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

Helly-gap of a graph and vertex eccentricities

Feodor F. Dragan and Heather M. Guarnera
Algorithmic Research Laboratory, Department of Computer Science
Kent State University, Kent, Ohio, USA
[email protected], [email protected]
Abstract

A new metric parameter for a graph, Helly-gap, is introduced. A graph GG is called α\alpha-weakly-Helly if any system of pairwise intersecting disks in GG has a nonempty common intersection when the radius of each disk is increased by an additive value α\alpha. The minimum α\alpha for which a graph GG is α\alpha-weakly-Helly is called the Helly-gap of GG and denoted by α(G)\alpha(G). The Helly-gap of a graph GG is characterized by distances in the injective hull (G)\mathcal{H}(G), which is a (unique) minimal Helly graph which contains GG as an isometric subgraph. This characterization is used as a tool to generalize many eccentricity related results known for Helly graphs (α(G)=0\alpha(G)=0), as well as for chordal graphs (α(G)1\alpha(G)\leq 1), distance-hereditary graphs (α(G)1\alpha(G)\leq 1) and δ\delta-hyperbolic graphs (α(G)2δ\alpha(G)\leq 2\delta), to all graphs, parameterized by their Helly-gap α(G)\alpha(G). Several additional graph classes are shown to have a bounded Helly-gap, including AT-free graphs and graphs with bounded tree-length, bounded chordality or bounded αi\alpha_{i}-metric.

Keywords: Helly-type property; injective hull; chordal graphs; δ\delta-hyperbolic graphs; tree-length; αi\alpha_{i}-metric, eccentricity; center; diameter; radius.

1 Introduction

All graphs in this paper are unweighted, undirected, connected, and without loops or multiple edges. In a graph G=(V(G),E(G))G=(V(G),E(G)), a disk DG(v,r)D_{G}(v,r) with radius rr and centered at a vertex vv consists of all vertices with distance at most rr from vv, i.e., DG(v,r)={uV(G):dG(u,v)r}D_{G}(v,r)=\{u\in V(G):d_{G}(u,v)\leq r\}. A graph GG is called Helly if every system of pairwise intersecting disks of GG has a non-empty intersection.

Definition 1 (Helly graph).

A graph GG is Helly if, for any system of disks ={DG(v,r(v)):vSV(G)}\mathcal{F}=\{D_{G}(v,r(v)):v\in S\subseteq V(G)\}, the following Helly property holds: if XYX\cap Y\neq\emptyset for every X,YX,Y\in\mathcal{F}, then vSDG(v,r(v))\underset{v\in S}{\bigcap}D_{G}(v,r(v))\neq\emptyset.

Helly graphs are well investigated. They have several characterizations and important features as established in [16, 17, 4, 44, 46, 5]. They are exactly the so-called absolute retracts of reflexive graphs and possess a certain elimination scheme [16, 17, 4, 44, 5]. They can be recognized in O(n2m)O(n^{2}m) time [16] where nn is the number of vertices and mm is the number of edges. The Helly property works as a compactness criterion on graphs [46]. More importantly, every graph is isometrically embeddable into a Helly graph [36, 27, 40] (see Section 3 for more details).

Many nice properties of Helly graphs are based on the eccentricity eG(v)e_{G}(v) of a vertex vv, which is defined as the maximum distance from vv to any other vertex of the graph (i.e., eG(v)=maxuV(G)dG(v,u)e_{G}(v)=\max_{u\in V(G)}d_{G}(v,u)). The minimum (maximum) eccentricity in a graph GG is the radius rad(G)rad(G) (diameter diam(G)diam(G)). The center C(G)C(G) of a graph GG is the set of all vertices with minimum eccentricity (i.e., C(G)={vV(G):eG(v)=rad(G)}C(G)=\{v\in V(G):e_{G}(v)=rad(G)\}. Graph’s diameter is tightly bounded in Helly graphs as 2rad(G)diam(G)2rad(G)12rad(G)\geq diam(G)\geq 2rad(G)-1 [16, 17]. Moreover, the eccentricity function in Helly graphs is unimodal [17], that is, any local minimum coincides with the global minimum; this is equivalent to the condition that, for any vertex vV(G)v\in V(G), eG(v)=dG(v,C(G))+rad(G)e_{G}(v)=d_{G}(v,C(G))+rad(G) holds [16]. The unimodality of the eccentricity function was recently used in [29, 28] to compute the radius, diameter and a central vertex of a Helly graph in subquadratic time. For a vertex vV(G)v\in V(G), let FG(v)F_{G}(v) denote the set of all vertices farthest from vv, that is, FG(v)={uV(G):eG(v)=dG(v,u)}F_{G}(v)=\{u\in V(G):e_{G}(v)=d_{G}(v,u)\}. For every vertex vv of a Helly graph GG, each vertex uFG(v)u\in F_{G}(v) satisfies eG(u)2rad(G)diam(C(G))e_{G}(u)\geq 2rad(G)-diam(C(G)) [16]. Although the center C(G)C(G) of a Helly graph GG may have an arbitrarily large diameter (as any Helly graph is the center of some Helly graph), C(G)C(G) induces a Helly graph and is isometric in GG [16]. Additionally, any power of a Helly graph is a Helly graph as well [16].

In this paper, we introduce a far reaching generalization of Helly graphs, α\alpha-weakly-Helly graphs. We define α\alpha-weakly-Helly graphs as those graphs which are “weakly” Helly with the following simple generalization: for any system of disks, if the disks pairwise intersect, then by expanding the radius of each disk by some integer α\alpha there forms a common intersection. Thus, Helly graphs are exactly 0-weakly-Helly graphs.

Definition 2 (α\alpha-weakly-Helly graph).

A graph GG is α\alpha-weakly-Helly if, for any system of disks ={DG(v,r(v)):vSV(G)}\mathcal{F}=\{D_{G}(v,r(v)):v\in S\subseteq V(G)\}, the following α\alpha-weakly-Helly property holds: if XYX\cap Y\neq\emptyset for every X,YX,Y\in\mathcal{F}, then vSDG(v,r(v)+α)\underset{v\in S}{\bigcap}D_{G}(v,r(v)+\alpha)\neq\emptyset.
Clearly, every graph is α\alpha-weakly-Helly for some α\alpha. We call the minimum α\alpha for which a graph GG is α\alpha-weakly-Helly the Helly-gap of GG, denoted by α(G)\alpha(G).

Interestingly, there are a few results in the literature which demonstrate that such a α\alpha-weakly-Helly property with bounded α\alpha is present in some graphs and in some metric spaces. In [14], Chepoi and Estellon showed that for every δ\delta-hyperbolic geodesic space (X,d)(X,d) (and for every δ\delta-hyperbolic graph GG), if disks of the family ={D(x,r(x)):vSX,S\mathcal{F}=\{D(x,r(x)):v\in S\subseteq X,S is compact}\} pairwise intersect, then the disks {D(x,r(x)+2δ):xS}\{D(x,r(x)+2\delta):x\in S\} have a nonempty common intersection (see also  [11]). That is, the disks in δ\delta-hyperbolic geodesic spaces and in δ\delta-hyperbolic graphs satisfy (2δ)(2\delta)-weakly-Helly property. Even earlier, Lenhart, Pollack et al. [41, Lemma 9] established that the disks of simple polygons endowed with the link distance satisfy a Helly-type property which implies 1-weakly-Helly property. For chordal graphs [20] and for distance-hereditary graphs [18] (as well as for a more general class of graphs [10]) the following similar result is known: if for a set MV(G)M\subseteq V(G) and radius function r:Mr:M\rightarrow\mathbb{N}, every two vertices x,yMx,y\in M satisfy dG(x,y)r(x)+r(y)+1d_{G}(x,y)\leq r(x)+r(y)+1, then there is a clique KK in GG such that dG(v,K)r(v)d_{G}(v,K)\leq r(v) for all vMv\in M. Hence, clearly, every chordal graph and every distance-hereditary graph is 1-weakly-Helly. Note that two disks DG(v,p)D_{G}(v,p) and DG(u,q)D_{G}(u,q) intersect if and only if dG(v,u)p+qd_{G}(v,u)\leq p+q.

There are numerous other approaches to generalize in some form the Helly graphs. One may loosen the restriction on the type of sets which must satisfy the Helly property. If one considers neighborhoods or cliques, rather than arbitrary disks, then one gets the neighborhood-Helly graphs and the clique-Helly graphs as superclasses of Helly graphs (see [42, 34, 6] and papers cited therein). One may also generalize the Helly-property that a family of sets satisfies. This can be accomplished by specifying a minimum number or size of sets that pairwise intersect, a minimum number of subfamilies that have a common intersection, or the size of the intersection (see [30] and papers cited therein). Far reaching examples include the (p,q)(p,q)-Helly property (see [3, 2] and papers cited therein) and the fractional Helly property (see [37, 7] and papers cited therein). There is also a related to ours notion of λ\lambda-hyperconvexity in metric spaces (X,d)(X,d) [31, 38], where λ\lambda is the smallest multiplicative constant such that for any system of disks ={D(x,r(x)):xSX}\mathcal{F}=\{D(x,r(x)):x\in S\subseteq X\}, the following property holds: if XYX\cap Y\neq\emptyset for every X,YX,Y\in\mathcal{F}, then xSD(x,λr(x))\underset{x\in S}{\bigcap}D(x,\lambda\cdot r(x))\neq\emptyset. Several classical metric spaces are λ\lambda-hyperconvex for a bounded λ\lambda (see [31] and papers cited therein). These include reflexive Banach spaces and dual Banach spaces (λ2\lambda\leq 2) and Hilbert spaces (λ2\lambda\leq\sqrt{2}). The classical Jung Theorem asserting that each subset SS of the Euclidean space 𝔼m\mathbb{E}^{m} with finite diameter DD is contained in a disk of radius at most m2(m+1)D\sqrt{\frac{m}{2(m+1)}}D belongs to this kind of results, too.

In this paper, we are interested in α\alpha-weakly-Helly graphs, where α\alpha describes an additive constant by which radius of each disk in the family of pairwise intersecting disks can be increased in order to obtain a nonempty common intersection of all expanded disks. By definition, any α\alpha-weakly-Helly graph is (α+1)(\alpha+1)-hyperconvex. We find that there are also an abundance of graph classes with a bounded Helly-gap α(G)\alpha(G). Here, we generalize many eccentricity related results known for Helly graphs (as well as for chordal graphs, distance-hereditary graphs and δ\delta-hyperbolic graphs) to all graphs, parameterized by their Helly-gap α(G)\alpha(G). We provide a characterization of weakly-Helly graphs with respect to their injective hulls and use this as a tool to prove those generalizations. Several additional well-known graph classes that are α\alpha-weakly-Helly for some constant α\alpha are also identified.

The main results of this paper can be summarized as follows. After providing in Section 2 necessary definitions, notations and some relevant results on Helly graphs, in Section 3 we give a characterization of α\alpha-weakly-Helly graphs through distances in injective hulls. The injective hull of a graph GG, denoted by (G)\mathcal{H}(G), is a (unique) minimal Helly graph which contains GG as an isometric subgraph [36, 27]. We show that GG is an α\alpha-weakly-Helly graph if and only if for every vertex hV((G))h\in V(\mathcal{H}(G)) there is a vertex vV(G)v\in V(G) such that d(G)(h,v)αd_{\mathcal{H}(G)}(h,v)\leq\alpha (Theorem 1). In Section 4, we relate the diameter, radius, and all eccentricities in GG to their counterparts in H:=(G)H:=\mathcal{H}(G). In particular, we show that eG(v)=eH(v)e_{G}(v)=e_{H}(v) for all vV(G)v\in V(G), diam(G)=diam(H)diam(G)=diam(H), and rad(G)α(G)rad(H)rad(G)rad(G)-\alpha(G)\leq rad(H)\leq rad(G). Additionally, we show diamG(C(G))2α(G)diamH(C(H))diamG(Cα(G)(G))+2α(G)diam_{G}(C(G))-2\alpha(G)\leq diam_{H}(C(H))\leq diam_{G}(C^{\alpha(G)}(G))+2\alpha(G), where C(G)={vV(G):eG(v)rad(G)+}C^{\ell}(G)=\{v\in V(G):e_{G}(v)\leq rad(G)+\ell\}. We also provide several bounds on α(G)\alpha(G) including its relation to diameter and radius as well as investigate the Helly-gap of powers of weakly-Helly graphs (Theorem 3). In particular, (2rad(G)diam(G))/2α(G)diam(G)/2\lfloor(2rad(G)-diam(G))/2\rfloor\leq\alpha(G)\leq\lfloor diam(G)/2\rfloor holds. The eccentricity function in α\alpha-weakly-Helly graphs is shown to exhibit the property that any vertex vCα(G)v\notin C^{\alpha}(G) has a nearby vertex, within distance 2α+12\alpha+1 from vv, with strictly smaller eccentricity (Theorem 4). In Section 5, we give upper and lower bounds on the eccentricity eG(v)e_{G}(v) of a vertex vv. We consider bounds based on the distance from vv to a closest vertex in Cα(G)(G)C^{\alpha(G)}(G), whether vv is farthest from some other vertex and if diamG(Cα(G)(G))diam_{G}(C^{\alpha(G)}(G)) is bounded. In particular, we show that |eG(v)dG(v,Cα(G)(G))rad(G)|α(G)|e_{G}(v)-d_{G}(v,C^{\alpha(G)}(G))-rad(G)|\leq\alpha(G) holds for any vertex vV(G)v\in V(G) (Theorem 5). We also prove the existence of a spanning tree TT of GG which gives an approximation of all vertex eccentricities in GG with an additive error depending only on α(G)\alpha(G) and diamG(Cα(G)(G))diam_{G}(C^{\alpha(G)}(G)) (Theorem 7). We find that in any shortest path to Cα(G)(G)C^{\alpha(G)}(G), the number of vertices with locality more than 1 does not exceed 2α(G)2\alpha(G), where the locality of a vertex vv is the minimum distance from vv to a vertex of strictly smaller eccentricity (Theorem 8). All these results greatly generalize some known facts about distance-hereditary graphs, chordal graphs, and δ\delta-hyperbolic graphs. Those graphs have bounded Helly gap. In Section 6, we identify several more (well-known) graph classes with a bounded Helly-gap, including kk-chordal graphs, AT-free graphs, rectilinear grids, graphs with a bounded αi\alpha_{i}-metric, and graphs with bounded tree-length or tree-breadth. We conclude with open questions in Section 7.

2 Preliminaries

All graphs G=(V(G),E(G))G=(V(G),E(G)) occurring in this paper are undirected, connected, and without loops or multiple edges. The length of a path from a vertex uu to a vertex vv is the number of edges in the path. The distance dG(u,v)d_{G}(u,v) between two vertices uu and vv is the length of a shortest path connecting them in GG. The distance from a vertex uu to a vertex set SV(G)S\subseteq V(G) is defined as dG(u,S)=minsSdG(u,s)d_{G}(u,S)=\min_{s\in S}d_{G}(u,s). The eccentricity of a vertex is defined as eG(v)=maxuV(G)dG(v,u)e_{G}(v)=\max_{u\in V(G)}d_{G}(v,u). The vertex set FG(v)F_{G}(v) for a vertex vV(G)v\in V(G) denotes the set of all vertices farthest from vv, that is, FG(v)={uV(G):eG(v)=dG(v,u)}F_{G}(v)=\{u\in V(G):e_{G}(v)=d_{G}(v,u)\}. A disk of radius kk centered at a set SS (or a vertex) is the set of vertices of distance at most kk from SS, that is, DG(S,k)={uV(G):dG(u,S)k}D_{G}(S,k)=\{u\in V(G):d_{G}(u,S)\leq k\}. We omit GG from subindices when the graph is known by context.

The kthk^{th} power GkG^{k} of a graph GG is a graph that has the same set of vertices, but in which two distinct vertices are adjacent if and only if their distance in GG is at most kk. A subgraph GG^{\prime} of a graph GG is called isometric (or a distance preserving subgraph) if for any two vertices x,yx,y of GG^{\prime}, dG(x,y)=dG(x,y)d_{G}(x,y)=d_{G^{\prime}}(x,y) holds. The interval I(x,y)I(x,y) is the set of all vertices belonging to a shortest (x,y)(x,y)-path, that is, I(x,y)={uV(G):d(x,u)+d(u,y)=d(x,y)}I(x,y)=\{u\in V(G):d(x,u)+d(u,y)=d(x,y)\}. The interval slice Sk(x,y)S_{k}(x,y) is the set of vertices on I(x,y)I(x,y) which are at distance kk from vertex xx. An interval I(x,y)I(x,y) is said to be κ\kappa-thin if, for any 0d(x,y)0\leq\ell\leq d(x,y) and any u,vS(x,y)u,v\in S_{\ell}(x,y), d(u,v)κd(u,v)\leq\kappa holds. The smallest κ\kappa for which all intervals of GG are κ\kappa-thin is called the interval thinness of GG and denoted by κ(G)\kappa(G). By G{x}G-\{x\} we denote an induced subgraph of GG obtained from GG by removing a vertex xV(G)x\in V(G).

The minimum (maximum) eccentricity in a graph GG is the radius rad(G)rad(G) (diameter diam(G)diam(G)). The center C(G)C(G) is the set of vertices with minimum eccentricity (i.e., C(G)={vV(G):e(v)=rad(G)}C(G)=\{v\in V(G):e(v)=rad(G)\}. It will be useful to define also Ck(G)={vV(G):e(v)rad(G)+k}C^{k}(G)=\{v\in V(G):e(v)\leq rad(G)+k\}; then C(G)=C0(G)C(G)=C^{0}(G). Let MV(G)M\subseteq V(G) be a subset of vertices of GG. We distinguish the eccentricity with respect to MM as follows: denote by eGM(v)e_{G}^{M}(v) the maximum distance from vertex vv to any vertex uMu\in M, i.e., eGM(v)=maxuMdG(v,u)e_{G}^{M}(v)=\max_{u\in M}d_{G}(v,u). We then define FGM(v)={uM:eGM(v)=dG(v,u)}F_{G}^{M}(v)=\{u\in M:e_{G}^{M}(v)=d_{G}(v,u)\}, radG(M)=minvV(G)eGM(v)rad_{G}(M)=\min_{v\in V(G)}e_{G}^{M}(v), and diamG(M)=maxvMeGM(v)diam_{G}(M)=\max_{v\in M}e_{G}^{M}(v). Let CG(M)={vV(G):eGM(v)radG(M)+}C_{G}^{\ell}(M)=\{v\in V(G):e_{G}^{M}(v)\leq rad_{G}(M)+\ell\} and CG(M)=CG0(M)C_{G}(M)=C_{G}^{0}(M). For simplicity, when M=V(G)M=V(G), we continue to use the notation rad(G)rad(G), C(G)C(G), and diam(G)diam(G).

The following results for Helly graphs will prove useful for α\alpha-weakly-Helly graphs.

Lemma 1.

[16] Let GG be a Helly graph. For every MV(G)M\subseteq V(G), the graph induced by the center CG(M)C_{G}(M) is Helly and is an isometric subgraph of GG.

Given this lemma, it will be convenient to also denote by CG(M)C_{G}(M) a subgraph of GG induced by CG(M)C_{G}(M).

Lemma 2.

[16, 4] Let GG be a Helly graph. If there are two distinct vertices w,xV(G)w,x\in V(G) such that D(w,1)D(x,1)D(w,1)\supseteq D(x,1), then G{x}G-\{x\} is Helly and an isometric subgraph of GG.

For a graph GG and a subset MV(G)M\subseteq V(G), the eccentricity function eGM()e_{G}^{M}(\cdot) is called unimodal if every vertex vV(G)CG(M)v\in V(G)\setminus C_{G}(M) has a neighbor uu such that eGM(u)<eGM(v)e_{G}^{M}(u)<e_{G}^{M}(v).

Lemma 3.

[16, 17] A graph GG is Helly if and only if the eccentricity function eGM()e_{G}^{M}(\cdot) is unimodal on GG for every MV(G)M\subseteq V(G).

The following two results were earlier proven in [16] only for M=V(G)M=V(G) and then later extended in [21] to all MV(G)M\subseteq V(G).

Lemma 4.

[21] Let GG be a Helly graph. For any MV(G)M\subseteq V(G), 2radG(M)1diamG(M)2radG(M)2rad_{G}(M)-1\leq diam_{G}(M)\leq 2rad_{G}(M). Moreover, radG(M)=diamG(M)/2rad_{G}(M)=\lceil diam_{G}(M)/2\rceil.

Lemma 5.

[21] Let GG be a Helly graph. For every vV(G)v\in V(G) and MV(G)M\subseteq V(G), eGM(v)=dG(v,CG(M))+radG(M)e_{G}^{M}(v)=d_{G}(v,C_{G}(M))+rad_{G}(M) holds.

Corollary 1.

For every Helly graph GG, any subset MV(G)M\subseteq V(G) and any integers 0\ell\geq 0 and k0k\geq 0, CG+k(M)=DG(CGk(M)+)=DG(CG(M)+k+).C^{\ell+k}_{G}(M)=D_{G}(C^{k}_{G}(M)+\ell)=D_{G}(C_{G}(M)+k+\ell). Furthermore, diamG(CG+k(M))diamG(CGk(M))+2.diam_{G}(C^{\ell+k}_{G}(M))\leq diam_{G}(C^{k}_{G}(M))+2\ell.

Proof.

Consider a vertex vv with eGM(v)=k++radG(M)e_{G}^{M}(v)=k+\ell+rad_{G}(M) and a vertex cc in CG(M)C_{G}(M) closest to vv. By Lemma 5, eGM(v)=dG(v,CG(M))+radG(M)=dG(v,c)+radG(M)=k++radG(M)e_{G}^{M}(v)=d_{G}(v,C_{G}(M))+rad_{G}(M)=d_{G}(v,c)+rad_{G}(M)=k+\ell+rad_{G}(M). Hence, for every vertex vv, dG(v,CG(M))=k+d_{G}(v,C_{G}(M))=k+\ell if and only if eGM(v)=k++radG(M)e_{G}^{M}(v)=k+\ell+rad_{G}(M).

Let uu be a vertex on a shortest path from vv to cc at distance kk from cc. By Lemma 5, eGM(u)=dG(u,c)+radG(M)=k+radG(M)e_{G}^{M}(u)=d_{G}(u,c)+rad_{G}(M)=k+rad_{G}(M) and hence eGM(v)=k++radG(M)=eGM(u)+e_{G}^{M}(v)=k+\ell+rad_{G}(M)=e_{G}^{M}(u)+\ell. Therefore, eGM(v)=k++radG(M)e_{G}^{M}(v)=k+\ell+rad_{G}(M) if and only if dG(v,CGk(M))=dG(v,u)=d_{G}(v,C^{k}_{G}(M))=d_{G}(v,u)=\ell.

Let x,yx,y be vertices of CG+k(M)C^{\ell+k}_{G}(M) with dG(x,y)=diamG(CG+k(M))d_{G}(x,y)=diam_{G}(C^{\ell+k}_{G}(M)). Since dG(v,CGk(M))d_{G}(v,C^{k}_{G}(M))\leq\ell for each v{x,y}v\in\{x,y\}, by the triangle inequality, we have dG(x,y)dG(x,CGk(M))+diamG(CGk(M))+dG(y,CGk(M))diamG(CGk(M))+2.d_{G}(x,y)\leq d_{G}(x,C^{k}_{G}(M))+diam_{G}(C^{k}_{G}(M))+d_{G}(y,C^{k}_{G}(M))\leq diam_{G}(C^{k}_{G}(M))+2\ell.

3 Distances in injective hulls characterize weakly-Helly graphs

Recall that the injective hull of GG, denoted by (G)\mathcal{H}(G), is a minimal Helly graph which contains GG as an isometric subgraph [36, 27]. It turns out that (G)\mathcal{H}(G) is unique for every GG [36]. When GG is known by context, we often let H:=(G)H:=\mathcal{H}(G). By an equivalent definition of an injective hull [27] (also called a tight span), each vertex fV((G))f\in V(\mathcal{H}(G)) can be represented as a vector with nonnegative integer values f(x)f(x) for each xV(G)x\in V(G), such that the following two properties hold:

x,yV(G)f(x)+f(y)dG(x,y)\forall x,y\in V(G)\ f(x)+f(y)\geq d_{G}(x,y) (1)
xV(G)yV(G)f(x)+f(y)=dG(x,y)\forall x\in V(G)\ \exists y\in V(G)\ f(x)+f(y)=d_{G}(x,y) (2)

Additionally, there is an edge between two vertices f,gV((G))f,g\in V(\mathcal{H}(G)) if and only if their Chebyshev distance is 1, i.e., maxxV(G)|f(x)g(x)|=1\max_{x\in V(G)}\lvert f(x)-g(x)\rvert=1. Thus, dH(f,g)=maxxV(G)|f(x)g(x)|d_{H}(f,g)=max_{x\in V(G)}\lvert f(x)-g(x)\rvert. Notice that if fV((G))f\in V(\mathcal{H}(G)), then {D(x,f(x)):xV(G)}\{D(x,f(x)):x\in V(G)\} is a family of pairwise intersecting disks. For a vertex zV(G)z\in V(G), define the distance function dzd_{z} by setting dz(x)=dG(z,x)d_{z}(x)=d_{G}(z,x) for any xV(G)x\in V(G). By the triangle inequality, each dzd_{z} belongs to V((G))V(\mathcal{H}(G)). An isometric embedding of GG into (G)\mathcal{H}(G) is obtained by mapping each vertex zz of GG to its distance vector dzd_{z}.

We classify every vertex vv in V((G))V(\mathcal{H}(G)) as either a real vertex or a Helly vertex. A vertex fV((G))f\in V(\mathcal{H}(G)) is a real vertex provided f=dzf=d_{z} for some zV(G)z\in V(G), i.e., there is a one-to-one correspondence between zV(G)z\in V(G) and its representative real vertex fV((G))f\in V(\mathcal{H}(G)) which uniquely satisfies f(z)=0f(z)=0 and f(x)=dG(z,x)f(x)=d_{G}(z,x) for all xV(G)x\in V(G). When working with (G)\mathcal{H}(G), we will use interchangeably the notation V(G)V(G) to represent the vertex set in GG as well as the vertex subset of (G)\mathcal{H}(G) which uniquely corresponds to the vertex set of GG. Then, a vertex vV((G))v\in V(\mathcal{H}(G)) is a real vertex if it belongs to V(G)V(G) and a Helly vertex otherwise. Equivalently, a vertex hV((G))h\in V(\mathcal{H}(G)) is a Helly vertex provided that h(x)1h(x)\geq 1 for all xV(G)x\in V(G), that is, a Helly vertex exists only in the injective hull (G)\mathcal{H}(G) and not in GG. We will show in Theorem 1 that a graph GG is α\alpha-weakly-Helly if and only if the distance from any Helly vertex in (G)\mathcal{H}(G) to a closest real vertex in V(G)V(G) is no more than α\alpha. We recently learned that this result was independently discovered by Chalopin et al. [8]. They use name coarse Helly property for α\alpha-weakly-Helly property used here.

The following properties will be useful to the main result of this section and to later sections. A vertex xx is called a peripheral vertex if I(y,x)I(y,z)I(y,x)\not\subset I(y,z) for some vertex yy and all vertices zxz\neq x. We show next that the peripheral vertices of (G)\mathcal{H}(G) are real vertices. This adheres to the intuitive notion that an injective hull contains all of the Helly vertices “between” the vertices of GG, so that the outermost vertices of (G)\mathcal{H}(G) are real.

Proposition 1.

Peripheral vertices of (G)\mathcal{H}(G) are real.

Proof.

By contradiction, suppose there is a peripheral Helly vertex uV((G))V(G)u\in V(\mathcal{H}(G))\setminus V(G). By definition, there is a vertex sV((G))s\in V(\mathcal{H}(G)) such that, for all xV((G))x\in V(\mathcal{H}(G)), I(s,u)I(s,x)I(s,u)\not\subset I(s,x). Let k:=d(u,s)k:=d(u,s). Consider pairwise intersecting disks D(s,k1)D(s,k-1) and D(x,1)D(x,1) for each xD(u,1)x\in D(u,1). By the Helly property, there exists vertex wV((G))w\in V(\mathcal{H}(G)) with d(w,s)=k1d(w,s)=k-1 and D(w,1)D(u,1)D(w,1)\supseteq D(u,1). By Lemma 2, (G){u}\mathcal{H}(G)-\{u\} is Helly and is an isometric subgraph of (G)\mathcal{H}(G). Since uu is a Helly vertex, GG is an isometric subgraph of (G){u}\mathcal{H}(G)-\{u\}, a contradiction with the minimality of (G)\mathcal{H}(G). ∎

Moreover, we show that any shortest path of (G)\mathcal{H}(G) is a subpath of a shortest path between real vertices, which will later prove a useful property of injective hulls.

Proposition 2.

Let HH be the injective hull of GG. For any shortest path P(x,y)P(x,y), where x,yV(H)x,y\in V(H), there is a shortest path P(x,y)P(x^{\prime},y^{\prime}), where x,yV(G)x^{\prime},y^{\prime}\in V(G) such that P(x,y)P(x,y)P(x^{\prime},y^{\prime})\supseteq P(x,y).

Proof.

If xx and yy are both real vertices, then the proposition is trivially true. Without loss of generality, let yy be a Helly vertex. Consider a breadth-first search layering where yy belongs to layer LiL_{i} of BFS(H,xH,x). Let yLky^{\prime}\in L_{k} be a vertex with yI(x,y)y\in I(x,y^{\prime}) that maximizes k=dH(x,y)k=d_{H}(x,y^{\prime}). Then, for any vertex zV((G))z\in V(\mathcal{H}(G)), I(x,y)I(x,z)I(x,y^{\prime})\not\subset I(x,z). By Proposition 1, yV(G)y^{\prime}\in V(G). If xV(G)x\notin V(G), then applying the previous step using BFS(H,yH,y^{\prime}) yields vertex xV(G)x^{\prime}\in V(G). ∎

We are now ready to prove the main result of this section.

Theorem 1.

For any vertex hV((G))h\in V(\mathcal{H}(G)) there is a real vertex vV(G)v\in V(G) such that d(G)(h,v)αd_{\mathcal{H}(G)}(h,v)\leq\alpha if and only if GG is an α\alpha-weakly-Helly graph.

Proof.

Suppose any Helly vertex in H:=(G)H:=\mathcal{H}(G) is within distance at most α\alpha from a vertex of GG. Consider in GG a family of pairwise intersecting disks G={DG(v,r(v)):vSV(G)}\mathcal{F}_{G}=\{D_{G}(v,r(v)):v\in S\subseteq V(G)\}. As HH contains GG as an isometric subgraph, the disks H={DH(v,r(v):vS}\mathcal{F}_{H}=\{D_{H}(v,r(v):v\in S\} are also pairwise intersecting in HH. By the Helly property, in HH there is a vertex xvSDH(v,r(v))x\in\bigcap_{v\in S}D_{H}(v,r(v)). By assumption, there is a vertex uV(G)u\in V(G) such that dH(u,x)αd_{H}(u,x)\leq\alpha. Thus uvSDG(v,r(v)+α)u\in\bigcap_{v\in S}D_{G}(v,r(v)+\alpha) and so GG is α\alpha-weakly-Helly.

Assume now that GG is α\alpha-weakly-Helly. Let hV(H)h\in V(H) be an arbitrary vertex represented as a vector with nonnegative integer values h(x)h(x) for each xV(G)x\in V(G) satisfying conditions (1) and (2) from the definition of an injective hull. Then, {DG(x,h(x)):xV(G)}\{D_{G}(x,h(x)):x\in V(G)\} is a family of pairwise intersecting disks in GG. By the α\alpha-weakly-Helly property, there is a real vertex zV(G)z\in V(G) belonging to xV(G)DG(x,h(x)+α)\bigcap_{x\in V(G)}D_{G}(x,h(x)+\alpha). To establish that dH(z,h)αd_{H}(z,h)\leq\alpha and complete the proof, we will show that maxtV(G)|z(t)h(t)|α\max_{t\in V(G)}\lvert z(t)-h(t)\rvert\leq\alpha.

First, we show that z(t)h(t)αz(t)-h(t)\leq\alpha for all tV(G)t\in V(G). As zz is a real vertex, by definition, z(t)=dG(z,t)z(t)=d_{G}(z,t) and, by the α\alpha-weakly-Helly property (recall that zxV(G)DG(x,h(x)+α)z\in\bigcap_{x\in V(G)}D_{G}(x,h(x)+\alpha)), z(t)h(t)+αz(t)\leq h(t)+\alpha. Next, we show that h(t)z(t)αh(t)-z(t)\leq\alpha for all tV(G)t\in V(G). Suppose that h(x)z(x)>αh(x)-z(x)>\alpha for some xV(G)x\in V(G). Then, for all vertices yxy\neq x, h(x)+h(y)>z(x)+α+h(y)z(x)+z(y)d(x,y)h(x)+h(y)>z(x)+\alpha+h(y)\geq z(x)+z(y)\geq d(x,y), a contradiction with condition (2). ∎

The injective hull is also useful to prove that α\alpha-weakly-Helly graphs are closed under pendant vertex addition.

Lemma 6.

Let G+{x}G+\{x\} be the graph obtained from GG by adding a vertex xx pendant to any fixed vertex vV(G)v\in V(G). Then, (G+{x})=(G)+{x}\mathcal{H}(G+\{x\})=\mathcal{H}(G)+\{x\}.

Proof.

As xx is pendant to vv, then all uV(G)u\in V(G) have dG+{x}(u,x)=dG(u,v)+1d_{G+\{x\}}(u,x)=d_{G}(u,v)+1. Let H1:=(G+{x})H_{1}:=\mathcal{H}(G+\{x\}) and H2:=(G)+{x}H_{2}:=\mathcal{H}(G)+\{x\}. Note that H2H_{2} is a Helly graph containing G+{x}G+\{x\} as an isometric subgraph. We first show that any hV(H1)h\in V(H_{1}) also belongs to V(H2)V(H_{2}). The statement clearly holds if hh is a real vertex, so assume that hh is a Helly vertex of H1H_{1} represented as a vector with nonnegative integer values for each uV(G+{x})u\in V(G+\{x\}) satisfying conditions (1) and (2) from the definition of an injective hull. We will show hh also satisfies the conditions under GG. By condition (1) under G+{x}G+\{x\}, and since GG is isometric in G+{x}G+\{x\}, for all u,yV(G)u,y\in V(G), h(u)+h(y)dG+{x}(u,y)=dG(u,y)h(u)+h(y)\geq d_{G+\{x\}}(u,y)=d_{G}(u,y). Thus, hh satisfies condition (1) in GG. By condition (2) under G+{x}G+\{x\}, for every uV(G)u\in V(G) there is a vertex yV(G+{x})y\in V(G+\{x\}) with h(u)+h(y)=dG+{x}(u,y)h(u)+h(y)=d_{G+\{x\}}(u,y). We claim that if y=xy=x, then h(x)=h(v)+1h(x)=h(v)+1 and so vertex vv also satisfies h(u)+h(v)=h(u)+h(y)1=dG+{x}(u,y)1=dG(u,v)h(u)+h(v)=h(u)+h(y)-1=d_{G+\{x\}}(u,y)-1=d_{G}(u,v). On one hand, h(x)h(v)+1h(x)\leq h(v)+1 since h(u)+h(x)=dG+{x}(u,x)=dG(u,v)+1h(u)+h(v)+1h(u)+h(x)=d_{G+\{x\}}(u,x)=d_{G}(u,v)+1\leq h(u)+h(v)+1. On the other hand, let zV(G){x}z\in V(G)\cup\{x\} be a vertex such that h(z)+h(v)=dG+{x}(z,v)h(z)+h(v)=d_{G+\{x\}}(z,v). Note that zxz\neq x as hh is not a real vertex and therefore h(v)=dH1(h,v)>0h(v)=d_{H_{1}}(h,v)>0 and h(z)=dH1(h,z)>0h(z)=d_{H_{1}}(h,z)>0. Then, h(x)dG+{x}(z,x)h(z)=dG+{x}(z,v)+1h(z)=h(v)+1h(x)\geq d_{G+\{x\}}(z,x)-h(z)=d_{G+\{x\}}(z,v)+1-h(z)=h(v)+1. With the claim established, hh satisfies condition (2) in GG. Thus, hV(H2)h\in V(H_{2}) and V(H1)V(H2)V(H_{1})\subseteq V(H_{2}). By minimality of (G)\mathcal{H}(G), V(H1)=V(H2)V(H_{1})=V(H_{2}). ∎

Corollary 2.

Let G+{x}G+\{x\} be the graph obtained from GG by adding a pendant vertex xx adjacent to any fixed vertex vV(G)v\in V(G). Then, α(G)=α(G+{x})\alpha(G)=\alpha(G+\{x\}).

4 Diameter, radius, and all eccentricities

We establish several bounds on all eccentricities, and the diameter and radius in particular, for α\alpha-weakly-Helly graphs. As an intermediate step, we relate the diameter, radius, and all eccentricities in GG to their counterparts in (G)\mathcal{H}(G).

4.1 Eccentricities and centers

First, we provide a few immediate consequences of Proposition 1 which establishes that farthest vertices in (G)\mathcal{H}(G) are real vertices. That is, for any vV((G))v\in V(\mathcal{H}(G)), F(G)(v)V(G)F_{\mathcal{H}(G)}(v)\subseteq V(G). It follows that all eccentricities in GG and the diameter of GG are preserved in (G)\mathcal{H}(G), including with respect to any subset MV(G)M\subseteq V(G) of real vertices.

Proposition 3.

Let HH be the injective hull of GG. For any MV(G)M\subseteq V(G) and vV(G)v\in V(G), eGM(v)=eHM(v)e_{G}^{M}(v)=e_{H}^{M}(v). Moreover, eG(v)=eH(v)e_{G}(v)=e_{H}(v).

Proposition 4.

Let HH be the injective hull of GG. For any MV(G)M\subseteq V(G), diamG(M)=diamH(M)diam_{G}(M)=diam_{H}(M). Moreover, diam(G)=diam(H)diam(G)=diam(H).

Proposition 5.

Let HH be the injective hull of an α\alpha-weakly-Helly graph GG. For any MV(G)M\subseteq V(G), radG(M)αradH(M)radG(M)rad_{G}(M)-\alpha\leq rad_{H}(M)\leq rad_{G}(M). In particular, rad(G)αrad(H)rad(G)rad(G)-\alpha\leq rad(H)\leq rad(G).

Proof.

By Proposition 3, the eccentricity of any vertex of GG is preserved in HH. Hence, radH(M)radG(M)rad_{H}(M)\leq rad_{G}(M). Consider any vertex hCH(M)h\in C_{H}(M). We have dH(h,x)radH(M)d_{H}(h,x)\leq rad_{H}(M) for all xMx\in M. By Theorem 1, there is a real vertex vV(G)v\in V(G) such that dH(v,h)αd_{H}(v,h)\leq\alpha. By the triangle inequality, radG(M)eGM(v)maxxM{dH(v,h)+dH(h,x)}radH(M)+αrad_{G}(M)\leq e_{G}^{M}(v)\leq max_{x\in M}\{d_{H}(v,h)+d_{H}(h,x)\}\leq rad_{H}(M)+\alpha. ∎

Though the diameter is preserved, the radius in (G)\mathcal{H}(G) may be smaller than the radius in GG by at most α(G)\alpha(G). In this case, the radius in (G)\mathcal{H}(G) is realized by Helly vertices which are not present in GG, resulting in different centers in (G)\mathcal{H}(G) and GG. In what follows, we establish that, for any MV(G)M\subseteq V(G), any vertex of CG(M)C_{G}(M) is close to a vertex of CH(M)C_{H}(M). Moreover, the diameter of the center of MM in GG is at most the diameter of the center of MM in the injective hull (G)\mathcal{H}(G) plus 2α(G)2\alpha(G).

Lemma 7.

Let HH be the injective hull of an α\alpha-weakly-Helly graph GG. For any MV(G)M\subseteq V(G) and integer 0\ell\geq 0, CG(M)DH(CH(M),α+)=DH(CH(M),α)C_{G}^{\ell}(M)\subseteq D_{H}(C_{H}(M),\alpha+\ell)=D_{H}(C^{\ell}_{H}(M),\alpha).

Proof.

Consider any vertex xCG(M)x\in C_{G}^{\ell}(M). By Proposition 5, eGM(x)radG(M)+radH(M)+α+e_{G}^{M}(x)\leq rad_{G}(M)+\ell\leq rad_{H}(M)+\alpha+\ell. By Proposition 3 and Lemma 5 (since HH is Helly), eGM(x)=eHM(x)=radH(M)+dH(x,CH(M))e_{G}^{M}(x)=e_{H}^{M}(x)=rad_{H}(M)+d_{H}(x,C_{H}(M)). Therefore, dH(x,CH(M))α+d_{H}(x,C_{H}(M))\leq\alpha+\ell. By Corollary 1, DH(CH(M),α+)=DH(CH(M),α).D_{H}(C_{H}(M),\alpha+\ell)=D_{H}(C^{\ell}_{H}(M),\alpha).

Theorem 2.

Let HH be the injective hull of an α\alpha-weakly-Helly graph GG. For any integer 0\ell\geq 0 and any MV(G)M\subseteq V(G), diamG(CG(M))2αdiamH(CH(M))diamG(CGα+(M))+2αdiam_{G}(C_{G}^{\ell}(M))-2\alpha\leq diam_{H}(C_{H}^{\ell}(M))\leq diam_{G}(C_{G}^{\alpha+\ell}(M))+2\alpha.

Proof.

Let dH(u,v)=diamH(CH(M))d_{H}(u,v)=diam_{H}(C_{H}^{\ell}(M)) for some u,vCH(M)u,v\in C_{H}^{\ell}(M). By Theorem 1, there is a real vertex uV(G)u^{*}\in V(G) at distance at most α\alpha from uu. By Proposition 3 and Proposition 5, eGM(u)=eHM(u)eHM(u)+αradH(M)++αradG(M)++αe_{G}^{M}(u^{*})=e_{H}^{M}(u^{*})\leq e_{H}^{M}(u)+\alpha\leq rad_{H}(M)+\ell+\alpha\leq rad_{G}(M)+\ell+\alpha. Similarly, there is a real vertex vV(G)v^{*}\in V(G) at distance at most α\alpha from vv with eGM(v)radG(M)++αe_{G}^{M}(v^{*})\leq rad_{G}(M)+\ell+\alpha. Both vertices v,uv^{*},u^{*} belong to CGα+(M)C_{G}^{\alpha+\ell}(M). By the triangle inequality, diamH(CH(M))=dH(u,v)2α+dH(u,v)2α+diamG(CGα+(M))diam_{H}(C_{H}^{\ell}(M))=d_{H}(u,v)\leq 2\alpha+d_{H}(u^{*},v^{*})\leq 2\alpha+diam_{G}(C_{G}^{\alpha+\ell}(M)).

On the other hand, by Lemma 7, any vertex of CG(M)C_{G}^{\ell}(M) has distance at most α\alpha to CH(M)C^{\ell}_{H}(M). Let x,yCG(M)x,y\in C_{G}^{\ell}(M) such that dG(x,y)=diamG(CG(M))d_{G}(x,y)=diam_{G}(C_{G}^{\ell}(M)). By the triangle inequality, dG(x,y)=dH(x,y)diamH(CH(M))+2αd_{G}(x,y)=d_{H}(x,y)\leq diam_{H}(C^{\ell}_{H}(M))+2\alpha. A small graph depicted on Figure 1 demonstrates that this inequality is tight. ∎

[Uncaptioned image]
Figure 1: A graph GG (left) and its injective hull HH (right) which show that inequalities in Proposition 5 and Theorem 2 are tight: α(G)=1\alpha(G)=1, rad(G)=2rad(G)=2, rad(H)=1rad(H)=1, diam(C(H))=0diam(C(H))=0, diam(C(G))=2diam(C(G))=2.
Corollary 3.

For each α\alpha-weakly-Helly graph GG with the injective hull H=(G)H={\mathcal{H}(G)} and every 0\ell\geq 0, C(H)V(G)C(G)C+α(H)V(G).C^{\ell}(H)\cap V(G)\subseteq C^{\ell}(G)\subseteq C^{\ell+\alpha}(H)\cap V(G).

[Uncaptioned image]
Figure 2: Inclusions for sets C(G)C(G), C(H)C(H), Cα(H)C^{\alpha}(H), C(G)C^{\ell}(G), and Cα+(H)C^{\alpha+\ell}(H) in H=(G)H=\mathcal{H}(G).
Proof.

Let vC(H)V(G)v\in C^{\ell}(H)\cap V(G). By Proposition 5 and Proposition 3, eG(v)=eH(v)rad(H)+rad(G)+e_{G}(v)=e_{H}(v)\leq rad(H)+\ell\leq rad(G)+\ell. Hence, vC(G)v\in C^{\ell}(G). Consider now a vertex vC(G)v\in C^{\ell}(G). It follows from Lemma 7 that dH(v,C(H))+αd_{H}(v,C(H))\leq\ell+\alpha. By Lemma 5, eH(v)=rad(H)+dH(v,C(H))rad(H)++αe_{H}(v)=rad(H)+d_{H}(v,C(H))\leq rad(H)+\ell+\alpha. Thus, vC+α(H)v\in C^{\ell+\alpha}(H). ∎

The results from this section on inclusions for central sets are illustrated in Figure 2.

4.2 Relation between Helly-gap and diameter, radius and graph powers

We obtain a few lower and upper bounds on the Helly-gap α(G)\alpha(G). The first follows directly from Proposition 5.

Corollary 4.

Let H:=(G)H:=\mathcal{H}(G). For any MV(G)M\subseteq V(G), α(G)radG(M)radH(M)\alpha(G)\geq rad_{G}(M)-rad_{H}(M).

If (G)\mathcal{H}(G) were given, one could compute α(G)\alpha(G) by computing the maximum distance from a Helly vertex to a closest real vertex. However, we provide in Corollary 5 and Lemma 8 upper and lower bounds which do not necessitate computing the injective hull.

Corollary 5.

Any graph GG is α\alpha-weakly-Helly for αdiam(G)/2\alpha\leq\lfloor diam(G)/2\rfloor.

Proof.

By contradiction, suppose α(G)>diam(G)/2\alpha(G)>\lfloor diam(G)/2\rfloor. Let H:=(G)H:=\mathcal{H}(G). By Theorem 1, in HH there exists a Helly vertex uu with dH(u,v)>diam(G)/2d_{H}(u,v)>\lfloor diam(G)/2\rfloor for all vV(G)v\in V(G). By Proposition 2, uu belongs to a shortest (x,y)(x,y)-path for two real vertices x,yV(G)x,y\in V(G). As GG is isometric in HH, dG(x,y)=dH(x,y)=dH(x,u)+dH(u,y)2(diam(G)/2+1)>diam(G)d_{G}(x,y)=d_{H}(x,y)=d_{H}(x,u)+d_{H}(u,y)\geq 2(\lfloor diam(G)/2\rfloor+1)>diam(G), a contradiction with dG(x,y)diam(G)d_{G}(x,y)\leq diam(G). ∎

Lemma 8.

Let HH be the injective hull of GG, MM be any subset of V(G)V(G) and k0k\geq 0 be an integer. If diamG(M)=2radG(M)kdiam_{G}(M)=2rad_{G}(M)-k, then radG(M)=radH(M)+k/2rad_{G}(M)=rad_{H}(M)+\lfloor k/2\rfloor. Moreover, α(G)(2radG(M)diamG(M))/2\alpha(G)\geq\lfloor(2rad_{G}(M)-diam_{G}(M))/2\rfloor, that is, 2radG(M)diamG(M)2radG(M)2α(G)12rad_{G}(M)\geq diam_{G}(M)\geq 2rad_{G}(M)-2\alpha(G)-1.

Proof.

By Proposition 4, 2radG(M)k=diamG(M)=diamH(M)2rad_{G}(M)-k=diam_{G}(M)=diam_{H}(M). Since HH is Helly, by Lemma 4, radH(M)=(2radG(M)k)/2=radG(M)k/2rad_{H}(M)=\lceil(2rad_{G}(M)-k)/2\rceil=rad_{G}(M)-\lfloor k/2\rfloor. By Corollary 4, α(G)k/2\alpha(G)\geq\lfloor k/2\rfloor. ∎

Remark.

Observe that the Helly-gap of a graph GG can be much larger than rad(G)rad(H)rad(G)-rad(H) and (2rad(G)diam(G))/2\lfloor(2rad(G)-diam(G))/2\rfloor. Consider a graph GG formed by a cycle C4kC_{4k} of size 4k4k and two paths of length kk each connected to opposite ends of the cycle, as illustrated in Fig. 3. By Lemma 8, α(C4k)k\alpha(C_{4k})\geq k. By Corollary 2, α(G)=α(C4k)k\alpha(G)=\alpha(C_{4k})\geq k. However, diam(G)=4kdiam(G)=4k and rad(G)=2krad(G)=2k, and by Lemma 8, rad(G)=rad(H)rad(G)=rad(H). In this case, rad(G)rad(H)=(2rad(G)diam(G))/2=0rad(G)-rad(H)=\lfloor(2rad(G)-diam(G))/2\rfloor=0.

[Uncaptioned image]
Figure 3: A graph GG with Helly-gap kk and rad(G)=rad(H)=2krad(G)=rad(H)=2k.
Corollary 6.

Let HH be the injective hull of GG and MV(G)M\subseteq V(G). Then, diamG(M)2radG(M)1diam_{G}(M)\geq 2rad_{G}(M)-1 if and only if CG(M)CH(M)C_{G}(M)\subseteq C_{H}(M).

Proof.

If diamG(M)2radG(M)1diam_{G}(M)\geq 2rad_{G}(M)-1, then by Lemma 8, radG(M)=radH(M)rad_{G}(M)=rad_{H}(M). Thus, any vCG(M)v\in C_{G}(M) has eHM(v)radH(M)e_{H}^{M}(v)\leq rad_{H}(M). On the other hand, if CG(M)CH(M)C_{G}(M)\subseteq C_{H}(M), since eccentricities are preserved in HH by Proposition 3, then also radG(M)=radH(M)rad_{G}(M)=rad_{H}(M). Since HH is Helly, by Lemma 4 and Proposition 4, diamG(M)=diamH(M)2radH(M)1=2radG(M)1diam_{G}(M)=diam_{H}(M)\geq 2rad_{H}(M)-1=2rad_{G}(M)-1 holds. ∎

Interestingly, the Helly-gap of a graph GG decreases in powers of GG.

Lemma 9.

Let GG be an α\alpha-weakly-Helly graph. For every integer k1k\geq 1, GkG^{k} is α/k\lceil\alpha/k\rceil-weakly-Helly.

Proof.

Let ={DGk(v,r(v)):vS}\mathcal{F}=\{D_{G^{k}}(v,r(v)):v\in S\} be a system of disks which pairwise intersect in GkG^{k} so that any two vertices u,vSu,v\in S satisfy dGk(u,v)r(u)+r(v)d_{G^{k}}(u,v)\leq r(u)+r(v). Then, their distances in GG satisfy dG(u,v)k(r(u)+r(v))d_{G}(u,v)\leq k(r(u)+r(v)). Consider now a corresponding system of disks in GG centered at same vertices, defined as ={DG(v,kr(v)):vS}\mathcal{M}=\{D_{G}(v,kr(v)):v\in S\}. \mathcal{M} is a family of pairwise intersecting disks of GG as any two vertices u,vSu,v\in S satisfy dG(u,v)kr(u)+kr(v)d_{G}(u,v)\leq kr(u)+kr(v). As GG is α\alpha-weakly-Helly, there exists a common vertex z{DG(v,kr(v)+α):vS}z\in\cap\{D_{G}(v,kr(v)+\alpha):v\in S\}. Since, for any vSv\in S, dG(z,v)kr(v)+αd_{G}(z,v)\leq kr(v)+\alpha, necessarily, dGk(z,v)r(v)+α/kd_{G^{k}}(z,v)\leq r(v)+\lceil\alpha/k\rceil. Hence, vertex zz intersects all disks of \mathcal{F} when the radii of each disk is extended by α/k\lceil\alpha/k\rceil. Therefore, GkG^{k} is α/k\lceil\alpha/k\rceil-weakly-Helly. ∎

The results of this subsection are summarized in Theorem 3.

Theorem 3.

Let GG be an arbitrary graph. Then, the following holds:

  • i)

    (2rad(G)diam(G))/2α(G)diam(G)/2\lfloor(2rad(G)-diam(G))/2\rfloor\leq\alpha(G)\leq\lfloor diam(G)/2\rfloor, and

  • ii)

    α(Gk)α(G)/k\alpha(G^{k})\leq\lceil\alpha(G)/k\rceil.

This Theorem 3 generalizes some known results for Helly graphs. Recall that, in Helly graphs, (2rad(G)diam(G))/2=0\lfloor(2rad(G)-diam(G))/2\rfloor=0 holds and that every power of a Helly graph is a Helly graph as well [16]. We know also (see also Section 6) that the Helly-gap of a chordal graph or a distance-hereditary graph is at most 1 and the Helly-gap of a δ\delta-hyperbolic graph is at most 2δ2\delta. Hence, Theorem 3 generalizes some known results on those graphs, too. For every chordal graph GG as well as for every distance-hereditary graph GG, (2rad(G)diam(G))/21\lfloor(2rad(G)-diam(G))/2\rfloor\leq 1 holds [9, 51, 18]. For every δ\delta-hyperbolic graph GG, (2rad(G)diam(G))/22δ\lfloor(2rad(G)-diam(G))/2\rfloor\leq 2\delta holds [11, 24].

4.3 The eccentricity function is almost unimodal in α\alpha-weakly-Helly graphs

The locality loc(v)loc(v) of a vertex vv is defined as the minimum distance from vv to a vertex of strictly smaller eccentricity. Recall that the eccentricity function eG()e_{G}(\cdot) is unimodal in GG if every vertex vV(G)C(G)v\in V(G)\setminus C(G) has a neighbor uu such that eG(u)<eG(v)e_{G}(u)<e_{G}(v), i.e., loc(v)=1loc(v)=1. In a graph GG with a unimodal eccentricity function, any local minimum of the eccentricity function (i.e., a vertex whose eccentricity is not larger than the eccentricity of any of its neighbors) is the global minimum of the eccentricity function in GG (i.e., it is a central vertex). Recall that Helly graphs are characterized by the property that every eccentricity function eGMe_{G}^{M} is unimodal for any MV(G)M\subseteq V(G); therefore, eGM(v)=d(v,CG(M))+radG(M)e_{G}^{M}(v)=d(v,C_{G}(M))+rad_{G}(M) holds (see Lemma 5). A natural question for α\alpha-weakly-Helly graphs is whether similar results on the unimodality of the eccentricity function hold up to a function of α\alpha, that is, if any vertex vV(G)Cα(G)v\in V(G)\setminus C^{\alpha}(G) has loc(v)f(α)loc(v)\leq f(\alpha). The following lemmas answer in the positive.

Lemma 10.

Let GG be an α\alpha-weakly-Helly graph and let MV(G)M\subseteq V(G). If there is a vertex vV(G)v\in V(G) such that eGM(v)>radG(M)+αe_{G}^{M}(v)>rad_{G}(M)+\alpha, then there is a vertex uDG(v,2α+1)u\in D_{G}(v,2\alpha+1) with eGM(u)<eGM(v)e_{G}^{M}(u)<e_{G}^{M}(v).

Proof.

Let S=M{v}S=M\cup\{v\}. Consider in GG a system of disks ={D(u,ρu):uS}\mathcal{F}=\{D(u,\rho_{u}):u\in S\}, where the radii are defined as ρw=eGM(v)1α\rho_{w}=e_{G}^{M}(v)-1-\alpha for any wMw\in M, and ρv=α+1\rho_{v}=\alpha+1. We assert that all disks of \mathcal{F} are pairwise intersecting. Clearly for any wMw\in M, disks D(v,ρv)D(v,\rho_{v}) and D(w,ρw)D(w,\rho_{w}) intersect as d(w,v)eGM(v)d(w,v)\leq e_{G}^{M}(v). We now show that for any w,wMw,w^{\prime}\in M the disks D(w,ρw)D(w,\rho_{w}) and D(w,ρw)D(w^{\prime},\rho_{w^{\prime}}) intersect.

Consider a vertex cCG(M)c\in C_{G}(M). By choice of vv, eGM(v)radG(M)+α+1=eGM(c)+α+1e_{G}^{M}(v)\geq rad_{G}(M)+\alpha+1=e_{G}^{M}(c)+\alpha+1. By the triangle inequality, for any two vertices w,wMw,w^{\prime}\in M we have d(w,w)d(w,c)+d(w,c)eGM(c)+eGM(c)(eGM(v)1α)+(eGM(v)1α)=ρw+ρwd(w,w^{\prime})\leq d(w,c)+d(w^{\prime},c)\leq e_{G}^{M}(c)+e_{G}^{M}(c)\leq(e_{G}^{M}(v)-1-\alpha)+(e_{G}^{M}(v)-1-\alpha)=\rho_{w}+\rho_{w^{\prime}}. Then, by the α\alpha-weakly-Helly property, the system \mathcal{F} of pairwise intersecting disks has a common intersection when radii of all disks are extended by α\alpha. Therefore, there is a vertex uu such that d(u,v)2α+1d(u,v)\leq 2\alpha+1 and d(u,w)eGM(v)1d(u,w)\leq e_{G}^{M}(v)-1 for all wMw\in M. ∎

For α=0\alpha=0 we obtain a result known for Helly graphs. As δ\delta-hyperbolic graphs are (2δ)(2\delta)-weakly-Helly (this follows from a result in [14], see also Section 6), Lemma 10 also greatly generalizes a result known for δ\delta-hyperbolic graphs (see [23]).

Lemma 11.

Let HH be the injective hull of an α\alpha-weakly-Helly graph GG, and let MV(G)M\subseteq V(G). If there is a vertex vV(G)v\in V(G) such that dH(v,CH(M))>αd_{H}(v,C_{H}(M))>\alpha, then there is a real vertex uDG(v,2α+1)u\in D_{G}(v,2\alpha+1) such that eGM(u)<eGM(v)e_{G}^{M}(u)<e_{G}^{M}(v).

Proof.

Let cCH(M)c\in C_{H}(M) be a vertex closest in HH to vv, and assume that dH(v,c)α+1d_{H}(v,c)\geq\alpha+1. Let wV(H)w\in V(H) be a vertex on a shortest (v,c)(v,c)-path of HH such that dH(v,w)=α+1d_{H}(v,w)=\alpha+1. Since HH is Helly, by Proposition 3 and Lemma 5, eGM(v)=eHM(v)=dH(v,CH(M))+radH(M)=dH(v,w)+dH(w,c)+radH(M)=α+1+eHM(w)e_{G}^{M}(v)=e_{H}^{M}(v)=d_{H}(v,C_{H}(M))+rad_{H}(M)=d_{H}(v,w)+d_{H}(w,c)+rad_{H}(M)=\alpha+1+e_{H}^{M}(w). Therefore, eHM(w)=eGM(v)α1e_{H}^{M}(w)=e_{G}^{M}(v)-\alpha-1. Since GG is α\alpha-weakly-Helly, by Theorem 1, there is a vertex uV(G)u\in V(G) such that dH(u,w)αd_{H}(u,w)\leq\alpha. Hence, eGM(u)=eHM(u)eHM(w)+αeGM(v)1e_{G}^{M}(u)=e_{H}^{M}(u)\leq e_{H}^{M}(w)+\alpha\leq e_{G}^{M}(v)-1. See Fig. 4 for an illustration. ∎

[Uncaptioned image]
Figure 4: Illustration to the proof of Lemma 11.

Theorem 4 summarizes the results of Lemma 10 and Lemma 11.

Theorem 4.

Let HH be the injective hull of an α\alpha-weakly-Helly graph GG, and let MV(G)M\subseteq V(G). If there is a vertex vV(G)v\in V(G) such that eGM(v)>radG(M)+αe_{G}^{M}(v)>rad_{G}(M)+\alpha or dH(v,CH(M))>αd_{H}(v,C_{H}(M))>\alpha, then there is a real vertex uDG(v,2α+1)u\in D_{G}(v,2\alpha+1) such that eGM(u)<eGM(v)e_{G}^{M}(u)<e_{G}^{M}(v).

5 Estimating all eccentricities

This section provides upper and lower bounds on the eccentricity of a vertex vv based on a variety of conditions: the distance from vv to a closest almost central vertex (i.e., a closest vertex with eccentricity at most radG(M)+αrad_{G}(M)+\alpha) and whether vv is a farthest vertex from some other vertex and if diam(CG2α(M))diam(C_{G}^{2\alpha}(M)) is bounded. We also prove that any α\alpha-weakly-Helly graph has an eccentricity approximating spanning tree where the additive approximation error depends on the diameter of the set Cα(G)C^{\alpha}(G). Finally, we describe the eccentricity terrain in α\alpha-weakly-Helly graphs, that is, how vertex eccentricities change along vertices of a shortest path to CGα(M)C_{G}^{\alpha}(M).

5.1 Using distances to CGα(M)C_{G}^{\alpha}(M)

Lemma 12.

Let GG be a graph, MV(G)M\subseteq V(G), and k0k\geq 0. For every vertex xV(G)x\in V(G), eGM(x)d(x,CGk(M))+radG(M)+ke_{G}^{M}(x)\leq d(x,C_{G}^{k}(M))+rad_{G}(M)+k holds.

Proof.

Let eGM(x)=d(x,y)e_{G}^{M}(x)=d(x,y) where yFGM(x)y\in F_{G}^{M}(x), and let xpCGk(M)x_{p}\in C_{G}^{k}(M) be closest to xx. By the triangle inequality, eGM(x)d(x,xp)+d(xp,y)d(x,CGk(M))+radG(M)+ke_{G}^{M}(x)\leq d(x,x_{p})+d(x_{p},y)\leq d(x,C_{G}^{k}(M))+rad_{G}(M)+k. ∎

Lemma 13.

Let GG be an α\alpha-weakly-Helly graph and MV(G)M\subseteq V(G). For every vertex xV(G)x\in V(G), eGM(x)d(x,CGα(M))+radG(M)αe_{G}^{M}(x)\geq d(x,C_{G}^{\alpha}(M))+rad_{G}(M)-\alpha holds.

Proof.

Let H:=(G)H:=\mathcal{H}(G) and hCH(M)h\in C_{H}(M) be a closest vertex to xx in HH. By Lemma 5 and Proposition 3, dH(x,h)=dH(x,CH(M))=eHM(x)radH(M)=eGM(x)radH(M)radG(M)radH(M)d_{H}(x,h)=d_{H}(x,C_{H}(M))=e_{H}^{M}(x)-rad_{H}(M)=e_{G}^{M}(x)-rad_{H}(M)\geq rad_{G}(M)-rad_{H}(M). Then, let yy be a vertex on a shortest (x,h)(x,h)-path with dH(h,y)=radG(M)radH(M)d_{H}(h,y)=rad_{G}(M)-rad_{H}(M). By Lemma 5, eHM(y)=dH(y,CH(M))+radH(M)=dH(y,h)+radH(M)=radG(M)e_{H}^{M}(y)=d_{H}(y,C_{H}(M))+rad_{H}(M)=d_{H}(y,h)+rad_{H}(M)=rad_{G}(M). By Theorem 1, there is a real vertex yV(G)y^{*}\in V(G) such that dH(y,y)αd_{H}(y,y^{*})\leq\alpha. Applying Proposition 3 one obtains eGM(y)=eHM(y)radG(M)+αe_{G}^{M}(y^{*})=e_{H}^{M}(y^{*})\leq rad_{G}(M)+\alpha. By the triangle inequality, dH(x,y)dH(x,y)dH(y,y)d(x,CGα(M))αd_{H}(x,y)\geq d_{H}(x,y^{*})-d_{H}(y,y^{*})\geq d(x,C_{G}^{\alpha}(M))-\alpha. Thus, eGM(x)=dH(x,h)+radH(M)=dH(x,y)+radG(M)d(x,CGα(M))+radG(M)αe_{G}^{M}(x)=d_{H}(x,h)+rad_{H}(M)=d_{H}(x,y)+rad_{G}(M)\geq d(x,C_{G}^{\alpha}(M))+rad_{G}(M)-\alpha. ∎

Applying Lemma 12 with k=αk=\alpha and Lemma 13, we obtain the following approximation of eccentricities in α\alpha-weakly-Helly graphs.

Theorem 5.

Let GG be an α\alpha-weakly-Helly graph. For every MV(G)M\subseteq V(G), every xV(G)x\in V(G) satisfies

d(x,CGα(M))+radG(M)αeGM(x)d(x,CGα(M))+radG(M)+α.d(x,C_{G}^{\alpha}(M))+rad_{G}(M)-\alpha\leq e_{G}^{M}(x)\leq d(x,C_{G}^{\alpha}(M))+rad_{G}(M)+\alpha.

If CGα(M)C_{G}^{\alpha}(M) is known in advance, then one obtains a linear time additive 2α2\alpha-approximation of all eccentricities by using a breadth-first search from CGα(M)C_{G}^{\alpha}(M) (BFS(CGα(M))BFS(C_{G}^{\alpha}(M))), to obtain all distances to CGα(M)C_{G}^{\alpha}(M), and a BFS(c)BFS(c) from any fixed cCGα(M)c\in C_{G}^{\alpha}(M) that has a neighbor not in CGα(M)C_{G}^{\alpha}(M), to compute radG(M)+αrad_{G}(M)+\alpha. Then, for each vertex vv, set an approximate eccentricity e^M(v)=dG(v,CGα(M))+radG(M)+α\hat{e}^{M}(v)=d_{G}(v,C_{G}^{\alpha}(M))+rad_{G}(M)+\alpha. By Theorem 5, eGM(v)e^M(v)eGM(v)+2αe_{G}^{M}(v)\leq\hat{e}^{M}(v)\leq e_{G}^{M}(v)+2\alpha.

By Corollary 1, any Helly graph GG satisfies CG(M)=D(CG(M),)C_{G}^{\ell}(M)=D(C_{G}(M),\ell). This also extends to α\alpha-weakly-Helly graphs in the following way.

Corollary 7.

Let GG be an α\alpha-weakly-Helly graph. For any MV(G)M\subseteq V(G) and any integer 0\ell\geq 0, CG(M)DG(CGα(M),α+)C_{G}^{\ell}(M)\subseteq D_{G}(C_{G}^{\alpha}(M),\alpha+\ell).

Proof.

If xCG(M)x\in C_{G}^{\ell}(M) then, by Theorem 5, d(x,CGα(M))+radG(M)αeGM(x)+radG(M)d(x,C_{G}^{\alpha}(M))+rad_{G}(M)-\alpha\leq e_{G}^{M}(x)\leq\ell+rad_{G}(M). Hence, d(x,CGα(M))α+.d(x,C_{G}^{\alpha}(M))\leq\alpha+\ell.

We can restate the lower bound on eGM(x)e_{G}^{M}(x) in Theorem 5 by using thinnes κ((G))\kappa(\mathcal{H}(G)) of metric intervals of a graph’s injective hull.

Lemma 14.

Let HH be the injective hull of a graph GG. For any MV(G)M\subseteq V(G), every xV(G)x\in V(G) satisfies eGM(x)dG(x,CGκ(H)(M))+radG(M)e_{G}^{M}(x)\geq d_{G}(x,C_{G}^{\kappa(H)}(M))+rad_{G}(M).

Proof.

We apply the same idea as in the proof of Lemma 13 with vertex xx, where eGM(x)=eHM(x)=dH(x,t)e_{G}^{M}(x)=e_{H}^{M}(x)=d_{H}(x,t) for some vertex tMt\in M. Let hCH(M)h\in C_{H}(M) be a closest vertex to xx, and let yy be a vertex on a shortest (x,h)(x,h)-path with dH(h,y)=radG(M)radH(M)d_{H}(h,y)=rad_{G}(M)-rad_{H}(M) and eHM(y)=radG(M)e_{H}^{M}(y)=rad_{G}(M). Then, since GG is isometric in HH, there is a shortest (x,t)(x,t)-path PP in HH such that each vertex vPv\in P belongs to V(G)V(G). Note also that, by Lemma 5, hh is on a shortest path from xx to tt in HH. Let yPy^{*}\in P such that dG(y,t)=radG(M)d_{G}(y^{*},t)=rad_{G}(M) so that yy and yy^{*} belong to the same interval slice SradG(M)(t,x)S_{rad_{G}(M)}(t,x) in HH. Therefore, dH(y,y)κ(H)d_{H}(y,y^{*})\leq\kappa(H) and, by Proposition 3, eGM(y)=eHM(y)eHM(y)+κ(H)=radG(M)+κ(H)e_{G}^{M}(y^{*})=e_{H}^{M}(y^{*})\leq e_{H}^{M}(y)+\kappa(H)=rad_{G}(M)+\kappa(H). Hence, eGM(x)=dH(x,y)+radG(M)=dH(x,y)+radG(M)dH(x,CGκ(H)(M))+radG(M)e_{G}^{M}(x)=d_{H}(x,y)+rad_{G}(M)=d_{H}(x,y^{*})+rad_{G}(M)\geq d_{H}(x,C_{G}^{\kappa(H)}(M))+rad_{G}(M) as yCGκ(H)(M)y^{*}\in C_{G}^{\kappa(H)}(M). ∎

As in a δ\delta-hyperbolic graph GG, κ((G))2δ{\kappa({\cal H}(G))}\leq 2\delta (see Subsection 6.2), Lemma 14 generalizes a known result for δ\delta-hyperbolic graphs: eG(x)dG(x,C2δ(G))+rad(G)e_{G}(x)\geq d_{G}(x,C^{2\delta}(G))+rad(G) for all xV(G)x\in V(G) [23]. Note that similar results are known for chordal graphs: eG(x)dG(x,C(G))+rad(G)1e_{G}(x)\geq d_{G}(x,C(G))+rad(G)-1 for every xV(G)x\in V(G) [26], and for distance-hereditary graphs: eG(x)=dG(x,C1(G))+rad(G)+1e_{G}(x)=d_{G}(x,C^{1}(G))+rad(G)+1 for every xV(G)C(G)x\in V(G)\setminus C(G)  [22].

5.2 A vertex furthest from some other vertex

Recall that in a Helly graph GG, for every vertex xVx\in V and every farthest vertex yF(x)y\in F(x), e(y)2rad(G)diam(C(G))e(y)\geq 2rad(G)-diam(C(G)) holds [16].

Theorem 6.

Let GG be an α\alpha-weakly-Helly graph. Then, for every MV(G)M\subseteq V(G) and every xV(G)x\in V(G), each vertex yFGM(x)y\in F_{G}^{M}(x) satisfies

eGM(y)2radG(M)diamG(CG2α(M))2α,e_{G}^{M}(y)\geq 2rad_{G}(M)-diam_{G}(C_{G}^{2\alpha}(M))-2\alpha,
eGM(y)2radG(M)diamG(CGα(M))4α.e_{G}^{M}(y)\geq 2rad_{G}(M)-diam_{G}(C_{G}^{\alpha}(M))-4\alpha.
Proof.

Let H:=(G)H:=\mathcal{H}(G) and let eGM(y)=dG(y,w)e_{G}^{M}(y)=d_{G}(y,w) for some wMw\in M. As eHM()e_{H}^{M}(\cdot) is unimodal, in HH there is a closest to xx vertex bhCH(M)b_{h}\in C_{H}(M) such that eGM(x)=dH(x,bh)+radH(M)e_{G}^{M}(x)=d_{H}(x,b_{h})+rad_{H}(M). Similarly, in HH there is a closest to yy vertex chCH(M)c_{h}\in C_{H}(M) such that eGM(y)=dH(y,ch)+radH(M)e_{G}^{M}(y)=d_{H}(y,c_{h})+rad_{H}(M). Then, dH(x,CH(M))=dH(x,bh)=eHM(x)radH(M)radG(M)radH(M)d_{H}(x,C_{H}(M))=d_{H}(x,b_{h})=e_{H}^{M}(x)-rad_{H}(M)\geq rad_{G}(M)-rad_{H}(M) and, by symmetry, dH(y,ch)radG(M)radH(M)d_{H}(y,c_{h})\geq rad_{G}(M)-rad_{H}(M). Now let cI(ch,y)c\in I(c_{h},y) be a vertex in HH such that dH(ch,c)=radG(M)radH(M)d_{H}(c_{h},c)=rad_{G}(M)-rad_{H}(M), and also bI(bh,x)b\in I(b_{h},x) be a vertex in HH such that dH(bh,b)=radG(M)radH(M)d_{H}(b_{h},b)=rad_{G}(M)-rad_{H}(M), as illustrated in Fig. 5. By Proposition 5, eHM(b)radG(M)radH(M)+eHM(bh)=radG(M)radH(M)+αe_{H}^{M}(b)\leq rad_{G}(M)-rad_{H}(M)+e_{H}^{M}(b_{h})=rad_{G}(M)\leq rad_{H}(M)+\alpha. By symmetry, eHM(c)radH(M)+αe_{H}^{M}(c)\leq rad_{H}(M)+\alpha, and so both bb and cc belong to CHα(M)C_{H}^{\alpha}(M). By the triangle inequality, radG(M)=dH(b,y)dH(b,c)+dH(c,y)=dH(b,c)+dH(w,y)radG(M)rad_{G}(M)=d_{H}(b,y)\leq d_{H}(b,c)+d_{H}(c,y)=d_{H}(b,c)+d_{H}(w,y)-rad_{G}(M). That is, eGM(y)=dH(w,y)2radG(M)dH(b,c)2radG(M)diamH(CHα(M))e_{G}^{M}(y)=d_{H}(w,y)\geq 2rad_{G}(M)-d_{H}(b,c)\geq 2rad_{G}(M)-diam_{H}(C_{H}^{\alpha}(M)). By Corollary 1, as HH is Helly, eGM(y)2radG(M)diamH(CHα(M))=2radG(M)diamH(CH(M))2αe_{G}^{M}(y)\geq 2rad_{G}(M)-diam_{H}(C_{H}^{\alpha}(M))=2rad_{G}(M)-diam_{H}(C_{H}(M))-2\alpha. Applying now Theorem 2 with =α\ell=\alpha, we get diamH(CHα(M))2α+diamG(CG2α(M))diam_{H}(C_{H}^{\alpha}(M))\leq 2\alpha+diam_{G}(C_{G}^{2\alpha}(M)) and hence eGM(y)2radG(M)diamG(CG2α(M))2αe_{G}^{M}(y)\geq 2rad_{G}(M)-diam_{G}(C_{G}^{2\alpha}(M))-2\alpha. Applying Theorem 2 with =0\ell=0, we get diamH(CH(M))2α+diamG(CGα(M))diam_{H}(C_{H}(M))\leq 2\alpha+diam_{G}(C_{G}^{\alpha}(M)) and hence eGM(y)2radG(M)diamG(CGα(M))4αe_{G}^{M}(y)\geq 2rad_{G}(M)-diam_{G}(C_{G}^{\alpha}(M))-4\alpha. ∎

It is known that for every vertex xx, any vertex yFG(x)y\in F_{G}(x) satisfies eG(y)2rad(G)3e_{G}(y)\geq 2rad(G)-3 if GG is a chordal graph or a distance-hereditary graph [18, 12] and eG(y)diam(G)2δ2rad(G)6δ1e_{G}(y)\geq diam(G)-2\delta\geq 2rad(G)-6\delta-1 if GG is a δ\delta-hyperbolic graph[11, 24]. Furthermore, for every chordal or distance-hereditary graph GG, diamG(C(G))3diam_{G}(C(G))\leq 3 holds [50, 51, 9] and for every δ\delta-hyperbolic graph GG, diamG(C2δ(G))8δ+1diam_{G}(C^{2\delta}(G))\leq 8\delta+1 and diamG(C(G))4δ+1diam_{G}(C(G))\leq 4\delta+1 hold [11, 24]. Recall that the diameter of the center of a Helly graph cannot be bounded. Although C(G)C(G) itself is Helly and is isometric in GG for a Helly graph GG [16], C(G)C(G) can have an arbitrarily large diameter as any Helly graph is the center of some Helly graph [16]. Thus, diamG(Cα(G)(G))diam_{G}(C^{\alpha(G)}(G)) cannot be bounded by a function of α(G)\alpha(G) (even for α(G)=0\alpha(G)=0, i.e., for Helly graphs).

[Uncaptioned image]
Figure 5: Illustration to the proof of Theorem 6.

5.3 Eccentricity approximating spanning tree

An eccentricity t-approximating spanning tree is a spanning tree TT such that eT(v)eG(v)te_{T}(v)-e_{G}(v)\leq t holds for all vV(G)v\in V(G). Such a tree tries to approximately preserve only distances from vv to its farthest vertices, allowing for a larger increase in distances to nearby vertices. Many classes of graphs have an eccentricity tt-approximating spanning tree, where tt is a small constant (e.g., see [43, 45, 26, 23, 13] and papers cited therein). We will show that α\alpha-weakly-Helly graphs have an eccentricity tt-approximating spanning tree where tt depends linearly on α\alpha and on the diameter of Cα(G)C^{\alpha}(G).

Lemma 15.

Let HH be the injective hull of an α\alpha-weakly-Helly graph GG. For every MV(G)M\subseteq V(G), there is a real vertex cV(G)c^{*}\in V(G) such that, for every vertex xV(G)x\in V(G), dG(x,c)+eGM(c)eGM(x)diamH(CH(M))2+2αd_{G}(x,c^{*})+e_{G}^{M}(c^{*})-e_{G}^{M}(x)\leq\lceil\frac{diam_{H}(C_{H}(M))}{2}\rceil+2\alpha holds.

Proof.

Let cV(H)c\in V(H) be a vertex belonging to CH(CH(M))C_{H}(C_{H}(M)). By Theorem 1, there is a real vertex cV(G)c^{*}\in V(G) such that dH(c,c)αd_{H}(c,c^{*})\leq\alpha.

We first establish a few useful inequalities. Since HH is Helly, by Lemma 1, CH(M)C_{H}(M) and CH(CH(M))C_{H}(C_{H}(M)) are also Helly and isometric in HH. Thus, by Lemma 4, radH(M)=diamH(M)/2rad_{H}(M)=\lceil diam_{H}(M)/2\rceil and radH(CH(M))=diamH(CH(M))/2rad_{H}(C_{H}(M))=\lceil diam_{H}(C_{H}(M))/2\rceil. Let cxCH(M)c_{x}\in C_{H}(M) be a closest vertex to xx. By Lemma 5, eGM(x)=eHM(x)=dH(x,CH(M))+radH(M)=dH(x,cx)+diamH(M)/2e_{G}^{M}(x)=e_{H}^{M}(x)=d_{H}(x,C_{H}(M))+rad_{H}(M)=d_{H}(x,c_{x})+\lceil diam_{H}(M)/2\rceil. By the triangle inequality, eGM(c)=eHM(c)eHM(c)+α=diamH(M)/2+αe_{G}^{M}(c^{*})=e_{H}^{M}(c^{*})\leq e_{H}^{M}(c)+\alpha=\lceil diam_{H}(M)/2\rceil+\alpha. Since GG is isometric in HH, dG(x,c)=dH(x,c)dH(x,cx)+dH(cx,c)+dH(c,c)dH(x,cx)+radH(CH(M))+α=dH(x,cx)+diamH(CH(M))/2+αd_{G}(x,c^{*})=d_{H}(x,c^{*})\leq d_{H}(x,c_{x})+d_{H}(c_{x},c)+d_{H}(c,c^{*})\leq d_{H}(x,c_{x})+rad_{H}(C_{H}(M))+\alpha=d_{H}(x,c_{x})+\lceil diam_{H}(C_{H}(M))/2\rceil+\alpha. We are now ready to prove the claim.

eGM(x)=eHM(x)\displaystyle e_{G}^{M}(x)=e_{H}^{M}(x) =dH(x,cx)+diamH(M)/2\displaystyle=d_{H}(x,c_{x})+\lceil diam_{H}(M)/2\rceil
(dG(x,c)diamH(CH(M))/2α)+diamH(M)/2\displaystyle\geq(d_{G}(x,c^{*})-\lceil diam_{H}(C_{H}(M))/2\rceil-\alpha)+\lceil diam_{H}(M)/2\rceil
dG(x,c)diamH(CH(M))/2+eGM(c)2α.\displaystyle\geq d_{G}(x,c^{*})-\lceil diam_{H}(C_{H}(M))/2\rceil+e_{G}^{M}(c^{*})-2\alpha.

Therefore, dG(x,c)+eGM(c)eGM(x)diamH(CH(M))/2+2αd_{G}(x,c^{*})+e_{G}^{M}(c^{*})-e_{G}^{M}(x)\leq diam_{H}(C_{H}(M))/2\rceil+2\alpha. ∎

Theorem 7.

Let GG be an α\alpha-weakly-Helly graph. For every MV(G)M\subseteq V(G), GG has a spanning tree TT such that, for all xV(G)x\in V(G), eTM(x)eGM(x)diamG(CGα(M))/2+3αe_{T}^{M}(x)-e_{G}^{M}(x)\leq\lceil diam_{G}(C_{G}^{\alpha}(M))/2\rceil+3\alpha.

Proof.

By Lemma 15, there exists a vertex cV(G)c^{*}\in V(G) such that any vertex xV(G)x\in V(G) satisfies dG(x,c)+eGM(c)eGM(x)diamH(CH(M))/2+2αd_{G}(x,c^{*})+e_{G}^{M}(c^{*})-e_{G}^{M}(x)\leq\lceil diam_{H}(C_{H}(M))/2\rceil+2\alpha. Let TT be a BFS tree of GG rooted at cc^{*}. Then, any vertex xV(G)x\in V(G) has eTM(x)eGM(x)dT(x,c)+eTM(c)eGM(x)e_{T}^{M}(x)-e_{G}^{M}(x)\leq d_{T}(x,c^{*})+e_{T}^{M}(c^{*})-e_{G}^{M}(x). As distances to cc^{*} are preserved in TT, one obtains eTM(x)eGM(x)diamH(CH(M))/2+2αe_{T}^{M}(x)-e_{G}^{M}(x)\leq\lceil diam_{H}(C_{H}(M))/2\rceil+2\alpha. Applying Theorem 2 with =0\ell=0, diamH(CH(M))diamG(CGα(M))+2αdiam_{H}(C_{H}(M))\leq diam_{G}(C_{G}^{\alpha}(M))+2\alpha. Hence, eTM(x)eGM(x)(diamG(CGα(M))+2α)/2)+2αe_{T}^{M}(x)-e_{G}^{M}(x)\leq\lceil(diam_{G}(C_{G}^{\alpha}(M))+2\alpha)/2)\rceil+2\alpha, establishing the desired result. ∎

It is known that every chordal graph and every distance-hereditary graph has an eccentricity 2-approximating spanning tree [45, 26, 19, 39] and every δ\delta-hyperbolic graph has an eccentricity (4δ+1)(4\delta+1)-approximating spanning tree [23]. These graph classes have additional nice properties that allowed to get for them sharper results than the one of Theorem 7 which holds for general α\alpha-weakly-Helly graphs.

5.4 Eccentricity terrain

We can describe the behavior of the eccentricity function in a graph in terms of graph’s eccentricity terrain, that is, how vertex eccentricities change along vertices of a shortest path to CGα(M)C_{G}^{\alpha}(M). We define an ordered pair of vertices (u,v)(u,v), where (u,v)E(u,v)\in E, as an up-edge if eGM(u)<eGM(v)e_{G}^{M}(u)<e_{G}^{M}(v), as a down-edge if eGM(u)>eGM(v)e_{G}^{M}(u)>e_{G}^{M}(v), and as a horizontal-edge if eGM(u)=eGM(v)e_{G}^{M}(u)=e_{G}^{M}(v). Let 𝕌(P(y,x))\mathbb{U}(P(y,x)), (P(y,x))\mathbb{H}(P(y,x)), and 𝔻(P(y,x))\mathbb{D}(P(y,x)) denote the number of up-edges, horizontal-edges, and down-edges along a shortest path P(y,x)P(y,x) from yV(G)y\in V(G) to xV(G)x\in V(G).

We note that Lemma 16 is shown in [23] for the eccentricity function eG()e_{G}(\cdot). For completeness, we provide a proof that it holds for eGM()e_{G}^{M}(\cdot) for any MV(G)M\subseteq V(G).

Lemma 16.

[23] Let GG be an arbitrary graph and MV(G)M\subseteq V(G). For any shortest path P(y,x)P(y,x) of GG from a vertex yy to a vertex xx the following holds:

  • i)

    𝔻(P(y,x))𝕌(P(y,x))=eGM(y)eGM(x)\mathbb{D}(P(y,x))-\mathbb{U}(P(y,x))=e_{G}^{M}(y)-e_{G}^{M}(x), and

  • ii)

    2𝕌(P(y,x))+(P(y,x))=dG(y,x)(eGM(y)eGM(x))2\mathbb{U}(P(y,x))+\mathbb{H}(P(y,x))=d_{G}(y,x)-(e_{G}^{M}(y)-e_{G}^{M}(x)).

Proof.

We use an induction on dG(y,v)d_{G}(y,v) for any vertex vP(y,x)v\in P(y,x). First, assume that vv is adjacent to yy. If (y,v)(y,v) is an up-edge, then eGM(y)eGM(v)=1e_{G}^{M}(y)-e_{G}^{M}(v)=-1 and 𝔻(P(y,v))𝕌(P(y,v))=1\mathbb{D}(P(y,v))-\mathbb{U}(P(y,v))=-1. If (y,v)(y,v) is a horizontal-edge, then eGM(y)eGM(v)=0e_{G}^{M}(y)-e_{G}^{M}(v)=0 and 𝔻(P(y,v))𝕌(P(y,v))=0\mathbb{D}(P(y,v))-\mathbb{U}(P(y,v))=0. If (y,v)(y,v) is a down-edge, then eGM(y)eGM(v)=1e_{G}^{M}(y)-e_{G}^{M}(v)=1 and 𝔻(P(y,v))𝕌(P(y,v))=1\mathbb{D}(P(y,v))-\mathbb{U}(P(y,v))=1. Now consider an arbitrary vertex vP(y,x)v\in P(y,x) and assume, by induction, that eGM(y)eGM(v)=𝔻(P(y,v))𝕌(P(y,v))e_{G}^{M}(y)-e_{G}^{M}(v)=\mathbb{D}(P(y,v))-\mathbb{U}(P(y,v)). Let vertex uP(y,x)u\in P(y,x) be adjacent to vv with d(y,u)=d(y,v)+1d(y,u)=d(y,v)+1. By definition, 𝔻(P(y,u))=𝔻(P(y,v))+𝔻((v,u))\mathbb{D}(P(y,u))=\mathbb{D}(P(y,v))+\mathbb{D}((v,u)) and 𝕌(P(y,u))=𝕌(P(y,v))+𝕌((v,u))\mathbb{U}(P(y,u))=\mathbb{U}(P(y,v))+\mathbb{U}((v,u)). We consider three cases based on the classification of edge (v,u)(v,u).

If (v,u)(v,u) is an up-edge, then eGM(u)=eGM(v)+1e_{G}^{M}(u)=e_{G}^{M}(v)+1, 𝕌((v,u))=1\mathbb{U}((v,u))=1, and 𝔻((v,u))=0=𝕌((v,u))1\mathbb{D}((v,u))=0=\mathbb{U}((v,u))-1. By the inductive hypothesis, 𝔻(P(y,u))=𝔻(P(y,v))+𝔻((v,u))=𝕌(P(y,v))+eGM(y)eGM(v)+𝕌((v,u))1=𝕌(P(y,u))+eGM(y)eGM(v)1=𝕌(P(y,u))+eGM(y)eGM(u)\mathbb{D}(P(y,u))=\mathbb{D}(P(y,v))+\mathbb{D}((v,u))=\mathbb{U}(P(y,v))+e_{G}^{M}(y)-e_{G}^{M}(v)+\mathbb{U}((v,u))-1=\mathbb{U}(P(y,u))+e_{G}^{M}(y)-e_{G}^{M}(v)-1=\mathbb{U}(P(y,u))+e_{G}^{M}(y)-e_{G}^{M}(u).

If (v,u)(v,u) is a horizontal-edge, then eGM(u)=eGM(v)e_{G}^{M}(u)=e_{G}^{M}(v), 𝕌((v,u))=0=𝔻((v,u))\mathbb{U}((v,u))=0=\mathbb{D}((v,u)). By the inductive hypothesis, 𝔻(P(y,u))=𝔻(P(y,v))+𝔻((v,u))=𝕌(P(y,v))+eGM(y)eGM(v)+𝕌((v,u))=𝕌(P(y,u))+eGM(y)eGM(v)=𝕌(P(y,u))+eGM(y)eGM(u)\mathbb{D}(P(y,u))=\mathbb{D}(P(y,v))+\mathbb{D}((v,u))=\mathbb{U}(P(y,v))+e_{G}^{M}(y)-e_{G}^{M}(v)+\mathbb{U}((v,u))=\mathbb{U}(P(y,u))+e_{G}^{M}(y)-e_{G}^{M}(v)=\mathbb{U}(P(y,u))+e_{G}^{M}(y)-e_{G}^{M}(u).

If (v,u)(v,u) is a down-edge, then eGM(u)=eGM(v)1e_{G}^{M}(u)=e_{G}^{M}(v)-1, 𝕌((v,u))=0\mathbb{U}((v,u))=0, and 𝔻((v,u))=𝕌((v,u))+1\mathbb{D}((v,u))=\mathbb{U}((v,u))+1. By the inductive hypothesis, 𝔻(P(y,u))=𝔻(P(y,v))+𝔻((v,u))=𝕌(P(y,v))+eGM(y)eGM(v)+𝕌((v,u))+1=𝕌(P(y,u))+eGM(y)eGM(v)+1=𝕌(P(y,u))+eGM(y)eGM(u)\mathbb{D}(P(y,u))=\mathbb{D}(P(y,v))+\mathbb{D}((v,u))=\mathbb{U}(P(y,v))+e_{G}^{M}(y)-e_{G}^{M}(v)+\mathbb{U}((v,u))+1=\mathbb{U}(P(y,u))+e_{G}^{M}(y)-e_{G}^{M}(v)+1=\mathbb{U}(P(y,u))+e_{G}^{M}(y)-e_{G}^{M}(u).

This completes the proof of (i) that 𝔻(P(y,x))𝕌(P(y,x))=eGM(y)eGM(x)\mathbb{D}(P(y,x))-\mathbb{U}(P(y,x))=e_{G}^{M}(y)-e_{G}^{M}(x). To complete the proof of (ii), it now suffices to see that 𝕌(P(y,x))+(P(y,x))+(𝕌(P(y,x))+eGM(y)eGM(x))=𝕌(P(y,x))+(P(y,x))+𝔻(P(y,x))=dG(y,x)\mathbb{U}(P(y,x))+\mathbb{H}(P(y,x))+(\mathbb{U}(P(y,x))+e_{G}^{M}(y)-e_{G}^{M}(x))=\mathbb{U}(P(y,x))+\mathbb{H}(P(y,x))+\mathbb{D}(P(y,x))=d_{G}(y,x). Hence, 2𝕌(P(y,x))+(P(y,x))=dG(y,x)(eGM(y)eGM(x))2\mathbb{U}(P(y,x))+\mathbb{H}(P(y,x))=d_{G}(y,x)-(e_{G}^{M}(y)-e_{G}^{M}(x)). ∎

Theorem 8.

Let GG be an α\alpha-weakly-Helly graph and MV(G)M\subseteq V(G). For any shortest path P(y,x)P(y,x) of GG from a vertex yCGα(M)y\notin C_{G}^{\alpha}(M) to a closest vertex xCGα(M)x\in C_{G}^{\alpha}(M), 2𝕌(P(y,x))+(P(y,x))2α2\mathbb{U}(P(y,x))+\mathbb{H}(P(y,x))\leq 2\alpha holds.

Proof.

Let xCGα(M)x\in C_{G}^{\alpha}(M) be closest to yy. Hence, eGM(x)=radG(M)+αe_{G}^{M}(x)=rad_{G}(M)+\alpha. Applying Theorem 5, eGM(y)dG(y,CGα(M))+radG(M)αe_{G}^{M}(y)\geq d_{G}(y,C_{G}^{\alpha}(M))+rad_{G}(M)-\alpha. Hence, dG(y,x)=dG(y,CGα(M))eGM(y)radG(M)+α=eGM(y)eGM(x)+2αd_{G}(y,x)=d_{G}(y,C_{G}^{\alpha}(M))\leq e_{G}^{M}(y)-rad_{G}(M)+\alpha=e_{G}^{M}(y)-e_{G}^{M}(x)+2\alpha. Applying Lemma 16, we obtain the desired result. ∎

As a consequence of Theorem 8, at most 2α(G)2\alpha(G) non-descending edges can occur along every shortest path from any vertex yCα(G)y\notin C^{\alpha}(G) to Cα(G)C^{\alpha}(G). Hence, in any shortest path to Cα(G)C^{\alpha}(G), the number of vertices with locality more than 1 does not exceed 2α(G)2\alpha(G). These kind of results were only known for chordal graphs and distance-hereditary graphs (at most one non-descending edge, i.e., horizontal-edge [26, 22]) and for δ\delta-hyperbolic graphs (at most 4δ4\delta non-descending edges [23]).

6 Classes of graphs having small Helly-gap

Let GG be a graph and let α:=α(G)\alpha:=\alpha(G) be its Helly-gap. We now focus on the upper bound on α\alpha for some special graph classes and identify those classes which exhibit a topological Helly-likeness, that is, classes which have small α\alpha. Recall that, by definition, Helly graphs are 0-weakly-Helly. It was also recently shown that each Helly vertex in the injective hull of a distance-hereditary graph is adjacent to a real vertex [35]; thus distance-hereditary graphs are 1-weakly-Helly (this result follows also from an earlier result on existence of rr-dominating cliques in distance-hereditary graphs[18]). Distance-hereditary graphs can be defined as the graphs where each induced path is a shortest path.

We consider several other graph classes and summarize our findings in Table 1.

Graph class Helly-gap
Helly α(G)=0\alpha(G)=0
Distance-hereditary α(G)1\alpha(G)\leq 1
kk-Chordal α(G)k/2\alpha(G)\leq\lfloor k/2\rfloor
Chordal α(G)1\alpha(G)\leq 1
AT-free α(G)2\alpha(G)\leq 2
n×nn\times n Rectilinear grid α(G)=1\alpha(G)=1
δ\delta-Hyperbolic α(G)2δ\alpha(G)\leq 2\delta
Cycle CnC_{n} α(G)=n/4\alpha(G)=\lfloor n/4\rfloor
αi\alpha_{i}-Metric α(G)i/2\alpha(G)\leq\lceil i/2\rceil
Tree-breadth tb(G)tb(G) α(G)tb(G)\alpha(G)\leq tb(G)
Tree-length tl(G)tl(G) α(G)tl(G)\alpha(G)\leq tl(G)
Tree-width tw(G)tw(G) no relation
Bridged unbounded
Table 1: The bound on the Helly-gap α(G)\alpha(G) for various graph classes.

6.1 Rectilinear grids

Let GG be an n×nn\times n rectilinear grid (the cartesian product of two paths of length nn). GG can be isometrically embedded into a 2n×2n2n\times 2n king grid HH (the strong product of two paths of length 2n2n, and a natural subclass of Helly graphs) wherein the four extreme vertices of GG are the midpoints of each side of HH (see Fig. 6). One obtains (G)\mathcal{H}(G) by removing from HH any peripheral vertices that are not present in GG. Therefore, each Helly vertex hh of (G)\mathcal{H}(G) corresponds to one of the n2n^{2} squares of the grid GG, where hh is adjacent to the four corners of the square and to each of the Helly vertices corresponding to the 2-4 adjacent squares. Hence, the Helly-gap of GG is 1.

[Uncaptioned image]
Figure 6: An n×nn\times n rectilinear grid (shown in bold blue lines) is isometrically embedded into a 2n×2n2n\times 2n king grid (shown in thin black lines).

6.2 Graphs of bounded hyperbolicity

Gromov [33] defines δ\delta-hyperbolic graphs via a simple 4-point condition: for any four vertices u,v,w,xu,v,w,x, the two larger of the three distance sums d(u,v)+d(w,x)d(u,v)+d(w,x), d(u,w)+d(v,x)d(u,w)+d(v,x), and d(u,x)+d(v,w)d(u,x)+d(v,w) differ by at most 2δ02\delta\geq 0. The smallest value δ\delta for which GG is δ\delta-hyperbolic is called the hyperbolicity of GG and is denoted by δ(G)\delta(G). As we have mentioned earlier, we omit GG from subindices when the graph is known by context.

Recall that κ(G)\kappa(G) denotes the smallest value κ\kappa for which all intervals of GG are κ\kappa-thin. The following lemma is a folklore and easy to show using the definition of hyperbolicity.

Lemma 17.

For any graph GG, κ(G)2δ(G)\kappa(G)\leq{2}\delta(G).

Proof.

Consider any interval I(x,y)I(x,y) in GG and arbitrary two vertices u,vSk(x,y)u,v\in S_{k}(x,y). Consider the three distance sums S1=d(x,y)+d(u,v)S_{1}=d(x,y)+d(u,v), S2=d(x,u)+d(y,v)S_{2}=d(x,u)+d(y,v), S3=d(x,v)+d(y,u)S_{3}=d(x,v)+d(y,u). As u,vSk(x,y)u,v\in S_{k}(x,y), we have S2=S3=d(x,y)S1S_{2}=S_{3}=d(x,y)\leq S_{1}. Hence, 2δ(G)S1S2=d(x,y)+d(u,v)d(x,y)=d(u,v)2\delta(G)\geq S_{1}-S_{2}=d(x,y)+d(u,v)-d(x,y)=d(u,v) for any two vertices from the same slice of GG, i.e., 2δ(G)κ(G)2\delta(G)\geq\kappa(G). ∎

We now show that the Helly-gap of any graph GG is at most κ((G))\kappa(\mathcal{H}(G)).

Lemma 18.

For any vertex hV((G))h\in V(\mathcal{H}(G)) there is a real vertex vV(G)v\in V(G) such that d(h,v)κ((G))d(h,v)\leq\kappa(\mathcal{H}(G)).

Proof.

Let H:=(G)H:=\mathcal{H}(G) and consider a vertex hV(H)h\in V(H). By Proposition 2, hh belongs to a shortest path PH(x,y)P_{H}(x,y) where x,yV(G)x,y\in V(G). Let dH(x,h)=d_{H}(x,h)=\ell. Because GG is isometric in HH, there is a shortest path PG(x,y):=(v0,.,vk)P_{G}(x,y):=(v_{0},....,v_{k}), where x=v0x=v_{0}, y=vky=v_{k} and all viv_{i} in PGP_{G} are real vertices. By assumption, dH(h,v)κ(H)d_{H}(h,v_{\ell})\leq\kappa(H). ∎

Corollary 8.

Any graph GG is α\alpha-weakly-Helly for ακ((G))\alpha\leq\kappa(\mathcal{H}(G)).

Proof.

The result follows from Theorem 1 and Lemma 18. ∎

Note that an n×nn\times n rectilinear grid gives an example of a graph where α=1\alpha=1 and κ((G))=2n\kappa(\mathcal{H}(G))=2n.

Urs Lang [40] has established that the δ\delta-hyperbolicity of a graph GG is preserved in H=(G)H=\mathcal{H}(G). As hyperbolicity is preserved in (G)\mathcal{H}(G), by Lemma 18 and Lemma 17, for any Helly vertex hV(H)h\in V(H) there is a real vertex vV(G)v\in V(G) such that dH(h,v)κ((G))2δ((G))=2δ(G)d_{H}(h,v)\leq\kappa(\mathcal{H}(G))\leq 2\delta(\mathcal{H}(G))=2\delta(G). Hence, δ\delta-hyperbolic graphs are 2δ2\delta-weakly-Helly. The latter result follows also from [14].

Remark.

Several graph classes are δ\delta-hyperbolic for a bounded value δ\delta, including kk-chordal graphs and graphs without an asteroidal triple (AT). Three pairwise non-adjacent vertices of a graph form an AT if every two of them are connected by a path that avoids the neighborhood of the third. A graph is kk-chordal provided it has no induced cycle of length greater than kk. It is known [49] that kk-chordal graphs have hyperbolicity at most k/22\frac{\lfloor k/2\rfloor}{2}. Therefore, kk-chordal graphs are k/2\lfloor k/2\rfloor-weakly-Helly. Chordal graphs, which are exactly the 3-chordal graphs, are 1-weakly-Helly (this result follows also from an earlier result on existence of rr-dominating cliques in chordal graphs[20]). The 2-weakly-Helly graphs also include all 5-chordal graphs such as AT-free graphs, all 4-chordal graphs such as weakly-chordal and cocomparability graphs, as well as all 1-hyperbolic graphs including interval graphs and permutation graphs. The definitions of graph classes not given here can be found in [6].

Remark.

Observe that α(G)\alpha(G) can be arbitrarily smaller than the δ\delta-hyperbolicity of a graph. Consider a (2r×2r)(2r\times 2r) king grid GG, i.e., the strong product of two paths each of even length 2r2r. King grids form a natural subclass of Helly graphs, and therefore (G)=G\mathcal{H}(G)=G. Thus, GG is 0-weakly-Helly, whereas δ(G)=r\delta(G)=r.

6.3 Cycles

Recall that CnC_{n} denotes a cycle of length nn.

Lemma 19.

The Helly-gap of a cycle CnC_{n} is n/4\lfloor n/4\rfloor.

Proof.

Let GG be a cycle of length nn. Clearly, diam(G)=n/2=rad(G)diam(G)=\lfloor n/2\rfloor=rad(G). By Proposition 5, GG is α\alpha-weakly-Helly for αdiam(G)/2=n/4\alpha\leq\lfloor diam(G)/2\rfloor=\lfloor n/4\rfloor. By Lemma 8, α(2rad(G)diam(G))/2=n/4\alpha\geq\lfloor(2rad(G)-diam(G))/2\rfloor=\lfloor n/4\rfloor. ∎

It should also be noted that for a self-centered graph (i.e., a graph where all vertex eccentricities are equal), from (2rad(G)diam(G))/2α(G)diam(G)/2\lfloor(2rad(G)-diam(G))/2\rfloor\leq\alpha(G)\leq\lfloor diam(G)/2\rfloor and since for a self-centered graph GG, rad(G)=diam(G)rad(G)=diam(G), we get α(G)=diam(G)/2=rad(G)/2\alpha(G)=\lfloor diam(G)/2\rfloor=\lfloor rad(G)/2\rfloor. Cycles are self-centered graphs.

6.4 Graphs with an αi\alpha_{i}-metric

Introduced by Chepoi and Yushmanov [9, 51], a graph GG is said to have an αi\alpha_{i}-metric if it satisfies the following: for any x,y,z,vV(G)x,y,z,v\in V(G) such that zI(x,y)z\in I(x,y) and yI(z,v)y\in I(z,v), dG(x,v)dG(x,y)+dG(y,v)id_{G}(x,v)\geq d_{G}(x,y)+d_{G}(y,v)-i holds. For every graph GG with an αi\alpha_{i}-metric, diam(G)2rad(G)i1diam(G)\geq 2rad(G)-i-1 holds [51]. Several graph classes have an αi\alpha_{i}-metric [51]. Ptolemaic graphs are exactly the graphs with α0\alpha_{0}-metric [51]. Chordal graphs are a subclass of graphs with α1\alpha_{1}-metric [9]. All graphs with α1\alpha_{1}-metric are characterized in [51]. In a private communication, discussing the results of this paper, Victor Chepoi asked if graphs with an αi\alpha_{i}-metric are α\alpha-weakly-Helly for an α\alpha depending only on ii. The following lemma answers this question in the affirmative.

Lemma 20.

Any graph GG with an αi\alpha_{i}-metric is α\alpha-weakly-Helly for αi/2\alpha\leq\lceil i/2\rceil.

Proof.

We prove by induction on the number kk of disks in a family of pairwise intersecting disks ={D(v,r(v)):vMV(G)}\mathcal{F}=\{D(v,r(v)):v\in M\subseteq V(G)\}. Let yMy\in M and pick a vertex cc which belongs to vM{y}D(v,r(v)+α)\bigcap_{v\in M\setminus\{y\}}D(v,r(v)+\alpha) such that cc is closest to yy. If d(c,y)r(y)+αd(c,y)\leq r(y)+\alpha, then cc is common to all disks of \mathcal{F} inflated by α\alpha and we are done. Assume now that d(c,y)>r(y)+αd(c,y)>r(y)+\alpha. Let cS1(c,y)c^{\prime}\in S_{1}(c,y). By choice of cc, there is a disk D(x,r(x))D(x,r(x))\in\mathcal{F} such that d(c,x)=r(x)+α=d(c,x)1d(c,x)=r(x)+\alpha=d(c^{\prime},x)-1. Hence, cI(x,c)c\in I(x,c^{\prime}) and cI(c,y)c^{\prime}\in I(c,y). Applying αi\alpha_{i}-metric to x,c,c,yx,c,c^{\prime},y, one obtains d(x,y)d(x,c)+d(c,y)i>(r(x)+α)+(r(y)+α)i=r(x)+r(y)+2αid(x,y)\geq d(x,c)+d(c,y)-i>(r(x)+\alpha)+(r(y)+\alpha)-i=r(x)+r(y)+2\alpha-i. If now i2αi\leq 2\alpha, we get d(x,y)>r(x)+r(y)d(x,y)>r(x)+r(y), contradicting the fact that disks D(x,r(x))D(x,r(x)) and D(y,r(y))D(y,r(y)) intersect. Thus, for every ii (even or odd) such that i2αi\leq 2\alpha, d(c,y)r(y)+αd(c,y)\leq r(y)+\alpha must hold. That is, GG is α\alpha-weakly-Helly for α=i/2\alpha=\lceil i/2\rceil (pick α=i/2\alpha=i/2, when ii is even, and α=(i+1)/2\alpha=(i+1)/2, when ii is odd). ∎

Remark.

Observe that a graph GG with an αi\alpha_{i}-metric can have Helly-gap α(G)\alpha(G) that is arbitrarily smaller than i/2\lceil i/2\rceil. Consider a (1×)(1\times\ell) rectilinear grid GG. Let (x,y)(x,y) and (z,v)(z,v) be edges on extreme ends of GG so that d(x,z)=d(y,v)=d(x,z)=d(y,v)=\ell and d(y,z)=d(x,v)=+1d(y,z)=d(x,v)=\ell+1. Then, zI(x,v)z\in I(x,v), vI(z,y)v\in I(z,y), and 1=d(x,y)d(x,v)+d(v,y)i=2+1i1=d(x,y)\geq d(x,v)+d(v,y)-i=2\ell+1-i. Thus, GG has an αi\alpha_{i}-metric for i2i\geq 2\ell whereas α(G)=1\alpha(G)=1.

6.5 Graphs of bounded tree-breadth, tree-length, or tree-width

A tree-decomposition (𝒯,T)(\mathcal{T},T) [48] for a graph G=(V,E)G=(V,E) is a family 𝒯={B1,B2,}\mathcal{T}=\{B_{1},B_{2},...\} of subsets of VV, called bags, such that 𝒯\mathcal{T} forms a tree TT with the bags in 𝒯\mathcal{T} as nodes which satisfy the following conditions: (i) each vertex is contained in a bag, (ii) for each edge (u,v)E(u,v)\in E, 𝒯\mathcal{T} has a bag BB with u,vBu,v\in B, and (iii) for each vertex vVv\in V, the bags containing vv induce a subtree of TT. The width of a tree decomposition is the size of its largest bag minus one. A tree decomposition has breadth ρ\rho if, for each bag BB, there is a vertex vv in GG such that BDG(v,ρ)B\subseteq D_{G}(v,\rho). A tree decomposition has length λ\lambda if the diameter in GG of each bag BB is at most λ\lambda. The tree-width tw(G)tw(G) [48], tree-breadth tb(G)tb(G) [25] and tree-length tl(G)tl(G) [15] are the minimum width, breadth, and length, respectively, among all possible tree decompositions of GG. By definition, tb(G)tl(G)2tb(G)tb(G)\leq tl(G)\leq 2tb(G), as for any graph GG and any set MV(G)M\subseteq V(G), radG(M)diamG(M)2radG(M)rad_{G}(M)\leq diam_{G}(M)\leq 2rad_{G}(M) holds.

Lemma 21.

Any graph GG is α\alpha-weakly-Helly for αtb(G)\alpha\leq tb(G) and αtl(G)\alpha\leq tl(G).

Proof.

Let GG have tree-breadth tb(G)ρtb(G)\leq\rho. Consider a corresponding tree-decomposition (𝒯,T)(\mathcal{T},T) of GG of breadth ρ\rho, and let ={DG(v,r(v)):vSV(G)}\mathcal{F}=\{D_{G}(v,r(v)):v\in S\subseteq V(G)\} be a family of disks of GG that pairwise intersect. Observe that bags containing vertices of a disk induce a subtree of TT and the subtrees of TT corresponding to disks of \mathcal{F} pairwise intersect in TT. If the subtrees of a tree pairwise intersect, then they have a common node in TT. Therefore, there is a bag B𝒯B\in\mathcal{T} such that each disk in \mathcal{F} intersects BB in GG. Let vertex ww be the center of bag BB, i.e., BDG(w,ρ)B\subseteq D_{G}(w,\rho). So, each vSv\in S satisfies wDG(v,r(v)+ρ)w\in D_{G}(v,r(v)+\rho). Hence, GG is α\alpha-weakly-Helly where αtb(G)tl(G)\alpha\leq tb(G)\leq tl(G). ∎

Remark.

While α:=α(G)\alpha:=\alpha(G) is upper bounded by tree-breadth and tree-length, α\alpha can be arbitrarily far from tree-width and arbitrarily smaller than tree-length. Consider the (r×r)(r\times r) rectilinear grid GG which is 1-weakly-Helly but tw(G)=rtw(G)=r (cf. [47]) and tl(G)=2rtl(G)=2r (cf. [15, 1]). On the other hand, let GG be a cycle of size 4k4k; the Helly-gap of GG is kk, yet tw(G)=2tw(G)=2.

6.6 Bridged graphs

A graph is bridged[32] if it contains no isometric cycles of length greater than 3. As any isometric subgraph is an induced subgraph, bridged graphs form a superclass of chordal graphs. Interestingly, we find that although chordal graphs are 1-weakly-Helly, the Helly-gap of a bridged graph can be arbitrarily large. Consider the bridged graph GG in Fig. 7 with each side of length s=4ks=4k for some integer kk. Clearly the disks centered at x,y,zx,y,z with radius 2k2k pairwise intersect and have no common intersection. However, only extending all radii by kk yields middle vertex uu as a common intersection. Thus, the Helly-gap of GG is at least kk, where kk can be arbitrarily large.

[Uncaptioned image]
Figure 7: Example that bridged graphs have arbitrarily large Helly-gap.

7 Conclusion

We introduced a new metric parameter for a graph, the Helly-gap α(G)\alpha(G). We characterized α\alpha-weakly-Helly graphs with respect to distances in their injective hulls and used this as a tool to generalize many eccentricity related results known for Helly graphs to much larger family of graphs, α\alpha-weakly-Helly graphs. In particular, we related the diameter, radius, center, and all eccentricities in GG to their counterparts in (G)\mathcal{H}(G). We provided estimates on α(G)\alpha(G) and on eccentricities based on a variety of conditions, thereby generalizing some eccentricity related results known not only for Helly graphs but also for distance-hereditary graphs, chordal graphs, and δ\delta-hyperbolic graphs. Several additional graph classes are identified which have a bounded Helly-gap.

Two interesting questions remain open. First, is there a characterization of α(G)\alpha(G) using only the radius and diameter of every MV(G)M\subseteq V(G)? Is it true that α(G)=maxMV(G)(2radG(M)diamG(M))/2\alpha(G)=\max_{M\subseteq V(G)}\lfloor(2rad_{G}(M)-diam_{G}(M))/2\rfloor? As it was recently shown [21], the equality maxMV(G)(2radG(M)diamG(M))/2=0\max_{M\subseteq V(G)}\lfloor(2rad_{G}(M)-diam_{G}(M))/2\rfloor=0 characterizes the Helly graphs (i.e., the graphs with α(G)=0\alpha(G)=0): GG is a Helly graph if and only if maxMV(G)(2radG(M)diamG(M))/2=0\max_{M\subseteq V(G)}\lfloor(2rad_{G}(M)-diam_{G}(M))/2\rfloor=0. Second, as α\alpha-weakly-Helly graphs are a far reaching superclass of δ\delta-hyperbolic graphs, under what additional conditions are α\alpha-weakly-Helly graphs reduced to δ\delta-hyperbolic graphs for some δ\delta dependent on α\alpha?

Acknowledgment: We are grateful to Victor Chepoi for reading an earlier version of this paper and providing a few interesting remarks.

References

  • [1] A. B. Adcock, B. D. Sullivan, and M. W. Mahoney. Tree decompositions and social graphs. Internet Mathematics, 12(5):315–361, May 2016.
  • [2] N. Alon, G. Kalai, J. Matoušek, and R. Meshulam. Transversal numbers for hypergraphs arising in geometry. Advances in Applied Mathematics, 29(1):79 – 101, 2002.
  • [3] N. Alon and D. J. Kleitman. Piercing convex sets and the Hadwiger-Debrunner (p, q)-problem. Advances in Mathematics, 96(1):103 – 112, 1992.
  • [4] H.-J. Bandelt and E. Pesch. Dismantling absolute retracts of reflexive graphs. Eur. J. Comb., 10(3):211–220, May 1989.
  • [5] H.-J. Bandelt and E. Prisner. Clique graphs and helly graphs. Journal of Combinatorial Theory, Series B, 51(1):34 – 45, 1991.
  • [6] A. Brandstädt, V. B. Le, and J. P. Spinrad. Graph Classes: A Survey. Society for Industrial and Applied Mathematics, 1999.
  • [7] I. Bárány and J. Matoušek. A fractional Helly theorem for convex lattice sets. Advances in Mathematics, 174(2):227 – 235, 2003.
  • [8] J. Chalopin, V. Chepoi, A. Genevois, H. Hirai, and D. Osajda. Helly groups. arXiv:2002.06895, 2020.
  • [9] V. Chepoi. Centers of triangulated graphs. Mathematical Notes of the Academy of Sciences of the USSR, 43:82–86, 1988.
  • [10] V. Chepoi. A note on r-dominating cliques. Discret. Math., 183(1-3):47–60, 1998.
  • [11] V. Chepoi, F. Dragan, B. Estellon, M. Habib, and Y. Vaxès. Diameters, centers, and approximating trees of delta-hyperbolic geodesic spaces and graphs. Symposium on Computational Geometry, 2008.
  • [12] V. Chepoi and F. F. Dragan. A linear-time algorithm for finding a central vertex of a chordal graph. In Algorithms - ESA, pages 159–170, 1994.
  • [13] V. Chepoi, F. F. Dragan, M. Habib, Y. Vaxès, and H. Alrasheed. Fast approximation of eccentricities and distances in hyperbolic graphs. J. Graph Algorithms Appl., 23(2):393–433, 2019.
  • [14] V. Chepoi and B. Estellon. Packing and covering delta-hyperbolic spaces by balls. In APPROX-RANDOM, pages 59–73, 2007.
  • [15] Y. Dourisboure and C. Gavoille. Tree-decompositions with bags of small diameter. Discrete Mathematics, 307(16):2008 – 2029, 2007. EuroComb ’03 - Graphs and Algorithms.
  • [16] F. F. Dragan. Centers of Graphs and the Helly Property (in Russian). PhD thesis, Moldava State University, Chişinău, 1989.
  • [17] F. F. Dragan. Conditions for coincidence of local and global minima for eccentricity function on graphs and the Helly property (in Russian). Studies in Applied Mathematics and Information Science, pages 49–56, 1990.
  • [18] F. F. Dragan. Dominating cliques in distance-hereditary graphs. In Algorithm Theory — SWAT, pages 370–381, 1994.
  • [19] F. F. Dragan. An eccentricity 2-approximating spanning tree of a chordal graph is computable in linear time. Inf. Process. Lett., 154, 2020.
  • [20] F. F. Dragan and A. Brandstädt. r-dominating cliques in graphs with hypertree structure. Discret. Math., 162(1-3):93–108, 1996.
  • [21] F. F. Dragan, G. Ducoffe, and H. M. Guarnera. Fast deterministic algorithms for computing all eccentricities in (hyperbolic) Helly graphs. Manuscript, 2020.
  • [22] F. F. Dragan and H. M. Guarnera. Eccentricity function in distance-hereditary graphs. CoRR, abs/1907.05445, 2019.
  • [23] F. F. Dragan and H. M. Guarnera. Eccentricity terrain of δ\delta-hyperbolic graphs. Journal of Computer and System Sciences, 2020.
  • [24] F. F. Dragan, M. Habib, and L. Viennot. Revisiting radius, diameter, and all eccentricity computation in graphs through certificates. CoRR, abs/1803.04660, 2018.
  • [25] F. F. Dragan and E. Köhler. An approximation algorithm for the tree t-spanner problem on unweighted graphs via generalized chordal graphs. Algorithmica, 69(4):884–905, 2014.
  • [26] F. F. Dragan, E. Köhler, and H. Alrasheed. Eccentricity approximating trees. Discrete Applied Mathematics, 232:142–156, 2017.
  • [27] A. W. Dress. Trees, tight extensions of metric spaces, and the cohomological dimension of certain groups: A note on combinatorial properties of metric spaces. Advances in Mathematics, 53(3):321 – 402, 1984.
  • [28] G. Ducoffe. Facility Location problems on Helly graphs and k-Helly graphs. Manuscript, 2020.
  • [29] G. Ducoffe and F. F. Dragan. A story of diameter, radius and Helly property. CoRR, abs/1910.10412, 2019.
  • [30] J. Eckhoff. Chapter 2.1 - Helly, Radon, and Carathéodory Type Theorems. In P. Gruber and J. Wills, editors, Handbook of Convex Geometry, pages 389 – 448. North-Holland, Amsterdam, 1993.
  • [31] R. Espínola and M. A. Khamsi. Introduction to Hyperconvex Spaces, pages 391–435. Springer Netherlands, Dordrecht, 2001.
  • [32] M. Farber. Bridged graphs and geodesic convexity. Discrete Mathematics, 66(3):249 – 257, 1987.
  • [33] M. Gromov. Hyperbolic Groups, pages 75–263. Springer, New York, NY, 1987.
  • [34] M. Groshaus, M. C. Lin, and J. L. Szwarcfiter. On neighborhood-Helly graphs. Discret. Appl. Math., 216:191–202, 2017.
  • [35] H. M. Guarnera, F. F. Dragan, and A. Leitert. Injective hulls of various graph classes. Manuscript, 2020.
  • [36] J. Isbell. Six theorems about injective metric spaces. Commentarii mathematici Helvetici, 39:65–76, 1964.
  • [37] M. Katchalski and A. Liu. A problem of geometry in RnR^{n}. Proceedings of the American Mathematical Society, 75(2):284–288, 1979.
  • [38] M. Khamsi, H. Knaust, N. Nguyen, and M. O’Neill. Lambda-hyperconvexity in metric spaces. Nonlinear Analysis: Theory, Methods & Applications, 43(1):21 – 31, 2001.
  • [39] D. Kratsch, H. Le, H. Müller, E. Prisner, and D. Wagner. Additive tree spanners. SIAM Journal on Discrete Mathematics, 17(2):332–340, 2003.
  • [40] U. Lang. Injective hulls of certain discrete metric spaces and groups. Journal of Topology and Analysis, 05(03):297–331, 2013.
  • [41] W. Lenhart, R. Pollack, J. Sack, R. Seidel, M. Sharir, S. Suri, G. T. Toussaint, S. Whitesides, and C. Yap. Computing the link center of a simple polygon. Discret. Comput. Geom., 3:281–293, 1988.
  • [42] M. C. Lin and J. L. Szwarcfiter. Faster recognition of clique-Helly and hereditary clique-Helly graphs. Inf. Process. Lett., 103(1):40–43, 2007.
  • [43] R. Nandakumar and K. Parthasarathy. Eccentricity-preserving spanning trees. J. Math. Phys Sci, 24:33–36, 1990.
  • [44] R. Nowakowski and I. Rival. The smallest graph variety containing all paths. Discrete Mathematics, 43(2):223 – 234, 1983.
  • [45] E. Prisner. Eccentricity-approximating trees in chordal graphs. Discrete Mathematics, 220(1-3):263–269, Jun 2000.
  • [46] A. Quilliot. On the helly property working as a compactness criterion on graphs. Journal of Combinatorial Theory, Series A, 40(1):186 – 193, 1985.
  • [47] N. Robertson and P. Seymour. Graph minors. iii. planar tree-width. Journal of Combinatorial Theory, Series B, 36(1):49 – 64, 1984.
  • [48] N. Robertson and P. Seymour. Graph minors. ii. algorithmic aspects of tree-width. Journal of Algorithms, 7(3):309 – 322, 1986.
  • [49] Y. Wu and C. Zhang. Hyperbolicity and chordality of a graph. The Electronic Journal of Combinatorics, 18(1):P43, 2011.
  • [50] H.-G. Yeh and G. J. Chang. Centers and medians of distance-hereditary graphs. Discrete Math., 265(1–3):297–310, Apr. 2003.
  • [51] S. Yushmanov and V. Chepoi. A general method of investigation of metric graph properties related to the eccentricity (in Russian). Mathematical Problems of Cybernetics (Moscow), 3:217–232, 1991.