Abstract 1 Introduction 2 Separators for intersection graphs of balls 3 Separators for intersection graphs of spheres 4 Further results 5 Algorithmic implementations References

Separators for Intersection Graphs of Spheres

Jacob Fox ORCID Department of Mathematics, Stanford University, Stanford, CA, USA    Jonathan Tidor ORCID Department of Mathematics, Princeton University, NJ, USA
Abstract

We prove the existence of optimal separators for intersection graphs of balls and spheres in any dimension d. One of our results is that if an intersection graph of n spheres in d has m edges, then it contains a balanced separator of size Od(m1/dn12/d). This bound is best possible in terms of the parameters involved. The same result holds if the balls and spheres are replaced by fat convex bodies and their boundaries.

Keywords and phrases:
graph separators, intersection graphs
Funding:
Jacob Fox: supported by NSF Awards DMS-2452737 and DMS-2154129.
Jonathan Tidor: supported in part by a Stanford Science Fellowship.
Copyright and License:
[Uncaptioned image] © Jacob Fox and Jonathan Tidor; licensed under Creative Commons License CC-BY 4.0
2012 ACM Subject Classification:
Theory of computation Computational geometry
; Mathematics of computing Graph theory
Related Version:
Full Version: https://arxiv.org/abs/2603.22204
Funding:
This material is based in part upon work supported by the National Science Foundation under Grant No. DMS-1928930, while the authors were in residence at the Simons Laufer Mathematical Sciences Institute in Berkeley, California, during the Spring 2025 semester.
Editors:
Hee-Kap Ahn, Michael Hoffmann, and Amir Nayyeri

1 Introduction

For a graph G=(V,E), a balanced separator is a vertex set 𝒳V such that each connected component of G𝒳 contains at most 23|V| vertices. In 1979, Lipton and Tarjan [15] showed that n-vertex planar graphs have balanced separators of size O(n). There are many different characterizations of planar graphs that each lead to a different extension of the Lipton–Tarjan separator theorem. For example, Wagner’s theorem [23] states that a graph is planar if and only if it does not contain K5 or K3,3 as a minor. The family of graphs that can be embedded in a fixed surface has a finite list of forbidden minors by the Robertson–Seymour theorem [19]. Gilbert, Hutchinson, and Tarjan [9] proved that every n-vertex graph which can be embedded in a surface of genus g has a separator of size O(gn). Later, Alon, Seymour, and Thomas [1] proved that for every graph H, every n-vertex graph that does not contain H as a minor has a separator of size at most OH(n).

Another important characterization of planar graphs is the Koebe–Andreev–Thurston circle packing theorem, which says that a graph is planar if and only if it is the intersection graph of interior-disjoint disks in the plane. (This means that there is a collection of interior-disjoint disks, one corresponding to each vertex, so that two vertices are connected by an edge if and only if the corresponding disks are tangent.) This leads to another way to try to extend the Lipton–Tarjan separator theorem, to intersection graphs of geometric objects. A string graph is an intersection graph of arcwise-connected sets in the plane, so planar graphs are string graphs. The Lipton–Tarjan separator theorem has been extended in several works studying separators for string graphs [6, 7, 12, 13, 17], and we now know that every string graph with m edges has a balanced separator of size O(m).

Graphs with a small balanced separator are poor expanders, and one can utilize the separator for many applications, including in fast divide-and-conquer type algorithms for many algorithmically hard graph problems (see, for example, [16]), extremal applications (see, for example, [8, 14]), and enumerative problems (see, for example, [5]).

In this paper, we are interested in finding small separators for intersection graphs of geometric objects in higher dimensions. Miller, Teng, Thurston, and Vavasis [18] proved a separator theorem for the intersection graph of balls in d in which no point is in more than k of the balls. Generalizing this, we prove that every intersection graph of n balls in d with m edges has a balanced separator of size Od(m1/dn12/d). This bound is best possible in terms of the parameters involved. (Here and throughout the paper, we use the notation Od() to hide a multiplicative constant that depends only on the dimension d.)

Proving a separator theorem for intersection graphs of spheres turns out to be more challenging as such graphs can have significantly more complicated combinatorial structure. Recently, Davies, Georgakopoulos, Hatzel, and McCarty [3] studied this problem under the constraint that the intersection graph is Kt,t-free, proving that if an intersection graph of n spheres in d is Kt,t-free, then it has a balanced separator of size O~d(t10n11/(2d+8)). We first study the problem in general. Despite their more complicated structure, we show that intersection graphs of spheres have the same size separators as intersection graphs of balls.

Theorem 1.1.

An intersection graph of n spheres in d with m edges has a balanced separator of size Od(m1/dn12/d).

As a corollary, we show that every Kt,t-free intersection graph of n spheres has a separator of size Od(t1/dn11/d) (see Corollary 3.9). This verifies a strengthening of a conjecture of Davies, Georgakopoulos, Hatzel and McCarty [3, Conjecture 20]. Both Theorem 1.1 and the corollary are best possible up to the constant factor depending on the dimension d (see Proposition 4.1). We also prove a refinement of Theorem 1.1 which gives a separator whose size depends on the degree sequence of the intersection graph. As a corollary, we prove that every Kt,t-free intersection graph of spheres in d has average degree Od(t). It follows that Kt,t-free intersection graphs of spheres in d have chromatic number Od(t).

We further study intersection graphs of other geometric objects in Section 4. We generalize our results from balls and spheres to fat convex bodies and their boundaries. Unlike the planar case, we show that both the fatness and convexity assumptions are necessary in dimensions three and larger: no non-trivial separator exists for the intersection graph of general convex bodies or general fat bodies.

Finally, as separators have many algorithmic applications, in Section 5 we show that there are efficient randomized algorithms for finding such separators of intersection graphs of balls or spheres.

2 Separators for intersection graphs of balls

Let be a collection of balls in d. The intersection graph of , denoted G(), is the graph with vertex set where two balls are adjacent if they intersect. The collection is k-ply if every point of d is contained in at most k balls. Miller, Teng, Thurston, and Vavasis [18] proved the following separator theorem for intersection graphs of balls.

Theorem 2.1 ([18]).

Let be a k-ply collection of n balls in d. Then G() has a balanced separator of size Od(k1/dn11/d).

A more recent proof of this result was given by Har-Peled [10]. Using his techniques, we are able to give a more refined version of the Miller–Teng–Thurston–Vavasis separator theorem. (Similar techniques also appeared in an earlier work of Smith and Wormald [20].)

Given a collection of balls , for each point x, the ply of x, denoted p(x), is the number of balls in which contain x. For each ball B, the ply of B, denoted p(B), is the maximum over p(x) over all points xB. With this definition we strengthen Theorem 2.1 as follows.

Theorem 2.2.

Let be a collection of n balls in d. Then G() has a balanced separator of size

Od((Bp(B)1d1)11d).

Since every ball has ply at most one more than its degree in the intersection graph, we have the following corollary which bounds the size of the separator in terms of the degree sequence of the graph, or simply its number of edges.

Corollary 2.3.

Let be a collection of n balls in d. Then G() has a balanced separator of size

Od((Bdeg(B)1d1)11d).

In particular, if G() has m edges, then it has a balanced separator of size Od(m1/dn12/d).

The second claim follows from the first via an application of Hölder’s inequality since

Bdeg(B)1d1(Bdeg(B))1d1(B1)d2d1=(2m)1d1nd2d1.

To construct a balanced separator, we first construct a separator with slightly worse quantitative properties and then iterate a bounded number of times to find the desired separator.

Definition 2.4.

For a graph G=(V,E) and c<1, a c-balanced separator is a vertex set 𝒳V such that each connected component of G𝒳 contains at most c|V| vertices.

Lemma 2.5.

Suppose graph G=(V,E) is such that every induced subgraph has a c-balanced separator of size at most s. Then G has a balanced separator of size at most logc(2/3)s.

Proof.

Set V1=V. Given a vertex set Vi, let 𝒳i be a c-balanced separator of G[Vi] of size at most s. Define Vi+1 to be the set of vertices comprising the largest connected component of G[Vi]𝒳i. By definition, |Vi+1|c|Vi|. Furthermore, each other connected components of G[Vi]𝒳i contains at most 12|Vi|23|V| vertices. Halt the process at the first i=r satisfying |Vr+1|23|V|, so rlogc(2/3). Defining 𝒳=𝒳1𝒳r, we have |𝒳|rslogc(2/3)s. Furthermore, every connected component of G𝒳 is either equal to Vr+1, or is a connected component of G[Vi]𝒳i other than Vi+1 for some i. Since all of these sets have size at most 23|V|, this implies that 𝒳 is a balanced separator for G.

Proof of Theorem 2.2.

We will give a randomized procedure that produces a (15d)-balanced separator 𝒳 for G() of the desired size. By Lemma 2.5, losing another constant factor depending only on d, this suffices to prove the desired result.

Let B0 be the ball in d with smallest radius that contains at least 5dn centers of balls of . (B0 is not necessarily an element of .) By perturbing B0 slightly, one can see that B0 contains exactly 5dn centers. After translating and rescaling the problem appropriately, assume that B0=B(1), the origin-centered ball of radius 1.

For some r[1,2], let 𝒳 be the set of balls B which intersect the origin-centered sphere of radius r. We claim that for any r[1,2], the set 𝒳 is a (15d)-balanced separator for G(). To see this, note that each connected component of 𝒳 consists of balls which are either entirely inside B(r) or entirely outside B(r). Since B(1)B(r) contains 5dn centers, there are at most (15d)n balls entirely outside B(r). Now B(2)B(r) can be covered by 4d translates of B(1) which, by the minimality of B0, each contain at most 5dn centers. Thus B(2) contains at most 4d5dn(15d)n centers, so there are at most this many balls entirely inside B(r).

We now show that there is a choice of r[1,2] so that 𝒳 is small. To do this, pick r[1,2] uniformly at random. We now bound the expected size of 𝒳. We do this by modifying by shrinking those balls that intersect B(2) so that they lie entirely in B(2). For each B which intersects B(2), let B be the line that passes through the origin and the center of B. Let B be the ball contained in B(2) whose center lies on B and which satisfies BB=BBB(2). (See Figure 1.) Define to be the set of these balls B (the balls in that do not intersect B(2) do not have any corresponding ball in ).

Figure 1: Constructing from .

Since each ball B lies inside the corresponding ball B, clearly we have p(B)p(B). Write p(x) for the number of balls of which contain the point x. Then p(B)=maxxBp(x). We now have

vol(B(2)) vol(BB)=BBp(x)dxp(x)=BBdxp(x)
BBdxp(B)=Bvol(B)p(B).

Thus, dividing by vol(B(1)), we conclude that

2dBr(B)dp(B).

Now for each B, the probability that B is in the separator is at most the length of the interval BBB(2). In particular, the probability is 0 if B does not intersect B(2). Now if B does intersect B(2), it corresponds to a ball B so that BBB(2)=BB is an interval of length 2r(B). Thus by Hölder’s inequality we conclude

𝔼[|𝒳|] 2Br(B)2(Br(B)dp(B))1/d(Bp(B)1/(d1))(d1)/d
4(Bp(B)1/(d1))(d1)/d4(Bp(B)1/(d1))(d1)/d.

 Remark 2.6.

Clearly Theorem 2.2 implies Theorem 2.1. We point out that even the weakest version of Corollary 2.3 is sufficient to imply Theorem 2.1. To see this, let be a k-ply collection of balls in d. Then the intersection graph G() is Od(k)-degenerate. (Recall that a graph is D-degenerate if there exist an ordering of the vertices so that each vertex is adjacent to at most D vertices which appear earlier in the ordering.) To see this, order the balls by radius. One can check that any ball Bj which intersects Bi and has radius satisfying rjri must occupy at least a (1/3)d-fraction of Bi, defined to be the dilate of Bi by a factor of 3. The k-ply assumption shows that there are at most 3dk such balls Bj. Thus G() has at most 3dkn edges so Theorem 2.1 follows from Corollary 2.3.

3 Separators for intersection graphs of spheres

The main difficulty in generalizing the results of the previous section to intersection graphs of spheres is that the notion of ply is not useful for sphere intersection graphs. In particular, one can nest spheres so that the corresponding collection of balls would have large ply yet the sphere intersection graph has few or no edges. Furthermore, there are dense graphs (in particular, complete bipartite graphs) that are intersection graphs of spheres but are 2-ply in the sense that no point lies on more than two spheres. (To see this, take two nested families of spheres that are arranged so that every sphere from one family intersects every sphere from the other family.) This shows that no analogue of the argument in Remark 2.6 applies to sphere intersection graphs. Despite these difficulties, we show that separators of the same size exist.

Theorem 3.1.

Let 𝒮 be a collection of n spheres in d. Then G(𝒮) has a balanced separator of size at most

Od((S𝒮deg(S)1d1)11d).

As with Corollary 2.3, this result implies Theorem 1.1 by an application of Hölder’s inequality. We start with an elementary lemma that is needed for the proof.

Lemma 3.2.

Let 𝒮 be a collection of spheres in d. Suppose that G(𝒮) has maximum degree at most Δ and that there exists a point pd contained in the interior of at least 2s spheres. Then there exists a set 𝒳𝒮 with |𝒳|Δ such that every connected component of G(𝒮)𝒳 has size at most |𝒮|s.

Proof.

Let 𝒮𝒮 be the set of spheres containing p in their interior. For any two spheres S,S𝒮, either they intersect or one contains the other in its interior. Consider the partial order on 𝒮 defined by containment. Take any linear extension of this poset and let S0𝒮 be the median element of this linear extension. This means that there are at least s1 spheres which either intersect S0 or are contained in its interior and there are at least s1 spheres which either intersect S0 or contain it in their interior.

Now let 𝒳=N(S0), the neighborhood of S0 in G(𝒮), i.e., the set of spheres which intersect S0. We claim that every connected component of G(𝒮)𝒳 has size at most |𝒮|s. To see this, note that no sphere in 𝒮𝒳 intersects S0, so any connected component of G(𝒮)𝒳 consists of spheres either entirely inside S0 or entirely outside S0. However, we showed that there are at least s spheres that are not of the first type (S0 and the s1 spheres which follow it in the linear extension) and at least s spheres that are not of the second type. Thus each connected component of G(𝒮)𝒳 has size at most |𝒮|s.

Proof of Theorem 3.1.

Set

Σ=Cd(S𝒮deg(S)1d1)11d.

Choosing cd>0 small in terms of d and Cd large in terms of cd, we will construct a (1cd)-balanced separator for G(𝒮) of size at most Σ. By Lemma 2.5, losing another constant factor depending only on d, this suffices to prove the desired result. Assume Σ<n, since the result is trivial otherwise.

Through the proof, if we have not already found the desired separator, we will define sets 𝒳i𝒮 for i=1,2,3,4 with |𝒳i|Σ/4 for each i so that 𝒳1𝒳2𝒳3𝒳4 is the desired separator.

Step 1:

First, define 𝒳1𝒮 to be the Σ/4 vertices of highest degree in G(𝒮). Let Δ be the maximum degree of the resulting graph G(𝒮)𝒳1. We claim that ΔΣ/4 and Δcdn. If the first does not hold, then G(𝒮) would have at least Σ/4 vertices of degree at least Σ/4, meaning that

Σ=Cd(S𝒮deg(S)1/(d1))11/dCd((Σ/4)(Σ/4)1/(d1))11/d=Cd4Σ.

Clearly this is impossible for Cd>4. Similarly, if the second fails then G(𝒳) would have at least Σ/4 vertices of degree at least cdn, meaning that

ΣCd((Σ/4)(cdn)1/(d1))11/d=Cdcd1/d411/dΣ11/dn1/d>Cdcd1/d411/dΣ,

where the final inequality holds since Σ<n. Again, this is impossible for Cd sufficiently large in terms of cd.

Step 2:

Write 𝒮1=𝒮𝒳1. The goal of this step is to pass to a subset 𝒮2𝒮1 such that every sphere S𝒮2 contains fewer than 4cdn centers of spheres in 𝒮2. Let S0𝒮1 be a sphere of minimal radius that contains 4cdn centers of spheres in 𝒮1. (If no such sphere exists, taking 𝒮2=𝒮1 completes this step.) Define 𝒮S0𝒮1 to be the set of spheres which contain S0 entirely within their interior. First suppose that |𝒮S0|2cdn. In this case we can apply Lemma 3.2 to find a separator 𝒳2𝒮1 such that |𝒳2|ΔΣ/4 and each connected component of G(𝒮1)𝒳2 has size at most (1cd)n. Then 𝒳1𝒳2 is a (1cd)-balanced separator for G(𝒮) of size at most Σ/2.

Thus we can assume that |𝒮S0|<2cdn. In this case, define 𝒳2=N(S0), the neighborhood of S0 in G(𝒮). Define 𝒮2𝒮1 to be the set of spheres which are entirely contained within S0. Clearly |𝒳2|ΔΣ/4.

Claim 3.3.

If G(𝒮)(𝒳1𝒳2) has a connected component of size more than (1cd)n, then it is also a connected component of G(𝒮2).

Proof.

Every connected component of G(𝒮)(𝒳1𝒳2) is either made up of spheres entirely inside S0 or entirely outside S0. The former are also connected components of G(𝒮2) while the latter are not. There are at least 4cdn centers of spheres of 𝒮1 inside of S0. Since at most 2cdn of these correspond to spheres surrounding S0 and at most Δ correspond to spheres which intersect S0, we see that there are at least 2cdnΔcdn spheres inside S0. Therefore any connected component outside of S0 has size at most (1cd)n.

We now construct a separator for G(𝒮2). We know that every sphere in 𝒮2 contains fewer than 4cdn centers of spheres in 𝒮2. We can also assume that |𝒮2|(1cd)n, otherwise 𝒳1𝒳2 is already the desired (1cd)-balanced separator.

Step 3:

Let Bin be a ball of minimal radius which contains at least 4cdn centers of spheres in 𝒮2. Since we showed that no sphere in 𝒮2 contains 4cdn centers, we see that no sphere in 𝒮2 contains Bin entirely in its interior. This means that each of the 4cdn spheres whose centers lie in Bin also intersect Bin.

Write R for the radius of Bin and define Bout to be the ball with the same center and radius 2R. By the minimality of Bin, we see that Bout contains at most 4d4cdn centers. In particular, choosing cd sufficiently small, there are at least cdn sphere centers outside of Bout. Let Srand be a sphere with the same center as Bin and Bout whose radius is chosen uniformly at random in [R,2R].

For a sphere S𝒮, write r(S) for the radius of S. Write r~(S) for the radius of the spherical cap formed by intersecting the ball bounded by S with Srand. (The radius of a spherical cap is the minimal r~(S) so that the cap is contained in a ball of radius r~(S).)

Choose some Cd sufficiently large in terms of d and let 𝒳3 consist of the spheres S𝒮2 which intersect Srand and which satisfy r~(S)CdR(deg(S)/Σ)1/(d1).

Claim 3.4.

𝔼[|𝒳3|]Σ/4.

Proof.

We need to upper bound the probability that a sphere S intersects Srand in a spherical cap of radius at most r~. If r~r(S), the probability is upper bounded by

2(r(S)r(S)2r~2)R2r~R.

To see this, note that this event is the same as the event that Srand intersects the line connecting the center of Bin and the center of S in one of the bold segments in Figure 2.

Now if r~>r(S), the probability we need to upper bound is just the probability that Srand and S intersect. This quantity is at most 2r(S)/R<2r~/R.

Since 𝒳3 is defined with the cutoff r~=CdR(deg(S)/Σ)1/(d1), for each S𝒮2 we have the bound

Pr[S intersects Srand and r~(S)CdR(deg(S)/Σ)1/(d1)]2Cd(deg(S)/Σ)1/(d1).

This implies that

𝔼[|𝒳3|] S𝒮22Cd(deg(S)/Σ)1/(d1)=2CdΣ1/(d1)Cdd/(d1)Σd/(d1)Σ/4,

where the equality is by the definition of Σ and the last inequality holds for Cd chosen appropriately large in terms of Cd.

Fix some choice of Srand for which |𝒳3|Σ/4.

Figure 2: Bounding Pr[S𝒳3].

Step 4:

Define 𝒮3=𝒮2𝒳3. We perform the following iterative process. Start with 𝒳4= and 𝒮4=𝒮3. We will define spheres S1,S2,,S. Suppose we have defined S1,,Si1. Define Si to be a sphere in 𝒮4 which intersects Srand and has maximal radius among such spheres. Place the neighborhood N(Si) in 𝒳4 and remove all spheres which intersect the ball bounded by Si from 𝒮4 (in particular we remove Si and N(Si) from 𝒮4). Repeat until no more spheres in 𝒮4 intersect Srand.

We claim that the balls bounded by S1,S2,,S are disjoint. For i<j we know that r(Si)r(Sj) and the ball bounded by Si is disjoint from the sphere Sj. The latter implies that Si,Sj do not intersect and that Si does not contain Sj; the former implies that Sj does not contain Si.

Claim 3.5.

|𝒳4|Σ/4.

Proof.

By the definition of 𝒳3, every sphere S𝒮3 which intersects Srand also satisfies r~(S)>CdR(deg(S)/Σ)1/(d1). Thus

|𝒳4|i=1deg(Si)<i=1Σ(r~(Si)CdR)d1.

Now since the balls bounded by S1,,S are disjoint, we have

i=1τd1r~(Si)d1σd(2R)d1

where τk,σk denote the k-volume of the unit ball in k and the (k1)-volume of the unit sphere in k, respectively. (In the above inequality, we use the fact that the (d1)-volume of a spherical cap of radius r~ on any sphere in d is at least the (d1)-volume of a ball of radius r~ in d1.) Thus we conclude that

|𝒳4|Σ2d1σdCdd1τd1Σ/4,

where the last inequality holds for Cd chosen sufficiently large in terms of d.

Claim 3.6.

𝒳=𝒳1𝒳2𝒳3𝒳4 is a (1cd)-balanced separator for G(𝒮).

Proof.

By Claim 3.3, it suffices to show that each connected component of G(𝒮2)(𝒳3𝒳4) has size at most (1cd)n. Note that any sphere in 𝒮2(𝒳3𝒳4) which crosses Srand must lie in the ball bounded by one of the Si. Furthermore, no sphere in 𝒮2(𝒳3𝒳4) crosses any of the Si. Therefore any connected component of G(𝒮2)(𝒳3𝒳4) must either lie entirely within some sphere Si, entirely within Srand, or entirely outside of Srand. We chose 𝒮2 so that any sphere in 𝒮2 contains at most 4cdn centers of spheres in 𝒮2. Thus the first type of connected component has size at most 4cdn. For the second type, note that Srand is contained within Bout which we argued earlier contains at most 4d4cdn<(1cd)n centers of spheres in 𝒮2. For the third type of connected component, we know that there are at least 4cdn spheres which intersect Bin. Since Bin is contained in the interior of Srand, we see that there are at most (14cd)n spheres outside of Srand, giving the desired bound on the third type of connected component. Since |𝒳|Σ, this completes the proof.

Intersection graphs of spheres are an example of semialgebraic graphs. Using known bounds on the number of edges in Kt,t-free semialgebraic graphs [21], we deduce the following weak bound for separators in Kt,t-free sphere intersection graphs.

Corollary 3.7.

Let 𝒮 be a collection of n spheres in d. If G(𝒮) is Kt,t-free, then it has a balanced separator of size Od(t2/d(d+2)n12/d(d+2)).

Proof.

Represent the sphere centered at xd with radius r by the point (x,r)d+1. Now two spheres (x,r) and (x,r) intersect if and only if (rr)2xx2(r+r)2. These are polynomial inequalities of degree O(1), meaning that G(𝒮) can be represented by a semialgebraic graph in d+1 of total degree O(1). (See [21, Definition 1.1] for the full definition.)

By [21, Theorem 1.6] (a bound on the Zarankiewicz problem for semialgebraic graphs), if G(𝒮) is Kt,t-free, then it has Od(t2/(d+2)n2(d+1)/(d+2)) edges. Applying Theorem 1.1 with this bound shows that G(𝒮) has a balanced separator of size

Od((t2/(d+2)n2(d+1)/(d+2))1/dn12/d)=Od(t2/d(d+2)n12/d(d+2)).

We can bootstrap this weak bound to get a better bound on the number of edges in a Kt,t-free sphere intersection graph. This follows from [6, Lemma 8] which turns a bound on the size of separators into a bound on the number of edges for any hereditary family of graphs. (See [8, Section 3] for another application of this method.)

Corollary 3.8.

Let 𝒮 be a collection of n spheres in d. If G(𝒮) is Kt,t-free, then it has Od(tn) edges.

Proof.

We aim to apply [6, Lemma 8]. This result requires a monotone function ϕ so that any n-vertex Kt,t-free intersection graph of spheres in d contains a balanced separator of size at most nϕ(n). By Corollary 3.7, we can take ϕ(n)=Cd(t/n)2/d(d+2) where Cd is the constant in that result. Clearly ϕ(n) is monotone.

Set n0=(12Cd)d(d+2)/2t so that ϕ(n0)=1/12. Then [6, Lemma 8] implies that every Kt,t-free intersection graph of spheres in d on nn0 vertices has at most Cn0n/2 edges where

C=i=0(1+ϕ((4/3)in0)).

We can bound

Cexp(i=0ϕ((4/3)in0))exp(2Cdi=0((34)itn0)2/d(d+2)).

As n0=Od(t), this geometric series sums to Od(1), showing that C=Od(1). This gives the desired bound on the number of edges when nn0.

Note that if n<n0, then we can trivially bound the number of edges by n2<n0n=Od(tn).

Combining Theorem 1.1 and Corollary 3.8, we immediately conclude the following strengthening of Corollary 3.7. This result improves upon the sphere separator result of Davies–Georgakopoulos–Hatzel–McCarty [3, Theorem 1], proving a strengthening of [3, Conjecture 20].

Corollary 3.9.

Let 𝒮 be a collection of n spheres in d. If G(𝒮) is Kt,t-free, then it has a balanced separator of size Od(t1/dn11/d).

 Remark 3.10.

The same proof also gives Corollaries 3.8 and 3.9 for intersection graphs of balls. However these results actually have significantly easier proofs. For example, let be a collection of balls in d. If is (2t1)-ply, then Theorem 2.1 already implies that G() has a balanced separator of size Od(t1/dn11/d). Otherwise, there is a point in 2t balls of , implying that G() contains K2t which contains Kt,t.

4 Further results

We start by giving a construction that shows that our main results are asymptotically optimal. We then generalize these results to a larger class of geometric objects.

Proposition 4.1.

For each n,d1 and nm(n2), there exists an intersection graph of n balls or spheres in d with at most m edges for which every balanced separator has size at least Ωd(m1/dn12/d).

Proof.

Let V={1,,k}d for k=n1/d. Define G=(V,E) where x,yV are adjacent if xyr. Choosing 1r=Θd((m/n)1/d) appropriately, we can ensure that G has Θd(m) edges. Note that G is both the intersection graph of the radius r/2 balls centered at the points of V and the intersection graph of the radius r/2 spheres centered at the points of V. We will show that every balanced separator of G has size at least Ωd(m1/dn12/d).

We first handle the case when r=1. Suppose that 𝒳 is a balanced separator with |𝒳|kd/3. Let AV be the union of some connected components of G𝒳, satisfying kd/3|A|2kd/3. When r=1, the graph G is known as the grid graph on {1,,k}d. Then the isoperimetric inequality on the grid graph [2] implies that there are at least kd1 edges between A and VA. However, all such edges must have one endpoint in 𝒳. Since the maximum degree of G is 2d, we conclude that |𝒳|kd1/2dΩd(n11/d), as desired.

To handle the general case, let r=r/(2d). For simplicity, assume that r is an integer that divides k, say k=rk. Divide {1,,k}d into (k)d subcubes of side length r. Note that every vertex in one subcube is adjacent in G to every vertex in an adjacent subcube. Similar to before, suppose that 𝒳 is a balanced separator with |𝒳|kd/12 and A is a union of connected components of G𝒳 with kd/3|A|2kd/3. Write B=V(𝒳A). We know |B|kd/4. Define A{1,,k}d where a point lies in A if the corresponding r××r subcube of {1,,k}d contains at least one point of A. Define B{1,,k}d similarly. First note that A and B are disjoint – if they shared a vertex then A,B would contain points in the same subcube which have distance at most dr<r.

Write X={1,,k}d(AB). Note that |A||A|/(r)d(k)d/3 and similarly |B||B|/(r)d(k)d/4. Thus |A|(k)d|B|3(k)d/4. Applying the isoperimetric inequality to the grid graph on {1,,k}d, we conclude that there are at least (k)d1 edges between A and BX. However, there are no edges between A,B because any two points in two adjacent subcube are at distance at most 2dr=r. Thus we conclude that |X|(k)d1/(2d). Note though that a point lies in X only if the corresponding subcube lies entirely in 𝒳. We thus see that |𝒳|(r)d(k)d1/(2d)=Θd(m1/dn12/d).

Our main results on ball and sphere separators, Theorems 2.2 and 3.1, hold for more general classes of geometric objects. We say that a set Bd is C-fat if B contains a ball of radius r and is contained inside a ball of radius Cr for some r. Our results hold for fat convex bodies and for the boundaries of fat convex bodies. (See [11, 4] for related results under slightly different hypotheses.)

Theorem 4.2.

Let be a collection of n C-fat convex bodies in d. Then G() has a balanced separator of size at most

Od,C((Bp(B)1d1)11d).

In particular, if G() has m edges, then it has a balanced separator of size Od(m1/dn12/d).

Proof.

For each B, pick an arbitrary point pint(B) that we call the center of B. With this definition, we proceed with the proof of Theorem 2.2 in the same way. The same randomized procedure produces a (15d)-balanced separator.

For each B, let xB0 be the set of radii r so that the origin-centered sphere of radius r intersects B. We wish to show that if xB[1,2] is non-empty, B contains a genuine ball B which is contained in B(2) such that r(B)ΩC(|xB[1,2]|).

First, if xB[0,2], then B is contained in B(2). Now B contains a point of norm max(xB) and a point of norm min(xB), i.e., it contains two points at distance at least max(xB)min(xB)|xB|. This means that any ball containing B must have radius at least |xB|/2. By hypothesis, B contains a ball of radius |xB|/2C. Letting B be this ball, we are done since in this case BB is contained in B(2).

Now suppose xB[0,2], but xB[0,2]. By the same argument, B contains some ball (say B~) of radius (max(xB)min(xB))/2C centered at some point p with pmax(xB). By assumption, B contains a point q with q=min(xB)2. (See Figure 3.) Then the convexity of B implies that B contains the convex hull of {q}B~. Elementary geometry implies that this convex hull contains a ball B of radius 2min(xB)2C+1|xB[1,2]|2C+1 that is contained in B(2). (To see this, dilate the convex hull of {q}B~ by a factor of (2q)/(p+r(B~)q) so that it lies within B(2).)

Thus we have constructed a set contained in B(2) so that 𝔼[|𝒳|]ΩC(Br(B)). Note that consists of genuine balls, so the rest of the proof applies unchanged.

Figure 3: The construction of B inside of BB(2).
Theorem 4.3.

Let 𝒮 be a collection of n boundaries of C-fat convex bodies in d. Then G(𝒮) has a balanced separator of size at most

Od,C((S𝒮deg(S)1d1)11d).

In particular, if G(𝒮) has m edges, then it has a balanced separator of size Od(m1/dn12/d).

Proof.

For each S𝒮, pick an arbitrary point in the interior of S to be its center. Consider the partial order on 𝒮 where SS if S is contained in the interior of S. Pick an arbitrary linear extension of this partial order. When we talk about the S𝒮 of maximal radius (or minimal radius) we mean maximal with respect to this linear extension. With these two definitions, we can proceed with the proof of Theorem 3.1 in the same way. The same randomized procedure produces a (1cd)-balanced separator.

The main difference in the analysis is understanding 𝔼[|𝒳3|]. For each S, define r~(S) so that the (d1)-volume of the intersection of the convex body bounded by S and the sphere Srand is τd1r~(S)d1. We claim that Pr[r~(S)r~]OC(r~/R).

Fix r~. Change coordinates so that Srand is centered at the origin. Write B for the convex body with boundary S. Let rmin=minxBx and rmax=maxxBx. Pick p,qB with p=rmin and q=rmax. Since pqrmaxrmin, we see that the smallest ball that contains B has radius at least (rmaxrmin)/2. By the C-fat hypothesis, B contains some ball B centered at a point o with radius (rmaxrmin)/2C. Now B contains the convex hull of {p,q}B which contains two cones with apexes p,q and with bases some disk of radius (rmaxrmin)/2C centered at o. This implies that for each ϵ>0, the body B contains a ball of radius ϵ(rmaxrmin)/2C centered at a point of distance rmin+ϵ(rmaxrmin) from the origin and another ball of the same radius centered at a point of distance rmaxϵ(rmaxrmin) from the origin. (See Figure 4.) In particular, if ϵ(rmaxrmin)/2Cr~, then the intersection of B and the spheres of radius rmin+ϵ(rmaxrmin) and rmaxϵ(rmaxrmin) both contain a spherical cap of radius at least ϵ(rmaxrmin)/2Cr~, so these intersections have (d1)-volume at least τd1r~d1. The first inequality holds for ϵ2Cr~/(rmaxrmin), so

Pr[r~(S)r~]2(rmaxrmin)2Cr~/(rmaxrmin)R=4Cr~R,

as desired.

This gives us the bound 𝔼[|𝒳3|]Σ/4 as in Claim 3.4. Then the bound |𝒳4|Σ/4 also holds since we defined r~(S) so that the proof of Claim 3.5 still goes through.

Figure 4: Bounding Pr[r~(S)r~].

With these results, Corollary 3.8 also holds for fat convex bodies if the bodies are also semialgebraic sets of bounded description complexity.

We also note that both of the assumptions, convexity and fatness, are necessary. Indeed, for d3, every graph is the intersection graph of convex bodies in d and also every graph is the intersection graph of fat connected bodies in d. The former fact is classical (see [22]); we prove the latter fact here.

Proposition 4.4.

For d3, ϵ>0, and any graph G, there exists a collection of (1+ϵ)-fat connected bodies in d whose intersection graph is G.

Proof.

Let Bk be the ball centered at the origin in d of radius rk=(2+2ϵ1)k. We will define Xk to be a (1+ϵ)-fat connected body that is contained in Bk. The set Xk will be a “balloon” consisting of a large sphere in int(Bk)int(Bk1) that is tangent to Bk1 at a point pk and a “string” in Bk1: a curve starting at pkBk1, passing through B0 at a point qk, and then continuing into B0. We will define these sets Xk so that they are disjoint in dint(B0). See Figure 5 for an illustration of this construction.

Suppose we have defined B0,,Bk1 and X1,,Xk1. Let pk be a point on Bk1. Say Bk1 has radius rk1. Let Yk be the ball of radius ϵ1rk1, centered at (1+ϵ1)pk. Then Bk1Yk is contained in the ball of radius (1+ϵ)ϵ1rk1 centered at ϵ1pk. Since rk=2(1+ϵ1)rk1, we see that Yk is contained in int(Bk). We will define Xk to be Ykγk where γk is a curve starting from pk which is contained in Bk1. Whatever our choice of γk, we have ensured that Xk is (1+ϵ)-fat and connected.

Pick qk to be a point in B0{q1,,qk1}. Let γk be a curve from pk to qk whose interior is contained in

int(Bk1)(B0i=1k1Xi).

If k=1, we set q1=p1 and let γ1 have length 0. For k>1, clearly there is enough space in dimension d3 to find such a γk.

Finally, define γk=γkδk where δk is a curve starting from qk whose interior is contained in int(B0) such that δk and δ intersect if and only if k, are adjacent in G. Again, in dimension d3, this is clearly possible.

Figure 5: The construction in Proposition 4.4.

5 Algorithmic implementations

Suppose we are given as a collection of centers and radii of balls. We give a randomized linear time algorithm that computes a small separator for G().

Theorem 5.1.

The separator for intersection graphs of balls given in Theorem 2.2 can be computed by a randomized linear time algorithm that succeeds with probability at least 1/2.

Proof.

Write 𝒫 for the centers of the balls of . Let r be the smallest radius such that there exists a ball B0 of radius r that contains at least 9dn points of 𝒫. Fix such a ball B0. Then for any p𝒫B0, the ball of radius 2r centered at p contains B0 and thus contains at least 9dn points of 𝒫.

Our first goal is to efficiently find an approximation for this ball B0. Fix a large constant Cd and sample Cd9d points p1,p2,𝒫 uniformly at random. For each point pi, compute the smallest ri such that the ball of radius ri centered at pi contains 9dn points of 𝒫. This can be computed in linear time for each pi. Say we succeed if miniri2r. By the discussion above, we succeed if some pi𝒫B0, which occurs with probability at least

1(19d)Cd9d1eCd.

Pick 1iCd9d such that ri is minimal. As in the proof of Theorem 2.2, pick rrand[ri,2ri] uniformly at random and let 𝒳 be the set of balls B which intersect the sphere of radius rrand centered at pi. This set can be computed in linear time. Now we still have

𝔼[|𝒳|]4(Bp(B)1d1)11d,

so by Markov’s inequality,

Pr[|𝒳|4Cd(Bp(B)1d1)11d]1Cd1.

Now the set 𝒳 is a (19d)-balanced separator for G(). To see this, note that the ball of radius ri centered at pi contains B0 which contains 9dn points of 𝒫. Furthermore, the ball of radius 2ri centered at pi can be covered by at most 8d translates of B0, so it contains at most 8d9dn<(19d)n points of 𝒫.

Thus we have a randomized linear time algorithm which computes a small (19d)-balanced separator for G() with probability at least 1Cd1. By Lemma 2.5, for Cd sufficiently large, we can iterate this algorithm log19d(2/3) times to produce a 23-balanced separator for G() still in linear time with success probability at least 12.

Suppose we are given 𝒮 as a list of centers and radii of the spheres. In quadratic time we can compute which pairs of spheres intersect and which pairs are contained in each other. From this we can run the algorithm described in Theorem 3.1 in quadratic time. The only potential difficulty is in selecting the minimal radius ball Bin, but this can be handled the same way is in Theorem 5.1.

Theorem 5.2.

The separator for intersection graphs of spheres given in Theorem 3.1 can be computed by a randomized quadratic time algorithm that succeeds with probability at least 1/2.

References

  • [1] Noga Alon, Paul Seymour, and Robin Thomas. A separator theorem for nonplanar graphs. J. Amer. Math. Soc., 3(4):801–808, 1990. doi:10.2307/1990903.
  • [2] Béla Bollobás and Imre Leader. Edge-isoperimetric inequalities in the grid. Combinatorica, 11(4):299–314, 1991. doi:10.1007/BF01275667.
  • [3] James Davies, Agelos Georgakopoulos, Meike Hatzel, and Rose McCarty. Strongly sublinear separators and bounded asymptotic dimension for sphere intersection graphs. In 41st International Symposium on Computational Geometry, volume 332 of LIPIcs. Leibniz Int. Proc. Inform., pages Art. No. 36, 16. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2025. doi:10.4230/lipics.socg.2025.36.
  • [4] Mark de Berg, Hans L. Bodlaender, Sándor Kisfaludi-Bak, Dániel Marx, and Tom C. van der Zanden. A framework for exponential-time-hypothesis-tight algorithms and lower bounds in geometric intersection graphs. SIAM J. Comput., 49(6):1291–1331, 2020. doi:10.1137/20M1320870.
  • [5] Zdeněk Dvořák and Serguei Norine. Small graph classes and bounded expansion. J. Combin. Theory Ser. B, 100(2):171–175, 2010. doi:10.1016/j.jctb.2009.06.001.
  • [6] Jacob Fox and János Pach. Separator theorems and Turán-type results for planar intersection graphs. Adv. Math., 219(3):1070–1080, 2008. doi:10.1016/j.aim.2008.06.002.
  • [7] Jacob Fox and János Pach. A separator theorem for string graphs and its applications. Combin. Probab. Comput., 19(3):371–390, 2010. doi:10.1017/S0963548309990459.
  • [8] Jacob Fox and János Pach. Applications of a new separator theorem for string graphs. Combin. Probab. Comput., 23(1):66–74, 2014. doi:10.1017/S0963548313000412.
  • [9] John R. Gilbert, Joan P. Hutchinson, and Robert Endre Tarjan. A separator theorem for graphs of bounded genus. J. Algorithms, 5(3):391–407, 1984. doi:10.1016/0196-6774(84)90019-1.
  • [10] Sariel Har-Peled. A simple proof of the existence of a planar separator, 2011. arXiv:1105.0103. arXiv:1105.0103.
  • [11] Sariel Har-Peled and Kent Quanrud. Approximation algorithms for polynomial-expansion and low-density graphs. SIAM J. Comput., 46(6):1712–1744, 2017. doi:10.1137/16M1079336.
  • [12] Tuukka Korhonen and Daniel Lokshtanov. Induced-minor-free graphs: separator theorem, subexponential algorithms, and improved hardness of recognition. In Proceedings of the 2024 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 5249–5275. SIAM, Philadelphia, PA, 2024. doi:10.1137/1.9781611977912.188.
  • [13] James R. Lee. Separators in region intersection graphs. In 8th Innovations in Theoretical Computer Science Conference, volume 67 of LIPIcs. Leibniz Int. Proc. Inform., pages Art. No. 1, 8. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2017. doi:10.4230/LIPIcs.ITCS.2017.1.
  • [14] Richard J. Lipton, Donald J. Rose, and Robert Endre Tarjan. Generalized nested dissection. SIAM J. Numer. Anal., 16(2):346–358, 1979. doi:10.1137/0716027.
  • [15] Richard J. Lipton and Robert Endre Tarjan. A separator theorem for planar graphs. SIAM J. Appl. Math., 36(2):177–189, 1979. doi:10.1137/0136016.
  • [16] Richard J. Lipton and Robert Endre Tarjan. Applications of a planar separator theorem. SIAM J. Comput., 9(3):615–627, 1980. doi:10.1137/0209046.
  • [17] Jiří Matoušek. Near-optimal separators in string graphs. Combin. Probab. Comput., 23(1):135–139, 2014. doi:10.1017/S0963548313000400.
  • [18] Gary L. Miller, Shang-Hua Teng, William Thurston, and Stephen A. Vavasis. Separators for sphere-packings and nearest neighbor graphs. J. ACM, 44(1):1–29, 1997. doi:10.1145/256292.256294.
  • [19] Neil Robertson and P. D. Seymour. Graph minors. XX. Wagner’s conjecture. J. Combin. Theory Ser. B, 92(2):325–357, 2004. doi:10.1016/j.jctb.2004.08.001.
  • [20] W.D. Smith and N.C. Wormald. Geometric separator theorems and applications. In Proceedings 39th Annual Symposium on Foundations of Computer Science (Cat. No.98CB36280), pages 232–243, 1998. doi:10.1109/SFCS.1998.743449.
  • [21] Jonathan Tidor and Hung-Hsun Hans Yu. Multilevel polynomial partitioning and semialgebraic hypergraphs: regularity, Turán, and Zarankiewicz results, 2024. arXiv:2407.20221.
  • [22] Heinrich Tietze. Über das Problem der Nachbargebiete im Raum. Monatsh. Math. Phys., 16(1):211–216, 1905. doi:10.1007/BF01693778.
  • [23] K. Wagner. Über eine Eigenschaft der ebenen Komplexe. Math. Ann., 114(1):570–590, 1937. doi:10.1007/BF01594196.