Introduction
Let F \mathbb{F} F be a field. We assume characteristic zero except where otherwise stated. For an n × n n \times n n × n matrix A = ( a i j ) A = (a_{ij}) A = ( a ij ) over F \mathbb{F} F , the permanent of A A A is
per A = ∑ σ ∈ S n ∏ i = 1 n a i , σ ( i ) , \operatorname{per} A = \sum_{\sigma\in S_n} \prod_{i=1}^{n} a_{i,\sigma(i)}, per A = σ ∈ S n ∑ i = 1 ∏ n a i , σ ( i ) ,
where S n S_n S n is the symmetric group on { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } . The permanent of the empty matrix is one.
Barvinok [1 ] proved that the permanent of a matrix of fixed rank can be computed in polynomial time. Flarup and Lyaudet [2 ] treated matrices of bounded weighted clique-width.
A cut I ∣ I c I \mid I^c I ∣ I c is one partition of the indices applied to both rows and columns. Its width is the sum of the ranks of A [ I , I c ] A[I,I^c] A [ I , I c ] and A [ I c , I ] A[I^c,I] A [ I c , I ] . The rank-width is the smallest possible maximum cut width over binary trees on the indices. In [3 § 4] , Lyaudet asked whether the permanent of a matrix of bounded rank-width be computed in polynomial time?
Theorem 3.1 gives an affirmative answer over any field. For each fixed bound on the rank-width, it computes the permanent from the entries of the matrix alone. The proof finds a tree of bounded width from the cut ranks and evaluates a finite sum by combining coefficient vectors at the nodes. For rational entries the algorithm has polynomial bit complexity. In characteristic p > 0 p > 0 p > 0 , Theorem 3.2 gives the bound p O ( k ) n O ( 1 ) p^{O(k)}n^{O(1)} p O ( k ) n O ( 1 ) , including the construction of a suitable tree. Thus, for fixed p p p , rank-width O ( log ( n + 1 ) ) O(\log(n + 1)) O ( log ( n + 1 )) is also allowed.
Cancellation in prime characteristic has been studied in terms of the rank of the whole matrix. Yu [4 ] proved the bound that per A = 0 \operatorname{per} A = 0 per A = 0 if n > ( p − 1 ) rank A n > (p - 1)\operatorname{rank} A n > ( p − 1 ) rank A . Glynn [5 ] gave another proof and a factorization at the boundary rank. Arvind, Chatterjee, Datta and Mukhopadhyay [6 Corollary 4.7] gave a parameterized algorithm over finite fields for matrices of bounded rank. Here the rank bounds concern submatrices across a tree, and the field may be infinite. The proof reduces the endpoint polynomials modulo the p p p th powers of their variables. For rational Toeplitz diagonal sequences this gives a fixed rational generating function for the permanents.
The ranks of the off-diagonal submatrices at successive cuts are used in the theory of quasiseparable matrices [7 , 8 ] . Bounding these two ranks allows the total rank to grow with the order of the matrix. For example, the matrix ( 2 − ∣ i − j ∣ ) i , j = 1 n (2^{-|i-j|})_{i,j=1}^{n} ( 2 − ∣ i − j ∣ ) i , j = 1 n has full rank, and both ranks at each nontrivial cut in the natural ordering are one.
Borchardt’s identity [9 ] expresses the permanent of a Cauchy matrix as a quotient of two determinants. Han and Krattenthaler [10 ] treated rectangular Cauchy permanents, and Khare and Sahi [11 ] obtained permanent and immanant identities for entrywise power series. These formulas do not directly give the evaluation bound considered here for
A i j = 1 x i − y j + P ( x i , y j ) , A_{ij} = \frac{1}{x_i-y_j}+P(x_i,y_j), A ij = x i − y j 1 + P ( x i , y j ) ,
where the two degrees of P P P are fixed and its coefficients and the nodes may vary with n n n . Theorem 3.3 gives a finite formula with polynomially many determinants, each of order at most 2 n 2n 2 n . Its proof expands a bordered Cauchy permanent according to the blocks and then uses polynomial interpolation. The resulting families need not have bounded rank-width.
For circulants with fixed jumps, Minc [12 , 13 ] gave recurrences for the permanent. Golin, Leung and Wang [14 ] gave a transfer matrix proof and treated jumps that vary linearly with the order. Codenotti and Resta [15 ] expressed the permanent of a three-shift circulant as a linear combination of four determinants and treated a class with four shifts. For further formulas for sparse Toeplitz and banded matrices, see [16 , 17 , 18 ] . Alekseyev and Khomovsky [19 , 20 ] give fixed-band recurrences and their symmetry reductions. The coefficient signs for three-shift circulants were studied in [21 , 22 ] . Rodtes and Anwar [23 ] gave a formula for the permanent of a block matrix in terms of submatrices of its blocks. For a circulant with four arbitrary shifts and one weight per shift, the question of a closed formula that can be evaluated in polynomial time uniformly in the shifts remains open.
Our main results are the following. Theorems 3.1 and 3.2 give the tree formula and its refinement in positive characteristic. Theorem 3.3 treats polynomial perturbations of Cauchy matrices in characteristic zero. These are the two general constructions used for the permanent. For cuts along an ordering, Theorem 3.4 and Corollary 3.5 give a single determinant. The subsequent results treat prescribed cycle weights and hook immanants. Theorems 4.4 , 4.7 , 4.9 and 4.11 collect the formulas for the corresponding matrix classes. Their corollaries give the specializations. In each polynomial-time assertion, the parameters that must be fixed are stated explicitly.
Preliminaries
Write [ n ] = { 1 , … , n } [n]=\{1,\ldots,n\} [ n ] = { 1 , … , n } and N = { 0 , 1 , 2 , … } \mathbb{N}=\{0,1,2,\ldots\} N = { 0 , 1 , 2 , … } . For sets of row and column indices I , J I,J I , J , A [ I , J ] A[I,J] A [ I , J ] is the submatrix in the given order. We use I c I^{c} I c for the complement in the relevant index set, and [ w ] f [w]f [ w ] f for the coefficient of the monomial w w w in f f f . The determinant of the empty matrix is one. Permuting rows or columns preserves the permanent.
For I ⊆ [ n ] I\subseteq[n] I ⊆ [ n ] , use the partition I ∣ I c I\mid I^{c} I ∣ I c for the row labels and for the column labels. The two crossing blocks are A [ I , I c ] A[I,I^{c}] A [ I , I c ] and A [ I c , I ] A[I^{c},I] A [ I c , I ] . In an ordering, we use I i = { 1 , … , i } I_i=\{1,\ldots,i\} I i = { 1 , … , i } . In a rooted binary tree on the indices, each leaf is an index and each node I I I is the set of leaves below it. Write
w A ( I ) = rank A [ I , I c ] + rank A [ I c , I ] , rw F ( A ) = min T max I ∈ T w A ( I ) . (1) w_A(I)=\operatorname{rank} A[I,I^{c}]+\operatorname{rank} A[I^{c},I],\qquad
\operatorname{rw}_{\mathbb{F}}(A)=\min_{T}\max_{I\in T}w_A(I). \tag*{(1)} w A ( I ) = rank A [ I , I c ] + rank A [ I c , I ] , rw F ( A ) = T min I ∈ T max w A ( I ) . (1)
The minimum is over all such trees on [ n ] [n] [ n ] . The ranks are taken over the field F \mathbb{F} F of the entries of A A A . In the terminology of Kanté and Rao [24 Definitions 4.1 and 4.3] , w A ( I ) w_A(I) w A ( I ) is the bi-cut-rank and rw F ( A ) \operatorname{rw}_{\mathbb{F}}(A) rw F ( A ) is the F \mathbb{F} F -bi-rank-width of the directed graph with an arc i → j i\to j i → j of weight a i j a_{ij} a ij whenever i ≠ j i\ne j i = j and a i j ≠ 0 a_{ij}\ne0 a ij = 0 . For a symmetric matrix the crossing blocks are transposes, so this rank-width is twice the width defined using one cut block. In formulas involving cut coordinates, r , s r,s r , s denote upper bounds on the two ranks. A cut block of rank zero has empty coordinates. In the Cauchy formulas, the locally specified r , s r,s r , s instead bound the two polynomial degrees.
For each node choose rank factorizations
A [ I , I c ] = R I F I , A [ I c , I ] T = C I G I . (2) A[I,I^c] = R_I F_I,\qquad A[I^c,I]^{\mathsf{T}} = C_I G_I. \tag*{(2)} A [ I , I c ] = R I F I , A [ I c , I ] T = C I G I . (2)
Here R I , C I R_I,C_I R I , C I have full column rank and rows indexed by I I I . The matrices F I , G I F_I,G_I F I , G I have full row rank and columns indexed by I c I^c I c . Put r I = rank R I r_I = \operatorname{rank} R_I r I = rank R I and s I = rank C I s_I = \operatorname{rank} C_I s I = rank C I . Let x I x_I x I and y I y_I y I be columns of r I r_I r I and s I s_I s I indeterminates, respectively. The row and column coordinates at i , j ∈ I i,j \in I i , j ∈ I are the linear forms
R I ( i , : ) x I , C I ( j , : ) y I . R_I(i,:)x_I,\qquad C_I(j,:)y_I. R I ( i , : ) x I , C I ( j , : ) y I .
The variables at the root are empty. Suppose U , V U,V U , V are the two children of I I I . Choose matrices B U V B_{UV} B U V and B V U B_{VU} B V U satisfying
A [ U , V ] = R U B U V C V T , A [ V , U ] = R V B V U C U T . (3) A[U,V] = R_U B_{UV} C_V^{\mathsf{T}},\qquad A[V,U] = R_V B_{VU} C_U^{\mathsf{T}}. \tag*{(3)} A [ U , V ] = R U B U V C V T , A [ V , U ] = R V B V U C U T . (3)
They exist because the columns and rows of each cross block lie in the respective spaces of the two factors. They are obtained using left inverses of the two coordinate matrices. For J = U , V J=U,V J = U , V define S J , I , T J , I S_{J,I},T_{J,I} S J , I , T J , I by
F J [ : , I c ] = S J , I F I , G J [ : , I c ] = T J , I G I . F_J[:,I^c] = S_{J,I}F_I,\qquad G_J[:,I^c] = T_{J,I}G_I. F J [ : , I c ] = S J , I F I , G J [ : , I c ] = T J , I G I .
The rows of each child factor span the corresponding cut row space. Restriction to the parent complement puts them in the row spaces of F I , G I F_I,G_I F I , G I . Write ρ I \rho_I ρ I for the substitution x J ↦ S J , I x I x_J \mapsto S_{J,I}x_I x J ↦ S J , I x I , y J ↦ T J , I y I y_J \mapsto T_{J,I}y_I y J ↦ T J , I y I , for both children.
The coordinate changes satisfy R J S J , I = R I [ J , : ] R_JS_{J,I} = R_I[J,:] R J S J , I = R I [ J , : ] and C J T J , I = C I [ J , : ] C_JT_{J,I} = C_I[J,:] C J T J , I = C I [ J , : ] . Indeed, factor the corresponding submatrix with columns in I c I^c I c in the two ways and use the full row rank of F I F_I F I or G I G_I G I .
2.1.
For this subsection let char F = p > 0 \operatorname{char}\mathbb{F}=p>0 char F = p > 0 . For nonnegative integers r , s r,s r , s , put
B p , r , s = { x α y β : 0 ≤ α i , β j < p , ∣ α ∣ = ∣ β ∣ } , D p , r , s = ∣ B p , r , s ∣ . (4) \mathcal{B}_{p,r,s}=\{x^\alpha y^\beta:0\leq\alpha_i,\beta_j<p,\ |\alpha|=|\beta|\},\qquad D_{p,r,s}=|\mathcal{B}_{p,r,s}|. \tag*{(4)} B p , r , s = { x α y β : 0 ≤ α i , β j < p , ∣ α ∣ = ∣ β ∣ } , D p , r , s = ∣ B p , r , s ∣. (4)
Empty lists of variables are allowed. Replacing β j \beta_j β j by p − 1 − β j p-1-\beta_j p − 1 − β j in the equality of the two degrees gives
D p , r , s = [ t s ( p − 1 ) ] ( 1 + t + ⋯ + t p − 1 ) r + s ≤ p r + s . (5) D_{p,r,s}=[t^{s(p-1)}](1+t+\cdots+t^{p-1})^{r+s}\leq p^{r+s}. \tag*{(5)} D p , r , s = [ t s ( p − 1 ) ] ( 1 + t + ⋯ + t p − 1 ) r + s ≤ p r + s . (5)
At a tree node I I I , calculations will take place in
R I = F [ x I , y I ] / ( x I , 1 p , … , x I , r I p , y I , 1 p , … , y I , s I p ) . (6) R_I=\mathbb{F}[x_I,y_I]/(x_{I,1}^p,\ldots,x_{I,r_I}^p,y_{I,1}^p,\ldots,y_{I,s_I}^p). \tag*{(6)} R I = F [ x I , y I ] / ( x I , 1 p , … , x I , r I p , y I , 1 p , … , y I , s I p ) . (6)
Each class has a unique representative with every exponent less than p p p . The monomials in B p , r I , s I \mathcal{B}_{p,r_I,s_I} B p , r I , s I are its balanced monomials. For a scalar b b b and two variables define
E b ; x , y = ∑ m = 0 p − 1 b m m ! ∂ x m ∂ y m . (7) E_{b;x,y}=\sum_{m=0}^{p-1}\frac{b^m}{m!}\partial_x^m\partial_y^m. \tag*{(7)} E b ; x , y = m = 0 ∑ p − 1 m ! b m ∂ x m ∂ y m . (7)
All denominators in this formula are nonzero in F \mathbb{F} F . Lemma 5.1 proves that these operations and the coordinate substitutions are well defined in the quotients.
2.2.
In this and the next subsection the field has characteristic zero, n ≥ 1 n\geq1 n ≥ 1 , and r , s ≥ 1 r,s\geq1 r , s ≥ 1 . Let the row nodes x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n be distinct, let the column nodes y 1 , … , y n y_1,\ldots,y_n y 1 , … , y n be distinct, and assume x i ≠ y j x_i\ne y_j x i = y j . Write C = ( ( x i − y j ) − 1 ) C=((x_i-y_j)^{-1}) C = (( x i − y j ) − 1 ) . Borchardt’s identity is
per C = det ( ( x i − y j ) − 2 ) det ( ( x i − y j ) − 1 ) . (8) \operatorname{per} C=\frac{\det((x_i-y_j)^{-2})}{\det((x_i-y_j)^{-1})}. \tag*{(8)} per C = det (( x i − y j ) − 1 ) det (( x i − y j ) − 2 ) . (8)
The denominator is nonzero by the Cauchy determinant formula. The following notation describes the coefficients used for adding a polynomial to this kernel. For a ∈ N r a\in\mathbb{N}^r a ∈ N r and b ∈ N s b\in\mathbb{N}^s b ∈ N s , write ∣ a ∣ = ∑ α a α |a| = \sum_{\alpha} a_{\alpha} ∣ a ∣ = ∑ α a α , a ! = ∏ α a α ! a! = \prod_{\alpha} a_{\alpha}! a ! = ∏ α a α ! , and z a = ∏ α z α a α z^a = \prod_{\alpha} z_{\alpha}^{a_{\alpha}} z a = ∏ α z α a α , with the same conventions for b b b . Inequalities between multiindices are componentwise. Put
I n = { ( a , b ) ∈ N r × N s : ∣ a ∣ = ∣ b ∣ ≤ n } , D n = ∣ I n ∣ = ∑ h = 0 n ( h + r − 1 r − 1 ) ( h + s − 1 s − 1 ) . (9) \mathcal{I}_n = \{(a,b) \in\mathbb{N}^r \times\mathbb{N}^s : |a| = |b| \le n\}, \qquad D_n = |\mathcal{I}_n| = \sum_{h=0}^{n} \binom{h+r-1}{r-1}\binom{h+s-1}{s-1}. \tag*{(9)} I n = {( a , b ) ∈ N r × N s : ∣ a ∣ = ∣ b ∣ ≤ n } , D n = ∣ I n ∣ = h = 0 ∑ n ( r − 1 h + r − 1 ) ( s − 1 h + s − 1 ) . (9)
In particular, D n = O r , s ( n r + s − 1 ) D_n = O_{r,s}(n^{r+s-1}) D n = O r , s ( n r + s − 1 ) .
For an r × s r \times s r × s matrix L L L , define
e L ( a , b ) = [ z a w b ] ∏ α = 1 r ∏ β = 1 s ( ∑ m = 0 n L α β m m ! z α m w β m ) , ( a , b ) ∈ I n . (10) e_L(a,b) = [z^a w^b]\prod_{\alpha=1}^{r}\prod_{\beta=1}^{s}\left(\sum_{m=0}^{n}\frac{L_{\alpha\beta}^{m}}{m!}z_{\alpha}^{m}w_{\beta}^{m}\right), \qquad(a,b) \in\mathcal{I}_n. \tag*{(10)} e L ( a , b ) = [ z a w b ] α = 1 ∏ r β = 1 ∏ s ( m = 0 ∑ n m ! L α β m z α m w β m ) , ( a , b ) ∈ I n . (10)
This is a coefficient of a finite product. Equivalently,
e L ( a , b ) = ∑ M ∈ N r × s , M 1 = a , M T 1 = b ∏ α , β L α β M α β M α β ! . (11) e_L(a,b) = \sum_{\substack{M\in\mathbb{N}^{r\times s},\ M\mathbf{1}=a,\ M^{T}\mathbf{1}=b}}\prod_{\alpha,\beta}\frac{L_{\alpha\beta}^{M_{\alpha\beta}}}{M_{\alpha\beta}!}. \tag*{(11)} e L ( a , b ) = M ∈ N r × s , M 1 = a , M T 1 = b ∑ α , β ∏ M α β ! L α β M α β . (11)
Here and below 0 0 = 1 0^0 = 1 0 0 = 1 . Formal exponentials will be used in proofs to abbreviate (10) ; no infinite series has to be evaluated.
2.3.
Choose auxiliary column nodes ζ 1 , … , ζ r \zeta_1,\ldots,\zeta_r ζ 1 , … , ζ r and auxiliary row nodes η 1 , … , η s \eta_1,\ldots,\eta_s η 1 , … , η s . We call this choice admissible if the entries of ( x , η ) (x,\eta) ( x , η ) are distinct, the entries of ( y , ζ ) (y,\zeta) ( y , ζ ) are distinct, and the two lists are disjoint. Put
U i α = 1 x i − ζ α , V β j = 1 η β − y j , W β α = 1 η β − ζ α . (12) U_{i\alpha} = \frac{1}{x_i-\zeta_\alpha}, \qquad V_{\beta j} = \frac{1}{\eta_\beta-y_j}, \qquad W_{\beta\alpha} = \frac{1}{\eta_\beta-\zeta_\alpha}. \tag*{(12)} U i α = x i − ζ α 1 , V β j = η β − y j 1 , W β α = η β − ζ α 1 . (12)
For ( a , b ) ∈ I n (a,b) \in\mathcal{I}_n ( a , b ) ∈ I n , form a row list consisting of ( x i , 0 ) (x_i,0) ( x i , 0 ) , 1 ≤ i ≤ n 1 \le i \le n 1 ≤ i ≤ n , followed, for each β \beta β , by ( η β , 0 ) , … , ( η β , b β − 1 ) (\eta_\beta,0),\ldots,(\eta_\beta,b_\beta-1) ( η β , 0 ) , … , ( η β , b β − 1 ) . If b β = 0 b_\beta=0 b β = 0 , this part of the list is empty. Form the column list in the same way from ( y j , 0 ) (y_j,0) ( y j , 0 ) and ( ζ α , 0 ) , … , ( ζ α , a α − 1 ) (\zeta_\alpha,0),\ldots,(\zeta_\alpha,a_\alpha-1) ( ζ α , 0 ) , … , ( ζ α , a α − 1 ) . For q = 1 , 2 q=1,2 q = 1 , 2 , define a matrix on these lists by
K q ( a , b ) ( u , ℓ ) , ( v , m ) = ( − 1 ) ℓ ( q + ℓ + m − 1 ) ! ( q − 1 ) ! ℓ ! m ! ( u − v ) − q − ℓ − m . (13) K_q(a,b)_{(u,\ell),(v,m)} = (-1)^\ell\frac{(q+\ell+m-1)!}{(q-1)!\ell!m!}(u-v)^{-q-\ell-m}. \tag*{(13)} K q ( a , b ) ( u , ℓ ) , ( v , m ) = ( − 1 ) ℓ ( q − 1 )! ℓ ! m ! ( q + ℓ + m − 1 )! ( u − v ) − q − ℓ − m . (13)
Its order is n + ∣ a ∣ ≤ 2 n n+|a| \le2n n + ∣ a ∣ ≤ 2 n . Define
R a , b = det K 2 ( a , b ) det K 1 ( a , b ) . (14) R_{a,b} = \frac{\det K_2(a,b)}{\det K_1(a,b)}. \tag*{(14)} R a , b = det K 1 ( a , b ) det K 2 ( a , b ) . (14)
Lemma 5.2 proves that the denominator is nonzero and identifies this quotient as a permanent with repeated nodes.
The following finite sums will be used in the main formula:
Q a , b = ∑ c ≤ a , d ≤ b ∣ c ∣ = ∣ d ∣ e − W T ( c , d ) R a − c , b − d ( a − c ) ! ( b − d ) ! . (15) Q_{a,b} = \sum_{\substack{c\le a,\ d\le b\\ |c|=|d|}} e^{-W^T}(c,d)\frac{R_{a-c,b-d}}{(a-c)!(b-d)!}. \tag*{(15)} Q a , b = c ≤ a , d ≤ b ∣ c ∣ = ∣ d ∣ ∑ e − W T ( c , d ) ( a − c )! ( b − d )! R a − c , b − d . (15)
D ( x , y ; η , ζ ; L ) = ∑ ( a , b ) ∈ I n a ! b ! e L ( a , b ) Q a , b . (16) D(x,y;\eta,\zeta;L) = \sum_{(a,b)\in\mathcal{I}_n} a!b! \, e_L(a,b)Q_{a,b}. \tag*{(16)} D ( x , y ; η , ζ ; L ) = ( a , b ) ∈ I n ∑ a ! b ! e L ( a , b ) Q a , b . (16)
Thus D D D is explicitly defined by determinants, factorials and finite sums. Its definition contains no unevaluated permanent.
2.4.
For successive matrices W 1 , … , W n W_1,\ldots,W_n W 1 , … , W n of the same order, let W W W have blocks indexed by 0 , … , n 0,\ldots,n 0 , … , n , with W i − 1 , i = W i W_{i-1,i}=W_i W i − 1 , i = W i and all other blocks zero. If e 0 , e n e_0,e_n e 0 , e n are chosen coordinate vectors in blocks 0 , n 0,n 0 , n , then
det ( I − W e n − e 0 T 0 ) = e 0 T ( I − W ) − 1 e n . (17) \det\begin{pmatrix} I-W & e_n \\ -e_0^{\mathsf T} & 0 \end{pmatrix}=e_0^{\mathsf T}(I-W)^{-1}e_n. \tag*{(17)} det ( I − W − e 0 T e n 0 ) = e 0 T ( I − W ) − 1 e n . (17)
Indeed det ( I − W ) = 1 \det(I-W)=1 det ( I − W ) = 1 , and the Schur complement gives the identity. The ( 0 , n ) (0,n) ( 0 , n ) block of the inverse is W 1 ⋯ W n W_1\cdots W_n W 1 ⋯ W n , by the finite expansion I + W + ⋯ + W n I+W+\cdots+W^n I + W + ⋯ + W n .
Main results
3.1.
Let T \mathcal{T} T be a rooted binary tree on [ n ] [n] [ n ] , with cut coordinates as in Section 2. For a node I I I with children U , V U,V U , V , let
D I = ∑ α , β ( B U V ) α β ∂ x U , α ∂ y V , β + ∑ α , β ( B V U ) α β ∂ x V , α ∂ y U , β . D_I=\sum_{\alpha,\beta}(B_{UV})_{\alpha\beta}\partial_{x_{U,\alpha}}\partial_{y_{V,\beta}}+\sum_{\alpha,\beta}(B_{VU})_{\alpha\beta}\partial_{x_{V,\alpha}}\partial_{y_{U,\beta}}. D I = α , β ∑ ( B U V ) α β ∂ x U , α ∂ y V , β + α , β ∑ ( B V U ) α β ∂ x V , α ∂ y U , β .
At node I I I , let M I \mathcal{M}_I M I be the monomials in x I , y I x_I,y_I x I , y I of bidegree ( h , h ) (h,h) ( h , h ) , 0 ≤ h ≤ ∣ I ∣ 0\le h\le|I| 0 ≤ h ≤ ∣ I ∣ . For u ∈ M U u\in\mathcal{M}_U u ∈ M U , v ∈ M V v\in\mathcal{M}_V v ∈ M V , w ∈ M I w\in\mathcal{M}_I w ∈ M I , the entry in characteristic zero is
t I ( u , v ; w ) = [ w ] ρ I ( ∑ ℓ = 0 ∣ I ∣ D I ℓ ℓ ! u v ) . (18) t_I(u,v;w)=[w]\,\rho_I\left(\sum_{\ell=0}^{|I|}\frac{D_I^\ell}{\ell!}uv\right). \tag*{(18)} t I ( u , v ; w ) = [ w ] ρ I ℓ = 0 ∑ ∣ I ∣ ℓ ! D I ℓ uv . (18)
Over any field the expression inside ρ I \rho_I ρ I can be formed using the integer coefficients in (5.2). At a leaf { i } \{i\} { i } put
l i ( w ) = [ w ] ( a i i + ( R { i } ( i , : ) x { i } ) ( C { i } ( i , : ) y { i } ) ) . l_i(w)=[w]\left(a_{ii}+(R_{\{i\}}(i,:)x_{\{i\}})(C_{\{i\}}(i,:)y_{\{i\}})\right). l i ( w ) = [ w ] ( a ii + ( R { i } ( i , : ) x { i } ) ( C { i } ( i , : ) y { i } ) ) .
The permanent is obtained by summing the contributions from all nodes, with the constant monomial at the root.
Theorem 3.1 (Answer to Lyaudet’s question). Let k k k be a fixed nonnegative integer and A A A a square matrix over a field F \mathbb{F} F . If rw F ( A ) ≤ k \operatorname{rw}_{\mathbb{F}}(A)\le k rw F ( A ) ≤ k , then per A \operatorname{per} A per A can be computed from the entries of A A A alone in polynomial time, with field operations and zero tests counted at unit cost.
For a tree T \mathcal{T} T as above,
per A = ∑ w I ∈ M I for all nodes I w { 1 , … , n } = 1 ∏ i = 1 n l i ( w { i } ) ∏ I with children U , V t I ( w U , w V ; w I ) . (19) \operatorname{per} A=
\sum_{\substack{w_I\in\mathcal{M}_I\ \text{for all nodes }I\\w_{\{1,\ldots,n\}}=1}}
\prod_{i=1}^{n}l_i(w_{\{i\}})
\prod_{\substack{I\ \text{with}\\\text{children }U,V}}t_I(w_U,w_V;w_I). \tag*{(19)} per A = w I ∈ M I for all nodes I w { 1 , … , n } = 1 ∑ i = 1 ∏ n l i ( w { i } ) I with children U , V ∏ t I ( w U , w V ; w I ) . (19)
For fixed r , s r,s r , s , the sum can be evaluated in polynomial time from A A A and T \mathcal{T} T . For rational entries both finding the tree and evaluating the formula have polynomial bit complexity.
To evaluate (3.2), let p I ( w ) p_I(w) p I ( w ) be the coefficient at node I I I . Start with p { i } = l i p_{\{i\}}=l_i p { i } = l i and use
p I ( w ) = ∑ u ∈ M U , v ∈ M V p U ( u ) p V ( v ) t I ( u , v ; w ) . p_I(w)=\sum_{\substack{u\in\mathcal{M}_U,\ v\in\mathcal{M}_V}}p_U(u)p_V(v)t_I(u,v;w). p I ( w ) = u ∈ M U , v ∈ M V ∑ p U ( u ) p V ( v ) t I ( u , v ; w ) .
The answer is p [ n ] ( 1 ) p_{[n]}(1) p [ n ] ( 1 ) . Each coefficient records partial matchings inside I I I , with linear factors for their unmatched rows and columns. A matching on I I I has unique restrictions to its two children and a unique set of edges between them. This is the counting identity behind the formula. The polynomial-time bound comes from the number of coefficients at a node; it does not require listing all assignments in the displayed sum. The full argument is given in Section 5.1.
3.2.
Assume char F = p > 0 \operatorname{char} \mathbb{F}=p>0 char F = p > 0 . Use the cut coordinates of (2) and (3) , and write B I = B p , r I , s I B_I=B_{p,r_I,s_I} B I = B p , r I , s I . At a node with children U , V U,V U , V , put
E I = ∏ a = 1 r U ∏ b = 1 s V E ( B U V ) a b ; x U , a , y V , b ∏ a = 1 r V ∏ b = 1 s U E ( B V U ) a b ; x V , a , y U , b . E_I=\prod_{a=1}^{r_U}\prod_{b=1}^{s_V}E_{(B_{UV})_{ab};x_{U,a},y_{V,b}}\prod_{a=1}^{r_V}\prod_{b=1}^{s_U}E_{(B_{VU})_{ab};x_{V,a},y_{U,b}}. E I = a = 1 ∏ r U b = 1 ∏ s V E ( B U V ) ab ; x U , a , y V , b a = 1 ∏ r V b = 1 ∏ s U E ( B V U ) ab ; x V , a , y U , b .
The factors commute. For u ∈ B U u\in B_U u ∈ B U , v ∈ B V v\in B_V v ∈ B V and w ∈ B I w\in B_I w ∈ B I , define
t I ( p ) ( u , v ; w ) = [ w ] ρ I ( E I ( u v ) ) t_I^{(p)}(u,v;w)=[w]_{\rho_I}\bigl(E_I(uv)\bigr) t I ( p ) ( u , v ; w ) = [ w ] ρ I ( E I ( uv ) )
in the quotient (6) . At a leaf use the same polynomial l i l_i l i as in Theorem 3.1, reduced in this quotient. The next theorem bounds the number of coefficients independently of the number of leaves below the node.
Theorem 3.2. Let A A A have order n n n over a field of characteristic p > 0 p>0 p > 0 . For every rooted binary tree on its indices,
per A = ∑ w I ∈ B I ( I ∈ T ) w [ n ] = 1 ∏ i = 1 n l i ( w { i } ) ∏ I ∈ T I has children U , V t I ( p ) ( w U , w V ; w I ) . (20) \operatorname{per} A=\sum_{\substack{w_I\in B_I\ (I\in\mathcal{T})\\w_{[n]}=1}}\prod_{i=1}^{n}l_i(w_{\{i\}})\prod_{\substack{I\in\mathcal{T}\\I\text{ has children }U,V}}t_I^{(p)}(w_U,w_V;w_I). \tag*{(20)} per A = w I ∈ B I ( I ∈ T ) w [ n ] = 1 ∑ i = 1 ∏ n l i ( w { i } ) I ∈ T I has children U , V ∏ t I ( p ) ( w U , w V ; w I ) . (20) At a node of width at most k k k , the formula uses at most p k p^k p k coefficients. If rw F ( A ) ≤ k \operatorname{rw}_{\mathbb{F}}(A)\leq k rw F ( A ) ≤ k , the permanent can be computed from A A A alone in p O ( k ) n O ( 1 ) p^{O(k)}n^{O(1)} p O ( k ) n O ( 1 ) field operations and zero tests. The exponent of n n n is independent of p p p and k k k . For fixed p p p , this is polynomial when k = O ( log ( n + 1 ) ) k=O(\log(n+1)) k = O ( log ( n + 1 )) .
The field need not be finite. The displayed sum is evaluated by combining the coefficient vectors at successive nodes, as in Theorem 3.1. A tree supplied with width k k k can be used directly. From the matrix alone, it suffices to find a tree of width at most 4 k 4k 4 k ; this does not change the stated bound.
3.3.
In characteristic zero, consider
A i j = 1 x i − y j + P ( x i , y j ) , P ( X , Y ) = ∑ a = 0 r − 1 ∑ b = 0 s − 1 H a b X a Y b . (21) A_{ij}=\frac{1}{x_i-y_j}+P(x_i,y_j),\qquad P(X,Y)=\sum_{a=0}^{r-1}\sum_{b=0}^{s-1}H_{ab}X^aY^b. \tag*{(21)} A ij = x i − y j 1 + P ( x i , y j ) , P ( X , Y ) = a = 0 ∑ r − 1 b = 0 ∑ s − 1 H ab X a Y b . (21)
The nodes satisfy the distinctness and disjointness assumptions in the preliminaries. The positive integers r , s r,s r , s bound the degrees in the two variables. We specify the interpolation data for the formula. Take the rational numbers
ξ α = α ( 1 ≤ α ≤ r ) , υ β = r + β ( 1 ≤ β ≤ s ) . \xi_\alpha=\alpha\quad(1\leq\alpha\leq r),\qquad\upsilon_\beta=r+\beta\quad(1\leq\beta\leq s). ξ α = α ( 1 ≤ α ≤ r ) , υ β = r + β ( 1 ≤ β ≤ s ) .
Define the constant matrices
( B x ) a α = − ξ α − a − 1 ( 0 ≤ a < r ) , ( B y ) β b = υ β − b − 1 ( 0 ≤ b < s ) . (B_x)_{a\alpha}=-\xi_\alpha^{-a-1}\quad(0\leq a<r),\qquad(B_y)_{\beta b}=\upsilon_\beta^{-b-1}\quad(0\leq b<s). ( B x ) a α = − ξ α − a − 1 ( 0 ≤ a < r ) , ( B y ) β b = υ β − b − 1 ( 0 ≤ b < s ) .
They are invertible Vandermonde matrices with nonzero column or row factors. For t ≠ 0 t\ne0 t = 0 , put
ζ α ( t ) = ξ α / t , η β ( t ) = υ β / t , T r ( t ) = diag ( t − 1 , … , t − r ) , T s ( t ) = diag ( t − 1 , … , t − s ) , H ( t ) = B x − 1 T r ( t ) H T s ( t ) B y − 1 . (22) \begin{aligned}
\zeta_\alpha(t)&=\xi_\alpha/t,\qquad\eta_\beta(t)=\upsilon_\beta/t,\\
T_r(t)&=\operatorname{diag}(t^{-1},\ldots,t^{-r}),\qquad T_s(t)=\operatorname{diag}(t^{-1},\ldots,t^{-s}),\\
H(t)&=B_x^{-1}T_r(t)HT_s(t)B_y^{-1}. \tag*{(22)}
\end{aligned} ζ α ( t ) T r ( t ) H ( t ) = ξ α / t , η β ( t ) = υ β / t , = diag ( t − 1 , … , t − r ) , T s ( t ) = diag ( t − 1 , … , t − s ) , = B x − 1 T r ( t ) H T s ( t ) B y − 1 . (22)
The H H H in (22) is the coefficient matrix in (21) , in increasing order of the two exponents. Put
m 0 = n ( r + s − 2 ) , d 0 = n ( 2 r + 2 s − 2 ) , (23) m_0=n(r+s-2),\qquad d_0=n(2r+2s-2), \tag*{(23)} m 0 = n ( r + s − 2 ) , d 0 = n ( 2 r + 2 s − 2 ) , (23)
and
q ( t ) = ∏ i = 1 n ∏ α = 1 r ( x i t − ξ α ) ∏ β = 1 s ∏ j = 1 n ( υ β − y j t ) . (24) q(t)=\prod_{i=1}^{n}\prod_{\alpha=1}^{r}(x_it-\xi_\alpha)\prod_{\beta=1}^{s}\prod_{j=1}^{n}(\upsilon_\beta-y_jt). \tag*{(24)} q ( t ) = i = 1 ∏ n α = 1 ∏ r ( x i t − ξ α ) β = 1 ∏ s j = 1 ∏ n ( υ β − y j t ) . (24)
In particular q ( 0 ) ≠ 0 q(0) \ne0 q ( 0 ) = 0 . Choose d 0 + 1 d_0+1 d 0 + 1 distinct nonzero rational numbers t 0 , … , t d 0 t_0,\ldots,t_{d_0} t 0 , … , t d 0 for which the auxiliary nodes in (22) are admissible. They can be chosen as positive integers of size O r , s ( n ) O_{r,s}(n) O r , s ( n ) . Indeed, there are at most 2 n ( r + s ) 2n(r+s) 2 n ( r + s ) forbidden nonzero values. Finally put
λ ℓ = [ t m 0 ] ∏ 0 ≤ j ≤ d 0 j ≠ ℓ t − t j t ℓ − t j . (25) \lambda_\ell=[t^{m_0}]\prod_{\substack{0\le j\le d_0\\j\ne\ell}}\frac{t-t_j}{t_\ell-t_j}. \tag*{(25)} λ ℓ = [ t m 0 ] 0 ≤ j ≤ d 0 j = ℓ ∏ t ℓ − t j t − t j . (25)
This is the coefficient of an explicitly given polynomial of degree d 0 d_0 d 0 .
The next theorem evaluates the permanent after a polynomial of fixed degree has been added to a Cauchy kernel.
Theorem 3.3. For the matrix A A A in (21) ,
per A = 1 q ( 0 ) ∑ ℓ = 0 d 0 λ ℓ t ℓ m 0 q ( t ℓ ) D ( x , y ; η ( t ℓ ) , ζ ( t ℓ ) ; H ( t ℓ ) ) . (26) \operatorname{per} A=\frac{1}{q(0)}\sum_{\ell=0}^{d_0}\lambda_\ell t_\ell^{m_0}q(t_\ell)D\bigl(x,y;\eta(t_\ell),\zeta(t_\ell);H(t_\ell)\bigr). \tag*{(26)} per A = q ( 0 ) 1 ℓ = 0 ∑ d 0 λ ℓ t ℓ m 0 q ( t ℓ ) D ( x , y ; η ( t ℓ ) , ζ ( t ℓ ) ; H ( t ℓ ) ) . (26) For fixed r , s r,s r , s , this formula can be evaluated in
O r , s ( n r + s + 3 + n 2 r + 2 s − 1 ) (27) O_{r,s}\left(n^{r+s+3}+n^{2r+2s-1}\right) \tag*{(27)} O r , s ( n r + s + 3 + n 2 r + 2 s − 1 ) (27) field operations and zero tests. Every determinant has order at most 2 n 2n 2 n . For rational input, the bit complexity is polynomial in n n n and the input bit length.
Only the bounds on the degrees are fixed here. The nodes and all the coefficients of P P P may vary with n n n . Formula (26) is nonrecursive. Its auxiliary evaluation points are chosen by the stated finite rule.
Suppose the two ranks at the initial interval cuts are bounded by r , s r,s r , s . We use the following generators for the upper and lower entries.
a i j = p i T B i + 1 ⋯ B j − 1 q j , a_{ij}=p_i^{\mathsf T}B_{i+1}\cdots B_{j-1}q_j, a ij = p i T B i + 1 ⋯ B j − 1 q j ,
a j i = u i T C i + 1 ⋯ C j − 1 v j ( i < j ) . (28) a_{ji}=u_i^{\mathsf T}C_{i+1}\cdots C_{j-1}v_j\qquad(i<j). \tag*{(28)} a j i = u i T C i + 1 ⋯ C j − 1 v j ( i < j ) . (28)
The first line uses vectors and square matrices of order r r r , and the second uses order s s s . An empty product is the identity.
Write d i = a i i d_i=a_{ii} d i = a ii . Put N = ⌊ n / 2 ⌋ N=\lfloor n/2\rfloor N = ⌊ n /2 ⌋ and let
M = { x α y β : ∣ α ∣ = ∣ β ∣ ≤ N } , D = ∣ M ∣ , \mathcal{M}=\{x^\alpha y^\beta:|\alpha|=|\beta|\le N\},\qquad D=|\mathcal{M}|, M = { x α y β : ∣ α ∣ = ∣ β ∣ ≤ N } , D = ∣ M ∣ ,
where x , y x,y x , y have r , s r,s r , s variables. Zero generators may be used on either side. Denote by Π \Pi Π the projection onto the span of M \mathcal{M} M . Define
D i = q i T ∂ x , E i = v i T ∂ y , L i = p i T x , M i = u i T y , D_i=q_i^{\mathsf T}\partial_x,\qquad E_i=v_i^{\mathsf T}\partial_y,\qquad L_i=p_i^{\mathsf T}x,\qquad M_i=u_i^{\mathsf T}y, D i = q i T ∂ x , E i = v i T ∂ y , L i = p i T x , M i = u i T y ,
Φ i f = f ( B i x , C i y ) , \Phi_i f=f(B_i x,C_i y), Φ i f = f ( B i x , C i y ) ,
U i f = Π ( Φ i ( D i E i f ) + L i Φ i ( D i f ) + M i Φ i ( E i f ) + ( d i + L i M i ) Φ i f ) . U_i f=\Pi\bigl(\Phi_i(D_iE_i f)+L_i\Phi_i(D_i f)+M_i\Phi_i(E_i f)+(d_i+L_iM_i)\Phi_i f\bigr). U i f = Π ( Φ i ( D i E i f ) + L i Φ i ( D i f ) + M i Φ i ( E i f ) + ( d i + L i M i ) Φ i f ) .
The matrices W i W_i W i , of order D D D , have entries
( W i ) f , g = [ g ] U i f ( f , g ∈ M ) . (29) (W_i)_{f,g}=[g]U_i f\qquad(f,g\in\mathcal{M}). \tag*{(29)} ( W i ) f , g = [ g ] U i f ( f , g ∈ M ) . (29)
D = ∑ h = 0 N ( h + r − 1 r − 1 ) ( h + s − 1 s − 1 ) . (30) D=\sum_{h=0}^{N}\binom{h+r-1}{r-1}\binom{h+s-1}{s-1}. \tag*{(30)} D = h = 0 ∑ N ( r − 1 h + r − 1 ) ( s − 1 h + s − 1 ) . (30)
Use the block matrix W W W of (17) for the matrices W i W_i W i , and let e 0 , e n e_0,e_n e 0 , e n select the constant monomial at the first and last cuts. The constant-to-constant entry of the product of the W i W_i W i counts the perfect matchings.
Theorem 3.4. The rank bounds imply the representation (28) . With the matrices just defined,
per A = ( W 1 ⋯ W n ) 1 , 1 = det ( I − W e n − e 0 T 0 ) . (31) \operatorname{per} A=(W_1\cdots W_n)_{1,1}=\det\begin{pmatrix} I-W & e_n \\ -e_0^{\mathsf T} & 0 \end{pmatrix}. \tag*{(31)} per A = ( W 1 ⋯ W n ) 1 , 1 = det ( I − W − e 0 T e n 0 ) . (31) The order of the determinant is ( n + 1 ) D + 1 (n+1)D+1 ( n + 1 ) D + 1 . For fixed r , s r,s r , s , evaluation from the entries of A A A takes O r , s ( n 2 + n r + s + 1 ) O_{r,s}(n^2+n^{r+s+1}) O r , s ( n 2 + n r + s + 1 ) arithmetic operations, and O ( n 2 ) O(n^2) O ( n 2 ) when r = s = 1 r=s=1 r = s = 1 .
3.5.
Now let char F = p > 0 \operatorname{char} F=p>0 char F = p > 0 in the generator representation (28) . Use the same five terms defining U i U_i U i , with reduction modulo the individual p p p th powers in place of Π \Pi Π . Denote this operator by U i ( p ) U_i^{(p)} U i ( p ) , and set
( W i ( p ) ) f , g = [ g ] U i ( p ) f , f , g ∈ B p , r , s . (W_i^{(p)})_{f,g}=[g]U_i^{(p)}f,\qquad f,g\in\mathcal{B}_{p,r,s}. ( W i ( p ) ) f , g = [ g ] U i ( p ) f , f , g ∈ B p , r , s .
The constant monomial is the first index. Form W ( p ) W^{(p)} W ( p ) from these matrices as in (17) . The basis now has size independent of n n n .
Corollary 3.5. For the generators (28) in characteristic p p p ,
per A = ( W 1 ( p ) ⋯ W n ( p ) ) 1 , 1 = det ( I − W ( p ) e n − e 0 T 0 ) . (32) \operatorname{per} A=(W_1^{(p)}\cdots W_n^{(p)})_{1,1}=\det\begin{pmatrix} I-W^{(p)} & e_n \\ -e_0^{\mathsf T} & 0 \end{pmatrix}. \tag*{(32)} per A = ( W 1 ( p ) ⋯ W n ( p ) ) 1 , 1 = det ( I − W ( p ) − e 0 T e n 0 ) . (32) The determinant has order ( n + 1 ) D p , r , s + 1 (n+1)D_{p,r,s}+1 ( n + 1 ) D p , r , s + 1 . From the supplied generators the product can be evaluated in p O ( r + s ) n p^{O(r+s)}n p O ( r + s ) n field operations.
3.6.
Write C ( σ ) C(\sigma) C ( σ ) for the set of cycles of σ \sigma σ , including its fixed points. We regard σ \sigma σ as the directed cycle cover with edges i → σ ( i ) i\to\sigma(i) i → σ ( i ) , each of weight a i , σ ( i ) a_{i,\sigma(i)} a i , σ ( i ) . A permutation consisting of one cycle gives a directed Hamiltonian cycle. Simultaneous permutation of rows and columns relabels the vertices and preserves all cycle lengths.
Fix d ≥ 1 d\geq1 d ≥ 1 , and let x = ( x 1 , … , x d ) x=(x_1,\ldots,x_d) x = ( x 1 , … , x d ) and z = ( z 1 , … , z d ) z=(z_1,\ldots,z_d) z = ( z 1 , … , z d ) be indeterminates. Put
F A ( x , z ) = ∑ σ ∈ S n ( ∏ i = 1 n a i , σ ( i ) ) ∏ C ∈ C ( σ ) ( ∑ c = 1 d z c x c ∣ C ∣ ) . (33) F_A(x,z)=\sum_{\sigma\in S_n}\left(\prod_{i=1}^{n}a_{i,\sigma(i)}\right)\prod_{C\in C(\sigma)}\left(\sum_{c=1}^{d}z_cx_c^{|C|}\right). \tag*{(33)} F A ( x , z ) = σ ∈ S n ∑ ( i = 1 ∏ n a i , σ ( i ) ) C ∈ C ( σ ) ∏ ( c = 1 ∑ d z c x c ∣ C ∣ ) . (33)
The index c c c is a color for a cycle. The variable z c z_c z c counts cycles of color c c c , and x c x_c x c counts their vertices.
For one color put P A ( z ) = F A ( ( 1 ) , ( z ) ) P_A(z)=F_A((1),(z)) P A ( z ) = F A (( 1 ) , ( z )) and C k ( A ) = [ z k ] P A ( z ) C_k(A)=[z^k]P_A(z) C k ( A ) = [ z k ] P A ( z ) . Thus C k ( A ) C_k(A) C k ( A ) is the sum of the weights of the permutations with k k k cycles. In particular, C 1 ( A ) C_1(A) C 1 ( A ) is the weighted Hamiltonian-cycle sum. The polynomial P A ( z ) P_A(z) P A ( z ) is the α \alpha α -permanent with its parameter left indeterminate [25 ] . The sign of a permutation with k k k cycles is ( − 1 ) n − k (-1)^{n-k} ( − 1 ) n − k , so
P A ( 1 ) = per A , P A ( − 1 ) = ( − 1 ) n det A . (34) P_A(1)=\operatorname{per} A,\qquad P_A(-1)=(-1)^n\det A. \tag*{(34)} P A ( 1 ) = per A , P A ( − 1 ) = ( − 1 ) n det A . (34)
The variables for the full polynomial record both ends of a path. We use the same cut factorizations and couplings as for the permanent. At node I I I , use d d d matrices Z I ( c ) Z_I^{(c)} Z I ( c ) of r I × s I r_I\times s_I r I × s I indeterminates. At a merge with children U , V U,V U , V , put
B = ( 0 B U V B V U 0 ) , S I = ( S U , I S V , I ) , T I = ( T U , I T V , I ) . B=\begin{pmatrix}0&B_{UV}\\B_{VU}&0\end{pmatrix},\qquad S_I=\begin{pmatrix}S_{U,I}\\S_{V,I}\end{pmatrix},\qquad T_I=\begin{pmatrix}T_{U,I}\\T_{V,I}\end{pmatrix}. B = ( 0 B V U B U V 0 ) , S I = ( S U , I S V , I ) , T I = ( T U , I T V , I ) .
For each c c c , use a matrix Z ( c ) Z^{(c)} Z ( c ) with rows from the coordinates of R U , R V R_U,R_V R U , R V and columns from those of C U , C V C_U,C_V C U , C V . Identify Z U ( c ) Z_U^{(c)} Z U ( c ) and Z V ( c ) Z_V^{(c)} Z V ( c ) with its diagonal blocks. Write ∂ α β ( c ) = ∂ / ∂ Z α β ( c ) \partial_{\alpha\beta}^{(c)}=\partial/\partial Z_{\alpha\beta}^{(c)} ∂ α β ( c ) = ∂ / ∂ Z α β ( c ) . Define
E I = ∑ c = 1 d ( z c ∑ α , β B α β ∂ α β ( c ) + ∑ α , β , γ , δ B α β Z γ δ ( c ) ∂ α δ ( c ) ∂ γ β ( c ) ) . (35) \mathcal{E}_I=\sum_{c=1}^{d}\left(z_c\sum_{\alpha,\beta}B_{\alpha\beta}\partial_{\alpha\beta}^{(c)}+\sum_{\alpha,\beta,\gamma,\delta}B_{\alpha\beta}Z_{\gamma\delta}^{(c)}\partial_{\alpha\delta}^{(c)}\partial_{\gamma\beta}^{(c)}\right). \tag*{(35)} E I = c = 1 ∑ d z c α , β ∑ B α β ∂ α β ( c ) + α , β , γ , δ ∑ B α β Z γ δ ( c ) ∂ α δ ( c ) ∂ γ β ( c ) . (35)
Let η I \eta_I η I be the substitution Z ( c ) ↦ S I Z I ( c ) T I T Z^{(c)} \mapsto S_I Z_I^{(c)}T_I^{\mathsf{T}} Z ( c ) ↦ S I Z I ( c ) T I T for all c c c . Let N I \mathcal{N}_I N I be the monomials in the entries of the Z I ( c ) Z_I^{(c)} Z I ( c ) of total degree at most ∣ I ∣ |I| ∣ I ∣ . For u ∈ N U u \in\mathcal{N}_U u ∈ N U , v ∈ N V v \in\mathcal{N}_V v ∈ N V , w ∈ N I w \in\mathcal{N}_I w ∈ N I , set
t I c y c ( u , v ; w ) = [ w ] η I ( ∑ ℓ = 0 ∣ I ∣ E I ℓ ℓ ! u v ) . (36) t_I^{\mathrm{cyc}}(u,v;w)=[w]\eta_I\left(\sum_{\ell=0}^{|I|}\frac{\mathcal{E}_I^\ell}{\ell!}uv\right). \tag*{(36)} t I cyc ( u , v ; w ) = [ w ] η I ℓ = 0 ∑ ∣ I ∣ ℓ ! E I ℓ uv . (36)
At a leaf { i } \{i\} { i } put
l i c y c ( w ) = [ w ] ∑ c = 1 d x c ( z c a i i + R { i } ( i , : ) Z { i } ( c ) C { i } ( i , : ) T ) . l_i^{\mathrm{cyc}}(w)=[w]\sum_{c=1}^{d}x_c\left(z_ca_{ii}+R_{\{i\}}(i,:)Z_{\{i\}}^{(c)}C_{\{i\}}(i,:)^{\mathsf{T}}\right). l i cyc ( w ) = [ w ] c = 1 ∑ d x c ( z c a ii + R { i } ( i , : ) Z { i } ( c ) C { i } ( i , : ) T ) .
For an ordering, put I i = { 1 , … , i } I_i=\{1,\ldots,i\} I i = { 1 , … , i } , N = ⌊ n / 2 ⌋ N=\lfloor n/2\rfloor N = ⌊ n /2 ⌋ and D d = ( N + d r s d r s ) D_d=\binom{N+drs}{drs} D d = ( d r s N + d r s ) . At each cut retain the monomials of degree at most N N N and pad the coefficient vectors with zeros to length D d D_d D d . For i ≥ 2 i\ge2 i ≥ 2 the transition from the monomial u u u at I i − 1 I_{i-1} I i − 1 to w w w at I i I_i I i has entry
( W i c y c ) u , w = ∑ v ∈ N { i } l i c y c ( v ) t I i c y c ( u , v ; w ) , \left(W_i^{\mathrm{cyc}}\right)_{u,w}=\sum_{v\in\mathcal{N}_{\{i\}}}l_i^{\mathrm{cyc}}(v)t_{I_i}^{\mathrm{cyc}}(u,v;w), ( W i cyc ) u , w = v ∈ N { i } ∑ l i cyc ( v ) t I i cyc ( u , v ; w ) ,
with children I i − 1 , { i } I_{i-1},\{i\} I i − 1 , { i } . At i = 1 i=1 i = 1 the entry from the constant state to w w w is l 1 c y c ( w ) l_1^{\mathrm{cyc}}(w) l 1 cyc ( w ) . Form the strictly upper block triangular matrix W c y c W^{\mathrm{cyc}} W cyc with block ( i − 1 , i ) (i-1,i) ( i − 1 , i ) equal to W i c y c W_i^{\mathrm{cyc}} W i cyc , as in (31) . Let e 0 , e n e_0,e_n e 0 , e n select the constant monomials at the first and last cuts. The endpoint variables retain the pairing of the start and end of each path. This allows cycles to be counted when they close.
Theorem 3.6. Assume that F \mathbb{F} F has characteristic zero. Suppose the two ranks at every node of T T T are bounded by r , s r,s r , s , respectively. Then
F A ( x , z ) = ∑ w I ∈ N I for all nodes I w { 1 , … , n } = 1 ∏ i = 1 n l i c y c ( w { i } ) ∏ I with children U , V t I c y c ( w U , w V ; w I ) . (37) F_A(x,z)=\sum_{\substack{w_I\in\mathcal{N}_I\ \text{for all nodes }I\\w_{\{1,\ldots,n\}}=1}}\prod_{i=1}^{n}l_i^{\mathrm{cyc}}(w_{\{i\}})\prod_{\substack{I\ \text{with}\\\text{children }U,V}}t_I^{\mathrm{cyc}}(w_U,w_V;w_I). \tag*{(37)} F A ( x , z ) = w I ∈ N I for all nodes I w { 1 , … , n } = 1 ∑ i = 1 ∏ n l i cyc ( w { i } ) I with children U , V ∏ t I cyc ( w U , w V ; w I ) . (37) For fixed d , r , s d,r,s d , r , s , all coefficients can be computed in polynomial time with polynomial bit complexity for rational entries. The existence of a suitable tree suffices for polynomial-time computation from A A A alone. If the rank bounds hold along an ordering, then
F A ( x , z ) = det ( I − W c y c e n − e 0 T 0 ) . (38) F_A(x,z)=\det\begin{pmatrix}I-W^{\mathrm{cyc}}&e_n\\-e_0^{\mathsf{T}}&0\end{pmatrix}. \tag*{(38)} F A ( x , z ) = det ( I − W cyc − e 0 T e n 0 ) . (38) The order is ( n + 1 ) ( ⌊ n / 2 ⌋ + d r s d r s ) + 1 (n+1)\binom{\lfloor n/2\rfloor+drs}{drs}+1 ( n + 1 ) ( d r s ⌊ n /2 ⌋ + d r s ) + 1 .
3.7.
Fix 1 ≤ k ≤ n 1\le k\le n 1 ≤ k ≤ n . Choose marks v 1 < ⋯ < v k v_1<\cdots<v_k v 1 < ⋯ < v k in [ n ] [n] [ n ] . Write A ( v 1 , … , v k ) \mathcal{A}(v_1,\ldots,v_k) A ( v 1 , … , v k ) for the ordered partitions into S 1 , T 1 , … , S k , T k S_1,T_1,\ldots,S_k,T_k S 1 , T 1 , … , S k , T k with v c ∈ S c v_c\in S_c v c ∈ S c and min ( S c ∪ T c ) = v c \min(S_c\cup T_c)=v_c min ( S c ∪ T c ) = v c . The sets T c T_c T c may be empty. The two sets of a pair will carry permanent and determinant weights, respectively.
Put N = ⌊ n / 2 ⌋ N=\lfloor n/2\rfloor N = ⌊ n /2 ⌋ and set
D + = ∑ h = 0 N ( h + r − 1 r − 1 ) ( h + s − 1 s − 1 ) , D − = ( r + s r ) , D k = ( k + 1 ) ( D + D − ) k . (39) D^+=\sum_{h=0}^{N}\binom{h+r-1}{r-1}\binom{h+s-1}{s-1},\qquad D^-=\binom{r+s}{r},\qquad D_k=(k+1)(D^+D^-)^k. \tag*{(39)} D + = h = 0 ∑ N ( r − 1 h + r − 1 ) ( s − 1 h + s − 1 ) , D − = ( r r + s ) , D k = ( k + 1 ) ( D + D − ) k . (39)
Ordinary polynomial variables commute. In the formulas involving determinants of color classes we also use exterior variables. They anticommute and have square zero. A left derivative with respect to an exterior variable removes its occurrence in an ordered monomial with sign ( − 1 ) j − 1 (-1)^{j-1} ( − 1 ) j − 1 if it is the j j j -th factor, and gives zero if it is absent. Different colors use separate tensor factors.
For the determinant in the ordered case, regard S c , T c S_c,T_c S c , T c as the two colors of pair c c c . At cut I I I use commuting variables x I ( c ) , y I ( c ) x_I^{(c)},y_I^{(c)} x I ( c ) , y I ( c ) for S c S_c S c , as in (18) . Use exterior variables u I ( c ) , v I ( c ) u_I^{(c)},v_I^{(c)} u I ( c ) , v I ( c ) for T c T_c T c , of respective lengths r I , s I r_I,s_I r I , s I . At a leaf { i } \{i\} { i } , the factors for the two choices of color are
L i ( c , + ) = a i i + ( R { i } ( i , : ) x { i } ( c ) ) ( C { i } ( i , : ) y { i } ( c ) ) , L_i^{(c,+)}=a_{ii}+(R_{\{i\}}(i,:)\,x_{\{i\}}^{(c)})(C_{\{i\}}(i,:)\,y_{\{i\}}^{(c)}), L i ( c , + ) = a ii + ( R { i } ( i , : ) x { i } ( c ) ) ( C { i } ( i , : ) y { i } ( c ) ) ,
L i ( c , − ) = − a i i − ( R { i } ( i , : ) u { i } ( c ) ) ( C { i } ( i , : ) v { i } ( c ) ) . (40) L_i^{(c,-)}=-a_{ii}-(R_{\{i\}}(i,:)\,u_{\{i\}}^{(c)})(C_{\{i\}}(i,:)\,v_{\{i\}}^{(c)}). \tag*{(40)} L i ( c , − ) = − a ii − ( R { i } ( i , : ) u { i } ( c ) ) ( C { i } ( i , : ) v { i } ( c ) ) . (40)
At a merge use the operator D I D_I D I of (18) on each commuting pair. On each exterior pair use the operator
∑ ( J , K ) ∈ { ( U , V ) , ( V , U ) } ∑ ( B J K ) α β ∂ v K , β ( c ) ∂ u J , α ( c ) . (41) \sum_{(J,K)\in\{(U,V),(V,U)\}}\sum(B_{JK})_{\alpha\beta}\partial_{v_{K,\beta}^{(c)}}\partial_{u_{J,\alpha}^{(c)}}. \tag*{(41)} ( J , K ) ∈ {( U , V ) , ( V , U )} ∑ ∑ ( B J K ) α β ∂ v K , β ( c ) ∂ u J , α ( c ) . (41)
Here the derivatives on exterior variables are left derivatives. Denote the sum of these operators over all colors by D I ( k ) D_I^{(k)} D I ( k ) . Apply ∑ j = 0 ∣ I ∣ ( D I ( k ) ) j / j ! \sum_{j=0}^{|I|}(D_I^{(k)})^j/j! ∑ j = 0 ∣ I ∣ ( D I ( k ) ) j / j ! to the product of the two child polynomials. Then apply the changes of coordinates in (18) to both types of variables.
Along the ordering 1 , … , n 1,\ldots,n 1 , … , n , also keep the number ℓ \ell ℓ of marks chosen so far. At the next index one may use either color of a pair c ≤ ℓ c\leq\ell c ≤ ℓ . If ℓ < k \ell<k ℓ < k , one may choose the index as the next mark, using L i ( ℓ + 1 , + ) L_i^{(\ell+1,+)} L i ( ℓ + 1 , + ) and increasing ℓ \ell ℓ by one. For each choice multiply by the corresponding factor in (3.23), merge with the next leaf, and keep commuting degrees at most N N N in the row variables and in the column variables of each pair. At the first index use the leaf factor alone. Denote by W i ( k ) W_i^{(k)} W i ( k ) the matrix of this linear map in the monomial bases, padded with zeros to order D k D_k D k . Form W ( k ) W^{(k)} W ( k ) with block ( i − 1 , i ) (i-1,i) ( i − 1 , i ) equal to W i ( k ) W_i^{(k)} W i ( k ) and all other blocks zero.
Let e 0 , e n e_0,e_n e 0 , e n select the constant states with ℓ = 0 , k \ell=0,k ℓ = 0 , k at the two ends. The marked partitions count each permutation with k k k cycles once, after cancellation of the other terms.
Theorem 3.7. Assume that F F F has characteristic zero. Suppose the two ranks at every node of T \mathcal{T} T are bounded by r , s r,s r , s , respectively. Then
C k ( A ) = ∑ v 1 < ⋯ < v k ∑ ( S 1 , T 1 , … , S k , T k ) ∈ A ( v 1 , … , v k ) ∏ c = 1 k per A [ S c , S c ] ( − 1 ) ∣ T c ∣ det A [ T c , T c ] . (42) C_k(A)=\sum_{v_1<\cdots<v_k}\sum_{(S_1,T_1,\ldots,S_k,T_k)\in A(v_1,\ldots,v_k)}\prod_{c=1}^{k}\operatorname{per} A[S_c,S_c]\,(-1)^{|T_c|}\det A[T_c,T_c]. \tag*{(42)} C k ( A ) = v 1 < ⋯ < v k ∑ ( S 1 , T 1 , … , S k , T k ) ∈ A ( v 1 , … , v k ) ∑ c = 1 ∏ k per A [ S c , S c ] ( − 1 ) ∣ T c ∣ det A [ T c , T c ] . (42) For fixed k , r , s k,r,s k , r , s , the sum can be evaluated on T \mathcal{T} T in polynomial time, with polynomial bit complexity for rational entries. If the rank bounds hold along an ordering, then
C k ( A ) = det ( I − W ( k ) e n − e 0 T 0 ) . (43) C_k(A)=\det\begin{pmatrix} I-W^{(k)} & e_n \\ -e_0^{\mathsf{T}} & 0 \end{pmatrix}. \tag*{(43)} C k ( A ) = det ( I − W ( k ) − e 0 T e n 0 ) . (43) The order is at most 1 + ( n + 1 ) D k = O k , r , s ( n k ( r + s − 1 ) + 1 ) 1+(n+1)D_k=O_{k,r,s}(n^{k(r+s-1)+1}) 1 + ( n + 1 ) D k = O k , r , s ( n k ( r + s − 1 ) + 1 ) . For k = 1 k=1 k = 1 this gives a determinant of order O r , s ( n r + s ) O_{r,s}(n^{r+s}) O r , s ( n r + s ) for the weighted Hamiltonian-cycle sum.
For k = 1 k=1 k = 1 the mark in (3.25) is 1. One can instead prescribe any vertex in S 1 S_1 S 1 and drop the minimum condition. For related permanent and determinant formulas for Hamiltonian cycles, see [26 ] .
3.8.
For a partition λ \lambda λ of n n n , let χ λ \chi_\lambda χ λ be the corresponding irreducible character of S n S_n S n . Its immanant
imm λ ( A ) = ∑ σ ∈ S n χ λ ( σ ) ∏ i a i , σ ( i ) . \operatorname{imm}_{\lambda}(A)=\sum_{\sigma\in S_n}\chi_{\lambda}(\sigma)\prod_i a_{i,\sigma(i)}. imm λ ( A ) = σ ∈ S n ∑ χ λ ( σ ) i ∏ a i , σ ( i ) .
The ordinary hook partitions are ( n − k , 1 k ) (n-k,1^k) ( n − k , 1 k ) . More generally, the ( p , q ) (p,q) ( p , q ) -hook is the condition λ p + 1 ≤ q \lambda_{p+1}\leq q λ p + 1 ≤ q , with missing parts taken to be zero. The Durfee size is the largest j j j with λ j ≥ j \lambda_j\geq j λ j ≥ j . A power-sum specialization of F A F_A F A gives all immanants whose partitions lie in a fixed ( p , q ) (p,q) ( p , q ) -hook.
Corollary 3.8. Under the hypotheses of Theorem 3.6 , for fixed nonnegative integers p p p , q q q with p + q > 0 p+q>0 p + q > 0 , all the immanants imm λ ( A ) \operatorname{imm}_{\lambda}(A) imm λ ( A ) with λ p + 1 ≤ q \lambda_{p+1}\leq q λ p + 1 ≤ q can be computed in polynomial time. In particular this holds for all ordinary hook immanants and for all immanants of bounded Durfee size. For rational entries the bit complexity is polynomial. For ordinary hooks,
F A ( ( 1 , − t ) , ( 1 , − 1 ) ) 1 + t = ∑ k = 0 n − 1 imm ( n − k , 1 k ) ( A ) t k . (44) \frac{F_A((1,-t),(1,-1))}{1+t}=\sum_{k=0}^{n-1}\operatorname{imm}_{(n-k,1^k)}(A)t^k. \tag*{(44)} 1 + t F A (( 1 , − t ) , ( 1 , − 1 )) = k = 0 ∑ n − 1 imm ( n − k , 1 k ) ( A ) t k . (44)
3.9. Determinant methods for counting Hamiltonian cycles on graphs of bounded treewidth were developed by Bodlaender, Cygan, Kratsch and Nederlof [27 ] . We compare the two calculations at a single cut. For the use of completion matrices in counting Hamiltonian cycles on graphs of bounded pathwidth, see Curticapean, Lindzey and Nederlof [28 ] . For h ≥ 1 h\geq1 h ≥ 1 let B h B_h B h be the set of nonnegative r × s r\times s r × s integer arrays M = ( m a b ) M=(m_{ab}) M = ( m ab ) of sum h h h . List h h h pairs ( a , b ) (a,b) ( a , b ) with these multiplicities, and let U M , V M U_M,V_M U M , V M have corresponding rows e a T , e b T e_a^{\mathsf T},e_b^{\mathsf T} e a T , e b T of lengths r , s r,s r , s , respectively. For M , M ′ ∈ B h M,M'\in B_h M , M ′ ∈ B h put
A M , M ′ = ( 0 U M U M ′ T V M ′ V M T 0 ) , K h ( z ) = ( P A M , M ′ ( z ) ) M , M ′ ∈ B h . A_{M,M'}=\begin{pmatrix}0&U_MU_{M'}^{\mathsf T}\\V_{M'}V_M^{\mathsf T}&0\end{pmatrix},\qquad K_h(z)=\left(P_{A_{M,M'}}(z)\right)_{M,M'\in B_h}. A M , M ′ = ( 0 V M ′ V M T U M U M ′ T 0 ) , K h ( z ) = ( P A M , M ′ ( z ) ) M , M ′ ∈ B h .
The first h h h vertices precede the last h h h . Put B h = ( h + r − 1 r − 1 ) ( h + s − 1 s − 1 ) B_h=\binom{h+r-1}{r-1}\binom{h+s-1}{s-1} B h = ( r − 1 h + r − 1 ) ( s − 1 h + s − 1 ) . The following proposition compares the ranks for the whole cycle polynomial and for its one-cycle coefficient.
Proposition 3.9. Assume that F \mathbb{F} F has characteristic zero. The two ranks of each A M , M ′ A_{M,M'} A M , M ′ at every initial interval cut are bounded by r , s r,s r , s , respectively. Moreover,
rank F ( z ) K h ( z ) = ( h + r s − 1 r s − 1 ) , B h ≤ rank F [ z ] K h ( z ) ≤ D − B h . (45) \operatorname{rank}_{\mathbb{F}(z)}K_h(z)=\binom{h+rs-1}{rs-1},\qquad B_h\leq\operatorname{rank}_{\mathbb{F}}[z]K_h(z)\leq D-B_h. \tag*{(45)} rank F ( z ) K h ( z ) = ( r s − 1 h + r s − 1 ) , B h ≤ rank F [ z ] K h ( z ) ≤ D − B h . (45)
In particular, a factorization of the full completion matrix into a left summary depending on M M M and a right summary depending on M ′ M' M ′ , with a fixed bilinear pairing over F ( z ) \mathbb{F}(z) F ( z ) , requires at least ( h + r s − 1 r s − 1 ) \binom{h+rs-1}{rs-1} ( r s − 1 h + r s − 1 ) coordinates. This is the number of degree h h h monomials in the endpoint variables of Theorem 3.6 for one color. For the one-cycle coefficient the rank has order h r + s − 2 h^{r+s-2} h r + s − 2 for fixed r , s r,s r , s .
Applications to structured matrices
4.1. For this subsection index rows and columns by 1 , … , n 1,\ldots,n 1 , … , n . Let R ( X , Y ) R(X,Y) R ( X , Y ) be a polynomial with fixed degrees in its two variables. The Cauchy theorem applies to the Hilbert matrix after adding R ( i , j ) R(i,j) R ( i , j ) .
Corollary 4.1. For B n = ( 1 / ( i + j − 1 ) + R ( i , j ) ) i , j = 1 n B_n=\left(1/(i+j-1)+R(i,j)\right)_{i,j=1}^{n} B n = ( 1/ ( i + j − 1 ) + R ( i , j ) ) i , j = 1 n , formula (26) applies with
x i = i , y j = 1 − j , P ( X , Y ) = R ( X , 1 − Y ) . x_i=i,\qquad y_j=1-j,\qquad P(X,Y)=R(X,1-Y). x i = i , y j = 1 − j , P ( X , Y ) = R ( X , 1 − Y ) .
It can be evaluated in polynomial time for fixed degrees of R R R .
For example, take R ( i , j ) = i j R(i,j)=ij R ( i , j ) = ij . Then P ( X , Y ) = X − X Y P(X,Y)=X-XY P ( X , Y ) = X − X Y , so r = s = 2 r=s=2 r = s = 2 and
H = ( 0 0 1 − 1 ) , B 2 = ( 2 5 / 2 5 / 2 13 / 3 ) . H=\begin{pmatrix}0&0\\1&-1\end{pmatrix},\qquad B_2=\begin{pmatrix}2&5/2\\5/2&13/3\end{pmatrix}. H = ( 0 1 0 − 1 ) , B 2 = ( 2 5/2 5/2 13/3 ) .
Consequently per B 2 = 179 / 12 \operatorname{per} B_2=179/12 per B 2 = 179/12 ; in order three the value is 397087 / 1080 397087/1080 397087/1080 . The same formula applies for every order.
The following rank bound explains why the Cauchy construction is needed in addition to the bounded rank-width construction.
Proposition 4.2. Let A A A be as in (21) , let n ≥ 2 n\geq2 n ≥ 2 , and put q 0 = min ( r , s ) q_0=\min(r,s) q 0 = min ( r , s ) . For every I ⊆ [ n ] I\subseteq[n] I ⊆ [ n ] ,
w A ( I ) ≥ 2 max { 0 , min ( ∣ I ∣ , n − ∣ I ∣ ) − q 0 } . w_A(I)\geq2\max\{0,\min(|I|,n-|I|)-q_0\}. w A ( I ) ≥ 2 max { 0 , min ( ∣ I ∣ , n − ∣ I ∣ ) − q 0 } .
Consequently, every rooted binary tree on [ n ] [n] [ n ] has width at least
2 max { 0 , ⌈ n / 3 ⌉ − q 0 } . 2\max\{0,\lceil n/3\rceil-q_0\}. 2 max { 0 , ⌈ n /3 ⌉ − q 0 } .
4.2.
For a square matrix B B B , its undirected support graph has edge i j ij ij , i ≠ j i \ne j i = j , if at least one of b i j , b j i b_{ij}, b_{ji} b ij , b j i is nonzero. A tree decomposition uses vertex bags on a tree. Each edge lies in a bag, and the bags containing any fixed vertex form a nonempty connected subtree. The two crossing ranks are unchanged by inversion. Thus a decomposition for B B B also gives one for B − 1 B^{-1} B − 1 .
Corollary 4.3. The following permanents can be computed in polynomial time and have the finite formulas of Theorem 3.1 or Theorem 3.4.
The permanent of B − 1 B^{-1} B − 1 , where B B B is a nonsingular banded matrix of fixed bandwidth. For a tridiagonal B B B , the calculation from B B B uses O ( n 2 ) O(n^2) O ( n 2 ) arithmetic operations.
The permanent of B − 1 B^{-1} B − 1 , where B B B is nonsingular and a tree decomposition with bags of bounded size is given for its undirected support graph.
The same rank bounds give polynomial-time cycle-polynomial formulas for these inverses. This also applies to the rational circulants in Corollary 4.5 and the banded block-constant matrices in Corollary 4.8.
4.3.
In the diagonal formulas below, letters for offsets and transfer orders are local to each family; they no longer denote the rank bounds of § 3 . For scalar and block diagonal formulas we start indices at zero. Write
P n = ( 1 j − i ≡ 1 ( m o d n ) ) , J d = ( 1 j − i = d ) , J n = ( 1 i + j = n − 1 ) . P_n=(\mathbf{1}_{j-i\equiv1\pmod n}),\qquad J_d=(\mathbf{1}_{j-i=d}),\qquad J_n=(\mathbf{1}_{i+j=n-1}). P n = ( 1 j − i ≡ 1 ( mod n ) ) , J d = ( 1 j − i = d ) , J n = ( 1 i + j = n − 1 ) .
Here 1 \mathbf{1} 1 denotes the indicator of the stated condition. The matrix P n P_n P n is the cyclic shift, J d J_d J d has ones on the ordinary diagonal of offset d d d , and right multiplication by J n J_n J n reverses columns. An open band of widths p , q p,q p , q has offsets − p , … , q -p,\ldots,q − p , … , q ; a cyclic band starting at zero has offsets 0 , … , w 0,\ldots,w 0 , … , w modulo the order. In a block band these offsets refer to block rows and columns. A Toeplitz matrix has entries t j − i t_{j-i} t j − i , and a Hankel matrix has entries h i + j h_{i+j} h i + j .
For n > w n>w n > w , write C n ( c 0 , … , c w ) = ∑ d = 0 w c d P n d C_n(c_0,\ldots,c_w)=\sum_{d=0}^{w}c_dP_n^d C n ( c 0 , … , c w ) = ∑ d = 0 w c d P n d . The two composition polynomials used below are
H m ( q 1 , … , q D ) = ∑ α ∈ N D ∑ j = 1 D j α j = m ∣ α ∣ ! α 1 ! ⋯ α D ! ∏ j = 1 D q j α j (46) H_m(q_1,\ldots,q_D)=\sum_{\substack{\alpha\in\mathbb{N}^D\\ \sum_{j=1}^{D}j\alpha_j=m}}\frac{|\alpha|!}{\alpha_1!\cdots\alpha_D!}\prod_{j=1}^{D}q_j^{\alpha_j} \tag*{(46)} H m ( q 1 , … , q D ) = α ∈ N D ∑ j = 1 D j α j = m ∑ α 1 ! ⋯ α D ! ∣ α ∣ ! j = 1 ∏ D q j α j (46)
L m ( q 1 , … , q D ) = ∑ α ∈ N D ∑ j = 1 D j α j = m m ( ∣ α ∣ − 1 ) ! α 1 ! ⋯ α D ! ∏ j = 1 D q j α j ( m ≥ 1 ) . (47) L_m(q_1,\ldots,q_D)=\sum_{\substack{\alpha\in\mathbb{N}^D\\ \sum_{j=1}^{D}j\alpha_j=m}}\frac{m(|\alpha|-1)!}{\alpha_1!\cdots\alpha_D!}\prod_{j=1}^{D}q_j^{\alpha_j}\qquad(m\ge1). \tag*{(47)} L m ( q 1 , … , q D ) = α ∈ N D ∑ j = 1 D j α j = m ∑ α 1 ! ⋯ α D ! m ( ∣ α ∣ − 1 )! j = 1 ∏ D q j α j ( m ≥ 1 ) . (47)
We put H m = 0 H_m=0 H m = 0 for m < 0 m<0 m < 0 , so H 0 = 1 H_0=1 H 0 = 1 . If λ 1 , … , λ D \lambda_1,\ldots,\lambda_D λ 1 , … , λ D are the roots of X D − ∑ j = 1 D q j X D − j X^D-\sum_{j=1}^{D}q_jX^{D-j} X D − ∑ j = 1 D q j X D − j , counted with multiplicity, then L m ( q 1 , … , q D ) = ∑ j = 1 D λ j m L_m(q_1,\ldots,q_D)=\sum_{j=1}^{D}\lambda_j^m L m ( q 1 , … , q D ) = ∑ j = 1 D λ j m . The coefficients of L m L_m L m count tilings of a labelled circle.
We use the same transfer for scalar and block diagonals. Let D 0 , … , D w D_0,\ldots,D_w D 0 , … , D w be b × b b\times b b × b matrices and let T b , w , k ( D 0 , … , D w ) T_{b,w,k}(D_0,\ldots,D_w) T b , w , k ( D 0 , … , D w ) be indexed by the k k k -subsets of { 0 , … , b w − 1 } \{0,\ldots,bw-1\} { 0 , … , b w − 1 } . Write D = [ D 0 ⋯ D w ] D=[D_0\ \cdots\ D_w] D = [ D 0 ⋯ D w ] . For states U , V U,V U , V its entry is
∑ S per D [ { 0 , … , b − 1 } , S ] , (48) \sum_S\operatorname{per}D[\{0,\ldots,b-1\},S], \tag*{(48)} S ∑ per D [{ 0 , … , b − 1 } , S ] , (48)
where S S S runs over the b b b -subsets of { 0 , … , b ( w + 1 ) − 1 } ∖ U \{0,\ldots,b(w+1)-1\}\setminus U { 0 , … , b ( w + 1 ) − 1 } ∖ U such that
{ 0 , … , b − 1 } ⊆ U ∪ S , V = { j − b : j ∈ U ∪ S , j ≥ b } . \{0,\ldots,b-1\}\subseteq U\cup S,\qquad V=\{j-b:j\in U\cup S,\ j\ge b\}. { 0 , … , b − 1 } ⊆ U ∪ S , V = { j − b : j ∈ U ∪ S , j ≥ b } .
For b = 1 b=1 b = 1 , we write T w , k T_{w,k} T w , k for this matrix. In this case a choice is a single displacement d ∉ U d\notin U d ∈ / U , of weight c d c_d c d , and
V = { j − 1 : j ∈ U ∪ { d } , j > 0 } , 0 ∈ U ∪ { d } . (49) V=\{j-1:j\in U\cup\{d\},\ j>0\},\qquad0\in U\cup\{d\}. \tag*{(49)} V = { j − 1 : j ∈ U ∪ { d } , j > 0 } , 0 ∈ U ∪ { d } . (49)
4.4.
For the signed formula, take cyclic offsets 0 , s , t , s + t 0,s,t,s+t 0 , s , t , s + t with all four residues distinct and gcd ( n , s , t ) = 1 \gcd(n,s,t)=1 g cd( n , s , t ) = 1 . The entries along each diagonal may vary with the row. For three offsets 0 , s , t 0,s,t 0 , s , t with the same gcd condition, omit the last diagonal. The allowed nonzero steps correspond to ( 1 , 0 ) , ( 0 , 1 ) , ( 1 , 1 ) (1,0),(0,1),(1,1) ( 1 , 0 ) , ( 0 , 1 ) , ( 1 , 1 ) in the lattice construction
Λ = { ( x , y ) ∈ Z 2 : s x + t y ≡ 0 ( m o d n ) } . \Lambda=\{(x,y)\in\mathbb{Z}^{2}:sx+ty\equiv0\pmod n\}. Λ = {( x , y ) ∈ Z 2 : s x + t y ≡ 0 ( mod n )} .
Put g 0 = gcd ( n , s ) g_0=\gcd(n,s) g 0 = g cd( n , s ) , h 0 = n / g 0 h_0=n/g_0 h 0 = n / g 0 , and u 0 ≡ − ( s / g 0 ) − 1 t ( m o d h 0 ) u_0\equiv-(s/g_0)^{-1}t\pmod{h_0} u 0 ≡ − ( s / g 0 ) − 1 t ( mod h 0 ) , 0 ≤ u 0 < h 0 0\leq u_0<h_0 0 ≤ u 0 < h 0 . Write b 1 = ( h 0 , 0 ) b_1=(h_0,0) b 1 = ( h 0 , 0 ) , b 2 = ( u 0 , g 0 ) b_2=(u_0,g_0) b 2 = ( u 0 , g 0 ) . For index i i i , let ρ i \rho_i ρ i be the representative of i i i under the map ( x , y ) ↦ s x + t y (x,y)\mapsto sx+ty ( x , y ) ↦ s x + t y in the rectangle 0 ≤ x < h 0 0\leq x<h_0 0 ≤ x < h 0 , 0 ≤ y < g 0 0\leq y<g_0 0 ≤ y < g 0 . For an edge i → j i\to j i → j of step δ \delta δ , write ρ i + δ − ρ j = m 1 b 1 + m 2 b 2 \rho_i+\delta-\rho_j=m_1b_1+m_2b_2 ρ i + δ − ρ j = m 1 b 1 + m 2 b 2 . For ϵ , η ∈ { 0 , 1 } \epsilon,\eta\in\{0,1\} ϵ , η ∈ { 0 , 1 } , let K ϵ η K_{\epsilon\eta} K ϵη keep the diagonal of A A A and multiply the entry on this edge by − ( − 1 ) ϵ m 1 + η m 2 -(-1)^{\epsilon m_1+\eta m_2} − ( − 1 ) ϵ m 1 + η m 2 .
For constant weights write C = a I + b P n s + c P n t + d P n s + t C=aI+bP_n^s+cP_n^t+dP_n^{s+t} C = a I + b P n s + c P n t + d P n s + t . Given r , z , v ≥ 0 r,z,v\geq0 r , z , v ≥ 0 with r + z + v ≤ n r+z+v\leq n r + z + v ≤ n , set
R = r + v , S = z + v , M = a n − r − z − v b r c z d v . R=r+v,\qquad S=z+v,\qquad M=a^{n-r-z-v}b^rc^zd^v. R = r + v , S = z + v , M = a n − r − z − v b r c z d v .
When only 0 , s , t 0,s,t 0 , s , t occur take v = 0 v=0 v = 0 . If n ∣ s R + t S n\mid sR+tS n ∣ s R + tS , put k = gcd ( R , S , ( s R + t S ) / n ) k=\gcd(R,S,(sR+tS)/n) k = g cd( R , S , ( s R + tS ) / n ) , with gcd ( 0 , 0 , 0 ) = 0 \gcd(0,0,0)=0 g cd( 0 , 0 , 0 ) = 0 .
For the residue formula, write n = b m n=bm n = bm and let C = ∑ s ∈ S a s P n s C=\sum_{s\in S}a_sP_n^s C = ∑ s ∈ S a s P n s . Take integers u , h , w u,h,w u , h , w and 0 ≤ d s ≤ w 0\leq d_s\leq w 0 ≤ d s ≤ w such that s − u ≡ d s h ( m o d m ) s-u\equiv d_sh\pmod m s − u ≡ d s h ( mod m ) . Set g = gcd ( m , h ) g=\gcd(m,h) g = g cd( m , h ) , L = m / g L=m/g L = m / g and require w < L w<L w < L . Form the blocks and transfers
D d = ∑ s : d s = d a s P b ( s − u − d h ) / m , T k = T b , w , k ( D 0 , … , D w ) . D_d=\sum_{s:d_s=d}a_sP_b^{(s-u-dh)/m},\qquad T_k=T_{b,w,k}(D_0,\ldots,D_w). D d = s : d s = d ∑ a s P b ( s − u − d h ) / m , T k = T b , w , k ( D 0 , … , D w ) .
On states, τ \tau τ replaces each position b ℓ + j b\ell+j b ℓ + j by b ℓ + ( j − h / g ( m o d b ) ) b\ell+(j-h/g\pmod b) b ℓ + ( j − h / g ( mod b )) , where 0 ≤ j < b 0\leq j<b 0 ≤ j < b . For the scalar band C n ( c 0 , … , c w ) C_n(c_0,\ldots,c_w) C n ( c 0 , … , c w ) , put
det ( I − z T w , k ) = 1 − ∑ j = 1 ( w k ) q k , j z j . \det(I-zT_{w,k})=1-\sum_{j=1}^{\binom{w}{k}}q_{k,j}z^j. det ( I − z T w , k ) = 1 − j = 1 ∑ ( k w ) q k , j z j .
For independent supported entries, let L ( A ) L(A) L ( A ) be the least number of determinants in an identity
per A = ∑ ν = 1 L ( A ) c ν det ( ε ν ∘ A ) , ε ν , i j ∈ { 1 , − 1 } . (50) \operatorname{per} A=\sum_{\nu=1}^{L(A)}c_\nu\det(\varepsilon_\nu\circ A),\qquad\varepsilon_{\nu,ij}\in\{1,-1\}. \tag*{(50)} per A = ν = 1 ∑ L ( A ) c ν det ( ε ν ∘ A ) , ε ν , ij ∈ { 1 , − 1 } . (50)
with constant coefficients c ν c_\nu c ν and constant signs, where ∘ \circ ∘ is entrywise multiplication. The additive support admits a four-determinant formula even with independent entries. The last part of the theorem compares it with another four-diagonal support.
Theorem 4.4. With the above notation, the following hold.
(1) For the additive support, or for its three-shift version,
per A = − det K 00 + det K 10 + det K 01 + det K 11 2 . (51) \operatorname{per} A=\frac{-\det K_{00}+\det K_{10}+\det K_{01}+\det K_{11}}{2}. \tag*{(51)} per A = 2 − det K 00 + det K 10 + det K 01 + det K 11 . (51) This uses O ( n 3 ) O(n^3) O ( n 3 ) arithmetic operations uniformly in s , t s,t s , t , with polynomial bit complexity for rational entries. A contribution to the monomial M M M requires n ∣ s R + t S n\mid sR+tS n ∣ s R + tS . Every permutation contributing to M M M has k k k nontrivial cycles, all with the same primitive displacement in Λ \Lambda Λ . In particular,
[ M ] per C = ( − 1 ) r + z + v − k [ M ] det C . (52) [M]\operatorname{per} C=(-1)^{r+z+v-k}[M]\det C. \tag*{(52)} [ M ] per C = ( − 1 ) r + z + v − k [ M ] det C . (52) (2) For the residue support,
per C = ( ∑ k = 0 b w ∑ ∣ U ∣ = k ( T k L ) U , τ U ) g . (53) \operatorname{per} C=\left(\sum_{k=0}^{bw}\sum_{\lvert U\rvert=k}(T_k^L)_{U,\tau U}\right)^g. \tag*{(53)} per C = k = 0 ∑ b w ∣ U ∣ = k ∑ ( T k L ) U , τ U g . (53) *For fixed b , w b,w b , w this uses O b , w ( log ( n + 1 ) ) O_{b,w}(\log(n+1)) O b , w ( log ( n + 1 )) arithmetic operations on the weights. If ∣ S ∣ \lvert S\rvert ∣ S ∣ is fixed and w ≥ 1 w\geq1 w ≥ 1 , the formula can still be evaluated in polynomial time when b w = O ( log ( n + 1 ) ) bw=O(\log(n+1)) b w = O ( log ( n + 1 )) .
(3) For n > w n>w n > w , the consecutive cyclic band satisfies
per C n ( c 0 , … , c w ) = ∑ k = 0 w tr ( T w , k n ) = ∑ k = 0 w L n ( q k , 1 , … , q k , ( w k ) ) . (54) \operatorname{per} C_n(c_0,\ldots,c_w)=\sum_{k=0}^{w}\operatorname{tr}(T_{w,k}^{n})=\sum_{k=0}^{w}\mathcal{L}_n\left(q_{k,1},\ldots,q_{k,\binom{w}{k}}\right). \tag*{(54)} per C n ( c 0 , … , c w ) = k = 0 ∑ w tr ( T w , k n ) = k = 0 ∑ w L n ( q k , 1 , … , q k , ( k w ) ) . (54)
(4) For the connected additive four-shift support, L ( A ) = 4 L(A)=4 L ( A ) = 4 . For the support { 0 , 1 , 2 , 4 } \{0,1,2,4\} { 0 , 1 , 2 , 4 } with n ≥ 6 n\geq6 n ≥ 6 ,
L ( A ) ≥ ⌊ 8 3 ⌋ ⌊ n / 5 ⌋ . L(A)\geq\left\lfloor\frac{8}{3}\right\rfloor^{\lfloor n/5\rfloor}. L ( A ) ≥ ⌊ 3 8 ⌋ ⌊ n /5 ⌋ .
Its permanent is given by the row-dependent trace formula (69) with b = 1 b=1 b = 1 and w = 4 w=4 w = 4 , using O ( n ) O(n) O ( n ) arithmetic operations.
The residue formula includes the unweighted linear-jump case with h = 1 h=1 h = 1 , u = 0 u=0 u = 0 and b , w b,w b , w fixed, as in [14 ] . For example, let n = 2 m n=2m n = 2 m , m ≥ 4 m\geq4 m ≥ 4 , and w = ⌊ log 2 m ⌋ w=\lfloor\log_2m\rfloor w = ⌊ log 2 m ⌋ . For the shifts 0 , 1 , m , m + w 0,1,m,m+w 0 , 1 , m , m + w , take b = 2 b=2 b = 2 , u = 0 u=0 u = 0 and h = 1 h=1 h = 1 . The only nonzero blocks are
D 0 = a 0 I 2 + a m P 2 , D 1 = a 1 I 2 , D w = a m + w P 2 . D_0=a_0I_2+a_mP_2,\qquad D_1=a_1I_2,\qquad D_w=a_{m+w}P_2. D 0 = a 0 I 2 + a m P 2 , D 1 = a 1 I 2 , D w = a m + w P 2 .
There are 2 2 w ≤ m 2 2^{2w}\leq m^2 2 2 w ≤ m 2 states in (53) .
For the five-diagonal formulas put
q 1 = c , q 2 = a e + b d , q 3 = a d 2 + b 2 e , q 4 = a 2 e 2 + a b d e , q 5 = − a 2 c e 2 , q 6 = − a 3 e 3 . (55) \begin{aligned}
q_1&=c,\qquad q_2=ae+bd,\qquad q_3=ad^2+b^2e,\\
q_4&=a^2e^2+abde,\qquad q_5=-a^2ce^2,\qquad q_6=-a^3e^3. \tag*{(55)}
\end{aligned} q 1 q 4 = c , q 2 = a e + b d , q 3 = a d 2 + b 2 e , = a 2 e 2 + ab d e , q 5 = − a 2 c e 2 , q 6 = − a 3 e 3 . (55)
For the positive formulas put x = a e x=ae x = a e , y = b d y=bd y = b d , h = a d 2 + b 2 e h=ad^2+b^2e h = a d 2 + b 2 e , and define
F ( z ) = c z ( 1 + x z 2 ) + x 2 z 4 + y z 2 ( 1 + x z 2 ) + h z 3 1 − x z 2 . (56) F(z)=cz(1+xz^2)+x^2z^4+\frac{yz^2(1+xz^2)+hz^3}{1-xz^2}. \tag*{(56)} F ( z ) = cz ( 1 + x z 2 ) + x 2 z 4 + 1 − x z 2 y z 2 ( 1 + x z 2 ) + h z 3 . (56)
For matrices with at most three cyclic offsets, or with four offsets of additive form, write D n ( A ) D_n(A) D n ( A ) for the following expression. Translate an offset to zero by shifting columns. For one offset use the product of its entries. For two offsets 0 , s 0,s 0 , s use
∏ O ( ∏ i ∈ O a i i + ∏ i ∈ O a i , i + s ) , (57) \prod_{O}\left(\prod_{i\in O}a_{ii}+\prod_{i\in O}a_{i,i+s}\right), \tag*{(57)} O ∏ ( i ∈ O ∏ a ii + i ∈ O ∏ a i , i + s ) , (57)
where O O O runs over the orbits of addition by s s s . For three or four offsets 0 , s , t 0,s,t 0 , s , t or 0 , s , t , s + t 0,s,t,s+t 0 , s , t , s + t , use the product of (51) on the classes modulo gcd ( n , s , t ) \gcd(n,s,t) g cd( n , s , t ) , with the divided offsets in each class.
For constant weights in the connected additive case define D ϵ η D_{\epsilon\eta} D ϵη in an algebraic closure of F \mathbb{F} F by
D ϵ η = ∏ X h 0 = ( − 1 ) ϵ X u 0 Y g 0 = ( − 1 ) η ( a − b X − c Y − d X Y ) . (58) D_{\epsilon\eta}=
\prod_{\substack{X^{h_0}=(-1)^\epsilon\\X^{u_0}Y^{g_0}=(-1)^\eta}}
(a-bX-cY-dXY). \tag*{(58)} D ϵη = X h 0 = ( − 1 ) ϵ X u 0 Y g 0 = ( − 1 ) η ∏ ( a − b X − c Y − d X Y ) . (58)
In the three-shift case put d = 0 d=0 d = 0 . Put f ϵ = X h 0 − ( − 1 ) ϵ f_\epsilon=X^{h_0}-(-1)^\epsilon f ϵ = X h 0 − ( − 1 ) ϵ and let R ϵ η R_{\epsilon\eta} R ϵη be the remainder of ( a − b X ) g 0 − ( − 1 ) η X − u 0 ( c + d X ) g 0 (a-bX)^{g_0}-(-1)^\eta X^{-u_0}(c+dX)^{g_0} ( a − b X ) g 0 − ( − 1 ) η X − u 0 ( c + d X ) g 0 modulo f ϵ f_\epsilon f ϵ , using the inverse of X X X in this quotient. Over C \mathbb{C} C , set ζ = exp ( π i / n ) \zeta=\exp(\pi i/n) ζ = exp ( π i / n ) , x ϵ = ζ ϵ g 0 x_\epsilon=\zeta^{\epsilon g_0} x ϵ = ζ ϵ g 0 and y ϵ η = ζ η h 0 − ϵ u 0 y_{\epsilon\eta}=\zeta^{\eta h_0-\epsilon u_0} y ϵη = ζ η h 0 − ϵ u 0 . The pairs in the product are ( x ϵ ζ 2 s j , y ϵ η ζ 2 t j ) (x_\epsilon\zeta^{2sj},y_{\epsilon\eta}\zeta^{2tj}) ( x ϵ ζ 2 s j , y ϵη ζ 2 t j ) for 0 ≤ j < n 0\leq j<n 0 ≤ j < n .
The corollary gives explicit low-width formulas, the product over cyclic components, and the product and resultant forms for constant weights.
Corollary 4.5.
For n ≥ 3 , 4 , 5 n \ge3,4,5 n ≥ 3 , 4 , 5 , respectively,
per C n ( a , b , c ) = a n + c n + ∑ j = 0 ⌊ n / 2 ⌋ n n − j ( n − j j ) ( a c ) j b n − 2 j . (59) \operatorname{per} C_n(a,b,c)=a^n+c^n+\sum_{j=0}^{\lfloor n/2\rfloor}\frac{n}{n-j}\binom{n-j}{j}(ac)^j b^{n-2j}. \tag*{(59)} per C n ( a , b , c ) = a n + c n + j = 0 ∑ ⌊ n /2 ⌋ n − j n ( j n − j ) ( a c ) j b n − 2 j . (59) per C n ( a , b , c , d ) = a n + d n + L n ( b , a c , a 2 d ) + L n ( c , b d , a d 2 ) . (60) \operatorname{per} C_n(a,b,c,d)=a^n+d^n+\mathcal{L}_n(b,ac,a^2d)+\mathcal{L}_n(c,bd,ad^2). \tag*{(60)} per C n ( a , b , c , d ) = a n + d n + L n ( b , a c , a 2 d ) + L n ( c , b d , a d 2 ) . (60) per C n ( a , b , c , d , e ) = a n + e n + L n ( b , a c , a 2 d , a 3 e ) + L n ( q 1 , … , q 6 ) + L n ( d , c e , b e 2 , a e 3 ) . (61) \begin{aligned}
\operatorname{per} C_n(a,b,c,d,e)={}&a^n+e^n+\mathcal{L}_n(b,ac,a^2d,a^3e)\\
&+\mathcal{L}_n(q_1,\ldots,q_6)+\mathcal{L}_n(d,ce,be^2,ae^3). \tag*{(61)}
\end{aligned} per C n ( a , b , c , d , e ) = a n + e n + L n ( b , a c , a 2 d , a 3 e ) + L n ( q 1 , … , q 6 ) + L n ( d , ce , b e 2 , a e 3 ) . (61) The expanded four-diagonal formula is
per C n ( a , b , c , d ) = a n + d n + ∑ u , v , w ∈ N u + 2 v + 3 w = n n ( u + v + w − 1 ) ! u ! v ! w ! ( a v + 2 w b u c v d w + a w b v c u d v + 2 w ) . (62) \operatorname{per} C_n(a,b,c,d)=a^n+d^n+\sum_{\substack{u,v,w\in\mathbb{N}\\u+2v+3w=n}}\frac{n(u+v+w-1)!}{u!v!w!}\left(a^{v+2w}b^u c^v d^w+a^w b^v c^u d^{v+2w}\right). \tag*{(62)} per C n ( a , b , c , d ) = a n + d n + u , v , w ∈ N u + 2 v + 3 w = n ∑ u ! v ! w ! n ( u + v + w − 1 )! ( a v + 2 w b u c v d w + a w b v c u d v + 2 w ) . (62) The middle term for five diagonals has the positive form
L n ( q 1 , … , q 6 ) = 21 2 ∣ n x ⌊ n / 2 ⌋ + n ∑ j = 1 n [ z n ] F ( z ) j j . (63) \mathcal{L}_n(q_1,\ldots,q_6)=21_{2\mid n}x^{\lfloor n/2\rfloor}+n\sum_{j=1}^{n}\frac{[z^n]F(z)^j}{j}. \tag*{(63)} L n ( q 1 , … , q 6 ) = 2 1 2 ∣ n x ⌊ n /2 ⌋ + n j = 1 ∑ n j [ z n ] F ( z ) j . (63) For g = gcd ( n , h ) g=\gcd(n,h) g = g cd( n , h ) and n / g > w n/g>w n / g > w ,
per ( ∑ d = 0 w c d P n u + d h ) = ( per C n / g ( c 0 , … , c w ) ) g . (64) \operatorname{per}\left(\sum_{d=0}^{w}c_dP_n^{u+dh}\right)=\left(\operatorname{per}C_{n/g}(c_0,\ldots,c_w)\right)^g. \tag*{(64)} per ( d = 0 ∑ w c d P n u + d h ) = ( per C n / g ( c 0 , … , c w ) ) g . (64) For n = 2 m n=2m n = 2 m , m > 2 m>2 m > 2 and gcd ( m , h ) = 1 \gcd(m,h)=1 g cd( m , h ) = 1 , (53) for a I + b P n h + c P n 2 h + d P n m aI+bP_n^h+cP_n^{2h}+dP_n^m a I + b P n h + c P n 2 h + d P n m uses D 0 = a I 2 + d P 2 D_0=aI_2+dP_2 D 0 = a I 2 + d P 2 , D 1 = b I 2 D_1=bI_2 D 1 = b I 2 , D 2 = c I 2 D_2=cI_2 D 2 = c I 2 . There are 16 states, and τ \tau τ switches the two labels in a block if h h h is odd.
In the cases defining D n ( A ) D_n(A) D n ( A ) , per A = D n ( A ) \operatorname{per} A=D_n(A) per A = D n ( A ) . In particular, for constant weights and g = gcd ( n , s , t ) g=\gcd(n,s,t) g = g cd( n , s , t ) , the component factors are equal:
per C = [ D n / g ( a I n / g + b P n / g s / g + c P n / g t / g + d P n / g ( s + t ) / g ) ] g . \operatorname{per} C=\left[D_{n/g}\left(aI_{n/g}+bP_{n/g}^{s/g}+cP_{n/g}^{t/g}+dP_{n/g}^{(s+t)/g}\right)\right]^g. per C = [ D n / g ( a I n / g + b P n / g s / g + c P n / g t / g + d P n / g ( s + t ) / g ) ] g . In (51) , for constant weights one may replace det K ϵ η \det K_{\epsilon\eta} det K ϵη by D ϵ η D_{\epsilon\eta} D ϵη . Also
D ϵ η = Res X ( f ϵ , R ϵ η ) , (65) D_{\epsilon\eta}=\operatorname{Res}_X(f_\epsilon,R_{\epsilon\eta}), \tag*{(65)} D ϵη = Res X ( f ϵ , R ϵη ) , (65) This uses O ( h 0 2 log ( g 0 + 1 ) ) O(h_0^2\log(g_0+1)) O ( h 0 2 log ( g 0 + 1 )) operations in F \mathbb{F} F . For s = 1 s=1 s = 1 , t = q t=q t = q , 2 ≤ q ≤ n − 2 2\le q\le n-2 2 ≤ q ≤ n − 2 , with z ϵ j = exp ( π i ( 2 j + ϵ ) / n ) z_{\epsilon j}=\exp(\pi i(2j+\epsilon)/n) z ϵ j = exp ( π i ( 2 j + ϵ ) / n ) , the product is
D ϵ η = ∏ j = 0 n − 1 ( a − b z ϵ j − ( − 1 ) ϵ + η c z ϵ j q − ( − 1 ) ϵ + η d z ϵ j q + 1 ) . D_{\epsilon\eta}=\prod_{j=0}^{n-1}\left(a-bz_{\epsilon j}-(-1)^{\epsilon+\eta}cz_{\epsilon j}^{q}-(-1)^{\epsilon+\eta}dz_{\epsilon j}^{q+1}\right). D ϵη = j = 0 ∏ n − 1 ( a − b z ϵ j − ( − 1 ) ϵ + η c z ϵ j q − ( − 1 ) ϵ + η d z ϵ j q + 1 ) . One may use ( − 1 ) η (-1)^\eta ( − 1 ) η in place of ( − 1 ) ϵ + η (-1)^{\epsilon+\eta} ( − 1 ) ϵ + η in all four products.
For Laurent polynomials f , g f,g f , g supported in fixed intervals and g ( P n ) g(P_n) g ( P n ) nonsingular, the permanent of f ( P n ) g ( P n ) − 1 f(P_n)g(P_n)^{-1} f ( P n ) g ( P n ) − 1 is the determinant (31) with bounded cut ranks. If P n P_n P n is replaced by P n h P_n^h P n h , use the formula in order n / gcd ( n , h ) n/\gcd(n,h) n / g cd( n , h ) and raise it to the power gcd ( n , h ) \gcd(n,h) g cd( n , h ) . In particular,
per ( I − ρ P n ) − 1 = ∑ k = 0 n − 1 ⟨ n k ⟩ ρ n k ( 1 − ρ n ) n , ρ n ≠ 1 , (66) \operatorname{per}(I-\rho P_n)^{-1}=\frac{\sum_{k=0}^{n-1}\left\langle\begin{matrix}n\\k\end{matrix}\right\rangle\rho^{nk}}{(1-\rho^n)^n},\qquad\rho^n\ne1, \tag*{(66)} per ( I − ρ P n ) − 1 = ( 1 − ρ n ) n ∑ k = 0 n − 1 ⟨ n k ⟩ ρ nk , ρ n = 1 , (66) Here ⟨ n k ⟩ \left\langle\begin{matrix}n\\k\end{matrix}\right\rangle ⟨ n k ⟩ is the Eulerian number, which equals ∑ j = 0 k + 1 ( − 1 ) j ( n + 1 j ) ( k + 1 − j ) n \sum_{j=0}^{k+1}(-1)^j\binom{n+1}{j}(k+1-j)^n ∑ j = 0 k + 1 ( − 1 ) j ( j n + 1 ) ( k + 1 − j ) n .
Apart from a n , d n a^n,d^n a n , d n , the coefficient of a monomial a α 0 b α 1 c α 2 d α 3 a^{\alpha_0}b^{\alpha_1}c^{\alpha_2}d^{\alpha_3} a α 0 b α 1 c α 2 d α 3 of degree n n n in (60) , with m = α 1 + 2 α 2 + 3 α 3 m=\alpha_1+2\alpha_2+3\alpha_3 m = α 1 + 2 α 2 + 3 α 3 , is
{ n ( α 1 + α 2 + α 3 − 1 ) ! α 1 ! α 2 ! α 3 ! , m = n , n ( α 0 + α 1 + α 2 − 1 ) ! α 0 ! α 1 ! α 2 ! , m = 2 n , 0 , otherwise . \begin{cases}
\dfrac{n(\alpha_1+\alpha_2+\alpha_3-1)!}{\alpha_1!\alpha_2!\alpha_3!}, & m=n,\\
\dfrac{n(\alpha_0+\alpha_1+\alpha_2-1)!}{\alpha_0!\alpha_1!\alpha_2!}, & m=2n,\\
0, & \text{otherwise}.
\end{cases} ⎩ ⎨ ⎧ α 1 ! α 2 ! α 3 ! n ( α 1 + α 2 + α 3 − 1 )! , α 0 ! α 1 ! α 2 ! n ( α 0 + α 1 + α 2 − 1 )! , 0 , m = n , m = 2 n , otherwise .
For example, [ a 4 b 2 d 2 ] per C 8 ( a , b , c , d ) = 12 [a^4b^2d^2]\operatorname{per} C_8(a,b,c,d)=12 [ a 4 b 2 d 2 ] per C 8 ( a , b , c , d ) = 12 . The six- and seven-diagonal sector orders are ( 1 , 5 , 10 , 10 , 5 , 1 ) (1,5,10,10,5,1) ( 1 , 5 , 10 , 10 , 5 , 1 ) and ( 1 , 6 , 15 , 20 , 15 , 6 , 1 ) (1,6,15,20,15,6,1) ( 1 , 6 , 15 , 20 , 15 , 6 , 1 ) , respectively. In general the total number of states is 2 w 2^w 2 w and the largest sector has order ( w ⌊ w / 2 ⌋ ) \binom{w}{\lfloor w/2\rfloor} ( ⌊ w /2 ⌋ w ) .
4.5.
Let G G G be a finite abelian group, and let S S S be a nonempty subset of G G G . Write P s P_s P s for the translation matrix with a one at ( i , i + s ) (i,i+s) ( i , i + s ) and put C S = ∑ s ∈ S a s P s C_S=\sum_{s\in S}a_sP_s C S = ∑ s ∈ S a s P s , where the a s a_s a s are independent. Put H = ⟨ S − S ⟩ H=\langle S-S\rangle H = ⟨ S − S ⟩ , m = ∣ H ∣ m=|H| m = ∣ H ∣ and g = [ G : H ] g=[G:H] g = [ G : H ] . Translating S S S into H H H and ordering by the cosets of H H H gives g g g equal blocks. We write ∣ det C S ∣ c o e f f |\det C_S|_{\mathrm{coeff}} ∣ det C S ∣ coeff for the polynomial obtained by taking the absolute values of its integer coefficients.
Equality with the permanent holds exactly when all permutations with the same shift multiplicities have the same sign.
Theorem 4.6. The equality per C S = ∣ det C S ∣ c o e f f \operatorname{per} C_S=|\det C_S|_{\mathrm{coeff}} per C S = ∣ det C S ∣ coeff holds in precisely the following cases.
∣ S ∣ ≤ 2 |S|\le2 ∣ S ∣ ≤ 2 .
∣ S ∣ = 3 |S|=3 ∣ S ∣ = 3 and one of the following holds.
g = 1 g=1 g = 1 .
S S S is an arithmetic progression, with either m m m even or g = 2 g=2 g = 2 .
After translation and an isomorphism H ≃ Z 7 H\simeq\mathbb{Z}_7 H ≃ Z 7 , the support is { 0 , 1 , 3 } \{0,1,3\} { 0 , 1 , 3 } .
∣ S ∣ = 4 |S|=4 ∣ S ∣ = 4 , g = 1 g=1 g = 1 , and the four elements partition into two pairs with the same sum.
In particular, for connected cyclic supports, the equality holds for every support of size at most three and for exactly the additive supports of size four. For C = a I 6 + b P 6 + c P 6 2 + d P 6 4 C=aI_6+bP_6+cP_6^2+dP_6^4 C = a I 6 + b P 6 + c P 6 2 + d P 6 4 , for example,
[ a b 2 c d 2 ] per C = 12 , [ a b 2 c d 2 ] det C = 0. [ab^2cd^2]\operatorname{per} C=12,\qquad[ab^2cd^2]\det C=0. [ a b 2 c d 2 ] per C = 12 , [ a b 2 c d 2 ] det C = 0.
The permutations ( 0 , 2 , 4 , 1 , 5 , 3 ) (0,2,4,1,5,3) ( 0 , 2 , 4 , 1 , 5 , 3 ) and ( 0 , 5 , 3 , 4 , 2 , 1 ) (0,5,3,4,2,1) ( 0 , 5 , 3 , 4 , 2 , 1 ) , in one-line notation with indices starting at zero, give opposite signs at this monomial.
4.6.
Consider first a cyclic block matrix M M M with m ≥ 2 m\ge2 m ≥ 2 block rows and columns of size b ≥ 2 b\ge2 b ≥ 2 . Its only supported blocks are A i A_i A i on the diagonal and B i B_i B i from i i i to i + 1 i+1 i + 1 modulo m m m . Each supported block has two consecutive internal cyclic offsets. For such a block D D D , take its offsets to be d , d + 1 d,d+1 d , d + 1 with 0 ≤ d < b 0\le d<b 0 ≤ d < b . Write D ( η ) D^{(\eta)} D ( η ) for the matrix that multiplies an entry of offset s s s in row j j j by η \eta η if j + s ≥ b j+s\ge b j + s ≥ b , leaving the other entries unchanged. Put E = diag ( ( − 1 ) j ) j = 0 b − 1 E=\operatorname{diag}((-1)^j)_{j=0}^{b-1} E = diag (( − 1 ) j ) j = 0 b − 1 . For ϵ = ± 1 \epsilon=\pm1 ϵ = ± 1 let F ϵ ( x ) F_\epsilon(x) F ϵ ( x ) be the determinant of the block array obtained by the replacements
A i ⟼ A i ( ( − 1 ) b ϵ ) , B i ⟼ E B i ( ϵ ) E , A_i\longmapsto A_i^{((-1)^{b}\epsilon)},\qquad B_i\longmapsto EB_i^{(\epsilon)}E, A i ⟼ A i (( − 1 ) b ϵ ) , B i ⟼ E B i ( ϵ ) E ,
with the last forward block also multiplied by ( − 1 ) m − 1 x (-1)^{m-1}x ( − 1 ) m − 1 x .
For an open block band with offsets − p , … , q -p,\ldots,q − p , … , q put w = p + q w=p+q w = p + q and B = { 0 , … , b p − 1 } B=\{0,\ldots,bp-1\} B = { 0 , … , b p − 1 } . Use T i = T b , w , b p T_i=T_{b,w,bp} T i = T b , w , b p on the blocks of block row i i i in order of increasing offset, filling out-of-range blocks with zeros. For a cyclic block band with offsets 0 , … , w 0,\ldots,w 0 , … , w , w < m w<m w < m , let T i , k T_{i,k} T i , k be the transfer on k k k -subsets from row i i i . For the block-constant case, take r r r row classes of sizes m i m_i m i and s s s column classes of sizes n j n_j n j , and let the entries of block i j ij ij be b i j b_{ij} b ij . The theorem gives four determinants for the first class and finite transfer and sum formulas for the other classes.
Theorem 4.7. (1) For the cyclic block matrix M M M ,
per M = F + ( 1 ) − F + ( − 1 ) + F − ( 1 ) + F − ( − 1 ) 2 . (67) \operatorname{per} M = \frac{F_{+}(1)-F_{+}(-1)+F_{-}(1)+F_{-}(-1)}{2}. \tag*{(67)} per M = 2 F + ( 1 ) − F + ( − 1 ) + F − ( 1 ) + F − ( − 1 ) . (67) The block size and the number of blocks may both grow.
(2) The open and cyclic block bands, respectively, satisfy
per A = ( T 0 ⋯ T m − 1 ) B , B . (68) \operatorname{per} A = (T_0\cdots T_{m-1})_{B,B}. \tag*{(68)} per A = ( T 0 ⋯ T m − 1 ) B , B . (68) per A = ∑ k = 0 b w tr ( T 0 , k ⋯ T m − 1 , k ) . (69) \operatorname{per} A = \sum_{k=0}^{bw}\operatorname{tr}(T_{0,k}\cdots T_{m-1,k}). \tag*{(69)} per A = k = 0 ∑ b w tr ( T 0 , k ⋯ T m − 1 , k ) . (69) (3) For the block-constant matrix A A A ,
per A = ( ∏ i m i ! ) ( ∏ j n j ! ) ∑ k i j ≥ 0 ∑ j k i j = m i , ∑ i k i j = n j ∏ i , j b i j k i j k i j ! . (70) \operatorname{per} A =
\left(\prod_i m_i!\right)
\left(\prod_j n_j!\right)
\sum_{\substack{k_{ij}\geq0\\ \sum_j k_{ij}=m_i,\ \sum_i k_{ij}=n_j}}
\prod_{i,j}\frac{b_{ij}^{k_{ij}}}{k_{ij}!}. \tag*{(70)} per A = ( i ∏ m i ! ) ( j ∏ n j ! ) k ij ≥ 0 ∑ j k ij = m i , ∑ i k ij = n j ∑ i , j ∏ k ij ! b ij k ij . (70) For fixed r , s r,s r , s , at most ( n + 1 ) ( r − 1 ) ( s − 1 ) (n+1)^{(r-1)(s-1)} ( n + 1 ) ( r − 1 ) ( s − 1 ) arrays occur in the sum.
For a block Toeplitz band the transfer is the same matrix T T T at every row, using the untruncated block entries. Let T ′ T' T ′ omit state B B B . A block circulant band has common sector matrices T k T_k T k . The corollary gives the resulting coefficient and trace formulas.
Corollary 4.8. For a block Toeplitz band,
per A = [ z m ] det ( I − z T ′ ) det ( I − z T ) . (71) \operatorname{per} A = [z^m]\frac{\det(I-zT')}{\det(I-zT)}. \tag*{(71)} per A = [ z m ] det ( I − z T ) det ( I − z T ′ ) . (71) Its order is ( b ( p + q ) b p ) \binom{b(p+q)}{bp} ( b p b ( p + q ) ) , which is polynomial in m m m for fixed b , p b,p b , p , even when q q q is as large as m − 1 m-1 m − 1 . For a block circulant band with common transfers T k T_k T k , the permanent is ∑ k tr ( T k m ) \sum_k\operatorname{tr}(T_k^m) ∑ k tr ( T k m ) . The orders and the number of transfers are fixed when b , w b,w b , w are fixed. A cyclic shift of the block columns allows any common starting offset. Reversing the block columns of a block Hankel matrix gives the block Toeplitz formula.
For a block circulant with just two cyclic block offsets, translate one offset to zero and order the block indices along the cycles of the remaining offset. When its supported blocks satisfy Theorem 4.7 (1), use (4.22) on each cycle of length at least two and multiply the results.
For a block-constant matrix with a fixed band of nonzero blocks on a common ordered partition, (31) gives a determinant of polynomial order, even when the number and the sizes of the blocks vary.
4.7.
For an open band with widths p , q p,q p , q , let w = p + q w=p+q w = p + q and B = { 0 , … , p − 1 } B=\{0,\ldots,p-1\} B = { 0 , … , p − 1 } . Form T i = T w , p T_i=T_{w,p} T i = T w , p with weights c d = a i , i + d − p c_d=a_{i,i+d-p} c d = a i , i + d − p , using zero for out-of-range entries. For A n = ∑ d = − p q t d J d A_n=\sum_{d=-p}^{q}t_dJ_d A n = ∑ d = − p q t d J d , use T = T w , p T=T_{w,p} T = T w , p with the constant weights t d − p t_{d-p} t d − p , of order D = ( p + q p ) D=\binom{p+q}{p} D = ( p p + q ) . Let T ′ T' T ′ omit state B B B and write
det ( I − z T ) = 1 − ∑ j = 1 D q j z j , det ( I − z T ′ ) = ∑ j = 0 D − 1 r j z j . \det(I-zT)=1-\sum_{j=1}^{D}q_jz^j,\qquad
\det(I-zT')=\sum_{j=0}^{D-1}r_jz^j. det ( I − z T ) = 1 − j = 1 ∑ D q j z j , det ( I − z T ′ ) = j = 0 ∑ D − 1 r j z j .
For a Toeplitz sequence with rational diagonal series, write
∑ ℓ ≥ 1 t ℓ z ℓ − 1 , ∑ ℓ ≥ 1 t − ℓ z ℓ − 1 . (72) \sum_{\ell\geq1}t_{\ell}z^{\ell-1},\qquad
\sum_{\ell\geq1}t_{-\ell}z^{\ell-1}. \tag*{(72)} ℓ ≥ 1 ∑ t ℓ z ℓ − 1 , ℓ ≥ 1 ∑ t − ℓ z ℓ − 1 . (72)
If the numerator and denominator degrees are bounded, we use constant generators t ℓ = p T B ℓ − 1 q t_\ell=p^{T}B^{\ell-1}q t ℓ = p T B ℓ − 1 q and t − ℓ = u T C ℓ − 1 v t_{-\ell}=u^{T}C^{\ell-1}v t − ℓ = u T C ℓ − 1 v for ℓ ≥ 1 \ell\geq1 ℓ ≥ 1 . For each order n n n , let T n T_n T n be the common matrix W i W_i W i of (29) , with d i = t 0 d_i=t_0 d i = t 0 and the degree bound ⌊ n / 2 ⌋ \lfloor n/2\rfloor ⌊ n /2 ⌋ . Let T n ′ T'_n T n ′ omit the constant monomial. The matrix T n T_n T n may grow with n n n ; the fixed-band matrix T T T above does not.
In characteristic p > 0 p>0 p > 0 , use instead the common matrix T p = W i ( p ) T_p=W_i^{(p)} T p = W i ( p ) of Corollary 3.5, with the same constant generators. Its order is D p , r , s D_{p,r,s} D p , r , s , and it is independent of n n n . Write T p ′ T'_p T p ′ for the matrix with the constant row and column deleted. The p p p in this notation is the characteristic, unrelated to an open bandwidth.
There is also a characteristic-zero family not requiring rational diagonal series. For a univariate polynomial P P P , use the one-based indices 1 , … , n 1,\ldots,n 1 , … , n and put
C n T = ( 1 α + j − i + P ( j − i ) ) i , j = 1 n , C_n^T=\left(\frac{1}{\alpha+j-i}+P(j-i)\right)_{i,j=1}^{n}, C n T = ( α + j − i 1 + P ( j − i ) ) i , j = 1 n ,
assuming that every denominator is nonzero. In the Cauchy formula take x i = − i x_i=-i x i = − i , y j = − α − j y_j=-\alpha-j y j = − α − j and the bivariate polynomial P ( X − Y − α ) P(X-Y-\alpha) P ( X − Y − α ) . A degree bound d d d for P P P permits r = s = d + 1 r=s=d+1 r = s = d + 1 .
For the padding formula, let S S S be the ordinary offsets of a matrix A A A of order n n n . Choose N ≥ n N\ge n N ≥ n with N = n N=n N = n or 0 ∈ S 0\in S 0 ∈ S , so that S S S modulo N N N has at most three elements or is an additive four-set. The cyclic formula D N D_N D N is given in Corollary 4.5 . For the direct signing, take offsets 0 , p , q , p + q 0,p,q,p+q 0 , p , q , p + q with p > 0 > q p>0>q p > 0 > q and p + q ≠ 0 p+q\ne0 p + q = 0 . Let g = gcd ( p , q ) g=\gcd(p,q) g = g cd( p , q ) , choose x , y ∈ { 1 , − 1 } x,y\in\{1,-1\} x , y ∈ { 1 , − 1 } with x q / g y − p / g = − 1 x^{q/g}y^{-p/g}=-1 x q / g y − p / g = − 1 , and let K K K be obtained by multiplying the entries on p , q , p + q p,q,p+q p , q , p + q by − x , − y , − x y -x,-y,-xy − x , − y , − x y , respectively. The next theorem collects the open-diagonal and dense Toeplitz formulas.
Theorem 4.9. With the above notation, the following hold.
(1) The open band and the corresponding Toeplitz band satisfy
per A = ( T 0 ⋯ T n − 1 ) B , B . (73) \operatorname{per} A=(T_0\cdots T_{n-1})_{B,B}. \tag*{(73)} per A = ( T 0 ⋯ T n − 1 ) B , B . (73) per A n = ( T n ) B , B = [ z n ] det ( I − z T ′ ) det ( I − z T ) = ∑ j = 0 D − 1 r j H n − j ( q 1 , … , q D ) . (74) \operatorname{per} A_n=(T^n)_{B,B}=[z^n]\frac{\det(I-zT')}{\det(I-zT)}=\sum_{j=0}^{D-1}r_j\mathcal{H}_{n-j}(q_1,\ldots,q_D). \tag*{(74)} per A n = ( T n ) B , B = [ z n ] det ( I − z T ) det ( I − z T ′ ) = j = 0 ∑ D − 1 r j H n − j ( q 1 , … , q D ) . (74) For fixed diagonal weights the generating function is
∑ n ≥ 0 per A n z n = det ( I − z T ′ ) det ( I − z T ) . (75) \sum_{n\ge0}\operatorname{per} A_nz^n=\frac{\det(I-zT')}{\det(I-zT)}. \tag*{(75)} n ≥ 0 ∑ per A n z n = det ( I − z T ) det ( I − z T ′ ) . (75) At fixed w w w the power takes O ( log ( n + 1 ) ) O(\log(n+1)) O ( log ( n + 1 )) arithmetic operations. At fixed p p p the order is polynomial in n n n for q ≤ n − 1 q\le n-1 q ≤ n − 1 , and (4.28) takes O p ( n p + 1 ) O_p(n^{p+1}) O p ( n p + 1 ) operations. Transposition gives the corresponding bound with the upper bandwidth fixed. For growing q q q , the stated bound uses (4.28), without expanding H n − j \mathcal{H}_{n-j} H n − j into its individual summands.
(2) For the rational diagonal series,
per A n = [ z n ] det ( I − z T n ′ ) det ( I − z T n ) . (76) \operatorname{per} A_n=[z^n]\frac{\det(I-zT'_n)}{\det(I-zT_n)}. \tag*{(76)} per A n = [ z n ] det ( I − z T n ) det ( I − z T n ′ ) . (76) The determinant orders are polynomial in n n n when the degrees in (4.27) are bounded. This is a coefficient formula for the specified n n n ; unlike (4.30), its rational function may depend on n n n .
(3) With the stated padding,
per A = D N ( A ⊕ I N − n ) . (77) \operatorname{per} A=D_N(A\oplus I_{N-n}). \tag*{(77)} per A = D N ( A ⊕ I N − n ) . (77) Any three ordinary diagonals are covered with N = n N=n N = n . For four offsets d 1 < d 2 < d 3 < d 4 d_1<d_2<d_3<d_4 d 1 < d 2 < d 3 < d 4 that include zero, one may take N = d 3 + d 4 − d 1 − d 2 N=d_3+d_4-d_1-d_2 N = d 3 + d 4 − d 1 − d 2 if N ≥ n N\ge n N ≥ n . Both applications use O ( n 3 ) O(n^3) O ( n 3 ) arithmetic operations.
(4) For the additive ordinary offsets,
per A = det K . (78) \operatorname{per} A=\det K. \tag*{(78)} per A = det K . (78) The cost is O ( n 3 ) O(n^3) O ( n 3 ) arithmetic operations uniformly in p , q p,q p , q .
(5) In characteristic p > 0 p > 0 p > 0 , rational diagonal series give
∑ n ≥ 0 per A n z n = det ( I − z T p ′ ) det ( I − z T p ) . (79) \sum_{n\geq0}\operatorname{per} A_n z^n = \frac{\det(I-zT'_p)}{\det(I-zT_p)}. \tag*{(79)} n ≥ 0 ∑ per A n z n = det ( I − z T p ) det ( I − z T p ′ ) . (79)
For fixed p , r , s p,r,s p , r , s , a specified value per A n = ( T p n ) 1 , 1 \operatorname{per} A_n=(T_p^n)_{1,1} per A n = ( T p n ) 1 , 1 can be evaluated in O ( log ( n + 1 ) ) O(\log(n+1)) O ( log ( n + 1 )) field operations after T p T_p T p has been formed.
(6) In characteristic zero, per C n T \operatorname{per} C_n^T per C n T is given by (3.9) with the nodes and polynomial specified above. For fixed degree of P P P , this is a polynomial-time determinant formula, also in bit complexity for rational input.
For row weights with period h h h before truncation, put M = T 0 ⋯ T h − 1 M=T_0\cdots T_{h-1} M = T 0 ⋯ T h − 1 , N r = T 0 ⋯ T r − 1 N_r=T_0\cdots T_{r-1} N r = T 0 ⋯ T r − 1 and N 0 = I N_0=I N 0 = I . For n = h k + r n=hk+r n = hk + r , 0 ≤ r < h 0\leq r<h 0 ≤ r < h , write
det ( I − z M ) = 1 − ∑ j = 1 D q j z j , e B T adj ( I − z M ) N r e B = ∑ j = 0 D − 1 r r , j z j . \det(I-zM)=1-\sum_{j=1}^{D}q_jz^j,\qquad e_B^{\mathsf T}\operatorname{adj}(I-zM)N_re_B=\sum_{j=0}^{D-1}r_{r,j}z^j. det ( I − z M ) = 1 − j = 1 ∑ D q j z j , e B T adj ( I − z M ) N r e B = j = 0 ∑ D − 1 r r , j z j .
where e B e_B e B is the vector of state B B B .
For A n = a J − 2 + b J − 1 + c I + d J 1 + e J 2 A_n=aJ_{-2}+bJ_{-1}+cI+dJ_1+eJ_2 A n = a J − 2 + b J − 1 + c I + d J 1 + e J 2 , use q j , x , y , h , F q_j,x,y,h,F q j , x , y , h , F from (4.10) and (4.11). Put B ( 0 , 0 ) = 1 B(0,0)=1 B ( 0 , 0 ) = 1 , B ( 0 , v ) = 0 B(0,v)=0 B ( 0 , v ) = 0 for v > 0 v>0 v > 0 , and B ( s , v ) = ( s + v − 1 v ) B(s,v)=\binom{s+v-1}{v} B ( s , v ) = ( v s + v − 1 ) for s > 0 s>0 s > 0 . In the symmetric case a = e = v a=e=v a = e = v , b = d = u b=d=u b = d = u , c = t c=t c = t , put
( s 1 , … , s 5 ) = ( t + v , u 2 − t v , v u 2 + t v 2 , v 4 − t v 3 , − v 5 ) . (s_1,\ldots,s_5)=(t+v,\ u^2-tv,\ vu^2+tv^2,\ v^4-tv^3,\ -v^5). ( s 1 , … , s 5 ) = ( t + v , u 2 − t v , v u 2 + t v 2 , v 4 − t v 3 , − v 5 ) .
For arbitrary tridiagonal entries d i , u i , l i d_i,u_i,l_i d i , u i , l i on the diagonal, superdiagonal and subdiagonal, put x i = u i l i x_i=u_il_i x i = u i l i . Let E E E range over the nonconsecutive subsets of { 1 , … , n − 1 } \{1,\ldots,n-1\} { 1 , … , n − 1 } , and let V ( E ) = ⋃ i ∈ E { i , i + 1 } V(E)=\bigcup_{i\in E}\{i,i+1\} V ( E ) = ⋃ i ∈ E { i , i + 1 } . For the separated diagonals − r , 0 , r -r,0,r − r , 0 , r set ℓ s = ⌊ ( n − s ) / r ⌋ + 1 \ell_s=\lfloor(n-s)/r\rfloor+1 ℓ s = ⌊( n − s ) / r ⌋ + 1 for 1 ≤ s ≤ min ( r , n ) 1\leq s\leq\min(r,n) 1 ≤ s ≤ min ( r , n ) . For a centrosymmetric tridiagonal matrix, let p j p_j p j be the sum in (4.45) restricted to its first j j j indices, with p 0 = 1 p_0=1 p 0 = 1 .
For geometric diagonals take t 0 = c t_0=c t 0 = c , t ℓ = a p ℓ − 1 t_\ell=ap^{\ell-1} t ℓ = a p ℓ − 1 , and t − ℓ = b q ℓ − 1 t_{-\ell}=bq^{\ell-1} t − ℓ = b q ℓ − 1 for ℓ ≥ 1 \ell\geq1 ℓ ≥ 1 . Put τ = p q \tau=pq τ = pq and N = ⌊ n / 2 ⌋ N=\lfloor n/2\rfloor N = ⌊ n /2 ⌋ . The matrix T n T_n T n in (4.31) is indexed by 0 , … , N 0,\ldots,N 0 , … , N and has the tridiagonal entries
( T n ) 0 , 0 = c , ( T n ) h , h = c τ h + h ( a q + b p ) τ h − 1 ( h ≥ 1 ) , (T_n)_{0,0}=c,\qquad(T_n)_{h,h}=c\tau^h+h(aq+bp)\tau^{h-1}\qquad(h\geq1), ( T n ) 0 , 0 = c , ( T n ) h , h = c τ h + h ( a q + b p ) τ h − 1 ( h ≥ 1 ) ,
( T n ) h , h + 1 = a b τ h , ( T n ) h + 1 , h = ( h + 1 ) 2 τ h ( 0 ≤ h < N ) . (80) (T_n)_{h,h+1}=ab\tau^h,\qquad(T_n)_{h+1,h}=(h+1)^2\tau^h\qquad(0\leq h<N). \tag*{(80)} ( T n ) h , h + 1 = ab τ h , ( T n ) h + 1 , h = ( h + 1 ) 2 τ h ( 0 ≤ h < N ) . (80)
Write T n ( z ) T_n(z) T n ( z ) for this matrix with c c c replaced by z c zc z c and the factor ( h + 1 ) 2 (h+1)^2 ( h + 1 ) 2 by ( h + 1 ) ( h + z ) (h+1)(h+z) ( h + 1 ) ( h + z ) , and let T n ′ ( z ) T'_n(z) T n ′ ( z ) omit index zero.
In characteristic three, the geometric family A n = ( t ∣ i − j ∣ ) A_n=(t^{|i-j|}) A n = ( t ∣ i − j ∣ ) has basis 1 , x y , x 2 y 2 1,xy,x^2y^2 1 , x y , x 2 y 2 . Put τ = t 2 \tau=t^2 τ = t 2 . Its fixed matrix is
T 3 = ( 1 τ 0 1 0 τ 2 0 τ − τ 2 ) . (81) T_3=
\begin{pmatrix}
1 & \tau& 0\\
1 & 0 & \tau^2\\
0 & \tau& -\tau^2
\end{pmatrix}. \tag*{(81)} T 3 = 1 1 0 τ 0 τ 0 τ 2 − τ 2 . (81)
For periodic bands the product over one period is fixed. The remaining formulas are the Hessenberg, pentadiagonal, tridiagonal and geometric specializations.
Corollary 4.10.
(1) For row period h h h ,
per A = [ z k ] e B T ( I − z M ) − 1 N r e B = ∑ j = 0 D − 1 r r , j H k − j ( q 1 , … , q D ) . (82) \operatorname{per} A=[z^k]e_B^{\mathsf T}(I-zM)^{-1}N_re_B=\sum_{j=0}^{D-1}r_{r,j}\mathcal{H}_{k-j}(q_1,\ldots,q_D). \tag*{(82)} per A = [ z k ] e B T ( I − z M ) − 1 N r e B = j = 0 ∑ D − 1 r r , j H k − j ( q 1 , … , q D ) . (82) (2) For A n = a J − 1 + b I + ∑ j = 1 r − 1 c j J j A_n=aJ_{-1}+bI+\sum_{j=1}^{r-1}c_jJ_j A n = a J − 1 + b I + ∑ j = 1 r − 1 c j J j ,
per A n = H n ( b , a c 1 , … , a r − 1 c r − 1 ) . (83) \operatorname{per} A_n=\mathcal{H}_n(b,ac_1,\ldots,a^{r-1}c_{r-1}). \tag*{(83)} per A n = H n ( b , a c 1 , … , a r − 1 c r − 1 ) . (83)
At offsets − 1 , 0 , 1 , 2 -1,0,1,2 − 1 , 0 , 1 , 2 with weights a , b , c , d a,b,c,d a , b , c , d , this is
per T = ∑ u , v , w ∈ N u + 2 v + 3 w = n ( u + v + w ) ! u ! v ! w ! b u ( a c ) v ( a 2 d ) w . (84) \operatorname{per} T=\sum_{\substack{u,v,w\in\mathbb{N}\\u+2v+3w=n}}\frac{(u+v+w)!}{u!v!w!}b^{u}(ac)^{v}(a^{2}d)^{w}. \tag*{(84)} per T = u , v , w ∈ N u + 2 v + 3 w = n ∑ u ! v ! w ! ( u + v + w )! b u ( a c ) v ( a 2 d ) w . (84)
(3) For the pentadiagonal case,
per A n = H n ( q 1 , … , q 6 ) − a e H n − 2 ( q 1 , … , q 6 ) . (85) \operatorname{per} A_n=\mathcal{H}_n(q_1,\ldots,q_6)-ae\,\mathcal{H}_{n-2}(q_1,\ldots,q_6). \tag*{(85)} per A n = H n ( q 1 , … , q 6 ) − a e H n − 2 ( q 1 , … , q 6 ) . (85)
The corresponding generating function is
∑ n ≥ 0 per A n z n = 1 − a e z 2 1 − ∑ j = 1 6 q j z j . (86) \sum_{n\geq0}\operatorname{per} A_nz^n=\frac{1-aez^2}{1-\sum_{j=1}^{6}q_jz^j}. \tag*{(86)} n ≥ 0 ∑ per A n z n = 1 − ∑ j = 1 6 q j z j 1 − a e z 2 . (86)
The positive coefficient and sum formulas are
per A n = [ z n ] ∑ j = 0 n F ( z ) j . (87) \operatorname{per} A_n=[z^n]\sum_{j=0}^{n}F(z)^j. \tag*{(87)} per A n = [ z n ] j = 0 ∑ n F ( z ) j . (87)
per A n = ∑ i , j , k , ℓ , u , v ∈ N i + 4 j + 2 k + 3 ℓ + 2 u + 2 v = n u ≤ i + k ( i + j + k + ℓ ) ! i ! j ! k ! ℓ ! ( i + k u ) B ( k + ℓ , v ) c i y k h ℓ x 2 j + u + v . (88) \operatorname{per} A_n=\sum_{\substack{i,j,k,\ell,u,v\in\mathbb{N}\\i+4j+2k+3\ell+2u+2v=n\\u\leq i+k}}\frac{(i+j+k+\ell)!}{i!j!k!\ell!}\binom{i+k}{u}B(k+\ell,v)c^iy^kh^\ell x^{2j+u+v}. \tag*{(88)} per A n = i , j , k , ℓ , u , v ∈ N i + 4 j + 2 k + 3 ℓ + 2 u + 2 v = n u ≤ i + k ∑ i ! j ! k ! ℓ ! ( i + j + k + ℓ )! ( u i + k ) B ( k + ℓ , v ) c i y k h ℓ x 2 j + u + v . (88)
Every term in (88) is nonnegative when the entries are nonnegative. For the symmetric specialization,
per A n = H n ( s 1 , … , s 5 ) − v H n − 1 ( s 1 , … , s 5 ) , (89) \operatorname{per} A_n=\mathcal{H}_n(s_1,\ldots,s_5)-v\mathcal{H}_{n-1}(s_1,\ldots,s_5), \tag*{(89)} per A n = H n ( s 1 , … , s 5 ) − v H n − 1 ( s 1 , … , s 5 ) , (89)
(4) For the tridiagonal case and the separated-diagonal case, respectively,
per A = ∑ E ∏ i ∈ E x i ∏ j ∉ V ( E ) d j . (90) \operatorname{per} A=\sum_{E}\prod_{i\in E}x_i\prod_{j\notin V(E)}d_j. \tag*{(90)} per A = E ∑ i ∈ E ∏ x i j ∈ / V ( E ) ∏ d j . (90)
per ( d I + u J r + l J − r ) = ∏ s = 1 min ( r , n ) ∑ k = 0 ⌊ ℓ s / 2 ⌋ ( ℓ s − k k ) d ℓ s − 2 k ( u l ) k . (91) \operatorname{per}(dI+uJ_r+lJ_{-r})=\prod_{s=1}^{\min(r,n)}\sum_{k=0}^{\lfloor\ell_s/2\rfloor}\binom{\ell_s-k}{k}d^{\ell_s-2k}(ul)^k. \tag*{(91)} per ( d I + u J r + l J − r ) = s = 1 ∏ m i n ( r , n ) k = 0 ∑ ⌊ ℓ s /2 ⌋ ( k ℓ s − k ) d ℓ s − 2 k ( u l ) k . (91)
For constant d i = d d_i=d d i = d , x i = x x_i=x x i = x , the first sum is ∑ k = 0 ⌊ n / 2 ⌋ ( n − k k ) d n − 2 k x k \sum_{k=0}^{\lfloor n/2\rfloor}\binom{n-k}{k}d^{n-2k}x^k ∑ k = 0 ⌊ n /2 ⌋ ( k n − k ) d n − 2 k x k . For a centrosymmetric tridiagonal matrix, the permanent is
p m 2 + x m p m − 1 2 ( n = 2 m ) , d m + 1 p m 2 + 2 x m p m p m − 1 ( n = 2 m + 1 , m ≥ 1 ) . p_m^2+x_mp_{m-1}^2\qquad(n=2m),\qquad d_{m+1}p_m^2+2x_mp_mp_{m-1}\qquad(n=2m+1,\ m\geq1). p m 2 + x m p m − 1 2 ( n = 2 m ) , d m + 1 p m 2 + 2 x m p m p m − 1 ( n = 2 m + 1 , m ≥ 1 ) .
(5) For the geometric diagonals the permanent is (4.31) with (4.35), and
P A n ( z ) = [ t n ] det ( I − t T n ′ ( z ) ) det ( I − t T n ( z ) ) . (92) P_{A_n}(z)=[t^n]\frac{\det(I-tT'_n(z))}{\det(I-tT_n(z))}. \tag*{(92)} P A n ( z ) = [ t n ] det ( I − t T n ( z )) det ( I − t T n ′ ( z )) . (92)
The determinant orders are at most ⌊ n / 2 ⌋ + 1 \lfloor n/2\rfloor+1 ⌊ n /2 ⌋ + 1 . If M ( n ) M(n) M ( n ) is the cost of multiplying polynomials of degree at most n n n , with M ( n ) / n M(n)/n M ( n ) / n nondecreasing, computing the permanent or evaluating P A n P_{A_n} P A n at a specified scalar takes O ( M ( n ) log ( n + 1 ) ) O(M(n)\log(n+1)) O ( M ( n ) log ( n + 1 )) arithmetic operations. All coefficients of P A n P_{A_n} P A n can be computed in O ( n M ( n ) log ( n + 1 ) ) O(nM(n)\log(n+1)) O ( n M ( n ) log ( n + 1 )) operations.
(6) Over any field of characteristic three, for A n = ( t ∣ i − j ∣ ) A_n=(t^{\lvert i-j\rvert}) A n = ( t ∣ i − j ∣ ) and τ = t 2 \tau=t^2 τ = t 2 ,
∑ n ≥ 0 per A n z n = 1 + τ 2 z − τ 3 z 2 1 + ( τ 2 − 1 ) z − ( τ + τ 2 + τ 3 ) z 2 . (93) \sum_{n\geq0}\operatorname{per} A_nz^n=\frac{1+\tau^2z-\tau^3z^2}{1+(\tau^2-1)z-(\tau+\tau^2+\tau^3)z^2}. \tag*{(93)} n ≥ 0 ∑ per A n z n = 1 + ( τ 2 − 1 ) z − ( τ + τ 2 + τ 3 ) z 2 1 + τ 2 z − τ 3 z 2 . (93)
The bound in Theorem 4.9 (5) counts field operations. Over F p ( t ) \mathbb{F}_p(t) F p ( t ) , the degrees of rational functions also enter the bit complexity. At t = 1 t=1 t = 1 , (93) becomes 1 + z − z 2 1+z-z^2 1 + z − z 2 , agreeing with per J n = n ! \operatorname{per} J_n=n! per J n = n ! in characteristic three.
For a J − 2 + b J − 1 + c I + d J u + e J v aJ_{-2}+bJ_{-1}+cI+dJ_u+eJ_v a J − 2 + b J − 1 + c I + d J u + e J v , with 0 < u < v < n 0<u<v<n 0 < u < v < n , take p = 2 p=2 p = 2 , q = v q=v q = v in (74) and zero weights on the other diagonals. For the sign formula in Theorem 4.9 (4), one example is
per ( a I + b J 2 + c J − 3 + d J − 1 ) = det ( a I + b J 2 − c J − 3 + d J − 1 ) . \operatorname{per}(aI+bJ_2+cJ_{-3}+dJ_{-1})=\det(aI+bJ_2-cJ_{-3}+dJ_{-1}). per ( a I + b J 2 + c J − 3 + d J − 1 ) = det ( a I + b J 2 − c J − 3 + d J − 1 ) .
For (77) , in order 50 the offsets − 21 , 0 , 13 , 24 -21,0,13,24 − 21 , 0 , 13 , 24 give N = 58 N=58 N = 58 .
For example, the matrix ( ρ ∣ i − j ∣ ) i , j = 1 n (\rho^{|i-j|})_{i,j=1}^n ( ρ ∣ i − j ∣ ) i , j = 1 n is obtained by taking c = 1 c=1 c = 1 and a = b = p = q = ρ a=b=p=q=\rho a = b = p = q = ρ . Its determinant is ( 1 − ρ 2 ) n − 1 (1-\rho^2)^{n-1} ( 1 − ρ 2 ) n − 1 , so this is a dense full-rank family for ρ ≠ 0 \rho\ne0 ρ = 0 and ρ 2 ≠ 1 \rho^2\ne1 ρ 2 = 1 . The two ranks across each nontrivial initial interval cut are one. In order 2 m 2m 2 m , the odd-row, even-column minor has determinant ρ m ( 1 − ρ 2 ) m − 1 \rho^m(1-\rho^2)^{m-1} ρ m ( 1 − ρ 2 ) m − 1 , so the rank is at least m m m after any diagonal change. For rational 0 < ρ < 1 0<\rho<1 0 < ρ < 1 , the weighted clique-width is at least n / 6 n/6 n /6 . For ρ = 2 \rho=2 ρ = 2 , (92) gives
P A 4 ( z ) = z 4 + 108 z 3 + 848 z 2 + 768 z . P_{A_4}(z)=z^4+108z^3+848z^2+768z. P A 4 ( z ) = z 4 + 108 z 3 + 848 z 2 + 768 z .
In (63) , the second term is obtained from (88) by multiplying each summand by n / ( i + j + k + ℓ ) n/(i+j+k+\ell) n / ( i + j + k + ℓ ) for n ≥ 1 n\ge1 n ≥ 1 . The pentadiagonal denominator and its symmetric reduction also occur in [19 , 20 ] . In the symmetric specialization both A n A_n A n and A n J n A_nJ_n A n J n are symmetric and centrosymmetric. At t = u = v = 1 t=u=v=1 t = u = v = 1 the generating function is ( 1 − z ) / ( 1 − 2 z − 2 z 3 + z 5 ) (1-z)/(1-2z-2z^3+z^5) ( 1 − z ) / ( 1 − 2 z − 2 z 3 + z 5 ) . Its coefficients for n = 0 , … , 10 n=0,\ldots,10 n = 0 , … , 10 are 1 , 1 , 2 , 6 , 14 , 31 , 73 , 172 , 400 , 932 , 2177 1,1,2,6,14,31,73,172,400,932,2177 1 , 1 , 2 , 6 , 14 , 31 , 73 , 172 , 400 , 932 , 2177 , the perfect-matching counts for biadjacency matrix ( 1 ∣ i − j ∣ ≤ 2 ) (1_{|i-j|\le2}) ( 1 ∣ i − j ∣ ≤ 2 ) . For related cycle continued fractions see Deb [29 ] .
More generally, when every offset is a multiple of r r r , the permanent is the product over the residue classes modulo r r r of the ordinary diagonal formulas with offsets divided by r r r . In the direct signing one may take ( x , y ) = ( − 1 , 1 ) (x,y)=(-1,1) ( x , y ) = ( − 1 , 1 ) if q / g q/g q / g is odd and ( 1 , − 1 ) (1,-1) ( 1 , − 1 ) otherwise. For constant weights, the formula reads
per ( a I + b J p + c J q + d J p + q ) = det ( a I − b x J p − c y J q − d x y J p + q ) . \operatorname{per}(aI+bJ_p+cJ_q+dJ_{p+q})=\det(aI-bxJ_p-cyJ_q-dxyJ_{p+q}). per ( a I + b J p + c J q + d J p + q ) = det ( a I − b x J p − cy J q − d x y J p + q ) .
4.8.
Let H H H have order n n n and put A = H J n A=HJ_n A = H J n . The anti-diagonal i + j = n − 1 − d i+j=n-1-d i + j = n − 1 − d of H H H is the ordinary diagonal of offset d d d in A A A . Column reversal preserves the permanent, so the open-diagonal formulas apply to H H H .
For the characteristic-zero reciprocal family, use one-based indices and put
C n H = ( 1 α + i + j + P ( i + j ) ) i , j = 1 n , C_n^H=\left(\frac{1}{\alpha+i+j}+P(i+j)\right)_{i,j=1}^n, C n H = ( α + i + j 1 + P ( i + j ) ) i , j = 1 n ,
where the denominators are nonzero and P P P is a polynomial. In (26) , take x i = i x_i=i x i = i , y j = − α − j y_j=-\alpha-j y j = − α − j and the bivariate polynomial P ( X − Y − α ) P(X-Y-\alpha) P ( X − Y − α ) .
Theorem 4.11. We have per H = per A \operatorname{per} H=\operatorname{per} A per H = per A , with the following specializations.
For anti-diagonals with − p ≤ d ≤ q -p\le d\le q − p ≤ d ≤ q , use (73) for arbitrary entries. If H = ( h i + j ) H=(h_{i+j}) H = ( h i + j ) , take t d = h n − 1 − d t_d=h_{n-1-d} t d = h n − 1 − d in (74) . The same fixed one-sided bound applies. For rational diagonal series of this t d t_d t d , the permanent is (76) .
For any three anti-diagonals, or for four satisfying the condition of Theorem 4.9 (3) in the offsets d d d ,
per H = D N ( ( H J n ) ⊕ I N − n ) . \operatorname{per} H=D_N\left((HJ_n)\oplus I_{N-n}\right). per H = D N ( ( H J n ) ⊕ I N − n ) . For offsets satisfying Theorem 4.9 (4), let K K K be its signed matrix for A A A . Then
per H = det K = ( − 1 ) n ( n − 1 ) / 2 det ( K J n ) . \operatorname{per} H=\det K=(-1)^{n(n-1)/2}\det(KJ_n). per H = det K = ( − 1 ) n ( n − 1 ) /2 det ( K J n ) . In characteristic zero, per C n H \operatorname{per} C_n^H per C n H is given by (26) with the specified data. For fixed degree of P P P , the arithmetic cost is polynomial, and so is the bit complexity for rational input.
(4) In characteristic p > 0 p > 0 p > 0 , if the reversed matrix A A A has generators of orders r r r , s s s , then its permanent and that of H H H are given by (32) , with cost p O ( r + s ) n p^{O(r+s)}n p O ( r + s ) n from the generators. For a reversed Toeplitz matrix with constant generators, this is ( T p n ) 1 , 1 (T_p^n)_{1,1} ( T p n ) 1 , 1 .
Column reversal gives the separated tridiagonal, pentadiagonal and geometric parameters of Corollary 4.10.
Corollary 4.12. For H = ( d I + u J r + l J − r ) J n H = (dI + uJ_r + lJ_{-r})J_n H = ( d I + u J r + l J − r ) J n , its permanent is the product in (91) . For H = ( a J − 2 + b J − 1 + c I + d J 1 + e J 2 ) J n H = (aJ_{-2} + bJ_{-1} + cI + dJ_1 + eJ_2)J_n H = ( a J − 2 + b J − 1 + c I + d J 1 + e J 2 ) J n , it is
H n ( q 1 , … , q 6 ) − a e H n − 2 ( q 1 , … , q 6 ) , \mathcal{H}_n(q_1,\ldots,q_6) - ae\mathcal{H}_{n-2}(q_1,\ldots,q_6), H n ( q 1 , … , q 6 ) − a e H n − 2 ( q 1 , … , q 6 ) , with q j q_j q j from (55) . Its positive form is (88) . For H i j = t n − 1 − i − j H_{ij} = t_{n-1-i-j} H ij = t n − 1 − i − j with t 0 = c t_0 = c t 0 = c , t ℓ = a p ℓ − 1 t_\ell= ap^{\ell-1} t ℓ = a p ℓ − 1 , t − ℓ = b q ℓ − 1 t_{-\ell} = bq^{\ell-1} t − ℓ = b q ℓ − 1 for ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 , use (76) with the tridiagonal matrix (80) .
For a fixed sequence h j h_j h j , the reversed diagonal sequence t d = h n − 1 − d t_d = h_{n-1-d} t d = h n − 1 − d can change with n n n . The last part of the theorem is an evaluation for each order; it does not assert a single rational generating function for all such Hankel matrices.
4.9.
Let B B B be an m × m m \times m m × m ( 0 , 1 ) (0,1) ( 0 , 1 ) matrix, and put
S = ( 0 B B T 0 ) , C = B ⊕ ( J m B J m ) S = \begin{pmatrix} 0 & B \\ B^{\mathsf{T}} & 0 \end{pmatrix}, \qquad C = B \oplus(J_mBJ_m) S = ( 0 B T B 0 ) , C = B ⊕ ( J m B J m )
and D = S ⊕ ( J 2 m S J 2 m ) D = S \oplus(J_{2m}SJ_{2m}) D = S ⊕ ( J 2 m S J 2 m ) . The proposition relates their permanents to the permanent of B B B .
Proposition 4.13. The matrix S S S is symmetric and C C C is centrosymmetric, with per S = per C = ( per B ) 2 \operatorname{per} S = \operatorname{per} C = (\operatorname{per} B)^2 per S = per C = ( per B ) 2 . The matrix D D D has both symmetries and per D = ( per B ) 4 \operatorname{per} D = (\operatorname{per} B)^4 per D = ( per B ) 4 .
Thus polynomial-time evaluation in either whole class would give polynomial-time evaluation of arbitrary ( 0 , 1 ) (0,1) ( 0 , 1 ) permanents. For the centrosymmetric matrix with rows ( a , b ) , ( b , a ) (a,b),(b,a) ( a , b ) , ( b , a ) , the permanent is a 2 + b 2 a^2+b^2 a 2 + b 2 , while the product from the two one-dimensional invariant subspaces is ( a + b ) ( a − b ) (a+b)(a-b) ( a + b ) ( a − b ) .
Proofs
Proof of Theorem 3.1
Proof of Theorem 3.1. We first prove the formula for a given tree. Take two copies of I I I , one for rows and one for columns. For a matching M M M between the two copies write its contribution as
∏ ( i , j ) ∈ M a i j ∏ i ∈ I i u n m a t c h e d a s a r o w R I ( i , : ) x I ∏ j ∈ I j u n m a t c h e d a s a c o l u m n C I ( j , : ) y I . (94) \prod_{(i,j)\in M} a_{ij}
\prod_{\substack{i\in I\\ i\ \mathrm{unmatched\ as\ a\ row}}} R_I(i,:)x_I
\prod_{\substack{j\in I\\ j\ \mathrm{unmatched\ as\ a\ column}}} C_I(j,:)y_I . \tag*{(94)} ( i , j ) ∈ M ∏ a ij i ∈ I i unmatched as a row ∏ R I ( i , : ) x I j ∈ I j unmatched as a column ∏ C I ( j , : ) y I . (94)
Let P I P_I P I be the sum of these contributions. A matching of a a a edges leaves ∣ I ∣ − a |I|-a ∣ I ∣ − a factors of each kind. Thus its monomials belong to M I M_I M I . At a leaf { i } \{i\} { i } the empty matching gives the two coordinate factors and the one-edge matching gives a i i a_{ii} a ii . This is the polynomial defining l i l_i l i .
Let U , V U,V U , V be the children of I I I . Take a matching from each child, and choose an unmatched row i i i in U U U and an unmatched column j j j in V V V . When the first sum in D I D_I D I differentiates their factors, the scalar obtained is
∑ α , β R U ( i , α ) ( B U V ) α β C V ( j , β ) = ( R U B U V C V T ) i j = a i j . \sum_{\alpha,\beta} R_U(i,\alpha)(B_{UV})_{\alpha\beta}C_V(j,\beta) = (R_UB_{UV}C_V^{\mathsf{T}})_{ij} = a_{ij}. α , β ∑ R U ( i , α ) ( B U V ) α β C V ( j , β ) = ( R U B U V C V T ) ij = a ij .
The other factors are untouched. The second sum gives a i j a_{ij} a ij for a row in V V V and a column in U U U . If several endpoint factors contain the same variable, the product rule still makes one choice of a factor at a time. It therefore counts choices of endpoints with their multiplicities. For example, with one row variable and one column variable,
∑ ℓ = 0 2 ( b ∂ x ∂ y ) ℓ ℓ ! x 2 y 2 = x 2 y 2 + 4 b x y + 2 b 2 . \sum_{\ell=0}^{2}\frac{(b\partial_x\partial_y)^\ell}{\ell!}x^2y^2=x^2y^2+4bxy+2b^2. ℓ = 0 ∑ 2 ℓ ! ( b ∂ x ∂ y ) ℓ x 2 y 2 = x 2 y 2 + 4 b x y + 2 b 2 .
There are four choices for one pair of factors and two bijections when both row factors are paired. The coefficients remain integers when the variables are replaced by longer lists.
For a fixed set of ℓ \ell ℓ disjoint cross edges, the repeated differentiation can add them in any of ℓ ! \ell! ℓ ! orders. Conversely, once a row or column factor has been removed, it is unavailable in the next differentiation. Thus every term after ℓ \ell ℓ applications comes from ℓ \ell ℓ disjoint cross edges. In characteristic zero the factor 1 / ℓ ! 1/\ell! 1/ ℓ ! counts each set once. There are at most ∣ I ∣ |I| ∣ I ∣ cross edges in a matching.
The same count can be used over any field. Write X = ( x U , x V ) X=(x_U,x_V) X = ( x U , x V ) and Y = ( y U , y V ) Y=(y_U,y_V) Y = ( y U , y V ) for the lists of variables, of lengths ρ = r U + r V \rho=r_U+r_V ρ = r U + r V and σ = s U + s V \sigma=s_U+s_V σ = s U + s V . Let B = ( b a b ) B=(b_{ab}) B = ( b ab ) be the block matrix with zero diagonal blocks and off-diagonal blocks B U V , B V U B_{UV},B_{VU} B U V , B V U . For a monomial X α Y β X^\alpha Y^\beta X α Y β , take a ρ × σ \rho\times\sigma ρ × σ array M = ( m a b ) M=(m_{ab}) M = ( m ab ) of nonnegative integers. Write p a = ∑ b m a b p_a=\sum_b m_{ab} p a = ∑ b m ab and q b = ∑ a m a b q_b=\sum_a m_{ab} q b = ∑ a m ab . The contraction inside ρ I \rho_I ρ I in (18) replaces the monomial by
∑ M ≥ 0 p ≤ α , q ≤ β ( ∏ a ( α a α a − p a , m a 1 , … , m a σ ) ) ( ∏ b ( β b β b − q b , m 1 b , … , m ρ b ) ) ⋅ ( ∏ a , b m a b ! b a b m a b ) X α − p Y β − q . (95) \sum_{\substack{M\geq0\\p\leq\alpha,\ q\leq\beta}}
\left(\prod_a \binom{\alpha_a}{\alpha_a-p_a,m_{a1},\ldots,m_{a\sigma}}\right)
\left(\prod_b \binom{\beta_b}{\beta_b-q_b,m_{1b},\ldots,m_{\rho b}}\right)
\cdot\left(\prod_{a,b}m_{ab}!\,b_{ab}^{m_{ab}}\right)X^{\alpha-p}Y^{\beta-q}. \tag*{(95)} M ≥ 0 p ≤ α , q ≤ β ∑ ( a ∏ ( α a − p a , m a 1 , … , m aσ α a ) ) ( b ∏ ( β b − q b , m 1 b , … , m ρ b β b ) ) ⋅ a , b ∏ m ab ! b ab m ab X α − p Y β − q . (95)
Here the multinomial coefficients and factorials are integers read in the field. To obtain the term for M M M , first distinguish the repeated factors in the monomial. Of the α a \alpha_a α a factors of type a a a , choose m a b m_{ab} m ab to be paired with factors of type b b b , for each b b b . The first multinomial counts these choices. The second multinomial chooses the factors on the other side. For each a , b a,b a , b , pair the selected factors in m a b ! m_{ab}! m ab ! ways and include b a b b_{ab} b ab for each pair. This counts every choice of disjoint edges once. It proves (5.2) and gives the same coefficients as the divided derivatives in characteristic zero.
The change of coordinates then turns, for example, R U ( i , : ) x U R_U(i,:)x_U R U ( i , : ) x U into R U ( i , : ) S U , I x I = R I ( i , : ) x I R_U(i,:)S_{U,I}x_I=R_I(i,:)x_I R U ( i , : ) S U , I x I = R I ( i , : ) x I . The column factors change in the same way. It follows that
P I = ρ I ( ∑ ℓ = 0 ∣ I ∣ D I ℓ ℓ ! P U P V ) . \mathcal{P}_I=\rho_I\left(\sum_{\ell=0}^{|I|}\frac{\mathcal{D}_I^\ell}{\ell!}\mathcal{P}_U\mathcal{P}_V\right). P I = ρ I ℓ = 0 ∑ ∣ I ∣ ℓ ! D I ℓ P U P V .
In positive characteristic the sum in parentheses is understood as (5.2) applied to each monomial. Each matching on I I I occurs in this identity through its unique restrictions to U , V U,V U , V and its unique set of cross edges.
If p J ( w ) = [ w ] P J p_J(w)=[w]P_J p J ( w ) = [ w ] P J , taking [ w ] [w] [ w ] at the parent gives
p I ( w ) = ∑ u ∈ M U , v ∈ M V p U ( u ) p V ( v ) t I ( u , v ; w ) . p_I(w)=\sum_{u\in\mathcal{M}_U,\ v\in\mathcal{M}_V}p_U(u)p_V(v)t_I(u,v;w). p I ( w ) = u ∈ M U , v ∈ M V ∑ p U ( u ) p V ( v ) t I ( u , v ; w ) .
Expanding this identity down the tree gives one sum over all node monomials, with a factor l i l_i l i at each leaf and t I t_I t I at each merge. At the root the coordinate variables are empty, so every matching with an unmatched row or column gives zero. Taking its constant coefficient gives (3.2).
For a node I I I , the number of coordinates is at most
∑ h = 0 ∣ I ∣ ( h + r − 1 r − 1 ) ( h + s − 1 s − 1 ) = O r , s ( ( ∣ I ∣ + 1 ) r + s − 1 ) . \sum_{h=0}^{|I|}\binom{h+r-1}{r-1}\binom{h+s-1}{s-1}=O_{r,s}\left((|I|+1)^{r+s-1}\right). h = 0 ∑ ∣ I ∣ ( r − 1 h + r − 1 ) ( s − 1 h + s − 1 ) = O r , s ( ( ∣ I ∣ + 1 ) r + s − 1 ) .
Zero-rank factors only reduce this number. There are n − 1 n-1 n − 1 merges. The product at a merge uses at most 2 ( r + s ) 2(r+s) 2 ( r + s ) variables of total degree at most 2 n 2n 2 n . The number of ordinary monomials is thus bounded by ( 2 n + 2 ( r + s ) 2 ( r + s ) ) \binom{2n+2(r+s)}{2(r+s)} ( 2 ( r + s ) 2 n + 2 ( r + s ) ) . For each monomial, the number of arrays M M M in (5.2) is at most ( n + 1 ) 4 r s (n+1)^{4rs} ( n + 1 ) 4 r s . The integer coefficients can be formed by binomial coefficient tables and multiplication. Polynomial multiplication, contraction and substitution of linear forms therefore take polynomially many field operations for fixed r , s r, s r , s . The rank data are obtained by elimination on submatrices of A A A .
It remains to find a tree from the matrix. If k = 0 k = 0 k = 0 , each singleton cut has width zero, so A A A is diagonal and its permanent is the product of its diagonal entries. Assume now that k ≥ 1 k \ge1 k ≥ 1 . For I ⊆ [ n ] I \subseteq[n] I ⊆ [ n ] , put f ( I ) = rank A [ I c , I ] f(I) = \operatorname{rank} A[I^{\mathrm{c}}, I] f ( I ) = rank A [ I c , I ] . Group column i i i of A A A with the i i i -th unit column. Let V I V_I V I be the span of the groups indexed by I I I . The selected unit columns span the coordinate subspace on I I I . After quotienting out this subspace, the remaining selected columns are the columns of A [ I c , I ] A[I^{\mathrm{c}}, I] A [ I c , I ] . It follows that dim V I = ∣ I ∣ + f ( I ) \dim V_I = |I| + f(I) dim V I = ∣ I ∣ + f ( I ) .
We have V I ∪ J = V I + V J V_{I\cup J} = V_I + V_J V I ∪ J = V I + V J and V I ∩ J ⊆ V I ∩ V J V_{I\cap J} \subseteq V_I \cap V_J V I ∩ J ⊆ V I ∩ V J . The dimension formula gives
dim V I + dim V J ≥ dim V I ∪ J + dim V I ∩ J . \dim V_I + \dim V_J \ge\dim V_{I\cup J} + \dim V_{I\cap J}. dim V I + dim V J ≥ dim V I ∪ J + dim V I ∩ J .
Subtracting ∣ I ∣ + ∣ J ∣ = ∣ I ∪ J ∣ + ∣ I ∩ J ∣ |I|+|J| = |I\cup J|+|I\cap J| ∣ I ∣ + ∣ J ∣ = ∣ I ∪ J ∣ + ∣ I ∩ J ∣ proves the same inequality for f f f . It also holds for I ↦ f ( I c ) I \mapsto f(I^{\mathrm{c}}) I ↦ f ( I c ) . Thus w A ( I ) = f ( I ) + f ( I c ) w_A(I) = f(I) + f(I^{\mathrm{c}}) w A ( I ) = f ( I ) + f ( I c ) is submodular, is symmetric under taking complements, and vanishes on the empty set. It can be evaluated by computing two ranks.
If rw F ( A ) ≤ k \operatorname{rw}_{\mathbb{F}}(A) \le k rw F ( A ) ≤ k , the branch decomposition algorithm of Korhonen and Oum [30 ] finds a tree of width at most k k k using w A w_A w A as its oracle. Rooting the decomposition at a subdivided edge gives a rooted binary tree of the same width. The one-index tree needs no construction. For fixed k k k the search takes polynomial time. Over Q \mathbb{Q} Q each rank query has polynomial bit cost. Both cut ranks on the resulting tree are bounded by k k k . The preceding evaluation therefore applies with r = s = max ( 1 , k ) r = s = \max(1,k) r = s = max ( 1 , k ) .
For rational inputs choose the coordinate bases from rows and columns of the original cut blocks. Solving the coordinate equations uses minors of order bounded in terms of r , s r, s r , s , so the resulting rational numbers have polynomial bit length. There are polynomially many of them. Take a common denominator Q Q Q for these numbers and the entries of A A A . A product in (94) has at most 2 n 2n 2 n factors, so its coefficients have a denominator dividing Q 2 n Q^{2n} Q 2 n . To bound the numerator, there are at most ( n + 1 ) n (n+1)^n ( n + 1 ) n partial choices of columns before imposing the matching condition, and each endpoint factor has at most max ( r , s ) \max(r,s) max ( r , s ) terms. The logarithm of the number of products being added is therefore polynomial in n n n . At a merge at most ∣ I ∣ |I| ∣ I ∣ differentiations are made, and the extra divisors ℓ ! \ell! ℓ ! all divide ∣ I ∣ ! |I|! ∣ I ∣ ! . The substitution contributes at most 2 n 2n 2 n additional coordinate factors to a product. These bounds also give polynomial bit length for the intermediate coefficients.
Proof of Theorem 3.2
We first check that reduction modulo the p p p th powers commutes with the operations used in the matching formula.
Lemma 5.1. Over a field of characteristic p > 0 p>0 p > 0 , the ideal generated by the p p p th powers of the variables is preserved by multiplication, ordinary differentiation, and linear substitution. The integer matching contraction descends to the quotient by this ideal and is the product of the operators (2.7).
Proof. Multiplication preserves every ideal. If z z z is one of the variables, then
∂ z ( z p f ) = p z p − 1 f + z p ∂ z f = z p ∂ z f . \partial_z(z^p f)=p z^{p-1}f+z^p\partial_z f=z^p\partial_z f. ∂ z ( z p f ) = p z p − 1 f + z p ∂ z f = z p ∂ z f .
Differentiation in another variable also preserves the ideal. For linear substitution, use
( a 1 z 1 + ⋯ + a d z d ) p = a 1 p z 1 p + ⋯ + a d p z d p . (a_1z_1+\cdots+a_dz_d)^p=a_1^p z_1^p+\cdots+a_d^p z_d^p. ( a 1 z 1 + ⋯ + a d z d ) p = a 1 p z 1 p + ⋯ + a d p z d p .
This identity holds over every field of characteristic p p p ; no assumption that the field has p p p elements is being made.
For one pair of coordinate types, match m m m of the a a a row factors to m m m of the b b b column factors. With edge weight c c c , the integer coefficient is
( a m ) ( b m ) m ! c m = ( a m ) ( b ) m c m , ( b ) m = b ( b − 1 ) ⋯ ( b − m + 1 ) . (96) \binom{a}{m}\binom{b}{m}m!c^m=\binom{a}{m}(b)_m c^m,\qquad(b)_m=b(b-1)\cdots(b-m+1). \tag*{(96)} ( m a ) ( m b ) m ! c m = ( m a ) ( b ) m c m , ( b ) m = b ( b − 1 ) ⋯ ( b − m + 1 ) . (96)
If m ≥ p m \ge p m ≥ p , this coefficient is divisible by p p p (when m > b m > b m > b the term is already zero). If m < p m < p m < p , it equals ( a ) m ( b ) m c m / m ! (a)_m(b)_m c^m/m! ( a ) m ( b ) m c m / m ! , the coefficient in (7) . It follows that the full integer contraction for this pair reduces to (7) . In particular, this statement also applies before high powers have been discarded.
Apply these pair contractions successively for all coordinate pairs. An endpoint used by one pair is no longer available to a later pair. Consequently each set of disjoint matches is counted once, grouped according to its numbers of matches of each type. This is the integer contraction in (95) . Each surviving operator is made from ordinary derivatives and scalars, so it preserves the ideal. Linear substitution then preserves the corresponding parent ideal. This proves the lemma. □
Proof of Theorem 3.2 . Before taking the quotient, associate with a node I I I the polynomial
P I ( x I , y I ) = ∑ S , T ⊆ I ∣ S ∣ = ∣ T ∣ per A [ S , T ] ∏ i ∈ I ∖ S ( R I ( i , : ) x I ) ∏ j ∈ I ∖ T ( C I ( j , : ) y I ) . (97) P_I(x_I,y_I)=\sum_{\substack{S,T\subseteq I\\ |S|=|T|}}\operatorname{per} A[S,T]\prod_{i\in I\setminus S}(R_I(i,:)x_I)\prod_{j\in I\setminus T}(C_I(j,:)y_I). \tag*{(97)} P I ( x I , y I ) = S , T ⊆ I ∣ S ∣ = ∣ T ∣ ∑ per A [ S , T ] i ∈ I ∖ S ∏ ( R I ( i , : ) x I ) j ∈ I ∖ T ∏ ( C I ( j , : ) y I ) . (97)
A term records a matching inside I I I and the unmatched rows and columns. Both numbers of unmatched endpoints are ∣ I ∣ − ∣ S ∣ |I|-|S| ∣ I ∣ − ∣ S ∣ . The polynomial therefore has equal degree in x I x_I x I and y I y_I y I .
At a leaf this is the polynomial defining l i l_i l i . At a merge, first choose a matching in each child. The remaining internal edges of the parent go between the two children. For instance, differentiating the row factor for i ∈ U i\in U i ∈ U and the column factor for j ∈ V j\in V j ∈ V with the coupling B U V B_{UV} B U V gives
R U ( i , : ) B U V C V ( j , : ) T = a i j . R_U(i,:)B_{UV}C_V(j,:)^{\mathsf T}=a_{ij}. R U ( i , : ) B U V C V ( j , : ) T = a ij .
The integer coefficients (96) count all possible choices of distinct endpoints. Each matching on I I I has a unique restriction to its children and a unique set of cross edges. After the substitutions ρ I \rho_I ρ I , every surviving endpoint has the parent coordinate. Hence the untruncated polynomials obey the integer matching contraction rule.
By Lemma 5.1, that rule can be applied after reduction modulo the p p p th powers at every stage. If P ‾ I \overline{P}_I P I is the reduced polynomial, it follows that
P ‾ I = ρ I ( E I ( P ‾ U P ‾ V ) ) . \overline{P}_I=\rho_I\left(E_I(\overline{P}_U\overline{P}_V)\right). P I = ρ I ( E I ( P U P V ) ) .
Expanding coefficients down the tree proves (20) . At the root there are no coordinate variables. In (97) , every term with an unmatched endpoint is zero, leaving exactly per A \operatorname{per} A per A .
The basis bound follows from (5) . At a merge there are at most 2 k 2k 2 k variables before substitution and at most k k k after substitution. A polynomial with all exponents less than p p p in d d d variables has at most p d p^d p d terms. Multiplication, the at most k 2 k^2 k 2 pair contractions, and substitution of linear forms can each be performed in p O ( k ) p^{O(k)} p O ( k ) operations by direct polynomial arithmetic in these bounded lists. For substitution, precompute each linear form to powers 0 , … , p − 1 0,\ldots,p-1 0 , … , p − 1 , reducing after each multiplication, and then substitute monomials one at a time. There are fewer than 2 n 2n 2 n nodes. Rank factorizations and the coordinate matrices use a polynomial number of field operations and zero tests, independent of k k k . This proves the bound when a tree is supplied. For k = 0 k=0 k = 0 , the matrix is diagonal and the assertion is immediate. The construction from the matrix uses an approximate decomposition. The proof of Theorem 3.1 shows that w A w_A w A is a symmetric submodular integer-valued function, with w A ( ∅ ) = 0 w_A(\varnothing)=0 w A ( ∅ ) = 0 . Its singleton values are at most two. The approximation theorem of Oum and Seymour [31 ] , in the form [30 Theorem 5.3] , finds a decomposition of width at most 3 k + c ′ 3k+c' 3 k + c ′ , where c ′ = max i w A ( { i } ) c'=\max_i w_A(\{i\}) c ′ = max i w A ({ i }) , or certifies that the width exceeds k k k . Its cost here is 2 O ( k ) n O ( 1 ) 2^{O(k)}n^{O(1)} 2 O ( k ) n O ( 1 ) , since two rank computations give the value oracle and submodular minimization has polynomial cost. Under the hypothesis rw ( A ) ≤ k \operatorname{rw}(A)\le k rw ( A ) ≤ k , we have c ′ ≤ k c'\le k c ′ ≤ k , so the output width is at most 4 k 4k 4 k . Subdivide an edge to root the tree. The preceding evaluation then costs p O ( 4 k ) n O ( 1 ) = p O ( k ) n O ( 1 ) p^{O(4k)}n^{O(1)}=p^{O(k)}n^{O(1)} p O ( 4 k ) n O ( 1 ) = p O ( k ) n O ( 1 ) . If k k k is not given, first check the diagonal case, then try successive positive integers until a tree is returned. This succeeds by the actual rank-width and has the same bound.
Proof of Theorem 3.3
The determinant quotient for repeated nodes
For balanced multiindices a , b a,b a , b , let C a , b C_{a,b} C a , b be the ordinary Cauchy matrix whose row nodes are x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n followed by b β b_\beta b β copies of each η β \eta_\beta η β , and whose column nodes are y 1 , … , y n y_1,\ldots,y_n y 1 , … , y n followed by a α a_\alpha a α copies of each ζ α \zeta_\alpha ζ α . This matrix has repeated rows and columns. The next identity is the confluent form of (2.8).
Lemma 5.2. For admissible auxiliary nodes and ( a , b ) ∈ I n (a,b) \in\mathcal{I}_n ( a , b ) ∈ I n ,
det K 1 ( a , b ) ≠ 0 , per C a , b = R a , b . \det K_1(a,b) \ne0,\qquad\operatorname{per} C_{a,b}=R_{a,b}. det K 1 ( a , b ) = 0 , per C a , b = R a , b .
Proof. First replace every repeated node by distinct indeterminates and apply (2.8) to the resulting Cauchy matrix. In each determinant divide by the Vandermonde product within each row cluster and within each column cluster. The same factors are removed from the numerator and denominator.
We describe the resulting specialization. If h h h rows have nodes u + ϵ 1 , … , u + ϵ h u+\epsilon_1,\ldots,u+\epsilon_h u + ϵ 1 , … , u + ϵ h , expand their entries in the ϵ i \epsilon_i ϵ i . Multilinearity of the determinant shows that terms with repeated Taylor orders vanish. The term of smallest total degree has Taylor orders 0 , 1 , … , h − 1 0,1,\ldots,h-1 0 , 1 , … , h − 1 . Dividing by ∏ i < j ( ϵ j − ϵ i ) \prod_{i<j}(\epsilon_j-\epsilon_i) ∏ i < j ( ϵ j − ϵ i ) and then setting all ϵ i = 0 \epsilon_i=0 ϵ i = 0 therefore replaces these rows by the divided derivatives of orders 0 , 1 , … , h − 1 0,1,\ldots,h-1 0 , 1 , … , h − 1 . Apply this argument to every row cluster, and then to every column cluster. The required divided derivative is
1 ℓ ! m ! ∂ v ℓ ∂ u m ( u − v ) − q = ( − 1 ) ℓ ( q + ℓ + m − 1 ) ! ( q − 1 ) ! ℓ ! m ! ( u − v ) − q − ℓ − m , \frac{1}{\ell!m!}\partial_v^\ell\partial_u^m(u-v)^{-q}
=(-1)^\ell\frac{(q+\ell+m-1)!}{(q-1)!\ell!m!}(u-v)^{-q-\ell-m}, ℓ ! m ! 1 ∂ v ℓ ∂ u m ( u − v ) − q = ( − 1 ) ℓ ( q − 1 )! ℓ ! m ! ( q + ℓ + m − 1 )! ( u − v ) − q − ℓ − m ,
which is exactly (2.13).
The Cauchy determinant formula also shows that the divided denominator has nonzero specialization. After removing the within-cluster Vandermonde factors, its numerator is a product of nonzero differences between distinct row nodes and between distinct column nodes, with their appropriate multiplicities. Its denominator is a product of differences between row and column nodes. Every such difference is nonzero by admissibility.
The permanent itself has no singularity when nodes within a row list or within a column list coalesce. Specializing Borchardt’s identity after the divisions consequently gives the claimed ratio. This argument takes place in a rational function field and by specialization holds over every field of characteristic zero. □
Expansion according to the blocks
The purpose of the auxiliary rows and columns is to collect many submatrix permanents in a polynomial number of coefficients. The following identity explains the sum in (2.16).
Lemma 5.3. For admissible auxiliary nodes and every r × s r\times s r × s matrix L L L ,
per ( C + U L V ) = D ( x , y ; η , ζ ; L ) . (98) \operatorname{per}(C+ULV)=D(x,y;\eta,\zeta;L). \tag*{(98)} per ( C + U L V ) = D ( x , y ; η , ζ ; L ) . (98)
Proof. Introduce the polynomial
Q ( z , w ) = ∑ I , J ⊆ [ n ] ∣ I ∣ = ∣ J ∣ per C [ I c , J c ] ∏ i ∈ I ( ∑ α U i α z α ) ∏ j ∈ J ( ∑ β V j β w β ) . (99) Q(z,w)=\sum_{\substack{I,J\subseteq[n]\\|I|=|J|}}\operatorname{per}C[I^c,J^c]\prod_{i\in I}\left(\sum_\alpha U_{i\alpha}z_\alpha\right)\prod_{j\in J}\left(\sum_\beta V_{j\beta}w_\beta\right). \tag*{(99)} Q ( z , w ) = I , J ⊆ [ n ] ∣ I ∣ = ∣ J ∣ ∑ per C [ I c , J c ] i ∈ I ∏ ( α ∑ U i α z α ) j ∈ J ∏ β ∑ V j β w β . (99)
Its monomials have the form z a w b z^a w^b z a w b with ( a , b ) ∈ I n (a,b)\in\mathcal{I}_n ( a , b ) ∈ I n . The subsets in (5.6) are used to prove the identity; we will evaluate its coefficients by determinants instead of enumerating these subsets.
Consider a perfect matching counted by per C a , b \operatorname{per} C_{a,b} per C a , b . Its edges between original rows and original columns give one of the submatrix terms in (99) . Each remaining original row uses an auxiliary column; each remaining original column uses an auxiliary row. Suppose the used auxiliary columns and rows have type multiplicities a ′ a' a ′ and b ′ b' b ′ . Put c = a − a ′ c=a-a' c = a − a ′ and d = b − b ′ d=b-b' d = b − b ′ .
The auxiliary columns are labelled even when their entries agree. For prescribed types, the number of ways of assigning them to the original rows is a ! / c ! a!/c! a ! / c ! . Similarly the auxiliary row assignments contribute b ! / d ! b!/d! b ! / d ! . The remaining auxiliary rows and columns are matched using W W W . If M α β M_{\alpha\beta} M α β records the number of edges from row type β \beta β to column type α \alpha α , the number of such labelled matchings is
c ! d ! ∏ α , β M α β ! \frac{c!d!}{\prod_{\alpha,\beta}M_{\alpha\beta}!} ∏ α , β M α β ! c ! d !
After including the weights and summing over M M M , their contribution is c ! d ! e W T ( c , d ) c!d!e_{W^{\mathsf T}}(c,d) c ! d ! e W T ( c , d ) . The product of the three factorial factors is a ! b ! a!b! a ! b ! . We have therefore proved, coefficient by coefficient,
∑ a , b per C a , b a ! b ! z a w b = exp ( z T W w ) Q ( z , w ) . (100) \sum_{a,b}\frac{\operatorname{per} C_{a,b}}{a!b!}z^{a}w^{b}
=\exp\left(z^{\mathsf T}Ww\right)Q(z,w). \tag*{(100)} a , b ∑ a ! b ! per C a , b z a w b = exp ( z T W w ) Q ( z , w ) . (100)
The sum here ranges over all balanced multiindices; it is a formal identity. To recover Q Q Q only degrees ∣ a ∣ = ∣ b ∣ ≤ n |a|=|b|\le n ∣ a ∣ = ∣ b ∣ ≤ n are needed.
Multiply (100) by exp ( − z T W T w ) \exp(-z^{\mathsf T}W^{\mathsf T}w) exp ( − z T W T w ) and compare these coefficients. Using Lemma 5.2 gives
[ z a w b ] Q ( z , w ) = ∑ c ≤ a , d ≤ b ∣ c ∣ = ∣ d ∣ e − W T ( c , d ) R a − c , b − d ( a − c ) ! ( b − d ) ! = Q a , b . [z^{a}w^{b}]Q(z,w)
=\sum_{\substack{c\le a,\ d\le b\\ |c|=|d|}}
e_{-W^{\mathsf T}}(c,d)
\frac{R_{a-c,b-d}}{(a-c)!(b-d)!}
=Q_{a,b}. [ z a w b ] Q ( z , w ) = c ≤ a , d ≤ b ∣ c ∣ = ∣ d ∣ ∑ e − W T ( c , d ) ( a − c )! ( b − d )! R a − c , b − d = Q a , b .
It remains to join the unmatched original rows and columns using U L V ULV U L V . First expand the permanent according to the entries taken from C C C and those taken from U L V ULV U L V . This is the usual formula
per ( C + U L V ) = ∑ I , J ⊆ [ n ] ∣ I ∣ = ∣ J ∣ per C [ I , J c ] per ( U L V ) [ I , J ] . \operatorname{per}(C+ULV)
=\sum_{\substack{I,J\subseteq[n]\\ |I|=|J|}}
\operatorname{per} C[I,J^{c}]\operatorname{per}(ULV)[I,J]. per ( C + U L V ) = I , J ⊆ [ n ] ∣ I ∣ = ∣ J ∣ ∑ per C [ I , J c ] per ( U L V ) [ I , J ] .
For each of the latter edges expand ( U L V ) i j = ∑ α , β U i α L α β V β j (ULV)_{ij}=\sum_{\alpha,\beta}U_{i\alpha}L_{\alpha\beta}V_{\beta j} ( U L V ) ij = ∑ α , β U i α L α β V β j . Once the U U U and V V V indices have been fixed, with multiplicities a , b a,b a , b , the sum over bijections between the unmatched rows and columns is a ! b ! e L ( a , b ) a!b!e_L(a,b) a ! b ! e L ( a , b ) . This follows from the same counting by the table M M M as above, now with weights L α β L_{\alpha\beta} L α β . The contributions of the C C C , U U U and V V V entries are exactly the coefficient Q a , b Q_{a,b} Q a , b in (99) . Summing gives (16) , proving (98) . □
Obtaining the polynomial entries
Let U ( t ) , V ( t ) U(t),V(t) U ( t ) , V ( t ) be the matrices in (12) , with the nodes from (22) , and put
A ( t ) = C + U ( t ) H ( t ) V ( t ) . A(t)=C+U(t)H(t)V(t). A ( t ) = C + U ( t ) H ( t ) V ( t ) .
We next show that A ( t ) A(t) A ( t ) has no pole at zero and specializes there to the matrix in (21) . All expansions in the argument are formal expansions at t = 0 t=0 t = 0 .
For each auxiliary column,
U ( t ) i α = t x i t − ξ α = − ∑ a ≥ 0 x i a ξ α − a − 1 t a + 1 . U(t)_{i\alpha}=\frac{t}{x_i t-\xi_\alpha}
=-\sum_{a\ge0}x_i^a\xi_\alpha^{-a-1}t^{a+1}. U ( t ) i α = x i t − ξ α t = − a ≥ 0 ∑ x i a ξ α − a − 1 t a + 1 .
The first r r r coefficient rows in this expansion are the rows of B x B_x B x . Consequently, for 0 ≤ a < r 0\le a<r 0 ≤ a < r ,
( U ( t ) B x − 1 T r ( t ) ) i a = x i a + O ( t r − a ) . \left(U(t)B_x^{-1}T_r(t)\right)_{ia}=x_i^a+O(t^{r-a}). ( U ( t ) B x − 1 T r ( t ) ) ia = x i a + O ( t r − a ) .
To see the cancellation explicitly, the coefficient row of t k + 1 t^{k+1} t k + 1 for k < r k<r k < r is x i k x_i^k x i k times row k k k of B x B_x B x . Multiplication by B x − 1 B_x^{-1} B x − 1 sends it to x i k x_i^k x i k times the k k k th unit row. Multiplication by T r ( t ) T_r(t) T r ( t ) then divides column a a a by t a + 1 t^{a+1} t a + 1 . Terms with k ≥ r k\ge r k ≥ r have exponent at least r − a > 0 r-a>0 r − a > 0 in that column. Likewise,
V ( t ) β j = t v β − y j t = ∑ b ≥ 0 v β b − 1 y j b + 1 , V(t)_{\beta j}=\frac{t}{v_{\beta}-y_jt}=\sum_{b\geq0}v_{\beta}^{b-1}y_j^{b+1}, V ( t ) β j = v β − y j t t = b ≥ 0 ∑ v β b − 1 y j b + 1 ,
so
( T s ( t ) B y − 1 V ( t ) ) b j = y j b + O ( t s − b ) ( 0 ≤ b < s ) . \left(T_s(t)B_y^{-1}V(t)\right)_{bj}=y_j^b+O\left(t^{s-b}\right)\qquad(0\leq b<s). ( T s ( t ) B y − 1 V ( t ) ) bj = y j b + O ( t s − b ) ( 0 ≤ b < s ) .
Using (22) , we conclude that
A ( t ) is regular at 0 , A ( 0 ) = A . (101) A(t)\text{ is regular at }0,\qquad A(0)=A. \tag*{(101)} A ( t ) is regular at 0 , A ( 0 ) = A . (101)
We now replace this specialization by a finite evaluation formula. Write
d i x ( t ) = ∏ α ( x i t − ξ α ) , d j y ( t ) = ∏ β ( v β − y j t ) . d_i^x(t)=\prod_{\alpha}(x_it-\xi_\alpha),\qquad d_j^y(t)=\prod_{\beta}(v_\beta-y_jt). d i x ( t ) = α ∏ ( x i t − ξ α ) , d j y ( t ) = β ∏ ( v β − y j t ) .
The expression
t r + s − 2 d i x ( t ) d j y ( t ) A ( t ) i j (102) t^{r+s-2}d_i^x(t)d_j^y(t)A(t)_{ij} \tag*{(102)} t r + s − 2 d i x ( t ) d j y ( t ) A ( t ) ij (102)
is a polynomial of degree at most 2 r + 2 s − 2 2r+2s-2 2 r + 2 s − 2 . Indeed, the Cauchy term has this degree bound immediately. A term of the perturbation containing H a b H_{ab} H ab has a factor t − a − b t^{-a-b} t − a − b before multiplication by t r + s − 2 t^{r+s-2} t r + s − 2 , because the two Cauchy factors contribute t 2 t^2 t 2 and the two diagonal factors contribute t − a − b − 2 t^{-a-b-2} t − a − b − 2 . The remaining denominators are one factor of d i x d_i^x d i x and one factor of d j y d_j^y d j y . After clearing them, its degree is at most 2 r + 2 s − 4 − a − b 2r+2s-4-a-b 2 r + 2 s − 4 − a − b , and its lowest possible exponent is nonnegative since a + b ≤ r + s − 2 a+b\leq r+s-2 a + b ≤ r + s − 2 .
Multiplying (102) over a permutation uses every d i x d_i^x d i x and every d j y d_j^y d j y once. Thus
F ( t ) = t m 0 q ( t ) per A ( t ) (103) F(t)=t^{m_0}q(t)\operatorname{per} A(t) \tag*{(103)} F ( t ) = t m 0 q ( t ) per A ( t ) (103)
is a polynomial of degree at most d 0 d_0 d 0 . By (101) ,
[ t m 0 ] F ( t ) = q ( 0 ) per A . (104) [t^{m_0}]F(t)=q(0)\operatorname{per} A. \tag*{(104)} [ t m 0 ] F ( t ) = q ( 0 ) per A . (104)
Lagrange interpolation at t 0 , … , t d 0 t_0,\ldots,t_{d_0} t 0 , … , t d 0 gives [ t m 0 ] F ( t ) = ∑ ℓ λ ℓ F ( t ℓ ) [t^{m_0}]F(t)=\sum_{\ell}\lambda_\ell F(t_\ell) [ t m 0 ] F ( t ) = ∑ ℓ λ ℓ F ( t ℓ ) . At each chosen point the auxiliary nodes are admissible, so Lemma 5.3 evaluates per A ( t ℓ ) \operatorname{per} A(t_\ell) per A ( t ℓ ) as D \mathcal{D} D . Substitution in (104) proves (26) . In particular, the final formula contains no limit.
Number of operations and bit lengths
At one evaluation point, there are D n D_n D n quotients in (14) . Their matrices have order at most 2 n 2n 2 n , so their construction and evaluation take O r , s ( D n n 3 ) O_{r,s}(D_n n^3) O r , s ( D n n 3 ) field operations. The factorials and the powers of each node difference needed in (13) can be computed in advance.
Each of the r s rs r s factors in (10) has at most n + 1 n+1 n + 1 terms. When multiplying them we retain only indices in I n \mathcal{I}_n I n . This takes O r , s ( n D n ) O_{r,s}(nD_n) O r , s ( n D n ) operations. The convolution (15) takes at most O ( D n 2 ) O(D_n^2) O ( D n 2 ) operations; the final sum (16) takes O ( D n ) O(D_n) O ( D n ) . Therefore one evaluation of D \mathcal{D} D takes
O r , s ( n r + s + 2 + n 2 r + 2 s − 2 ) O_{r,s}\left(n^{r+s+2}+n^{2r+2s-2}\right) O r , s ( n r + s + 2 + n 2 r + 2 s − 2 )
operations. There are d 0 + 1 = O r , s ( n ) d_0+1=O_{r,s}(n) d 0 + 1 = O r , s ( n ) evaluation points. Choosing them, forming H ( t ℓ ) H(t_\ell) H ( t ℓ ) and q ( t ℓ ) q(t_\ell) q ( t ℓ ) , and expanding the products in (25) add a polynomial number of operations bounded by (27) . This proves the field operation bound.
For completeness, suppose the input is rational. The chosen integers t ℓ t_\ell t ℓ have O r , s ( log ( n + 1 ) ) O_{r,s}(\log(n+1)) O r , s ( log ( n + 1 )) bits. Each entry of K q K_q K q is obtained from a rational node difference by a power of order O ( n ) O(n) O ( n ) and a factorial quotient of order O ( n ) O(n) O ( n ) . Its numerator and denominator therefore have polynomial bit length. Clearing the entry denominators in each determinant and using the determinant expansion bounds the resulting integer bit lengths by a polynomial as well. Exact fraction-free elimination consequently evaluates these determinants in polynomial bit complexity.
The coefficient products in (10) involve only r s rs r s different weights, exponents at most n n n , and factorials at most n ! n! n ! . A common denominator of polynomial bit length can be chosen for each coefficient array. The determinant quotients have polynomial bit length and there are only polynomially many of them. Their products and sums in (15) –(26) have common denominators of polynomial bit length. The same holds for the Lagrange coefficients, whose factors are differences of integers of size O r , s ( n ) O_{r,s}(n) O r , s ( n ) . Performing the rational operations exactly proves the asserted bit bound. □
The matrix classes and the rank bound
For Corollary 4.1, put x i = i x_i=i x i = i and y j = 1 − j y_j=1-j y j = 1 − j . Then x i − y j = i + j − 1 x_i-y_j=i+j-1 x i − y j = i + j − 1 , and R ( i , j ) = R ( x i , 1 − y j ) R(i,j)=R(x_i,1-y_j) R ( i , j ) = R ( x i , 1 − y j ) . For Theorem 4.9(6), use x i = − i x_i=-i x i = − i and y j = − α − j y_j=-\alpha-j y j = − α − j . Their difference is α + j − i \alpha+j-i α + j − i , and x i − y j − α = j − i x_i-y_j-\alpha=j-i x i − y j − α = j − i . For Theorem 4.11(3), use x i = i x_i=i x i = i and y j = − α − j y_j=-\alpha-j y j = − α − j ; the corresponding difference minus α \alpha α is i + j i+j i + j . In each case both lists consist of distinct nodes, and their disjointness is exactly the nonvanishing of the denominators. Substituting in Theorem 3.3 proves the three formulas and their degree and complexity assertions.
Proof of Proposition 4.2. Every square submatrix of C C C is a Cauchy matrix on distinct, disjoint nodes and has nonzero determinant. Hence each rectangular submatrix of C C C has the largest possible rank. The matrix ( P ( x i , y j ) ) (P(x_i,y_j)) ( P ( x i , y j )) factors as X H Y XHY X H Y , where X i a = x i a X_{ia}=x_i^a X ia = x i a and Y b j = y j b Y_{bj}=y_j^b Y bj = y j b , and has rank at most q 0 q_0 q 0 . The inequality rank ( B + D ) ≥ rank B − rank D \operatorname{rank}(B+D)\ge\operatorname{rank}B-\operatorname{rank}D rank ( B + D ) ≥ rank B − rank D now gives
rank A [ I , I c ] ≥ min ( ∣ I ∣ , n − ∣ I ∣ ) − q 0 . \operatorname{rank} A[I,I^c] \ge\min(|I|,n-|I|)-q_0. rank A [ I , I c ] ≥ min ( ∣ I ∣ , n − ∣ I ∣ ) − q 0 .
The same argument applies in the reverse direction. Both ranks are nonnegative, proving the first assertion.
Every rooted binary tree with n ≥ 2 n\ge2 n ≥ 2 leaves has a node with between n / 3 n/3 n /3 and 2 n / 3 2n/3 2 n /3 leaves. To verify this, start at the root and descend to a child having more than 2 n / 3 2n/3 2 n /3 leaves for as long as such a child exists. The last node reached has more than 2 n / 3 2n/3 2 n /3 leaves and each child has at most 2 n / 3 2n/3 2 n /3 . Its larger child has more than n / 3 n/3 n /3 leaves and gives the required node. For this node both ∣ I ∣ |I| ∣ I ∣ and n − ∣ I ∣ n-|I| n − ∣ I ∣ are at least ⌈ n / 3 ⌉ \lceil n/3\rceil ⌈ n /3 ⌉ . The second assertion follows. □
Proof of Theorem 3.4
Let I i = { 1 , … , i } I_i=\{1,\ldots,i\} I i = { 1 , … , i } . Choose a basis F i F_i F i of the row space of A [ I i , I i c ] A[I_i,I_i^c] A [ I i , I i c ] , with zero rows appended to give r r r rows. Restrict F i − 1 F_{i-1} F i − 1 to columns i + 1 , … , n i+1,\ldots,n i + 1 , … , n . Each restricted row belongs to the row space of F i F_i F i , so there is a matrix B i B_i B i with
F i − 1 [ : , { i + 1 , … , n } ] = B i F i . F_{i-1}[:,\{i+1,\ldots,n\}]=B_iF_i. F i − 1 [ : , { i + 1 , … , n }] = B i F i .
Let q i q_i q i be the column of F i − 1 F_{i-1} F i − 1 with index i i i , and choose p i p_i p i so that the part of row i i i to its right is p i T F i p_i^{\mathsf T}F_i p i T F i . If j < i j<i j < i , applying the restriction identity at cuts j + 1 , … , i − 1 j+1,\ldots,i-1 j + 1 , … , i − 1 gives
A [ j , { i , … , n } ] = p j T B j + 1 ⋯ B i − 1 F i − 1 . A[j,\{i,\ldots,n\}]=p_j^{\mathsf T}B_{j+1}\cdots B_{i-1}F_{i-1}. A [ j , { i , … , n }] = p j T B j + 1 ⋯ B i − 1 F i − 1 .
Taking column i i i gives the upper formula in (28) . Applying the construction to A T A^{\mathsf T} A T gives the lower formula. In forming F i F_i F i , the restricted previous basis and the new row already span the required space. Elimination involves at most r + 1 r+1 r + 1 rows of length n − i n-i n − i . On the lower side there are at most s + 1 s+1 s + 1 such rows. Summing these costs over the cuts gives O r , s ( n 2 ) O_{r,s}(n^2) O r , s ( n 2 ) .
Consider a matching between the rows and columns of I i − 1 I_{i-1} I i − 1 . We give it the product of its entries, together with a factor
p j T B j + 1 ⋯ B i − 1 x for each unmatched row j , p_j^{\mathsf T}B_{j+1}\cdots B_{i-1}x \qquad\text{for each unmatched row }j, p j T B j + 1 ⋯ B i − 1 x for each unmatched row j ,
u k T C k + 1 ⋯ C i − 1 y for each unmatched column k . u_k^{\mathsf T}C_{k+1}\cdots C_{i-1}y \qquad\text{for each unmatched column }k. u k T C k + 1 ⋯ C i − 1 y for each unmatched column k .
Let f f f be the sum of these products. When D i D_i D i acts on the factor for row j j j , the result is p j T B j + 1 ⋯ B i − 1 q i = a j i p_j^{\mathsf T}B_{j+1}\cdots B_{i-1}q_i=a_{ji} p j T B j + 1 ⋯ B i − 1 q i = a j i . In the product rule, that factor is removed and all other factors are retained. Thus a summand of D i f D_if D i f uses the edge from row j j j to column i i i . In the same way E i E_i E i removes the factor for an earlier column k k k and gives the entry a i k a_{ik} a ik . On any factor that remains, changing x x x to B i x B_ix B i x or y y y to C i y C_iy C i y appends B i B_i B i or C i C_i C i to its product. This is the change to the next cut.
If both new vertices are joined to earlier vertices, the two differentiations give Φ i ( D i E i f ) \Phi_i(D_iE_if) Φ i ( D i E i f ) . If only column i i i is joined to an earlier row, the new row remains unmatched and gives L i Φ i ( D i f ) L_i\Phi_i(D_if) L i Φ i ( D i f ) . If only row i i i is joined to an earlier column, the new column gives M i Φ i ( E i f ) M_i\Phi_i(E_i f) M i Φ i ( E i f ) . In the two cases with no edge to an earlier vertex, either the new row and column are matched to each other, giving d i Φ i f d_i\Phi_i f d i Φ i f , or both are left unmatched, giving L i M i Φ i f L_iM_i\Phi_i f L i M i Φ i f . Every matching on I i I_i I i restricts to one matching on I i − 1 I_{i-1} I i − 1 and belongs to exactly one of these five cases. In particular the changes of coordinates act on the retained factors before the new factors L i , M i L_i,M_i L i , M i are inserted.
After i i i steps, if there are h h h unmatched rows there are also h h h unmatched columns. A completion has to match these rows to distinct columns outside I i I_i I i and these columns to distinct rows outside I i I_i I i . Hence h ≤ i h \le i h ≤ i and h ≤ n − i h \le n-i h ≤ n − i . A term with h > N h>N h > N at any cut cannot contribute to the final constant term. This justifies Π \Pi Π and gives the operator U i U_i U i . Starting at the constant monomial, the coefficient of a monomial g g g after step i i i is obtained by multiplying the preceding coefficient row vector by W i W_i W i . At the last cut the constant coefficient counts precisely the perfect matchings. This gives the product in (31) , and (17) gives the determinant. To count the states, a monomial with degrees h , h h,h h , h is a choice of a weak composition of h h h into r r r parts and one into s s s parts. Summing their product over 0 ≤ h ≤ N 0\le h\le N 0 ≤ h ≤ N gives (30) .
We apply U i U_i U i to its input polynomial without forming the matrix W i W_i W i . The numbers of monomials of degrees h , h h,h h , h , or of degrees differing by one after a differentiation, are O r , s ( ( h + 1 ) r + s − 2 ) O_{r,s}((h+1)^{r+s-2}) O r , s (( h + 1 ) r + s − 2 ) . There are O r , s ( n r + s − 1 ) O_{r,s}(n^{r+s-1}) O r , s ( n r + s − 1 ) monomials in all. Row and column elimination reduce any of the fixed-order matrices B i , C i B_i,C_i B i , C i to a diagonal matrix of zeros and ones. Reversing the invertible operations expresses its substitution as a bounded number of permutations, scalings, elementary additions of one variable to another with a scalar multiple, and zero substitutions. For example, a substitution x j ↦ x j + λ x k x_j\mapsto x_j+\lambda x_k x j ↦ x j + λ x k replaces x j a x_j^a x j a by its binomial expansion with at most a + 1 ≤ n + 1 a+1\le n+1 a + 1 ≤ n + 1 terms. The cost of a step is therefore O r , s ( n r + s ) O_{r,s}(n^{r+s}) O r , s ( n r + s ) . When r = s = 1 r=s=1 r = s = 1 , each linear substitution is just a scaling and the cost of a step is O ( n ) O(n) O ( n ) . Together with the construction of the generators, these are the required bounds.
Proof of Corollary 3.5
Process the indices in order. At index i i i , the new row and column may both be matched to previous endpoints, exactly one may be so matched, they may form the diagonal edge, or both may remain unmatched. These are respectively the first, next two, diagonal, and L i M i L_iM_i L i M i terms of U i ( p ) U_i^{(p)} U i ( p ) . Each derivative chooses a previous endpoint, and Φ i \Phi_i Φ i advances the remaining endpoint factors by B i B_i B i or C i C_i C i . The products of generators are exactly (28) . Thus the constant-to-constant entry counts complete matchings. Lemma 5.1 permits reduction after each step. Finally, det ( I − W ( p ) ) = 1 \det(I-W^{(p)})=1 det ( I − W ( p ) ) = 1 and ( I − W ( p ) ) − 1 = I + W ( p ) + ⋯ + ( W ( p ) ) n (I-W^{(p)})^{-1}=I+W^{(p)}+\cdots+(W^{(p)})^n ( I − W ( p ) ) − 1 = I + W ( p ) + ⋯ + ( W ( p ) ) n . A Schur complement proves the second identity. Each step acts on a fixed basis of size D p , r , s D_{p,r,s} D p , r , s , with p O ( r + s ) p^{O(r+s)} p O ( r + s ) work to construct and apply its matrix.
Proof of Theorem 3.6
A partial cover of I I I consists of directed edges inside I I I such that each vertex has at most one incoming and one outgoing edge. Its components are directed paths and cycles. We include an isolated vertex among the paths, with the same start and end. Give each component a color c c c , include x c x_c x c for every vertex of that color, and multiply by the edge weights. A cycle also contributes z c z_c z c . For a path of color c c c starting at j j j and ending at i i i , include the factor
R I ( i , : ) Z I ( c ) C I ( j , : ) T . (105) R_I(i,:)Z_I^{(c)}C_I(j,:)^{\mathsf T}. \tag*{(105)} R I ( i , : ) Z I ( c ) C I ( j , : ) T . (105)
Let Q I Q_I Q I be the sum of the resulting weights. For a single vertex i i i of color c c c , the isolated path gives x c R I { i : } Z I ( c ) C I { i : } T x_cR_I\{i\,: \}Z_I^{(c)}C_I\{i\,: \}^{\mathsf T} x c R I { i : } Z I ( c ) C I { i : } T . The loop gives x c z c a i i x_c z_c a_{ii} x c z c a ii . Adding over c c c gives the coefficients l i c y c l_i^{\mathrm{cyc}} l i cyc .
At a merge we view the child variables as the two diagonal blocks of Z ( c ) Z^{(c)} Z ( c ) . Write r i , c j r_i,c_j r i , c j for the row vectors of the end i i i and the start j j j , padded with zeros on the coordinates of the other child. Thus r i B c j T r_iB c_j^{\mathsf T} r i B c j T is a i j a_{ij} a ij when i , j i,j i , j belong to different children, and is zero when they belong to the same child. This follows from (3) and the zero diagonal blocks of B B B .
In the first term of (35) , ∂ α β ( c ) \partial_{\alpha\beta}^{(c)} ∂ α β ( c ) applied to r i Z ( c ) c j T r_i Z^{(c)} c_j^{\mathsf T} r i Z ( c ) c j T gives r i ( α ) c j ( β ) r_i(\alpha)c_j(\beta) r i ( α ) c j ( β ) . After summation the contribution is z c r i B c j T z_c r_i B c_j^{\mathsf T} z c r i B c j T . This adds the edge from i i i to j j j , removes the two endpoints, and records the newly closed cycle. In the second term take paths from j j j to i i i and from j ′ j' j ′ to i ′ i' i ′ , both of color c c c . Each path factor is linear in Z ( c ) Z^{(c)} Z ( c ) . To add the edge i → j ′ i \to j' i → j ′ , the derivatives take r i ( α ) c j ( δ ) r_i(\alpha)c_j(\delta) r i ( α ) c j ( δ ) from the first factor and r i ′ ( γ ) c j ′ ( β ) r_{i'}(\gamma)c_{j'}(\beta) r i ′ ( γ ) c j ′ ( β ) from the second. The sums with B α β B_{\alpha\beta} B α β and Z γ δ ( c ) Z_{\gamma\delta}^{(c)} Z γ δ ( c ) give
( r i B c j ′ T ) ( r i ′ Z ( c ) c j T ) . (r_i B c_{j'}^{\mathsf T})(r_{i'} Z^{(c)} c_j^{\mathsf T}). ( r i B c j ′ T ) ( r i ′ Z ( c ) c j T ) .
The first scalar is the edge weight, and the other factor records the new path from j j j to i ′ i' i ′ . Taking the derivatives on the factors in the other order accounts for the choice i ′ → j i' \to j i ′ → j . The derivatives cannot take both factors from the same path. The two paths must have the same color because each summand uses a single c c c .
Every partial cover on I I I restricts to partial covers on U , V U,V U , V , with the colors inherited from its components. The edges lost on restriction are the crossing edges. Starting with these restrictions, they can be restored in any order. At an intermediate step the end and start needed by the next edge are still available because that edge has yet to be added. A cycle is recorded only when its last edge has been restored, so its z c z_c z c is included once in every ordering. Consequently ℓ \ell ℓ crossing edges are counted ℓ ! \ell! ℓ ! times by E I ℓ E_I^\ell E I ℓ . The coordinate change takes a remaining factor to
r i S I Z I ( c ) T I T c j T = R I ( i , : ) Z I ( c ) C I ( j , : ) T . r_i S_I Z_I^{(c)} T_I^{\mathsf T} c_j^{\mathsf T}
= R_I(i,:) Z_I^{(c)} C_I(j,:)^{\mathsf T}. r i S I Z I ( c ) T I T c j T = R I ( i , : ) Z I ( c ) C I ( j , : ) T .
We obtain the identity
Q I = η I ( ∑ ℓ = 0 ∣ I ∣ E I ℓ ℓ ! Q U Q V ) . Q_I=\eta_I\left(\sum_{\ell=0}^{|I|}\frac{E_I^\ell}{\ell!}Q_UQ_V\right). Q I = η I ℓ = 0 ∑ ∣ I ∣ ℓ ! E I ℓ Q U Q V .
Taking coefficients in the endpoint monomials gives the products of ℓ i c y c \ell_i^{\mathrm{cyc}} ℓ i cyc and t i c y c t_i^{\mathrm{cyc}} t i cyc in (37) . The root has no endpoint variables, so only closed covers survive there. Each closed cover is a permutation, and summing its colors gives the cycle factors in (33) .
For one color, set the cycle and vertex markers to one and substitute Z I ( 1 ) = x I y I T Z_I^{(1)}=x_Iy_I^{\mathsf T} Z I ( 1 ) = x I y I T in the partial-cover polynomial. Each path contributes the independent row and column factors in (94) . The starts and ends are exactly the unmatched columns and rows in the associated matching. This identifies the two partial polynomials in this specialization.
The matrix of endpoint variables at a merge has at most 2 r 2r 2 r rows and 2 s 2s 2 s columns for each color. Together with the markers there are at most 4 d r s + 2 d 4drs+2d 4 d r s + 2 d variables. Each family of markers has degree at most n n n , as do the endpoint variables in the product of the child polynomials. The number of coefficients is polynomial in n n n for fixed d d d , r r r , s s s . The rational bit bound can be seen as in the permanent calculation. In a partial cover there are at most n n n edges and at most two coordinate factors per path. The sum of the edge count and twice the path count is at most 2 n 2n 2 n . There are at most ( n + 1 ) n (n+1)^n ( n + 1 ) n choices of outgoing edges before the indegree restriction, at most d n d^n d n colorings, and a bounded number of coordinate terms per factor. A common denominator of polynomial bit length raised to a power at most 2 n 2n 2 n suffices for the state coefficients. The linear substitutions and the factorials at a merge give polynomial bit length for the intermediate coefficients as well.
Along an ordering, the restriction of a full permutation to I i I_i I i has, say, h h h paths. The h h h outgoing edges from their ends must use distinct targets outside I i I_i I i , so h ≤ min ( i , n − i ) h\leq\min(i,n-i) h ≤ min ( i , n − i ) . Each path contributes one endpoint variable to a monomial. We can therefore project at every cut to total degree at most N N N without losing any term that can survive at the root. In at most d r s drs d r s endpoint variables the number of retained monomials is at most ( N + d r s d r s ) \binom{N+drs}{drs} ( d r s N + d r s ) . When I i I_i I i is obtained by merging I i − 1 I_{i-1} I i − 1 with the leaf { i } \{i\} { i } , the leaf polynomial is fixed and the operation is linear in the coefficient vector at I i − 1 I_{i-1} I i − 1 . Its coefficients are the stated entries of W i c y c W_i^{\mathrm{cyc}} W i cyc . The first leaf starts from the constant state. Thus the constant-to-constant entry is F A F_A F A , and (17) gives (38) with the stated order.
If a tree with the stated rank bounds exists, then rw F ( A ) ≤ r + s \operatorname{rw}_{F}(A) \le r+s rw F ( A ) ≤ r + s . The construction in the proof of Theorem 3.1 finds a tree on which both ranks are at most r + s r+s r + s . These bounds suffice for the polynomial-time calculation of the cycle polynomial.
Proof of Theorem 3.7
Proof of Theorem 3.7. Fix v 1 , … , v k v_{1},\ldots,v_{k} v 1 , … , v k in (3.25). When we expand the products of permanents and determinants, a permutation can occur for a choice of the sets only if each of its cycles lies in a single S c S_{c} S c or T c T_{c} T c . A cycle of length a a a in S c S_{c} S c contributes its usual weight. In T c T_{c} T c its sign in the determinant is ( − 1 ) a − 1 (-1)^{a-1} ( − 1 ) a − 1 , and it accounts for a factor ( − 1 ) a (-1)^{a} ( − 1 ) a in ( − 1 ) ∣ T c ∣ (-1)^{|T_{c}|} ( − 1 ) ∣ T c ∣ . The total extra sign for that cycle is therefore − 1 -1 − 1 .
If the permutation has a cycle with no mark, choose such a cycle by its least vertex. Move the whole cycle from S c S_{c} S c to T c T_{c} T c , or from T c T_{c} T c to S c S_{c} S c , within the pair it occupies. The union S c ∪ T c S_{c}\cup T_{c} S c ∪ T c is unchanged and the cycle contains none of the prescribed v c v_{c} v c . The new choice therefore satisfies the same restrictions. This operation pairs two occurrences of the same weight with opposite signs. If every cycle has a mark, the cycle containing v c v_{c} v c has to lie in S c S_{c} S c . Two marks cannot lie in that cycle, since different marks are assigned to different S c S_{c} S c ’s. The minimum condition requires v c v_{c} v c to be the least vertex of its cycle. The permutation then occurs once with positive sign. Summing over marks gives (3.25). With one mark fixed and the minimum condition dropped, the same pairing leaves exactly the single cycles.
For the determinant factor at a merge, fix one determinant color and let L ⊆ I L\subseteq I L ⊆ I be the vertices having this color. Omit the color superscript on u I , v I u_{I},v_{I} u I , v I . Introduce exterior variables θ i , ϕ i \theta_{i},\phi_{i} θ i , ϕ i for i ∈ L i\in L i ∈ L , anticommuting with one another and with the cut variables. Set Θ L = ∏ i ∈ L ( θ i ϕ i ) \Theta_{L}=\prod_{i\in L}(\theta_{i}\phi_{i}) Θ L = ∏ i ∈ L ( θ i ϕ i ) . The order of these pairs is immaterial, since each pair has even degree. With interior variables written before the cut variables, consider
( − 1 ) ∣ L ∣ [ Θ L ] exp ( ∑ i , j ∈ L θ i a i j ϕ j + ∑ i ∈ L θ i R I ( i , : ) u I + ∑ j ∈ L ( C I ( j , : ) v I ) ϕ j ) . (106) (-1)^{|L|}[\Theta_{L}]\exp\left(\sum_{i,j\in L}\theta_{i}a_{ij}\phi_{j}+\sum_{i\in L}\theta_{i}R_{I}(i,:)u_{I}+\sum_{j\in L}(C_{I}(j,:)v_{I})\phi_{j}\right). \tag*{(106)} ( − 1 ) ∣ L ∣ [ Θ L ] exp i , j ∈ L ∑ θ i a ij ϕ j + i ∈ L ∑ θ i R I ( i , : ) u I + j ∈ L ∑ ( C I ( j , : ) v I ) ϕ j . (106)
All terms in the exponent have even degree and its exponential is a finite sum. If a term uses a a a internal edges, it uses a a a of the θ \theta θ ’s and a a a of the ϕ \phi ϕ ’s in those edges. For it to survive extraction of Θ L \Theta_{L} Θ L , the other ∣ L ∣ − a |L|-a ∣ L ∣ − a occurrences of each kind must come from the cut terms. In particular the u I u_{I} u I -degree and the v I v_{I} v I -degree are equal. At a leaf the coefficient before the factor − 1 -1 − 1 is
a i i + ( R { i } ( i , : ) u { i } ) ( C { i } ( i , : ) v { i } ) . a_{ii}+\left(R_{\{i\}}(i,:)u_{\{i\}}\right)\left(C_{\{i\}}(i,:)v_{\{i\}}\right). a ii + ( R { i } ( i , : ) u { i } ) ( C { i } ( i , : ) v { i } ) .
Indeed the two factors linear in the cut variables give θ i ( R { i } ( i , : ) u { i } ) ( C { i } ( i , : ) v { i } ) ϕ i \theta_{i}\left(R_{\{i\}}(i,:)u_{\{i\}}\right)\left(C_{\{i\}}(i,:)v_{\{i\}}\right)\phi_{i} θ i ( R { i } ( i , : ) u { i } ) ( C { i } ( i , : ) v { i } ) ϕ i . Moving ϕ i \phi_{i} ϕ i past the two cut factors has positive sign. This proves the second line of (3.23). At the root there are no cut terms. A surviving product is ∏ i ∈ L ( θ i a i , σ ( i ) ϕ σ ( i ) ) \prod_{i\in L}(\theta_{i}a_{i,\sigma(i)}\phi_{\sigma(i)}) ∏ i ∈ L ( θ i a i , σ ( i ) ϕ σ ( i ) ) for a permutation of L L L . The even factors can be ordered by i i i , and putting the ϕ \phi ϕ ’s in their natural places multiplies the coefficient by sgn ( σ ) \operatorname{sgn}(\sigma) sgn ( σ ) . The orderings of the even factors cancel the factorial in the exponential. The resulting root value is ( − 1 ) ∣ L ∣ det A [ L , L ] (-1)^{|L|}\det A[L,L] ( − 1 ) ∣ L ∣ det A [ L , L ] .
For the merge let H H H be the sum of the two child exponents. The two sets of interior variables are disjoint. Taking the two interior coefficients of exp H \exp H exp H therefore gives the product of the child polynomials, including their two signs. The interior coefficients have even degree in the interior variables, so passing cut derivatives across them introduces no sign. We can consequently perform the differentiation before taking the interior coefficient.
For an ordered pair of different children J , K J,K J , K put
A α = ∑ i ∈ L ∩ J R J ( i , α ) θ i , B β = ∑ j ∈ L ∩ K C K ( j , β ) ϕ j . A_{\alpha}=\sum_{i\in L\cap J}R_{J}(i,\alpha)\theta_{i},\qquad B_{\beta}=\sum_{j\in L\cap K}C_{K}(j,\beta)\phi_{j}. A α = i ∈ L ∩ J ∑ R J ( i , α ) θ i , B β = j ∈ L ∩ K ∑ C K ( j , β ) ϕ j .
By the rule for a left derivative, ∂ u J , α H = − A α \partial_{u_{J,\alpha}}H=-A_{\alpha} ∂ u J , α H = − A α and ∂ v K , β H = B β \partial_{v_{K,\beta}}H=B_{\beta} ∂ v K , β H = B β . As H H H is even, it commutes with its derivatives, and ∂ ξ exp H = ( ∂ ξ H ) exp H \partial_{\xi}\exp H=(\partial_{\xi}H)\exp H ∂ ξ exp H = ( ∂ ξ H ) exp H . It follows that
∂ v K , β ∂ u J , α exp H = ∂ v K , β ( − A α exp H ) = A α B β exp H . \partial_{v_{K,\beta}}\partial_{u_{J,\alpha}}\exp H=\partial_{v_{K,\beta}}(-A_{\alpha}\exp H)=A_{\alpha}B_{\beta}\exp H. ∂ v K , β ∂ u J , α exp H = ∂ v K , β ( − A α exp H ) = A α B β exp H .
The last sign is positive because the v v v -derivative crosses the odd factor A α A_\alpha A α . Multiplying by the entries of B J K B_{JK} B J K and summing over the indices produces ∑ i ∈ L ∩ J , j ∈ L ∩ K θ i a i j φ j \sum_{i\in L\cap J,j\in L\cap K}\theta_i a_{ij}\varphi_j ∑ i ∈ L ∩ J , j ∈ L ∩ K θ i a ij φ j by (3) . This is exactly the missing part of the internal exponent. It contains no cut variables and has even degree, so repeating the derivative operator multiplies by successive powers of this part. Summing with the factorial denominators inserts its exponential. Terms of power greater than ∣ I ∣ |I| ∣ I ∣ vanish, since they repeat an interior variable. The final substitutions give the coordinate terms at I I I , by R J S J , I = R I [ J , : ] R_JS_{J,I}=R_I[J,:] R J S J , I = R I [ J , : ] and C J T J , I = C I [ J , : ] C_JT_{J,I}=C_I[J,:] C J T J , I = C I [ J , : ] . This proves the exterior merge rule.
For each permanent color the same merge is given by the proof of Theorem 3.1 , with the matchings restricted to the vertices of that color. For fixed marks, a marked vertex v c v_c v c is assigned the permanent color of pair c c c . An unmarked vertex i i i may take either color of a pair with v c < i v_c<i v c < i . Multiplication of the leaf sums lists the allowed assignments, and linearity of each merge sums their contributions in the coefficient arrays. There are ( n k ) \binom{n}{k} ( k n ) choices of marks. For fixed k , r , s k,r,s k , r , s , the number of variables at each merge is bounded, commuting degrees are at most 2 n 2n 2 n , and exterior exponents are zero or one. This gives polynomially many coefficients and operations. Over the rationals the factors in each state coefficient come from the input and the cut coordinates, with at most 2 n 2n 2 n per product. The number of color assignments is at most ( 2 k ) n (2k)^n ( 2 k ) n . The same common-denominator argument as for (94) , with the extra factorials from a single merge, gives the rational bit bound.
In the ordered calculation we can sum over marks in the transfer itself. A pair can first appear only at its least vertex, in its permanent color. This is exactly the transition that increases ℓ \ell ℓ , and all other transitions use a pair already introduced. Thus an allowed partition determines a unique sequence of these choices, with ℓ = 0 \ell=0 ℓ = 0 at the start and ℓ = k \ell=k ℓ = k at the end. At cut I i I_i I i a term that can contribute has at most min ( i , n − i ) \min(i,n-i) min ( i , n − i ) unmatched rows of any color. Projection to degree at most N N N in the commuting variables therefore preserves the root coefficient.
For one permanent color the count of surviving monomials is at most D + D^+ D + . For one determinant color, if the two exterior degrees are h , h h,h h , h , there are at most ( r h ) ( s h ) \binom{r}{h}\binom{s}{h} ( h r ) ( h s ) choices of variables. Summing gives
∑ h = 0 min ( r , s ) ( r h ) ( s h ) = ∑ h ( r h ) ( s s − h ) = ( r + s s ) = D − . \sum_{h=0}^{\min(r,s)}\binom{r}{h}\binom{s}{h}
=\sum_{h}\binom{r}{h}\binom{s}{s-h}
=\binom{r+s}{s}=D^-. h = 0 ∑ m i n ( r , s ) ( h r ) ( h s ) = h ∑ ( h r ) ( s − h s ) = ( s r + s ) = D − .
The middle sum counts the choices of s s s elements from two sets of sizes r , s r,s r , s , according to how many come from the first set. Including all k k k pairs and ℓ \ell ℓ , the number of coordinates is at most ( k + 1 ) ( D + D − ) k = D k (k+1)(D^+D^-)^k=D_k ( k + 1 ) ( D + D − ) k = D k . The constant-to-constant entry of the ordered transfer, with the prescribed end values of ℓ \ell ℓ , is C k ( A ) C_k(A) C k ( A ) . Equation (17) gives (43) . The summand for h h h in D + D^+ D + is O r , s ( ( h + 1 ) r + s − 2 ) O_{r,s}((h+1)^{r+s-2}) O r , s (( h + 1 ) r + s − 2 ) , so D + = O r , s ( n r + s − 1 ) D^+=O_{r,s}(n^{r+s-1}) D + = O r , s ( n r + s − 1 ) . This gives the stated order and, at k = 1 k=1 k = 1 , the Hamiltonian-cycle bound.
Proof of Corollary 3.8
Take d = p + q d=p+q d = p + q in (33) . In the first p p p colors use z c = 1 z_c=1 z c = 1 . In color p + j p+j p + j use z p + j = − 1 z_{p+j}=-1 z p + j = − 1 and x p + j = − y j x_{p+j}=-y_j x p + j = − y j . For a cycle of length ℓ \ell ℓ , its factor becomes
p ℓ ( x ∣ y ) = ∑ i = 1 p x i ℓ + ( − 1 ) ℓ − 1 ∑ j = 1 q y j ℓ . p_\ell(x\mid y)=\sum_{i=1}^{p}x_i^\ell+(-1)^{\ell-1}\sum_{j=1}^{q}y_j^\ell. p ℓ ( x ∣ y ) = i = 1 ∑ p x i ℓ + ( − 1 ) ℓ − 1 j = 1 ∑ q y j ℓ .
For a cycle type μ \mu μ , the product of these power sums is p μ ( x ∣ y ) p_\mu(x\mid y) p μ ( x ∣ y ) . The Frobenius character formula expands p μ p_\mu p μ as ∑ λ ⊢ n χ λ ( μ ) s λ \sum_{\lambda\vdash n}\chi_\lambda(\mu)s_\lambda ∑ λ ⊢ n χ λ ( μ ) s λ . Applying the specialization to this identity for each permutation, its coefficient of s λ ( x ∣ y ) s_\lambda(x\mid y) s λ ( x ∣ y ) is the defining weighted character sum. Thus the specialized F A F_A F A is
∑ λ ⊢ n imm λ ( A ) s λ ( x ∣ y ) . (107) \sum_{\lambda\vdash n}\operatorname{imm}_\lambda(A)s_\lambda(x\mid y). \tag*{(107)} λ ⊢ n ∑ imm λ ( A ) s λ ( x ∣ y ) . (107)
The Schur specialization of Berele and Regev [32 ] gives
s λ ( x ∣ y ) = ∑ μ ⊆ λ s μ ( x ) s λ ′ / μ ′ ( y ) s_{\lambda}(x\mid y)=\sum_{\mu\subseteq\lambda}s_{\mu}(x)s_{\lambda^{\prime}/\mu^{\prime}}(y) s λ ( x ∣ y ) = μ ⊆ λ ∑ s μ ( x ) s λ ′ / μ ′ ( y )
where primes denote conjugate partitions. A term with s μ ( x ) ≠ 0 s_{\mu}(x)\ne0 s μ ( x ) = 0 has μ p + 1 = 0 \mu_{p+1}=0 μ p + 1 = 0 . If λ p + 1 > q \lambda_{p+1}>q λ p + 1 > q , column p + 1 p+1 p + 1 of λ ′ / μ ′ \lambda^{\prime}/\mu^{\prime} λ ′ / μ ′ then has length λ p + 1 \lambda_{p+1} λ p + 1 . A column of a semistandard tableau uses distinct entries, so this skew Schur polynomial in q q q variables vanishes. Hence only the partitions in the ( p , q ) (p,q) ( p , q ) -hook contribute to (107) .
For a partition in the hook, the largest possible x x x -degree is λ 1 + ⋯ + λ p \lambda_1+\cdots+\lambda_p λ 1 + ⋯ + λ p . It is attained only for μ = ( λ 1 , … , λ p ) \mu=(\lambda_1,\ldots,\lambda_p) μ = ( λ 1 , … , λ p ) . In the conjugate skew diagram the remaining boxes are the conjugate of the tail ( λ p + 1 , … ) (\lambda_{p+1},\ldots) ( λ p + 1 , … ) , up to horizontal translation. The component of this largest x x x -degree is therefore
s ( λ 1 , … , λ p ) ( x ) s ( λ p + 1 , λ p + 2 , … ) ′ ( y ) . s_{(\lambda_1,\ldots,\lambda_p)}(x)s_{(\lambda_{p+1},\lambda_{p+2},\ldots)^{\prime}}(y). s ( λ 1 , … , λ p ) ( x ) s ( λ p + 1 , λ p + 2 , … ) ′ ( y ) .
The second partition has at most q q q parts, and both Schur factors are nonzero. If a linear relation among the specialized Schur functions existed, take the largest x x x -degree of their highest components among the terms with nonzero coefficients. At this degree one obtains a relation among distinct products of Schur basis elements in the two alphabets. Such products are independent, and the two partitions determine λ \lambda λ . This rules out the relation.
There are at most ( n + 1 ) p + q (n+1)^{p+q} ( n + 1 ) p + q choices of the two partitions, by choosing their parts. For fixed p , q p,q p , q their polynomials have polynomially many monomials. Their coefficients can be obtained using Jacobi–Trudi with the complete functions having generating series
∏ j = 1 q ( 1 + y j t ) ∏ i = 1 p ( 1 − x i t ) . \frac{\prod_{j=1}^{q}(1+y_jt)}{\prod_{i=1}^{p}(1-x_it)}. ∏ i = 1 p ( 1 − x i t ) ∏ j = 1 q ( 1 + y j t ) .
For instance one can evaluate each polynomial on an interpolation grid with n + 1 n+1 n + 1 values in every variable and recover its coefficients. The arrays are linearly independent by the preceding argument, so solving (107) in the monomial basis recovers all the immanants. The numbers of rows and columns of this system are polynomial in n n n . The Schur coefficients count tableaux with a fixed alphabet, so their bit lengths are polynomial. For rational A A A the same holds for the other side of the system by Theorem 3.6 , and rational elimination gives the asserted bit bound.
For the ordinary hook λ = ( n − k , 1 k ) \lambda=(n-k,1^k) λ = ( n − k , 1 k ) , put x = 1 x=1 x = 1 and y = t y=t y = t . In the displayed Schur sum, μ \mu μ is a single row, and the skew diagram in the second factor must have at most one cell per column. Its first column has n − k − ∣ μ ∣ n-k-|\mu| n − k − ∣ μ ∣ cells and every other column has at most one. Thus ∣ μ ∣ = n − k |\mu|=n-k ∣ μ ∣ = n − k or n − k − 1 n-k-1 n − k − 1 . Their two weights are t k , t k + 1 t^k,t^{k+1} t k , t k + 1 , respectively. This gives s ( n − k , 1 k ) ( 1 ∣ t ) = ( 1 + t ) t k s_{(n-k,1^k)}(1\mid t)=(1+t)t^k s ( n − k , 1 k ) ( 1 ∣ t ) = ( 1 + t ) t k . The markers in this specialization are those of (44) , which follows on dividing (107) by 1 + t 1+t 1 + t .
Proof of Proposition 3.9
Write a left vertex type as ( a , b ) (a,b) ( a , b ) and a right type as ( a ′ , b ′ ) (a',b') ( a ′ , b ′ ) . An edge from left to right has weight one if a = a ′ a=a' a = a ′ and zero otherwise. An edge from right to left has weight one if b ′ = b b'=b b ′ = b and zero otherwise. The upper and lower cut blocks, with zero rows or columns omitted, are submatrices of U M U M ′ T U_MU_{M'}^{\mathsf T} U M U M ′ T and V M V M ′ T V_MV_{M'}^{\mathsf T} V M V M ′ T , respectively. Indeed, at any initial cut all left vertices precede all right vertices in the ordering, and an edge inside one half has weight zero. This proves the rank bounds r , s r,s r , s .
Every nonzero cycle alternates between the two halves. There are 2 h 2h 2 h vertices, so the number of cycles is at most h h h , with equality only when all cycles have length two. Such a cycle requires both type coordinates to agree. Thus the coefficient of z h z^h z h in P A M , M ′ ( z ) P_{A_M,M'}(z) P A M , M ′ ( z ) vanishes unless M = M ′ M=M' M = M ′ . If they are equal, the two-cycles pair the m a b m_{ab} m ab left vertices of each type bijectively with the m a b m_{ab} m ab right vertices of that type. This gives the diagonal entry ∏ a , b m a b ! \prod_{a,b}m_{ab}! ∏ a , b m ab ! . If the matrix K h ( z ) K_h(z) K h ( z ) has size d h d_h d h , the coefficient of z d h z^{d_h} z d h in its determinant is the determinant of this diagonal matrix. It is nonzero in characteristic zero. The rank is therefore d h d_h d h , the number of arrays with r s rs r s entries and sum h h h . This is the first equality of (45) .
For the one-cycle lower bound, a perfect matching in each direction is possible only when the left and right lists have the same first-coordinate totals and the same second-coordinate totals. The forward matching uses every right vertex once and preserves the first coordinate. The backward matching gives the second condition. There are B h = ( h + r − 1 r − 1 ) ( h + s − 1 s − 1 ) B_h = \binom{h+r-1}{r-1}\binom{h+s-1}{s-1} B h = ( r − 1 h + r − 1 ) ( s − 1 h + s − 1 ) pairs of totals.
For one such pair take any list with these totals and write it as ( a i , b i ) (a_i,b_i) ( a i , b i ) , 1 ≤ i ≤ h 1 \le i \le h 1 ≤ i ≤ h . Use the right list ( a i , b i + 1 ) (a_i,b_{i+1}) ( a i , b i + 1 ) with b h + 1 = b 1 b_{h+1}=b_1 b h + 1 = b 1 . Then left vertex i i i can go to right vertex i i i , and this vertex can go to left vertex i + 1 i+1 i + 1 . All these edges have weight one and make one cycle. Using this choice for every pair of totals gives a submatrix of [ z ] K h ( z ) [z]K_h(z) [ z ] K h ( z ) with zero entries between different pairs and positive entries on the diagonal. This gives rank at least B h B_h B h .
For the upper bound use the one-cycle formula with the first left vertex in the permanent color. At the cut between the halves the couplings are the identities on the first and second coordinate spaces. The sum over allowed colors within the left half gives a polynomial depending only on M M M . The analogous right polynomial depends only on M ′ M' M ′ . The merge uses these two polynomials bilinearly and is independent of the two lists. If j j j left vertices have the determinant color, all j j j of their rows and columns are unmatched within the half. Their exterior degrees are j , j j,j j , j . The remaining h − j h-j h − j left vertices give commuting degrees h − j , h − j h-j,h-j h − j , h − j . The number of left coordinates is therefore bounded by
∑ j = 0 min ( h , r , s ) ( r j ) ( s j ) ( h − j + r − 1 r − 1 ) ( h − j + s − 1 s − 1 ) ≤ B h ∑ j ( r j ) ( s j ) = B h D − . \sum_{j=0}^{\min(h,r,s)}
\binom{r}{j}\binom{s}{j}
\binom{h-j+r-1}{r-1}
\binom{h-j+s-1}{s-1}
\le B_h \sum_j \binom{r}{j}\binom{s}{j}
= B_hD^{-}. j = 0 ∑ m i n ( h , r , s ) ( j r ) ( j s ) ( r − 1 h − j + r − 1 ) ( s − 1 h − j + s − 1 ) ≤ B h j ∑ ( j r ) ( j s ) = B h D − .
In matrix terms the pairing factors the one-cycle coefficient matrix through this coordinate space. Its rank is at most the dimension of that space, giving the other inequality of (45) .
Proof of Corollary 4.3
For complementary index sets I , J I,J I , J and nonsingular B B B , we use the nullity identity [33 ]
rank B [ I , J ] = rank B − 1 [ I , J ] . (108) \operatorname{rank} B[I,J] = \operatorname{rank} B^{-1}[I,J]. \tag*{(108)} rank B [ I , J ] = rank B − 1 [ I , J ] . (108)
To see it here, regard a vector in the kernel of B [ I , J ] B[I,J] B [ I , J ] as a full vector v v v supported on J J J . The I I I -coordinates of B v Bv B v vanish, so B v Bv B v is also supported on J J J . Its image under B − 1 B^{-1} B − 1 has zero I I I -coordinates, which means that its restriction to J J J lies in the kernel of B − 1 [ I , J ] B^{-1}[I,J] B − 1 [ I , J ] . Conversely, starting from that kernel and multiplying by B − 1 B^{-1} B − 1 gives the original kernel. The two maps are inverse. Since the matrices have the same number of columns, equality of nullities gives (5.15).
We also have
rank ( A B ) [ I , J ] ≤ rank A [ I , J ] + rank B [ I , J ] . (109) \operatorname{rank}(AB)[I,J] \le\operatorname{rank} A[I,J] + \operatorname{rank} B[I,J]. \tag*{(109)} rank ( A B ) [ I , J ] ≤ rank A [ I , J ] + rank B [ I , J ] . (109)
Indeed ( A B ) [ I , J ] = A [ I , I ] B [ I , J ] + A [ I , J ] B [ J , J ] (AB)[I,J] = A[I,I]B[I,J] + A[I,J]B[J,J] ( A B ) [ I , J ] = A [ I , I ] B [ I , J ] + A [ I , J ] B [ J , J ] , and the two terms have ranks at most the two ranks on the right.
If B B B has open lower and upper widths p , q p,q p , q , the upper cross block at a cut after i i i has entries only in the last q q q rows of I I I . The lower block has entries only in the first p p p rows after the cut. Their ranks are at most q , p q,p q , p , and (5.15) gives the same bounds for the inverse. This proves the first assertion using Theorem 3.4.
For the tridiagonal cost, write θ i = det B [ 1 : i , 1 : i ] \theta_i=\det B[1:i,1:i] θ i = det B [ 1 : i , 1 : i ] and φ i = det B [ i : n , i : n ] \varphi_i=\det B[i:n,i:n] φ i = det B [ i : n , i : n ] , with θ 0 = φ n + 1 = 1 \theta_0=\varphi_{n+1}=1 θ 0 = φ n + 1 = 1 . Expanding a leading minor in its last row gives
θ i = b i i θ i − 1 − b i , i − 1 b i − 1 , i θ i − 2 ( i ≥ 2 ) . \theta_i=b_{ii}\theta_{i-1}-b_{i,i-1}b_{i-1,i}\theta_{i-2}\qquad(i\ge2). θ i = b ii θ i − 1 − b i , i − 1 b i − 1 , i θ i − 2 ( i ≥ 2 ) .
The trailing minors satisfy the reversed formula, so all of these numbers take O ( n ) O(n) O ( n ) operations. For i ≤ j i\le j i ≤ j , the cofactor giving ( B − 1 ) i j (B^{-1})_{ij} ( B − 1 ) ij deletes row j j j and column i i i . Divide the remaining rows into 1 , … , i − 1 1,\ldots,i-1 1 , … , i − 1 , i , … , j − 1 i,\ldots,j-1 i , … , j − 1 , and j + 1 , … , n j+1,\ldots,n j + 1 , … , n . The corresponding column groups are 1 , … , i − 1 1,\ldots,i-1 1 , … , i − 1 , i + 1 , … , j i+1,\ldots,j i + 1 , … , j , and j + 1 , … , n j+1,\ldots,n j + 1 , … , n . The resulting matrix is block lower triangular. Its middle block is lower triangular with diagonal b k , k + 1 b_{k,k+1} b k , k + 1 , i ≤ k < j i \le k < j i ≤ k < j . This gives
( B − 1 ) i j = ( − 1 ) i + j θ i − 1 ϕ j + 1 θ n ∏ k = i j − 1 b k , k + 1 . (B^{-1})_{ij}=\frac{(-1)^{i+j}\theta_{i-1}\phi_{j+1}}{\theta_n}\prod_{k=i}^{j-1}b_{k,k+1}. ( B − 1 ) ij = θ n ( − 1 ) i + j θ i − 1 ϕ j + 1 k = i ∏ j − 1 b k , k + 1 .
For i = j i=j i = j the middle block is empty. The lower entries are obtained by transposing the argument, with the lower diagonal in the product. See also [34 ] . For fixed i i i one can successively multiply the factors as j j j increases. Thus all the entries of B − 1 B^{-1} B − 1 take O ( n 2 ) O(n^2) O ( n 2 ) operations, with division only by θ n ≠ 0 \theta_n \ne0 θ n = 0 . Its ordered cut ranks are at most one, so Theorem 3.4 gives the same bound for computing its permanent.
For part (2), root the tree decomposition and assign each vertex to the highest bag containing it. This bag is unique because the bags containing the vertex form a connected subtree. At a bag D D D take the vertex sets assigned below its individual children and the singleton vertices assigned to D D D itself. Combine these sets by binary unions. Applying this construction from the leaves of the decomposition upward, omitting empty sets and suppressing unary nodes, gives a binary tree on the vertices.
Consider a cut arising from one of the binary unions at D D D . A vertex outside D D D has all its bags in one component of the decomposition tree with D D D deleted. If two such vertices are adjacent, some bag contains both, so they lie in the same component. Each component is entirely on one side of the cut for vertices outside D D D , because each child set was included or excluded as a whole. Thus a crossing edge has an endpoint in D D D . In the block from I I I to J J J this means that a nonzero entry is in a row belonging to D ∩ I D \cap I D ∩ I or a column belonging to D ∩ J D \cap J D ∩ J . Splitting off the first group of rows bounds the rank by ∣ D ∩ I ∣ + ∣ D ∩ J ∣ = ∣ D ∣ |D \cap I|+|D \cap J|=|D| ∣ D ∩ I ∣ + ∣ D ∩ J ∣ = ∣ D ∣ . The same holds for the reverse block. By (108) the inverse has the same bounds on this tree. Theorem 3.1 now applies. In particular a tree support graph has a decomposition with bags of size at most two, by using a bag for each vertex and its parent in a rooting of that graph.
The scalar and block band formulas
We prove the transfer formulas in Theorem 4.4 (3), Theorem 4.9 (1) and Theorem 4.7 (2).
Consider the b b b rows in block row i i i . Their available columns lie in w + 1 w+1 w + 1 consecutive blocks. Number these columns relative to the first block, from 0 0 0 to b ( w + 1 ) − 1 b(w+1)-1 b ( w + 1 ) − 1 . The earlier rows could have used columns in the first w w w blocks of this window, and we denote their set by U U U . They cannot have used the last block, since the rightmost allowed block increases by one at each step. Choose b b b unused columns S S S for the current b b b rows. For this fixed set the sum of the weights of the bijections from the rows to S S S is per D ( { 0 , … , b − 1 } , S ) D(\{0,\ldots,b-1\},S) D ({ 0 , … , b − 1 } , S ) . In particular, the weight of a transition already includes all orders of assigning these columns to the b b b rows.
At the next step the first block has left the window. Its columns must all lie in U ∪ S U \cup S U ∪ S , since no later row can fill a missing one. Remove those b b b columns from U ∪ S U \cup S U ∪ S and subtract b b b from the other relative indices. The resulting set is V V V in (48) . As ∣ S ∣ = b |S|=b ∣ S ∣ = b and S ∩ U S \cap U S ∩ U is empty, we have ∣ V ∣ = ∣ U ∣ |V|=|U| ∣ V ∣ = ∣ U ∣ . Thus the transition stays in the sector indexed by this common size.
For an open band, the first block in the window for row i i i is i − p i-p i − p . At row zero declare the p p p blocks with negative indices occupied. This gives the initial state B B B , and prevents the choice of any column with negative index. After the last row, the window starts at m − p m-p m − p . State B B B now means that all blocks still in the window below m m m are full and those from m m m onward are empty. Every earlier block has already left the window and was full when it left. Even with the untruncated weights, no column of index at least m m m can be used on a path from B B B to B B B . Such a column would still lie in the window at the final step and would appear in the terminal state outside B B B . Hence the walks from B B B to B B B , together with the choices counted in each transition, give exactly the matchings in (73) and (68) .
For the scalar cost with p p p fixed, after i ≥ p i \ge p i ≥ p rows the state specifies p p p used columns beyond the first i − p i-p i − p columns. There are O p ( n p ) O_p(n^p) O p ( n p ) possible states. If the next column to leave the window is unused, the current row must use it and there is at most one transition. If that column is already used, only p − 1 p-1 p − 1 other columns have to be specified in the state. There are O p ( n p − 1 ) O_p(n^{p-1}) O p ( n p − 1 ) such states when p ≥ 1 p\ge1 p ≥ 1 , and at most n n n choices for the current row from each. Thus the work at a row is O p ( n p ) O_p(n^p) O p ( n p ) . For p = 0 p=0 p = 0 only the forced choice occurs. During the first p p p rows at most p p p actual columns are chosen, so the transitions at each of these steps also number O p ( n p ) O_p(n^p) O p ( n p ) . Summing over the rows proves the operation bound.
For a cyclic band, each chosen block displacement has a unique representative in 0 , … , w 0,\ldots,w 0 , … , w because w < m w<m w < m . Repeat the choices on block rows indexed by Z \mathbb{Z} Z , using these displacements and the same internal column labels in every period. This is a matching of the lifted columns. In fact, each target column has exactly one preimage modulo the period, and precisely one translate of that preimage reaches the given lifted column. The occupied sets at successive cuts give a closed walk of length m m m in a fixed sector.
Conversely, start from a closed walk and repeat it in both directions. At each transition keep the actual choice of columns counted by that transition. After w w w steps none of the columns in an initial state can remain in the window. Thus each occupied position has arisen from a column used at an earlier step, and no column is chosen a second time. The condition on the departing block makes each column used at least once. Passing back to the cyclic indices gives the required matching. Taking traces counts all initial states and yields (4.9) and (4.24). For a scalar matching an edge with displacement d d d appears in the occupied sets at the d d d cuts between its row and column. The sum of the displacements is therefore n k nk nk when each of the n n n cuts has k k k occupied positions.
To derive the scalar coefficient formulas, let Q ( z ) = ∑ j = 1 D q j z j Q(z)=\sum_{j=1}^{D}q_jz^j Q ( z ) = ∑ j = 1 D q j z j , where det ( I − z T ) = 1 − Q ( z ) \det(I-zT)=1-Q(z) det ( I − z T ) = 1 − Q ( z ) . The trace of T m T^m T m is the sum of the m m m -th powers of its eigenvalues, counted with multiplicity. Factoring 1 − Q ( z ) 1-Q(z) 1 − Q ( z ) over an algebraic closure gives
− log ( 1 − Q ( z ) ) = ∑ m ≥ 1 z m m tr ( T m ) . -\log(1-Q(z))=\sum_{m\geq1}\frac{z^m}{m}\operatorname{tr}(T^m). − log ( 1 − Q ( z )) = m ≥ 1 ∑ m z m tr ( T m ) .
In the formal expansion ∑ ℓ ≥ 1 Q ( z ) ℓ / ℓ \sum_{\ell\geq1}Q(z)^\ell/\ell ∑ ℓ ≥ 1 Q ( z ) ℓ / ℓ , a term indexed by α \alpha α with ∣ α ∣ = ℓ |\alpha|=\ell ∣ α ∣ = ℓ has multinomial coefficient ℓ ! / ∏ j α j ! \ell!/\prod_j\alpha_j! ℓ ! / ∏ j α j ! before division by ℓ \ell ℓ . Taking m [ z m ] m[z^m] m [ z m ] gives (4.2). Equivalently, if the α j \alpha_j α j specify numbers of tiles of length j j j , mark one of the ∣ α ∣ |\alpha| ∣ α ∣ boundaries. There are m m m choices of its position and ∣ α ∣ ! / ∏ j α j ! |\alpha|!/\prod_j\alpha_j! ∣ α ∣ ! / ∏ j α j ! ordered lists of tiles starting there. Dividing by ∣ α ∣ |\alpha| ∣ α ∣ counts the unmarked tilings.
For the open case, the formal inverse is ( I − z T ) − 1 = ∑ m ≥ 0 z m T m (I-zT)^{-1}=\sum_{m\geq0}z^mT^m ( I − z T ) − 1 = ∑ m ≥ 0 z m T m . The adjugate formula gives its B , B B,B B , B entry as det ( I − z T ′ ) / det ( I − z T ) \det(I-zT')/\det(I-zT) det ( I − z T ′ ) / det ( I − z T ) . Expanding ( 1 − Q ( z ) ) − 1 = ∑ ℓ ≥ 0 Q ( z ) ℓ (1-Q(z))^{-1}=\sum_{\ell\geq0}Q(z)^\ell ( 1 − Q ( z ) ) − 1 = ∑ ℓ ≥ 0 Q ( z ) ℓ in the same way, now without division by ℓ \ell ℓ , gives (4.1). Multiplication by the numerator yields (4.29) and (4.30). The block formula (4.26) uses the same inverse entry with ∣ B ∣ = b p |B|=bp ∣ B ∣ = b p .
With a row period h h h , the product for n = h k + r n=hk+r n = hk + r is M k N r M^kN_r M k N r . Its series in k k k , at fixed r r r , has B , B B,B B , B entry e B T ( I − z M ) − 1 N r e B e_B^{\mathsf T}(I-zM)^{-1}N_re_B e B T ( I − z M ) − 1 N r e B . Expanding the denominator as above gives (4.37). At fixed width and fixed block size the matrices have fixed order, so repeated squaring gives the stated cost for their powers.
The residue case of Theorem 4.4
5.12. The residue case of Theorem 4.4. Shift columns first so that u u u is subtracted from all the jumps. Since g g g divides both m m m and h h h , each new jump is a multiple of g g g . After grouping rows and columns by their residue classes modulo g g g , the matrix is a direct sum. Translation of both indices carries any class to any other and preserves the weights, so the component permanents are equal.
In the component with residue c c c , write the indices in the form
c + t h + j m ( m o d n ) , 0 ≤ t < L , 0 ≤ j < b . c+th+jm \qquad(\bmod\ n),\qquad0\leq t<L,\quad0\leq j<b. c + t h + j m ( mod n ) , 0 ≤ t < L , 0 ≤ j < b .
To check that these labels are distinct, reduce an equality first modulo m m m . It gives L ∣ t − t ′ L \mid t-t' L ∣ t − t ′ , since L = m / g L=m/g L = m / g and gcd ( h / g , L ) = 1 \gcd(h/g,L)=1 g cd( h / g , L ) = 1 . Hence t = t ′ t=t' t = t ′ , and equality modulo b m bm bm then gives j = j ′ j=j' j = j ′ . There are L b = n / g Lb=n/g L b = n / g such labels, so they exhaust the component.
In the lifted block indices t ∈ Z t\in\mathbb{Z} t ∈ Z , a jump s − u = d h + ℓ m s-u=dh+\ell m s − u = d h + ℓ m goes from ( t , j ) (t,j) ( t , j ) to ( t + d , j + ℓ ) (t+d,j+\ell) ( t + d , j + ℓ ) , with the second coordinate read modulo b b b . This gives the blocks D d D_d D d of Theorem 4.4 (2). Since L h = m ( h / g ) Lh=m(h/g) L h = m ( h / g ) , the label ( t + L , j − h / g ) (t+L,j-h/g) ( t + L , j − h / g ) represents the same column as ( t , j ) (t,j) ( t , j ) . Thus the relative positions of occupied columns after a period are those of τ U \tau U τ U if the starting state is U U U .
As in the cyclic band proof, a matching of the component lifts to a walk, now from U U U to τ U \tau U τ U . For the converse, extend such a walk periodically using the change j ↦ j − h / g j\mapsto j-h/g j ↦ j − h / g after every period. The blocks D d D_d D d are circulant in their internal labels, so rotating all internal labels preserves every transition and its weight. Each column is therefore used once by the extended walk, by the same departing-block condition as in the ordinary cyclic case. The result descends to the component. Summing the entries of T k L T_k^L T k L indexed by U , τ U U,\tau U U , τ U , and then over k k k , gives the component permanent. Taking its g g g -th power proves (4.8).
There are 2 b w 2^{bw} 2 b w states across all the sectors. For a fixed state, form the transitions by choosing one column for each of the b b b rows in the current block. Each row has at most ∣ S ∣ |S| ∣ S ∣ choices. Discard choices with a repeated column or a violation of the conditions in (4.3), and add the weights of the remaining choices to their corresponding entries. There are at most ∣ S ∣ b |S|^b ∣ S ∣ b choices for each state. Repeated squaring then takes at most O ( 2 3 b w log ( n + 1 ) ) O(2^{3bw}\log(n+1)) O ( 2 3 b w log ( n + 1 )) operations for all sectors together. If w ≥ 1 w\ge1 w ≥ 1 , then b ≤ b w b\le bw b ≤ b w , so both bounds are polynomial in n n n when ∣ S ∣ |S| ∣ S ∣ is fixed and b w = O ( log ( n + 1 ) ) bw=O(\log(n+1)) b w = O ( log ( n + 1 )) .
The parameters for this bound can be found by testing the divisors b b b of n n n , the residues h h h mod m m m , and u ∈ S u\in S u ∈ S . Indeed, in any representation, one can take the offset with the smallest d s d_s d s as u u u and subtract that smallest value from every d s d_s d s . Discard a test if any congruence d s h ≡ s − u ( m o d m ) d_sh\equiv s-u\pmod m d s h ≡ s − u ( mod m ) has no solution. Otherwise take the unique solution in 0 ≤ d s < L 0\le d_s<L 0 ≤ d s < L for each s s s and use their maximum for w w w .
For (4.19) use b = 1 b=1 b = 1 , so that each D d D_d D d is the scalar c d c_d c d and τ \tau τ is the identity. In the example of order 2 m 2m 2 m there are two internal labels, and the change of label is subtraction of h h h modulo two. The jump m m m exchanges those labels inside a block, giving the d P 2 dP_2 d P 2 contribution to D 0 D_0 D 0 .
The signed formulas of Theorem 4.4
For the torus determinant method, see [15 , 35 ] . The map ( x , y ) ↦ s x + t y ( m o d n ) (x,y)\mapsto sx+ty\pmod n ( x , y ) ↦ s x + t y ( mod n ) is onto by the gcd condition. If ( x , y ) ∈ Λ (x,y)\in\Lambda ( x , y ) ∈ Λ , reduction modulo g 0 g_0 g 0 gives t y ≡ 0 ( m o d g 0 ) ty\equiv0\pmod{g_0} t y ≡ 0 ( mod g 0 ) , and gcd ( t , g 0 ) = 1 \gcd(t,g_0)=1 g cd( t , g 0 ) = 1 gives g 0 ∣ y g_0\mid y g 0 ∣ y . Write y = g 0 m y=g_0m y = g 0 m . Dividing the congruence by g 0 g_0 g 0 gives x ≡ u 0 m ( m o d h 0 ) x\equiv u_0m\pmod{h_0} x ≡ u 0 m ( mod h 0 ) . Thus each element of Λ \Lambda Λ is an integral combination of b 1 , b 2 b_1,b_2 b 1 , b 2 . Conversely these two vectors satisfy the congruence. To reduce any lattice point to the specified rectangle, subtract a multiple of b 2 b_2 b 2 to put y y y between 0 0 0 and g 0 − 1 g_0-1 g 0 − 1 , and then subtract a multiple of b 1 b_1 b 1 to put x x x between 0 0 0 and h 0 − 1 h_0-1 h 0 − 1 . The congruence just proved also shows that two points of this rectangle in the same coset must coincide. This proves the description of ρ i \rho_i ρ i .
Join each point of Z 2 \mathbb{Z}^2 Z 2 to its east, north and northeast neighbors. These segments have disjoint interiors, and their translates give the supported off-diagonal edges on the quotient torus. The nontrivial cycles of a contributing permutation are therefore disjoint simple curves. The displacement of a lift of each cycle is the sum of its steps. The steps have nonnegative coordinates and each increases at least one coordinate, so the displacement is a nonzero element of Λ \Lambda Λ . In particular the curve is essential.
Cutting a torus along an essential simple curve gives a cylinder. An arc joining the two boundary copies of a point on the curve closes after gluing, and its intersection number with the original curve is 1 1 1 or − 1 -1 − 1 . The homology class of the original curve therefore cannot be a nontrivial integer multiple. For two disjoint curves the intersection number is zero. In lattice coordinates this intersection number is the determinant of their two coordinate vectors. Their primitive classes are consequently equal up to sign. In our lift both vectors have nonnegative coordinates, so they have the same sign. Thus the total displacement of k k k nontrivial cycles is k k k times one primitive vector. In the basis b 1 , b 2 b_1,b_2 b 1 , b 2 , if the total has coordinates H 1 , H 2 H_1,H_2 H 1 , H 2 , this says k = gcd ( H 1 , H 2 ) k=\gcd(H_1,H_2) k = g cd( H 1 , H 2 ) , with the stated convention for the empty collection.
For the monomial M M M , the number of east steps plus northeast steps is R R R , and the number of north steps plus northeast steps is S S S . Their total ( R , S ) (R,S) ( R , S ) lies in Λ \Lambda Λ , giving n ∣ s R + t S n\mid sR+tS n ∣ s R + tS . An integer j > 0 j>0 j > 0 divides this vector in Λ \Lambda Λ precisely when j j j divides R , S R,S R , S and ( R / j , S / j ) (R/j,S/j) ( R / j , S / j ) still satisfies the defining congruence. The last condition is that j j j divide ( s R + t S ) / n (sR+tS)/n ( s R + tS ) / n . This proves the formula for k k k in ordinary coordinates. If the nontrivial cycle lengths are m 1 , … , m k m_1,\ldots,m_k m 1 , … , m k , their signs multiply to ( − 1 ) ∑ i ( m i − 1 ) (-1)^{\sum_i(m_i-1)} ( − 1 ) ∑ i ( m i − 1 ) . As ∑ i m i = r + z + v \sum_i m_i=r+z+v ∑ i m i = r + z + v , this is ( − 1 ) r + z + v − k (-1)^{r+z+v-k} ( − 1 ) r + z + v − k , independent of the contributing permutation. Taking coefficients proves (52) .
Now keep arbitrary row weights. In the determinant expansion a nontrivial cycle of length m m m has sign ( − 1 ) m − 1 (-1)^{m-1} ( − 1 ) m − 1 . Its m m m off-diagonal entries supply m m m additional minus signs, leaving − 1 -1 − 1 . The remaining signs in K ϵ η K_{\epsilon\eta} K ϵη come from the lattice coordinates of the edge differences. Around a cycle the differences ρ i − ρ j \rho_i-\rho_j ρ i − ρ j telescope. Thus the sums of the edge coordinates over all cycles are H 1 , H 2 H_1,H_2 H 1 , H 2 , and the ratio of the signed determinant term to the unsigned permanent term is
( − 1 ) k ( − 1 ) ϵ H 1 + η H 2 . (-1)^k(-1)^{\epsilon H_1+\eta H_2}. ( − 1 ) k ( − 1 ) ϵ H 1 + η H 2 .
Since k = gcd ( H 1 , H 2 ) k=\gcd(H_1,H_2) k = g cd( H 1 , H 2 ) , the sign ( − 1 ) k (-1)^k ( − 1 ) k is positive exactly when both coordinates are even. The same parity rule is expressed by
( − 1 ) k = − 1 + ( − 1 ) H 1 + ( − 1 ) H 2 + ( − 1 ) H 1 + H 2 2 . (-1)^k=\frac{-1+(-1)^{H_1}+(-1)^{H_2}+(-1)^{H_1+H_2}}{2}. ( − 1 ) k = 2 − 1 + ( − 1 ) H 1 + ( − 1 ) H 2 + ( − 1 ) H 1 + H 2 .
In the signed combination (51) , the multiplier is the product of ( − 1 ) k (-1)^k ( − 1 ) k with this last expression, and is therefore one. Fixed points have their original weights throughout. This proves (51) term by term. Each supported entry is signed using two lattice coordinates, and the four determinants take O ( n 3 ) O(n^3) O ( n 3 ) arithmetic operations. Fraction-free elimination gives polynomial bit complexity over the rationals.
For constant weights, fix a solution ( X 0 , Y 0 ) (X_0,Y_0) ( X 0 , Y 0 ) of the two root equations in (58) over an algebraic closure, and put χ ( x , y ) = X 0 x Y 0 y \chi(x,y)=X_0^xY_0^y χ ( x , y ) = X 0 x Y 0 y . Its values on b 1 , b 2 b_1,b_2 b 1 , b 2 are ( − 1 ) ϵ , ( − 1 ) η (-1)^\epsilon,(-1)^\eta ( − 1 ) ϵ , ( − 1 ) η . With Q i i = χ ( ρ i ) Q_{ii}=\chi(\rho_i) Q ii = χ ( ρ i ) on the diagonal, the entry on an edge with step δ \delta δ acquires the factor χ ( ρ i ) χ ( δ ) χ ( ρ j ) − 1 \chi(\rho_i)\chi(\delta)\chi(\rho_j)^{-1} χ ( ρ i ) χ ( δ ) χ ( ρ j ) − 1 under the following conjugation:
K ϵ η = Q ( a I − b X 0 P n s − c Y 0 P n t − d X 0 Y 0 P n s + t ) Q − 1 . K_{\epsilon\eta}=Q(aI-bX_0P_n^s-cY_0P_n^t-dX_0Y_0P_n^{s+t})Q^{-1}. K ϵη = Q ( a I − b X 0 P n s − c Y 0 P n t − d X 0 Y 0 P n s + t ) Q − 1 .
This factor is χ ( ρ i + δ − ρ j ) \chi(\rho_i+\delta-\rho_j) χ ( ρ i + δ − ρ j ) , which is exactly the remaining sign for that edge. The diagonal is unchanged.
For each n n n -th root of unity ω \omega ω , the vector with i i i -th entry ω i \omega^i ω i is an eigenvector of P n P_n P n with eigenvalue ω \omega ω . The conjugate circulant therefore has the factor a − b X − c Y − d X Y a-bX-cY-dXY a − b X − c Y − d X Y for ( X , Y ) = ( X 0 ω s , Y 0 ω t ) (X,Y)=(X_0\omega^s,Y_0\omega^t) ( X , Y ) = ( X 0 ω s , Y 0 ω t ) . These pairs satisfy the two root equations since b 1 , b 2 ∈ Λ b_1,b_2\in\Lambda b 1 , b 2 ∈ Λ . They are distinct, as equality for two roots gives a ratio with s s s -th, t t t -th and n n n -th powers all one. There are h 0 g 0 = n h_0g_0=n h 0 g 0 = n solutions of the two root equations. This proves (58) . For the explicit parametrization in the corollary, x ϵ h 0 = ζ ϵ n x_\epsilon^{h_0}=\zeta^{\epsilon n} x ϵ h 0 = ζ ϵ n and x ϵ u 0 y η g 0 = ζ η n x_\epsilon^{u_0}y_\eta^{g_0}=\zeta^{\eta n} x ϵ u 0 y η g 0 = ζ η n , so the chosen starting pair satisfies the root equations.
For fixed X X X , the equation for Y Y Y is Y g 0 = ( − 1 ) η X − u 0 Y^{g_0}=(-1)^\eta X^{-u_0} Y g 0 = ( − 1 ) η X − u 0 . Multiplying a − b X − ( c + d X ) Y a-bX-(c+dX)Y a − b X − ( c + d X ) Y over those roots, by the monic polynomial for Y Y Y , gives
( a − b X ) g 0 − ( − 1 ) η X − u 0 ( c + d X ) g 0 . (a-bX)^{g_0}-(-1)^\eta X^{-u_0}(c+dX)^{g_0}. ( a − b X ) g 0 − ( − 1 ) η X − u 0 ( c + d X ) g 0 .
The identity is polynomial in a − b X , c + d X a-bX,c+dX a − b X , c + d X , including when the latter is zero. Since X X X is invertible modulo f ϵ f_\epsilon f ϵ , the value at each root of f ϵ f_\epsilon f ϵ is unchanged when the expression is replaced by its remainder. The product over those roots is the resultant in (65) . Modular powering uses O ( log ( g 0 + 1 ) ) O(\log(g_0+1)) O ( log ( g 0 + 1 )) multiplications of polynomials of degree at most h 0 h_0 h 0 . The classical Euclidean resultant calculation uses O ( h 0 2 ) O(h_0^2) O ( h 0 2 ) arithmetic operations. For s = 1 , t = q s=1,t=q s = 1 , t = q , one has g 0 = 1 , h 0 = n , u 0 = n − q g_0=1,h_0=n,u_0=n-q g 0 = 1 , h 0 = n , u 0 = n − q . The second root equation then gives Y = ( − 1 ) η X q − n = ( − 1 ) ϵ + η X q Y=(-1)^\eta X^{q-n}=(-1)^{\epsilon+\eta}X^q Y = ( − 1 ) η X q − n = ( − 1 ) ϵ + η X q . Using the exponent η \eta η instead reindexes the two products with ϵ = 1 \epsilon=1 ϵ = 1 , exchanging D 10 D_{10} D 10 and D 11 D_{11} D 11 . Their coefficients in (51) agree.
Finally, after translating an offset to zero, the three- and four-offset cases of D n ( A ) D_n(A) D n ( A ) separate into classes modulo gcd ( n , s , t ) \gcd(n,s,t) g cd( n , s , t ) . Inside one class, division of the indices and offsets by this gcd gives the connected case already proved. Taking the product gives the formula for arbitrary row weights, and constant weights give equal factors. For two shifts 0 , s 0,s 0 , s , in an orbit of addition by s s s , a row choosing the s s s -edge uses the next row’s diagonal column. That next row must also use the s s s -edge, and the choice continues around the orbit. The only other choice is to use the diagonal throughout the orbit. This gives the two terms in (4.12).
The specializations in Corollary 4.5 and Corollary 4.10
In sector k = 1 k=1 k = 1 , the total displacement is n n n . Along any permutation cycle, the sum of the displacements is a multiple of n n n , since the lifted walk returns to the same residue. For a nontrivial cycle this sum is positive. Hence there is one nontrivial cycle in this sector, and its steps go once around the circle in their positive order. The steps of lengths d d d are intervals tiling the circle. A step of length d d d has weight c d c_d c d , and the d − 1 d-1 d − 1 vertices skipped between its ends are fixed, with weight c 0 c_0 c 0 each. Its total contribution is c 0 d − 1 c d c_0^{d-1}c_d c 0 d − 1 c d .
If there are α d \alpha_d α d intervals of length d d d , put H = ∑ d α d H=\sum_d\alpha_d H = ∑ d α d . A tiling with a marked boundary is specified by the position of that boundary and the ordered list of intervals from it. Thus there are n H ! ∏ d α d ! \frac{nH!}{\prod_d\alpha_d!} ∏ d α d ! n H ! marked tilings. Each tiling has H H H boundaries that can be marked. Summing the unmarked weights gives L n ( c 1 , c 0 c 2 , … , c 0 w − 1 c w ) L_n(c_1,c_0c_2,\ldots,c_0^{w-1}c_w) L n ( c 1 , c 0 c 2 , … , c 0 w − 1 c w ) . The sectors 0 , w 0,w 0 , w use only steps 0 , w 0,w 0 , w , respectively, and give c 0 n , c w n c_0^n,c_w^n c 0 n , c w n . Transposition changes a displacement d d d into − d -d − d , and a column shift then changes it into w − d w-d w − d . The total displacement n k nk nk becomes n ( w − k ) n(w-k) n ( w − k ) . This gives the formula for sector w − 1 w-1 w − 1 by reversing the weights in sector one.
For w = 2 w=2 w = 2 , the numbers of steps of lengths one and two are n − 2 j n-2j n − 2 j and j j j . In the marked-boundary count, H = n − j H=n-j H = n − j , so the factor becomes n ( n − j ) n − j \frac{n^{(n-j)}}{n-j} n − j n ( n − j ) . This gives (4.14). For w = 3 w=3 w = 3 the two middle sectors are the winding-one count and its reversal, giving (4.15). If the numbers of steps of lengths one, two and three are u , v , w u,v,w u , v , w in that count, the product of the weights is a v + 2 w b u c v d w a^{v+2w}b^u c^v d^w a v + 2 w b u c v d w . Reversing the coefficients gives the second monomial in (4.17), with the same factorial factor. For a prescribed monomial, the total displacement distinguishes the two middle sectors. Reading its step counts in the appropriate sector gives the stated individual coefficients.
For the open one-subdiagonal band, number rows and columns from one. Suppose row one chooses column j + 1 j+1 j + 1 . If j ≥ 1 j\geq1 j ≥ 1 , column one can now only be filled by row two, since rows after it cannot reach column one. After that, column two can only be filled by row three if j ≥ 2 j\geq2 j ≥ 2 . Continuing, row k + 1 k+1 k + 1 must fill column k k k for 1 ≤ k ≤ j 1\leq k\leq j 1 ≤ k ≤ j . The first j + 1 j+1 j + 1 indices are thus matched among themselves and contribute a j c j a^jc_j a j c j . For j = 0 j=0 j = 0 row one gives b b b . Removing the first interval leaves a band of the same type. Conversely, any composition into intervals of the allowed lengths gives one such choice in each interval. The product of the interval weights, summed over the compositions, gives (4.38), and for lengths one, two and three its multinomial expansion is (4.39).
For width four, the only sector not yet counted is sector two. Use the order 01, 02, 03, 12, 13, 23 for its two-element sets. For example, from 01 the choices 2, 3, 4 give the new sets 01, 02, 03, respectively. From 12 the only choice is zero, which gives 01. Filling the other rows in the same way gives
T 4 , 2 = ( c d e 0 0 0 b 0 0 d e 0 0 b 0 c 0 e a 0 0 0 0 0 0 a 0 0 0 0 0 0 0 a 0 0 ) . T_{4,2}=
\begin{pmatrix}
c & d & e & 0 & 0 & 0 \\
b & 0 & 0 & d & e & 0 \\
0 & b & 0 & c & 0 & e \\
a & 0 & 0 & 0 & 0 & 0 \\
0 & a & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & a & 0 & 0
\end{pmatrix}. T 4 , 2 = c b 0 a 0 0 d 0 b 0 a 0 e 0 0 0 0 0 0 d c 0 0 a 0 e 0 0 0 0 0 0 e 0 0 0 .
The lower right three-by-three block of I − z T 4 , 2 I-zT_{4,2} I − z T 4 , 2 is unit lower triangular. Eliminating the lower left block, the remaining determinant is
det ( 1 − c z − d z 2 − e z − b z − a d z 2 1 − a e z 2 0 − a c z 2 − a 2 e z 3 − b z 1 ) \det\begin{pmatrix} 1-cz & -dz^{2} & -ez \\ -bz-adz^{2} & 1-aez^{2} & 0 \\ -acz^{2}-a^{2}ez^{3} & -bz & 1 \end{pmatrix} det 1 − cz − b z − a d z 2 − a c z 2 − a 2 e z 3 − d z 2 1 − a e z 2 − b z − ez 0 1
For the numerator with state 01 deleted, the same elimination leaves the lower right two-by-two minor of this last matrix. Its determinant is 1 − a e z 2 1-aez^{2} 1 − a e z 2 . For the denominator, put x = a e x=ae x = a e , y = b d y=bd y = b d , h = a d 2 + b 2 e h=ad^{2}+b^{2}e h = a d 2 + b 2 e . Expansion of the three-by-three determinant gives
( 1 − x z 2 ) ( 1 − c z ( 1 + x z 2 ) − x z 4 ) − y z 2 ( 1 + x z 2 ) − h z 3 (1-xz^{2})(1-cz(1+xz^{2})-xz^{4})-yz^{2}(1+xz^{2})-hz^{3} ( 1 − x z 2 ) ( 1 − cz ( 1 + x z 2 ) − x z 4 ) − y z 2 ( 1 + x z 2 ) − h z 3
This is 1 − ∑ j = 1 6 q j z j 1-\sum_{j=1}^{6}q_{j}z^{j} 1 − ∑ j = 1 6 q j z j with the parameters in (55) . It is also ( 1 − x z 2 ) ( 1 − F ( z ) ) (1-xz^{2})(1-F(z)) ( 1 − x z 2 ) ( 1 − F ( z )) with F F F as in (56) . The trace expansion gives (61) , and the open quotient gives (85) and (86) .
The factor 1 − x z 2 1-xz^{2} 1 − x z 2 cancels in the open quotient, so the coefficient of z n z^{n} z n is that of ∑ ℓ = 0 n F ( z ) ℓ \sum_{\ell=0}^{n}F(z)^{\ell} ∑ ℓ = 0 n F ( z ) ℓ . To expand it, take i , j , k , ℓ i,j,k,\ell i , j , k , ℓ copies of the terms c z ( 1 + x z 2 ) cz(1+xz^{2}) cz ( 1 + x z 2 ) , x z 4 xz^{4} x z 4 , y z 2 ( 1 + x z 2 ) / ( 1 − x z 2 ) yz^{2}(1+xz^{2})/(1-xz^{2}) y z 2 ( 1 + x z 2 ) / ( 1 − x z 2 ) , and h z 3 / ( 1 − x z 2 ) hz^{3}/(1-xz^{2}) h z 3 / ( 1 − x z 2 ) , respectively. Their multinomial factor is ( i + j + k + ℓ ) ! / ( i ! j ! k ! ℓ ! ) (i+j+k+\ell)!/(i!j!k!\ell!) ( i + j + k + ℓ )! / ( i ! j ! k ! ℓ !) . The remaining power series are ( 1 + x z 2 ) i + k (1+xz^{2})^{i+k} ( 1 + x z 2 ) i + k and ( 1 − x z 2 ) − k − ℓ (1-xz^{2})^{-k-\ell} ( 1 − x z 2 ) − k − ℓ . Taking terms u , v u,v u , v from these series gives ( i + k u ) B ( k + ℓ , v ) \binom{i+k}{u}B(k+\ell,v) ( u i + k ) B ( k + ℓ , v ) and the power z i + 4 j + 2 k + 3 ℓ + 2 u + 2 v z^{i+4j+2k+3\ell+2u+2v} z i + 4 j + 2 k + 3 ℓ + 2 u + 2 v . This is (88) . In the cyclic formula the logarithm of the denominator splits into the logarithms of 1 − x z 2 1-xz^{2} 1 − x z 2 and 1 − F ( z ) 1-F(z) 1 − F ( z ) . The first gives 2 x n / 2 2x^{n/2} 2 x n /2 for even n n n and zero for odd n n n . For the second, the term with i + j + k + ℓ i+j+k+\ell i + j + k + ℓ factors is multiplied by n / ( i + j + k + ℓ ) n/(i+j+k+\ell) n / ( i + j + k + ℓ ) . This proves (63) .
For the symmetric specialization a = e = v a=e=v a = e = v , b = d = u b=d=u b = d = u , c = t c=t c = t , the numerator is 1 − v 2 z 2 1-v^{2}z^{2} 1 − v 2 z 2 . Substitution into the denominator just computed and division by 1 + v z 1+vz 1 + v z gives 1 − ∑ j = 1 5 s j z j 1-\sum_{j=1}^{5}s_{j}z^{j} 1 − ∑ j = 1 5 s j z j . The numerator after division is 1 − v z 1-vz 1 − v z , giving (89) by the ordinary composition expansion. The numerical values in the corollary come from this quotient at t = u = v = 1 t=u=v=1 t = u = v = 1 . Each nonzero product in the permanent of the resulting ( 0 , 1 ) (0,1) ( 0 , 1 ) matrix represents one perfect matching of its bipartite graph.
For a tridiagonal matrix, a nontrivial permutation cycle can only be a transposition of neighbors. Indeed, at the least vertex in such a cycle the preceding and following vertices must both be its immediate successor in the path. A cycle of length greater than two would repeat that successor. Thus the possible transpositions form a set E E E of disjoint path edges. Their weights are x i x_i x i , and the remaining vertices use the diagonal. This proves (90) . For k k k transpositions, subtract j − 1 j-1 j − 1 from the j j j -th selected edge index, for 1 ≤ j ≤ k 1\le j\le k 1 ≤ j ≤ k . This is a bijection with the k k k -subsets of { 1 , … , n − k } \{1,\ldots,n-k\} { 1 , … , n − k } , giving the binomial factor for constant weights. For offsets divisible by r r r , ordering by residues modulo r r r separates the matrix into blocks. In the class of s s s of the offsets − r , 0 , r -r,0,r − r , 0 , r become − 1 , 0 , 1 -1,0,1 − 1 , 0 , 1 on ℓ s \ell_s ℓ s indices. Multiplying their formulas gives (91) .
For even order 2 m 2m 2 m with centrosymmetry, if edge m m m is unused the two halves contribute p m 2 p_m^2 p m 2 . If it is used, it has weight x m x_m x m and leaves two paths of length m − 1 m-1 m − 1 , giving x m p m − 1 2 x_mp_{m-1}^2 x m p m − 1 2 . For odd order 2 m + 1 2m+1 2 m + 1 , fixing the central vertex gives d m + 1 p m 2 d_{m+1}p_m^2 d m + 1 p m 2 . Matching it to its left neighbor has weight x m x_m x m and leaves paths of lengths m − 1 , m m-1,m m − 1 , m . Reflection gives the same contribution when it is matched to its right neighbor. This proves the two reflection formulas.
The rational circulants of Corollary 4.5
Consider a Laurent polynomial in P n P_n P n supported on offsets − p , … , q -p,\ldots,q − p , … , q . At an initial cut after i i i , a forward edge from the initial interval to its complement either moves by a positive offset and starts in one of its last q q q rows, or wraps by a negative offset and starts in one of its first p p p rows. Thus the upper cross block has rank at most p + q p+q p + q . The same argument after transposition bounds the rank of the opposite block by p + q p+q p + q . Applying (108) to g ( P n ) g(P_n) g ( P n ) bounds the cut ranks of its inverse. By (109) , the two ranks of f ( P n ) g ( P n ) − 1 f(P_n)g(P_n)^{-1} f ( P n ) g ( P n ) − 1 are bounded by the sums of the corresponding bounds for f f f and g g g . Equation (31) therefore applies with constant rank parameters.
The permutation P n h P_n^h P n h has gcd ( n , h ) \gcd(n,h) g cd( n , h ) orbits, all of size n / gcd ( n , h ) n/\gcd(n,h) n / g cd( n , h ) . Ordering each orbit along its cycle makes its restriction the shift of that order. Both Laurent polynomials and the inverse are block diagonal in this ordering with equal blocks. Multiplicativity of the permanent over a direct sum gives the stated power.
For the Eulerian formula, multiplication by I − ρ P n I-\rho P_n I − ρ P n gives
( I − ρ P n ) ∑ d = 0 n − 1 ρ d P n d = ( 1 − ρ n ) I . (I-\rho P_n)\sum_{d=0}^{n-1}\rho^d P_n^d=(1-\rho^n)I. ( I − ρ P n ) d = 0 ∑ n − 1 ρ d P n d = ( 1 − ρ n ) I .
For a permutation σ \sigma σ , let d i ∈ { 0 , … , n − 1 } d_i\in\{0,\ldots,n-1\} d i ∈ { 0 , … , n − 1 } be the residue of σ ( i ) − i \sigma(i)-i σ ( i ) − i . Then d i = σ ( i ) − i d_i=\sigma(i)-i d i = σ ( i ) − i if σ ( i ) ≥ i \sigma(i)\ge i σ ( i ) ≥ i , and d i = σ ( i ) − i + n d_i=\sigma(i)-i+n d i = σ ( i ) − i + n otherwise. The differences σ ( i ) − i \sigma(i)-i σ ( i ) − i sum to zero. Thus the term for σ \sigma σ in the inverse permanent has weight
( 1 − ρ n ) − n ρ n ∣ { i : σ ( i ) < i } ∣ . (1-\rho^n)^{-n}\rho^{n|\{i:\sigma(i)<i\}|}. ( 1 − ρ n ) − n ρ n ∣ { i : σ ( i ) < i } ∣ .
The number of permutations with k k k such positions is the Eulerian number ⟨ n k ⟩ \left\langle{n \atop k}\right\rangle ⟨ k n ⟩ [36 ] . Summing the weights proves (66) .
The number of signed determinants
Write the row vertices first and the column vertices second in the skew-adjacency matrix of the bipartite support graph. Its Pfaffian is the determinant of its upper-right block, up to one overall sign depending only on the order. Changing the orientation of an edge changes the corresponding sign in that block. Thus the least number in (50) is the Pfaffian number of the support graph. For any fixed collection of signings, comparing monomial coefficients gives a rational linear system for the coefficients of the determinants. Its consistency is unchanged by extension of the ground field in characteristic zero. We may therefore use the real Pfaffian results for the lower bounds.
For the connected additive support, denote the neighborhood of a row set X X X by X + S X+S X + S . For 1 ≤ ∣ X ∣ ≤ n − 2 1\le|X|\le n-2 1 ≤ ∣ X ∣ ≤ n − 2 we claim ∣ X + S ∣ ≥ ∣ X ∣ + 2 |X+S|\ge|X|+2 ∣ X + S ∣ ≥ ∣ X ∣ + 2 . This follows at once if X + S X+S X + S is the whole group. If the stabilizer D D D of X + S X+S X + S is trivial, Kneser’s inequality gives ∣ X + S ∣ ≥ ∣ X ∣ + ∣ S ∣ − 1 = ∣ X ∣ + 3 |X+S|\ge|X|+|S|-1=|X|+3 ∣ X + S ∣ ≥ ∣ X ∣ + ∣ S ∣ − 1 = ∣ X ∣ + 3 . If D D D is nontrivial and proper, connectedness means S S S meets at least two D D D -cosets. Indeed, if it met only one, every difference of two shifts would lie in D D D . Kneser’s inequality now gives
∣ X + S ∣ ≥ ∣ X + D ∣ + ∣ S + D ∣ − ∣ D ∣ ≥ ∣ X ∣ + ∣ D ∣ ≥ ∣ X ∣ + 2. |X+S|\ge|X+D|+|S+D|-|D|\ge|X|+|D|\ge|X|+2. ∣ X + S ∣ ≥ ∣ X + D ∣ + ∣ S + D ∣ − ∣ D ∣ ≥ ∣ X ∣ + ∣ D ∣ ≥ ∣ X ∣ + 2.
Fix any two disjoint edges and remove their rows and columns. A set of surviving rows has lost at most two neighbors. The inequality proves Hall’s condition on the remaining graph. Its matching, together with the two prescribed edges, is a perfect matching. The connected support graph is therefore a brace.
A Pfaffian brace is either the Heawood graph or is obtained from planar braces by trisums along four-cycles [37 ] . Here we only need the degree consequence of this description. A planar brace on v ≥ 4 v\ge4 v ≥ 4 vertices has minimum degree at least three. Since it is bipartite, Euler’s inequality gives at most 2 v − 4 2v-4 2 v − 4 edges. If v 3 v_3 v 3 vertices have degree three, the degree sum is at least 4 v − v 3 4v-v_3 4 v − v 3 and at most 4 v − 8 4v-8 4 v − 8 . Thus v 3 ≥ 8 v_3\ge8 v 3 ≥ 8 . In a trisum the vertices off the common four-cycle retain their degrees, and the three operands meet only on that cycle. Each operand has vertices off the cycle by the definition of a trisum. By induction each operand has at least eight degree-three vertices, at most four of which are on the common cycle. In particular the result has degree-three vertices. The Heawood graph itself is cubic. Our graph is four-regular and has more than four vertices, so it cannot be a Pfaffian brace. Norine’s theorem [38 ] implies that a graph admitting a combination of at most three signings is Pfaffian. Thus at least four determinants are required, and (51) supplies four.
For the support { 0 , 1 , 2 , 4 } \{0,1,2,4\} { 0 , 1 , 2 , 4 } , subtract two from the offsets by a column shift. The available offsets are now − 2 , − 1 , 0 , 2 -2,-1,0,2 − 2 , − 1 , 0 , 2 . Divide the ordered rows and columns into ⌊ n / 5 ⌋ \lfloor n/5\rfloor ⌊ n /5 ⌋ groups of five, leaving the same remaining indices on the two sides. In a group use row labels r 0 , … , r 4 r_0,\ldots,r_4 r 0 , … , r 4 and column labels c 0 , … , c 4 c_0,\ldots,c_4 c 0 , … , c 4 . Take the K 3 , 3 K_{3,3} K 3 , 3 on rows r 1 , r 2 , r 3 r_1,r_2,r_3 r 1 , r 2 , r 3 and columns c 1 , c 2 , c 3 c_1,c_2,c_3 c 1 , c 2 , c 3 , except that the edges r 1 c 2 r_1c_2 r 1 c 2 and r 2 c 3 r_2c_3 r 2 c 3 are replaced by
r 1 − c 0 − r 0 − c 2 , r 2 − c 4 − r 4 − c 3 . r_1-c_0-r_0-c_2,\qquad r_2-c_4-r_4-c_3. r 1 − c 0 − r 0 − c 2 , r 2 − c 4 − r 4 − c 3 .
The three offsets on the first path are − 1 , 0 , 2 -1,0,2 − 1 , 0 , 2 , and on the second are 2 , 0 , − 1 2,0,-1 2 , 0 , − 1 . The seven retained direct edges have offsets in { − 2 , − 1 , 0 , 2 } \{-2,-1,0,2\} { − 2 , − 1 , 0 , 2 } . Thus every group contains a subdivision of K 3 , 3 K_{3,3} K 3 , 3 of the required kind. After deleting the vertices of their union, the remaining rows and columns can be matched on the diagonal. The union is consequently conformal, and [39 Corollary 4.3] gives the stated lower bound.
To see that there is a finite upper bound, regard each signed determinant coefficient as a function of the independent edge signs. For a fixed matching, averaging the signed determinants against the product of its edge signs isolates that matching monomial, up to its permutation sign. Every other monomial has an edge in the symmetric difference and averages to zero. Thus the signed determinants span the matching monomials. There are at most 4 n 4^n 4 n such monomials. For evaluation of the permanent itself, use (69) with b = 1 b=1 b = 1 , w = 4 w=4 w = 4 . There are a fixed number of transfer entries at each row, so the cost is O ( n ) O(n) O ( n ) arithmetic operations.
Proof of Theorem 4.6
For a monomial in the shift variables, the permanent coefficient counts all its contributing permutations. The determinant coefficient is the sum of their signs. Its absolute value equals the count precisely when all those signs agree. A common translation of the offsets permutes the columns and changes every determinant sign by the same factor. An automorphism of the group amounts to a simultaneous ordering of rows and columns. We may use both operations when testing the equality.
We first consider one connected component, so H = G H=G H = G . With one shift there is just one permutation. With two shifts, translate them to 0 , s 0,s 0 , s . The rows on each orbit of addition by s s s use either all diagonal entries or all s s s -entries, as in (57) . The number of orbits of the second kind is determined by the multiplicity of s s s . Since the orbits have the same size, the sign is determined. For three shifts write S = { 0 , s , t } S=\{0,s,t\} S = { 0 , s , t } , and for four shifts with equal pair sums write S = { 0 , s , t , s + t } S=\{0,s,t,s+t\} S = { 0 , s , t , s + t } . In both cases
Λ = { ( x , y ) ∈ Z 2 : x s + y t = 0 in G } . \Lambda=\{(x,y)\in\mathbb{Z}^2:xs+yt=0\text{ in }G\}. Λ = {( x , y ) ∈ Z 2 : x s + y t = 0 in G } .
has index ∣ G ∣ |G| ∣ G ∣ , and the quotient grid is the torus used in the proof of (52) . A monomial specifies the total lifted displacement. Its divisibility in Λ \Lambda Λ gives the number of nontrivial cycles by the primitive-vector argument. It also specifies how many vertices are moved. These two numbers determine the sign, giving the desired equality in these cases.
Suppose ∣ S ∣ = 4 |S|=4 ∣ S ∣ = 4 and no disjoint pairs in S S S have the same sum. A nonzero period would have an orbit size dividing four. In that case there is a period b b b of order two, and S S S is the union of pairs { x , x + b } , { y , y + b } \{x,x+b\},\{y,y+b\} { x , x + b } , { y , y + b } . The cross pairs { x , y + b } , { x + b , y } \{x,y+b\},\{x+b,y\} { x , y + b } , { x + b , y } have the same sum, a contradiction. Thus S S S is aperiodic. Also there is at most one z z z with z − S = S z-S=S z − S = S . Indeed the product of the reflections u ↦ z − u u\mapsto z-u u ↦ z − u and u ↦ z ′ − u u\mapsto z'-u u ↦ z ′ − u translates by z − z ′ z-z' z − z ′ .
If there is such a reflection with a nonfixed pair, choose a cyclic ordering s 0 , s 1 , s 2 , s 3 s_0,s_1,s_2,s_3 s 0 , s 1 , s 2 , s 3 of S S S in which that pair is adjacent. If every element is fixed, translating one to zero gives a set { 0 , a , b , c } \{0,a,b,c\} { 0 , a , b , c } in an elementary abelian two-group. The three nonzero elements generate this group and are independent, since a + b = c a+b=c a + b = c would give equal pair sums. The cycle ( 0 a + b ) (0\ a+b) ( 0 a + b ) uses a , b a,b a , b twice each and has negative sign. The two transpositions ( a ) ( c c + b ) (a)(c\ c+b) ( a ) ( c c + b ) have the same multiplicities and positive sign. All remaining vertices are fixed, so this gives the required cancellation. When there is no reflection, any cyclic ordering will do.
With the chosen ordering put y i = s i + s i + 1 y_i=s_i+s_{i+1} y i = s i + s i + 1 , using subscripts modulo four, and Y = { y 0 , y 1 , y 2 , y 3 } Y=\{y_0,y_1,y_2,y_3\} Y = { y 0 , y 1 , y 2 , y 3 } . Adjacent sums cannot be equal since they share one summand. Equality of opposite sums would contradict the condition on disjoint pairs. The four sums are therefore distinct, and they satisfy y 0 + y 2 = y 1 + y 3 y_0+y_2=y_1+y_3 y 0 + y 2 = y 1 + y 3 . Match s i s_i s i to y i y_i y i in one matching and to y i − 1 y_{i-1} y i − 1 in the other. The offsets used are s i + 1 s_{i+1} s i + 1 in the first case and s i − 1 s_{i-1} s i − 1 in the second. Thus each shift is used once, and the assignments to the four columns differ by a four-cycle. To extend both matchings using the same edges on the remaining vertices, it suffices by Hall’s theorem to prove
∣ ( X + S ) ∖ Y ∣ ≥ ∣ X ∣ ( X ⊆ G ∖ S ) . (110) \left|(X+S)\setminus Y\right| \ge|X| \qquad(X \subseteq G\setminus S). \tag*{(110)} ∣ ( X + S ) ∖ Y ∣ ≥ ∣ X ∣ ( X ⊆ G ∖ S ) . (110)
For X ≠ ∅ X \ne\varnothing X = ∅ write Δ = ∣ X + S ∣ − ∣ X ∣ \Delta=|X+S|-|X| Δ = ∣ X + S ∣ − ∣ X ∣ . Since at most four columns are deleted, we only need consider Δ ≤ 3 \Delta\le3 Δ ≤ 3 . If ∣ X ∣ = 1 |X|=1 ∣ X ∣ = 1 , then X + S X+S X + S is a translate of S S S and differs from the additive set Y Y Y . Thus at least one of its elements remains. If X + S = G X+S=G X + S = G , then Δ ≥ 4 \Delta\ge4 Δ ≥ 4 because X ⊆ G ∖ S X \subseteq G\setminus S X ⊆ G ∖ S . If X + S = G ∖ { z } X+S=G\setminus\{z\} X + S = G ∖ { z } , then Δ ≤ 3 \Delta\le3 Δ ≤ 3 requires X = G ∖ S X=G\setminus S X = G ∖ S and Δ = 3 \Delta=3 Δ = 3 . Since z − s ∉ X z-s\notin X z − s ∈ / X for each s ∈ S s\in S s ∈ S , we have z − S = S z-S=S z − S = S . The chosen ordering places a pair with sum z z z next to each other, so z ∈ Y z\in Y z ∈ Y . Only three elements of Y Y Y are then deleted from X + S X+S X + S , and (110) follows.
Suppose next that the stabilizer D D D of X + S X+S X + S is nontrivial and proper. Kneser’s inequality [40 , 41 ] gives
Δ ≥ ∣ X + D ∣ − ∣ X ∣ + ∣ S + D ∣ − ∣ D ∣ . \Delta\ge|X+D|-|X|+|S+D|-|D|. Δ ≥ ∣ X + D ∣ − ∣ X ∣ + ∣ S + D ∣ − ∣ D ∣.
The connected set S S S meets at least two D D D -cosets. If ∣ D ∣ = 2 |D|=2 ∣ D ∣ = 2 , meeting only two cosets would fill both, making S S S additive. Meeting at least three gives Δ ≥ 4 \Delta\ge4 Δ ≥ 4 . If ∣ D ∣ ≥ 4 |D|\ge4 ∣ D ∣ ≥ 4 , meeting at least two gives the same conclusion. The remaining case with Δ ≤ 3 \Delta\le3 Δ ≤ 3 has ∣ D ∣ = 3 |D|=3 ∣ D ∣ = 3 , two cosets met by S S S , and X + D = X X+D=X X + D = X . A split of S S S into two points in each coset would again make it additive. Indeed, the two differences in a group of order three agree up to sign, and ordering the pairs with equal differences gives equal cross sums. Hence, after translation, S = D ∪ { t } S=D\cup\{t\} S = D ∪ { t } . Translating X , S X,S X , S by the same element translates both X + S X+S X + S and Y Y Y by twice that element, preserving Hall’s inequality. In G / D G/D G / D the image of X X X omits 0 , t ˉ 0,\bar{t} 0 , t ˉ , since X + D = X X+D=X X + D = X and X X X avoids S S S . Its sum with { 0 , t ˉ } \{0,\bar{t}\} { 0 , t ˉ } therefore omits t ˉ \bar{t} t ˉ . Two points of Y Y Y are sums of t t t with a neighbor in D D D and lie in that missing coset. Thus
∣ ( X + S ) ∩ Y ∣ ≤ 2 < Δ = 3. |(X+S)\cap Y| \le2 < \Delta=3. ∣ ( X + S ) ∩ Y ∣ ≤ 2 < Δ = 3.
It remains to take aperiodic X + S X+S X + S , with ∣ X ∣ ≥ 2 |X|\ge2 ∣ X ∣ ≥ 2 and ∣ X + S ∣ ≤ ∣ G ∣ − 2 |X+S|\le|G|-2 ∣ X + S ∣ ≤ ∣ G ∣ − 2 . Kneser’s inequality gives Δ = 3 \Delta=3 Δ = 3 , so ∣ X + S ∣ = ∣ X ∣ + ∣ S ∣ − 1 |X+S|=|X|+|S|-1 ∣ X + S ∣ = ∣ X ∣ + ∣ S ∣ − 1 . This equality and the aperiodicity of X + S X+S X + S are the two hypotheses in Kemperman’s theorem [41 Theorem C] . An elementary pair with both sets of size at least two is either a pair of progressions, or lies in subgroup cosets whose sum fills a coset or omits one point. Here a four-point progression is additive. In the coset cases the subgroup would have to be G G G , since the differences of S S S generate G G G , and the size bound excludes these cases. Thus the theorem gives a proper nonzero subgroup D D D . Each of X , S X,S X , S is a union of full D D D -cosets and one distinguished part contained in a further coset. In the quotient, the sum of the two distinguished cosets has a unique representation. All other cosets in X + S X+S X + S are full because at least one of their summands is full.
The set S S S has a full coset, since its differences generate G G G , and its distinguished part cannot itself be full, since then X + S X+S X + S would be periodic. Write 4 = a ∣ D ∣ + b 4=a|D|+b 4 = a ∣ D ∣ + b , where a ≥ 1 a\ge1 a ≥ 1 is the number of full cosets and 1 ≤ b < ∣ D ∣ 1\le b<|D| 1 ≤ b < ∣ D ∣ is the size of the distinguished part. Since ∣ D ∣ ≥ 2 |D|\ge2 ∣ D ∣ ≥ 2 , the only possibility is ∣ D ∣ = 3 |D|=3 ∣ D ∣ = 3 , a = b = 1 a=b=1 a = b = 1 . Thus S = D ∪ { t } S=D\cup\{t\} S = D ∪ { t } after translation. Write X 0 X_0 X 0 for its distinguished part, and use bars for quotient images. The exceptional coset of X + S X+S X + S has exactly ∣ X 0 ∣ |X_0| ∣ X 0 ∣ points, since its only source is X 0 + { t } X_0+\{t\} X 0 + { t } . Consequently
∣ X + S ∣ = 3 ( ∣ X ˉ + { 0 , t ˉ } ∣ − 1 ) + ∣ X 0 ∣ , |X+S|=3(|\bar{X}+\{0,\bar{t}\}|-1)+|X_0|, ∣ X + S ∣ = 3 ( ∣ X ˉ + { 0 , t ˉ } ∣ − 1 ) + ∣ X 0 ∣ ,
∣ X ∣ = 3 ( ∣ X ˉ ∣ − 1 ) + ∣ X 0 ∣ . |X|=3(|\bar{X}|-1)+|X_0|. ∣ X ∣ = 3 ( ∣ X ˉ ∣ − 1 ) + ∣ X 0 ∣.
Their difference is three, giving ∣ X ˉ + { 0 , t ˉ } ∣ = ∣ X ˉ ∣ + 1 |\bar{X}+\{0,\bar{t}\}|=|\bar{X}|+1 ∣ X ˉ + { 0 , t ˉ } ∣ = ∣ X ˉ ∣ + 1 .
The quotient is cyclic with generator t ˉ \bar{t} t ˉ . The set X ˉ \bar{X} X ˉ omits zero. If it also omits t ˉ \bar{t} t ˉ , then its sum with { 0 , t ˉ } \{0,\bar{t}\} { 0 , t ˉ } omits t ˉ \bar{t} t ˉ , and again two points of Y Y Y are absent from X + S X+S X + S . If X ˉ \bar{X} X ˉ contains t ˉ \bar{t} t ˉ , this must be the distinguished coset of X 0 X_0 X 0 because t t t itself is excluded. In the cyclic order along t ˉ \bar{t} t ˉ , the number of new elements obtained by adding the translate X ˉ + t ˉ \bar{X}+\bar{t} X ˉ + t ˉ is the number of maximal intervals of X ˉ \bar{X} X ˉ . The displayed cardinality equality gives one interval. It begins at t ˉ \bar{t} t ˉ because zero is absent. If 2 t ˉ 2\bar{t} 2 t ˉ also belonged to it, the sum 2 t ˉ 2\bar{t} 2 t ˉ would have the representations t ˉ + t ˉ \bar{t}+\bar{t} t ˉ + t ˉ and $2\bar{t}+0, contrary to uniqueness for the distinguished cosets. Thus X ‾ = { t } \overline{X}=\{t\} X = { t } . If the quotient has order at least three, X + S X+S X + S omits D D D , which contains the two adjacent sums of elements of S S S that avoid t t t . If the quotient has order two, then ∣ X ∣ ≥ 2 |X|\ge2 ∣ X ∣ ≥ 2 and X ⊆ ( t + D ) ∖ { t } X\subseteq(t+D)\setminus\{t\} X ⊆ ( t + D ) ∖ { t } give ∣ X + S ∣ = 5 = ∣ G ∣ − 1 |X+S|=5=|G|-1 ∣ X + S ∣ = 5 = ∣ G ∣ − 1 , already treated. This proves (5.17) in every case. The common extension keeps the two signs opposite and gives the same shift counts.
For ∣ S ∣ ≥ 5 |S|\ge5 ∣ S ∣ ≥ 5 , take any five elements of S S S . If they contain a nonadditive four-set, translate it into the subgroup generated by its differences and use the construction just proved in one coset. In other cosets use a fixed translation. Returning to the original offsets changes both signs by the same factor and leaves equal multiplicities. Otherwise every four-subset of the five is additive. For three fixed elements a , b , c a,b,c a , b , c , the possible completions are a + b − c a+b-c a + b − c , a + c − b a+c-b a + c − b , b + c − a b+c-a b + c − a . The two other elements are distinct choices from these three. Relabeling and translating a a a to zero, the five elements can be written as { 0 , x , − x , c , c + x } \{0,x,-x,c,c+x\} { 0 , x , − x , c , c + x } . The two completions of { 0 , x , − x } \{0,x,-x\} { 0 , x , − x } must then be 2 x , − 2 x 2x,-2x 2 x , − 2 x . If c = − 2 x c=-2x c = − 2 x , then c + x = − x c+x=-x c + x = − x repeats a point. Hence c = 2 x c=2x c = 2 x , and 3 x = − 2 x 3x=-2x 3 x = − 2 x shows 5 x = 0 5x=0 5 x = 0 . The five elements are a translate of a group of order five. On that group the permutations i ↦ 2 i i\mapsto2i i ↦ 2 i and i ↦ − i i\mapsto-i i ↦ − i both use every shift once, since the differences from i i i are i i i and − 2 i -2i − 2 i . The first has a four-cycle on the nonzero elements, and the second has two transpositions there. Their signs are opposite, as required.
Now suppose g ≥ 2 g\ge2 g ≥ 2 . A cancellation in one connected block remains a cancellation after filling the other blocks with the same matching. For at most two shifts, all shift orbits have the same size, so the sign is still determined by the total multiplicities. It remains to consider repeated blocks for three shifts and for additive four-shift supports.
For three shifts on H H H , write the set as { 0 , s , t } \{0,s,t\} { 0 , s , t } . Let Λ = { ( x , y ) : x s + y t = 0 in H } \Lambda=\{(x,y):xs+yt=0\text{ in }H\} Λ = {( x , y ) : x s + y t = 0 in H } . We show that every lattice point ( r , z ) (r,z) ( r , z ) with r , z ≥ 0 r,z\ge0 r , z ≥ 0 and r + z ≤ m r+z\le m r + z ≤ m occurs as a total displacement. The zero point comes from the identity. For a nonzero point write ( r , z ) = k ( a , b ) (r,z)=k(a,b) ( r , z ) = k ( a , b ) with ( a , b ) (a,b) ( a , b ) primitive in Λ \Lambda Λ , and write ( a , b ) = e ( p , q ) (a,b)=e(p,q) ( a , b ) = e ( p , q ) , where p , q p,q p , q are relatively prime. Completing ( a , b ) (a,b) ( a , b ) to a basis of Λ \Lambda Λ gives a second vector whose determinant with ( a , b ) (a,b) ( a , b ) is ± m \pm m ± m . Hence ℓ ( x , y ) = p y − q x \ell(x,y)=py-qx ℓ ( x , y ) = p y − q x takes exactly the values ( m / e ) Z (m/e)\mathbb{Z} ( m / e ) Z on Λ \Lambda Λ .
Start in the strip 0 ≤ ℓ < p + q 0\le\ell<p+q 0 ≤ ℓ < p + q , at level zero. Take an east step if ℓ ≥ q \ell\ge q ℓ ≥ q and a north step otherwise. The new level is ℓ − q \ell-q ℓ − q or ℓ + p \ell+p ℓ + p , respectively. Both rules are addition of p p p modulo p + q p+q p + q within the strip. The levels are distinct until p + q p+q p + q steps have been made, since gcd ( p , p + q ) = 1 \gcd(p,p+q)=1 g cd( p , p + q ) = 1 . The resulting displacement has level zero, is parallel to ( p , q ) (p,q) ( p , q ) , and has coordinate sum p + q p+q p + q . It is therefore ( p , q ) (p,q) ( p , q ) . After e e e periods the displacement is ( a , b ) (a,b) ( a , b ) , so the walk closes modulo Λ \Lambda Λ . No earlier repetition is possible. Equal cosets within the strip must have equal levels, since the width is at most m / e m/e m / e , and equal levels in different periods are separated by a multiple of ( p , q ) (p,q) ( p , q ) . The first positive multiple in Λ \Lambda Λ is e ( p , q ) e(p,q) e ( p , q ) . Thus the projected walk is a simple cycle. Every integer is a possible starting level since p , q p,q p , q are coprime. Take k k k such strips starting at 0 , p + q , … , ( k − 1 ) ( p + q ) 0,p+q,\ldots,(k-1)(p+q) 0 , p + q , … , ( k − 1 ) ( p + q ) . Their total width is at most m / e m/e m / e , by r + z ≤ m r+z\le m r + z ≤ m . Levels from different strips are therefore distinct modulo m / e m/e m / e , and their projected cycles are disjoint. Fixing all other vertices realizes ( r , z ) (r,z) ( r , z ) .
If u , v u,v u , v form a lattice basis and u , v , u + v u,v,u+v u , v , u + v lie in this triangle, use displacements 0 , u + v 0,u+v 0 , u + v in two blocks, and u , v u,v u , v in the same two blocks for a second term. Both terms have the same total shift counts. Each of u , v , u + v u,v,u+v u , v , u + v is primitive, so the first term has one nontrivial cycle and the second has two. The signs are opposite.
We check when such a basis can be found. Suppose a difference s s s has order h < m h<m h < m , and translate the support to { 0 , s , t } \{0,s,t\} { 0 , s , t } . Put d = m / h d=m/h d = m / h . The image of t t t generates H / ⟨ s ⟩ H/\langle s\rangle H / ⟨ s ⟩ of order d d d . Thus any lattice vector has second coordinate divisible by d d d . There is a unique 0 ≤ v < h 0\le v<h 0 ≤ v < h with v s + d t = 0 vs+dt=0 v s + d t = 0 , so ( h , 0 ) , ( v , d ) (h,0),(v,d) ( h , 0 ) , ( v , d ) form a basis. Their sum is in the triangle if h + v + d ≤ h d h+v+d\le hd h + v + d ≤ h d . For d ≥ 3 d\ge3 d ≥ 3 this follows from v ≤ h − 1 v\le h-1 v ≤ h − 1 and ( h − 1 ) ( d − 2 ) ≥ 1 (h-1)(d-2)\ge1 ( h − 1 ) ( d − 2 ) ≥ 1 . For d = 2 d=2 d = 2 the only failure is v = h − 1 v=h-1 v = h − 1 . Then ( h − 1 ) s + 2 t = 0 (h-1)s+2t=0 ( h − 1 ) s + 2 t = 0 gives 2 t = s 2t=s 2 t = s , so the support is a progression.
Otherwise every difference generates H H H . It follows that H H H is cyclic, and its order is odd, since three residues modulo an even number cannot all have odd pairwise differences. We can take the support as { 0 , 1 , q } \{0,1,q\} { 0 , 1 , q } with 2 ≤ q ≤ ( m + 1 ) / 2 2 \le q \le(m+1)/2 2 ≤ q ≤ ( m + 1 ) /2 , using q ↦ 1 − q q \mapsto1-q q ↦ 1 − q if needed. The endpoints q = 2 , ( m + 1 ) / 2 q=2,(m+1)/2 q = 2 , ( m + 1 ) /2 give progressions. For the other values write m = k q + v m=kq+v m = k q + v , 0 ≤ v < q 0 \le v < q 0 ≤ v < q . The vectors ( v , k ) , ( v + q , k − 1 ) (v,k),(v+q,k-1) ( v , k ) , ( v + q , k − 1 ) belong to Λ \Lambda Λ , since each gives x + q y = m x+qy=m x + q y = m . Their determinant is − m -m − m , so they form a basis. Their sum is in the triangle precisely when
v + q − 1 ≤ k ( q − 2 ) . v+q-1 \le k(q-2). v + q − 1 ≤ k ( q − 2 ) .
For k ≥ 3 k \ge3 k ≥ 3 , q ≥ 4 q \ge4 q ≥ 4 , the left side is at most 2 q − 2 2q-2 2 q − 2 , and the right side is at least 3 q − 6 3q-6 3 q − 6 . For k ≥ 3 k \ge3 k ≥ 3 , q = 3 q=3 q = 3 , the only failure is k = 3 k=3 k = 3 , v = 2 v=2 v = 2 , namely m = 11 m=11 m = 11 , q = 3 q=3 q = 3 . Replacing q q q by ( 1 − q ) − 1 = 5 (1-q)^{-1}=5 ( 1 − q ) − 1 = 5 modulo eleven gives a basis satisfying the inequality. For k = 2 k=2 k = 2 a failure requires v = q − 1 v=q-1 v = q − 1 or v = q − 2 v=q-2 v = q − 2 . If m = 3 q − 1 m=3q-1 m = 3 q − 1 , replace q q q by q − 1 = 3 q^{-1}=3 q − 1 = 3 and use the preceding cases. If m = 3 q − 2 m=3q-2 m = 3 q − 2 , replace q q q by 1 − q − 1 1-q^{-1} 1 − q − 1 , whose representative is ( m − 1 ) / 2 (m-1)/2 ( m − 1 ) /2 . Then the new remainder is one, and the inequality holds for m ≥ 9 m \ge9 m ≥ 9 . These replacements are obtained by permuting and rescaling the three offsets, so they preserve the sign question.
The remaining support is { 0 , 1 , 3 } \{0,1,3\} { 0 , 1 , 3 } in Z 7 \mathbb{Z}_7 Z 7 . Listing the nonzero points r , z ≥ 0 r,z \ge0 r , z ≥ 0 with r + z ≤ 7 r+z \le7 r + z ≤ 7 and r + 3 z ≡ 0 ( m o d 7 ) r+3z \equiv0 \pmod7 r + 3 z ≡ 0 ( mod 7 ) gives ( 7 , 0 ) , ( 0 , 7 ) , ( 4 , 1 ) , ( 1 , 2 ) , ( 2 , 4 ) (7,0),(0,7),(4,1),(1,2),(2,4) ( 7 , 0 ) , ( 0 , 7 ) , ( 4 , 1 ) , ( 1 , 2 ) , ( 2 , 4 ) . For these points the numbers of nontrivial cycles are 1 , 1 , 1 , 1 , 2 1,1,1,1,2 1 , 1 , 1 , 1 , 2 by (52) . The numbers of moved vertices are 7 , 7 , 5 , 3 , 6 7,7,5,3,6 7 , 7 , 5 , 3 , 6 , so all signs are positive. This remains true for any number of equal blocks.
For a three-point progression translate its middle shift to zero. The nonzero steps are then the two neighboring steps on a cycle of length m m m . A nontrivial permutation cycle is a transposition or a full rotation in one direction. Let u , v u,v u , v count the blocks using the two rotations, and let j j j count the transpositions. The two nonzero shift counts are m u + j mu+j m u + j , m v + j mv+j m v + j , and the sign is ( − 1 ) ( m − 1 ) ( u + v ) + j (-1)^{(m-1)(u+v)+j} ( − 1 ) ( m − 1 ) ( u + v ) + j . At fixed shift counts, any changes satisfy Δ u = Δ v = − Δ j / m \Delta u=\Delta v=-\Delta j/m Δ u = Δ v = − Δ j / m . For even m m m both Δ j \Delta j Δ j and Δ u + Δ v \Delta u+\Delta v Δ u + Δ v are even, so the sign is fixed. For odd m m m in two blocks we have 0 ≤ j ≤ m − 1 0 \le j \le m-1 0 ≤ j ≤ m − 1 ; a change of j j j by a multiple of m m m is possible only if it is zero. With three blocks, the two rotations and an identity give the shift counts ( m , m ) (m,m) ( m , m ) and positive sign. The same counts come from m m m transpositions spread over three blocks. For m = 2 a + 1 m=2a+1 m = 2 a + 1 , take a , a , 1 a,a,1 a , a , 1 transpositions, which fit in those blocks. Their sign is negative. This gives cancellation for g ≥ 3 g \ge3 g ≥ 3 , using identity permutations in any extra blocks.
Finally consider repeated blocks for { 0 , s , t , s + t } \{0,s,t,s+t\} { 0 , s , t , s + t } . The triples { 0 , s , t } \{0,s,t\} { 0 , s , t } and { 0 , s , − t } \{0,s,-t\} { 0 , s , − t } occur up to translation and generate H H H . A cancellation in two blocks for either triple also works for the four shifts. Otherwise both triples are progressions or one is the order-seven case. If both are progressions, zero cannot be the middle point of { 0 , s , t } \{0,s,t\} { 0 , s , t } , since s + t ≠ 0 s+t \ne0 s + t = 0 . Interchanging s , t s,t s , t if needed gives t = 2 s t=2s t = 2 s . The other triple is then { 0 , s , − 2 s } \{0,s,-2s\} { 0 , s , − 2 s } ; its midpoint condition gives 4 s = 0 4s=0 4 s = 0 or 5 s = 0 5s=0 5 s = 0 . Since s s s generates H H H and the four offsets are distinct, m = 4 m=4 m = 4 or 5 5 5 .
For m = 4 m=4 m = 4 label H H H by 0 , 1 , 2 , 3 0,1,2,3 0 , 1 , 2 , 3 . Compare the identity in one block and ( 3 , 2 , 0 , 1 ) (3,2,0,1) ( 3 , 2 , 0 , 1 ) in the other with the pair ( 2 , 1 , 0 , 3 ) , ( 3 , 1 , 2 , 0 ) (2,1,0,3),(3,1,2,0) ( 2 , 1 , 0 , 3 ) , ( 3 , 1 , 2 , 0 ) , in one-line notation. The four-cycle in the first pair has shifts 3 , 1 , 2 , 2 3,1,2,2 3 , 1 , 2 , 2 , while the two transpositions in the second have shifts 2 , 2 2,2 2 , 2 and 3 , 1 3,1 3 , 1 on their moved vertices. Both pairs give a 4 b c 2 d a^4bc^2d a 4 b c 2 d . The first pair has negative sign and the second positive.
In the order-seven case replace t t t by − t -t − t if necessary, which translates the four-set. With a change of generator the exceptional triple is { 0 , 1 , t } \{0,1,t\} { 0 , 1 , t } with t = 3 t=3 t = 3 or 5 5 5 . For t = 3 t=3 t = 3 , multiplying the four-set by two and adding one gives { 0 , 1 , 2 , 3 } \{0,1,2,3\} { 0 , 1 , 2 , 3 } . For t = 5 t=5 t = 5 , adding two does so. It therefore remains to treat consecutive four shifts at odd m = 2 j + 1 ≥ 5 m=2j+1 \ge5 m = 2 j + 1 ≥ 5 , including m = 5 m=5 m = 5 .
In two blocks use the pairs of monomials
( a m , b 2 c m − 4 d 2 ) , ( a j + 1 c j − 1 d , a j b 2 c j − 2 d ) . (a^m,\ b^2c^{m-4}d^2),\qquad(a^{j+1}c^{j-1}d,\ a^jb^2c^{j-2}d). ( a m , b 2 c m − 4 d 2 ) , ( a j + 1 c j − 1 d , a j b 2 c j − 2 d ) .
Their products agree. The second monomial in the first pair has winding number two, and reversal of offsets makes it a winding-one monomial with m − 2 m-2 m − 2 moved vertices. Since m m m is odd, the column shift in the reversal has positive sign, as does the resulting cycle of odd length m − 2 m-2 m − 2 .
In the second pair both monomials have winding number one, with j j j and j + 1 j+1 j + 1 moved vertices, respectively. Their signs multiply to − 1 -1 − 1 . The occurrence of all these monomials follows from (60) , or directly by the winding-one compositions. Thus there is cancellation in two blocks, and additional blocks can be filled with the identity. This completes the classification.
The signed and block-constant cases of Theorem 4.7
5.18. For square matrices X , Y X,Y X , Y of the same order, expand each product in per ( X + Y ) \operatorname{per}(X+Y) per ( X + Y ) according to the rows that use an entry of Y Y Y . If those rows are I I I and their columns are J J J , the remaining rows must use the remaining columns. Hence
per ( X + Y ) = ∑ ∣ I ∣ = ∣ J ∣ per X [ I c , J c ] per Y [ I , J ] . \operatorname{per}(X+Y)=\sum_{|I|=|J|}\operatorname{per}X[I^c,J^c]\operatorname{per}Y[I,J]. per ( X + Y ) = ∣ I ∣ = ∣ J ∣ ∑ per X [ I c , J c ] per Y [ I , J ] .
Apply this identity to the diagonal and forward parts of M M M . Let I i I_i I i be the rows of block i i i using its forward block, and J i J_i J i the columns of block i i i filled by a forward block. Since J i J_i J i is filled from I i − 1 I_{i-1} I i − 1 , we have ∣ J i ∣ = ∣ I i − 1 ∣ |J_i|=|I_{i-1}| ∣ J i ∣ = ∣ I i − 1 ∣ . The remaining rows and columns at block i i i use A i A_i A i , so b − ∣ I i ∣ = b − ∣ J i ∣ b-|I_i|=b-|J_i| b − ∣ I i ∣ = b − ∣ J i ∣ . Thus every ∣ I i ∣ |I_i| ∣ I i ∣ and ∣ J i ∣ |J_i| ∣ J i ∣ equals a common k k k . For fixed sets, the sum of the matching weights is
∏ i = 0 m − 1 per A i [ I i c , J i c ] per B i [ I i , J i + 1 ] , (111) \prod_{i=0}^{m-1}\operatorname{per}A_i[I_i^c,J_i^c]\operatorname{per}B_i[I_i,J_{i+1}], \tag*{(111)} i = 0 ∏ m − 1 per A i [ I i c , J i c ] per B i [ I i , J i + 1 ] , (111)
with indices read modulo m m m .
We first compare a minor of a single block with its signed determinant. Let its order be r r r and let d , d + 1 d,d+1 d , d + 1 be the two internal offsets. For a contributing matching, list its rows in increasing order. The lifted targets are j + d j+d j + d or j + d + 1 j+d+1 j + d + 1 . As j j j increases, the targets are nondecreasing. Equality would repeat a column, and the difference between the first and last targets is at most b b b . Equality in this last bound would also repeat a column modulo b b b . Thus the targets are strictly increasing and span less than b b b . If e e e of them are at least b b b , they form a suffix. On reducing modulo b b b , each of these e e e targets moves ahead of all the other r − e r-e r − e targets. There are e ( r − e ) e(r-e) e ( r − e ) inversions and no others, so the term has sign ( − 1 ) e ( r − e ) = ( − 1 ) e ( r − 1 ) (-1)^{e(r-e)}=(-1)^{e(r-1)} ( − 1 ) e ( r − e ) = ( − 1 ) e ( r − 1 ) . Multiplying each wrapped entry by ( − 1 ) r − 1 (-1)^{r-1} ( − 1 ) r − 1 cancels that sign. Therefore
per D [ I , J ] = det D ( ( − 1 ) r − 1 [ I , J ] ) , ∣ I ∣ = ∣ J ∣ = r , \operatorname{per}D[I,J]=\det D\bigl((-1)^{r-1}[I,J]\bigr),\qquad|I|=|J|=r, per D [ I , J ] = det D ( ( − 1 ) r − 1 [ I , J ] ) , ∣ I ∣ = ∣ J ∣ = r ,
with both values one when r = 0 r=0 r = 0 .
Now expand the determinant F ϵ ( x ) F_\epsilon(x) F ϵ ( x ) by the same sets I i , J i I_i,J_i I i , J i of size k k k . In a block, moving a chosen row set I i I_i I i to the end while preserving its order takes
∑ j ∈ I i ( b − 1 − j ) − ( k 2 ) \sum_{j\in I_i}(b-1-j)-\binom{k}{2} j ∈ I i ∑ ( b − 1 − j ) − ( 2 k )
interchanges. The first sum counts every row after a chosen row, and the subtraction removes the pairs of chosen rows. The columns give the same expression with J i J_i J i . The terms depending only on b , k b,k b , k occur twice in every block and cancel in parity. The total factor from these row and column orderings is thus ( − 1 ) ∑ i ( ∑ I i + ∑ J i ) (-1)^{\sum_i(\sum I_i+\sum J_i)} ( − 1 ) ∑ i ( ∑ I i + ∑ J i ) . In the forward minors, the left and right factors E E E supply exactly the same sign, since the forward columns in B i B_i B i are J i + 1 J_{i+1} J i + 1 . The two factors cancel.
After this reordering, separate the diagonal groups from the forward groups in both rows and columns. These two shuffles have the same sign and cancel. In the forward part the column groups are cyclically shifted by one group of size k k k . Moving such a group past the other m − 1 m-1 m − 1 groups has sign ( − 1 ) ( m − 1 ) k 2 = ( − 1 ) ( m − 1 ) k (-1)^{(m-1)k^2}=(-1)^{(m-1)k} ( − 1 ) ( m − 1 ) k 2 = ( − 1 ) ( m − 1 ) k . This cancels the sign from multiplying the last forward block by ( − 1 ) m − 1 x (-1)^{m-1}x ( − 1 ) m − 1 x , which also supplies x k x^k x k .
Choose ϵ = ( − 1 ) k − 1 \epsilon=(-1)^{k-1} ϵ = ( − 1 ) k − 1 . The twist in a forward minor is then the one for order k k k . In a diagonal minor it is ( − 1 ) b ϵ = ( − 1 ) b − k + 1 (-1)^b\epsilon=(-1)^{b-k+1} ( − 1 ) b ϵ = ( − 1 ) b − k + 1 , the one for order b − k b-k b − k . The coefficient [ x k ] F ϵ ( x ) [x^k]F_\epsilon(x) [ x k ] F ϵ ( x ) is therefore the sum of (5.18) over the sets of that size. For odd k k k use F + F_+ F + , and for even k k k use F − F_{-} F − . The odd coefficients of F + F_{+} F + sum to ( F + ( 1 ) − F + ( − 1 ) ) / 2 (F_{+}(1)-F_{+}(-1))/2 ( F + ( 1 ) − F + ( − 1 )) /2 ; the even coefficients of F − F_{-} F − sum to ( F − ( 1 ) + F − ( − 1 ) ) / 2 (F_{-}(1)+F_{-}(-1))/2 ( F − ( 1 ) + F − ( − 1 )) /2 . Adding them proves (4.22).
For a block circulant with two offsets, a column shift makes one of them zero. Order the block indices along the cycles of addition by the other offset. Rows from one cycle can use only columns in that cycle. The matrix is therefore a direct sum of cyclic block matrices of the kind just considered. For block Hankel matrices, reversing the order of the block columns changes a block depending on i + j i+j i + j into one depending on i − j i-j i − j , with no change to its internal row and column order. The permanent is preserved by this column permutation.
For the banded block-constant case in Corollary 4.8, the rows inside a single block are identical. At a cut between blocks, only the last q q q row blocks can have entries into later blocks. Their restricted rows span a space of dimension at most q q q . At a cut inside a block there is at most one further row type from that block. The upper ranks are therefore at most q + 1 q+1 q + 1 . Transposition gives at most p + 1 p+1 p + 1 for the reverse cut ranks. Theorem 3.4 gives the determinant assertion.
Finally, for (4.25) fix the numbers k i j k_{ij} k ij of rows in class i i i using columns in class j j j . Within row class i i i , the number of ways to assign the column classes is m i ! / ∏ j k i j ! m_i!/\prod_j k_{ij}! m i ! / ∏ j k ij ! . For column class j j j , its n j n_j n j columns can be assigned to the selected rows in n j ! n_j! n j ! ways. Every such matching has weight ∏ i , j b i j k i j \prod_{i,j} b_{ij}^{k_{ij}} ∏ i , j b ij k ij . Multiplying these factors and summing over the arrays with the required row and column sums gives (4.25). The entries outside the first r − 1 r-1 r − 1 rows and first s − 1 s-1 s − 1 columns of the array are determined by its margins. This gives the stated bound on the number of arrays.
The open-diagonal and Hankel formulas
For (4.32), embed A A A as the first diagonal block of A ⊕ I N − n A\oplus I_{N-n} A ⊕ I N − n . Any contributing permutation matches the two blocks separately. Its restriction to the added block is the identity, with weight one, so the permanent is per A \operatorname{per} A per A . The allowed ordinary offsets of A A A , reduced modulo N N N , give its cyclic offsets in the enlarged matrix. The only added entries are diagonal and thus allowed when 0 ∈ S 0\in S 0 ∈ S . When N = n N=n N = n there are no added entries. By the hypotheses on these cyclic offsets, Corollary 4.5(3) applies. For the four displayed offsets, the chosen N N N makes d 1 + d 2 ≡ d 3 + d 4 ( m o d N ) d_1+d_2\equiv d_3+d_4\pmod{N} d 1 + d 2 ≡ d 3 + d 4 ( mod N ) . If the residues are distinct, subtracting d 1 d_1 d 1 makes the second offset the sum of the third and fourth offsets. If any coincide there are at most three offsets. Since each d i d_i d i lies between 1 − n 1-n 1 − n and n − 1 n-1 n − 1 , the chosen N N N is O ( n ) O(n) O ( n ) . This gives the stated cost in these cases.
For (4.33), at least one of q / g , p / g q/g,p/g q / g , p / g is odd, so the stated choice of x , y x,y x , y satisfies x q / g y − p / g = − 1 x^{q/g}y^{-p/g}=-1 x q / g y − p / g = − 1 . Work in one residue class modulo g g g . On the grid, the map ( u , v ) ↦ p u + q v (u,v)\mapsto pu+qv ( u , v ) ↦ p u + q v has kernel generated by ( q / g , − p / g ) (q/g,-p/g) ( q / g , − p / g ) . Its image is g Z g\mathbb{Z} g Z . Thus the grid modulo that kernel represents the vertices of the residue class, up to an initial translation, and the east, north and northeast edges are the three nonzero offsets. The finite open support retains only those edges with both ends in the matrix.
The quotient is a cylinder. Each nontrivial cycle of a contributing permutation gives an embedded simple curve in it. Its lifted displacement is nonzero, since the steps increase at least one coordinate, and is a multiple of the kernel generator. A simple essential curve on a cylinder has winding number 1 1 1 or − 1 -1 − 1 . One can see this by identifying the cylinder with the punctured plane and using the index of a simple closed curve about the missing point. Thus the product of the x , y x,y x , y factors on the cycle is ( x q / g y − p / g ) ± 1 = − 1 (x^{q/g}y^{-p/g})^{\pm1}=-1 ( x q / g y − p / g ) ± 1 = − 1 . The permutation sign of a cycle of length a a a , together with its off-diagonal minus signs, is another factor ( − 1 ) a − 1 ( − 1 ) a = − 1 (-1)^{a-1}(-1)^a=-1 ( − 1 ) a − 1 ( − 1 ) a = − 1 . They cancel. A fixed point has its original weight, so the determinant and permanent agree term by term on each residue class and hence on the whole matrix.
For the rational series, write one of them as ∑ j ≥ 0 γ j z j = P ( z ) / Q ( z ) \sum_{j\geq0}\gamma_j z^j=P(z)/Q(z) ∑ j ≥ 0 γ j z j = P ( z ) / Q ( z ) with Q ( 0 ) = 1 Q(0)=1 Q ( 0 ) = 1 . Take d ≥ 1 d\geq1 d ≥ 1 at least deg Q \deg Q deg Q and larger than deg P \deg P deg P , and write Q ( z ) = 1 + q 1 z + ⋯ + q d z d Q(z)=1+q_1z+\cdots+q_dz^d Q ( z ) = 1 + q 1 z + ⋯ + q d z d , using zeros as needed. The coefficients in Q ( z ) ∑ j γ j z j Q(z)\sum_j\gamma_jz^j Q ( z ) ∑ j γ j z j give
γ j = − ∑ a = 1 d q a γ j − a ( j ≥ d ) . \gamma_j=-\sum_{a=1}^{d}q_a\gamma_{j-a}\qquad(j\geq d). γ j = − a = 1 ∑ d q a γ j − a ( j ≥ d ) .
The vector ( γ j , … , γ j + d − 1 ) T (\gamma_j,\ldots,\gamma_{j+d-1})^{\mathsf{T}} ( γ j , … , γ j + d − 1 ) T therefore advances by a fixed matrix, with its last row ( − q d , … , − q 1 ) (-q_d,\ldots,-q_1) ( − q d , … , − q 1 ) and its other rows shifting the coordinates. Its first coordinate is γ j \gamma_j γ j . This gives the constant generators on both sides of (72) , also for polynomial series or repeated roots. For a cut after k k k , for example,
t j − i = ( p T B k − i ) ( B j − k − 1 q ) ( i ≤ k < j ) . t_{j-i}=(p^{\mathsf{T}}B^{k-i})(B^{j-k-1}q)\qquad(i\leq k<j). t j − i = ( p T B k − i ) ( B j − k − 1 q ) ( i ≤ k < j ) .
Thus the upper cut block has rank bounded by the order of B B B . The lower block is treated with the other constant generator. Since the diagonal weight is also constant, (31) gives ( T n n ) 1 , 1 (T_n^n)_{1,1} ( T n n ) 1 , 1 . The bound on the degrees of the generators fixes the number of coordinate variables, while their allowed degrees grow with n n n . The adjugate formula for the constant-state entry of ( I − z T n ) − 1 (I-zT_n)^{-1} ( I − z T n ) − 1 gives (76) .
In characteristic p > 0 p>0 p > 0 , the same companion constructions give constant generators. Corollary 3.5 uses the fixed basis B p , r , s B_{p,r,s} B p , r , s , so its common matrix T p T_p T p is independent of the order. Thus per A n = ( T p n ) 1 , 1 \operatorname{per} A_n=(T_p^n)_{1,1} per A n = ( T p n ) 1 , 1 for every n ≥ 0 n\geq0 n ≥ 0 . Summing the formal geometric series gives
∑ n ≥ 0 per A n z n = ( ( I − z T p ) − 1 ) 1 , 1 . \sum_{n\geq0}\operatorname{per} A_n z^n=((I-zT_p)^{-1})_{1,1}. n ≥ 0 ∑ per A n z n = (( I − z T p ) − 1 ) 1 , 1 .
The cofactor identity proves (79) . The denominator has constant coefficient one. Repeated squaring gives the stated field-operation bound.
For the characteristic-three example, on x h y h x^h y^h x h y h the diagonal coefficient of the ordered operator is ( 1 + 2 h ) τ h (1+2h)\tau^h ( 1 + 2 h ) τ h , the coefficient at degree h + 1 , h + 1 h+1,h+1 h + 1 , h + 1 is τ h + 1 \tau^{h+1} τ h + 1 , and that at degree h − 1 , h − 1 h-1,h-1 h − 1 , h − 1 is h 2 τ h − 1 h^2\tau^{h-1} h 2 τ h − 1 . Reading these for h = 0 , 1 , 2 h=0,1,2 h = 0 , 1 , 2 in characteristic three gives (81) . The term x 3 y 3 x^3y^3 x 3 y 3 is zero in the quotient and cannot contribute at a later step by Lemma 5.1 . The two determinants in (79) give (93) .
For the geometric parameters choose B = p B=p B = p , C = q C=q C = q , D i = ∂ x D_i=\partial_x D i = ∂ x , E i = ∂ y E_i=\partial_y E i = ∂ y , L i = a x L_i=ax L i = a x , M i = b y M_i=by M i = b y . On a monomial x h y h x^h y^h x h y h , the substitution alone multiplies it by τ h \tau^h τ h . The two single derivatives followed by insertion of a x ax a x or b y by b y give h a q τ h − 1 haq\tau^{h-1} ha q τ h − 1 and h b p τ h − 1 hbp\tau^{h-1} hb p τ h − 1 at degree h , h h,h h , h , when h ≥ 1 h\geq1 h ≥ 1 . The double derivative gives h 2 τ h − 1 h^2\tau^{h-1} h 2 τ h − 1 at degree h − 1 , h − 1 h-1,h-1 h − 1 , h − 1 . The two other terms give c τ h c\tau^h c τ h at degree h , h h,h h , h and a b τ h ab\tau^h ab τ h at degree h + 1 , h + 1 h+1,h+1 h + 1 , h + 1 . These are precisely the entries of (80) . At h = 0 h=0 h = 0 the derivatives vanish. At the upper end the projection removes degree N + 1 , N + 1 N+1,N+1 N + 1 , N + 1 .
For I − z T n I-zT_n I − z T n , expanding a leading tridiagonal minor in its last row gives a continuant relation. If Δ h \Delta_h Δ h is the leading determinant through index h h h , then for h ≥ 1 h\geq1 h ≥ 1
Δ h = ( 1 − z ( T n ) h , h ) Δ h − 1 − z 2 ( T n ) h − 1 , h ( T n ) h , h − 1 Δ h − 2 , Δ − 1 = 1 , Δ 0 = 1 − z ( T n ) 0 , 0 . \Delta_h=(1-z(T_n)_{h,h})\Delta_{h-1}-z^2(T_n)_{h-1,h}(T_n)_{h,h-1}\Delta_{h-2},\qquad\Delta_{-1}=1,\quad\Delta_0=1-z(T_n)_{0,0}. Δ h = ( 1 − z ( T n ) h , h ) Δ h − 1 − z 2 ( T n ) h − 1 , h ( T n ) h , h − 1 Δ h − 2 , Δ − 1 = 1 , Δ 0 = 1 − z ( T n ) 0 , 0 .
Thus the determinants for T n , T n ′ T_n,T'_n T n , T n ′ can be formed by products of two-by-two polynomial matrices of degree at most two. Multiply consecutive groups of factors in a balanced tree. At a level with groups of length a a a , there are O ( n / a ) O(n/a) O ( n / a ) products of degree O ( a ) O(a) O ( a ) . The assumption on M ( n ) M(n) M ( n ) gives O ( M ( n ) ) O(M(n)) O ( M ( n )) operations at each level. Inverting the denominator series by doubling the precision and multiplying by the numerator also lies within O ( M ( n ) log ( n + 1 ) ) O(M(n)\log(n+1)) O ( M ( n ) log ( n + 1 )) .
For the cycle polynomial, at the cut after vertex k k k , give a path with end i i i and start j j j the factors a p k − i ap^{k-i} a p k − i and b q k − j bq^{k-j} b q k − j . These are the weights of the edges from i i i to k + 1 k+1 k + 1 and from k + 1 k+1 k + 1 to j j j . A retained end gains a factor p p p at the next cut and a retained start gains a factor q q q . Suppose there are h h h old paths. If the new vertex is isolated, it contributes a b ab ab from its new endpoints, with τ h \tau^h τ h from the old ones. A loop instead contributes z c τ h zc\tau^h z c τ h . If an edge from an old end enters the new vertex, that end is consumed and the new end contributes a a a . There are h h h choices and the retained endpoints gain p h − 1 q h p^{h-1}q^h p h − 1 q h , giving h a q τ h − 1 haq\tau^{h-1} ha q τ h − 1 . The reverse choice gives h b p τ h − 1 hbp\tau^{h-1} hb p τ h − 1 . In the two-edge case the old end and start can lie on the same path in h h h ways, closing a cycle, or on distinct paths in h ( h − 1 ) h(h-1) h ( h − 1 ) ways, joining them. In both cases h − 1 h-1 h − 1 paths remain and the retained endpoints gain τ h − 1 \tau^{h-1} τ h − 1 . The coefficient for this step is h ( h − 1 + z ) τ h − 1 h(h-1+z)\tau^{h-1} h ( h − 1 + z ) τ h − 1 . These cases give T n ( z ) T_n(z) T n ( z ) . A complete cover starts and ends with no paths. Taking the corresponding entry of T n ( z ) n T_n(z)^n T n ( z ) n , then using the same adjugate identity, proves (92) . For the full coefficient list, P A n P_{A_n} P A n has degree at most n n n . Evaluating it at 0 , 1 , … , n 0,1,\ldots,n 0 , 1 , … , n and interpolating gives the stated bound for all its coefficients.
For ( ρ ∣ i − j ∣ ) (\rho^{\lvert i-j\rvert}) ( ρ ∣ i − j ∣ ) , the factorization ρ j − i = ρ k − i ρ j − k \rho^{j-i}=\rho^{k-i}\rho^{j-k} ρ j − i = ρ k − i ρ j − k at a cut shows that the two cut ranks are one for ρ ≠ 0 \rho\ne0 ρ = 0 . For its determinant, subtract ρ \rho ρ times row i − 1 i-1 i − 1 from row i i i , for i = n , … , 2 i=n,\ldots,2 i = n , … , 2 . In row i i i this zeros columns j < i j<i j < i , and the new diagonal entry is 1 − ρ 2 1-\rho^2 1 − ρ 2 . The first diagonal entry remains one.
For the odd-row, even-column submatrix, write its entries as ρ ∣ 2 a − 1 − 2 b ∣ \rho^{\lvert2a-1-2b\rvert} ρ ∣ 2 a − 1 − 2 b ∣ , 1 ≤ a , b ≤ m 1\le a,b\le m 1 ≤ a , b ≤ m . Subtract ρ 2 \rho^2 ρ 2 times column b − 1 b-1 b − 1 from column b b b in decreasing order of b b b . When a < b a<b a < b , the original entry in column b b b is ρ 2 \rho^2 ρ 2 times the one in column b − 1 b-1 b − 1 . The diagonal entries after subtraction are ρ \rho ρ and ρ ( 1 − ρ 2 ) \rho(1-\rho^2) ρ ( 1 − ρ 2 ) for b ≥ 2 b\ge2 b ≥ 2 . This gives the stated minor and is unaffected by changing the full matrix diagonal.
For the weighted clique-width bound, consider the binary disjoint-union tree of an expression with n ≥ 2 n\ge2 n ≥ 2 vertices. There is a proper subexpression with between n / 3 n/3 n /3 and 2 n / 3 2n/3 2 n /3 vertices. Indeed, starting at the root, descend through a child with more than 2 n / 3 2n/3 2 n /3 vertices while one exists. At the final union at least one of its two children has at least n / 3 n/3 n /3 vertices and both have at most 2 n / 3 2n/3 2 n /3 . Vertices with the same label in that child can never be separated by subsequent relabelings. They therefore get equal weights to each fixed vertex outside the child. For 0 < ρ < 1 0<\rho<1 0 < ρ < 1 , a weight ρ ∣ i − j ∣ \rho^{\lvert i-j\rvert} ρ ∣ i − j ∣ determines the distance to that fixed vertex j j j , and at most two vertices have a given distance. Thus each label accounts for at most two of the at least n / 3 n/3 n /3 vertices. This gives at least n / 6 n/6 n /6 labels.
Finally, for an anti-diagonal support, ( H J n ) i j = H i , n − 1 − j (HJ_n)_{ij}=H_{i,n-1-j} ( H J n ) ij = H i , n − 1 − j . If H i j = h i + j H_{ij}=h_{i+j} H ij = h i + j , the new entry is h n − 1 + i − j = t j − i h_{n-1+i-j}=t_{j-i} h n − 1 + i − j = t j − i , with t d = h n − 1 − d t_d=h_{n-1-d} t d = h n − 1 − d . The same index calculation carries arbitrary entries on an anti-diagonal into the corresponding ordinary diagonal. This proves the transfer and padding formulas for H H H . For the direct signing, K K K is the signed matrix for H J n HJ_n H J n , so per H = det K \operatorname{per} H=\det K per H = det K . The reversal permutation has n ( n − 1 ) / 2 n(n-1)/2 n ( n − 1 ) /2 inversions. Consequently det ( K J n ) = ( − 1 ) n ( n − 1 ) / 2 det K \det(KJ_n)=(-1)^{n(n-1)/2}\det K det ( K J n ) = ( − 1 ) n ( n − 1 ) /2 det K , giving the signed Hankel formula. The special cases of Corollary 4.12 use the same substitution in the three displayed Toeplitz families. In positive characteristic, apply Corollary 3.5 to the reversed matrix. This proves Theorem 4.11 (4).
Proof of Proposition 4.13
Proof of Proposition 4.13. In a contributing permutation for S S S , the first m m m rows must use the last m m m columns and the last m m m rows must use the first m m m columns. These are independent choices, so per S = per B per B T = ( per B ) 2 \operatorname{per} S=\operatorname{per} B\operatorname{per} B^{\mathsf T}=(\operatorname{per} B)^2 per S = per B per B T = ( per B ) 2 . For C C C each set of rows uses the columns of its own block. Both blocks have permanent per B \operatorname{per} B per B by row and column reversal, which gives per C = ( per B ) 2 \operatorname{per} C=(\operatorname{per} B)^2 per C = ( per B ) 2 . Conjugating C C C by J 2 m J_{2m} J 2 m reverses both blocks and exchanges them, so C C C is centrosymmetric.
Both blocks of D = S ⊕ ( J 2 m S J 2 m ) D=S\oplus(J_{2m}SJ_{2m}) D = S ⊕ ( J 2 m S J 2 m ) are symmetric. Reversing the full row and column orders exchanges them after internal reversal, so D D D is also centrosymmetric. Its permanent is ( per S ) 2 = ( per B ) 4 (\operatorname{per} S)^2=(\operatorname{per} B)^4 ( per S ) 2 = ( per B ) 4 . For ( 0 , 1 ) (0,1) ( 0 , 1 ) matrices these values determine per B \operatorname{per} B per B by taking the nonnegative integer square or fourth root, respectively.
References
Department of Mathematics, Sukkur IBA University, Sukkur 65200, Pakistan
Email address: fazeel.anwar@iba-suk.edu.pk