AI Use: Discussions with GPT-6 Astra suggested the explicit polynomial used to separate homogeneous algebraic branching programs (ABPs) of fixed width and degree from the Zariski closure of this class. This polynomial was found after the matrix encoding of polynomials through translations of universal ABPs and our previously established variants of Burnside Theorems for semigroup congruences were given to the model. The proofs of Theorem 2 and Corollary 1 were developed through an interaction with GPT-6. We also used GPT-6 to help develop several examples.
Funding: Shirmohammadi is supported by the ERC Synergy grant Project 101224640 (VePaSS). Worrell is supported by EPSRC Fellowship EP/X033813/1. Guttenberg is supported by the ERC Consolidator grant Polaris 101231447.
Introduction
In 1902, Burnside asked whether a finitely generated group is necessarily finite if every element has finite order, that is, if some positive power of each element equals the identity. Novikov and Adian showed in 1968 that the answer is negative even when the orders of all elements are bounded by a common constant [1 ] . By contrast, Schur proved that the answer is positive for groups of invertible matrices over the complex numbers [2 ] . Kaplansky generalised Schur’s result from the complex numbers to arbitrary fields (see [3 ] ). McNaughton and Zalcstein further extended this conclusion to matrix semigroups over arbitrary fields [4 ] . They proved that a finitely generated matrix semigroup is finite whenever each element is torsion, meaning that its powers take only finitely many distinct values. Among other things, this shows that infiniteness of finitely generated matrix semigroups is semi-decidable (a witness of infiniteness is a matrix of infinite order) and hence finiteness of finitely generated matrix semigroups is decidable.
The positive answer to the Burnside problem for matrix groups generalises to quotients of matrix groups by Zariski closed normal subgroups, since such quotients can be embedded in matrix groups in higher ambient dimension. This finiteness result is a key ingredient in the first proposed algorithm for computing the Zariski closure of a finitely generated matrix group [5 ] , resolving an open problem in quantum computation. In subsequent joint work, again relying on such a Burnside finiteness theorem, we obtained an elementary algorithm for computing this closure [6 ] .
In this paper, we study Burnside-type questions for quotients of matrix semigroups. Our approach is based on congruences, that is, equivalence relations of matrix semigroups that are compatible with matrix multiplication. Given a finite set X X X of rational matrices and a congruence ∼ \sim ∼ on the semigroup ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ they generate, we ask for sufficient conditions that
the quotient ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ be finite, and
there be a uniform bound on the length of the shortest product of generators from X X X representing each equivalence class.
The classical results on the Burnside problem for matrix semigroups do not apply here, since the quotient need not admit a faithful finite-dimensional representation in a monoid of matrices with entries in a field.1 We establish three quantitative finiteness criteria for such quotients and present two applications to border complexity.
Border complexity, also known as approximate complexity, studies polynomials and tensors obtained as limits of parametrised families of bounded complexity. The distinction between exact and approximate computability is central to geometric complexity theory, where questions about computational power are expressed through polynomial and tensor orbit closures [7 , 8 , 9 ] . This distinction is already visible in small-width algebraic branching programs, as approximation can increase the expressive power in these models [10 ] . Our two applications study these distinctions through semigroup congruences:
First, for homogeneous algebraic branching programs, we separate the class of polynomials computable by algebraic branching programs of fixed width and degree from its Zariski closure.
Second, we construct a finitely generated matrix monoid under whose action the orbit closure of the permanent is not contained in that of the determinant. Curiously, under the congruence of mutual reducibility, the orbit closure of the determinant has a finite quotient whereas that of the permanent is not finite.
Both of these applications use the fact that a finitely generated matrix semigroup can have a finite quotient under a congruence, while its Zariski closure has infinitely many classes under the same congruence. These applications connect semigroup theory with border complexity and provide a new toolkit for studying a broader range of algebraic computational models.
The central tool in our proof is the notion of factorisation forest for matrix semigroups, which we developed in recent works [11 , 12 ] . The original notion of factorization forest for semigroups was introduced by Imre Simon [13 ] and was later used in [14 ] to give an algorithm for deciding finiteness of matrix semigroups with entries in the tropical semiring. The notion has become central in the algebraic theory of formal languages (see, for example, [15 , 16 ] ) and has played a significant role in the resolution of a number of open problems (e.g., the proof of Courcelle’s conjecture that for a class of graphs of bounded treewidth a property is MSO-definable if and only if it is recognisable by a tree automaton [17 ] ). In particular, factorisation forests provide a means to translate results on groups to the setting of semigroups. Our variants generalise Simon’s factorisation from finite semigroups to (infinite) matrix semigroups [11 ] and matrix semigroup quotients. In our factorisation trees, the role of idempotent elements in the classical version is replaced by (rank-)stable matrices and generalisations. We recently used one of these variants in the context of finite matrix semigroups, to give a polynomial-space algorithm for the finiteness problem for finitely generated matrix semigroups, improving a previous NExp-upper bound. This was a corollary of our polynomial bit-size bound on matrix entries in finite monoids [12 ] .
Our Contribution
In this section we state our main technical contributions, which concern congruences on finitely generated semigroups ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Each result assumes a uniform bound N N N on the number of equivalence classes of a certain family of group-like sub-semigroups, and infers a bound on the number of classes of the ambient semigroup. The bound depends on N N N and the ambient dimension d d d , with the first result additionally depending on the cardinality of the set X X X of generators. There are three variants, according to the kind of sub-semigroups on which the uniform bound is applied. Theorem 3 considers subgroups of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ , Theorem 1 considers stable sub-semigroups of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ (that is, sub-semigroups in which all matrices share the same kernel and the same image that moreover have intersection zero), finally Theorem 2 considers rank-stable sub-semigroups of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ , that is, sub-semigroups in which all matrices have the same rank. Every subgroup of a matrix semigroup is a stable sub-semigroup and a stable semigroup of rank r r r is isomorphic to a sub-semigroup of the general linear group in dimension r r r . We state a Burnside theorem for quotients of matrix semigroups by closed congruences in Corollary 1.
For positive integers N , d , m N,d,m N , d , m , define
F d ( N , m ) : = [ m ( N + 1 ) 2 ] ( 2 d + 3 + 1 ) d d − 1 F_d(N,m) := \left[m(N+1)^2\right]^{(2^{d+3}+1)d^{d-1}} F d ( N , m ) := [ m ( N + 1 ) 2 ] ( 2 d + 3 + 1 ) d d − 1
This yields the asymptotic bound ( m ( N + 1 ) ) 2 O ( d 2 ) (m(N+1))^{2^{O(d^2)}} ( m ( N + 1 ) ) 2 O ( d 2 ) , which is polynomial in m m m and N N N for fixed d d d .
Theorem 1. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite , and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is finite if and only if there is an N ≥ 1 N \ge1 N ≥ 1 such that ∣ S / ∼ ∣ ≤ N |S/\sim| \le N ∣ S / ∼ ∣ ≤ N for every stable subsemigroup S ≤ ⟨ X ⟩ S \le\langle X\rangle S ≤ ⟨ X ⟩ . The cardinality and the diameter of the quotient are at most F d ( N , ∣ X ∣ ) F_d(N,|X|) F d ( N , ∣ X ∣ ) .
:::`
We say that a congruence ∼ \sim ∼ defined on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ is relatively Zariski closed in ⟨ X ⟩ × ⟨ X ⟩ \langle X\rangle\times\langle X\rangle ⟨ X ⟩ × ⟨ X ⟩ when there exists a Zariski-closed set R ⊆ Mat d ( Q ) × Mat d ( Q ) R \subseteq\operatorname{Mat}_{d}(\mathbb{Q}) \times\operatorname{Mat}_{d}(\mathbb{Q}) R ⊆ Mat d ( Q ) × Mat d ( Q ) such that, for all a , b ∈ ⟨ X ⟩ a,b \in\langle X\rangle a , b ∈ ⟨ X ⟩ , a ∼ b a \sim b a ∼ b if and only if ( a , b ) ∈ R (a,b) \in R ( a , b ) ∈ R . The following finiteness result for torsion quotients follows from Theorem 1. This gives a Burnside theorem for quotients of matrix semigroups by closed congruences.
Corollary 1. Let X ⊂ Mat d ( Q ) X \subset\operatorname{Mat}_{d}(\mathbb{Q}) X ⊂ Mat d ( Q ) be finite, and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Suppose that the relation ∼ \sim ∼ is relatively Zariski closed in ⟨ X ⟩ × ⟨ X ⟩ \langle X\rangle\times\langle X\rangle ⟨ X ⟩ × ⟨ X ⟩ . If ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is torsion, then it is finite.
For positive integers N , d N,d N , d , define
G d ( N ) : = ( N + 1 ) ( H + 1 ) d − 1 ≤ min { 2 O ( d 2 N ) , ( N + 1 ) 2 O ( d ) } . G_{d}(N) := (N+1)(H+1)^{d}-1 \le\min\left\{2^{O(d^{2}N)}, (N+1)^{2^{O(d)}}\right\}. G d ( N ) := ( N + 1 ) ( H + 1 ) d − 1 ≤ min { 2 O ( d 2 N ) , ( N + 1 ) 2 O ( d ) } .
where
H : = ( N + D 2 + 1 D 2 ) and D : = ( d ⌊ d / 2 ⌋ ) . H := \binom{N+D^{2}+1}{D^{2}} \qquad\text{and} \qquad D := \binom{d}{\lfloor d/2\rfloor}. H := ( D 2 N + D 2 + 1 ) and D := ( ⌊ d /2 ⌋ d ) .
Theorem 2. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_{d}(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is finite if and only if there is an N ≥ 1 N \ge1 N ≥ 1 such that ∣ S / ∼ ∣ ≤ N |S/\sim| \le N ∣ S / ∼ ∣ ≤ N for every rank-stable subsemigroup S ≤ ⟨ X ⟩ S \le\langle X\rangle S ≤ ⟨ X ⟩ . The diameter of the quotient is at most G d ( N ) G_{d}(N) G d ( N ) .
The uniform boundedness assumption in Theorem 2 is stronger than in Theorem 1, since it applies to all rank-stable semigroups, not just stable semigroups. Correspondingly, the theorem yields a better bound on the diameter of ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ : while the bound in Theorem 1 is doubly exponential in d 2 d^{2} d 2 and depends on the number ∣ X ∣ |X| ∣ X ∣ of generators, the bound in Theorem 2 is doubly exponential only in d d d and is independent of the number of generators.
The next result, Theorem 3, obtains a diameter bound on ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ that is only singly exponential in d d d , but assumes that the congruence ∼ \sim ∼ is defined on the Zariski closure of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ , with a uniform bound on the number of classes of this closure met by each subgroup.
Theorem 3. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_{d}(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Assume that there exists N ≥ 1 N \ge1 N ≥ 1 such that ∣ G / ∼ ∣ ≤ N |G/\sim| \le N ∣ G / ∼ ∣ ≤ N for every group G ≤ ⟨ X ⟩ G \le\langle X\rangle G ≤ ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is finite and has diameter at most 2 2 d ( d + 3 ) N d 2^{2d(d+3)}N^{d} 2 2 d ( d + 3 ) N d .
We now apply the viewpoint of matrix semigroup congruences and quotients to questions in border complexity, focusing on two lines of applications. In each application, we construct a finitely generated matrix semigroup and a congruence on its Zariski closure. We show that any group in the closure lies in a single congruence class, so Theorem 3 applies and gives a finite quotient of the generated semigroup. We then show that the closure has infinitely many classes, and relate this distinction or the congruence classes to the polynomial families (or their orbits) under study.
Separation of homogeneous (commutative) ABPs and their closures. Let w ≥ 2 w \ge2 w ≥ 2 and d ≥ 3 d \ge3 d ≥ 3 . Recall that a homogeneous algebraic branching program (ABP) of width at most w w w and degree d d d computes a polynomial through matrix multiplication, that is,
f ( x ) = L 1 ( x ) ⋯ L d ( x ) , f(\boldsymbol{x}) = L_{1}(\boldsymbol{x}) \cdots L_{d}(\boldsymbol{x}), f ( x ) = L 1 ( x ) ⋯ L d ( x ) ,
where L 1 L_{1} L 1 is a 1 × w 1 \times w 1 × w matrix, L d L_{d} L d is a w × 1 w \times1 w × 1 matrix, and the remaining matrices are w × w w \times w w × w . The entries of the matrices L i L_{i} L i are homogeneous linear forms in the input variables x = ( x 1 , … , x m ) \boldsymbol{x}=(x_{1},\ldots,x_{m}) x = ( x 1 , … , x m ) , with coefficients in Q \mathbb{Q} Q . ABPs and their border complexity have been studied in recent years [10 , 18 , 19 , 20 ] .
Homogeneous models and their connections with algebraic geometry have also received recent attention [21 , 22 ] . In our constructions, we keep both the width w w w and the degree d d d fixed and use homogeneous linear forms as used in [22 ] .
Write Q ‾ [ x ] d \overline{\mathbb{Q}}[x]_d Q [ x ] d for the vector space of homogeneous degree-d d d polynomials in the variables x x x , with coefficients in Q ‾ \overline{\mathbb{Q}} Q . Denote by A w , d \mathcal{A}_{w,d} A w , d the set of polynomials Q ‾ [ x ] d \overline{\mathbb{Q}}[x]_d Q [ x ] d computed by the ABPs defined above. As usual, viewing each polynomial as its vector of coefficients, we take the Zariski closure of the set A w , d \mathcal{A}_{w,d} A w , d of polynomials in the vector space of homogeneous degree-d d d polynomials over Q ‾ \overline{\mathbb{Q}} Q ; we denote this closure by A ‾ w , d \overline{\mathcal{A}}_{w,d} A w , d .
Below, we prove the separation of A w , d \mathcal{A}_{w,d} A w , d and its closure through a semigroup congruence. Roughly speaking, we use translation of the degree-d d d and width-w w w universal ABP to encode polynomials (in A w , d \mathcal{A}_{w,d} A w , d ) in a finitely generated matrix semigroup. We construct a congruence on its closure whose zero class consists exactly of the encodings of polynomials in A w , d \mathcal{A}_{w,d} A w , d . We give an explicit polynomial F w , d F_{w,d} F w , d with integer coefficients whose encoding is not congruent with zero, implying that F w , d ∈ A ‾ w , d ∖ A w , d F_{w,d} \in\overline{\mathcal{A}}_{w,d} \setminus\mathcal{A}_{w,d} F w , d ∈ A w , d ∖ A w , d .
Theorem 4. Let w ≥ 2 w \ge2 w ≥ 2 and d ≥ 3 d \ge3 d ≥ 3 and m ≥ 2 w + 7 m \ge2w+7 m ≥ 2 w + 7 . The class A w , d \mathcal{A}_{w,d} A w , d of width-w w w degree-d d d homogenous ABPs is separated from A ‾ w , d \overline{\mathcal{A}}_{w,d} A w , d .
Separation of Det n \operatorname{Det}_n Det n from Perm n \operatorname{Perm}_n Perm n under a subclass of reductions
Let n ≥ 3 n \ge3 n ≥ 3 . Consider the matrix [ x i , j ] 1 ≤ i , j ≤ n [x_{i,j}]_{1\le i,j\le n} [ x i , j ] 1 ≤ i , j ≤ n whose entries are the formal variables x i , j x_{i,j} x i , j , and X : = ( x 1 , 1 , … , x n , n ) X := (x_{1,1},\ldots,x_{n,n}) X := ( x 1 , 1 , … , x n , n ) be the vector of these n 2 n^2 n 2 variables. We denote by Det n \operatorname{Det}_n Det n and Perm n \operatorname{Perm}_n Perm n the determinant and permanent polynomials, respectively, defined by
Det n ( X ) : = ∑ σ ∈ S n sgn ( σ ) ∏ i = 1 n x i , σ ( i ) and Perm n ( X ) : = ∑ σ ∈ S n ∏ i = 1 n x i , σ ( i ) , \operatorname{Det}_n(X) := \sum_{\sigma\in S_n}\operatorname{sgn}(\sigma)\prod_{i=1}^{n}x_{i,\sigma(i)}
\quad\text{and}\quad
\operatorname{Perm}_n(X) := \sum_{\sigma\in S_n}\prod_{i=1}^{n}x_{i,\sigma(i)}, Det n ( X ) := σ ∈ S n ∑ sgn ( σ ) i = 1 ∏ n x i , σ ( i ) and Perm n ( X ) := σ ∈ S n ∑ i = 1 ∏ n x i , σ ( i ) ,
where S n S_n S n is the set of permutations of { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } and sgn ( σ ) \operatorname{sgn}(\sigma) sgn ( σ ) denotes the sign of σ \sigma σ .
Below, we use the standard action of Mat n 2 ( Q ‾ ) \operatorname{Mat}_{n^2}(\overline{\mathbb{Q}}) Mat n 2 ( Q ) on the polynomial ring Q ‾ [ X ] \overline{\mathbb{Q}}[X] Q [ X ] , defined by
( h ◃ L ) ( x ) : = h ( x L ) , (h \triangleleft L)(x) := h(xL), ( h ◃ L ) ( x ) := h ( xL ) ,
where L ∈ Mat n 2 ( Q ‾ ) L \in\operatorname{Mat}_{n^2}(\overline{\mathbb{Q}}) L ∈ Mat n 2 ( Q ) and h ∈ Q ‾ [ X ] h \in\overline{\mathbb{Q}}[X] h ∈ Q [ X ] . This is the action of linear substitutions in the n 2 n^2 n 2 variables; as defined in [8 Section 3.2] .
We construct a finitely generated monoid ⟨ Y n ⟩ ≤ Mat n 2 ( Q ‾ ) \langle Y_n\rangle\le\operatorname{Mat}_{n^2}(\overline{\mathbb{Q}}) ⟨ Y n ⟩ ≤ Mat n 2 ( Q ) such that
Perm ‾ n ◃ ⟨ Y n ⟩ ⊈ Det ‾ n ◃ ⟨ Y n ⟩ . \overline{\operatorname{Perm}}_n \triangleleft\langle Y_n\rangle\not\subseteq\overline{\operatorname{Det}}_n \triangleleft\langle Y_n\rangle. Perm n ◃ ⟨ Y n ⟩ ⊆ Det n ◃ ⟨ Y n ⟩ .
Specifically, these orbit closures behave differently under the mutual reducibility congruence ≈ \approx ≈ . We write f ≈ g f \approx g f ≈ g if there exist L , M ∈ ⟨ Y n ⟩ L,M \in\langle Y_n\rangle L , M ∈ ⟨ Y n ⟩ such that f = g ◃ L f = g \triangleleft L f = g ◃ L and g = f ◃ M g = f \triangleleft M g = f ◃ M . Using Theorem 3, we show that both generated orbits have finite quotients under ≈ \approx ≈ . However, the orbit closure of the determinant has a finite quotient, whereas that of the permanent has an infinite quotient.
All proofs and details omitted from the main text are provided in the appendix.
Preliminaries
Write N : = { 0 , 1 , 2 , … } \mathbb{N} := \{0,1,2,\ldots\} N := { 0 , 1 , 2 , … } , Z \mathbb{Z} Z , and Q \mathbb{Q} Q for the sets of natural numbers, integers, and rational numbers, respectively. We denote by Q ‾ \overline{\mathbb{Q}} Q the field of algebraic numbers, that is, the roots of nonzero polynomials with rational coefficients.
Let S S S be a semigroup. Given s ∈ S s \in S s ∈ S , the subsemigroup generated by s s s is ⟨ s ⟩ : = { s , s 2 , s 3 , … } \langle s\rangle:= \{s, s^{2}, s^{3}, \ldots\} ⟨ s ⟩ := { s , s 2 , s 3 , … } . We call s s s torsion if ⟨ s ⟩ \langle s\rangle ⟨ s ⟩ is finite. In this case, the idempotent order (or simply order) of s s s is the least positive integer n n n such that s n = s 2 n s^{n}=s^{2n} s n = s 2 n , or equivalently, such that s n s^{n} s n is idempotent. The semigroup S S S is called torsion if all its elements are torsion, and uniformly torsion of order n n n if s n = s 2 n s^{n}=s^{2n} s n = s 2 n holds for all s ∈ S s \in S s ∈ S .
Matrix semigroups and their congruences. Given a positive integer d d d and a field F F F , write Mat d ( F ) \operatorname{Mat}_{d}(F) Mat d ( F ) for the set of d × d d \times d d × d matrices with entries in F F F , and GL d ( F ) \operatorname{GL}_{d}(F) GL d ( F ) for the group of invertible matrices. Given a ∈ Mat d ( F ) a \in\operatorname{Mat}_{d}(F) a ∈ Mat d ( F ) , write ker ( a ) \operatorname{ker}(a) ker ( a ) and im ( a ) \operatorname{im}(a) im ( a ) for the kernel and image of the linear transformation x ↦ x a x \mapsto xa x ↦ x a on F d F^{d} F d , respectively. We call a a a rank-stable if rank ( a ) = rank ( a 2 ) \operatorname{rank}(a)=\operatorname{rank}(a^{2}) rank ( a ) = rank ( a 2 ) . A set X ⊆ Mat d ( F ) X \subseteq\operatorname{Mat}_{d}(F) X ⊆ Mat d ( F ) is called rank-stable if rank ( a ) = rank ( b ) = rank ( a b ) \operatorname{rank}(a)=\operatorname{rank}(b)=\operatorname{rank}(ab) rank ( a ) = rank ( b ) = rank ( ab ) for all a , b ∈ X a,b \in X a , b ∈ X . A rank-stable set is called stable if all its elements have the same kernel and the same image.
Example 1. Let
A = ( 0 1 0 0 ) , A x = ( 1 x 0 0 ) . A=\begin{pmatrix}0 & 1 \\ 0 & 0\end{pmatrix}, \qquad A_{x}=\begin{pmatrix}1 & x \\ 0 & 0\end{pmatrix}. A = ( 0 0 1 0 ) , A x = ( 1 0 x 0 ) . For x ≠ y x \ne y x = y , the set { A x , A y } \{A_{x}, A_{y}\} { A x , A y } is rank-stable but not stable, as the two matrices have the same kernel but different images. The matrix A A A is nilpotent, since A 2 = 0 A^{2}=0 A 2 = 0 . Although singleton sets trivially satisfy the common kernel and image condition, { A } \{A\} { A } is not stable because A A A is not rank-stable. ♣
Given a finite set X ⊆ Mat d ( F ) X \subseteq\operatorname{Mat}_{d}(F) X ⊆ Mat d ( F ) , write ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ for the multiplicative semigroup of all nonempty products of matrices from X X X . The length of an element a ∈ ⟨ X ⟩ a \in\langle X\rangle a ∈ ⟨ X ⟩ , with respect to X X X , is the least positive integer k k k such that a = a 1 ⋯ a k a=a_{1}\cdots a_{k} a = a 1 ⋯ a k with a i ∈ X a_{i} \in X a i ∈ X ; when the generating set is clear, we simply refer to this quantity as the length of a a a . The diameter of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ with respect to X X X is the least n ∈ N n \in\mathbb{N} n ∈ N such that every element has length at most n n n , and is infinite if no such n n n exists. By the extension of the Burnside–Schur theorem to semigroups [4 ] , every finitely generated uniformly torsion subsemigroup of Mat d ( F ) \operatorname{Mat}_{d}(F) Mat d ( F ) is finite. In particular, its diameter with respect to any finite generating set is finite.
A congruence on a semigroup is an equivalence relation compatible with multiplication, that is, a 1 ∼ b 1 a_{1} \sim b_{1} a 1 ∼ b 1 and a 2 ∼ b 2 a_{2} \sim b_{2} a 2 ∼ b 2 implies a 1 a 2 ∼ b 1 b 2 a_{1}a_{2} \sim b_{1}b_{2} a 1 a 2 ∼ b 1 b 2 . If ∼ \sim ∼ is a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ , its quotient is a semigroup generated by X / ∼ X/{\sim} X / ∼ . The length of a congruence class is the smallest length of an element in that class: the class of a a a has length equal to the least positive integer k k k such that a ∼ a 1 ⋯ a k a \sim a_{1}\cdots a_{k} a ∼ a 1 ⋯ a k with a i ∈ X a_{i} \in X a i ∈ X . The diameter of ⟨ X ⟩ / ∼ \langle X\rangle/{\sim} ⟨ X ⟩ / ∼ with respect to X / ∼ X/{\sim} X / ∼ is the least n ∈ N n \in\mathbb{N} n ∈ N such that every class has length at most n n n , and is infinite if no such n n n exists. For a subsemigroup S ≤ ⟨ X ⟩ S \leq\langle X\rangle S ≤ ⟨ X ⟩ , we write S / ∼ S/{\sim} S / ∼ for the quotient by the restriction of ∼ \sim ∼ to S S S ; its size counts the classes of ⟨ X ⟩ / ∼ \langle X\rangle/{\sim} ⟨ X ⟩ / ∼ that meet S S S .
Zariski Topology. We recall basic notions from algebraic geometry, and refer the reader to [23 , 24 ] for an extended background.
An affine algebraic variety (or simply a variety) in Q ‾ m \overline{\mathbb{Q}}^{m} Q m is the common zero set of a collection of polynomials in Q [ x 1 , … , x m ] \mathbb{Q}[x_{1},\ldots,x_{m}] Q [ x 1 , … , x m ] . The Zariski topology on Q m \mathbb{Q}^{m} Q m is the topology whose closed sets are these varieties. Given a subset S ⊆ Q ‾ m S \subseteq\overline{\mathbb{Q}}^{m} S ⊆ Q m , its Zariski closure, denoted by S ‾ \overline{S} S , is the smallest Zariski-closed set containing S S S .
We recall several facts following standard properties [23 ] . Finite sets and vector spaces are Zariski closed, and the set Z m \mathbb{Z}^{m} Z m is Zariski dense in Q m \mathbb{Q}^{m} Q m , meaning that Z m ‾ = Q m ‾ \overline{\mathbb{Z}^{m}}=\overline{\mathbb{Q}^{m}} Z m = Q m . Moreover, closure commutes with finite unions. Polynomial maps are continuous in this topology, meaning that,
φ ( S ‾ ) ⊆ φ ( S ) ‾ \varphi(\overline{S}) \subseteq\overline{\varphi(S)} φ ( S ) ⊆ φ ( S )
for every polynomial map φ : Q ‾ m → Q ‾ n \varphi: \overline{\mathbb{Q}}^m \to\overline{\mathbb{Q}}^n φ : Q m → Q n and S ⊆ Q ‾ m S \subseteq\overline{\mathbb{Q}}^m S ⊆ Q m (the equality holds when φ \varphi φ is an injective linear map). By continuity of matrix multiplication, the Zariski closure of a matrix semigroup is again a semigroup. Moreover, the closure of a nonempty subsemigroup (inside a linear algebraic group) is a group [11 Fact 9] .
Burnside Problems for Matrix Semigroup Quotients
The classical Burnside problem asks whether finitely generated torsion groups must be finite. The answer is negative for arbitrary groups, even with bounded exponent [1 ] , but positive for linear groups in characteristic zero by Schur’s theorem [25 ] . McNaughton and Zalcstein extended the latter conclusion to matrix semigroups [4 ] . We are interested in such Burnside questions concerning quotients of matrix semigroups by congruences, which in general may not themselves be linear.
Throughout this section, fix a finite nonempty set X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) with size ∣ X ∣ |X| ∣ X ∣ . We measure lengths of matrices with respect to X X X , and lengths of congruence classes with respect to X / ∼ X/{\sim} X / ∼ . Our first two theorems concern congruences on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ , whereas for the last theorem, the congruence is defined on its Zariski closure in Mat d ( Q ) \operatorname{Mat}_d(\mathbb{Q}) Mat d ( Q ) .
Factorisation of matrix products. For a matrix product a 1 ⋯ a k a_1 \cdots a_k a 1 ⋯ a k , a factor is a consecutive subproduct a i ⋯ a j a_i \cdots a_j a i ⋯ a j with $1 \leq i \leq j \leq k$; it is proper if it is not the full product. Given such a matrix product a 1 ⋯ a k a_1 \cdots a_k a 1 ⋯ a k , a factorisation tree groups its factors recursively. The leaves of the tree are labelled with the matrices a 1 a_1 a 1 to a k a_k a k in the exact order from left to right, whereas each internal node of the tree is labelled by the product of its children’s labels. The height of the tree is the length of the longest path from the root to a leaf.
The original notion of factorization tree was introduced by Imre Simon [13 ] in his resolution of the star-height problem for regular languages (see also [15 ] ). Simon’s original factorisation tree establishes a bound on the height of factorisation trees over finite monoids, which depends on the size of the monoid. In a recent work [11 ] , we adapt this notion from finite monoids to the infinite monoid Mat d ( Q ) \operatorname{Mat}_d(\mathbb{Q}) Mat d ( Q ) . In our variant, the role of idempotent elements in the classical version is replaced by rank-stable matrices². We have proved in particular that every matrix product in Mat d ( Q ) \operatorname{Mat}_d(\mathbb{Q}) Mat d ( Q ) , no matter its length, has a factorization tree of height bounded only by d ( d + 3 ) d(d+3) d ( d + 3 ) . We then refine this notion slightly, in [12 ] , for finite matrix monoids in Mat d ( Q ) \operatorname{Mat}_d(\mathbb{Q}) Mat d ( Q ) , replacing rank-stability with stability of the children (see more details below). We combined the height bound on such factorisation trees, together with Artin–Wedderburn decomposition, to obtain the first singly exponential bound on the cardinality of finite rational matrix monoids, in terms of the total encoding size of their generators.
Uniform bound on stable subsemigroups
For positive integers N , d , m N,d,m N , d , m , define
F d ( N , m ) : = [ m ( N + 1 ) 2 ] ( 2 d 2 + 3 ) d d − 1 (1) F_d(N,m) := \left[m(N+1)^2\right]^{(2d^2+3)d^{d-1}} \tag*{(1)} F d ( N , m ) := [ m ( N + 1 ) 2 ] ( 2 d 2 + 3 ) d d − 1 (1)
This yields the asymptotic bound ( m ( N + 1 ) ) 2 O ( d 2 ) (m(N+1))^{2^{O(d^2)}} ( m ( N + 1 ) ) 2 O ( d 2 ) , which is polynomial in m m m and N N N for fixed d d d .
Theorem 1. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite, and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/{\sim} ⟨ X ⟩ / ∼ is finite if and only if there is an N ≥ 1 N \geq1 N ≥ 1 such that ∣ S / ∼ ∣ ≤ N |S/{\sim}| \leq N ∣ S / ∼ ∣ ≤ N for every stable subsemigroup S ⊆ ⟨ X ⟩ S \subseteq\langle X\rangle S ⊆ ⟨ X ⟩ . The cardinality and the diameter of the quotient are at most F d ( N , ∣ X ∣ ) F_d(N,|X|) F d ( N , ∣ X ∣ ) .
Before explaining the proof, we give a family in dimension three showing that the diameter must depend on both N N N and ∣ X ∣ |X| ∣ X ∣ .
Example 2. Let N , m ≥ 2 N,m\ge2 N , m ≥ 2 with B : = m 2 + 1 B:=m^2+1 B := m 2 + 1 . Let X m = { A 1 , … , A m } ⊆ Mat 3 ( Q ) X_m=\{A_1,\ldots,A_m\}\subseteq\operatorname{Mat}_3(\mathbb{Q}) X m = { A 1 , … , A m } ⊆ Mat 3 ( Q ) where
A i = ( B 0 B i 0 1 0 0 i 0 ) . A_i=\begin{pmatrix} B & 0 & Bi \\ 0 & 1 & 0 \\ 0 & i & 0 \end{pmatrix}. A i = B 0 0 0 1 i B i 0 0 . We define the congruence ∼ \sim ∼ by the following matrix product
W = ( A 1 ⋯ A m ) N = ( A 1 ⋯ A m ) ⋯ ( A 1 ⋯ A m ) ⏟ N repetitions . W=(A_1\cdots A_m)^N=\underbrace{(A_1\cdots A_m)\cdots(A_1\cdots A_m)}_{N\ \text{repetitions}}. W = ( A 1 ⋯ A m ) N = N repetitions ( A 1 ⋯ A m ) ⋯ ( A 1 ⋯ A m ) . Let each distinct matrix appearing as a factor of W W W be in a separate singleton class, and group every other product in ⟨ X m ⟩ \langle X_m\rangle ⟨ X m ⟩ in a single class (we refer to it as the large class).
The crux of the construction is that the A i A_i A i freely generate the semigroup ⟨ X m ⟩ \langle X_m\rangle ⟨ X m ⟩ , in other words, every matrix a ∈ ⟨ X m ⟩ a\in\langle X_m\rangle a ∈ ⟨ X m ⟩ has a unique matrix product a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k with the a i a_i a i 's in X m X_m X m (the uniqueness comes from the fact that the product encodes a number in the base B B B ). The kernel and image of a product comes from a 1 ∈ X m a_1\in X_m a 1 ∈ X m and a k ∈ X m a_k\in X_m a k ∈ X m , respectively. Intuitively speaking, by the above, every stable subsemigroup has ker ( A i ) \operatorname{ker}(A_i) ker ( A i ) and im ( A j ) \operatorname{im}(A_j) im ( A j ) for some pair ( i , j ) (i,j) ( i , j ) with 1 ≤ i , j ≤ m 1\le i,j\le m 1 ≤ i , j ≤ m . Notice that there are at most N N N distinct factors of W W W starting with A i A_i A i and ending with A j A_j A j . This implies that the stable subsemigroup quotient with ∼ \sim ∼ meets at most N N N singleton classes and possibly the large class, giving the bound N + 1 N+1 N + 1 . Moreover, the stable subsemigroup generated by the pair ( 1 , m ) (1,m) ( 1 , m ) will meet this exact bound. The diameter bound N ∣ X m ∣ N|X_m| N ∣ X m ∣ follows from the freeness of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ and the length of the singleton class of W W W . The details are provided in Appendix G. ♣
Proof Sketch of Theorem 1. We outline the main ideas of the proof here; the full details, including proofs of the claims, are provided in Appendix B. We focus on the non-trivial direction of the proof, obtaining the stated bound on ∣ ⟨ X ⟩ / ∼ ∣ |\langle X\rangle/\sim| ∣ ⟨ X ⟩ / ∼ ∣ from the uniform bound on N N N on ∣ S / ∼ ∣ |S/\sim| ∣ S / ∼ ∣ for stable S ≤ ⟨ X ⟩ S\le\langle X\rangle S ≤ ⟨ X ⟩ .
The proof of the theorem relies on the notion of forest factorisation from [11 ] . Since we will also use a second variant used in [12 ] , we refer to this one as the rank-stable forest factorisation . Given a matrix product, in its rank-stable forest factorisation every node with more than two children is rank-stable. It is shown in [11 Proposition 3] that every matrix product a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k in Mat d ( Q ) \operatorname{Mat}_d(\mathbb{Q}) Mat d ( Q ) has a rank-stable factorization tree of height at most d ( d + 3 ) d(d+3) d ( d + 3 ) . To obtain this bound, we have shown in [11 Proposition 7] that if rank ( a 1 ) = ⋯ = rank ( a k ) = rank ( a 1 ⋯ a k ) \operatorname{rank}(a_1)=\cdots=\operatorname{rank}(a_k)=\operatorname{rank}(a_1\cdots a_k) rank ( a 1 ) = ⋯ = rank ( a k ) = rank ( a 1 ⋯ a k ) holds, such a tree can be chosen with height at most d + 2 d+2 d + 2 . The following claim follows from a truncation procedure on such trees, see Figure 1 .
Figure 1. In the factorisation tree of matrix product, along each path from the root to a leaf, we truncate the tree at the first nodes with more than two children. Here, the factors P 1 P_1 P 1 and P 2 P_2 P 2 are rank-stable.
Claim 3.1. Let a 1 , … , a k ∈ Mat d ( Q ) a_1,\ldots,a_k\in\operatorname{Mat}_d(\mathbb{Q}) a 1 , … , a k ∈ Mat d ( Q ) be such that rank ( a 1 ⋯ a k ) = rank ( a i ) \operatorname{rank}(a_1\cdots a_k)=\operatorname{rank}(a_i) rank ( a 1 ⋯ a k ) = rank ( a i ) for all 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k . Then a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k can be partitioned into at most 2 d + 2 2^{d+2} 2 d + 2 factors, each of which is either a single matrix among the a i a_i a i 's or a rank-stable matrix product.
To prove finiteness of the quotient, we bound its number of classes by decreasing induction on rank. Let T ⊆ ⟨ X ⟩ T\subseteq\langle X\rangle T ⊆ ⟨ X ⟩ be finite and all its matrices have rank r r r . The essential step is to count the number of congruence classes of matrix products over T T T , whose resulting product has rank r r r . Such rank-preserving products keep the kernel of the first matrix and the image of the last matrix, so at most ∣ T ∣ 2 |T|^2 ∣ T ∣ 2 pairs ( K , U ) (K,U) ( K , U ) of kernel and image can occur. Observe that, whenever K ∩ U = { 0 } K\cap U=\{0\} K ∩ U = { 0 } , the matrices in ⟨ T ⟩ \langle T\rangle ⟨ T ⟩ with kernel K K K and image U U U form a stable subsemigroup, whose quotient by assumption has at most N N N classes. Applying Claim 3.1 now gives the following bound.
Claim 3.2. Let T ⊆ ⟨ X ⟩ T \subseteq\langle X\rangle T ⊆ ⟨ X ⟩ be finite and all its matrices have rank r r r . The rank-r r r matrices in ⟨ T ⟩ \langle T\rangle ⟨ T ⟩ meet at most ( 1 + ∣ T ∣ + N ∣ T ∣ 2 ) 2 d + 2 (1 + |T| + N|T|^2)^{2d+2} ( 1 + ∣ T ∣ + N ∣ T ∣ 2 ) 2 d + 2 congruence classes.
We note in passing that congruent matrices can have different ranks. For each class c c c in the congruence, we define ρ ( c ) : = min { rank ( a ) : a ∈ c } \rho(c) := \min\{\operatorname{rank}(a) : a \in c\} ρ ( c ) := min { rank ( a ) : a ∈ c } . We count classes with ρ ( c ) ≥ r \rho(c) \ge r ρ ( c ) ≥ r by decreasing induction on r r r . For the base case, observe that the invertible matrices in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ form a stable subsemigroup whose quotient has at most N N N classes. Hence there are at most N N N classes with ρ ( c ) = d \rho(c) = d ρ ( c ) = d .
Suppose there are at most n n n classes c c c with ρ ( c ) > r \rho(c) > r ρ ( c ) > r . Let a 1 ⋯ a k a_1 \cdots a_k a 1 ⋯ a k with a i ∈ X a_i \in X a i ∈ X be a matrix product, whose class [ a 1 ⋯ a k ] ∼ [a_1 \cdots a_k]_{\sim} [ a 1 ⋯ a k ] ∼ has minimum rank r r r . Recall that matrix multiplication cannot increase rank, so ρ ( b c ) ≤ min { ρ ( b ) , ρ ( c ) } \rho(bc) \le\min\{\rho(b), \rho(c)\} ρ ( b c ) ≤ min { ρ ( b ) , ρ ( c )} for all congruence classes b , c b, c b , c . This implies that every factor of a 1 ⋯ a k a_1 \cdots a_k a 1 ⋯ a k belongs to a class with minimum rank at least r r r . Write a 1 ⋯ a k a_1 \cdots a_k a 1 ⋯ a k as y 1 c 1 ⋯ y ℓ c ℓ y ℓ + 1 y_1c_1 \cdots y_\ell c_\ell y_{\ell+1} y 1 c 1 ⋯ y ℓ c ℓ y ℓ + 1 with each c i ∈ X c_i \in X c i ∈ X and ρ ( [ y i c i ] ∼ ) = r \rho([y_i c_i]_{\sim}) = r ρ ([ y i c i ] ∼ ) = r , and the y i y_i y i are either an empty product or ρ ( [ y i ] ∼ ) > r \rho([y_i]_{\sim}) > r ρ ([ y i ] ∼ ) > r . See Figure 2 for an illustration.
Figure 2. The factor y i c i y_i c_i y i c i can be replaced by a rank-r r r congruent matrix b i b_i b i .
Since factors y i c i y_i c_i y i c i belong to a class of the form [ c i ] ∼ [c_i]_{\sim} [ c i ] ∼ or [ y i ] ∼ [ c i ] ∼ [y_i]_{\sim}[c_i]_{\sim} [ y i ] ∼ [ c i ] ∼ , where c i ∈ X c_i \in X c i ∈ X and ρ ( [ y i ] ∼ ) > r \rho([y_i]_{\sim}) > r ρ ([ y i ] ∼ ) > r , there are at most t : = ∣ X ∣ ( n + 1 ) t := |X|(n + 1) t := ∣ X ∣ ( n + 1 ) such classes. We choose one matrix product b i b_i b i of rank r r r in each of these classes [ y i c i ] ∼ [y_i c_i]_{\sim} [ y i c i ] ∼ to form a finite set T ⊆ ⟨ X ⟩ T \subseteq\langle X\rangle T ⊆ ⟨ X ⟩ where ∣ T ∣ ≤ t |T| \le t ∣ T ∣ ≤ t . Then
a 1 ⋯ a k ∼ b 1 ⋯ b ℓ y ℓ + 1 . a_1 \cdots a_k \sim b_1 \cdots b_\ell y_{\ell+1}. a 1 ⋯ a k ∼ b 1 ⋯ b ℓ y ℓ + 1 .
Since each b i b_i b i has rank r r r and ρ ( b 1 ⋯ b ℓ y ℓ + 1 ) = r \rho(b_1 \cdots b_\ell y_{\ell+1}) = r ρ ( b 1 ⋯ b ℓ y ℓ + 1 ) = r we get r ≤ rank ( b 1 ⋯ b ℓ y ℓ + 1 ) ≤ rank ( b 1 ⋯ b ℓ ) ≤ r r \le\operatorname{rank}(b_1 \cdots b_\ell y_{\ell+1}) \le\operatorname{rank}(b_1 \cdots b_\ell) \le r r ≤ rank ( b 1 ⋯ b ℓ y ℓ + 1 ) ≤ rank ( b 1 ⋯ b ℓ ) ≤ r , which means that rank ( b 1 ⋯ b ℓ ) = r \operatorname{rank}(b_1 \cdots b_\ell) = r rank ( b 1 ⋯ b ℓ ) = r . Claim 3.2 now bounds the number of possible classes [ b 1 ⋯ b ℓ ] ∼ [b_1 \cdots b_\ell]_{\sim} [ b 1 ⋯ b ℓ ] ∼ by ( 1 + t + N t 2 ) 2 d + 2 (1 + t + Nt^2)^{2d+2} ( 1 + t + N t 2 ) 2 d + 2 . There are at most n + 1 n + 1 n + 1 choices for the suffix class, including the empty suffix. Hence the number of classes with minimum rank exactly r r r is at most
( n + 1 ) ( 1 + t + N t 2 ) 2 d + 2 , (n + 1)(1 + t + Nt^2)^{2d+2}, ( n + 1 ) ( 1 + t + N t 2 ) 2 d + 2 ,
@】【。 where t = ∣ X ∣ ( n + 1 ) t = |X|(n + 1) t = ∣ X ∣ ( n + 1 ) . Adding the at most n n n classes with larger minimum rank completes the induction step. By a summation over ranks, we arrive to the bound given in (1) for the cardinality of the quotient (and thus for its diameter as well).
Uniform bound on rank-stable subsemigroups.
The next theorem differs from Theorem 1 by requiring a uniform bound on rank-stable subsemigroups rather than stable subsemigroups. The next example shows that these uniform bounds can be different.
Example 3. Let m , d ≥ 2 m, d \ge2 m , d ≥ 2 . Let X = { A 1 , … , A m } ⊆ Mat d ( Q ) X = \{A_1,\ldots,A_m\} \subseteq\operatorname{Mat}_d(\mathbb{Q}) X = { A 1 , … , A m } ⊆ Mat d ( Q ) where
A i = ( 1 0 i 0 0 0 0 I d − 2 0 ) . A_i =
\begin{pmatrix}
1 & 0 & i \\
0 & 0 & 0 \\
0 & I_{d-2} & 0
\end{pmatrix}. A i = 1 0 0 0 0 I d − 2 i 0 0 . where I d − 2 I_{d-2} I d − 2 is the identity matrix in dimension d − 2 d-2 d − 2 . Let ∼ \sim ∼ be the equality congruence. Since ( x 1 , … , x d ) A i = ( x 1 , x 3 , … , x d , i x 1 ) (x_1,\ldots,x_d)A_i = (x_1,x_3,\ldots,x_d,ix_1) ( x 1 , … , x d ) A i = ( x 1 , x 3 , … , x d , i x 1 ) , every product of k k k generators has rank max { 1 , d − k } \max\{1,d-k\} max { 1 , d − k } . Hence, the rank-stable subsemigroups contain only rank one matrices. The rank-one matrices in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ , say obtained by products a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k with a i ∈ X a_i \in X a i ∈ X , are equal to their suffix factor of length d − 1 d-1 d − 1 , as shown below; given k ≥ d − 1 k \ge d-1 k ≥ d − 1 we have
a 1 ⋯ a k = a k − d + 2 ⋯ a k = A i 1 ⋯ A i d − 1 = ( 1 i 1 ⋯ i d − 1 0 0 ⋯ 0 ⋮ ⋮ ⋮ 0 0 ⋯ 0 ) . a_1\cdots a_k = a_{k-d+2}\cdots a_k = A_{i_1}\cdots A_{i_{d-1}} =
\begin{pmatrix}
1 & i_1 & \cdots& i_{d-1} \\
0 & 0 & \cdots& 0 \\
\vdots& \vdots& & \vdots\\
0 & 0 & \cdots& 0
\end{pmatrix}. a 1 ⋯ a k = a k − d + 2 ⋯ a k = A i 1 ⋯ A i d − 1 = 1 0 ⋮ 0 i 1 0 ⋮ 0 ⋯ ⋯ ⋯ i d − 1 0 ⋮ 0 . Observe that S = ⟨ a 1 ⋯ a d − 1 : a i ∈ X ⟩ S = \langle a_1\cdots a_{d-1} : a_i \in X\rangle S = ⟨ a 1 ⋯ a d − 1 : a i ∈ X ⟩ is a rank-stable subsemigroup of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ ; moreover, every rank-stable (and stable) subsemigroup is contained in S S S . Hence the bound on the cardinalities of quotients of rank-stable subsemigroups is ∣ X ∣ d − 1 |X|^{d-1} ∣ X ∣ d − 1 . Notice that the stable subsemigroups are all singletons, as each matrix in S S S is idempotent and has a distinct image:
im ( A i 1 ⋯ A i d − 1 ) = Span Q { ( 1 , i 1 , … , i d − 1 ) } . \operatorname{im}\left(A_{i_1}\cdots A_{i_{d-1}}\right) = \operatorname{Span}_{\mathbb{Q}}\{(1,i_1,\ldots,i_{d-1})\}. im ( A i 1 ⋯ A i d − 1 ) = Span Q {( 1 , i 1 , … , i d − 1 )} . This means that the smallest bound on the quotient images of stable subsemigroups in this example is 1. We provide more details in Appendix § 13 . ♣
For positive integers N , d N,d N , d , define
G d ( N ) : = ( N + 1 ) ( H + 1 ) d − 1 ≤ min { 2 O ( d 2 N ) , ( N + 1 ) 2 O ( d ) } . (2) G_d(N) := (N+1)(H+1)^d - 1 \le\min\left\{2^{O(d^2N)}, (N+1)^{2^{O(d)}}\right\}. \tag*{(2)} G d ( N ) := ( N + 1 ) ( H + 1 ) d − 1 ≤ min { 2 O ( d 2 N ) , ( N + 1 ) 2 O ( d ) } . (2)
where
H : = ( N + D 2 + 1 D 2 ) and D : = ( d ⌊ d / 2 ⌋ ) . H := \binom{N+D^2+1}{D^2}
\qquad\text{and}\qquad
D := \binom{d}{\lfloor d/2\rfloor}. H := ( D 2 N + D 2 + 1 ) and D := ( ⌊ d /2 ⌋ d ) .
Theorem 2. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is finite if and only if there is an N ≥ 1 N \ge1 N ≥ 1 such that ∣ S / ∼ ∣ ≤ N |S/\sim| \le N ∣ S / ∼ ∣ ≤ N for every rank-stable subsemigroup S ≤ ⟨ X ⟩ S \le\langle X\rangle S ≤ ⟨ X ⟩ . The diameter of the quotient is at most G d ( N ) G_d(N) G d ( N ) .
Proof Sketch of Theorem 2. The proof of this theorem relies on exterior algebra. We provide a gentle introduction to the necessary background, together with the detailed proof, in Appendix § 10 . We bound class lengths using the following claim.
Claim 3.3. Let a 1 , … , a k ∈ ⟨ X ⟩ a_1,\ldots,a_k\in\langle X\rangle a 1 , … , a k ∈ ⟨ X ⟩ be such that k > H k>H k > H and rank ( a 1 ⋯ a k ) = rank ( a i ) \operatorname{rank}(a_1\cdots a_k)=\operatorname{rank}(a_i) rank ( a 1 ⋯ a k ) = rank ( a i ) for all 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k . There exist i , j i,j i , j with 1 ≤ i + 1 < j ≤ k 1\le i+1<j\le k 1 ≤ i + 1 < j ≤ k such that a 1 ⋯ a i a j ⋯ a k ∼ a 1 ⋯ a k a_1\cdots a_i a_j\cdots a_k\sim a_1\cdots a_k a 1 ⋯ a i a j ⋯ a k ∼ a 1 ⋯ a k .
To prove Claim 3.3, let r r r be the common rank of the matrices a 1 , … , a k a_1,\ldots,a_k a 1 , … , a k and their product. We call the position between a i a_i a i and a i + 1 a_{i+1} a i + 1 cut i i i , and form a graph whose vertices are the k − 1 k-1 k − 1 cuts. For two distinct cuts i , j i,j i , j , consider the products a i a j + 1 a_i a_{j+1} a i a j + 1 and a j a i + 1 a_j a_{i+1} a j a i + 1 . We connect i i i and j j j by an edge when both products are rank preserving; see Figure 3 for an illustration. An edge between cuts i < j i<j i < j means that the middle factor b = a i + 1 ⋯ a j b=a_{i+1}\cdots a_j b = a i + 1 ⋯ a j can be deleted or repeated while preserving the rank of the matrix product.
Figure 3. There is an edge between cuts i i i and j j j when rank ( a i a j + 1 ) = rank ( a j a i + 1 ) = r \operatorname{rank}(a_i a_{j+1})=\operatorname{rank}(a_j a_{i+1})=r rank ( a i a j + 1 ) = rank ( a j a i + 1 ) = r . For cuts i < j i<j i < j , such an edge means the middle factor b b b can be deleted or repeated while preserving rank r r r . The deletion uses im ( a i ) ∩ ker ( a j + 1 ) = { 0 } \operatorname{im}(a_i)\cap\ker(a_{j+1})=\{0\} im ( a i ) ∩ ker ( a j + 1 ) = { 0 } , and repetition uses im ( a j ) ∩ ker ( a i + 1 ) = { 0 } \operatorname{im}(a_j)\cap\ker(a_{i+1})=\{0\} im ( a j ) ∩ ker ( a i + 1 ) = { 0 } .
We first show that in the resulting graph a set of cuts with no edges between them, also known as an independent set, cannot be too large.
Claim 3.4. Each independent set of the cut graph has size at most ( d r ) 2 ≤ D 2 \binom{d}{r}^{2}\le D^{2} ( r d ) 2 ≤ D 2 .
The proof of this claim is the heart of the argument, and this is where we use exterior algebra. The idea is to encode the image and kernel at each cut by a rank-one linear map on W : = ⋀ r Q d W:=\bigwedge^{r}\mathbb{Q}^{d} W := ⋀ r Q d . For each cut i i i , denote this map by T i : W → W T_i:W\to W T i : W → W . Write End ( W ) \operatorname{End}(W) End ( W ) for the vector space of all linear maps from W W W to itself, so T i ∈ End ( W ) T_i\in\operatorname{End}(W) T i ∈ End ( W ) . Our construction also uses a linear functional f i : End ( W ) → Q f_i:\operatorname{End}(W)\to\mathbb{Q} f i : End ( W ) → Q , which assigns a rational number to each such map. The functionals f i f_i f i are constructed such that f i ( T j ) ≠ 0 f_i(T_j)\ne0 f i ( T j ) = 0 if and only if rank ( a i a j + 1 ) = rank ( a j a i + 1 ) = r \operatorname{rank}(a_i a_{j+1})=\operatorname{rank}(a_j a_{i+1})=r rank ( a i a j + 1 ) = rank ( a j a i + 1 ) = r .
Now let I I I be an independent set. For each i ∈ I i\in I i ∈ I , the functional f i f_i f i therefore vanishes on every map T j T_j T j with j ∈ I ∖ { i } j\in I\setminus\{i\} j ∈ I ∖ { i } but not on T i T_i T i . To obtain the claimed bound on ∣ I ∣ |I| ∣ I ∣ , it suffices to show that the maps T i T_i T i with i ∈ I i\in I i ∈ I , are linearly independent in End ( W ) \operatorname{End}(W) End ( W ) . As usual, this means that a relation ∑ i ∈ I α i T i = 0 \sum_{i\in I}\alpha_iT_i=0 ∑ i ∈ I α i T i = 0 can hold only when each α i \alpha_i α i is zero. The bound will then follow from the fact that End ( W ) \operatorname{End}(W) End ( W ) has dimension ( d r ) 2 \binom{d}{r}^{2} ( r d ) 2 .
Consider such a relation ∑ i ∈ I α i T i = 0 \sum_{i\in I}\alpha_iT_i=0 ∑ i ∈ I α i T i = 0 and fix j ∈ I j\in I j ∈ I . Evaluating f j f_j f j at both sides and using its linearity gives
0 = f j ( ∑ i ∈ I α i T i ) = ∑ i ∈ I α i f j ( T i ) = α j f j ( T j ) . 0=f_j\left(\sum_{i\in I}\alpha_iT_i\right)=\sum_{i\in I}\alpha_i f_j(T_i)=\alpha_j f_j(T_j). 0 = f j ( i ∈ I ∑ α i T i ) = i ∈ I ∑ α i f j ( T i ) = α j f j ( T j ) .
The last equality holds because f j ( T i ) = 0 f_j(T_i)=0 f j ( T i ) = 0 for every i ∈ I ∖ { j } i\in I\setminus\{j\} i ∈ I ∖ { j } . Since f j ( T j ) ≠ 0 f_j(T_j)\ne0 f j ( T j ) = 0 , we obtain α j = 0 \alpha_j=0 α j = 0 . This holds for every j ∈ I j\in I j ∈ I , proving linear independence and hence the claimed bound in Claim 3.4.
To bound the diameter of the quotient, let the matrix product a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k be such that a i ∈ X a_i\in X a i ∈ X and k k k is equal to the length of the class [ a 1 ⋯ a k ] ∼ [a_1\cdots a_k]_{\sim} [ a 1 ⋯ a k ] ∼ . Define L d : = N L_d:=N L d := N and L r : = ( H + 1 ) L r + 1 + H L_r:=(H+1)L_{r+1}+H L r := ( H + 1 ) L r + 1 + H for 0 ≤ r < d 0\le r<d 0 ≤ r < d . We prove by decreasing induction on r r r that every matrix product a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k , chosen as above, with rank ( a 1 ⋯ a k ) ≥ r \operatorname{rank}(a_1\cdots a_k)\ge r rank ( a 1 ⋯ a k ) ≥ r satisfies k ≤ L r k\le L_r k ≤ L r . (Here the induction is on the rank of the chosen matrix product, compared to Theorem 1 where the minimum rank of its congruence class was used.)
For the base case with rank d d d , all prefixes lie in the rank-stable subsemigroup of invertible matrices in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Similar to Claim 3.3, we have k ≤ N = L d k \le N = L_d k ≤ N = L d , as otherwise two prefixes would be congruent, contradicting the minimality of k k k .
For the induction step, suppose rank ( a 1 ⋯ a k ) = r < d \operatorname{rank}(a_1 \cdots a_k) = r < d rank ( a 1 ⋯ a k ) = r < d . Decompose the product as y 0 ( c 1 y 1 ) ⋯ ( c ℓ y ℓ ) y_0(c_1y_1)\cdots(c_\ell y_\ell) y 0 ( c 1 y 1 ) ⋯ ( c ℓ y ℓ ) , where c i ∈ X c_i \in X c i ∈ X , each nonempty y i y_i y i has rank greater than r r r , and each c i y i c_i y_i c i y i has rank r r r . Observe that each nonempty factor y i y_i y i is itself shortest in its class [ y i ] ∼ [y_i]_{\sim} [ y i ] ∼ , as otherwise, replacing it by a shorter congruent product would contradict the minimality of k k k . The induction hypothesis therefore bounds each y i y_i y i by L r + 1 L_{r+1} L r + 1 matrices. Since the factors c i y i c_i y_i c i y i and their product all have rank r r r , so ℓ ≤ H \ell\le H ℓ ≤ H must hold by Claim 3.3. Again as otherwise, it would contradict the minimality of k k k . This implies
k ≤ L r + 1 + ℓ ( L r + 1 + 1 ) ≤ L r + 1 + H ( L r + 1 + 1 ) = L r , k \le L_{r+1} + \ell(L_{r+1} + 1) \le L_{r+1} + H(L_{r+1} + 1) = L_r, k ≤ L r + 1 + ℓ ( L r + 1 + 1 ) ≤ L r + 1 + H ( L r + 1 + 1 ) = L r ,
where solving this recurrence gives L 0 = ( N + 1 ) ( H + 1 ) d − 1 L_0 = (N + 1)(H + 1)^d - 1 L 0 = ( N + 1 ) ( H + 1 ) d − 1 , proving the diameter bound.
Example 4. Let k ≥ 1 k \ge1 k ≥ 1 . Let X k = { A 1 , … , A k } ⊆ Mat 2 k ( Q ) X_k = \{A_1,\ldots,A_k\} \subseteq\operatorname{Mat}_{2k}(\mathbb{Q}) X k = { A 1 , … , A k } ⊆ Mat 2 k ( Q ) , where
A i = diag ( E 10 , … , E 10 ⏟ i − 1 blocks , E 01 , I 2 , … , I 2 ⏟ k − i blocks ) with E 10 = ( 0 0 1 0 ) and E 01 = ( 0 1 0 0 ) . A_i = \operatorname{diag}\left(\underbrace{E_{10},\ldots,E_{10}}_{i-1\ \text{blocks}}, E_{01}, \underbrace{I_2,\ldots,I_2}_{k-i\ \text{blocks}}\right)
\quad\text{with}\quad
E_{10} = \begin{pmatrix} 0 & 0 \\ 1 & 0 \end{pmatrix}
\quad\text{and}\quad
E_{01} = \begin{pmatrix} 0 & 1 \\ 0 & 0 \end{pmatrix}. A i = diag i − 1 blocks E 10 , … , E 10 , E 01 , k − i blocks I 2 , … , I 2 with E 10 = ( 0 1 0 0 ) and E 01 = ( 0 0 1 0 ) .
Each A i A_i A i consists of k k k diagonal blocks of size 2 × 2 2 \times2 2 × 2 , and all its entries belong to { 0 , 1 } \{0,1\} { 0 , 1 } . Let the congruence be equality. We show that ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ is finite and every rank-stable subsemigroup is a singleton, while the matrix
T k : = diag ( E 01 , … , E 01 ⏟ k blocks ) T_k := \operatorname{diag}\left(\underbrace{E_{01},\ldots,E_{01}}_{k\ \text{blocks}}\right) T k := diag k blocks E 01 , … , E 01
has length 2 k − 1 2^k - 1 2 k − 1 . Hence, in dimension d = 2 k d = 2k d = 2 k , the quotient has diameter at least 2 d / 2 − 1 2^{d/2} - 1 2 d /2 − 1 , even though the uniform bound on rank-stable subsemigroups is N = 1 N = 1 N = 1 .
To understand the construction, represent the two bit values by the row vectors e 0 = ( 1 , 0 ) e_0 = (1,0) e 0 = ( 1 , 0 ) and e 1 = ( 0 , 1 ) e_1 = (0,1) e 1 = ( 0 , 1 ) . Each diagonal block records one bit, with the first block representing the least significant bit. The two matrices act as follows:
e 1 E 10 = e 0 , e 0 E 01 = e 1 , e 0 E 10 = e 1 E 01 = 0. e_1E_{10} = e_0,\qquad e_0E_{01} = e_1,\qquad e_0E_{10} = e_1E_{01} = 0. e 1 E 10 = e 0 , e 0 E 01 = e 1 , e 0 E 10 = e 1 E 01 = 0.
So multiplication by A i A_i A i changes the bit pattern
( 1 , … , 1 , 0 , b i + 1 , … , b k ) to ( 0 , … , 0 , 1 , b i + 1 , … , b k ) . (1,\ldots,1,0,b_{i+1},\ldots,b_k)
\quad\text{to}\quad
(0,\ldots,0,1,b_{i+1},\ldots,b_k). ( 1 , … , 1 , 0 , b i + 1 , … , b k ) to ( 0 , … , 0 , 1 , b i + 1 , … , b k ) .
Now consider a product representing P k P_k P k , starting with e 0 e_0 e 0 in every block. Every tracked row must end at e 1 e_1 e 1 , and none can become zero along the way, since subsequent multiplication cannot restore a zero row. The next generator must therefore correspond to the least significant bit that is still zero. Its action is exactly a binary increment, as it resets the preceding 1-bits to 0, changes this 0-bit to 1, and leaves the remaining bits unchanged. Reaching the all-one pattern from the all-zero pattern thus requires 2 k − 1 2^k - 1 2 k − 1 increments. Following these increments gives a product equal to P k P_k P k , so its length is exactly 2 k − 1 2^k - 1 2 k − 1 .
The rank-stable matrices in ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ are diagonal, with 0 and 1 along the diagonal. For two such matrices a , b a,b a , b , the condition rank ( a ) = rank ( b ) = rank ( a b ) \operatorname{rank}(a) = \operatorname{rank}(b) = \operatorname{rank}(ab) rank ( a ) = rank ( b ) = rank ( ab ) forces a = b a = b a = b . Thus every rank-stable subsemigroup is a singleton, giving the uniform bound N = 1 N = 1 N = 1 . The details are given in Appendix § 13 .
Uniform bound on groups in the Zariski closure
The next theorem assumes a uniform bound on the quotient images of groups in the Zariski closure of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ , where the congruence is also defined on the closure. Write ⟨ X ⟩ ‾ \overline{\langle X\rangle} ⟨ X ⟩ for this closure in Mat d ( Q ) \operatorname{Mat}_d(\mathbb{Q}) Mat d ( Q ) .
Theorem 3. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite and ∼ \sim ∼ be a congruence on ⟨ X ⟩ ‾ \overline{\langle X\rangle} ⟨ X ⟩ . Assume that there exists N ≥ 1 N \ge1 N ≥ 1 such that ∣ G / ∼ ∣ ≤ N |G/\sim| \le N ∣ G / ∼ ∣ ≤ N for every group G ≤ ⟨ X ⟩ ‾ G \le\overline{\langle X\rangle} G ≤ ⟨ X ⟩ . Then ⟨ X ⟩ ‾ / ∼ \overline{\langle X\rangle}/\sim ⟨ X ⟩ / ∼ is finite and has diameter at most 2 2 d ( d + 3 ) N d 2^{2d(d+3)N^d} 2 2 d ( d + 3 ) N d .
To obtain the diameter bound, we return to the factorisation trees used in Theorem 1. There, we retained a bounded number of factors and counted their possible classes. Here, we use the full tree to bound the length of each class. Our starting point is the refinement of rank-stable factorisation trees in [12 Theorem 5] . There, we strengthened the rank-stability condition to stability for finite matrix monoids. In this finite setting, the children at every node with at least three children generate a group. Below, we adapt this construction by using the closure and allowing matrix products to change within their congruence classes (more details below).
Before we proceed, we note that Example 4 provides a lower bound for this theorem as well.
Example 5. Consider the matrices X k ⊆ Mat 2 k ( Q ) X_k \subseteq\operatorname{Mat}_{2k}(\mathbb{Q}) X k ⊆ Mat 2 k ( Q ) from Example 4, with equality as the congruence. Since ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ is finite, it is Zariski closed. Recall that every matrix group is a rank-stable subsemigroup. As all rank-stable subsemigroups in this example are singletons (see Example 4), every group in the closure is trivial. Thus Theorem 3 applies with N = 1 N=1 N = 1 . However, as shown in Example 4, the matrix T k T_k T k has length 2 k − 1 2^k-1 2 k − 1 . Writing d = 2 k d=2k d = 2 k , the quotient therefore has diameter at least 2 d / 2 − 1 2^{d/2}-1 2 d /2 − 1 . Hence an exponential dependence on d d d is necessary, even when all groups in the closure are trivial.
We can also obtain an infinite Zariski closure for the semigroup closure by adding one coordinate to each generator, while preserving the quotient and the bound N = 1 N=1 N = 1 . Let Y k = { B 1 , … , B k } ⊆ Mat 2 k + 1 ( Q ) Y_k=\{B_1,\ldots,B_k\}\subseteq\operatorname{Mat}_{2k+1}(\mathbb{Q}) Y k = { B 1 , … , B k } ⊆ Mat 2 k + 1 ( Q ) be such that B i = diag ( 2 , A i ) B_i=\operatorname{diag}(2,A_i) B i = diag ( 2 , A i ) . We have that B i 1 ⋯ B i ℓ = diag ( 2 ℓ , A i 1 ⋯ A i ℓ ) B_{i_1}\cdots B_{i_\ell}=\operatorname{diag}(2^\ell,A_{i_1}\cdots A_{i_\ell}) B i 1 ⋯ B i ℓ = diag ( 2 ℓ , A i 1 ⋯ A i ℓ ) . Observe that both ⟨ Y k ⟩ \langle Y_k\rangle ⟨ Y k ⟩ and its Zariski closure C k : = ⟨ Y k ⟩ ‾ C_k:=\overline{\langle Y_k\rangle} C k := ⟨ Y k ⟩ are infinite.
We define a congruence ∼ \sim ∼ by forgetting the added coordinate. Since ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ is finite, the set
{ diag ( t , B ) : t ∈ Q , B ∈ ⟨ X k ⟩ } \{\operatorname{diag}(t,B):t\in\mathbb{Q},\ B\in\langle X_k\rangle\} { diag ( t , B ) : t ∈ Q , B ∈ ⟨ X k ⟩} is Zariski closed and contains C k C_k C k . Hence projection onto the lower-right block defines a semigroup homomorphism π : C k → ⟨ X k ⟩ \pi:C_k\to\langle X_k\rangle π : C k → ⟨ X k ⟩ . Define a ∼ b a\sim b a ∼ b if π ( a ) = π ( b ) \pi(a)=\pi(b) π ( a ) = π ( b ) . This gives a congruence on the closure. For every group G ≤ C k G\le C_k G ≤ C k , its image π ( G ) \pi(G) π ( G ) is a group in ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ . By Example 4, every such group is trivial. Consequently, ∣ G / ∼ ∣ = 1 |G/\sim|=1 ∣ G / ∼ ∣ = 1 , and Theorem 3 applies with N = 1 N=1 N = 1 .
Moreover, projection induces an isomorphism ⟨ Y k ⟩ / ∼ ≅ ⟨ X k ⟩ \langle Y_k\rangle/\sim\cong\langle X_k\rangle ⟨ Y k ⟩ / ∼≅ ⟨ X k ⟩ , sending the class of B i B_i B i to A i A_i A i . By the above, the class mapping to T k T_k T k has length 2 k − 1 2^k-1 2 k − 1 . Again, writing d = 2 k + 1 d=2k+1 d = 2 k + 1 , the quotient thus has diameter at least $2^{(d-1)/2}-1.
Proof Sketch of Theorem 3. We outline the main ideas here; the full details, including proofs of the claims, are provided in Appendix E.
To prove finiteness of the quotient and bound its diameter, it suffices to bound the length of every congruence class. For each a ∈ ⟨ X ⟩ a\in\langle X\rangle a ∈ ⟨ X ⟩ , we will construct a tree whose leaves belong to X X X and whose leaf product is congruent to a a a . We then shorten this tree and bound its number of leaves. We will need the following claim.
Claim 3.5. Given a , b ∈ ⟨ X ⟩ a,b\in\langle X\rangle a , b ∈ ⟨ X ⟩ , the condition im ( a ) = im ( a b ) \operatorname{im}(a)=\operatorname{im}(ab) im ( a ) = im ( ab ) implies a ∼ a b n a\sim ab^n a ∼ a b n for some n ≥ 1 n\ge1 n ≥ 1 .
We use the fact that the Zariski closure of a subsemigroup inside an algebraic group is itself a group. Let a ∈ ⟨ X ⟩ a \in\langle X\rangle a ∈ ⟨ X ⟩ be rank-stable. The matrices with kernel ker ( a ) \ker(a) ker ( a ) and image im ( a ) \operatorname{im}(a) im ( a ) form an algebraic group, whose identity I im ( a ) I_{\operatorname{im}(a)} I im ( a ) is the projection onto im ( a ) \operatorname{im}(a) im ( a ) along ker ( a ) \ker(a) ker ( a ) . Therefore, the Zariski closure of ⟨ a ⟩ \langle a\rangle ⟨ a ⟩ inside this group is a group contained in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . By assumption, its quotient has at most N N N elements, so a n ∼ I im ( a ) a^n \sim I_{\operatorname{im}(a)} a n ∼ I im ( a ) for some 1 ≤ n ≤ N 1 \le n \le N 1 ≤ n ≤ N .
Now suppose im ( a ) = im ( a b ) \operatorname{im}(a) = \operatorname{im}(ab) im ( a ) = im ( ab ) . This means that b b b maps im ( a ) \operatorname{im}(a) im ( a ) onto itself, so intuitively speaking b b b restricts to an invertible map on im ( a ) \operatorname{im}(a) im ( a ) . This gives im ( a ) = im ( a ) b d ⊆ im ( b d ) \operatorname{im}(a) = \operatorname{im}(a)b^d \subseteq\operatorname{im}(b^d) im ( a ) = im ( a ) b d ⊆ im ( b d ) . But b d b^d b d is rank-stable, and its identity I im ( b d ) I_{\operatorname{im}(b^d)} I im ( b d ) fixes every vector in im ( a ) \operatorname{im}(a) im ( a ) . Consequently, a b d n = a ( b d ) n ∼ a I im ( b d ) = a ab^dn = a(b^d)^n \sim aI_{\operatorname{im}(b^d)} = a a b d n = a ( b d ) n ∼ a I im ( b d ) = a , which proves our Claim 3.5.
We next use Claim 3.5 to modify a rank-stable factorisation tree. We follow the strategy of the proof of [12 Theorem 5] , where rank-stable factorisation trees for finite matrix monoids are modified so that the children at every node with at least three children generate a group. Here, we use Claim 3.5 to modify products within their congruence classes, so that these children generate a stable subsemigroup.
Following the strategy, the starting tree for this procedure is defined in [12 Theorem 7] . Such a tree is a rank-stable factorisation tree, where every node with at least three children, besides the rank-stability of the parent, must have children with the same rank as the parent. Thus, if the children are a 1 , … , a s a_1,\ldots,a_s a 1 , … , a s and the parent is a = a 1 ⋯ a s a = a_1 \cdots a_s a = a 1 ⋯ a s , then rank ( a 1 ) = ⋯ = rank ( a s ) = rank ( a ) = rank ( a 2 ) \operatorname{rank}(a_1) = \cdots= \operatorname{rank}(a_s) = \operatorname{rank}(a) = \operatorname{rank}(a^2) rank ( a 1 ) = ⋯ = rank ( a s ) = rank ( a ) = rank ( a 2 ) . Such a tree has height at most d ( d + 3 ) d(d+3) d ( d + 3 ) and reduced height at most d d d , where the reduced height counts the maximum number of nodes with at least three children on a path from the root to a leaf [12 Theorem 7] .
We now explain how to modify a node a a a with at least three children; see Figure 4 for an illustration. Recall that the children all have the same rank as a a a , however, they may have different kernels and images. Our aim is to replace their product by a congruent product whose factors all have the kernel and image of a a a . Since a a a is rank-stable, these factors will generate a stable subsemigroup.
Figure 4. We illustrate the insertions and regrouping procedure at a node a = a 1 a 2 a 3 a=a_1a_2a_3 a = a 1 a 2 a 3 with three children, where rank ( a i ) = rank ( a ) = rank ( a 2 ) \operatorname{rank}(a_i)=\operatorname{rank}(a)=\operatorname{rank}(a^2) rank ( a i ) = rank ( a ) = rank ( a 2 ) . At cut i i i with i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } , the products y i y_i y i and x i x_i x i have the kernel and image of the prefix a 1 ⋯ a i a_1\cdots a_i a 1 ⋯ a i and suffix a i + 1 ⋯ a 3 a_{i+1}\cdots a_3 a i + 1 ⋯ a 3 , respectively. The shaded powers are inserted first at the right cut, then at the left. Every child in (iv) has kernel ker ( a ) \operatorname{ker}(a) ker ( a ) and image im ( a ) \operatorname{im}(a) im ( a ) . Throughout, the child product is congruent to the unchanged root label a a a .
To obtain these factors, we insert suitable products between the children and then regroup. Roughly speaking, consider a cut that divides a 1 ⋯ a s a_1 \cdots a_s a 1 ⋯ a s into a prefix product c 1 c_1 c 1 and a suffix product c 2 c_2 c 2 , so that a = c 1 c 2 a = c_1c_2 a = c 1 c 2 . To construct a suitable insertion, we focus on the kernels and images of c 1 c_1 c 1 and c 2 c_2 c 2 . The key observation is that these subspaces can be obtained from a bounded number of their children. In particular, by [12 Lemma 8] there are products y y y and x x x , each using at most 2 d + 1 2^d+1 2 d + 1 children, such that
ker ( y ) = ker ( c 1 ) im ( y ) = im ( c 1 ) , and ker ( x ) = ker ( c 2 ) im ( x ) = im ( c 2 ) . \ker(y) = \ker(c_1) \qquad\operatorname{im}(y) = \operatorname{im}(c_1), \qquad\text{and} \qquad\ker(x) = \ker(c_2) \qquad\operatorname{im}(x) = \operatorname{im}(c_2). ker ( y ) = ker ( c 1 ) im ( y ) = im ( c 1 ) , and ker ( x ) = ker ( c 2 ) im ( x ) = im ( c 2 ) .
The rank conditions on the node imply that b : = x y b := xy b := x y satisfies im ( c 1 b ) = im ( c 1 ) \operatorname{im}(c_1b) = \operatorname{im}(c_1) im ( c 1 b ) = im ( c 1 ) . Claim 3.5 therefore gives c 1 ∼ c 1 b n c_1 \sim c_1b^n c 1 ∼ c 1 b n for some n ≥ 1 n \ge1 n ≥ 1 , and hence a = c 1 c 2 ∼ c 1 ( x y ) n c 2 a = c_1c_2 \sim c_1(xy)^nc_2 a = c 1 c 2 ∼ c 1 ( x y ) n c 2 . We insert such powers ( x y ) n (xy)^n ( x y ) n starting at the rightmost cut and moving left, as shown in Figure 4 . At each step, the powers already inserted lie to the right of the current cut, so the prefix is still c 1 c_1 c 1 and we can apply the relation c 1 ∼ c 1 b n c_1 \sim c_1b^n c 1 ∼ c 1 b n . We then regroup the resulting product into factors whose kernel and image are those of a a a . As explained, the construction is such that each new factor consists of at most two auxiliary products and some (original) child a i a_i a i , and therefore uses at most 2 ( 2 d + 1 ) + 1 ≤ 2 d + 2 2(2^d+1)+1 \le2^{d+2} 2 ( 2 d + 1 ) + 1 ≤ 2 d + 2 matrices among the children a i a_i a i ’s. Since a a a is rank-stable, these factors generate a stable subsemigroup. The details of this regrouping are given in Appendix § 11 .
Applying this modification throughout the tree results a tree notion as follows. Let ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . A stable ∼ \sim ∼ -factorisation tree of a ∈ ⟨ X ⟩ a \in\langle X\rangle a ∈ ⟨ X ⟩ is a finite ordered tree with nodes in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . The root of the tree is a a a and leaves lie in X X X . The nodes c c c of the tree, except leaves, have children c 1 , … , c s c_1,\ldots,c_s c 1 , … , c s with s ≥ 2 s\ge2 s ≥ 2 such that (1) c ∼ c 1 ⋯ c s c\sim c_1\cdots c_s c ∼ c 1 ⋯ c s and (2) the matrices c i c_i c i generate a stable subsemigroup when s ≥ 3 s\ge3 s ≥ 3 . A stable ∼ \sim ∼ -factorisation forest is a choice of tree for every a ∈ ⟨ X ⟩ a\in\langle X\rangle a ∈ ⟨ X ⟩ . The construction sketched above gives the following bounds.
Lemma 1. Suppose ∼ \sim ∼ is a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ such that, for all a , b ∈ ⟨ X ⟩ a,b\in\langle X\rangle a , b ∈ ⟨ X ⟩ , the equality im ( a ) = im ( a b ) \operatorname{im}(a)=\operatorname{im}(ab) im ( a ) = im ( ab ) implies a ∼ a b n a\sim ab^n a ∼ a b n for some n ≥ 1 n\ge1 n ≥ 1 . Then ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ admits a stable ∼ \sim ∼ -factorisation forest whose trees have height at most 2 d ( d + 3 ) 2d(d+3) 2 d ( d + 3 ) and reduced height at most d d d .
To obtain these bounds, observe that each new factor obtained after regrouping is built from at most 2 d + 2 2^{d+2} 2 d + 2 matrices among the children a i a_i a i of a a a , using a binary tree of height at most d + 2 d+2 d + 2 . Here we were modifying an arbitrary node a a a of the tree, so its children a i a_i a i may themselves be products of matrices in X X X . We process the tree from the leaves upwards, so that for each a i a_i a i (re)appearing while processing a a a , we attach a copy of the already modified subtree rooted at a i a_i a i . All in all, each such local modification adds at most d + 3 d+3 d + 3 levels without increasing the reduced height. Since a path in the starting tree contains at most d d d nodes with at least three children, the resulting height is at most d ( d + 3 ) + d ( d + 3 ) = 2 d ( d + 3 ) d(d+3)+d(d+3)=2d(d+3) d ( d + 3 ) + d ( d + 3 ) = 2 d ( d + 3 ) , and the reduced height remains at most d d d .
As the last step, we shorten these trees using the assumed group bound. Consider a node whose children generate a stable subsemigroup S S S . Taking the Zariski closure of S S S inside the group of matrices with its common kernel and image gives a group G ≤ ⟨ X ⟩ G\le\langle X\rangle G ≤ ⟨ X ⟩ containing S S S . By assumption, ∣ G / ∼ ∣ ≤ N |G/\sim|\le N ∣ G / ∼ ∣ ≤ N , so S S S meets at most N N N congruence classes. Clearly, if this node had more than N N N children, two nonempty prefix products of its children are congruent. We can therefore delete the children between these prefixes while preserving the congruence class, and repeat until the number of children is at most N N N .
The remaining binary nodes contribute at most a factor of two per level to the number of leaves. The shortened nodes contribute at most a factor of N N N , and at most d d d such nodes occur on any path from the root to a leaf. These two facts combined together yield the bound 2 2 d ( d + 3 ) N d 2^{2d(d+3)}N^d 2 2 d ( d + 3 ) N d for the number of leaves in the final tree. These leaves all lie in X X X , and notably their product is congruent to the root of the tree. This establishes that every congruence class has length at most 2 2 d ( d + 3 ) N d 2^{2d(d+3)}N^d 2 2 d ( d + 3 ) N d . This proves the diameter bound and, since X X X is finite, the finiteness of the quotient.
Application to Border Complexity
Separation of homogeneous ABPs and their closures
Let w ≥ 2 w \ge2 w ≥ 2 and d ≥ 3 d \ge3 d ≥ 3 . A homogeneous algebraic branching program (ABP) of width at most w w w and degree d d d computes a polynomial
f ( x ) = L 1 ( x ) ⋯ L d ( x ) , f(x) = L_1(x) \cdots L_d(x), f ( x ) = L 1 ( x ) ⋯ L d ( x ) ,
where L 1 L_1 L 1 is a 1 × w 1 \times w 1 × w matrix, L d L_d L d is a w × 1 w \times1 w × 1 matrix, and the remaining matrices are w × w w \times w w × w . The entries of the L i L_i L i are homogeneous linear forms in the input variables x = ( x 1 , … , x m ) x = (x_1,\ldots,x_m) x = ( x 1 , … , x m ) , with coefficients in Q \mathbb{Q} Q . We note in passing that, in this definition, we have exactly d d d matrix factors. ABPs and their border complexity have been studied in recent years [10 , 18 , 19 , 20 ] . Homogeneous models and their connections with algebraic geometry have also received recent attention [21 , 22 ] . In our constructions, we keep both the width w w w and the degree d d d fixed and use homogeneous linear forms as used in [22 ] .
Write Q [ x ] d \mathbb{Q}[x]_d Q [ x ] d for the vector space of homogeneous degree-d d d polynomials in the variables x x x , with coefficients in Q \mathbb{Q} Q . Denote by A w , d \mathcal{A}_{w,d} A w , d the set of polynomials Q ‾ [ x ] d \overline{\mathbb{Q}}[x]_d Q [ x ] d computed by the ABPs defined above. As usual, viewing each polynomial as its vector of coefficients, we take the Zariski closure of the set A w , d \mathcal{A}_{w,d} A w , d of polynomials in the vector space of homogeneous degree-d d d polynomials over Q ‾ \overline{\mathbb{Q}} Q ; we denote this closure by A ‾ w , d \overline{\mathcal{A}}_{w,d} A w , d .
Below, we prove the following separation and describe it through a semigroup congruence.
Theorem 4. Let w ≥ 2 w \ge2 w ≥ 2 and d ≥ 3 d \ge3 d ≥ 3 and m ≥ 2 w + 7 m \ge2w+7 m ≥ 2 w + 7 . The class A w , d \mathcal{A}_{w,d} A w , d of width-w w w degree-d d d homogeneous ABPs is separated from A ‾ w , d \overline{\mathcal{A}}_{w,d} A w , d .
We will give an explicit polynomial F w , d F_{w,d} F w , d with integer coefficients such that that F w , d ∈ A ‾ w , d ∖ A w , d F_{w,d} \in\overline{\mathcal{A}}_{w,d} \setminus\mathcal{A}_{w,d} F w , d ∈ A w , d ∖ A w , d .
Proof Sketch of Theorem 4. We first construct a matrix semigroup and a congruence whose zero class consists solely of an encoding of polynomials in A w , d \mathcal{A}_{w,d} A w , d . We then give a polynomial in A ‾ w , d ∖ A w , d \overline{\mathcal{A}}_{w,d} \setminus\mathcal{A}_{w,d} A w , d ∖ A w , d and show that its encoding belongs to a different class. The details including the proofs of the claims are provided in Appendices F.
To construct the semigroup, we use a notion of universal ABPs. For this aim, we replace each coefficient of each linear form in entries of the ABP by an independent parameter θ i \theta_i θ i ; let the vector θ = ( θ 1 , … , θ ν ) \boldsymbol{\theta} = (\theta_1,\ldots,\theta_\nu) θ = ( θ 1 , … , θ ν ) denote the introduced formal parameters in this way, where ν = m ( 2 w + ( d − 2 ) w 2 ) \nu= m(2w+(d-2)w^2) ν = m ( 2 w + ( d − 2 ) w 2 ) . Denote the polynomial computed by this universal ABP by U ( θ , x ) U(\theta,x) U ( θ , x ) , which is linear in each θ i \theta_i θ i . Choosing values for the parameters produces all ABPs in A w , d \mathcal{A}_{w,d} A w , d :
A w , d = { U ( t , x ) : t ∈ Q ‾ ν } . \mathcal{A}_{w,d} = \{U(t,x) : t \in\overline{\mathbb{Q}}^\nu\}. A w , d = { U ( t , x ) : t ∈ Q ν } .
Below, we use translations to describe the choices of parameters θ \boldsymbol{\theta} θ . For t ∈ Q ‾ ν \mathbf{t} \in\overline{\mathbb{Q}}^\nu t ∈ Q ν , the translation θ ↦ θ + t \boldsymbol{\theta} \mapsto\boldsymbol{\theta}+\mathbf{t} θ ↦ θ + t replaces each θ i \theta_i θ i by θ i + t i \theta_i+t_i θ i + t i , leaving the input variables x \mathbf{x} x unchanged. Starting from the zero assignment, this gives the assignment θ = t \boldsymbol{\theta}=\mathbf{t} θ = t . Our aim is to represent these translations by matrices and then recover the polynomial computed by the corresponding ABP.
Recall that each parameter θ i \theta_i θ i belongs to one layer of the universal ABP, and each produced monomial will have total degree d d d in θ \boldsymbol{\theta} θ . The combination of translations might however produce monomials with smaller total degree, for example, ( θ i + t i ) ( θ j + t j ) = θ i θ j + t j θ i + t i θ j + t i t j (\theta_i+t_i)(\theta_j+t_j)=\theta_i\theta_j+t_j\theta_i+t_i\theta_j+t_it_j ( θ i + t i ) ( θ j + t j ) = θ i θ j + t j θ i + t i θ j + t i t j . We therefore collect all monomials obtained by choosing at most one parameter from each layer into a row vector v ( θ ) v(\boldsymbol{\theta}) v ( θ ) . The number of such monomials is L θ = ( 1 + m w ) 2 ( 1 + m w 2 ) d − 2 L_\theta=(1+mw)^2(1+mw^2)^{d-2} L θ = ( 1 + m w ) 2 ( 1 + m w 2 ) d − 2 ; we then fix an order on them, placing the constant monomial 1 1 1 at first. This allows us define ρ t ∈ Mat L θ ( Q ‾ ) \rho_{\mathbf{t}}\in\operatorname{Mat}_{L_\theta}(\overline{\mathbb{Q}}) ρ t ∈ Mat L θ ( Q ) , for t ∈ Q ‾ ν \mathbf{t}\in\overline{\mathbb{Q}}^\nu t ∈ Q ν , by v ( θ ) ρ t = v ( θ + t ) v(\boldsymbol{\theta})\rho_{\mathbf{t}}=v(\boldsymbol{\theta}+\mathbf{t}) v ( θ ) ρ t = v ( θ + t ) . Composing translations gives
ρ 0 = I L θ , ρ s ρ t = ρ s + t , ρ t − 1 = ρ − t . \rho_{\mathbf{0}}=I_{L_\theta},\qquad\rho_{\mathbf{s}}\rho_{\mathbf{t}}=\rho_{\mathbf{s}+\mathbf{t}},\qquad\rho_{\mathbf{t}}^{-1}=\rho_{-\mathbf{t}}. ρ 0 = I L θ , ρ s ρ t = ρ s + t , ρ t − 1 = ρ − t .
Recall that the universal polynomial is homogenous of degree d d d in variables x \mathbf{x} x . There are L x : = ( m + d − 1 d ) L_x:=\binom{m+d-1}{d} L x := ( d m + d − 1 ) many such homogenous monomials, we again fix an order on them. Write coeff \operatorname{coeff} coeff for the coefficient map in this order:
coeff : Q ‾ [ θ ] [ x ] d ⟶ Q ‾ [ θ ] 1 × L x \operatorname{coeff}:\overline{\mathbb{Q}}[\boldsymbol{\theta}][\mathbf{x}]_d\longrightarrow\overline{\mathbb{Q}}[\boldsymbol{\theta}]^{1\times L_x} coeff : Q [ θ ] [ x ] d ⟶ Q [ θ ] 1 × L x
which extracts coefficients with respect to x \mathbf{x} x in the fixed order of monomials in x \mathbf{x} x . By the above, since the universal polynomial lie in Q ‾ [ θ ] [ x ] d \overline{\mathbb{Q}}[\boldsymbol{\theta}][\mathbf{x}]_d Q [ θ ] [ x ] d , and since each of its coefficients is a Q \mathbb{Q} Q -linear combination of coordinates in v ( θ ) v(\boldsymbol{\theta}) v ( θ ) , there is a fixed matrix C U ∈ Q L θ × L x C_U\in\mathbb{Q}^{L_\theta\times L_x} C U ∈ Q L θ × L x such that coeff ( U ( θ , x ) ) = v ( θ ) C U \operatorname{coeff}(U(\boldsymbol{\theta},\mathbf{x}))=v(\boldsymbol{\theta})C_U coeff ( U ( θ , x )) = v ( θ ) C U . Translating the parameters gives coeff ( U ( θ + t , x ) ) = v ( θ + t ) C U = v ( θ ) ρ t C U \operatorname{coeff}(U(\boldsymbol{\theta}+\mathbf{t},\mathbf{x}))=v(\boldsymbol{\theta}+\mathbf{t})C_U=v(\boldsymbol{\theta})\rho_{\mathbf{t}}C_U coeff ( U ( θ + t , x )) = v ( θ + t ) C U = v ( θ ) ρ t C U . Moreover, evaluating at θ = 0 \boldsymbol{\theta}=\mathbf{0} θ = 0 yields
v ( 0 ) ρ t C U = coeff ( U ( t , x ) ) , (3) v(\mathbf{0})\rho_{\mathbf{t}}C_U=\operatorname{coeff}(U(\mathbf{t},\mathbf{x})), \tag*{(3)} v ( 0 ) ρ t C U = coeff ( U ( t , x )) , (3)
where v ( 0 ) = ( 1 , 0 ) v(\mathbf{0})=(1,0) v ( 0 ) = ( 1 , 0 ) is the value of the monomial row at the zero parameter assignment. Intuitively speaking, multiplication of v ( 0 ) v(\mathbf{0}) v ( 0 ) by ρ t \rho_{\mathbf{t}} ρ t changes it to v ( t ) v(\mathbf{t}) v ( t ) , and then multiplication by C U C_U C U gives the coefficient vector of the resulting ABP; see Figure 5 for an example.
Figure 5. A small instance of the universal ABP with width 1 and degree 3, with input variables x , y x,y x , y . For simplicity of the example, we fix all parameters to 0 or 1, except θ 1 , θ 2 \theta_1,\theta_2 θ 1 , θ 2 . This three-layer ABP computes U ( θ 1 , θ 2 , x , y ) = ( θ 1 x ) ( θ 2 x + y ) x U(\theta_1,\theta_2,x,y)=(\theta_1x)(\theta_2x+y)x U ( θ 1 , θ 2 , x , y ) = ( θ 1 x ) ( θ 2 x + y ) x . We use v ( θ ) = ( 1 , θ 1 , θ 2 , θ 1 θ 2 ) v(\theta)=(1,\theta_1,\theta_2,\theta_1\theta_2) v ( θ ) = ( 1 , θ 1 , θ 2 , θ 1 θ 2 ) and order the input monomials as ( x 3 , x 2 y , x y 2 , y 3 ) (x^3,x^2y,xy^2,y^3) ( x 3 , x 2 y , x y 2 , y 3 ) . At t = ( 2 , 3 ) t=(2,3) t = ( 2 , 3 ) , multiplication by ρ t \rho_t ρ t gives v ( t ) = ( 1 , 2 , 3 , 6 ) v(t)=(1,2,3,6) v ( t ) = ( 1 , 2 , 3 , 6 ) , whose coordinates are the values of 1 , θ 1 , θ 2 , θ 1 θ 2 1,\theta_1,\theta_2,\theta_1\theta_2 1 , θ 1 , θ 2 , θ 1 θ 2 at θ = t \theta=t θ = t . The multiplication by C U C_U C U extracts the coefficients of 6 x 3 + 2 x 2 y 6x^3+2x^2y 6 x 3 + 2 x 2 y as shown above.
By (3) in the above, we can expresses the evaluation of the universal ABP at a parameter as a matrix product; below, we place the defined matrices into one common dimension, so that we can work within a matrix semigroup. To this aim, we consider 4 blocks of size 1 1 1 , L θ L_\theta L θ , L θ L_\theta L θ , L x + L θ L_x+L_\theta L x + L θ , and write L t o t : = 1 + 3 L θ + L x L_{\mathrm{tot}}:=1+3L_\theta+L_x L tot := 1 + 3 L θ + L x for the total length of these blocks. We define a linear map ι : Q ‾ [ x ] d → Mat L t o t ( Q ‾ ) \iota:\overline{\mathbb{Q}}[\mathbf{x}]_d\to\operatorname{Mat}_{L_{\mathrm{tot}}}(\overline{\mathbb{Q}}) ι : Q [ x ] d → Mat L tot ( Q ) as follows, for all polynomials f ∈ Q ‾ [ x ] d f\in\overline{\mathbb{Q}}[\mathbf{x}]_d f ∈ Q [ x ] d ,
ι ( f ) : = ( 0 1 × 1 0 1 × L θ 0 1 × L θ ( coeff ( f ) , 0 1 × L θ ) 0 L θ × 1 0 L θ × L θ 0 L θ × L θ 0 L θ × ( L x + L θ ) 0 L θ × 1 0 L θ × L θ 0 L θ × L θ 0 L θ × ( L x + L θ ) 0 ( L x + L θ ) × 1 0 ( L x + L θ ) × L θ 0 ( L x + L θ ) × L θ 0 ( L x + L θ ) × ( L x + L θ ) ) , \iota(f):=
\begin{pmatrix}
0_{1\times1} & 0_{1\times L_\theta} & 0_{1\times L_\theta} & \left(\operatorname{coeff}(f),0_{1\times L_\theta}\right) \\
0_{L_\theta\times1} & 0_{L_\theta\times L_\theta} & 0_{L_\theta\times L_\theta} & 0_{L_\theta\times(L_x+L_\theta)} \\
0_{L_\theta\times1} & 0_{L_\theta\times L_\theta} & 0_{L_\theta\times L_\theta} & 0_{L_\theta\times(L_x+L_\theta)} \\
0_{(L_x+L_\theta)\times1} & 0_{(L_x+L_\theta)\times L_\theta} & 0_{(L_x+L_\theta)\times L_\theta} & 0_{(L_x+L_\theta)\times(L_x+L_\theta)}
\end{pmatrix}, ι ( f ) := 0 1 × 1 0 L θ × 1 0 L θ × 1 0 ( L x + L θ ) × 1 0 1 × L θ 0 L θ × L θ 0 L θ × L θ 0 ( L x + L θ ) × L θ 0 1 × L θ 0 L θ × L θ 0 L θ × L θ 0 ( L x + L θ ) × L θ ( coeff ( f ) , 0 1 × L θ ) 0 L θ × ( L x + L θ ) 0 L θ × ( L x + L θ ) 0 ( L x + L θ ) × ( L x + L θ ) ,
where as defined above coeff ( f ) \operatorname{coeff}(f) coeff ( f ) is a row vector with dimension L x L_x L x . Observe that the map ι \iota ι is injective, since the coefficient vector determines the polynomial.
We need to encode v ( 0 ) ρ t C U = coeff ( U ( t , x ) ) v(\mathbf{0})\rho_{\mathbf{t}}C_U=\operatorname{coeff}(U(\mathbf{t},\mathbf{x})) v ( 0 ) ρ t C U = coeff ( U ( t , x )) given in (3) in this higher dimensional space, where we represent these operations by square matrices of dimension L t o t L_{\mathrm{tot}} L tot , with the following correspondence:
v ( 0 ) ρ t C U ↓ ↓ ↓ V 0 Γ t C \begin{array}{ccccc}
v(0)&\rho_t&C_U\\
\downarrow&\downarrow&\downarrow\\
V_0&\Gamma_t&C
\end{array} v ( 0 ) ↓ V 0 ρ t ↓ Γ t C U ↓ C
where V 0 V_0 V 0 contains the initial row v ( 0 ) v(0) v ( 0 ) , Γ t \Gamma_t Γ t contains the translation matrix, and C C C contains the coefficient matrix C U C_U C U . Using the same four block sizes as for ι ( f ) \iota(f) ι ( f ) , we define Γ t : = diag ( 1 , ρ t , ρ t , I L x + L θ ) \Gamma_t:=\operatorname{diag}(1,\rho_t,\rho_t,I_{L_x+L_\theta}) Γ t := diag ( 1 , ρ t , ρ t , I L x + L θ ) , and
V 0 : = ( 0 v ( 0 ) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ) and C : = ( 0 0 0 0 0 0 0 ( C U , 0 ) 0 0 0 ( 0 , I L θ ) 0 0 0 0 ) . V_0:=\begin{pmatrix}
0&v(0)&0&0\\
0&0&0&0\\
0&0&0&0\\
0&0&0&0
\end{pmatrix}
\qquad\text{and}\qquad
C:=\begin{pmatrix}
0&0&0&0\\
0&0&0&(C_U,0)\\
0&0&0&(0,I_{L_\theta})\\
0&0&0&0
\end{pmatrix}. V 0 := 0 0 0 0 v ( 0 ) 0 0 0 0 0 0 0 0 0 0 0 and C := 0 0 0 0 0 0 0 0 0 0 0 0 0 ( C U , 0 ) ( 0 , I L θ ) 0 .
It is straightforward to verify that multiplying V 0 V_0 V 0 by Γ t \Gamma_t Γ t replaces its nonzero row block by v ( 0 ) ρ t = v ( t ) v(0)\rho_t=v(t) v ( 0 ) ρ t = v ( t ) . Then multiplication by C C C places ( v ( t ) C U , 0 ) = ( coeff ( U ( t , x ) ) , 0 ) (v(t)C_U,0)=(\operatorname{coeff}(U(t,x)),0) ( v ( t ) C U , 0 ) = ( coeff ( U ( t , x )) , 0 ) in the top-right block. All other blocks are zero. Consequently,
V 0 Γ t C = ι ( U ( t , x ) ) . (4) V_0\Gamma_tC=\iota(U(t,x)). \tag*{(4)} V 0 Γ t C = ι ( U ( t , x )) . (4)
In particular, since U ( 0 , x ) = 0 U(0,x)=0 U ( 0 , x ) = 0 , we have V 0 C = 0 V_0C=0 V 0 C = 0 ; see Figure 6 for an illustration.
Figure 6. The ABP from Figure 5 with t = ( 2 , 3 ) t=(2,3) t = ( 2 , 3 ) is encoded by square matrices. Here L θ = L x = 4 L_\theta=L_x=4 L θ = L x = 4 , so the four block sizes are 1 , 4 , 4 , 8 1,4,4,8 1 , 4 , 4 , 8 and the matrices have dimension 17 17 17 . The shaded blocks contain the rows displayed above, with four trailing zeros in the final block; all unshaded blocks are zero. Thus V 0 Γ t C = ι ( 6 x 3 + 2 x 2 y ) V_0\Gamma_tC=\iota(6x^3+2x^2y) V 0 Γ t C = ι ( 6 x 3 + 2 x 2 y ) .
We now use these matrices to construct a finitely generated semigroup. The translation matrices form a group, since
Γ 0 = I L t o t , Γ s Γ t = Γ s + t , Γ t − 1 = Γ − t . \Gamma_0=I_{L_{\mathrm{tot}}},\qquad\Gamma_s\Gamma_t=\Gamma_{s+t},\qquad\Gamma_t^{-1}=\Gamma_{-t}. Γ 0 = I L tot , Γ s Γ t = Γ s + t , Γ t − 1 = Γ − t .
To obtain finitely many generators, we take the translations by the standard basis vectors and their negatives, together with V 0 V_0 V 0 and C C C . Thus, writing e 1 , … , e ν e_1,\ldots,e_\nu e 1 , … , e ν for the standard basis vectors, define
X : = { Γ e i , Γ − e i : 1 ≤ i ≤ ν } ∪ { V 0 , C } . X:=\{\Gamma_{e_i},\Gamma_{-e_i}:1\leq i\leq\nu\}\cup\{V_0,C\}. X := { Γ e i , Γ − e i : 1 ≤ i ≤ ν } ∪ { V 0 , C } .
All these matrices belong to Mat L t o t ( Q ) \operatorname{Mat}_{L_{\mathrm{tot}}}(\mathbb{Q}) Mat L tot ( Q ) . Indeed, C U C_U C U is rational, and the translation matrices are rational at rational parameter values. Moreover, { Γ e i , Γ − e i : 1 ≤ i ≤ ν } \{\Gamma_{e_i},\Gamma_{-e_i}:1\leq i\leq\nu\} { Γ e i , Γ − e i : 1 ≤ i ≤ ν } give Γ n \Gamma_n Γ n for every n ∈ Z n\in\mathbb{Z} n ∈ Z .
The non-zero blocks of V 0 V_0 V 0 and C C C restrict wrong encodings, as
Γ t V 0 = V 0 and C Γ t = C and V 0 2 = C 2 = C V 0 = 0. \Gamma_tV_0=V_0 \qquad\text{and} \qquad C\Gamma_t=C \qquad\text{and} \qquad V_0^2=C^2=CV_0=0. Γ t V 0 = V 0 and C Γ t = C and V 0 2 = C 2 = C V 0 = 0.
Together with Equation (4) , these identities show that every element of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ has one of the forms
Γ n and V 0 Γ n and Γ n C and ι ( U ( n , x ) ) , \Gamma_n \qquad\text{and} \qquad V_0\Gamma_n \qquad\text{and} \qquad\Gamma_nC \qquad\text{and} \qquad\iota(U(n,x)), Γ n and V 0 Γ n and Γ n C and ι ( U ( n , x )) ,
with n ∈ Z ν n\in\mathbb{Z}^{\nu} n ∈ Z ν . Conversely, every matrix in this list belongs to ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . We now take the Zariski closure of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ in Mat L t o t ( Q ‾ ) \operatorname{Mat}_{L_{\mathrm{tot}}}(\overline{\mathbb{Q}}) Mat L tot ( Q ) . The map t ↦ ι ( U ( t , x ) ) t\mapsto\iota(U(t,x)) t ↦ ι ( U ( t , x )) is polynomial, and Z ν \mathbb{Z}^{\nu} Z ν is Zariski dense in Q ‾ ν \overline{\mathbb{Q}}^{\nu} Q ν . Consequently,
{ ι ( U ( n , x ) ) : n ∈ Z ν } ‾ = ι ( A w , d ‾ ) = ι ( A w , d ‾ ) , \overline{\{\iota(U(n,x)):n\in\mathbb{Z}^{\nu}\}}=\iota(\overline{\mathcal{A}_{w,d}})=\iota(\overline{\mathcal{A}_{w,d}}), { ι ( U ( n , x )) : n ∈ Z ν } = ι ( A w , d ) = ι ( A w , d ) ,
where the last equality uses that ι \iota ι is a linear embedding. The proof of the following claim is in Appendix F.
Claim 4.1. ⟨ X ⟩ ‾ = { Γ t , V 0 Γ t , Γ t C : t ∈ Q ‾ ν } ∪ ˙ ι ( A w , d ‾ ) \overline{\langle X\rangle}=\{\Gamma_t,V_0\Gamma_t,\Gamma_tC:t\in\overline{\mathbb{Q}}^{\nu}\}\mathbin{\dot{\cup}}\iota(\overline{\mathcal{A}_{w,d}}) ⟨ X ⟩ = { Γ t , V 0 Γ t , Γ t C : t ∈ Q ν } ∪ ˙ ι ( A w , d ) .
Informally speaking, each of these families { Γ t , V 0 Γ t , Γ t C : t ∈ Q ‾ ν } \{\Gamma_t,V_0\Gamma_t,\Gamma_tC:t\in\overline{\mathbb{Q}}^{\nu}\} { Γ t , V 0 Γ t , Γ t C : t ∈ Q ν } is already Zariski closed. For example, in the case of Γ t \Gamma_t Γ t and V 0 Γ t V_0\Gamma_t V 0 Γ t , the parameter values can be recovered from matrix entries, and all remaining entries are polynomial functions of them.
We are now in a position to define the congruence ∼ \sim ∼ . Our aim is to identify translations with the identity, so that the parameter assignment disappears when passing to the quotient. Let ∼ \sim ∼ be the least semigroup congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ , with respect to inclusion, satisfying Γ t ∼ I L t o t \Gamma_t\sim I_{L_{\mathrm{tot}}} Γ t ∼ I L tot for every t ∈ Q ‾ ν t\in\overline{\mathbb{Q}}^{\nu} t ∈ Q ν . The compatibility of ∼ \sim ∼ with multiplication gives
ι ( U ( t , x ) ) = V 0 Γ t C ∼ V 0 C = 0 \iota(U(t,x))=V_0\Gamma_tC\sim V_0C=0 ι ( U ( t , x )) = V 0 Γ t C ∼ V 0 C = 0
implying that the encoding of every computable polynomial with ABPs belongs to the zero class.
Claim 4.2. The class of zero matrix consists of the encoding of polynomials computable by ABPs, that is,
[ 0 ] ∼ = ι ( A w , d ) . [0]_{\sim}=\iota(\mathcal{A}_{w,d}). [ 0 ] ∼ = ι ( A w , d ) .
We next count the classes that meet ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ .
Claim 4.3. The quotient ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ has exactly four classes, represented by I L t o t I_{L_{\mathrm{tot}}} I L tot , V 0 V_0 V 0 , C C C , and 0 0 0 .
This finiteness here follows from Theorem 3. Indeed, every group in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ is either contained in the translation group or consists only of 0 \mathbf{0} 0 . By the definition of ∼ \sim ∼ , each such group has a singleton quotient, so the theorem applies with N = 1 N=1 N = 1 . Our particular construction is also such that the finiteness and the count on the classes also directly follows from the definition and the multiplication rules.
We now turn to the separation of A w , d \mathcal{A}_{w,d} A w , d and A ‾ w , d \overline{\mathcal{A}}_{w,d} A w , d . Claim 4.2 says that a computable polynomial with ABPs belongs to [ 0 ] ∼ [0]_{\sim} [ 0 ] ∼ ; below, we construct a polynomial in A ‾ w , d \overline{\mathcal{A}}_{w,d} A w , d whose encoding does not belong to this class. (We show an example in Figure 7 showing the underlying phenomenon, following Figure 5 and Figure 6 .)
Figure 7. Closure and congruence for U ( θ 1 , θ 2 , x , y ) = θ 1 θ 2 x 3 + θ 1 x 2 y U(\theta_1,\theta_2,x,y)=\theta_1\theta_2x^3+\theta_1x^2y U ( θ 1 , θ 2 , x , y ) = θ 1 θ 2 x 3 + θ 1 x 2 y given in Figure 5 and Figure 6 . For ε ≠ 0 \varepsilon\ne0 ε = 0 , the parameter assignment ( θ 1 , θ 2 ) = ( ε , ε − 1 ) (\theta_1,\theta_2)=(\varepsilon,\varepsilon^{-1}) ( θ 1 , θ 2 ) = ( ε , ε − 1 ) gives x 3 + ε x 2 y x^3+\varepsilon x^2y x 3 + ε x 2 y . In the diagram, we see that the pair ( 1 , ε ) (1,\varepsilon) ( 1 , ε ) approaching ( 1 , 0 ) (1,0) ( 1 , 0 ) . The point ( 1 , ε ) (1,\varepsilon) ( 1 , ε ) represents an output of the family, while the point ( 1 , 0 ) (1,0) ( 1 , 0 ) is present only in its closure. Recall that we construct the congruence ∼ e x \sim_{\mathrm{ex}} ∼ ex to be the least congruence on the closure of the matrix semigroup constructed from this family satisfying Γ t ∼ e x I 17 \Gamma_t\sim_{\mathrm{ex}}I_{17} Γ t ∼ ex I 17 for every t ∈ Q ‾ 2 t\in\overline{\mathbb{Q}}^2 t ∈ Q 2 . The zero class consists exactly of the encoding of polynomials computable by this two-parameter family of ABPs. Indeed, ι ( x 3 + ε x 2 y ) = V 0 Γ ( ε , ε − 1 ) C ∼ e x V 0 C = 0 \iota(x^3+\varepsilon x^2y)=V_0\Gamma_{(\varepsilon,\varepsilon^{-1})}C\sim_{\mathrm{ex}}V_0C=\mathbf{0} ι ( x 3 + ε x 2 y ) = V 0 Γ ( ε , ε − 1 ) C ∼ ex V 0 C = 0 . At ε = 0 \varepsilon=0 ε = 0 , this encoded output specializes to ι ( x 3 ) \iota(x^3) ι ( x 3 ) , which belongs to the matrix closure but lies outside the zero class (producing x 3 x^3 x 3 would require both θ 1 = 0 \theta_1=0 θ 1 = 0 and θ 1 θ 2 = 1 \theta_1\theta_2=1 θ 1 θ 2 = 1 ). This is due to the fact that the zero class need not be Zariski closed.
For simplicity of construction, we rename the input variables by x , y , a , b , u , v , z , p , q , r , s x,y,a,b,u,v,z,p,q,r,s x , y , a , b , u , v , z , p , q , r , s and c i , z i c_i,z_i c i , z i with 1 ≤ i ≤ w − 2 1\le i\le w-2 1 ≤ i ≤ w − 2 (for w = 2 w=2 w = 2 , there is no c i c_i c i and z i z_i z i ). We first define the polynomials for width two as follows, for all k ≥ 0 k\ge0 k ≥ 0 ,
F 2 , 2 k + 3 : = a ξ k 2 + b η k 2 + z Δ k ( u x + v y ) , F_{2,2k+3}:=a\xi_k^2+b\eta_k^2+z\Delta^k(ux+vy), F 2 , 2 k + 3 := a ξ k 2 + b η k 2 + z Δ k ( ux + v y ) ,
F 2 , 2 k + 4 : = a ξ k + 1 ξ k + b η k + 1 η k + Δ k + 1 ( u x + v y ) , k ≥ 0 F_{2,2k+4}:=a\xi_{k+1}\xi_k+b\eta_{k+1}\eta_k+\Delta^{k+1}(ux+vy),\qquad k\ge0 F 2 , 2 k + 4 := a ξ k + 1 ξ k + b η k + 1 η k + Δ k + 1 ( ux + v y ) , k ≥ 0
where Δ : = p s − q r \Delta:=ps-qr Δ := p s − q r and
( ξ k η k ) : = G k ( x y ) with G : = ( p q r s ) . \begin{pmatrix}\xi_k\\\eta_k\end{pmatrix}:=G^k\begin{pmatrix}x\\y\end{pmatrix}
\qquad\text{with}\qquad
G:=\begin{pmatrix}p&q\\r&s\end{pmatrix}. ( ξ k η k ) := G k ( x y ) with G := ( p r q s ) .
For general width w w w , define
F w , d : = F 2 , d + ∑ i = 1 w − 2 c i z i d − 1 . (5) F_{w,d}:=F_{2,d}+\sum_{i=1}^{w-2}c_i z_i^{d-1}. \tag*{(5)} F w , d := F 2 , d + i = 1 ∑ w − 2 c i z i d − 1 . (5)
Observe that polynomials F w , d F_{w,d} F w , d are homogeneous of degree d d d and have integer coefficients. For example,
F w , 3 = a x 2 + b y 2 + z ( u x + v y ) + ∑ i = 1 w − 2 c i z i 2 . F_{w,3}=ax^2+by^2+z(ux+vy)+\sum_{i=1}^{w-2}c_i z_i^2. F w , 3 = a x 2 + b y 2 + z ( ux + v y ) + i = 1 ∑ w − 2 c i z i 2 .
We first prove that these polynomials belong to the closure A w , d \mathcal{A}_{w,d} A w , d .
Claim 4.4. F w , d ∈ A ‾ w , d F_{w,d} \in\overline{\mathcal{A}}_{w,d} F w , d ∈ A w , d .
The main idea in proving Claim 4.4 is to give an ABP depending on a nonzero parameter ε \varepsilon ε , which specializes to F w , d F_{w,d} F w , d when ε = 0 \varepsilon= 0 ε = 0 . Below, we consider width 2 2 2 . For the odd degree d = 2 k + 3 d = 2k + 3 d = 2 k + 3 , consider the matrix product
( x + ε v , y − ε u ) ( G T ) k ( a z / ε − z / ε b ) G k ( x y ) (x+\varepsilon v,\ y-\varepsilon u)(G^{\mathsf{T}})^k
\begin{pmatrix}
a & z/\varepsilon\\
-z/\varepsilon& b
\end{pmatrix}
G^k
\begin{pmatrix}
x\\
y
\end{pmatrix} ( x + ε v , y − ε u ) ( G T ) k ( a − z / ε z / ε b ) G k ( x y ) and for even degree d = 2 k + 4 d = 2k + 4 d = 2 k + 4 , use
( x + ε v , y − ε u ) ( G T ) k + 1 ( a + r / ε s / ε − p / ε b − q / ε ) G k ( x y ) . (x+\varepsilon v,\ y-\varepsilon u)(G^{\mathsf{T}})^{k+1}
\begin{pmatrix}
a+r/\varepsilon& s/\varepsilon\\
-p/\varepsilon& b-q/\varepsilon
\end{pmatrix}
G^k
\begin{pmatrix}
x\\
y
\end{pmatrix}. ( x + ε v , y − ε u ) ( G T ) k + 1 ( a + r / ε − p / ε s / ε b − q / ε ) G k ( x y ) . Observe that these two given products define width-2 2 2 ABPs with degree d d d whenever ε ≠ 0 \varepsilon\ne0 ε = 0 ; moreover, the result of the products have the form F 2 , d + ε H 2 , d F_{2,d}+\varepsilon H_{2,d} F 2 , d + ε H 2 , d , where H 2 , d H_{2,d} H 2 , d is homogeneous of degree d d d and independent of ε \varepsilon ε . For example, at degree 3 3 3 we have
( x + ε v , y − ε u ) ( a z / ε − z / ε b ) ( x y ) = F 2 , 3 + ε ( a v x − b u y ) . (x+\varepsilon v,\ y-\varepsilon u)
\begin{pmatrix}
a & z/\varepsilon\\
-z/\varepsilon& b
\end{pmatrix}
\begin{pmatrix}
x\\
y
\end{pmatrix}
= F_{2,3}+\varepsilon(avx-buy). ( x + ε v , y − ε u ) ( a − z / ε z / ε b ) ( x y ) = F 2 , 3 + ε ( a v x − b u y ) . Consequently, F 2 , d F_{2,d} F 2 , d belongs to A 2 , d \mathcal{A}_{2,d} A 2 , d , proving Claim 4.4 for w = 2 w=2 w = 2 ; see Appendix F for details and discussion for other widths.
We have shown that F w , d ∈ A ‾ w , d F_{w,d} \in\overline{\mathcal{A}}_{w,d} F w , d ∈ A w , d . It remains to prove that F w , d ∉ A w , d F_{w,d} \notin\mathcal{A}_{w,d} F w , d ∈ / A w , d , for which we will use Claim 4.2 about the zero congruent class.
Claim 4.5. ι ( F w , d ) ≁ 0 \iota(F_{w,d}) \not\sim0 ι ( F w , d ) ∼ 0 .
Suppose, towards a contradiction, that ι ( F w , d ) ∼ 0 \iota(F_{w,d}) \sim0 ι ( F w , d ) ∼ 0 . By Claim 4.2, F w , d F_{w,d} F w , d is computed by an ABP of width at most w w w and degree d d d . We examine the first and last layers of this ABP, where each consists of at most w w w homogeneous linear forms in x \mathbf{x} x . The polynomial F w , d F_{w,d} F w , d vanishes whenever all such forms in either layer vanish. The main algebraic step is to show that each of these forms is a linear combination of variables x x x , y y y and c i c_i c i , z i z_i z i with 1 ≤ i ≤ w − 2 1 \le i \le w-2 1 ≤ i ≤ w − 2 . This follows from a property of the linear subspaces on which F w , d F_{w,d} F w , d vanishes, proved in Appendix F. Since every term produced in the ABP contains a form from the first layer and a form from the last layer, every monomial must contain at least two occurrences of the variables x x x , y y y and c i c_i c i , z i z_i z i . However, the terms involving u x + v y ux+vy ux + v y in F w , d F_{w,d} F w , d contain exactly one of these variables. They are nonzero and cannot be cancelled by the other summands, a contradiction.
We conclude that ι ( F w , d ) ≁ 0 \iota(F_{w,d}) \not\sim0 ι ( F w , d ) ∼ 0 , proving Claim 4.5. By Claim 4.2, this implies F w , d ∉ A w , d F_{w,d} \notin\mathcal{A}_{w,d} F w , d ∈ / A w , d . Together with Claim 4.4, we obtain
F w , d ∈ A ‾ w , d ∖ A w , d , F_{w,d} \in\overline{\mathcal{A}}_{w,d} \setminus\mathcal{A}_{w,d}, F w , d ∈ A w , d ∖ A w , d , which completes the proof of Theorem 4.
Separation of Det n \operatorname{Det}_{n} Det n from Perm n \operatorname{Perm}_{n} Perm n under a subclass of reductions
Let n ≥ 3 n \ge3 n ≥ 3 . Consider the matrix
[ x i , j ] 1 ≤ i , j ≤ n = ( x 1 , 1 ⋯ x 1 , n ⋮ ⋱ ⋮ x n , 1 ⋯ x n , n ) . [x_{i,j}]_{1\le i,j\le n} =
\begin{pmatrix}
x_{1,1} & \cdots& x_{1,n}\\
\vdots& \ddots& \vdots\\
x_{n,1} & \cdots& x_{n,n}
\end{pmatrix}. [ x i , j ] 1 ≤ i , j ≤ n = x 1 , 1 ⋮ x n , 1 ⋯ ⋱ ⋯ x 1 , n ⋮ x n , n .
whose entries are the formal variables x i , j x_{i,j} x i , j . We write x i : = ( x i , 1 , … , x i , n ) \boldsymbol{x}_i := (x_{i,1},\ldots,x_{i,n}) x i := ( x i , 1 , … , x i , n ) for the i i i -th row of this matrix, and let X : = ( x 1 , … , x n ) X := (\boldsymbol{x}_1,\ldots,\boldsymbol{x}_n) X := ( x 1 , … , x n ) be the vector of its n 2 n^2 n 2 variables. We denote by Det n \operatorname{Det}_n Det n and Perm n \operatorname{Perm}_n Perm n the determinant and permanent polynomials, respectively, defined by
Det n ( X ) : = ∑ σ ∈ S n sgn ( σ ) ∏ i = 1 n x i , σ ( i ) and Perm n ( X ) : = ∑ σ ∈ S n ∏ i = 1 n x i , σ ( i ) . \operatorname{Det}_n(X) := \sum_{\sigma\in\mathcal{S}_n} \operatorname{sgn}(\sigma)\prod_{i=1}^{n} x_{i,\sigma(i)}
\qquad\text{and}\qquad
\operatorname{Perm}_n(X) := \sum_{\sigma\in\mathcal{S}_n}\prod_{i=1}^{n} x_{i,\sigma(i)}. Det n ( X ) := σ ∈ S n ∑ sgn ( σ ) i = 1 ∏ n x i , σ ( i ) and Perm n ( X ) := σ ∈ S n ∑ i = 1 ∏ n x i , σ ( i ) .
where S n \mathcal{S}_n S n is the set of permutations of { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } and sgn ( σ ) \operatorname{sgn}(\sigma) sgn ( σ ) denotes the sign of σ \sigma σ .
Below, we use the standard action of Mat n 2 ( Q ‾ ) \operatorname{Mat}_{n^2}(\overline{\mathbb{Q}}) Mat n 2 ( Q ) on the polynomial ring Q ‾ [ X ] \overline{\mathbb{Q}}[X] Q [ X ] , defined by
( h ◃ L ) ( x ) : = h ( x L ) . (h\triangleleft L)(x) := h(xL). ( h ◃ L ) ( x ) := h ( xL ) .
where L ∈ Mat n 2 ( Q ‾ ) L\in\operatorname{Mat}_{n^2}(\overline{\mathbb{Q}}) L ∈ Mat n 2 ( Q ) and h ∈ Q ‾ [ X ] h\in\overline{\mathbb{Q}}[X] h ∈ Q [ X ] . This is the action of linear substitutions in the n 2 n^2 n 2 variables; as defined in [8 Section 3.2] . It satisfies
h ◃ ( L 1 L 2 ) = ( h ◃ L 1 ) ◃ L 2 and h ◃ I n 2 = h . h\triangleleft(L_1L_2)=(h\triangleleft L_1)\triangleleft L_2
\qquad\text{and}\qquad
h\triangleleft I_{n^2}=h. h ◃ ( L 1 L 2 ) = ( h ◃ L 1 ) ◃ L 2 and h ◃ I n 2 = h .
where I n 2 I_{n^2} I n 2 is the identity matrix of dimension n 2 n^2 n 2 . Curiously, under the congruence of mutual reducibility, the determinant orbits closure has a finite quotient whereas that of permanent is not finite.
We construct a finitely generated monoid ⟨ Y n ⟩ ≤ Mat n 2 ( Q ‾ ) \langle Y_n\rangle\le\operatorname{Mat}_{n^2}(\overline{\mathbb{Q}}) ⟨ Y n ⟩ ≤ Mat n 2 ( Q ) such that
Perm n ◃ ⟨ Y n ⟩ ‾ ⊈ Det n ◃ ⟨ Y n ⟩ ‾ . \overline{\operatorname{Perm}_n\triangleleft\langle Y_n\rangle}\not\subseteq\overline{\operatorname{Det}_n\triangleleft\langle Y_n\rangle}. Perm n ◃ ⟨ Y n ⟩ ⊆ Det n ◃ ⟨ Y n ⟩ .
Notably, under the congruence of mutual reducibility, the determinant orbits closure has a finite quotient whereas that of permanent is not finite. For simplicity, to construct the finite set Y n ⊆ Mat n 2 ( Q ‾ ) Y_n\subseteq\operatorname{Mat}_{n^2}(\overline{\mathbb{Q}}) Y n ⊆ Mat n 2 ( Q ) , we first define X n ⊆ Mat n ( Q ) X_n\subseteq\operatorname{Mat}_n(\mathbb{Q}) X n ⊆ Mat n ( Q ) and use the embedding
ȷ ( M ) : = diag ( M , … , M ⏟ n b l o c k s ) ∈ Mat n 2 ( Q ‾ ) . \jmath(M):=\operatorname{diag}(\underbrace{M,\ldots,M}_{n\ \mathrm{blocks}})\in\operatorname{Mat}_{n^2}(\overline{\mathbb{Q}}). ( M ) := diag ( n blocks M , … , M ) ∈ Mat n 2 ( Q ) .
Then X ȷ ( M ) = ( x 1 M , … , x n M ) X_{\jmath(M)}=(\boldsymbol{x}_1M,\ldots,\boldsymbol{x}_nM) X ( M ) = ( x 1 M , … , x n M ) . Thus ȷ ( M ) \jmath(M) ( M ) acts on every row x i \boldsymbol{x}_i x i of the matrix [ x i , j ] 1 ≤ i , j ≤ n [x_{i,j}]_{1\le i,j\le n} [ x i , j ] 1 ≤ i , j ≤ n of variables, and
Det n ◃ ȷ ( M ) = det ( x 1 M ⋮ x n M ) and Perm n ◃ ȷ ( M ) = per ( x 1 M ⋮ x n M ) . \operatorname{Det}_n\triangleleft\jmath(M)=\det\begin{pmatrix}
\boldsymbol{x}_1M\\
\vdots\\
\boldsymbol{x}_nM
\end{pmatrix}
\qquad\text{and}\qquad
\operatorname{Perm}_n\triangleleft\jmath(M)=\operatorname{per}\begin{pmatrix}
\boldsymbol{x}_1M\\
\vdots\\
\boldsymbol{x}_nM
\end{pmatrix}. Det n ◃ ( M ) = det x 1 M ⋮ x n M and Perm n ◃ ( M ) = per x 1 M ⋮ x n M .
Let us first introduce the generators X n X_n X n . Recall that e j e_j e j denote the j j j -th standard basis vector of Q n \mathbb{Q}^n Q n , and let 1 \mathbf{1} 1 be the row vector with all entries being ones. For 2 ≤ j ≤ n 2\le j\le n 2 ≤ j ≤ n , define
R j : = ( e j − e 1 ) 1 = ( − 1 0 ( j − 2 ) × n 1 0 ( n − j ) × n ) , R_j:=(e_j-e_1)\mathbf{1}=
\begin{pmatrix}
-1\\
\mathbf{0}_{(j-2)\times n}\\
\mathbf{1}\\
\mathbf{0}_{(n-j)\times n}
\end{pmatrix}, R j := ( e j − e 1 ) 1 = − 1 0 ( j − 2 ) × n 1 0 ( n − j ) × n ,
where the first row of R j R_j R j consists entirely of − 1 -1 − 1 , the j j j -th row consists entirely of 1 1 1 , and zero otherwise. Since 1 ( e j − e 1 ) = 0 \mathbf{1}(e_j-e_1)=0 1 ( e j − e 1 ) = 0 , we have R i R j = 0 n × n R_iR_j=0_{n\times n} R i R j = 0 n × n for all 2 ≤ i , j ≤ n 2\le i,j\le n 2 ≤ i , j ≤ n . Consider the finite set
X n : = { I n , 2 I n } ∪ { I n + R j , R j : 2 ≤ j ≤ n } ⊆ Mat n ( Q ) . X_n:=\{I_n,2I_n\}\cup\{I_n+R_j,R_j:2\le j\le n\}\subseteq\operatorname{Mat}_n(\mathbb{Q}). X n := { I n , 2 I n } ∪ { I n + R j , R j : 2 ≤ j ≤ n } ⊆ Mat n ( Q ) .
From this set, as mentioned above, we define Y n : = { ȷ ( M ) : M ∈ X n } Y_n:=\{\jmath(M):M\in X_n\} Y n := { ( M ) : M ∈ X n } . Observe that elements of X n X_n X n and thus those of Y n Y_n Y n commute. Moreover, ȷ \jmath preserves multiplication, so ⟨ Y n ⟩ = { ȷ ( M ) : M ∈ ⟨ X n ⟩ } \langle Y_n\rangle=\{\jmath(M):M\in\langle X_n\rangle\} ⟨ Y n ⟩ = { ( M ) : M ∈ ⟨ X n ⟩} . A simple computation shows that
⟨ X n ⟩ = { 2 k ( I n + ∑ j = 2 n m j R j ) : k , m j ∈ N } ∪ { 2 k R j : k ∈ N , 2 ≤ j ≤ n } ∪ { 0 } . (6) \langle X_n\rangle=\left\{2^k\left(I_n+\sum_{j=2}^{n}m_jR_j\right):k,m_j\in\mathbb{N}\right\}\cup\{2^kR_j:k\in\mathbb{N},\ 2\le j\le n\}\cup\{0\}. \tag*{(6)} ⟨ X n ⟩ = { 2 k ( I n + j = 2 ∑ n m j R j ) : k , m j ∈ N } ∪ { 2 k R j : k ∈ N , 2 ≤ j ≤ n } ∪ { 0 } . (6)
Taking the Zariski closure in Mat n ( Q ‾ ) \operatorname{Mat}_n(\overline{\mathbb{Q}}) Mat n ( Q ) , we obtain
⟨ X n ⟩ ‾ = { t I n + ∑ j = 2 n u j R j : t , u j ∈ Q ‾ } . \overline{\langle X_n\rangle}=\left\{tI_n+\sum_{j=2}^{n}u_jR_j:t,u_j\in\overline{\mathbb{Q}}\right\}. ⟨ X n ⟩ = { t I n + j = 2 ∑ n u j R j : t , u j ∈ Q } .
But then since ȷ \jmath is a linear embedding ⟨ Y n ⟩ ‾ = { ȷ ( M ) : M ∈ ⟨ X n ⟩ ‾ } \overline{\langle Y_n\rangle}=\{\jmath(M):M\in\overline{\langle X_n\rangle}\} ⟨ Y n ⟩ = { ( M ) : M ∈ ⟨ X n ⟩ } holds.
We work with the following congruence ∼ \sim ∼ . Given M , N ∈ ⟨ X n ⟩ ‾ M,N\in\overline{\langle X_n\rangle} M , N ∈ ⟨ X n ⟩ , we define M ∼ N M\sim N M ∼ N if and only if M = N U M=NU M = N U and N = M V N=MV N = M V for some U , V ∈ ⟨ X n ⟩ ‾ U,V\in\overline{\langle X_n\rangle} U , V ∈ ⟨ X n ⟩ . We again note in passing that this is a congruence thanks to commutativity of ⟨ X n ⟩ ‾ \overline{\langle X_n\rangle} ⟨ X n ⟩ . We extend it to the image of ȷ \jmath , in the natural way, by defining ȷ ( M ) ∼ ȷ ( N ) \jmath(M)\sim\jmath(N) ( M ) ∼ ( N ) if and only if M ∼ N M\sim N M ∼ N . Consequently, we can conclude that
⟨ Y n ⟩ ‾ / ∼ ≅ ⟨ X n ⟩ ‾ / ∼ and ⟨ Y n ‾ ⟩ / ∼ ≅ ⟨ X n ⟩ ‾ / ∼ . \overline{\langle Y_n\rangle}/\sim\cong\overline{\langle X_n\rangle}/\sim
\qquad\text{and}\qquad
\langle\overline{Y_n}\rangle/\sim\cong\overline{\langle X_n\rangle}/\sim. ⟨ Y n ⟩ / ∼≅ ⟨ X n ⟩ / ∼ and ⟨ Y n ⟩ / ∼≅ ⟨ X n ⟩ / ∼ .
Observe that every group in ⟨ X n ⟩ ‾ \overline{\langle X_n\rangle} ⟨ X n ⟩ lies in one congruence class, as the inverses provide the multipliers required in both directions. Theorem 3 therefore applies with N = 1 N=1 N = 1 and gives finiteness of ⟨ X n ⟩ ‾ / ∼ \overline{\langle X_n\rangle}/\sim ⟨ X n ⟩ / ∼ , and hence of ⟨ Y n ⟩ ‾ / ∼ \overline{\langle Y_n\rangle}/\sim ⟨ Y n ⟩ / ∼ .
In the sequel, we consider the notion of mutual reducibility ≈ \approx ≈ among polynomials. Formally, for all f , g ∈ Q ‾ [ X ] f,g\in\overline{\mathbb{Q}}[X] f , g ∈ Q [ X ] , write f ≈ g f\approx g f ≈ g if and only if f = g ◃ M f=g\triangleleft M f = g ◃ M and g = f ◃ N g=f\triangleleft N g = f ◃ N for some M , N ∈ ⟨ Y n ⟩ ‾ M,N\in\overline{\langle Y_n\rangle} M , N ∈ ⟨ Y n ⟩ . This relation is an equivalence relation on Q ‾ [ X ] \overline{\mathbb{Q}}[X] Q [ X ] . Crucially, definitions given above imply that M ∼ N M\sim N M ∼ N if and only if
f ◃ ȷ ( M ) ≈ f ◃ ȷ ( N ) f\triangleleft\jmath(M)\approx f\triangleleft\jmath(N) f ◃ ( M ) ≈ f ◃ ( N )
holds for all f ∈ Q ‾ [ X ] f\in\overline{\mathbb{Q}}[X] f ∈ Q [ X ] . Therefore, finiteness of the matrix quotient implies that f ◃ ⟨ Y n ⟩ f\triangleleft\langle Y_n\rangle f ◃ ⟨ Y n ⟩ has only finitely many mutual-reducibility classes, since
∣ ( f ◃ ⟨ Y n ⟩ ) / ≈ ∣ ≤ ∣ ⟨ X n ⟩ / ∼ ∣ < ∞ . |(f\triangleleft\langle Y_n\rangle)/\approx|\leq|\langle X_n\rangle/\sim|<\infty. ∣ ( f ◃ ⟨ Y n ⟩) / ≈ ∣ ≤ ∣ ⟨ X n ⟩ / ∼ ∣ < ∞.
It is worth noting that the converse implication need not hold, as different matrix classes may become indistinguishable when acting on a particular polynomial. This is where the determinant Det n \operatorname{Det}_n Det n and permanent Perm n \operatorname{Perm}_n Perm n will differ. Consider the orbits of these two polynomials D n : = Det n ◃ ⟨ Y n ⟩ D_n:=\operatorname{Det}_n\triangleleft\langle Y_n\rangle D n := Det n ◃ ⟨ Y n ⟩ and P n : = Perm n ◃ ⟨ Y n ⟩ P_n:=\operatorname{Perm}_n\triangleleft\langle Y_n\rangle P n := Perm n ◃ ⟨ Y n ⟩ , and take the closures in the coefficient space of homogeneous degree-n n n polynomials over Q ‾ \overline{\mathbb{Q}} Q . Then
For the determinant, observe that det ( M ) = t n \det(M)=t^n det ( M ) = t n for matrices M = t I n + ∑ j = 2 n u j R j M=tI_n+\sum_{j=2}^{n}u_jR_j M = t I n + ∑ j = 2 n u j R j in the closure ⟨ X n ⟩ ‾ \overline{\langle X_n\rangle} ⟨ X n ⟩ . Hence, Det n ◃ ȷ ( M ) = t n Det n \operatorname{Det}_n\triangleleft\jmath(M)=t^n\operatorname{Det}_n Det n ◃ ( M ) = t n Det n . Using (6) above, we get D n = { 2 k n Det n : k ∈ N } ∪ { 0 } D_n=\{2^{kn}\operatorname{Det}_n:k\in\mathbb{N}\}\cup\{0\} D n = { 2 k n Det n : k ∈ N } ∪ { 0 } , which simply gives D n ‾ = Q ‾ Det n \overline{D_n}=\overline{\mathbb{Q}}\operatorname{Det}_n D n = Q Det n . This in turn proves that D n ‾ \overline{D_n} D n and its closure D n ‾ \overline{D_n} D n have exactly two mutual-reducibility classes, as (1) all nonzero multiples of Det n \operatorname{Det}_n Det n are mutually reducible using scalar matrices from ⟨ Y n ⟩ ‾ \overline{\langle Y_n\rangle} ⟨ Y n ⟩ and (2) the zero polynomial has its own class.
The permanent have different number of the mutual-reducibility classes; let us go slowly here. For 2 ≤ j ≤ n 2 \le j \le n 2 ≤ j ≤ n , after applying J ( R j ) \mathcal{J}(R_j) J ( R j ) , every column of the matrix of variables [ x i , j ] 1 ≤ i , j ≤ n [x_{i,j}]_{1 \le i,j \le n} [ x i , j ] 1 ≤ i , j ≤ n will become the difference between columns j j j and 1 1 1 . Hence
Q j : = Perm n ◃ J ( R j ) = n ! ∏ i = 1 n ( x i , j − x i , 1 ) Q_j := \operatorname{Perm}_n \triangleleft\mathcal{J}(R_j) = n! \prod_{i=1}^{n}(x_{i,j} - x_{i,1}) Q j := Perm n ◃ J ( R j ) = n ! i = 1 ∏ n ( x i , j − x i , 1 )
is a nonzero polynomial. More generally, we get
Perm n ◃ J ( ∑ j = 2 n u j R j ) = n ! ∏ i = 1 n ( ∑ j = 2 n u j ( x i , j − x i , 1 ) ) . \operatorname{Perm}_n \triangleleft\mathcal{J}\left(\sum_{j=2}^{n}u_jR_j\right) = n! \prod_{i=1}^{n}\left(\sum_{j=2}^{n}u_j(x_{i,j} - x_{i,1})\right). Perm n ◃ J ( j = 2 ∑ n u j R j ) = n ! i = 1 ∏ n ( j = 2 ∑ n u j ( x i , j − x i , 1 ) ) .
Moreover, for M = t I n + ∑ j v j R j M = tI_n + \sum_j v_jR_j M = t I n + ∑ j v j R j in ⟨ X n ⟩ ‾ \overline{\langle X_n\rangle} ⟨ X n ⟩ , we have M ( ∑ j = 2 n u j R j ) = t ( ∑ j = 2 n u j R j ) M(\sum_{j=2}^{n}u_jR_j) = t(\sum_{j=2}^{n}u_jR_j) M ( ∑ j = 2 n u j R j ) = t ( ∑ j = 2 n u j R j ) , implying that
( Perm n ◃ J ( ∑ j = 2 n u j R j ) ) ◃ J ( M ) = Perm n ◃ t n ( ∑ j = 2 n u j R j ) . \left(\operatorname{Perm}_n \triangleleft\mathcal{J}\left(\sum_{j=2}^{n}u_jR_j\right)\right) \triangleleft\mathcal{J}(M) = \operatorname{Perm}_n \triangleleft t^n\left(\sum_{j=2}^{n}u_jR_j\right). ( Perm n ◃ J ( j = 2 ∑ n u j R j ) ) ◃ J ( M ) = Perm n ◃ t n ( j = 2 ∑ n u j R j ) .
We are now in a position to describe the mutual-reducibility classes in P n \mathcal{P}_n P n . Observe that generated invertible matrices give polynomials mutually reducible to Perm n \operatorname{Perm}_n Perm n , as the inverses belong to the matrix closure. However, the nonzero singular generated matrices give multiples of the Q j Q_j Q j polynomial, which are pairwise incomparable with respect to ≈ \approx ≈ . Moreover, none of the Q j Q_j Q j polynomial is comparable to Perm n \operatorname{Perm}_n Perm n , as Q j Q_j Q j contains the monomial ∏ i = 1 n x i , j \prod_{i=1}^{n}x_{i,j} ∏ i = 1 n x i , j , whereas the permanent does not. (More in detail, the identities above show that every further reduction of Q j Q_j Q j by a matrix in ⟨ Y n ⟩ ‾ \overline{\langle Y_n\rangle} ⟨ Y n ⟩ is a scalar multiple of Q j Q_j Q j . Since Perm n \operatorname{Perm}_n Perm n is not such a multiple, it cannot be obtained from Q j Q_j Q j . Thus they are not mutually reducible.) Together with the class of zero polynomial, we get n + 1 n+1 n + 1 mutual-reducibility classes in P n \mathcal{P}_n P n . In contrast, we will show that P ‾ n \overline{\mathcal{P}}_n P n has infinitely many classes. For each λ ∈ Q \lambda\in\mathbb{Q} λ ∈ Q , consider
g λ : = Perm n ◃ J ( R 2 + λ R 3 ) = n ! ∏ i = 1 n ( x i , 2 − x i , 1 + λ ( x i , 3 − x i , 1 ) ) , g_\lambda:= \operatorname{Perm}_n \triangleleft\mathcal{J}(R_2 + \lambda R_3) = n! \prod_{i=1}^{n}(x_{i,2} - x_{i,1} + \lambda(x_{i,3} - x_{i,1})), g λ := Perm n ◃ J ( R 2 + λ R 3 ) = n ! i = 1 ∏ n ( x i , 2 − x i , 1 + λ ( x i , 3 − x i , 1 )) ,
where J ( R 2 + λ R 3 ) \mathcal{J}(R_2 + \lambda R_3) J ( R 2 + λ R 3 ) belongs to ⟨ Y n ⟩ ‾ \overline{\langle Y_n\rangle} ⟨ Y n ⟩ . Since the orbit map is polynomial, it follows that g λ ∈ P ‾ n g_\lambda\in\overline{\mathcal{P}}_n g λ ∈ P n . We claim that these polynomials g λ g_\lambda g λ are not multiple of each other. Assume otherwise g λ = c g μ g_\lambda= c g_\mu g λ = c g μ holds, then we can set x i , 2 = 1 x_{i,2} = 1 x i , 2 = 1 and all variables in the i i i -th row i i i to zero, for every i ≥ 2 i \ge2 i ≥ 2 . The resulting equality is
x 1 , 2 − x 1 , 1 + λ ( x 1 , 3 − x 1 , 1 ) = c ( x 1 , 2 − x 1 , 1 + μ ( x 1 , 3 − x 1 , 1 ) ) . x_{1,2} - x_{1,1} + \lambda(x_{1,3} - x_{1,1}) = c(x_{1,2} - x_{1,1} + \mu(x_{1,3} - x_{1,1})). x 1 , 2 − x 1 , 1 + λ ( x 1 , 3 − x 1 , 1 ) = c ( x 1 , 2 − x 1 , 1 + μ ( x 1 , 3 − x 1 , 1 )) .
But comparing the coefficients of x 1 , 2 x_{1,2} x 1 , 2 and x 1 , 3 x_{1,3} x 1 , 3 gives c = 1 c = 1 c = 1 and λ = μ \lambda= \mu λ = μ . Hence, the g λ g_\lambda g λ are pairwise non-equivalent under ≈ \approx ≈ . It follows that the polynomial g 1 g_1 g 1 is not mutually reducible to polynomials in P n \mathcal{P}_n P n . Consequently, g 1 ∈ P ‾ n ∖ P n g_1 \in\overline{\mathcal{P}}_n \setminus\mathcal{P}_n g 1 ∈ P n ∖ P n holds.
By the above, we can conclude that
∣ D n / ≈ ∣ = ∣ D ‾ n / ≈ ∣ = 2 and ∣ P n / ≈ ∣ = n + 1 , and ∣ P ‾ n / ≈ ∣ = ∞ . |\mathcal{D}_n/\approx| = |\overline{\mathcal{D}}_n/\approx| = 2 \qquad\text{and} \qquad|\mathcal{P}_n/\approx| = n+1, \qquad\text{and} \qquad|\overline{\mathcal{P}}_n/\approx| = \infty. ∣ D n / ≈ ∣ = ∣ D n / ≈ ∣ = 2 and ∣ P n / ≈ ∣ = n + 1 , and ∣ P n / ≈ ∣ = ∞.
This immediately proves that
Perm ‾ n ◃ ⟨ Y n ⟩ ⊈ Det ‾ n ◃ ⟨ Y n ⟩ . \overline{\operatorname{Perm}}_n \triangleleft\langle Y_n\rangle\not\subseteq\overline{\operatorname{Det}}_n \triangleleft\langle Y_n\rangle. Perm n ◃ ⟨ Y n ⟩ ⊆ Det n ◃ ⟨ Y n ⟩ .
An explicit witness is g 1 g_1 g 1 , as it contains the monomial ∏ i = 1 n x i , 2 \prod_{i=1}^{n}x_{i,2} ∏ i = 1 n x i , 2 with coefficient n ! n! n ! , whereas every polynomial in D ‾ n \overline{\mathcal{D}}_n D n has coefficient zero at this monomial.
Acknowledgment
We would like to thank Nikhil Balaji (IIT Delhi) and Sebastien Tavenas (CNRS) for helpful discussions about the border complexity and algebraic branching programs.
Background on Exterior Algebra
We recall the exterior-algebra background used in the proofs, following the presentation in [11 Appendix C] . The encoding of subspaces as exterior powers, discussed below, underlies the factorisation result used in Theorem 1 and in the proofs of Theorems 2 and 3. Throughout the appendix, matrices act on row vectors. An empty prefix or suffix in a decomposition is interpreted as the identity matrix I d I_d I d .
Encoding subspaces. Let V = Q d V = \mathbb{Q}^{d} V = Q d and 0 ≤ r ≤ d 0 \le r \le d 0 ≤ r ≤ d . The exterior power Λ r V \Lambda^{r} V Λ r V is the vector space generated by expressions v 1 ∧ ⋯ ∧ v r v_1 \wedge\cdots\wedge v_r v 1 ∧ ⋯ ∧ v r , subject to multilinearity and the alternating rule: a product with two equal vectors is zero. In particular, interchanging two vectors changes the sign. If e 1 , … , e d e_1,\ldots,e_d e 1 , … , e d is the standard basis of V V V , then
{ e i 1 ∧ ⋯ ∧ e i r : 1 ≤ i 1 < ⋯ < i r ≤ d } \{e_{i_1} \wedge\cdots\wedge e_{i_r} : 1 \le i_1 < \cdots< i_r \le d\} { e i 1 ∧ ⋯ ∧ e i r : 1 ≤ i 1 < ⋯ < i r ≤ d }
is a basis of Λ r V \Lambda^{r} V Λ r V , so its dimension is ( d r ) \binom{d}{r} ( r d ) . We use the convention Λ 0 V = Q \Lambda^{0} V = \mathbb{Q} Λ 0 V = Q , with empty wedge equal to 1 1 1 . The direct sum Λ V : = ⨁ r = 0 d Λ r V \Lambda V := \bigoplus_{r=0}^{d} \Lambda^{r} V Λ V := ⨁ r = 0 d Λ r V is the exterior algebra of V V V and has dimension 2 d 2^{d} 2 d .
The wedge v 1 ∧ ⋯ ∧ v r v_1 \wedge\cdots\wedge v_r v 1 ∧ ⋯ ∧ v r is nonzero exactly when v 1 , … , v r v_1,\ldots,v_r v 1 , … , v r are linearly independent. Thus an r r r -dimensional subspace U U U can be represented by a nonzero vector
u U : = u 1 ∧ ⋯ ∧ u r ∈ Λ r V , u_U := u_1 \wedge\cdots\wedge u_r \in\Lambda^{r} V, u U := u 1 ∧ ⋯ ∧ u r ∈ Λ r V ,
where u 1 , … , u r u_1,\ldots,u_r u 1 , … , u r is any basis of U U U . Changing this basis multiplies u U u_U u U by a nonzero scalar. Two subspaces of dimension r r r are equal exactly when their representing vectors are proportional. Moreover, for subspaces U , K ≤ V U,K \le V U , K ≤ V , their representing wedges satisfy
u U ∧ u K ≠ 0 ⟺ U ∩ K = { 0 } . u_U \wedge u_K \ne0 \quad\Longleftrightarrow\quad U \cap K = \{0\}. u U ∧ u K = 0 ⟺ U ∩ K = { 0 } .
Indeed, the wedge on the left is nonzero exactly when the two chosen bases together are linearly independent.
Testing whether a product preserves rank. A linear functional on a Q \mathbb{Q} Q -vector space is a linear map from that space to Q \mathbb{Q} Q . The following construction turns the intersection test on subspaces, above, into evaluation of a linear functional representing one subspace on a vector representing a second subspace.
Claim A.1. Let K ≤ V K \le V K ≤ V have dimension d − r d-r d − r . There is a linear functional φ K : Λ r V → Q \varphi_K : \Lambda^{r} V \to\mathbb{Q} φ K : Λ r V → Q such that, for every r r r -dimensional subspace U U U ,
φ K ( u U ) ≠ 0 ⟺ U ∩ K = { 0 } . \varphi_K(u_U) \ne0 \quad\Longleftrightarrow\quad U \cap K = \{0\}. φ K ( u U ) = 0 ⟺ U ∩ K = { 0 } . Proof. Choose a surjective linear map ψ K : V → Q r \psi_K : V \to\mathbb{Q}^{r} ψ K : V → Q r with kernel K K K . It induces the linear map
⋀ r ψ K : Λ r V ⟶ Λ r Q r ≅ Q . \bigwedge\nolimits^{r}\psi_K : \Lambda^{r} V \longrightarrow\Lambda^{r}\mathbb{Q}^{r} \cong\mathbb{Q}. ⋀ r ψ K : Λ r V ⟶ Λ r Q r ≅ Q . Explicitly, after identifying the wedge of the standard basis of Q r \mathbb{Q}^{r} Q r with 1 ∈ Q 1 \in\mathbb{Q} 1 ∈ Q , this is the functional
φ K ( v 1 ∧ ⋯ ∧ v r ) = det ( ψ K ( v 1 ) ⋮ ψ K ( v r ) ) . \varphi_K(v_1 \wedge\cdots\wedge v_r)
=
\det\begin{pmatrix}
\psi_K(v_1) \\
\vdots\\
\psi_K(v_r)
\end{pmatrix}. φ K ( v 1 ∧ ⋯ ∧ v r ) = det ψ K ( v 1 ) ⋮ ψ K ( v r ) .
Multilinearity and alternation of the determinant make this a well-defined linear map. For a basis u 1 , … , u r u_1,\ldots,u_r u 1 , … , u r of U U U , φ K ( u 1 ∧ ⋯ ∧ u r ) \varphi_K(u_1\wedge\cdots\wedge u_r) φ K ( u 1 ∧ ⋯ ∧ u r ) is nonzero exactly when ψ K \psi_K ψ K restricts to an isomorphism on U U U , equivalently when U ∩ K = { 0 } U\cap K=\{0\} U ∩ K = { 0 } . The assertion remains true when r = 0 r=0 r = 0 , since the determinant of the empty matrix is 1.
For rank-r r r matrices a , b a,b a , b , restriction of b b b to im ( a ) \operatorname{im}(a) im ( a ) gives
rank ( a b ) = r − dim ( im ( a ) ∩ ker ( b ) ) . (7) \operatorname{rank}(ab)=r-\dim\left(\operatorname{im}(a)\cap\operatorname{ker}(b)\right). \tag*{(7)} rank ( ab ) = r − dim ( im ( a ) ∩ ker ( b ) ) . (7)
Consequently, the product a b ab ab has rank r r r exactly when φ ker ( b ) ( u im ( a ) ) ≠ 0 \varphi_{\operatorname{ker}(b)}(u_{\operatorname{im}(a)})\ne0 φ ker ( b ) ( u im ( a ) ) = 0 . We denote by End ( W ) \operatorname{End}(W) End ( W ) the vector space of all linear maps from a finite-dimensional vector space W W W to itself with addition and scalar multiplication defined pointwise. Note that dim End ( W ) = ( dim W ) 2 \dim\operatorname{End}(W)=(\dim W)^2 dim End ( W ) = ( dim W ) 2 .
Kernels and images of rank-preserving products. The following elementary facts are used in all three proofs.
Claim A.2. If matrices a 1 , … , a k ∈ Mat d ( Q ) a_1,\ldots,a_k\in\operatorname{Mat}_d(\mathbb{Q}) a 1 , … , a k ∈ Mat d ( Q ) and their product a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k all have rank r r r , then, for 1 ≤ i ≤ j ≤ k 1\le i\le j\le k 1 ≤ i ≤ j ≤ k ,
rank ( a i ⋯ a j ) = r , ker ( a i ⋯ a j ) = ker ( a i ) , im ( a i ⋯ a j ) = im ( a j ) . \operatorname{rank}(a_i\cdots a_j)=r,\qquad\operatorname{ker}(a_i\cdots a_j)=\operatorname{ker}(a_i),\qquad\operatorname{im}(a_i\cdots a_j)=\operatorname{im}(a_j). rank ( a i ⋯ a j ) = r , ker ( a i ⋯ a j ) = ker ( a i ) , im ( a i ⋯ a j ) = im ( a j ) . Conversely, if all a i a_i a i have rank r r r and rank ( a i a i + 1 ) = r \operatorname{rank}(a_i a_{i+1})=r rank ( a i a i + 1 ) = r for 1 ≤ i < k 1\le i<k 1 ≤ i < k , then their product has rank r r r . In particular, any nonempty rank-stable set generates a rank-stable subsemigroup. A matrix is rank-stable exactly when its kernel and image have intersection { 0 } \{0\} { 0 } .
Proof. The rank cannot increase under multiplication, so the rank of a factor lies between that of the full product and that of each of its constituent matrices. Thus every factor has rank r r r . For row actions, ker ( a i ) ⊆ ker ( a i ⋯ a j ) \operatorname{ker}(a_i)\subseteq\operatorname{ker}(a_i\cdots a_j) ker ( a i ) ⊆ ker ( a i ⋯ a j ) and im ( a i ⋯ a j ) ⊆ im ( a j ) \operatorname{im}(a_i\cdots a_j)\subseteq\operatorname{im}(a_j) im ( a i ⋯ a j ) ⊆ im ( a j ) ; equality of dimensions gives the asserted identities.
For the converse, we use induction on the number of matrices. The partial product a 1 ⋯ a i a_1\cdots a_i a 1 ⋯ a i has image im ( a i ) \operatorname{im}(a_i) im ( a i ) , which is disjoint from ker ( a i + 1 ) \operatorname{ker}(a_{i+1}) ker ( a i + 1 ) by (7) . The next multiplication therefore preserves rank. If every pair of matrices in a set preserves their common rank, this argument applies to every product over that set, including products of two such products. Finally, apply (7) with b = a b=a b = a and use dim ker ( a ) + dim im ( a ) = d \dim\operatorname{ker}(a)+\dim\operatorname{im}(a)=d dim ker ( a ) + dim im ( a ) = d .
Detailed Proof of Theorem 1
Fix a finite nonempty set X ⊆ Mat d ( Q ) X\subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) and a congruence ∼ \sim ∼ on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . For the two claims below, assume that every stable subsemigroup S ≤ ⟨ X ⟩ S\le\langle X\rangle S ≤ ⟨ X ⟩ satisfies ∣ S / ∼ ∣ ≤ N |S/\sim|\le N ∣ S / ∼ ∣ ≤ N . We first obtain the bound for products whose factors and resulting matrix all have the same rank, and then use it in the decreasing induction on the minimum rank of a class.
Claim 3.1. Let a 1 , … , a k ∈ Mat d ( Q ) a_1,\ldots,a_k\in\operatorname{Mat}_d(\mathbb{Q}) a 1 , … , a k ∈ Mat d ( Q ) be such that rank ( a 1 ⋯ a k ) = rank ( a i ) \operatorname{rank}(a_1\cdots a_k)=\operatorname{rank}(a_i) rank ( a 1 ⋯ a k ) = rank ( a i ) for all 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k . Then a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k can be partitioned into at most 2 d + 2 2^{d+2} 2 d + 2 factors, each of which is either a single matrix among the a i a_i a i ’s or a rank-stable matrix product.
Proof. By [11 Proposition 7] , with the definition of factorisation tree in § 4 of that paper, the product has a factorisation tree of height at most d + 2 d+2 d + 2 . At every node with at least three children, the matrix labelling the node is rank-stable. Along each path from the root to a leaf, stop at the first node with at least three children and remove its descendants. The remaining tree has at most two children per internal node and height at most d + 2 d+2 d + 2 . Hence it has at most 2 d + 2 2^{d+2} 2 d + 2 leaves. Each remaining leaf is either an original leaf, labelled by one of the a i a_i a i , or a node at which we stopped, labelled by a rank-stable matrix. These labels are products of consecutive factors that partition a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k , as required. □
Claim 3.2. Let T ⊆ ⟨ X ⟩ T\subseteq\langle X\rangle T ⊆ ⟨ X ⟩ be finite and all its matrices have rank r r r . The rank-r r r matrices in ⟨ T ⟩ \langle T\rangle ⟨ T ⟩ meet at most ( 1 + ∣ T ∣ + N ∣ T ∣ 2 ) 2 d + 2 (1+|T|+N|T|^2)^{2^{d+2}} ( 1 + ∣ T ∣ + N ∣ T ∣ 2 ) 2 d + 2 congruence classes.
Proof. Write t = ∣ T ∣ t=|T| t = ∣ T ∣ and h = 2 d + 2 h=2^{d+2} h = 2 d + 2 . If T T T is empty there are no products to count. Otherwise, every rank-r r r product over T T T has the kernel of its first matrix and the image of its last matrix, by Claim A.2 . There are therefore at most t 2 t^2 t 2 possible pairs ( K , U ) (K,U) ( K , U ) of kernel and image.
For such a pair with K ∩ U = { 0 } K\cap U=\{0\} K ∩ U = { 0 } , consider
S K , U : = { a ∈ ⟨ T ⟩ : ker ( a ) = K , im ( a ) = U } . S_{K,U}:=\{a\in\langle T\rangle:\ker(a)=K,\ \operatorname{im}(a)=U\}. S K , U := { a ∈ ⟨ T ⟩ : ker ( a ) = K , im ( a ) = U } . If this set is nonempty, it is a stable subsemigroup. Indeed, for a , b ∈ S K , U a,b\in S_{K,U} a , b ∈ S K , U , the intersection im ( a ) ∩ ker ( b ) = U ∩ K \operatorname{im}(a)\cap\ker(b)=U\cap K im ( a ) ∩ ker ( b ) = U ∩ K is zero. Thus a b ab ab has rank r r r , kernel K K K , and image U U U . Our hypothesis gives ∣ S K , U / ∼ ∣ ≤ N |S_{K,U}/\sim|\leq N ∣ S K , U / ∼ ∣ ≤ N . Hence the congruence class of a rank-stable product a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k is determined by the triple of (1) ker ( a 1 ) = : K \ker(a_1)=:K ker ( a 1 ) =: K , (2) im ( a k ) = : U \operatorname{im}(a_k)=:U im ( a k ) =: U and (3) its induced element in S K , U / ∼ S_{K,U}/\sim S K , U / ∼ . Therefore the rank-stable matrices of rank r r r in ⟨ T ⟩ \langle T\rangle ⟨ T ⟩ meet at most N t 2 Nt^2 N t 2 classes in total.
By Claim 3.1, every rank-r r r product over T T T is a product of at most h h h factors, each either in T T T or a rank-stable matrix of rank r r r . There are at most t + N t 2 t+Nt^2 t + N t 2 choices for the class of each factor. Since ∼ \sim ∼ is a congruence, the number of classes of the resulting products is at most
∑ j = 1 h ( t + N t 2 ) j ≤ ( 1 + t + N t 2 ) h . \sum_{j=1}^{h}(t+Nt^2)^j\leq(1+t+Nt^2)^h. j = 1 ∑ h ( t + N t 2 ) j ≤ ( 1 + t + N t 2 ) h . □
Theorem 1. Let X ⊆ Mat d ( Q ) X\subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite, and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is finite if and only if there is an N ≥ 1 N\geq1 N ≥ 1 such that ∣ S / ∼ ∣ ≤ N |S/\sim|\leq N ∣ S / ∼ ∣ ≤ N for every stable subsemigroup S ≤ ⟨ X ⟩ S\leq\langle X\rangle S ≤ ⟨ X ⟩ . The cardinality and the diameter of the quotient are at most F d ( N , ∣ X ∣ ) F_d(N,|X|) F d ( N , ∣ X ∣ ) .
Proof. If the quotient is finite, take N = ∣ ⟨ X ⟩ / ∼ ∣ N=|\langle X\rangle/\sim| N = ∣ ⟨ X ⟩ / ∼ ∣ . For the converse, assume the uniform bound N N N on stable subsemigroups. Write m = ∣ X ∣ m=|X| m = ∣ X ∣ , h = 2 d + 2 h=2^{d+2} h = 2 d + 2 , and, for each equivalence class c c c , define
ρ ( c ) : = min { rank ( a ) : a ∈ c } . \rho(c):=\min\{\operatorname{rank}(a):a\in c\}. ρ ( c ) := min { rank ( a ) : a ∈ c } . Choosing a minimum-rank representative of either of two classes shows that
ρ ( b c ) ≤ min { ρ ( b ) , ρ ( c ) } . (8) \rho(bc)\leq\min\{\rho(b),\rho(c)\}. \tag*{(8)} ρ ( b c ) ≤ min { ρ ( b ) , ρ ( c )} . (8) In particular, every factor of a product whose class has minimum rank r r r belongs to a class with minimum rank at least r r r .
Let M r M_r M r be the number of classes c c c with ρ ( c ) ≥ r \rho(c)\geq r ρ ( c ) ≥ r . We show by decreasing induction, for 1 ≤ r ≤ d 1\leq r\leq d 1 ≤ r ≤ d , that these numbers are finite and satisfy an explicit recurrence. For the base case, the invertible matrices in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ form a stable subsemigroup, if there are any. Every class with minimum rank d d d meets this subsemigroup, so M d ≤ N M_d\leq N M d ≤ N .
Suppose that r < d r<d r < d and that M r + 1 = n M_{r+1}=n M r + 1 = n is finite. Consider all classes with minimum rank r r r of either of the forms
[ x ] ∼ or b [ x ] ∼ , x ∈ X , ρ ( b ) > r . [x]_{\sim}\quad\text{or}\quad b[x]_{\sim},\qquad x\in X,\quad\rho(b)>r. [ x ] ∼ or b [ x ] ∼ , x ∈ X , ρ ( b ) > r . @page 30/51 starts_mid=1 ends_mid=1 Along each path from the root to a leaf, stop at the first node with at least three children and remove its descendants. The remaining tree has at most two children per internal node and height at most d + 2 d+2 d + 2 . Hence it has at most 2 d + 2 2^{d+2} 2 d + 2 leaves. Each remaining leaf is either an original leaf, labelled by one of the a i a_i a i , or a node at which we stopped, labelled by a rank-stable matrix. These labels are products of consecutive factors that partition a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k , as required. □
Claim 3.2. Let T ⊆ ⟨ X ⟩ T\subseteq\langle X\rangle T ⊆ ⟨ X ⟩ be finite and all its matrices have rank r r r . The rank-r r r matrices in ⟨ T ⟩ \langle T\rangle ⟨ T ⟩ meet at most ( 1 + ∣ T ∣ + N ∣ T ∣ 2 ) 2 d + 2 (1+|T|+N|T|^2)^{2^{d+2}} ( 1 + ∣ T ∣ + N ∣ T ∣ 2 ) 2 d + 2 congruence classes.
Proof. Write t = ∣ T ∣ t=|T| t = ∣ T ∣ and h = 2 d + 2 h=2^{d+2} h = 2 d + 2 . If T T T is empty there are no products to count. Otherwise, every rank-r r r product over T T T has the kernel of its first matrix and the image of its last matrix, by Claim A.2 . There are therefore at most t 2 t^2 t 2 possible pairs ( K , U ) (K,U) ( K , U ) of kernel and image.
For such a pair with K ∩ U = { 0 } K\cap U=\{0\} K ∩ U = { 0 } , consider
S K , U : = { a ∈ ⟨ T ⟩ : ker ( a ) = K , im ( a ) = U } . S_{K,U}:=\{a\in\langle T\rangle:\ker(a)=K,\ \operatorname{im}(a)=U\}. S K , U := { a ∈ ⟨ T ⟩ : ker ( a ) = K , im ( a ) = U } . If this set is nonempty, it is a stable subsemigroup. Indeed, for a , b ∈ S K , U a,b\in S_{K,U} a , b ∈ S K , U , the intersection im ( a ) ∩ ker ( b ) = U ∩ K \operatorname{im}(a)\cap\ker(b)=U\cap K im ( a ) ∩ ker ( b ) = U ∩ K is zero. Thus a b ab ab has rank r r r , kernel K K K , and image U U U . Our hypothesis gives ∣ S K , U / ∼ ∣ ≤ N |S_{K,U}/\sim|\leq N ∣ S K , U / ∼ ∣ ≤ N . Hence the congruence class of a rank-stable product a 1 ⋯ a k a_1\cdots a_k a 1 ⋯ a k is determined by the triple of (1) ker ( a 1 ) = : K \ker(a_1)=:K ker ( a 1 ) =: K , (2) im ( a k ) = : U \operatorname{im}(a_k)=:U im ( a k ) =: U and (3) its induced element in S K , U / ∼ S_{K,U}/\sim S K , U / ∼ . Therefore the rank-stable matrices of rank r r r in ⟨ T ⟩ \langle T\rangle ⟨ T ⟩ meet at most N t 2 Nt^2 N t 2 classes in total.
By Claim 3.1, every rank-r r r product over T T T is a product of at most h h h factors, each either in T T T or a rank-stable matrix of rank r r r . There are at most t + N t 2 t+Nt^2 t + N t 2 choices for the class of each factor. Since ∼ \sim ∼ is a congruence, the number of classes of the resulting products is at most
∑ j = 1 h ( t + N t 2 ) j ≤ ( 1 + t + N t 2 ) h . \sum_{j=1}^{h}(t+Nt^2)^j\leq(1+t+Nt^2)^h. j = 1 ∑ h ( t + N t 2 ) j ≤ ( 1 + t + N t 2 ) h . Theorem 1. Let X ⊆ Mat d ( Q ) X\subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite, and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is finite if and only if there is an N ≥ 1 N\geq1 N ≥ 1 such that ∣ S / ∼ ∣ ≤ N |S/\sim|\leq N ∣ S / ∼ ∣ ≤ N for every stable subsemigroup S ≤ ⟨ X ⟩ S\leq\langle X\rangle S ≤ ⟨ X ⟩ . The cardinality and the diameter of the quotient are at most F d ( N , ∣ X ∣ ) F_d(N,|X|) F d ( N , ∣ X ∣ ) .
Proof. If the quotient is finite, take N = ∣ ⟨ X ⟩ / ∼ ∣ N=|\langle X\rangle/\sim| N = ∣ ⟨ X ⟩ / ∼ ∣ . For the converse, assume the uniform bound N N N on stable subsemigroups. Write m = ∣ X ∣ m=|X| m = ∣ X ∣ , h = 2 d + 2 h=2^{d+2} h = 2 d + 2 , and, for each equivalence class c c c , define
ρ ( c ) : = min { rank ( a ) : a ∈ c } . \rho(c):=\min\{\operatorname{rank}(a):a\in c\}. ρ ( c ) := min { rank ( a ) : a ∈ c } . Choosing a minimum-rank representative of either of two classes shows that
ρ ( b c ) ≤ min { ρ ( b ) , ρ ( c ) } . (9) \rho(bc)\leq\min\{\rho(b),\rho(c)\}. \tag*{(9)} ρ ( b c ) ≤ min { ρ ( b ) , ρ ( c )} . (9) In particular, every factor of a product whose class has minimum rank r r r belongs to a class with minimum rank at least $r.
Let M r M_r M r be the number of classes c c c with ρ ( c ) ≥ r \rho(c)\geq r ρ ( c ) ≥ r . We show by decreasing induction, for 1 ≤ r ≤ d 1\leq r\leq d 1 ≤ r ≤ d , that these numbers are finite and satisfy an explicit recurrence. For the base case, the invertible matrices in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ form a stable subsemigroup, if there are any. Every class with minimum rank d d d meets this subsemigroup, so M d ≤ N M_d\leq N M d ≤ N .
Suppose that r < d r<d r < d and that M r + 1 = n M_{r+1}=n M r + 1 = n is finite. Consider all classes with minimum rank r r r of either of the forms
[ x ] ∼ or b [ x ] ∼ , x ∈ X , ρ ( b ) > r . [x]_{\sim}\quad\text{or}\quad b[x]_{\sim},\qquad x\in X,\quad\rho(b)>r. [ x ] ∼ or b [ x ] ∼ , x ∈ X , ρ ( b ) > r .
There are at most t : = m ( n + 1 ) t := m(n+1) t := m ( n + 1 ) such classes. Choose one rank-r r r matrix in each and let T ⊆ ⟨ X ⟩ T \subseteq\langle X\rangle T ⊆ ⟨ X ⟩ be the finite set of these representatives. This set is fixed for the entire induction step, and ∣ T ∣ ≤ t |T| \leq t ∣ T ∣ ≤ t .
Let a 1 ⋯ a k a_1 \cdots a_k a 1 ⋯ a k , with a i ∈ X a_i \in X a i ∈ X , represent any class whose minimum rank is r r r . It admits a decomposition
a 1 ⋯ a k = ( y 1 c 1 ) ⋯ ( y ℓ c ℓ ) y ℓ + 1 , (10) a_1 \cdots a_k = (y_1c_1)\cdots(y_\ell c_\ell)y_{\ell+1}, \tag*{(10)} a 1 ⋯ a k = ( y 1 c 1 ) ⋯ ( y ℓ c ℓ ) y ℓ + 1 , (10)
where c i ∈ X c_i \in X c i ∈ X , each [ y i c i ] ∼ [y_i c_i]_{\sim} [ y i c i ] ∼ has minimum rank r r r , and every y i y_i y i , including the final suffix, is either empty or belongs to a class of minimum rank greater than r r r . To obtain this decomposition, take the shortest initial factor whose class has minimum rank r r r and separate its last generator c 1 c_1 c 1 . Repeat on the remaining suffix for as long as its class has minimum rank r r r . All factors have minimum rank at least r r r by (8), so the procedure has precisely the stated properties. Since the full class has minimum rank r r r , at least one factor y i c i y_i c_i y i c i is selected.
Each class [ y i c i ] ∼ [y_i c_i]_{\sim} [ y i c i ] ∼ has a representative b i ∈ T b_i \in T b i ∈ T . Replacing these factors gives
a 1 ⋯ a k ∼ b 1 ⋯ b ℓ y ℓ + 1 . a_1 \cdots a_k \sim b_1 \cdots b_\ell y_{\ell+1}. a 1 ⋯ a k ∼ b 1 ⋯ b ℓ y ℓ + 1 .
The product b 1 ⋯ b ℓ b_1 \cdots b_\ell b 1 ⋯ b ℓ has rank at most r r r because each b i b_i b i has rank r r r . It cannot have smaller rank: multiplying it by the remaining suffix, when present, would give a matrix of rank less than r r r in the original class. Thus its rank is exactly r r r . Claim 3.2 bounds its possible classes by ( 1 + t + N t 2 ) h (1+t+Nt^2)^h ( 1 + t + N t 2 ) h . There are at most n + 1 n+1 n + 1 choices for the suffix class, including the empty suffix as one possibility. Consequently,
M r ≤ n + ( n + 1 ) ( 1 + t + N t 2 ) h , n = M r + 1 , t = m ( n + 1 ) . (11) M_r \leq n+(n+1)(1+t+Nt^2)^h,\qquad n=M_{r+1},\qquad t=m(n+1). \tag*{(11)} M r ≤ n + ( n + 1 ) ( 1 + t + N t 2 ) h , n = M r + 1 , t = m ( n + 1 ) . (11)
This completes the induction. There is at most one additional class with minimum rank zero, since the zero matrix is unique. Hence the whole quotient has at most M 1 + 1 M_1+1 M 1 + 1 classes.
We now verify the stated numerical bound. Put K r = M r + 1 K_r=M_r+1 K r = M r + 1 and A = ( N + 3 ) m 2 A=(N+3)m^2 A = ( N + 3 ) m 2 . Since m ≥ 1 m\geq1 m ≥ 1 and K r + 1 ≥ 1 K_{r+1}\geq1 K r + 1 ≥ 1 ,
2 + m K r + 1 + N m 2 K r + 1 2 ≤ A K r + 1 2 . 2+mK_{r+1}+Nm^2K_{r+1}^2 \leq AK_{r+1}^2. 2 + m K r + 1 + N m 2 K r + 1 2 ≤ A K r + 1 2 .
Using 1 + z h ≤ ( 1 + z ) h 1+z^h\leq(1+z)^h 1 + z h ≤ ( 1 + z ) h for z ≥ 0 z\geq0 z ≥ 0 in (10), we obtain
K r ≤ A h K r + 1 2 h + 1 , A K r ≤ ( A K r + 1 ) 2 h + 1 . K_r\leq A^hK_{r+1}^{2h+1},\qquad\sqrt{A}K_r\leq\left(\sqrt{A}K_{r+1}\right)^{2h+1}. K r ≤ A h K r + 1 2 h + 1 , A K r ≤ ( A K r + 1 ) 2 h + 1 .
Iterating d − 1 d-1 d − 1 times and using K d ≤ N + 1 K_d\leq N+1 K d ≤ N + 1 gives
∣ ⟨ X ⟩ / ∼ ∣ ≤ K 1 ≤ [ A ( N + 1 ) ] ( 2 h + 1 ) d − 1 ≤ [ m ( N + 1 ) 2 ] ( 2 d + 3 + 1 ) d − 1 = F d ( N , m ) , \begin{aligned}
|\langle X\rangle/\sim|\leq K_1\leq\left[\sqrt{A}(N+1)\right]^{(2h+1)^{d-1}} \\
&\leq\left[m(N+1)^2\right]^{(2^{d+3}+1)^{d-1}}=F_d(N,m),
\end{aligned} ∣ ⟨ X ⟩ / ∼ ∣ ≤ K 1 ≤ [ A ( N + 1 ) ] ( 2 h + 1 ) d − 1 ≤ [ m ( N + 1 ) 2 ] ( 2 d + 3 + 1 ) d − 1 = F d ( N , m ) ,
where the last inequality uses N + 3 ≤ N + 1 \sqrt{N+3}\leq N+1 N + 3 ≤ N + 1 for N ≥ 1 N\geq1 N ≥ 1 . This also covers the case d = 1 d=1 d = 1 , when no induction step is needed.
Finally, a product of minimum length in its class has pairwise incongruent nonempty prefix products. Otherwise, the factor between two congruent prefixes could be omitted to obtain a shorter representative. Its length is therefore at most the cardinality of the quotient. This proves the diameter bound as well. Since ( 2 d + 3 + 1 ) d − 1 = 2 O ( d 2 ) (2^{d+3}+1)^{d-1}=2^{O(d^2)} ( 2 d + 3 + 1 ) d − 1 = 2 O ( d 2 ) , the asserted asymptotic form follows. ∎
Detailed Proof of Corollary 1
Corollary 1. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite, and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Suppose that the relation ∼ \sim ∼ is relatively Zariski closed in ⟨ X ⟩ × ⟨ X ⟩ \langle X\rangle\times\langle X\rangle ⟨ X ⟩ × ⟨ X ⟩ . If ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is torsion, then it is finite.
Proof. By Theorem 1, it suffices to give a uniform bound for ∣ S / ∼ ∣ |S/\sim| ∣ S / ∼ ∣ , where S ≤ ⟨ X ⟩ S \leq\langle X\rangle S ≤ ⟨ X ⟩ is a stable subgroup. The case that all matrices in S S S have rank 0 is immediate. Otherwise, restriction of the elements of S S S to their common image identifies S S S with a subsemigroup of GL r ( Q ) \operatorname{GL}_r(\mathbb{Q}) GL r ( Q ) , where 1 ≤ r ≤ d 1 \leq r \leq d 1 ≤ r ≤ d , and preserves relative closedness of the congruence. Taking closures over Q ‾ \overline{\mathbb{Q}} Q within the indicated groups, put
G = S ‾ ≤ GL r ( Q ‾ ) , H = { ( a , b ) ∈ S × S ‾ : a ∼ b } ≤ G × G . G=\overline{S}\leq\operatorname{GL}_r(\overline{\mathbb{Q}}), \qquad H=\{(a,b)\in\overline{S\times S}:a\sim b\}\leq G\times G. G = S ≤ GL r ( Q ) , H = {( a , b ) ∈ S × S : a ∼ b } ≤ G × G .
Being, respectively, closures of subsemigroups of GL r ( Q ‾ ) \operatorname{GL}_r(\overline{\mathbb{Q}}) GL r ( Q ) and GL r ( Q ‾ ) × GL r ( Q ‾ ) \operatorname{GL}_r(\overline{\mathbb{Q}})\times\operatorname{GL}_r(\overline{\mathbb{Q}}) GL r ( Q ) × GL r ( Q ) , G G G and H H H are algebraic groups by [11 Lemma 9] . Since H H H contains the diagonal of G G G , the subgroup N = { g ∈ G : ( g , 1 ) ∈ H } N=\{g\in G:(g,1)\in H\} N = { g ∈ G : ( g , 1 ) ∈ H } is closed and normal, and ( g , h ) ∈ H (g,h)\in H ( g , h ) ∈ H if and only if g h − 1 ∈ N gh^{-1}\in N g h − 1 ∈ N . Relative closedness therefore shows that the quotient map q : G → G / N q:G\to G/N q : G → G / N induces an isomorphism S / ∼ ≅ q ( S ) S/\sim\cong q(S) S / ∼≅ q ( S ) . The quotient G / N G/N G / N is a linear algebraic group [24 Theorem 11.5] , and q ( S ) q(S) q ( S ) is a torsion subgroup thereof.
The remaining problem is that while Burnside would now show finiteness of the quotient, we require a uniform bound N N N . Hence we have to prove that (1) the quotient is uniformly torsion, and (2) that the quotient can be embedded into a not-too-high dimension d ′ d' d ′ . In particular, while the quotient is again linear algebraic by [24 Theorem 11.5] , it does not state a bound on the corresponding matrix dimension d ′ d' d ′ , which we now have to prove in our setting.
To see (1), let
E d = lcm { n ≥ 1 : φ ( n ) ≤ d ! } . E_d=\operatorname{lcm}\{n\geq1:\varphi(n)\leq d!\}. E d = lcm { n ≥ 1 : φ ( n ) ≤ d !} .
where φ \varphi φ is Euler’s totient function. This is finite by the well-known identity φ ( n ) ≥ n 2 \varphi(n)\geq\sqrt{\frac{n}{2}} φ ( n ) ≥ 2 n , wherefore the lcm is taken over a finite set. For a ∈ S a\in S a ∈ S , let C a C_a C a be the closure of the cyclic group generated by a a a . Let C a ∘ C_a^\circ C a ∘ be the identity component of C a C_a C a . By [26 Theorem 3.1 and Remark 2.3] , the size of the quotient ∣ C a / C a ∘ ∣ |C_a/C_a^\circ| ∣ C a / C a ∘ ∣ equals the order of the torsion subgroup of the multiplicative group generated by the eigenvalues of a a a . This torsion subgroup is cyclic and lies in the splitting field of the characteristic polynomial of a a a , whose degree is at most r ! ≤ d ! r!\leq d! r ! ≤ d ! . Its order m m m thus satisfies φ ( m ) ≤ d ! \varphi(m)\leq d! φ ( m ) ≤ d ! , so m ∣ E d m\mid E_d m ∣ E d and a E d ∈ C a ∘ a^{E_d}\in C_a^\circ a E d ∈ C a ∘ . Since q ( a ) q(a) q ( a ) has finite order, q ( C a ) q(C_a) q ( C a ) is finite and q ( C a ∘ ) = 1 q(C_a^\circ)=1 q ( C a ∘ ) = 1 . Consequently q ( a ) E d = 1 q(a)^{E_d}=1 q ( a ) E d = 1 for every a ∈ S a\in S a ∈ S . Burnside’s theorem for linear groups of finite exponent [27 Theorem 8.1.11(i)] implies that q ( S ) q(S) q ( S ) is finite. As q ( S ) q(S) q ( S ) is dense in G / N G/N G / N , it follows that G / N = q ( S ) G/N=q(S) G / N = q ( S ) .
To see (2): By [28 Theorem 1.1] , there is a finite subgroup F ≤ G F\leq G F ≤ G with q ( F ) = G / N q(F)=G/N q ( F ) = G / N . Jordan’s theorem [29 Theorem 1.1] yields an abelian subgroup A ≤ F A\leq F A ≤ F with [ F : A ] ≤ J ( d ) [F:A]\leq J(d) [ F : A ] ≤ J ( d ) , where J ( d ) J(d) J ( d ) depends only on d d d . Simultaneous diagonalization shows that A A A has at most r ≤ d r\leq d r ≤ d generators. Since in an abelian group every element can be identified with a word g 1 n 1 ⋯ g r n r g_1^{n_1}\cdots g_r^{n_r} g 1 n 1 ⋯ g r n r where g 1 , … , g r g_1,\ldots,g_r g 1 , … , g r are the generators and n i ∈ N n_i\in\mathbb{N} n i ∈ N , we obtain ∣ q ( A ) ∣ ≤ E d d ≤ E d d |q(A)|\leq E_d^d\leq E_d^d ∣ q ( A ) ∣ ≤ E d d ≤ E d d , and
∣ S / ∼ ∣ = ∣ G / N ∣ ≤ [ F : A ] ∣ q ( A ) ∣ ≤ J ( d ) E d d . |S/\sim|=|G/N|\leq[F:A]|q(A)|\leq J(d)E_d^d. ∣ S / ∼ ∣ = ∣ G / N ∣ ≤ [ F : A ] ∣ q ( A ) ∣ ≤ J ( d ) E d d .
This bound depends only on d d d , so Theorem 1 now applies. □
Detailed Proof of Theorem 2
Fix a finite nonempty set X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_{d}(\mathbb{Q}) X ⊆ Mat d ( Q ) and a congruence ∼ \sim ∼ on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Throughout the argument below, assume that ∣ S / ∼ ∣ ≤ N |S/\sim| \leq N ∣ S / ∼ ∣ ≤ N for every rank-stable subsemigroup S ≤ ⟨ X ⟩ S \leq\langle X\rangle S ≤ ⟨ X ⟩ , and put
D = ( d ⌊ d / 2 ⌋ ) , H = ( N + D 2 + 1 D 2 ) . D = \binom{d}{\lfloor d/2 \rfloor}, \qquad H = \binom{N + D^{2} + 1}{D^{2}}. D = ( ⌊ d /2 ⌋ d ) , H = ( D 2 N + D 2 + 1 ) .
We first prove the pruning claim from the overview. Its proof uses the exterior-algebra background in Appendix § 7 to find sufficiently many mutually joined cuts. The uniform bound on rank-stable subsemigroups then supplies two congruent prefixes.
Claim 3.3. Let a 1 , … , a k ∈ ⟨ X ⟩ a_{1},\ldots,a_{k} \in\langle X\rangle a 1 , … , a k ∈ ⟨ X ⟩ be such that k > H k > H k > H and rank ( a 1 ⋯ a k ) = rank ( a i ) \operatorname{rank}(a_{1}\cdots a_{k}) = \operatorname{rank}(a_{i}) rank ( a 1 ⋯ a k ) = rank ( a i ) for all 1 ≤ i ≤ k 1 \leq i \leq k 1 ≤ i ≤ k . There exist i , j i,j i , j with 1 ≤ i + 1 < j ≤ k 1 \leq i+1 < j \leq k 1 ≤ i + 1 < j ≤ k such that a 1 ⋯ a i a j ⋯ a k ∼ a 1 ⋯ a k a_{1}\cdots a_{i}a_{j}\cdots a_{k} \sim a_{1}\cdots a_{k} a 1 ⋯ a i a j ⋯ a k ∼ a 1 ⋯ a k .
The cut graph and its independent sets. Let r r r be the common rank of a 1 , … , a k a_{1},\ldots,a_{k} a 1 , … , a k and their product. The cut graph has vertices 1 , … , k − 1 1,\ldots,k-1 1 , … , k − 1 . Distinct cuts i , j i,j i , j are joined when
rank ( a i a j + 1 ) = rank ( a j a i + 1 ) = r . (12) \operatorname{rank}(a_{i}a_{j+1})=\operatorname{rank}(a_{j}a_{i+1})=r. \tag*{(12)} rank ( a i a j + 1 ) = rank ( a j a i + 1 ) = r . (12)
Equivalently, both im ( a i ) ∩ ker ( a j + 1 ) \operatorname{im}(a_{i}) \cap\ker(a_{j+1}) im ( a i ) ∩ ker ( a j + 1 ) and im ( a j ) ∩ ker ( a i + 1 ) \operatorname{im}(a_{j}) \cap\ker(a_{i+1}) im ( a j ) ∩ ker ( a i + 1 ) are zero. For i = j i=j i = j , the corresponding rank equality also holds, because the original product preserves rank; we do not add loops to the graph.
Claim 3.4. Each independent set of the cut graph has size at most ( d r ) 2 ≤ D 2 \binom{d}{r}^{2} \leq D^{2} ( r d ) 2 ≤ D 2 .
Proof. Put W = ⋀ r Q d W = \bigwedge^{r}\mathbb{Q}^{d} W = ⋀ r Q d . For each cut i i i , choose a wedge u i ∈ W u_{i} \in W u i ∈ W representing the r r r -dimensional space im ( a i ) \operatorname{im}(a_{i}) im ( a i ) . By Claim A.1 , there is a linear functional φ i : W → Q \varphi_{i}: W \to\mathbb{Q} φ i : W → Q associated with ker ( a i + 1 ) \ker(a_{i+1}) ker ( a i + 1 ) such that
φ i ( u j ) ≠ 0 ⟺ im ( a j ) ∩ ker ( a i + 1 ) = { 0 } ⟺ rank ( a j a i + 1 ) = r . (13) \varphi_{i}(u_{j}) \ne0 \Longleftrightarrow\operatorname{im}(a_{j}) \cap\ker(a_{i+1}) = \{0\} \Longleftrightarrow\operatorname{rank}(a_{j}a_{i+1}) = r. \tag*{(13)} φ i ( u j ) = 0 ⟺ im ( a j ) ∩ ker ( a i + 1 ) = { 0 } ⟺ rank ( a j a i + 1 ) = r . (13)
In particular, φ i ( u i ) ≠ 0 \varphi_{i}(u_{i}) \ne0 φ i ( u i ) = 0 .
Define the rank-one map T i : W → W T_{i}: W \to W T i : W → W and the linear functional f i : End ( W ) → Q f_{i}: \operatorname{End}(W) \to\mathbb{Q} f i : End ( W ) → Q by
T i ( z ) = φ i ( z ) u i , f i ( T ) = φ i ( T ( u i ) ) . T_{i}(z) = \varphi_{i}(z)u_{i}, \qquad f_{i}(T) = \varphi_{i}(T(u_{i})). T i ( z ) = φ i ( z ) u i , f i ( T ) = φ i ( T ( u i )) .
The map T i T_{i} T i has image Q u i \mathbb{Q}u_{i} Q u i , and f i f_{i} f i is linear because evaluation at u i u_{i} u i and the map φ i \varphi_{i} φ i are both linear. For all cuts i , j i,j i , j , we have
f i ( T j ) = φ j ( u i ) φ i ( u j ) . (14) f_{i}(T_{j}) = \varphi_{j}(u_{i})\varphi_{i}(u_{j}). \tag*{(14)} f i ( T j ) = φ j ( u i ) φ i ( u j ) . (14)
Thus f i ( T i ) = φ i ( u i ) 2 ≠ 0 f_{i}(T_{i}) = \varphi_{i}(u_{i})^{2} \ne0 f i ( T i ) = φ i ( u i ) 2 = 0 , whereas for distinct i , j i,j i , j , the value f i ( T j ) f_{i}(T_{j}) f i ( T j ) is nonzero exactly when i , j i,j i , j are joined.
Let I I I be an independent set. We show that the maps T i T_{i} T i , i ∈ I i \in I i ∈ I , are linearly independent in End ( W ) \operatorname{End}(W) End ( W ) . Suppose that ∑ i ∈ I α i T i \sum_{i\in I}\alpha_{i}T_{i} ∑ i ∈ I α i T i is the zero map, and fix j ∈ I j \in I j ∈ I . Evaluating f j f_{j} f j at this relation gives
0 = f j ( ∑ i ∈ I α i T i ) = ∑ i ∈ I α i f j ( T i ) = α j φ j ( u j ) 2 . 0 = f_{j}\left(\sum_{i\in I}\alpha_{i}T_{i}\right) = \sum_{i\in I}\alpha_{i}f_{j}(T_{i}) = \alpha_{j}\varphi_{j}(u_{j})^{2}. 0 = f j ( i ∈ I ∑ α i T i ) = i ∈ I ∑ α i f j ( T i ) = α j φ j ( u j ) 2 .
All terms other than the j j j th vanish by independence of I I I . Since φ j ( u j ) ≠ 0 \varphi_{j}(u_{j}) \ne0 φ j ( u j ) = 0 , we obtain α j = 0 \alpha_{j} = 0 α j = 0 . This holds for every j ∈ I j \in I j ∈ I , proving linear independence. Therefore
∣ I ∣ ≤ dim End ( W ) = ( d r ) 2 ≤ D 2 . |I| \leq\dim\operatorname{End}(W) = \binom{d}{r}^{2} \leq D^{2}. ∣ I ∣ ≤ dim End ( W ) = ( r d ) 2 ≤ D 2 .
The conventions in Appendix A cover the case r = 0 r = 0 r = 0 as well. In that case all a i a_i a i are zero and the cut graph is complete.
Proof of Claim 3.3. Write R ( p , q ) R(p,q) R ( p , q ) for the least integer such that every graph on that many vertices has either a clique of size p p p or an independent set of size q q q . We use the following classical bound [30 , 31 ] for this
R ( p , q ) ≤ ( p + q − 2 p − 1 ) . (15) R(p,q) \le\binom{p+q-2}{p-1}. \tag*{(15)} R ( p , q ) ≤ ( p − 1 p + q − 2 ) . (15)
For completeness, the recurrence R ( p , q ) ≤ R ( p − 1 , q ) + R ( p , q − 1 ) R(p,q) \le R(p-1,q) + R(p,q-1) R ( p , q ) ≤ R ( p − 1 , q ) + R ( p , q − 1 ) follows by separating the neighbours and non-neighbours of a vertex. Together with R ( 1 , q ) = R ( p , 1 ) = 1 R(1,q) = R(p,1) = 1 R ( 1 , q ) = R ( p , 1 ) = 1 , induction and Pascal’s identity give the bound (14). In the cut graph the number of vertices is k − 1 ≥ H k-1 \ge H k − 1 ≥ H and
R ( N + 2 , D 2 + 1 ) ≤ ( N + D 2 + 1 D 2 ) = H . R(N+2,D^2+1) \le\binom{N+D^2+1}{D^2} = H. R ( N + 2 , D 2 + 1 ) ≤ ( D 2 N + D 2 + 1 ) = H .
Claim 3.4 excludes an independent set of size D 2 + 1 D^2+1 D 2 + 1 . Hence there is a clique of size N + 2 N+2 N + 2 , i.e., there are mutually joined cuts i 0 < i 1 < ⋯ < i N + 1 i_0 < i_1 < \cdots< i_{N+1} i 0 < i 1 < ⋯ < i N + 1 .
For 1 ≤ j ≤ N + 1 1 \le j \le N+1 1 ≤ j ≤ N + 1 , put
b j : = a i j − 1 + 1 ⋯ a i j . b_j := a_{i_{j-1}+1}\cdots a_{i_j}. b j := a i j − 1 + 1 ⋯ a i j .
By Claim A.2, each b j b_j b j has rank r r r , with kernel ker ( a i j − 1 + 1 ) \ker(a_{i_{j-1}+1}) ker ( a i j − 1 + 1 ) and image im ( a i j ) \operatorname{im}(a_{i_j}) im ( a i j ) . We check that the b j b_j b j generate a rank-stable subsemigroup. For any p , q ∈ { 1 , … , N + 1 } p,q \in\{1,\ldots,N+1\} p , q ∈ { 1 , … , N + 1 } , the intersection
im ( b p ) ∩ ker ( b q ) = im ( a i p ) ∩ ker ( a i q − 1 + 1 ) \operatorname{im}(b_p) \cap\ker(b_q) = \operatorname{im}(a_{i_p}) \cap\ker(a_{i_{q-1}+1}) im ( b p ) ∩ ker ( b q ) = im ( a i p ) ∩ ker ( a i q − 1 + 1 )
is zero: if the cuts i p i_p i p and i q − 1 i_{q-1} i q − 1 are distinct, this follows from the existence of an edge between them; if they coincide, it follows from rank preservation at that cut in the original product. Thus all products b p b q b_p b_q b p b q have rank r r r , and Claim A.2 shows that S : = ⟨ b 1 , … , b N + 1 ⟩ S := \langle b_1,\ldots,b_{N+1}\rangle S := ⟨ b 1 , … , b N + 1 ⟩ is rank-stable.
By hypothesis, ∣ S / ∼ ∣ ≤ N |S/\sim| \le N ∣ S / ∼ ∣ ≤ N . Hence there are two non-empty prefixes of b 1 , … , b N + 1 b_1,\ldots,b_{N+1} b 1 , … , b N + 1 with congruent products, say
b 1 ⋯ b p ∼ b 1 ⋯ b q , 1 ≤ p < q ≤ N + 1. b_1\cdots b_p \sim b_1\cdots b_q,\qquad1 \le p < q \le N+1. b 1 ⋯ b p ∼ b 1 ⋯ b q , 1 ≤ p < q ≤ N + 1.
Multiplying by the surrounding factors in the original product of the a i a_i a i gives
a 1 ⋯ a k ∼ a 1 ⋯ a i p a i q + 1 ⋯ a k . a_1\cdots a_k \sim a_1\cdots a_{i_p}a_{i_q+1}\cdots a_k. a 1 ⋯ a k ∼ a 1 ⋯ a i p a i q + 1 ⋯ a k .
Taking i = i p i=i_p i = i p and j = i q + 1 j=i_q+1 j = i q + 1 proves the claim.
Theorem 2. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is finite if and only if there is an N ≥ 1 N \ge1 N ≥ 1 such that ∣ S / ∼ ∣ ≤ N |S/\sim| \le N ∣ S / ∼ ∣ ≤ N for every rank-stable subsemigroup S ≤ ⟨ X ⟩ S \le\langle X\rangle S ≤ ⟨ X ⟩ . The diameter of the quotient is at most G d ( N ) G_d(N) G d ( N ) .
Proof. The forward implication is immediate by taking the cardinality of the quotient ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ as a uniform bound on rank-stable subsemigroups. For the converse, assume a uniform bound N N N on rank-stable subsemigroups. Since X X X is finite, to prove finiteness of the quotient, it suffices to bound the length of every equivalence class.
Let a 1 ⋯ a k a_1 \cdots a_k a 1 ⋯ a k , with a i ∈ X a_i \in X a i ∈ X , be a product whose length k k k equals the length of its equivalence class. Every nonempty factor of this displayed product is itself shortest in its class since replacing a factor by a shorter congruent product would contradict the minimality of k k k . Define
L d = N , L r = ( H + 1 ) L r + 1 + H ( 0 ≤ r < d ) . L_d = N,\qquad L_r = (H+1)L_{r+1}+H \qquad(0 \le r < d). L d = N , L r = ( H + 1 ) L r + 1 + H ( 0 ≤ r < d ) .
We prove by decreasing induction on r r r that every such product with rank at least r r r has k ≤ L r k \le L_r k ≤ L r .
For the base case, if the product has rank d d d , all its nonempty prefixes are invertible. The invertible matrices in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ form a rank-stable subsemigroup whose quotient has at most N N N classes. If k > N k>N k > N , two of the k k k nonempty prefixes would be congruent. Omitting the factor between them would give a shorter representative of the same class. Hence k ≤ N = L d k \le N=L_d k ≤ N = L d .
For the induction step, products of rank greater than r r r are already covered by L r + 1 ≤ L r L_{r+1} \le L_r L r + 1 ≤ L r , so suppose the rank is r < d r<d r < d . Decompose the product as
a 1 ⋯ a k = y 0 ( c 1 y 1 ) ⋯ ( c ℓ y ℓ ) , (16) a_1 \cdots a_k = y_0(c_1y_1)\cdots(c_\ell y_\ell), \tag*{(16)} a 1 ⋯ a k = y 0 ( c 1 y 1 ) ⋯ ( c ℓ y ℓ ) , (16)
where c i ∈ X c_i \in X c i ∈ X , each nonempty y i y_i y i has rank greater than r r r , and each c i y i c_i y_i c i y i has rank r r r . Such a decomposition can be obtained from right to left: take the shortest suffix of rank r r r and separate its first generator; repeat on the remaining prefix while its rank is r r r . All nonempty factors have rank at least r r r , so the procedure ends with an empty prefix or one of rank greater than r r r . The suffix after the separated generator has rank greater than r r r , or is empty, by the choice of a shortest suffix.
Every nonempty y i y_i y i is shortest in its own class, so the induction hypothesis bounds its number of generators by L r + 1 L_{r+1} L r + 1 . There is at least one factor c i y i c_i y_i c i y i , and their product has rank exactly r r r : its rank is at most r r r , while multiplication on the left by y 0 y_0 y 0 gives the original rank-r r r product. If ℓ > H \ell>H ℓ > H , then Claim 3.3 would remove a nonempty consecutive list of these factors. Substituting in (15) would give a shorter congruent product over X X X , a contradiction. Thus ℓ ≤ H \ell\le H ℓ ≤ H , and
k ≤ L r + 1 + ℓ ( L r + 1 + 1 ) ≤ L r + 1 + H ( L r + 1 + 1 ) = L r . k \le L_{r+1}+\ell(L_{r+1}+1)\le L_{r+1}+H(L_{r+1}+1)=L_r. k ≤ L r + 1 + ℓ ( L r + 1 + 1 ) ≤ L r + 1 + H ( L r + 1 + 1 ) = L r .
This proves the induction. Since L r + 1 = ( H + 1 ) ( L r + 1 + 1 ) L_r+1=(H+1)(L_{r+1}+1) L r + 1 = ( H + 1 ) ( L r + 1 + 1 ) , it yields
L 0 = ( N + 1 ) ( H + 1 ) d − 1 = G d ( N ) . L_0=(N+1)(H+1)^d-1=G_d(N). L 0 = ( N + 1 ) ( H + 1 ) d − 1 = G d ( N ) .
Every class has length at most G d ( N ) G_d(N) G d ( N ) , proving both the diameter bound and finiteness. To verify the two asymptotic estimates, put M = D 2 ≤ 4 d M=D^2\le4^d M = D 2 ≤ 4 d . Counting multisets by ordered lists gives
H = ( M + N + 1 N + 1 ) ≤ ( M + 1 ) N + 1 , H = ( M + N + 1 M ) ≤ ( N + 2 ) M . H=\binom{M+N+1}{N+1}\le(M+1)^{N+1},\qquad
H=\binom{M+N+1}{M}\le(N+2)^M. H = ( N + 1 M + N + 1 ) ≤ ( M + 1 ) N + 1 , H = ( M M + N + 1 ) ≤ ( N + 2 ) M .
The first inequality implies log 2 ( G d ( N ) + 1 ) = O ( d 2 N ) \log_2(G_d(N)+1)=O(d^2N) log 2 ( G d ( N ) + 1 ) = O ( d 2 N ) ; the second implies log 2 ( G d ( N ) + 1 ) = O ( d 4 d log 2 ( N + 1 ) ) \log_2(G_d(N)+1)=O(d4^d\log_2(N+1)) log 2 ( G d ( N ) + 1 ) = O ( d 4 d log 2 ( N + 1 )) . Therefore
G d ( N ) ≤ min { 2 O ( d 2 N ) , ( N + 1 ) 2 O ( d ) } . G_d(N)\le\min\left\{2^{O(d^2N)},\,(N+1)^{2^{O(d)}}\right\}. G d ( N ) ≤ min { 2 O ( d 2 N ) , ( N + 1 ) 2 O ( d ) } .
Detailed Proof of Theorem 3
In this section, all Zariski closures are taken over Q ‾ \overline{\mathbb{Q}} Q . Put C : = ⟨ X ⟩ ‾ ⊆ Mat d ( Q ‾ ) C:=\overline{\langle X\rangle}\subseteq\operatorname{Mat}_d(\overline{\mathbb{Q}}) C := ⟨ X ⟩ ⊆ Mat d ( Q ) , where X ⊆ Mat d ( Q ) X\subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) is finite and nonempty. Kernels and images of rational matrices can be computed over Q \mathbb{Q} Q and then extended to Q ‾ \overline{\mathbb{Q}} Q ; their ranks and the condition that an intersection of two subspaces is zero are unchanged. For the absorption claim and the final theorem proof, ∼ \sim ∼ is defined on all of C C C , and every group G ≤ C G\le C G ≤ C satisfies ∣ G / ∼ ∣ ≤ N |G/\sim|\le N ∣ G / ∼ ∣ ≤ N .
Groups in the closure and the absorption claim. We first justify precisely which group in the closure contains a stable subsemigroup. In general, the identity of such a group need not be I d I_d I d but is rather an idempotent matrix.
Lemma 2. Let K , U ≤ Q ‾ d K,U \le\overline{\mathbb{Q}}^d K , U ≤ Q d satisfy Q ‾ d = K ⊕ U \overline{\mathbb{Q}}^d = K \oplus U Q d = K ⊕ U . The matrices with kernel K K K and image U U U form an algebraic group
H ( K , U ) : = { a ∈ Mat d ( Q ‾ ) : ker ( a ) = K , im ( a ) = U } . \mathcal{H}(K,U) := \{a \in\operatorname{Mat}_d(\overline{\mathbb{Q}}) : \operatorname{ker}(a) = K,\ \operatorname{im}(a) = U\}. H ( K , U ) := { a ∈ Mat d ( Q ) : ker ( a ) = K , im ( a ) = U } . The identity of this group is the projection e e e onto U U U along K K K . If S S S is a nonempty subsemigroup of C ∩ H ( K , U ) C \cap\mathcal{H}(K,U) C ∩ H ( K , U ) , its Zariski closure inside H ( K , U ) \mathcal{H}(K,U) H ( K , U ) is a group contained in C C C .
Proof. Restriction to U U U identifies H ( K , U ) \mathcal{H}(K,U) H ( K , U ) with GL ( U ) \operatorname{GL}(U) GL ( U ) : an invertible linear map on U U U extends uniquely by zero on K K K . In a basis adapted to K ⊕ U K \oplus U K ⊕ U , these are precisely the matrices with zero block on K K K and invertible block on U U U . This also shows that H ( K , U ) \mathcal{H}(K,U) H ( K , U ) is locally closed in matrix space and that its group operations are regular. The identity extends the identity on U U U by zero on K K K . When U = { 0 } U = \{0\} U = { 0 } , this is the one-element group { 0 } \{0\} { 0 } .
We use the fact that a nonempty Zariski-closed subsemigroup of an algebraic group is a group; see also [11 Lemma 9] . By continuity of matrix multiplication, the closure of a subsemigroup is again a subsemigroup, both in matrix space and inside an algebraic group. To see this, write S ‾ \overline{S} S for the closure in the space concerned. For s ∈ S ‾ s \in\overline{S} s ∈ S , continuity of left multiplication and s S ⊆ S sS \subseteq S s S ⊆ S give s S ‾ ⊆ S ‾ s\overline{S} \subseteq\overline{S} s S ⊆ S . For b ∈ S ‾ b \in\overline{S} b ∈ S , we then have S ‾ b ⊆ S ‾ \overline{S}b \subseteq\overline{S} S b ⊆ S ; continuity of right multiplication gives S ‾ b ⊆ S ‾ \overline{S}b \subseteq\overline{S} S b ⊆ S . In particular, C C C is a semigroup. The closure of S S S inside H ( K , U ) \mathcal{H}(K,U) H ( K , U ) is thus a closed subsemigroup of an algebraic group, and hence a group. Since C ∩ H ( K , U ) C \cap\mathcal{H}(K,U) C ∩ H ( K , U ) is relatively closed and contains S S S , it contains this relative closure as well.
Claim 3.5. Given a , b ∈ ⟨ X ⟩ a,b \in\langle X\rangle a , b ∈ ⟨ X ⟩ , the condition im ( a ) = im ( a b ) \operatorname{im}(a) = \operatorname{im}(ab) im ( a ) = im ( ab ) implies a ∼ a b n a \sim ab^n a ∼ a b n for some n ≥ 1 n \ge1 n ≥ 1 .
Proof. First let c ∈ ⟨ X ⟩ c \in\langle X\rangle c ∈ ⟨ X ⟩ be rank-stable. Its kernel and image have intersection zero, so Lemma 2 applies to ⟨ c ⟩ \langle c\rangle ⟨ c ⟩ inside H ( ker ( c ) , im ( c ) ) \mathcal{H}(\operatorname{ker}(c),\operatorname{im}(c)) H ( ker ( c ) , im ( c )) . It gives a group G ≤ C G \le C G ≤ C containing c c c , with identity e c e_c e c , the projection onto im ( c ) \operatorname{im}(c) im ( c ) along ker ( c ) \operatorname{ker}(c) ker ( c ) . The restriction of a semigroup congruence to a group has a group as its quotient: the classes of the identity and inverses yield the identity and inverses in the quotient. Thus G / ∼ G/\sim G / ∼ is a finite group of order at most N N N . The order of the class of c c c gives an integer 1 ≤ q ≤ N 1 \le q \le N 1 ≤ q ≤ N such that
c q ∼ e c . (17) c^q \sim e_c. \tag*{(17)} c q ∼ e c . (17) Now suppose im ( a ) = im ( a b ) \operatorname{im}(a) = \operatorname{im}(ab) im ( a ) = im ( ab ) and put U = im ( a ) U = \operatorname{im}(a) U = im ( a ) . The equality says U b = U Ub = U U b = U , so b b b restricts to a surjective, and hence invertible, linear map on U U U . Iterating gives
U = U b d ⊆ im ( b d ) . U = Ub^d \subseteq\operatorname{im}(b^d). U = U b d ⊆ im ( b d ) . The matrix c = b d c = b^d c = b d is rank-stable. Indeed, the descending sequence of images of powers of b b b stabilises in at most d d d steps. Hence im ( b d ) = im ( b 2 d ) \operatorname{im}(b^d) = \operatorname{im}(b^{2d}) im ( b d ) = im ( b 2 d ) . The projection e c e_c e c fixes im ( b d ) \operatorname{im}(b^d) im ( b d ) pointwise, and therefore fixes U U U . Multiplying (16) by a a a gives
a b d q = a c q ∼ a e c = a . ab^{dq} = ac^q \sim ae_c = a. a b d q = a c q ∼ a e c = a . This proves the claim, with exponent $dq \ge 1.
Short products with the same kernel and image. We next make precise the exterior-algebra step used to construct insertions at a cut. The following is the form of [12 Lemma 8] needed here. We include its proof, which also gives the slightly stronger bound ( d r ) + 1 \binom{d}{r}+1 ( r d ) + 1 .
Lemma 3. Let a 1 , … , a ℓ ∈ Mat d ( Q ) a_1,\ldots,a_\ell\in\operatorname{Mat}_d(\mathbb{Q}) a 1 , … , a ℓ ∈ Mat d ( Q ) and their product a a a all have rank r r r . There are indices $1=i_1<\cdots<i_t=\ell$, with t ≤ ( d r ) + 1 ≤ 2 d + 1 t\le\binom{d}{r}+1\le2^d+1 t ≤ ( r d ) + 1 ≤ 2 d + 1 , such that a i 1 ⋯ a i t a_{i_1}\cdots a_{i_t} a i 1 ⋯ a i t has the same kernel and image as a a a . For ℓ = 1 \ell=1 ℓ = 1 , take the single index $i_1=1.
Proof. Choose a shortest subsequence that includes both endpoints and whose product still has rank r r r ; denote the matrices in this subsequence by b 1 , … , b t b_1,\ldots,b_t b 1 , … , b t . Such a subsequence exists because the full sequence has rank r r r . The product b 1 ⋯ b t b_1\cdots b_t b 1 ⋯ b t has kernel ker ( a 1 ) \ker(a_1) ker ( a 1 ) and image im ( a ℓ ) \operatorname{im}(a_\ell) im ( a ℓ ) , i.e., the kernel and image are those of a a a by Claim A.2. It remains to bound t t t . There is nothing to prove if t = 1 t=1 t = 1 .
For 1 ≤ i < t 1\le i<t 1 ≤ i < t , choose a wedge u i u_i u i representing im ( b i ) \operatorname{im}(b_i) im ( b i ) , and let φ i \varphi_i φ i be the functional associated with ker ( b i + 1 ) \ker(b_{i+1}) ker ( b i + 1 ) by Claim A.1. Since the selected product preserves rank, φ i ( u i ) ≠ 0 \varphi_i(u_i)\ne0 φ i ( u i ) = 0 . For 1 ≤ i < j < t 1\le i<j<t 1 ≤ i < j < t , we must have φ j ( u i ) = 0 \varphi_j(u_i)=0 φ j ( u i ) = 0 . Otherwise, b i b j + 1 b_i b_{j+1} b i b j + 1 would have rank r r r . All other adjacent joins in
b 1 ⋯ b i b j + 1 ⋯ b t b_1\cdots b_i b_{j+1}\cdots b_t b 1 ⋯ b i b j + 1 ⋯ b t
already preserve rank, so Claim A.2 would make this a shorter rank-r r r product retaining both endpoints, contrary to the choice of the subsequence.
The square matrix ( φ j ( u i ) ) 1 ≤ j , i < t (\varphi_j(u_i))_{1\le j,i<t} ( φ j ( u i ) ) 1 ≤ j , i < t is therefore triangular with nonzero diagonal, and is invertible. In particular, u 1 , … , u t − 1 u_1,\ldots,u_{t-1} u 1 , … , u t − 1 are linearly independent: applying the φ j \varphi_j φ j to a relation among these vectors gives an invertible linear system for its coefficients. Since they lie in ⋀ r Q d \bigwedge^r\mathbb{Q}^d ⋀ r Q d , we obtain
t − 1 ≤ ( d r ) , t-1\le\binom{d}{r}, t − 1 ≤ ( r d ) ,
as required. □
Insertions and regrouping at one node. For the next lemma and the forest construction, only the conclusion of Claim 3.5 is needed. Thus we assume that ∼ \sim ∼ is a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ satisfying
im ( p ) = im ( p b ) ⟹ p ∼ p b n for some n ≥ 1 , p , b ∈ ⟨ X ⟩ . (18) \operatorname{im}(p)=\operatorname{im}(pb)\quad\Longrightarrow\quad p\sim pb^n\text{ for some }n\ge1,\qquad p,b\in\langle X\rangle. \tag*{(18)} im ( p ) = im ( p b ) ⟹ p ∼ p b n for some n ≥ 1 , p , b ∈ ⟨ X ⟩ . (18)
In the theorem, we apply this to the restriction of the congruence on C C C . The construction follows the strategy of [12 Proposition 9] , replacing equalities obtained from finiteness there by the congruences in (17).
Lemma 4. Let a 1 , … , a s ∈ ⟨ X ⟩ a_1,\ldots,a_s\in\langle X\rangle a 1 , … , a s ∈ ⟨ X ⟩ , where s ≥ 3 s\ge3 s ≥ 3 , and put a = a 1 ⋯ a s a=a_1\cdots a_s a = a 1 ⋯ a s . Suppose
rank ( a 1 ) = ⋯ = rank ( a s ) = rank ( a ) = rank ( a 2 ) . \operatorname{rank}(a_1)=\cdots=\operatorname{rank}(a_s)=\operatorname{rank}(a)=\operatorname{rank}(a^2). rank ( a 1 ) = ⋯ = rank ( a s ) = rank ( a ) = rank ( a 2 ) . There are matrices u 1 , … , u t ∈ ⟨ X ⟩ u_1,\ldots,u_t\in\langle X\rangle u 1 , … , u t ∈ ⟨ X ⟩ , with t ≥ 3 t\ge3 t ≥ 3 , such that a ∼ u 1 ⋯ u t a\sim u_1\cdots u_t a ∼ u 1 ⋯ u t , every u j u_j u j has the kernel and image of a a a , and each u j u_j u j is a product of at most 2 d + 2 2^{d+2} 2 d + 2 matrices chosen, with repetition allowed, from a 1 , … , a s a_1,\ldots,a_s a 1 , … , a s . In particular, the u j u_j u j generate a stable subsemigroup.
Proof. Write r = rank ( a ) r=\operatorname{rank}(a) r = rank ( a ) , K = ker ( a ) K=\ker(a) K = ker ( a ) , and U = im ( a ) U=\operatorname{im}(a) U = im ( a ) . Since a a a is rank-stable, K ∩ U = { 0 } K\cap U=\{0\} K ∩ U = { 0 } . At each cut i ∈ { 1 , … , s − 1 } , i\in\{1,\ldots,s-1\}, i ∈ { 1 , … , s − 1 } , let
p i = a 1 ⋯ a i , q i = a i + 1 ⋯ a s . p_i=a_1\cdots a_i,\qquad q_i=a_{i+1}\cdots a_s. p i = a 1 ⋯ a i , q i = a i + 1 ⋯ a s .
Apply Lemma 3 to p i p_i p i and q i q_i q i . It supplies products y i y_i y i and x i x_i x i , respectively, each using at most 2 d + 1 2^d+1 2 d + 1 of the corresponding child matrices, such that
ker ( y i ) = K , im ( y i ) = im ( a i ) , ker ( x i ) = ker ( a i + 1 ) , im ( x i ) = U . (19) \begin{aligned}
\ker(y_i)&=K, & \operatorname{im}(y_i)&=\operatorname{im}(a_i), \\
\ker(x_i)&=\ker(a_{i+1}), & \operatorname{im}(x_i)&=U. \tag*{(19)}
\end{aligned} ker ( y i ) ker ( x i ) = K , = ker ( a i + 1 ) , im ( y i ) im ( x i ) = im ( a i ) , = U . (19)
These auxiliary products reproduce the kernels and images; they need not be congruent to the prefixes or suffixes.
Claim E.1. For every cut i i i , im ( p i x i y i ) = im ( p i ) \operatorname{im}(p_i x_i y_i)=\operatorname{im}(p_i) im ( p i x i y i ) = im ( p i ) .
Proof. The product p i x i p_i x_i p i x i has rank r r r , since im ( p i ) = im ( a i ) \operatorname{im}(p_i)=\operatorname{im}(a_i) im ( p i ) = im ( a i ) and ker ( x i ) = ker ( a i + 1 ) \operatorname{ker}(x_i)=\operatorname{ker}(a_{i+1}) ker ( x i ) = ker ( a i + 1 ) , and the original join a i a i + 1 a_i a_{i+1} a i a i + 1 preserves rank. Its image is U U U . The next multiplication by y i y_i y i also preserves rank, since U ∩ ker ( y i ) = U ∩ K = { 0 } U \cap\operatorname{ker}(y_i)=U \cap K=\{0\} U ∩ ker ( y i ) = U ∩ K = { 0 } . The resulting image is im ( y i ) = im ( a i ) = im ( p i ) \operatorname{im}(y_i)=\operatorname{im}(a_i)=\operatorname{im}(p_i) im ( y i ) = im ( a i ) = im ( p i ) .
By (17), choose n i ≥ 1 n_i \ge1 n i ≥ 1 with p i ∼ p i ( x i y i ) n i p_i \sim p_i(x_i y_i)^{n_i} p i ∼ p i ( x i y i ) n i . We insert these powers in order from cut s − 1 s-1 s − 1 to cut 1. When processing cut i i i , every power already inserted is to its right, so the product still begins with the original prefix p i p_i p i . Multiplication by the remaining suffix therefore preserves the congruence used for that insertion. After all cuts have been processed, we obtain
a ∼ a 1 ( x 1 y 1 ) n 1 a 2 ( x 2 y 2 ) n 2 ⋯ a s − 1 ( x s − 1 y s − 1 ) n s − 1 a s . (20) a \sim a_1(x_1y_1)^{n_1}a_2(x_2y_2)^{n_2}\cdots a_{s-1}(x_{s-1}y_{s-1})^{n_{s-1}}a_s. \tag*{(20)} a ∼ a 1 ( x 1 y 1 ) n 1 a 2 ( x 2 y 2 ) n 2 ⋯ a s − 1 ( x s − 1 y s − 1 ) n s − 1 a s . (20)
Regroup the product on the right into the following ordered list of factors. Start with a 1 x 1 a_1x_1 a 1 x 1 . For i = 1 , … , s − 1 i=1,\ldots,s-1 i = 1 , … , s − 1 , append n i − 1 n_i-1 n i − 1 copies of y i x i y_i x_i y i x i , followed by y i a i + 1 x i + 1 y_i a_{i+1}x_{i+1} y i a i + 1 x i + 1 if i < s − 1 i<s-1 i < s − 1 , and by y s − 1 a s y_{s-1}a_s y s − 1 a s if i = s − 1 i=s-1 i = s − 1 . When n i = 1 n_i=1 n i = 1 , no copy of y i x i y_i x_i y i x i is appended. This list has t = s + ∑ i = 1 s − 1 ( n i − 1 ) ≥ s t=s+\sum_{i=1}^{s-1}(n_i-1)\ge s t = s + ∑ i = 1 s − 1 ( n i − 1 ) ≥ s factors, and its product is exactly the right-hand side of (20), using ( x i y i ) n i = x i ( y i x i ) n i − 1 y i (x_i y_i)^{n_i}=x_i(y_i x_i)^{n_i-1}y_i ( x i y i ) n i = x i ( y i x i ) n i − 1 y i as a regrouping identity.
Claim E.2. Every factor in this list has rank r r r , kernel K K K , and image U U U .
Proof. For a 1 x 1 a_1x_1 a 1 x 1 , the join preserves rank by ker ( x 1 ) = ker ( a 2 ) \operatorname{ker}(x_1)=\operatorname{ker}(a_2) ker ( x 1 ) = ker ( a 2 ) , and the endpoints give kernel K K K and image U U U . For y i x i y_i x_i y i x i , the join is the original cut i i i expressed through (18)–(19), so it preserves rank and has the same endpoints K , U K,U K , U . For y i a i + 1 x i + 1 y_i a_{i+1}x_{i+1} y i a i + 1 x i + 1 , the first join preserves rank because im ( y i ) = im ( a i ) \operatorname{im}(y_i)=\operatorname{im}(a_i) im ( y i ) = im ( a i ) , and the second because ker ( x i + 1 ) = ker ( a i + 2 ) \operatorname{ker}(x_{i+1})=\operatorname{ker}(a_{i+2}) ker ( x i + 1 ) = ker ( a i + 2 ) . Its kernel is that of y i y_i y i and its image that of x i + 1 x_{i+1} x i + 1 , again K , U K,U K , U . The final factor y s − 1 a s y_{s-1}a_s y s − 1 a s has the required properties by a similar argument.
Call these factors u 1 , … , u t u_1,\ldots,u_t u 1 , … , u t . As K K K and U U U are complementary, products of the u j u_j u j also have kernel K K K , image U U U , and rank r r r . They therefore generate a stable subsemigroup. Each u j u_j u j consists of at most two auxiliary products and at most one original child matrix. Thus it uses at most
2 ( 2 d + 1 ) + 1 ≤ 2 d + 2 2(2^d+1)+1\le2^{d+2} 2 ( 2 d + 1 ) + 1 ≤ 2 d + 2
matrices among a 1 , … , a s a_1,\ldots,a_s a 1 , … , a s , completing the proof.
Modifying the whole tree. Recall that a stable ∼ \sim ∼ -factorisation tree has its root labelled by the given matrix, its leaves have labels in X X X , and every internal label is congruent to the product of its children’s labels. An internal node has at least two children; if it has at least three, their labels generate a stable subsemigroup. Its reduced height is the maximum number of nodes with at least three children on a path from the root to a leaf.
Lemma 1. Suppose ∼ \sim ∼ is a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ such that, for all a , b ∈ ⟨ X ⟩ a,b\in\langle X\rangle a , b ∈ ⟨ X ⟩ , the equality im ( a ) = im ( a b ) \operatorname{im}(a)=\operatorname{im}(ab) im ( a ) = im ( ab ) implies a ∼ a b n a\sim ab^n a ∼ a b n for some n ≥ 1 n\ge1 n ≥ 1 . Then ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ admits a stable ∼ \sim ∼ -factorisation forest whose trees have height at most 2 d ( d + 3 ) 2d(d+3) 2 d ( d + 3 ) and reduced height at most d d d .
Proof. For each a ∈ ⟨ X ⟩ a\in\langle X\rangle a ∈ ⟨ X ⟩ , choose a product over X X X equal to a a a . By [12 Theorem 7] , it has a rank-stable factorisation tree of height at most d ( d + 3 ) d(d+3) d ( d + 3 ) and reduced height at most d d d such that, at every node with at least three children, the child labels a 1 , … , a s a_1,\ldots,a_s a 1 , … , a s and the parent label satisfy
a = a 1 ⋯ a s , rank ( a 1 ) = ⋯ = rank ( a s ) = rank ( a ) = rank ( a 2 ) . a=a_1\cdots a_s,\qquad\operatorname{rank}(a_1)=\cdots=\operatorname{rank}(a_s)=\operatorname{rank}(a)=\operatorname{rank}(a^2). a = a 1 ⋯ a s , rank ( a 1 ) = ⋯ = rank ( a s ) = rank ( a ) = rank ( a 2 ) .
This is the rank-stable factorisation tree of that paper. The extra rank equalities on the children are part of this starting-tree result; they do not assert stability of the children.
We process the tree from the leaves upwards, keeping the root label of each processed subtree unchanged. A leaf needs no modification. At a binary node, attach the already modified child subtrees to the same parent. Their root labels remain the original two matrices, whose product is the parent label. At a node with at least three children, apply Lemma 4 to its original child labels. Replace its children by u 1 , … , u t u_1,\ldots,u_t u 1 , … , u t . Their product is congruent to the unchanged parent label, and they generate a stable subsemigroup.
For each u j u_j u j , choose the product of at most 2 d + 2 2^{d+2} 2 d + 2 original child matrices given by the lemma. Assemble this product using a binary tree of height at most d + 2 d+2 d + 2 , labelling its internal nodes by the exact products of their children. For each occurrence of a i a_i a i among its leaves, attach a copy of the already modified subtree rooted at a i a_i a i . If the product has one matrix, attach that subtree directly. All eventual leaves belong to X X X . Every new binary node satisfies equality with its child product, and every other internal node satisfies the required congruence. Thus the result is a stable ∼ \sim ∼ -factorisation tree rooted at a a a .
We verify the height bounds. At a modified node, the new top node replaces the original node with at least three children. Below it, only binary levels are inserted before reaching a copy of one of the original child subtrees. Consequently the modification adds at most d + 3 d+3 d + 3 levels and introduces no additional node with at least three children along a path. Every path in the modified tree follows a corresponding sequence of original child subtrees, even when some of them have been copied. The starting path has length at most d ( d + 3 ) d(d+3) d ( d + 3 ) and passes through at most d d d nodes with at least three children. Hence the new height is at most
d ( d + 3 ) + d ( d + 3 ) = 2 d ( d + 3 ) , d(d+3)+d(d+3)=2d(d+3), d ( d + 3 ) + d ( d + 3 ) = 2 d ( d + 3 ) ,
and its reduced height is at most d d d . Choosing such a tree for each a a a gives the asserted forest.
Theorem 3. Let X ⊆ Mat d ( Q ) X \subseteq\operatorname{Mat}_d(\mathbb{Q}) X ⊆ Mat d ( Q ) be finite and ∼ \sim ∼ be a congruence on ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Assume that there exists N ≥ 1 N \ge1 N ≥ 1 such that ∣ G / ∼ ∣ ≤ N |G/\sim| \le N ∣ G / ∼ ∣ ≤ N for every group G ≤ ⟨ X ⟩ G \le\langle X\rangle G ≤ ⟨ X ⟩ . Then ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ is finite and has diameter at most 2 2 d ( d + 3 ) N d 2^{2d(d+3)}N^d 2 2 d ( d + 3 ) N d .
Proof. Claim 3.5 shows that the restriction of ∼ \sim ∼ to ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ satisfies the hypothesis of Lemma 1. For every a ∈ ⟨ X ⟩ a \in\langle X\rangle a ∈ ⟨ X ⟩ , choose the resulting stable ∼ \sim ∼ -factorisation tree. Its height is at most h : = 2 d ( d + 3 ) h := 2d(d+3) h := 2 d ( d + 3 ) and its reduced height at most d d d . We prune this tree while preserving the class represented at each surviving node.
Consider a node with at least three children. The child labels generate a stable subsemigroup S ≤ ⟨ X ⟩ S \le\langle X\rangle S ≤ ⟨ X ⟩ with some common kernel and image K , U K,U K , U satisfying K ∩ U = { 0 } K \cap U = \{0\} K ∩ U = { 0 } . By Lemma 2, the relative closure of S S S inside H ( K , U ) \mathcal{H}(K,U) H ( K , U ) is a group G ≤ C G \le C G ≤ C . Therefore ∣ S / ∼ ∣ ≤ ∣ G / ∼ ∣ ≤ N |S/\sim| \le|G/\sim| \le N ∣ S / ∼ ∣ ≤ ∣ G / ∼ ∣ ≤ N . If the node has more than N N N children, two of their nonempty prefix products are congruent. We can delete the intervening children and their subtrees without changing the class of the parent product. Repeating this operation leaves at most N N N children. All retained child labels still belong to S S S , so the same argument remains applicable. Perform this pruning at each such node. Unary nodes may remain when N = 1 N=1 N = 1 ; this is harmless for counting leaves and for the congruence between a label and its child product.
For clarity, mark the nodes that originally had at least three children, even if pruning leaves them with one or two. Every unmarked internal node has at most two children; every marked node has at most N N N children. The height is at most h h h , and every path contains at most d d d marked nodes. A tree of height at most h h h with at most q q q marked nodes on each path has at most 2 h N q 2^hN^q 2 h N q leaves. This follows by induction on height: at an unmarked root sum at most two bounds 2 h − 1 N q 2^{h-1}N^q 2 h − 1 N q , and at a marked root sum at most N N N bounds 2 h N q − 1 2^hN^{q-1} 2 h N q − 1 ; a leaf contributes one. Applying this with q = d q=d q = d bounds the number of leaves by
2 2 d ( d + 3 ) N d . (21) 2^{2d(d+3)}N^d. \tag*{(21)} 2 2 d ( d + 3 ) N d . (21)
Every remaining leaf is labelled by a matrix in X X X . Induction from the leaves upwards, using the congruence at each internal node, shows that their ordered product is congruent to the root label a a a . Thus every class has length at most (21). This proves the diameter estimate and, since X X X is finite, the finiteness of ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ .
Detailed Proof of Theorem 4
Theorem 4. Let w ≥ 2 w \ge2 w ≥ 2 and d ≥ 3 d \ge3 d ≥ 3 and m ≥ 2 w + 7 m \ge2w+7 m ≥ 2 w + 7 . The class A w , d \mathcal{A}_{w,d} A w , d of width-w w w degree-d d d homogeneous ABPs is separated from A ‾ w , d \overline{\mathcal{A}}_{w,d} A w , d .
Proof. Recall that v ( θ ) v(\theta) v ( θ ) lists all parameter monomials obtained by choosing at most one parameter from each layer. Translating such a monomial produces a linear combination of monomials in the same list. This proves the existence and uniqueness of the matrix ρ t \rho_t ρ t satisfying
v ( θ ) ρ t = v ( θ + t ) . v(\theta)\rho_t = v(\theta+t). v ( θ ) ρ t = v ( θ + t ) .
Its entries are polynomials in t t t with integer coefficients. Moreover, the entry in the row indexed by the constant monomial and the column indexed by θ i \theta_i θ i is t i t_i t i . Thus the parameter values can be recovered from its entries. Composing translations gives
ρ s ρ t = ρ s + t , ρ 0 = I L θ , ρ t − 1 = ρ − t . \rho_s\rho_t=\rho_{s+t},\qquad\rho_0=I_{L_\theta},\qquad\rho_t^{-1}=\rho_{-t}. ρ s ρ t = ρ s + t , ρ 0 = I L θ , ρ t − 1 = ρ − t .
Recall also the coefficient-evaluation identity from the overview:
v ( 0 ) ρ t C U = coeff U ( t , x ) . (22) v(0)\rho_t C_U=\operatorname{coeff}_U(t,x). \tag*{(22)} v ( 0 ) ρ t C U = coeff U ( t , x ) . (22)
We record the multiplication rules used below. Block multiplication, together with (22) , gives
Γ s Γ t = Γ s + t , Γ t V 0 = V 0 , C Γ t = C , V 0 2 = 0 , C 2 = 0 , C V 0 = 0 , \begin{aligned}
\Gamma_s\Gamma_t&=\Gamma_{s+t},\qquad\Gamma_tV_0=V_0,\qquad C\Gamma_t=C,\\
V_0^2&=0,\qquad C^2=0,\qquad CV_0=0,
\end{aligned} Γ s Γ t V 0 2 = Γ s + t , Γ t V 0 = V 0 , C Γ t = C , = 0 , C 2 = 0 , C V 0 = 0 ,
and
( V 0 Γ s ) ( Γ t C ) = ι ( U ( s + t , x ) ) . (23) (V_0\Gamma_s)(\Gamma_tC)=\iota\bigl(U(s+t,x)\bigr). \tag*{(23)} ( V 0 Γ s ) ( Γ t C ) = ι ( U ( s + t , x ) ) . (23)
For every homogeneous polynomial f f f of degree d d d , we also have
Γ t ι ( f ) = ι ( f ) Γ t = ι ( f ) . \Gamma_t\iota(f)=\iota(f)\Gamma_t=\iota(f). Γ t ι ( f ) = ι ( f ) Γ t = ι ( f ) .
Multiplication of ι ( f ) \iota(f) ι ( f ) , on either side, by V 0 Γ t V_0\Gamma_t V 0 Γ t , Γ t C \Gamma_tC Γ t C , or another encoded polynomial gives 0 0 0 . Among two factors of the forms V 0 Γ s V_0\Gamma_s V 0 Γ s and Γ t C \Gamma_tC Γ t C , the only potentially nonzero product is the one in (23) .
These rules show that every element of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ has one of the forms
Γ n , V 0 Γ n , Γ n C , ι ( U ( n , x ) ) , n ∈ Z ν . (24) \Gamma_n,\qquad V_0\Gamma_n,\qquad\Gamma_nC,\qquad\iota\bigl(U(n,x)\bigr),\qquad n\in\mathbb{Z}^{\nu}. \tag*{(24)} Γ n , V 0 Γ n , Γ n C , ι ( U ( n , x ) ) , n ∈ Z ν . (24)
Conversely, each matrix in this list belongs to ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . Indeed, the generators Γ e i \Gamma_{e_i} Γ e i and Γ − e i \Gamma_{-e_i} Γ − e i produce every integer translation. They also produce the identity, and V 0 V_0 V 0 , C C C are generators. The last family is obtained by multiplying V 0 V_0 V 0 , an integer translation, and C C C . The zero matrix is included, since U ( 0 , x ) = 0 U(0,x)=0 U ( 0 , x ) = 0 .
Claim 4.1. ⟨ X ⟩ ‾ = { Γ t ‾ , V 0 Γ t ‾ , Γ t C ‾ : t ∈ Q ‾ ν } ∪ ˙ ι ( A w , d ‾ ) \overline{\langle X\rangle}=\{\overline{\Gamma_t},\overline{V_0\Gamma_t},\overline{\Gamma_tC}:t\in\overline{\mathbb{Q}}^\nu\}\mathbin{\dot\cup}\iota(\overline{\mathcal{A}_{w,d}}) ⟨ X ⟩ = { Γ t , V 0 Γ t , Γ t C : t ∈ Q ν } ∪ ˙ ι ( A w , d ) .
Proof. We take the closures of the four families in Equation (24) . First, Z ν \mathbb{Z}^\nu Z ν is Zariski dense in Q ‾ ν \overline{\mathbb{Q}}^\nu Q ν : a polynomial vanishing on all integer tuples is zero, by applying the corresponding univariate fact successively to its variables. It follows that the integer points in each of our parameterized families have the same closure as the family obtained using all algebraic parameter values. More explicitly, any polynomial in matrix entries which vanishes on the integer-parameter family pulls back to a polynomial vanishing on Z ν \mathbb{Z}^\nu Z ν , and hence on the entire parameter space.
The family { Γ t : t ∈ Q ‾ ν } \{\Gamma_t:t\in\overline{\mathbb{Q}}^\nu\} { Γ t : t ∈ Q ν } is closed. Its parameters occur as matrix entries, and all remaining entries are polynomial functions of them. It is therefore a graph of polynomial functions, with additional fixed coordinates. The same reasoning applies to V 0 Γ t V_0\Gamma_t V 0 Γ t , whose nonzero block is v ( t ) v(t) v ( t ) and contains every coordinate t i t_i t i . It also applies to Γ t C \Gamma_tC Γ t C : its block in block row 3 and block column 4 is ( 0 , ρ t ) (0,\rho_t) ( 0 , ρ t ) , so the parameters are again recoverable from matrix entries.
For the output family, the map t ↦ U ( t , x ) t\mapsto U(t,x) t ↦ U ( t , x ) is polynomial and has image A w , d \mathcal{A}_{w,d} A w , d . The map ι \iota ι identifies the coefficient space Q ‾ [ x ] d \overline{\mathbb{Q}}[x]_d Q [ x ] d with a linear subspace of the matrix space. Its inverse on that subspace reads off the coefficient row, so this identification preserves Zariski closures. Consequently,
{ ι ( U ( n , x ) ) : n ∈ Z ν } ‾ = ι ( A w , d ‾ ) . \overline{\{\iota(U(n,x)):n\in\mathbb{Z}^\nu\}}=\iota(\overline{\mathcal{A}_{w,d}}). { ι ( U ( n , x )) : n ∈ Z ν } = ι ( A w , d ) .
The closure of a finite union is the union of its closures, which gives the claimed equality. The four families are disjoint, as can be seen from their blocks. A translation has top-left entry 1. A matrix V 0 Γ t V_0\Gamma_t V 0 Γ t has the nonzero block v ( t ) v(t) v ( t ) in position ( 1 , 2 ) (1,2) ( 1 , 2 ) . A matrix Γ t C \Gamma_tC Γ t C has the nonzero block ( 0 , ρ t ) (0,\rho_t) ( 0 , ρ t ) in position ( 3 , 4 ) (3,4) ( 3 , 4 ) . An encoded polynomial has zero entries outside block ( 1 , 4 ) (1,4) ( 1 , 4 ) . These observations also justify the disjoint-union notation in the claim. □
Recall that ∼ \sim ∼ is the least congruence on ⟨ X ⟩ ‾ \overline{\langle X\rangle} ⟨ X ⟩ satisfying Γ t ∼ I L t o t \Gamma_t\sim I_{L_{\mathrm{tot}}} Γ t ∼ I L tot for every t ∈ Q ‾ ν t\in\overline{\mathbb{Q}}^\nu t ∈ Q ν . We describe all its classes in order to prove the next two claims.
Claim 4.2. The class of zero matrix consists of the encoding of polynomials computable by ABPs, that is,
[ 0 ] ∼ = ι ( A w , d ) . [0]_\sim=\iota(\mathcal{A}_{w,d}). [ 0 ] ∼ = ι ( A w , d ) .
Proof. The defining relations imply
V 0 Γ t ∼ V 0 , Γ t C ∼ C , ι ( U ( t , x ) ) = V 0 Γ t C ∼ V 0 C = 0. V_0\Gamma_t\sim V_0,\qquad\Gamma_tC\sim C,\qquad\iota(U(t,x))=V_0\Gamma_tC\sim V_0C=0. V 0 Γ t ∼ V 0 , Γ t C ∼ C , ι ( U ( t , x )) = V 0 Γ t C ∼ V 0 C = 0.
Since every polynomial in A w , d \mathcal{A}_{w,d} A w , d is an output of the universal ABP, this proves ι ( A w , d ) ⊆ [ 0 ] ∼ \iota(\mathcal{A}_{w,d})\subseteq[0]_\sim ι ( A w , d ) ⊆ [ 0 ] ∼ . To prove the reverse inclusion, consider the equivalence relation ≈ \approx ≈ whose classes are
{ Γ t : t ∈ Q ‾ ν } , { V 0 Γ t : t ∈ Q ‾ ν } , \{\Gamma_t:t\in\overline{\mathbb{Q}}^\nu\},\qquad
\{V_0\Gamma_t:t\in\overline{\mathbb{Q}}^\nu\}, { Γ t : t ∈ Q ν } , { V 0 Γ t : t ∈ Q ν } ,
{ Γ t C : t ∈ Q ‾ ν } , ι ( A w , d ) , \{\Gamma_tC:t\in\overline{\mathbb{Q}}^\nu\},\qquad
\iota(\mathcal{A}_{w,d}), { Γ t C : t ∈ Q ν } , ι ( A w , d ) ,
together with the singleton { ι ( f ) } \{\iota(f)\} { ι ( f )} for each f ∈ A w , d ‾ ∖ A w , d f\in\overline{\mathcal{A}_{w,d}}\setminus\mathcal{A}_{w,d} f ∈ A w , d ∖ A w , d . Claim 4.1 shows that these sets partition the entire matrix closure.
This partition defines a semigroup congruence. Indeed, the multiplication rules above determine the family of a product from the families of its factors. The only product of two nontranslation auxiliary families which may be nonzero is ( V 0 Γ s ) ( Γ t C ) (V_0\Gamma_s)(\Gamma_tC) ( V 0 Γ s ) ( Γ t C ) ; its output always belongs to ι ( A w , d ) \iota(\mathcal{A}_{w,d}) ι ( A w , d ) . Translations fix every encoded polynomial, and all other matrices annihilate encoded polynomials on either side. Thus replacing either factor by an ≈ \approx ≈ -equivalent matrix preserves the ≈ \approx ≈ -class of the product. Furthermore, ≈ \approx ≈ satisfies Γ t ≈ I L t o t \Gamma_t \approx I_{L_{\mathrm{tot}}} Γ t ≈ I L tot . By the minimality of ∼ \sim ∼ , we have ∼ ⊆ ≈ \sim\subseteq\approx ∼⊆≈ . In particular, the ∼ \sim ∼ -class of zero is contained in its ≈ \approx ≈ -class, namely ι ( A w , d ) \iota(\mathcal{A}_{w,d}) ι ( A w , d ) . This proves the desired equality.
The relations at the beginning of the proof also show that each of the four displayed sets lies in one ∼ \sim ∼ -class. Since the remaining ≈ \approx ≈ -classes are singletons, we in fact have ∼ = ≈ \sim= \approx ∼=≈ . In particular, every encoded polynomial outside ι ( A w , d ) \iota(\mathcal{A}_{w,d}) ι ( A w , d ) forms a singleton class.
Claim 4.3. The quotient ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ has exactly four classes, represented by I L t o t I_{L_{\mathrm{tot}}} I L tot , V 0 V_0 V 0 , C C C , and 0 \mathbf{0} 0 .
Proof. By Equation (24) , every element of ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ belongs to one of the four classes described above. Their representatives I L t o t I_{L_{\mathrm{tot}}} I L tot , V 0 V_0 V 0 , C C C , 0 \mathbf{0} 0 all belong to ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ : the identity is a product of opposite translations, and 0 = V 0 2 \mathbf{0}=V_0^2 0 = V 0 2 . The four classes are distinct by the explicit description of ∼ \sim ∼ . Hence the quotient has exactly four elements.
We also verify the application of Theorem 3 in the overview. The translations form a group. Every other matrix in the closure has square zero, by its block form. If a matrix group contains such a matrix, it contains 0 \mathbf{0} 0 and must be the singleton group { 0 } \{\mathbf{0}\} { 0 } . To see the last assertion, if its identity is e e e , the inverse of 0 \mathbf{0} 0 in the group would give e = 0 e=\mathbf{0} e = 0 , after which every group element equals 0 \mathbf{0} 0 . Thus every group in the closure is contained in the translation group or is { 0 } \{\mathbf{0}\} { 0 } . Its quotient under ∼ \sim ∼ is therefore a singleton. The theorem applies with N = 1 N=1 N = 1 and gives finiteness of ⟨ X ⟩ / ∼ \langle X\rangle/\sim ⟨ X ⟩ / ∼ , consistently with the exact count just proved.
For clarity, write
G : = ( p q r s ) , Δ : = p s − q r , ( ξ k η k ) : = G k ( x y ) . G := \begin{pmatrix} p & q \\ r & s \end{pmatrix}, \qquad\Delta:= ps-qr, \qquad\begin{pmatrix} \xi_k \\ \eta_k \end{pmatrix} := G^k \begin{pmatrix} x \\ y \end{pmatrix}. G := ( p r q s ) , Δ := p s − q r , ( ξ k η k ) := G k ( x y ) .
We had defined the polynomials F w , d F_{w,d} F w , d as follows.
F 2 , 2 k + 3 = a ξ k 2 + b η k 2 + z Δ k ( u x + v y ) , F_{2,2k+3}=a\xi_k^2+b\eta_k^2+z\Delta^k(ux+vy), F 2 , 2 k + 3 = a ξ k 2 + b η k 2 + z Δ k ( ux + v y ) ,
F 2 , 2 k + 4 = a ξ k + 1 ξ k + b η k + 1 η k + Δ k + 1 ( u x + v y ) , F_{2,2k+4}=a\xi_{k+1}\xi_k+b\eta_{k+1}\eta_k+\Delta^{k+1}(ux+vy), F 2 , 2 k + 4 = a ξ k + 1 ξ k + b η k + 1 η k + Δ k + 1 ( ux + v y ) ,
and
F w , d = F 2 , d + ∑ i = 1 w − 2 c i z i d − 1 . F_{w,d}=F_{2,d}+\sum_{i=1}^{w-2}c_i z_i^{d-1}. F w , d = F 2 , d + i = 1 ∑ w − 2 c i z i d − 1 .
These polynomials use 2 w + 7 2w+7 2 w + 7 variables (When more variables are available, we regard these polynomials as independent of the additional variables).
Claim 4.4. F w , d ∈ A ‾ w , d F_{w,d}\in\overline{\mathcal{A}}_{w,d} F w , d ∈ A w , d .
Proof. We first construct the approximations at width two. Introduce the two matrices
D : = ( a 0 0 b ) , J : = ( 0 1 − 1 0 ) . D := \begin{pmatrix} a & 0 \\ 0 & b \end{pmatrix}, \qquad J := \begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix}. D := ( a 0 0 b ) , J := ( 0 − 1 1 0 ) . For any 2 × 2 2\times2 2 × 2 matrix M M M over a commutative ring, direct multiplication gives
M T J M = ( det M ) J . (25) M^{\mathsf T}JM=(\det M)J. \tag*{(25)} M T J M = ( det M ) J . (25) We also have
( x + ε v , y − ε u ) J ( x y ) = ε ( u x + v y ) . (26) (x+\varepsilon v,\ y-\varepsilon u)J\begin{pmatrix} x \\ y \end{pmatrix}=\varepsilon(ux+vy). \tag*{(26)} ( x + ε v , y − ε u ) J ( x y ) = ε ( ux + v y ) . (26)
These identities explain the cancellation of the negative powers of ε \varepsilon ε in both approximations. Suppose first that d = 2 k + 3 d=2k+3 d = 2 k + 3 . The product in the overview is
( x + ε v , y − ε u ) ( G T ) k ( D + ε − 1 z J ) G k ( x y ) . (x+\varepsilon v,y-\varepsilon u)(G^{\mathsf{T}})^k(D+\varepsilon^{-1}zJ)G^k\begin{pmatrix}x\\y\end{pmatrix}. ( x + ε v , y − ε u ) ( G T ) k ( D + ε − 1 z J ) G k ( x y ) .
Its part involving D D D equals
a ξ k 2 + b η k 2 + ε ( v , − u ) ( G T ) k D G k ( x y ) . a\xi_k^2+b\eta_k^2+\varepsilon(v,-u)(G^{\mathsf{T}})^kDG^k\begin{pmatrix}x\\y\end{pmatrix}. a ξ k 2 + b η k 2 + ε ( v , − u ) ( G T ) k D G k ( x y ) .
By Equations (25) and (26), its remaining part is exactly z Δ k ( u x + v y ) z\Delta^k(ux+vy) z Δ k ( ux + v y ) . The output is therefore F 2 , 2 k + 3 + ε H 2 , 2 k + 3 F_{2,2k+3}+\varepsilon H_{2,2k+3} F 2 , 2 k + 3 + ε H 2 , 2 k + 3 , where
H 2 , 2 k + 3 : = ( v , − u ) ( G T ) k D G k ( x y ) . H_{2,2k+3}:=(v,-u)(G^{\mathsf{T}})^kDG^k\begin{pmatrix}x\\y\end{pmatrix}. H 2 , 2 k + 3 := ( v , − u ) ( G T ) k D G k ( x y ) .
Now suppose that d = 2 k + 4 d=2k+4 d = 2 k + 4 . The middle matrix of the approximation can be written as
( a + r / ε s / ε − p / ε b − q / ε ) = D + ε − 1 adj ( G T ) J , \begin{pmatrix}a+r/\varepsilon&s/\varepsilon\\-p/\varepsilon&b-q/\varepsilon\end{pmatrix}=D+\varepsilon^{-1}\operatorname{adj}(G^{\mathsf{T}})J, ( a + r / ε − p / ε s / ε b − q / ε ) = D + ε − 1 adj ( G T ) J ,
where adj \operatorname{adj} adj denotes the adjugate matrix. The identity G T adj ( G T ) = Δ I 2 G^{\mathsf{T}}\operatorname{adj}(G^{\mathsf{T}})=\Delta I_2 G T adj ( G T ) = Δ I 2 gives
( G T ) k + 1 adj ( G T ) J G k = Δ k + 1 J . (G^{\mathsf{T}})^{k+1}\operatorname{adj}(G^{\mathsf{T}})JG^k=\Delta^{k+1}J. ( G T ) k + 1 adj ( G T ) J G k = Δ k + 1 J .
Consequently, the part with coefficient ε − 1 \varepsilon^{-1} ε − 1 contributes Δ k + 1 ( u x + v y ) \Delta^{k+1}(ux+vy) Δ k + 1 ( ux + v y ) . The part involving D D D contributes
a ξ k + 1 ξ k + b η k + 1 η k + ε ( v , − u ) ( G T ) k + 1 D G k ( x y ) . a\xi_{k+1}\xi_k+b\eta_{k+1}\eta_k+\varepsilon(v,-u)(G^{\mathsf{T}})^{k+1}DG^k\begin{pmatrix}x\\y\end{pmatrix}. a ξ k + 1 ξ k + b η k + 1 η k + ε ( v , − u ) ( G T ) k + 1 D G k ( x y ) .
Thus the output is F 2 , 2 k + 4 + ε H 2 , 2 k + 4 F_{2,2k+4}+\varepsilon H_{2,2k+4} F 2 , 2 k + 4 + ε H 2 , 2 k + 4 , where
H 2 , 2 k + 4 : = ( v , − u ) ( G T ) k + 1 D G k ( x y ) . H_{2,2k+4}:=(v,-u)(G^{\mathsf{T}})^{k+1}DG^k\begin{pmatrix}x\\y\end{pmatrix}. H 2 , 2 k + 4 := ( v , − u ) ( G T ) k + 1 D G k ( x y ) .
For every nonzero algebraic ε \varepsilon ε , all factors in these products have homogeneous linear entries. When matrix powers are expanded into successive factors, the odd-degree product has 1 + k + 1 + k + 1 = 2 k + 3 1+k+1+k+1=2k+3 1 + k + 1 + k + 1 = 2 k + 3 factors, and the even-degree product has 1 + ( k + 1 ) + 1 + k + 1 = 2 k + 4 1+(k+1)+1+k+1=2k+4 1 + ( k + 1 ) + 1 + k + 1 = 2 k + 4 factors. Powers with exponent zero contribute no factors. Both products therefore define ABPs of width at most two and exactly the required degree.
For general w w w , add w − 2 w-2 w − 2 paths, sharing only the source and sink with the width-two ABP and with each other. The i i i -th path has d d d edges labelled c i , z i , … , z i c_i,z_i,\ldots,z_i c i , z i , … , z i . The resulting ABP has width at most w w w and output
F w , d + ε H 2 , d . F_{w,d}+\varepsilon H_{2,d}. F w , d + ε H 2 , d .
This also covers w = 2 w=2 w = 2 , when there are no additional paths.
The coefficients of this output are polynomials in ε \varepsilon ε . For every nonzero algebraic ε \varepsilon ε , the output lies in A w , d \mathcal{A}_{w,d} A w , d . Since Q ∖ { 0 } \mathbb{Q}\setminus\{0\} Q ∖ { 0 } is Zariski dense in the affine line, its value at ε = 0 \varepsilon=0 ε = 0 lies in A w , d \mathcal{A}_{w,d} A w , d . That value is F w , d F_{w,d} F w , d , proving the claim. □
We now establish the algebraic property used for the lower bound. The aim is to control the common zero sets of the first and last layers of a putative ABP. We begin with two linear-algebra calculations and then apply them to the polynomial F w , d F_{w,d} F w , d .
Claim F.1. Let U U U be a linear subspace of Mat 2 ( Q ‾ ) \operatorname{Mat}_2(\overline{\mathbb{Q}}) Mat 2 ( Q ) consisting of singular matrices. Then dim U ≤ 2 \dim U \le2 dim U ≤ 2 . If dim U = 2 \dim U = 2 dim U = 2 , then U U U contains a rank-one idempotent.
Proof. If U U U is nonzero, choose a nonzero matrix in U U U . It has rank one. After invertible changes of row and column bases, we may assume that it is
( 1 0 0 0 ) \begin{pmatrix} 1 & 0 \\ 0 & 0 \end{pmatrix} ( 1 0 0 0 )
For any B = ( b i j ) ∈ U B=(b_{ij})\in U B = ( b ij ) ∈ U , the determinant of
( 1 0 0 0 ) + t B \begin{pmatrix} 1 & 0 \\ 0 & 0 \end{pmatrix}+tB ( 1 0 0 0 ) + tB
vanishes for every t t t . The coefficient of t t t shows that b 22 = 0 b_{22}=0 b 22 = 0 . Since B B B itself is singular, we also have b 12 b 21 = 0 b_{12}b_{21}=0 b 12 b 21 = 0 .
The two entries b 12 b_{12} b 12 and b 21 b_{21} b 21 are linear functions on U U U . Their product is identically zero, so one of them is identically zero. Thus every matrix in U U U has either its second row zero or its second column zero, with the same alternative holding throughout U U U . In particular, dim U ≤ 2 \dim U\le2 dim U ≤ 2 .
Returning to the original bases, this says that U U U is contained in a space of one of the forms
{ u φ : φ ∈ Q ‾ 1 × 2 } , { u φ : u ∈ Q ‾ 2 × 1 } , \{u\varphi:\varphi\in\overline{\mathbb{Q}}^{1\times2}\},\qquad
\{u\varphi:u\in\overline{\mathbb{Q}}^{2\times1}\}, { u φ : φ ∈ Q 1 × 2 } , { u φ : u ∈ Q 2 × 1 } ,
where the column u u u is fixed and nonzero in the first case, and the row φ \varphi φ is fixed and nonzero in the second. If dim U = 2 \dim U=2 dim U = 2 , it equals the corresponding displayed space. We may then choose u , φ u,\varphi u , φ with φ u = 1 \varphi u=1 φ u = 1 . The matrix u φ u\varphi u φ has rank one and satisfies ( u φ ) 2 = u φ (u\varphi)^2=u\varphi ( u φ ) 2 = u φ , as required.
The polynomial F 2 , d F_{2,d} F 2 , d is linear in the four variables a , b , u , v a,b,u,v a , b , u , v . Write
F 2 , d = a Q a + b Q b + u Q u + v Q v , (27) F_{2,d}=aQ_a+bQ_b+uQ_u+vQ_v, \tag*{(27)} F 2 , d = a Q a + b Q b + u Q u + v Q v , (27)
where the four coefficient polynomials depend only on x , y , z , p , q , r , s x,y,z,p,q,r,s x , y , z , p , q , r , s . For d = 2 k + 3 d=2k+3 d = 2 k + 3 , they are
Q a = ξ k 2 , Q b = η k 2 , Q u = z Δ k x , Q v = z Δ k y . Q_a=\xi_k^2,\qquad Q_b=\eta_k^2,\qquad Q_u=z\Delta^k x,\qquad Q_v=z\Delta^k y. Q a = ξ k 2 , Q b = η k 2 , Q u = z Δ k x , Q v = z Δ k y .
For d = 2 k + 4 d=2k+4 d = 2 k + 4 , they are
Q a = ξ k + 1 ξ k , Q b = η k + 1 η k , Q u = Δ k + 1 x , Q v = Δ k + 1 y . Q_a=\xi_{k+1}\xi_k,\qquad Q_b=\eta_{k+1}\eta_k,\qquad Q_u=\Delta^{k+1}x,\qquad Q_v=\Delta^{k+1}y. Q a = ξ k + 1 ξ k , Q b = η k + 1 η k , Q u = Δ k + 1 x , Q v = Δ k + 1 y .
All four are homogeneous of degree d − 1 d-1 d − 1 .
Claim F.2. Let L ⊆ Q ‾ 7 L\subseteq\overline{\mathbb{Q}}^7 L ⊆ Q 7 be a linear subspace in the coordinates x , y , z , p , q , r , s x,y,z,p,q,r,s x , y , z , p , q , r , s , with dim L ≥ 5 \dim L\ge5 dim L ≥ 5 . If
dim span Q { Q a ∣ L , Q b ∣ L , Q u ∣ L , Q v ∣ L } ≤ dim L − 5 , \dim\operatorname{span}_{\mathbb{Q}}\{Q_a|_L,Q_b|_L,Q_u|_L,Q_v|_L\}\le\dim L-5, dim span Q { Q a ∣ L , Q b ∣ L , Q u ∣ L , Q v ∣ L } ≤ dim L − 5 , then L = { x = y = 0 } L=\{x=y=0\} L = { x = y = 0 } .
Proof. The restrictions in the claim are polynomial functions on L L L . We consider its three possible dimensions.
If dim L = 7 \dim L=7 dim L = 7 , then L L L is the whole space. The polynomials Q a Q_a Q a and Q b Q_b Q b are linearly independent: setting G = I 2 G=I_2 G = I 2 gives x 2 x^2 x 2 and y 2 y^2 y 2 , respectively. Both have degree two in the pair of variables x , y x,y x , y , whereas the nonzero polynomial Q u Q_u Q u has degree one in that pair. Hence these three polynomials are linearly independent. This contradicts the assumed bound dim L − 5 = 2 \dim L-5=2 dim L − 5 = 2 .
Suppose next that dim L = 6 \dim L=6 dim L = 6 , so that L L L is a hyperplane. Write Q u = γ x Q_u=\gamma x Q u = γ x and Q v = γ y Q_v=\gamma y Q v = γ y , where
γ = { z Δ k , d = 2 k + 3 , Δ k + 1 , d = 2 k + 4. \gamma=
\begin{cases}
z\Delta^k, & d=2k+3,\\
\Delta^{k+1}, & d=2k+4.
\end{cases} γ = { z Δ k , Δ k + 1 , d = 2 k + 3 , d = 2 k + 4.
If γ ∣ L ≠ 0 \gamma|_L \ne0 γ ∣ L = 0 and x ∣ L , y ∣ L x|_L,y|_L x ∣ L , y ∣ L are linearly independent, then Q u ∣ L , Q v ∣ L Q_u|_L,Q_v|_L Q u ∣ L , Q v ∣ L are linearly independent. Indeed, the coordinate ring of a linear subspace is an integral domain, so multiplication by the nonzero polynomial γ ∣ L \gamma|_L γ ∣ L preserves linear independence.
If γ ∣ L ≠ 0 \gamma|_L \ne0 γ ∣ L = 0 but x ∣ L , y ∣ L x|_L,y|_L x ∣ L , y ∣ L are dependent, then L L L is defined by a nonzero linear equation in x , y x,y x , y alone. One independent coordinate remains from this pair, and all the other coordinates are free. At least one of Q a ∣ L , Q b ∣ L Q_a|_L,Q_b|_L Q a ∣ L , Q b ∣ L is nonzero, as is seen by setting G = I 2 G=I_2 G = I 2 . It has degree two in the remaining coordinate from x , y x,y x , y , whereas a nonzero member of Q u ∣ L , Q v ∣ L Q_u|_L,Q_v|_L Q u ∣ L , Q v ∣ L has degree one in it. These two restrictions are therefore linearly independent.
Finally, suppose that γ ∣ L = 0 \gamma|_L=0 γ ∣ L = 0 . The polynomial Δ = p s − q r \Delta=ps-qr Δ = p s − q r has no linear factor: viewed as a polynomial in p p p , its coefficients s s s and − q r -qr − q r have no common nonconstant divisor. Thus Δ \Delta Δ cannot vanish identically on a hyperplane. The only remaining possibility is odd degree and L = { z = 0 } L=\{z=0\} L = { z = 0 } . On this hyperplane Q a , Q b Q_a,Q_b Q a , Q b are still linearly independent, again by setting G = I 2 G=I_2 G = I 2 . In every case the span has dimension at least two, contradicting the assumed bound dim L − 5 = 1 \dim L-5=1 dim L − 5 = 1 .
It remains to consider dim L = 5 \dim L=5 dim L = 5 . The dimension bound now says that all four coefficient polynomials vanish identically on L L L . Suppose first that Δ ∣ L ≠ 0 \Delta|_L\ne0 Δ ∣ L = 0 . At every point of L L L where G G G is invertible, the odd-degree equalities ξ k 2 = η k 2 = 0 \xi_k^2=\eta_k^2=0 ξ k 2 = η k 2 = 0 force x = y = 0 x=y=0 x = y = 0 . In even degree, the equalities Δ k + 1 x = Δ k + 1 y = 0 \Delta^{k+1}x=\Delta^{k+1}y=0 Δ k + 1 x = Δ k + 1 y = 0 have the same consequence. Such points form a dense subset of L L L , so the linear functions x x x and y y y vanish on all of L L L . Both L L L and { x = y = 0 } \{x=y=0\} { x = y = 0 } have dimension five, and therefore they are equal.
Suppose instead that Δ ∣ L = 0 \Delta|_L=0 Δ ∣ L = 0 . Project L L L onto the four coordinates of G G G , and denote the image by U U U . The kernel of this projection has dimension at most three, since only x , y , z x,y,z x , y , z are forgotten. Thus dim U ≥ 2 \dim U\ge2 dim U ≥ 2 . Every matrix in U U U is singular, so Claim F.1 gives dim U = 2 \dim U=2 dim U = 2 and a rank-one idempotent E ∈ U E\in U E ∈ U . The projection kernel has dimension three and hence contains all choices of x , y , z x,y,z x , y , z with G = 0 G=0 G = 0 . It follows that, for each fixed G ∈ U G\in U G ∈ U , all values of x , y , z x,y,z x , y , z occur in L L L .
Choose a nonzero column ( x y ) \begin{pmatrix}x\\y\end{pmatrix} ( x y ) fixed by E E E , and take G = E G=E G = E . Then G j ( x y ) = ( x y ) G^j\begin{pmatrix}x\\y\end{pmatrix}=\begin{pmatrix}x\\y\end{pmatrix} G j ( x y ) = ( x y ) for every j ≥ 0 j\ge0 j ≥ 0 . In both parity cases this gives Q a = x 2 Q_a=x^2 Q a = x 2 and Q b = y 2 Q_b=y^2 Q b = y 2 , which are not both zero at this point. This contradicts their vanishing on L L L . The claim follows.
We now return to all 2 w + 7 2w+7 2 w + 7 variables of F w , d F_{w,d} F w , d . Let W 0 W_0 W 0 be the linear subspace defined by
x = y = 0 , c i = z i = 0 ( 1 ≤ i ≤ w − 2 ) . x=y=0,\qquad c_i=z_i=0\quad(1\le i\le w-2). x = y = 0 , c i = z i = 0 ( 1 ≤ i ≤ w − 2 ) .
It has dimension nine; its free coordinates are a , b , u , v , z , p , q , r , s a,b,u,v,z,p,q,r,s a , b , u , v , z , p , q , r , s .
Claim F.3. Every linear subspace of codimension at most w w w on which F w , d F_{w,d} F w , d vanishes contains W 0 W_0 W 0 .
Proof. It is enough to consider codimension exactly w w w . Indeed, a subspace of smaller codimension contains one of codimension w w w , and vanishing is preserved when passing to a smaller subspace. Let W W W have codimension w w w , so that dim W = w + 7 \dim W=w+7 dim W = w + 7 .
Project W W W by forgetting the coefficient variables a , b , u , v , c 1 , … , c w − 2 a,b,u,v,c_1,\ldots,c_{w-2} a , b , u , v , c 1 , … , c w − 2 , and write L L L for its image. These are w + 2 w+2 w + 2 forgotten coordinates. The projection kernel has dimension w + 7 − dim L w+7-\dim L w + 7 − dim L . Each vector in this kernel gives a linear relation, with constant coefficients, among the following polynomial functions on L L L :
Q a ∣ L , Q b ∣ L , Q u ∣ L , Q v ∣ L , z 1 d − 1 ∣ L , … , z w − 2 d − 1 ∣ L . Q_a|_L,\ Q_b|_L,\ Q_u|_L,\ Q_v|_L,\ z_1^{d-1}|_L,\ldots,z_{w-2}^{d-1}|_L. Q a ∣ L , Q b ∣ L , Q u ∣ L , Q v ∣ L , z 1 d − 1 ∣ L , … , z w − 2 d − 1 ∣ L .
To see this, take any point of W W W and add a kernel vector. Only the coefficient variables change. Subtracting the two zero values of F w , d F_{w,d} F w , d gives the indicated relation. Every point of L L L has a lift to W W W , so the relation holds throughout L L L . Independent kernel vectors give independent relation vectors. Consequently,
dim span { Q a ∣ L , Q b ∣ L , Q u ∣ L , Q v ∣ L , z 1 d − 1 ∣ L , … , z w − 2 d − 1 ∣ L } ≤ dim L − 5. (28) \dim\operatorname{span}\{Q_a|_L,Q_b|_L,Q_u|_L,Q_v|_L,z_1^{d-1}|_L,\ldots,z_{w-2}^{d-1}|_L\}\leq\dim L-5. \tag*{(28)} dim span { Q a ∣ L , Q b ∣ L , Q u ∣ L , Q v ∣ L , z 1 d − 1 ∣ L , … , z w − 2 d − 1 ∣ L } ≤ dim L − 5. (28)
All spans in this proof are over Q \mathbb{Q} Q .
Next project L L L onto the seven coordinates x , y , z , p , q , r , s x,y,z,p,q,r,s x , y , z , p , q , r , s . Write L 0 L_0 L 0 for its image and T T T for the projection kernel. The functions Q a , Q b , Q u , Q v Q_a,Q_b,Q_u,Q_v Q a , Q b , Q u , Q v vanish on T T T , since all their variables are zero there. The coordinate functions z i z_i z i span the dual of T T T . Choose dim T \dim T dim T of them which are linearly independent. Their ( d − 1 ) (d-1) ( d − 1 ) -st powers are also linearly independent: in these coordinates they are distinct monomials. Thus restriction to T T T maps the span in Equation (28) onto a space of dimension at least dim T \dim T dim T . The kernel of this restriction contains the span of Q a ∣ L , Q b ∣ L , Q u ∣ L , Q v ∣ L Q_a|_L,Q_b|_L,Q_u|_L,Q_v|_L Q a ∣ L , Q b ∣ L , Q u ∣ L , Q v ∣ L . Since these functions depend only on the projection to L 0 L_0 L 0 , their span has the same dimension as the span of their restrictions to L 0 L_0 L 0 . We obtain
dim T + dim span { Q a ∣ L 0 , Q b ∣ L 0 , Q u ∣ L 0 , Q v ∣ L 0 } ≤ dim L − 5. \dim T+\dim\operatorname{span}\{Q_a|_{L_0},Q_b|_{L_0},Q_u|_{L_0},Q_v|_{L_0}\}\leq\dim L-5. dim T + dim span { Q a ∣ L 0 , Q b ∣ L 0 , Q u ∣ L 0 , Q v ∣ L 0 } ≤ dim L − 5.
Using dim L = dim T + dim L 0 \dim L=\dim T+\dim L_0 dim L = dim T + dim L 0 , this becomes
dim span { Q a ∣ L 0 , Q b ∣ L 0 , Q u ∣ L 0 , Q v ∣ L 0 } ≤ dim L 0 − 5. \dim\operatorname{span}\{Q_a|_{L_0},Q_b|_{L_0},Q_u|_{L_0},Q_v|_{L_0}\}\leq\dim L_0-5. dim span { Q a ∣ L 0 , Q b ∣ L 0 , Q u ∣ L 0 , Q v ∣ L 0 } ≤ dim L 0 − 5.
In particular, dim L 0 ≥ 5 \dim L_0\geq5 dim L 0 ≥ 5 . Claim F.2 therefore gives L 0 = { x = y = 0 } L_0=\{x=y=0\} L 0 = { x = y = 0 } . Hence x = y = 0 x=y=0 x = y = 0 on W W W .
On W W W , the polynomial F w , d F_{w,d} F w , d now reduces to ∑ i = 1 w − 2 c i z i d − 1 \sum_{i=1}^{w-2}c_i z_i^{d-1} ∑ i = 1 w − 2 c i z i d − 1 . Let V V V be the projection of W W W onto the coordinates ( c i , z i ) i = 1 w − 2 (c_i,z_i)_{i=1}^{w-2} ( c i , z i ) i = 1 w − 2 . We show that dim V ≤ w − 2 \dim V\leq w-2 dim V ≤ w − 2 . Project V V V onto its z i z_i z i -coordinates, and let j j j be the dimension of this image. The kernel has dimension dim V − j \dim V-j dim V − j and supplies that many independent linear relations among the restrictions of z 1 d − 1 , … , z w − 2 d − 1 z_1^{d-1},\ldots,z_{w-2}^{d-1} z 1 d − 1 , … , z w − 2 d − 1 to the image. Their span therefore has dimension at most ( w − 2 ) − ( dim V − j ) (w-2)-(\dim V-j) ( w − 2 ) − ( dim V − j ) . On the other hand, j j j independent coordinate functions have independent ( d − 1 ) (d-1) ( d − 1 ) -st powers, so the same span has dimension at least j j j . Thus
j ≤ ( w − 2 ) − ( dim V − j ) , and hence dim V ≤ w − 2. j\leq(w-2)-(\dim V-j),\qquad\text{and hence}\qquad\dim V\leq w-2. j ≤ ( w − 2 ) − ( dim V − j ) , and hence dim V ≤ w − 2.
For w = 2 w=2 w = 2 , these coordinate spaces are zero-dimensional and the same conclusion holds.
It follows that the kernel of the projection W → V W\to V W → V has dimension at least ( w + 7 ) − ( w − 2 ) = 9 (w+7)-(w-2)=9 ( w + 7 ) − ( w − 2 ) = 9 . We already know that x = y = 0 x=y=0 x = y = 0 on W W W . In this kernel all c i , z i c_i,z_i c i , z i also vanish, so the kernel is contained in the nine-dimensional space W 0 W_0 W 0 . It must therefore equal W 0 W_0 W 0 . In particular, W 0 ⊆ W W_0\subseteq W W 0 ⊆ W , as required. □
We now prove that F w , d F_{w,d} F w , d cannot be computed by ABPs of width w w w and degree d d d .
Claim 4.5. ι ( F w , d ) ≁ 0 \iota(F_{w,d})\not\sim0 ι ( F w , d ) ∼ 0 .
Proof. Suppose that ι ( F w , d ) ∼ 0 \iota(F_{w,d})\sim0 ι ( F w , d ) ∼ 0 . By Claim 4.2 and the injectivity of ι \iota ι , the polynomial F w , d F_{w,d} F w , d is computed by an ABP of width at most w w w and degree d d d . If m > 2 w + 7 m>2w+7 m > 2 w + 7 , set all the unused variables to zero. The resulting ABP still computes F w , d F_{w,d} F w , d , has exactly d d d homogeneous linear matrix factors, and has width at most w w w . We therefore work in the 2 w + 7 2w+7 2 w + 7 variables used in its definition.
Let the first and last layers consist of the linear forms ℓ 1 , … , ℓ w \ell_1,\ldots,\ell_w ℓ 1 , … , ℓ w and r 1 , … , r w r_1,\ldots,r_w r 1 , … , r w , allowing zero forms if necessary. The polynomial F w , d F_{w,d} F w , d vanishes on the common zero set of the ℓ i \ell_i ℓ i , and also on the common zero set of the r i r_i r i . Each of these zero sets is a linear subspace of codimension at most w w w . Claim F.3 implies that both contain W 0 W_0 W 0 . Hence every ℓ i \ell_i ℓ i and every r i r_i r i vanishes on W 0 W_0 W 0 . Since these forms are linear, each belongs to the linear span of
x , y , c 1 , z 1 , … , c w − 2 , z w − 2 . x,y,c_1,z_1,\ldots,c_{w-2},z_{w-2}. x , y , c 1 , z 1 , … , c w − 2 , z w − 2 .
Every path in the ABP contains one first-layer form and one last-layer form. After expansion, each of its monomials therefore has total degree at least two in the displayed variables. Taking the sum over all paths cannot create a monomial of smaller degree in those variables.
However, in F w , d F_{w,d} F w , d the sum of the monomials having degree exactly one in those variables is
{ z Δ k ( u x + v y ) , d = 2 k + 3 , Δ k + 1 ( u x + v y ) , d = 2 k + 4. \begin{cases}
z\Delta^k(ux+vy), & d=2k+3,\\
\Delta^{k+1}(ux+vy), & d=2k+4.
\end{cases} { z Δ k ( ux + v y ) , Δ k + 1 ( ux + v y ) , d = 2 k + 3 , d = 2 k + 4.
It is nonzero in both cases. The summands involving a , b a,b a , b have degree two in x , y x,y x , y , and each c i z i d − 1 c_i z_i^{d-1} c i z i d − 1 has degree d d d in the displayed variables. Thus none of them can cancel this degree-one part. This contradicts the ABP representation and proves ι ( F w , d ) ≁ 0 \iota(F_{w,d})\not\sim0 ι ( F w , d ) ∼ 0 .
By Claim 4.2, the last claim gives F w , d ∉ A w , d F_{w,d}\notin\mathcal{A}_{w,d} F w , d ∈ / A w , d . Claim 4.4 gives F w , d ∈ A ‾ w , d F_{w,d}\in\overline{\mathcal{A}}_{w,d} F w , d ∈ A w , d . This proves Theorem 4.
For completeness, the description of ∼ \sim ∼ also determines the number of classes in the matrix closure. For every nonzero λ ∈ Q ‾ \lambda\in\overline{\mathbb{Q}} λ ∈ Q , the polynomial λ F w , d \lambda F_{w,d} λ F w , d is still in A ‾ w , d ∖ A w , d \overline{\mathcal{A}}_{w,d}\setminus\mathcal{A}_{w,d} A w , d ∖ A w , d . Indeed, scalar multiplication preserves the ABP class by scaling its first layer, and hence preserves its closure; an ABP for λ F w , d \lambda F_{w,d} λ F w , d could be scaled back to one for F w , d F_{w,d} F w , d . The distinct matrices ι ( λ F w , d ) \iota(\lambda F_{w,d}) ι ( λ F w , d ) therefore give infinitely many singleton classes. Consequently,
∣ ⟨ X ⟩ / ∼ ∣ = 4 , ∣ ⟨ X ⟩ ‾ / ∼ ∣ = ∞ . |\langle X\rangle/\sim|=4,\qquad|\overline{\langle X\rangle}/\sim|=\infty. ∣ ⟨ X ⟩ / ∼ ∣ = 4 , ∣ ⟨ X ⟩ / ∼ ∣ = ∞.
Detailed Proofs of the Examples
We collect here the calculations and proofs for the examples in the preliminaries and overview. The final example gives an infinite Zariski closure to which Theorem 3 applies.
Example 1
For the matrices in that example, direct multiplication gives A x A y = A y A_xA_y=A_y A x A y = A y for all x , y x,y x , y . All these matrices have rank one, so every pair preserves rank, including when x = y x=y x = y . Under the row-vector convention,
ker ( A x ) = { ( 0 , t ) : t ∈ Q } , im ( A x ) = Span Q { ( 1 , x ) } . \ker(A_x)=\{(0,t):t\in\mathbb{Q}\},\qquad\operatorname{im}(A_x)=\operatorname{Span}_{\mathbb{Q}}\{(1,x)\}. ker ( A x ) = {( 0 , t ) : t ∈ Q } , im ( A x ) = Span Q {( 1 , x )} .
The images are distinct when x ≠ y x\ne y x = y , so { A x , A y } \{A_x,A_y\} { A x , A y } is rank-stable but not stable. Finally, A A A has rank one and A 2 = 0 A^2=0 A 2 = 0 . It is therefore not rank-stable, and the singleton { A } \{A\} { A } is not a stable set.
Example 2
Recall that N , m ≥ 2 N,m\ge2 N , m ≥ 2 , B = m 2 + 1 B=m^2+1 B = m 2 + 1 , and
A i = ( B 0 B i 0 1 0 0 i 0 ) ( 1 ≤ i ≤ m ) , W = ( A 1 ⋯ A m ) N . A_i=
\begin{pmatrix}
B & 0 & Bi\\
0 & 1 & 0\\
0 & i & 0
\end{pmatrix}
\quad(1\le i\le m),\qquad W=(A_1\cdots A_m)^N. A i = B 0 0 0 1 i B i 0 0 ( 1 ≤ i ≤ m ) , W = ( A 1 ⋯ A m ) N .
We first verify uniqueness of representation and identify the kernels and images of products.
Claim G.1. For i 1 , … , i k ∈ { 1 , … , m } i_1,\ldots,i_k \in\{1,\ldots,m\} i 1 , … , i k ∈ { 1 , … , m } ,
A i 1 ⋯ A i k = ( B k ∑ t = 1 k − 1 B t i t i t + 1 B k i k 0 1 0 0 i 1 0 ) . (29) A_{i_1}\cdots A_{i_k}=
\begin{pmatrix}
B^k & \sum_{t=1}^{k-1}B^t i_ti_{t+1} & B^k i_k\\
0 & 1 & 0\\
0 & i_1 & 0
\end{pmatrix}. \tag*{(29)} A i 1 ⋯ A i k = B k 0 0 ∑ t = 1 k − 1 B t i t i t + 1 1 i 1 B k i k 0 0 . (29) The displayed matrix determines the entire sequence i 1 , … , i k i_1,\ldots,i_k i 1 , … , i k . Every such matrix has rank two, kernel Span Q { ( 0 , − i 1 , 1 ) } \operatorname{Span}_{\mathbb{Q}}\{(0,-i_1,1)\} Span Q {( 0 , − i 1 , 1 )} , and image { ( x , y , i k x ) : x , y ∈ Q } \{(x,y,i_kx):x,y\in\mathbb{Q}\} {( x , y , i k x ) : x , y ∈ Q } .
Proof. The formula holds for k = 1 k=1 k = 1 , with empty sum zero. Multiplication on the right by A j A_j A j multiplies the first entry by B B B , adds B k i k j B^k i_kj B k i k j to the second entry of the first row, and gives last entry B k + 1 j B^{k+1}j B k + 1 j . The other two rows remain as displayed. This proves the formula by induction.
The first diagonal entry determines k k k , and the entry in position ( 3 , 2 ) (3,2) ( 3 , 2 ) determines i 1 i_1 i 1 . For k ≥ 2 k\ge2 k ≥ 2 , the base-B B B digits of the entry in position ( 1 , 2 ) (1,2) ( 1 , 2 ) determine all numbers i t i t + 1 i_ti_{t+1} i t i t + 1 , because 1 ≤ i t i t + 1 ≤ m 2 < B 1\le i_ti_{t+1}\le m^2<B 1 ≤ i t i t + 1 ≤ m 2 < B . Starting with i 1 i_1 i 1 , successive division by the already recovered i t i_t i t determines i t + 1 i_{t+1} i t + 1 . This also covers uniqueness for k = 1 k=1 k = 1 using i 1 i_1 i 1 alone. Finally, the second and third rows are dependent and the first row is independent of them, proving rank two. Solving for a row vector annihilated by the matrix gives the stated kernel. The first row divided by B k B^k B k , together with ( 0 , 1 , 0 ) (0,1,0) ( 0 , 1 , 0 ) , spans exactly the stated image. □
Let I I I be the set of matrices in ⟨ X m ⟩ \langle X_m\rangle ⟨ X m ⟩ that do not occur as factors of the displayed product W W W . Uniqueness of representation implies that I I I is a two-sided ideal. Indeed, if a product containing the unique representation of some z ∈ I z\in I z ∈ I occurred as a factor of W W W , the representation of z z z would itself occur as a factor of W W W , a contradiction. Thus multiplying a matrix of I I I on either side by any matrix of ⟨ X m ⟩ \langle X_m\rangle ⟨ X m ⟩ keeps it in I I I . Collapsing I I I to one class and leaving every matrix outside I I I in its own singleton class is consequently a semigroup congruence.
Claim G.2. Every stable subsemigroup meets at most N + 1 N+1 N + 1 classes of this congruence, and one meets exactly N + 1 N+1 N + 1 classes.
Proof. By Claim G.1, a common kernel fixes the first generator index i i i , and a common image fixes the last generator index j j j . For fixed i , j i,j i , j , a factor of W W W is determined by its length: the indices follow the periodic sequence 1 , … , m , 1 , … , m , … 1,\ldots,m,1,\ldots,m,\ldots 1 , … , m , 1 , … , m , … . Its length lies in one residue class modulo m m m , as its first and last indices are fixed. Among 1 , … , N m 1,\ldots,Nm 1 , … , N m , there are exactly N N N lengths in each residue class, so there are at most N N N distinct such factors of W W W . Any stable subsemigroup therefore meets at most N N N singleton classes and possibly the large class I I I .
For equality, put C 0 = A 1 ⋯ A m C_0=A_1\cdots A_m C 0 = A 1 ⋯ A m and consider ⟨ C 0 ⟩ \langle C_0\rangle ⟨ C 0 ⟩ . All of its matrices have the same kernel and image, by (29), and all products have rank two. Thus it is stable. The matrices C 0 , … , C 0 N C_0,\ldots,C_0^N C 0 , … , C 0 N are distinct factors of W W W , and C 0 N + 1 ∈ I C_0^{N+1}\in I C 0 N + 1 ∈ I . It follows that this stable subsemigroup meets exactly N + 1 N+1 N + 1 classes. □
The class of W W W is a singleton and W W W has length N m Nm N m , by uniqueness of representation. Every other singleton class is represented by a factor of W W W of length at most N m Nm N m . The large class contains A 1 2 A_1^2 A 1 2 : since m ≥ 2 m\ge2 m ≥ 2 , the periodic product W W W never has two consecutive occurrences of A 1 A_1 A 1 . That class has length at most 2 ≤ N m 2\le Nm 2 ≤ N m . Hence the diameter of the quotient is exactly N m = N ∣ X m ∣ Nm=N|X_m| N m = N ∣ X m ∣ , while its smallest uniform stable bound is N + 1 N+1 N + 1 .
Example 3
The generators act on row vectors by
( x 1 , … , x d ) A i = ( x 1 , x 3 , … , x d , i x 1 ) . (x_1,\ldots,x_d)A_i=(x_1,x_3,\ldots,x_d,ix_1). ( x 1 , … , x d ) A i = ( x 1 , x 3 , … , x d , i x 1 ) .
Iterating this action, for 1 ≤ k ≤ d − 1 1 \le k \le d-1 1 ≤ k ≤ d − 1 , gives
( x 1 , … , x d ) A i 1 ⋯ A i k = ( x 1 , x k + 2 , … , x d , i 1 x 1 , … , i k x 1 ) , (x_1,\ldots,x_d)A_{i_1}\cdots A_{i_k}=(x_1,x_{k+2},\ldots,x_d,i_1x_1,\ldots,i_kx_1), ( x 1 , … , x d ) A i 1 ⋯ A i k = ( x 1 , x k + 2 , … , x d , i 1 x 1 , … , i k x 1 ) ,
where the middle list is empty when k = d − 1 k=d-1 k = d − 1 . The rank is therefore d − k d-k d − k in this range. Every generator preserves the first coordinate, whereas every product of d − 1 d-1 d − 1 generators depends only on that coordinate. Consequently, for k ≥ d − 1 k \ge d-1 k ≥ d − 1 ,
a 1 ⋯ a k = a k − d + 2 ⋯ a k , a_1\cdots a_k=a_{k-d+2}\cdots a_k, a 1 ⋯ a k = a k − d + 2 ⋯ a k ,
and its rank is one. This proves the rank formula max { 1 , d − k } \max\{1,d-k\} max { 1 , d − k } .
Let S S S be the set of rank-one matrices in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . They are precisely the m d − 1 m^{d-1} m d − 1 distinct matrices
P ( i 1 , … , i d − 1 ) = ( 1 i 1 ⋯ i d − 1 0 0 ⋯ 0 ⋮ ⋮ ⋮ 0 0 ⋯ 0 ) , 1 ≤ i j ≤ m . P(i_1,\ldots,i_{d-1})=
\begin{pmatrix}
1 & i_1 & \cdots& i_{d-1}\\
0 & 0 & \cdots& 0\\
\vdots& \vdots& & \vdots\\
0 & 0 & \cdots& 0
\end{pmatrix},
\qquad1 \le i_j \le m. P ( i 1 , … , i d − 1 ) = 1 0 ⋮ 0 i 1 0 ⋮ 0 ⋯ ⋯ ⋯ i d − 1 0 ⋮ 0 , 1 ≤ i j ≤ m .
For P , Q ∈ S P,Q \in S P , Q ∈ S , multiplication gives P Q = Q PQ=Q P Q = Q . Hence S S S is a rank-stable subsemigroup, and every matrix in it is idempotent. If a product a a a of k k k generators has rank greater than one, then k < d − 1 k<d-1 k < d − 1 , and
rank ( a 2 ) = max { 1 , d − 2 k } < d − k = rank ( a ) . \operatorname{rank}(a^2)=\max\{1,d-2k\}<d-k=\operatorname{rank}(a). rank ( a 2 ) = max { 1 , d − 2 k } < d − k = rank ( a ) .
It cannot belong to any rank-stable subsemigroup. Thus all rank-stable subsemigroups are contained in S S S . For the equality congruence, the smallest uniform bound on their quotient sizes is exactly ∣ S ∣ = m d − 1 |S|=m^{d-1} ∣ S ∣ = m d − 1 .
The images of the matrices in S S S are the lines spanned by ( 1 , i 1 , … , i d − 1 ) (1,i_1,\ldots,i_{d-1}) ( 1 , i 1 , … , i d − 1 ) . These lines are pairwise distinct: proportional vectors whose first coordinate is one must be equal. Every stable subsemigroup is contained in S S S , so the common-image condition forces it to be a singleton. Conversely, every singleton in S S S is stable, since its matrix is idempotent. The smallest uniform stable bound is therefore exactly one.
Example 4
Index the rows and columns of a 2 × 2 2 \times2 2 × 2 matrix by 0 , 1 0,1 0 , 1 . Write E u v E_{uv} E uv for the matrix with a single 1 1 1 in position ( u , v ) (u,v) ( u , v ) and zeros elsewhere. Thus E u v E s t = E u t E_{uv}E_{st}=E_{ut} E uv E s t = E u t when v = s v=s v = s , and is zero otherwise. Recall that X k = { A 1 , … , A k } X_k=\{A_1,\ldots,A_k\} X k = { A 1 , … , A k } , where the j j j th diagonal block of A i A_i A i is E 10 E_{10} E 10 for j < i j<i j < i , E 01 E_{01} E 01 for j = i j=i j = i , and I 2 I_2 I 2 for j > i j>i j > i . The target is T k = diag ( E 01 , … , E 01 ) T_k=\operatorname{diag}(E_{01},\ldots,E_{01}) T k = diag ( E 01 , … , E 01 ) .
Every diagonal block of every product belongs to
U = { 0 , I 2 , E 00 , E 01 , E 10 , E 11 } , \mathcal{U}=\{0,I_2,E_{00},E_{01},E_{10},E_{11}\}, U = { 0 , I 2 , E 00 , E 01 , E 10 , E 11 } ,
which is closed under multiplication. All off-diagonal blocks are zero. Thus ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ is finite, with at most 6 k 6^k 6 k matrices, and all entries of all products belong to { 0 , 1 } \{0,1\} { 0 , 1 } .
Claim G.3. The matrix T k T_k T k has length exactly 2 k − 1 2^k-1 2 k − 1 with respect to X k X_k X k .
Proof. Consider any product representing T k T_k T k . Starting with e 0 = ( 1 , 0 ) e_0=(1,0) e 0 = ( 1 , 0 ) in each diagonal block, track its image under the successive generators. At the end, each tracked row equals e 1 = ( 0 , 1 ) e_1=(0,1) e 1 = ( 0 , 1 ) . None can become zero at an intermediate step, since subsequent multiplication cannot restore a zero row. At every step the tracked rows therefore determine a bit pattern ( b 1 , … , b k ) (b_1,\ldots,b_k) ( b 1 , … , b k ) , where the row in block j j j is e b j e_{b_j} e b j . Initially all bits are zero.
Multiplication by A i A_i A i keeps all tracked rows nonzero exactly when b 1 = ⋯ = b i − 1 = 1 b_1=\cdots=b_{i-1}=1 b 1 = ⋯ = b i − 1 = 1 and b i = 0 b_i=0 b i = 0 . It then changes the first i − 1 i-1 i − 1 bits to zero, bit i i i to one, and leaves the higher bits unchanged. Its index must therefore be the least index whose bit is zero. The integer ∑ j = 1 k 2 j − 1 b j \sum_{j=1}^{k}2^{j-1}b_j ∑ j = 1 k 2 j − 1 b j increases by exactly one at each step. Reaching the all-one pattern from the all-zero pattern requires 2 k − 1 2^k-1 2 k − 1 steps. This proves the lower bound.
To exhibit a product of that length, define matrices recursively by C 1 = A 1 C_1=A_1 C 1 = A 1 and C j = C j − 1 A j C j − 1 C_j=C_{j-1}A_jC_{j-1} C j = C j − 1 A j C j − 1 for 2 ≤ j ≤ k 2\le j\le k 2 ≤ j ≤ k . Using E 01 E 10 E 01 = E 01 E_{01}E_{10}E_{01}=E_{01} E 01 E 10 E 01 = E 01 , induction shows that C j C_j C j has its first j j j diagonal blocks equal to E 01 E_{01} E 01 and its remaining blocks equal to I 2 I_2 I 2 . Hence C k = T k C_k=T_k C k = T k . Expanding the recursion into generators gives lengths n 1 = 1 n_1=1 n 1 = 1 and n j = 2 n j − 1 + 1 n_j=2n_{j-1}+1 n j = 2 n j − 1 + 1 , so n k = 2 k − 1 n_k=2^k-1 n k = 2 k − 1 .
Claim G.4. Every rank-stable subsemigroup of ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ is a singleton.
Proof. Among the matrices in U U U , the only ones whose rank drops on squaring are E 01 E_{01} E 01 and E 10 E_{10} E 10 . Ranks add across diagonal blocks and cannot increase on squaring. Consequently every rank-stable matrix in ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ is diagonal, with only zeros and ones on its diagonal. For two such matrices p , q p,q p , q , let their supports be the sets of diagonal positions containing a one. Their ranks are the sizes of their supports, and rank ( p q ) \operatorname{rank}(pq) rank ( pq ) is the size of the intersection. Thus rank ( p ) = rank ( q ) = rank ( p q ) \operatorname{rank}(p)=\operatorname{rank}(q)=\operatorname{rank}(pq) rank ( p ) = rank ( q ) = rank ( pq ) forces the supports, and hence the matrices, to be equal. Every rank-stable subsemigroup is therefore a singleton. Such subsemigroups exist, for example { A 1 2 } \{A_1^2\} { A 1 2 } , since A 1 2 = diag ( 0 , I 2 , … , I 2 ) A_1^2=\operatorname{diag}(0,I_2,\ldots,I_2) A 1 2 = diag ( 0 , I 2 , … , I 2 ) is idempotent.
With equality as congruence, the smallest uniform bound on rank-stable subsemigroups is thus N = 1 N=1 N = 1 . By Claim G.3, the diameter is at least 2 k − 1 = 2 d / 2 − 1 2^k-1=2^{d/2}-1 2 k − 1 = 2 d /2 − 1 in dimension d = 2 k d=2k d = 2 k , as asserted.
Example 5
We first consider the finite-closure case. The semigroup ⟨ X k ⟩ \langle X_k\rangle ⟨ X k ⟩ is finite, so it is Zariski closed in Mat 2 k ( Q ) \operatorname{Mat}_{2^k}(\mathbb{Q}) Mat 2 k ( Q ) : singletons are closed, and a finite union of closed sets is closed. Every matrix group is rank-stable. Indeed, if its identity is e e e , then g = e g = g e g=eg=ge g = e g = g e implies rank ( g ) ≤ rank ( e ) \operatorname{rank}(g)\le\operatorname{rank}(e) rank ( g ) ≤ rank ( e ) , and e = g g − 1 e=gg^{-1} e = g g − 1 gives the reverse inequality; the same applies to every product in the group. Claim G.4 therefore implies that every group in this closure is trivial. The equality congruence satisfies the hypothesis of Theorem 3 with N = 1 N=1 N = 1 , while T k T_k T k still has length 2 k − 1 2^k-1 2 k − 1 . In dimension d = 2 k d=2k d = 2 k , the diameter is therefore at least 2 d / 2 − 1 2^{d/2}-1 2 d /2 − 1 .
For the infinite-closure case, add one coordinate to each generator and put
B i = diag ( 2 , A i ) , Y k = { B 1 , … , B k } ⊆ Mat 2 k + 1 ( Q ) . B_i=\operatorname{diag}(2,A_i),\qquad Y_k=\{B_1,\ldots,B_k\}\subseteq\operatorname{Mat}_{2k+1}(\mathbb{Q}). B i = diag ( 2 , A i ) , Y k = { B 1 , … , B k } ⊆ Mat 2 k + 1 ( Q ) .
Write C k ‾ : = ⟨ Y k ⟩ ‾ \overline{C_k}:=\overline{\langle Y_k\rangle} C k := ⟨ Y k ⟩ for the closure in Mat 2 k + 1 ( Q ) \operatorname{Mat}_{2k+1}(\mathbb{Q}) Mat 2 k + 1 ( Q ) . Every product of ℓ \ell ℓ generators from Y k Y_k Y k has the form
B i 1 ⋯ B i ℓ = diag ( 2 ℓ , A i 1 ⋯ A i ℓ ) . B_{i_1}\cdots B_{i_\ell}=\operatorname{diag}(2^\ell,A_{i_1}\cdots A_{i_\ell}). B i 1 ⋯ B i ℓ = diag ( 2 ℓ , A i 1 ⋯ A i ℓ ) .
In particular, powers of any B i B_i B i are distinct, so the generated semigroup and its closure are infinite. Let S k = ⟨ X k ⟩ S_k=\langle X_k\rangle S k = ⟨ X k ⟩ , which is finite. The set
Z k : = { diag ( t , B ) : t ∈ Q ‾ , B ∈ S k } Z_k:=\{\operatorname{diag}(t,B):t\in\overline{\mathbb{Q}},\ B\in S_k\} Z k := { diag ( t , B ) : t ∈ Q , B ∈ S k }
is a finite union of affine lines, and hence is Zariski closed. It contains ⟨ Y k ⟩ \langle Y_k\rangle ⟨ Y k ⟩ , so C k ⊆ Z k C_k\subseteq Z_k C k ⊆ Z k . Projection onto the lower-right block defines a semigroup homomorphism π : C k → S k \pi:C_k\to S_k π : C k → S k . Define a ∼ b a\sim b a ∼ b if π ( a ) = π ( b ) \pi(a)=\pi(b) π ( a ) = π ( b ) . This is a congruence on the entire closure, identifying matrices exactly when their lower-right blocks are equal.
If G ≤ C k G\leq C_k G ≤ C k is a group, its homomorphic image π ( G ) \pi(G) π ( G ) is a matrix group in S k S_k S k . Every such group is trivial, as proved above. Thus ∣ G / ∼ ∣ = ∣ π ( G ) ∣ = 1 |G/\sim|=|\pi(G)|=1 ∣ G / ∼ ∣ = ∣ π ( G ) ∣ = 1 . The restriction of π \pi π to ⟨ Y k ⟩ \langle Y_k\rangle ⟨ Y k ⟩ is onto S k S_k S k , and induces an isomorphism
⟨ Y k ⟩ / ∼ ≅ S k \langle Y_k\rangle/\sim\ \cong\ S_k ⟨ Y k ⟩ / ∼ ≅ S k
that sends the class of B i B_i B i to A i A_i A i . Representations in these corresponding generators have the same lengths. The class mapping to T k T_k T k therefore has length exactly 2 k − 1 2^k-1 2 k − 1 . Thus Theorem 3 applies with N = 1 N=1 N = 1 and infinite closure, while in dimension d = 2 k + 1 d=2k+1 d = 2 k + 1 the quotient has diameter at least 2 ( d − 1 ) / 2 − 1 2^{(d-1)/2}-1 2 ( d − 1 ) /2 − 1 .
In fact this closure contains an infinite group whose quotient is a singleton. Put E = A 1 2 = diag ( 0 , I 2 , … , I 2 ) E=A_1^2=\operatorname{diag}(0,I_2,\ldots,I_2) E = A 1 2 = diag ( 0 , I 2 , … , I 2 ) . For every n ≥ 2 n\geq2 n ≥ 2 , the matrix B 1 n B_1^n B 1 n equals diag ( 2 n , E ) \operatorname{diag}(2^n,E) diag ( 2 n , E ) . These infinitely many points are Zariski dense in the line { diag ( t , E ) : t ∈ Q ‾ } \{\operatorname{diag}(t,E):t\in\overline{\mathbb{Q}}\} { diag ( t , E ) : t ∈ Q } , because a nonzero polynomial in one variable has only finitely many roots. Hence C k C_k C k contains the group
{ diag ( t , E ) : t ∈ Q ‾ ∖ { 0 } } , \{\operatorname{diag}(t,E):t\in\overline{\mathbb{Q}}\setminus\{0\}\}, { diag ( t , E ) : t ∈ Q ∖ { 0 }} ,
whose identity is diag ( 1 , E ) \operatorname{diag}(1,E) diag ( 1 , E ) and whose quotient under ∼ \sim ∼ is a singleton.
Both cases show that exponential dependence on d d d is necessary. These lower bounds do not establish the optimality of the exponent O ( d 2 ) O(d^2) O ( d 2 ) in the theorem’s upper bound.