For any m pairwise commuting matrices \{A_i\}_{i=1}^{m}\subset \M_n(\F), we prove \dim_\F \F[A_1,\cdots, A_m]\leq (2^m-1)(n-1)+1, answering Bergman's question on the existence of a linear bound for a fixed number m.
Introduction
Let F be a field. For pairwise commuting matrices A1,…,Am∈Mn(F), this paper is concerned with linear upper bounds for the dimension of F[A1,…,Am].
Gerstenhaber problem
Gerstenhaber problem. For a single matrix A1, the Cayley–Hamilton theorem says that the dimension of F[A1] is at most n. For two commuting matrices A1,A2, the dimension of F[A1,A2] is again at most n, which was proved by Motzkin and Taussky [23] and, independently, by Gerstenhaber [11], both arguments resting on the irreducibility of the variety of commuting pairs. Linear-algebraic and commutative-algebraic proofs were found later [2, 20, 32, 3]. For four or more matrices, the corresponding bound fails [11]. Correspondingly, the variety of commuting m-tuples of n×n matrices is reducible when m,n≥4[11, 12].
The case of three matrices lies between these two regimes and remains open.
Gerstenhaber Problem. Is dimFF[A,B,C]≤n for any pairwise commuting A,B,C∈Mn(F)?
We refer to [29, 15] for surveys. For n≤10, the answer is affirmative: in this range, the variety of commuting triples of n×n matrices over an algebraically closed field is irreducible [13, 17, 25, 14, 30, 31], and Gerstenhaber’s argument for pairs then applies. For n≥29, the variety of commuting triples is reducible [17, 24] (see also [19] for its components), so the geometric approach cannot settle this problem for all n. Affirmative answers are known for various special classes of triples, for example, when one of the matrices has square zero [17] or small nullity [13], and for several combinatorially defined classes [26, 27, 9, 8]. Holbrook and O’Meara [16] have proposed a computational strategy for resolving the problem.
Key words and phrases. commuting matrices; algebras of matrices; Gerstenhaber problem; Bergman question.
Bergman question
It is well-known that every commutative subalgebra of Mn(F) has dimension at most ⌊4n2⌋+1, by theorems of Schur [28] and Jacobson [18] (see also [22]), and this bound is sharp; the extremal algebras, however, require on the order of n2 generators. For a fixed number of generators, subquadratic bounds arose in the representation theory of p-adic groups, in connection with Bernstein’s uniform admissibility theorem [4]. It states that m pairwise-commuting complex n×n matrices generate an algebra of dimension at most n2−21−m (see [5 §4.10][6 Proposition 20]).
In the other direction, Bergman [3 §6] considered the supremum rm of the ratios dimFA/dimFV over all commutative subalgebras A⊂EndF(V) generated by m elements. Explicitly, define
All the bounds above grow faster than linearly in n, so they do not address Bergman’s question.
Linear Bound
Our main result answers the Bergman question affirmatively.
Theorem 1.1. Let F be any field and let m,n≥1. For any A1,…,Am∈Mn(F) commute pairwise, then
(i) dimFF[A1,…,Am]≤(2m−1)(n−1)+1.
(ii) limm→∞cm(F)1/m=limm→∞rm1/m=2.
(iii) if char(F)=0 (resp. a prime p), then cm(F)=cm(Q) (resp. cm(Fp)).
Sketch of Proof of (i) and (ii)
(i) For the upper bound, the first ingredient is the maximal-rank argument. Put V=Fn and R=F[A1,…,Am], and choose v so that W=Rv has maximal dimension, we have AnnR(v) sends V into W and annihilates W. Write P=F[x1,…,xm], and let Q, H, J be the annihilators in P of V, W, V/W, respectively. Then
H⊆J,HJ⊆Q,dimF(P/H)=dimFW.
The second ingredient is the colength inequality
(dimF(P/KL))1/m≤(dimF(P/K))1/m+(dimF(P/L))1/m.
If r=dimFW and s=dimF(P/J), these ingredients yield
dimFR≤(r1/m+s1/m)m≤(2m−1)r+s,
because s≤r. Induction applies to the algebra induced on V/W.
(ii) The exponential growth rate follows from constructing a lower bound. Write m=p+q with p,q≥2. From the coefficients of the product of two polynomials of degrees p−1 and q−1, together with one additional bilinear monomial, we obtain a graded algebra A with m degree-one generators. Its ideal of relations is generated by a single homogeneous polynomial of total degree
e=(p−1m−2).
Formal shift operators then give a faithful action of A/A>d on a space of dimension (m−1d+m−1). The ratio of the dimensions of the algebra to this space tends to e as d→∞. Balancing p−1 and q−1 shows that cm(F)≥(⌊2m−2⌋m−2), and the central binomial coefficient has exponential rate 2.
Proof of Theorem 1.1(i)
Colength inequality ingredient
Let P=F[x1,…,xm]. For an ideal I in P, we define
ℓ(I):=dimF(P/I).
We say I has finite colength if ℓ(I) is finite. If I is a monomial ideal, i.e., I is generated by some monomials xα=x1α1⋯xmαm with an m-tuple α=(α1,…,αm)∈Nm, we define
Γ(I)=xα∈I⋃(α+R≥0m).
Lemma 2.1. An ideal I of P has finite colength if and only if there exist ai∈N+ with xiai∈I for each i=1,…,m.
Proof. If xiai∈/I for all ai≥1. Then the images of 1,xi,xi2,… in P/I are linearly independent, so ℓ(I)=∞. Conversely, assume xiai∈I for each i, and let xα∈/I. Then αi≤ai−1 for each i since xiai∣xα implies xα∈I. So the monomials outside I are contained in {x1α1⋯xmαm:0≤αi≤ai−1}, a set of at most ∏iai elements. Therefore, ℓ(I)≤∏iai<∞. □
Lemma 2.2. Let K,L be two monomial ideals of P that have finite colength. Then
(i) Γ(K) is closed.
(ii) ℓ(K)=Vol(R≥0m∖Γ(K)).
(iii) KL has finite colength.
(iv) Γ(KL)=Γ(K)+Γ(L).
Proof. (i) By Lemma 2.1, there exist some ai∈N+ such that xiai∈K. Let ej=(δ1j,…,δmj) and δij is the Kronecker delta which is 1 if i=j, 0 otherwise. Then
Γ(K)=(j=1⋃m(ajej+R≥0m))∪xα∈Kαi≤ai−1 for all i⋃(α+R≥0m)
since if αj≥aj, then α≥ajej coordinate-wise, and hence α+R≥0m⊂ajej+R≥0m, otherwise α∈∏i{0,1,…,ai−1}. Then Γ(K) is a finite union of closed subsets. Therefore, Γ(K) is closed.
(ii) For α∈R≥0m, we claim α∈Γ(K) if and only if x⌊α⌋∈K where ⌊α⌋=(⌊α1⌋,…,⌊αm⌋). If α∈Γ(K), then α=β+u with xβ∈K and u∈R≥0m, so βi≤αi for each i. Since βi are integers, βi≤⌊αi⌋, and x⌊α⌋=xβx⌊α⌋−β∈K. Conversely, if x⌊α⌋∈K, then α∈⌊α⌋+R≥0m⊂Γ(K).
It follows that
R≥0m∖Γ(K)=xα∈/K⋃(α+[0,1)m),
since for any α∈R≥0m, α∈⌊α⌋+[0,1)m, this m-cube lies in the complement if and only if x⌊α⌋∈/K, and distinct ⌊α⌋+[0,1)m are disjoint. By Lemma 2.1, there exist finitely many such m-cubes, each of volume 1, so the right-hand side is measurable and
Vol(R≥0m∖Γ(K))=#({xα:xα∈/K})=ℓ(K).
(iii) By Lemma 2.1, there exist ai,bi∈N+ such that xiai∈K and xibi∈L. Then xiai+bi=xiaixibi∈KL for each i. By Lemma 2.1, KL has finite colength.
(iv) If xα∈K and xβ∈L, then xα+β=xαxβ∈KL, so
(α+R≥0m)+(β+R≥0m)=(α+β)+R≥0m⊂Γ(KL).
It follows that Γ(K)+Γ(L)⊂Γ(KL). Conversely, let γ∈Γ(KL). Then γ=α+u for xα∈KL and u∈R≥0m. The ideal KL is spanned by products xα′xβ′ with xα′ a monomial of K and xβ′ a monomial of L as an F-vector space. Since the spanning set consists of monomials and monomials are linearly independent, the monomial xα is equal to one of them. Let xα=xα′xβ′ with xα′∈K,xβ′∈L. Therefore, γ=α′+(β′+u)∈Γ(K)+Γ(L), and hence Γ(KL)⊂Γ(K)+Γ(L). So (iv) holds. □
A monomial order is a well order ≤ on the set of monomials in P that satisfies mm1≤mm2 if m1≤m2 for any monomials m,m1,m2∈P. Fix a monomial ordering ≤ on P. For a nonzero polynomial f∈P, we denote by LM(f) its greatest term with respect to ≤, with its coefficient removed.
Lemma 2.3. Let f,g∈P be nonzero. Then LM(fg)=LM(f)LM(g).
Proof. Every monomial m1m2 occurring in fg satisfies m1m2≤LM(f)m2≤LM(f)LM(g) by the definition of monomial order. If there exists a monomial n1n2 occurring in fg such that n1n2=LM(f)LM(g) with n1<LM(f), then n1n2<LM(f)n2≤LM(f)LM(g)=n1n2, which is a contradiction. □
Proposition 2.4. Let K,L be two ideals of P that have finite colength. Then
(1)
ℓ(KL)m1≤ℓ(K)m1+ℓ(L)m1.
Proof. If one of the ideals K,L,KL is equal to P, the inequality is immediate. If one of them is 0, its colength is infinite, which contradicts the assumption that K,L,KL have finite colength. Assume that K,L,KL are nonzero and proper.
First assume that K and L are monomial ideals. Let Θ=Sm−1∩R≥0m where Sm−1 is the unit sphere in Rm, and let dθ be the surface measure on Θ. For θ∈Θ, set
ρK(θ)=min{t≥0:tθ∈Γ(K)}.
By Lemma 2.2 (i), Γ(K) is closed. By Lemma 2.1, K contains a power of every variable; then each ray in R≥0m meets Γ(K). Therefore, ρK(θ) is well-defined. The function ρK is bounded by mmax{ai:1≤i≤m}, where xiai∈K are as in Lemma 2.1, and it is measurable since {θ:ρK(θ)≤t}={θ:tθ∈Γ(K)} is closed for t≥0. In polar coordinates, by Lemma 2.2 (ii),
Since ρK(θ)θ∈Γ(K) and ρL(θ)θ∈Γ(L), by Lemma 2.2 (iv), their sum is in Γ(KL). By Lemma 2.2 (iii), KL has finite colength. Therefore, ρKL(θ)≤ρK(θ)+ρL(θ) for θ∈Θ. Apply Minkowski’s inequality to Lm(Θ,mdθ), the inequality (1) holds.
Now let K and L be arbitrary ideals. Let in(K) be the initial ideal of K, the monomial ideal generated by LM(f) for any f∈K.
The monomials outside in(K) form a basis of P/K under the canonical projection. For independence, let r be a combination of such monomials with r∈K. If r=0, then LM(r) is one of them, hence outside in(K). But r∈K implies LM(r)∈in(K), which is a contradiction. Therefore, r=0, and the monomials outside in(K) are linearly independent. For spanning, let f∈P with LM(f)∈in(K). Then LM(f)=xαLM(h) for some nonzero h∈K, by Lemma 2.3, LM(f)=LM(xαh). Let g=xαh∈K, the element f1=f−LC(f)g, where LC(f) is the coefficient of LM(f) in f, satisfies f1=0 or LM(f1)<LM(f). Iterating, the leading monomial strictly decreases at each step. This terminates since only finitely many monomials are smaller than a given one. The final remainder r is an F-linear combination of monomials outside in(K), and f−r∈K. Therefore, f=r+K in P/K. Therefore, ℓ(K)=ℓ(in(K)).
Lemma 2.5 (Flanders–Atkinson lemma, See [10 Lemma 2.2]). Let F be an infinite field, and let V be a finite-dimensional vector space over F. For a commutative unital algebra R⊂EndF(V), choose v∈V such that r=dimF(Rv) is maximal, and let I=AnnR(v). Then
(i) IV⊂Rv,
(ii) I(Rv)=0,
(iii) I2=0.
Proof. Let b1v,…,brv be a basis of Rv, where b1,…,br∈R. Fix a∈I, w∈V, and for any t∈F, the vectors
b1(v+tw),…,br(v+tw),a(v+tw)
are contained in R(v+tw) of dimension at most r by the choice of v. Therefore, their exterior product vanishes. Since av=0, this gives the polynomial identity
(b1v+tb1w)∧⋯∧(brv+tbrw)∧taw=0.
The field is infinite, so the coefficient of t is zero, i.e.,
b1v∧⋯∧brv∧aw=0.
Since b1v,…,brv are independent, b1v∧⋯∧brv=0, and hence
aw∈Span{b1v,…,brv}=Rv.
Since the choice of a,w is arbitrary, IV⊂Rv, so (i) holds. Condition (ii) holds since commutativity gives a(bv)=b(av)=0 for b∈R. By (i), we have I2V=0. Since R is an algebra of F-linear maps, I2=0, and hence (iii) holds.
Proof of Theorem 1.1 (i)
Proof of Theorem 1.1(i). Let V=Fn, and let R=F[A1,…,Am]. It is easy to see that
dimFF[A1,…,Am]=dimFF[A1,…,Am].
Then we can replace F by its algebraic closure. Without loss of generality, we assume F is algebraically closed. So F is infinite, and hence Lemma 2.5 is available.
Fix m∈N+. We prove this theorem by induction on n=dimFV. For n=1, the matrices A1,…,Am can be regarded as the elements in F. Then F[A1,…,Am]=F and
dimFF[A1,…,Am]=1=(2m−1)(1−1)+1.
Therefore, this theorem holds for n=1. Assuming that this theorem holds for the positive integers that are strictly less than n, we show that this theorem holds for n. Choose the vector v∈V such that r=dimF(Rv) is maximal, and let W=Rv. Since R has an identity, r≥1.
If W=V, then the F-linear map
φ:R→V
a↦av.
is surjective. If b∈Ker(φ), then bv=0. Commutativity gives b(cv)=c(bv)=0 for any c∈R, i.e., bV=0. Since R is an algebra of F-linear maps, b=0. Then Ker(φ)={0}. It follows that φ is also injective. Therefore, φ is an isomorphism between R and V. So we have
dimFR=dimFV=n≤(2m−1)(n−1)+1,
the last inequality holds since 2m−1≥1 for m∈N+.
Assume dimFV/W>0. Regard V as a P=F[x1,…,xm]-module via the action
f(x1,…,xm)⋅v=f(A1,…,Am)v.
Then
R≅P/AnnP(V).(1)
Commutativity and W=Pv give AnnP(W)=AnnP(v), so f↦fv induces P/AnnP(W)≅W, and hence
ℓ(AnnP(W))=dimF(W)=r.(2)
The image of AnnP(W)=AnnP(v) in R=P/AnnP(V) is AnnR(v), by Lemma 2.5(i), AnnP(W)V⊂W, and hence AnnP(W)⊂AnnP(V/W). It follows that
ℓ(AnnP(V/W))≤ℓ(AnnP(W))=r.(3)
Let RV/W be the image of P in EndF(V/W), i.e., the algebra induced by R on V/W, generated by the identity and by the quotient map induced by A1,…,Am. The action map P→EndF(V/W) has kernel AnnP(V/W) and image RV/W, by the rank-nullity theorem and the inductive hypothesis,
where the sum is empty when m=1. This completes the induction. □
Proof of Theorem 1.1(ii)
Preliminaries
Fix integers p,q≥2 and set m=p+q. Let
B=r≥0⨁Br,Br=F[x0,…,xp−1,z0,…,zq−1](r,r),
where the subscript (r,r) denotes the bi-degree, with degree r with respect to the x-variables and degree r with respect to the z-variables. For d≥0, let Vd be an F-vector space with the basis
(iv) for any subalgebra A=⨁r≥0Ar of B, the map ρd induces a faithful action of A/A>d on Vd.
Proof. (i) Counting the number of bases of vα,β is equivalent to counting the number of tuples (α0,⋯,αp−1,β1,⋯,βq−1) satisfying
α0+⋯+αp−1+β0+⋯+βq−1=d.
The latter can be counted by the stars-and-bars method; there are (m−1d+m−1) in total. Therefore,
dimFVd=(m−1d+m−1).
(ii) It is easy to see that CijCkl=CklCij.
(iii) Let xμzν∈Br with r>d, and let vα,β be a basis element of Vd. Then ∣α∣+∣β∣=d. Since ∣ν∣=r>d≥∣β∣, there exists some 0≤i0≤q−1 such that βi0<νi0, and hence xμzν annihilates vα,β. Therefore, Br kills Vd for r>d, i.e., B>d⊂Ker(ρd).
For the reverse inclusion, let P=∑i=0sPi=∑i=0dPi+∑j=d+1sPj∈Ker(ρd) where Pi∈Bi. Then the part ∑j=d+1sPj kills Vd, which implies ∑j=d+1sPj∈Ker(ρd). For the remainder part P′=∑i=0dPi, assume, by contradiction, that P′=0, i.e., P′=∑i=0rPi, where r≤d, Pr=0. Let
Pr′(x,z)=∣ν∣=r∑Pν(x)zν
and choose ν0 with Pν0=0. For any γ∈Np with ∣γ∣=d−r, and the basis element vγ,ν0. Then
ρd(P′)vγ,ν0=∣μ∣=r∑cμvγ+μ,0=0
since ∣ν∣=∣ν0∣=r and ν≤ν0 implies ν=ν0. For i<r,
ρd(Pi′)vγ,ν0=∣ν∣=i∑cνvγ+μ,ν0−ν
where ν0−ν=0. The terms of index <r contain only basis elements whose β-part is nonzero. However, the terms of index r contain basis elements with β-part 0. Then ρd(P)vγ,ν0=ρd(P′)vγ,ν0=0, which contradicts the fact that P∈Ker(ρd). It follows that P′=0, i.e., P=∑j=d+1sPj∈B>d, and hence Ker(ρd)⊂B>d.
(iv) By (iii), A∩Ker(ρd)=A>d, so Ker(ρd∣A)=A>d, and hence ρd induces a faithful action of A/A>d on Vd.
A Subalgebra of B
Lemma 3.2. Let m=p+q, s=p+q−2, and let
hl=0≤i<p0≤j<qi+j=l∑xizj(0≤l≤s),u=xp−1zq−2.
and
A=F[h0,…,hs,u].
Then
(i) the elements h0,…,hs are algebraically independent.
(ii) there is a monic polynomial f∈F[H0,…,Hs,U] of degree e=(p−1m−2) in U and homogeneous of total degree e such that
A≅F[H0,…,Hs,U]/(f).
under Hi↦hi, U↦u. Consequently,
A=i=0⨁e−1F[h0,…,hs]ui.
(iii) Let Rd=ρd(A)=F[ρd(h0),…,ρd(hs),ρd(u)]. Then
Rd≅A/A>d,
and
dimFRd=j=0∑e−1(m−1d−j+m−1)=(md+m)−(md+m−e).
Proof. (i) Let a,γ1,…,γs be algebraically independent indeterminates over F with a=zq−1xp−1, and put
where degf(t)=p−1 and degg(t)=q−1, comparing the coefficient of tl gives
i+j=l∑xizj=(−1)s−laεs−l(γ1,…,γs)
for 0≤l≤s, where εs−l is the (s−l)-th elementary symmetric polynomial. For simplicity, denote εi(γ1,…,γs) as εi. By Artin’s Theorem (see [21 Corollary 3.11]), the field extension
F(γ1,…,γs)/F(γ1,…,γs)Sn
is finite, so they have the same transcendence degree. Since the transcendence degree of F(γ1,…,γs) is s, by the fundamental theorem of symmetric polynomials, the transcendence degree of
F(γ1,…,γs)Sn≅F(ε1,…,εs)
is also s. Therefore, ε1,…,εs are algebraically independent. If a,ε1,…,εs are algebraically dependent over F, there exists a nonzero polynomial P∈F[y0,…,ys] such that
P(a,ε1,…,εs)=i=0∑kQi(ε1,…,εs)ai=0
where Qi∈F[y1,…,ys]. Since a is an indeterminate over F, this equality implies
Q1(ε1,…,εs)=⋯=Qs(ε1,…,εs)=0,
Since ε1,…,εs are algebraically independent, these polynomials Q1,…,Qs are zero polynomial in F[y1,…,ys]. It follows that P=0 in F[y0,…,ys], which contradicts that P is nonzero. Therefore, a,ε1,…,εs are algebraically independent. By the definition of h0,…,hs, we have
hs=a,hs−j=(−1)jaεj(γ1,…,γs).
Therefore, h0,…,hs are algebraically independent.
Since Pi is symmetric and homogeneous of degree i, by the fundamental theorem of symmetric polynomials, Pi(γ1,…,γs) is a linear combination of some ∏j=1sεjbj for some bj≥0 satisfying i=∑j=1sjbj. By Formula (3.2), we have
with Pi homogeneous of degree i. Consider the factor U+a∑i∈{p,…,s}γi=U−u in Formula (7), it follows that F∗(u)=0. Therefore, F(h0,…,hs,u)=0.
Regard F(h0,…,hs) as a subfield of F(a,γ1,…,γs). By Formula (3.2), the field F(h0,…,hs) is generated over F by a and the aεi, and hence also by a and the εi since εi=(−1)ihshs−i. Every permutation of γ1,…,γs fixes a and each εi, so it fixes K, and it sends u through the e distinct elements
−ai∈I∑γi
for some ∣I∣=q−1. Therefore, if G(U)∈F(h0,…,hs)[U] is nonzero and G(u)=0, applying these permutations shows that G has at least e distinct roots, so degUG≥e. Since F∗ has degree e and is monic, it is the minimal polynomial of u over F(h0,…,hs).
Let G(H0,…,Hs,U)∈F[H0,…,Hs,U] satisfy G(h0,…,hs,u)=0. Dividing by the monic polynomial F in the variable U, we have
G=QF+R,
where Q,R∈F[H0,…,Hs][U] and degUR<e. Then R(h0,…,hs,u)=0.
Assume that R=∑i=0kQi(H0,…,Hs)Ui with Qi∈F[H0,…,Hs]. If R=0, then there exists some Qi0=0. Substitute Hj↦hj, then R(h0,…,hs,U)∈F(h0,…,hs)[U] of degree <e. By the minimality of the degree of F, we have each Qi(h0,…,hs)=0. It follows that Qi0(h0,…,hs)=0, which contradicts the fact h0,…,hs are algebraically independent. Therefore, G=QF for some Q∈F[H0,…,Hs,U].
(iii) By Lemma 3.1, we have Ker(ρd∣A)=A>d. Therefore,
Rd≅A/Ker(ρd∣A)=A/A>d,(8)
and Lemma 3.1 show that Rd acts faithfully on Vd.
Let F[h0,⋯,hs]≤d−j be the F-subspace of polynomials of degree ≤d−j. A monomial of total degree at most d−j in s+1=m−1 variables is determined by its exponent vector α=(α0,⋯,αs) such that
α0+⋯+αs≤d−j.(9)
Add one variable αs+1 such that
α0+⋯+αs+αs+1=d−j,(10)
then counting the number of tuples (α0,⋯,αs+1) satisfying Formula (9) is equivalent to counting the number of tuples (α0,⋯,αs+1) satisfying Formula (10). The latter can be counted by the stars-and-bars method; there are (m−1d−j+m−1) in total. Therefore,
By Formula (12), dimFRd is the coefficient of the term tm−1 on the left-hand side, and on the right-hand side, tm−1 has coefficient (md+m)−(md+m−e). Therefore,
dimFRd=(md+m)−(md+m−e).
Proof of Theorem 1.1(ii)
Proof of Theorem 1.1(ii). Fix p,q. By Lemma 3.2, there exists an algebra Rd=F[ρd(h0),…,ρd(hs),ρd(u)] such that
dimFRd=(md+m)−(md+m−e),(13)
where e=(p−1m−2). Then, by Lemma 3.1 and Formula (13),
Lemma 4.1. Let K be a subfield of a field F, and let A1,…,Am∈Mn(F) commute pairwise. Then there exists a finite extension L/K and pairwise commuting matrices A1,…,Am∈Mn(L) such that
dimLL[A1,…,Am]≥dimFF[A1,…,Am].
Proof. Let M1,…,Md be an F-basis of F[A1,…,Am]. Regarding each Mi as a column vector of its n2 entries, we obtain an n2×d matrix over F of rank d, and hence there exists a nonzero d×d minor Δ. Let
B=K[all entries of A1,…,Am,{dij},Δ−1]⊂F.
where {dij} is the set of coefficients of one linear combination of M1,…,Md under the monomials generated A1,…,Am. Only finitely many entries occur, so B is a finitely generated K-algebra. Let m be a maximal ideal of B, and let L=B/m. By the weak version of Hilbert’s Nullstellensatz (see [1 Corollary 7.10]), L is a finite extension of K.
Let π:B→L be the canonical quotient map, and for each i, let Ai be the matrix whose (s,t)-entry is the image under π of the (s,t)-entry of Ai. Since Ai and Aj commute, we have AiAj−AjAi=0. Each entry of the right-hand side is a polynomial in the entries of Ai and Aj, and π is a ring homomorphism. Therefore, every entry of AiAj−AjAi is zero, i.e., AiAj−AjAi=0. It follows that A1,…,Am commute pairwise. Since Δ⋅Δ−1=1 holds in B, we have π(Δ)⋅π(Δ−1)=π(Δ⋅Δ−1)=π(1F)=1L. Then π(Δ) is a unit of L.
For each i, let Mi be the matrix whose (i,j)-entry is the image under π of the (i,j)-entry of Mi. We arrange M1,…,Md in the same way as M1,…,Md. Then the minor Δ corresponding to the position Δ is the image of Δ under π. Since Δ=π(Δ) is a unit of L, i.e., Δ=0, and hence Δ is a nonzero d×d minor. It follows that M1,…,Md are linearly independent. Therefore, we have
dimLL[A1,…,Am]≥d=dimFF[A1,…,Am].
Lemma 4.2. Let L be a field, and let A1,…,Am∈Mn(L) commute pairwise with characteristic polynomials p1,…,pm respectively. Consider the splitting field E over L of p1,…,pm, and let V=En. Then
(i) E/L is a finite extension.
(ii) the algebra E[A1,…,Am] can be decomposed as a direct sum of local E-algebras, i.e.,
E[A1,…,Am]=i⨁Ri.
(iii) V can be decomposed into the corresponding simultaneous generalized eigenspaces
V=j⨁Vj.
(iv) each Rj acts faithfully on Vj.
Proof. (i) Let P(x)=∏i=1mpi(x)∈L[x]. Then E is the splitting field of P over L given by adding finitely many roots of finitely many polynomials. Therefore, E/L is a finite extension. (ii) The E-algebra E[A1,…,Am] has finite dimension over E, hence an Artinian ring. By the structure theorem for Artinian rings, it can be decomposed as a direct sum of local E-algebras
E[A1,…,Am]=i=1⨁sRi.
(iii) Let ei be the idempotent element of Ri. Then
i=1∑sei=1,eiej={ei,0,i=j,i=j.(16)
and Ri=E[A1,…,Am]ei. Let Vi=eiV, by Formula (16), we have
V=i=1⨁sVi.
Since Ri is integral over E, E⊂E[Aiej]⊂Rj, and hence Rj is integral over E[Aiej]. Since Rj is local, by the going-up theorem (see [1 Theorem 5.11]), E[Aiej] has a unique maximal ideal. Let mij(x) be the minimal polynomial of Aiej on Vj. Then
E[Aiej]≅E[x]/(mij(x)E[x]).
Since the left-hand side is local, there exists a unique maximal ideal of the right-hand side. This implies that mij(x)=p(x)t for some irreducible polynomial p(x). Let λ be a root of p(x). Then λ is a root of mij(x), i.e, λ is the eigenvalue of Aiej on Vj, and hence λ∈E. Therefore, p(x)=(x−λ)k, and hence mij(x)=(x−λ)kt. It follows that Aiej has only one eigenvalue on Vj.
(iv) Assume that Aei∈Ri acts as zero in Vi for some A∈E[A1,…,Am]. For any j=1,…,s, Aeiejv=A(0)=0 if i=j, Aeiejv=Aeiv=0 if i=j. So Aei acts as zero on V. Since E[A1,…,Am] acts faithfully on V, and Aei∈E[A1,…,Am], hence Aei=0. Therefore, Ri acts faithfully on Vi. □
Proof of Theorem 1.1 (iii)
Proof of Theorem 1.1 (iii). Let K be the prime subfield of F. For any pairwise commuting matrices A1,…,Am∈Mn(K), it is easy to see
dimKK[A1,…,Am]=dimFF[A1,…,Am]≤cm(F).
Since the choice of A1,…,Am is arbitrary, taking the supremum gives
cm(K)≤cm(F).
For the reverse inequality, let A1,…,Am∈Mn(F) commute pairwise. By Lemma 4.1, there exists a finite extension L/K and pairwise commuting matrices A1,…,Am∈Mn(L) such that
dimLL[A1,…,Am]≥dimFF[A1,…,Am].(17)
Let E be the splitting field over L of the characteristic polynomials of A1,…,Am. Since L/K is finite, so is E/K. By Lemma 4.2, there exist decompositions
E[A1,…,Am]=i⨁Ri,V=j⨁Vj
such that Rj acts faithfully on Vj. It follows that
and Rj can be regarded as the subalgebra of EndE(Vj) for each j. Since ndimEVj≤1, by Formula (18), there exists at least one component W=Vj0 such that
dimEWdimERj0≥ndimEE[A1,…,Am],(19)
otherwise the right-hand side is strictly less than the left-hand side in Formula (18). Let λi be the eigenvalue of Aiej0W in W, and let Ni=Aiej0W−λi. Then N1,…,Nm are pairwise commuting nilpotent operators on W, and Rj0=E[N1,…,Nm].
Since E/L and L/K are finite extensions, the field extension E/K is finite. The field K is the prime subfield of F, so K is perfect, and hence the finite extension L/K is separable. By the primitive element theorem, there exists an element α∈E such that E=K(α). Regard W as an K-vector space, define
B1=αI+N1,Bi=Ni(2≤i≤m).
These m operators commute pairwise. The part αI is semisimple over K since its minimal polynomial is separable, and it commutes with N1. By Jordan-Chevalley decomposition, the uniqueness of the decomposition gives that the semisimple part of B1 is αI, and hence there exists a polynomial g(x)∈K[x] such that αI=g(B1). Therefore,
αI∈K[B1]andK[B1,…,Bm]=E[N1,…,Nm]⊂EndK(W).
The action is faithful since restriction of scalars does not identify distinct endomorphisms. Therefore, by Formula (17), we have
Since the choice of A1,⋯,Am is arbitrary, taking the supremum gives cm(F)≤cm(K). □
Acknowledgement.
Acknowledgement. We would like to thank Prof. Kaiming Zhao for his encouragement and helpful suggestions, as well as for informing us of an interesting and beautiful recent work [7]. We would also like to thank the referee for his/her detailed comments and kindness.
AI disclosure.
AI disclosure. During the preparation of this manuscript, the authors used LLMs to assist in drafting and checking parts of the mathematical arguments, organize references, and improve the clarity and readability of the exposition. All mathematical statements and proofs in the final manuscript were written, reviewed, and verified by the authors, who take full responsibility for the content of this manuscript.
SCHOOL OF MATHEMATICAL SCIENCES, HEBEI NORMAL UNIVERSITY, SHIJIAZHUANG, HEBEI 050024, CHINA. Email address:[email protected]
SCHOOL OF SCIENCE AND ENGINEERING, THE CHINESE UNIVERSITY OF HONG KONG, SHENZHEN, GUANGDONG 518172, CHINA. Email address:[email protected]
SCHOOL OF MATHEMATICAL SCIENCES, HEBEI NORMAL UNIVERSITY, SHIJIAZHUANG, HEBEI 050024, CHINA. Email address:[email protected]
SCHOOL OF SCIENCE AND ENGINEERING, THE CHINESE UNIVERSITY OF HONG KONG, SHENZHEN, GUANGDONG 518172, CHINA. Email address:[email protected]
References
[1]M. F. Atiyah and I. G. Macdonald. Introduction to commutative algebra. Addison-Wesley Publishing Co., Reading, Mass.-London-Don Mills, Ont., 1969.
[2]José Barría and P. R. Halmos. Vector bases for two commuting matrices. Linear and Multilinear Algebra, 27(3):147–157, 1990.DOI
[3]George M. Bergman. Commuting matrices, and modules over artinian local rings. arXiv preprint arXiv:1309.0053, 2013.arxiv.org/abs/1309.0053
[4]I. N. Bernstein. All reductive p-adic groups are tame. Functional Analysis and Its Applications, 8(2):91–93, April 1974.
[5]I. N. Bernstein and A. V. Zelevinsky. Representations of the group GL(n, F), where F is a local non-Archimedean field. Uspehi Mat. Nauk, 31(3(189)):5–70, 1976.
[6]J. Bernstein. Representations of p-adic groups. Lectures at Harvard University, Fall 1992, written by K. E. Ruelumhart, unpublished notes, 1992.
[7]Cunguang Cheng, Gengqiang Liu, Kaiming Zhao, and Yueqiang Zhao. Commutative subalgebras of matrix algebras. Manuscript in preparation, 2026.
[8]Ron Cherny, Tam An Le Quang, and Matthew Satriano. On combinatorial algebras generated by three commuting matrices. arXiv preprint arXiv:2511.22092, 2025.arxiv.org/abs/2511.22092
[9]Ron Cherny, Matthew Satriano, and Yohan Song. On the algebra generated by three commuting matrices: combinatorial cases. Electron. J. Combin., 31(4):Paper No. 4.43, 20, 2024.arxiv.org/abs/2402.16334
[10]Clément de Seguins Pazzis. Local linear dependence seen through duality II. Linear Algebra Appl., 462:133–185, 2014.DOI
[11]Murray Gerstenhaber. On dominance and varieties of commuting matrices. Ann. of Math. (2), 73:324–348, 1961.DOI
[12]Robert M. Guralnick. A note on commuting pairs of matrices. Linear and Multilinear Algebra, 31(1–4):71–75, 1992.DOI
[13]Robert M. Guralnick and B. A. Sethuraman. Commuting pairs and triples of matrices and related varieties. Linear Algebra Appl., 310(1–3):139–148, 2000.DOI
[14]Yongho Han. Commuting triples of matrices. Electron. J. Linear Algebra, 13:274–343, 2005.DOI
[15]J. Holbrook and K. C. O’Meara. Some thoughts on Gerstenhaber’s theorem. Linear Algebra Appl., 466:267–295, 2015.DOI
[16]John Holbrook and Kevin C. O’Meara. A computing strategy and programs to resolve the gerstenhaber problem for commuting triples of matrices. arXiv preprint arXiv:2006.08588, 2020.arxiv.org/abs/2006.08588
[17]John Holbrook and Matjaž Omladič. Approximating commuting operators. Linear Algebra Appl., 327(1–3):131–149, 2001.DOI
[19]Joachim Jelisieiew and Klemen Šivic. Components and singularities of Quot schemes and varieties of commuting matrices. J. Reine Angew. Math., 788:129–187, 2022.arxiv.org/abs/2106.13137
[20]Thomas J. Laffey and Susan Lazarus. Two-generated commutative matrix subalgebras. Linear Algebra Appl., 147:249–273, 1991.DOI
[22]M. Mirzakhani. A simple proof of a theorem of Schur. Amer. Math. Monthly, 105(3):260–262, 1998.DOI
[23]T. S. Motzkin and Olga Taussky. Pairs of matrices with property L. II. Trans. Amer. Math. Soc., 80:387–401, 1955.DOI
[24]Nham V. Ngo and Klemen Šivic. On varieties of commuting nilpotent matrices. Linear Algebra Appl., 452:237–262, 2014.arxiv.org/abs/1308.4438
[25]Matjaž Omladič. A variety of commuting triples. Linear Algebra Appl., 383:233–245, 2004.DOI
[26]Jenna Rajchgot and Matthew Satriano. New classes of examples satisfying the three matrix analog of Gerstenhaber’s theorem. J. Algebra, 516:245–270, 2018.arxiv.org/abs/1711.10109