betaFeedback

WorksMathematics

Property A and the identification of the maximal and reduced uniform Roe algebras

Abstract

Let XX be a discrete metric space of bounded geometry. We prove that XX has property A if and only if the canonical map λ:Cu,max∗(X)→Cu∗(X)λ:C^*_{u,max}(X)\to C^*_u(X) is an isomorphism.

Introduction

Roe algebras arose in J. Roe’s study of higher indices of elliptic operators on open manifolds [11]. The reduced uniform Roe algebra appears in [12 Section 4.4], and its maximal counterpart follows the construction of the maximal Roe algebra by Gong, Wang and Yu [8 Sections 3–4].

For a discrete metric space XX of bounded geometry, the identity on the algebraic uniform Roe algebra extends to the canonical quotient map

λ:Cu,max⁡∗(X)⟶Cu∗(X).\lambda: C^*_{u,\max}(X) \longrightarrow C^*_u(X).

Whether λ\lambda is an isomorphism is related to the large-scale geometry of XX. Yu introduced property A as a coarse analogue of amenability [17]. Špakula and Willett proved that property A implies that λ\lambda is an isomorphism and asked whether the converse holds [15 Proposition 1.3 and Section 1]. They also showed that λ\lambda induces an isomorphism on KK-theory when XX coarsely embeds into Hilbert space [15 Theorem 1.4]. The box space of F2\mathbb{F}_2 constructed by Arzhantseva, Guentner and Špakula admits such an embedding and fails property A [3]. For that space, Špakula and Willett explicitly asked whether λ\lambda itself is an isomorphism [15 Example 1.16].

Question 1.1. Let XX be a discrete metric space with bounded geometry. If the canonical quotient map λ\lambda is an isomorphism, must XX have Yu’s property A?

Skandalis, Tu and Yu associated the coarse groupoid G(X)G(X) to XX and recast the coarse assembly map as a groupoid Baum–Connes assembly map with coefficients [14 Proposition 4.8]. The maximal and reduced uniform Roe algebras identify with the corresponding groupoid C∗C^*-algebras [15 Remark 1.5]

Cu,max⁡∗(X)≅Cmax⁡∗(G(X)),Cu∗(X)≅Cr∗(G(X)).C^*_{u,\max}(X) \cong C^*_{\max}(G(X)), \qquad C^*_u(X) \cong C^*_r(G(X)).

Under these identifications, λ\lambda is the canonical groupoid quotient. Skandalis, Tu and Yu also showed that XX has property A if and only if G(X)G(X) is topologically amenable [14 Theorem 5.3]. The question thus asks whether weak containment forces topological amenability for coarse groupoids.

Willett’s 2015 paper grew out of an attempt to answer this question for the box space of SL⁡(2,Z)\operatorname{SL}(2,\mathbb{Z}) in Špakula and Willett’s Example 1.15 [16 Remark 3.5]. He constructed a Higson–Lafforgue–Skandalis groupoid with finite quotients of F2F_2 as its finite fibers and F2F_2 as the fiber at infinity. The non-amenability of the fiber at infinity forces the groupoid to fail topological amenability [16 Lemma 2.4]. Using Lubotzky and Shalom’s property FD, Willett showed that the regular representations of the chosen finite quotients recover the full norm on C[F2]\mathbb{C}[F_2]. This yields equality of the maximal and reduced groupoid C∗C^*-algebras [16 Lemmas 2.7 and 2.8]. In Remark 3.5 of the same paper, he asked whether weak containment can also hold for a non-amenable coarse groupoid.

Alekseev and Finn–Sell subsequently constructed non-amenable principal étale groupoids with weak containment and second-countable unit spaces [1 Section 2]. The coarse groupoid of an infinite XX has the non-second-countable Stone–Čech compactification βX\beta X as its unit space [14 Section 3]. Ara, Li, Lledó and Wu recorded the coarse groupoid question as open [2 Example 4.6]. For coarse disjoint unions of large-girth expanders, Buss, Echterhoff and Willett distinguished the two uniform Roe completions by comparing the spectra of the graph Laplacian [5 proof of Lemma 4.7]. We answer Question 1.1 affirmatively. Together with Špakula and Willett’s result [15 Proposition 1.3], this gives the following main theorem.

Theorem 1.2. Let XX be a discrete metric space with bounded geometry. Then, XX has Yu’s property A if and only if the canonical quotient map

λ:Cu,max⁡∗(X)→Cu∗(X)\lambda: C^*_{u,\max}(X) \to C^*_u(X)

is an isomorphism.

The proof was inspired by the work of Roe–Willett [13], in particular about the construction of the sequence of graphs using the failure off property A.

When XX fails property A, we use Elek’s result on uniform local amenability [6] to construct locally expanding graphs. We then construct large-girth graphs on subsets of their vertex sets, sign the edges using the theorem of Marcus, Spielman and Srivastava [10], and form the signed Laplacian LσL_\sigma. This gives an operator T∈Cu,max⁡∗(X)T \in C^*_{u,\max}(X) with

∥T∥max⁡=1,∥λ(T)∥<1.\lVert T\rVert_{\max} = 1,\qquad\lVert\lambda(T)\rVert< 1.

Together with Špakula and Willett’s forward implication [15 Proposition 1.3], this yields our main result.

This paper is organized as follows. In Section 2, we briefly recall the two uniform Roe completions, Yu’s property A, and the decomposition of finite-propagation operators into weighted partial translations. In § 3, we introduce the graph terminology and constructs large-girth graphs from local expansion. § 4 constructs a state on the maximal completion from signed paths. Section 5 establishes the norm estimates and proves Theorems 1.2.

AI disclosure. We thank ChatGPT-6 for assistance in understanding the result of Marcus, Spielman and Srivastava [10] and for language editing during the preparation of this manuscript. All main results and mathematical proofs were developed by the authors.

Uniform Roe algebras

In this section, we first briefly recall the definitions of maximal and reduced uniform Roe algebras and the concept of Yu’s property A.

Uniform Roe algebras

Let XX be a discrete metric space with bounded geometry. We write BX(x,R)=BR(x)={z∈X:d(x,z)≤R}B_X(x,R)=B_R(x)=\{z\in X:d(x,z)\le R\} and set

NX(R)=sup⁡x∈X∣BX(x,R)∣<∞,R>0.N_X(R)=\sup_{x\in X}|B_X(x,R)|<\infty,\qquad R>0.

Let HX=ℓ2(X)H_X=\ell^2(X). A bounded linear operator T∈B(ℓ2(X))T\in B(\ell^2(X)) can be expressed as an (X×X)(X\times X)-matrix

T=(Tx,y)x,y∈X,T=(T_{x,y})_{x,y\in X},

where Tx,y∈CT_{x,y}\in\mathbb{C} for any x,y∈Xx,y\in X. The propagation of TT is defined as

prop⁡(T)=sup⁡{d(x,y):Tx,y≠0},prop⁡(0)=0.\operatorname{prop}(T)=\sup\{d(x,y):T_{x,y}\ne0\},\qquad\operatorname{prop}(0)=0.

We say that the operator TT has finite propagation if

prop⁡(T)<∞.\operatorname{prop}(T)<\infty.

Now let us recall the definitions of the algebraic uniform Roe algebras.

Definition 2.1. Let XX be a discrete metric space with bounded geometry. The algebraic uniform Roe algebra, denoted by Cu[X]C_u[X], is the ∗*-algebra of all bounded operators T∈B(ℓ2(X))T\in B(\ell^2(X)) with finite propagation.

For any two operators T=(Tx,y)x,y∈XT=(T_{x,y})_{x,y\in X}, S=(Sx,y)x,y∈X∈Cu[X]S=(S_{x,y})_{x,y\in X}\in C_u[X], the product TSTS is defined as

(TS)x,z=∑y∈XTx,ySy,z,(TS)_{x,z}=\sum_{y\in X}T_{x,y}S_{y,z},

for any x,z∈Xx,z\in X.

For every r>0r>0, there exists a constant Cr>0C_r>0 such that

∥π(T)∥≤Cr⋅sup⁡x,y∈X∣Tx,y∣\|\pi(T)\|\le C_r\cdot\sup_{x,y\in X}|T_{x,y}|

for any ∗*-representation π:Cu[X]→B(H)\pi:C_u[X]\to B(H) and for any T∈Cu[X]T\in C_u[X] with $\operatorname{prop}(T)\le r [15, Lemma 1.10]. As a result, the following maximal norm is finite:

∥T∥max=sup⁡{∥π(T)∥:π:Cu[X]→B(H) is a *-representation}.\lVert T\rVert_{\mathrm{max}}=\sup\{\lVert\pi(T)\rVert:\pi:C_u[X]\to B(H)\text{ is a *-representation}\}.

The maximal and reduced uniform Roe algebras can be viewed as coarse analogues of the maximal and reduced group C∗C^*-algebras.

Definition 2.2 ([15 Definitions 1.9 and 1.11]). Let XX be a discrete metric space with bounded geometry.

  1. The uniform Roe algebra Cu∗(X)C_u^*(X) is defined to be the completion of Cu[X]C_u[X] under the norm on B(ℓ2(X))B(\ell^2(X));

  2. The maximal uniform Roe algebra Cu,max∗(X)C_{u,\mathrm{max}}^*(X) is defined to be the completion of Cu[X]C_u[X] under the maximal norm.

For the index-theoretic origins of Roe algebras, see [11]. In Roe’s terminology, Cu∗(X)C_u^*(X) is the uniform translation C∗C^*-algebra; see [12 Section 4.4].

By the universality of the maximal norm, the identity map on Cu[X]C_u[X] continuously extends to the quotient map

λ:Cu,max∗(X)→Cu∗(X).\lambda:C_{u,\mathrm{max}}^*(X)\to C_u^*(X).

We use ∥⋅∥red\lVert\cdot\rVert_{\mathrm{red}} for the concrete operator norm.

We use the following equivalent formulation of Yu’s property A [17].

Definition 2.3 ([15 Definition 2.1]). Let XX be a discrete metric space with bounded geometry, and let ℓ2(X)+\ell^2(X)_+ denote the nonnegative real-valued functions in ℓ2(X)\ell^2(X). We say that XX has Yu’s property A if for every R>0R>0 and ϵ>0\epsilon>0, there exists a map ξ:X→ℓ2(X)+\xi:X\to\ell^2(X)_+ satisfying

  1. for each x∈Xx\in X, ∥ξx∥2=1\lVert\xi_x\rVert_2=1;

  2. there exists an S>0S>0, such that Supp⁡(ξx)⊆BS(x)\operatorname{Supp}(\xi_x)\subseteq B_S(x), for all x∈Xx\in X;

  3. for all x,y∈Xx,y\in X with d(x,y)≤Rd(x,y)\le R, ∥ξx−ξy∥2<ϵ\lVert\xi_x-\xi_y\rVert_2<\epsilon.

The following result of Špakula and Willett gives the forward implication of Theorem 1.2.

Theorem 2.4 ([15 Proposition 1.3]). Let XX be a discrete metric space with bounded geometry. If XX has Yu’s property A, then the quotient map

λ:Cu,max∗(X)→Cu∗(X)\lambda:C_{u,\mathrm{max}}^*(X)\to C_u^*(X)

is an isomorphism.

Partial translations

A partial translation is a bijection t:A→Bt:A\to B between subsets of XX with sup⁡x∈AdX(x,t(x))<∞\sup_{x\in A}d_X(x,t(x))<\infty. Its operator is

Vtδx={δt(x),x∈A,0,x∉A.V_t\delta_x= \begin{cases} \delta_{t(x)}, & x\in A,\\ 0, & x\notin A. \end{cases}

For f∈ℓ∞(X)f \in\ell^\infty(X), write MfM_f for diagonal multiplication by ff.

Every finite-propagation operator is a finite sum of operators MfVtM_fV_t. Indeed, if the propagation is at most RR, each row and column contains at most NX(R)N_X(R) nonzero entries. Enumerate the nonzero entries and color them greedily with 2NX(R)−12N_X(R)-1 colors, forbidding repetitions of a color in the same row or column. At most 2(NX(R)−1)2(N_X(R)-1) colors are forbidden at each step. Within each color, the nonzero entries define a partial translation from column indices to row indices, and their values define the coefficient ff on its range.

Local expansion and large-girth graphs

Let XX be a discrete metric space without property A. We first construct a sequence of finite graphs with uniform local expansion. Based on this sequence, we construct a sequence of graphs with degrees five or six, uniformly bounded ambient edge lengths, and girth tending to infinity.

Preliminaries on graphs

In this subsection, we recall some background on graph theory.

A simple graph is a pair G=(V(G),E(G))G=(V(G),E(G)), where V(G)V(G) is its vertex set and E(G)E(G) is a set of unordered pairs of distinct vertices. We write (x,y)(x,y) for the edge with endpoints xx and yy. The expressions (x,y)(x,y) and (y,x)(y,x) denote the same undirected edge. The endpoints of an edge are adjacent, or neighbors. The neighbor set and degree of xx are

NG(x)={y:(x,y)∈E(G)},dG(x)=∣NG(x)∣.N_G(x)=\{y:(x,y)\in E(G)\},\qquad d_G(x)=|N_G(x)|.

An edge is incident to each of its endpoints. A graph has bounded degree if there is a finite constant δ\delta such that dG(x)≤δd_G(x)\leq\delta for every vertex x∈V(G)x\in V(G).

A subgraph FF of GG is a graph whose vertex set V(F)V(F) and edge set E(F)E(F) are subsets of V(G)V(G) and E(G)E(G), respectively, and every edge of E(F)E(F) has endpoints in V(F)V(F). For W⊆V(G)W\subseteq V(G), the induced subgraph G[W]G[W] has vertex set WW and all edges of GG with both endpoints in WW. A simple graph on WW has vertex set WW, with an edge set specified separately.

A matching is a set of edges no two of which share an endpoint. A proper edge coloring assigns colors to edges so that edges sharing an endpoint have different colors. Each color class is therefore a matching.

A walk of length kk is a finite sequence

x0,e1,x1,e2,…,ek,xk,x_0,e_1,x_1,e_2,\ldots,e_k,x_k,

where eje_j has endpoints xj−1x_{j-1} and xjx_j. A walk is closed if x0=xkx_0=x_k. A trail is a walk with no repeated edge, and a path is a walk with pairwise distinct vertices. A graph is connected if any two vertices can be joined by a path. Its components are the maximal connected subgraphs. A cycle is a closed trail of positive length in which x0,…,xk−1x_0,\ldots,x_{k-1} are pairwise distinct. Cycles are identified under change of starting vertex and reversal. Every closed trail of positive length contains a cycle, by taking a shortest nonempty closed subtrail.

The girth girth⁡(G)\operatorname{girth}(G) is the least length of a cycle, with value ∞\infty if there is no cycle. A forest is a graph with no cycles. A tree is a connected forest. In a tree, there is a unique path between any two vertices, since distinct paths with the same endpoints would contain a cycle in their union.

For vertices x,yx,y in the same component, the graph distance dGgr(x,y)d^{\mathrm{gr}}_G(x,y) is defined to be the minimum length of a path joining them. For vertices in different components, we set dGgr(x,y)=∞d^{\mathrm{gr}}_G(x,y)=\infty.

If V(G)⊆XV(G)\subseteq X, the ambient length of the edge (x,y)(x,y) is dX(x,y)d_X(x,y). A bound RR on all ambient edge lengths implies

dX(x,y)≤RdGgr(x,y)d_X(x,y) \le R d^{\mathrm{gr}}_G(x,y)

when x,yx,y are in the same component.

For nonempty E⊆XE\subseteq X, we write diam⁡X(E)=sup⁡{dX(x,y):x,y∈E}\operatorname{diam}_X(E)=\sup\{d_X(x,y):x,y\in E\}.

Definition 3.1. Let (X,dX)(X,d_X) be a discrete metric space with bounded geometry and R>0R>0. The scale-RR graph ΓR\Gamma_R has vertex set XX and an edge (x,y)(x,y) for each pair of distinct vertices with dX(x,y)≤Rd_X(x,y)\le R.

Definition 3.2. Let GG be a graph and E⊆V(G)E\subseteq V(G). The internal vertex boundary and edge boundary are defined as

∂GintE={x∈E: some y∈V(G)∖E satisfies (x,y)∈E(G)},\partial^{\mathrm{int}}_G E=\{x\in E:\text{ some }y\in V(G)\setminus E\text{ satisfies }(x,y)\in E(G)\},
∂GE={e∈E(G):e has exactly one endpoint in E}.\partial_G E=\{e\in E(G):e\text{ has exactly one endpoint in }E\}.

Every internal boundary vertex is incident to a crossing edge, so ∣∂GintE∣≤∣∂GE∣|\partial^{\mathrm{int}}_G E|\le|\partial_G E|. If FF is a subgraph, ∂FintE\partial^{\mathrm{int}}_F E counts only neighbors in V(F)∖EV(F)\setminus E and edges belonging to FF.

Local expansion

Brodzki, Niblo, Špakula, Willett and Wright introduced uniform local amenability (ULA) and proved its coarse invariance [4 Definition 2.2 and Lemma 2.6]. We use Elek’s formulation [6, 7] for graphs with bounded degree.

Definition 3.3 ([6 Definition 1.2]). Let GG be an infinite graph of bounded degree. The graph GG is said to be uniformly locally amenable (ULA) if, for every h>0h>0, there is K<∞K<\infty such that each nonempty finite subgraph FF contains a nonempty E⊆V(F)E\subseteq V(F) with

∣E∣≤K,∣∂FintE∣≤h∣E∣.|E|\le K,\qquad|\partial^{\mathrm{int}}_F E|\le h|E|.

For metric spaces with bounded geometry, Brodzki, Niblo, Špakula, Willett and Wright showed that Yu’s property A implies ULA [4 Proposition 3.2 and Lemma 2.7], while Elek proved the converse [6 Theorem 1]. For the equivalence in the graph-metric formulation, see also [7 Theorem 1].

The following construction is inspired by the arguments in Roe–Willett [13]. Failure of ULA gives rise to a sequence of finite graphs with a uniform lower bound on the ratio of internal boundary size to cardinality for small vertex sets.

Lemma 3.4. Suppose XX has bounded geometry without Yu’s property A. Then, there are constants R>0R > 0, h∈(0,1)h \in(0,1) and an integer δ≥2\delta\ge2, and a sequence (Gn)n(G_n)_n of finite connected simple graphs with nonempty, pairwise disjoint vertex sets (Wn)n(W_n)_n in XX, such that

  1. each GnG_n has degrees at most δ\delta and ambient edge lengths at most RR;

  1. for every nn and every nonempty subset E⊆WnE \subseteq W_n with diam⁡X(E)≤2nR\operatorname{diam}_X(E) \le2nR, we have that

∣∂GnintE∣>h∣E∣.(1)\left|\partial^{\mathrm{int}}_{G_n} E\right| > h|E|. \tag*{(1)}

Proof. It is obvious that the space XX is infinite. Suppose that every scale graph ΓR\Gamma_R has ULA. Fix R,ε>0R,\varepsilon> 0. Since ΓR\Gamma_R has bounded degree, ULA gives nonnegative ℓ1\ell^1 unit vectors axa_x supported in graph balls of a common radius SS, with ∥ax−ay∥1<ε2\|a_x-a_y\|_1 < \varepsilon^2 on edges. Each such ball about xx is contained in BX(x,RS)B_X(x,RS), and distinct points at ambient distance at most RR are adjacent in ΓR\Gamma_R. Thus ξx=ax\xi_x=\sqrt{a_x} gives the required ℓ2\ell^2 unit vectors at scale RR, since ∥ξx−ξy∥22≤∥ax−ay∥1\|\xi_x-\xi_y\|_2^2 \le\|a_x-a_y\|_1. As RR and ε\varepsilon are arbitrary, XX has property A, a contradiction. Therefore, there must be an R>0R > 0, such that the scaled graph ΓR\Gamma_R does not have ULA.

Fix R>0R > 0 for which ΓR\Gamma_R does not have ULA. There is h0∈(0,1)h_0 \in(0,1) such that for every positive integer KK there is a nonempty finite subgraph FF of ΓR\Gamma_R satisfying

∣∂FintA∣>h0∣A∣(A⊆V(F), 1≤∣A∣≤K).(2)\left|\partial^{\mathrm{int}}_F A\right| > h_0|A| \qquad(A \subseteq V(F),\ 1 \le|A| \le K). \tag*{(2)}

In particular, ∣V(F)∣>K|V(F)| > K, since V(F)V(F) has empty internal boundary in FF. Let

h=h0/2,δ=max⁡{2,NX(R)−1},h=h_0/2,\qquad\delta=\max\{2,N_X(R)-1\},

where

NX(R)=sup⁡x∈X∣BX(x,R)∣.N_X(R)=\sup_{x\in X}|B_X(x,R)|.

We construct (Gn)n(G_n)_n by induction. We start with G1G_1. Choose F1F_1 satisfying (2) with K=NX(2R)K=N_X(2R) and take a connected component G1G_1 with vertex set W1W_1. A nonempty subset E⊂W1E\subset W_1 of ambient diameter at most 2R2R lies in BX(x,2R)B_X(x,2R) for any x∈Ex\in E, so ∣E∣≤K|E|\le K. Its internal boundary is the same in G1G_1 and F1F_1, so (2) gives the required bound with constant hh.

For the construction of G2G_2, take M=W1M=W_1 and k=NX(4R)k=N_X(4R). Choose an integer b>∣M∣/hb>|M|/h and a new F2F_2 satisfying (2) with K=b+kK=b+k. First remove S0=V(F2)∩MS_0=V(F_2)\cap M. Let SS be the set of vertices removed so far and write FS=F2[V(F2)∖S]F_S=F_2[V(F_2)\setminus S] for the remaining induced subgraph. Whenever FSF_S contains a nonempty subset EE with

∣E∣≤k,∣∂FSintE∣≤h∣E∣,|E|\le k,\qquad\left|\partial^{\mathrm{int}}_{F_S}E\right|\le h|E|,

add EE to SS and repeat.

To see why this leaves a nonempty graph, count the internal boundary of SS in the original graph F2F_2. When EE is added, a vertex in SS can remain on the boundary only if it was already there. The vertices of EE on the new boundary are exactly those with a neighbor outside S∪ES\cup E, namely the vertices in ∂FintE\partial^{\mathrm{int}}_F E. Thus each deletion adds at most h∣E∣h|E| boundary vertices. The initial boundary has size at most ∣M∣|M|, and the sets removed at successive steps are disjoint. follows that

∣∂F2intS∣≤∣M∣+h∣S∣.|\partial^{\mathrm{int}}_{F_2}S| \le|M|+h|S|.

Initially ∣S∣≤∣M∣<b|S|\le|M|<b. If ∣S∣|S| first reaches or exceeds bb, then b≤∣S∣≤b+k=Kb\le|S|\le b+k=K, since the last step adds at most kk vertices. Thus (2) applies to SS. But ∣M∣<hb≤h∣S∣|M|<hb\le h|S|, so

∣∂F2intS∣≤∣M∣+h∣S∣<2h∣S∣=h0∣S∣,|\partial^{\mathrm{int}}_{F_2}S|\le|M|+h|S|<2h|S|=h_0|S|,

contrary to (2). Hence ∣S∣<b|S|<b throughout. The process terminates since F2F_2 is finite and each step removes at least one vertex. Since ∣V(F2)∣>K|V(F_2)|>K, the final FSF_S is nonempty.

Take a connected component G2G_2 of the final FSF_S and put W2=V(G2)W_2=V(G_2). The initial deletion gives W2∩W1=∅W_2\cap W_1=\varnothing. If ∅≠E⊆W2\varnothing\ne E\subseteq W_2 has ambient diameter at most 4R4R, then E⊆BX(x,4R)E\subseteq B_X(x,4R) for any x∈Ex\in E, so ∣E∣≤k|E|\le k. The stopping condition gives ∣∂FSintE∣>h∣E∣|\partial^{\mathrm{int}}_{F_S}E|>h|E|. Every neighbor in FSF_S of a vertex of EE lies in G2G_2, so ∂G2intE=∂FSintE\partial^{\mathrm{int}}_{G_2}E=\partial^{\mathrm{int}}_{F_S}E.

Now suppose that G1,…,GnG_1,\ldots,G_n have been constructed. Repeat the construction of G2G_2 with M=⋃i=1nWiM=\bigcup_{i=1}^{n}W_i and k=NX(2(n+1)R)k=N_X(2(n+1)R). The resulting component Gn+1G_{n+1} avoids MM and satisfies the required boundary estimate with the same constant hh. Each GnG_n is a subgraph of ΓR\Gamma_R, so the degree bound DD and the ambient edge length bound RR hold uniformly in nn.

Definition 3.5. For a finite simple graph GG, the adjacency matrix AGA_G acts on ℓ2(V(G))\ell^2(V(G)) and is defined by

(AG)xy={1,(x,y)∈E(G),0,otherwise.(AGf)(x)=∑y∈NG(x)f(y).(A_G)_{xy}= \begin{cases} 1, & (x,y)\in E(G),\\ 0, & \text{otherwise}. \end{cases} \qquad (A_Gf)(x)=\sum_{y\in N_G(x)}f(y).

The degree matrix is DG=diag⁡(dG(x))D_G=\operatorname{diag}(d_G(x)), and the combinatorial Laplacian is ΔG=DG−AG\Delta_G=D_G-A_G.

With each undirected edge counted once, its quadratic form is

⟨f,ΔGf⟩=∑(x,y)∈E(G)∣f(x)−f(y)∣2.\langle f,\Delta_Gf\rangle=\sum_{(x,y)\in E(G)}|f(x)-f(y)|^2.

Let (Gn)n(G_n)_n be the graphs with vertex sets (Wn)n(W_n)_n in Lemma 3.4. On ℓ2(Wn)\ell^2(W_n), consider the following operators

Δn=DGn−AGn,Pn=I−Δn2δ.\Delta_n=D_{G_n}-A_{G_n},\qquad P_n=I-\frac{\Delta_n}{2\delta}.

Each PnP_n is symmetric and positive semidefinite, with nonnegative entries and row sums one. Consider the constant

ρ=1−h24δ2∈(0,1).\rho=1-\frac{h^2}{4\delta^2}\in(0,1).

For the operator PnP_n, we have the following estimates on the diagonals of its powers PnℓP_n^\ell, for ℓ≤n\ell\le n.

Lemma 3.6. For x∈Wnx\in W_n and 1≤ℓ≤n1\le\ell\le n,

(Pnℓ)xx≤ρℓ.(3)(P_n^\ell)_{xx}\le\rho^\ell. \tag*{(3)}

Proof. Fix x∈Wnx\in W_n and 1≤ℓ≤n1\le\ell\le n. Let B=Wn∩BX(x,ℓR)B=W_n\cap B_X(x,\ell R) and let Pn,BP_{n,B} be the compression of PnP_n to ℓ2(B)\ell^2(B). For any function f∈ℓ2(B)f \in\ell^2(B), it extends by zero to WnW_n. List the distinct values of ∣f∣2|f|^2 as t1<⋯<tst_1 < \cdots< t_s, put t0=0t_0 = 0, and set

Ei={u:∣f(u)∣2≥ti}.E_i = \{u : |f(u)|^2 \ge t_i\}.

Since diam⁡X(B)≤2nR\operatorname{diam}_X(B) \le2nR, by the Lemma 3.4, we have that ∣∂GnEi∣≥h∣Ei∣|\partial_{G_n} E_i| \ge h|E_i|. Summing these inequalities with weights ti−ti−1t_i-t_{i-1} gives rise to the inequality

h∥f∥22≤∑i=1s(ti−ti−1)∣∂GnEi∣=∑(u,v)∈E(Gn)∣∣f(u)∣2−∣f(v)∣2∣.h\|f\|_2^2 \le\sum_{i=1}^{s}(t_i-t_{i-1})|\partial_{G_n} E_i| = \sum_{(u,v)\in E(G_n)} \bigl||f(u)|^2-|f(v)|^2\bigr|.

By Cauchy–Schwarz and the degree bound, the last sum is at most

(∑(u,v)∈E(Gn)∣f(u)−f(v)∣2)1/2(∑(u,v)∈E(Gn)(∣f(u)∣+∣f(v)∣)2)1/2≤⟨f,Δnf⟩1/2(2δ∥f∥22)1/2.\left(\sum_{(u,v)\in E(G_n)}|f(u)-f(v)|^2\right)^{1/2} \left(\sum_{(u,v)\in E(G_n)}(|f(u)|+|f(v)|)^2\right)^{1/2} \le\langle f,\Delta_n f\rangle^{1/2}(2\delta\|f\|_2^2)^{1/2}.

It follows that ⟨f,Δnf⟩≥h22δ∥f∥22\langle f,\Delta_n f\rangle\ge\frac{h^2}{2\delta}\|f\|_2^2, and hence

0≤⟨f,Pn,Bf⟩=∥f∥22−12δ⟨f,Δnf⟩≤ρ∥f∥22.0 \le\langle f,P_{n,B}f\rangle= \|f\|_2^2-\frac{1}{2\delta}\langle f,\Delta_n f\rangle\le\rho\|f\|_2^2.

Thus ∥Pn,B∥≤ρ\|P_{n,B}\| \le\rho.

Each multiplication by PnP_n enlarges the support by at most RR in the ambient metric. Hence PnjδxP_n^j\delta_x is supported in BB for 0≤j≤ℓ0 \le j \le\ell. Projecting onto ℓ2(B)\ell^2(B) after each step therefore leaves these vectors unchanged, so Pn,Bℓδx=PnℓδxP_{n,B}^{\ell}\delta_x=P_n^{\ell}\delta_x. Accordingly, we have that

(Pnℓ)xx=⟨δx,Pn,Bℓδx⟩≤∥Pn,B∥ℓ≤ρℓ.(P_n^\ell)_{xx}=\langle\delta_x,P_{n,B}^{\ell}\delta_x\rangle\le\|P_{n,B}\|^\ell\le\rho^\ell.

We use the estimates on the entries of PnmP_n^m to define a sequence of new graph and then removing certain vertices and edges gives us a sequence of graphs with large girth and degree at most 6.

Lemma 3.7. Let (Gn)n(G_n)_n be as in Lemma 3.4, with vertex sets (Wn)n(W_n)_n. There exist a constant L<∞L < \infty, an integer n0n_0 and a sequence (Hn)n≥n0(H_n)_{n\ge n_0} of simple graphs with nonempty vertex sets (Un)n≥n0(U_n)_{n\ge n_0} such that Un⊆WnU_n \subseteq W_n for each n≥n0n \ge n_0 and

dHn(x)∈{5,6}(x∈Un),girth⁡(Hn)⟶∞,dX(x,y)≤L((x,y)∈E(Hn)).d_{H_n}(x) \in\{5,6\}\quad(x\in U_n),\qquad\operatorname{girth}(H_n)\longrightarrow\infty,\qquad d_X(x,y)\le L\quad((x,y)\in E(H_n)).

Proof. Fix an integer m≥1m \ge1 with ρm≤1/20\rho^m \le1/20. For n≥4mn \ge4m, denote

Q=Pnm,qn=⌊n/m⌋+1,N=∣Wn∣.Q=P_n^m,\qquad q_n=\lfloor n/m\rfloor+1,\qquad N=|W_n|.

The matrix QQ is symmetric and positive semidefinite, with nonnegative entries and row sums one. By Lemma 3.6,

0≤Qxy≤QxxQyy≤120.0\le Q_{xy}\le\sqrt{Q_{xx}Q_{yy}}\le\frac{1}{20}.

Let JJ be the simple graph on WnW_n with edge set

E(J)={(x,y):x≠y, Qxy>0}.E(J)=\{(x,y):x\ne y,\ Q_{xy}>0\}.

Since PnP_n has propagation at most RR, every edge of JJ has ambient length at most mRmR.

For a graph KK, write C(K)C(K) for its cycles of length less than qnq_n. For u∈[0,1]E(J)u\in[0,1]^{E(J)}, define

F(u)=∑e∈E(J)ue−∑C∈C(J)∏e∈E(C)ue.F(u)=\sum_{e\in E(J)}u_e-\sum_{C\in C(J)}\prod_{e\in E(C)}u_e.

Set a(x,y)=10Qxya_{(x,y)}=10Q_{xy}, so that ae∈[0,1/2]a_e\in[0,1/2]. By symmetry and the row sums of QQ,

∑e∈E(J)ae=5∑x∈Wn(1−Qxx)≥194N.\sum_{e\in E(J)}a_e=5\sum_{x\in W_n}(1-Q_{xx})\ge\frac{19}{4}N.

Choose a starting vertex and a direction for each kk-cycle of JJ. Distinct cycles then give distinct terms in the expansion of tr⁡(Qk)\operatorname{tr}(Q^k). Each such term is the product of the corresponding edge weights aea_e divided by 10k10^k. Since the trace expansion has nonnegative terms and mk≤nmk\le n for k<qnk<q_n, we obtain that

∑C∈C(J)∣E(C)∣=k∏e∈E(C)ae≤10ktr⁡(Qk)≤(10ρm)kN≤2−kN(3≤k<qn).(4)\sum_{\substack{C\in C(J)\\ |E(C)|=k}}\prod_{e\in E(C)}a_e \le10^k\operatorname{tr}(Q^k) \le(10\rho^m)^kN \le2^{-k}N \qquad(3\le k<q_n). \tag*{(4)}

Consequently,

F(a)≥(194−∑k=3∞2−k)N=92N>4N.F(a)\ge\left(\frac{19}{4}-\sum_{k=3}^{\infty}2^{-k}\right)N =\frac{9}{2}N>4N.

With the other coordinates fixed, FF is affine in ueu_e. Thus one of the choices ue=0u_e=0 and ue=1u_e=1 preserves or increases FF. Applying this successively to all coordinates of aa gives b∈{0,1}E(J)b\in\{0,1\}^{E(J)} with F(b)≥F(a)F(b)\ge F(a). Let J0J_0 be the graph on WnW_n with edges {e∈E(J):be=1}\{e\in E(J):b_e=1\}. Then

∣E(J0)∣−∣C(J0)∣=F(b)>4N.|E(J_0)|-|C(J_0)|=F(b)>4N.

For each C∈C(J0)C\in C(J_0) choose eC∈E(C)e_C\in E(C), and let J1J_1 be obtained by deleting the set {eC:C∈C(J0)}\{e_C:C\in C(J_0)\}. Hence

∣E(J1)∣≥∣E(J0)∣−∣C(J0)∣>4N,girth⁡(J1)≥qn.|E(J_1)|\ge|E(J_0)|-|C(J_0)|>4N, \qquad\operatorname{girth}(J_1)\ge q_n.

Repeatedly delete vertices of current degree at most four from J1J_1. If all vertices were deleted in the order v1,…,vNv_1,\ldots,v_N, counting each edge when its first endpoint is deleted would result in

∣E(J1)∣=∑i=1NdJ1[{vi,…,vN}](vi)≤4N,|E(J_1)|=\sum_{i=1}^{N}d_{J_1[\{v_i,\ldots,v_N\}]}(v_i)\le4N,

and this is a contradiction. The remaining graph KK therefore has a nonempty vertex set Un⊆WnU_n\subseteq W_n and satisfies

min⁡v∈UndK(v)≥5,girth⁡(K)≥qn.\min_{v\in U_n}d_K(v)\ge5, \qquad\operatorname{girth}(K)\ge q_n.

Starting from KK, delete edges whose two endpoints both have current degree greater than five until no such edge remains. Write K′K' for the resulting graph and put

S={v∈Un:dK′(v)>5}.S = \{v \in U_n : d_{K'}(v) > 5\}.

Then SS is independent and dK′(u)=5d_{K'}(u) = 5 for u∈Un∖Su \in U_n \setminus S. For each v∈Sv \in S, partition 2rv2r_v of its incident edges into pairs, where

rv=⌊dK′(v)−52⌋.r_v = \left\lfloor\frac{d_{K'}(v)-5}{2} \right\rfloor.

Each pair forms a length-two path u−v−wu-v-w with distinct endpoints in Un∖SU_n \setminus S. The chosen paths are pairwise edge-disjoint, since the pairs at each vv are disjoint and each edge of K′K' meets at most one vertex of SS.

Let P\mathcal{P} consist of the chosen length-two paths together with the length-one paths given by the unused edges of K′K'. For a path p∈Pp \in\mathcal{P} with endpoints x,yx,y, write ℓ(p)\ell(p) for its length and set ∂p=(x,y)\partial p=(x,y). Define

V(Hn)=Un,E(Hn)={∂p:p∈P}.V(H_n)=U_n,\qquad E(H_n)=\{\partial p:p\in\mathcal{P}\}.

If distinct p,p′∈Pp,p'\in\mathcal{P} had the same endpoints, traversing pp and returning along p′p' would give a closed trail in K′K'. Since a nonempty closed trail contains a cycle, this would give

qn≤girth⁡(K′)≤ℓ(p)+ℓ(p′)≤4,q_n \leq\operatorname{girth}(K') \leq\ell(p)+\ell(p') \leq4,

contrary to qn≥5q_n\geq5. Thus the endpoint map is injective and HnH_n is simple.

For u∈Un∖Su\in U_n\setminus S, let sus_u be the number of chosen length-two paths with endpoint uu. Each such path replaces one edge incident to uu by one shortcut. For v∈Sv\in S, each of the rvr_v chosen paths removes two edges incident to vv. Hence

dHn(u)=dK′(u)−su+su=5(u∈Un∖S),d_{H_n}(u)=d_{K'}(u)-s_u+s_u=5 \qquad(u\in U_n\setminus S),
dHn(v)=dK′(v)−2rv∈{5,6}(v∈S).d_{H_n}(v)=d_{K'}(v)-2r_v\in\{5,6\} \qquad(v\in S).

Let pep_e be the unique path representing e∈E(Hn)e\in E(H_n). For a cycle CC in HnH_n, concatenating the paths pep_e in cycle order gives a closed trail in K′K', since the paths are pairwise edge-disjoint. Therefore

qn≤∑e∈E(C)ℓ(pe)≤2∣E(C)∣,girth⁡(Hn)≥qn2⟶∞.q_n \leq\sum_{e\in E(C)}\ell(p_e)\leq2|E(C)|,\qquad\operatorname{girth}(H_n)\geq\frac{q_n}{2}\longrightarrow\infty.

Finally, if e=(x,y)∈E(Hn)e=(x,y)\in E(H_n), the triangle inequality along pep_e gives

dX(x,y)≤mRℓ(pe)≤2mR.d_X(x,y)\leq mR\ell(p_e)\leq2mR.

The conclusion follows with n0=4mn_0=4m and L=2mRL=2mR.

We would like to remark here that the shortcut construction follows from the arguments in [9 Lemma 14].

A state on the maximal uniform Roe algebra

In this section, we construct a state on the maximal uniform Roe algebra using products of edge signs along short paths, for any given signing of graphs. The state vanishes on the signed Laplacian and provides the lower bound for the maximal norm in § 5.

Let (Hn)n(H_n)_n satisfy the conclusion of Lemma 3.7, with pairwise disjoint vertex sets (Un)n(U_n)_n. The degrees are five or six, the ambient edge lengths are at most LL, and girth⁡(Hn)→∞\operatorname{girth}(H_n) \to\infty. After discarding a finite initial segment, we have girth⁡(Hn)>6\operatorname{girth}(H_n) > 6 for every remaining nn. Set U=⋃nUnU = \bigcup_n U_n and p=1Up = 1_U for this sequence.

A signing of a graph HH is a map

σ:E(H)→{−1,1}.\sigma: E(H) \to\{-1,1\}.

Its signed adjacency matrix has entry σ((x,y))\sigma((x,y)) when (x,y)(x,y) is an edge and zero otherwise. For a sequence of signings σ=(σn)n\sigma= (\sigma_n)_n, let DD and BσB_\sigma be the block diagonal operators with blocks DHnD_{H_n} and the signed adjacency matrices, respectively, extended by zero off UU. The degree diagonal has norm at most six. For BσB_\sigma, Cauchy–Schwarz bounds ∣(Bσf)(x)∣2\lvert(B_\sigma f)(x)\rvert^2 by six times the sum of ∣f(y)∣2\lvert f(y)\rvert^2 over neighbors yy of xx. Summing over xx counts each ∣f(y)∣2\lvert f(y)\rvert^2 at most six times, so ∥Bσ∥≤6\lVert B_\sigma\rVert\le6. The ambient edge length bound gives propagation at most LL. Define

Lσ=D−Bσ,T=p−Lσ/12.L_\sigma= D - B_\sigma,\qquad T = p - L_\sigma/12.

The sign of a path is the product of its edge signs, with value one on the empty path. Since girth⁡(Hn)→∞\operatorname{girth}(H_n) \to\infty, choose integers Jn≥1J_n \ge1 with Jn→∞J_n \to\infty and 3⋅2Jn<girth⁡(Hn)3 \cdot2^{J_n} < \operatorname{girth}(H_n). Write dn=dHngrd_n = d_{H_n}^{\mathrm{gr}}. For 1≤j≤Jn1 \le j \le J_n, let rj=2jr_j = 2^j, and define

Kn,j(x,y)={sign of the shortest path from x to y,dn(x,y)≤rj,0,dn(x,y)>rj.K_{n,j}(x,y) = \begin{cases} \text{sign of the shortest path from } x \text{ to } y, & d_n(x,y) \le r_j,\\ 0, & d_n(x,y) > r_j. \end{cases}

Two distinct paths of length at most rjr_j with the same endpoints would contain a cycle of length at most 2rj2r_j in their union. The girth bound therefore makes the sign in the first case well defined. Define

Kn=1Jn∑j=1JnKn,j.K_n = \frac{1}{J_n}\sum_{j=1}^{J_n} K_{n,j}.

Then KnK_n is real and symmetric, its entries have absolute value at most one, Kn(x,x)=1K_n(x,x) = 1, and Kn(x,y)=σn((x,y))K_n(x,y) = \sigma_n((x,y)) on each edge of HnH_n.

Lemma 4.1. Let MM be a positive integer. For any S⊆UnS \subseteq U_n with ∣S∣≤M\lvert S\rvert\le M, there is a positive semidefinite matrix on SS that differs from Kn∣S×SK_n|_{S \times S} entrywise by at most (M2)/Jn\binom{M}{2}/J_n.

Proof. Let ISI_S be the set of indices jj in the average defining KnK_n for which no pair x,y∈Sx,y \in S satisfies rj<dn(x,y)≤2rjr_j < d_n(x,y) \le2r_j. Since 2rj=rj+12r_j = r_{j+1}, the intervals (rj,2rj](r_j,2r_j] are disjoint. Each unordered pair in SS therefore excludes at most one index, so at most (M2)\binom{M}{2} indices lie outside ISI_S.

Fix j∈ISj \in I_S. If dn(x,y)≤rjd_n(x,y) \le r_j and dn(y,z)≤rjd_n(y,z) \le r_j, then dn(x,z)≤2rjd_n(x,z) \le2r_j by the triangle inequality. By the definition of ISI_S, we have that dn(x,z)≤rjd_n(x,z) \le r_j. Thus the relation dn(x,y)≤rjd_n(x,y) \le r_j on SS is an

equivalence relation. Therefore, Kn,j(x,y)=0K_{n,j}(x,y)=0 for vertices in distinct classes.

Fix a root oo in an equivalence class CC and write sxs_x for the sign of the shortest path from oo to x∈Cx \in C. For x,y∈Cx,y \in C, the three shortest paths between o,x,yo,x,y have a connected union with at most 3rj<girth⁡(Hn)3r_j < \operatorname{girth}(H_n) edges, so their union is a tree. The signs on the common initial segment of the paths from oo to xx and from oo to yy cancel in sxsys_xs_y, leaving the sign along the path from xx to yy. Hence Kn,j(x,y)=sxsyK_{n,j}(x,y)=s_xs_y on CC. Each block is therefore positive semidefinite, and so is Kn,j∣S×SK_{n,j}|_{S\times S}.

Now set

AS=1Jn∑j∈ISKn,j∣S×S.A_S=\frac{1}{J_n}\sum_{j\in I_S}K_{n,j}|_{S\times S}.

Obviously, ASA_S is positive semidefinite. Since each entry of Kn,jK_{n,j} has absolute value at most one, for x,y∈Sx,y\in S we have

∣Kn(x,y)−AS(x,y)∣≤1Jn∑j∉IS∣Kn,j(x,y)∣≤(M2)Jn.|K_n(x,y)-A_S(x,y)|\leq\frac{1}{J_n}\sum_{j\notin I_S}|K_{n,j}(x,y)|\leq\frac{\binom{M}{2}}{J_n}.

where the sum is over the indices in the average defining KnK_n.

Now, we are ready to define the state using ultrafilter. A free ultrafilter ω\omega on the infinite index set II of (Hn)n(H_n)_n is a family of subsets of II closed under finite intersections and passage to supersets. It contains every cofinite subset of II and exactly one of EE and I∖EI\setminus E for each E⊆IE\subseteq I. For a bounded sequence (an)n∈I(a_n)_{n\in I}, its ultralimit

a=lim⁡n→ωana=\lim_{n\to\omega}a_n

is defined by

{n∈I:∣an−a∣<ε}∈ωfor every ε>0.\{n\in I:|a_n-a|<\varepsilon\}\in\omega\qquad\text{for every }\varepsilon>0.

The existence of this limit is guaranteed by the compactness of a closed disk containing the sequence, and it is obvious that the limit is unique.

Lemma 4.2. Fix a free ultrafilter ω\omega on the indices. The formula

Φn(A)=1∣Un∣∑x,y∈UnAxyKn(x,y),Φ(A)=lim⁡n→ωΦn(A),A∈Cu[X],(5)\Phi_n(A)=\frac{1}{|U_n|}\sum_{x,y\in U_n}A_{xy}K_n(x,y),\qquad\Phi(A)=\lim_{n\to\omega}\Phi_n(A),\qquad A\in C_u[X], \tag*{(5)}

defines a functional Φ\Phi on Cu[X]C_u[X] that extends uniquely to a state on Cu,max⁡∗(X)C^*_{u,\max}(X) satisfying

Φ(p)=1,Φ(Lσ)=0.\Phi(p)=1,\qquad\Phi(L_\sigma)=0.

Proof. Fix AA with propagation at most RA>0R_A>0. Each row and column has at most M=NX(RA)M=N_X(R_A) nonzero entries. At most M∣Un∣M|U_n| terms can contribute to Φn(A)\Phi_n(A), and each has absolute value at most ∥A∥red\|A\|_{\mathrm{red}}, since ∣Kn(x,y)∣≤1|K_n(x,y)|\leq1. Thus Φn(A)\Phi_n(A) is bounded in nn for each AA, and the fixed ultralimit defines a linear functional Φ\Phi. Since Kn(x,x)=1K_n(x,x)=1, we have Φn(1)=1\Phi_n(1)=1 for all nn, hence Φ(1)=1\Phi(1)=1. Fix nn and denote

Sz={x∈Un:Azx≠0}.S_z=\{x\in U_n:A_{zx}\ne0\}.

The row bound gives ∣Sz∣≤M|S_z|\le M, while the column bound shows that only finitely many rows z∈Xz\in X can meet UnU_n. Expanding the matrix product gives rise to

Φn(A∗A)=1∣Un∣∑z∈X∑x,y∈SzAzx‾AzyKn(x,y).\Phi_n(A^*A)=\frac{1}{|U_n|}\sum_{z\in X}\sum_{x,y\in S_z}\overline{A_{zx}}A_{zy}K_n(x,y).

For every intermediate row index z∈Xz\in X, the set SzS_z is contained in the ambient ball BX(z,RA)B_X(z,R_A). Its cardinality is therefore bounded by MM, so Lemma 4.1 applies even when the vertices of SzS_z are far apart in the graph metric. Replacing the kernel on SzS_z by its positive semidefinite approximation makes the inner sum nonnegative. The original sum is real since KnK_n is real and symmetric. The absolute value of the change is at most ϵn(∑x∈Sz∣Azx∣)2\epsilon_n\left(\sum_{x\in S_z}|A_{zx}|\right)^2, where ϵn=(M2)/Jn\epsilon_n=\binom{M}{2}/J_n. By Cauchy–Schwarz inequality, we have that

∑z∈X(∑x∈Sz∣Azx∣)2≤M∑x∈Un∥Aδx∥22≤M∣Un∣∥A∥red2.\sum_{z\in X}\left(\sum_{x\in S_z}|A_{zx}|\right)^2\le M\sum_{x\in U_n}\|A\delta_x\|_2^2\le M|U_n|\|A\|_{\mathrm{red}}^2.

Consequently,

Φn(A∗A)≥−Mϵn∥A∥red2.\Phi_n(A^*A)\ge-M\epsilon_n\|A\|_{\mathrm{red}}^2.

For each fixed AA, the number MM is fixed while Jn→∞J_n\to\infty, so the error tends to zero. The ultralimit preserves inequalities between bounded real sequences, hence Φ(A∗A)≥0\Phi(A^*A)\ge0.

For the algebraic GNS construction, quotient by the null space of the positive form ⟨a,b⟩Φ=Φ(a∗b)\langle a,b\rangle_\Phi=\Phi(a^*b). If g=Mfg=M_f is a diagonal contraction or g=Vtg=V_t is a partial translation operator, then 1−g∗g=b∗b1-g^*g=b^*b for a bounded diagonal operator bb. We may take b=M1−∣f∣2b=M_{\sqrt{1-|f|^2}} in the first case and b=1X∖dom⁡tb=1_{X\setminus\operatorname{dom}t} in the second. It follows that

Φ(a∗a)−Φ((ga)∗(ga))=Φ((ba)∗(ba))≥0.\Phi(a^*a)-\Phi((ga)^*(ga))=\Phi((ba)^*(ba))\ge0.

Thus these left multiplications preserve the null space and act contractively on the quotient. By the decomposition in § 2.2, the left action of A=∑iMfiVtiA=\sum_i M_{f_i}V_{t_i} is bounded by ∑i∥fi∥∞\sum_i\|f_i\|_\infty. Left multiplication by A∗A^* is its adjoint for the form, since Φ(a∗Ab)=Φ((A∗a)∗b)\Phi(a^*Ab)=\Phi((A^*a)^*b). Completing gives a bounded ∗*-representation π\pi. Since [1][1] is a unit vector,

∣Φ(A)∣=∣⟨[1],π(A)[1]⟩∣≤∥π(A)∥≤∥A∥max⁡.|\Phi(A)|=|\langle[1],\pi(A)[1]\rangle|\le\|\pi(A)\|\le\|A\|_{\max}.

The representation extends to Cu,max⁡∗(X)C^*_{u,\max}(X) by the definition of the maximal norm, and its vector state at [1][1] is the extension of Φ\Phi.

The values of KnK_n on the diagonal and on the edges determine the evaluations in (5) on DD and BσB_\sigma. On the diagonal Kn(x,x)=1K_n(x,x)=1, so Φn(D)\Phi_n(D) is the average degree of HnH_n. For each edge (x,y)(x,y), the two ordered entries of BσB_\sigma contribute σn((x,y))2=1\sigma_n((x,y))^2=1 each. Hence

Φn(D)=Φn(Bσ)=2∣E(Hn)∣∣Un∣,Φn(p)=1.\Phi_n(D)=\Phi_n(B_\sigma)=\frac{2|E(H_n)|}{|U_n|},\qquad\Phi_n(p)=1.

Taking the ultralimit gives Φ(p)=1\Phi(p)=1 and Φ(Lσ)=Φ(D)−Φ(Bσ)=0\Phi(L_\sigma)=\Phi(D)-\Phi(B_\sigma)=0.

Proof of the main theorem

In this section, we shall prove our main theorem. Before proving, we shall recall the following theorem of Marcus–Spielman–Srivastava [10] to estimate the maximal norm of the operator of the signed adjacency matrices.

Theorem 5.1 ([10 Theorems 3.2, 3.7, 4.4 and 5.2]). Let HH be a nonempty finite simple graph with degree at most d≥2d \ge2. Then, there is a signing

σ:E(H)→{−1,1},\sigma:E(H)\to\{-1,1\},

such that the largest eigenvalue of its signed adjacency matrix BσB_\sigma satisfies

λmax⁡(Bσ)≤2d−1.\lambda_{\max}(B_\sigma)\le2\sqrt{d-1}.

For the graphs constructed in the Lemma 3.7, we take d=6d=6. The above eigenvalue bound gives an upper bound on BσB_\sigma in the concrete operator order. The minimum degree of vertices is five, and 5>255>2\sqrt{5}, so subtracting BσB_\sigma from the degree diagonal ensures a strictly positive lower bound for the signed Laplacian on ℓ2(U)\ell^2(U).

Theorem 5.2. Let XX be a discrete metric space of bounded geometry. If XX does not have property A, then the quotient map λ:Cu,max⁡∗(X)→Cu∗(X)\lambda:C^*_{u,\max}(X)\to C^*_u(X) is not an isomorphism.

Proof. Suppose that XX does not have property A. Let (Hn)n(H_n)_n be the sequence of graphs obtained in Lemmas 3.4 and 3.7. By Theorem 5.1, for any a signing σn\sigma_n on each HnH_n, the signed adjacency matrix has largest eigenvalue at most 252\sqrt{5}. Let

Lσ=D−Bσ,T=p−Lσ12.L_\sigma=D-B_\sigma,\qquad T=p-\frac{L_\sigma}{12}.

Since the disjoint union of the HnH_n has maximum degree six, a greedy edge coloring uses at most eleven colors. Each color class is a matching. Orient its edges and let WiW_i be the resulting partial translation operator weighted by the edge signs. Each WiW_i belongs to Cu[X]C_u[X] because its displacement is at most LL. Writing qi=WiWi∗q_i=W_iW_i^*, we have qiWi=Wiq_iW_i=W_i. By summing over the matchings, we have that

D±Bσ=∑i(qi±Wi)∗(qi±Wi).D\pm B_\sigma=\sum_i(q_i\pm W_i)^*(q_i\pm W_i).

Therefore,

0≤Lσ≤2D0\le L_\sigma\le2D

in Cu,max⁡∗(X)C^*_{u,\max}(X). The diagonal 6p−D6p-D has a square root in Cu[X]C_u[X], so 2D≤12p2D\le12p, and hence

0≤T≤p.0\le T\le p.

In the reduced completion, the signing bound gives Bσ≤25 pB_\sigma\le2\sqrt{5}\,p, while D≥5pD\ge5p. It follows that

0≤λ(T)≤cλ(p),c=1−5−2512<1.0 \le\lambda(T) \le c\lambda(p), \qquad c=1-\frac{5-2\sqrt{5}}{12}<1.

On the maximal completion, the state Φ\Phi in Lemma 4.2 staisfies that

Φ(p)=1,Φ(Lσ)=0.\Phi(p)=1,\qquad\Phi(L_\sigma)=0.

Thus, Φ(T)=1\Phi(T)=1. It follows from 0≤T≤p0\le T\le p that

∥T∥max⁡=1.\lVert T\rVert_{\max}=1.

Set a=(T−c1)+a=(T-c1)_+ by continuous functional calculus. Since TT is positive and has maximal norm one, its spectrum in Cu,max⁡∗(X)C^*_{u,\max}(X) contains 11, so ∥a∥max⁡=1−c>0\lVert a\rVert_{\max}=1-c>0. On the other hand, 0≤λ(T)≤c10\le\lambda(T)\le c1, and therefore

λ(a)=(λ(T)−c1)+=0.\lambda(a)=(\lambda(T)-c1)_+=0.

Thus aa is a nonzero element of ker⁡λ\ker\lambda. □

Together with [15 Proposition 1.3], we have the following:

Theorem 5.3. Let XX be a discrete metric space with bounded geometry. Then, XX has property A if and only if the canonical quotient map

λ:Cu,max⁡∗(X)⟶Cu∗(X)\lambda:C^*_{u,\max}(X)\longrightarrow C^*_u(X)

is an isomorphism.

Jintao Deng

Department of mathematics, Center of Noncommutative Geometry, Shanghai University of Finance and Economics

Email: [email protected]

Geng Tian

Department of Mathematics, Liaoning University, Liaoning, China 110036

Email: [email protected]

References

  1. [1]V. Alekseev and M. Finn-Sell, Non-amenable principal groupoids with weak containment, Int. Math. Res. Not. IMRN (2018), 2332–2340.
  2. [2]P. Ara, K. Li, F. Lledó and J. Wu, Amenability and uniform Roe algebras, J. Math. Anal. Appl. 459 (2018), 686–716.DOI
  3. [3]G. Arzhantseva, E. Guentner and J. Špakula, Coarse non-amenability and coarse embed-dings, Geom. Funct. Anal. 22 (2012), 22–36.
  4. [4]J. Brodzki, G. A. Niblo, J. Špakula, R. Willett and N. Wright, Uniform local amenability, J. Noncommut. Geom. 7 (2013), 583–603.arxiv.org/abs/1203.6169
  5. [5]A. Buss, S. Echterhoff and R. Willett, The maximal injective crossed product, Ergodic Theory Dynam. Systems 40 (2020), 2995–3014.DOI
  6. [6]G. Elek, Uniform local amenability implies property A, Proc. Amer. Math. Soc. 149 (2021), 2573–2577.arxiv.org/abs/1912.00806
  7. [7]G. Elek and Á. Timár, Strong almost finiteness, J. Funct. Anal. 289 (2025), Article 111116.arxiv.org/abs/2308.14554
  8. [8]G. Gong, Q. Wang and G. Yu, Geometrization of the strong Novikov conjecture for residually finite groups, J. Reine Angew. Math. 621 (2008), 159–189.DOI
  9. [9]G. Kun, Expanders have a spanning Lipschitz subgraph with large girth, arXiv:1303.4982v2 (2013).arxiv.org/abs/1303.4982
  10. [10]A. W. Marcus, D. A. Spielman and N. Srivastava, Interlacing families I: Bipartite Ramanujan graphs of all degrees, Ann. of Math. 182 (2015), 307–325.arxiv.org/abs/1304.4132
  11. [11]J. Roe, An index theorem on open manifolds. I, J. Differential Geom. 27 (1988), 87–113.DOI
  12. [12]J. Roe, Lectures on coarse geometry, University Lecture Series, vol. 31, Amer. Math. Soc., Providence, RI, 2003.DOI
  13. [13]J. Roe and R. Willett, Ghostbusting and property A, J. Funct. Anal. 266 (2014), no. 3, 1674–1684.
  14. [14]G. Skandalis, J.-L. Tu and G. Yu, The coarse Baum–Connes conjecture and groupoids, Topology 41 (2002), 807–834.DOI
  15. [15]J. Špakula and R. Willett, Maximal and reduced Roe algebras of coarsely embeddable spaces, J. Reine Angew. Math. 678 (2013), 35–68.arxiv.org/abs/1110.1531
  16. [16]R. Willett, A non-amenable groupoid whose maximal and reduced C*-algebras are the same, Münster J. Math. 8 (2015), 241–252.arxiv.org/abs/1504.05615
  17. [17]G. Yu, The coarse Baum–Connes conjecture for spaces which admit a uniform embedding into Hilbert space, Invent. Math. 139 (2000), 201–240.DOI

Paper details

vlnc:W_czm8zh53
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents