This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Graphs with nonnegative Ricci curvature and maximum degree at most 3

Fengwen Han Fengwen, Han: School of Mathematical Sciences, Fudan University, Shanghai 200433, China [email protected]  and  Tao Wang Tao, Wang: School of Mathematical Sciences, Fudan University, Shanghai 200433, China [email protected]
Abstract.

In this paper, we classify graphs with nonnegative Lin-Lu-Yau-Ollivier Ricci curvature, maximum degree at most 3 and diameter at least 6.

1. Introduction

Ricci curvature is one of the most important geometric quantities in Riemannian geometry. For discrete settings, various definitions of Ricci curvature were developed over the past few decades. In this paper, we focus on Olliver Ricci curvature and its variations. The original Ollivier Ricci curvature was developed on metric spaces and Markov chains, which can be applied on normalized graphs; see [Oll09, Oll07]. In 2011, Lin-Lu-Yau [LLY11] modified Ollivier’s definition. This version was widely used in the research of normalized graphs. Münch and Wojciechowski [MW19] extended Lin-Lu-Yau’s definition to general weighted graphs. Bakry-Émery curvature is another important Ricci curvature on graphs, readers may refer to [Sch99, LY10] for more details.

There are significant geometric and analytic results for Riemannian manifolds with nonnegative Ricci curvature. As discrete analogs, graphs with nonnegative Ricci curvature are also worthy to study. A basic method is to explore their local geometric structures. Many papers on this subject were published recently. Cushing et al. [CKL+19] classified all 3-regular graphs with nonnegative Bakry-Émery curvature. They also classified all 3-regular graphs with nonnegative original Ollivier Ricci curvature. Under the setting of Lin-Lu-Yau-Ollivier Ricci curvature, there are many works on the classification of Ricci-flat graphs, i.e. graphs whose Ricci curvature vanishes on every edge; see [CKL+18, LLY14, CP18, HLYY18, BLY21].

It is interesting to study harmonic functions on graphs. For the continuous case, Yau [Yau75] showed that positive harmonic functions on a complete, noncompact Riemannian manifold with nonnegative Ricci curvature are constant, which is known as Liouville property of harmonic functions. For the discrete case, there are also some known results. Jost, Münch and Rose [JMR19] proved that bounded harmonic functions are constant under the nonnegative Lin-Lu-Yau-Ollivier Ricci curvature condition. We also refer to [Hua19, BHL+15, HLLY19] for Liouville properties under the other curvature conditions. However, the strong Liouville property of graphs with non-negative Lin-Lu-Yau-Ollivier Ricci curvature is still unknown. As an application of the main results, we prove that it is true for those graphs with maximum degree at most three.

Let 𝒢\mathcal{G} be the class of combinatorial graphs with following properties:

1) Lin-Lu-Yau-Ollivier Ricci curvatures on all edges of G𝒢G\in\mathcal{G} are nonnegative;

2) Degrees of all vertices of G𝒢G\in\mathcal{G} are at most 3;

3) Diameters of the graphs in 𝒢\mathcal{G} are at least 6.

The precise definition of 𝒢\mathcal{G} is given in section 2. We classify all graphs in 𝒢\mathcal{G} by the following two theorems.

Theorem 1.1.

Let G𝒢G\in\mathcal{G} be a finite graph. Then GG is isomorphic to a finite path, a simple cycle, a prism graph, a Möbius ladder, a particular graph as shown in Figure 1, or a quasi-ladder graph as shown in Figure 2.

Refer to caption
(a) Path
Refer to caption
(b) Cycle
Refer to caption
(c) Prism graph
Refer to caption
(d) Möbius ladder
Refer to caption
(e) Particular graph
Figure 1.
Refer to caption
Figure 2. Quasi-ladder
Theorem 1.2.

Let G𝒢G\in\mathcal{G} be an infinite graph. Then GG is isomorphic to one of the following graphs.

Refer to caption
(a) Both-side infinite line
Refer to caption
(b) One-side infinite line
Refer to caption
(c) Infinite ladder
Refer to caption
(d) Infinite quasi-ladder 1
Refer to caption
(e) Infinite quasi-ladder 2
Refer to caption
(f) Infinite quasi-ladder 3
Refer to caption
(g) Infinite quasi-ladder 4
Refer to caption
(h) Infinite quasi-ladder 5
Refer to caption
(i) Infinite quasi-ladder 6
Refer to caption
(j) Infinite quasi-ladder 7
Figure 3.
Remark 1.3.
  1. (1)

    Theorem 1.1 and Theorem 1.2 also hold for normalized graphs; see Theorem 2.5.

  2. (2)

    For normalized graphs, if we change the nonegative Lin-Lu-Yau-Ollivier Ricci curvature constraint to nonnegative Ollivier Ricci curvature, we also get a classification; see Theorem 2.6.

  3. (3)

    It is well known that manifolds with nonnegative Ricci curvature satisfy Cheeger-Gromoll splitting theorem [CG72]. As a discrete analog, the splitting theorem also holds for Graphs in 𝒢\mathcal{G}, i.e. if a both-side infinite line is contained in G𝒢G\in\mathcal{G}, then GG is the Cartesian product of a line and a graph.

Bai, Lu and Yau [BLY21] classified all Ricci-flat graphs with maximum degree at most 4. Compared with our results, they require a more stringent curvature condition but relax the maximum degree condition from 3 to 4, which causes a huge increase in the complexity of the problem.

The proofs of the main results are shown in section 4. We give a sketch here. Let P=v0v1vlP=v_{0}v_{1}\cdots v_{l} be a diameter of G𝒢G\in\mathcal{G} with l6l\geq 6. The local structure along PP can be characterized in Lemma 3.1 , Lemma 3.3 and Lemma 3.4. Moreover, the subgraph induced by PP and its neighbors is a quasi-ladder. We divide GG into four parts as shown in Figure 4. If LL is connected to RR via OO, we will get a special cycle (we call it a geodesic cycle in this paper) by Lemma 4.1, and we prove GG has to be a cycle, a prism or a Möbius ladder by Lemma 4.2. If LL is not connected to RR via OO, then O=O=\emptyset and GG is a quasi-ladder, otherwise we will find a longer geodesic path. For infinite graphs, the proof is similar. For graphs with small diameter, say l=4 or 5l=4\text{ or }5, the structures are more complicated. However, since the maximum degree is at most 3, graphs with nonnegative curvature and small diameter can be enumerated by computers, which we are not going to explore in this paper.

Refer to caption
Figure 4.

The paper is organized as follows: In next section, we introduce some notions on graphs and the definition of Ricci curvature. Section 3 gives the local structures of graph along a diameter, which serve as basic tools in the following discussions. In section 4, we prove the main results of this paper. In section 5, as an applications of the main results, we show that all positive harmonic functions on G𝒢G\in\mathcal{G} are constant.

2. Preliminaries

2.1. Basic notions

Throughout this paper, we always assume that G(V,E)G(V,E) is an undirected, connected, locally finite, simple graph with vertex set VV and edge set EE. Let |G|:=|V||G|:=|V| be the number of vertices of the graph. For any vertices u,vVu,v\in V, we call vv a neighbor of uu if {u,v}E\{u,v\}\in E, denoted by uvu\sim v. For notational simplicity, we write uvuv for the unordered pair {u,v}\{u,v\}. For A,BVA,B\subset V, let E(A,B):={uvE:uA,vB}E(A,B):=\{uv\in E:u\in A,v\in B\}. The degree of a vertex vVv\in V is defined via deg(v):=#{uV:uv}\deg(v):=\#\{u\in V:u\sim v\}. A graph H(V,E)H(V^{\prime},E^{\prime}) is called a subgraph of G(V,E)G(V,E) if VVV^{\prime}\subseteq V and EEE^{\prime}\subseteq E. A subgraph HH is called an induced subgraph of GG if E={{u,v}E:u,vV}E^{\prime}=\{\{u,v\}\in E:u,v\in V^{\prime}\}, and is denoted by G(V)G(V^{\prime}). For A,BVA,B\subseteq V^{\prime}, let dH(A,B):=inf{n:Av0vnB}d_{H}(A,B):=\inf\{n:A\ni v_{0}\sim\cdots v_{n}\in B\} be the combinatorial distance between AA and BB in subgraph H(V,E)H(V^{\prime},E^{\prime}). If H=GH=G, we omit the subscript, i.e. d(A,B):=dG(A,B)d(A,B):=d_{G}(A,B). If A={u}A=\{u\} or B={v}B=\{v\} is a single vertex subset, we will omit the braces, i.e. dH(u,v):=dH({u},{v})d_{H}(u,v):=d_{H}(\{u\},\{v\}).

A walk WW of length ll is a sequence of vertices v0,v1,,vlv_{0},v_{1},\cdots,v_{l} such that vivi+1v_{i}\sim v_{i+1} for i=0,1,,l1i=0,1,\cdots,l-1. We write W=v0v1vlW=v_{0}v_{1}\cdots v_{l} for short. If v0=vlv_{0}=v_{l}, we call it a closed walk. Moreover, a closed walk W=v0v1vlW=v_{0}v_{1}\cdots v_{l} is called a cycle if vivjv_{i}\neq v_{j} for all 1i,jl1\leq i,j\leq l. We also regard a cycle C=v0v1vlC=v_{0}v_{1}\cdots v_{l} as a subgraph with vertex set V(C)={v0,v1vl1}V(C)=\{v_{0},v_{1}\cdots v_{l-1}\} and edge set E(C)={{vi,vi+1}:i=0,1,l1}E(C)=\{\{v_{i},v_{i+1}\}:i=0,1,\cdots l-1\}. A walk W=v0v1vlW=v_{0}v_{1}\cdots v_{l} is called a path if vivjv_{i}\neq v_{j} for all 0i,jl0\leq i,j\leq l. We also regard a path P=v0v1vlP=v_{0}v_{1}\cdots v_{l} as a subgraph with vertex set V(P)={v0,v1vl}V(P)=\{v_{0},v_{1}\cdots v_{l}\} and edge set E(P)={{vi,vi+1}:i=0,1,l1}E(P)=\{\{v_{i},v_{i+1}\}:i=0,1,\cdots l-1\}.

Let HH be a subgraph of GG. We say a path P=v0v1vlP=v_{0}v_{1}\cdots v_{l} is geodesic in HH if dH(v0,vl)=ld_{H}(v_{0},v_{l})=l. It is easy to know that P=v0v1vlP=v_{0}v_{1}\cdots v_{l} is a geodesic path of HH if and only if dH(vi,vj)=|ij|d_{H}(v_{i},v_{j})=|i-j| for all 0i,jl0\leq i,j\leq l. We say HH is a geodesic subgraph if all geodesic paths in HH are also geodesic paths in GG. It is not hard to verify that HH is a geodesic subgraph of GG if and only if dH(u,v)=dG(u,v)d_{H}(u,v)=d_{G}(u,v) for all u,vHu,v\in H. A diameter path of GG is a longest geodesic path PP in GG. We let diam(G)\mathrm{diam}(G) denote the length of a diameter path in GG, i.e. diam(G)=sup{d(u,v):u,vV}\mathrm{diam}(G)=\sup\{d(u,v):u,v\in V\}.

A graph G(V,E)G(V,E) is called rooted at pp if a vertex pVp\in V is specially appointed, denoted by (G,p)(G,p). We define r:V0r:V\to\mathbb{N}^{0} via r(u)=d(p,u)r(u)=d(p,u). If r(u)>0r(u)>0, there always exists a vuv\sim u such that r(v)=r(u)1r(v)=r(u)-1 and a geodesic path pvup\cdots vu with length equals to r(u)r(u). Given a geodesic path P=v0v1vlP=v_{0}v_{1}\cdots v_{l} in GG, we always regard GG as a graph rooted at v0v_{0} if not specified.

Let G(V,E)G(V,E) be a graph with degree at most 3, P=v0v1vlP=v_{0}v_{1}\cdots v_{l} be a geodesic path. For all 1il11\leq i\leq l-1, viv_{i} has at most one extra neighbor as vi1v_{i-1} and vi+1v_{i+1} are two neighbors of viv_{i}. We define a state function s:{vi:1il1}{2,3,30,3+}s:\{v_{i}:1\leq i\leq l-1\}\to\{2,3^{-},3^{0},3^{+}\} via

s(vi)={2,ifdeg(vi)=2,3,if vi has an extra neighbor ui and r(ui)<r(vi),30,if vi has an extra neighbor ui and r(ui)=r(vi),3+,if vi has an extra neighbor ui and r(ui)>r(vi).\displaystyle\begin{split}s(v_{i})=\begin{cases}2,\ \ \ \ \ \ \ \ \ &\text{if}\deg(v_{i})=2,\\ 3^{-},&\text{if }v_{i}\text{ has an extra neighbor }u_{i}\text{ and }r(u_{i})<r(v_{i}),\\ 3^{0},&\text{if }v_{i}\text{ has an extra neighbor }u_{i}\text{ and }r(u_{i})=r(v_{i}),\\ 3^{+},&\text{if }v_{i}\text{ has an extra neighbor }u_{i}\text{ and }r(u_{i})>r(v_{i}).\end{cases}\end{split}

As figures are quite important in the discussions of this paper, we explain the drawing style here. If not specified, a horizontal segment with vertices represents a geodesic path as shown in (a) of the following figure. If viv_{i} has an extra neighbor uiu_{i} and r(ui)<r(vi)r(u_{i})<r(v_{i}), i.e. s(vi)=3s(v_{i})=3^{-}, we draw the figure as shown in (b) such that uiu_{i} has the same abscissa with vi1v_{i-1} as r(ui)=r(vi1)r(u_{i})=r(v_{i-1}). Figure (c) and (d) show the drawings when r(ui)=r(vi)r(u_{i})=r(v_{i}) and r(ui)>r(vi)r(u_{i})>r(v_{i}) respectively.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 5.

2.2. Ricci curvature on weighted graphs

Let

w:E(0,),{x,y}w(x,y)=w(y,x)w:E\to\mathbb{(}0,\infty),\ \{x,y\}\mapsto w(x,y)=w(y,x)

be the edge weight function,

m:V(0,),xm(x)m:V\to(0,\infty),\ x\mapsto m(x)

be the vertex weight function. We call the quadruple G=(V,E,m,w)G=(V,E,m,w) a weighted graph. Unless otherwise specified, we always regard a graph G(V,E)G(V,E) as a combinatorial graph, i.e. w1w\equiv 1 and m1m\equiv 1. Our results also hold for normalized graphs, i.e. w1w\equiv 1 and m(u)=deg(u)m(u)=\deg(u) for all uVu\in V; see Theorem 2.5.

We define the Laplacian Δ:VV\Delta:\mathbb{R}^{V}\to\mathbb{R}^{V} via

Δf(u):=1m(u)vuw(u,v)(f(v)f(u)).\Delta f(u):=\frac{1}{m(u)}\sum_{v\sim u}w(u,v)(f(v)-f(u)).

A function ff on GG is called harmonic if Δf0\Delta f\equiv 0.

For a weighted graph G(V,E,w,m)G(V,E,w,m), let Deg(u):=vuw(u,v)m(u)\mathrm{Deg}(u):=\sum_{v\sim u}\frac{w(u,v)}{m(u)}. We define a probability measure

μuε(v)={1εDeg(u):v=u;εw(u,v)/m(u):otherwise.\displaystyle\begin{split}\mu_{u}^{\varepsilon}(v)=\begin{cases}1-\varepsilon\mathrm{Deg}(u)\ \ \ \ \ \ \ &:v=u;\\ \varepsilon w(u,v)/m(u)\ &:\text{otherwise}.\end{cases}\end{split}

Particularly, for combinatorial graphs, we have

μuε(v)={1εdeg(u):v=u;ε:otherwise.\displaystyle\begin{split}\mu_{u}^{\varepsilon}(v)=\begin{cases}1-\varepsilon\deg(u)\ \ \ \ \ \ \ &:v=u;\\ \varepsilon&:\text{otherwise}.\end{cases}\end{split}

For normalized graphs, we have

μuε(v)={1ε:v=u;ε/deg(u):otherwise.\displaystyle\begin{split}\mu_{u}^{\varepsilon}(v)=\begin{cases}1-\varepsilon\ \ \ \ \ \ \ \ \ \ &:v=u;\\ \varepsilon/\deg(u)\ &:\text{otherwise}.\end{cases}\end{split}
Definition 2.1.

(Wasserstein distance) Let G(V,E)G(V,E) be a graph, μ\mu and ν\nu be two probability measures on GG. The Wasserstein distance W(μ,ν)W(\mu,\nu) is given by

W(μ,ν):=infAu,vVA(u,v)d(u,v),W(\mu,\nu):=\inf_{A}\sum_{u,v\in V}A(u,v)d(u,v),

where AA is a coupling between μ\mu and ν\nu, i.e. a mapping A:V×V[0,1]A:V\times V\to[0,1] s.t.

vVA(u,v)=μ(u) and uVA(u,v)=ν(v).\sum_{v\in V}A(u,v)=\mu(u)\text{ and }\sum_{u\in V}A(u,v)=\nu(v).
Definition 2.2.

For any vertices u,vG(V,E,w,m)u,v\in G(V,E,w,m), the ε\varepsilon-Ollivier-Ricci curvature κε(u,v)\kappa_{\varepsilon}(u,v) is defined via

κε(u,v):=1W(μuε,μvε)d(u,v).\kappa_{\varepsilon}(u,v):=1-\frac{W(\mu_{u}^{\varepsilon},\mu_{v}^{\varepsilon})}{d(u,v)}.

Specifically, the Ollivier Ricci curvature on a normalized graph is defined via

κO(u,v):=κ1(u,v).\kappa^{O}(u,v):=\kappa_{1}(u,v).

The Lin-Lu-Yau-Ollivier Ricci curvature κ(u,v)\kappa(u,v) is defined via

κ(u,v):=limε0+1εκε(u,v).\kappa(u,v):=\lim_{\varepsilon\to 0^{+}}\frac{1}{\varepsilon}\kappa_{\varepsilon}(u,v).

For fixed vertices u,vEu,v\in E, it was shown in [BCL+18] that the function εκε(u,v)\varepsilon\mapsto\kappa_{\varepsilon}(u,v) is concave and piecewise linear, so the limitation exists and κ\kappa is well defined. There are also equivalent limit-free definitions via coupling function or Laplacian on graphs; see [BHLY20, MW19]. When we say Ricci curvature or curvature in the following, we always mean the Lin-Lu-Yau-Ollivier Ricci curvature defined above.

For combinatorial graphs, Münch and Wojciechowsk [MW19, Theorem 2.6] gave an easier way to calculate Ricci curvatures on edges as shown in the following theorem.

Theorem 2.3.

(Transport and combinatorial graphs) Let G=(V,E)G=(V,E) be a combinatorial graph and let uvu\sim v be adjacent vertices. Let Buv:=B1(u)B1(v)B_{uv}:=B_{1}(u)\cap B_{1}(v), Buv:=B1(u)B1(v)B_{u}^{v}:=B_{1}(u)\setminus B_{1}(v) and Bvu:=B1(v)B1(u)B_{v}^{u}:=B_{1}(v)\setminus B_{1}(u). Let Φuv:={ϕ:D(ϕ)BuvR(ϕ)Bvu:ϕbijective}\Phi_{uv}:=\{\phi:D(\phi)\subseteq B_{u}^{v}\to R(\phi)\subseteq B_{v}^{u}\ :\phi\ bijective\}. For ϕΦuv\phi\in\Phi_{uv} write D(ϕ)c:=BuvD(ϕ)D(\phi)^{c}:=B_{u}^{v}\setminus D(\phi) and R(ϕ)c:=BvuR(ϕ)R(\phi)^{c}:=B_{v}^{u}\setminus R(\phi). Then

κ(u,v)=#BuvinfϕΦuv(#D(ϕ)c+#R(ϕ)c+wD(ϕ)[d(w,ϕ(w))1]).\kappa(u,v)=\#B_{uv}-\inf_{\phi\in\Phi_{uv}}\left(\#D(\phi)^{c}+\#R(\phi)^{c}+\sum_{w\in D(\phi)}[d(w,\phi(w))-1]\right).
Refer to caption
Figure 6.

We always suppose #D(ϕ)c=0\#D(\phi)^{c}=0 or #R(ϕ)c=0\#R(\phi)^{c}=0 when applying Theorem 2.3. If ϕ\phi is an optimal map such that uD(ϕ)cu\in D(\phi)^{c} and vR(ϕ)cv\in R(\phi)^{c}, we extend ϕ\phi by adding an extra map ϕ(u)=v\phi(u)=v. It is easy to verify that the new ϕ\phi is still an optimal map. So our hypothesis is fair. The following is a direct corollary of Theorem 2.3.

Corollary 2.4.

Let G=(V,E)G=(V,E) be a combinatorial graph and let uvu\sim v be adjacent vertices. Let Nw:=B1(w){u,v}N_{w}:=B_{1}(w)\setminus\{u,v\} be the set of extra neighbors of ww, w{u,v}w\in\{u,v\}. If |Nu|+|Nv|3|N_{u}|+|N_{v}|\geq 3 and d(Nu,Nv)3d(N_{u},N_{v})\geq 3, we have κ(u,v)<0\kappa(u,v)<0.

Let κ(G):=inf{u,v}Eκ(u,v)\kappa(G):=\inf_{\{u,v\}\in E}\kappa(u,v) be the Ricci curvature of GG and deg(G):=supuVdeg(u)\deg(G):=\sup_{u\in V}\deg(u) the maximum degree of GG. The class of graphs we will classify in this paper is defined by

𝒢:={G(V,E):deg(G)3,κ(G)0,diam(G)6}.\mathcal{G}:=\{G(V,E):\deg(G)\leq 3,\kappa(G)\geq 0,\mathrm{diam}(G)\geq 6\}.

Cushing et al. [CKL+19] provide a powerful online tool, the Graph Curvature Calculator, to calculate various of curvatures on graphs. Readers can find it on https://www.mas.ncl.ac.uk/graph-curvature/. By using this tool, it is convenient to check that the Ricci curvatures of graphs in Figure 1, 2 and 3 are nonnegative. We also use it to prove the following theorems.

Theorem 2.5.

Let G=(V,E)G=(V,E) be a graph with deg(G)3\deg(G)\leq 3. Let GCG_{C} be the combinatorial graph and GNG_{N} be the normalized graph obtained by equipping GG with combinatorial weight and normalized weight respectively. Then κ(GC)0\kappa(G_{C})\geq 0 if and only if κ(GN)0\kappa(G_{N})\geq 0.

Proof.

For any uvEuv\in E, let κC(uv)\kappa_{C}(uv) be the curvature of uvuv when GG is equipped with combinatorial weight and κN(uv)\kappa_{N}(uv) be the curvature of uvuv when GG is equipped with normalized weight. If deg(u)=1\deg(u)=1 or deg(v)=1\deg(v)=1, we have κC(uv)0\kappa_{C}(uv)\geq 0 and κN(uv)0\kappa_{N}(uv)\geq 0 by a direct calculation. If deg(u)=deg(v)\deg(u)=\deg(v), we have κC(uv)=deg(u)κN(uv)\kappa_{C}(uv)=\deg(u)\cdot\kappa_{N}(uv) by the definition of Ricci curvature. So we only need to consider the case that deg(u)=2\deg(u)=2 and deg(v)=3\deg(v)=3, i.e. uu has an extra neighbor uu^{\prime}, vv has two extra neighbors vv^{\prime} and v′′v^{\prime\prime}. Then κC(uv)\kappa_{C}(uv) and κN(uv)\kappa_{N}(uv) depend only on the distances d(u,v)d(u^{\prime},v^{\prime}) and d(u,v′′)d(u^{\prime},v^{\prime\prime}). By a direct calculation, we have κC(uv)<0\kappa_{C}(uv)<0 \Leftrightarrow d(u,v)=d(u,v′′)=3d(u^{\prime},v^{\prime})=d(u^{\prime},v^{\prime\prime})=3 \Leftrightarrow κN(uv)<0\kappa_{N}(uv)<0. This completes the proof. ∎

It is noteworthy that Theorem 2.5 does not hold for graphs with maximum degree larger than 3. The minimum Ricci curvatures of the following graphs are 0, 0.5, -0.133 as normalized graphs and -1, -1, 0 as combinatorial graphs respectively. 111Graph (a) built by Bai-Lu-Yau [BLY21] is Ricci-flat as a normalized graph. Graphs (b) and (c) are pointed out to us by Florentin Münch.

Refer to caption
Refer to caption
Refer to caption
Figure 7.
Theorem 2.6.

Let GG be a normalized graph with nonnegative Ollivier Ricci curvature, maximum degree at most 3 and diameter at least 6. If GG is finite, then it is isometric to a finite path, a simple cycle, a prism graph, a Möbius ladder or a quasi-ladder as shown in the following figure. If GG is infinite, then it is isometric to (a), (b), (c), (d), (e), (f), (g) or (h) of Figure 3.

Refer to caption
Figure 8. Quasi-ladder
Proof.

For normalized graphs, κ0(u,v)0\kappa_{0}(u,v)\equiv 0 for all u,vVu,v\in V. If κO(u,v)0\kappa^{O}(u,v)\geq 0, we always have κ(u,v)0\kappa(u,v)\geq 0, as the function εκε(u,v)\varepsilon\mapsto\kappa_{\varepsilon}(u,v) is concave and piecewise linear; see [BCL+18, Theorem 1.1]. This indicates that all normalized graphs with nonnegative Ollivier Ricci curvature are also graphs with nonnegative Lin-Lu-Yau-Ollivier Ricci curvature. We check the Ollivier Ricci curvatures of graphs in Figure 1, 2, 3 by using the Graph Curvature Calculator and get the result. ∎

3. The local structures

In this section, we study the local structure on a geodesic path of a graph G𝒢G\in\mathcal{G}.

Lemma 3.1.

Let P=v0vivi+1vi+2vi+3vlP=v_{0}\cdots v_{i}v_{i+1}v_{i+2}v_{i+3}\cdots v_{l} be a geodesic path of G𝒢G\in\mathcal{G}.

Case (2,2)(2,2): If (s(vi+1),s(vi+2))=(2,2)(s(v_{i+1}),s(v_{i+2}))=(2,2), then GG contains the following as a subgraph.

Refer to caption
Figure 9. Case (2,2)(2,2)

Case (2,3)(2,3^{-}): If (s(vi+1),s(vi+2))=(2,3)(s(v_{i+1}),s(v_{i+2}))=(2,3^{-}), then GG contains one of the followings as a subgraph.

Refer to caption
Refer to caption
Refer to caption
Figure 10. Case (2,3)(2,3^{-})

Case (2,30)(2,3^{0}): If (s(vi+1),s(vi+2))=(2,30)(s(v_{i+1}),s(v_{i+2}))=(2,3^{0}), then GG contains the following as a subgraph.

Refer to caption
Figure 11. Case (2,30)(2,3^{0})

Case (2,3+)(2,3^{+}): The situation (s(vi+1),s(vi+2))=(2,3+)(s(v_{i+1}),s(v_{i+2}))=(2,3^{+}) can not exist.

Case (3,2)(3^{-},2): The situation (s(vi+1),s(vi+2))=(3,2)(s(v_{i+1}),s(v_{i+2}))=(3^{-},2) can not exist.

Case (30,2)(3^{0},2): If (s(vi+1),s(vi+2))=(30,2)(s(v_{i+1}),s(v_{i+2}))=(3^{0},2), then GG contains the following as a subgraph.

Refer to caption
Figure 12. Case (30,2)(3^{0},2)

Case (3+,2)(3^{+},2): If (s(vi+1),s(vi+2))=(3+,2)(s(v_{i+1}),s(v_{i+2}))=(3^{+},2), then GG contains one of the followings as a subgraph.

Refer to caption
Refer to caption
Refer to caption
Figure 13. Case (3+,2)(3^{+},2)

Case (3,3)(3^{-},3^{-}): If (s(vi+1),s(vi+2))=(3,3)(s(v_{i+1}),s(v_{i+2}))=(3^{-},3^{-}), then GG contains one of the followings as a subgraph.

Refer to caption
Refer to caption
Figure 14. Case (3,3)(3^{-},3^{-})

Case (3,30)(3^{-},3^{0}): The situation (s(vi+1),s(vi+2))=(3,30)(s(v_{i+1}),s(v_{i+2}))=(3^{-},3^{0}) can not exist.

Case (3,3+)(3^{-},3^{+}): The situation (s(vi+1),s(vi+2))=(3,3+)(s(v_{i+1}),s(v_{i+2}))=(3^{-},3^{+}) can not exist.

Case (30,3)(3^{0},3^{-}): If (s(vi+1),s(vi+2))=(30,3)(s(v_{i+1}),s(v_{i+2}))=(3^{0},3^{-}), then GG contains one of the followings as a subgraph.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 15. Case (30,3)(3^{0},3^{-})

Case (30,30)(3^{0},3^{0}): If (s(vi+1),s(vi+2))=(30,30)(s(v_{i+1}),s(v_{i+2}))=(3^{0},3^{0}), then GG contains one of the followings as a subgraph.

Refer to caption
Refer to caption
Figure 16. Case (30,30)(3^{0},3^{0})

Case (30,3+)(3^{0},3^{+}): The situation (s(vi+1),s(vi+2))=(30,3+)(s(v_{i+1}),s(v_{i+2}))=(3^{0},3^{+}) can not exist.

Case (3+,3)(3^{+},3^{-}): If (s(vi+1),s(vi+2))=(3+,3)(s(v_{i+1}),s(v_{i+2}))=(3^{+},3^{-}), then GG contains one of the followings as a subgraph.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 17. Case (3+,3)(3^{+},3^{-})

Case (3+,30)(3^{+},3^{0}): If (s(vi+1),s(vi+2))=(3+,30)(s(v_{i+1}),s(v_{i+2}))=(3^{+},3^{0}), then GG contains one of the followings as a subgraph.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 18. Case (3+,30)(3^{+},3^{0})

Case (3+,3+)(3^{+},3^{+}): If (s(vi+1),s(vi+2))=(3+,3+)(s(v_{i+1}),s(v_{i+2}))=(3^{+},3^{+}), then GG contains one of the followings as a subgraph.

Refer to caption
Refer to caption
Figure 19. Case (3+,3+)(3^{+},3^{+})
Proof.

We only show the proof of case (2,3)(2,3^{-}), proofs for other cases are similar. Without loss of generality, let i=0i=0. Then GG has a subgraph as shown in the following figure with deg(v1)=2\deg(v_{1})=2.

Refer to caption
Figure 20.

By Theorem 2.3 we have

κ(v1,v2)=#Bv1v2infϕΦv1v2(#D(ϕ)c+#R(ϕ)c+wD(ϕ)[d(w,ϕ(w))1]),\kappa(v_{1},v_{2})=\#B_{v_{1}v_{2}}-\inf_{\phi\in\Phi_{v_{1}v_{2}}}(\#D(\phi)^{c}+\#R(\phi)^{c}+\sum_{w\in D(\phi)}[d(w,\phi(w))-1]),

where Bv1v2={v1,v2}B_{v_{1}v_{2}}=\{v_{1},v_{2}\}, Φv1v2={all maps from {v0} to {v3,u2}}\Phi_{v_{1}v_{2}}=\{\text{all maps from $\{v_{0}\}$ to $\{v_{3},u_{2}\}$}\}, #D(ϕ)c=0\#D(\phi)^{c}=0, and #R(ϕ)c=1\#R(\phi)^{c}=1. Then we have

0κ(v1,v2)=2infϕ(v0){v3,u2}(0+1+d(v0,ϕ(v0))1)=max{2d(v0,v3),2d(v0,u2)}.\begin{split}0&\leq\kappa(v_{1},v_{2})\\ &=2-\inf_{\phi(v_{0})\in\{v_{3},u_{2}\}}(0+1+d(v_{0},\phi(v_{0}))-1)\\ &=\max\{2-d(v_{0},v_{3}),2-d(v_{0},u_{2})\}.\end{split}

As d(v0,v3)=3d(v_{0},v_{3})=3, we have d(v0,u2)=1d(v_{0},u_{2})=1 or d(v0,u2)=2d(v_{0},u_{2})=2. If d(v0,u2)=1d(v_{0},u_{2})=1, GG contain (a) of Figure 10 as a subgraph. If d(v0,u2)=2d(v_{0},u_{2})=2, GG contains (b) or (c) of Figure 10 as a subgraph. This completes the proof. ∎

By applying case (,3+)(\cdot,3^{+}) or case (3,)(3^{-},\cdot) of Lemma 3.1 repeatedly, we get the following corollary.

Corollary 3.2.

Let P=v0v1vlP=v_{0}v_{1}\cdots v_{l} be a geodesic path in G𝒢G\in\mathcal{G}, l4l\geq 4.

1) If there exists a i{1,2,,l1}i\in\{1,2,\cdots,l-1\} such that s(vi)=3+s(v_{i})=3^{+}, then s(vj)=3+s(v_{j})=3^{+} for all 1ji1\leq j\leq i and GG has a subgraph (b) or (c) of the following figure.

2) If there exists a i{2,,l1}i\in\{2,\cdots,l-1\} such that s(vi)=3s(v_{i})=3^{-}, then s(vj)=3s(v_{j})=3^{-} for all ijl1i\leq j\leq l-1 and GG has a subgraph (e) or (f) of the following figure.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 21.

As a supplement to Lemma 3.1, we explore the local structures on long geodesic paths in the following two lemmas,

Lemma 3.3.

Let P=v0vivi+1vi+2vi+3vlP=v_{0}\cdots v_{i}v_{i+1}v_{i+2}v_{i+3}\cdots v_{l} be a geodesic path with l6l\geq 6. If GG contains (b) of Figure 16 as a subgraph, then GG is isometric to the particular graph (e) of Figure 1.

Proof.

If i=2i=2, GG contains a subgraph (a) of the following figure. By Corollary 3.2, GG contains a subgraph (b). By applying case (3+,30)(3^{+},3^{0}) of Lemma 3.1 on v1v2v3v4v_{1}v_{2}v_{3}v_{4}, GG contains a subgraph (c). By applying case (3+,30)(3^{+},3^{0}) of Lemma 3.1 on v1u1u3u5v_{1}u_{1}u_{3}u_{5}, GG contains a subgraph (d). Moreover, deg(v0)=deg(v6)=1\deg(v_{0})=\deg(v_{6})=1 by Corollary 2.4. So GG is isometric to graph (d) of the following figure, which is isometric to (e) of Figure 1. If i2i\neq 2, it is impossible to get a graph with nonnegative Ricci curvature by a similar argument. This proves the lemma.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 22.

Lemma 3.4.

Let P=v0vivi+1vi+2vi+3vlP=v_{0}\cdots v_{i}v_{i+1}v_{i+2}v_{i+3}\cdots v_{l} be a geodesic path with l6l\geq 6. Then none of the followings can be a subgraph of GG: (c) in Figure 10; (c) in Figure 13; (c), (d) and (e) in Figure 15; (c), (d), (e), (f) and (g) in Figure 17; (c), (d) and (e) in Figure 18.

Proof.

We only show (c) of Figure 13 can not be a subgraph of GG. Proofs for other cases are similar. We prove by contradiction. Suppose i1i\leq 1, then i+5li+5\leq l and GG contains a subgraph (a) of the following figure with deg(vi+2)=2\deg(v_{i+2})=2 by Corollary 3.2. This is impossible as deg(ui+3)3\deg(u_{i+3})\leq 3 and ui+3u_{i+3} must has an extra neighbor wi+3w_{i+3} with r(wi+3)<r(ui+3)r(w_{i+3})<r(u_{i+3}). Suppose 2il42\leq i\leq l-4. Then GG contains a subgraph (b) of the following figure with deg(vi+2)=2\deg(v_{i+2})=2 by Corollary 3.2. Applying case (30,3)(3^{0},3^{-}) of Lemma 3.1 on P=v0wi+3ui+3vi+3vi+4P^{\prime}=v_{0}\cdots w_{i+3}u_{i+3}v_{i+3}v_{i+4}\cdots, we get a contradiction. Suppose i=l3i=l-3. Then GG contains a subgraph (c) of the following figure with deg(vl1)=2\deg(v_{l-1})=2 by Corollary 3.2. We have κ(ul,ul2)<0\kappa(u_{l},u_{l-2})<0 by Theorem 2.3. This completes the proof.

Refer to caption
Refer to caption
Refer to caption
Figure 23.

4. Classification of graphs

Let G𝒢G\in\mathcal{G} be a finite graph and P=v0v1vlP=v_{0}v_{1}\cdots v_{l} be a diameter of GG. Then PP is a geodesic path. We explore all the available structures of GG separately by the following four cases.
Case (i): For all 1il11\leq i\leq l-1, s(vi)=2s(v_{i})=2. We discuss this case in Theorem 4.3.
Case (ii): There exists a viv_{i} such that s(vi)=30s(v_{i})=3^{0}. We discuss this case in Theorem 4.5.
Case (iii): There exists a viv_{i} such that s(vi)=3s(v_{i})=3^{-}. We discuss this case in Theorem 4.6.
Case (iv): There exists a viv_{i} such that s(vi)=3+s(v_{i})=3^{+}. We discuss this case in Theorem 4.7.

We firstly introduce two lemmas.

Lemma 4.1.

Let G(V,E)G(V,E) be a finite graph. Suppose V=ABC1C2V=A\sqcup B\sqcup C_{1}\sqcup C_{2} satisfying:
(1) E(A,B)=E(C1,C2)=E(A,B)=E(C_{1},C_{2})=\emptyset;
(2) The induced subgraphs G(C1)G(C_{1}) and G(C2)G(C_{2}) are complete graphs, i.e. for all u,vCiu,v\in C_{i}, uvu\sim v in G(Ci)G(C_{i}), i=1,2i=1,2;
(3) The induced subgraphs GA=G(AC1C2)G_{A}=G(A\sqcup C_{1}\sqcup C_{2}) and GB=G(BC1C2)G_{B}=G(B\sqcup C_{1}\sqcup C_{2}) are connected graphs.
Then GG contains a geodesic cycle CC with |C|dGA(C1,C2)+dGB(C1,C2)|C|\geq d_{G_{A}}(C_{1},C_{2})+d_{G_{B}}(C_{1},C_{2}).

Refer to caption
Refer to caption
Figure 24.
Proof.

Let 𝒞=\mathcal{C}={WW is a closed walk : W=P1PAP2PBW=P_{1}P_{A}P_{2}P_{B}, where P1P_{1}, P2P_{2}, PAP_{A} and PBP_{B} are walks in C1C_{1}, C2C_{2}, AA and BB respectively}. Let C𝒞C\in\mathcal{C} with shortest length, then it is a cycle. Moreover, we can prove that CC is a geodesic cycle. Suppose not, there are u,vCu,v\in C such that dG(u,v)<dC(u,v)d_{G}(u,v)<d_{C}(u,v). We only prove the case that uAu\in A and vBv\in B, proofs for other cases are similar.

Let PGP_{G} be a shortest walk between uu and vv in GG and PCiP_{C_{i}} be a shortest walk between uu and vv in CC crossing CiC_{i}, i=1, 2i=1,\ 2. Note that PC1P_{C_{1}} and PC2P_{C_{2}} are disjoint paths from uu to vv and CC is the union of them.

Next, we claim that PGP_{G} has one of the following forms: PAP1PAP2PBP_{A}P_{1}P_{A}^{\prime}P_{2}P_{B}, PAP1PBP2PBP_{A}P_{1}P_{B}P_{2}P_{B}^{\prime}, PAP2PAP1PBP_{A}P_{2}P_{A}^{\prime}P_{1}P_{B}, PAP2PBP1PBP_{A}P_{2}P_{B}P_{1}P_{B}^{\prime}, PAP1PBP_{A}P_{1}P_{B} or PAP2PBP_{A}P_{2}P_{B}. In fact, it is easy to see that PGC1P_{G}\cap C_{1} is connected in PGP_{G} and |PGC1|2|P_{G}\cap C_{1}|\leq 2 as C1C_{1} is complete and PGP_{G} is a shortest walk. The same conclusion holds for PGC2P_{G}\cap C_{2}. So the claim holds.

Now, if PG=PAP1PAP2PBP_{G}=P_{A}P_{1}P_{A}^{\prime}P_{2}P_{B}, then PGP_{G} and PC1P_{C_{1}} form a closed walk. By connecting C1C_{1} part in PGP_{G} and C1C_{1} part in PC1P_{C_{1}}, we get a new closed walk C𝒞C^{\prime}\in\mathcal{C} with a shorter length. This contradicts to the fact that CC is a shortest walk. Proofs for other forms are similar. Then we proved dG(u,v)=dC(u,v)d_{G}(u,v)=d_{C}(u,v) for all u,vCu,v\in C, i.e. CC is a geodesic cycle. The inequality |C|dGA(C1,C2)+dGB(C1,C2)|C|\geq d_{G_{A}}(C_{1},C_{2})+d_{G_{B}}(C_{1},C_{2}) is obtained directly from the construction of CC. ∎

Lemma 4.2.

Suppose G𝒢G\in\mathcal{G} contains a geodesic cycle CC with |C|8|C|\geq 8, then |C|11|C|\geq 11 and GG must be a cycle, a prism graph or a Möbius ladder.

Proof.

Suppose |C|=8|C|=8. We assume that there exists a vertex, say v0v_{0}, has an extra neighbor u0u_{0}, otherwise G=CG=C. Note that every path PP in CC with |P|5|P|\leq 5 is geodesic in GG.

We claim that GG is a prism or Möbius ladder if u0v2u_{0}\sim v_{2}. By applying local structure (3,)(3^{-},\cdot) on v1v2v3v4v_{1}v_{2}v_{3}v_{4} in (G,v0)(G,v_{0}), we get that GG contains (a) or (b) of the following figure as a subgraph. By symmetric, GG has a subgraph (c) or (d) in the following figure. By doing this repeatedly, we get that GG has a subgraph (e). Finally we get a prism graph or a Möbius ladder.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 25.

We claim that u0u_{0} is not a neighbor of v1v_{1}. Suppose not, GG has a subgraph (a) of the following figure. Applying local structure (30,)(3^{0},\cdot) on v0v1v2v3v_{0}v_{1}v_{2}v_{3} in (G,v0)(G,v_{0}), we know that GG has a subgraph (b) or (c) of the following figure. This is impossible by the symmetry of the geodesic cycle. So our claim is true.

Refer to caption
Refer to caption
Refer to caption
Figure 26.

The previous claim also shows u0v7u_{0}\nsim v_{7}. By applying local structure (3+,)(3^{+},\cdot) on v7v0v1v2v_{7}v_{0}v_{1}v_{2} in (G,v7)(G,v_{7}), we know that GG has one of the following subgraphs.

Refer to caption
Refer to caption
Refer to caption
Figure 27.

By applying local structures on CC as we did previously, we get the following results. If GG has a subgraph (c) in Figure 27, then GG is isomorphic to (a) of Figure 28. Otherwise GG must be isometric to (b), (c) or (d) of the following figure if GG is not a prism graph or Möbius ladder.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 28.

Note that the diameters of all the possible graphs we get are less than 6, thus they are not elements of 𝒢\mathcal{G}. It follows that |C|8|C|\neq 8. We also have |C|9,10|C|\neq 9,10 with similar arguments. When we do the previous steps on a geodesic cycle CC with |C|11|C|\geq 11, the graphs with the same forms shown in Figure 28 contain edges with negative curvature. This indicates that GG must be a cycle, prism graph or Möbius ladder.

Theorem 4.3.

Let P=v0v1vlP=v_{0}v_{1}\cdots v_{l} be a diameter of G𝒢G\in\mathcal{G} with l6l\geq 6. If for all 1il11\leq i\leq l-1, s(vi)=2s(v_{i})=2, then G=PG=P or isometric to a cycle.

Proof.

Let GB:=G(V{v0,v1,,vl})G_{B}:=G(V\setminus\{v_{0},v_{1},\cdots,v_{l}\}). If v0v_{0} is not connected to vlv_{l} via GBG_{B}, then G=PG=P, otherwise we will get a longer diameter. If v0v_{0} is connected to vlv_{l} via GBG_{B}. By setting C1={v0}C_{1}=\{v_{0}\} and C2={vl}C_{2}=\{v_{l}\} and applying Lemma 4.1, we get a geodesic cycle CC with |C|12|C|\geq 12. Then the theorem holds by Lemma 4.2. ∎

Lemma 4.4.

Let P=v0v1vlP=v_{0}v_{1}\cdots v_{l}, l6l\geq 6 be a diameter of G𝒢G\in\mathcal{G}. If there exists 3il33\leq i\leq l-3 such that s(vi)=30s(v_{i})=3^{0}, then s(vj)=30s(v_{j})=3^{0} for all 3il33\leq i\leq l-3, and GG contains a subgraph shown in the following figure.

Refer to caption
Figure 29.
Proof.

If l=6l=6, it is trivial. Suppose l7l\geq 7 and i+1l3i+1\leq l-3.

Firstly we prove s(vi+1)=30s(v_{i+1})=3^{0}. If s(vi+1)=2s(v_{i+1})=2, by applying local structure (30,2)(3^{0},2) on vi1vivi+1vi+2v_{i-1}v_{i}v_{i+1}v_{i+2} and Corollary 3.2, we know that GG contains the following figure as a subgraph with deg(vi+1)=2\deg(v_{i+1})=2. As r(ui)>0r(u_{i})>0, there is a vertex wiuiw_{i}\sim u_{i} with r(wi)<r(ui)r(w_{i})<r(u_{i}). By applying local structure (30,3+)(3^{0},3^{+}) on wiuiui+2ui+3w_{i}u_{i}u_{i+2}u_{i+3}, we get a contradiction. If s(vi+1)=3s(v_{i+1})=3^{-}, by applying local structure (30,3)(3^{0},3^{-}) on vi1vivi+1vi+2v_{i-1}v_{i}v_{i+1}v_{i+2} and Corollary 3.2, we also get a contradiction. It follows that s(vi+1)=30s(v_{i+1})=3^{0}. By applying local structure (30,30)(3^{0},3^{0}) on vi1vivi+1vi+2v_{i-1}v_{i}v_{i+1}v_{i+2}, we have uiui+1u_{i}\sim u_{i+1}. If i13i-1\geq 3, with the similar argument, we can also prove that s(vi1)=30s(v_{i-1})=3^{0}.

Refer to caption
Figure 30.

By taking the previous steps repeatedly, we complete the proof. ∎

Theorem 4.5.

Let P=v0v1vlP=v_{0}v_{1}\cdots v_{l}, l6l\geq 6, be a diameter of G𝒢G\in\mathcal{G}. If there exists a i{1,2,,l1}i\in\{1,2,\cdots,l-1\} such that s(vi)=30s(v_{i})=3^{0}, then GG is a quasi-ladder or isometric to the particular graph (d) in Figure 1.

Proof.

Let k=min{i:s(vi)=30,1il1}k=\min\{i:s(v_{i})=3^{0},1\leq i\leq l-1\}. By Lemma 4.4, we have k{1,2,3,l2,l1}k\in\{1,2,3,l-2,l-1\}.

Case (i): k=1k=1. GG contains (a) of the following figure as a subgraph. We claim that s(vj)=30s(v_{j})=3^{0} or 33^{-} for all 2jl32\leq j\leq l-3. If s(v2)=2s(v_{2})=2, by applying local structure (30,2)(3^{0},2) on v0v1v2v3v_{0}v_{1}v_{2}v_{3}, we have that GG has a subgraph (b) of the following figure. By applying local structure (30,3+)(3^{0},3^{+}) on v0u0u3u4v_{0}u_{0}u_{3}u_{4}, we get a contradiction. If s(v2)=3+s(v_{2})=3^{+}, we also get a contradiction by applying (30,3+)(3^{0},3^{+}) on v0v1v2v3v_{0}v_{1}v_{2}v_{3}. So s(v2)=30s(v_{2})=3^{0} or 33^{-}.

Refer to caption
Refer to caption
Figure 31.

If s(v2)=30s(v_{2})=3^{0}, we have u0u2u_{0}\sim u_{2} as shown in (a) of the following figure by applying local structure (30,30)(3^{0},3^{0}) on v0v1v2v3v_{0}v_{1}v_{2}v_{3}. The previous argument tells us that s(v3)2,3+s(v_{3})\neq 2,3^{+}. It is not hard to verify that s(v3)3s(v_{3})\neq 3^{-} neither. So s(v3)=30s(v_{3})=3^{0} and u2u3u_{2}\sim u_{3}. By Lemma 4.4, GG has a subgraph (a) of the following figure. Moreover, deg(v0)=2\deg(v_{0})=2, otherwise v0v_{0} has an extra neighbor w0w_{0}. Then deg(w0)=1\deg(w_{0})=1 by Corollary 2.4 and dG(w0,vl)=l+1d_{G}(w_{0},v_{l})=l+1, which is contrary to diam(G)=l\mathrm{diam}(G)=l. If s(v2)=3s(v_{2})=3^{-}, by Corollary 3.2 we have s(vi)=3s(v_{i})=3^{-} for i{2,3,,l1}i\in\{2,3,\cdots,l-1\} and GG has a subgraph (b) of the following figure. Moreover, degG(u0)=2\deg_{G}(u_{0})=2 and u0v1v2vlu_{0}v_{1}v_{2}\cdots v_{l} is also a diameter of GG. We can redraw (b) as shown in (c) of the following figure, which contains (a) as a subgraph. So we only need to consider the situation that GG contains a subgraph (a).

Refer to caption
Refer to caption
Refer to caption
Figure 32.

As we see, the left side of GG is fixed, we start to study the structure of the right side. If deg(vl2)=2\deg(v_{l-2})=2, GG contains a subgraph (a) of the following figure by applying local structure (30,2)(3^{0},2) on vl4vl3vl2vl1v_{l-4}v_{l-3}v_{l-2}v_{l-1}. If deg(vl2)=3\deg(v_{l-2})=3, we have s(vl2)=30s(v_{l-2})=3^{0} and ul3ul2u_{l-3}\sim u_{l-2}. With similar arguments, we classify all possible structures of the right side of GG as shown in the following figure.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 33.

Case (ii): k=2k=2. By applying (,30)(\cdot,3^{0}) on v0v1v2v3v_{0}v_{1}v_{2}v_{3}, GG contains a subgraph (a), (b) or (c) shown in the following figure. Obviously situation (c) is contained in (b) as u0u2v2v3u_{0}u_{2}v_{2}v_{3}\cdots in (c) is also a diameter. We only need to consider the situation that GG has a subgraph (a) or (b). It is easy to verify that deg(v0)=1\deg(v_{0})=1 in (a) and deg(v0)=deg(v1)=deg(u0)=2\deg(v_{0})=\deg(v_{1})=\deg(u_{0})=2 in (b). So the left-hand side is fixed. The right-hand side of GG is the same with that in case (i).

Refer to caption
Refer to caption
Refer to caption
Figure 34.

Case (iii): k=3k=3. If l7l\geq 7, the left side of GG is shown as following. The available structures of the right side are the same with that in case (i). If l=6l=6, GG has an additional available structure shown in (d) of Figure 1.

Refer to caption
Figure 35.

Case (iv): k=l2k=l-2. It is easy to verify this case can not happen.

Case (v): k=l1k=l-1. GG has a subgraph (a) or (b) shown in the following figure. As vlvl1v0v_{l}v_{l-1}\cdots v_{0} is also a diameter, graphs in this case have been discussed in the previous cases.

Refer to caption
Refer to caption
Figure 36.

By previous discussions, we prove that GG must be a quasi-ladder graph with the following structures or isometric to the particular graph (d) in Figure 1.

Refer to caption
(a) Quasi-ladder
Refer to caption
(b) Forms of L
Refer to caption
(c) Forms of R
Figure 37.

Theorem 4.6.

Let P=v0v1vlP=v_{0}v_{1}\cdots v_{l}, l6l\geq 6, be a diameter of G𝒢G\in\mathcal{G}. If for all 1il1,s(vi)30,1\leq i\leq l-1,s(v_{i})\neq 3^{0}, and there exists viv_{i} such that s(vi)=3s(v_{i})=3^{-}, then GG is a prism graph, a Möbius ladder or a quasi-ladder.

Proof.

Let k=min{i:1il1,s(vi)=3}k=\min\{i:1\leq i\leq l-1\ ,\ s(v_{i})=3^{-}\}, then k2k\geq 2.

Case (i): If k=2k=2, GG has a subgraph (a) of the following figure. By Corollary 3.2, GG has a subgraph (b) or (c) of the following figure. By switching u0u_{0} and v1v_{1} in (b), we get (b) is the same with (c). So we only consider the case that GG contains a subgraph (b).

Refer to caption
Refer to caption
Refer to caption
Figure 38.

Let C1=G({v0,v1})C_{1}=G(\{v_{0},v_{1}\}), GA=G({u0,u3,,ul2,v2,v3,vl2})G_{A}=G(\{u_{0},u_{3},\cdots,u_{l-2},v_{2},v_{3}\cdots,v_{l-2}\}), C2=G({vl1,ul1})C_{2}=G(\{v_{l-1},u_{l-1}\}) and GB=G(V{u0,u3,,ul1,v0,v1,vl})G_{B}=G(V\setminus\{u_{0},u_{3},\cdots,u_{l-1},v_{0},v_{1}\cdots,v_{l}\}). If C1C_{1} is connected to C2C_{2} via GBG_{B}, we get a geodesic cycle CC with |C|8|C|\geq 8 by Lemma 4.1 and GG must be a prism graph or Möbius ladder by Lemma 4.2. In the following, we suppose that C1C_{1} is not connected to C2C_{2} via GBG_{B}. With a similar argument in the proof of Theorem 4.5, the left side of GG has one of the following forms.

Refer to caption
Refer to caption
Figure 39.

We start to deal with the right-hand side. If deg(ul1)=2\deg(u_{l-1})=2, we have deg(vl)=1\deg(v_{l})=1 by Corollary 2.4 and the structure of the right-hand side is (a) of the following figure. If ul1u_{l-1} has an extra neighbor wl1w_{l-1}, we have r(ul1)<r(wl1)r(u_{l-1})<r(w_{l-1}). Otherwise, wl1w_{l-1} has an extra neighbor ww^{\prime} and κ(w,wl1)<0\kappa(w^{\prime},w_{l-1})<0 by Corollary 2.4. If deg(wl1)=1\deg(w_{l-1})=1, we have deg(vl)=1\deg(v_{l})=1 and the right-hand side structure is shown in (b) of the following figure. With similar arguments, we find all possible forms of the right side of GG as shown in the following figure.

Refer to caption
Refer to caption
Refer to caption
Refer to caption
Refer to caption
Figure 40.

Case (ii): k=3k=3. GG has a subgraph (a) of the following figure. The left side of GG must be (a) or (b) of the following figure and all possible forms of the right side have been shown in Figure 40.

Refer to caption
Refer to caption
Figure 41.

Case (iii): 4kl24\leq k\leq l-2. GG has a subgraph shown in the following figure which is the same with case (ii) by switching ui2u_{i-2} and vi1v_{i-1}, 4ik4\leq i\leq k.

Refer to caption
Figure 42.

Case (iv): j=l1j=l-1. GG contains a subgraph shown in the following figure which is the same with (a) of Figure 41 if ul2vlu_{l-2}\sim v_{l}, or (b) of Figure 41 if ul2vlu_{l-2}\nsim v_{l} as P=vlvl1v0P^{\prime}=v_{l}v_{l-1}\cdots v_{0} is also a diameter.

Refer to caption
Figure 43.

By previous discussions, we prove that if GG is not a prism graph or a Möbius ladder, then it has one of the following forms, which are isometric to graphs in Figure 45.

Refer to caption
Refer to caption
(a) Forms of L
Refer to caption
(b) Forms of R
Figure 44.
Refer to caption
(a) Quasi-ladder
Refer to caption
(b) Forms of L
Refer to caption
(c) Forms of R
Figure 45.

Theorem 4.7.

Let P=v0v1vlP=v_{0}v_{1}\cdots v_{l}, l6l\geq 6, be a diameter of G𝒢G\in\mathcal{G}. Suppose that for all 1il11\leq i\leq l-1, s(vi)30,s(vi)3s(v_{i})\neq 3^{0},s(v_{i})\neq 3^{-}, and there exists s(vi)=3+s(v_{i})=3^{+}, then GG is a prism graph, a Möbius ladder or a quasi-ladder.

Proof.

In fact we will prove that graphs considered in this theorem have been discussed in the previous theorems. Let k=max{i:1il1,s(vi)=3+}k=\max\{i:1\leq i\leq l-1,\ s(v_{i})=3^{+}\}, then k=l2k=l-2 or l1l-1.

Case (i): k=l2k=l-2. Then deg(vl1)=2\deg(v_{l-1})=2 and GG has a subgraph shown in figure (a). By applying local structure (3+,2)(3^{+},2) on vl3vl2vl1vlv_{l-3}v_{l-2}v_{l-1}v_{l}, we get that GG has a subgraph (b) or (c). We have discussed those graphs by changing P=v0v1vlP=v_{0}v_{1}\cdots v_{l} to P=vlvl1v0P^{\prime}=v_{l}v_{l-1}\cdots v_{0}.

Refer to caption
Refer to caption
Refer to caption
Figure 46.

Case (ii): k=l1k=l-1. GG contains one of the followings as a subgraph, which have been discussed in the previous theorems by changing P=v0v1vlP=v_{0}v_{1}\cdots v_{l} to P=vlvl1v0P^{\prime}=v_{l}v_{l-1}\cdots v_{0}.

Refer to caption
Refer to caption
Figure 47.

Now, we give the proofs of Theorem 1.1 and 1.2.

Proof of Theorem 1.1.

By applying Theorem 4.3, 4.5, 4.6 and 4.7, we obtain the conclusion. ∎

Proof of Theorem 1.2.

As GG is an infinite graph, we can find an infinite geodesic path PP which is isometric to \mathbb{Z} or +\mathbb{Z}^{+}. By applying the arguments used for finite graphs, we get the conclusion. ∎

5. Applications

In this section, as an application of the main results, we establish the following strong Liouville property.

Theorem 5.1.

Let G𝒢G\in\mathcal{G} be an infinite graph, then all positive harmonic functions on GG are constant.

Proof.

By Theorem 1.2, GG is a line, half line, infinite ladder or infinite quasi-ladder. As line and infinite ladder are finitely generated Cayley graphs of Abelian groups, all positive harmonic functions on them are constant; see [CD60]. Besides, Hua and Münch also proved all positive harmonic functions are constant on graphs with two ends and nonnegative Ricci curvature, which provides another proof for the strong Liouville property of line and infinite ladder; see [HM21, Theorem 5.9]. For a half line, there is a stronger conclusion. In fact all harmonic functions on a half line are constant. So, we only consider the quasi-ladder case.

Let ff be a positive harmonic function on GG. If ff is not a constant, there exist i0i\geq 0, such that f(xi)f(yi)f(x_{i})\neq f(y_{i}). Without loss of generality, we assume i=0i=0. By choosing suitable constants aa and bb such that g:=af+bg:=af+b satisfies g(x0)=1g(x_{0})=1 and g(y0)=0g(y_{0})=0, we get a harmonic function gg which is bounded from below or above.

[Uncaptioned image]

For any n+n\in\mathbb{Z}^{+}, we have

g(xn)=13(g(xn1)+g(xn+1)+g(yn)),g(x_{n})=\frac{1}{3}(g(x_{n-1})+g(x_{n+1})+g(y_{n})),
g(yn)=13(g(yn1)+g(yn+1)+g(xn)).g(y_{n})=\frac{1}{3}(g(y_{n-1})+g(y_{n+1})+g(x_{n})).

Let zn=g(xn)g(yn)z_{n}=g(x_{n})-g_{(}y_{n}), hn=g(xn)+g(yn)h_{n}=g(x_{n})+g(y_{n}). We have 4zn=zn1+zn+14z_{n}=z_{n-1}+z_{n+1}, 2hn=hn1+hn+12h_{n}=h_{n-1}+h_{n+1}. On the one hand, it is easy to check that hn=O(n)h_{n}=O(n). On the other hand, znzn1z_{n}\geq z_{n-1} by the maximum principle. Thus zn+13znz_{n+1}\geq 3z_{n}, which means that znC3nz_{n}\geq C\cdot 3^{n} for some constant C>0C>0. It follows that g(yn)g(y_{n})\to-\infty and g(xn)+g(x_{n})\to+\infty as n+n\to+\infty. We get a contradiction as gg is bounded from below or above. ∎

Acknowledgments

The authors are grateful to Prof. Bobo Hua for his guidance and support. The authors would also like to thank Florentin Münch and David Cushing for their helpful advice.

References

  • [BCL+18] D. P. Bourne, D. Cushing, S. Liu, F. Münch, and N. Peyerimhoff. Ollivier-Ricci idleness functions of graphs. SIAM J. Discrete Math., 32(2):1408–1424, 2018.
  • [BHL+15] Frank Bauer, Paul Horn, Yong Lin, Gabor Lippner, Dan Mangoubi, and Shing-Tung Yau. Li-Yau inequality on graphs. J. Differential Geom., 99(3):359–405, 2015.
  • [BHLY20] Shuliang Bai, An Huang, Linyuan Lu, and Shing-Tung Yau. On the sum of ricci-curvatures for weighted graphs. arXiv preprint arXiv:2001.01776, 2020.
  • [BLY21] Shuliang Bai, Linyuan Lu, and Shing-Tung Yau. Ricci-flat graphs with maximum degree at most 4. arXiv preprint arXiv:2103.00941, 2021.
  • [CD60] Gustave Choquet and Jacques Deny. Sur l’équation de convolution μ=μσ\mu=\mu\ast\sigma. C. R. Acad. Sci. Paris, 250:799–801, 1960.
  • [CG72] Jeff Cheeger and Detlef Gromoll. The splitting theorem for manifolds of nonnegative Ricci curvature. J. Differential Geometry, 6:119–128, 1971/72.
  • [CKL+18] David Cushing, Riikka Kangaslampi, Yong Lin, Shiping Liu, Linyuan Lu, and Shing-Tung Yau. Ricci-flat cubic graphs with girth five. arXiv preprint arXiv:1802.02982, 2018.
  • [CKL+19] David Cushing, Riikka Kangaslampi, Valtteri Lipiäinen, Shiping Liu, and George W Stagg. The graph curvature calculator and the curvatures of cubic graphs. Experimental Mathematics, pages 1–13, 2019.
  • [CP18] Hee Je Cho and Seong-Hun Paeng. Classification of α\alpha-Ricci flat graphs with girth at least five. Discrete Math., 341(10):2894–2902, 2018.
  • [HLLY19] Paul Horn, Yong Lin, Shuang Liu, and Shing-Tung Yau. Volume doubling, Poincaré inequality and Gaussian heat kernel estimate for non-negatively curved graphs. J. Reine Angew. Math., 757:89–130, 2019.
  • [HLYY18] Weihua He, Jun Luo, Chao Yang, and Wei Yuan. Ricci-flat graphs with girth four. arXiv preprint arXiv:1807.07253, 2018.
  • [HM21] Bobo Hua and Florentin Münch. Every salami has two ends. arXiv preprint arXiv:2105.11887, 2021.
  • [Hua19] Bobo Hua. Liouville theorem for bounded harmonic functions on manifolds and graphs satisfying non-negative curvature dimension condition. Calc. Var. Partial Differential Equations, 58(2):Paper No. 42, 8, 2019.
  • [JMR19] Jürgen Jost, Florentin Münch, and Christian Rose. Liouville property and non-negative ollivier curvature on graphs. arXiv preprint arXiv:1903.10796, 2019.
  • [LLY11] Yong Lin, Linyuan Lu, and Shing-Tung Yau. Ricci curvature of graphs. Tohoku Math. J. (2), 63(4):605–627, 2011.
  • [LLY14] Yong Lin, Linyuan Lu, and S.-T. Yau. Ricci-flat graphs with girth at least five. Comm. Anal. Geom., 22(4):671–687, 2014.
  • [LY10] Yong Lin and Shing-Tung Yau. Ricci curvature and eigenvalue estimate on locally finite graphs. Math. Res. Lett., 17(2):343–356, 2010.
  • [MW19] Florentin Münch and Radosław K. Wojciechowski. Ollivier Ricci curvature for general graph Laplacians: heat equation, Laplacian comparison, non-explosion and diameter bounds. Adv. Math., 356:106759, 45, 2019.
  • [Oll07] Yann Ollivier. Ricci curvature of metric spaces. C. R. Math. Acad. Sci. Paris, 345(11):643–646, 2007.
  • [Oll09] Yann Ollivier. Ricci curvature of Markov chains on metric spaces. J. Funct. Anal., 256(3):810–864, 2009.
  • [Sch99] Michael Schmuckenschläger. Curvature of nonlocal Markov generators. In Convex geometric analysis (Berkeley, CA, 1996), volume 34 of Math. Sci. Res. Inst. Publ., pages 189–197. Cambridge Univ. Press, Cambridge, 1999.
  • [Yau75] Shing Tung Yau. Harmonic functions on complete Riemannian manifolds. Comm. Pure Appl. Math., 28:201–228, 1975.