∑ ℓ w ℓ J ℓ . \sum_{i=1}^{m} W_iB_i=1+\sum_{\ell}w_\ell J_\ell. i = 1 ∑ m W i B i = 1 + ℓ ∑ w ℓ J ℓ . where W i ≥ 0 W_i\ge0 W i ≥ 0 is the weight collected by the i i i -th clause of C C C . Not all W i W_i W i vanish, since the right-hand side is at least 1 1 1 . With a : = max i W i a:=\max_i W_i a := max i W i , equalising as in the proof of Theorem 3.2 gives a vcl − 1 = ∑ ℓ w ℓ J ℓ + ∑ i ( a − W i ) B i a\operatorname{vcl}^{-1}=\sum_\ell w_\ell J_\ell+\sum_i(a-W_i)B_i a vcl − 1 = ∑ ℓ w ℓ J ℓ + ∑ i ( a − W i ) B i , a conical junta with at most 2 n k + m 2nk+m 2 nk + m conjunctions. Dividing by a a a and applying Lemma 3.3 gives an Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b proof with ϵ = 1 / a \epsilon=1/a ϵ = 1/ a .
Hrubeš’ simulation [21 Prop. 5.8] bounds AMC proof size, a circuit measure, by O ( ( m + k ) n ) O((m+k)n) O (( m + k ) n ) ; the corollary counts product terms instead, O ( n k + m ) O(nk+m) O ( nk + m ) of them and so O ( n 2 k + n m ) O(n^2k+nm) O ( n 2 k + nm ) wires as a Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ circuit. Both simulations have polynomial size, and we record the corollary to fix the position of Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b among the standard systems.
Size lower bounds transfer. Atserias and Hakoniemi [2 ] work in the twin variables modulo I n \mathcal{I}_n I n and measure a sums-of-squares refutation by the number of twin-variable monomials in the polynomials under the squares and in the multipliers of the constraint equations, multipliers of the Boolean axioms being uncharged. In this measure an r r r -term Σ Π Σ -AMC b \Sigma\Pi\Sigma\text{-AMC}_{b} ΣΠΣ -AMC b proof is a sums-of-squares refutation of { B i = 0 } \{B_i=0\} { B i = 0 } with at most r + m r+m r + m monomials: dividing the identity viol C − ϵ = ∑ j η j Q j \operatorname{viol}_{C}-\epsilon=\sum_{j}\eta_j Q_j viol C − ϵ = ∑ j η j Q j of (1) by ϵ \epsilon ϵ and using Q j 2 ≡ Q j Q_j^2\equiv Q_j Q j 2 ≡ Q j gives
− 1 ≡ ∑ j = 1 r ( η j / ϵ Q j ) 2 + ∑ i = 1 m ( − 1 ϵ ) B i ( m o d I n ) , -1\equiv\sum_{j=1}^{r}\left(\sqrt{\eta_j/\epsilon}\,Q_j\right)^2+\sum_{i=1}^{m}\left(-\frac{1}{\epsilon}\right)B_i\pmod{\mathcal{I}_n}, − 1 ≡ j = 1 ∑ r ( η j / ϵ Q j ) 2 + i = 1 ∑ m ( − ϵ 1 ) B i ( mod I n ) ,
with one monomial under each square and one constant multiplier per clause; a conjunction counts as a single monomial, and there is no expansion over x ˉ i = 1 − x i \bar{x}_i=1-x_i x ˉ i = 1 − x i . Lower bounds come from their size–degree trade-off, Theorem 3.6.
Theorem 3.6 ([2 ] ). If a system of degree-k k k polynomial constraints in n n n Boolean variables has a sums-of-squares refutation of monomial size at most s s s , then it has one of degree O ( n log s ) + k O(\sqrt{n\log s})+k O ( n log s ) + k .
A degree lower bound may be invoked only for the system of equations it was proved for. For random 3-SAT their system is the clause system itself: their equation p D = 1 p_D=1 p D = 1 for a clause D D D is B D = 1 − p D = 0 B_D=1-p_D=0 B D = 1 − p D = 0 . For Tseitin it is not: they use the parity equations P v = 0 P_v=0 P v = 0 , P v : = ∏ e ∋ v ( 1 − 2 x e ) − ( − 1 ) c ( v ) P_v:=\prod_{e\ni v}(1-2x_e)-(-1)^{c(v)} P v := ∏ e ∋ v ( 1 − 2 x e ) − ( − 1 ) c ( v ) , for which [18 ] supplies the degree lower bound. The clause system converts at no cost. If B v , a B_{v,a} B v , a is the falsifying conjunction of an assignment a a a to δ ( v ) \delta(v) δ ( v ) violating the constraint at v v v , then ∏ e ∋ v ( 1 − 2 x e ) = − ( − 1 ) c ( v ) \prod_{e\ni v}(1-2x_e)=-(-1)^{c(v)} ∏ e ∋ v ( 1 − 2 x e ) = − ( − 1 ) c ( v ) on its support, so
B v , a P v ≡ κ v B v , a ( m o d I n ) , κ v : = − 2 ( − 1 ) c ( v ) ≠ 0 , B_{v,a}P_v\equiv\kappa_vB_{v,a}\pmod{\mathcal{I}_n},\qquad\kappa_v:=-2(-1)^{c(v)}\ne0, B v , a P v ≡ κ v B v , a ( mod I n ) , κ v := − 2 ( − 1 ) c ( v ) = 0 ,
so in the refutation above the constant multiplier of B v , a B_{v,a} B v , a can be traded for the one-monomial multiplier − B v , a / ( ϵ κ v ) -B_{v,a}/(\epsilon\kappa_v) − B v , a / ( ϵ κ v ) of P v P_v P v . An r r r -term Σ Π Σ -AMC b \Sigma\Pi\Sigma\text{-AMC}_{b} ΣΠΣ -AMC b proof of T S G \mathrm{TS}_{G} TS G thus gives a sums-of-squares refutation of the parity system with at most r + m r+m r + m monomials. With these inputs the trade-off gives monomial size 2 Ω ( n ) 2^{\Omega(n)} 2 Ω ( n ) in both cases [2 Cor. 11 and Cor. 14] . Hence:
Corollary 3.7 (Σ Π Σ -AMC b \Sigma\Pi\Sigma\text{-AMC}_{b} ΣΠΣ -AMC b size lower bounds, uniformly in ϵ \epsilon ϵ ). Let G G G be a constant-degree expander on n n n vertices. Every Σ Π Σ -AMC b \Sigma\Pi\Sigma\text{-AMC}_{b} ΣΠΣ -AMC b proof of T S G \mathrm{TS}_{G} TS G has size 2 Ω ( n ) 2^{\Omega(n)} 2 Ω ( n ) , and so does every Σ Π Σ -AMC b \Sigma\Pi\Sigma\text{-AMC}_{b} ΣΠΣ -AMC b proof of the negation of a random 3-SAT instance as in [2 Cor. 14] , in both cases simultaneously for every ϵ > 0 \epsilon>0 ϵ > 0 .
What has been studied in this normal form is its degree : the least width of a conical junta representing viol C − 1 \operatorname{viol}_{C}-1 viol C − 1 , that is the case ϵ = 1 \epsilon=1 ϵ = 1 of (1), which Göös, Jain and Watson treat under the name of the “∃ \exists ∃ -1 game” and for which they prove degree lower bounds for Tseitin formulas [17 ] , and whose restriction to integer coefficients Filmus, Mahajan, Sood and Vinyals relate to the complexity of MaxSAT resolution [14 ] ; see [16 §1.2] for the comparison. The size of a conical junta of unrestricted width, which is the measure of this paper, seems to have received little attention.
An upper bound for the pigeonhole principle
The fragment is thus at least as strong as Sherali–Adams, and the polynomial-size Sherali–Adams refutation of the unary pigeonhole principle, noted by Rhodes [25 ] (see also [10 ] ), already gives a polynomial Σ Π Σ -AMC b \Sigma\Pi\Sigma\text{-AMC}_{b} ΣΠΣ -AMC b proof by Corollary 3.4. We include the construction below, obtained independently of the simulation, because it is explicit: O ( n 4 ) O(n^4) O ( n 4 ) subcube terms with coefficients η ∈ { 0 , 1 } \eta\in\{0,1\} η ∈ { 0 , 1 } and no dependence on the occurrence convention.
The two arguments share their core. One of the three families of terms below consists of the juntas ∏ i x i , k 0 \prod_i x^0_{i,k} ∏ i x i , k 0 , that is ∏ i ( 1 − x i , k ) \prod_i(1-x_{i,k}) ∏ i ( 1 − x i , k ) , saying that hole k k k is empty, and these are also the multipliers carrying the Sherali–Adams refutation. The other two families are the bookkeeping demanded by the exact pointwise identity (1) , which an inequality-based refutation may discard.
We consider the DNF encoding of P H P n n + 1 \mathrm{PHP}^{n+1}_n PHP n n + 1 in the variables { x i , k 0 , x i , k 1 } i ∈ [ n + 1 ] , k ∈ [ n ] \{x^0_{i,k},x^1_{i,k}\}_{i\in[n+1],\,k\in[n]} { x i , k 0 , x i , k 1 } i ∈ [ n + 1 ] , k ∈ [ n ] , so that χ ( 1 ) = ∏ i , k ( x i , k 0 + x i , k 1 ) \chi(1)=\prod_{i,k}(x^0_{i,k}+x^1_{i,k}) χ ( 1 ) = ∏ i , k ( x i , k 0 + x i , k 1 ) . For a Boolean assignment σ \sigma σ set
c k ( σ ) : = ∣ { i : σ ( i , k ) = 1 } ∣ , d i ( σ ) : = ∣ { k : σ ( i , k ) = 1 } ∣ , E i p i g ( σ ) : = 1 [ d i ( σ ) = 0 ] . c_k(\sigma):=|\{i:\sigma(i,k)=1\}|,\qquad d_i(\sigma):=|\{k:\sigma(i,k)=1\}|,\qquad E_i^{\mathrm{pig}}(\sigma):=\mathbf{1}[d_i(\sigma)=0]. c k ( σ ) := ∣ { i : σ ( i , k ) = 1 } ∣ , d i ( σ ) := ∣ { k : σ ( i , k ) = 1 } ∣ , E i pig ( σ ) := 1 [ d i ( σ ) = 0 ] .
The tautology is
P H P n n + 1 : = ⋁ i ∈ [ n + 1 ] ⋀ k ∈ [ n ] ¬ x i , k ∨ ⋁ k ∈ [ n ] ⋁ i < i ′ ( x i , k ∧ x i ′ , k ) , \mathrm{PHP}^{n+1}_n:=\bigvee_{i\in[n+1]}\bigwedge_{k\in[n]}\neg x_{i,k}\ \vee\ \bigvee_{k\in[n]}\bigvee_{i<i'}(x_{i,k}\wedge x_{i',k}), PHP n n + 1 := i ∈ [ n + 1 ] ⋁ k ∈ [ n ] ⋀ ¬ x i , k ∨ k ∈ [ n ] ⋁ i < i ′ ⋁ ( x i , k ∧ x i ′ , k ) ,
one term per empty pigeon and one per collision in a hole, so that ¬ P H P n n + 1 \neg\mathrm{PHP}^{n+1}_n ¬ PHP n n + 1 is the usual CNF with the pigeon clauses ⋁ k x i , k \bigvee_k x_{i,k} ⋁ k x i , k and the hole clauses ¬ x i , k ∨ ¬ x i ′ , k \neg x_{i,k}\vee\neg x_{i',k} ¬ x i , k ∨ ¬ x i ′ , k . Counting satisfied terms,
[ x σ ] χ ( P H P n n + 1 ) = ∑ i = 1 n + 1 E i p i g ( σ ) + ∑ k = 1 n ( c k ( σ ) 2 ) . [\mathbf{x}^{\sigma}]\chi(\mathrm{PHP}^{n+1}_n)=\sum_{i=1}^{n+1}E_i^{\mathrm{pig}}(\sigma)+\sum_{k=1}^{n}\binom{c_k(\sigma)}{2}. [ x σ ] χ ( PHP n n + 1 ) = i = 1 ∑ n + 1 E i pig ( σ ) + k = 1 ∑ n ( 2 c k ( σ ) ) .
Theorem 3.8. P H P n n + 1 \mathrm{PHP}^{n+1}_n PHP n n + 1 admits an Σ Π Σ - A M C b ϵ \Sigma\Pi\Sigma\text{-}\mathrm{AMC}^{\epsilon}_b ΣΠΣ - AMC b ϵ proof of size O ( n 4 ) O(n^4) O ( n 4 ) , for every ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] , using only coefficients η ∈ { 0 , 1 } \eta\in\{0,1\} η ∈ { 0 , 1 } together with a single coefficient 1 − ϵ 1-\epsilon 1 − ϵ .
Proof. Write t + : = max { t , 0 } t_+:=\max\{t,0\} t + := max { t , 0 } . The combinatorial identity ( c 2 ) = ( c − 1 ) + + ( ( c − 1 ) + 2 ) \binom{c}{2}=(c-1)_++\binom{(c-1)_+}{2} ( 2 c ) = ( c − 1 ) + + ( 2 ( c − 1 ) + ) , valid for every integer c ≥ 0 c\ge0 c ≥ 0 , applied to each hole gives
∑ k ( c k 2 ) = ∑ k : c k ≥ 1 ( c k − 1 ) + ∑ k ( ( c k − 1 ) + 2 ) = ∑ i d i − ( n − E h o l ) + ∑ k ( ( c k − 1 ) + 2 ) , \sum_k\binom{c_k}{2}=\sum_{k:c_k\ge1}(c_k-1)+\sum_k\binom{(c_k-1)_+}{2}=\sum_i d_i-(n-E^{\mathrm{hol}})+\sum_k\binom{(c_k-1)_+}{2}, k ∑ ( 2 c k ) = k : c k ≥ 1 ∑ ( c k − 1 ) + k ∑ ( 2 ( c k − 1 ) + ) = i ∑ d i − ( n − E hol ) + k ∑ ( 2 ( c k − 1 ) + ) ,
where E h o l : = ∣ { k : c k = 0 } ∣ E^{\mathrm{hol}}:=|\{k:c_k=0\}| E hol := ∣ { k : c k = 0 } ∣ and we used ∑ k c k = ∑ i d i \sum_k c_k=\sum_i d_i ∑ k c k = ∑ i d i (edge conservation); here and in what follows the fixed assignment σ \sigma σ is suppressed from c k c_k c k , d i d_i d i , E i p i g E_i^{\mathrm{pig}} E i pig and E h o l E^{\mathrm{hol}} E hol . Substituting,
[ x σ ] ( χ ( P H P n n + 1 ) − χ ( 1 ) ) = ∑ i E i p i g + ∑ i d i + E h o l − n + ∑ k ( ( c k − 1 ) + 2 ) − 1 = ∑ i ( E i p i g + d i − 1 ) + E h o l + ∑ k ( ( c k − 1 ) + 2 ) , \begin{aligned}
[\mathbf{x}^{\sigma}]\left(\chi(\mathrm{PHP}^{n+1}_n)-\chi(1)\right)=\sum_iE_i^{\mathrm{pig}}+\sum_i d_i+E^{\mathrm{hol}}-n+\sum_k\binom{(c_k-1)_+}{2}-1 \\
&=\sum_i(E_i^{\mathrm{pig}}+d_i-1)+E^{\mathrm{hol}}+\sum_k\binom{(c_k-1)_+}{2},
\end{aligned} [ x σ ] ( χ ( PHP n n + 1 ) − χ ( 1 ) ) = i ∑ E i pig + i ∑ d i + E hol − n + k ∑ ( 2 ( c k − 1 ) + ) − 1 = i ∑ ( E i pig + d i − 1 ) + E hol + k ∑ ( 2 ( c k − 1 ) + ) ,
using ∑ i 1 = n + 1 \sum_i1=n+1 ∑ i 1 = n + 1 to fold the − n − 1 -n-1 − n − 1 into the per-pigeon sum. Checking cases, E i p i g + d i − 1 E_i^{\mathrm{pig}}+d_i-1 E i pig + d i − 1 is 0 if d i ∈ { 0 , 1 } d_i\in\{0,1\} d i ∈ { 0 , 1 } and equals d i − 1 d_i-1 d i − 1 if d i ≥ 1 d_i\ge1 d i ≥ 1 , so F p i g F_{\mathrm{pig}} F pig counts excess hole assignments per pigeon. All three terms are nonnegative, so, writing
F p i g : = ∑ σ ( ∑ i ( d i − 1 ) + ) x σ , E h o l e s : = ∑ σ E h o l x σ , T h o l e s : = ∑ σ ∑ k ( ( c k − 1 ) + 2 ) x σ F_{\mathrm{pig}}:=\sum_{\sigma}\left(\sum_i(d_i-1)_+\right)\mathbf{x}^{\sigma},\qquad E_{\mathrm{holes}}:=\sum_{\sigma}E^{\mathrm{hol}}\mathbf{x}^{\sigma},\qquad T_{\mathrm{holes}}:=\sum_{\sigma}\sum_k\binom{(c_k-1)_+}{2}\mathbf{x}^{\sigma} F pig := σ ∑ ( i ∑ ( d i − 1 ) + ) x σ , E holes := σ ∑ E hol x σ , T holes := σ ∑ k ∑ ( 2 ( c k − 1 ) + ) x σ
(each coefficient evaluated at the indexing σ \sigma σ ), we obtain the coefficientwise identity χ ( P H P n n + 1 ) − χ ( 1 ) = F p i g + E h o l e s + T h o l e s \chi(\mathrm{PHP}^{n+1}_n)-\chi(1)=F_{\mathrm{pig}}+E_{\mathrm{holes}}+T_{\mathrm{holes}} χ ( PHP n n + 1 ) − χ ( 1 ) = F pig + E holes + T holes and hence
χ ( P H P n n + 1 ) − ϵ ⋅ χ ( 1 ) = ( 1 − ϵ ) ⋅ χ ( 1 ) + F p i g + E h o l e s + T h o l e s . (17) \chi(\mathrm{PHP}^{n+1}_n)-\epsilon\cdot\chi(1)=(1-\epsilon)\cdot\chi(1)+F_{\mathrm{pig}}+E_{\mathrm{holes}}+T_{\mathrm{holes}}. \tag*{(17)} χ ( PHP n n + 1 ) − ϵ ⋅ χ ( 1 ) = ( 1 − ϵ ) ⋅ χ ( 1 ) + F pig + E holes + T holes . (17)
Each family decomposes into bounded-bias rank-1 tensors as follows. Below, “free” means that the variable v v v contributes the factor ( x v 0 + x v 1 ) (x_v^0+x_v^1) ( x v 0 + x v 1 ) .
Floor. A single rank-1 tensor ( 1 − ϵ ) ⋅ χ ( 1 ) (1-\epsilon)\cdot\chi(1) ( 1 − ϵ ) ⋅ χ ( 1 ) . One term.
F p i g F_{\mathrm{pig}} F pig (excess pigeons ). For each i ∈ [ n + 1 ] i\in[n+1] i ∈ [ n + 1 ] and each pair of holes k < k ′ k<k' k < k ′ , the indicator “k k k is the first hole of pigeon i i i and k ′ k' k ′ is also a hole of i i i ” is
F i , k , k ′ ( X ) : = ( ∏ p < k x i , p 0 ) ⋅ x i , k 1 ⋅ x i , k ′ 1 ⋅ ∏ v f r e e ( x v 0 + x v 1 ) . F_{i,k,k'}(X):=\left(\prod_{p<k}x_{i,p}^0\right)\cdot x_{i,k}^1\cdot x_{i,k'}^1\cdot\prod_{v\ \mathrm{free}}(x_v^0+x_v^1). F i , k , k ′ ( X ) := p < k ∏ x i , p 0 ⋅ x i , k 1 ⋅ x i , k ′ 1 ⋅ v free ∏ ( x v 0 + x v 1 ) .
Summing over ( i , k , k ′ ) (i,k,k') ( i , k , k ′ ) recovers F p i g F_{\mathrm{pig}} F pig , whose coefficient at σ \sigma σ is ∑ i ( d i − 1 ) + \sum_i(d_i-1)_+ ∑ i ( d i − 1 ) + . Number of terms: ( n + 1 ) ( n 2 ) = O ( n 3 ) (n+1)\binom{n}{2}=O(n^3) ( n + 1 ) ( 2 n ) = O ( n 3 ) .
E h o l e s E_{\mathrm{holes}} E holes (empty holes ). For each k ∈ [ n ] k\in[n] k ∈ [ n ] ,
E k ( X ) : = ( ∏ i ∈ [ n + 1 ] x i , k 0 ) ⋅ ∏ v f r e e ( x v 0 + x v 1 ) , E_k(X):=\left(\prod_{i\in[n+1]}x_{i,k}^0\right)\cdot\prod_{v\ \mathrm{free}}(x_v^0+x_v^1), E k ( X ) := i ∈ [ n + 1 ] ∏ x i , k 0 ⋅ v free ∏ ( x v 0 + x v 1 ) ,
and summing over k k k recovers E h o l e s E_{\mathrm{holes}} E holes . Number of terms: n n n .
T h o l e s T_{\mathrm{holes}} T holes (excess pairs ). For each hole k k k , each first pigeon i i i in k k k , and each pair of further pigeons i < l < l ′ i<l<l' i < l < l ′ :
T k , i , l , l ′ ( X ) : = ( ∏ q < i x q , k 0 ) ⋅ x i , k 1 ⋅ x l , k 1 ⋅ x l ′ , k 1 ⋅ ∏ v f r e e ( x v 0 + x v 1 ) , T_{k,i,l,l'}(X):=\left(\prod_{q<i}x_{q,k}^0\right)\cdot x_{i,k}^1\cdot x_{l,k}^1\cdot x_{l',k}^1\cdot\prod_{v\ \mathrm{free}}(x_v^0+x_v^1), T k , i , l , l ′ ( X ) := ( q < i ∏ x q , k 0 ) ⋅ x i , k 1 ⋅ x l , k 1 ⋅ x l ′ , k 1 ⋅ v free ∏ ( x v 0 + x v 1 ) ,
and summing recovers T h o l e s T_{\mathrm{holes}} T holes , whose coefficient at σ \sigma σ is ∑ k ( ( c k − 1 ) + 2 ) \sum_k\binom{(c_k-1)_+}{2} ∑ k ( 2 ( c k − 1 ) + ) . Number of terms: n ⋅ ( n + 1 3 ) = O ( n 4 ) n\cdot\binom{n+1}{3}=O(n^4) n ⋅ ( 3 n + 1 ) = O ( n 4 ) .
All the tensors listed are subcube terms in the sense of Definition 2.5, and their total number is O ( n 4 ) O(n^4) O ( n 4 ) .
Tseitin: a direct tensor rank lower bound
Our principal direct lower bound is proved here. It is not a reduction to the extension complexity of a polytope, and it cannot pass through a bipartite flattening: by Corollary 3.34 every bipartite flattening of the perturbed Tseitin target has nonnegative rank O ( n 3 ) O(n^3) O ( n 3 ) , at every ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] , so the route of § 4 is empty here whatever the mixed-product depth. We adapt instead Rothvoß’ hyperplane separation method [26 ] directly to monotone tensor rank and to the regime ϵ < 1 \epsilon<1 ϵ < 1 .
The argument runs in two stages. Sections 3.3.1–3.3.6 carry it out in the bounded-bias fragment Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b , where a rank-one term is a subcube and the defect-stratum bookkeeping is available; § 3.3.7 then removes the bias restriction, by refining an arbitrary product distribution exactly into subcubes and summing the refinements with their actual weights. The outcome, Theorem 3.24, is the same 2 Ω ( ϵ h ˉ n ) 2^{\Omega(\epsilon\bar{h}n)} 2 Ω ( ϵ h ˉ n ) bound, with h ˉ : = min { h , 1 } \bar{h}:=\min\{h,1\} h ˉ := min { h , 1 } for the edge expansion h h h defined below, for all of Σ Π Σ - A M C \Sigma\Pi\Sigma\text{-}\mathrm{AMC} ΣΠΣ - AMC , that is, for every general depth-three AMC proof. § 3.3.8 then shows why it cannot reach all the way: Tseitin has a polynomial-size arbitrary-bias proof once ϵ \epsilon ϵ is exponentially small.
One caveat belongs at the front. At bounded bias the transfer of Corollary 3.7 already gives 2 Ω ( n ) 2^{\Omega(n)} 2 Ω ( n ) on these graphs for every ϵ > 0 \epsilon>0 ϵ > 0 , so Theorem 3.21 is not the strongest Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_b ΣΠΣ - AMC b statement available for Tseitin, and Theorem 3.24 implies it up to the polynomial prefactor, since rk + , b ⊗ ≥ rk + ⊗ \operatorname{rk}^{\otimes}_{+,b}\ge\operatorname{rk}^{\otimes}_+ rk + , b ⊗ ≥ rk + ⊗ . The contribution here is the arbitrary-bias bound, which no transfer currently reaches; the bounded-bias argument is kept because the refinement of § 3.3.7 reuses it verbatim.
Setting
Let G = ( V , E ) G=(V,E) G = ( V , E ) be a graph on n = ∣ V ∣ n=|V| n = ∣ V ∣ vertices with a charge function c : V → F 2 c:V\to\mathbb{F}_2 c : V → F 2 of odd total charge ∑ v c ( v ) ≡ 1 ( m o d 2 ) \sum_v c(v)\equiv1\pmod{2} ∑ v c ( v ) ≡ 1 ( mod 2 ) . The Tseitin formula asserts ⨁ e ∋ v x e = c ( v ) \bigoplus_{e\ni v}x_e=c(v) ⨁ e ∋ v x e = c ( v ) for every v v v ; it is unsatisfiable, and T S G \mathrm{TS}_G TS G denotes the DNF, in the variables { x e } e ∈ E \{x_e\}_{e\in E} { x e } e ∈ E , whose terms are the local assignments violating one of these constraints. For a total assignment x ∈ { 0 , 1 } E x\in\{0,1\}^{E} x ∈ { 0 , 1 } E its defect set is
∂ x + c ∈ F 2 V , ( ∂ x ) v : = ⨁ e ∋ v x e \partial x+c\in\mathbb{F}_2^V,\qquad(\partial x)_v:=\bigoplus_{e\ni v}x_e ∂ x + c ∈ F 2 V , ( ∂ x ) v := e ∋ v ⨁ x e
the set of violated constraints, and its defect count k ( x ) k(x) k ( x ) is the size of that set. Since χ ( T S G ) \chi(\mathrm{TS}_G) χ ( TS G ) has coefficient k ( x ) k(x) k ( x ) at x x x^x x x , we study
T ϵ : = χ ( T S G ) − ϵ ⋅ χ ( 1 ) , [ x x ] T ϵ = k ( x ) − ϵ . T^\epsilon:=\chi(\mathrm{TS}_G)-\epsilon\cdot\chi(1),\qquad[x^x]T^\epsilon=k(x)-\epsilon. T ϵ := χ ( TS G ) − ϵ ⋅ χ ( 1 ) , [ x x ] T ϵ = k ( x ) − ϵ .
We write Q k Q_k Q k for the stratum of assignments with exactly k k k defects, and Z ( G ) : = ker ∂ Z(G):=\ker\partial Z ( G ) := ker ∂ for the cycle space of G G G , whose dimension is ∣ E ∣ − n + 1 |E|-n+1 ∣ E ∣ − n + 1 when G G G is connected and ∣ E ∣ − n + p |E|-n+p ∣ E ∣ − n + p when G G G has p p p components.
Throughout Section 3.3 the graph G G G is d d d -regular and connected with d ≥ 3 d\ge3 d ≥ 3 and edge expansion
h : = min 0 < ∣ S ∣ ≤ n / 2 ∣ E ( S , V ∖ S ) ∣ ∣ S ∣ > 0 ; h:=\min_{0<|S|\le n/2}\frac{|E(S,V\setminus S)|}{|S|}>0; h := 0 < ∣ S ∣ ≤ n /2 min ∣ S ∣ ∣ E ( S , V ∖ S ) ∣ > 0 ;
h > 0 h>0 h > 0 for every connected G G G , and what makes a family of such graphs an expander family is a bound h = Ω ( 1 ) h=\Omega(1) h = Ω ( 1 ) holding uniformly as n n n grows; that is the regime in which the bounds below are read. Note that h ≤ d h\le d h ≤ d , since ∣ E ( S , V ∖ S ) ∣ ≤ d ∣ S ∣ |E(S,V\setminus S)|\le d|S| ∣ E ( S , V ∖ S ) ∣ ≤ d ∣ S ∣ for every S S S . We also assume G G G is bridgeless, which holds whenever h > 1 2 h>\frac{1}{2} h > 2 1 .
Lemma 3.9 (Defect strata). Let G G G carry a charge c c c .
For every x ∈ { 0 , 1 } E x\in\{0,1\}^{E} x ∈ { 0 , 1 } E the defect count k ( x ) k(x) k ( x ) has the same parity as ∑ v c ( v ) \sum_v c(v) ∑ v c ( v ) ; so Q k = ∅ Q_k=\varnothing Q k = ∅ unless k ≡ ∑ v c ( v ) ( m o d 2 ) k\equiv\sum_v c(v)\pmod{2} k ≡ ∑ v c ( v ) ( mod 2 ) , and when the formula is unsatisfiable every total assignment has an odd number of defects.
If moreover G G G is connected, on n n n vertices, then ∣ Q k ∣ = ( n k ) 2 ∣ E ∣ − n + 1 |Q_k|=\binom{n}{k}2^{|E|-n+1} ∣ Q k ∣ = ( k n ) 2 ∣ E ∣ − n + 1 for every k ≡ ∑ v c ( v ) ( m o d 2 ) k\equiv\sum_v c(v)\pmod{2} k ≡ ∑ v c ( v ) ( mod 2 ) .
Proof. Working in F 2 \mathbb{F}_2 F 2 , ∑ v ( ∂ x ) v = ∑ v ⨁ e ∋ v x e = ∑ e 2 x e = 0 \sum_v(\partial x)_v=\sum_v\bigoplus_{e\ni v}x_e=\sum_e2x_e=0 ∑ v ( ∂ x ) v = ∑ v ⨁ e ∋ v x e = ∑ e 2 x e = 0 , because each edge is counted at its two endpoints; hence ∑ v ( ∂ x + c ) v = ∑ v c ( v ) \sum_v(\partial x+c)_v=\sum_v c(v) ∑ v ( ∂ x + c ) v = ∑ v c ( v ) , which is (1). In particular im ∂ \operatorname{im}\partial im ∂ lies in the even-weight subspace of F 2 V \mathbb{F}_2^V F 2 V , and for connected G G G it is all of it: the edge set of a path from u u u to v v v has boundary e u + e v e_u+e_v e u + e v , and such vectors span that subspace, of dimension n − 1 n-1 n − 1 . By rank–nullity dim ker ∂ = ∣ E ∣ − n + 1 \dim\ker\partial=|E|-n+1 dim ker ∂ = ∣ E ∣ − n + 1 , so the defect vector ∂ x + c \partial x+c ∂ x + c takes each value of the parity of c c c at exactly 2 ∣ E ∣ − n + 1 2^{|E|-n+1} 2 ∣ E ∣ − n + 1 assignments, and ( n k ) \binom{n}{k} ( k n ) of those values have weight k k k .
Bounded-bias rank-one tensors as partial assignments
A rank-one monotone tensor in ⨂ e ∈ E ⟨ x e 0 , x e 1 ⟩ + \bigotimes_{e\in E}\langle x_e^0,x_e^1\rangle_+ ⨂ e ∈ E ⟨ x e 0 , x e 1 ⟩ + is a product
R = η ⋅ ∏ e ∈ E ( p e x e 1 + ( 1 − p e ) x e 0 ) , p e ∈ [ 0 , 1 ] , η ≥ 0 , R=\eta\cdot\prod_{e\in E}\left(p_e x_e^1+(1-p_e)x_e^0\right),\qquad p_e\in[0,1],\ \eta\ge0, R = η ⋅ e ∈ E ∏ ( p e x e 1 + ( 1 − p e ) x e 0 ) , p e ∈ [ 0 , 1 ] , η ≥ 0 ,
and it defines a probability distribution D R D_R D R on total assignments, edge e e e being set to 1 independently with probability p e p_e p e . We write
ω k ( R ) : = η ⋅ Pr D R [ x ∈ Q k ] \omega_k(R):=\eta\cdot\Pr_{D_R}[x\in Q_k] ω k ( R ) := η ⋅ D R Pr [ x ∈ Q k ]
for the mass that R R R places on the k k k -th defect stratum, and extend ω k \omega_k ω k additively to sums of rank-one tensors. Until Section 3.3.7 all biases are bounded , i.e. p e ∈ { 0 , 1 , 1 2 } p_e \in\{0,1,\frac{1}{2}\} p e ∈ { 0 , 1 , 2 1 } , which is Definition 2.5 ; we call such a term uniform . Its residual graph G ′ ( R ) G'(R) G ′ ( R ) has edge set { e : p e = 1 2 } \{e:p_e=\frac{1}{2}\} { e : p e = 2 1 } and residual charges c ′ ( v ) = c ( v ) ⊕ ⨁ e ∋ v , p e ∈ { 0 , 1 } p e c'(v)=c(v)\mathbin{\oplus}\bigoplus_{e\ni v,\ p_e\in\{0,1\}}p_e c ′ ( v ) = c ( v ) ⊕ ⨁ e ∋ v , p e ∈ { 0 , 1 } p e ; its components (isolated vertices included) are called odd or even according to the parity of their total residual charge. For a uniform term, every assignment in the support of R R R has the same value η 2 − ∣ E ( G ′ ( R ) ) ∣ \eta2^{-|E(G'(R))|} η 2 − ∣ E ( G ′ ( R )) ∣ , so
ω k ( R ) = η 2 − ∣ E ( G ′ ( R ) ) ∣ ∣ Q k ( R ) ∣ , Q k ( R ) : = Q k ∩ supp R . (18) \omega_k(R)=\eta2^{-|E(G'(R))|}|Q_k(R)|,\qquad Q_k(R):=Q_k\cap\operatorname{supp}R. \tag*{(18)} ω k ( R ) = η 2 − ∣ E ( G ′ ( R )) ∣ ∣ Q k ( R ) ∣ , Q k ( R ) := Q k ∩ supp R . (18)
Proposition 3.10 (Odd-component characterisation). Let R R R be a uniform rank-one tensor and let ℓ = ℓ ( R ) \ell=\ell(R) ℓ = ℓ ( R ) be the number of odd components of G ′ ( R ) G'(R) G ′ ( R ) . Then the minimum defect count over the support of R R R equals ℓ \ell ℓ , and ℓ \ell ℓ is odd. In particular
ω 1 ( R ) = 0 ⟺ ℓ ≥ 3. \omega_1(R)=0\Longleftrightarrow\ell\ge3. ω 1 ( R ) = 0 ⟺ ℓ ≥ 3.
Proof. Fix the deterministic edges. By Lemma 3.9(2) applied to each component X X X of G ′ ( R ) G'(R) G ′ ( R ) , the defect set of an assignment in the support of R R R meets X X X in an arbitrary subset of X X X whose parity is that of the residual charge of X X X , independently across components. The minimum total is therefore one defect in each odd component and none in the even ones, that is ℓ \ell ℓ . Lemma 3.9(1) makes every realisable defect count odd, so ℓ \ell ℓ is odd. Finally ω 1 ( R ) > 0 \omega_1(R)>0 ω 1 ( R ) > 0 iff the minimum is 1 1 1 iff ℓ = 1 \ell=1 ℓ = 1 .
The witness and the plan of the proof
Fix an odd parameter k ∗ ≥ 5 k^*\ge5 k ∗ ≥ 5 and define the symmetric functional W ϵ W^\epsilon W ϵ on assignments by ⟨ W ϵ , x x ⟩ : = W k ( x ) ϵ \langle W^\epsilon,x^x\rangle:=W^\epsilon_{k(x)} ⟨ W ϵ , x x ⟩ := W k ( x ) ϵ with
W 1 ϵ : = − ∞ , W 3 ϵ : = 2 ϵ ∣ Q 3 ∣ , W k ∗ ϵ : = − 1 ( k ∗ − 1 ) ∣ Q k ∗ ∣ , W k ϵ : = 0 for k ∉ { 1 , 3 , k ∗ } . (19) W^\epsilon_1:=-\infty,\qquad W^\epsilon_3:=\frac{2}{\epsilon|Q_3|},\qquad W^\epsilon_{k^*}:=-\frac{1}{(k^*-1)|Q_{k^*}|},\qquad W^\epsilon_k:=0\text{ for }k\notin\{1,3,k^*\}. \tag*{(19)} W 1 ϵ := − ∞ , W 3 ϵ := ϵ ∣ Q 3 ∣ 2 , W k ∗ ϵ := − ( k ∗ − 1 ) ∣ Q k ∗ ∣ 1 , W k ϵ := 0 for k ∈ / { 1 , 3 , k ∗ } . (19)
The value − ∞ -\infty − ∞ is a convention recording that ⟨ W ϵ , ⋅ ⟩ \langle W^\epsilon,\cdot\rangle ⟨ W ϵ , ⋅ ⟩ is evaluated only on tensors with no Q 1 Q_1 Q 1 mass.
The rest of the argument compares the score of W ϵ W^\epsilon W ϵ on the whole target with its score on a single term. Let T = ∑ j ≤ r R j T=\sum_{j\le r}R_j T = ∑ j ≤ r R j be a decomposition into uniform rank-one tensors and let P P P collect the terms with ω 1 > 0 \omega_1>0 ω 1 > 0 , H H H the rest. § 3.3.4 bounds the Q 3 Q_3 Q 3 mass that P P P can carry, by bounding ω 3 ( R ) / ω 1 ( R ) \omega_3(R)/\omega_1(R) ω 3 ( R ) / ω 1 ( R ) for a single term (Lemma 3.12 ); what is left satisfies ω 3 ( H ) ≥ 2 ϵ ∣ Q 3 ∣ \omega_3(H)\ge2\epsilon|Q_3| ω 3 ( H ) ≥ 2 ϵ ∣ Q 3 ∣ (Lemma 3.13 ) and hence ⟨ W ϵ , H ⟩ ≥ 1 \langle W^\epsilon,H\rangle\ge1 ⟨ W ϵ , H ⟩ ≥ 1 (Lemma 3.14 ). § 3.3.5 bounds the other side: a term of H H H whose Q 3 Q_3 Q 3 mass is at least a 1 / τ 1/\tau 1/ τ fraction of 2 ϵ ∣ Q 3 ∣ 2\epsilon|Q_3| 2 ϵ ∣ Q 3 ∣ has a residual component containing more than half the vertices, and expansion then forces on it so much Q k ∗ Q_{k^*} Q k ∗ mass that ⟨ W ϵ , R j ⟩ ≤ 0 \langle W^\epsilon,R_j\rangle\le0 ⟨ W ϵ , R j ⟩ ≤ 0 ; every other term scores less than 4 / τ 4/\tau 4/ τ . Comparing the two gives r ≥ τ / 4 r\ge\tau/4 r ≥ τ /4 in § 3.3.6 .
Remark 3.11 (Why arbitrary biases are excluded here). The restriction p e ∈ { 0 , 1 , 1 2 } p_e\in\{0,1,\frac{1}{2}\} p e ∈ { 0 , 1 , 2 1 } is essential to the argument of Sections 3.3.4–3.3.6, and not for convenience. For general p e ∈ ( 0 , 1 ) p_e\in(0,1) p e ∈ ( 0 , 1 ) the ratio ω 3 ( R ) / ω 1 ( R ) \omega_3(R)/\omega_1(R) ω 3 ( R ) / ω 1 ( R ) is unbounded . Take R R R whose residual graph has exactly two components: an odd one C C C , and an even one consisting of a single free edge e e e of bias p e p_e p e whose two endpoints are satisfied when e e e is unset and defective when it is set. Defect counts add across components, and C C C contributes an odd number while e e e contributes 0 0 0 or 2 2 2 . Writing μ j ( C ) \mu_j(C) μ j ( C ) for the weight of C C C on its j j j -defect patterns, as in the proof of Lemma 3.12 , a total of one defect forces j C = 1 j_C=1 j C = 1 and e e e unset, so ω 1 ( R ) = η μ 1 ( C ) ( 1 − p e ) \omega_1(R)=\eta\mu_1(C)(1-p_e) ω 1 ( R ) = η μ 1 ( C ) ( 1 − p e ) while ω 3 ( R ) = η [ μ 3 ( C ) ( 1 − p e ) + μ 1 ( C ) p e ] \omega_3(R)=\eta[\mu_3(C)(1-p_e)+\mu_1(C)p_e] ω 3 ( R ) = η [ μ 3 ( C ) ( 1 − p e ) + μ 1 ( C ) p e ] . Hence ω 3 ( R ) / ω 1 ( R ) ≥ p e / ( 1 − p e ) → ∞ \omega_3(R)/\omega_1(R)\ge p_e/(1-p_e)\to\infty ω 3 ( R ) / ω 1 ( R ) ≥ p e / ( 1 − p e ) → ∞ as p e → 1 p_e\to1 p e → 1 , the weights of C C C cancelling from the bound. The contribution to ω 3 \omega_3 ω 3 of a term with one odd component can therefore not be controlled by ω 1 \omega_1 ω 1 alone. § 3.3.7 lifts the restriction all the same, not by controlling that ratio — it really is unbounded — but by refining the term into uniform ones and charging each refinement separately.
The deficit after deleting the one-defect terms
Deleting the terms with ω 1 > 0 \omega_1 > 0 ω 1 > 0 is of use only if they cannot carry away most of the Q 3 Q_3 Q 3 mass with them. Their total Q 1 Q_1 Q 1 mass is fixed, at ( 1 − ϵ ) ∣ Q 1 ∣ (1-\epsilon)|Q_1| ( 1 − ϵ ) ∣ Q 1 ∣ , so what they can carry is governed by the ratio ω 3 ( R ) / ω 1 ( R ) \omega_3(R)/\omega_1(R) ω 3 ( R ) / ω 1 ( R ) of a single term. The extreme case is
χ ( T S G ) = ∑ v ∈ V 1 [ v defective ] , (20) \chi(\mathrm{TSG})=\sum_{v\in V}\mathbf{1}[v\text{ defective}], \tag*{(20)} χ ( TSG ) = v ∈ V ∑ 1 [ v defective ] , (20)
whose terms, refined into subcubes, achieve the largest such ratio. The next lemma computes it and shows that no term does better.
Lemma 3.12 (Ratio bound at bounded bias). Let R R R be a uniform rank-one tensor with ω 1 ( R ) > 0 \omega_1(R)>0 ω 1 ( R ) > 0 ; equivalently (Proposition 3.10) G ′ ( R ) G'(R) G ′ ( R ) has exactly one odd component C C C , the remaining components D 1 , … , D s D_1,\ldots,D_s D 1 , … , D s being even. Then
ω 3 ( R ) ω 1 ( R ) = ( ∣ C ∣ 3 ) ( ∣ C ∣ 1 ) + ∑ i = 1 s ( ∣ D i ∣ 2 ) = ( ∣ C ∣ − 1 ) ( ∣ C ∣ − 2 ) 6 + ∑ i = 1 s ∣ D i ∣ ( ∣ D i ∣ − 1 ) 2 ≤ 3 ∣ Q 3 ∣ ∣ Q 1 ∣ . \frac{\omega_3(R)}{\omega_1(R)}
=
\frac{\binom{|C|}{3}}{\binom{|C|}{1}}
+\sum_{i=1}^{s}\binom{|D_i|}{2}
=
\frac{(|C|-1)(|C|-2)}{6}
+\sum_{i=1}^{s}\frac{|D_i|(|D_i|-1)}{2}
\leq\frac{3|Q_3|}{|Q_1|}. ω 1 ( R ) ω 3 ( R ) = ( 1 ∣ C ∣ ) ( 3 ∣ C ∣ ) + i = 1 ∑ s ( 2 ∣ D i ∣ ) = 6 ( ∣ C ∣ − 1 ) ( ∣ C ∣ − 2 ) + i = 1 ∑ s 2 ∣ D i ∣ ( ∣ D i ∣ − 1 ) ≤ ∣ Q 1 ∣ 3∣ Q 3 ∣ . with equality iff ∣ C ∣ = 1 |C|=1 ∣ C ∣ = 1 and there is a single even component (so ∣ D 1 ∣ = n − 1 |D_1|=n-1 ∣ D 1 ∣ = n − 1 ). More generally, for every odd k k k ,
ω k ( R ) ω 1 ( R ) ≤ ( n − 1 k − 1 ) = k ∣ Q k ∣ ∣ Q 1 ∣ . (21) \frac{\omega_k(R)}{\omega_1(R)}
\leq
\binom{n-1}{k-1}
=
\frac{k|Q_k|}{|Q_1|}. \tag*{(21)} ω 1 ( R ) ω k ( R ) ≤ ( k − 1 n − 1 ) = ∣ Q 1 ∣ k ∣ Q k ∣ . (21)
Proof. Component factorisation. For a component X X X of G ′ ( R ) G'(R) G ′ ( R ) let μ j ( X ) : = 2 − ∣ E ( X ) ∣ ∣ Q j ( X ) ∣ \mu_j(X):=2^{-|E(X)|}|Q_j(X)| μ j ( X ) := 2 − ∣ E ( X ) ∣ ∣ Q j ( X ) ∣ be its unnormalised weight on j j j -defect internal patterns. By Lemma 3.9(2) applied to X X X , each realisable internal defect set has 2 ∣ E ( X ) ∣ − ∣ X ∣ + 1 2^{|E(X)|-|X|+1} 2 ∣ E ( X ) ∣ − ∣ X ∣ + 1 preimages, so
μ j ( X ) = ( ∣ X ∣ j ) 2 ∣ X ∣ − 1 \mu_j(X)=\frac{\binom{|X|}{j}}{2^{|X|-1}} μ j ( X ) = 2 ∣ X ∣ − 1 ( j ∣ X ∣ )
if j j j matches the parity of X X X , μ j ( X ) = 0 \mu_j(X)=0 μ j ( X ) = 0 otherwise.
By (18) and edge-disjointness of the components, ω j ( R ) = η ∑ j C + ∑ i j D i = j μ j C ( C ) ∏ i μ j D i ( D i ) \omega_j(R)=\eta\sum_{j_C+\sum_i j_{D_i}=j}\mu_{j_C}(C)\prod_i\mu_{j_{D_i}}(D_i) ω j ( R ) = η ∑ j C + ∑ i j D i = j μ j C ( C ) ∏ i μ j D i ( D i ) , each j X j_X j X ranging over values of the parity of X X X .
The two weights. With the single odd component C C C and even components D i D_i D i , the only way to total one defect is j C = 1 j_C=1 j C = 1 with all j D i = 0 j_{D_i}=0 j D i = 0 ; to total three, either j C = 3 j_C=3 j C = 3 and the rest 0 0 0 , or j C = 1 j_C=1 j C = 1 and exactly one j D i = 2 j_{D_i}=2 j D i = 2 . Hence
ω 1 = η μ 1 ( C ) ∏ i μ 0 ( D i ) , ω 3 = η [ μ 3 ( C ) ∏ i μ 0 ( D i ) + μ 1 ( C ) ∑ i μ 2 ( D i ) ∏ j ≠ i μ 0 ( D j ) ] , \omega_1=\eta\mu_1(C)\prod_i\mu_0(D_i),\qquad
\omega_3=\eta\left[\mu_3(C)\prod_i\mu_0(D_i)+\mu_1(C)\sum_i\mu_2(D_i)\prod_{j\ne i}\mu_0(D_j)\right], ω 1 = η μ 1 ( C ) i ∏ μ 0 ( D i ) , ω 3 = η μ 3 ( C ) i ∏ μ 0 ( D i ) + μ 1 ( C ) i ∑ μ 2 ( D i ) j = i ∏ μ 0 ( D j ) ,
and dividing (the η \eta η and the ∏ i μ 0 ( D i ) \prod_i\mu_0(D_i) ∏ i μ 0 ( D i ) cancel) gives the displayed identity.
Maximisation over ∣ C ∣ + ∑ i ∣ D i ∣ = n |C|+\sum_i|D_i|=n ∣ C ∣ + ∑ i ∣ D i ∣ = n with ∣ C ∣ ≥ 1 |C|\geq1 ∣ C ∣ ≥ 1 , ∣ D i ∣ ≥ 1 |D_i|\geq1 ∣ D i ∣ ≥ 1 . Since
( a + b 2 ) = ( a 2 ) + ( b 2 ) + a b > ( a 2 ) + ( b 2 ) \binom{a+b}{2}=\binom{a}{2}+\binom{b}{2}+ab>\binom{a}{2}+\binom{b}{2} ( 2 a + b ) = ( 2 a ) + ( 2 b ) + ab > ( 2 a ) + ( 2 b )
for a , b ≥ 1 a,b\geq1 a , b ≥ 1 , merging two even components strictly increases ∑ i ( ∣ D i ∣ 2 ) \sum_i\binom{|D_i|}{2} ∑ i ( 2 ∣ D i ∣ ) ; so the maximum has a single even component D D D with ∣ D ∣ = n − ∣ C ∣ |D|=n-|C| ∣ D ∣ = n − ∣ C ∣ , and the objective becomes
f ( κ ) = ( κ − 1 ) ( κ − 2 ) 6 + ( n − κ ) ( n − κ − 1 ) 2 f(\kappa)=\frac{(\kappa-1)(\kappa-2)}{6}+\frac{(n-\kappa)(n-\kappa-1)}{2} f ( κ ) = 6 ( κ − 1 ) ( κ − 2 ) + 2 ( n − κ ) ( n − κ − 1 )
for κ : = ∣ C ∣ ∈ [ 1 , n ] \kappa:=|C|\in[1,n] κ := ∣ C ∣ ∈ [ 1 , n ] . Both terms are convex in κ \kappa κ (leading coefficients 1 6 \frac{1}{6} 6 1 and 1 2 \frac{1}{2} 2 1 , so f ′ ′ = 4 3 > 0 f''=\frac{4}{3}>0 f ′′ = 3 4 > 0 ), so f f f attains its maximum on [ 1 , n ] [1,n] [ 1 , n ] at an endpoint; since f ( 1 ) = ( n − 1 2 ) f(1)=\binom{n-1}{2} f ( 1 ) = ( 2 n − 1 ) and f ( n ) = 1 3 f ( 1 ) f(n)=\frac{1}{3}f(1) f ( n ) = 3 1 f ( 1 ) , the maximum is f ( 1 ) f(1) f ( 1 ) , at κ = 1 \kappa=1 κ = 1 . Both reductions are strict, so equality holds iff ∣ C ∣ = 1 |C|=1 ∣ C ∣ = 1 and there is one even component. Finally, by Lemma 3.9(2), 3 ∣ Q 3 ∣ / ∣ Q 1 ∣ = 3 ( n 3 ) / n = ( n − 1 ) ( n − 2 ) / 2 = f ( 1 ) 3|Q_3|/|Q_1|=3\binom{n}{3}/n=(n-1)(n-2)/2=f(1) 3∣ Q 3 ∣/∣ Q 1 ∣ = 3 ( 3 n ) / n = ( n − 1 ) ( n − 2 ) /2 = f ( 1 ) .
The all-strata form. All realisable defect sets of R R R have the same number of preimages, so ω k ( R ) / ω 1 ( R ) = ∣ F k ∣ / ∣ C ∣ \omega_k(R)/\omega_1(R)=|\mathcal{F}_k|/|C| ω k ( R ) / ω 1 ( R ) = ∣ F k ∣/∣ C ∣ , where F k \mathcal{F}_k F k is the family of k k k -subsets of V V V having odd intersection with C C C and even intersection with every even component, and where the denominator counts the ∣ C ∣ |C| ∣ C ∣ admissible singleton defect sets. Every S ∈ F k S \in\mathcal{F}_k S ∈ F k meets C C C , so counting pairs ( v , S ) (v,S) ( v , S ) with v ∈ C ∩ S v \in C \cap S v ∈ C ∩ S and forgetting the parity restrictions gives ∣ F k ∣ ≤ ∣ C ∣ ( n − 1 k − 1 ) |\mathcal{F}_k| \le|C|\binom{n-1}{k-1} ∣ F k ∣ ≤ ∣ C ∣ ( k − 1 n − 1 ) . Lemma 3.9 gives ∣ Q k ∣ / ∣ Q 1 ∣ = ( n k ) / n |Q_k|/|Q_1| = \binom{n}{k}/n ∣ Q k ∣/∣ Q 1 ∣ = ( k n ) / n and k ( n k ) / n = ( n − 1 k − 1 ) k\binom{n}{k}/n = \binom{n-1}{k-1} k ( k n ) / n = ( k − 1 n − 1 ) .
Lemma 3.13 (Deficit lemma). Let T ϵ = P + H T^\epsilon= P + H T ϵ = P + H be a decomposition into uniform rank-one tensors, where P P P is the sum of all terms with ω 1 > 0 \omega_1 > 0 ω 1 > 0 and H H H the sum of the remaining terms. Then ω 3 ( P ) ≤ 3 ( 1 − ϵ ) ∣ Q 3 ∣ \omega_3(P) \le3(1-\epsilon)|Q_3| ω 3 ( P ) ≤ 3 ( 1 − ϵ ) ∣ Q 3 ∣ and consequently
ω 3 ( H ) ≥ 2 ϵ ∣ Q 3 ∣ . \omega_3(H) \ge2\epsilon|Q_3|. ω 3 ( H ) ≥ 2 ϵ ∣ Q 3 ∣. More generally ω k ( H ) ≥ ϵ ( k − 1 ) ∣ Q k ∣ \omega_k(H) \ge\epsilon(k-1)|Q_k| ω k ( H ) ≥ ϵ ( k − 1 ) ∣ Q k ∣ for every odd k ≥ 3 k \ge3 k ≥ 3 , and if G − v G-v G − v is connected for every vertex v v v this is tight for all k k k simultaneously.
Proof. Every term of H H H has zero Q 1 Q_1 Q 1 mass, so ω 1 ( P ) = ω 1 ( T ϵ ) \omega_1(P) = \omega_1(T^\epsilon) ω 1 ( P ) = ω 1 ( T ϵ ) , and the coefficient of T ϵ T^\epsilon T ϵ at every x ∈ Q 1 x \in Q_1 x ∈ Q 1 is k ( x ) − ϵ = 1 − ϵ k(x)-\epsilon= 1-\epsilon k ( x ) − ϵ = 1 − ϵ , so ω 1 ( P ) = ( 1 − ϵ ) ∣ Q 1 ∣ \omega_1(P) = (1-\epsilon)|Q_1| ω 1 ( P ) = ( 1 − ϵ ) ∣ Q 1 ∣ . Summing the all-strata bound (21) of Lemma 3.12 over the terms of P P P gives ω k ( P ) ≤ k ∣ Q k ∣ ∣ Q 1 ∣ ω 1 ( P ) = ( 1 − ϵ ) k ∣ Q k ∣ \omega_k(P) \le\frac{k|Q_k|}{|Q_1|}\omega_1(P) = (1-\epsilon)k|Q_k| ω k ( P ) ≤ ∣ Q 1 ∣ k ∣ Q k ∣ ω 1 ( P ) = ( 1 − ϵ ) k ∣ Q k ∣ . Since ω k ( T ϵ ) = ( k − ϵ ) ∣ Q k ∣ \omega_k(T^\epsilon) = (k-\epsilon)|Q_k| ω k ( T ϵ ) = ( k − ϵ ) ∣ Q k ∣ and H = T ϵ − P H = T^\epsilon- P H = T ϵ − P ,
ω k ( H ) ≥ ( k − ϵ − ( 1 − ϵ ) k ) ∣ Q k ∣ = ϵ ( k − 1 ) ∣ Q k ∣ , \omega_k(H) \ge\bigl(k-\epsilon-(1-\epsilon)k\bigr)|Q_k| = \epsilon(k-1)|Q_k|, ω k ( H ) ≥ ( k − ϵ − ( 1 − ϵ ) k ) ∣ Q k ∣ = ϵ ( k − 1 ) ∣ Q k ∣ ,
which for k = 3 k=3 k = 3 is the displayed bound.
Tightness, assuming G − v G-v G − v connected for every v v v . Refine each 1 [ v defective ] 1[v\text{ defective}] 1 [ v defective ] of (20) into subcubes, one for every assignment of δ ( v ) \delta(v) δ ( v ) violating the constraint at v v v , the other edges free, and decompose
T ϵ = ( 1 − ϵ ) T 0 + ϵ T 1 , T^\epsilon= (1-\epsilon)T^0+\epsilon T^1, T ϵ = ( 1 − ϵ ) T 0 + ϵ T 1 ,
with T 0 = χ ( T S G ) T^0=\chi(T_{\mathrm{SG}}) T 0 = χ ( T SG ) the left side of (20) and T 1 = χ ( T S G ) − χ ( 1 ) T^1=\chi(T_{\mathrm{SG}})-\chi(1) T 1 = χ ( T SG ) − χ ( 1 ) , as these subcubes scaled by 1 − ϵ 1-\epsilon 1 − ϵ together with any decomposition of ϵ T 1 \epsilon T^1 ϵ T 1 into uniform terms. Every subcube of the first group has positive Q 1 Q_1 Q 1 mass. Its free edges are exactly the edges of G − v G-v G − v , and the charges that G − v G-v G − v inherits once δ ( v ) \delta(v) δ ( v ) is pinned have even total, because the pinned pattern violates the constraint at v v v while ∑ u c ( u ) \sum_u c(u) ∑ u c ( u ) is odd; so Q 0 ( G − v ) ≠ ∅ Q_0(G-v) \ne\varnothing Q 0 ( G − v ) = ∅ by Lemma 3.9(2), and any such assignment of the free edges, taken together with the pinned pattern, is an assignment of G G G whose only defect is v v v , hence a point of Q 1 Q_1 Q 1 inside the subcube. Every term of the second group has ω 1 = 0 \omega_1=0 ω 1 = 0 , since the coefficients of T 1 T^1 T 1 vanish on Q 1 Q_1 Q 1 . The split therefore returns the two groups as P P P and H H H , and ω k ( H ) = ϵ ( k − 1 ) ∣ Q k ∣ \omega_k(H)=\epsilon(k-1)|Q_k| ω k ( H ) = ϵ ( k − 1 ) ∣ Q k ∣ for every odd k k k , which is the bound. Without the hypothesis the first group may contain subcubes with no one-defect mass — two copies of K 5 K_5 K 5 each missing an edge, joined through a new vertex v v v adjacent to the four endpoints, is connected, bridgeless and 4-regular while G − v G-v G − v is not — and those fall into H H H instead, so the decomposition no longer certifies equality.
Lemma 3.14. Let P P P and H H H be as in Lemma 3.13 and let W ϵ W^\epsilon W ϵ be the witness (19) , with its odd parameter k ∗ ≥ 5 k^* \ge5 k ∗ ≥ 5 . Then ⟨ W ϵ , H ⟩ ≥ 4 − k ∗ − ϵ k ∗ − 1 ≥ 1 \langle W^\epsilon,H\rangle\ge4-\frac{k^*-\epsilon}{k^*-1} \ge1 ⟨ W ϵ , H ⟩ ≥ 4 − k ∗ − 1 k ∗ − ϵ ≥ 1 .
Proof. H H H has no mass on Q 1 Q_1 Q 1 , so only the strata 3 and k ∗ k^* k ∗ contribute. By Lemma 3.13 , ω 3 ( H ) ≥ 2 ϵ ∣ Q 3 ∣ \omega_3(H) \ge2\epsilon|Q_3| ω 3 ( H ) ≥ 2 ϵ ∣ Q 3 ∣ , and H ≤ T ϵ H \le T^\epsilon H ≤ T ϵ entrywise gives ω k ∗ ( H ) ≤ ( k ∗ − ϵ ) ∣ Q k ∗ ∣ \omega_{k^*}(H) \le(k^*-\epsilon)|Q_{k^*}| ω k ∗ ( H ) ≤ ( k ∗ − ϵ ) ∣ Q k ∗ ∣ . Hence
⟨ W ϵ , H ⟩ ≥ 2 ϵ ∣ Q 3 ∣ ⋅ 2 ϵ ∣ Q 3 ∣ − ( k ∗ − ϵ ) ∣ Q k ∗ ∣ ( k ∗ − 1 ) ∣ Q k ∗ ∣ = 4 − k ∗ − ϵ k ∗ − 1 . \langle W^\epsilon,H\rangle\ge\frac{2}{\epsilon|Q_3|}\cdot2\epsilon|Q_3|-\frac{(k^*-\epsilon)|Q_{k^*}|}{(k^*-1)|Q_{k^*}|}=4-\frac{k^*-\epsilon}{k^*-1}. ⟨ W ϵ , H ⟩ ≥ ϵ ∣ Q 3 ∣ 2 ⋅ 2 ϵ ∣ Q 3 ∣ − ( k ∗ − 1 ) ∣ Q k ∗ ∣ ( k ∗ − ϵ ) ∣ Q k ∗ ∣ = 4 − k ∗ − 1 k ∗ − ϵ .
For k ∗ ≥ 5 k^* \ge5 k ∗ ≥ 5 and ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] we have k ∗ − ϵ k ∗ − 1 ≤ 5 4 \frac{k^*-\epsilon}{k^*-1} \le\frac{5}{4} k ∗ − 1 k ∗ − ϵ ≤ 4 5 , so the right side is at least 11 4 ≥ 1 \frac{11}{4} \ge1 4 11 ≥ 1 .
From a Q 3 Q_3 Q 3 -cover to a giant residual component
The heart of the argument is the estimate showing that a uniform term which contributes substantially to ω 3 \omega_3 ω 3 must also contribute so much to ω k ∗ \omega_{k^*} ω k ∗ that the witness sees it as nonpositive. This is where expansion enters.
Definition 3.15 (Pruning property). Let ρ 0 ≥ 0 \rho_0 \ge0 ρ 0 ≥ 0 . We say that G G G has the ρ 0 \rho_0 ρ 0 -pruning property if for every spanning subgraph G ′ ⊆ G G' \subseteq G G ′ ⊆ G whose cycle-rank drop ρ : = dim Z ( G ) − dim Z ( G ′ ) \rho:= \dim Z(G) - \dim Z(G') ρ := dim Z ( G ) − dim Z ( G ′ ) is at most ρ 0 \rho_0 ρ 0 , the graph G ′ G' G ′ has a (necessarily unique) component with more than n / 2 n/2 n /2 vertices.
Some hypothesis of this kind is needed: a chain of b ≥ 2 b \ge2 b ≥ 2 equally sized dense blocks joined consecutively by two edges — two, since a single joining edge would be a bridge — has no component above n / 2 n/2 n /2 once the 2 ( b − 1 ) 2(b-1) 2 ( b − 1 ) connecting edges are deleted, and that deletion creates b − 1 b-1 b − 1 new components, so its cycle-rank cost is only 2 ( b − 1 ) − ( b − 1 ) = b − 1 2(b-1) - (b-1) = b-1 2 ( b − 1 ) − ( b − 1 ) = b − 1 , a constant for constant b b b . Such a graph therefore fails the ρ 0 \rho_0 ρ 0 -pruning property for every ρ 0 ≥ b − 1 \rho_0 \ge b-1 ρ 0 ≥ b − 1 ; it cannot satisfy the property up to a linear cycle-rank drop. The next lemma shows that the property holds, with ρ 0 \rho_0 ρ 0 as large as a constant fraction of n n n , for expanders whose girth exceeds 4 / h 4/h 4/ h — for instance for the Ramanujan families of [24 ] , whose girth is Θ ( log n ) \Theta(\log n) Θ ( log n ) .
Lemma 3.16 (Pruning). Let G G G be connected, bridgeless and d d d -regular with d ≥ 3 d \ge3 d ≥ 3 , of edge expansion h > 0 h > 0 h > 0 and girth g > 4 / h g > 4/h g > 4/ h , where g g g is the length of a shortest cycle. Then G G G has the ρ 0 \rho_0 ρ 0 -pruning property for every ρ 0 < min { h , 2 } n / 8 \rho_0 < \min\{h,2\}n/8 ρ 0 < min { h , 2 } n /8 .
Proof. Let G ′ G' G ′ be spanning with components X 1 , … , X p X_1,\ldots,X_p X 1 , … , X p of sizes s 1 , … , s p s_1,\ldots,s_p s 1 , … , s p , and suppose for contradiction that s i ≤ n / 2 s_i \le n/2 s i ≤ n /2 for every i i i . Every edge of G G G leaving X i X_i X i is absent from G ′ G' G ′ ; writing F F F for the set of edges of G G G not in G ′ G' G ′ we have ∣ F ∣ ≥ 1 2 ∑ i ∣ δ ( X i ) ∣ |F| \ge\frac{1}{2}\sum_i |\delta(X_i)| ∣ F ∣ ≥ 2 1 ∑ i ∣ δ ( X i ) ∣ , and
ρ = dim Z ( G ) − dim Z ( G ′ ) = ( ∣ E ∣ − n + 1 ) − ( ∣ E ∣ − ∣ F ∣ − n + p ) = ∣ F ∣ − p + 1 ≥ ∑ i = 1 p ( 1 2 ∣ δ ( X i ) ∣ − 1 ) + 1. \rho= \dim Z(G) - \dim Z(G') = (|E|-n+1)-(|E|-|F|-n+p)
= |F|-p+1 \ge\sum_{i=1}^{p}\left(\frac{1}{2}|\delta(X_i)|-1\right)+1. ρ = dim Z ( G ) − dim Z ( G ′ ) = ( ∣ E ∣ − n + 1 ) − ( ∣ E ∣ − ∣ F ∣ − n + p ) = ∣ F ∣ − p + 1 ≥ i = 1 ∑ p ( 2 1 ∣ δ ( X i ) ∣ − 1 ) + 1.
Since G G G is bridgeless, ∣ δ ( X i ) ∣ ≥ 2 |\delta(X_i)| \ge2 ∣ δ ( X i ) ∣ ≥ 2 and every summand is nonnegative. Since s i ≤ n / 2 s_i \le n/2 s i ≤ n /2 , expansion gives ∣ δ ( X i ) ∣ ≥ h s i |\delta(X_i)| \ge hs_i ∣ δ ( X i ) ∣ ≥ h s i .
Call X i X_i X i large if s i ≥ 4 / h s_i \ge4/h s i ≥ 4/ h and small otherwise. For a large component 1 2 ∣ δ ( X i ) ∣ − 1 ≥ h 2 s i − 1 ≥ h 4 s i \frac{1}{2}|\delta(X_i)|-1 \ge\frac{h}{2}s_i-1 \ge\frac{h}{4}s_i 2 1 ∣ δ ( X i ) ∣ − 1 ≥ 2 h s i − 1 ≥ 4 h s i . For a small component we have s i < 4 / h < g s_i < 4/h < g s i < 4/ h < g , so G [ X i ] G[X_i] G [ X i ] contains no cycle and, being connected, is a tree with s i − 1 s_i-1 s i − 1 edges; d d d -regularity then gives ∣ δ ( X i ) ∣ = d s i − 2 ( s i − 1 ) ≥ s i + 2 |\delta(X_i)| = ds_i - 2(s_i-1) \ge s_i+2 ∣ δ ( X i ) ∣ = d s i − 2 ( s i − 1 ) ≥ s i + 2 , whence 1 2 ∣ δ ( X i ) ∣ − 1 ≥ 1 2 s i \frac{1}{2}|\delta(X_i)|-1 \ge\frac{1}{2}s_i 2 1 ∣ δ ( X i ) ∣ − 1 ≥ 2 1 s i . Consequently
ρ ≥ h 4 ∑ l a r g e s i and ρ ≥ 1 2 ∑ s m a l l s i . \rho\ge\frac{h}{4}\sum_{\mathrm{large}} s_i
\qquad\text{and}\qquad
\rho\ge\frac{1}{2}\sum_{\mathrm{small}} s_i. ρ ≥ 4 h large ∑ s i and ρ ≥ 2 1 small ∑ s i .
and since the two sums add up to n n n , one of them is at least n / 2 n/2 n /2 , so
ρ ≥ min { h / 4 , 1 / 2 } ⋅ n 2 = min { h , 2 } n 8 > ρ 0 , \rho\ge\min\{h/4,1/2\}\cdot\frac{n}{2}
= \frac{\min\{h,2\}n}{8} > \rho_0, ρ ≥ min { h /4 , 1/2 } ⋅ 2 n = 8 min { h , 2 } n > ρ 0 ,
contradicting ρ ≤ ρ 0 \rho\le\rho_0 ρ ≤ ρ 0 . Hence some s i > n / 2 s_i > n/2 s i > n /2 ; uniqueness is immediate, two components of size > n / 2 >n/2 > n /2 being impossible.
Lemma 3.17 (Cost of separating a giant component). Let G G G be connected and bridgeless with edge expansion h > 0 h > 0 h > 0 , let B ⊆ V B \subseteq V B ⊆ V induce a connected subgraph with ∣ B ∣ > n / 2 |B| > n/2 ∣ B ∣ > n /2 , and put X : = V ∖ B X := V \setminus B X := V ∖ B . Then every F ⊆ E F \subseteq E F ⊆ E containing E ( X , B ) E(X,B) E ( X , B ) satisfies
dim Z ( G ) − dim Z ( G − F ) ≥ h 2 ∣ X ∣ . \dim Z(G) - \dim Z(G-F) \ge\frac{h}{2}|X|. dim Z ( G ) − dim Z ( G − F ) ≥ 2 h ∣ X ∣.
Proof. Deleting edges only lowers dim Z \dim Z dim Z , so we may take F = E ( X , B ) F=E(X,B) F = E ( X , B ) . Let X 1 , … , X q X_1,\ldots,X_q X 1 , … , X q be the components of G [ X ] G[X] G [ X ] ; each has all its external edges going to B B B , and B B B is connected, so G − F G-F G − F is the disjoint union of B , X 1 , … , X q B,X_1,\ldots,X_q B , X 1 , … , X q and
dim Z ( G ) − dim Z ( G − F ) = ∣ F ∣ − q = ∑ i = 1 q ( ∣ E ( X i , B ) ∣ − 1 ) . \dim Z(G)-\dim Z(G-F)=|F|-q=\sum_{i=1}^{q}\left(|E(X_i,B)|-1\right). dim Z ( G ) − dim Z ( G − F ) = ∣ F ∣ − q = i = 1 ∑ q ( ∣ E ( X i , B ) ∣ − 1 ) .
Each X i X_i X i has ∣ X i ∣ ≤ ∣ X ∣ < n / 2 |X_i|\leq|X|<n/2 ∣ X i ∣ ≤ ∣ X ∣ < n /2 , so expansion gives ∣ E ( X i , B ) ∣ ≥ h ∣ X i ∣ |E(X_i,B)|\geq h|X_i| ∣ E ( X i , B ) ∣ ≥ h ∣ X i ∣ , while bridgelessness gives ∣ E ( X i , B ) ∣ ≥ 2 |E(X_i,B)|\geq2 ∣ E ( X i , B ) ∣ ≥ 2 ; hence ∣ E ( X i , B ) ∣ − 1 ≥ 1 2 ∣ E ( X i , B ) ∣ ≥ h 2 ∣ X i ∣ |E(X_i,B)|-1\geq\frac{1}{2}|E(X_i,B)|\geq\frac{h}{2}|X_i| ∣ E ( X i , B ) ∣ − 1 ≥ 2 1 ∣ E ( X i , B ) ∣ ≥ 2 h ∣ X i ∣ . Summing over i i i gives the claim. □ \square □
Lemma 3.18 (Defect-ratio estimate). Let k ∗ k^* k ∗ be odd with 5 ≤ k ∗ ≤ n / 2 5\leq k^*\leq n/2 5 ≤ k ∗ ≤ n /2 and let R R R be a uniform rank-one tensor whose residual graph has exactly three odd components and a component B B B with ∣ B ∣ ≥ n − n / k ∗ |B|\geq n-n/k^* ∣ B ∣ ≥ n − n / k ∗ . Then
∣ Q 3 ∣ ⋅ ∣ Q k ∗ ( R ) ∣ ∣ Q k ∗ ∣ ⋅ ∣ Q 3 ( R ) ∣ ≥ e − 4 k ∗ ( k ∗ − 1 ) 6 . \frac{|Q_3|\cdot|Q_{k^*}(R)|}{|Q_{k^*}|\cdot|Q_3(R)|}\geq e^{-4}\frac{k^*(k^*-1)}{6}. ∣ Q k ∗ ∣ ⋅ ∣ Q 3 ( R ) ∣ ∣ Q 3 ∣ ⋅ ∣ Q k ∗ ( R ) ∣ ≥ e − 4 6 k ∗ ( k ∗ − 1 ) .
Proof. Write C 1 , C 2 , C 3 C_1,C_2,C_3 C 1 , C 2 , C 3 for the three odd components, X : = V ∖ B X:=V\setminus B X := V ∖ B and m : = ∣ B ∣ m:=|B| m := ∣ B ∣ , so that ∣ X ∣ = n − m ≤ n / k ∗ |X|=n-m\leq n/k^* ∣ X ∣ = n − m ≤ n / k ∗ and m ≥ n − n / k ∗ ≥ 4 5 n ≥ n / 2 ≥ k ∗ m\geq n-n/k^*\geq\frac{4}{5}n\geq n/2\geq k^* m ≥ n − n / k ∗ ≥ 5 4 n ≥ n /2 ≥ k ∗ , using k ∗ ≥ 5 k^*\geq5 k ∗ ≥ 5 and k ∗ ≤ n / 2 k^*\leq n/2 k ∗ ≤ n /2 . At most one of C 1 , C 2 , C 3 C_1,C_2,C_3 C 1 , C 2 , C 3 equals B B B , so X X X contains at least two of them; without loss of generality C 1 , … , C t C_1,\ldots,C_t C 1 , … , C t with t ∈ { 2 , 3 } t\in\{2,3\} t ∈ { 2 , 3 } lie in X X X , and B B B is either the remaining odd component C 3 C_3 C 3 (if t = 2 t=2 t = 2 ) or an even component (if t = 3 t=3 t = 3 ).
Case A: B = C 3 B=C_3 B = C 3 . We may place one defect in each of C 1 , C 2 C_1,C_2 C 1 , C 2 and k ∗ − 2 k^*-2 k ∗ − 2 defects in C 3 C_3 C 3 — a valid pattern, k ∗ − 2 k^*-2 k ∗ − 2 being odd and matching the odd residual charge of C 3 C_3 C 3 — and none in the even components. The 3-defect support points are exactly one defect per odd component, so the shared factors ∣ Q 1 ( C 1 ) ∣ , ∣ Q 1 ( C 2 ) ∣ |Q_1(C_1)|,|Q_1(C_2)| ∣ Q 1 ( C 1 ) ∣ , ∣ Q 1 ( C 2 ) ∣ and those of the even components cancel and Lemma 3.9 gives
∣ Q k ∗ ( R ) ∣ ∣ Q 3 ( R ) ∣ ≥ ∣ Q k ∗ − 2 ( C 3 ) ∣ ∣ Q 1 ( C 3 ) ∣ = ( m k ∗ − 2 ) m = ( m − 1 ) ( m − 2 ) ⋯ ( m − k ∗ + 3 ) ( k ∗ − 2 ) ! . \frac{|Q_{k^*}(R)|}{|Q_3(R)|}\geq\frac{|Q_{k^*-2}(C_3)|}{|Q_1(C_3)|}=\frac{\binom{m}{k^*-2}}{m}=\frac{(m-1)(m-2)\cdots(m-k^*+3)}{(k^*-2)!}. ∣ Q 3 ( R ) ∣ ∣ Q k ∗ ( R ) ∣ ≥ ∣ Q 1 ( C 3 ) ∣ ∣ Q k ∗ − 2 ( C 3 ) ∣ = m ( k ∗ − 2 m ) = ( k ∗ − 2 )! ( m − 1 ) ( m − 2 ) ⋯ ( m − k ∗ + 3 ) .
Using ( n 3 ) / ( n k ∗ ) = k ∗ ! 6 ⋅ 1 ( n − 3 ) ⋯ ( n − k ∗ + 1 ) \binom{n}{3}/\binom{n}{k^*}=\frac{k^*!}{6}\cdot\frac{1}{(n-3)\cdots(n-k^*+1)} ( 3 n ) / ( k ∗ n ) = 6 k ∗ ! ⋅ ( n − 3 ) ⋯ ( n − k ∗ + 1 ) 1 and k ∗ ! / ( k ∗ − 2 ) ! = k ∗ ( k ∗ − 1 ) k^*!/(k^*-2)!=k^*(k^*-1) k ∗ ! / ( k ∗ − 2 )! = k ∗ ( k ∗ − 1 ) ,
∣ Q 3 ∣ ⋅ ∣ Q k ∗ ( R ) ∣ ∣ Q k ∗ ∣ ⋅ ∣ Q 3 ( R ) ∣ ≥ k ∗ ( k ∗ − 1 ) 6 ⋅ ( m − 1 ) ⋯ ( m − k ∗ + 3 ) ( n − 3 ) ⋯ ( n − k ∗ + 1 ) . \frac{|Q_3|\cdot|Q_{k^*}(R)|}{|Q_{k^*}|\cdot|Q_3(R)|}\geq\frac{k^*(k^*-1)}{6}\cdot\frac{(m-1)\cdots(m-k^*+3)}{(n-3)\cdots(n-k^*+1)}. ∣ Q k ∗ ∣ ⋅ ∣ Q 3 ( R ) ∣ ∣ Q 3 ∣ ⋅ ∣ Q k ∗ ( R ) ∣ ≥ 6 k ∗ ( k ∗ − 1 ) ⋅ ( n − 3 ) ⋯ ( n − k ∗ + 1 ) ( m − 1 ) ⋯ ( m − k ∗ + 3 ) .
The second factor is a product of k ∗ − 3 k^*-3 k ∗ − 3 terms, each m − 1 − j n − 3 − j ≥ 1 − ∣ X ∣ n − k ∗ + 1 ≥ 1 − 2 ∣ X ∣ n \frac{m-1-j}{n-3-j}\geq1-\frac{|X|}{n-k^*+1}\geq1-\frac{2|X|}{n} n − 3 − j m − 1 − j ≥ 1 − n − k ∗ + 1 ∣ X ∣ ≥ 1 − n 2∣ X ∣ (using j ≤ k ∗ − 4 j\leq k^*-4 j ≤ k ∗ − 4 and k ∗ ≤ n / 2 k^*\leq n/2 k ∗ ≤ n /2 ), so it is at least ( 1 − 2 ∣ X ∣ n ) k ∗ ≥ e − 4 ∣ X ∣ k ∗ / n ≥ e − 4 \left(1-\frac{2|X|}{n}\right)^{k^*}\geq e^{-4|X|k^*/n}\geq e^{-4} ( 1 − n 2∣ X ∣ ) k ∗ ≥ e − 4∣ X ∣ k ∗ / n ≥ e − 4 , since ∣ X ∣ ≤ n / k ∗ |X|\leq n/k^* ∣ X ∣ ≤ n / k ∗ and 1 − u ≥ e − 2 u 1-u\geq e^{-2u} 1 − u ≥ e − 2 u for u ∈ [ 0 , 1 2 ] u\in[0,\frac{1}{2}] u ∈ [ 0 , 2 1 ] .
Case B: B B B even. All three odd components lie in X X X ; placing one defect in each C i C_i C i and k ∗ − 3 k^*-3 k ∗ − 3 defects in B B B (valid, k ∗ − 3 k^*-3 k ∗ − 3 being even) and none elsewhere gives ∣ Q k ∗ ( R ) ∣ / ∣ Q 3 ( R ) ∣ ≥ ∣ Q k ∗ − 3 ( B ) ∣ / ∣ Q 0 ( B ) ∣ = ( m k ∗ − 3 ) |Q_{k^*}(R)|/|Q_3(R)|\geq|Q_{k^*-3}(B)|/|Q_0(B)|=\binom{m}{k^*-3} ∣ Q k ∗ ( R ) ∣/∣ Q 3 ( R ) ∣ ≥ ∣ Q k ∗ − 3 ( B ) ∣/∣ Q 0 ( B ) ∣ = ( k ∗ − 3 m ) , whence
∣ Q 3 ∣ ⋅ ∣ Q k ∗ ( R ) ∣ ∣ Q k ∗ ∣ ⋅ ∣ Q 3 ( R ) ∣ ≥ k ∗ ( k ∗ − 1 ) ( k ∗ − 2 ) 6 ⋅ m ( m − 1 ) ⋯ ( m − k ∗ + 4 ) ( n − 3 ) ( n − 4 ) ⋯ ( n − k ∗ + 1 ) . \frac{|Q_3|\cdot|Q_{k^*}(R)|}{|Q_{k^*}|\cdot|Q_3(R)|}\geq\frac{k^*(k^*-1)(k^*-2)}{6}\cdot\frac{m(m-1)\cdots(m-k^*+4)}{(n-3)(n-4)\cdots(n-k^*+1)}. ∣ Q k ∗ ∣ ⋅ ∣ Q 3 ( R ) ∣ ∣ Q 3 ∣ ⋅ ∣ Q k ∗ ( R ) ∣ ≥ 6 k ∗ ( k ∗ − 1 ) ( k ∗ − 2 ) ⋅ ( n − 3 ) ( n − 4 ) ⋯ ( n − k ∗ + 1 ) m ( m − 1 ) ⋯ ( m − k ∗ + 4 ) .
a product of k ∗ − 3 k^*-3 k ∗ − 3 ratios each at least 1 − 2 ∣ X ∣ n 1-\frac{2|X|}{n} 1 − n 2∣ X ∣ by the same estimate, so at least e − 4 k ∗ ( k ∗ − 1 ) ( k ∗ − 2 ) / 6 e^{-4k^*(k^*-1)(k^*-2)/6} e − 4 k ∗ ( k ∗ − 1 ) ( k ∗ − 2 ) /6 — stronger than needed. □ \square □
Lemma 3.19 (Q 3 Q_3 Q 3 mass forces a giant residual component). Let G G G be connected, bridgeless, d d d -regular with d ≥ 3 d\geq3 d ≥ 3 and edge expansion h > 0 h>0 h > 0 , let ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] and τ ≥ 2 \tau\geq2 τ ≥ 2 , and put
K : = log 2 τ + log 2 3 − ϵ 2 ϵ , Λ : = 2 K h . K:=\log_2\tau+\log_2\frac{3-\epsilon}{2\epsilon},\qquad\Lambda:=\frac{2K}{h}. K := log 2 τ + log 2 2 ϵ 3 − ϵ , Λ := h 2 K .
Assume G G G has the K K K -pruning property. Then every uniform rank-one tensor R R R with R ≤ T ε R \le T^{\varepsilon} R ≤ T ε entrywise, ω 1 ( R ) = 0 \omega_{1}(R) = 0 ω 1 ( R ) = 0 and ω 3 ( R ) ≥ 2 ε ∣ Q 3 ∣ / τ \omega_{3}(R) \ge2\varepsilon|Q_{3}|/\tau ω 3 ( R ) ≥ 2 ε ∣ Q 3 ∣/ τ has exactly three odd residual components, and G ′ ( R ) G'(R) G ′ ( R ) has a component B B B with ∣ V ∖ B ∣ ≤ Λ |V \setminus B| \le\Lambda ∣ V ∖ B ∣ ≤ Λ .
Proof. By Proposition 3.10, ω 1 ( R ) = 0 \omega_{1}(R) = 0 ω 1 ( R ) = 0 gives ℓ ≥ 3 \ell\ge3 ℓ ≥ 3 , while ω 3 ( R ) > 0 \omega_{3}(R) > 0 ω 3 ( R ) > 0 forces the minimum defect count ℓ ≤ 3 \ell\le3 ℓ ≤ 3 . So G ′ ( R ) G'(R) G ′ ( R ) has exactly three odd components C 1 , C 2 , C 3 C_{1}, C_{2}, C_{3} C 1 , C 2 , C 3 , the remaining components { D j } \{D_{j}\} { D j } being even.
Step 1: the cycle-rank drop is at most K K K . By (18) every support point of R R R has the same value, and [ x T ] T ε = 3 − ε [\mathbf{x}^{T}]T^{\varepsilon} = 3 - \varepsilon [ x T ] T ε = 3 − ε for x ∈ Q 3 x \in Q_{3} x ∈ Q 3 , so R ≤ T ε R \le T^{\varepsilon} R ≤ T ε gives η 2 − ∣ E ( G ′ ( R ) ) ∣ ≤ 3 − ε \eta2^{-|E(G'(R))|} \le3 - \varepsilon η 2 − ∣ E ( G ′ ( R )) ∣ ≤ 3 − ε and hence ω 3 ( R ) ≤ ( 3 − ε ) ∣ Q 3 ( R ) ∣ \omega_{3}(R) \le(3 - \varepsilon)|Q_{3}(R)| ω 3 ( R ) ≤ ( 3 − ε ) ∣ Q 3 ( R ) ∣ . Counting the 3-defect support points by components (one defect in each C i C_{i} C i , none in the D j D_{j} D j ) and using Lemma 3.9 within each component,
∣ Q 3 ( R ) ∣ = ∣ C 1 ∣ ∣ C 2 ∣ ∣ C 3 ∣ ⋅ 2 dim Z ( G ′ ( R ) ) = ∣ Q 3 ∣ ⋅ ∣ C 1 ∣ ∣ C 2 ∣ ∣ C 3 ∣ ( n 3 ) ⋅ 2 − ρ , |Q_{3}(R)| = |C_{1}||C_{2}||C_{3}| \cdot2^{\dim Z(G'(R))} = |Q_{3}| \cdot\frac{|C_{1}||C_{2}||C_{3}|}{\binom{n}{3}} \cdot2^{-\rho}, ∣ Q 3 ( R ) ∣ = ∣ C 1 ∣∣ C 2 ∣∣ C 3 ∣ ⋅ 2 d i m Z ( G ′ ( R )) = ∣ Q 3 ∣ ⋅ ( 3 n ) ∣ C 1 ∣∣ C 2 ∣∣ C 3 ∣ ⋅ 2 − ρ ,
where
ρ : = dim Z ( G ) − dim Z ( G ′ ( R ) ) ≥ 0 , \rho:= \dim Z(G) - \dim Z(G'(R)) \ge0, ρ := dim Z ( G ) − dim Z ( G ′ ( R )) ≥ 0 ,
where we also used ∣ Q 3 ∣ = ( n 3 ) 2 dim Z ( G ) |Q_{3}| = \binom{n}{3}2^{\dim Z(G)} ∣ Q 3 ∣ = ( 3 n ) 2 d i m Z ( G ) and Z ( G ′ ( R ) ) ⊆ Z ( G ) Z(G'(R)) \subseteq Z(G) Z ( G ′ ( R )) ⊆ Z ( G ) . Since ∣ C 1 ∣ ∣ C 2 ∣ ∣ C 3 ∣ ≤ ( n 3 ) |C_{1}||C_{2}||C_{3}| \le\binom{n}{3} ∣ C 1 ∣∣ C 2 ∣∣ C 3 ∣ ≤ ( 3 n ) , combining with the hypothesis ω 3 ( R ) ≥ 2 ε ∣ Q 3 ∣ / τ \omega_{3}(R) \ge2\varepsilon|Q_{3}|/\tau ω 3 ( R ) ≥ 2 ε ∣ Q 3 ∣/ τ gives
2 ε ∣ Q 3 ∣ τ ≤ ( 3 − ε ) ∣ Q 3 ∣ 2 − ρ , \frac{2\varepsilon|Q_{3}|}{\tau} \le(3-\varepsilon)|Q_{3}|2^{-\rho}, τ 2 ε ∣ Q 3 ∣ ≤ ( 3 − ε ) ∣ Q 3 ∣ 2 − ρ ,
i.e.
ρ ≤ log 2 τ + log 2 3 − ε 2 ε = K . \rho\le\log_{2}\tau+ \log_{2}\frac{3-\varepsilon}{2\varepsilon} = K. ρ ≤ log 2 τ + log 2 2 ε 3 − ε = K .
Step 2: a giant residual component, and the ratio. Since ρ ≤ K \rho\le K ρ ≤ K , the K K K -pruning property supplies a unique component B B B of G ′ ( R ) G'(R) G ′ ( R ) with ∣ B ∣ > n / 2 |B| > n/2 ∣ B ∣ > n /2 . Set X : = V ∖ B X := V \setminus B X := V ∖ B and F : = E ( X , B ) F := E(X,B) F := E ( X , B ) . No edge of G ′ ( R ) G'(R) G ′ ( R ) crosses F F F , so Z ( G ′ ( R ) ) ⊆ Z ( G − F ) Z(G'(R)) \subseteq Z(G - F) Z ( G ′ ( R )) ⊆ Z ( G − F ) and Lemma 3.17 gives
h 2 ∣ X ∣ ≤ dim Z ( G ) − dim Z ( G − F ) ≤ ρ ≤ K , \frac{h}{2}|X| \le\dim Z(G) - \dim Z(G-F) \le\rho\le K, 2 h ∣ X ∣ ≤ dim Z ( G ) − dim Z ( G − F ) ≤ ρ ≤ K ,
that is
∣ X ∣ ≤ 2 K h = Λ . |X| \le\frac{2K}{h} = \Lambda. ∣ X ∣ ≤ h 2 K = Λ.
This is the assertion. □
Lemma 3.20 (Local sign of the Tseitin witness). Let k ∗ ≥ 24 e 4 / ε k^{*} \ge24e^{4}/\varepsilon k ∗ ≥ 24 e 4 / ε be odd with k ∗ ≤ n / 2 k^{*} \le n/2 k ∗ ≤ n /2 , and let R ≤ T ε R \le T^{\varepsilon} R ≤ T ε be a uniform rank-one term with ω 1 ( R ) = 0 \omega_{1}(R) = 0 ω 1 ( R ) = 0 . Then ⟨ W ε , R ⟩ ≤ 0 \langle W^{\varepsilon}, R\rangle\le0 ⟨ W ε , R ⟩ ≤ 0 in each of the following two cases: the residual graph of R R R has at least five odd components; or it has exactly three, and some component B B B satisfies
∣ V ∖ B ∣ ≤ n / k ∗ . (22) |V \setminus B| \le n/k^{*}. \tag*{(22)} ∣ V ∖ B ∣ ≤ n / k ∗ . (22)
Proof. If there are at least five odd components then ω 3 ( R ) = 0 \omega_{3}(R) = 0 ω 3 ( R ) = 0 by Proposition 3.10, while the only other finite nonzero coefficient of W ε W^{\varepsilon} W ε is negative.
Suppose there are exactly three odd components. Then (22) is the hypothesis ∣ B ∣ ≥ n − n / k ∗ |B| \ge n - n/k^{*} ∣ B ∣ ≥ n − n / k ∗ of Lemma 3.18 , which gives
∣ Q 3 ∣ ∣ Q k ∗ ( R ) ∣ ∣ Q k ∗ ∣ ∣ Q 3 ( R ) ∣ ≥ e − 4 k ∗ ( k ∗ − 1 ) 6 , \frac{|Q_{3}||Q_{k^{*}}(R)|}{|Q_{k^{*}}||Q_{3}(R)|} \ge e^{-4}\frac{k^{*}(k^{*}-1)}{6}, ∣ Q k ∗ ∣∣ Q 3 ( R ) ∣ ∣ Q 3 ∣∣ Q k ∗ ( R ) ∣ ≥ e − 4 6 k ∗ ( k ∗ − 1 ) ,
and since R R R is uniform its masses have the same ratio as the corresponding support sizes by (18) , so
⟨ W ε , R ⟩ ≤ ω 3 ( R ) ∣ Q 3 ∣ ( 2 ε − e − 4 k ∗ 6 ) ≤ 0 \langle W^{\varepsilon}, R\rangle\le\frac{\omega_{3}(R)}{|Q_{3}|}\left(\frac{2}{\varepsilon} - \frac{e^{-4}k^{*}}{6}\right) \le0 ⟨ W ε , R ⟩ ≤ ∣ Q 3 ∣ ω 3 ( R ) ( ε 2 − 6 e − 4 k ∗ ) ≤ 0
by k ∗ ≥ 24 e 4 / ε k^{*} \ge24e^{4}/\varepsilon k ∗ ≥ 24 e 4 / ε . □
:::
The bounded-bias lower bound
Theorem 3.21 (Bounded-bias Tseitin lower bound). There is an absolute constant c 0 > 0 c_0 > 0 c 0 > 0 such that the following holds. Let G = ( V , E ) G = (V,E) G = ( V , E ) be connected, bridgeless and d d d -regular with d ≥ 3 d \ge3 d ≥ 3 , on n n n vertices, with edge expansion h > 0 h > 0 h > 0 ; write h ˉ : = min { h , 1 } \bar h := \min\{h,1\} h ˉ := min { h , 1 } ; let the Tseitin charge have odd total parity; and let ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] . Let τ ≥ 2 \tau\ge2 τ ≥ 2 satisfy
log 2 τ + log 2 2 ϵ ≤ c 0 h ˉ n , (23) \log_2 \tau+ \log_2 \frac{2}{\epsilon} \le c_0 \bar h n, \tag*{(23)} log 2 τ + log 2 ϵ 2 ≤ c 0 h ˉ n , (23) and assume G G G has the K K K -pruning property for K = log 2 τ + log 2 3 − ϵ 2 ϵ K = \log_2 \tau+ \log_2 \frac{3-\epsilon}{2\epsilon} K = log 2 τ + log 2 2 ϵ 3 − ϵ , which holds in particular when the girth of G G G exceeds 4 / h 4/h 4/ h . Then every decomposition T ϵ = ∑ j = 1 r R j T^\epsilon= \sum_{j=1}^{r} R_j T ϵ = ∑ j = 1 r R j into uniform (bounded-bias) rank-one tensors has r ≥ τ / 4 r \ge\tau/4 r ≥ τ /4 . In particular, if the girth exceeds 4 / h 4/h 4/ h and ϵ h ˉ n ≥ C log 2 ( 2 / ϵ ) \epsilon\bar h n \ge C\log_2(2/\epsilon) ϵ h ˉ n ≥ C log 2 ( 2/ ϵ ) for a sufficiently large absolute constant C C C , then
rk + , b ⊗ ( χ ( T S G ) − ϵ ⋅ χ ( 1 ) ) = 2 Ω ( ϵ h ˉ n ) , \operatorname{rk}^{\otimes}_{+,b}\bigl(\chi(TS_G)-\epsilon\cdot\chi(1)\bigr)=2^{\Omega(\epsilon\bar h n)}, rk + , b ⊗ ( χ ( T S G ) − ϵ ⋅ χ ( 1 ) ) = 2 Ω ( ϵ h ˉ n ) , and hence every Σ Π Σ - A M C ϵ \Sigma\Pi\Sigma\text{-}\mathrm{AMC}^{\epsilon} ΣΠΣ - AMC ϵ proof of T S G TS_G T S G has this lower bound. For bounded degree, h ≤ d h \le d h ≤ d gives h ˉ = Θ ( h ) \bar h = \Theta(h) h ˉ = Θ ( h ) , and the bound reads 2 Ω ( ϵ h n ) 2^{\Omega(\epsilon h n)} 2 Ω ( ϵ hn ) .
Proof. Take c 0 : = 1 / ( 200 e 4 ) c_0 := 1/(200e^4) c 0 := 1/ ( 200 e 4 ) and let k ∗ k^* k ∗ be the least odd integer with k ∗ ≥ 24 e 4 / ϵ k^* \ge24e^4/\epsilon k ∗ ≥ 24 e 4 / ϵ ; then k ∗ ≤ 25 e 4 / ϵ k^* \le25e^4/\epsilon k ∗ ≤ 25 e 4 / ϵ . Since τ ≥ 2 \tau\ge2 τ ≥ 2 and ( 3 − ϵ ) / ( 2 ϵ ) ≥ 1 (3-\epsilon)/(2\epsilon) \ge1 ( 3 − ϵ ) / ( 2 ϵ ) ≥ 1 for ϵ ≤ 1 \epsilon\le1 ϵ ≤ 1 , we have K ≥ 1 K \ge1 K ≥ 1 , and (23) gives K ≤ log 2 τ + log 2 2 ϵ ≤ c 0 h ˉ n K \le\log_2\tau+\log_2\frac{2}{\epsilon} \le c_0\bar h n K ≤ log 2 τ + log 2 ϵ 2 ≤ c 0 h ˉ n . In particular K ≤ c 0 h ˉ n < min { h , 2 } n / 8 K \le c_0\bar h n < \min\{h,2\}n/8 K ≤ c 0 h ˉ n < min { h , 2 } n /8 , since c 0 < 1 / 8 c_0 < 1/8 c 0 < 1/8 and h ˉ = min { h , 1 } ≤ min { h , 2 } \bar h=\min\{h,1\}\le\min\{h,2\} h ˉ = min { h , 1 } ≤ min { h , 2 } , so Lemma 3.16 supplies the K K K -pruning property whenever the girth exceeds 4 / h 4/h 4/ h . Hence Λ : = 2 K / h ˉ ≤ 2 c 0 ϵ n \Lambda:= 2K/\bar h \le2c_0\epsilon n Λ := 2 K / h ˉ ≤ 2 c 0 ϵ n and Λ ≥ 2 K / h ˉ ≥ 2 \Lambda\ge2K/\bar h \ge2 Λ ≥ 2 K / h ˉ ≥ 2 , so
k ∗ Λ ≤ 25 e 4 ϵ ⋅ 2 c 0 ϵ n = 50 e 4 c 0 n = n 4 ≤ n , k ∗ ≤ k ∗ Λ 2 ≤ n 8 ≤ n 2 , k^*\Lambda\le\frac{25e^4}{\epsilon}\cdot2c_0\epsilon n = 50e^4c_0n = \frac{n}{4}\le n,
\qquad
k^* \le\frac{k^*\Lambda}{2}\le\frac{n}{8}\le\frac{n}{2}, k ∗ Λ ≤ ϵ 25 e 4 ⋅ 2 c 0 ϵ n = 50 e 4 c 0 n = 4 n ≤ n , k ∗ ≤ 2 k ∗ Λ ≤ 8 n ≤ 2 n ,
so k ∗ k^* k ∗ and Λ \Lambda Λ satisfy the hypotheses of Lemma 3.19 — note that the Λ \Lambda Λ appearing there is 2 K / h ≤ 2 K / h ˉ 2K/h \le2K/\bar h 2 K / h ≤ 2 K / h ˉ , so the bounds just derived apply a fortiori — and we use the witness W ϵ W^\epsilon W ϵ of (19).
Step 1: eliminate the one-defect terms. Let P P P be the sum of those R j R_j R j with ω 1 ( R j ) > 0 \omega_1(R_j)>0 ω 1 ( R j ) > 0 and H H H the sum of the rest, so T ϵ = P + H T^\epsilon=P+H T ϵ = P + H . By Lemma 3.14, ⟨ W ϵ , H ⟩ ≥ 1 \langle W^\epsilon,H\rangle\ge1 ⟨ W ϵ , H ⟩ ≥ 1 ; the functional is applied to H H H only, as (19) records.
Step 2: bound each remaining term. Every R j R_j R j occurring in H H H satisfies ω 1 ( R j ) = 0 \omega_1(R_j)=0 ω 1 ( R j ) = 0 and R j ≤ T ϵ R_j\le T^\epsilon R j ≤ T ϵ entrywise. Write ℓ j \ell_j ℓ j for its number of odd residual components; by Proposition 3.10, ℓ j ≥ 3 \ell_j\ge3 ℓ j ≥ 3 and ℓ j \ell_j ℓ j is odd. We distinguish three cases.
ℓ j = 3 \ell_j=3 ℓ j = 3 and ω 3 ( R j ) ≥ 2 ϵ ∣ Q 3 ∣ / τ \omega_3(R_j)\ge2\epsilon|Q_3|/\tau ω 3 ( R j ) ≥ 2 ϵ ∣ Q 3 ∣/ τ . Lemma 3.19 supplies a component B B B of G ′ ( R j ) G'(R_j) G ′ ( R j ) with ∣ V ∖ B ∣ ≤ Λ ≤ n / k ∗ |V\setminus B|\le\Lambda\le n/k^* ∣ V ∖ B ∣ ≤ Λ ≤ n / k ∗ , so again ⟨ W ϵ , R j ⟩ ≤ 0 \langle W^\epsilon,R_j\rangle\le0 ⟨ W ϵ , R j ⟩ ≤ 0 by Lemma 3.20.
ℓ j = 3 \ell_j=3 ℓ j = 3 and ω 3 ( R j ) < 2 ϵ ∣ Q 3 ∣ / τ \omega_3(R_j)<2\epsilon|Q_3|/\tau ω 3 ( R j ) < 2 ϵ ∣ Q 3 ∣/ τ . Then ⟨ W ϵ , R j ⟩ ≤ W 3 ϵ ω 3 ( R j ) < 2 ϵ ∣ Q 3 ∣ ⋅ 2 ϵ ∣ Q 3 ∣ τ = 4 τ \langle W^\epsilon,R_j\rangle\le W^\epsilon_3\omega_3(R_j)<\frac{2}{\epsilon|Q_3|}\cdot\frac{2\epsilon|Q_3|}{\tau}=\frac{4}{\tau} ⟨ W ϵ , R j ⟩ ≤ W 3 ϵ ω 3 ( R j ) < ϵ ∣ Q 3 ∣ 2 ⋅ τ 2 ϵ ∣ Q 3 ∣ = τ 4 .
Step 3: count. Let N N N be the number of terms of the third kind. All other terms contribute at most 0 0 0 , so 1 ≤ ⟨ W ϵ , H ⟩ ≤ N ⋅ 4 τ 1\le\langle W^\epsilon,H\rangle\le N\cdot\frac{4}{\tau} 1 ≤ ⟨ W ϵ , H ⟩ ≤ N ⋅ τ 4 , giving N ≥ τ / 4 N\ge\tau/4 N ≥ τ /4 and hence r ≥ N ≥ τ / 4 r\ge N\ge\tau/4 r ≥ N ≥ τ /4 .
Finally, taking τ \tau τ as large as (23) permits gives log 2 r ≥ c 0 ϵ h ˉ n − log 2 2 ϵ − 2 \log_2r\ge c_0\epsilon\bar h n-\log_2\frac{2}{\epsilon}-2 log 2 r ≥ c 0 ϵ h ˉ n − log 2 ϵ 2 − 2 , which is Ω ( ϵ h ˉ n ) \Omega(\epsilon\bar h n) Ω ( ϵ h ˉ n ) whenever ϵ h ˉ n ≥ C log 2 ( 2 / ϵ ) \epsilon\bar h n\ge C\log_2(2/\epsilon) ϵ h ˉ n ≥ C log 2 ( 2/ ϵ ) for a suitable constant C C C . Under the girth assumption the required pruning property holds for this choice of τ \tau τ , which proves the final assertion.
Removing the bias restriction by weighted refinement
By Remark 3.11 the defect-ratio lemma does not apply to an arbitrary-bias term. We refine each term exactly into uniform ones, where it does apply, and charge the refinements by weight rather than by number: those the witness can score on fix many edges and carry little weight, and the total weight of a term is capped by entrywise domination.
For q ∈ [ − 1 , 1 ] q \in[-1,1] q ∈ [ − 1 , 1 ] let
ν q ( x ) : = 1 + q ( − 1 ) x 2 , x ∈ { 0 , 1 } , \nu_q(x) := \frac{1+q(-1)^x}{2}, \qquad x \in\{0,1\}, ν q ( x ) := 2 1 + q ( − 1 ) x , x ∈ { 0 , 1 } ,
and let b ( q ) = 0 b(q)=0 b ( q ) = 0 if q ≥ 0 q \ge0 q ≥ 0 and b ( q ) = 1 b(q)=1 b ( q ) = 1 if q < 0 q<0 q < 0 . Thus ν q \nu_q ν q is the Bernoulli distribution with bias p = ( 1 − q ) / 2 p=(1-q)/2 p = ( 1 − q ) /2 , and b ( q ) b(q) b ( q ) is a modal bit: ν q ( b ( q ) ) = ( 1 + ∣ q ∣ ) / 2 \nu_q(b(q))=(1+|q|)/2 ν q ( b ( q )) = ( 1 + ∣ q ∣ ) /2 is the larger of the two values. We write δ b \delta_b δ b for the point mass at b ∈ { 0 , 1 } b \in\{0,1\} b ∈ { 0 , 1 } .
Lemma 3.22 (Canonical bounded-bias refinement). For q ∈ [ − 1 , 1 ] q \in[-1,1] q ∈ [ − 1 , 1 ] , putting a = ∣ q ∣ a=|q| a = ∣ q ∣ ,
ν q = a δ b ( q ) + ( 1 − a ) ν 0 . (24) \nu_q=a\delta_{b(q)}+(1-a)\nu_0. \tag*{(24)} ν q = a δ b ( q ) + ( 1 − a ) ν 0 . (24) Consequently, for q = ( q e ) e ∈ E q=(q_e)_{e\in E} q = ( q e ) e ∈ E put a e = ∣ q e ∣ a_e=|q_e| a e = ∣ q e ∣ and, for F ⊆ E F\subseteq E F ⊆ E ,
α q ( F ) : = ∏ e ∈ F a e ∏ e ∉ F ( 1 − a e ) . \alpha_q(F):=\prod_{e\in F}a_e\prod_{e\notin F}(1-a_e). α q ( F ) := e ∈ F ∏ a e e ∈ / F ∏ ( 1 − a e ) . If μ q = ⨂ e ∈ E ν q e \mu_q=\bigotimes_{e\in E}\nu_{q_e} μ q = ⨂ e ∈ E ν q e , and μ q , F \mu_{q,F} μ q , F is the uniform distribution which fixes every e ∈ F e\in F e ∈ F to b ( q e ) b(q_e) b ( q e ) and leaves every e ∉ F e\notin F e ∈ / F unbiased, then
μ q = ∑ F ⊆ E α q ( F ) μ q , F , ∑ F ⊆ E α q ( F ) 2 ∣ F ∣ = ∏ e ∈ E ( 1 + a e ) . (25) \mu_q=\sum_{F\subseteq E}\alpha_q(F)\mu_{q,F}, \qquad\sum_{F\subseteq E}\alpha_q(F)2^{|F|}=\prod_{e\in E}(1+a_e). \tag*{(25)} μ q = F ⊆ E ∑ α q ( F ) μ q , F , F ⊆ E ∑ α q ( F ) 2 ∣ F ∣ = e ∈ E ∏ ( 1 + a e ) . (25) Terms of zero weight in the first sum are omitted.
Proof. If q ≥ 0 q\ge0 q ≥ 0 then ν q = q δ 0 + ( 1 − q ) ν 0 \nu_q=q\delta_0+(1-q)\nu_0 ν q = q δ 0 + ( 1 − q ) ν 0 ; if q ≤ 0 q\le0 q ≤ 0 then ν q = ( − q ) δ 1 + ( 1 + q ) ν 0 \nu_q=(-q)\delta_1+(1+q)\nu_0 ν q = ( − q ) δ 1 + ( 1 + q ) ν 0 . This is (24). Tensoring these identities gives the first identity in (25), and the second is
∑ F ∏ e ∈ F ( 2 a e ) ∏ e ∉ F ( 1 − a e ) = ∏ e ( 1 − a e + 2 a e ) = ∏ e ( 1 + a e ) . \sum_{F}\prod_{e\in F}(2a_e)\prod_{e\notin F}(1-a_e)=\prod_e(1-a_e+2a_e)=\prod_e(1+a_e). F ∑ e ∈ F ∏ ( 2 a e ) e ∈ / F ∏ ( 1 − a e ) = e ∏ ( 1 − a e + 2 a e ) = e ∏ ( 1 + a e ) .
□ \square □
The next observation is the weight-sensitive ingredient. Write m : = ∣ E ∣ m:=|E| m := ∣ E ∣ and put
Λ 0 : = ∑ x T ϵ ( x ) = 2 m − 1 ( n − 2 ϵ ) , \Lambda_0:=\sum_x T^\epsilon(x)=2^{m-1}(n-2\epsilon), Λ 0 := x ∑ T ϵ ( x ) = 2 m − 1 ( n − 2 ϵ ) ,
the sum being 2 m ( n / 2 − ϵ ) 2^m(n/2-\epsilon) 2 m ( n /2 − ϵ ) because flipping one edge at v v v toggles the defect at v v v , so each vertex is defective on exactly half the edge cube.
Lemma 3.23 (Modal domination). Suppose
T ϵ = ∑ j = 1 r η j μ q j , λ j : = η j / Λ 0 , T^\epsilon=\sum_{j=1}^{r}\eta_j\mu_{q_j}, \qquad\lambda_j:=\eta_j/\Lambda_0, T ϵ = j = 1 ∑ r η j μ q j , λ j := η j / Λ 0 , is an arbitrary-bias decomposition. Put a j , e = ∣ q j , e ∣ a_{j,e}=|q_{j,e}| a j , e = ∣ q j , e ∣ and Z j = ∏ e ( 1 + a j , e ) Z_j=\prod_e(1+a_{j,e}) Z j = ∏ e ( 1 + a j , e ) . Then
λ j Z j ≤ M ϵ : = 2 ( n − ϵ ) n − 2 ϵ for every j . (26) \lambda_j Z_j\le M_\epsilon:=\frac{2(n-\epsilon)}{n-2\epsilon}\qquad\text{for every }j. \tag*{(26)} λ j Z j ≤ M ϵ := n − 2 ϵ 2 ( n − ϵ ) for every j . (26)
Proof. At the modal assignment x j ∗ = ( b ( q j , e ) ) e x_j^*=(b(q_{j,e}))_e x j ∗ = ( b ( q j , e ) ) e each factor contributes ν q j , e ( b ( q j , e ) ) = ( 1 + a j , e ) / 2 \nu_{q_{j,e}}(b(q_{j,e}))=(1+a_{j,e})/2 ν q j , e ( b ( q j , e )) = ( 1 + a j , e ) /2 , so μ q j ( x j ∗ ) = 2 − m Z j \mu_{q_j}(x_j^*)=2^{-m}Z_j μ q j ( x j ∗ ) = 2 − m Z j . Every summand is nonnegative, so η j μ q j ( x j ∗ ) ≤ T ϵ ( x j ∗ ) ≤ n − ϵ \eta_j\mu_{q_j}(x_j^*)\le T^\epsilon(x_j^*)\le n-\epsilon η j μ q j ( x j ∗ ) ≤ T ϵ ( x j ∗ ) ≤ n − ϵ . Substituting the value of Λ 0 \Lambda_0 Λ 0 and rearranging gives (26).
Theorem 3.24 (Arbitrary-bias Tseitin lower bound). Let G = ( V , E ) G=(V,E) G = ( V , E ) be connected, bridgeless and d d d -regular with d ≥ 3 d\ge3 d ≥ 3 , on n n n vertices, with edge expansion h > 0 h>0 h > 0 and girth greater than 4 / h 4/h 4/ h . Put h ˉ = min { h , 1 } \bar h=\min\{h,1\} h ˉ = min { h , 1 } , and let the Tseitin charge have odd total parity. If 0 < ϵ ≤ 1 0<\epsilon\le1 0 < ϵ ≤ 1 and
ϵ n ≥ 50 e 4 (27) \epsilon n\ge50e^4 \tag*{(27)} ϵ n ≥ 50 e 4 (27) then
r k ⊕ ( χ ( T S G ) − ϵ ⋅ χ ( 1 ) ) ≥ ϵ 2 ( n − ϵ ) 2 ϵ h ˉ n / ( 50 e 4 ) . (28) \mathrm{rk}_{\oplus}\bigl(\chi(\mathrm{TS}_G)-\epsilon\cdot\chi(1)\bigr)\ge\frac{\epsilon}{2(n-\epsilon)}2^{\epsilon\bar h n/(50e^4)}. \tag*{(28)} rk ⊕ ( χ ( TS G ) − ϵ ⋅ χ ( 1 ) ) ≥ 2 ( n − ϵ ) ϵ 2 ϵ h ˉ n / ( 50 e 4 ) . (28) The same bound holds for the size of every Σ Π Σ \mathord{\Sigma\Pi\Sigma} ΣΠΣ -AMCϵ ^\epsilon ϵ proof of T S G \mathrm{TS}_G TS G , that is, every general depth-three AMCϵ ^\epsilon ϵ proof.
Remark 3.25. The girth hypothesis enters only through Lemma 3.16 : what the proof uses is the K K K -pruning property for K = ϵ h ˉ n / ( 50 e 4 ) K=\epsilon\bar h n/(50e^4) K = ϵ h ˉ n / ( 50 e 4 ) , supplied by Lemma 3.16 as in Theorem 3.21 , where K K K is chosen differently.
Proof. Let k ∗ k^* k ∗ be the least odd integer at least 24 e 4 / ϵ 24e^4/\epsilon 24 e 4 / ϵ , and use the witness W ϵ W^\epsilon W ϵ of (19) with this choice of k ∗ k^* k ∗ . Then k ∗ ≤ 25 e 4 / ϵ k^*\le25e^4/\epsilon k ∗ ≤ 25 e 4 / ϵ , and (27) gives k ∗ ≤ n / 2 k^*\le n/2 k ∗ ≤ n /2 . Set
X 0 : = ϵ n 25 e 4 , K : = h ˉ n 50 e 4 . (29) X_0:=\frac{\epsilon n}{25e^4},\qquad K:=\frac{\bar h n}{50e^4}. \tag*{(29)} X 0 := 25 e 4 ϵ n , K := 50 e 4 h ˉ n . (29)
Thus k ∗ X 0 ≤ n k^*X_0\le n k ∗ X 0 ≤ n .
Consider an arbitrary-bias decomposition
T ϵ = ∑ j = 1 r η j μ q j . (30) T^\epsilon=\sum_{j=1}^{r}\eta_j\mu_{q_j}. \tag*{(30)} T ϵ = j = 1 ∑ r η j μ q j . (30)
Apply Lemma 3.22 to every summand and write
R j , F : = η j α j ( F ) μ j , F , α j ( F ) : = α q j ( F ) . R_{j,F}:=\eta_j\alpha_j(F)\mu_{j,F},\qquad\alpha_j(F):=\alpha_{q_j}(F). R j , F := η j α j ( F ) μ j , F , α j ( F ) := α q j ( F ) .
This is an exact decomposition of T ϵ T^\epsilon T ϵ into uniform rank-one terms. Put into P P P all refinements with positive one-defect mass and into H H H all remaining refinements, and write H = ∑ j = 1 r H j H=\sum_{j=1}^{r}H_j H = ∑ j = 1 r H j , where H j H_j H j is the sum of the refinements of the j j j -th original term which occur in H H H . Lemmas 3.13 and 3.14 ask only that T ϵ = P + H T^\epsilon=P+H T ϵ = P + H be a decomposition into uniform terms split by one-defect mass, and say nothing about how many terms there are; they therefore apply verbatim to the refinement and give
⟨ W ϵ , H ⟩ ≥ 1. (31) \langle W^\epsilon,H\rangle\ge1. \tag*{(31)} ⟨ W ϵ , H ⟩ ≥ 1. (31)
In particular P P P needs no further attention: (31) is a bound on H H H alone, and it is Lemma 3.13 that keeps the Q 3 Q_3 Q 3 mass of H H H large however large P P P is and however many refinements it contains. Since ⟨ W ϵ , H ⟩ = ∑ j < r ⟨ W ϵ , H j ⟩ \langle W^\epsilon,H\rangle=\sum_{j<r}\langle W^\epsilon,H_j\rangle ⟨ W ϵ , H ⟩ = ∑ j < r ⟨ W ϵ , H j ⟩ , a bound ⟨ W ϵ , H j ⟩ ≤ c \langle W^\epsilon,H_j\rangle\le c ⟨ W ϵ , H j ⟩ ≤ c holding for each j j j turns (31) into 1 ≤ r c 1\le rc 1 ≤ r c , a lower bound on the number r r r of original terms; the number of refinements never enters.
Fix an original term j j j . Every refinement of it which occurs in H H H is entrywise at most T ϵ T^\epsilon T ϵ , because it is a nonnegative summand of the exact refined decomposition, and it has an odd number ℓ ≥ 3 \ell\ge3 ℓ ≥ 3 of odd residual components. By Lemma 3.20 its witness pairing is nonpositive if ℓ ≥ 5 \ell\ge5 ℓ ≥ 5 , or if ℓ = 3 \ell= 3 ℓ = 3 and the complement of its largest residual component has size at most X 0 X_0 X 0 . Call all remaining refinements bad . Since ℓ \ell ℓ is odd, these two cases leave only ℓ = 3 \ell= 3 ℓ = 3 : a bad refinement has exactly three odd components and either has no component larger than n / 2 n/2 n /2 , or has a unique largest component B B B , with X = V ∖ B X = V \setminus B X = V ∖ B satisfying ∣ X ∣ > X 0 |X| > X_0 ∣ X ∣ > X 0 .
We claim that every bad refinement, with fixed-edge set F F F , has cycle-rank drop
ρ ( F ) : = dim Z ( G ) − dim Z ( G − F ) ≥ K . (32) \rho(F) := \dim Z(G) - \dim Z(G - F) \ge K. \tag*{(32)} ρ ( F ) := dim Z ( G ) − dim Z ( G − F ) ≥ K . (32)
Since ϵ , h ˉ ≤ 1 \epsilon,\bar{h} \le1 ϵ , h ˉ ≤ 1 and 50 e 4 > 8 50e^4 > 8 50 e 4 > 8 , we have K < min { h , 2 } n / 8 K < \min\{h,2\}n/8 K < min { h , 2 } n /8 , so Lemma 3.16 gives G G G the K K K -pruning property. If G − F G-F G − F has no component larger than n / 2 n/2 n /2 , the contrapositive of that property gives ρ ( F ) > K \rho(F) > K ρ ( F ) > K at once. If it has a component B B B larger than n / 2 n/2 n /2 , put X = V ∖ B X = V \setminus B X = V ∖ B ; all edges between X X X and B B B belong to F F F , so Lemma 3.17 gives
ρ ( F ) ≥ h 2 ∣ X ∣ > h ϵ n 50 e 4 ≥ K . \rho(F) \ge\frac{h}{2}|X| > \frac{h\epsilon n}{50e^4} \ge K. ρ ( F ) ≥ 2 h ∣ X ∣ > 50 e 4 h ϵ n ≥ K .
This proves (32).
Suppose now that F F F is a bad refinement and its three odd components are C 1 , C 2 , C 3 C_1,C_2,C_3 C 1 , C 2 , C 3 . Put
g ( F ) : = ∣ C 1 ∣ ∣ C 2 ∣ ∣ C 3 ∣ ( n 3 ) ≤ 1. g(F) := \frac{|C_1||C_2||C_3|}{\binom{n}{3}} \le1. g ( F ) := ( 3 n ) ∣ C 1 ∣∣ C 2 ∣∣ C 3 ∣ ≤ 1.
Write Q 3 ( j , F ) : = Q 3 ∩ supp μ j , F Q_3(j,F) := Q_3 \cap\operatorname{supp}\mu_{j,F} Q 3 ( j , F ) := Q 3 ∩ supp μ j , F . The one-defect-per-odd-component count and Lemma 3.9 give
∣ Q 3 ( j , F ) ∣ ∣ Q 3 ∣ = 2 − ρ ( F ) g ( F ) . (33) \frac{|Q_3(j,F)|}{|Q_3|} = 2^{-\rho(F)}g(F). \tag*{(33)} ∣ Q 3 ∣ ∣ Q 3 ( j , F ) ∣ = 2 − ρ ( F ) g ( F ) . (33)
Using η j = Λ 0 λ j \eta_j = \Lambda_0\lambda_j η j = Λ 0 λ j , the fact that μ j , F \mu_{j,F} μ j , F is uniform on the subcube fixing the ∣ F ∣ |F| ∣ F ∣ edges of F F F and so has coefficient 2 − ( m − ∣ F ∣ ) 2^{-(m-|F|)} 2 − ( m − ∣ F ∣ ) at each of its 2 m − ∣ F ∣ 2^{m-|F|} 2 m − ∣ F ∣ support points, and (33), we obtain the exact identity
ω 3 ( R j , F ) ∣ Q 3 ∣ = n − 2 ϵ 2 λ j α j ( F ) 2 ∣ F ∣ − ρ ( F ) g ( F ) . (34) \frac{\omega_3(R_{j,F})}{|Q_3|} = \frac{n-2\epsilon}{2}\lambda_j\alpha_j(F)2^{|F|-\rho(F)}g(F). \tag*{(34)} ∣ Q 3 ∣ ω 3 ( R j , F ) = 2 n − 2 ϵ λ j α j ( F ) 2 ∣ F ∣ − ρ ( F ) g ( F ) . (34)
Let H j b a d H_j^{\mathrm{bad}} H j bad be the sum of the bad refinements of the j j j -th original term. Summing (34), using (32), g ( F ) ≤ 1 g(F) \le1 g ( F ) ≤ 1 , the second identity in (25), and Lemma 3.23, gives
ω 3 ( H j b a d ) ∣ Q 3 ∣ ≤ n − 2 ϵ 2 λ j 2 − K ∑ F α j ( F ) 2 ∣ F ∣ = n − 2 ϵ 2 λ j Z j 2 − K ≤ n − 2 ϵ 2 M ϵ 2 − K . \frac{\omega_3(H_j^{\mathrm{bad}})}{|Q_3|}
\le\frac{n-2\epsilon}{2}\lambda_j2^{-K}\sum_F\alpha_j(F)2^{|F|}
= \frac{n-2\epsilon}{2}\lambda_jZ_j2^{-K}
\le\frac{n-2\epsilon}{2}M\epsilon2^{-K}. ∣ Q 3 ∣ ω 3 ( H j bad ) ≤ 2 n − 2 ϵ λ j 2 − K F ∑ α j ( F ) 2 ∣ F ∣ = 2 n − 2 ϵ λ j Z j 2 − K ≤ 2 n − 2 ϵ M ϵ 2 − K .
All nonbad refinements in H j H_j H j pair nonpositively with the witness, and the negative k ∗ k^* k ∗ -stratum can only decrease the pairing of a bad refinement. Therefore
⟨ W ϵ , H j ⟩ ≤ 2 ϵ ∣ Q 3 ∣ ω 3 ( H j b a d ) ≤ 2 ( n − ϵ ) ϵ 2 − K . (35) \langle W^\epsilon,H_j\rangle\le\frac{2}{\epsilon|Q_3|}\omega_3(H_j^{\mathrm{bad}})
\le\frac{2(n-\epsilon)}{\epsilon}2^{-K}. \tag*{(35)} ⟨ W ϵ , H j ⟩ ≤ ϵ ∣ Q 3 ∣ 2 ω 3 ( H j bad ) ≤ ϵ 2 ( n − ϵ ) 2 − K . (35)
Finally sum (35) over the r r r original terms and combine it with (31):
1 ≤ ⟨ W ϵ , H ⟩ ≤ r 2 ( n − ϵ ) ϵ 2 − K . 1 \le\langle W^\epsilon,H\rangle\le r\frac{2(n-\epsilon)}{\epsilon}2^{-K}. 1 ≤ ⟨ W ϵ , H ⟩ ≤ r ϵ 2 ( n − ϵ ) 2 − K .
Rearranging and substituting the definition of K K K gives (28).
For the proof-system conclusion, Corollary 2.9 turns every general depth-three monotone expression computing T ϵ T^\epsilon T ϵ into an arbitrary-bias tensor decomposition with at most as many rank-one terms as the proof has size.
Corollary 3.26 (Depth-three regimes). Let G G G range over an explicit family of fixed-degree, bridgeless expanders with h = Ω ( 1 ) h=\Omega(1) h = Ω ( 1 ) and girth greater than 4 / h 4/h 4/ h , and give each graph any explicitly chosen odd charge. Then every Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCϵ _\epsilon ϵ proof of T S G \mathrm{TSG} TSG has size at least
ϵ 2 ( n − ϵ ) 2 ϵ h ˉ n 50 e 4 . \frac{\epsilon}{2(n-\epsilon)}2^{\frac{\epsilon\bar{h}n}{50e^4}}. 2 ( n − ϵ ) ϵ 2 50 e 4 ϵ h ˉ n . In particular:
if ϵ = ω ( log n / n ) \epsilon=\omega(\log n/n) ϵ = ω ( log n / n ) , the size is 2 Ω ( ϵ h n ) = n ω ( 1 ) 2^{\Omega(\epsilon h n)}=n^{\omega(1)} 2 Ω ( ϵ hn ) = n ω ( 1 ) ;
if ϵ = C log 2 n / n \epsilon=C\log^2 n/n ϵ = C log 2 n / n , then the explicit estimate in Theorem 3.24 is n C h / ( 50 e 4 ) − 2 + o ( 1 ) n^{Ch/(50e^4)-2+o(1)} n C h / ( 50 e 4 ) − 2 + o ( 1 ) ;
for constant ϵ \epsilon ϵ , the size is 2 Ω ( n ) 2^{\Omega(n)} 2 Ω ( n ) .
Such families exist. For distinct primes p , q ≡ 1 ( m o d 4 ) p,q\equiv1\pmod{4} p , q ≡ 1 ( mod 4 ) with p p p a quadratic residue mod q q q , the graphs X p , q X^{p,q} X p , q of Lubotzky, Phillips and Sarnak [24 ] are ( p + 1 ) (p+1) ( p + 1 ) -regular on n = q ( q 2 − 1 ) / 2 n=q(q^2-1)/2 n = q ( q 2 − 1 ) /2 vertices, Ramanujan, so that every nontrivial eigenvalue is at most 2 p 2\sqrt{p} 2 p , and of girth exceeding 2 3 log p n \frac{2}{3}\log_p n 3 2 log p n . The spectral bound h ≥ ( d − λ ) / 2 h\ge(d-\lambda)/2 h ≥ ( d − λ ) /2 [19 ] gives h ≥ ( p − 1 ) 2 / 2 h\ge(\sqrt{p}-1)^2/2 h ≥ ( p − 1 ) 2 /2 , which for p ≥ 5 p\ge5 p ≥ 5 exceeds 1 2 \frac{1}{2} 2 1 and so also makes these graphs bridgeless; and their girth, being Θ ( log n ) \Theta(\log n) Θ ( log n ) , exceeds the constant 4 / h 4/h 4/ h for all large n n n .
A polynomial-size proof at exponentially small shift
The bounds above cannot be extended to all ϵ > 0 \epsilon>0 ϵ > 0 : below an exponentially small threshold the Tseitin target has an explicit arbitrary-bias decomposition with O d ( n 3 ) O_d(n^3) O d ( n 3 ) terms.
Work in edge-sign variables z e ∈ { − 1 , 1 } z_e\in\{-1,1\} z e ∈ { − 1 , 1 } and write s v ( z ) = ( − 1 ) c ( v ) ∏ e ∋ v z e s_v(z)=(-1)^{c(v)}\prod_{e\ni v}z_e s v ( z ) = ( − 1 ) c ( v ) ∏ e ∋ v z e , so that B v : = ( 1 − s v ) / 2 B_v:=(1-s_v)/2 B v := ( 1 − s v ) /2 is the defect indicator at v v v and T ϵ = ∑ v ∈ V B v − ϵ T^\epsilon=\sum_{v\in V}B_v-\epsilon T ϵ = ∑ v ∈ V B v − ϵ . By Lemma 2.6 the rank-one terms of (2) are the nonnegative products ∏ e ∈ E g e ( z e ) \prod_{e\in E}g_e(z_e) ∏ e ∈ E g e ( z e ) , so the least number of those summing to T ϵ T^\epsilon T ϵ is rk + ( T ϵ ) \operatorname{rk}_{+}(T^\epsilon) rk + ( T ϵ ) .
Lemma 3.27 (Telescoping the global inconsistency). Order V = { v 1 , … , v n } V=\{v_1,\ldots,v_n\} V = { v 1 , … , v n } and put H 0 : = 1 H_0:=1 H 0 := 1 and H i : = ∏ j ≤ i s v j H_i:=\prod_{j\le i}s_{v_j} H i := ∏ j ≤ i s v j . Then ∑ i = 1 n B v i H i − 1 = 1 \sum_{i=1}^{n}B_{v_i}H_{i-1}=1 ∑ i = 1 n B v i H i − 1 = 1 , and consequently
T ϵ = ∑ i = 1 n B v i ( 1 − ϵ H i − 1 ) . (36) T^\epsilon=\sum_{i=1}^{n}B_{v_i}(1-\epsilon H_{i-1}). \tag*{(36)} T ϵ = i = 1 ∑ n B v i ( 1 − ϵ H i − 1 ) . (36) Each H i − 1 H_{i-1} H i − 1 is ± \pm ± the parity of the cut δ ( { v 1 , … , v i − 1 } ) \delta(\{v_1,\ldots,v_{i-1}\}) δ ({ v 1 , … , v i − 1 }) .
Proof. Odd total charge gives H n = ( − 1 ) ∑ v c ( v ) ∏ e z e 2 = − 1 H_n=(-1)^{\sum_v c(v)}\prod_e z_e^2=-1 H n = ( − 1 ) ∑ v c ( v ) ∏ e z e 2 = − 1 , and 2 B v i H i − 1 = ( 1 − s v i ) H i − 1 = H i − 1 − H i 2B_{v_i}H_{i-1}=(1-s_{v_i})H_{i-1}=H_{i-1}-H_i 2 B v i H i − 1 = ( 1 − s v i ) H i − 1 = H i − 1 − H i , so the sum telescopes to ( H 0 − H n ) / 2 = 1 (H_0-H_n)/2=1 ( H 0 − H n ) /2 = 1 . Substituting ϵ = ϵ ∑ i B v i H i − 1 \epsilon=\epsilon\sum_iB_{v_i}H_{i-1} ϵ = ϵ ∑ i B v i H i − 1 into T ϵ = ∑ i B v i − ϵ T^\epsilon=\sum_iB_{v_i}-\epsilon T ϵ = ∑ i B v i − ϵ gives (36). In H i − 1 H_{i-1} H i − 1 every edge internal to { v 1 , … , v i − 1 } \{v_1,\ldots,v_{i-1}\} { v 1 , … , v i − 1 } occurs twice and cancels. □
Everything therefore reduces to decomposing a very weakly biased parity.
Lemma 3.28 (Weak parity has quadratic product rank). For s ≥ 2 s\ge2 s ≥ 2 put ℓ : = s − 1 \ell:=s-1 ℓ := s − 1 and
δ s : = ℓ ! s 2 2 ℓ ℓ 2 ℓ = 2 − Θ ( s log s ) . \delta_s:=\frac{\ell!}{s^2 2^\ell\ell^{2\ell}}=2^{-\Theta(s\log s)}. δ s := s 2 2 ℓ ℓ 2 ℓ ℓ ! = 2 − Θ ( s l o g s ) . If ∣ θ ∣ ≤ δ s |\theta|\le\delta_s ∣ θ ∣ ≤ δ s then 1 + θ ∏ i = 1 s z i 1+\theta\prod_{i=1}^{s}z_i 1 + θ ∏ i = 1 s z i is a sum of exactly s 2 s^2 s 2 nonnegative products. The construction is a finite-difference moment-matching argument in the spirit of the mixtures of product distributions of [9 ] .
Proof. For j = 0 , … , ℓ j=0,\ldots,\ell j = 0 , … , ℓ set t j : = j / ℓ 2 t_j:=j/\ell^2 t j := j / ℓ 2 and a j : = ( − 1 ) ℓ − j ( ℓ j ) 2 − ℓ a_j:=(-1)^{\ell-j}\binom{\ell}{j}2^{-\ell} a j := ( − 1 ) ℓ − j ( j ℓ ) 2 − ℓ , let Π j , k : = ∏ i = k ℓ ( 1 + t j z i ) \Pi_{j,k}:=\prod_{i=k}^{\ell}(1+t_jz_i) Π j , k := ∏ i = k ℓ ( 1 + t j z i ) for 1 ≤ k ≤ ℓ + 1 1\le k\le\ell+1 1 ≤ k ≤ ℓ + 1 , so that Π j , ℓ + 1 = 1 \Pi_{j,\ell+1}=1 Π j , ℓ + 1 = 1 , and put
P j : = Π j , 1 , R j , k : = ( 1 − k t j z k ) Π j , k + 1 ( 1 ≤ k ≤ ℓ ) . P_j:=\Pi_{j,1},\qquad R_{j,k}:=(1-kt_jz_k)\Pi_{j,k+1}\qquad(1\le k\le\ell). P j := Π j , 1 , R j , k := ( 1 − k t j z k ) Π j , k + 1 ( 1 ≤ k ≤ ℓ ) .
Since Π j , k = ( 1 + t j z k ) Π j , k + 1 \Pi_{j,k}=(1+t_jz_k)\Pi_{j,k+1} Π j , k = ( 1 + t j z k ) Π j , k + 1 one has t j z k Π j , k + 1 = Π j , k − Π j , k + 1 t_jz_k\Pi_{j,k+1}=\Pi_{j,k}-\Pi_{j,k+1} t j z k Π j , k + 1 = Π j , k − Π j , k + 1 , so R j , k = ( k + 1 ) Π j , k + 1 − k Π j , k R_{j,k}=(k+1)\Pi_{j,k+1}-k\Pi_{j,k} R j , k = ( k + 1 ) Π j , k + 1 − k Π j , k and P j + ∑ k = 1 ℓ R j , k = ℓ + 1 = s P_j+\sum_{k=1}^{\ell}R_{j,k}=\ell+1=s P j + ∑ k = 1 ℓ R j , k = ℓ + 1 = s . Expanding P j = ∑ S ⊆ [ ℓ ] t j ∣ S ∣ z S P_j=\sum_{S\subseteq[\ell]}t_j^{|S|}z_S P j = ∑ S ⊆ [ ℓ ] t j ∣ S ∣ z S , the coefficient of z S z_S z S in ∑ j a j P j \sum_j a_jP_j ∑ j a j P j is ∑ j a j t j ∣ S ∣ \sum_j a_jt_j^{|S|} ∑ j a j t j ∣ S ∣ , and substituting t j = j / ℓ 2 t_j=j/\ell^2 t j = j / ℓ 2 gives, for every p ≥ 0 p\ge0 p ≥ 0 ,
∑ j = 0 ℓ a j t j p = 1 2 ℓ ℓ 2 p ∑ j = 0 ℓ ( − 1 ) ℓ − j ( ℓ j ) j p = Δ ℓ [ x p ] ( 0 ) 2 ℓ ℓ 2 p . \sum_{j=0}^{\ell}a_jt_j^p=\frac{1}{2^\ell\ell^{2p}}\sum_{j=0}^{\ell}(-1)^{\ell-j}\binom{\ell}{j}j^p=\frac{\Delta^\ell[x^p](0)}{2^\ell\ell^{2p}}. j = 0 ∑ ℓ a j t j p = 2 ℓ ℓ 2 p 1 j = 0 ∑ ℓ ( − 1 ) ℓ − j ( j ℓ ) j p = 2 ℓ ℓ 2 p Δ ℓ [ x p ] ( 0 ) .
where Δ f ( x ) : = f ( x + 1 ) − f ( x ) \Delta f(x):=f(x+1)-f(x) Δ f ( x ) := f ( x + 1 ) − f ( x ) is the forward difference with unit step and Δ ℓ \Delta^\ell Δ ℓ its ℓ \ell ℓ -fold iterate; writing Δ = S − I \Delta=S-I Δ = S − I for the shift ( S f ) ( x ) = f ( x + 1 ) (Sf)(x)=f(x+1) ( S f ) ( x ) = f ( x + 1 ) and expanding the binomial in the commuting operators S , I S,I S , I gives Δ ℓ f ( 0 ) = ∑ j = 0 ℓ ( − 1 ) ℓ − j ( ℓ j ) f ( j ) \Delta^\ell f(0)=\sum_{j=0}^{\ell}(-1)^{\ell-j}\binom{\ell}{j}f(j) Δ ℓ f ( 0 ) = ∑ j = 0 ℓ ( − 1 ) ℓ − j ( j ℓ ) f ( j ) . Applied to f = ∑ i ≤ m c i x i f=\sum_{i\le m}c_ix^i f = ∑ i ≤ m c i x i of degree m ≥ 1 m\ge1 m ≥ 1 , the operator Δ \Delta Δ gives ∑ i ≤ m c i [ ( x + 1 ) i − x i ] \sum_{i\le m}c_i[(x+1)^i-x^i] ∑ i ≤ m c i [( x + 1 ) i − x i ] , in which each bracket has degree i − 1 i-1 i − 1 ; so Δ f \Delta f Δ f has degree exactly m − 1 m-1 m − 1 with leading coefficient c m m c_m m c m m , and Δ \Delta Δ annihilates constants. Iterating ℓ \ell ℓ times therefore annihilates every polynomial of degree < ℓ <\ell < ℓ and returns the constant c ℓ ℓ ! c_\ell\ell! c ℓ ℓ ! on one of degree ℓ \ell ℓ : in particular Δ ℓ [ x p ] ( 0 ) = 0 \Delta^\ell[x^p](0)=0 Δ ℓ [ x p ] ( 0 ) = 0 for p < ℓ p<\ell p < ℓ and = ℓ ! =\ell! = ℓ ! for p = ℓ p=\ell p = ℓ . Hence only S = [ ℓ ] S=[\ell] S = [ ℓ ] survives and ∑ j a j P j = ℓ ! 2 ℓ ℓ 2 ℓ ∏ i ≤ ℓ z i \sum_j a_jP_j=\frac{\ell!}{2^\ell\ell^{2\ell}}\prod_{i\le\ell}z_i ∑ j a j P j = 2 ℓ ℓ 2 ℓ ℓ ! ∏ i ≤ ℓ z i . Combining the two, for every u ∈ [ − 1 , 1 ] u\in[-1,1] u ∈ [ − 1 , 1 ] ,
1 s 2 ∑ j = 0 ℓ [ ( 1 + u a j z s ) P j + ∑ k = 1 ℓ R j , k ] = 1 + u δ s ∏ i = 1 s z i , (37) \frac{1}{s^2}\sum_{j=0}^{\ell}\left[(1+ua_jz_s)P_j+\sum_{k=1}^{\ell}R_{j,k}\right]=1+u\delta_s\prod_{i=1}^{s}z_i, \tag*{(37)} s 2 1 j = 0 ∑ ℓ [ ( 1 + u a j z s ) P j + k = 1 ∑ ℓ R j , k ] = 1 + u δ s i = 1 ∏ s z i , (37)
a sum of s ( ℓ + 1 ) = s 2 s(\ell+1)=s^2 s ( ℓ + 1 ) = s 2 products. Every one-coordinate factor is nonnegative on { − 1 , 1 } \{-1,1\} { − 1 , 1 } , since t j ≤ 1 / ℓ t_j\le1/\ell t j ≤ 1/ ℓ , k t j ≤ 1 kt_j\le1 k t j ≤ 1 and ∣ u a j ∣ ≤ 1 |ua_j|\le1 ∣ u a j ∣ ≤ 1 . Take u = θ / δ s u=\theta/\delta_s u = θ / δ s .
Theorem 3.29 (Tiny-shift upper bound). Let G G G have maximum degree d d d and odd total Tseitin charge, put m : = ∣ E ∣ ≥ 2 m:=|E|\ge2 m := ∣ E ∣ ≥ 2 and ϵ 0 ( m ) : = δ m = 2 − Θ ( m log m ) \epsilon_0(m):=\delta_m=2^{-\Theta(m\log m)} ϵ 0 ( m ) := δ m = 2 − Θ ( m l o g m ) . Then
rk + ⊗ ( χ ( T S G ) − ϵ ⋅ χ ( 1 ) ) ≤ 2 d − 1 ∣ V ∣ ∣ E ∣ 2 for every 0 < ϵ ≤ ϵ 0 ( m ) , \operatorname{rk}_{+}^{\otimes}\bigl(\chi(\mathrm{TSG})-\epsilon\cdot\chi(1)\bigr)\le2^{d-1}|V||E|^2\qquad\text{for every }0<\epsilon\le\epsilon_0(m), rk + ⊗ ( χ ( TSG ) − ϵ ⋅ χ ( 1 ) ) ≤ 2 d − 1 ∣ V ∣∣ E ∣ 2 for every 0 < ϵ ≤ ϵ 0 ( m ) , which for fixed degree is O ( n 3 ) O(n^3) O ( n 3 ) product terms, hence an Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCϵ ^\epsilon ϵ proof of T S G \mathrm{TSG} TSG of that size.
Proof. In (36) the i i i -th factor 1 − ϵ H i − 1 1-\epsilon H_{i-1} 1 − ϵ H i − 1 is 1 + θ ∏ e ∈ S z e 1+\theta\prod_{e\in S}z_e 1 + θ ∏ e ∈ S z e with S = δ ( { v 1 , … , v i − 1 } ) S=\delta(\{v_1,\ldots,v_{i-1}\}) S = δ ({ v 1 , … , v i − 1 }) and ∣ θ ∣ = ϵ |\theta|=\epsilon ∣ θ ∣ = ϵ . Here ∣ S ∣ ≤ m |S|\le m ∣ S ∣ ≤ m , and δ s \delta_s δ s decreases in s s s , so the hypothesis ϵ ≤ δ m \epsilon\le\delta_m ϵ ≤ δ m gives ϵ ≤ δ ∣ S ∣ \epsilon\le\delta_{|S|} ϵ ≤ δ ∣ S ∣ . For ∣ S ∣ ≥ 2 |S|\ge2 ∣ S ∣ ≥ 2 therefore Lemma 3.28 writes that factor as a sum of ∣ S ∣ 2 ≤ m 2 |S|^2\le m^2 ∣ S ∣ 2 ≤ m 2 nonnegative products in the variables z e z_e z e , e ∈ S e\in S e ∈ S (with g e ≡ 1 g_e\equiv1 g e ≡ 1 off S S S ). The two degenerate cuts are handled directly and more cheaply: S = ∅ S=\varnothing S = ∅ , which occurs at i = 1 i=1 i = 1 , leaves the constant 1 − ϵ > 0 1-\epsilon>0 1 − ϵ > 0 , a single nonnegative product, and ∣ S ∣ = 1 |S|=1 ∣ S ∣ = 1 leaves 1 + θ z e 1+\theta z_e 1 + θ z e with ∣ θ ∣ ≤ 1 |\theta|\le1 ∣ θ ∣ ≤ 1 , again a single one. Each B v i B_{v_i} B v i is the sum of the 2 deg v i − 1 2^{\deg v_i-1} 2 d e g v i − 1 locally violating complete patterns π \pi π on δ ( v i ) \delta(v_i) δ ( v i ) , and each such pattern is itself the nonnegative product ∏ e ∈ δ ( v i ) 1 [ z e = π e ] \prod_{e\in\delta(v_i)}\mathbf{1}[z_e=\pi_e] ∏ e ∈ δ ( v i ) 1 [ z e = π e ] . Nonnegative products are closed under multiplication, since at each edge the two factors combine into the single nonnegative one-coordinate function g e ( z e ) 1 [ z e = π e ] g_e(z_e)\mathbf{1}[z_e=\pi_e] g e ( z e ) 1 [ z e = π e ] ; so every product obtained above is by Lemma 2.6 a rank-one set-multilinear term. Summing over i i i gives at most n 2 d − 1 m 2 n2^{d-1}m^2 n 2 d − 1 m 2 products. □ \square □
Two consequences. The first is a separation inside depth three.
Corollary 3.30 (Bounded and arbitrary bias separate at tiny shift). Let G G G be a constant-degree expander and ϵ ≤ 2 − Θ ( n log n ) \epsilon\le2^{-\Theta(n\log n)} ϵ ≤ 2 − Θ ( n l o g n ) . Every Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCb _b b proof of T S G \mathrm{TS}_G TS G has size 2 Ω ( n ) 2^{\Omega(n)} 2 Ω ( n ) by Corollary 3.7, while Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC has one of size O ( n 3 ) O(n^3) O ( n 3 ) . The two fragments are therefore exponentially separated.
The second concerns the quantifier over ϵ \epsilon ϵ .
Remark 3.31 (Tseitin cannot settle the ϵ \epsilon ϵ quantifier). By Theorem 3.29 no superpolynomial lower bound on Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC proof size for T S G \mathrm{TS}_G TS G holds uniformly in ϵ > 0 \epsilon> 0 ϵ > 0 , whatever the method; and by Theorem 3.33 no superpolynomial lower bound on AMCϵ ^\epsilon ϵ proof size holds for T S G \mathrm{TS}_G TS G at any ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] once the depth is unrestricted, nor at any fixed mixed-product depth. An unconditional AMC lower bound must therefore be sought on a different formula. Section 3.5 shows how little has to change: an index gadget of constant size applied to T S G \mathrm{TS}_G TS G itself is hard at every ϵ ≥ 2 − Ω ( N ) \epsilon\ge2^{-\Omega(N)} ϵ ≥ 2 − Ω ( N ) . What the Tseitin bounds of this section do settle is the behaviour above ϵ = Ω ( log n / n ) \epsilon= \Omega(\log n/n) ϵ = Ω ( log n / n ) , and the true threshold between that scale and 2 − Θ ( n log n ) 2^{-\Theta(n\log n)} 2 − Θ ( n l o g n ) is open.
A proof of mixed-product depth zero at every shift
Both bounds of § 3.3 are bounds on rk + ⊗ \operatorname{rk}^{\otimes}_{+} rk + ⊗ , that is on depth three, and § 3.3.8 shows that even there they cannot reach every ϵ \epsilon ϵ . Once the depth is not restricted the picture changes completely: T S G \mathrm{TS}_G TS G has polynomial-size proofs at every ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] , and they can be arranged to have mixed-product depth zero with respect to any bipartition of the edges. The ingredient is that the parities which the telescoping of Lemma 3.27 leaves behind have small monotone set-multilinear formulas; what Lemma 3.28 had to buy with an exponentially small θ \theta θ is free as soon as logarithmic depth is allowed.
Lemma 3.32 (Parity polynomials). For S ⊆ E S \subseteq E S ⊆ E and b ∈ { 0 , 1 } b \in\{0,1\} b ∈ { 0 , 1 } put
P S b : = ∑ σ ∈ { 0 , 1 } S ⨁ e ∈ S σ e = b ∏ e ∈ S x e σ e , P_S^b := \sum_{\substack{\sigma\in\{0,1\}^{S}\\ \bigoplus_{e\in S}\sigma_e=b}} \prod_{e\in S} x_e^{\sigma_e}, P S b := σ ∈ { 0 , 1 } S ⨁ e ∈ S σ e = b ∑ e ∈ S ∏ x e σ e , the set-multilinear tensor of the indicator of ⨁ e ∈ S σ e = b \bigoplus_{e\in S}\sigma_e=b ⨁ e ∈ S σ e = b , with one variable block { x e 0 , x e 1 } \{x_e^0,x_e^1\} { x e 0 , x e 1 } per edge e ∈ S e \in S e ∈ S . Each of P S 0 P_S^0 P S 0 and P S 1 P_S^1 P S 1 is computed by a monotone syntactically set-multilinear formula with O ( ∣ S ∣ 2 ) O(|S|^2) O ( ∣ S ∣ 2 ) leaves and depth O ( log ∣ S ∣ ) O(\log|S|) O ( log ∣ S ∣ ) .
Proof. For a partition S = A ⊥ B S=A \mathbin{\perp} B S = A ⊥ B ,
P S 0 = P A 0 P B 0 + P A 1 P B 1 , P S 1 = P A 0 P B 1 + P A 1 P B 0 . (38) P_S^0=P_A^0P_B^0+P_A^1P_B^1,\qquad P_S^1=P_A^0P_B^1+P_A^1P_B^0. \tag*{(38)} P S 0 = P A 0 P B 0 + P A 1 P B 1 , P S 1 = P A 0 P B 1 + P A 1 P B 0 . (38)
Halving S S S repeatedly therefore gives both polynomials, starting from x e 0 , x e 1 x_e^0,x_e^1 x e 0 , x e 1 at ∣ S ∣ = 1 |S|=1 ∣ S ∣ = 1 . Writing L ( s ) L(s) L ( s ) for the total number of leaves of the two formulas when ∣ S ∣ = s |S|=s ∣ S ∣ = s , one has L ( 1 ) = 2 L(1)=2 L ( 1 ) = 2 and L ( s ) ≤ 4 L ( s / 2 ) L(s)\le4L(s/2) L ( s ) ≤ 4 L ( s /2 ) , so L ( s ) = O ( s 2 ) L(s)=O(s^2) L ( s ) = O ( s 2 ) , and the depth grows by O ( 1 ) O(1) O ( 1 ) per halving.
Theorem 3.33 (Polynomial size at mixed-product depth zero, for every shift). Let G G G have maximum degree d d d and odd total Tseitin charge, and put m : = ∣ E ∣ ≥ 2 m:=|E|\ge2 m := ∣ E ∣ ≥ 2 . Fix a bipartition of the edge blocks into X ⊔ Y \mathcal{X}\sqcup\mathcal{Y} X ⊔ Y . Then for every ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] the target T = χ ( T S G ) − ϵ ⋅ χ ( 1 ) T=\chi(\mathrm{TS}_G)-\epsilon\cdot\chi(1) T = χ ( TS G ) − ϵ ⋅ χ ( 1 ) is computed by a monotone syntactically set-multilinear formula with O ( 2 d n m 2 ) O(2^{d}nm^2) O ( 2 d n m 2 ) leaves, of depth O ( d + log m ) O(d+\log m) O ( d + log m ) , whose mixed-product depth with respect to that bipartition is zero.
Proof. Write D i : = δ ( { v 1 , … , v i − 1 } ) D_i:=\delta(\{v_1,\ldots,v_{i-1}\}) D i := δ ({ v 1 , … , v i − 1 }) and C i − 1 : = ∑ j < i c ( v j ) C_{i-1}:=\sum_{j<i}c(v_j) C i − 1 := ∑ j < i c ( v j ) , so that H i − 1 = ( − 1 ) C i − 1 ∏ e ∈ D i z e H_{i-1}=(-1)^{C_{i-1}}\prod_{e\in D_i}z_e H i − 1 = ( − 1 ) C i − 1 ∏ e ∈ D i z e by Lemma 3.27. Expand B v i B_{v_i} B v i into the 2 deg v i − 1 2^{\deg v_i-1} 2 d e g v i − 1 local assignments π \pi π of δ ( v i ) \delta(v_i) δ ( v i ) that violate the charge at v i v_i v i . Fixing such a π \pi π fixes the coordinates of D i ∩ δ ( v i ) D_i\cap\delta(v_i) D i ∩ δ ( v i ) . Put S i : = D i ∖ δ ( v i ) S_i:=D_i\setminus\delta(v_i) S i := D i ∖ δ ( v i ) and b ( i , π ) : = b(i,\pi):= b ( i , π ) := C i − 1 + ∑ e ∈ D i ∩ δ ( v i ) π e C_{i-1}+\sum_{e\in D_i\cap\delta(v_i)}\pi_e C i − 1 + ∑ e ∈ D i ∩ δ ( v i ) π e mod 2; the factor 1 − ϵ H i − 1 1-\epsilon H_{i-1} 1 − ϵ H i − 1 then equals 1 − ϵ 1-\epsilon 1 − ϵ at the assignments of S i S_i S i of parity b ( i , π ) b(i,\pi) b ( i , π ) and 1 + ϵ 1+\epsilon 1 + ϵ at the others. The term indexed by ( i , π ) (i,\pi) ( i , π ) in the resulting expansion of (36) thus involves δ ( v i ) \delta(v_i) δ ( v i ) , S i S_i S i and the remaining blocks separately, so it is a product of tensors:
T ϵ = ∑ i = 1 n ∑ π ( ∏ e ∈ δ ( v i ) x e π e ) ( ( 1 − ϵ ) P S i b ( i , π ) + ( 1 + ϵ ) P S i 1 − b ( i , π ) ) ∏ e ∉ δ ( v i ) ∪ S i ( x e 0 + x e 1 ) , (39) T^\epsilon=\sum_{i=1}^{n}\sum_{\pi}\left(\prod_{e\in\delta(v_i)}x_e^{\pi_e}\right)\left((1-\epsilon)P_{S_i}^{b(i,\pi)}+(1+\epsilon)P_{S_i}^{1-b(i,\pi)}\right)\prod_{e\notin\delta(v_i)\cup S_i}(x_e^0+x_e^1), \tag*{(39)} T ϵ = i = 1 ∑ n π ∑ e ∈ δ ( v i ) ∏ x e π e ( ( 1 − ϵ ) P S i b ( i , π ) + ( 1 + ϵ ) P S i 1 − b ( i , π ) ) e ∈ / δ ( v i ) ∪ S i ∏ ( x e 0 + x e 1 ) , (39)
with nonnegative coefficients because ϵ ≤ 1 \epsilon\le1 ϵ ≤ 1 , and syntactically set-multilinear because sums have children of equal block support and products of disjoint block support. Of its at most n 2 d − 1 n2^{d-1} n 2 d − 1 summands each costs O ( m 2 ) O(m^2) O ( m 2 ) leaves and depth O ( log m ) O(\log m) O ( log m ) by Lemma 3.32 , whence the stated size and depth O ( d + log m ) O(d+\log m) O ( d + log m ) .
For the mixed-product depth put A : = S i ∩ X A:=S_i\cap\mathcal{X} A := S i ∩ X and B : = S i ∩ Y B:=S_i\cap\mathcal{Y} B := S i ∩ Y and apply (38) to both parity polynomials of the ( i , π ) (i,\pi) ( i , π ) summand of (39) : its bracket becomes ∑ b 1 , b 2 c b 1 b 2 P A b 1 P B b 2 \sum_{b_1,b_2}c_{b_1b_2}P_A^{b_1}P_B^{b_2} ∑ b 1 , b 2 c b 1 b 2 P A b 1 P B b 2 , one term for each of the four pairs, with c b 1 b 2 = 1 − ϵ c_{b_1b_2}=1-\epsilon c b 1 b 2 = 1 − ϵ when b 1 ⊕ b 2 = b ( i , π ) b_1\oplus b_2=b(i,\pi) b 1 ⊕ b 2 = b ( i , π ) and 1 + ϵ 1+\epsilon 1 + ϵ otherwise. Every fixed literal x e π e x_e^{\pi_e} x e π e and every free block x e 0 + x e 1 x_e^0+x_e^1 x e 0 + x e 1 involves a single block, so lies on one side; distributing these over the four terms and computing P A b 1 P_A^{b_1} P A b 1 , P B b 2 P_B^{b_2} P B b 2 inside their own sides by Lemma 3.32 writes the summand as a sum of four products c b 1 b 2 F ⋅ G c_{b_1b_2}F\cdot G c b 1 b 2 F ⋅ G , with F F F built from variables of blocks in X \mathcal{X} X alone and G G G from blocks in Y \mathcal{Y} Y alone. The support of a gate is the union of the blocks occurring below it, so no gate inside F F F or inside G G G is mixed; and the top gate of each term multiplies the two one-sided polynomials F F F and G G G . No product gate therefore has two mixed children, so the mixed-product depth is zero.
For fixed degree the formula of Theorem 3.33 has O ( n 3 ) O(n^3) O ( n 3 ) leaves and depth O ( log n ) O(\log n) O ( log n ) . Taking it together with the circuits of § 2.2 for χ ( T S G ) \chi(T_{\mathrm{SG}}) χ ( T SG ) and χ ( 1 ) \chi(1) χ ( 1 ) yields one monotone circuit, still of size O ( n 3 ) O(n^3) O ( n 3 ) and depth O ( log n ) O(\log n) O ( log n ) , which computes the three polynomials χ ( 1 ) \chi(1) χ ( 1 ) , χ ( T S G ) \chi(T_{\mathrm{SG}}) χ ( T SG ) and χ ( T S G ) − ϵ ⋅ χ ( 1 ) \chi(T_{\mathrm{SG}})-\epsilon\cdot\chi(1) χ ( T SG ) − ϵ ⋅ χ ( 1 ) simultaneously. By Theorem 2.3 (b), χ ( 1 ) ⪯ χ ( T S G ) \chi(1)\preceq\chi(T_{\mathrm{SG}}) χ ( 1 ) ⪯ χ ( T SG ) then has an AMC derivation of size O ( n 3 ) O(n^3) O ( n 3 ) . The depth bound refers to the displayed certificate formula.
Corollary 3.34 (The flattening bound is empty for Tseitin). For fixed degree and every ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] , every bipartite flattening of T ϵ T^\epsilon T ϵ has nonnegative rank O ( n 3 ) O(n^3) O ( n 3 ) . So (6) gives no superpolynomial bound for T S G T_{\mathrm{SG}} T SG at any mixed-product depth, and in particular the exponential bounds of § 3.3 cannot be obtained from a flattening.
Proof. Theorem 3.33 supplies, for each bipartition, a formula of mixed-product depth zero with L = O ( n 3 ) L=O(n^3) L = O ( n 3 ) leaves, so Theorem 2.11 gives rk + ( Flat ( T ϵ ) ) ≤ L 2 = O ( n 3 ) \operatorname{rk}_{+}(\operatorname{Flat}(T^\epsilon))\le L^2=O(n^3) rk + ( Flat ( T ϵ )) ≤ L 2 = O ( n 3 ) .
Remark 3.35 (Depth three versus mixed-product depth zero). Depth three implies mixed-product depth zero (Remark 2.12), and the converse now fails exponentially: on a constant-degree expander at constant ϵ \epsilon ϵ , Theorem 3.24 gives rk + ⊗ ( T ϵ ) = 2 Ω ( n ) \operatorname{rk}_{+}^{\otimes}(T^\epsilon)=2^{\Omega(n)} rk + ⊗ ( T ϵ ) = 2 Ω ( n ) while Theorem 3.33 gives a formula of mixed-product depth zero with O ( n 3 ) O(n^3) O ( n 3 ) leaves. At mixed-product depth zero the flattening bound (6) is therefore strictly weaker than the tensor rank bound, and mixed-product depth organises a hierarchy of its own, not the hierarchy of ordinary depth: the formula above has depth O ( log n ) O(\log n) O ( log n ) and mixed-product depth zero simultaneously. The ϵ \epsilon ϵ -sensitive Tseitin bounds measure tensor rank, which is depth three, and nothing beyond it.
Lifting Sherali–Adams degree to Sigma Pi Sigma-AMC size at exponentially small shift
Remark 3.31 leaves the quantifier over ϵ \epsilon ϵ to be settled on a formula other than Tseitin. We give one whose admissible range reaches exponentially small shifts, and it is obtained from Tseitin by an index gadget. The statement is about rk + ⊗ \operatorname{rk}_{+}^{\otimes} rk + ⊗ , hence about depth-three proof size, and only about that: unlike § 4 it yields no bipartite flattening bound and therefore does not feed Theorem 2.11 .
Instances, Fourier notation and local expectation functionals. This subsection and Section 4.4 both use a Tseitin base instance that fools a Sherali–Adams relaxation, and the objects involved are fixed here once, for both.
A normalized Boolean Max-k k k -CSP instance on m m m variables is identified with the function I : { − 1 , 1 } m → [ 0 , 1 ] I:\{-1,1\}^{m}\to[0,1] I : { − 1 , 1 } m → [ 0 , 1 ] sending an assignment to the fraction of its constraints that the assignment satisfies; opt ( I ) : = max x I ( x ) \operatorname{opt}(I):=\max_x I(x) opt ( I ) := max x I ( x ) , so that I I I is unsatisfiable exactly when opt ( I ) < 1 \operatorname{opt}(I)<1 opt ( I ) < 1 .
For a finite coordinate set V V V and α ⊆ V \alpha\subseteq V α ⊆ V write x α : = ∏ i ∈ α x i x^\alpha:=\prod_{i\in\alpha}x_i x α := ∏ i ∈ α x i , with x ∅ ≡ 1 x^\varnothing\equiv1 x ∅ ≡ 1 . These 2 ∣ V ∣ 2^{|V|} 2 ∣ V ∣ monomials are an orthonormal basis of the real functions on { − 1 , 1 } V \{-1,1\}^{V} { − 1 , 1 } V for the inner product ⟨ f , g ⟩ = E [ f g ] \langle f,g\rangle=\mathbb{E}[fg] ⟨ f , g ⟩ = E [ f g ] , every expectation here and below being over a uniform x ∈ { − 1 , 1 } V x\in\{-1,1\}^{V} x ∈ { − 1 , 1 } V . Every such g g g therefore has a unique expansion g = ∑ α ⊆ V g ^ ( α ) x α g=\sum_{\alpha\subseteq V}\widehat{g}(\alpha)x^\alpha g = ∑ α ⊆ V g ( α ) x α with coefficients g ^ ( α ) = E [ g x α ] \widehat{g}(\alpha)=\mathbb{E}[gx^\alpha] g ( α ) = E [ g x α ] , its Fourier coefficients . The Fourier degree of g g g is the largest ∣ α ∣ |\alpha| ∣ α ∣ with g ^ ( α ) ≠ 0 \widehat{g}(\alpha)\ne0 g ( α ) = 0 , and
g ≤ d : = ∑ ∣ α ∣ ≤ d g ^ ( α ) x α g^{\le d}:=\sum_{|\alpha|\le d}\widehat{g}(\alpha)x^\alpha g ≤ d := ∣ α ∣ ≤ d ∑ g ( α ) x α
is the degree-d d d truncation of g g g . We write ( m ≤ d ) : = ∑ j ≤ d ( m j ) \binom{m}{\le d}:=\sum_{j\le d}\binom{m}{j} ( ≤ d m ) := ∑ j ≤ d ( j m ) for the number of monomials of degree at most d d d on m m m coordinates.
A degree-d d d local expectation functional is a linear functional E ~ \widetilde{\mathbb{E}} E on the real functions on { − 1 , 1 } m \{-1,1\}^{m} { − 1 , 1 } m of Fourier degree at most d d d with E ~ 1 = 1 \widetilde{\mathbb{E}}1=1 E 1 = 1 which is nonnegative on every nonnegative function depending on at most d d d of the coordinates. Such a functional is a relaxed substitute for an average over genuine assignments: it need not come from a distribution, but it cannot be told apart from one by looking at d d d coordinates at a time. The tilde separates it from the genuine expectation E \mathbb{E} E under the uniform measure, which keeps that meaning throughout; in particular the Fourier coefficients g ^ ( α ) \widehat{g}(\alpha) g ( α ) above are taken against E \mathbb{E} E and not against E ~ \widetilde{\mathbb{E}} E . A d d d -junta is a function depending on at most d d d coordinates. Only nonnegative ones occur below, as in the LP framework of [5 ] , and d d d -junta always means a nonnegative one here. These are not the conical juntas of Section 2.3.
These functionals are the degree-d d d Sherali–Adams solutions, which is what connects the selector lift below and the planted family of Section 4.4 to the Sherali–Adams hierarchy. A d d d -junta f f f depending on the d d d -set T T T is ∑ σ f ( σ ) 1 [ x ∣ T = σ ] \sum_\sigma f(\sigma)1[x|_T=\sigma] ∑ σ f ( σ ) 1 [ x ∣ T = σ ] with nonnegative coefficients, and each indicator 1 [ x ∣ T = σ ] 1[x|_T=\sigma] 1 [ x ∣ T = σ ] is a product of at most d d d literals, hence of degree at most d d d . Nonnegativity on the d d d -juntas therefore says exactly that p T ( σ ) : = E ~ 1 [ x ∣ T = σ ] ≥ 0 p_T(\sigma):=\widetilde{\mathbb{E}}1[x|_T=\sigma]\ge0 p T ( σ ) := E 1 [ x ∣ T = σ ] ≥ 0 for every d d d -set T T T and every σ \sigma σ . As these indicators sum to 1 1 1 and E ~ 1 = 1 \widetilde{\mathbb{E}}1=1 E 1 = 1 , each p T p_T p T is a probability distribution on { − 1 , 1 } T \{-1,1\}^{T} { − 1 , 1 } T , whose moments are the numbers E ~ x α \widetilde{\mathbb{E}}x^\alpha E x α with α ⊆ T \alpha\subseteq T α ⊆ T . Distributions belonging to different d d d -sets agree on the overlap, but need not be the marginals of any single distribution on all the coordinates. Finally, SA d ( I ) \operatorname{SA}_d(I) SA d ( I ) is the degree-d d d Sherali–Adams value of I I I , the maximum of E ~ I \widetilde{\mathbb{E}}I E I over the degree-d d d local expectation functionals, so that SA d ( I ) = 1 \operatorname{SA}_d(I)=1 SA d ( I ) = 1 says the degree-d d d relaxation is fooled into valuing I I I at one.
The base. Let I 0 I_0 I 0 be an unsatisfiable normalized Max-k k k -CSP instance on y 1 , … , y m ∈ { − 1 , 1 } y_1,\ldots,y_m\in\{-1,1\} y 1 , … , y m ∈ { − 1 , 1 } with L L L constraints, let B ℓ B_\ell B ℓ be the indicator that the ℓ \ell ℓ -th constraint is violated, and put ν : = ∑ ℓ ≤ L B ℓ = L ( 1 − I 0 ) \nu:=\sum_{\ell\le L}B_\ell=L(1-I_0) ν := ∑ ℓ ≤ L B ℓ = L ( 1 − I 0 ) , the number of violated constraints, so that 1 ≤ ν ≤ L 1\le\nu\le L 1 ≤ ν ≤ L . Fix k ≤ d < m k\le d<m k ≤ d < m and assume SA d ( I 0 ) = 1 \operatorname{SA}_d(I_0)=1 SA d ( I 0 ) = 1 , witnessed by a degree-d d d local expectation functional E ~ \widetilde{\mathbb{E}} E ; in terms of ν \nu ν this reads
E ~ 1 = 1 , E ~ ν = 0. \widetilde{\mathbb{E}}1=1,\qquad\widetilde{\mathbb{E}}\nu=0. E 1 = 1 , E ν = 0.
The lift. Fix b ≥ 1 b\ge1 b ≥ 1 and put q : = 2 b q:=2^b q := 2 b . As in Section 3.3.8 we use ± 1 \pm1 ± 1 variables, and the lifted formula has two kinds of them: for each base variable u u u there are b b b selector bits , taking values in { − 1 , 1 } \{-1,1\} { − 1 , 1 } , whose joint value is read as an address a u ∈ [ q ] a_u\in[q] a u ∈ [ q ] ; and q q q table bits z u , 1 , … , z u , q ∈ { − 1 , 1 } z_{u,1},\ldots,z_{u,q}\in\{-1,1\} z u , 1 , … , z u , q ∈ { − 1 , 1 } private to u u u . Thus there are
N = m ( q + b ) N = m(q+b) N = m ( q + b )
variables in all. An assignment to them is a pair ( a , z ) (a,z) ( a , z ) , where a = ( a 1 , … , a m ) ∈ [ q ] m a=(a_1,\ldots,a_m)\in[q]^m a = ( a 1 , … , a m ) ∈ [ q ] m collects the address of every base variable as read off its own selector bits, and it induces the base assignment
y u : = z u , a u ∈ { − 1 , 1 } , y_u := z_{u,a_u}\in\{-1,1\}, y u := z u , a u ∈ { − 1 , 1 } ,
the entry of u u u ’s table at u u u ’s own address; y y y is a function of the lifted assignment, not a variable of the lifted formula. Equivalently y u = IND q ( a u , z u ) y_u=\operatorname{IND}_q(a_u,z_u) y u = IND q ( a u , z u ) for the index gadget IND q ( x , y ) : = y x \operatorname{IND}_q(x,y):=y_x IND q ( x , y ) := y x on [ q ] × { − 1 , 1 } q [q]\times\{-1,1\}^q [ q ] × { − 1 , 1 } q , the address x = a u x=a_u x = a u being given in binary by u u u ’s b b b selector bits; the construction below is thus the composition I 0 ∘ IND q m I_0\circ\operatorname{IND}_q^m I 0 ∘ IND q m , one gadget per base variable.
For each constraint with scope { u 1 , … , u t } \{u_1,\ldots,u_t\} { u 1 , … , u t } , each tuple of addresses ( a u 1 , … , a u t ) ∈ [ q ] t (a_{u_1},\ldots,a_{u_t})\in[q]^t ( a u 1 , … , a u t ) ∈ [ q ] t and each pattern ( σ 1 , … , σ t ) ∈ { − 1 , 1 } t (\sigma_1,\ldots,\sigma_t)\in\{-1,1\}^t ( σ 1 , … , σ t ) ∈ { − 1 , 1 } t violating that constraint, take the conjunction which fixes the t b tb t b selector bits of the scope to those addresses and fixes the t t t addressed table bits z u 1 , a u 1 , … , z u t , a u t z_{u_1,a_{u_1}},\ldots,z_{u_t,a_{u_t}} z u 1 , a u 1 , … , z u t , a u t to that pattern, and let A A A be the disjunction of all of them. Such a conjunction holds at ( a , z ) (a,z) ( a , z ) exactly when the scope has those addresses and the induced y y y falsifies that constraint through that pattern; so each violated base constraint contributes exactly one satisfied conjunction and each satisfied one none, giving
c A ( a , z ) = V ( y ) = V ( z 1 , a 1 , … , z m , a m ) . (40) c_A(a,z)=\mathcal{V}(y)=\mathcal{V}(z_{1,a_1},\ldots,z_{m,a_m}). \tag*{(40)} c A ( a , z ) = V ( y ) = V ( z 1 , a 1 , … , z m , a m ) . (40)
Hence A A A is a tautology, with at most L 2 k q k L2^kq^k L 2 k q k terms of width at most k ( b + 1 ) k(b+1) k ( b + 1 ) .
Fixing the selector bits to spell a given a a a restricts the lifted instance, by (40), to I 0 I_0 I 0 evaluated on the m m m addressed table bits z 1 , a 1 , … , z m , a m z_{1,a_1},\ldots,z_{m,a_m} z 1 , a 1 , … , z m , a m . The lift is therefore q m q^m q m copies of I 0 I_0 I 0 , one for each address tuple and none of them distinguished, and a purported decomposition of c A − ϵ c_A-\epsilon c A − ϵ has to serve all of them at once; the proof below selects a copy only after seeing the decomposition.
Theorem 3.36 (Selector-lift lower bound). With I 0 I_0 I 0 , E ~ \widetilde{E} E , k k k , d d d , q q q and A A A as above, for every ϵ > 0 \epsilon>0 ϵ > 0 for which c A − ϵ c_A-\epsilon c A − ϵ is nonnegative,
rk + ⊗ ( χ ( A ) − ϵ ⋅ χ ( 1 ) ) ≥ ϵ q d + 1 L ( m ≤ d ) . (41) \operatorname{rk}_{+}^{\otimes}\bigl(\chi(A)-\epsilon\cdot\chi(1)\bigr)\ge\frac{\epsilon q^{d+1}}{L\binom{m}{\le d}}. \tag*{(41)} rk + ⊗ ( χ ( A ) − ϵ ⋅ χ ( 1 ) ) ≥ L ( ≤ d m ) ϵ q d + 1 . (41) More generally the same bound holds for the number of terms in any identity c A ( a , z ) − ϵ = ∑ j = 1 r λ j ( a ) Q j ( z ) c_A(a,z)-\epsilon=\sum_{j=1}^{r}\lambda_j(a)Q_j(z) c A ( a , z ) − ϵ = ∑ j = 1 r λ j ( a ) Q j ( z ) in which each λ j \lambda_j λ j is an arbitrary nonnegative function of all the selector bits and each Q j Q_j Q j a nonnegative product over the individual table bits.
Proof. We prove the general form of the statement, for an identity c A ( a , z ) − ϵ = ∑ j ≤ r λ j ( a ) Q j ( z ) c_A(a,z)-\epsilon=\sum_{j\le r}\lambda_j(a)Q_j(z) c A ( a , z ) − ϵ = ∑ j ≤ r λ j ( a ) Q j ( z ) with each λ j ≥ 0 \lambda_j\ge0 λ j ≥ 0 arbitrary on the selector bits and each Q j Q_j Q j a nonnegative product over the table bits; a tensor decomposition into r r r rank-one set-multilinear terms is the special case in which each λ j \lambda_j λ j is itself a product over the selector bits. All coordinates are ± 1 \pm1 ± 1 -valued and each Q j Q_j Q j is written in the corresponding form: discard identically zero terms and normalise each table factor to have uniform average one, absorbing its average into λ j \lambda_j λ j , so that
Q j ( z ) = ∏ u = 1 m ∏ t = 1 q ( 1 + β j , u , t z u , t ) , ∣ β j , u , t ∣ ≤ 1. Q_j(z)=\prod_{u=1}^{m}\prod_{t=1}^{q}(1+\beta_{j,u,t}z_{u,t}),\qquad\lvert\beta_{j,u,t}\rvert\le1. Q j ( z ) = u = 1 ∏ m t = 1 ∏ q ( 1 + β j , u , t z u , t ) , ∣ β j , u , t ∣ ≤ 1.
Every nonnegative product over the table bits can be written this way, by Lemma 2.6, the scalar η ′ \eta' η ′ of (3) being what has been absorbed into λ j \lambda_j λ j , and Q j Q_j Q j is 2 m q 2^{mq} 2 m q times a product distribution, in the bias parametrisation of Lemma 3.22. Write h j , u , t : = ∣ β j , u , t ∣ h_{j,u,t}:=\lvert\beta_{j,u,t}\rvert h j , u , t := ∣ β j , u , t ∣ and let Z j : = max z Q j ( z ) = ∏ u , t ( 1 + h j , u , t ) Z_j:=\max_zQ_j(z)=\prod_{u,t}(1+h_{j,u,t}) Z j := max z Q j ( z ) = ∏ u , t ( 1 + h j , u , t ) be the value of Q j Q_j Q j at its modal assignment z u , t = sgn β j , u , t z_{u,t}=\operatorname{sgn}\beta_{j,u,t} z u , t = sgn β j , u , t . All summands are nonnegative and c A − ϵ ≤ L c_A-\epsilon\le L c A − ϵ ≤ L , so evaluating at a modal table assignment for Q j Q_j Q j gives
λ j ( a ) Z j ≤ L for every j and every a , (42) \lambda_j(a)Z_j\le L\qquad\text{for every }j\text{ and every }a, \tag*{(42)} λ j ( a ) Z j ≤ L for every j and every a , (42)
an estimate uniform in the selector assignment; the selector may therefore be chosen later.
Localising at a selector. Fix a ∈ [ q ] m a\in[q]^m a ∈ [ q ] m , that is, pass to one copy. Write y u : = z u , a u y_u:=z_{u,a_u} y u := z u , a u for the addressed bits. Now average the identity c A ( a , z ) − ϵ = ∑ j λ j ( a ) Q j ( z ) c_A(a,z)-\epsilon=\sum_j\lambda_j(a)Q_j(z) c A ( a , z ) − ϵ = ∑ j λ j ( a ) Q j ( z ) over the unaddressed table bits, independently and uniformly. On the right each factor 1 + β j , u , t z u , t 1+\beta_{j,u,t}z_{u,t} 1 + β j , u , t z u , t with t ≠ a u t\ne a_u t = a u averages to 1 1 1 , since E z u , t = 0 \mathbb{E}z_{u,t}=0 E z u , t = 0 ; only the m m m addressed factors survive. On the left, by (40), c A − ϵ c_A-\epsilon c A − ϵ already depends on the addressed bits alone, and equals V ( y ) − ϵ \mathcal{V}(y)-\epsilon V ( y ) − ϵ there. Hence
V ( y ) − ϵ = ∑ j = 1 r λ j ( a ) Q j , a ( y ) , Q j , a ( y ) : = ∏ u = 1 m ( 1 + β j , u , a u y u ) . (43) \mathcal{V}(y)-\epsilon=\sum_{j=1}^{r}\lambda_j(a)Q_{j,a}(y),\qquad Q_{j,a}(y):=\prod_{u=1}^{m}(1+\beta_{j,u,a_u}y_u). \tag*{(43)} V ( y ) − ϵ = j = 1 ∑ r λ j ( a ) Q j , a ( y ) , Q j , a ( y ) := u = 1 ∏ m ( 1 + β j , u , a u y u ) . (43)
The base functional E ~ \widetilde{\mathbb{E}} E is defined only in Fourier degree at most d d d ; extend it to every function on { − 1 , 1 } m \{-1,1\}^m { − 1 , 1 } m by
Λ ( g ) : = E ~ ( g ≤ d ) . \Lambda(g):=\widetilde{\mathbb{E}}(g^{\le d}). Λ ( g ) := E ( g ≤ d ) .
Then Λ \Lambda Λ is linear and agrees with E ~ \widetilde{\mathbb{E}} E in degree at most d d d , so Λ ( V − ϵ ) = E ~ V − ϵ = − ϵ \Lambda(\mathcal{V}-\epsilon)=\widetilde{\mathbb{E}}\mathcal{V}-\epsilon=-\epsilon Λ ( V − ϵ ) = E V − ϵ = − ϵ because deg V ≤ k ≤ d \deg\mathcal{V}\le k\le d deg V ≤ k ≤ d . Applying Λ \Lambda Λ to (43) therefore gives ∑ j λ j ( a ) Λ ( Q j , a ) = − ϵ \sum_j\lambda_j(a)\Lambda(Q_{j,a})=-\epsilon ∑ j λ j ( a ) Λ ( Q j , a ) = − ϵ .
Testing one localised product. For F ⊆ [ m ] F\subseteq[m] F ⊆ [ m ] and σ ∈ { − 1 , 1 } F \sigma\in\{-1,1\}^F σ ∈ { − 1 , 1 } F put ρ F , σ ( y ) : = ∏ u ∈ F ( 1 + σ u y u ) = 2 ∣ F ∣ 1 [ y F = σ ] \rho_{F,\sigma}(y):=\prod_{u\in F}(1+\sigma_u y_u)=2^{|F|}\mathbf{1}[y_F=\sigma] ρ F , σ ( y ) := ∏ u ∈ F ( 1 + σ u y u ) = 2 ∣ F ∣ 1 [ y F = σ ] , which is nonnegative, supported on the subcube y F = σ y_F=\sigma y F = σ and of average E ρ F , σ = 1 \mathbb{E}\rho_{F,\sigma}=1 E ρ F , σ = 1 . If ∣ F ∣ ≤ d |F|\le d ∣ F ∣ ≤ d then Λ ( ρ F , σ ) ≥ 0 \Lambda(\rho_{F,\sigma})\ge0 Λ ( ρ F , σ ) ≥ 0 by local nonnegativity. For arbitrary F F F , expanding the product gives ρ F , σ = ∑ α ⊆ F σ α y α \rho_{F,\sigma}=\sum_{\alpha\subseteq F}\sigma^\alpha y^\alpha ρ F , σ = ∑ α ⊆ F σ α y α with σ α : = ∏ u ∈ α σ u ∈ { − 1 , 1 } \sigma^\alpha:=\prod_{u\in\alpha}\sigma_u\in\{-1,1\} σ α := ∏ u ∈ α σ u ∈ { − 1 , 1 } , so
Λ ( ρ F , σ ) = ∑ α ⊆ F ∣ α ∣ ≤ d σ α E ~ y α , \Lambda(\rho_{F,\sigma})=\sum_{\substack{\alpha\subseteq F\\|\alpha|\le d}}\sigma^\alpha\widetilde{\mathbb{E}}y^\alpha, Λ ( ρ F , σ ) = α ⊆ F ∣ α ∣ ≤ d ∑ σ α E y α ,
a sum of at most ( m ≤ d ) \binom{m}{\le d} ( ≤ d m ) terms whatever ∣ F ∣ |F| ∣ F ∣ is. Each term is at most one in absolute value: 1 ± x α 1\pm x^\alpha 1 ± x α is a nonnegative function of ∣ α ∣ |\alpha| ∣ α ∣ coordinates, so
∣ E ~ x α ∣ ≤ E ~ 1 = 1 ( ∣ α ∣ ≤ d ) . (44) \left|\widetilde{\mathbb{E}}x^\alpha\right|\le\widetilde{\mathbb{E}}1=1\qquad(|\alpha|\le d). \tag*{(44)} E x α ≤ E 1 = 1 ( ∣ α ∣ ≤ d ) . (44)
Hence ∣ Λ ( ρ F , σ ) ∣ ≤ ( m ≤ d ) |\Lambda(\rho_{F,\sigma})|\le\binom{m}{\le d} ∣Λ ( ρ F , σ ) ∣ ≤ ( ≤ d m ) .
For the fixed j j j and a a a of this step abbreviate β u : = β j , u , a u \beta_u:=\beta_{j,u,a_u} β u := β j , u , a u , σ u : = sgn ( β u ) \sigma_u:=\operatorname{sgn}(\beta_u) σ u := sgn ( β u ) and h u : = h j , u , a u = ∣ β u ∣ h_u:=h_{j,u,a_u}=|\beta_u| h u := h j , u , a u = ∣ β u ∣ , so that β u = h u σ u \beta_u=h_u\sigma_u β u = h u σ u and hence
1 + β u y u = ( 1 − h u ) + h u ( 1 + σ u y u ) . 1+\beta_u y_u=(1-h_u)+h_u(1+\sigma_u y_u). 1 + β u y u = ( 1 − h u ) + h u ( 1 + σ u y u ) .
Substitute this in each of the m m m factors of Q j , a Q_{j,a} Q j , a and expand the product. Each resulting term is indexed by the set F ⊆ [ m ] F\subseteq[m] F ⊆ [ m ] of coordinates at which the second summand was taken, and equals ∏ u ∈ F h u ∏ u ∉ F ( 1 − h u ) \prod_{u\in F}h_u\prod_{u\notin F}(1-h_u) ∏ u ∈ F h u ∏ u ∈ / F ( 1 − h u ) times ρ F , σ F \rho_{F,\sigma_F} ρ F , σ F , where σ F : = ( σ u ) u ∈ F \sigma_F:=(\sigma_u)_{u\in F} σ F := ( σ u ) u ∈ F . These coefficients are the probabilities of the random set F F F containing each u u u independently with probability h u h_u h u , so
Q j , a = E F [ ρ F , σ F ] . Q_{j,a}=\mathbb{E}_F[\rho_{F,\sigma_F}]. Q j , a = E F [ ρ F , σ F ] .
A coordinate with β u = 0 \beta_u=0 β u = 0 has h u = 0 h_u=0 h u = 0 and so never enters F F F , which makes the choice of σ u \sigma_u σ u there immaterial. Apply Λ \Lambda Λ and split this expectation at ∣ F ∣ = d |F|=d ∣ F ∣ = d : the terms with ∣ F ∣ ≤ d |F|\le d ∣ F ∣ ≤ d contribute nonnegatively, and each term with ∣ F ∣ > d |F|>d ∣ F ∣ > d is at least − ( m ≤ d ) -\binom{m}{\le d} − ( ≤ d m ) . Writing D : = d + 1 D:=d+1 D := d + 1 and τ j ( a ) : = Pr [ ∣ F ∣ ≥ D ] \tau_j(a):=\Pr[|F|\ge D] τ j ( a ) := Pr [ ∣ F ∣ ≥ D ] ,
Λ ( Q j , a ) ≥ − ( m ≤ d ) τ j ( a ) . (45) \Lambda(Q_{j,a})\ge-\binom{m}{\le d}\tau_j(a). \tag*{(45)} Λ ( Q j , a ) ≥ − ( ≤ d m ) τ j ( a ) . (45)
Choosing a common selector. Bound τ j ( a ) \tau_j(a) τ j ( a ) first. For an integer s ≥ 0 s\ge0 s ≥ 0 the binomial coefficient ( s D ) \binom{s}{D} ( D s ) is 0 when s < D s<D s < D and at least 1 when s ≥ D s\ge D s ≥ D , so ( ∣ F ∣ D ) \binom{|F|}{D} ( D ∣ F ∣ ) dominates the indicator of the event ∣ F ∣ ≥ D |F|\ge D ∣ F ∣ ≥ D ; and ( ∣ F ∣ D ) \binom{|F|}{D} ( D ∣ F ∣ ) counts the D D D -element subsets of F F F . Hence
τ j ( a ) = Pr [ ∣ F ∣ ≥ D ] ≤ E F ( ∣ F ∣ D ) = ∑ U ⊆ [ m ] ∣ U ∣ = D Pr [ U ⊆ F ] = ∑ U ⊆ [ m ] ∣ U ∣ = D ∏ u ∈ U h u , \tau_j(a)=\Pr[|F|\ge D]\le\mathbb{E}_F\binom{|F|}{D}=\sum_{\substack{U\subseteq[m]\\|U|=D}}\Pr[U\subseteq F]=\sum_{\substack{U\subseteq[m]\\|U|=D}}\prod_{u\in U}h_u, τ j ( a ) = Pr [ ∣ F ∣ ≥ D ] ≤ E F ( D ∣ F ∣ ) = U ⊆ [ m ] ∣ U ∣ = D ∑ Pr [ U ⊆ F ] = U ⊆ [ m ] ∣ U ∣ = D ∑ u ∈ U ∏ h u ,
the coordinates entering F F F independently. Now let a a a vary: draw each a u a_u a u independently and uniformly from [ q ] [q] [ q ] . Since h u = h j , u , a u h_u=h_{j,u,a_u} h u = h j , u , a u depends on a u a_u a u we write it out in full, and averaging, with independence used once more,
E a τ j ( a ) ≤ ∑ ∣ U ∣ = D ∏ u ∈ U 1 q ∑ t ≤ q h j , u , t = q − D ∑ ∣ U ∣ = D ∏ u ∈ U ( ∑ t ≤ q h j , u , t ) ≤ q − D ∏ u , t ( 1 + h j , u , t ) = q − D Z j , \mathbb{E}_a\tau_j(a)\le\sum_{|U|=D}\prod_{u\in U}\frac{1}{q}\sum_{t\le q}h_{j,u,t}=q^{-D}\sum_{|U|=D}\prod_{u\in U}\left(\sum_{t\le q}h_{j,u,t}\right)\le q^{-D}\prod_{u,t}(1+h_{j,u,t})=q^{-D}Z_j, E a τ j ( a ) ≤ ∣ U ∣ = D ∑ u ∈ U ∏ q 1 t ≤ q ∑ h j , u , t = q − D ∣ U ∣ = D ∑ u ∈ U ∏ ( t ≤ q ∑ h j , u , t ) ≤ q − D u , t ∏ ( 1 + h j , u , t ) = q − D Z j ,
the last inequality because expanding the middle sum gives distinct squarefree monomials — one factor h j , u , t h_{j,u,t} h j , u , t from each of D D D distinct blocks u u u — each occurring with coefficient one in the expansion of ∏ u , t ( 1 + h j , u , t ) \prod_{u,t}(1+h_{j,u,t}) ∏ u , t ( 1 + h j , u , t ) , all of whose monomials are nonnegative. Dividing by Z j Z_j Z j and summing over j j j gives E a ∑ j τ j ( a ) / Z j ≤ r q − D \mathbb{E}_a\sum_j\tau_j(a)/Z_j\le rq^{-D} E a ∑ j τ j ( a ) / Z j ≤ r q − D , so some selector assignment a ∗ a^* a ∗ — depending on the whole purported decomposition — satisfies
∑ j = 1 r τ j ( a ∗ ) Z j ≤ r q − D . (46) \sum_{j=1}^{r}\frac{\tau_j(a^*)}{Z_j}\le rq^{-D}. \tag*{(46)} j = 1 ∑ r Z j τ j ( a ∗ ) ≤ r q − D . (46)
Contradiction at the selected copy. At a = a ∗ a=a^* a = a ∗ , combining (43), (45), (42) and (46) in that order,
− ϵ = ∑ j λ j ( a ∗ ) Λ ( Q j , a ∗ ) ≥ − ( m ≤ d ) ∑ j λ j ( a ∗ ) τ j ( a ∗ ) ≥ − L ( m ≤ d ) ∑ j τ j ( a ∗ ) Z j ≥ − L ( m ≤ d ) r q − D , \begin{aligned}
-\epsilon=\sum_j\lambda_j(a^*)\Lambda(Q_{j,a^*})\ge-\binom{m}{\le d}\sum_j\lambda_j(a^*)\tau_j(a^*) \\
&\ge-L\binom{m}{\le d}\sum_j\frac{\tau_j(a^*)}{Z_j}\ge-L\binom{m}{\le d}rq^{-D},
\end{aligned} − ϵ = j ∑ λ j ( a ∗ ) Λ ( Q j , a ∗ ) ≥ − ( ≤ d m ) j ∑ λ j ( a ∗ ) τ j ( a ∗ ) ≥ − L ( ≤ d m ) j ∑ Z j τ j ( a ∗ ) ≥ − L ( ≤ d m ) r q − D ,
which rearranges to (41). A nonnegative rank-one tensor over the raw Boolean variables is in particular of the displayed form, so the bound applies to rk + ⊗ \operatorname{rk}^{\otimes}_{+} rk + ⊗ and, by Corollary 2.9 , to the size of every Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMCϵ ^\epsilon ϵ proof of A A A . □
The base we use is Tseitin itself.
Corollary 3.37 (Lifted Tseitin is hard at exponentially small shift). Fix an explicit family of connected d G d_G d G -regular expanders with d G = O ( 1 ) d_G=O(1) d G = O ( 1 ) and edge expansion h = Ω ( 1 ) h=\Omega(1) h = Ω ( 1 ) , and give each member an explicitly computable odd charge, for instance a single charged vertex. There are constants c > 0 c>0 c > 0 and w w w , depending only on d G d_G d G and h h h , and an explicit family of tautological DNFs A N : = Ind q ( TSG ) A_N:=\operatorname{Ind}_q(\operatorname{TSG}) A N := Ind q ( TSG ) on N N N Boolean variables with O ( N ) O(N) O ( N ) terms of width at most w w w , such that for all sufficiently large members
2 − c N ≤ ϵ ≤ 1 ⟹ rk χ ⊗ ( A N ) − ϵ ⋅ χ ( 1 ) ≥ 2 c N . 2^{-cN}\le\epsilon\le1\qquad\Longrightarrow\qquad\operatorname{rk}^{\otimes}_{\chi}(A_N)-\epsilon\cdot\chi(1)\ge2^{cN}. 2 − c N ≤ ϵ ≤ 1 ⟹ rk χ ⊗ ( A N ) − ϵ ⋅ χ ( 1 ) ≥ 2 c N . Every depth-three AMCϵ ^\epsilon ϵ proof of A N A_N A N therefore has at least 2 c N 2^{cN} 2 c N product terms throughout that range of shifts.
Proof. Take I 0 : = T S G I_0 := \mathrm{TSG} I 0 := TSG as base: m = ∣ E ∣ = d G n / 2 m = |E| = d_G n/2 m = ∣ E ∣ = d G n /2 variables, L = n L = n L = n constraints of arity d G d_G d G , and unsatisfiable. By [18 ] its sum-of-squares refutation degree is Ω ( n ) \Omega(n) Ω ( n ) , hence so is its Sherali–Adams refutation degree, and linear programming duality for Sherali–Adams turns the latter into a degree-d d d local expectation functional with d = α m d = \alpha m d = α m for a constant α = α ( d G , h ) > 0 \alpha= \alpha(d_G,h) > 0 α = α ( d G , h ) > 0 . Each B ℓ B_\ell B ℓ is a nonnegative d G d_G d G -junta and E ~ V = 0 \widetilde{\mathbb{E}}V = 0 E V = 0 , so E ~ B ℓ = 0 \widetilde{\mathbb{E}}B_\ell= 0 E B ℓ = 0 for every ℓ \ell ℓ : on each vertex’s scope the local distribution satisfies that vertex’s parity with probability one. (Each nonnegative j j j -junta is a nonnegative combination of subcube indicators, which are idempotent and hence squares of degree-j j j polynomials; a degree-2 d 2d 2 d sum-of-squares solution therefore restricts to a degree-d d d local expectation functional, and the factor 2 2 2 is absorbed into α \alpha α .)
Put b : = ⌈ 4 / α ⌉ b := \lceil4/\alpha\rceil b := ⌈ 4/ α ⌉ , q : = 2 b q := 2^b q := 2 b , κ : = q + b \kappa:= q+b κ := q + b and c : = κ − 1 c := \kappa^{-1} c := κ − 1 , all constants determined by d G d_G d G and h h h . The lift has N = κ m N = \kappa m N = κm variables, at most L 2 d G q d G q = O ( N ) L2^{d_Gq}d_G^{q} = O(N) L 2 d G q d G q = O ( N ) terms of width at most w : = d G ( b + 1 ) w := d_G(b+1) w := d G ( b + 1 ) , and is constructible in polynomial time. Since ( m ≤ d ) ≤ 2 m \binom{m}{\le d} \le2^m ( ≤ d m ) ≤ 2 m and b ( d + 1 ) ≥ 4 m b(d+1) \ge4m b ( d + 1 ) ≥ 4 m , (41) gives r ≥ ϵ 2 b ( d + 1 ) / ( L ( m ≤ d ) ) ≥ ϵ 2 3 m / L r \ge\epsilon2^{b(d+1)}/(L\binom{m}{\le d}) \ge\epsilon2^{3m}/L r ≥ ϵ 2 b ( d + 1 ) / ( L ( ≤ d m ) ) ≥ ϵ 2 3 m / L ; here L = n ≤ m ≤ 2 m L = n \le m \le2^m L = n ≤ m ≤ 2 m , so r ≥ ϵ 2 2 m r \ge\epsilon2^{2m} r ≥ ϵ 2 2 m , and ϵ ≥ 2 − m \epsilon\ge2^{-m} ϵ ≥ 2 − m gives r ≥ 2 m = 2 c N r \ge2^m = 2^{cN} r ≥ 2 m = 2 c N .
Remark 3.38 (Limits of the construction). The argument acts on products over the individual table bits, and a general nonnegative column factor of a bipartite flattening need not factor over them, so it gives nothing for rk + \operatorname{rk}_+ rk + and nothing at mixed-product depth above zero. For T S G \mathrm{TS}_G TS G itself no such bound exists at all, by Theorem 3.33 ; whether Ind q ( T S G ) \operatorname{Ind}_q(\mathrm{TS}_G) Ind q ( TS G ) is likewise easy at mixed-product depth zero we do not know. § 4.4 runs the same skeleton with copies of the same Tseitin base planted in a shared variable set instead of private tables, and trades the reach in ϵ \epsilon ϵ for exactly those two properties; the comparison is set out there.
Lower bounds at bounded mixed-product depth
We now feed the flattening bound (6) . What it asks for is a DNF tautology A A A , a bipartition of the blocks of χ ( A ) \chi(A) χ ( A ) , and a submatrix of M X I ∣ X I ( χ ( A ) ) − ϵ ⋅ J X I ∣ X I M_{X_I|X_I}(\chi(A))-\epsilon\cdot J_{X_I|X_I} M X I ∣ X I ( χ ( A )) − ϵ ⋅ J X I ∣ X I of large nonnegative rank; monotonicity of rk + \operatorname{rk}_+ rk + under submatrices then bounds the leaves of every monotone formula of mixed-product depth t t t computing χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A)-\epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) , and at t = 0 t=0 t = 0 gives the Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -A M C ϵ \mathrm{AMC}^{\epsilon} AMC ϵ proof size bound (5). Every such submatrix has the form M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J with M M M a submatrix of the unperturbed flattening and J J J the all-ones matrix of matching dimensions.
We isolate two ways of producing one, differing in where the rank lower bound comes from. In Recipe I the rows are a family of valid inequalities of a combinatorial polytope and the rank is its approximate extension complexity; in Recipe II (§ 4.4 ) the rows are the instances of a planted constraint satisfaction family and the rank comes from a Sherali–Adams integrality gap, no polytope being named. The two have disjoint hypotheses and different admissible ranges of ϵ \epsilon ϵ , and share only the encoding lemma, Proposition 2.14, so we state each interface once and present the constructions as instances.
Recipe I is stated in the language of approximate extension complexity, which we fix here. Let P P P be a polytope with vertex set V V V and let F \mathcal{F} F be a family of inequalities a ˉ f ⋅ x ˉ − b ≥ 0 \bar{a}_f\cdot\bar{x}-b\ge0 a ˉ f ⋅ x ˉ − b ≥ 0 valid at every vertex. Write P δ F P^\mathcal{F}_\delta P δ F for the outer relaxation cut out by a ˉ f ⋅ x ˉ − b + δ ≥ 0 \bar{a}_f\cdot\bar{x}-b+\delta\ge0 a ˉ f ⋅ x ˉ − b + δ ≥ 0 , f ∈ F f\in\mathcal{F} f ∈ F , so that P 0 F P^\mathcal{F}_0 P 0 F is the body cut out by F \mathcal{F} F itself. Let S S S be the slack matrix of F \mathcal{F} F at V V V and put M : = S + b ⋅ J M := S+b\cdot J M := S + b ⋅ J , so that M − ϵ ⋅ J = S + ( b − ϵ ) J M-\epsilon\cdot J=S+(b-\epsilon)J M − ϵ ⋅ J = S + ( b − ϵ ) J is the slack matrix at V V V of the relaxed family defining P b − ϵ F P^\mathcal{F}_{b-\epsilon} P b − ϵ F . Then
xc ( P , P b − ϵ F ) ≤ rk + ( M − ϵ ⋅ J ) ≤ xc ( P , P b − ϵ F ) + 1 , xc ( P , P ′ ) : = min { xc ( Q ) : P ⊆ Q ⊆ P ′ } . (47) \operatorname{xc}(P,P^\mathcal{F}_{b-\epsilon}) \le\operatorname{rk}_+(M-\epsilon\cdot J) \le\operatorname{xc}(P,P^\mathcal{F}_{b-\epsilon})+1,\qquad\operatorname{xc}(P,P') := \min\{\operatorname{xc}(Q): P\subseteq Q\subseteq P'\}. \tag*{(47)} xc ( P , P b − ϵ F ) ≤ rk + ( M − ϵ ⋅ J ) ≤ xc ( P , P b − ϵ F ) + 1 , xc ( P , P ′ ) := min { xc ( Q ) : P ⊆ Q ⊆ P ′ } . (47)
a nonnegative factorization of the perturbed slack matrix is an extended formulation of a body squeezed between P P P and the relaxation, and conversely up to one further term, the constant slack such a formulation may need [21 Lem. 3.4] , taken there from [3 ] . Equality can fail: for P = { 1 } ⊂ R P = \{1\} \subset\mathbb{R} P = { 1 } ⊂ R with the single inequality x ≥ 1 x \ge1 x ≥ 1 and 0 < ϵ < 1 0 < \epsilon< 1 0 < ϵ < 1 the perturbed slack matrix is [ 1 − ϵ ] [1-\epsilon] [ 1 − ϵ ] , of nonnegative rank one, while Q = P Q = P Q = P is cut out by an equality and needs no inequality at all. Recipe I uses only the left inequality, and Q Q Q is allowed to be an unbounded polyhedron, as the outer body of § 4.3 is.
The construction behind the first interface is this. Specialise the above to a combinatorial setting: a finite ground set E E E , a family H \mathcal{H} H of subsets of E E E whose indicator vectors χ H \chi_H χ H are the vertices of P P P , and for F \mathcal{F} F a family of covering inequalities ⟨ a f , x ⟩ ≥ b \langle a_f,x\rangle\ge b ⟨ a f , x ⟩ ≥ b — covering in the usual sense of nonnegative coefficients and the ≥ \ge ≥ direction, and restricted further so that every a f a_f a f lies in { 0 , 1 } E \{0,1\}^{E} { 0 , 1 } E and the right hand side b b b is the same throughout the family. For such an a f a_f a f the left hand side ⟨ a f , χ H ⟩ \langle a_f,\chi_H\rangle ⟨ a f , χ H ⟩ is just the overlap count ∣ supp ( a f ) ∩ H ∣ |\operatorname{supp}(a_f) \cap H| ∣ supp ( a f ) ∩ H ∣ , and the slack of f f f at χ H \chi_H χ H is that count less b b b . So we want a DNF whose valid slice — the submatrix of its flattening carried by the valid encodings, in the sense of Proposition 2.14 — is the overlap count ∣ supp ( a f ) ∩ H ∣ |\operatorname{supp}(a_f) \cap H| ∣ supp ( a f ) ∩ H ∣ : its perturbation is then the slack matrix of a nested pair of bodies, up to the shift by b b b , and (47) applies to it. Proposition 2.14 turns any list of witness conjunctions into such a DNF, the valid slice counting how many of them a valid pair satisfies. We choose the list element by element: a group of conjunctions for each e ∈ E e \in E e ∈ E , arranged so that exactly one member of the group fires when e e e lies in supp ( a f ) \operatorname{supp}(a_f) supp ( a f ) and in H H H at once, and none fires otherwise. One of the two memberships is free: the encoding of a member of H \mathcal{H} H includes its indicator vector, so e ∈ H e \in H e ∈ H is recorded by a single variable. The other must be read off whatever encoding F \mathcal{F} F carries, by a test of constant size — and read without multiplicity , since a second firing for the same e e e would inflate the slice. Invalid encodings on either side are absorbed by the two invalidity DNFs.
Proposition 4.1 (Recipe I: covering polytopes). Let E E E be a finite ground set, H \mathcal{H} H a family of subsets of E E E , and F \mathcal{F} F a family of inequalities ⟨ a f , x ⟩ ≥ b \langle a_f,x\rangle\ge b ⟨ a f , x ⟩ ≥ b with a f ∈ { 0 , 1 } E a_f \in\{0,1\}^{E} a f ∈ { 0 , 1 } E and a common right hand side b ∈ Z > 0 b \in\mathbb{Z}_{>0} b ∈ Z > 0 , all valid at every χ H \chi_H χ H , H ∈ H H \in\mathcal{H} H ∈ H ; put P : = conv { χ H : H ∈ H } P := \operatorname{conv}\{\chi_H : H \in\mathcal{H}\} P := conv { χ H : H ∈ H } . Encode the members of F \mathcal{F} F over a set X X X of Boolean variables and those of H \mathcal{H} H over a disjoint set Y Y Y containing their indicator bits, γ e ∈ Y \gamma_e \in Y γ e ∈ Y recording e ∈ H e \in H e ∈ H . Either encoding may carry auxiliary variables beyond these, in which case their consistency with the encoded object is part of validity. Assume:
(I.1) for every e ∈ E e \in E e ∈ E there are conjunctions ρ e , 1 , … , ρ e , c \rho_{e,1},\ldots,\rho_{e,c} ρ e , 1 , … , ρ e , c of literals in X X X , with c = O ( 1 ) c = O(1) c = O ( 1 ) , of which exactly one holds at the encoding of f f f when e ∈ supp ( a f ) e \in\operatorname{supp}(a_f) e ∈ supp ( a f ) and none holds otherwise;
(I.2) invalid encodings are recognised exactly by DNFs I R ( X ) I_R(X) I R ( X ) and I C ( Y ) I_C(Y) I C ( Y ) with polynomially many terms.
Then
A : = I R ∨ I C ∨ ⋁ e ∈ E ⋁ j = 1 c ( ρ e , j ∧ γ e ) A := I_R \vee I_C \vee\bigvee_{e \in E}\bigvee_{j=1}^{c}(\rho_{e,j} \wedge\gamma_e) A := I R ∨ I C ∨ e ∈ E ⋁ j = 1 ⋁ c ( ρ e , j ∧ γ e ) is a tautological DNF with ∣ I R ∣ + ∣ I C ∣ + c ∣ E ∣ |I_R| + |I_C| + c|E| ∣ I R ∣ + ∣ I C ∣ + c ∣ E ∣ terms whose valid slice is M ( f , H ) = ∣ supp ( a f ) ∩ H ∣ M(f,H) = |\operatorname{supp}(a_f) \cap H| M ( f , H ) = ∣ supp ( a f ) ∩ H ∣ , so that M = S + b ⋅ J M = S + b \cdot J M = S + b ⋅ J for the slack matrix S S S of F \mathcal{F} F at H \mathcal{H} H ; and every monotone formula of mixed-product depth at most t t t computing χ ( A ) − ϵ ⋅ χ ( 1 ) \chi(A) - \epsilon\cdot\chi(1) χ ( A ) − ϵ ⋅ χ ( 1 ) has at least
( min { χ ( Q ) : P ⊆ Q ⊆ P b − ϵ F } ) 1 / 2 t \left(\min\left\{\chi(Q) : P \subseteq Q \subseteq P^{\mathcal{F}}_{b-\epsilon}\right\}\right)^{1/2^t} ( min { χ ( Q ) : P ⊆ Q ⊆ P b − ϵ F } ) 1/ 2 t leaves.
Proof. On a valid pair ( f , H ) (f,H) ( f , H ) the two invalidity DNFs are unsatisfied, and by (I.1) exactly one witness term ρ e , j ∧ γ e \rho_{e,j} \wedge\gamma_e ρ e , j ∧ γ e is satisfied for each e ∈ supp ( a f ) ∩ H e \in\operatorname{supp}(a_f) \cap H e ∈ supp ( a f ) ∩ H and none for any other e e e ; the number of satisfied terms is therefore ∣ supp ( a f ) ∩ H ∣ = ⟨ a f , χ H ⟩ ≥ b ≥ 1 |\operatorname{supp}(a_f) \cap H| = \langle a_f, \chi_H\rangle\ge b \ge1 ∣ supp ( a f ) ∩ H ∣ = ⟨ a f , χ H ⟩ ≥ b ≥ 1 . Proposition 2.14 now gives the term count, the tautology, and the valid slice M M M . Finally M − ϵ ⋅ J = S + ( b − ϵ ) J M - \epsilon\cdot J = S + (b - \epsilon)J M − ϵ ⋅ J = S + ( b − ϵ ) J is the slack matrix of the defining inequalities of P b − ϵ F P^F_{b-\epsilon} P b − ϵ F evaluated at the points χ H \chi_H χ H , so the left inequality of (47) bounds its nonnegative rank below by the displayed minimum, and Corollary 2.15 converts that into the leaf bound.
The right hand side matters. With b = 1 b = 1 b = 1 and ϵ ∈ ( 0 , 1 ) \epsilon\in(0,1) ϵ ∈ ( 0 , 1 ) the relaxation P b − ϵ F P^F_{b-\epsilon} P b − ϵ F loosens each inequality by less than a unit; with b > 1 b > 1 b > 1 it loosens by more, so the same range of ϵ \epsilon ϵ demands a bound against a weaker relaxation. Odd cut inequalities have b = 1 b = 1 b = 1 ; subtour elimination constraints, for instance, have b = 2 b = 2 b = 2 .
Sections 4.1–4.3 instantiate Recipe I with E E E the edge set of K 2 n K_{2n} K 2 n , H \mathcal{H} H the perfect matchings or the odd factors, and F \mathcal{F} F the odd cut inequalities x ˉ ( δ ( U ) ) ≥ 1 \bar{x}(\delta(U)) \ge1 x ˉ ( δ ( U )) ≥ 1 . Section 4.4 states its own interface, Proposition 4.5, which shares only Proposition 2.14 with this one.
Matching, perfect matching and odd factor polytopes
Let G = ( V , E ) G = (V,E) G = ( V , E ) be a graph. The matching polytope is
P M ( G ) = conv { χ M ∈ { 0 , 1 } E ⊂ R E ∣ M ⊆ E is a matching } , P_M(G) = \operatorname{conv}\{\chi_M \in\{0,1\}^{E} \subset\mathbb{R}^{E} \mid M \subseteq E \text{ is a matching}\}, P M ( G ) = conv { χ M ∈ { 0 , 1 } E ⊂ R E ∣ M ⊆ E is a matching } ,
which is described by x e ≥ 0 x_e \ge0 x e ≥ 0 for e ∈ E e \in E e ∈ E , x ( δ ( v ) ) ≤ 1 x(\delta(v)) \le1 x ( δ ( v )) ≤ 1 for v ∈ V v \in V v ∈ V , and x ( E ( U ) ) ≤ ∣ U ∣ − 1 2 x(E(U)) \le\frac{|U|-1}{2} x ( E ( U )) ≤ 2 ∣ U ∣ − 1 for U ⊆ V U \subseteq V U ⊆ V with ∣ U ∣ |U| ∣ U ∣ odd.
If ∣ V ∣ |V| ∣ V ∣ is even, then the perfect matching polytope P P M ( G ) P_{\mathrm{PM}}(G) P PM ( G ) is the face of P M ( G ) P_M(G) P M ( G ) cut out by x ( δ ( v ) ) = 1 x(\delta(v)) = 1 x ( δ ( v )) = 1 , v ∈ V v \in V v ∈ V :
P P M ( G ) = conv { χ M ∈ { 0 , 1 } E ⊂ R E ∣ M ⊆ E is a perfect matching } , P_{\mathrm{PM}}(G) = \operatorname{conv}\{\chi_M \in\{0,1\}^{E} \subset\mathbb{R}^{E} \mid M \subseteq E \text{ is a perfect matching}\}, P PM ( G ) = conv { χ M ∈ { 0 , 1 } E ⊂ R E ∣ M ⊆ E is a perfect matching } ,
which is described by
x e ≥ 0 x_e \ge0 x e ≥ 0 , e ∈ E e \in E e ∈ E , together with x ( δ ( v ) ) = 1 x(\delta(v)) = 1 x ( δ ( v )) = 1 , v ∈ V v \in V v ∈ V ;
x ( δ ( U ) ) ≥ 1 x(\delta(U)) \ge1 x ( δ ( U )) ≥ 1 , U ⊆ V U \subseteq V U ⊆ V , ∣ U ∣ |U| ∣ U ∣ odd.
For an arbitrary graph neither list is irredundant, and the equalities are not facets.
An odd factor of G G G is a spanning subgraph in which every vertex has odd degree; equivalently, a T T T -join for T = V T = V T = V . This forces ∣ V ∣ |V| ∣ V ∣ to be even. The odd factor polytope is
P O F ( G ) = conv { χ H ∈ { 0 , 1 } E : H ⊆ E is an odd factor of G } . P_{\mathrm{OF}}(G) = \operatorname{conv}\{\chi_H \in\{0,1\}^{E} : H \subseteq E \text{ is an odd factor of } G\}. P OF ( G ) = conv { χ H ∈ { 0 , 1 } E : H ⊆ E is an odd factor of G } .
The odd cut inequalities x ( δ ( U ) ) ≥ 1 x(\delta(U)) \ge1 x ( δ ( U )) ≥ 1 , ∣ U ∣ |U| ∣ U ∣ odd, are valid for it — this is checked in Section 4.3 — but they do not describe it, even together with 0 ≤ x e ≤ 1 0 \le x_e \le1 0 ≤ x e ≤ 1 : in K 4 K_4 K 4 the indicator of a four-cycle satisfies all of them, while every vertex of the four-cycle has even degree, so it is not an odd factor and, being integral, not a convex combination of odd factors either. What those inequalities together with x ≥ 0 x \ge0 x ≥ 0 do describe, by Edmonds and Johnson [12 ] , is the dominant P O F ( G ) + R ≥ 0 E P_{\mathrm{OF}}(G) + \mathbb{R}^{E}_{\ge0} P OF ( G ) + R ≥ 0 E . Nothing below needs a facet description: only the validity of the odd cut inequalities is used, and that every perfect matching is a 1-regular odd factor, so that P P M ( G ) P_{\mathrm{PM}}(G) P PM ( G ) is the face of P O F ( G ) P_{\mathrm{OF}}(G) P OF ( G ) cut out by x ( δ ( v ) ) = 1 x(\delta(v)) = 1 x ( δ ( v )) = 1 for all v ∈ V v \in V v ∈ V .
Write S P M ( G ) S_{\mathrm{PM}}(G) S PM ( G ) and S O F ( G ) S_{\mathrm{OF}}(G) S OF ( G ) for the odd cut slack matrices of the two bodies: the slacks of the odd cut inequalities at the vertices, a submatrix of the slack matrix of any description containing them. They share the same rows — indexed by odd cuts δ ( U ) \delta(U) δ ( U ) , ∣ U ∣ |U| ∣ U ∣ odd — and the same entry rule
S ( U , H ) = ∣ δ ( U ) ∩ H ∣ − 1 , S(U,H) = |\delta(U) \cap H| - 1, S ( U , H ) = ∣ δ ( U ) ∩ H ∣ − 1 ,
their columns ranging respectively over perfect matchings and over all odd factors of G G G . Since every perfect matching is an odd factor, S P M ( G ) S_{\mathrm{PM}}(G) S PM ( G ) is a column submatrix of S O F ( G ) S_{\mathrm{OF}}(G) S OF ( G ) ; by monotonicity of rk + \operatorname{rk}_{+} rk + under submatrices, rk + ( S O F ( G ) ) ≥ rk + ( S P M ( G ) ) \operatorname{rk}_{+}(S_{\mathrm{OF}}(G)) \ge\operatorname{rk}_{+}(S_{\mathrm{PM}}(G)) rk + ( S OF ( G )) ≥ rk + ( S PM ( G )) .
For the complete graph K 2 n K_{2n} K 2 n — the vertex count must be even for a perfect matching to exist, and we fix it as 2 n 2n 2 n throughout § 4.2 and § 4.3 and Appendix § 7 — we denote the corresponding polytopes simply by P P M P_{\mathrm{PM}} P PM , P O F P_{\mathrm{OF}} P OF . By Rothvoß [26 ] , rk + ( S P M ) = 2 Ω ( n ) \operatorname{rk}_{+}(S_{\mathrm{PM}}) = 2^{\Omega(n)} rk + ( S PM ) = 2 Ω ( n ) and therefore rk + ( S O F ) = 2 Ω ( n ) \operatorname{rk}_{+}(S_{\mathrm{OF}}) = 2^{\Omega(n)} rk + ( S OF ) = 2 Ω ( n ) . Writing M P M M_{\mathrm{PM}} M PM , M O F M_{\mathrm{OF}} M OF for the corresponding edge-count matrices, so that S P M = M P M − J S_{\mathrm{PM}} = M_{\mathrm{PM}} - J S PM = M PM − J and S O F = M O F − J S_{\mathrm{OF}} = M_{\mathrm{OF}} - J S OF = M OF − J , the same column-submatrix inclusion and Braun and Pokutta’s [4 ] ϵ \epsilon ϵ -perturbed bound give
rk + ( M O F − ϵ ⋅ J ) ≥ rk + ( M P M − ϵ ⋅ J ) = 2 Ω ( n ) (48) \operatorname{rk}_{+}(M_{\mathrm{OF}}-\epsilon\cdot J) \ge\operatorname{rk}_{+}(M_{\mathrm{PM}}-\epsilon\cdot J) = 2^{\Omega(n)} \tag*{(48)} rk + ( M OF − ϵ ⋅ J ) ≥ rk + ( M PM − ϵ ⋅ J ) = 2 Ω ( n ) (48)
for every constant ϵ ∈ ( 0 , 1 ) \epsilon\in(0,1) ϵ ∈ ( 0 , 1 ) , remaining superpolynomial down to ϵ = c 0 log log n / log n \epsilon= c_{0}\log\log n/\log n ϵ = c 0 log log n / log n for any fixed c 0 > 8 c_{0} > 8 c 0 > 8 . Appendix § 7 makes the constants explicit and establishes this sufficient range.
Perfect matching and odd factor DNFs
We define a DNF tautology A P M ( x ˉ , y ˉ , z ˉ ) A_{\mathrm{PM}}(\bar{x},\bar{y},\bar{z}) A PM ( x ˉ , y ˉ , z ˉ ) phrasing “for every odd cut x ˉ \bar{x} x ˉ , every perfect matching y ˉ \bar{y} y ˉ intersects δ ( x ˉ ) \delta(\bar{x}) δ ( x ˉ ) ” (z ˉ \bar{z} z ˉ are auxiliary variables). Specifically, A P M ( x ˉ , y ˉ , z ˉ ) A_{\mathrm{PM}}(\bar{x},\bar{y},\bar{z}) A PM ( x ˉ , y ˉ , z ˉ ) is the disjunction of the DNFs ¬ B o d d ( x ˉ , z ˉ ) \neg B_{\mathrm{odd}}(\bar{x},\bar{z}) ¬ B odd ( x ˉ , z ˉ ) , ¬ C m a t c h ( y ˉ ) \neg C_{\mathrm{match}}(\bar{y}) ¬ C match ( y ˉ ) , ¬ C V -cover ( y ˉ ) \neg C_{V\text{-cover}}(\bar{y}) ¬ C V -cover ( y ˉ ) and D i n t e r s e c t s ( x ˉ , y ˉ ) D_{\mathrm{intersects}}(\bar{x},\bar{y}) D intersects ( x ˉ , y ˉ ) , saying that either x ˉ \bar{x} x ˉ is not odd, or y ˉ \bar{y} y ˉ is not a matching, or y ˉ \bar{y} y ˉ does not cover V V V , or y ˉ \bar{y} y ˉ intersects δ ( x ˉ ) \delta(\bar{x}) δ ( x ˉ ) , respectively. These DNFs are in turn defined as follows.
¬ B o d d ( x ˉ , z ˉ ) \neg B_{\mathrm{odd}}(\bar{x},\bar{z}) ¬ B odd ( x ˉ , z ˉ ) : “x ˉ \bar{x} x ˉ is not odd”. The standard encoding of x 1 ⊕ ⋯ ⊕ x 2 n = 0 x_{1} \oplus\cdots\oplus x_{2n} = 0 x 1 ⊕ ⋯ ⊕ x 2 n = 0 as a DNF: extra Boolean variables z ˉ = { z x 1 ⊕ ⋯ ⊕ x i } i > 1 \bar{z} = \{z_{x_{1}\oplus\cdots\oplus x_{i}}\}_{i>1} z ˉ = { z x 1 ⊕ ⋯ ⊕ x i } i > 1 (we write z x 1 ≡ x 1 z_{x_{1}} \equiv x_{1} z x 1 ≡ x 1 ) represent values of partial XORs, and the terms are { z x 1 ⊕ ⋯ ⊕ x i a ∧ z x 1 ⊕ ⋯ ⊕ x i + 1 b } a , b ∈ { 0 , 1 } , i ∈ [ 2 n − 1 ] \{z_{x_{1}\oplus\cdots\oplus x_{i}}^{a} \wedge z_{x_{1}\oplus\cdots\oplus x_{i+1}}^{b}\}_{a,b\in\{0,1\},\,i\in[2n-1]} { z x 1 ⊕ ⋯ ⊕ x i a ∧ z x 1 ⊕ ⋯ ⊕ x i + 1 b } a , b ∈ { 0 , 1 } , i ∈ [ 2 n − 1 ] together with ¬ z x 1 ⊕ ⋯ ⊕ x 2 n \neg z_{x_{1}\oplus\cdots\oplus x_{2n}} ¬ z x 1 ⊕ ⋯ ⊕ x 2 n , where c ˉ : = 1 − c \bar{c} := 1-c c ˉ := 1 − c . A term of the first kind is satisfied exactly when the guessed partial XOR at step i + 1 i+1 i + 1 is inconsistent with the one at step i i i , so ¬ B o d d \neg B_{\mathrm{odd}} ¬ B odd is satisfied precisely when z ˉ \bar{z} z ˉ misrepresents the partial XORs of x ˉ \bar{x} x ˉ or when x ˉ \bar{x} x ˉ has even parity. This is what makes an odd x ˉ \bar{x} x ˉ with a consistent z ˉ \bar{z} z ˉ — the rows used in the lower bound below — contribute nothing to χ ( ¬ B o d d ) \chi(\neg B_{\mathrm{odd}}) χ ( ¬ B odd ) .
¬ C m a t c h ( y ˉ ) \neg C_{\mathrm{match}}(\bar{y}) ¬ C match ( y ˉ ) : “y ˉ \bar{y} y ˉ is not a matching”. Terms { y i , j ∧ y j , k } \{y_{i,j} \wedge y_{j,k}\} { y i , j ∧ y j , k } for j ∈ [ 2 n ] j \in[2n] j ∈ [ 2 n ] and i < k i < k i < k in [ 2 n ] ∖ { j } [2n]\setminus\{j\} [ 2 n ] ∖ { j } , one for each pair of distinct edges sharing the vertex j j j . Distinctness of i i i and k k k is needed: the term with i = k i=k i = k is y i , j y_{i,j} y i , j itself, which is satisfied by every nonempty matching, so it would contribute to the coefficient at a valid pair and the perturbed valid slice computed in § 4.3 would not be M P M − ϵ ⋅ J M_{\mathrm{PM}}-\epsilon\cdot J M PM − ϵ ⋅ J .
¬ C V -cover ( y ˉ ) \neg C_{V\text{-cover}}(\bar{y}) ¬ C V -cover ( y ˉ ) : “y ˉ \bar{y} y ˉ does not cover V V V ”. Terms { ⋀ j ∈ [ 2 n ] ∖ { i } ¬ y i , j } i ∈ [ 2 n ] \{\bigwedge_{j\in[2n]\setminus\{i\}}\neg y_{i,j}\}_{i\in[2n]} { ⋀ j ∈ [ 2 n ] ∖ { i } ¬ y i , j } i ∈ [ 2 n ] .
D i n t e r s e c t s ( x ˉ , y ˉ ) D_{\mathrm{intersects}}(\bar{x},\bar{y}) D intersects ( x ˉ , y ˉ ) : “y ˉ \bar{y} y ˉ intersects δ ( x ˉ ) \delta(\bar{x}) δ ( x ˉ ) ”. Terms { x i ∧ y i , j ∧ ¬ x j } \{x_i \wedge y_{i,j} \wedge\neg x_j\} { x i ∧ y i , j ∧ ¬ x j } .
The tautology A O F ( x ˉ , y ˉ , z ˉ , w ˉ ) A_{\mathrm{OF}}(\bar{x},\bar{y},\bar{z},\bar{w}) A OF ( x ˉ , y ˉ , z ˉ , w ˉ ) asserts the analogous principle for odd factors: “for every odd cut x ˉ \bar{x} x ˉ , every odd factor y ˉ \bar{y} y ˉ intersects δ ( x ˉ ) \delta(\bar{x}) δ ( x ˉ ) .” It is a tautology by the parity argument ∑ v ∈ V deg H ( v ) = 2 ∣ E H ( U ) ∣ + ∣ δ ( U ) ∩ H ∣ \sum_{v\in V}\deg_H(v)=2|E_H(U)|+|\delta(U)\cap H| ∑ v ∈ V deg H ( v ) = 2∣ E H ( U ) ∣ + ∣ δ ( U ) ∩ H ∣ : an odd number of odd summands is odd, so ∣ δ ( U ) ∩ H ∣ |\delta(U)\cap H| ∣ δ ( U ) ∩ H ∣ is odd and in particular positive. A O F A_{\mathrm{OF}} A OF is the disjunction of ¬ B o d d ( x ˉ , z ˉ ) \neg B_{\mathrm{odd}}(\bar{x},\bar{z}) ¬ B odd ( x ˉ , z ˉ ) , ¬ C o d d f a c t o r ( y ˉ , w ˉ ) \neg C_{\mathrm{oddfactor}}(\bar{y},\bar{w}) ¬ C oddfactor ( y ˉ , w ˉ ) and D i n t e r s e c t s ( x ˉ , y ˉ ) D_{\mathrm{intersects}}(\bar{x},\bar{y}) D intersects ( x ˉ , y ˉ ) , saying that either x ˉ \bar{x} x ˉ is not odd, or some vertex has even degree in y ˉ \bar{y} y ˉ , or y ˉ \bar{y} y ˉ intersects δ ( x ˉ ) \delta(\bar{x}) δ ( x ˉ ) . The disjuncts ¬ B o d d \neg B_{\mathrm{odd}} ¬ B odd and D i n t e r s e c t s D_{\mathrm{intersects}} D intersects are as in A P M A_{\mathrm{PM}} A PM above, and
(2′) ¬ C o d d f a c t o r ( y ˉ , w ˉ ) \neg C_{\mathrm{oddfactor}}(\bar{y},\bar{w}) ¬ C oddfactor ( y ˉ , w ˉ ) : the disjunction over v ∈ V v\in V v ∈ V of the standard encoding of ⨁ j : { v , j } ∈ E y v , j = 0 \bigoplus_{j:\{v,j\}\in E} y_{v,j}=0 ⨁ j : { v , j } ∈ E y v , j = 0 as a DNF; for each v v v , extra Boolean variables w ˉ v = { w v , i } i > 1 \bar{w}_{v}=\{w_{v,i}\}_{i>1} w ˉ v = { w v , i } i > 1 represent the partial XORs of the incident edge variables y v , j 1 , y v , j 2 , … y_{v,j_1},y_{v,j_2},\ldots y v , j 1 , y v , j 2 , … (with j 1 , j 2 , … j_1,j_2,\ldots j 1 , j 2 , … an enumeration of the neighbours of v v v ), and the terms have the same shape as those of ¬ B o d d \neg B_{odd} ¬ B o dd with x x x replaced by these incidence variables.
Compared to A P M A_{\mathrm{PM}} A PM , the matching DNF ¬ C m a t c h \neg C_{\mathrm{match}} ¬ C match is dropped entirely and the covering DNF ¬ C V -cover \neg C_{V\text{-cover}} ¬ C V -cover is replaced by a union of ∣ V ∣ |V| ∣ V ∣ parity gadgets structurally identical to ¬ B o d d \neg B_{odd} ¬ B o dd ; the resulting DNF A O F A_{\mathrm{OF}} A OF is a disjunction of parity/intersection gadgets of a single shape, giving its encoding a more uniform block structure than A P M A_{\mathrm{PM}} A PM . In view of the previous subsection, A O F A_{\mathrm{OF}} A OF yields the same Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC lower bounds as A P M A_{\mathrm{PM}} A PM . Both DNFs have polynomially many terms.
Lower bounds for A P M A_{\mathrm{PM}} A PM and A O F A_{\mathrm{OF}} A OF
Both bounds are instances of Recipe I, and it remains only to name its data and check its hypotheses. Take E E E the edge set of K 2 n K_{2n} K 2 n , H \mathcal{H} H the perfect matchings, respectively the odd factors, and F \mathcal{F} F the odd cut inequalities x ˉ ( δ ( U ) ) ≥ 1 \bar{x}(\delta(U)) \ge1 x ˉ ( δ ( U )) ≥ 1 indexed by the odd U ⊆ V U \subseteq V U ⊆ V , so that a f = χ δ ( U ) a_f=\chi_{\delta(U)} a f = χ δ ( U ) and b = 1 b=1 b = 1 , and P = conv { χ H : H ∈ H } P=\operatorname{conv}\{\chi_H:H\in\mathcal{H}\} P = conv { χ H : H ∈ H } is P P M P_{\mathrm{PM}} P PM , respectively P O F P_{\mathrm{OF}} P OF . Members of F \mathcal{F} F are encoded over X : = x ˉ ∪ z ˉ X:=\bar{x}\cup\bar{z} X := x ˉ ∪ z ˉ , with x ˉ \bar{x} x ˉ the indicator of U U U and z ˉ \bar{z} z ˉ the partial XORs, and members of H \mathcal{H} H over Y : = y ˉ Y:=\bar{y} Y := y ˉ , respectively y ˉ ∪ w ˉ \bar{y}\cup\bar{w} y ˉ ∪ w ˉ ; this is the split of the variables that induces the flattening M U x ∣ U y ( χ ( A ) − ϵ ⋅ χ ( 1 ) ) M_{U_x\mid U_y}(\chi(A)-\epsilon\cdot\chi(1)) M U x ∣ U y ( χ ( A ) − ϵ ⋅ χ ( 1 )) used below, with U x : = X U_x:=X U x := X and U y : = Y U_y:=Y U y := Y .
The inequalities of F \mathcal{F} F hold at every χ H \chi_H χ H . No perfect matching lies inside a set of odd size, so a perfect matching M M M has an edge leaving U U U and ∣ δ ( U ) ∩ M ∣ ≥ 1 |\delta(U)\cap M|\ge1 ∣ δ ( U ) ∩ M ∣ ≥ 1 . For an odd factor H H H , the identity ∑ v ∈ U deg H ( v ) = 2 ∣ E H ( U ) ∣ + ∣ δ ( U ) ∩ H ∣ \sum_{v\in U}\deg_H(v)=2|E_H(U)|+|\delta(U)\cap H| ∑ v ∈ U deg H ( v ) = 2∣ E H ( U ) ∣ + ∣ δ ( U ) ∩ H ∣ exhibits ∣ δ ( U ) ∩ H ∣ |\delta(U)\cap H| ∣ δ ( U ) ∩ H ∣ as a sum of ∣ U ∣ |U| ∣ U ∣ odd numbers minus an even number; as ∣ U ∣ |U| ∣ U ∣ is odd, ∣ δ ( U ) ∩ H ∣ |\delta(U)\cap H| ∣ δ ( U ) ∩ H ∣ is odd and in particular positive.
Condition (I.1). For an edge e = { i , j } e=\{i,j\} e = { i , j } put ρ e , 1 : = x i ∧ ¬ x j \rho_{e,1}:=x_i\wedge\neg x_j ρ e , 1 := x i ∧ ¬ x j and ρ e , 2 : = x j ∧ ¬ x i \rho_{e,2}:=x_j\wedge\neg x_i ρ e , 2 := x j ∧ ¬ x i , so c = 2 c=2 c = 2 . They are contradictory, and their disjunction holds at the encoding of U U U exactly when e e e has one endpoint in U U U and one outside, that is exactly when e ∈ δ ( U ) = supp ( a f ) e\in\delta(U)=\operatorname{supp}(a_f) e ∈ δ ( U ) = supp ( a f ) . Take γ e : = y i , j \gamma_e:=y_{i,j} γ e := y i , j , which holds exactly when e e e belongs to the encoded H H H . The resulting witness terms ρ e , j ∧ γ e \rho_{e,j}\wedge\gamma_e ρ e , j ∧ γ e are precisely the terms x i ∧ y i , j ∧ ¬ x j x_i\wedge y_{i,j}\wedge\neg x_j x i ∧ y i , j ∧ ¬ x j of D i n t e r s e c t s D_{\mathrm{intersects}} D intersects .
Condition (I.2). Invalid rows are recognised by I R : = ¬ B o d d I_R:=\neg B_{odd} I R := ¬ B o dd : its terms of the first kind are satisfied exactly when z ˉ \bar{z} z ˉ misrepresents some partial XOR of x ˉ \bar{x} x ˉ , and ¬ z x 1 ⊕ ⋯ ⊕ x 2 n \neg z_{x_1\oplus\cdots\oplus x_{2n}} ¬ z x 1 ⊕ ⋯ ⊕ x 2 n exactly when the recorded total parity is even, so ¬ B o d d \neg B_{odd} ¬ B o dd holds precisely at those ( x ˉ , z ˉ ) (\bar{x},\bar{z}) ( x ˉ , z ˉ ) that do not encode an odd U U U with its partial XORs. Invalid columns are recognised by I C : = ¬ C m a t c h ∨ ¬ C V -cover I_C:=\neg C_{\mathrm{match}}\vee\neg C_{V\text{-cover}} I C := ¬ C match ∨ ¬ C V -cover , since y ˉ \bar{y} y ˉ fails to encode a perfect matching exactly when two of its edges share a vertex or some vertex is left uncovered; for A O F A_{\mathrm{OF}} A OF by I C : = ¬ C o d d f a c t o r I_C:=\neg C_{\mathrm{oddfactor}} I C := ¬ C oddfactor , whose parity gadgets are satisfied exactly when w ˉ \bar{w} w ˉ misrepresents a partial XOR at some vertex or some vertex has even degree in y ˉ \bar{y} y ˉ . Both have polynomially many terms.
The DNF that Recipe I builds from this data — the disjunction of the two invalidity DNFs with the witness terms — is therefore exactly A P M A_{\mathrm{PM}} A PM , respectively A O F A_{\mathrm{OF}} A OF , of Section 4.2. So both are tautologies, and their valid slice is the ground set overlap count M ( U , H ) = ∣ δ ( U ) ∩ H ∣ M(U,H)=|\delta(U)\cap H| M ( U , H ) = ∣ δ ( U ) ∩ H ∣ : the valid rows biject with the odd subsets of V V V and the valid columns with the perfect matchings, respectively the odd factors, which are exactly the row and column indices of M P M M_{\mathrm{PM}} M PM , respectively M O F M_{\mathrm{OF}} M OF . The perturbed valid slice is M P M − ϵ ⋅ J M_{\mathrm{PM}}-\epsilon\cdot J M PM − ϵ ⋅ J , respectively M O F − ϵ ⋅ J M_{\mathrm{OF}}-\epsilon\cdot J M OF − ϵ ⋅ J . Recipe I now gives the following, connecting proof size directly to the nonnegative rank of the odd cut slack matrices of P P M P_{\mathrm{PM}} P PM and P O F P_{\mathrm{OF}} P OF .
Theorem 4.2. Let Φ \Phi Φ be a monotone formula, of mixed-product depth at most t t t relative to the partition ( U x , U y ) (U_x,U_y) ( U x , U y ) above, computing χ ( A P M ) − ϵ ⋅ χ ( 1 ) \chi(A_{\mathrm{PM}})-\epsilon\cdot\chi(1) χ ( A PM ) − ϵ ⋅ χ ( 1 ) or χ ( A O F ) − ϵ ⋅ χ ( 1 ) \chi(A_{\mathrm{OF}})-\epsilon\cdot\chi(1) χ ( A OF ) − ϵ ⋅ χ ( 1 ) . Then its number of leaves is at least
rk + ( M P M − ϵ ⋅ J ) 1 / 2 t , \operatorname{rk}_{+}\left(M_{\mathrm{PM}}-\epsilon\cdot J\right)^{1/2^t}, rk + ( M PM − ϵ ⋅ J ) 1/ 2 t ,
where M P M − ϵ ⋅ J M_{\mathrm{PM}}-\epsilon\cdot J M PM − ϵ ⋅ J is the slack matrix, at the odd cut inequalities F \mathcal{F} F , of the pair ( P P M , P P M , 1 − ϵ F ) (P_{\mathrm{PM}},P^{\mathcal{F}}_{\mathrm{PM},1-\epsilon}) ( P PM , P PM , 1 − ϵ F ) , and correspondingly for A O F A_{\mathrm{OF}} A OF .
Proof. The valid slice identified above is M P M − ϵ ⋅ J M_{\mathrm{PM}}-\epsilon\cdot J M PM − ϵ ⋅ J , respectively M O F − ϵ ⋅ J M_{\mathrm{OF}}-\epsilon\cdot J M OF − ϵ ⋅ J . Write F F F for the polynomial computed by Φ \Phi Φ and L L L for its leaf count. Monotonicity under submatrices and Theorem 2.11 give
rk + ( M P M − ϵ ⋅ J ) ≤ rk + ( Flat ( F ) ) ≤ L 2 t . \operatorname{rk}_{+}(M_{\mathrm{PM}}-\epsilon\cdot J)\leq\operatorname{rk}_{+}(\operatorname{Flat}(F))\leq L^{2^{t}}. rk + ( M PM − ϵ ⋅ J ) ≤ rk + ( Flat ( F )) ≤ L 2 t .
For the odd factor target, the same argument applies because the perfect matching matrix is a column submatrix of the odd factor matrix. Taking the 2 − t 2^{-t} 2 − t power proves the claim. Since b = 1 b=1 b = 1 , the slack-matrix interpretation is the relaxation obtained by loosening each odd cut inequality by 1 − ϵ 1-\epsilon 1 − ϵ .
Corollary 4.3. For constant ϵ ∈ ( 0 , 1 ) \epsilon\in(0,1) ϵ ∈ ( 0 , 1 ) the leaf count is at least 2 Ω ( n / 2 t ) 2^{\Omega(n/2^{t})} 2 Ω ( n / 2 t ) . For ϵ ≥ c 0 log log n / log n \epsilon\geq c_{0}\log\log n/\log n ϵ ≥ c 0 log log n / log n with c 0 > 8 c_{0}>8 c 0 > 8 fixed it is at least 2 n 1 − 8 / c 0 + o ( 1 ) / 2 t 2^{n^{1-8/c_{0}+o(1)}/2^{t}} 2 n 1 − 8/ c 0 + o ( 1 ) / 2 t , and in particular superpolynomial for every constant t t t . At t = 0 t=0 t = 0 , both Σ Π Σ - A M C ϵ \Sigma\Pi\Sigma\text{-}\mathrm{AMC}^{\epsilon} ΣΠΣ - AMC ϵ and Σ Π Σ - A M C b ϵ \Sigma\Pi\Sigma\text{-}\mathrm{AMC}^{\epsilon}_{b} ΣΠΣ - AMC b ϵ satisfy the exponential bound in the constant-shift regime and the second bound in the vanishing-shift regime.
Proof. Insert rk + ( M P M − ϵ ⋅ J ) = 2 Ω ( n ) \operatorname{rk}_{+}(M_{\mathrm{PM}}-\epsilon\cdot J)=2^{\Omega(n)} rk + ( M PM − ϵ ⋅ J ) = 2 Ω ( n ) from (48) for constant ϵ \epsilon ϵ ; in the vanishing regime (57) gives log rk + ( M P M − ϵ ⋅ J ) ≥ n 1 − 8 / c 0 + o ( 1 ) \log\operatorname{rk}_{+}(M_{\mathrm{PM}}-\epsilon\cdot J)\geq n^{1-8/c_{0}+o(1)} log rk + ( M PM − ϵ ⋅ J ) ≥ n 1 − 8/ c 0 + o ( 1 ) , and raising the rank to the power 2 − t 2^{-t} 2 − t divides this logarithm by 2 t 2^{t} 2 t . The case t = 0 t=0 t = 0 is Remark 2.12.
By Proposition 2.13 the restriction to formulas cannot be removed as stated. Braun and Pokutta state their per-chunk constant only as a positive constant depending on k k k and on their perturbation parameter. Since the threshold ϵ ≥ c 0 log log n / log n \epsilon\geq c_{0}\log\log n/\log n ϵ ≥ c 0 log log n / log n , c 0 > 8 c_{0}>8 c 0 > 8 , rests on uniformity in k k k , Appendix A unpacks their proof, makes the dependence on both parameters explicit, and proves the stated sufficient threshold.
An explicit DNF hard at inverse-polynomial shift
Sections 4.1–4.3 took the DNF from a combinatorial polytope and drew the rank bound from its approximate extension complexity. Here we change the DNF family instead, and obtain a much smaller admissible shift: an explicit polynomial-size tautological DNF whose perturbed flattening has superpolynomial nonnegative rank for every inverse-polynomial ϵ ≥ K − γ \epsilon\geq K^{-\gamma} ϵ ≥ K − γ with γ < 1 / 4 \gamma<1/4 γ < 1/4 fixed. The witness conjunctions of Proposition 2.14 record a violated constraint of a planted Tseitin copy on a constant-degree expander, and the rank lower bound comes from a Sherali–Adams integrality gap rather than from the extension complexity of a fixed body.
The skeleton and the Tseitin base are those already used in Section 3.5: plant many copies of a base instance whose degree-d d d Sherali–Adams relaxation is fooled, so that the copy to be tested is chosen only after the purported decomposition, then test that decomposition with the local expectation functional witnessing the failure. One design choice differs, and everything else follows from it. There each base variable carried a private table, so every address tuple already held a copy and no work was needed to find one; here the copies are planted into a shared pool of N N N variables, and a copy the factors cannot see must first be located. That costs an imported random-restriction lemma, whose N − 1 / 4 N^{-1/4} N − 1/4 is what caps the admissible shift at K − γ K^{-\gamma} K − γ with γ < 1 / 4 \gamma<1/4 γ < 1/4 rather than the 2 − Ω ( N ) 2^{-\Omega(N)} 2 − Ω ( N ) reached there. What it buys is the converse: the argument runs on marginals of arbitrary nonnegative factors rather than on products over private bits, so it bounds rk + \operatorname{rk}_{+} rk + and survives mixed-product depth above zero, neither of which the lift does.
The mechanism is that of Chan, Lee, Raghavendra and Steurer [5 ] ; what is added here is the dependence of the admissible shift on all parameters, which is what fixes the range of ϵ \epsilon ϵ . The only input taken from their work is a random-restriction lemma, quoted below in the form we use. The matrix. Fix an unsatisfiable constraint satisfaction instance I 0 I_0 I 0 on m m m variables, writing I 0 ( x ) I_0(x) I 0 ( x ) for the fraction of its constraints that x x x satisfies, and let I ϕ I_\phi I ϕ be the copy of I 0 I_0 I 0 planted along an injection ϕ : [ m ] ↪ [ N ] \phi:[m]\hookrightarrow[N] ϕ : [ m ] ↪ [ N ] — one copy for each injection. The matrix whose nonnegative rank is bounded is
A ϕ , x η = 1 − η − I ϕ ( x ) , ϕ an injection , x ∈ { − 1 , 1 } N , A^\eta_{\phi,x}=1-\eta-I_\phi(x),\qquad\phi\text{ an injection},\quad x\in\{-1,1\}^{N}, A ϕ , x η = 1 − η − I ϕ ( x ) , ϕ an injection , x ∈ { − 1 , 1 } N ,
the slack at the assignment x x x of the valid inequality I ϕ ( x ) ≤ 1 − η I_\phi(x)\le1-\eta I ϕ ( x ) ≤ 1 − η . It is nonnegative exactly because planting does not change the optimum, so that every copy has opt ( I ϕ ) = opt ( I 0 ) ≤ 1 − η \operatorname{opt}(I_\phi)=\operatorname{opt}(I_0)\le1-\eta opt ( I ϕ ) = opt ( I 0 ) ≤ 1 − η ; and by the factorization theorem a nonnegative factorization of it of rank T T T is a linear program with T T T inequalities certifying I ϕ ≤ 1 − η I_\phi\le1-\eta I ϕ ≤ 1 − η for all planted copies simultaneously. Proposition 4.4 says that no such program of size N d / 2 N^{d/2} N d /2 exists once I 0 I_0 I 0 fools the degree-d d d Sherali–Adams relaxation, and § 4.4.2 realises A η A^\eta A η as the perturbed valid slice of a DNF.
Read geometrically, A η A^\eta A η is again a slack matrix of a nested pair of convex bodies, as in § 4.3 and Appendix § 7 . Lift an assignment x x x to the vector v ( x ) v(x) v ( x ) whose coordinates are the indicators 1 [ x ∣ T = σ ] \mathbf{1}[x|_T=\sigma] 1 [ x ∣ T = σ ] , one for each k k k -subset T ⊆ [ N ] T\subseteq[N] T ⊆ [ N ] and each pattern σ ∈ { − 1 , 1 } T \sigma\in\{-1,1\}^{T} σ ∈ { − 1 , 1 } T . Every constraint of I ϕ I_\phi I ϕ is a sum of such indicators on its own scope, so I ϕ I_\phi I ϕ is linear in this lift: I ϕ ( x ) = ⟨ a ϕ , v ( x ) ⟩ I_\phi(x)=\langle a_\phi,v(x)\rangle I ϕ ( x ) = ⟨ a ϕ , v ( x )⟩ with a ϕ ≥ 0 a_\phi\ge0 a ϕ ≥ 0 . The inner body is the convex hull of the v ( x ) v(x) v ( x ) and the outer one is cut out by the inequalities ⟨ a ϕ , y ⟩ ≤ 1 − η \langle a_\phi,y\rangle\le1-\eta ⟨ a ϕ , y ⟩ ≤ 1 − η , one per planted copy, the two being nested precisely because opt ( I 0 ) < 1 − η \operatorname{opt}(I_0)<1-\eta opt ( I 0 ) < 1 − η . Nothing below appeals to this reading, and in particular no polytope is named: the rows here are the instances of a family, not the facets of one fixed body, and it is the family that the argument consumes.
Why the witness cannot be fixed in advance. The matrices produced by Proposition 2.14 are entrywise strictly positive, so by § 2.4 a usable witness must be signed; and a signed witness fixed in advance would have to defeat every admissible rank-one matrix at once.
The argument below does not fix one. It supplies instead a whole family of tests, one for each m m m -element set S ⊆ [ N ] S\subseteq[N] S ⊆ [ N ] — each useless on its own, certifying nothing against a general nonnegative function and only something against functions that are nearly juntas on S S S — and a counting principle, Lemma B.1, guaranteeing that any family of at most N d / 2 N^{d/2} N d /2 nonnegative functions of mean one and bounded sup-norm has a single S S S on whose coordinates all of their marginals are nearly juntas. The assumed factorization then selects the test that defeats it. Where the lift of Section 3.5 located a copy by averaging over selector assignments, here it is this counting principle that does so.
The planted family and its rank
Normalized Max-k k k -CSP instances, opt \operatorname{opt} opt , the Fourier notation, the degree-d d d local expectation functionals, the d d d -juntas and S A d \mathrm{SA}_d SA d are those of Section 3.5, and are used here without further comment. One object is new. For an injection ϕ : [ m ] ↪ [ N ] \phi:[m]\hookrightarrow[N] ϕ : [ m ] ↪ [ N ] , the planted copy I ϕ I_\phi I ϕ of an instance I I I on m m m variables is the instance on N N N variables obtained by renaming the variable u u u as ϕ ( u ) \phi(u) ϕ ( u ) and leaving the other N − m N-m N − m variables unconstrained, so that I ϕ ( x ) = I ( x ϕ ( 1 ) , … , x ϕ ( m ) ) I_\phi(x)=I(x_{\phi(1)},\ldots,x_{\phi(m)}) I ϕ ( x ) = I ( x ϕ ( 1 ) , … , x ϕ ( m ) ) for x ∈ { − 1 , 1 } N x\in\{-1,1\}^{N} x ∈ { − 1 , 1 } N , and opt ( I ϕ ) = opt ( I ) \operatorname{opt}(I_\phi)=\operatorname{opt}(I) opt ( I ϕ ) = opt ( I ) . Finally we put
Δ ( N , m , d ) : = ( m ≤ d ) ( 4 d m log 2 N N − 1 / 4 + N − d / 2 ) . (49) \Delta(N,m,d):=\binom{m}{\le d}\left(4d\sqrt{m\log_2N}\,N^{-1/4}+N^{-d/2}\right). \tag*{(49)} Δ ( N , m , d ) := ( ≤ d m ) ( 4 d m log 2 N N − 1/4 + N − d /2 ) . (49)
Proposition 4.4 (Quantitative planting bound). Let I 0 I_0 I 0 be a normalized Boolean Max-k k k -CSP instance on m m m variables, where 3 ≤ m 3\le m 3 ≤ m and k ≤ d ≤ m k\le d\le m k ≤ d ≤ m , such that
opt ( I 0 ) ≤ s < 1 , S A d ( I 0 ) = 1. \operatorname{opt}(I_0)\le s<1,\qquad\mathrm{SA}_d(I_0)=1. opt ( I 0 ) ≤ s < 1 , SA d ( I 0 ) = 1.
For N ≥ 4 m N \ge4m N ≥ 4 m let Φ \Phi Φ be the set of injections [ m ] ↪ [ N ] [m] \hookrightarrow[N] [ m ] ↪ [ N ] , and for ϕ ∈ Φ \phi\in\Phi ϕ ∈ Φ let I ϕ I_\phi I ϕ be the copy of I 0 I_0 I 0 planted according to ϕ \phi ϕ . For 0 < η < 1 − s 0 < \eta< 1-s 0 < η < 1 − s define the nonnegative matrix
A ϕ , x η = 1 − η − I ϕ ( x ) , ϕ ∈ Φ , x ∈ { − 1 , 1 } N . A^\eta_{\phi,x}=1-\eta-I_\phi(x), \qquad\phi\in\Phi,\quad x\in\{-1,1\}^{N}. A ϕ , x η = 1 − η − I ϕ ( x ) , ϕ ∈ Φ , x ∈ { − 1 , 1 } N .
If η > Δ ( N , m , d ) \eta> \Delta(N,m,d) η > Δ ( N , m , d ) , with Δ \Delta Δ as in (49) , then rk + ( A η ) > N d / 2 \operatorname{rk}_{+}(A^\eta)>N^{d/2} rk + ( A η ) > N d /2 .
The mechanism is this. Suppose A η A^\eta A η had a nonnegative factorization A ϕ , x η = ∑ i ≤ R λ i ( ϕ ) q i ( x ) A^\eta_{\phi,x}=\sum_{i\le R}\lambda_i(\phi)q_i(x) A ϕ , x η = ∑ i ≤ R λ i ( ϕ ) q i ( x ) with R ≤ N d / 2 R\le N^{d/2} R ≤ N d /2 , each column factor normalized to mean one and its mean absorbed into λ i \lambda_i λ i , so that the q i q_i q i are densities. The proof localizes its factors to a set of m m m coordinates, evaluates the identity at the row ϕ \phi ϕ that plants I 0 I_0 I 0 on exactly those coordinates, and applies the functional witnessing SA d ( I 0 ) = 1 \operatorname{SA}_d(I_0)=1 SA d ( I 0 ) = 1 . The left side is then the function 1 − η − I 0 1-\eta-I_0 1 − η − I 0 of those m m m coordinates, on which the functional returns ( 1 − η ) − E ~ I 0 = − η (1-\eta)-\widetilde{\mathbb{E}}I_0=-\eta ( 1 − η ) − E I 0 = − η : the failure of the degree-d d d Sherali–Adams relaxation to certify I 0 ≤ 1 − η I_0\le1-\eta I 0 ≤ 1 − η . On the right it can be made only slightly negative, and the two terms of Δ ( N , m , d ) \Delta(N,m,d) Δ ( N , m , d ) are the two reasons why.
The functional is nonnegative on d d d -juntas, so a factor whose marginal on those m m m coordinates is a d d d -junta costs nothing at all, and the whole question is how far the factors are from that. Call a factor flat if ∥ q i ∥ ∞ ≤ N d \|q_i\|_\infty\le N^d ∥ q i ∥ ∞ ≤ N d . The lemma we import from [5 ] produces a single m m m -set on which the marginal of every flat factor splits as a d d d -junta plus an error whose Fourier coefficients of degree at most d d d are all bounded by 4 d m log 2 N N − 1 / 4 4d\sqrt{m\log_2N}\,N^{-1/4} 4 d m log 2 N N − 1/4 . Its higher coefficients are left unbounded, and need not be bounded: the functional reaches the error only through its degree-d d d truncation, which is a combination of the ( m ≤ d ) \binom{m}{\le d} ( ≤ d m ) monomials of degree at most d d d with exactly the coefficients the lemma controls. On each of those monomials the functional has value at most one in absolute value by (44), so the error costs at most ( m ≤ d ) \binom{m}{\le d} ( ≤ d m ) times that bound; and averaging the identity over x x x gives ∑ i λ i ( ϕ ) = E x A ϕ , x η ≤ 1 \sum_i\lambda_i(\phi)=\mathbb{E}_xA^\eta_{\phi,x}\le1 ∑ i λ i ( ϕ ) = E x A ϕ , x η ≤ 1 , the q i q_i q i having mean one and A η A^\eta A η being at most one. So this is the first term of Δ ( N , m , d ) \Delta(N,m,d) Δ ( N , m , d ) .
The lemma applies only to flat factors, and the rest are held down by their coefficients instead. Each summand lies pointwise below A η ≤ 1 A^\eta\le1 A η ≤ 1 , so λ i ( ϕ ) ∥ q i ∥ ∞ ≤ 1 \lambda_i(\phi)\|q_i\|_\infty\le1 λ i ( ϕ ) ∥ q i ∥ ∞ ≤ 1 and a factor failing flatness has λ i ( ϕ ) < N − d \lambda_i(\phi)<N^{-d} λ i ( ϕ ) < N − d . Each factor enters the identity multiplied by its coefficient, so by linearity such a factor contributes λ i ( ϕ ) \lambda_i(\phi) λ i ( ϕ ) times the functional’s value on its marginal, hence at most N − d N^{-d} N − d times that value; and ( m ≤ d ) \binom{m}{\le d} ( ≤ d m ) bounds it for every factor. It comes from the same expansion: what the functional is applied to is the factor’s marginal on the chosen m m m -set, which is again a density; a density q q q has ∣ q ^ ( α ) ∣ = ∣ E [ q χ α ] ∣ ≤ E q = 1 |\widehat q(\alpha)|=|\mathbb{E}[q\,\chi^\alpha]|\le\mathbb{E}q=1 ∣ q ( α ) ∣ = ∣ E [ q χ α ] ∣ ≤ E q = 1 for every α \alpha α ; and only the coefficients of degree at most d d d reach the functional, of which there are ( m ≤ d ) \binom{m}{\le d} ( ≤ d m ) . There are at most N d / 2 N^{d/2} N d /2 factors in all, so together these cost at most ( m ≤ d ) N d / 2 N − d \binom{m}{\le d}N^{d/2}N^{-d} ( ≤ d m ) N d /2 N − d — the second term. The right hand side is therefore at least − Δ ( N , m , d ) -\Delta(N,m,d) − Δ ( N , m , d ) , while the left is − η -\eta − η ; so η > Δ ( N , m , d ) \eta>\Delta(N,m,d) η > Δ ( N , m , d ) contradicts the assumed factorization, and none with R ≤ N d / 2 R\le N^{d/2} R ≤ N d /2 exists.
The imported lemma is the only place where the ambient dimension N N N enters, and its N − 1 / 4 N^{-1/4} N − 1/4 , visible in (49) , is what fixes the ceiling 1 / 4 1/4 1/4 in Theorem 4.6 . Appendix § 8 states the lemma and gives the proof.
From rank bounds to hard DNFs
It remains to convert the rank bound into a statement about DNFs. That conversion is the only step of § 4.4 in which the planting plays no part: it is the encoding of Proposition 2.14 applied to the planted family, and nothing more. Proposition 2.14 is also the only thing shared with Proposition 4.1 ; the interface below names no polytope, and its row structure is demanded by the route rather than incidental to it.
Proposition 4.5 (Recipe II: planted Max-k k k -CSP instances). Let I 0 I_0 I 0 be a normalized Boolean Max-k k k -CSP instance on m m m variables with L L L constraints and opt ( I 0 ) < 1 \operatorname{opt}(I_0)<1 opt ( I 0 ) < 1 , and let N ≥ 4 m N\ge4m N ≥ 4 m . There is a tautological DNF A m , N ( Z , X ) A_{m,N}(Z,X) A m , N ( Z , X ) in K : = ( m + 1 ) N K:=(m+1)N K := ( m + 1 ) N variables, constructible from I 0 I_0 I 0 in time polynomial in its output size with at most
m + m ( N 2 ) + ( m 2 ) N + 2 k L N k = O ( L N k + m N 2 ) (50) m+m\binom{N}{2}+\binom{m}{2}N+2^kLN^k=O(LN^k+mN^2) \tag*{(50)} m + m ( 2 N ) + ( 2 m ) N + 2 k L N k = O ( L N k + m N 2 ) (50) terms and maximum term width N N N , every witness term having width 2 k 2k 2 k , whose valid slice in the Z ∣ X Z\mid X Z ∣ X flattening is the violated constraint count
V ϕ , x = # { constraints of I ϕ violated by x } = L ( 1 − I ϕ ( x ) ) , V_{\phi,x}=\#\{\text{constraints of }I_\phi\text{ violated by }x\}=L(1-I_\phi(x)), V ϕ , x = # { constraints of I ϕ violated by x } = L ( 1 − I ϕ ( x )) , the rows ranging over the injections ϕ : [ m ] ↪ [ N ] \phi:[m]\hookrightarrow[N] ϕ : [ m ] ↪ [ N ] and the columns over all assignments x ∈ { − 1 , 1 } N x\in\{-1,1\}^N x ∈ { − 1 , 1 } N . If moreover SA d ( I 0 ) = 1 \operatorname{SA}_d(I_0)=1 SA d ( I 0 ) = 1 for some k ≤ d ≤ m k\le d\le m k ≤ d ≤ m and ϵ \epsilon ϵ satisfies
Δ ( N , m , d ) < ϵ L < 1 − opt ( I 0 ) , (51) \Delta(N,m,d)<\frac{\epsilon}{L}<1-\operatorname{opt}(I_0), \tag*{(51)} Δ ( N , m , d ) < L ϵ < 1 − opt ( I 0 ) , (51) then rk + ( V − ϵ ⋅ J ) > N d / 2 \operatorname{rk}_+(V-\epsilon\cdot J)>N^{d/2} rk + ( V − ϵ ⋅ J ) > N d /2 , and every monotone formula of mixed-product depth at most t t t computing χ ( A m , N ) − ϵ ⋅ χ ( 1 ) \chi(A_{m,N})-\epsilon\cdot\chi(1) χ ( A m , N ) − ϵ ⋅ χ ( 1 ) has more than N d / 2 t + 1 N^{d/2^{t+1}} N d / 2 t + 1 leaves.
Proof. Introduce Boolean variables z u , a z_{u,a} z u , a for u ∈ [ m ] u\in[m] u ∈ [ m ] , a ∈ [ N ] a\in[N] a ∈ [ N ] , and x a x_a x a for a ∈ [ N ] a\in[N] a ∈ [ N ] ; a valid Z Z Z sets exactly one z u , a = 1 z_{u,a}=1 z u , a = 1 for each u u u , these locations being distinct, and so encodes an injection ϕ : [ m ] ↪ [ N ] \phi:[m]\hookrightarrow[N] ϕ : [ m ] ↪ [ N ] by ϕ ( u ) : = a \phi(u):=a ϕ ( u ) := a . The DNF A m , N A_{m,N} A m , N is:
Invalid-injection terms. One term ⋀ a = 1 N ¬ z u , a \bigwedge_{a=1}^N\neg z_{u,a} ⋀ a = 1 N ¬ z u , a for every u u u (“u u u has no location”), one term z u , a ∧ z u , b z_{u,a}\wedge z_{u,b} z u , a ∧ z u , b for every u u u and a < b a<b a < b (“u u u has two locations”), and one term z u , a ∧ z v , a z_{u,a}\wedge z_{v,a} z u , a ∧ z v , a for every u < v u<v u < v and every a a a (“two base variables collide”).
Violated-constraint terms. For every constraint of I 0 I_0 I 0 , on base variables u 1 , … , u k u_1,\ldots,u_k u 1 , … , u k say, every tuple of distinct locations a 1 , … , a k ∈ [ N ] a_1,\ldots,a_k\in[N] a 1 , … , a k ∈ [ N ] , and every one of the at most 2 k 2^k 2 k local assignments violating that constraint, the term fixing z u 1 , a 1 = ⋯ = z u k , a k = 1 z_{u_1,a_1}=\cdots=z_{u_k,a_k}=1 z u 1 , a 1 = ⋯ = z u k , a k = 1 together with the corresponding literals of x a 1 , … , x a k x_{a_1},\ldots,x_{a_k} x a 1 , … , x a k .
This is an instance of Proposition 2.14 with valid rows R R R the Z Z Z encoding injections, valid columns C C C all of { 0 , 1 } N \{0,1\}^N { 0 , 1 } N , and witness conjunctions the violated-constraint terms: on a valid pair ( ϕ , x ) (\phi,x) ( ϕ , x ) all invalidity terms vanish, and for each constraint of I ϕ I_\phi I ϕ violated by x x x exactly one of its local terms is satisfied, while no local term of a constraint satisfied by x x x is. So A m , N A_{m,N} A m , N is a tautology as soon as every planted copy is unsatisfiable, which holds because opt ( I ϕ ) = opt ( I 0 ) < 1 \operatorname{opt}(I_\phi)=\operatorname{opt}(I_0)<1 opt ( I ϕ ) = opt ( I 0 ) < 1 , the valid slice is V V V as displayed, and the term count and widths are as stated.
Assume now (51) and put η : = ϵ / L \eta:=\epsilon/L η := ϵ / L . Entrywise
( V − ϵ ⋅ J ) ϕ , x = L ( 1 − η − I ϕ ( x ) ) , (52) (V-\epsilon\cdot J)_{\phi,x}=L(1-\eta-I_\phi(x)), \tag*{(52)} ( V − ϵ ⋅ J ) ϕ , x = L ( 1 − η − I ϕ ( x )) , (52)
so V − ϵ ⋅ J V-\epsilon\cdot J V − ϵ ⋅ J is a positive multiple of the matrix A η A^\eta A η of Proposition 4.4: the right hand inequality of (51) supplies its hypothesis 0 < η < 1 − s 0<\eta<1-s 0 < η < 1 − s with s = opt ( I 0 ) s=\operatorname{opt}(I_0) s = opt ( I 0 ) , and the left hand inequality is its hypothesis η > Δ ( N , m , d ) \eta>\Delta(N,m,d) η > Δ ( N , m , d ) . Hence rk + ( V − ϵ ⋅ J ) > N d / 2 \operatorname{rk}_+(V-\epsilon\cdot J)>N^{d/2} rk + ( V − ϵ ⋅ J ) > N d /2 , positive scaling not affecting rk + \operatorname{rk}_+ rk + . The leaf bound is Corollary 2.15. □ \square □
For the base instance we reuse Tseitin from Corollary 3.37.
Theorem 4.6 (Explicit inverse-polynomial-shift DNF). For every fixed 0 < γ 0 < 1 / 4 0 < \gamma_0 < 1/4 0 < γ 0 < 1/4 , there is an explicit infinite family of tautological DNFs A K A_K A K on K K K variables, with K O ( 1 ) K^{O(1)} K O ( 1 ) terms and constructible in deterministic polynomial time in the output size, such that for all sufficiently large members of the family and every
K − γ 0 ≤ ϵ < 1 , K^{-\gamma_0} \le\epsilon< 1, K − γ 0 ≤ ϵ < 1 , the flattening of χ ( A K ) − ϵ ⋅ χ ( 1 ) \chi(A_K) - \epsilon\cdot\chi(1) χ ( A K ) − ϵ ⋅ χ ( 1 ) has nonnegative rank
K Ω ( log K ) . K^{\Omega(\log K)}. K Ω ( l o g K ) . Consequently, by Theorem 2.11, every monotone formula of mixed-product depth at most t t t computing this AMC target has at least
exp ( Ω ( ( log K ) 2 2 t ) ) \exp\left(\Omega\left(\frac{(\log K)^2}{2^t}\right)\right) exp ( Ω ( 2 t ( log K ) 2 ) ) leaves; in particular the bound is superpolynomial for every constant t t t . The family and the implicit constants may depend on γ 0 \gamma_0 γ 0 .
Proof. Fix 0 < γ 0 < 1 / 4 0 < \gamma_0 < 1/4 0 < γ 0 < 1/4 . We instantiate Proposition 4.5 with the Tseitin parity constraints on the explicit connected d G d_G d G -regular expanders of Corollary 3.37, with the same odd charges. Let I 0 I_0 I 0 be the normalized instance, with n : = ∣ V ( G ) ∣ n := |V(G)| n := ∣ V ( G ) ∣ , m : = ∣ E ( G ) ∣ = d G n / 2 m := |E(G)| = d_G n/2 m := ∣ E ( G ) ∣ = d G n /2 , L : = n L := n L := n and arity k : = d G = O ( 1 ) k := d_G = O(1) k := d G = O ( 1 ) . Then
opt ( I 0 ) = 1 − 1 n and SA d ( I 0 ) = 1 \operatorname{opt}(I_0) = 1 - \frac{1}{n}
\qquad\text{and}\qquad
\operatorname{SA}_d(I_0) = 1 opt ( I 0 ) = 1 − n 1 and SA d ( I 0 ) = 1
for d = ⌊ α m ⌋ d = \lfloor\alpha m \rfloor d = ⌊ α m ⌋ and a constant α > 0 \alpha> 0 α > 0 . The first identity follows from Lemma 3.9: every assignment has at least one defect, and single-defect assignments exist because G G G is connected. The second is the Sherali–Adams gap established in the proof of Corollary 3.37.
Fix a constant C > 4 / ( 1 − 4 γ 0 ) C > 4/(1 - 4\gamma_0) C > 4/ ( 1 − 4 γ 0 ) , set N = ⌈ 2 C m ⌉ N = \lceil2^{Cm}\rceil N = ⌈ 2 C m ⌉ , and let A K : = A m , N A_K := A_{m,N} A K := A m , N be the DNF supplied by Proposition 4.5 , on K = ( m + 1 ) N K = (m+1)N K = ( m + 1 ) N variables. By (50) with k = d G k = d_G k = d G it has O ( L N d G + m N 2 ) = K O ( 1 ) O(LN^{d_G} + mN^2) = K^{O(1)} O ( L N d G + m N 2 ) = K O ( 1 ) terms.
It remains to verify (51) throughout the stated range of ϵ \epsilon ϵ . Its right hand inequality is exactly ϵ < 1 \epsilon< 1 ϵ < 1 , since η = ϵ / L = ϵ / n \eta= \epsilon/L = \epsilon/n η = ϵ / L = ϵ / n and 1 − opt ( I 0 ) = 1 / n 1 - \operatorname{opt}(I_0) = 1/n 1 − opt ( I 0 ) = 1/ n . For the left hand one, ( m d ) ≤ 2 m \binom{m}{d} \le2^m ( d m ) ≤ 2 m and so
Δ ( N , m , d ) ≤ 2 m poly ( m , log N ) N − 1 / 4 + 2 m N − d / 2 . \Delta(N,m,d) \le2^m \operatorname{poly}(m,\log N)N^{-1/4} + 2^mN^{-d/2}. Δ ( N , m , d ) ≤ 2 m poly ( m , log N ) N − 1/4 + 2 m N − d /2 .
Here log 2 N = C m + O ( 1 ) \log_2 N = Cm + O(1) log 2 N = C m + O ( 1 ) , so the first term is poly ( m ) 2 m ( 1 − C / 4 ) \operatorname{poly}(m)\,2^{m(1-C/4)} poly ( m ) 2 m ( 1 − C /4 ) and the second is 2 m − C α m 2 / 2 2^{m-C\alpha m^2/2} 2 m − C α m 2 /2 ; also log K = log N + O ( log m ) = ( 1 + o ( 1 ) ) log N \log K = \log N + O(\log m) = (1 + o(1))\log N log K = log N + O ( log m ) = ( 1 + o ( 1 )) log N , so ϵ / L ≥ K − γ 0 / L = 2 − C γ 0 m ( 1 + o ( 1 ) ) \epsilon/L \ge K^{-\gamma_0}/L = 2^{-C\gamma_0m(1+o(1))} ϵ / L ≥ K − γ 0 / L = 2 − C γ 0 m ( 1 + o ( 1 )) . Since C ( 1 / 4 − γ 0 ) > 1 C(1/4 - \gamma_0) > 1 C ( 1/4 − γ 0 ) > 1 , the inequality ϵ / L > Δ ( N , m , d ) \epsilon/L > \Delta(N,m,d) ϵ / L > Δ ( N , m , d ) holds uniformly for ϵ ≥ K − γ 0 \epsilon\ge K^{-\gamma_0} ϵ ≥ K − γ 0 and all sufficiently large members of the family. Proposition 4.5 therefore gives rk + ( V − ϵ ⋅ J ) > N d / 2 \operatorname{rk}_+(V - \epsilon\cdot J) > N^{d/2} rk + ( V − ϵ ⋅ J ) > N d /2 , and rk + \operatorname{rk}_+ rk + is monotone under submatrices.
Finally, log N = Θ ( m ) \log N = \Theta(m) log N = Θ ( m ) gives log K = Θ ( m ) \log K = \Theta(m) log K = Θ ( m ) , so d = Θ ( m ) = Θ ( log K ) d = \Theta(m) = \Theta(\log K) d = Θ ( m ) = Θ ( log K ) and log N = Θ ( log K ) \log N = \Theta(\log K) log N = Θ ( log K ) ; hence
log rk + ≥ ( d / 2 ) log N = Θ ( ( log K ) 2 ) , \log\operatorname{rk}_+ \ge(d/2)\log N = \Theta((\log K)^2), log rk + ≥ ( d /2 ) log N = Θ (( log K ) 2 ) ,
which is the stated rank K Ω ( log K ) K^{\Omega(\log K)} K Ω ( l o g K ) . After pruning zero gates, Lemma 2.8 permits applying Theorem 2.11 to any monotone formula for the target. Taking R = N d / 2 R = N^{d/2} R = N d /2 in its leaf bound R 1 / 2 t R^{1/2^t} R 1/ 2 t gives the claimed formula lower bound, since log R = Θ ( ( log K ) 2 ) \log R = \Theta((\log K)^2) log R = Θ (( log K ) 2 ) .
Conclusion and open problems
Our lower bounds for arbitrary rank-one summands hold over prescribed ranges of shifts and for restricted classes of certificates, either depth three or bounded mixed-product depth. These are independent restrictions. For the subcube fragment, the Sherali–Adams equivalence also gives lower bounds uniform in ϵ \epsilon ϵ .
A bound uniform in ϵ \epsilon ϵ at depth three. An unconditional Σ Π Σ \Sigma\Pi\Sigma ΣΠΣ -AMC lower bound asks for an explicit tautological DNF family A N A_N A N with
min ϵ > 0 rk + ⊗ ( χ ( A N ) − ϵ ⋅ χ ( 1 ) ) = N ω ( 1 ) , \min_{\epsilon>0}\operatorname{rk}_{+}^{\otimes}\bigl(\chi(A_N)-\epsilon\cdot\chi(1)\bigr)=N^{\omega(1)}, ϵ > 0 min rk + ⊗ ( χ ( A N ) − ϵ ⋅ χ ( 1 ) ) = N ω ( 1 ) ,
the minimum being over all admissible shifts at once. All three routes of this paper stop short of it. The flattening recipes of § 4 need ϵ ≥ K − γ \epsilon\ge K^{-\gamma} ϵ ≥ K − γ with γ < 1 / 4 \gamma< 1/4 γ < 1/4 ; the selector lift of Section 3.5 reaches ϵ ≥ 2 − c N \epsilon\ge2^{-cN} ϵ ≥ 2 − c N for a constant c > 0 c>0 c > 0 and says nothing at ϵ = 2 − N 2 \epsilon=2^{-N^2} ϵ = 2 − N 2 ; and the Tseitin bounds of § 3.3 cannot be pushed to every shift, since Theorem 3.29 gives T S G TS_G T S G proofs of size O ( n 3 ) O(n^3) O ( n 3 ) at ϵ = 2 − Θ ( n log n ) \epsilon=2^{-\Theta(n\log n)} ϵ = 2 − Θ ( n l o g n ) . The lifted family Ind q ( T S G ) \operatorname{Ind}_q(TS_G) Ind q ( T S G ) and the planted family A m , N A_{m,N} A m , N of § 4.4 are candidates; Theorem 3.29 does not apply to either.
Unbounded fan-in depth four. No superpolynomial lower bound is available for A M C ϵ AMC^\epsilon A M C ϵ proofs given by a monotone formula of ordinary depth four with unbounded fan-in, at any ϵ \epsilon ϵ , and the flattening bound does not apply to one, since bracketing such a formula into binary gates can make its mixed-product depth grow with n n n (Figure 1 ). Even a bound at depth four and constant ϵ \epsilon ϵ would be new.
The Tseitin threshold. For a constant-degree expander, Theorem 3.24 gives rk + ⊗ ( T ϵ ) ≥ 2 Ω ( ϵ n ) \operatorname{rk}_{+}^{\otimes}(T^\epsilon)\ge2^{\Omega(\epsilon n)} rk + ⊗ ( T ϵ ) ≥ 2 Ω ( ϵ n ) when ϵ = ω ( log n / n ) \epsilon=\omega(\log n/n) ϵ = ω ( log n / n ) , hence an exponential bound at constant shift and a superpolynomial bound throughout that range. The rank is O ( n 3 ) O(n^3) O ( n 3 ) at ϵ = 2 − Θ ( n log n ) \epsilon=2^{-\Theta(n\log n)} ϵ = 2 − Θ ( n l o g n ) by Theorem 3.29 . Where between the two the transition occurs is open, as is whether the quadratic product rank of Lemma 3.28 is optimal for a weakly biased parity.
Small shifts through the planting route. The ceiling γ < 1 / 4 \gamma<1/4 γ < 1/4 of Theorem 4.6 comes from the sampling scale in Lemma B.1 , not from the encoding; a sharper planting estimate would enlarge the admissible range by the same exponent. Reaching ϵ = K − ω ( 1 ) \epsilon=K^{-\omega(1)} ϵ = K − ω ( 1 ) by this route would need a Sherali–Adams certificate lower bound at superpolynomially small gap.
Flattenings of the lifted families. The selector lift bounds rk + ⊗ \operatorname{rk}_{+}^{\otimes} rk + ⊗ but gives nothing for rk + \operatorname{rk}_{+} rk + of a flattening, and whether Ind q ( T S G ) \operatorname{Ind}_q(TS_G) Ind q ( T S G ) is easy at mixed-product depth zero—as T S G TS_G T S G itself is, by Theorem 3.33 —we do not know. An answer either way would say which of the two features of § 4.4 , the reach in ϵ \epsilon ϵ or the applicability above depth three, is intrinsic to the shared-pool planting.
Which systems does AMC simulate? The simulation established here already holds at depth three, where the bounded-bias fragment is equivalent to Sherali–Adams up to clause width (Corollary 3.4 ). Whether AMC, at depth three or at any depth, p-simulates Cutting Planes, sums-of-squares or Polynomial Calculus is open, as is whether any standard system p-simulates AMC.
The bounded-bias fragment
The equivalence with Sherali–Adams gives lower bounds for Σ Π Σ - A M C b \Sigma\Pi\Sigma\text{-}\mathrm{AMC}_{b} ΣΠΣ - AMC b at every shift (Corollary 3.7). At ϵ = 2 − Θ ( n log n ) \epsilon= 2^{-\Theta(n \log n)} ϵ = 2 − Θ ( n l o g n ) , Σ Π Σ - A M C \Sigma\Pi\Sigma\text{-}\mathrm{AMC} ΣΠΣ - AMC has exponentially smaller proofs of Tseitin (Corollary 3.30). It remains open whether the fragments separate at constant ϵ \epsilon ϵ .
The ϵ \epsilon ϵ threshold in the Braun–Pokutta bound
This appendix justifies the claim of § 4.3 : the bound of Braun and Pokutta [4 ] makes rk + ( M − ϵ ⋅ J ) \operatorname{rk}_{+}(M-\epsilon\cdot J) rk + ( M − ϵ ⋅ J ) at least 2 Ω ( 1 ) 2^{\Omega(1)} 2 Ω ( 1 ) for every ϵ ≥ c 0 log log n / log n \epsilon\ge c_{0}\log\log n/\log n ϵ ≥ c 0 log log n / log n with c 0 > 8 c_{0}>8 c 0 > 8 fixed. We establish this sufficient range without claiming that the constant 8 8 8 is optimal. They state their per-chunk constant and the auxiliary constant B B B only as positive constants depending on k k k and on the perturbation. Their perturbation parameter is called ϵ \epsilon ϵ there and is our δ = 1 − ϵ \delta=1-\epsilon δ = 1 − ϵ below, and their goodness threshold, called δ \delta δ there, is our t t t . Since the threshold computation rests on uniformity in k k k , we unpack their proof [4 Proposition 3.2, Sections 3.3 and 3.4] and make the dependence on the chunk parameter k k k and the perturbation δ \delta δ explicit. Throughout, M M M is the matrix indexed by perfect matchings (columns) and odd cuts δ ( U ) \delta(U) δ ( U ) (rows) of K 2 n K_{2n} K 2 n , with entry ∣ δ ( U ) ∩ M ∣ |\delta(U)\cap M| ∣ δ ( U ) ∩ M ∣ , and J J J is the all-ones matrix; logarithms are base 2.
Reduction to the Braun–Pokutta slack matrix. Write S + δ S^{+\delta} S + δ for the perturbed slack matrix of [4 Section 3] , whose entry for a matching M M M and an odd cut U U U is
S M , U + δ = ∣ δ ( U ) ∩ M ∣ − 1 + δ 2 , 0 ≤ δ < 1. S^{+\delta}_{M,U}=\frac{|\delta(U)\cap M|-1+\delta}{2},\qquad0\le\delta<1. S M , U + δ = 2 ∣ δ ( U ) ∩ M ∣ − 1 + δ , 0 ≤ δ < 1.
Since M − ϵ ⋅ J M-\epsilon\cdot J M − ϵ ⋅ J has entries ∣ δ ( U ) ∩ M ∣ − ϵ = 2 S M , U + ( 1 − ϵ ) |\delta(U)\cap M|-\epsilon=2S^{+(1-\epsilon)}_{M,U} ∣ δ ( U ) ∩ M ∣ − ϵ = 2 S M , U + ( 1 − ϵ ) , we have M − ϵ ⋅ J = 2 ( S + ( 1 − ϵ ) ) T M-\epsilon\cdot J=2(S^{+(1-\epsilon)})^{\mathsf{T}} M − ϵ ⋅ J = 2 ( S + ( 1 − ϵ ) ) T , so the relevant Braun–Pokutta parameter is δ = 1 − ϵ \delta=1-\epsilon δ = 1 − ϵ and rk + ( M − ϵ ⋅ J ) = rk + ( S + ( 1 − ϵ ) ) \operatorname{rk}_{+}(M-\epsilon\cdot J)=\operatorname{rk}_{+}(S^{+(1-\epsilon)}) rk + ( M − ϵ ⋅ J ) = rk + ( S + ( 1 − ϵ ) ) (the factor 2 and transposition are immaterial for nonnegative rank).
The bound and its chunk structure. Their key estimate is [4 Proposition 3.2] : for an odd integer k ≥ 7 k\ge7 k ≥ 7 and 0 ≤ δ < 1 − 4 / k 0\le\delta<1-4/k 0 ≤ δ < 1 − 4/ k , partition the vertices of K 2 n K_{2n} K 2 n minus a reserved 3-matching H H H (6 vertices) into m m m chunks T 1 , … , T m T_{1},\ldots,T_{m} T 1 , … , T m of 2 ( k − 3 ) 2(k-3) 2 ( k − 3 ) vertices each, plus a remainder chunk L L L of fewer than 2 ( k − 3 ) 2(k-3) 2 ( k − 3 ) vertices, so that
m = 2 n − 6 − ∣ L ∣ 2 ( k − 3 ) ≥ n − 3 k − 3 − 1 = Θ ( n / k ) , m=\frac{2n-6-|L|}{2(k-3)}\ge\frac{n-3}{k-3}-1=\Theta(n/k), m = 2 ( k − 3 ) 2 n − 6 − ∣ L ∣ ≥ k − 3 n − 3 − 1 = Θ ( n / k ) ,
and bound the common information C [ ⋅ ] C[\cdot] C [ ⋅ ] of S + δ S^{+\delta} S + δ [4 Definition 2.9] from below chunk by chunk:
log rk + ( S + δ ) ≥ C [ M ; U ∣ T , N , H , E ] ≥ c k , δ ⋅ m , \log\operatorname{rk}_{+}(S^{+\delta})\ge C[M;U\mid T,N,H,E]\ge c_{k,\delta}\cdot m, log rk + ( S + δ ) ≥ C [ M ; U ∣ T , N , H , E ] ≥ c k , δ ⋅ m ,
using log rk + ≥ C [ ⋅ ] \log\operatorname{rk}_{+}\ge C[\cdot] log rk + ≥ C [ ⋅ ] [4 Lemma 2.11] . Under δ = 1 − ϵ \delta=1-\epsilon δ = 1 − ϵ the admissibility constraint δ < 1 − 4 / k \delta<1-4/k δ < 1 − 4/ k reads
ϵ > 4 k . \epsilon>\frac{4}{k}. ϵ > k 4 .
The per-chunk constant, explicitly. Unwrapping the proof of [4 Proposition 3.2] in [4 Section 3.3] gives the following admissible choice of constant:
c k , δ = B k , δ ( 2 log e ) t 2 min { Pr [ N = 0 ∣ E ] α 2 Pr [ U = C ( H ) ∣ E 0 ] , Pr [ N = 1 ∣ E ] / 4 Pr [ δ ( C ) ∩ M = H ∣ E 0 ] } , α = 1 ( 2 k − 7 ) ! ! . (53) c_{k,\delta}=B_{k,\delta}(2\log e)t^{2}\min\left\{\frac{\Pr[N=0\mid E]\alpha^{2}}{\Pr[U=C(H)\mid E_{0}]},\frac{\Pr[N=1\mid E]/4}{\Pr[\delta(C)\cap M=H\mid E_{0}]}\right\},\qquad\alpha=\frac{1}{(2k-7)!!}. \tag*{(53)} c k , δ = B k , δ ( 2 log e ) t 2 min { Pr [ U = C ( H ) ∣ E 0 ] Pr [ N = 0 ∣ E ] α 2 , Pr [ δ ( C ) ∩ M = H ∣ E 0 ] Pr [ N = 1 ∣ E ] /4 } , α = ( 2 k − 7 )!! 1 . (53)
where α \alpha α is the reciprocal of the number of perfect matchings on 2 k − 6 2k-6 2 k − 6 vertices, 0 < t < 1 0 < t < 1 0 < t < 1 is a free parameter of their proof (the goodness threshold of [4 Definition 3.3] , denoted δ \delta δ there), the squares ( t α ) 2 (t\alpha)^2 ( t α ) 2 and ( t / 2 ) 2 (t/2)^2 ( t /2 ) 2 are the gaps fed into Pinsker’s inequality [4 Lemma 2.2] , and C C C , D D D , H H H , N N N , E 0 E_0 E 0 , E E E are the objects of the local (single-chunk) case of their proof on K 2 k K_{2k} K 2 k : C ⊔ D C \sqcup D C ⊔ D is a bipartition of the vertex set into two k k k -sets, H H H a uniformly random 3-matching between C C C and D D D , N N N a fair coin, and E 0 = { U ∈ { C , C ( H ) } , H ⊆ M } E_0 = \{U \in\{C,C(H)\},\ H \subseteq M\} E 0 = { U ∈ { C , C ( H )} , H ⊆ M } , E = E 0 ∧ { U = C ( H ) E = E_0 \land\{U = C(H) E = E 0 ∧ { U = C ( H ) if N = 0 N = 0 N = 0 ; δ ( C ) ∩ M = H \delta(C) \cap M = H δ ( C ) ∩ M = H if N = 1 } N = 1\} N = 1 } . The constant B k , δ B_{k,\delta} B k , δ , a lower bound on the probability of the pair ( M , U ) (M,U) ( M , U ) being “bad” for the seed, is computed in [4 Section 3.4] :
B k , δ = ( 2 − ( 3 k − 4 ) ( 1 − β ) − 6 ( k − 4 ) k ( k − 1 ) ) γ A k β Pr [ E 0 ] . (54) B_{k,\delta}=\left(2-(3k-4)(1-\beta)-\frac{6(k-4)}{k(k-1)}\right)\frac{\gamma A}{k\beta\Pr[E_0]}. \tag*{(54)} B k , δ = ( 2 − ( 3 k − 4 ) ( 1 − β ) − k ( k − 1 ) 6 ( k − 4 ) ) k β Pr [ E 0 ] γ A . (54)
with
β = ( 1 − t 1 + t ) 2 , A = ( ( k − 4 ) ! ! ) 2 , γ = 1 ( k / 2 − 1 + δ ) ( 2 k − 1 ) ! ! 2 2 k − 1 , \beta=\left(\frac{1-t}{1+t}\right)^2,\qquad A=((k-4)!!)^2,\qquad\gamma=\frac{1}{(k/2-1+\delta)(2k-1)!!\,2^{2k-1}}, β = ( 1 + t 1 − t ) 2 , A = (( k − 4 )!! ) 2 , γ = ( k /2 − 1 + δ ) ( 2 k − 1 )!! 2 2 k − 1 1 ,
A A A counting the perfect matchings of K 2 k K_{2k} K 2 k that cross the cut ( C , D ) (C,D) ( C , D ) exactly along a fixed 3-matching, and γ \gamma γ being the normalization of the distribution induced by S + δ S^{+\delta} S + δ on K 2 k K_{2k} K 2 k .
These expressions evaluate as follows. Every perfect matching M ′ ⊇ H M' \supseteq H M ′ ⊇ H meets δ ( C ( H ) ) \delta(C(H)) δ ( C ( H )) in exactly the three edges of H H H , and meets δ ( C ) \delta(C) δ ( C ) in between 3 and k k k edges; the matchings with δ ( C ) ∩ M ′ = H \delta(C) \cap M' = H δ ( C ) ∩ M ′ = H are exactly the A A A matchings counted above. Hence
Pr [ E ∣ N = 0 ] = ( 2 + δ ) γ ( 2 k − 7 ) ! ! , Pr [ E ∣ N = 1 ] = 2 ( 2 + δ ) γ ( ( k − 4 ) ! ! ) 2 , \Pr[E\mid N=0]=(2+\delta)\gamma(2k-7)!!,\qquad\Pr[E\mid N=1]=2(2+\delta)\gamma((k-4)!!)^2, Pr [ E ∣ N = 0 ] = ( 2 + δ ) γ ( 2 k − 7 )!! , Pr [ E ∣ N = 1 ] = 2 ( 2 + δ ) γ (( k − 4 )!! ) 2 ,
2 ( 2 + δ ) γ ( 2 k − 7 ) ! ! ≤ Pr [ E 0 ] ≤ ( k + 1 + 2 δ ) γ ( 2 k − 7 ) ! ! . 2(2+\delta)\gamma(2k-7)!!\leq\Pr[E_0]\leq(k+1+2\delta)\gamma(2k-7)!!. 2 ( 2 + δ ) γ ( 2 k − 7 )!! ≤ Pr [ E 0 ] ≤ ( k + 1 + 2 δ ) γ ( 2 k − 7 )!! .
Consequently Pr [ N = 0 ∣ E ] ≥ 1 / 2 \Pr[N=0\mid E]\geq1/2 Pr [ N = 0 ∣ E ] ≥ 1/2 , the second branch of the minimum in (53) is ≥ 1 / 4 ≫ α 2 \geq1/4\gg\alpha^2 ≥ 1/4 ≫ α 2 , so the minimum is always the first branch, and γ \gamma γ cancels against Pr [ E 0 ] \Pr[E_0] Pr [ E 0 ] . The parenthesized factor in (54) never exceeds 2 and is positive only when 1 − β = O ( 1 / k ) 1-\beta=O(1/k) 1 − β = O ( 1/ k ) , i.e. t = O ( 1 / k ) t=O(1/k) t = O ( 1/ k ) ; the choice t = 1 / ( 12 k ) t=1/(12k) t = 1/ ( 12 k ) makes it ≥ 1 / 7 \geq1/7 ≥ 1/7 (and β ≥ 20 / 21 \beta\geq20/21 β ≥ 20/21 ) for every odd k ≥ 7 k\geq7 k ≥ 7 and every δ ∈ [ 0 , 1 − 4 / k ] \delta\in[0,1-4/k] δ ∈ [ 0 , 1 − 4/ k ] . Fixing this choice of t t t , uniformly in δ ∈ [ 0 , 1 − 4 / k ] \delta\in[0,1-4/k] δ ∈ [ 0 , 1 − 4/ k ] ,
1 1500 k 4 ( ( k − 4 ) ! ! ) 2 ( ( 2 k − 7 ) ! ! ) 3 ≤ c k , δ ≤ ( ( k − 4 ) ! ! ) 2 ( ( 2 k − 7 ) ! ! ) 3 , i.e. c k , δ = 2 − 2 k log k + O ( k ) . (55) \frac{1}{1500k^4}\frac{((k-4)!!)^2}{((2k-7)!!)^3}\leq c_{k,\delta}\leq\frac{((k-4)!!)^2}{((2k-7)!!)^3},\qquad\text{i.e.}\qquad c_{k,\delta}=2^{-2k\log k+O(k)}. \tag*{(55)} 1500 k 4 1 (( 2 k − 7 )!! ) 3 (( k − 4 )!! ) 2 ≤ c k , δ ≤ (( 2 k − 7 )!! ) 3 (( k − 4 )!! ) 2 , i.e. c k , δ = 2 − 2 k l o g k + O ( k ) . (55)
Two features matter for the threshold computation: the dominant term is the squared gap α 2 = ( ( 2 k − 7 ) ! ! ) − 2 \alpha^2=((2k-7)!!)^{-2} α 2 = (( 2 k − 7 )!! ) − 2 contributed by Pinsker’s inequality, and the perturbation δ \delta δ enters this choice of c k , δ c_{k,\delta} c k , δ only through the sufficient restriction δ < 1 − 4 / k \delta<1-4/k δ < 1 − 4/ k and through polynomial-in-k k k factors (the term k / 2 − 1 + δ k/2-1+\delta k /2 − 1 + δ in γ \gamma γ , and Pr [ E 0 ] \Pr[E_0] Pr [ E 0 ] ).
Optimizing the trade-off. For k = o ( n ) k=o(n) k = o ( n ) , combining (55) with m = Θ ( n / k ) m=\Theta(n/k) m = Θ ( n / k ) gives
log rk + ( M − ϵ ⋅ J ) ≥ c k , 1 − ϵ ⋅ m = n k 2 − 2 k log k + O ( k ) , k > 4 / ϵ . (56) \log\operatorname{rk}_{+}(M-\epsilon\cdot J)\geq c_{k,1-\epsilon}\cdot m=\frac{n}{k}2^{-2k\log k+O(k)},\qquad k>4/\epsilon. \tag*{(56)} log rk + ( M − ϵ ⋅ J ) ≥ c k , 1 − ϵ ⋅ m = k n 2 − 2 k l o g k + O ( k ) , k > 4/ ϵ . (56)
Choose the smallest odd k ≥ 7 k\geq7 k ≥ 7 with k > 4 / ϵ k>4/\epsilon k > 4/ ϵ , so k = Θ ( max { 1 , 1 / ϵ } ) k=\Theta(\max\{1,1/\epsilon\}) k = Θ ( max { 1 , 1/ ϵ }) . For constant ϵ ∈ ( 0 , 1 ) \epsilon\in(0,1) ϵ ∈ ( 0 , 1 ) this gives log rk + ( M − ϵ ⋅ J ) = Ω ( n ) \log\operatorname{rk}_{+}(M-\epsilon\cdot J)=\Omega(n) log rk + ( M − ϵ ⋅ J ) = Ω ( n ) . In general put ϵ = c 0 log log n / log n \epsilon=c_0\log\log n/\log n ϵ = c 0 log log n / log n with c 0 c_0 c 0 constant, and let k k k be the smallest odd integer above 4 / ϵ 4/\epsilon 4/ ϵ ; then k = n o ( 1 ) k=n^{o(1)} k = n o ( 1 ) , log k = ( 1 + o ( 1 ) ) log log n \log k=(1+o(1))\log\log n log k = ( 1 + o ( 1 )) log log n and 2 k log k = ( 8 / c 0 + o ( 1 ) ) log n 2k\log k=(8/c_0+o(1))\log n 2 k log k = ( 8/ c 0 + o ( 1 )) log n , so that (56) reads
log rk + ( M − ϵ ⋅ J ) ≥ n 1 − 8 / c 0 + o ( 1 ) . (57) \log\operatorname{rk}_{+}(M-\epsilon\cdot J)\geq n^{1-8/c_0+o(1)}. \tag*{(57)} log rk + ( M − ϵ ⋅ J ) ≥ n 1 − 8/ c 0 + o ( 1 ) . (57)
For c 0 > 8 c_0>8 c 0 > 8 the exponent is positive and the conclusion is rk + ( M − ϵ ⋅ J ) ≥ 2 n Ω ( 1 ) \operatorname{rk}_{+}(M-\epsilon\cdot J)\geq2^{n^{\Omega(1)}} rk + ( M − ϵ ⋅ J ) ≥ 2 n Ω ( 1 ) , far beyond superpolynomial; for c 0 < 8 c_0<8 c 0 < 8 the right hand side is o ( 1 ) o(1) o ( 1 ) and nothing is asserted. Superpolynomiality itself needs only c k , 1 − ϵ ⋅ m = ω ( log n ) c_{k,1-\epsilon}\cdot m=\omega(\log n) c k , 1 − ϵ ⋅ m = ω ( log n ) , so the argument delivers it throughout ϵ ≥ c 0 log log n / log n \epsilon\geq c_0\log\log n/\log n ϵ ≥ c 0 log log n / log n for every fixed c 0 > 8 c_0>8 c 0 > 8 , and a fortiori whenever ϵ = ω ( log log n / log n ) \epsilon=\omega(\log\log n/\log n) ϵ = ω ( log log n / log n ) .
Scope of the constant. The restriction δ < 1 − 4 / k \delta< 1 - 4/k δ < 1 − 4/ k is sufficient for the estimate used above. Before substituting this restriction, [4 Section 3.4] gives the badness lower bound
( 2 + δ − 6 ( k − 1 + δ ) / ( k ( k − 1 ) ) + 3 δ β ) γ A Pr [ E 0 ] . \left(2+\delta-\frac{6(k-1+\delta)/(k(k-1))+3\delta}{\beta}\right)\frac{\gamma A}{\Pr[E_0]}. ( 2 + δ − β 6 ( k − 1 + δ ) / ( k ( k − 1 )) + 3 δ ) Pr [ E 0 ] γ A .
At δ = 1 − ϵ \delta=1-\epsilon δ = 1 − ϵ , its parenthesized factor tends, as β → 1 \beta\to1 β → 1 , to 2 ϵ − 6 ( k − ϵ ) / ( k ( k − 1 ) ) 2\epsilon-6(k-\epsilon)/(k(k-1)) 2 ϵ − 6 ( k − ϵ ) / ( k ( k − 1 )) . This is positive whenever ϵ > 3 k / ( k ( k − 1 ) + 3 ) \epsilon> 3k/(k(k-1)+3) ϵ > 3 k / ( k ( k − 1 ) + 3 ) , which permits smaller chunks than k > 4 / ϵ k > 4/\epsilon k > 4/ ϵ . Thus the sufficient constant 8 8 8 above is not an optimality claim for the full Braun–Pokutta argument. We do not optimize the sharper estimate here.
Proof of the quantitative planting bound
We prove Proposition 4.4 , in the notation fixed in Sections 3.5 and § 4.4.1 . The one estimate we import is the following.
Lemma B.1 (Random restriction, [5 Lem. 3.5] , in the form used here). Let 3 ≤ m 3 \le m 3 ≤ m , let d ≤ m d \le m d ≤ m , let N ≥ 4 m N \ge4m N ≥ 4 m , and put γ : = 4 d m log 2 N N − 1 / 4 \gamma:= 4d\sqrt{m\log_2 N}N^{-1/4} γ := 4 d m log 2 N N − 1/4 . Let Q Q Q be a family of at most N d / 2 N^{d/2} N d /2 densities q : { − 1 , 1 } N → R ≥ 0 q:\{-1,1\}^{N}\to\mathbb{R}_{\ge0} q : { − 1 , 1 } N → R ≥ 0 with ∥ q ∥ ∞ ≤ N d \lVert q\rVert_{\infty}\le N^d ∥ q ∥ ∞ ≤ N d . Then there is a single m m m -element set S ⊆ [ N ] S\subseteq[N] S ⊆ [ N ] such that the marginal on S S S of every q ∈ Q q\in Q q ∈ Q can be written as c + e c+e c + e , where c c c is a d d d -junta on { − 1 , 1 } S \{-1,1\}^{S} { − 1 , 1 } S and the error e : { − 1 , 1 } S → R e:\{-1,1\}^{S}\to\mathbb{R} e : { − 1 , 1 } S → R , which is of no prescribed sign, has all of its low-degree Fourier coefficients small:
∣ e ^ ( α ) ∣ ≤ γ for every α ⊆ S with ∣ α ∣ ≤ d . \lvert\widehat{e}(\alpha)\rvert\le\gamma\qquad\text{for every }\alpha\subseteq S\text{ with }\lvert\alpha\rvert\le d. ∣ e ( α )∣ ≤ γ for every α ⊆ S with ∣ α ∣ ≤ d .
Here a density means a nonnegative function of mean one, and the marginal on S S S of a function q q q on { − 1 , 1 } N \{-1,1\}^{N} { − 1 , 1 } N is ( π S q ) ( y ) : = E [ q ( x ) ∣ x ∣ S = y ] (\pi_Sq)(y):=\mathbb{E}[q(x)\mid x|_S=y] ( π S q ) ( y ) := E [ q ( x ) ∣ x ∣ S = y ] , which is again a density and which fixes every function already depending only on the coordinates in S S S . The hypothesis on an individual density is stated in [5 ] as a lower bound on the entropy of the associated measure; the sup-norm bound above is the sufficient condition through which it is applied, since ∥ q ∥ ∞ ≤ N d \lVert q\rVert_{\infty}\le N^d ∥ q ∥ ∞ ≤ N d forces that entropy to be at least N − d log 2 N N-d\log_2 N N − d log 2 N . The cardinality bound N d / 2 N^{d/2} N d /2 is a hypothesis of the lemma and not an artefact of the application, and within that range the error γ \gamma γ does not depend on the size of the family at all.
Proof of Proposition 4.4. Write γ \gamma γ as in Lemma B.1, take the uniform measure μ \mu μ on X : = { − 1 , 1 } N \mathcal{X}:=\{-1,1\}^{N} X := { − 1 , 1 } N , and abbreviate A : = A η A:=A^{\eta} A := A η , A ϕ : = A ( ϕ , ⋅ ) A_{\phi}:=A(\phi,\cdot) A ϕ := A ( ϕ , ⋅ ) . The matrix is nonnegative because opt ( I ϕ ) = opt ( I 0 ) ≤ s < 1 − η \operatorname{opt}(I_{\phi})=\operatorname{opt}(I_0)\le s<1-\eta opt ( I ϕ ) = opt ( I 0 ) ≤ s < 1 − η , and I ϕ ≥ 0 I_{\phi}\ge0 I ϕ ≥ 0 with η > 0 \eta>0 η > 0 gives A ≤ 1 A\le1 A ≤ 1 pointwise, hence also ∫ A ϕ d μ ≤ 1 \int A_{\phi}\,\mathrm{d}\mu\le1 ∫ A ϕ d μ ≤ 1 . As S A d ( I 0 ) S A_d(I_0) S A d ( I 0 ) is a maximum over degree-d d d local expectation functionals, fix one, E ~ \widetilde{\mathbb{E}} E , attaining it, so E ~ I 0 = 1 \widetilde{\mathbb{E}}I_0=1 E I 0 = 1 ; for a function g g g on an m m m -element coordinate set put
Λ ( g ) : = E ~ ( g ≤ d ) , \Lambda(g):=\widetilde{\mathbb{E}}(g^{\le d}), Λ ( g ) := E ( g ≤ d ) ,
a linear functional defined on all functions there, the truncation being what extends E ~ \widetilde{\mathbb{E}} E past its degree bound. Fix for each m m m -set S ⊆ [ N ] S\subseteq[N] S ⊆ [ N ] a bijection [ m ] → S [m]\to S [ m ] → S and use it to transport E ~ \widetilde{\mathbb{E}} E , and hence Λ \Lambda Λ , to functions on { − 1 , 1 } S \{-1,1\}^{S} { − 1 , 1 } S .
Suppose, for contradiction, that A A A has a nonnegative factorization
A ( ϕ , x ) = ∑ i ≤ R λ i ( ϕ ) q i ( x ) , λ i : Φ → R ≥ 0 , q i : X → R ≥ 0 , (58) A(\phi,x)=\sum_{i\le R}\lambda_i(\phi)q_i(x),\qquad\lambda_i:\Phi\to\mathbb{R}_{\ge0},\qquad q_i:\mathcal{X}\to\mathbb{R}_{\ge0}, \tag*{(58)} A ( ϕ , x ) = i ≤ R ∑ λ i ( ϕ ) q i ( x ) , λ i : Φ → R ≥ 0 , q i : X → R ≥ 0 , (58)
with R ≤ N d / 2 R\le N^{d/2} R ≤ N d /2 . Discard the factors with λ i = 0 \lambda_i=0 λ i = 0 or q i = 0 q_i=0 q i = 0 , which changes nothing and only decreases R R R ; as μ \mu μ has full support every surviving q i q_i q i has ∫ q i d μ > 0 \int q_i\,\mathrm{d}\mu>0 ∫ q i d μ > 0 , so absorbing that integral into λ i \lambda_i λ i makes each q i q_i q i a density. Integrating (58) in x x x and using the pointwise bound on A A A ,
∑ i ≤ R λ i ( ϕ ) = ∫ A ϕ d μ ≤ 1 for every ϕ , (59) \sum_{i\le R}\lambda_i(\phi)=\int A_\phi\,\mathrm{d}\mu\le1 \qquad\text{for every }\phi, \tag*{(59)} i ≤ R ∑ λ i ( ϕ ) = ∫ A ϕ d μ ≤ 1 for every ϕ , (59)
and since all summands of (58) are nonnegative each is bounded pointwise by A A A , so
λ i ( ϕ ) ∥ q i ∥ ∞ = ∥ λ i ( ϕ ) q i ∥ ∞ ≤ ∥ A ϕ ∥ ∞ ≤ 1 for every i , ϕ . (60) \lambda_i(\phi)\lVert q_i\rVert_\infty=\lVert\lambda_i(\phi)q_i\rVert_\infty\le\lVert A_\phi\rVert_\infty\le1 \qquad\text{for every }i,\phi. \tag*{(60)} λ i ( ϕ ) ∥ q i ∥ ∞ = ∥ λ i ( ϕ ) q i ∥ ∞ ≤ ∥ A ϕ ∥ ∞ ≤ 1 for every i , ϕ . (60)
Choosing the coordinate set. Split the factors at the level N d N^d N d into G : = { i : ∥ q i ∥ ∞ ≤ N d } \mathcal{G}:=\{i:\lVert q_i\rVert_\infty\le N^d\} G := { i : ∥ q i ∥ ∞ ≤ N d } and the spiky ones B \mathcal{B} B , for which (60) gives λ i ( ϕ ) < N − d \lambda_i(\phi)<N^{-d} λ i ( ϕ ) < N − d at every row. Now { q i } i ∈ G \{q_i\}_{i\in\mathcal{G}} { q i } i ∈ G is a family of at most R ≤ N d / 2 R\le N^{d/2} R ≤ N d /2 densities of sup-norm at most N d N^d N d , which is exactly the hypothesis of Lemma B.1 at these N , m , d N,m,d N , m , d ; the standing assumption N ≥ 4 m N\ge4m N ≥ 4 m is that lemma’s. It supplies a single m m m -set S ⊆ [ N ] S\subseteq[N] S ⊆ [ N ] and splittings π S q i = c i + e i \pi_Sq_i=c_i+e_i π S q i = c i + e i , i ∈ G i\in\mathcal{G} i ∈ G , with c i c_i c i a d d d -junta on { − 1 , 1 } S \{-1,1\}^S { − 1 , 1 } S and ∣ e ^ i ( α ) ∣ ≤ γ |\widehat e_i(\alpha)|\le\gamma ∣ e i ( α ) ∣ ≤ γ for ∣ α ∣ ≤ d |\alpha|\le d ∣ α ∣ ≤ d .
The hard row on that set. Let ϕ S ∈ Φ \phi_S\in\Phi ϕ S ∈ Φ be the chosen bijection [ m ] → S [m]\to S [ m ] → S , regarded as an injection into [ N ] [N] [ N ] . Every m m m -set is the image of such a row, which is the whole point of indexing the rows by all injections. Since A ϕ S = 1 − η − I ϕ S A_{\phi_S}=1-\eta-I_{\phi_S} A ϕ S = 1 − η − I ϕ S depends only on the coordinates in S S S , marginalizing (58) at ϕ S \phi_S ϕ S gives
1 − η − I 0 = π S A ϕ S = ∑ i ≤ R λ i ( ϕ S ) π S q i . 1-\eta-I_0=\pi_SA_{\phi_S}=\sum_{i\le R}\lambda_i(\phi_S)\pi_Sq_i. 1 − η − I 0 = π S A ϕ S = i ≤ R ∑ λ i ( ϕ S ) π S q i .
under the identification, π S \pi_S π S being linear. Apply Λ \Lambda Λ . On the left, I 0 I_0 I 0 is an average of indicators of k k k -ary constraints and so has Fourier degree at most k ≤ d k\le d k ≤ d ; the function 1 − η − I 0 1-\eta-I_0 1 − η − I 0 therefore equals its own degree-d d d truncation and
Λ ( 1 − η − I 0 ) = ( 1 − η ) E ~ 1 − E ~ I 0 = − η . \Lambda(1-\eta-I_0)=(1-\eta)\widetilde{\mathbb{E}}1-\widetilde{\mathbb{E}}I_0=-\eta. Λ ( 1 − η − I 0 ) = ( 1 − η ) E 1 − E I 0 = − η .
On the right, split the sum at G \mathcal{G} G and B \mathcal{B} B and bound the three pieces. A d d d -junta has Fourier degree at most d d d , so Λ ( c i ) = E ~ ( c i ) ≥ 0 \Lambda(c_i)=\widetilde{\mathbb{E}}(c_i)\ge0 Λ ( c i ) = E ( c i ) ≥ 0 by the defining nonnegativity of a local expectation functional. Expanding Λ ( e i ) = ∑ ∣ α ∣ ≤ d e ^ i ( α ) E ~ x α \Lambda(e_i)=\sum_{|\alpha|\le d}\widehat e_i(\alpha)\widetilde{\mathbb{E}}x^\alpha Λ ( e i ) = ∑ ∣ α ∣ ≤ d e i ( α ) E x α and using (44) together with the bound from the lemma gives ∣ Λ ( e i ) ∣ ≤ ( m ≤ d ) γ |\Lambda(e_i)|\le\binom{m}{\le d}\gamma ∣Λ ( e i ) ∣ ≤ ( ≤ d m ) γ ; the same expansion for a density q q q , whose Fourier coefficients satisfy ∣ q ^ ( α ) ∣ ≤ E ∣ q ∣ = 1 |\widehat q(\alpha)|\le\mathbb{E}|q|=1 ∣ q ( α ) ∣ ≤ E ∣ q ∣ = 1 , gives ∣ Λ ( π S q ) ∣ ≤ ( m ≤ d ) |\Lambda(\pi_Sq)|\le\binom{m}{\le d} ∣Λ ( π S q ) ∣ ≤ ( ≤ d m ) . Hence, using (59) for the middle sum and λ i ( ϕ S ) < N − d \lambda_i(\phi_S)<N^{-d} λ i ( ϕ S ) < N − d with ∣ B ∣ ≤ N d / 2 |\mathcal{B}|\le N^{d/2} ∣ B ∣ ≤ N d /2 for the last,
− η = ∑ i ∈ G λ i ( ϕ S ) Λ ( c i ) + ∑ i ∈ G λ i ( ϕ S ) Λ ( e i ) + ∑ i ∈ B λ i ( ϕ S ) Λ ( π S q i ) -\eta=\sum_{i\in\mathcal{G}}\lambda_i(\phi_S)\Lambda(c_i)+\sum_{i\in\mathcal{G}}\lambda_i(\phi_S)\Lambda(e_i)+\sum_{i\in\mathcal{B}}\lambda_i(\phi_S)\Lambda(\pi_Sq_i) − η = i ∈ G ∑ λ i ( ϕ S ) Λ ( c i ) + i ∈ G ∑ λ i ( ϕ S ) Λ ( e i ) + i ∈ B ∑ λ i ( ϕ S ) Λ ( π S q i )
≥ 0 − ( m ≤ d ) γ − N d / 2 N − d ( m ≤ d ) , \ge0-\binom{m}{\le d}\gamma-N^{d/2}N^{-d}\binom{m}{\le d}, ≥ 0 − ( ≤ d m ) γ − N d /2 N − d ( ≤ d m ) ,
whose right hand side is − Δ ( N , m , d ) -\Delta(N,m,d) − Δ ( N , m , d ) by (49). So η ≤ Δ ( N , m , d ) \eta\le\Delta(N,m,d) η ≤ Δ ( N , m , d ) , contrary to hypothesis.