This paper investigates bipartitions of graphs, confirming a conjecture for some realizations, providing counterexamples to refine bounds, and studying edge distributions with new bounds and answers to open questions.
Contribution
It proves the Bollobás-Scott conjecture for certain graph realizations, presents counterexamples suggesting a revised bound, and establishes optimal bounds on bipartition edge distributions.
Findings
01
Confirmed the Bollobás-Scott conjecture for some graph realizations.
02
Provided infinite counterexamples indicating the need for a revised bound.
03
Established optimal bounds on bipartition edge sums and answered related open questions.
Abstract
Bollob\'{a}s and Scott [5] conjectured that every graph G has a balanced bipartite spanning subgraph H such that for each v∈V(G), dH(v)≥(dG(v)−1)/2. In this paper, we show that every graphic sequence has a realization for which this Bollob\'{a}s-Scott conjecture holds, confirming a conjecture of Hartke and Seacrest [10]. On the other hand, we give an infinite family of counterexamples to this Bollob\'{a}s-Scott conjecture, which indicates that ⌊(dG(v)−1)/2⌋ (rather than (dG(v)−1)/2) is probably the correct lower bound. We also study bipartitions V1,V2 of graphs with a fixed number of edges. We provide a (best possible) upper bound on e(V1)λ+e(V2)λ for any real λ≥1 (the case λ=2 is a question of Scott [13]) and answer a question of Scott [13] on max{e(V1),e(V2)}.
K=d_{\ell}^{\prime}=\left\{\begin{array}[]{ll}k-1&\text{ if }d^{\prime}_{\ell}=d_{\ell}-1,\\
k&\text{ if }d^{\prime}_{\ell}=d_{\ell}.\end{array}\right.
K=d_{\ell}^{\prime}=\left\{\begin{array}[]{ll}k-1&\text{ if }d^{\prime}_{\ell}=d_{\ell}-1,\\
k&\text{ if }d^{\prime}_{\ell}=d_{\ell}.\end{array}\right.
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsLimits and Structures in Graph Theory · graph theory and CDMA systems · Advanced Graph Theory Research
Full text
On problems about judicious bipartitions of graphs
Yuliang Ji
Jie Ma
Juan Yan
Xingxing Yu
School of Mathematical Sciences, University of Science and Technology of China,
Hefei, Anhui 230026, China. Email: [email protected] of Mathematical Sciences, University of Science and Technology of China,
Hefei, Anhui 230026, China. Email: [email protected]. Partially supported by NSFC projects 11501539 and 11622110.College of Mathematics and Systems Science, Xinjiang University,
Urumqi, Xinjiang 830046, China. Email: [email protected]. Partially supported by NSFC project 11501486.School of Mathematics, Georgia Institute of Technology, Atlanta, GA 30332, USA. Email: [email protected].
Partially supported by NSF grants DMS–1265564 and DMS-1600387.
Abstract:
Bollobás and Scott [5] conjectured that every graph G has a balanced bipartite spanning subgraph
H such that for each v∈V(G), dH(v)≥(dG(v)−1)/2. In this
paper, we show that every graphic sequence has a realization for which this Bollobás-Scott conjecture holds,
confirming a conjecture of Hartke and Seacrest [10].
On the other hand, we give an infinite family of counterexamples to this Bollobás-Scott conjecture, which
indicates that ⌊(dG(v)−1)/2⌋ (rather than
(dG(v)−1)/2) is probably the correct lower bound.
We also study bipartitions V1,V2 of graphs with a fixed number of edges.
We provide a (best possible) upper bound on
e(V1)λ+e(V2)λ for any real λ≥1
(the case λ=2 is a question of Scott [13]) and
answer a question of Scott [13] on
max{e(V1),e(V2)}.
For any positive integer k, let [k]:={1,…,k}.
Let G be a graph and V1,…,Vk be a
partition of V(G). When k=2, such a partition is said to be a bipartition of G.
A subgraph H of a graph G is said to be a bisection of G if H is a bipartite spanning subgraph of G and the partition sets of H
differ in size by at most one.
For i,j∈[k], we use e(Vi) to denote
the number of edges of G with both ends in Vi and use e(Vi,Vj) to denote the number of edges between Vi and Vj.
Judicious partitioning problems for graphs ask for partitions of graphs
that bound a number of quantities simultaneously, such as
all e(Vi) and e(Vi,Vj). There has been extensive research on this type of problems over the past two decades.
As an attempt to better understand how edges of a graph are
distributed,
we study several judicious bipartitioning problems. Specifically, we study a conjecture of Bollobás and Scott [5] and its
degree sequence version conjectured by Hartke and Seacrest
[10]. We also study
two questions of Scott [13] on bipartitions V1,V2 of a
graph with m edges, bounding e(V1)2+e(V2)2 and
max{e(V1),e(V2)} in terms of m.
For a graph G and for any v∈V(G), we use dG(v) to denote the
degree of the vertex v in G.
It is well known that if H is a maximum bipartite spanning subgraph of a graph G, then
dH(v)≥dG(v)/2 for each v∈V(G).
This, however, may not be true if one requires H to be a bisection, as observed by Bollobás and Scott [5] by considering the complete bipartite graphs K2ℓ+1,m for m≥2ℓ+3.
In an attempt to obtain a similar result for bisections, Bollobás and Scott [5] conjectured that every graph G has a bisection H such that
[TABLE]
This conjecture for regular graphs was made by Häggkvist [8] in
1978, and variations of this problem were studied by Ban and Linial [2].
Hartke and Seacrest [10] studied a degree sequence version of this Bollobás-Scott conjecture.
A nondecreasing sequence π (of nonnegative integers) is said to be
graphic if it is the degree sequence of some finite simple graph G; and
such G is called a realization of the sequence π.
Hartke and Seacrest [10] proved that for any graphic sequence π with even length,
π has a realization G which admits a bisection H such that for all
v∈V(G), dH(v)≥⌊(dG(v)−1)/2⌋.
They further conjectured that for any graphic sequence π with even length,
π has a realization G for which (1) holds.
We prove this Hartke-Seacrest conjecture for all graphic sequences.
For a graph G and a labeling of its vertices V(G)={v1,…,vn}, we define the parity bisection of G to be the
bisection with partition sets V1 and V2,
where Vi={vj∈V(G):j≡imod2} for each i∈[2], and
E(H)={uv∈E(G):u∈V1\mboxandv∈V2}.
Theorem 1.1**.**
Let π=(d1,…,dn) be any graphic sequence with d1≥⋯≥dn.
Then there exists a realization G of π with V(G)={v1,…,vn}
and dG(vi)=di for i∈[n], such that if H denotes the parity
bisection of G then
dH(vi)≥(dG(vi)−1)/2 for i∈[n].
The bound in Theorem 1.1 is best possible, as shown
by the following example given by
Hartke and Seacrest [10].
Let G be the join of a clique K on k vertices and an independent
set I on n−k vertices, where n is even and k<n/2 is odd.
It is not hard to show that G in fact is the unique realization of
the sequence π=(d1,…,dn) with d1=⋯=dk=n−1 and dk+1=⋯=dn=k.
Let H be an arbitrary bisection of G with parts A,B and, without loss of generality,
assume that ∣A∩V(K)∣≤k/2. Since k<n/2, there must exist a
vertex v∈B∩I. So dH(v)=∣A∩V(K)∣≤⌊k/2⌋. Since dG(v)=k and k is odd, we see that dH(v)≤(dG(v)−1)/2.
The second result in this paper gives indication that perhaps the
lower bound in the original
Bollobás-Scott conjecture was meant to be
⌊(dG(v)−1)/2⌋ (rather than (dG(v)−1)/2).
Proposition 1.2**.**
Let r1,r2,r3 be pairwise distinct odd integers such that for every i∈[3],
ri∈/{1,⌊(r1+r2+r3)/2⌋,⌈(r1+r2+r3)/2⌉}.
Then for any bisection H of the complete 3-partite graph
G:=Kr1,r2,r3, there always exists a vertex v with
dH(v)<(dG(v)−1)/2.
This result will follow from a more general result, Proposition 3.3, on all
complete multipartite graphs. We remark here that, for each complete multipartite graph G, it is
easy (as we will see in Section 3) to find a bisection H of G
such that dH(v)≥⌊(dG(v)−1)/2⌋ for all v∈V(G).
However, for general graphs, even the following weaker version of the
Bollobaás-Scott conjecture seems
quite difficult to prove (or disprove).
Conjecture 1.3**.**
There exists some absolute constant c>0 such that every graph G
has a bisection H with dH(v)≥(dG(v)−c)/2 for all v∈V(G).
We now turn our discussion to problems on general bipartitions.
Answering a question of Erdős, Edwards [6] showed in 1973
that every graph with m edges admits a bipartition V1,V2 such that
e(V1,V2)≥m/2+t(m)/2, where
[TABLE]
This bound is best possible for the complete graphs of odd order.
Bollobás and Scott [3] extended Edwards’ bound by showing that
every graph G with m edges has a bipartition V1,V2 simultaneously satisfying
e(V1,V2)≥m/2+t(m)/2 and
max{e(V1),e(V2)}≤m/4+t(m)/4,
where both bounds are tight for the complete graphs of odd order.
Scott [13] provided an interesting viewpoint by introducing
norm for partitions.
For a real number λ>0 and a bipartition V1,V2 of a graph G, define the ℓλ-norm of
(V1,V2) to be (e(V1)λ+e(V2)λ)1/λ.
Then to maximize e(V1,V2) is equivalent to minimize the ℓ1-norm of (V1,V2),
while minimizing max{e(V1),e(V2)} is the same as minimizing the ℓ∞-norm of (V1,V2).
It is natural to consider other norms. In particular,
Scott asked for the maximum of
[TABLE]
over graphs G with m
edges, see Problem 3.18 in [13].
We provide an answer to this question by proving the following general result.
Theorem 1.4**.**
Let m be any positive integer and λ≥1 be any real number.
Then, for any graph G with m edges,
[TABLE]
Moreover, the equality holds if and only if G is a complete graph of odd order.
We also consider analogous questions for k-partitions in Section 4.
Though Edward’s bound is tight for all integers m=(2n), Erdős [7] conjectured that
the difference between Edwards’ bound and the truth can still be
arbitrarily large for other m.
This was confirmed by Alon [1]: every graph with m=n2/2 edges
admits a bipartition V1,V2 such that e(V1,V2)≥m/2+t(m)/2+Ω(m1/4). Bollobás and Scott [5, 13] made a similar conjecture
for max{e(V1),e(V2)}: for certain m, max{e(V1),e(V2)} can be arbitrary far from m/4+t(m)/4. Ma and Yu
[12] proved that every graph with m=n2/2 edges
admits a bipartition V1,V2 such that max{e(V1),e(V2)}≤m/4+t(m)/4−Ω(m1/4).
Another result in the same spirit was given by Hofmeister and Lefmann [9] that any graph with (2kn) edges
has a k-partition V1,...,Vk with ∑i=1ke(Vi)≤k(2n),
which beats the trivial upper bound k1(2nk).
Motivated by these results, Scott asked the following question: does every graph G with
(2kn) edges have a vertex partition into k sets, each of
which contains at most (2n) edges? ( See Problem 3.9 in [13].)
We show that the answer to this question is negative for k=2.
Theorem 1.5**.**
There exist infinitely many positive integers n and for each such n there is a graph
G with (22n) edges, such that, for every bipartition
V1,V2 of G, max{e(V1),e(V2)}≥(2n)+5n/48.
This paper is organized as follows. We prove Theorem 1.1 in Section 2,
and then investigate complete multipartite graphs for
the Bollobás-Scott conjecture in Section 3.
In Section 4, we discuss the questions of Scott and complete the proofs of Theorems 1.4 and 1.5.
2 Hartke-Seacrest conjecture
In this section, we prove Theorem 1.1.
We need two operations on a sequence of non-negative integers.
Let π=(d1,…,dn) with d1≥⋯≥dn.
By removing di from π and subtracting 1 from the
di remaining elements of π with lowest indices, we obtain a new sequence
π′=(d1′,…,di−1′,di+1′,…,dn′), and we say that π′ is obtained from π by
laying offdi. This operation was introduced by Kleitman and Wang [11],
and they proved the following.
For any i∈[n], the sequence π=(d1,…,dn) with d1≥…≥dn is graphic if and only if
the sequence π′ obtained from π by laying off di is graphic.
It is easy to see that the sequence π′ obtained from π by
laying off di need not be non-increasing.
To avoid this issue, Hartke and Seacrest [10] introduced a
variation of the above operation.
Choose a fixed i∈[n].
Let s be the smallest value among the di
largest elements in π, not including the ith element of π
(namely, di itself). Let S={j∈[n]−{i}:dj>s}. Note that ∣S∣<di. Let T be the
set of di−∣S∣ largest indices j with j=i and dj=s.
Then by laying off di with order, we remove di from π
and subtract 1 from dj for all j∈S∪T. If
π′=(d1′,…di−1′,di+1′,…,dn′) denotes the new
sequence, then it has the monotone property d1′≥⋯≥dn′.
Clearly, the sequence obtained from π by laying off di with order is just a permutation of the
sequence obtained from π by laying off di. So the following is true.
For any i∈[n], the sequence π=(d1,…,dn) with d1≥…≥dn is graphic if and only if
the sequence obtained from π by laying off di with order is graphic.
We give a brief outline of our proof of Theorem 1.1. We
choose two consecutive elements dℓ and dℓ+1 of π.
Using Lemma 2.2 we obtain a new graphic sequence π′′ of
length n−2 by first laying off dℓ+1 with order and then laying
off dℓ with order. By induction, π′′ has an (n−2)-vertex
realization F whose parity bisection J has the desired
property. We then show that one can form G from F by adding two new vertices (for dℓ and dℓ+1)
and choosing their neighbors, so that the parity bisection of G
satisfies Theorem 1.1.
We apply induction on the length n of the graphic
sequence π=(d1,…,dn) with d1≥…≥dn.
The assertion is trivial when n=1,2.
So we may assume that n≥3 and the assertion holds for all graphic sequences with length less than n.
Then there exist two consecutive elements of π that are identical;
so let ℓ∈[n−1] be fixed such that
[TABLE]
Let π′=(d1′,...,dℓ′,dℓ+2′,...,dn′) be the sequence obtained
from π by laying off dℓ+1 with order. Let
π′′=(d1′′,...,dℓ−1′′,dℓ+2′′,...,dn′′) be the sequence
obtained from π′ by laying off dl′ with order.
By Lemma 2.2, π′ and π′′ both are graphic sequences.
Let ω=(f1,...,fn−2) be the sequence obtained from
π with dℓ and dℓ+1 removed,
and re-indexed so that the indices are consecutive, i.e., fi=di
for i∈[ℓ−1] and fi=di+2 for i∈[n−2]∖[ℓ−1].
Let ω′=(f1′,...,fn−2′) be the sequence obtained from π′
with dℓ′ removed, and re-indexed so that the indices are consecutive. Also,
let ω′′=(f1′′,...,fn−2′′) be the sequence obtained from
π′′ by re-indexing so that the indices are consecutive.
Note that ω′′ is a graphic sequence.
To turn a realization of ω′′ to a realization of π,
we need to track the changes between fi and fi′′ for all i∈[n−2]. Note that 0≤fi−fi′′≤2.
Let
[TABLE]
and
[TABLE]
So K=∑i∈[n−2]∣fi−fi′∣=∑i∈[n−2]∣fi′−fi′′∣; hence
[TABLE]
We now prove two claims asserting certain properties on X1 and
X2. For convenience, we introduce some notation.
For nonempty sets A and B of integers, we write A<B if the maximum integer in A is less than the minimum integer in B.
A set S of integers is consecutive if it consists of consecutive
integers. A sequence of pairwise disjoint sets, A1,...,At, of integers is said to be consecutive if A1∪...∪At is consecutive and, for any i,j∈[t] with i<j and Ai and Aj
nonempty, we have Ai<Aj.
Claim 1. There exist consecutive sets R1,R2,R1′,R2′,Q such
that X1=R1′∪R2′ and X2=R1∪R2
such that
(a)
the sequence R1,R1′,Q,R2′ is consecutive,
(b)
either R2=∅ or R2=Q, and
(c)
fi′′=fj′′+1 for all i∈R1′,j∈R2′.
Proof of Claim 1..
Let s be the minimum of the largest K numbers in ω=(f1,...,fn−2).
(Note that this s is the same as the s in the definition of laying off dℓ+1 with order from π.)
In order to keep
track whether fi′=fi or fi′=fi−1 and whether fi′′=fi′ or
fi′′=fi′−1, we divide [n−2] into six pairwise disjoint sets:
[TABLE]
By the definitions of π′ and ω′, we see that A,B,C,D,E,F is consecutive and
[TABLE]
Thus, it is easy to see that A∪B∪D={i∈[n−2]:fi′=fi−1}; so ∣A∣+∣B∣+∣D∣=K.
Let Y={i∈[n−2]:fi′′=fi′−1}. Then it follows that
[TABLE]
To complete our proof of Claim 1, we distinguish four cases based on
relations among the sizes of B,C,D,E.
First, suppose ∣C∣≥∣B∣+∣D∣. Let C′′ consist of the last ∣B∣+∣D∣ integers in C, and C′:=C∖C′′.
Then we see that Y=A∪C′′. Let R1=A, R2=∅, R1′=B,
R2′=C′′∪D and Q=C′.
It is easy to check that X1=R1′∪R2′ and X2=R1∪R2,
and (a) and (b) holds. Note that fi′′=s for i∈R1′, and
fj′′=s−1 for j∈R2′; so (c) holds.
Next, suppose ∣D∣≤∣C∣<∣B∣+∣D∣. Let B′′ consist of the last ∣B∣+∣D∣−∣C∣ integers in B, and B′=B∖B′′.
Then Y=A∪B′′∪C. Let R1=A, R2=Q=B′′, R1′=B′ and
R2′=C∪D.
It is easy to check that X1=R1′∪R2′ and X2=R1∪R2,
and that (a) and (b) holds.
Note that fi′′=s for i∈R1′, and fj′′=s−1 for j∈R2′;
so (c) holds.
Now assume ∣C∣<∣D∣≤∣C∣+∣E∣.
Let E′′ consist of the last ∣D∣−∣C∣ integers in E, and E′=E∖E′′.
Then Y=A∪B∪C∪E′′. Let R1=A∪B, R2=∅,
R1′=C∪D, R2′=E′′, and Q=E′.
It is easy to check that X1=R1′∪R2′ and X2=R1∪R2,
and (a) and (b) holds.
Note that fi′′=s−1 for i∈R1′ and fj′′=s−2 for j∈R2′; so (c) holds.
Finally we consider the case when ∣D∣>∣C∣+∣E∣.
Let D′′ consist of the last ∣D∣−∣C∣−∣E∣ integers in D, and D′=D∖D′′.
Then Y=A∪B∪C∪D′′∪E.
Let R1=A∪B, R2=Q=D′′, R1′=C∪D′ and R2′=E. It is easy to check that X1=R1′∪R2′ and X2=R1∪R2,
and (a) and (b) holds.
Note that fi′′=s−1 for i∈R1′ and fj′′=s−2 for j∈R2′; so (c) holds.
Let I1={i∈[n−2]:i≡1mod2} and I2={i∈[n−2]:i≡0mod2}.
By (2), we see ∣X1∣ must be even. So ∣R1′∣ and ∣R2′∣ are of the same parity.
Since both R1′ and R2′ are consecutive, ∣X1∩I1∣−∣X1∩I2∣∈{−2,0,2}.
Now suppose ∣X1∩I1∣−∣X1∩I2∣=0.
If R2=∅, then X2=R1 is a consecutive set and thus
∣X2∩I1∣ and ∣X2∩I2∣ differ by at most one. So we may
assume R2=∅. Then R2=Q by Claim 1. As the sequence R1,R1′,Q,R2′ is consecutive,
we see that X1∪X2 is a consecutive set; so ∣(X1∪X2)∩I1∣ and ∣(X1∪X2)∩I2∣ differ by at most one.
Hence, since ∣X1∩I1∣−∣X1∩I2∣=0, ∣∣X2∩I1∣−∣X2∩I2∣∣≤1.
We are ready to construct a realization of π=(d1,...,dn).
Recall that ω′′=(f1′′,...,fn−2′′) is a graphic sequence.
By induction hypothesis, there exists a realization F of ω′′
with V(F)={w1,...,wn−2} and dF(wi)=fi′′ for i∈[n−2], such that the parity bisection J of F satisfies
[TABLE]
Let Wj={wi:i≡jmod2} for j∈[2].
In what follows, we will construct a graph G as the realization of
π such that its parity bisection H of G satisfies
dH(v)≥(dG(v)−1)/2 for all v∈V(G),
by adding two new vertices a,b (so V(G)=V(F)∪{a,b}) and some
edges from these two vertices to F (which we will describe in three
separate cases).
Notice that if K=k−1, then we would add the edge ab as well; so
for convenience, let
[TABLE]
We write V(G)={v1,...,vn} such that
vi=wi for i<ℓ, {vℓ,vℓ+1}={a,b}, and
vi=wi−2 for ℓ+1<i≤n.
In view of Claim 2, we consider the following three cases. In each of
these three cases, we use a to represent the vertex in {vℓ,vℓ+1} with odd index.
So the parity partition of V(G) is
[TABLE]
Case 1. ∣X1∩I1∣−∣X1∩I2∣=0.
We know F⊆G and V(G)=V(F)∪{a,b}, and we need to add edges at a and
b to form G, a realization of π. Add
ab if ϵ=1, avi for all i∈X2∪(X1∩I2), and bvj for all j∈X2∪(X1∩I1).
Since ∣X1∩I1∣=∣X1∩I2∣, G is a realization of π.
Let H denote the parity bisection of G; so V1,V2 are the
partition sets of H. We need to show that dH(v)≥(dG(v)−1)/2 for all v∈V(G).
For each wi with i∈/X1∪X2, its
neighborhoods in F,G are the same; so by (3),
dH(wi)=dJ(wi)≥(dF(wi)−1)/2=(dG(wi)−1)/2.
For vertices wi with i∈X2, we have
dG(wi)=dF(wi)+2 and dH(wi)=dJ(wi)+1;
so by (3), dH(wi)=dJ(wi)+1≥(dF(wi)−1)/2+1=(dG(wi)−1)/2.
For vertices wi with i∈X1, we have
dG(wi)=dF(wi)+1 and dH(wi)=dJ(wi)+1; so
by (3), dH(wi)=dJ(wi)+1≥(dF(wi)−1)/2+1>(dG(wi)−1)/2.
For the vertex a, we have dG(a)=∣X2∣+∣X1∩I2∣+ϵ and dH(a)=∣X2∩I2∣+∣X1∩I2∣+ϵ.
Note that in this case, by Claim 2, we have ∣∣X2∩I1∣−∣X2∩I2∣∣≤1,
which implies that
[TABLE]
Hence, dH(a)≥(dG(a)−1)/2.
Similarly, for the vertex b, we have dG(b)=∣X2∣+∣X1∩I1∣+ϵ and dH(b)=∣X2∩I1∣+∣X1∩I1∣+ϵ.
Note, by Claim 2, ∣∣X2∩I1∣−∣X2∩I2∣∣≤1; so
[TABLE]
Hence, dH(b)≥(dG(b)−1)/2.
Case 2. ∣X1∩I2∣−∣X1∩I1∣=2.
Recall that X1=R1′∪R2′, where each Ri′ is consecutive.
Thus it follows that ∣Ri′∩I2∣=∣Ri′∩I1∣+1 for i∈[2].
Since the sequence R1,R1′,Q,R2′ is consecutive and starts from the integer 1,
we see that ∣R1∩I2∣=∣R1∩I1∣−1 and ∣Q∩I2∣=∣Q∩I1∣−1. Therefore, since R2=∅ or R2=Q (by (b) of Claim 1), we
have
[TABLE]
We claim that there exists some z∈X1∩I2 with
dJ(wz)≥dF(wz)/2.
To see this, choose x∈R1′∩I2 and y∈R2′∩I2.
By (3) ,
we have dJ(wx)≥(dF(wx)−1)/2 and dJ(wy)≥(dF(wy)−1)/2. By (c)
of Claim 1, dF(wx)=dF(wy)+1.
Observe that for any vertex u of F, dF(u) and 2dJ(u)−dF(u) are of the same parity;
so 2dJ(wx)−dF(wx) and 2dJ(wy)−dF(wy) must have different parities.
Therefore there exists z∈{x,y} such that dJ(wz)≥dF(wz)/2, proving the claim.
We now add edges at a and
b to form G from F: add ab if ϵ=1, avi for all i∈X2∪(X1∩I2)∖{z},
and bvj for all j∈X2∪(X1∩I1)∪{z}.
Since ∣X1∩I2∣=∣X1∩I1∣+2, G is a realization of π. Next we show that the parity
bisection H of G satisfies the property that dH(v)≥(dG(v)−1)/2 for all v∈V(G).
For each wi with i∈/X1∪X2, its
neighborhoods in F,G are the same; so by (3),
dH(wi)=dJ(wi)≥(dF(wi)−1)/2=(dG(wi)−1)/2.
For each wi with i∈X2, dG(wi)=dF(wi)+2 and dH(wi)=dJ(wi)+1.
Hence by (3) and the way we choose z, dH(wi)=dJ(wi)+1≥(dF(wi)−1)/2+1=(dG(wi)−1)/2.
For wi with i∈X1∖{z},
we have dG(wi)=dF(wi)+1 and dH(wi)=dJ(wi)+1;
so by (3), dH(wi)=dJ(wi)+1≥(dF(wi)−1)/2+1>(dG(wi)−1)/2.
The vertex wz satisfies
dG(wz)=dF(wz)+1 and dH(wz)=dJ(wz).
Hence by (3), dH(wz)=dJ(wz)≥dF(wz)/2=(dG(wz)−1)/2.
For the vertex a, by definition we have
dG(a)=∣X2∣+∣X1∩I2∣−1+ϵ and dH(a)=∣X2∩I2∣+∣X1∩I2∣−1+ϵ.
By (4) and the fact that ∣X1∩I2∣≥2,
[TABLE]
Hence, dH(a)≥(dG(a)−1)/2.
For the vertex b, we have
dG(b)=∣X2∣+∣X1∩I1∣+1+ϵ and dH(b)=∣X2∩I1∣+∣X1∩I1∣+ϵ.
This, together with (4), imply that
[TABLE]
Hence, dH(b)≥(dG(b)−1)/2.
Case 3. ∣X1∩I1∣−∣X1∩I2∣=2.
In this case, we have ∣Ri′∩I1∣=∣Ri′∩I2∣+1 for i∈[2]
(as R1′ and R2′ are consecutive).
Because R1,R1′,Q,R2′ is consecutive, it follows that
∣R1∩I1∣=∣R1∩I2∣ and ∣Q∩I1∣=∣Q∩I2∣−1. Since
R2=∅ or R2=Q (by (b) of Claim 1),
[TABLE]
Since ∣X1∣ is even and ∣Ri′∩I1∣=∣Ri′∩I2∣+1 for i∈[2], there exist x∈R1′∩I1 and y∈R2′∩I1. By (3) and (c) of Claim 1,
we have dJ(wx)≥(dF(wx)−1)/2, dJ(wy)≥(dF(wy)−1)/2,
and dF(wx)=dF(wy)+1.
Since for any vertex u of F, dF(u) and 2dJ(u)−dF(u) are of the same parity,
2dJ(wx)−dF(wx) and 2dJ(wy)−dF(wy) must have different parities.
Therefore there exists z∈{x,y} such that dJ(wz)≥dF(wz)/2.
We now add edges at a and b to form the graph G: add
ab if ϵ=1, avi for all i∈X2∪(X1∩I2)∪{z},
and bvj for all j∈X2∪(X1∩I1)∖{z}.
Since ∣X1∩I1∣=∣X1∩I2∣+2, G is a realization of π.
We need to verify that dH(v)≥(dG(v)−1)/2 for all v∈V(G).
If v=wi for some i∈/X1∪X2, then dG(wi)=dF(wi)
and dH(wi)=dJ(wi); so by (3), dH(wi)=dJ(wi)≥(dF(wi)−1)/2=(dG(wi)−1)/2.
If v=wi for some i∈X2, then dG(wi)=dF(wi)+2 and
dH(wi)=dJ(wi)+1; again by (3), dH(wi)=dJ(wi)+1≥(dF(wi)−1)/2+1=(dG(wi)−1)/2.
If v=wi for some i∈X1∖{z},
then dG(wi)=dF(wi)+1 and dH(wi)=dJ(wi)+1; so
by (3), dH(wi)=dJ(wi)+1≥(dF(wi)−1)/2+1>(dG(wi)−1)/2.
Now suppose v=wz. Note that dG(wz)=dF(wz)+1 and dH(wz)=dJ(wz).
So dH(wz)=dJ(wz)≥dF(wz)/2=(dG(wz)−1)/2.
Suppose v=a. Note that
dG(a)=∣X2∣+∣X1∩I2∣+1+ϵ and dH(a)=∣X2∩I2∣+∣X1∩I2∣+ϵ.
By (5),
[TABLE]
So dH(a)≥(dG(a)−1)/2.
Finally, suppose v=b. We have
dG(b)=∣X2∣+∣X1∩I1∣−1+ϵ and dH(b)=∣X2∩I1∣+∣X1∩I1∣−1+ϵ.
By (5) and the fact that ∣X1∩I1∣≥2,
[TABLE]
So dH(b)≥(dG(b)−1)/2.
3 Complete multipartite graphs
For convenience, we say that a bisection H of a graph G is good if for each v∈V(G), 2dH(v)≥dG(v)−1.
Thus the Bollobás-Scott conjecture says that every graph contains a good bisection.
Here we discuss which complete multipartite graphs have good bisections.
Throughout the rest of this section, let G:=Kr1,…,rk, and let X1,…,Xk denote the
partition sets of G with ∣Xi∣=ri for all i∈[k].
First, we note that whenever ∣V(G)∣ is even, G has a good bisection.
Since ∣V(G)∣ is even, V(G) has a partition V1,V2 such that ∣V1∣=∣V2∣, ∣∣Xi∩V1∣−∣Xi∩V2∣∣=1 if ∣Xi∣ is odd, and ∣Xi∩V1∣=∣Xi∩V2∣ if ∣Xi∣
is even. Let H denote the maximum bisection of G with partition
sets V1 and V2. For any v∈V(G), v∈Xi for some i∈[k]. Note that dG(v)=∣V(G)∣−∣Xi∣ and dH(v)≥(∣V(G)∣−∣Xi∣−1)/2=(dG(v)−1)/2.
We will see that this is not always the case when ∣V(G)∣ is odd.
The main result of this section is a necessary and sufficient condition for a complete multipartite graph with odd order
to contain a good bisection. As a consequence, we show that for many complete multipartite graphs G,
G (and even G minus an edge) does not have a good bisection.
(However, it is not hard to see that such G does have a bisection
H such that for each v∈V(G), dH(v)≥⌊(dG(v)−1)/2⌋.)
For a bisection H of G with partition sets V1 and V2,
we say that XicrossesH if Xi∩Vj=∅ for
j∈[2].
For a subset W⊆V(G), let W=V(G)∖W.
We need two easy lemmas.
Lemma 3.1**.**
Let G=Kr1,…,rk and let Xi, i∈[k], be the
partition sets of G. Suppose H is a good bisection of G with
partition sets V1 and V2. Then the following statements hold for i∈[k].
(i)
If Xi crosses H and ∣Xi∣ is even then ∣Xi∩V1∣=∣Xi∩V2∣ and
∣∣Xi∩V1∣−∣Xi∩V2∣∣≤1.
(ii)
If Xi crosses H and ∣Xi∣ is odd then ∣∣Xi∩V1∣−∣Xi∩V2∣∣=1 and
∣∣Xi∩V1∣−∣Xi∩V2∣∣≤2.
Proof.
Suppose Xi crosses H. Then there exist vj∈Xi∩Vj for
j∈[2]. Note that dG(v1)=dG(v2)=∣Xi∣.
Thus, since H is a good bisection, dH(vj)=∣Xi∩V3−j∣≥⌊∣Xi∣/2⌋ for j∈[2].
Notice that ∣Xi∩V1∣+∣Xi∩V2∣=∣Xi∣.
So ∣∣Xi∩V1∣−∣Xi∩V2∣∣≤1.
If ∣Xi∣ is even then ∣Xi∩V1∣=∣Xi∩V2∣, and (i) holds. So assume
∣Xi∣ is odd. Since ∣∣V1∣−∣V2∣∣≤1, ∣∣Xi∩V1∣−∣Xi∩V2∣∣≤2, and (ii) holds.
Lemma 3.2**.**
Let G=Kr1,…,rk with ∣V(G)∣ odd, let X1,…,Xk be the partition sets of G,
and let H be a good bisection of G with partition sets V1,V2 such that ∣V1∣=∣V2∣+1. Let
[TABLE]
and
[TABLE]
Let W1=⋃Xi∈X1Xi, W0=⋃Xi∈X0′Xi, ∣X1∣=t, and ∣X0′∣=t′.
Then
(i)
∣W1∩V1∣−∣W1∩V2∣=t,
(ii)
∣W0∩V1∣−∣W0∩V2∣=2t′, and
(iii)
∣W1∪W0∩V1∣−∣W1∪W0∩V2∣=−(t+2t′−1).
Proof.
First, we prove (i). If W1=∅, then ∣W1∩V1∣−∣W1∩V2∣=0=t.
So assume W1=∅ and let Xi∈X1. Then ∣Xi∣
is odd and Xi crosses H. Hence, since ∣V(G)∣ is odd, ∣Xi∣ is even.
By Lemma 3.1, ∣Xi∩V1∣=∣Xi∩V2∣.
Because ∣V1∣=∣V2∣+1, ∣Xi∩V1∣−∣Xi∩V2∣=1.
Hence, ∣W1∩V1∣−∣W1∩V2∣=∑Xi∈X1(∣Xi∩V1∣−∣Xi∩V2∣)=t.
We now prove (ii). If W0=∅ then t′=0 and the result
holds trivially. So assume W0=∅ and let Xi∈X0′.
Then ∣Xi∣ is even and Xi crosses H. Since ∣V(G)∣ is odd, ∣Xi∣ is odd.
By Lemma 3.1, ∣∣Xi∩V1∣−∣Xi∩V2∣∣=1,
and by the definition of X0′, ∣∣Xi∩V1∣−∣Xi∩V2∣∣=2.
Therefore, because ∣V1∣=∣V2∣+1, ∣Xi∩V1∣−∣Xi∩V2∣=2
(as well as ∣Xi∩V1∣−∣Xi∩V2∣=−1).
Hence ∣W0∩V1∣−∣W0∩V2∣=∑Xi∈X0′(∣Xi∩V1∣−∣Xi∩V2∣)=2t′.
It is easy to see that (iii) follows from (i), (ii) and the assumption ∣V1∣=∣V2∣+1.
We now give a necessary and sufficient condition for a complete
multipartite graph with odd order to admit a good bisection.
Let G=Kr1,…,rk with partition sets Xi, i∈[k], such that
∣Xi∣=ri. Let X={Xi:i∈[k]}, S1={Xi:i∈[k]\mboxand∣Xi∣≡1mod2} and S0={Xi:i∈[k]\mboxand∣Xi∣≡0mod2}.
For any A⊆X, let s(A)=∑Xi∈A∣Xi∣. We say that A is good if
there exists A′⊆A such that
[TABLE]
where m=∣S1∖A∣ and n is a nonnegative integer with
n≤∣S0∖A∣.
Proposition 3.3**.**
Let G=Kr1,…,rk with partition sets X1,…,Xk, and assume ∣V(G)∣ is odd.
Let X={Xi:i∈[k]}. Then G has a good bisection if and only if X has a good subset.
Proof.
First, we prove that if G has a good bisection, then X has a good subset.
Let H be a good bisection of G and let V1,V2 be the corresponding partition sets of H.
Since ∣V(G)∣ is odd, we may assume that ∣V1∣=∣V2∣+1. Let S={Xi:i∈[k]\mboxandXi\mboxcrossesH}, {\cal X}_{1}=\{X_{i}:i\in[k],X_{i}\mbox{ crosses H, and }|X_{i}|\equiv 1\mod 2\}, {\cal X}^{\prime}_{0}=\{X_{i}:i\in[k],X_{i}\mbox{ crosses H, }|X_{i}|\equiv 0\mod 2,\mbox{ and }||X_{i}\cap V_{1}|-|X_{i}\cap V_{2}||=2\}, and {\cal X}_{0}^{\prime\prime}=\{X_{i}:i\in[k],X_{i}\mbox{ crosses H, }|X_{i}|\equiv 0\mod 2,\mbox{ and }|X_{i}\cap V_{1}|=|X_{i}\cap V_{2}|\}.
By Lemma 3.1(ii) and the assumption ∣V1∣=∣V2∣+1, for every
Xi∈X crossing H and ∣Xi∣ even,
∣∣Xi∩V1∣−∣Xi∩V2∣∣=2 or ∣Xi∩V1∣=∣Xi∩V2∣. Hence, S=X1∪X0′∪X0′′.
Let A=X∖S.
Moreover, let W1=⋃Xi∈X1Xi, W0=⋃Xi∈X0′Xi,
W0′=⋃Xi∈X0′′Xi and W2=⋃Xi∈AXi.
Then V(H)=W1∪W0∪W0′∪W2. By the definition of X0′′, ∣W0′∩V1∣=∣W0′∩V2∣.
Let ∣X1∣=t and ∣X0′∣=t′; then
by Lemma 3.2(iii), ∣(W0′∪W2)∩V2∣−∣(W0′∪W2)∩V1∣=t+2t′−1.
Combining these two equalities, we get
[TABLE]
Since Xi does not cross H for any Xi⊆A, there exists A′⊆A such
that W2∩V2=⋃Xi∈A′Xi and W2∩V1=⋃Xi∈A∖A′Xi.
Now, ∣W2∩V2∣=s(A′) and ∣W2∩V1∣=s(A)−s(A′).
Thus s(A′)=s(A)/2+(t+2t′−1)/2 by (6).
Note that t=∣X1∣=∣S1∖A∣ and
t′=∣X0′∣≤∣X0′∪X0′′∣=∣S0∖A∣.
So A is a good subset of X.
Now, we prove that if X has a good subset, then G has a good bisection.
Let A be a good subset of X.
Then there exists A′⊆A such that s(A′)=s(A)/2+(m+2n−1)/2,
where m=∣S1∖A∣ and n≤∣S0∖A∣.
Let S0′⊆S0∖A with ∣S0′∣=n, and let S0′′=(S0∖A)∖S0′.
We partition V(G) into V1 and V2 such that
•
∣Xi∩V1∣−∣Xi∩V2∣=1 if Xi∈S1∖A,
•
∣Xi∩V1∣−∣Xi∩V2∣=2 if Xi∈S0′,
•
∣Xi∩V1∣−∣Xi∩V2∣=0 if Xi⊆S0′′,
and
•
Xi⊆V1 if Xi∈A∖A′, and
Xi⊆V2 if Xi∈A′.
Then
[TABLE]
Let H be the bisection of G with partition sets V1 and V2
and edge set E(H)={uv∈E(G):u∈V1\mboxandv∈V2}.
Next, we show that H is a good bisection of G.
Note that, for each v∈Xi⊆V(G),
dG(v)=∣Xi∣,
dH(v)=∣Xi∩V1∣ if v∈V2, and dH(v)=∣Xi∩V2∣ if v∈V1.
Also note that
[TABLE]
If v∈Xi for some Xi∈S1∖A, then
∣Xi∩V1∣−∣Xi∩V2∣=1−1=0; so
dH(v)=dG(v)/2.
If v∈Xi for some Xi∈S0′ then
∣Xi∩V1∣−∣Xi∩V2∣=1−2=−1; so
dH(v)≥(dG(v)−1)/2.
If v∈Xi for some Xi∈S0′′ then
∣Xi∩V1∣−∣Xi∩V2∣=1−0=1; so
dH(v)≥(dG(v)−1)/2.
If v∈Xi∩V2 for some Xi∈A then
∣Xi∩V1∣−∣Xi∩V2∣=1+∣Xi∣>0; so
dH(v)=∣Xi∩V1∣≥dG(v)/2.
Finally, suppose v∈Xi∩V1 for some Xi∈A. Then
∣Xi∩V1∣−∣Xi∩V2∣=1−∣Xi∣, i.e.,
∣Xi∩V2∣−∣Xi∩V1∣=∣Xi∣−1≥0. This
implies that dH(v)=∣Xi∩V2∣≥dG(v)/2.
Let G=Kr1,r2,r3 and X={X1,X2,X3} such that
X1,X2,X3 are the partition sets of G and ∣Xi∣=ri for i∈[3]. Let S0={Xi:i∈[3]\mboxand∣Xi∣≡0mod2} and
S1={Xi:i∈[3]\mboxand∣Xi∣≡1mod2}. Then
S0=∅ and S1=X.
If X has no good subset then the assertion follows from
Proposition 3.3. So assume that A is a good subset of
X with A′⊆A such that s(A′)=s(A)/2+(m+2n−1)/2, where m=∣S1∖A∣=3−∣A∣ and
n≤∣S0∖A∣=0.
So
[TABLE]
It is easy to see that A=∅. Since ri≥3≥∣A∣, A′=∅ and, hence, ∣A∣=1. Since r1,r2,r3 are all distinct, ∣A∣=2. So
∣A∣=3. Now a straightforward analysis shows that
for some i∈[3], ri∈{⌊(r1+r2+r3)/2⌋,⌈(r1+r2+r3)/2⌉}. This is a contradiction.
Proposition 3.3 characterizes those complete multipartite graphs
which do not have a good bisection. The next result says that there are more such
examples.
Proposition 3.4**.**
Let G=Kr1,…,rk where ri≥7 for every i∈[k].
Suppose G does not have a good bisection. Then for any edge e∈E(G),
G−e does not have a good bisection.
Proof.
Assume, to the contrary, that G′=G−e has a good bisection H′ with
partition sets V1,V2.
We may assume that E(H′)={xy∈E(G′):x∈V1\mboxandy∈V2}.
Then for every vertex v∈V(G′),
dH′(v)≥(dG′(v)−1)/2.
Let H be the bisection of G with partition sets V1 and V2
such that E(H)={xy∈E(G):x∈V1\mboxandy∈V2}.
Let e=uw.
Then dG(v)=dG′(v) for all v∈V(G)∖{u,w},
and dG(v)=dG′(v)+1 for v∈{u,w}.
Also, we have dH(v)=dH′(v) for all v∈V(G)∖{u,w},
and dH′(v)≤dH(v)≤dH′(v)+1 for v∈{u,w}.
Since H is not a good bisection of G, there exists a vertex v∈V(G) such that
dH(v)<(dG(v)−1)/2.
So we have
[TABLE]
which implies that dG′(v) is odd (since dH(v) is an integer),
dG(v)=dG′(v)+1, and
(dG′(v)−1)/2=dH′(v)=dH(v). Since
dG(v)=dG′(v)+1, v∈{u,w}.
Assume, without loss of
generality, that v=u∈Xi∩V1, where X1,…,Xk are the
partition sets of G. (Then w∈/Xi as uw∈E(G).)
So dG′(u)=∣Xi∣−1, ∣Xi∣ is even, and
w∈V1.
Thus,
∣V2∩Xi∣=dH′(u)=(dG′(u)−1)/2=∣Xi∣/2−1.
Therefore,
∣V1∩Xi∣=∣Xi∣−∣V2∩Xi∣=∣Xi∣/2+1.
So ∣V1∩Xi∣−∣V2∩Xi∣=2.
Because ∣∣V1∣−∣V2∣∣=1, ∣∣V1∩Xi∣−∣V2∩Xi∣∣≤3.
Since ∣Xi∣≥7, ∣V1∩Xi∣≥2.
Therefore, there exists a vertex v1∈V1∩Xi such that v1=u.
Also, v1=w since w∈/Xi.
Then dG′(v1)=dG(v1)=∣Xi∣ is even.
Thus dH′(v1)=∣V2∩Xi∣=∣Xi∣/2−1<(dG′(v1)−1)/2.
This contradicts the assumption that H′ is a good bisection of G′.
4 Scott’s questions on bipartitions
In this section, we address two questions of Scott [13] on bipartitions of graphs.
First, we prove Theorem 1.4 on ℓλ-norm of
bipartitions (with λ≥1), for which
we need a result of Bollobás and Scott [3] on judicious
bipartitions. Recall the definition of t(m)=m/2+1/16−1/4.
Lemma 4.1** (Bollobás and Scott).**
Let G be a graph with m edges. Then there exists a
bipartition V(G)=V1∪V2 such that e(V1,V2)≥m/2+t(m)/2 and max{e(V1),e(V2)}≤m/4+t(m)/4.
Moreover, if for every such bipartition V(G)=V1∪V2 it always holds that max{e(V1),e(V2)}=m/4+t(m)/4,
then G must be a complete graph of odd order.
Let λ≥1 and let G be a graph with m edges.
By Lemma 4.1, there is a bipartition V1,V2 of V(G) such
that e(V1,V2)≥m/2+t(m)/2 and, for i∈[2], e(Vi)≤m/4+t(m)/4=(2t(m)+1).
Note that
[TABLE]
Without loss of generality, we assume that e(V1)≥e(V2). Then
e(V2)≤t(m)2/2.
We claim that e(V1)λ+e(V2)λ≤(2t(m))λ+(2t(m)+1)λ.
This is true if e(V2)≤(2t(m)). So we may assume
(2t(m))≤e(V2)≤t(m)2/2. For λ≥1,
the function f(x)=(t(m)2−x)λ+xλ is strictly decreasing when (2t(m))≤x≤t(m)2/2.
Therefore,
[TABLE]
Now assume that for every bipartition V1,V2 of V(G), we have
e(V1)λ+e(V2)λ=(2t(m))λ+(2t(m)+1)λ. Then it follows from the above arguments, e(V2)=(2t(m)) and e(V1)+e(V2)=t(m)2.
So max{e(V1),e(V2)}=e(V1)=t(m)2−(2t(m))=m/4+t(m)/4.
By Lemma 4.1, G is a complete graph of odd order.
Remark. From the above proof, we see that actually
V(G) has a bipartition V1,V2 such that e(V1)λ+e(V2)λ≤(2t(m))λ+(2t(m)+1)λ for all λ≥1.
To extend Theorem 1.4 to k-partitions for k≥3, we need the following result of Xu and Yu [14] on
k-partitions.
Lemma 4.2** (Xu and Yu).**
Let G be a graph with m edges and let k≥3 be a positive
integer. Then there exists a k-partition V(G)=V1∪...∪Vk such that
[TABLE]
and for i∈[k],
[TABLE]
We now determine the ℓλ-norm (where λ≥1) for k-partitions,
up to an additive term O(mλ−1). The proof is similar to the bipartition case.
Theorem 4.3**.**
Let k≥3 be an integer and λ≥1 be a real number. Then any graph G with m edges has a k-partition
V(G)=V1∪...∪Vk such that
[TABLE]
Proof.
Let G be a graph with m edges.
By Lemma 4.2, there exists a k-partition V(G)=V1∪...∪Vk such that
[TABLE]
and for i∈[k],
[TABLE]
Without loss of generality, let e(V1)≥e(V2)≥…≥e(Vk) and let α:=e(V1)−m/k2.
If α≤−k2k−1t(m) then
[TABLE]
So we may assume that
[TABLE]
Note that we may assume ∑i=1ke(Vi)=km−kk−1t(m)+817k. Also note
that ∑i=1ke(Vi)λ increases if we replace e(Vk)
by e(Vk)−1 and replaces e(Vi) by e(Vi)+1, for any i∈[k−1].
Therefore, we may further assume that
e(V1)=...=e(Vk−1)=k2m+α and
e(Vk)=k2m−kk−1t(m)+817k−(k−1)α.
So by (7), we have
[TABLE]
where the second inequality holds because the expression in the second
line is an increasing function of α, for −k2k−1t(m)≤α≤k2k−1t(m).
We remark that the bound in the above theorem is tight up to the term
O(mλ−1),
by considering the complete graph Kks which has m=(2ks) edges.
Thus s=(2t(m)+1)/k.
The minimum ∑i=1ke(Vi)λ over all k-partitions
V1,…,Vk of V(Kks)
is attained when ∣Vi∣=s for i∈[k], and this minimum value equals
[TABLE]
Using 2t(m)2+t(m)=m and t(m)=Θ(m), we see that
[TABLE]
It would be interesting to find the optimal upper bound in Theorem 4.3.
We believe that the extremal graphs for ℓλ-norms of k-partitions (where λ≥1)
should be the complete graphs Kkn+⌊k/2⌋.
We formulate the following question.
For fixed λ≥1 and integer k≥2, let s:=s(m) be such that m=(2ks+⌊k/2⌋) and let
[TABLE]
Question 4.4**.**
Fix any real λ≥1 and integer k≥2.
For any positive integer m, is it true that
[TABLE]
for all graphs G with m edges,
with equality if and only if m=(2ks+⌊k/2⌋) for some integer s?
Does the equality hold only for Kks+⌊k/2⌋ (modulo some isolated vertices)?
A result of Bollobás and Scott [4] shows that this is true for λ=1 and any k.
Theorem 1.4 provides an affirmative answer for the case k=2.
We now turn to the following question of Scott [13].
Question 4.5**.**
Does every graph G with
(2kn) edges have a vertex partition into k sets, each of
which contains at most (2n) edges?
We give a negative answer to this question in the case k=2. For this
we need to show that there exist an infinite sequence of pairs of
integers with certain properties.
Lemma 4.6**.**
There are pairs (ai,bi) of integers for all i≥0 such that
(i)
ai≥36* and ai is even, and bi≥21 and bi is odd,*
(ii)
3bi(bi−1)=ai(ai−1), and
(iii)
bi≤7ai/12.
Proof.
We recursively define integer pairs (ni,ti) as follows, such that
the desired sequence {(ai,bi)}i≥0 will be a subsequence of
{(ni,ti)}i≥0.
Let
[TABLE]
and, for i≥1, let
[TABLE]
For convenience, we write
[TABLE]
for
i≥0, and
[TABLE]
for i≥1.
We claim that αi=0 for i≥0 and that βi=0 for i≥1.
By a direct calculation, we see that α0=0, α1=0 and β1=0.
Now assume for some i≥1, we have αj=0 for j∈[i]∪{0}, and βj=0 for j∈[i].
Using (8) and the definition of αi+1 and
βi+1, we have
[TABLE]
Thus, the claim follows from induction.
From (8), we see that both {ni}i≥0 and
{ti}i≥0 are increasing sequences; so ni≥36 and ti≥21 for i≥0.
Moreover, ti≤7ni/12 for i≥0. For otherwise,
ti>7ni/12 for some i. Then i≥1 and
[TABLE]
which is larger than ni(ni−1) (since ni≥36), a contradiction.
Using (8), it is easy to observe that
ni is even if and only if i≡0,3mod4,
and that ti is odd if and only if i≡0,1mod4.
Therefore letting ai=n4i and bi=t4i for i≥0, we see
that the pairs (ai,bi) satisfy all requirements (i), (ii) and
(iii).
By Lemma 4.6, there exist infinitely many pairs (2n,t)
of positive integers such that t is odd, t≤7n/6, and
3t(t−1)=2n(2n−1).
Let G be the union of three pairwise disjoint copies of the clique Kt. Then
∣V(G)∣=3t and
[TABLE]
Let V1,V2 be a bipartition of V(G). Without loss of generality, we assume that ∣V1∣≥∣V2∣.
Then G[V1] is the disjoint union of three cliques, say Ka,Kb, and
Kc. (We set K0=∅.) Hence, G[V2] is the disjoint union of cliques Kt−a,Kt−b and Kt−c.
As t is odd, we have
[TABLE]
Choose integers a′,b′,c′ such that a′≤a,b′≤b,c′≤c and
[TABLE]
We also need an easy property of binomial coefficients that for any integers m−n≥2,
It seems likely that similar result holds for general k-partitions,
though we are not able to construct such graphs due to difficulties in
proving a more general version of Lemma 4.6.
Bibliography14
The reference list from the paper itself. Each links out to its DOI / PubMed record.
1[1] N. Alon, Bipartite subgraphs, Cambinatorica 16 (1996), 301–311.
2[2] A. Ban and N. Linial, Internal partitions of regular graphs, J. Graph Theory 83 (2016), 5–18.
3[3] B. Bollobás and A.D. Scott, Exact bounds for judicious partitions of graphs, Combinatorica 19 (1999), 473-486.
4[4] B. Bollobás and A.D. Scott, Better bounds for max cut, in: Contemporary Combinatorics , in: Bolyai Soc. Math. Stud. , vol. 10, 2002, pp. 185–246.
5[5] B. Bollobás and A. D. Scott, Problems and results on judicious partitions, Random Struct. Alg. 21 (2002), 414–430.
6[6] C.S. Edwards, Some extremal properites of bipartite subgraphs, Canad. J. Math. 25 (1973), 475–485.
7[7] P. Erdős, Some recent problems in Combinatorics and Graph Theory, in Combinatorics and Computing (Proc. 26th Southeastern International Conference on Graph Theory, Boca Raton, 1995) Congressus Numerantium 112, (1995).
8[8] R. Häggkvist, Problems, Combinatorics, North-Holland, Amsterdam, New York, 1978, p. 1203.