We provide a new class of finitely generated groups whose word-length functions give rise to compact quantum metric spaces. Kaad's recent theory of matrix compact quantum metric spaces is instrumental in obtaining this new class, and along the way we also show that any length function on an amenable group which gives rise to a compact quantum metric space in fact gives rise to a matrix compact quantum metric space, and that the corresponding result is false for the free group on two generators.
Introduction
Over the past decade, substantial progress has been made in developing Rieffel’s theory of compact quantum metric spaces, which provides a far-reaching non-commutative generalisation of classical compact metric spaces, with strong ties to Connes’ non-commutative geometry and theoretical physics. Recent advances in the field [1, 13, 19, 6], suggest that basing the theory on operator systems, rather than order unit spaces or full C∗-algebras, provides a good blend of flexibility and rigidity, and with this in mind it is also very natural to seek a matricial version of the theory. A first approach to a matricial theory was developed by Kaad in [8], and the resulting theory of matrix compact quantum metric spaces has, among other things, the advantage that it admits a natural tensor product construction, compatible with the external product of spectral triples. The present paper continues the investigation of matrix compact quantum metric spaces, with particular emphasis on the examples arising from groups. These examples fit into the more general connection between compact quantum metric spaces and non-commutative (differential) geometry, which we now describe in more detail. Firstly, a compact quantum metric space consists of an operator system X together with a seminorm L:X→[0,∞), such that the associated Monge-Kantorovič metric
mkL(φ,ψ):=sup{∣φ(x)−ψ(x)∣∣x∈X,L(x)≤1},φ,ψ∈S(X),
induces the weak* topology on the state space S(X). The seminorm L may be viewed as a generalisation of the example arising from a classical compact metric space (M,d), where X is the algebra of Lipschitz continuous functions and L the seminorm that assigns the (smallest) Lipschitz constant to the function in question. In the same vein, a matrix compact quantum metric space consists of an operator system X together with a family of seminorms Ln:Mn(X)→[0,∞) satisfying a natural set of axioms; see § 2 for details.
On the other hand, the fundamental object in Connes’ non-commutative geometry [4] is the notion of a spectral triple, and from a spectral triple (A,H,D) one obtains a natural seminorm LD:A→[0,∞) given by LD(a):=∥[D,a]∥. In many interesting examples, but not always, this gives rise to a compact quantum metric space. From a discrete group G equipped with a proper length function ℓ:G→[0,∞), one obtains a natural spectral triple (C[G],ℓ2(G),Dℓ), and even in this situation it has proven very difficult to determine when one obtains a compact quantum metric structure from the associated seminorm LDℓ. On the positive side, this has been verified for hyperbolic groups and groups of polynomial growth in [14] and [5], respectively, and only very recently the first counterexample for a word-length function was found by Klisse [10]. The aim of the present paper is two-fold: First, we answer, negatively, a question posed in [8], asking if any compact quantum metric space is in fact automatically a matrix compact quantum metric space. Kaad conjectured that this is not the case, and we confirm this conjecture by showing that the free group on two generators provides a natural counterexample.
Theorem A (see Theorem 3.6). For the free group F2 on two generators, a and b, the word-length function arising from the generating set S:={a,a−1,b,b−1} gives rise to a compact quantum metric space, which is not a matrix compact quantum metric space.
Contrasting this, we provide an affirmative answer in the presence of amenability.
Theorem B (see Theorem 3.1). If G is an amenable group equipped with a proper length function yielding a compact quantum metric space, then it is automatically a matrix compact quantum metric space.
Secondly, exploiting the tensor product construction for matrix compact quantum metric spaces, we provide a new class of groups for which one does obtain a compact quantum metric structure.
Theorem C (see Theorem 4.10). Let G and K be finitely generated groups equipped with word-length functions ℓG and ℓK. If (C[G],LDℓG) is a compact quantum metric space and K is either finite, abelian or the infinite dihedral group, then (C[G×K],LDℓ) is a compact quantum metric space for the length function ℓ(g,k)=ℓG(g)+ℓK(k).
Theorem C does indeed provide genuinely new examples, in that one may, for instance, take G to be any non-elementary hyperbolic group and K to be the integers, then G×K is neither hyperbolic (as it contains a copy of Z2) nor of polynomial growth (as G must have exponential growth), and is thus not covered by the existing results in [14] and [5].
Notation. Throughout the article, all C∗-algebras are unital, and all spectral triples are faithfully and unitally represented. Moreover, inner products are assumed linear in the first variable. Algebraic tensor products are denoted by ⊙, while ⊗min will denote the minimal tensor product of C∗-algebras, while ⊗ will be used for tensor products in the category of Hilbert spaces. A unital completely positive map between operator systems will be referred to as a UCP map.
Acknowledgments. The present paper grew out of a series of interesting discussions with Jens Kaad and Nadia S. Larsen in the fall of 2025, and it is a pleasure to thank them both for sharing their insights. We also wish to acknowledge that the fundamental idea of geodesic cancellation, which plays a pivotal role in the proof of Theorem C, was originally discovered by Jens in the case of the integers. That Theorem 3.1 and Theorem 4.2 should be true, was also suggested by Jens. The first named author would also like to thank the University of Southern Denmark for its hospitality during a visit in November 2025, during which some of this work was conducted. The first named author was funded by the Research Council of Norway projects 324944 and 359602. The second named author was funded by the Independent Research Fund Denmark through grant no. 1026-00371B and by the European Commission through the MSCA Staff Exchanges grant no. 101086394.
AI declaration. The present paper consists of a mixture of human-generated results and AI-assisted proofs. As indicated above, the fundamental idea of geodesic cancellation leading to Theorem C was discovered and utilised by the authors (without AI assistance but through conversations with Jens Kaad and Nadia S. Larsen) in the case where K is Z or Z2, but the full class of groups covered in Theorem C was then obtained via ChatGPT through a carefully designed research prompt. Similarly, Theorem B was obtained without any AI assistance, whereas the counterexample in Theorem A was obtained through a dialogue with ChatGPT, to explore if Theorem B could be extended to also cover the class of hyperbolic groups (which is then shown to be impossible by Theorem A).
Preliminaries
The present section introduces the necessary background concerning compact quantum metric spaces and their matrix analogues. In the interest of brevity, we will start by introducing the matrix version.
Matrix compact quantum metric spaces
Following [8], we choose the operator system approach to (matrix) compact quantum metric spaces. By an operator system, we mean a unital, ∗-invariant subspace X of a unital C∗-algebra A. Note that we do not require it to be norm closed. As defined, the ambient C∗-algebra is part of the definition of the operator system X, but we remark that the theory also admits an abstract description; see [15] for details.
We will tacitly identify the scalars inside an operator system X as C⋅1X. A Lipschitz seminorm on an operator system X, is a ∗-invariant seminorm L:Dom(L)→[0,∞), defined on a dense operator-subsystem Dom(L)⊆X, which vanishes on the unit. As is common in the field, we will sometimes regard such a seminorm as an extended seminorm on the whole operator system by setting it to ∞ outside its domain. To discuss matrix compact quantum metric spaces, we will adopt the terminology developed in [8]. A Lipschitz operator seminorm on an operator system X is a family L=(Ln)n≥1 of ∗-invariant seminorms on Mn(X), which vanish on Mn(C), such that
for all x∈Mn(X), y∈Mm(X) and scalar matrices α,β∈Mn(C).
Notation 2.1. For an operator system X and an n∈N, we will denote by [⋅]:Mn(X)→Mn(X)/Mn(C) the natural quotient map. Moreover, for a linear map T:X→Y between operator systems, we denote by Tn:Mn(X)→Mn(Y) its entrywise matrix amplification.
Definition 2.1 ([8]). A pair (X,L) consisting of an operator system X and a Lipschitz operator seminorm L=(Ln)n∈N is said to be a matrix compact quantum metric space (MCQMS) if the following conditions hold:
There exists an R>0 such that
∥[x]∥≤R⋅Ln(x),for all n≥1,x∈Mn(X).
For every ε>0 there exist an operator system Yε, a unital complete isometry ιε:X→Yε, and a finite-rank unital completely positive map Φε:X→Yε satisfying
∥(ιε)n(x)−(Φε)n(x)∥≤ε⋅Ln(x)for all n≥1,x∈Mn(X).
Note in particular that R and the approximating maps are independent of the matrix size n. For a single Lipschitz seminorm L, (X,L) is called a compact quantum metric space (CQMS) if these conditions hold at n=1, with L1=L, and where ιε is only required to be a unital linear isometry and Φε a bounded, unital positive finite-rank linear map. In this case, L is called a Lip-norm.
Note first that any MCQMS is a CQMS by restricting the matrix approximations to the n=1 case, see also [8 Proposition 5.4]. Furthermore, the notion of CQMS from Definition 2.1 agrees with Rieffel’s usual definition by [8 Theorem 3.1]. In particular, a Lipschitz seminorm L:X→[0,∞) is a Lip-norm, if and only if the Monge-Kantorovič metric
mkL(ϕ,ψ):=sup{∣ϕ(x)−ψ(x)∣∣x∈X,L(x)≤1},ϕ,ψ∈S(X),
induces the weak∗ topology on the state space S(X). Given a classical compact metric space (M,d), one obtains [8 Proposition 5.7] an MCQMS by setting Dom(Ln):={f:M→Mn(C)∣f Lipschitz continuous} and
Ln(f):=sup{d(p,q)∥f(p)−f(q)∥p,q∈M,p=q}.
Here we tacitly identify C(M,Mn(C)) with Mn(C(M)).
We end this section, by recalling the following well known useful criterion.
Proposition 2.2. Fix μ∈S(X), and put
EL,μ={x∈X∣L(x)≤1,μ(x)=0}.
Then (X,L) is a CQMS if and only if EL,μ is totally bounded in norm. In this case, kerL=C, the radius Rμ:=supx∈EL,μ∥x∥ is finite, and
∥x−μ(x)1∥≤Rμ⋅L(x)
for all x∈X. Moreover, for every s,t≥0, the set {x∈X∣L(x)≤s,∣μ(x)∣≤t} is totally bounded.
Proof. The statement on total boundedness of EL,μ follows from [14 Proposition 1.3]. It is well known that if (X,L) is a CQMS, then kerL=C; see e.g. [9 Lemma 2.2] for a proof. The statement ∥x−μ(x)1∥≤RμL(x) follows from noting that x−μ(x)1∈L(x)⋅EL,μ. Total boundedness of {x∈X∣L(x)≤s,∣μ(x)∣≤t} now follows by noting that it is contained in the totally bounded set sEL,μ+{c1∣∣c∣≤t}.
Many interesting non-commutative examples of (M)CQMS arise from spectral triples [4], and these are the primary objects of study in the present paper. In fact, we will be focusing (almost) exclusively on the spectral triples that arise from length functions on discrete groups, for which the non-commutative metric geometry is already very complex [14, 5, 10].
Spectral triples and length functions on groups
To set the notation, we recall that a unital spectral triple (A,H,D) consists of a unital ∗-algebra faithfully represented on the Hilbert space H together with a self-adjoint, possibly unbounded operator D on H with compact resolvent, such that A preserves DomD and each commutator [D,a] is bounded. Given a spectral triple (A,H,D), we denote the C∗-closure of A by A. Note that one obtains a semi-norm LD on A by setting
LD(a)=∥[D,a]∥,
which in many interesting cases, but not always [10, 11], gives rise to a CQMS. A spectral triple (A,H,D) is called a spectral metric space if its commutator seminorm is a Lip-norm, i.e., if (A,LD) is a CQMS. Amplifying to matrices, we obtain a seminorm LD,n:Mn(A)→[0,∞) by setting
LD,n(x):=∥[1⊗D,x]∥
and (A,H,D) is called a matrix spectral metric space if (A,(LD,n)n≥1) is an MCQMS; see [8]. We will need to amplify unbounded operators as well. If D:Dom(D)→H is an unbounded operator on a Hilbert space H and K is some auxiliary Hilbert space then we obtain an unbounded operator D⊙1:Dom(D)⊙K→H⊗K defined by (D⊙1)(ξ⊗η)=Dξ⊗η. If D is furthermore selfadjoint, then D⊙1 is essentially selfadjoint and we denote its selfadjoint closure by D⊗1. For two spectral triples (A,HA,DA),(B,HB,DB), put
L(1)(z)=∥[DA⊙1,z]∥,L(2)(z)=∥[1⊙DB,z]∥,
for z∈A⊙B. Note that these are indeed well defined, since for z=∑k=1nak⊗bk∈A⊙B preserves Dom(D)⊙HB, and on this subspace we have
[DA⊙1,z]=k=1∑n[DA,ak]⊗bk,
which is bounded (and similarly for L(2)).
The classical commutative example, upon which the definition of a spectral triple is modelled, arises by choosing A to be the algebra of smooth functions on a compact, connected spin manifold and D to be the associated Dirac operator. An interesting source of non-commutative examples is obtained by considering length functions on discrete groups, as was already realised by Connes in his seminal paper [3].
Definition 2.3. By a length function function on a discrete group G we will mean a function ℓ:G→[0,∞) with
ℓ(g)=0 if and only if g=e,
ℓ(g−1)=ℓ(g) for all g∈G,
ℓ(gh)≤ℓ(g)+ℓ(h) for all g,h∈G.
We say ℓ is proper if the balls Bℓ(t):={g∈G∣ℓ(g)≤t} are finite for all t≥0.
Assume that ℓ is proper. We represent C[G]⊆C∗(G) on ℓ2(G) through the left regular representation λgδh=δgh and denote by (δg)g∈G the standard orthonormal basis in ℓ2(G). The relation Dℓ:span{δg∣g∈G}→ℓ2(G) defined by Dℓδg:=ℓ(g)⋅δg is essentially selfadjoint, and its closure Dℓ has compact resolvent by properness of ℓ. Each λg preserves the domain DomDℓ={ξ∈ℓ2(G)∣∑gℓ(g)2∣ξ(g)∣2<∞}, and
[Dℓ,λg]δh=(ℓ(gh)−ℓ(h))δgh,(1)
and hence ∥[Dℓ,λg]∥≤ℓ(g). Thus, (C[G],ℓ2(G),Dℓ) is a unital spectral triple. Set
Lℓ,n(x):=∥[1n⊗Dℓ,x]∥,andLℓ:=Lℓ,1.
We record that it is known that (C[G],ℓ2(G),Dℓ) is a spectral metric space if ℓ has bounded doubling [5], or if G is word-hyperbolic and ℓ is the word length function on G[14] — or more generally a length function whose induced metric is hyperbolic and weakly geodesic [2]. A recent example by Klisse [10] shows that a word length on (Z/2)≀F2 does not yield a spectral metric space.
For discrete groups G and K with length functions ℓG and ℓK, respectively, we obtain a natural length function on G×K by setting
ℓ(g,k):=ℓG(g)+ℓK(k).
Note that the Dirac operator Dℓ agrees with DℓG⊗1+1⊗DℓK on the dense subspace cc(G)⊙cc(K)⊆ℓ2(G)⊗ℓ2(K). Controlling ∥[DℓG⊗1,z]∥ and ∥[1⊗DℓK,z]∥ in terms of ∥[Dℓ,z]∥ will turn out to be a crucial step in our strategy for exhibiting a CQMS structure on the product group G×K using ℓ=ℓG+ℓK. We will employ a geodesic cancellation argument, see § 4.2.
Lastly, note that if ℓG and ℓK are word length functions arising from symmetric generating sets SG and SK for G and K, respectively, then the length function on G×K defined above is also a word length function corresponding to the generating set {(s,e),(e,t)∣s∈SG,t∈SK}.
Groups and matrix compact quantum metric spaces
In this section we investigate the question of whether any CQMS is in fact an MCQMS. This question already appears in [8], in which Kaad conjectures that this is not generally the case. We first show that in the context of length functions on amenable groups, one does indeed automatically obtain an MCQMS from a CQMS, and then proceed to show that the free group on two generators provides a counter example.
Theorem 3.1. Let G be an amenable, countable discrete group and ℓ a proper length function on G. If (C[G],Lℓ) is a CQMS, then (C[G],(Lℓ,n)n≥1) is an MCQMS.
The proof is a modification of the argument for [8 Theorem 5.4]. Write A=Cr∗(G), H=ℓ2(G), and D=Dℓ, and let Δ:C[G]→C[G]⊙C[G] be the coproduct determined by Δ(λg)=λg⊗λg. We shall use the isometric ∗-isomorphism
That ιp is indeed an isometry follows by applying the isomorphism Mp(E)≅Mp(C)⊗minE and associativity of the minimal tensor product. We use the same notation for its restrictions to the relevant subalgebras. We also use the isometric map
Indeed, κp is the matrix amplification of the injective unital ∗-homomorphism T↦1⊗T and hence isometric. For ω∈A∗ and z∈Mp(C[G]⊙C[G]), these identifications give the identity
(idA⊗ω)p(z)=(idMp(A)⊗ω)ιp−1(z).(2)
We note the following lemma.
Lemma 3.2. For every p≥1, every μ∈Mp(A)∗, and every x∈Mp(C[G]), we have
Lℓ((μ⊗id)ιp−1Δp(x))≤∥μ∥⋅Lℓ,p(x).
Proof. Consider the Kac–Takesaki unitary W∈B(H⊗H) given by
W(δs⊗δt)=δt−1s⊗δt.
It satisfies W∗(1⊗λg)W=λg⊗λg. Consequently, Δ extends to an injective unital ∗-homomorphism A→A⊗minA, given by a↦W∗(1⊗a)W. Both W and W∗ preserve the dense subspace C:=cc(G)⊙cc(G)⊆ℓ2(G)⊗ℓ2(G) and commute thereon with 1⊗D. For a∈C[G], it follows that
[1⊗D,Δ(a)]=W∗(1⊗[D,a])W.(3)
The identity holds on C, and hence for the bounded extensions. Denote by eij the standard matrix units in Mp(C) and write λgeij for the matrix in Mp(C[G]) whose (i,j)-entry is λg and whose other entries are zero. For μ∈Mp(A)∗, we get
for every x∈Mp(C[G]). Since any bounded functional is completely bounded with cb-norm equal to its norm [15 Proposition 3.8], we have ∥μ⊗id∥≤∥μ∥ on Mp(A)⊗minB(H). Using this bound together with the fact that ιp−1 and κp are isometries and W is unitary, we now get
The Følner property gives χn(λg)=∣Fn∩gFn∣/∣Fn∣→1 for every g∈G. Thus the trivial representation ϵ(∑gagλg):=∑gag is the pointwise limit of χn on C[G] and, by amenability, ϵ extends to a state on A. By density, we therefore have limnχn=ϵ in the weak-∗ topology. The CQMS hypothesis implies that δn:=mkLℓ(ϵ,χn)→0 and by definition of mkLℓ we have
∣(ϵ−χn)(b)∣≤δn⋅Lℓ(b)for b∈C[G].(5)
Consider now the maps βn=(id⊗χn)Δ:C[G]→C[G]. They are UCP, and they have finite rank because βn(λg)=χn(λg)λg and χn(λg)=0 if g∈/FnFn−1. We now show that these maps give matricial approximations. Fix p≥1, x∈Mp(C[G]), and μ∈Mp(A)∗ with ∥μ∥≤1. The identity (id⊗ϵ)Δ=id gives
for all n≥0 and all p≥1. For a given ε>0, choose n with δn≤ε. Then (id,βn) is a matricial ε-approximation in C[G], with the same maps for every matrix size.
Finally, F0={e} gives χ0=τ and β0(a)=τ(a)1 where τ denotes the canonical trace on A. Applying (6) with n=0 yields
∥[x]∥≤∥x−(β0)p(x)∥≤mkLℓ(ε,τ)⋅Lℓ,p(x),
and R:=mkLℓ(ε,τ) is finite because mkLℓ metrizes the weak-* topology on the compact state space. This proves uniformly finite matrix diameter and completes the proof. □
Note in particular that the examples considered in [5] have bounded doubling and are therefore amenable. As noted therein, any finitely generated group with polynomial growth has bounded doubling with respect to any word length. We have the following immediate corollary.
Corollary 3.3. If G is a finitely generated discrete group and ℓ is a proper length function with bounded doubling, then (C[G],(Lℓ,n)n≥1) is an MCQMS.
In addition to groups of polynomial growth, it is also known that word-length functions on hyperbolic groups yield compact quantum metric spaces [14]. Most hyperbolic groups are non-amenable, and as we shall now see the situation is quite different in this case. Indeed, we proceed to show that F2 with its standard word length function does not yield an MCQMS, despite it being hyperbolic and thus yielding a CQMS by the results in [14]. Before proceeding to the proof of this result, we recall the rank-one case of the projection-commutator identity in [18 Example 6.2].
Lemma 3.4. Let B=B∗∈MN(C), and suppose that v∈CN is a unit vector. Then the projection p=vv∗ satisfies
∥[B,p]∥2=∥Bv∥2−∣v∗Bv∣2.
In particular, if v=N−1/2(1,…,1)t and B=diag(b1,…,bN) with bj∈R, then
∥[B,p]∥2=N1j=1∑Nbj2−(N1j=1∑Nbj)2.
Proof. By [18 Example 6.2] we have ∥[P,A]∥=max{∥PA(1−P)∥,∥(1−P)AP∥} for a projection P and a general element A in a C∗-algebra. For P=p=vv∗ and A=B=B∗ the two terms coincide, and we get
Remark 3.5. Considering X={1,…,N} as a probability space with respect to the uniform probability measure μ, then if b:X→R is some random variable, we have that
The righthand side of the last formula in Lemma 3.4 may therefore be interpreted as the variance of a random variable, a point of view that will be important in the proof of the following result.
Theorem 3.6. Let G=F2 be the free group on two generators, which we denote a and b, and let ℓ be the word length for the symmetric generating set{a,a−1,b,b−1}. Then (C[F2],Lℓ) is a CQMS, but (C[F2],(Lℓ,n)n≥1) is not an MCQMS. More precisely, for every operator system Y, every linear complete isometry ι:C[F2]→Y, and every finite-rank linear map Φ:C[F2]→Y bounded with respect to the reduced C∗-norm, we have
Then WR∗PRWR=[NR1]g,h∈SR=pR⊗1. Note that WR and WR∗ preserve the domain of INR⊗D. On the summand CNR⊗Cδx the operator WR∗(INR⊗D)WR acts as Bx:=diagg∈SRℓ(gx). Using that WR is unitary, Lemma 3.4 then gives
where in the last equality we consider the function SR∋g↦ℓ(gx)∈R as a random variable on SR endowed with the uniform probability measure; see e.g. Remark 3.5.
For x∈F2 and g∈SR, denote by C(g,x) the number of cancellation pairs in the reduced product gx. Then
ℓ(gx)=R+ℓ(x)−2C(g,x).
We now fix an x∈F2 and consider the random variable C:SR→R given by C(g):=C(g,x). If the reduced word for x is t1⋯tm, the event C≥k, for 1≤k≤min(R,m), prescribes the suffix tk−1⋯t1−1 to g. There are 3R−k such words, so
P(C≥k)=4⋅3R−13R−k=4⋅3k−11.
For larger k the probability is zero. We now note that C2=∑k=1∞(2k−1)1C≥k; indeed, if C(g)=j then it holds that j2=∑k=1j(2k−1). We thus obtain
Finally, set XR=PR/3. Then Lℓ,NR(XR)≤1, and ι being a complete isometry implies
∥ιNR(XR)−ΦNR(XR)∥≥31(1−∥ΦNR(PR)∥).
Letting R→∞ proves (3.6). In particular, the finite rank UCP maps needed to establish an MCQMS cannot give an approximation with error constant smaller than 1/3. □
Cartesian products
Our next aim is to show how one may use the theory of matrix compact quantum metric spaces to expand on the class of groups with length functions that yield compact quantum metric spaces. As mentioned already, the only structural results in this direction cover the classes of hyperbolic groups [14] and of groups of polynomial growth [5]. We first develop some general machinery that can be used to verify that a certain length function on a cartesian product of two groups yields a CQMS, and then proceed to verify that the relevant assumptions are met in a number of instances.
A tensor product criterion
We first recall some standard properties of the minimal tensor product. Let A, B, and C be unital C∗-algebras. For every w∈C⊗minB, we have
∥w∥=sup{∥(idC⊗ψ)(w)∥∣n≥1,ψ∈UCP(B,Mn(C))}(12)
which may be seen by representing B faithfully on a Hilbert space and considering the UCP maps arising from compressions to finite-dimensional subspaces.
By injectivity of the minimal tensor product, a unital complete isometry j:A→C induces a complete isometry j⊗idB:A⊗minB→C⊗minB, see [12 Theorem 4.6]. Finally, every bounded linear functional f∈C∗ satisfies ∥f∥cb=∥f∥ by [15 Proposition 3.8]. The completely bounded mapping property of the minimal tensor product (see e.g. [16]) therefore gives
∥(f⊗idB)(w)∥≤∥f∥∥w∥,w∈C⊗minB.
Proposition 4.1. Let (X,LX) and (Y,LY) be CQMSs and let M be a Lipschitz seminorm on X⊙Y. Suppose that there exist cX,cY>0, such that
LX((id⊗g)z)≤cX∥g∥M(z)(g∈Y∗),
LY((f⊗id)z)≤cY∥f∥M(z)(f∈X∗).
Suppose, moreover, that for every ε>0 there exist a unital C∗-algebra Cε, a unital complete isometry ιε:X→Cε, and a bounded linear map Φε:X→Cε of finite rank satisfying
∥((ιε−Φε)⊗id)z∥≤εM(z),z∈X⊙Y.(13)
Then (X⊙Y,M) is a CQMS.
Here we consider X⊙Y with the operator system structure induced by the embedding into the minimal tensor product of their enveloping C∗-algebras.
Proof. Choose states μ∈S(X) and ν∈S(Y). By Proposition 2.2 it suffices to show that
EM,μ⊗ν={z∈X⊙Y∣M(z)≤1,(μ⊗ν)(z)=0}
is totally bounded in norm. Fix ε>0 and the maps ιε,Φε:X→Cε ensured by the assumptions. Choose a basis e1,…,er of RanΦε. We may then write Φε(x)=∑i=1rfi(x)ei with fi∈X∗. Defining bi(z)=(fi⊗id)z, we then have
(Φε⊗id)z=i∑ei⊗bi(z),andLY(bi(z))≤cY∥fi∥,
for z∈EM,μ⊗ν. For z∈EM,μ⊗ν, we now define az:=(id⊗ν)z. Then LX(az)≤cX and μ(az)=0. Setting
ELX,μ:={x∈X∣LX(x)≤1,μ(x)=0}
and RX:=supx∈ELX,μ∥x∥, we have that az/cX∈ELX,μ and
we deduce that the former set is totally bounded for i=1,…,r. Hence, (Φε⊗id)(EM,μ⊗ν) is totally bounded as well.
We may therefore choose z1,…,zN∈EM,μ⊗ν such that (Φε⊗id)z1,…,(Φε⊗id)zN form an ε-net in (Φε⊗id)(EM,μ⊗ν). Since ιε⊗id is isometric, for each z∈EM,μ⊗ν and a suitable i∈{1,…,N} we have
by using (4.2). It follows that EM,μ⊗ν is totally bounded.
We now specialise to spectral triples, which are the main case of interest in the present paper. For the statement, recall that a spectral triple (A,HA,DA) is called a (matrix) spectral metric space if the commutator seminorm yields a (M)CQMS; see § 2.2 for details
Theorem 4.2. Suppose (A,HA,DA) is a matrix spectral metric space and (B,HB,DB) is a spectral metric space. For z=∑kak⊗bk∈A⊙B, set
P(z)=k∑[DA,ak]⊗bkandQ(z)=k∑ak⊗[DB,bk].
If M is a Lipschitz seminorm on A⊙B for which there exist cA,cB>0 such that
∥P(z)∥≤cAM(z)and∥Q(z)∥≤cBM(z),(14)
for all z∈A⊙B, then (A⊙B,M) is a CQMS. The analogous assertion holds with the two factors interchanged.
Proof. We verify the hypotheses of Proposition 4.1. Denote by A and B the respective C∗-completions, and put LDA,n(x)=∥[1n⊗DA,x]∥. Let δ>0 be given, and use that (A,HA,DA) is a matrix spectral metric space to obtain a C∗-algebra C along with maps ι,Φ:A→C such that
∥ιn(x)−Φn(x)∥≤δ⋅LDA,n(x)(15)
for all n≥1 and all x∈Mn(A). Here ι is a complete isometry and Φ is UCP with finite-dimensional range; in particular both maps extend to A.
Fix now an n∈N, a UCP map ψ:B→Mn(C) as well as a z∈A⊙B, and set xψ=(id⊗ψ)z. Under the canonical matrix identification A⊙Mn(C)≅Mn(A)≅Mn(C)⊙A we then have
where the first inequality follows from (15) and the last inequality from (16). Since this holds for all n∈N and all ψ∈UCP(B,Mn(C)), (12) and (14) imply
∥((ι−Φ)⊗id)z∥≤δ⋅∥P(z)∥≤δcA⋅M(z).
Given ε>0, choosing δ=ε/cA we see that (13) is satisfied. Moreover,
and so Proposition 4.1 applies and shows that (A⊙B,M) is a CQMS. It is clear that the argument is symmetric in the choice of matrix factor. □
Cancellation in one factor
We now return to the setting of groups and length functions to provide a criterion ensuring that the assumptions in Theorem 4.2 are satisfied. Suppose therefore G and K are countable discrete groups equipped with proper length functions ℓG and ℓK, and consider the length function ℓG×K:=ℓG+ℓK on G×K. Suppressing the natural isomorphism ℓ2(G×K)≅ℓ2(G)⊗ℓ2(K), we then have
DℓG×K=DℓG⊗1+1⊗DℓK,
as unbounded operators on, say, cc(G)⊙cc(K). For notational convenience, we set
which are all extendable to bounded operators. We make no notational difference between these densely defined operators and their bounded extensions. The strategy in the sequel will be to average certain sequences of unitary conjugates of T(x) which fix T1(x) and make T2(x) converge strongly to zero. This is formalised by means of the following definition.
Definition 4.3. A finite cancellation datum for (K,ℓK) consists of a finite nonempty set I, numbers wi≥0 with ∑iwi=1, and sequences (gi(n))n⊆K indexed by i∈I, such that, for every fixed m,s∈K,
i∈I∑wi(ℓK(msgi(n))−ℓK(sgi(n)))⟶n→∞0.(17)
It is eventually exact if this expression is zero for all sufficiently large n, with the threshold allowed to depend on m and s.
Lemma 4.4. If K admits a finite cancellation datum then for every p≥1 and every x∈Mp(C[G×K]), it holds that
∥T1(x)∥≤∥T(x)∥and∥T2(x)∥≤2∥T(x)∥.
For the proof, we will need the right translations as well. For t∈K, we therefore let ρtδs=δst be the right translation, which is a unitary operator. It commutes with every left translation λm, and just like for left translation, every ρt preserves the domain of DℓK.
Proof. Write x=∑m∈Fxm⊗λm with xm∈Mp(C[G]) for some finite set F⊆K. To simplify notation, we will denote 1p⊗1ℓ2(G)⊗ρt by Vt in the following. Since Vt operator commutes with x and D1, it will fix T1(x) when conjugating, that is Vt∗T1(x)Vt=T1(x). For η∈Cp⊗ℓ2(G) and s∈K, we have
Such vectors are dense, and ∥Ψn(T2(x))∥≤∥T2(x)∥, so we have strong convergence everywhere. Consequently, Ψn(T2(x))→nT1(x) strongly. Contractivity then gives ∥T1(x)ξ∥≤∥T(x)∥∥ξ∥ for every ξ, which proves the first inequality in the lemma statement. The second then follows by writing T2(x)=T(x)−T1(x). □
Corollary 4.5. If (C[G],ℓ2(G),DℓG) is a spectral metric space and if (C[K],ℓ2(K),DℓK) is a matrix spectral metric space and if K admits a finite cancellation datum, then (C[G×K],ℓ2(G×K),DℓG×K) is a spectral metric space.
Proof. By Lemma 4.4, the assumptions in Theorem 4.2 are satisfied and the result follows. □
To illustrate cancellation, we include the following easy lemma.
Lemma 4.6. For any length function on a finite group K, one obtains an exact cancellation datum by setting I=K, wk=∣K∣1 and gk(n)=k for all n∈N.
Proof. For fixed m,s the maps t↦mst and t↦st both permute K, and so the sums of their lengths are equal. □
For abelian and infinite dihedral groups, we construct geodesic rays along which the length defects eventually stabilize. Averaging over the two directions and a finite set of residues then gives cancellation. The construction is related to Rieffel’s horofunction methods [17]. For an equally weighted family it suffices to prove, for each fixed h,
We now proceed to show that certain natural groups admit a finite cancellation datum.
Finitely generated abelian groups
We first show that finitely generated abelian groups have a finite cancelation datum.
Proposition 4.7.Every finite symmetric word length on a finitely generated abelian group K has an eventually exact cancellation datum. If K is infinite, the sequences can be chosen as samples of geodesic rays.
Proof. Since Lemma 4.6 covers the case for finite K, we assume K is infinite in the sequel. We will write K additively and denote its rank by d and its torsion subgroup by Tor(K). Suppose S is the given generating set, and consider the quotient map
π:K→K/Tor(K)≅Zd⊆Rd
for some d≥1. The set P:=conv(π(S)) is then a polytope of dimension d, which is centrally symmetric and it contains 0 in its interior. Recall that a facet of P is a face of dimension d−1. Choose a facet F together with a supporting linear functional σ:Rd→R; that is σ(p)≤1 for x∈P and F=P∩{ξ∈Rd∣σ(ξ)=1}; see e.g. [Zie95] for background on polytopes. Since 0 is in the interior of P, the maximum value for σ is positive. To ease notation we will also write σ(w)=σ(π(w)) for w∈K. Now define
V:={v∈S∣π(v)∈F},q:=∣V∣,z:=v∈V∑vandΛ:=⟨V⟩≤K.
Note that if w=s1+⋯+sk with si∈S, then σ(w)=∑iσ(si)≤k. Taking the minimum over all representing words, we obtain ℓK(w)≥σ(w). Since z is defined as the sum of q letters from V, nz can be represented by nq such letters. Since σ(nz)=nq, we deduce ℓK(nz)=nq for all n≥0. As S=−S, we deduce ℓK(−w)=ℓK(w) for every w∈K. If n<0, this means ℓK(nz)=ℓK((−n)z)=(−n)q, and so
ℓK(nz)=∣n∣qfor all n∈Z.(19)
Let γ:N0→K be the path obtained by successively adding the letters of this repeated listing of elements of V, starting at 0; i.e. writing V={s1,…,sq} the path is defined as:
If r<t, the subword from γ(r) to γ(t) has t−r letters, each of which takes the value 1 under σ. Hence
t−r=σ(γ(t)−γ(r))≤ℓK(γ(t)−γ(r))≤t−r.
It follows that γ is geodesic, and γ(nq)=nz.
We proceed to show that K/Λ is finite. Note that F=conv(π(V)) and that it has dimension d−1, and hence π(V) contains d affinely independent points u1,…,ud. By affine independence, a relation ∑j=1dtjuj=0 with real coefficients satisfying ∑j=1dtj=0 forces every tj to vanish. In fact, these points are also linearly independent. We may see this by applying σ to any relation ∑j=1dtjuj=0 and noting that since σ(uj)=1, this gives ∑j=1dtj=0. Affine independence then forces t1=⋯=td=0. Consequently, π(Λ) contains d linearly independent integer vectors, w1,…,wd. The matrix A whose columns consist of w1,…,wd is therefore invertible and its inverse is given by A−1=(det(A))−1Adj(A), where the adjugate matrix Adj(A) has integer entries. We therefore have that
det(A)⋅Zd=A⋅Adj(A)⋅Zd⊆AZd⊆π(Λ),
and it follows that π(Λ) has finite index in Zd. The surjection K/Λ→Zd/π(Λ) has kernel (Λ+Tor(K))/Λ which, by the second isomorphism theorem, is isomorphic to Tor(K)/Λ∩Tor(K), and thus finite. Finiteness of K/Λ now follows from noting that both kernel and range are finite.
For fixed w∈K, consider the integer-valued sequence
bn(w)=ℓK(nz+w)−nq.
Then ℓK((n+1)z+w)≤ℓK(nz+w)+q implies bn+1(w)≤bn(w), so the sequence is non-increasing. Using (19) we also have nq=ℓK(nz)≤ℓK(nz+w)+ℓK(w), and so we deduce −ℓK(w)≤bn(w)≤ℓK(w), forcing the sequence to eventually stabilize at a value, which we denote by b(w). We will need to understand how b changes under translations with elements from Λ. For v∈V, we have ℓK(v)=1. Moreover, as z−v=∑u∈V∖{v}u is a sum of q−1 generators, we have ℓK(z−v)≤q−1. Writing (n+1)z+w=(nz+w+v)+(z−v), we use the triangle inequality to calculate
ℓK(nz+w+v)≤ℓK(nz+w)+1,
ℓK((n+1)z+w)≤ℓK(nz+w+v)+q−1.
Thus
bn+1(w)+1≤bn(w+v)≤bn(w)+1.
As bn eventually stabilizes to b, we see that for sufficiently large n, both the outer terms equal b(w)+1, and so b(w+v)=b(w)+1. We also note that this yields b(w)−1=b(w−v+v)−1=b(w−v). Every a∈Λ can be written as a=∑v∈Vkvv for kv∈Z. By iterating these identities we get
It follows that b(w+a)−σ(w+a)=b(w)−σ(w), and so f([w]):=b(w)−σ(w) is a well-defined function on K/Λ. Thus, for all sufficiently large n (the threshold depending on w) we have
ℓK(nz+w)=nq+σ(w)+f([w]).(20)
With the analysis above at our disposal, we now obtain the claimed cancellation datum as follows: Choose as set of representatives, R, for K/Λ, put M=∣R∣, and define 2M sequences as follows:
gr+(n):=γ(nq)+r=nz+r,for r∈R,
gr−(n):=−γ(nq)−r=−nz−r,for r∈R.
Fix h∈K, and choose n0∈N such that (4.9) holds for all n≥n0 and all w in the finite set {r,r±h∣r∈R}. Then
For the second line, symmetry gives ℓK(h−nz−r)=ℓK(nz+r−h) and ℓK(−nz−r)=ℓK(nz+r). In both lines the f-terms sum to zero because [r]↦[r±h] permutes the finite quotient K/Λ. The two sums therefore cancel, giving An(h)=0, where An is given as in (18). The weighted sum in (17) is then
2MAn(m+s)−An(s)=0.
Thus, the weights 1/(2M) give the required eventually exact datum. □
The infinite dihedral group
Next we show that also the infinite dihedral group D∞ admits a finite cancellation datum.
Proposition 4.8. Let K be the infinite dihedral group
K=⟨a,b∣b2=e,bab=a−1⟩,
and let ℓK be the word length for any finite symmetric generating set S. Then there is an eventually exact cancellation datum whose sequences are samples of geodesic rays.
Proof. Write elements of K as ajbε for j∈Z and ε∈{0,1}, and note that
(ajbδ)(akbε)=aj+(−1)δkbδ+ε,
where the exponent of b is understood to be modulo 2. Now define
T={t∈Z:at∈S}andP=max({∣t∣:t∈T}∪{0}),
Q={r∈Z:arb∈S}andΔ=maxQ−minQ.
Note that T=−T. Also, Q is nonempty since K is non-abelian, and ν:=max{P,Δ/2}>0 because otherwise S would generate a group of order at most 2. Now choose positive integers q,L by
(q,L)={(P,1),(Δ,2),if P≥Δ/2,if P<Δ/2.
Then q=νL. Note that if (q,L)=(P,1), then aq=aP∈S as T=−T. If (q,L)=(Δ,2), then (amaxQb)(aminQb)=aΔ=aq and (amaxQb)(aminQb) has length 2. Thus, in either case, there is a word of length L representing aq. We first prove
ℓK(anq)=∣n∣L,for all n∈Z.
To this end, put c:=(maxQ+minQ)/2 and define X:K→R by X(akbε):=k−cε. We now wish to analyse how much the value of X changes upon multiplying the input by a generator. We first note that
X(akbεat)−X(akbε)=(−1)εt,
X(akbεarb)−X(akbε)=(−1)ε(r−c).
If at∈S then ∣t∣≤P≤ν and if arb∈S then ∣r−c∣≤Δ/2≤ν. So, upon multiplying a given element akbε∈K by a generator from the right, the absolute change in the value of X is at most ν. Also, X(e)=0, so telescoping along a word of length k representing h gives ∣X(h)∣≤kν. Thus ℓK(h)≥∣X(h)∣/ν. If h=anq for some n∈Z, this reads ℓK(anq)≥∣n∣q/ν=∣n∣L, which proves the lower bound. But, as we saw above aq can be represented by a word of length L so anq can be expressed as a product of ∣n∣L generators, which provides the upper bound.
For σ∈{±1}, 0≤r<q and ε∈{0,1}, set
gσ,r,ε(n):=aσnq+rbε.
We note that
gσ,r,ε(n)−1gσ,r,ε(k)=a(−1)εσ(k−n)q.
The length of this element is ∣n−k∣L by what we noted above. In other words, the distance between gσ,r,ε(n) and gσ,r,ε(k) in the induced left-invariant word metric is ∣n−k∣L. Repeating a shortest word for a(−1)εσq from the initial point arbε interpolates these samples to a geodesic ray. Indeed, the interpolating path has length ∣k−n∣L between any two samples. A shorter path between two intermediate vertices, together with the remaining portions of the interpolating path, would contradict the distance formula for two enclosing samples. For fixed (σ,r,ε), the integers
fn=ℓK(aσnq+rbε)−nL
are non-increasing, since left multiplication by aσq increases the length by at most L. Using subadditivity and the reverse triangle inequality we obtain
nL−ℓK(arbε)≤ℓK(aσnq+rbε)≤nL+ℓK(arbε).
It follows that fn is bounded and integer-valued, and must therefore stabilize. We may thus find an N0 and constants Cσ,r,ε∈Z such that
ℓK(gσ,r,ε(n))=nL+Cσ,r,εfor all n≥N0,(21)
for all 4q indices (σ,r,ε).
With this, we may now finally verify cancellation. Fix h=ajbδ. For every r∈{0,…,q−1}, there exist unique tr∈Z and r′∈{0,…,q−1} such that
j+(−1)δr=qtr+r′.
We now define σ′:=(−1)δσ and ε′:=δ+ε(mod2). Then
hgσ,r,ε(n)=gσ′,r′,ε′(n+σ′tr).
Note that tr is independent of σ and ϵ. For n≥N0+maxr∣tr∣, both n and n+σ′tr are at least N0, and so (4.10) yields
The map of indices (σ,r,ϵ) to (σ′,r′,ϵ′) is a permutation: from its image one recovers σ=(−1)δσ′, ϵ=ϵ′−δ modulo two, and the unique r∈{0,…,q−1} congruent to (−1)δ(r′−j) modulo q. Consequently the sums of the two families of constants agree. For fixed r,ϵ, summing over σ∈{1,−1} cancels the carry terms because tr is fixed and σ′ takes both signs. Thus
σ,r,ϵ∑(ℓK(hgσ,r,ϵ(n))−ℓK(gσ,r,ϵ(n)))=0(22)
for every sufficiently large n. This is An(h)=0, so the weights 1/(4q) give (4.6).
Products with hyperbolic groups
In this section we show how the analysis carried out above leads to new examples of length functions on groups which give rise to compact quantum metric spaces. We first summarize the upshot of the previous section as follows:
Proposition 4.9. Let G be any group with a proper length function ℓG. Let K be either
a finite group with any length function, or
a finitely generated abelian group with any finite symmetric word length, or
the infinite dihedral group with any finite symmetric word length.
Then the bounds of Lemma 4.4 hold at every matrix level, with constants 1 and 2 independent of the group, generating set, and matrix size.
Proof. This is immediate by Lemma 4.4, Lemma 4.6, Proposition 4.7 and Proposition 4.8. □
Theorem 4.10. Let G be a finitely generated group equipped with a word-length function ℓG for which (C[G],LℓG) is a CQMS, and let K1,…,Km be a family or groups, each of which is either finite, finitely generated abelian or isomorphic to D∞. Equipping each Kj with a word length function ℓKj and
Proof. Each Kj is amenable and yields a CQMS by [5]. Theorem 3.1 therefore makes every Kj an MCQMS. For m=1, the result therefore follows by Corollary 4.5. The result now follows by induction on m. □
Corollary 4.11. If H is word-hyperbolic and K1,…,Km are as in Theorem 4.10, then H×K1×⋯×Km, with length function coming from the union of the chosen generating sets for the factors, yields a CQMS.
Remark 4.12. Corollary 4.11 covers groups of the form G=H×F×D∞k×Zd for H word hyperbolic and F finite, and with the length function as described in the result. In particular it covers the case d=0, F a finite Coxeter group, and H a non-elementary hyperbolic Coxeter group, and k≥1. As such, we have the first examples of word length functions on non-amenable, non-hyperbolic Coxeter groups with CQMS structures for these word-lengths. Indeed, these groups contain H and a copy of Z2, and are therefore neither amenable nor word-hyperbolic. They are thus in particular not covered by the results of [5] or [14].
References
Are Austad, Department of Mathematics, University of Oslo, P.O.Box 1053 Blindern, 0316 Oslo, Norway
[1][AKK22] Konrad Aguilar, Jens Kaad, and David Kyed. The Podleś spheres converge to the sphere. Comm. Math. Phys., 392(3):1029–1061, 2022.DOI
[2][Aus26] Are Austad. Compact quantum metric spaces from weakly geodesic length functions on hyperbolic groups, 2026.arxiv.org/abs/2608.28303
[3][Con89] Alain Connes. Compact metric spaces, Fredholm modules, and hyperfiniteness. Ergodic Theory Dynam. Syst., 9(2):207–220, 1989.
[4][Con94] Alain Connes. Noncommutative Geometry. Academic Press, San Diego, 1994.
[5]Michael Christ and Marc A. Rieffel. Nilpotent group C*-algebras as compact quantum metric spaces. Canad. Math. Bull., 60(1):77–94, 2017.
[6]Alain Connes and Walter D. van Suijlekom. Spectral truncations in noncommutative geometry and operator systems. Commun. Math. Phys., 383(3):2021–2067, 2021.DOI
[7]Uffe Haagerup. An example of a nonnuclear C*-algebra, which has the metric approximation property. Invent. Math., 50(3):279–293, 1978/79.
[8]Jens Kaad. External products of spectral metric spaces. Pure Appl. Funct. Anal., 9(5):1175–1211, 2024.arxiv.org/abs/2304.03979
[9]Jens Kaad and David Kyed. Dynamics of compact quantum metric spaces. Ergodic Theory and Dynamical Systems, 41(7):2069–2109, 2021.DOI
[10]Mario Klisse. Word-length spectral triples of (Z/2Z) ≀ F_d are not metric, 2026.arxiv.org/abs/2608.12080
[11]David Kyed and Ryszard Nest. Finiteness of metrics on state spaces. Bull. Lond. Math. Soc., 56(1):288–295, 2024.DOI
[12]Ali Kavruk, Vern I. Paulsen, Ivan G. Todorov, and Mark Tomforde. Tensor products of operator systems. J. Funct. Anal., 261(2):267–299, 2011.arxiv.org/abs/0910.2210
[13]Malte Leimbach. Convergence of Peter–Weyl truncations of compact quantum groups. Journal of Noncommutative Geometry, 2025. Published online first.arxiv.org/abs/2409.16698
[14]Narutaka Ozawa and Marc A. Rieffel. Hyperbolic group C*-algebras and free-product C*-algebras as compact quantum metric spaces. Canad. J. Math., 57(5):1056–1079, 2005.
[15]Vern Paulsen. Completely bounded maps and operator algebras, volume 78 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2002.DOI
[16]Gilles Pisier. Introduction to Operator Space Theory, volume 294 of London Mathematical Society Lecture Note Series. Cambridge University Press, 2003.DOI
[17]Marc A Rieffel. Group C*-algebras as compact quantum metric spaces. Documenta Mathematica, 7:605–651, 2002.DOI
[18]Marc A. Rieffel. Standard deviation is a strongly Leibniz seminorm. New York J. Math., 20:35–56, 2014.
[19]Walter D. van Suijlekom. Gromov-Hausdorff convergence of state spaces for spectral truncations. J. Geom. Phys., 162:11, 2021. Id/No 104075.DOI