This paper studies a function related to bipartite graph concatenation, exploring its properties, symmetries, and discontinuities, and raises open questions about its behavior.
Contribution
It introduces the function (x,y) for bipartite graph concatenation, analyzes its properties, symmetry, and points of discontinuity, and proposes new questions and conjectures.
Findings
01
(x,y) is symmetric in x and y
02
(x,y) has a discontinuity at x=y=1/k for integers k>1
03
The paper raises open questions about the function's properties
Abstract
Let x,y∈(0,1] and let A,B,C be disjoint nonempty subsets of a graph G, where every vertex in A has at least x∣B∣ neighbours in B, and every vertex in B has at least y∣C∣ neighbours in C. We denote by ϕ(x,y) the maximum z such that, in all such graphs G, there is a vertex v in C that is joined to at least z∣A∣ vertices in A by two-edge paths. The function ϕ is interesting, and we investigate some of its properties. For instance, we show that it is symmetric in x and y, and that it has a discontinuity at x=y=1/k for all integers k>1. We raise a number of questions and conjectures.
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 · Advanced Graph Theory Research · Graph theory and applications
Full text
Concatenating bipartite graphs
Maria Chudnovsky
Princeton University, Princeton, NJ 08544, USA
Patrick Hompe
Princeton University, Princeton, NJ 08544, USA
Alex Scott
Mathematical Institute, University of Oxford, Oxford OX2 6GG, UK
Paul Seymour
Princeton University, Princeton, NJ 08544, USA
Sophie Spirkl
Rutgers University, New Brunswick, NJ 08901, USA
Supported by NSF grant DMS 1763817 and US Army Research Office Grant W911NF-16-1-0404.Supported by a Leverhulme Trust Research
Fellowship.Supported by ONR grant N00014-14-1-0084 and NSF
grants DMS-1265563 and DMS-1800053.This material is based upon work supported by the National Science
Foundation under Award No. DMS-1802201.
(September 21, 2018; revised )
Abstract
Let x,y∈(0,1]; and let A,B,C be disjoint nonempty subsets of a graph G, where every vertex in A has at least x∣B∣
neighbours in B, and every vertex in B has at least y∣C∣ neighbours in C. We denote by ϕ(x,y)
the maximum z such that, in all such graphs G,
there is a vertex v∈C that is joined to at least z∣A∣ vertices in A by two-edge paths.
The function ϕ is interesting, and we investigate some of its properties. For instance, we show that
•
ϕ(x,y)=ϕ(y,x) for all x,y; and
•
for each integer k>1, there is a discontinuity in ϕ(x,x) when x=1/k: ϕ(x,x)≤1/k when x≤1/k, and
ϕ(x,x)≥2k(k−1)2k−1 when x>1/k.
We raise several questions and
conjectures.
1 Introduction
All graphs in this paper are finite, and have no loops or multiple edges.
We denote the semi-open interval {x:0<x≤1} of real numbers by (0,1].
Let x,y∈(0,1]; and let A,B,C be disjoint nonempty subsets of a graph G, where every vertex in A has at least x∣B∣
neighbours in B, and every vertex in B has at least y∣C∣ neighbours in C. If we ask for a real number z
such that we can guarantee that some vertex in A can reach at least z∣C∣ vertices in C by two-edge paths,
then z must be at most y, since perhaps all the vertices in B have the same neighbours in C. But in the reverse
direction the question becomes much more interesting; that is, we ask for z such that some vertex in C can reach
at least
z∣A∣ vertices in A by two-edge paths. Then there might well be values of z>max(x,y) with this property.
Let us say this more precisely. A tripartition of a graph G is a partition (A,B,C) of V(G) where A,B,C are
all nonempty stable sets. For x,y∈(0,1], we say a graph G is (x,y)-constrained, via a tripartition (A,B,C),
if
•
every vertex in A has at least x∣B∣
neighbours in B;
•
every vertex in B has at least y∣C∣ neighbours in C; and
•
there are no edges between A and C.
For v∈V(G), N(v) denotes its set of neighbours, and N2(v) is the set of vertices with distance exactly two
from v. We write NA2(v) for N2(v)∩A, and so on.
A first observation:
1.1
Let x,y∈(0,1], and let Z be the set of all z∈(0,1] such that,
for every graph G, if G is (x,y)-constrained via (A,B,C)
then ∣NA2(v)∣≥z∣A∣ for some v∈C. Then sup{z∈Z} belongs to Z.
Proof. Let z′=sup{z∈Z}, and let G be an (x,y)-constrained graph, via (A,B,C). We must show that
∣NA2(v)∣≥z′∣A∣ for some v∈C.
We may assume that z′>0; so there exists
z with 0<z<z′, such that ⌈z∣A∣⌉=⌈z′∣A∣⌉. Since z′=sup{z∈Z}
and z<z′, and Z is an initial interval of (0,1], it follows that z∈Z,
and so ∣NA2(v)∣≥z∣A∣ for some v∈C. Consequently ∣NA2(v)∣≥⌈z∣A∣⌉≥z′∣A∣,
as required. This proves 1.1.
We define ϕ(x,y) to be sup{z∈Z}, as defined in 1.1. The objective of this paper
is to study the properties of the function ϕ.
We have a trivial lower bound:
1.2
ϕ(x,y)≥max(x,y)* for all x,y>0.*
Proof. Let G be (x,y)-constrained, via (A,B,C). Since every vertex in A has at least x∣B∣ neighbours in B, and
B=∅, there exists u∈B with at least x∣A∣ neighbours in A; let v∈C be adjacent to u
(this is possible since y>0), and then ∣NA2(v)∣≥x∣A∣. Consequently ϕ(x,y)≥x. Now every
vertex in A can reach at least y∣C∣ vertices in C by two-edge paths (since x>0); and so by averaging, some
vertex in C can reach at least y∣A∣ vertices in A by two-edge paths. Hence ϕ(x,y)≥y.
This proves 1.2.
And a trivial upper bound:
1.3
For all x,y∈(0,1],
[TABLE]
for every integer k≥1.
Proof. Let x,y∈(0,1], and let k≥1 be an integer. Let A,B,C be three disjoint sets each of cardinality k,
where A={a1,…,ak}, B={b1,…,bk} and C={c1,…,ck}. Make a graph G with vertex set
A∪B∪C as follows. Let g=⌈kx⌉, and for 1≤i≤k make ai adjacent to
bi,bi+1,…,bi+g−1 (reading subscripts modulo k). Now let h=⌈ky⌉, and for
1≤i≤k make bi adjacent to
ci,ci+1,…,ci+h−1 (reading subscripts modulo k). Then G is (x,y)-constrained via (A,B,C);
and for 1≤i≤k, NA2(ci)={ai,ai−1,…,ai−g−h+2} (again, reading subscripts modulo k).
Consequently ϕ(x,y)≤(g+h−1)/k. This proves 1.3.
In particular, we have:
1.4
For every integer k≥1, if x,y>0 and max(x,y)=1/k then ϕ(x,y)=1/k.
Proof. From 1.2, ϕ(x,y)≥1/k; and the graph consisting of k disjoint three-vertex paths shows that ϕ(x,y)≤1/k.
(This also follows from 1.3, since ⌈kx⌉,⌈ky⌉=1.)
This proves 1.4.
What makes the function ϕ interesting is that for some values of x,y, 1.2 is far from best possible,
and indeed 1.3 seems closer to the truth. We were originally motivated by the hope of extending
Kneser’s theorem from additive group theory [5] to a general graph-theoretic setting, and
a corresponding wild conjecture that
the bound in 1.3 is always best possible, that is, that for all x,y∈(0,1], there
is an integer k>0 with ϕ(x,y)=k⌈kx⌉+⌈ky⌉−1.
This turns out to be false, but perhaps not ridiculously false; maybe something
like it is true.
There are two other related problems:
•
Let us say G is (x,y)-biconstrained (via(A,B,C)) if G is (x,y)-constrained via (A,B,C), and in addition
–
every vertex in B has at least x∣A∣ neighbours in A, and
–
every vertex in C has at least y∣B∣ neighbours in B.
•
Say G is (x,y)-exact (via(A,B,C)) if
G is (x,y)-constrained via (A,B,C), and in addition there exist x′≥x and y′≥y such that
–
every vertex in A has exactly x′∣B∣ neighbours in B;
–
every vertex in B has exactly x′∣A∣ neighbours in A;
–
every vertex in B has exactly y′∣C∣ neighbours in C; and
–
every vertex in C has exactly y′∣B∣ neighbours in B.
We shall sometime use “mono-constrained” to clarify that we mean the (x,y)-constrained case and not the (x,y)-biconstrained
case.
Let ψ(x,y) be the analogue of ϕ(x,y) for biconstrained graphs; that is, the maximum z such that for
all G, if G is (x,y)-biconstrained via (A,B,C), then ∣NA2(v)∣≥z∣A∣ for some v∈C. (As before, this
maximum exists.) Similarly, let ξ(x,y) be the analogue of ϕ and ψ for the exact case.
Then we have
1.5
For all x,y∈(0,1],
[TABLE]
for every integer k≥1.
The proof of the non-trivial part of this is the same as the proof of 1.3.
One might hope that ψ (and even more ξ) are better-behaved than ϕ.
Let us see an example. Start with the graph of figure 1. Each vertex has a number
written next to it in the figure; replace each vertex v by a set Xv of new vertices
of the specified cardinality, and for each edge uv of the figure make every vertex in Xu adjacent
to every vertex in Xv. This results in a graph with 81 vertices, divided into three sets of 27 corresponding to
the three rows of the figure; call these A,B,C. The graph produced is (13/27,1/9)-biconstrained via (A,B,C), and
yet ∣NA2(v)∣=13 for every vertex v∈C; so this proves that ψ(13/27,1/9)≤13/27 (and therefore
equality holds, by 1.2). This shows that there need not exist an integer k with
ψ(x,y)=k⌈kx⌉+⌈ky⌉−1. The same graph, used from bottom to top, shows that
ψ(1/9,13/27)=13/27.
The example is not yet (13/27,1/9)-exact, because some vertices in A have three, four or five neighbours in
B, and vice versa. We can make it exact as follows. For each edge uv of the figure with u in the second row
and v in the third, the two sets Xu,Xv have the same cardinality, one of three, four, five. Delete some edges
between Xu and Xv such that every vertex in Xu has exactly three neighbours in Xv and vice versa. Then
the modified graph is (13/27,1/9)-exact, and shows that ξ(13/27,1/9)=13/27. Consequently, even for the
supposedly nicest function ξ of our three functions, there is not always an integer k with
ξ(x,y)=k⌈kx⌉+⌈ky⌉−1.
So what can we prove about the functions ϕ and ψ? For which x,y,z is ϕ(x,y)≥z, or ψ(x,y)≥z?
In order to make the question a little more manageable,
we focus on seven special cases, x=y, and z=1/2,2/3,1/3,3/4,2/5,3/5, but in each case the results for ϕ and for ψ are
quite different. The paper is organized as follows:
•
We begin with a proof that ϕ(x,y)=ϕ(y,x) for all x,y.
•
Then we give some general upper bounds on ϕ(x,y) and ψ(x,y), particularly focussing on the
case when x=y.
•
Next we consider when ϕ(x,y)≥1/2, or ψ(x,y)≥1/2. There are several theorems that this is true
for certain pairs (x,y), and their union fills a good part of the (x,y)-square. We also give a number of constructions
that shows the statement is not true for certain pairs (x,y). Ideally this would fill the complementary part
of the square, but there is an “undecided” band of varying width down the middle.
•
Then we do the same for 2/3 instead of 1/2; and then for 1/3,3/4,2/5,3/5.
•
Finally, we discuss some other questions and approaches.
In this section we prove that ϕ(x,y)=ϕ(y,x) for all x,y. The argument uses linear programming, and we need
some preparation.
We denote the set of real numbers by R, and the non-negative reals numbers by R+.
A weighted graph(G,w) consists of a graph G together with a function w:V(G)→R+.
If X⊆V(G), we denote ∑v∈Xw(v) by w(X).
Let (G,w) be a weighted graph, and (A,B,C) a tripartition of G. If x,y∈(0,1], a weighted graph (G,w)
is (x,y)-constrained via (A,B,C), if:
•
∑v∈Aw(v)=∑v∈Bw(v)=∑v∈Cw(v)=1;
•
for each v∈A, w(N(v)∩B)≥x; and
•
for each v∈B, w(N(v)∩C)≥y.
Similarly, we say (G,w) is (x,y)-biconstrained via (A,B,C), if in addition:
•
for each v∈B, w(N(v)∩A)≥x; and
•
for each v∈C, w(N(v)∩B)≥y.
To make the graph of figure 1 into an appropriate weighted graph, divide
all the numbers by 27.
2.1
For x,y,z∈(0,1], the following are equivalent:
•
ϕ(x,y)≥z;
•
w(NA2(v))≥z* for some v∈C, for every weighted graph (G,w)
that is (x,y)-constrained via a tripartition (A,B,C).*
Similarly, the following are equivalent:
•
ψ(x,y)≥z;
•
w(NA2(v))≥z* for some v∈C, for every weighted graph (G,w)
that is (x,y)-biconstrained via a tripartition (A,B,C).*
Proof. To prove the “if” direction of the first statement, let G be (x,y)-constrained via (A,B,C). Define
w(v)=1/∣A∣, for each v∈A, and w(v)=1/∣B∣ for v∈B and similarly for v∈C. Then (G,w)
is an (x,y)-constrained weighted graph, and the claim follows. The “if” direction of the second statement is proved similarly.
For the “only if” direction, let (G,w) be a weighted graph, (x,y)-constrained via (A,B,C), and suppose such a weighted graph
can be chosen with w(NA2(v))<z for each v∈C. Consequently we may choose (G,w) such that in addition,
w is rational-valued. Choose an integer N>0 such that Nw(v) is an integer for each v∈G.
For each v∈V(G), take a set Xv of Nw(v) new vertices; and make a graph G′ with vertex set
⋃v∈V(G)Xv, by making every vertex of Xu adjacent to every vertex of Xv for all adjacent u,v∈V(G).
Let A′=⋃v∈AXv, and define B′,C′ similarly; then (A′,B′,C′) is a tripartition of G′, and
G′ is (x,y)-constrained via (A′,B′,C′). Since in G, w(NA2(v))<z for each v∈C, it follows that in G′,
∣NA′2(v′)∣<z∣A′∣ for each v′∈C′, a contradiction. The “only if” direction of the second statement is similar.
This proves 2.1.
Let G be a graph with a bipartition (A,B),
and let w:B→R+ be some function. We define
w(A→B) to mean the minimum, over all u∈A, of w(N(u)) (taking
w(A→B)=0 if A=∅).
2.2
Let G be a graph with a bipartition (A,B),
and let w:B→R+ be some function
such that w(B)=1.
Then either
•
there is a function w′:B→R+, such that w′(B)=1 and
w′(A→B)≥w(A→B), and such that w′(v)=0 for some v∈B; or
•
there is a function f:A→R+, such that f(A)=1 and
f(B→A)≥w(A→B).
Proof. We may assume that A=∅. If some vertex in A has no neighbour in B, then
w(A→B)=0 and the second bullet holds; so we assume that each vertex in A has a neighbour in B.
Let x=w(A→B). The function w′, defined by w′(v)=1/∣B∣ for each v∈B, satisfies
w′(A→B)>0, since
every vertex in A has a neighbour in B. Thus we may assume that x>0, replacing w by w′ if necessary.
Let M be the 0/1-matrix (auv:u∈A,v∈B), where auv=1 if and only if u,v are adjacent.
Let 1A∈RA be the vector of all 1’s, and define
1B similarly. Then w∈R+B satisfies:
•
1BTw=1; and
•
Mw≥x1A.
Consequently b=w/x satisfies b∈R+B, and
•
1BTb=1/x; and
•
Mb≥1A.
Choose q∈R+B with Mq≥1A, with 1BTq minimum. (This is possible by compactness.)
Thus 1BTq≤1/x.
Since Mq≥1A and G has an edge, it follows that 1BTq>0; let
1/y=1BTq, and define
w′=yq. Then y≥x, and 1BTw′=1 and Mw′≥y1A, and so we may assume that w′(v)>0 for each v∈B,
because otherwise the first bullet holds.
Now q minimizes 1BTq subject to the linear programme
q∈R+B and Mq≥1A. From the
linear programming duality theorem,
there exists p∈R+A such that pTM≤1BT, and pT1A=1BTq=1/y.
Define f=yp. Then f:A→R+ satisfies f(A)=1, and
f(N(v))≤y for each v∈B.
Let v′∈B;
we claim that f(N(v′))=y. This follows from the “complementary slackness” principle, but we give the argument
in full, as follows. Let s=w′(v′)(y−f(N(v′))). Thus s≥0, and we will show s=0. We have
[TABLE]
Consequently s=0, as claimed. Hence f satisfies the second bullet. This proves 2.2.
Proof. Let z=ϕ(x,y), and choose a weighted graph (G,w) that is (x,y)-constrained via (A,B,C), such that
w(NA2(v))≤z for each z∈C. Moreover, choose G with ∣V(G)∣ minimum. If there is a function
w′:B→R+, such that w′(B)=1 and
w′(A→B)≥w(A→B), and such that w′(v)=0 for some v∈B, then we may replace
w by a new weight function, changing w to w′ on B and otherwise keeping w unchanged, and then we may delete the vertex
v∈B with w′(v)=0, contrary to the minimality of ∣V(G)∣. Thus there is no such w′, and so by 2.2,
there is a function f:A→R+, such that f(A)=1 and
f(B→A)≥w(A→B)≥x. Similarly, there is a function
g:B→R+, such that g(B)=1 and
g(C→B)≥y. Let H be the graph with bipartition (A,C) in which u∈A and v∈C are adjacent
if u∈/NA2(v) in G. Thus, in H, w(C→A)≥1−z; and so from 2.2 and the minimality of ∣V(G)∣,
there is a function h:C→R+, such that h(C)=1 and
(in H) h(A→C)≥1−z. Let w′ be defined by the union of f,g and h in the natural sense;
then (G,w′) is a weighted graph and is (y,x)-constrained via (C,B,A), and w′(NC2(v))≤z for each v∈A.
This proves that ϕ(y,z)≤z, and so proves 2.3.
We remark that we have not been able to prove an analogue of 2.3 for the biconstrained case, or for the exact case,
although
we have no counterexample for either one.
There is another useful application of 2.2, the following:
2.4
Let (G,w) be an (x,y)-constrained weighted graph, via (A,B,C), such that w(NA2(v))≤z for each v∈C.
Suppose that there exists X⊆A with ∣X∣<z−1 such that ⋃v∈XNC2(v)=C. Then there exists u∈A
and a weighted graph (G′,w′) such that
•
G′* is obtained from G by deleting u;*
•
(G′,w′)* is (x,y)-constrained via (A′,B,C), where A′=A∖{u};*
•
in G′, w′(NA′2(v))≤z for all v∈C; and
•
w′(u)=w(u)* for all u∈B∪C.*
Proof. Suppose not. Let H be the graph with bipartition (A,C), in which u∈A and v∈C are adjacent
if u∈/NA2(v) in G. Then by 2.2, applied to H, there is a function
h:C→R+, such that h(C)=1 and
(in H) h(A→C)≥1−z. Consequently, in G, h(NC2(v))≤z for each v∈A. In particular,
h(NC2(v))≤z for each v∈X, and so h(C)≤z∣A∣<1, a contradiction. This proves 2.4.
Let x,y,z∈(0,1]. We say that (x,y,z) is triangular if no triangle-free graph G admits a tripartition
A,B,C
of V(G) with the following properties:
•
A,B,C are nonempty stable sets;
•
every vertex in A has at least x∣B∣ neighbours in B;
•
every vertex in B has at least y∣C∣ neighbours in C; and
•
every vertex in C has at least z∣A∣ neighbours in A.
It is possible to reformulate results about ϕ(x,y) in terms of triangular triples, because we have:
2.5
For x,y,z∈(0,1], ϕ(x,y)>1−z if and only if (x,y,z) is triangular.
Consequently the three statements ϕ(x,y)≤1−z, ϕ(z,x)≤1−y, and ϕ(y,z)≤1−x are equivalent.
Proof. Suppose that (x,y,z) is not triangular. Then there is a triangle-free graph G with a tripartition (A,B,C),
satisfying the three bullets in the definition of “triangular”. Let H be the subgraph of G
with V(H)=V(G), obtained by deleting all edges between A and C. If v∈C, then NA2(v) (defined
with respect to H) contains only vertices in A that are nonadjacent to v in G, since G is triangle-free;
and so ∣NA2(v)∣≤∣A∣−z∣A∣, since in G, v has at least z∣A∣ neighbours in A. Consequently ϕ(x,y)≤1−z.
For the reverse implication, suppose that ϕ(x,y)≤1−z, and let H be (x,y)-constrained via (A,B,C),
such that ∣NA2(v)∣≤∣A∣−z∣A∣ for each v∈C. Make a graph G by adding certain edges to H, namely
for each v∈C and u∈A, add an edge uv if u∈/NA2(v). Then G is triangle-free, and every vertex
v∈C is adjacent in G to at least ∣A∣−(1−z)∣A∣=z∣A∣ vertices in A; and so (x,y,z) is not triangular.
In particular, (x,y,z) is triangular if and only if (z,x,y) is triangular; so it follows that
ϕ(x,y)≤1−z if and only if ϕ(z,x)≤1−y, and similarly if and only if ϕ(y,z)≤1−x.
This proves 2.5.
We call the equivalence of the second statement of 2.5 “rotating”.
3 Constructions
In this section we construct some graphs to prove upper bounds on ϕ(x,y) or ψ(x,y) for certain values of x,y.
We begin with:
3.1
Let x,y∈(0,1], and let z∈(0,1] such that z/(1−z)=ϕ(x/(1−x),y/(1−y)); then ϕ(x,y)≤z.
Proof. Let (G′,w′) be a weighted graph that is (x/(1−x),y/(1−y))-constrained via some tripartition (A′,B′,C′), such that w′(NA′2(v))≤z/(1−z)
for each v∈C′. Add three new vertices a,b,c to G′, and two edges ab and bc, forming G. Define w by
[TABLE]
Then G is (x,y)-constrained via (A′∪{a},B′∪{b},C′∪{c}) and shows that ϕ(x,y)≤z.
This proves 3.1.
3.2
Let k≥0 be an integer, and let x,y∈(0,1] with 1−kxx+1−kyy≤1, with strict inequality if x or y
is irrational; then ϕ(x,y)<k+11.
Proof. By increasing x and y if necessary, we may assume that x,y are rational.
Suppose first that k=0; then we may assume that x+y=1. Choose an integer k≥1
such that kx (and hence ky) is an integer. By 1.3,
[TABLE]
This completes the proof for k=0. For general k
we proceed by induction on k. We may assume that k>0; let x,y∈(0,1] with 1−kxx+1−kyy≤1,
with strict inequality if x or y
is irrational. Let x′=x/(1−x), and y′=y/(1−y). Thus x′,y′∈(0,1] with
[TABLE]
with strict inequality if x′ or y′
is irrational. From the inductive hypothesis, ϕ(x′,y′)<1/k. Let z satisfy
z/(1−z)=ϕ(x′,y′); then z/(1−z)<1/k, and so z<1/(k+1). From 3.1, ϕ(x,y)≤z<1/(k+1).
This proves 3.2.
3.3
Let k≥0 be an integer, and let x,y∈(0,1] with x+ky≤1 and kx+y≤1, with strict inequality in both if x or y
is irrational; then ψ(x,y)<k1.
Proof. Again, we may assume that x,y are rational. Let s=max(x,y); thus, s<1/k. Choose an integer N≥1
such that p=xN/(1−(k−1)s) and q=yN/(1−(k−1)s) are integers. (It follows that p+q≤N, from the hypothesis.)
Let G be a graph with vertex set partitioned into three sets A,B,C,
with ∣A∣=N+k and ∣B∣=∣C∣=N+k−1; let
[TABLE]
Let G have the following edges:
•
for 1≤i≤N, ai is adjacent to bi,bi+1,…,bi+p−1 reading subscripts modulo N;
•
for 1≤i≤N, bi is adjacent to ci,ci+1,…,ci+q−1 reading subscripts modulo N;
•
for 1≤i≤k−1, ai′ is adjacent to bi′, and bi′ is adjacent to ci′.
•
a∗ is adjacent to bi for 1≤i≤N.
(Thus, this is the same as in the proof of 3.2, except for the extra vertex a∗.)
Let r satisfy
[TABLE]
Thus r>0.
For each v∈V(G), define w(v) as follows:
•
w(v)=(k−1)r(N+1)/N for v∈{a1,…,aN}; w(v)=1/k−r for v∈{a1′,…,ak−1′};
•
w(a∗)=1/k−N(k−1)r;
•
w(v)=(1−(k−1)s)/N for v∈{b1,…,bN}; w(v)=s for v∈{b1′,…,bk−1′}; and
•
w(v)=(1−(k−1)y)/N for v∈{c1,…,cN}; w(v)=y for v∈{c1′,…,ck−1′}.
Then (G,w) is a weighted graph. We claim it is (x,y)-biconstrained via (A,B,C), and
w(NA2(v))<1/k for each v∈C. To see this we must verify:
[TABLE]
The first and third hold with equality from the definitions of p,q, and the second follows since y≤s. The fourth follows from
the definition of r. For the fifth, on substituting for p and simplifying, we need to show that
r(k−1)(N−x(N+1)/(1−(k−1)s))≤1/k−x, and this follows from the definition of r. Finally, the sixth
simplifies to (p+q−1)(N+1)/N<N, and this is true since p+q≤N. Consequently ψ(x,y)<k1, by 2.1.
This proves 3.3.
We also need the next three results:
3.4
Let x′,y′,z′∈(0,1), with ψ(x′,y′)≤z′. If x,y,z∈(0,1] satisfy x≤1/(2−x′), y≤y′/(1+y′),
x+(1−x′)y/y′≤1, z≥1/(2−z′), and
[TABLE]
*then ψ(x,y)≤z.
*
Proof. Since x′≤z′ (because ψ(x′,y′)≤z′) it follows that
[TABLE]
Let G′ be (x,y)-biconstrained via (A,B,C), such that ∣NA2(w)∣≤z′∣A∣ for all w∈C.
Add three vertices a,b,c to the graph, and edges from a to every vertex in B, edges from b to every vertex in A,
and an edge between b and c. Let this new graph be G.
Assign weights as follows:
[TABLE]
We will choose p,q such that the weighted graph (G,w) is (x,y)-biconstrained via (A∪{a},B∪{b},C∪{c})
and w(NA∪{a}2(w))≤z for all w∈C∪{c}. The conditions
are: 1−p≥x, 1−q≥x, q≥y, (1−q)x′+q≥x, (1−p)x′+p≥x, (1−y)y′≥y, (1−q)y′≥y, 1−p≤z,
and (1−p)z′+p≤z. These are equivalent to the following:
[TABLE]
Thus, it suffices to show that the lower bound on p is at most the upper bound on p, and the same for q. We obtain
eight conditions, which simplify to those given in the theorem statement. This proves 3.4.
3.5
Let x′,y′,z′∈(0,1), with ψ(x′,y′)≤z′. If x,y∈(0,1] satisfy y≤1/(2−y′), x≤x′/(1+x′), x≤x′z,
(1−y′)x/x′+y≤1, z≥1/(2−z′), and x≤(z−z′)/(1−z′), then ψ(x,y)≤z.
Proof. Let G′ be (x,y)-biconstrained via (A,B,C), such that ∣NA2(w)∣≤z′∣A∣ for all w∈C. Add three vertices
a,b,c to G′, with an edge from a to b, edges from b to every vertex in C, and edges from c to every vertex in B.
Let this new graph be G. Assign weights as follows:
[TABLE]
The conditions that the weighted graph (G,w) is (x,y)-biconstrained via (A∪{a},B∪{b∣],C∪{c})
with w(NA∪{a}A2(w))≤z for all w∈C∪{c} can be written as follows:
[TABLE]
We need to check that the lower bound for p is at most the upper bound for p, and the same for q. This gives
eight conditions, which simplify (using that 1−y′>x′, since ψ(x′,y′)<1) to those given in the theorem. This proves 3.5.
3.6
Let s,t≥1 be integers with s/t≤1/2. Let x,y∈(0,1], satisfying tx/s+y≤1, x+ty/s≤1, and either sy≤x
or sx≤y. Furthermore, if either x or y is irrational, let strict inequality hold in all of these, that is, tx/s+y<1,
x+ty/s<1, and either sy<x
or sx<y. Then ψ(x,y)<s/t.
Proof. By increasing x or y if necessary, we may assume that x,y are both rational.
Let k+1=st. In terms of k, the hypotheses become k≥1, (k+1)x+y≤1, x+(k+1)y≤1, and either sy≤x
or sx≤y.
Suppose first that sy≤x.
Choose an integer N≥1 such that p=xN/(1−kx) and q=yN/(1−kx) are integers, and thus p+q≤(x+y)N/(x+y)=N.
Let G1 be the graph with vertices {a1,...,aN,a∗,b1,...,bN,c1,...,cN} where (reading subscripts modulo N)
each ai is adjacent to
bi,...,bi+p−1, each bi is adjacent to ci,...,ci+q−1, and a∗ is adjacent to all of the bi.
Let m=t−s. Let G2 be the graph with vertex set {a1′,...,am′,b1′,...,bm′,c1′,...,cm′}, where each
ai′ is adjacent to bi′,...,bi+s−1′ (reading subscripts modulo m), and each bi′ is adjacent to ci′.
Let G be the disjoint union of G1 and G2. Let A={a1,...,aN,a∗,a1′,...,am′}, and
B={b1,...,bN,b1′,...,bm′} and define C similarly.
Let r satisfy (k+1)rN=k+11−x. Assign weights as follows:
[TABLE]
This defines a weighted graph (G,w), (x,y)-biconstrained via (A,B,C), such that w(NA2(v))<s/t=1/(k+1) for all v∈C,
and so
ψ(x,y)<s/t, as desired.
Now suppose sy>x, and consequently sx≤y. Choose an integer N≥1 such that p=xN/(1−ky), q=yN/(1−ky) are integers,
and thus p+q≤N. Let G1 be as before. Let m=b−a, and let G2 be the graph with vertex set
{a1′,...,am′,b1′,...,bm′,c1′,...,cm′}, where each ai′ is adjacent to bi′, and each bi′ is adjacent to
ci′,...,ci+a−1′, reading subscripts modulo m (thus, this is the earlier graph G2 flipped).
Let G be the disjoint union of G1 and G2, and define A,B,C as before. Let r>0 satisfy (k+1)rN≤1/(k+1)−x and
r≤1/(k+1)−y. Assign weights as follows:
[TABLE]
Then (G,w) is a weighted graph, and is (x,y)-biconstrained via (A,B,C), and w(NA2(v))<s/t for all v∈C, showing
that
ψ(x,y)<s/t. This proves 3.6.
3.7
Let x,y,z∈(0,1], with y≤1/2<x. If ϕ(2−1/x,y/(1−y))≤2−1/z, then ϕ(x,y)≤z.
Proof. Let x′=(2x−1)/x, and y′=y/(1−y), and let z′=ϕ(2−1/x,y/(1−y)). Let G′ be a graph that is
(x,y)-constrained via (A,B,C), such that
∣NA2(w)∣≤z′∣A∣ for all w∈C. Add three vertices a,b,c to the graph, and edges from a to every vertex
in B, edges from b to every vertex in A, and an edge between b and c. Let this new graph be G.
Assign weights w(v)(v∈V(G)) as follows:
[TABLE]
Then the weighted graph (G,w) is (x,y)-constrained via (A∪{a},B∪{b},C∪{c}), since xx′+(1−x)=x and (1−y)y′=y. Moreover,
w(NA2(w))≤(1−z)+zz′≤z
for all w∈C; and w(NA∪{a}2(c))=z.
Thus ϕ(x,y)≤z. This proves 3.7.
4 Biconstrained graphs
In this section we prove some lower bounds on ψ(x,y). On the diagonal x=y, ψ(x,y) behaves
perfectly; it turns out that for all x, ψ(x,x)=1/k, where k is the largest integer with 1/k≥x.
That follows from:
4.1
For all integers k≥1,
if x,y∈(0,1] with x+ky>1 and kx+1−(k−1)yx≥1, then ψ(x,y)≥1/k.
Proof. By 1.2 we may assume that x,y<1/k.
Let G be (x,y)-biconstrained, via (A,B,C).
We must show that ∣NA2(v)∣≥∣A∣/k for some v∈C. Suppose not.
Choose K⊆C with ∣K∣≤k, and subject to that with ∣K∣ maximum such that the sets N(v)(v∈K) are
pairwise disjoint. Let
I⊆A be the union of the sets NA2(v)(v∈K), and let J⊆B be the union of the sets N(v)(v∈K).
It follows that
(1) ∣A∖I∣>(1−∣K∣/k)∣A∣, and ∣B∖J∣≤(1−∣K∣y)∣B∣.
If ∣K∣=k, then by (1), ∣B∖J∣≤(1−ky)∣B∣<x∣B∣, and since every vertex in A has x∣B∣ neighbours in B,
it follows that every vertex in A has a neighbour in J, that is, I=A, contrary to (1). Thus ∣K∣<k.
Since each vertex in A∖I has at least
x∣B∣ neighbours in B, and they all belong to B∖J, some vertex t∈B∖J has at least
[TABLE]
neighbours
in A∖I by (1). Since ∣K∣≤k−1 and ky<1, and therefore ∣K∣y<1, it follows that
[TABLE]
and so t has at least k(1−(k−1)y)x∣A∣ neighbours in A∖I.
Let u∈C be adjacent to t.
From the maximality of K,
u has a neighbour w∈N(v) for some v∈K. Since w has at least x∣A∣ neighbours
in I, it follows that
For all integers k≥1,
if x,y∈(0,1] with x+ky>1 and kx+y≥1, then ψ(x,y)≥1/k.
Proof. If k=1 the result is easy (and follows from 5.2 below), so we assume that k≥2; and hence we may assume that
x,y<1/k≤1/2 by 1.2. By 4.1 we may assume (for a contradiction) that kx+1−(k−1)yx<1.
Consequently kx+1−(k−1)(1−kx)x<1. Let t=1−kx. Then 1−(k−1)t1−t<kt, and so
k(k−1)t2−(k+1)t+1<0. This is quadratic in t, with discriminant (k+1)2−4k(k−1), and the latter is negative if k>2;
so we may assume that k=2. Then 2t2−3t+1<0, so (2t−1)(t−1)<0, that is, 1/2<t<1. But t=1−2x, so 1/2<1−2x<1, that is,
x<1/4. But 2x+y≥1 and y<1/2, a contradiction. This proves 4.2.
Consequently we have:
4.3
For all x≥0, ψ(x,x)=1/k, where k is the largest integer with 1/k≥x.
Let k≥1 be an integer, let (k−1)/k2≤y≤1, and let (A,B,C) be a tripartition of a graph G, such that:
•
every vertex in B has at least y∣C∣ neighbours in C; and
•
∣NA2(v)∣<∣A∣/k* for each v∈C.*
Then there exist v1,…,vk∈A such that N(vi)∩N(vj)=∅ for 1≤i<j≤k.
Proof. If some vertex v in A has degree zero, then we may take v1=⋯=vk=v. So we assume that
every vertex in A has a neighbour in B.
For each v∈A, let c(v)=∣NC2(v)∣, and let A(v)⊆A be the set of vertices in A that have a neighbour in
N(v). Let ∣A(v)∣=a(v).
(1) For each v∈A, c(v)>kya(v)∣C∣/∣A∣.
If we choose u∈NC2(v) independently at random, then since every vertex in A(v)
has at least y∣C∣ second neighbours in NC2(v), the probability that a given vertex w∈A(v) belongs to NA2(u)
is at least y∣C∣/c(v), and so the expectation of ∣NA2(u)∣ is at least (y∣C∣/c(v))a(v). On the other hand, the
expectation of ∣NA2(u)∣ is less than ∣A∣/k.
This proves (1).
Let H be the graph with vertex set A, in which distinct u,v are adjacent if (in G) u,v have a common neighbour in B.
Thus every vertex v has degree a(v)−1 in H. So
2∣E(H)∣=∑v∈A(a(v)−1);
but
[TABLE]
Consequently
[TABLE]
By Turán’s theorem, H has a stable set of cardinality k. This proves 4.4.
4.5
Let k≥1 be an integer, and let x,y∈(0,1] where y≥(k−1)/k2 and kx+y>1.
Let G be (x,y)-constrained via (A,B,C), such that
every vertex in C has at least y∣B∣ neighbours in B.
Then ∣NA2(v)∣≥∣A∣/k for some v∈C. Consequently ψ(x,y)≥1/k.
Proof. Suppose not; then there is a weighted graph (G′,w), (x,y)-constrained via some tripartition (A′,B′,C′),
such that
•
for each v∈C′, w(N(v))≥y∣B′∣; and
•
for each v∈C′, w(NA′2(v))<1/k.
Choose such a weighted graph (G′,w) with ∣V(G′)∣ minimum, and let z<1/k such that
w(NA′2(v))≤z for each v∈C′. By 4.4, there exist v1,…,vk∈A′
such that N(v1),…,N(vk) are pairwise disjoint. Consequently
w(N(v1)∪⋯∪N(vk))≥kx;
and since w(N(u))≥y>1−kx for each u∈C′, it follows that
⋃v∈XNC′2(v)=C′ where X={v1,…,vk}. But ∣X∣<z−1, contrary to 2.4 and the minimality of ∣V(G′)∣.
This proves the first claim, and the second follows. This proves 4.5.
It is awkward to express the biconstrained problem in the language of triangular triples, but we can do so as follows.
For x,y,z∈(0,1] we say that (x∗,y,z) is triangular if no triangle-free graph G
admits a tripartition (A,B,C) that satisfies the three bullets of the previous definition, and in addition
satisfies
•
every vertex in B has at least x∣A∣ neighbours in A.
Similarly, we say (x∗,y∗,z) is triangular if no triangle-free graph G
admits a tripartition (A,B,C) that satisfies the three bullets of the previous definition, and in addition
satisfies
•
every vertex in B has at least x∣A∣ neighbours in A; and
•
every vertex in C has at least y∣B∣ neighbours in B;
and so on.
Then we have:
4.6
For x,y,z∈(0,1], ψ(x,y)>1−z if and only if (x∗,y∗,z) is triangular.
This is not the same as saying that (x∗,y∗,z∗) is triangular, so we need to keep track of the asterisks
if we rotate; but still it can be useful, as we shall see.
5 The mono-constrained case
In this section we are mostly concerned with ϕ(x,y) when x=y.
We know that ψ behaves well on the diagonal x=y, because of 4.3, so what about ϕ?
More generally, what about an analogue of 4.1 or 4.2
with ψ replaced by ϕ?
If we replace ψ by ϕ in 4.1, it
becomes false, even with k=2, because ϕ(3/10,4/11)≤4/9,
as the graph of figure 2 shows (the sets A,B,C are the rows,
and the numbers on the vertices are used as in figure 1).
But as far as we know, 4.2 might hold with ψ replaced by ϕ.
Let us state this as a conjecture:
5.1
*Conjecture: ***
For all integers k≥1,
if x,y∈(0,1] with x+ky>1 and kx+y≥1, then ϕ(x,y)≥1/k.
On the other hand, we have not even been able to prove what is presumably the simplest nontrivial case of this, namely that ϕ(x,y)≥1/2
for all x,y with x,y>1/3. But we do have several results approaching 5.1.
First, it is true with k=1; we have the trivial:
5.2
For x,y∈(0,1], if x+y>1, or x+y=1 and x is irrational, then ϕ(x,y)=1.
Proof. Let G be (x,y)-constrained via (A,B,C). Then some vertex v∈C has at least y∣B∣ neighbours in B, and strictly
more if y is irrational; and so NA2(v)=A, as every
vertex in A has at least x∣B∣ neighbours in B. This proves 5.2.
This is tight, in that if x+y=1 and x,y are rational, then ϕ(x,y)<1. We omit the proof, which is easy.
5.1 implies that ϕ(x,x)≥1/2 if x>1/3. We have not been able to prove this, but we can show that
ϕ(x,x)>3/7 if x>1/3. That is implied by the following:
5.3
Let k≥2 be an integer; then for x,y∈(0,1], if y>1/k then
[TABLE]
Indeed if k=2, then ϕ(x,y)≥2x−x2 (which is larger).
Proof. Let G be (x,y)-constrained via (A,B,C). If x is irrational
then G is (x,y)-constrained via (A,B,C),
for some rational x′>x; so we may assume that x is rational, by increasing x if necessary.
Suppose that k=2, and choose v1,v2∈B independently and uniformly at random. For each u∈A,
the probability that u is adjacent to at least one of v1,v2 is at least 2x−x2, since u has at least x∣B∣
neighbours in B; and so we may choose v1,v2 such that at least (2x−x2)∣A∣ vertices in A are adjacent to at least
one of them. But v1,v2 have a common neighbour in C, since y>1/2, and the claim follows.
Thus we may assume that k≥3. By 1.2, ϕ(x,y)≥x, and so we may assume that
[TABLE]
that is, x<1/(k−1). Consequently x≤(k−2)/(k−1) since k≥3.
Define
[TABLE]
These are all non-negative, and
p is rational with denominator T say; and by replacing each vertex by T copies,
we may assume that p∣A∣
is an integer. Since x≤(k−2)/(k−1) it follows that s≤1.
For 1≤i≤k−1, we define vi∈B, and a subset Pi of NA(vi) with ∣Pi∣=p∣A∣, inductively, as follows.
Let Q=P1∪⋯∪Pi−1.
(1) There exists vi∈B such that sa+b≥x(s∣Q∣+∣A∣−∣Q∣), where a=∣NA(vi)∩Q∣ and
b=∣NA(vi)∖Q∣.
Suppose not; then summing over all v∈B, we deduce that
[TABLE]
But the first sum is s times the number of edges between Q and B, and so at least xs∣B∣⋅∣Q∣; and the second
is similarly at least x∣B∣(∣A∣−∣Q∣), a contradiction. This proves (1).
Let vi be as in (1). Thus sa+b≥x(s∣Q∣+∣A∣−∣Q∣)≥x(1−(1−s)(k−2)p)∣A∣. In particular, since
[TABLE]
there exists Pi⊆NA(vi) of cardinality p∣A∣.
Also, since a≤(k−2)p∣A∣, and so
[TABLE]
it follows that
[TABLE]
and so
[TABLE]
for 1≤h<i. This completes the inductive definition of v1,…,vk−1 and P1,…,Pk−1.
Let P=P1∪⋯∪Pk−1. Then ∣P∣≤(k−1)p∣A∣.
Since every vertex in A∖P has at least x∣B∣ neighbours in B, there exists vk∈B
with at least x(∣A∣−∣P∣)≥x(1−(k−1)p)∣A∣ neighbours in A∖P. Let Pk be its set of neighbours in
A∖P. Then for all i with 1≤i≤k−1,
[TABLE]
Consequently, for all distinct
v,v′∈{v1,…,vk}, ∣NA(v)∪NA(v′)∣≥m∣A∣. But since y>1/k, some two of v1,…,vk
have a common neighbour u∈C, and so ∣NA2(u)∣≥m. This proves 5.3.
We deduce from 5.3 a version of 4.3 for the mono-constrained case:
5.4
For y∈(0,1], if y>1/k where k≥2 is an integer, then
ϕ(1/k,y)≥2k2−4k+12k−3.
Consequently ϕ(1/k,y)≥1/k+1/(2k2)+Ω(k−3).
5.4 tells us in particular that ϕ(x,x)≥2k2−4k+12k−3>1/k when x>1/k (if k≥2 is an integer), and since
ϕ(1/k,1/k)=1/k, there is a discontinuity in ϕ(x,x) when x=1/k,
and the limit of ϕ(x,x) as x→1/k from above is different from ϕ(1/k,1/k).
What happens when x→1/k from below? The next results investigate this. We will show that
if x is sufficiently close to 1/k from below, then ϕ(x,x)=1/k.
5.5
If k>0 is an integer and x∈(0,1] satisfies (1−x)k<x, then ϕ(x,x)≥1/k. In particular,
if x>0.382 then ϕ(x,x)≥1/2, and if x>0.318 then ϕ(x,x)>1/3.
Proof. Let G be (x,x)-constrained via (A,B,C). If we choosing k vertices from C uniformly at random,
the number of vertices in B nonadjacent to all of them is at most (1−x)k∣B∣ in expectation; and so there
exist v1,…,vk∈C such that at most (1−x)k∣B∣ vertices in B are nonadjacent to all of them.
Since (1−x)k∣B∣<x∣B∣, it follows that the sets NA2(vi)(1≤i≤k) have union A, and so one of them
has cardinality at least ∣A∣/k. This proves 5.5.
The proof of 5.5 is very simple, but the result is not of any value.
It is of no use when k≥4 because then (1−x)k<x implies x>1/k; and
we will prove in 6.6 and 8.3 that ϕ(x,x)≥1/2 when x>0.352202, and ϕ(x,x)≥1/3
when x≥0.28231, which are stronger than
5.5 when k=2,3.
Here is another approach to the same question, more successful for larger values of k.
5.6
*Let k≥1 be an integer, and let x≥1/k−ε where ε=1/(13k3). Then ϕ(x,x)≥1/k.
*
Proof. We may assume that x=1/k−ε. By 5.2 we may assume that k≥2. We leave the reader to check that
•
1/(2k)−ε>6k2ε;
•
x>1/(k+1); and
•
(2kx−1)/(2k−1)>(kε)/(x+kε).
(These are inequalities we will need later.)
Let G be (x,x)-constrained via (A,B,C), and suppose that ∣NA2(v)∣<∣A∣/k for each v∈C.
Let
P be the set of vertices in B that have at most (1/k−2kε)∣A∣
neighbours in A.
(1) ∣P∣≤∣B∣/(2k).
Every vertex in B has fewer than ∣A∣/k neighbours in A, and so the number of edges between A and B
is at most ∣P∣(1/k−2kε)∣A∣+(∣B∣−∣P∣)∣A∣/k. On the other hand, the number of such edges is at least (1/k−ε)∣A∣⋅∣B∣; and so
[TABLE]
which simplifies to
2k∣P∣≤∣B∣. This proves (1).
(2) If u,v∈B∖P have a common neighbour in C, then ∣NA(u)∖NA(v)∣≤2kε∣A∣.
Since u,v∈B∖P have a common neighbour in C, it follows that ∣NA(u)∪NA(v)∣≤∣A∣/k. But
∣NA(u)∣≥(1/k−2kε)∣A∣ since u∈B∖P, and so ∣NA(u)∖NA(v)∣≤2kε∣A∣. This proves (2).
(3) There exist v1,…,vk∈B∖P such that for 1≤i<j≤k, there
are at least (1/(2k)−ε)∣A∣/k vertices in A that are adjacent to vj and not to vi.
Choose v1,…,vk∈B∖P as follows. Choose v1∈B∖P arbitrarily.
Inductively, suppose we have defined v1,…,vi where i<k. Each has at most ∣A∣/k
neighbours in A, and so the set of vertices in A adjacent to
one of v1,…,vi has cardinality at most (i/k)∣A∣≤(1−1/k)∣A∣. Let D be the set of vertices in A nonadjacent to each of v1,…,vi;
then ∣D∣≥∣A∣/k. Since, by (1), each vertex in D has at least x∣B∣−∣P∣≥(1/(2k)−ε)∣B∣
neighbours in B∖P, there exists vi+1∈B∖P with at least
(1/(2k)−ε)∣A∣/k neighbours in D. This completes the inductive definition. We see that for 1≤i<j≤k, there
are at least (1/(2k)−ε)∣A∣/k vertices in A that are adjacent to vj and not to vi. This proves (3).
Let H be the bipartite graph G[(B∖P)∪C].
(4) For 1≤i<j≤k, vi and vj belong to distinct components of H.
From (2), the sets NC(v1),…,NC(vk) are pairwise disjoint, because (1/(2k)−ε)∣A∣/k>2kε∣A∣.
Suppose that there is a path of H joining some two of v1,…,vk, and take the shortest such path Q;
between vi and vj say, where j>i. Let Q have
m vertices in B, say u1,…,um in order where u1=vi. We claim that m≤4. For suppose that m≥5.
From the minimality of the length of Q, u3 has no common neighbour in C
with any of v1,…,vk, and so the sets NC(u3),NC(v1),…,NC(vk) are pairwise disjoint, which is impossible
since x>1/(k+1). Thus m≤4.
By applying (2) to each pair of consecutive members of V(Q)∩B, we deduce that
[TABLE]
But ∣NA(vj)∖NA(vi)∣≥(1/(2k)−ε)∣A∣/k, and so
(1/(2k)−ε)∣A∣/k≤6kε∣A∣,
a contradiction. This proves (4).
For 1≤i≤k, let Hi be the component of H containing vi, and let V(Hi)∩B=Bi
and V(Hi)∩C=Ci. If there exists v∈B∖P that does not belong to any of B1,…,Bk, then
the sets NC(v),NC(v1),…,NC(vk) are pairwise disjoint, which is impossible since they all have
cardinality at least x∣C∣, and (k+1)x>1. Consequently the sets B1,…,Bk and P form a partition of B.
(5) For 1≤i≤k there exists ui∈Ci adjacent to at least 1+k(k−1)ε(1−kε)∣Bi∣
vertices in Bi.
For 1≤i≤k, since vi has at least x∣C∣ neighbours in C, it follows that ∣Ci∣≥x∣C∣.
Let 1≤i≤k. Since C1,…,Ck are pairwise disjoint, and the union of the sets Cj(j∈{1,…,k}∖{i})
has cardinality at least (k−1)x∣C∣, it follows that
[TABLE]
There are at least x∣Bi∣⋅∣C∣ edges between Bi and Ci, and so some vertex in Ci
has at least
[TABLE]
neighbours in Bi. By substituting x=1/k−ε, this proves (5).
For 1≤i≤k, let Ai=NA2(ui). Since ∣Ai∣<∣A∣/k for 1≤i≤k, there exists v∈A
that is in none of A1,…,Ak. Now v has at least x∣B∣ neighbours in B, and they all
belong to B1∪⋯∪Bk except for at most ∣P∣ of them. Consequently there exists i∈{1,…,k}
such that v has at least (x∣B∣−∣P∣)∣Bi∣/∣B∖P∣ neighbours in Bi. Since v∈/Ai,
it follows that
[TABLE]
Since x∣B∣≤∣B∣ and ∣P∣≤∣B∣/(2k) by (1), it follows that
[TABLE]
and so
(2kx−1)/(2k−1)≤kε/(x+kε),
a contradiction.
This proves 5.6.
For comparison, in figure 3 we give graphs of the function ψ(x,x) (which we know completely, because of 4.3),
and the function ϕ(x,x) (which we only know partially, from 5.6, 5.4 and 1.4.)
The next result is a useful general lower bound on ϕ(x,y).
5.7
For x,y,z∈(0,1], if y>1/2 and 4x2y(1−z)≥(z−x)2 then ϕ(x,y)≥z. If in addition 4x2y(1−z)>(z−x)2 then ϕ(x,y)>z.
Proof. Let G be (x,y)-constrained via (A,B,C), and suppose that ∣NA2(w)∣<z∣A∣ for each w∈C.
There are at least xy∣A∣⋅∣B∣⋅∣C∣ two-edge paths between A and C, and so there is a vertex
w∈C that is an end of at least xy∣A∣⋅∣B∣ such paths. Let w be an end of exactly xq∣A∣⋅∣B∣ such paths; thus y≤q.
Let B1=NB(w), and let t=∣B1∣/∣B∣. Since ∣NA2(w)∣≤z∣A∣,
there
exists A1⊆A including NA2(w) with ∣A1∣=z∣A∣ (we may assume the latter is an integer.) For each u∈A1
let u have exactly d(u)∣B∣ neighbours in B1, and therefore at least (x−d(u))∣B∣ neighbours in B∖B1.
It follows that
[TABLE]
Let v1∈B1 and v2∈B∖B1, and let A(v1,v2)=NA(v1)∪NA(v2). For every such choice of v1,v2, since
y>1/2, there is a vertex w in C adjacent to both v1,v2, and since ∣NA2(w)∣<z∣A∣, it follows that
∣A(v1,v2)∣<z∣A∣. Let us choose v1∈B1 and v2∈B∖B1, uniformly at random.
It follows that the expected value of ∣A(v1,v2)∣ is less than z∣A∣. The expected value of ∣A(v1,v2)∩A1∣
is at least
[TABLE]
and the expected value of ∣A(v1,v2)∖A1∣ is at least
[TABLE]
Consequently the sum of these two is less than z∣A∣, and so
[TABLE]
Since ∑u∈A1d(v)=xq∣A∣, this simplifies to
[TABLE]
Now since ∑u∈A1d(v)=xq∣A∣ and ∣A1∣=z∣A∣, it follows from the Cauchy-Schwarz inequality that ∑u∈A1d(u)2≥x2q2∣A∣/z.
Consequently
[TABLE]
This can be rewritten as:
[TABLE]
Since the first term above is a square, it is nonnegative, and so, since q≥y, it follows that
[TABLE]
contrary to the hypothesis.
This proves the first statement of the theorem, and the second is immediate by slightly increasing z. This proves 5.7.
6 When is ϕ(x,y) or ψ(x,y)≥1/2?
Another way to approach the problem is to ask, given some value z, for which x,y∈(0,1] is ϕ(x,y)≥z? Or we could ask the
same question for ψ, or ask when ϕ(x,y)>z. For instance:
6.1
If k≥1 is an integer, then for x,y∈(0,1], ϕ(x,y)>1/k if and only if max(x,y)>1/k.
This follows trivially from 1.3 and 1.2. And the same holds with ϕ replaced by ψ.
But deciding when ψ(x,y)≥1/k or ϕ(x,y)≥1/k seems to be much less obvious. In this section
we discuss when ψ(x,y) or ϕ(x,y) is at least 1/2; and in later sections we look at when they are at least 2/3,
and at least 1/3.
For x,y∈(0,1], we say (temporarily) that
(x,y) is good if ψ(x,y)≥1/2, and bad otherwise.
The “map” of good and bad points is shown in the left half of figure 4. The solid black curve borders
the known bad points, and the dotted curve borders the good points; between them is undecided. The borders are complicated, and we
have indicated in the figure which theorem is responsible for each stretch of border.
Let us explain some of the details. First, if max(x,y)≥1/2, then
(x,y) is good; and all pairs (x,y) with
x+2y,2x+y≤1 are bad, by 3.3.
We searched by computer to find other examples of bad pairs (x,y),
and found about 12 maximal such pairs of rationals, with numerator and denominator at most 100.
In fact we only searched for pairs (x,y) where the corresponding (x,y)-biconstrained graph
is similar to the graph obtained from figure 1, that is, it is obtained by “blowing up”
the vertices of another graph in which the graph between two of the three parts is a matching. All these examples
not only show that ψ(x,y)<1/2, but also that ψ(y,x)<1/2, and ξ(x,y)<1/2. In particular,
for every bad pair (x,y) we found by computer search, (y,x) is another.
This is just an artifact of our method of search, and is not evidence
that the set of all bad pairs is closed under switching x and y (though it might be; it is for ϕ, by 2.3).
Anyway, for each bad pair the computer found, all pairs it dominates are also bad, and that gave us a step function
bordering the area of the known bad points. We improved on this; we were able to smooth out some of the
steps of the step function, by means of
3.3 and 6.5, so the step function the computer found now only survives towards the ends of the solid
black curve in the figure. (These “fills” are not invariant under switching x and y.)
We give the coordinates of some bad pairs that we find particularly interesting.
The apparent asymmetry between x and y in the left half of the figure is just asymmetry among what we have been able to prove;
we have no proof of asymmetry. The right half of figure 4 does the same for ϕ. Here
there is symmetry exchanging x and y, by 2.3, and so
we only “explain” half of the border.
A graph has radius at most r if there is a vertex u such that every vertex has distance at most r from u.
We will need the following theorem of Erdős, Saks and Sós [1]:
6.2
Let G be a connected graph with radius at least r, where r≥1 is an integer. Then G has an induced path with 2r−1 vertices,
and consequently has a stable set of cardinality at least r.
When we have more than one graph defined using the same vertices, we speak of “H-distance” to mean distance in the graph H,
and so on.
6.3
Let x,y∈(0,1], such that
[TABLE]
Then ψ(x,y)≥1/2.
Proof. Let G be (x,y)-biconstrained via (A,B,C), and suppose that ∣NA2(v)∣<∣A∣/2 for each v∈C.
Then 1.2 implies that x,y<1/2.
Suppose that y>1/4. The given inequality implies that 56x2−64x+17<0, and so x>.41.
Since 2x+y>1, 4.5 implies that y≤1/4, a contradiction. Thus y≤1/4.
We leave the reader to verify that the following are consequences of the given inequality:
•
3−3yx>1−2x; and in particular, 1−yx>1−2x, so from 4.1 it follows that x+2y≤1;
•
3−6yy>1−2x; and so 2−2yy>1−2x, since y≤1/4; and
•
x+3y>1.
(We found the easiest way to check these is to have a computer plot the various curves.)
Let H be the bipartite graph G[B∪C].
(1) If v,v′∈C have H-distance at most 2t where t>0 is an integer, then
[TABLE]
Take a path P of H joining v and v′, of length at most 2t.
Let the vertices of P in C
be
[TABLE]
in order. For 1≤i≤t let ui∈B be adjacent to vi−1 and vi.
Then for 1≤i≤t, NA2(vi−1)∩NA2(vi) includes NA(ui) and hence has cardinality at least x∣A∣;
and since
∣NA2(vi)∣<∣A∣/2, it follows that ∣NA2(vi)∖NA2(vi−1)∣<(1/2−x)∣A∣. But the union of the t
sets NA2(vi)∖NA2(vi−1) includes NA2(v′)∖NA2(v), and so the latter has cardinality less than
t(1/2−x)∣A∣. This proves (1).
(2) There do not exist v1,…,v4∈C, pairwise with no common neighbour in B.
Then every three of N(v1),…,N(v4) have union of cardinality at least 3y; and since 3y>1−x,
every vertex in A has a neighbour in at least
two of N(v1),…,N(v4). Consequently every vertex in A belongs to at least two of NA2(v1),…,NA2(v4), and so
one of NA2(v1),…,NA2(v4) has cardinality at least ∣A∣/2, a contradiction. This proves (2).
(3) H* has at least two components.*
Suppose not, and let H′ be the graph with vertex set C in which v,v′ are adjacent if they have a common neighbour
in H. By (2), it follows that H′ has no stable set of cardinality four, and so has radius at most three by 6.2.
Choose v∈C such that every vertex in C has H′-distance at most three from v. Let B1=NB(v) and A1=NA2(v).
Every vertex in A∖A1 has at least x∣B∣ neighbours in B∖B1, and so some vertex u∈B∖B1
has at least x(∣B∣/∣B∖B1∣)∣A∖A1∣ neighbours in A∖A1. Let A2 be the set of neighbours of u
in A∖A1. Since ∣B∖B1∣≤(1−y)∣B∣ and ∣A∖A1∣>∣A∣/2, and 3−3yx>1−2x,
it follows that
[TABLE]
Let v′∈C be adjacent to u.
Since the H′-distance from v to v′ is at most three, the H-distance from v to v′
is at most six. By (1), ∣NA2(v′)∖NA2(v)∣<3(1/2−x)∣A∣, a contradiction.
This proves (3).
(4) If H′ is a component of H then ∣V(H′)∩B∣≤(1−x)∣B∣.
Suppose that ∣V(H′)∩B∣>(1−x)∣B∣; then every vertex in A has a neighbour in V(H′). By (2), there do not exist three vertices
in C∩V(H′) pairwise with no common neighbour, and so by 6.2, it follows that there is a vertex v∈C∩V(H′)
with H′-distance at most four from every vertex in C∩V(H′). Let
A′=NA2(v); then ∣A′∣<∣A∣/2. Since every vertex in A∖A′ has at least y∣C∣ second neighbours in C∩V(H′),
and ∣C∩V(H′)∣≤(1−y)∣C∣, some vertex v′∈C∩V(H′) has at least (y/(1−y))∣A∖A′∣ second neighbours
in A∖A′. By (1),
(y/(1−y))∣A∖A′∣<2(1/2−x)∣A∣. But ∣A′∣<∣A∣/2, so y/(4(1−y))≤1/2−x, a contradiction.
This
proves (4).
(5) Some component H′ of H satisfies (1−x)∣B∣≥∣V(H′)∩B∣≥x∣B∣.
By (2) and (3), H has either two or three components. If H has only two components, then they both satisfy (5), by (4); so we assume there are three.
Let the components of H be H1,H2,H3, and for 1≤i≤3, let
V(Hi)∩B=Bi and V(Hi)∩C=Ci;
and let ∣Bi∣/∣B∣=bi and ∣Ci∣/∣C∣=ci. Suppose that b1,b2,b3<x.
Consequently every vertex in A
has neighbours in at least two of B1,B2,B3. For 1≤i≤3, let Ai be the set of vertices in A
with a neighbour in Bi. Thus every vertex in A belongs to at least two of A1,A2,A3, so from the symmetry
we may assume that ∣A1∣≥2∣A∣/3.
By (2), every two vertices in C1
have a common neighbour in B. Choose v∈C1, and let
A′=NA2(v); then ∣A′∣≤∣A∣/2. Since every vertex in A1 has at least y∣C∣ second neighbours in C1,
some vertex v′ in C1 has at least (y/c1)∣A1∖A′∣ second neighbours
in A1∖A′. By (1),
(y/c1)∣A1∖A′∣<(1/2−x)∣A∣. But ∣A′∣<∣A∣/2, so ∣A1∖A′∣≥∣A∣/6; and c1≤1−2y,
so y/(6(1−2y))≤1/2−x, a contradiction. This proves (5).
(6) Every vertex in C has at most (1−x−y)∣B∣ neighbours in B. Consequently
[TABLE]
for each component H′ of H.
Suppose that v∈C has more than (1−x−y)∣B∣ neighbours in B. Choose v′∈C in a different component of H; so
v,v′ have no common neighbour in B. Consequently
[TABLE]
and so every vertex in A
has a neighbour in N(v)∪N(v′). But then one of ∣NA2(v)∣,∣NA2(v′)∣≥∣A∣/2, a contradiction. This proves
the first assertion. Let H′ be a component of H. Then H′ has at least y∣C∣⋅∣V(H′)∩B∣ edges, and at most
(1−x−y)∣B∣⋅∣V(H′)∩C∣ edges, so the second claim follows. This proves (6).
Let H′ be as in (5), and take the union of the other (one or two) components of H′. We obtain
nonnull subgraphs H1,H2 of H, pairwise vertex-disjoint and with union H, such that
∣V(Hi)∩B∣≥x∣B∣ for i=1,2.
For i=1,2, let
V(Hi)∩B=Bi and V(Hi)∩C=Ci;
and let ∣Bi∣/∣B∣=bi and ∣Ci∣/∣C∣=ci. Thus b1,b2≥x. From (6), bi≤(1−x−y)ci/y for i=1,2; and
c1,c2≥y, since every vertex
in Bi has at least y∣C∣ neighbours in Ci. Also b1+b2=c1+c2=1.
For i=1,2 let Ai be the set of vertices v∈A that have more than (bi−y)∣B∣ neighbours in Bi.
Let A0=A∖(A1∪A2). Hence if u∈A0, then since u has at least x∣B∣ neighbours in B,
u has at least (x+y−b2)∣B∣ neighbours in B1, and at least (x+y−b1)∣B∣ neighbours in B2.
Since A1,A2 and A0 have union A, we may assume that ∣A1∣+∣A0∣/2≥∣A∣/2.
Now A1⊆NA2(v) for each v∈C1, since if u∈A1, then u
has more than (b1−y)∣B∣ neighbours in B1, and v has at least y∣B∣ neighbours in B. Consequently
∣NA2(v)∩A0∣<∣A0∣/2 for each v∈C1.
Let us choose v∈C1 uniformly at random; then the expected
number of second neighbours of v in A0 is less than ∣A0∣/2, and so for some vertex u∈A0, the probability
that u∈NA2(v) is less than 1/2. Let D be the set of neighbours of u in B1.
Then ∣D∣≥(x+y−b2)∣B∣,
and the probability that v has a neighbour in D is less than 1/2. Thus more than ∣C1∣/2 vertices in C1
have no neighbour in D. On the other hand, the expectation of the number of neighbours of v in D is at least
∣D∣y/c1; and so there exists v∈C1 with more than 2∣D∣y/c1 neighbours in D. Also there exists v′∈C1
with no neighbours in D. It follows that
[TABLE]
Some vertex in A0 is not a second neighbour of either of v,v′, and so
[TABLE]
Consequently
y+2(x+y+b1−1)y/c1≤1−x−y. Now c1≤(1−x−2y+yb1)/(1−x−y) since
[TABLE]
. So
[TABLE]
that is,
[TABLE]
But b1≥x, contrary to the hypothesis. This proves 6.3.
6.4
Let x,y∈(0,1], such that x+2y>1, x≥1/4 and y≥1/3. Then ψ(x,y)≥1/2.
Proof. The only lower bound constraints on y are y≥1−2x and y≥1/3, and these are both satisfied if y=0.38 since x≥1/4. Hence we may
assume that y≤0.38, by replacing y by min(y,0.38). Consequently y2−3y+1>0, and so
[TABLE]
Let G be (x,y)-biconstrained via (A,B,C), and suppose that ∣NA2(w)∣<∣A∣/2 for each w∈C.
Choose w1,w2,w3∈C uniformly at random. The expected number of vertices in B nonadjacent to all of w1,w2,w3
is at most (1−y)3∣B∣<x∣B∣; so we may choose w1,w2,w3 such that
fewer than x∣B∣ vertices in B are nonadjacent to all of w1,w2,w3. For i=1,2,3 let Ai=NA2(wi). Thus A1∪A2∪A3=A.
In particular, one of A2,A3, say A2, includes at least half of A∖A1; and since ∣A1∣<∣A∣/2, it follows that
∣A2∖A1∣>∣A∣/4. Since ∣A2∣<∣A∣/2, it follows that ∣A1∩A2∣<∣A∣/4<x∣A∣; and so NB(w1),NB(w2) are disjoint (because
any common neighbour would have at least x∣A∣ neighbours in A, all belonging to A1∩A2). Hence
[TABLE]
and so A1∪A2=A, contradicting that ∣A1∣,∣A2∣<∣A∣/2. This proves 6.4.
6.5
Let x,y∈(0,1], such that x≤13/27 and y≤1/7 and 3x+5y≤2. Then ψ(x,y)≤13/27. If in addition y≤1/8,
then ψ(y,x)<1/2.
Proof. We claim that, for both statements of the theorem, we may assume that 3x+5y=2.
By increasing x, we may assume that either x=13/27 or 3x+5y=2;
and if x=13/27 then y≤1/9, since 3x+5y≤2,
and by increasing y we may assume that 3x+5y=2. This proves our claim.
Since 3x+5y=2 and x≤13/27, it follows that y≥1/9; and since y≤1/7 it follows that x≥3/7.
We return to the graph of figure 1. Let A,B,C be the three rows of vertices, in order where A is
the top row.
We need to adjust the vertex weights. Define p=1/2−x/2−y and r=(x−y)/2.
With the vertices in the same order as the figure,
take vertex weights as follows:
[TABLE]
One can check (it takes some time and we omit the details) that this defines an (x,y)-biconstrained weighted graph
showing that ψ(x,y)≤13/27. For the second statement, take the same graph and same vertex weighting, except
replace the third row (of all one-sevenths) in the table above, by
[TABLE]
where p′=1/2−3y, and r′=3y/2. This weighted graph is (y,x)-biconstrained via (C,B,A), and shows
that ψ(y,x)<1/2.
(Again, we leave the reader to check that this works.)
This proves 6.5.
Now the mono-constrained case: for which pairs (x,y) is ϕ(x,y)≥1/2? Now we have symmetry between x and y,
and we found some examples of pairs (x,y) with ϕ(x,y)<1/2 on a computer searching randomly.
(Conjecture 5.1
says that all points above both the lines x+2y=1 and 2x+y=1 should be good, and indeed, all the maximal examples
the computer found lie
in the wedges between the lines.)
Let x,y∈(0,1], such that 2x2y≥(1−x−y)2. Then ϕ(x,y)≥1/2.
Proof. Suppose that ϕ(x,y)=1/2−ε where ε>0. Then by 2.3 and 2.5 we have
ϕ(x,1/2+ε)≤1−y. Let y′=1/2+ε and z=1−y. Since y′>1/2 and ϕ(x,y′)≤z, the
second statement of 5.7 implies that 4x2y′(1−z)≤(z−x)2, and so 2x2y<(1−x−y)2, a contradiction. This
proves 6.6.
In particular, 6.6 implies that ϕ(x,x)≥1/2 if x≥0.352202, which is stronger than 5.5 when k=2.
6.7
Let x,y∈(0,1] with y<1/2 and y<(1−x)2/(2−4x+6x2); then ϕ(x,y)<1/2.
Proof. By increasing y slightly if necessary, we may assume that s=(1/y−2)1/2 is rational.
Thus s≤1/y−1, s2≤1/y−2 and
1+2/s≤1/x, since y≤1/2 and y≤(1−x)2/(2−4x+6x2).
Choose an integer N≥2 such that sN is an integer.
Choose a graph G1 that is (s,1−s)-constrained via a tripartition (A1,B1,C1),
such that ∣A1∣=∣B1∣=∣C1∣=N
and NAi2(v)=A1 for each v∈C1. (It is easy to see that such a graph exists, for instance, one of the
graphs used in 1.3.) Let G2 be isomorphic to G1, and let (A2,B2,C3) be the corresponding
tripartition. Take the disjoint union of G1 and G2, and add edges to make every vertex in B1
adjacent to every vertex in C2. Add three more vertices a,b,c,
where a is adjacent to b, b is adjacent to every vertex in C1, and c is adjacent to every vertex in B2,
forming G.
We define a weighting w of G as follows. Let p=1/(2N) and q=1/2−1/(4N2).
Define w by:
[TABLE]
Define A=A1∪A2∪{a} and define B,C similarly.
Then the weighted graph (G,w) is (x,y)-constrained via (A,B,C), and proves that ϕ(x,y)<1/2.
(To see the latter, observe that, for instance, if v∈C1 then
When is ϕ(x,y)≥2/3; or the same question for ψ? In this section we
say what we know about these.
7.1
If x>1/2 then ψ(x,1/3)≥2/3.
Proof. Let G be (x,1/3)-biconstrained via (A,B,C), and suppose for a contradiction that
∣NA2(v)∣<2∣A∣/3 for all v∈C. By averaging, there exists v0∈A such that
∣NC2(v0)∣<2∣C∣/3. Let B0=N(v0) and C0=NC2(v0). Hence ∣B0∣≥x∣B∣, and ∣C0∣<2∣C∣/3,
and there are no edges between B0 and C∖C0, and every vertex in C0 has a neighbour in B0.
Choose v1∈C0. Thus N(v1)∩B0=∅. Let B1=N(v1) and A1=NA2(v1).
So ∣A1∣≥x∣A∣, and ∣A1∣<2∣A∣/3. Every vertex v∈A∖A1 has a neighbour in B0,
since ∣B∖B0∣<∣B∣/2<∣N(v)∣. Consequently every vertex in A∖A1 has at least ∣C∣/3≥∣C0∣/2
second neighbours in C0, and by averaging it follows that some vertex v2∈C0
has at least ∣A∖A1∣/2 second neighbours in A∖A1. Let B2=N(v2) and A2=NA2(v2).
Then ∣A2∖A1∣≥∣A∖A1∣/2≥∣A∣/6. If there exists u∈B1∩B2, then since
u has at least x∣A∣ neighbours in A1, and they all belong to A2, it follows that
[TABLE]
a contradiction. Consequently B1∩B2=∅.
In particular, ∣B1∪B2∣≥2∣B∣/3, and so every vertex in A has a neighbour in B1∪B2;
and so A1∪A2=A. Since ∣A1∣,∣A2∣<2∣A∣/3, it follows that ∣A1∩A2∣<∣A∣/3. For i=1,2,
choose bi∈Bi∩B0. Then N(bi)∩A⊆Ai for i=1,2, and so
∣N(b1)∩N(b2)∩A∣<∣A∣/3. Consequently ∣(N(b1)∪N(b2))∩A∣>2∣A∣/3.
Since b1,b2∈B0 and they each have at least ∣C∣/3 neighbours in C0, and ∣C0∣<2∣C∣/3,
it follows that they have a common neighbour v∈C0. But then N(b1)∪N(b2)∩A⊆NA2(v),
and so ∣NA2(v)∣≥2∣A∣/3, a contradiction. This proves 7.1.
7.2
If y>1/2 then ψ(1/3,y)≥2/3.
Proof. Let G be (1/3,y)-biconstrained via (A,B,C).
There exists v1∈C with at least y∣B∣ neighbours in B (in fact, every vertex in C has this property).
Let B1=N(v1) and A1=NA2(v1). Thus ∣B1∣≥y∣B∣. Since every vertex in
B∖B1 has at least y∣C∣ neighbours in C, some vertex v2∈C has at least y∣B∖B1∣
neighbours in B∖B1. Let B2=N(v2) and A2=NA2(v2). Thus ∣B2∖B1∣≥y∣B∖B1∣,
and so
[TABLE]
In particular, since every vertex in A has at least ∣B∣/3 neighbours in B, it follows that A1∪A2=A.
But B1∩B2=∅, since ∣B1∣,∣B2∣≥y∣B∣>∣B∣/2, and so there exists b∈B1∩B2;
and since b has at least ∣A∣/3 neighbours in A, and they all belong to A1∩A2, it follows that
∣A1∩A2∣≥∣A∣/3. Since ∣A1∪A3∣=∣A∣, it follows that ∣A1∣+∣A2∣≥4∣A∣/3, and so
one of ∣A1∣,∣A2∣ is at least 2∣A∣/3. This proves 7.2.
The last two results are closely related, via reformulation into triangular language.
First, we need some shorthand for results of the form “if x′>x then (x′,y,z) is triangular”;
let us say “(x+,y,z) is triangular” to mean “(x′,y,z) is triangular for all x′>x”, and treat the
other two coordinates similarly. We will mix the two systems of notation, in expressions such as
“(x+∗,y+,z) is triangular”, meaning “(x′∗,y+,z) is triangular for all x′>x”.
Thus, in triangular language, we have the following.
•
(1/2+∗,1/3∗,1/3+) is triangular: because 7.1 says that (x∗,1/3∗,1/3+) is triangular when x>1/2.
•
(1/2+,1/3+,1/3∗) is triangular: because
the proof of 7.2 did not use that
every vertex in C has at least x∣B∣ neighbours in B, and so it proves that (1/3∗,1/2+,1/3+) is triangular,
and rotating gives that (1/2+,1/3+,1/3∗) is triangular.
•
(1/2+,1/3+∗,1/3∗) is triangular; this follows from 4.1 with k=2 and rotating.
•
(1/2+,1/3∗,1/3+∗) is triangular; this also follows from 4.1 with k=2 and rotating.
These four statements are similar, but no two are equivalent, and
it would be good to find a common strengthening. Note, however, that (1/2+∗,1/3∗,1/3∗) is not triangular,
and indeed (2/3∗,1/3∗,1/3∗) is not triangular. We have not been able to decide whether
(1/2+,1/3+,1/3) and (1/2+,1/3,1/3+) are triangular, or indeed whether (1/2+∗,1/3+,1/3+) is triangular.
Pursuing this further, what about (1/2+,1/3+,x) when x<1/3 (perhaps with some sprinkling of asterisks)?
How small can x be such that the triple remains triangular? We have examples that show that (5/9,5/14,4/13)
and (4/7,3/8,2/7) are not triangular, and (3/5∗,2/5,1/4) is not triangular,
but as far as we know, (1/2+∗,1/3+∗,1/5) might be triangular.
This extends to weighted graphs in the natural way.
For instance, the weighted graph of figure 6 (identify the vertices on the left with those on the right, in order)
shows that (4/7,2/7,3/8) is not triangular.
The graph of figure 6 shows that ϕ(4/7,2/7)≤5/8, and so
we studied ψ(4/7,2/7), and proved the following.
7.3
If x,y∈(0,1] such that max(x,y)>1/2, x≥1/3, x+2y>1, and 3x+y/(1−y)>2, then
ψ(x,y)≥2/3.
Proof. Let G be (x,y)-biconstrained, via (A,B,C), and suppose for a contradiction that ∣NA2(v)∣<2∣A∣/3
for each v∈C. By 7.2, y≤1/2 since x≥1/3; and so x>1/2 since max(x,y)>1/2.
Hence y<1/3 by 7.1. Also x<2/3, by 1.2.
(1) For all v1,v2∈C, if ∣N(v1)∪N(v2)∣>(1−x)∣B∣ then N(v1)∩N(v2)=∅,
NA2(v1)∪NA2(v2)=A, and ∣NA2(v1)∩NA2(v2)∣<∣A∣/3.
Every vertex in A has a neighbour in N(u)∪N(v), and so NA2(u)∪NA2(v)=A.
Since ∣NA2(u)∣<2∣A∣/3 and ∣NA2(v)∣<2∣A∣/3 it follows that ∣NA2(u)∩NA2(v)∣<∣A∣/3,
and so there is no vertex in N(u)∩N(v) (since any such vertex would have at least x∣A∣ neighbours in A,
all belonging to NA2(u)∩NA2(v)). This proves (1).
(2) There exist v1,v2∈C with N(v1)∩N(v2)=∅.
Choose v1∈C. Since every vertex in A∖NA2(v1) has at least x∣B∣ second neighbours in
B∖N(v1),
some vertex u2∈B∖N(v1) has at least (x/(3(1−y)))∣A∣ neighbours in A∖NA2(v1).
Let v2∈C be adjacent to u2. If v1,v2 have a common neighbour u1, then since NA(u1)⊆NA2(v2),
it follows that ∣NA2(v2)∣≥(x/(3(1−y))+x)∣A∣, and so x/(3(1−y))+x<2/3, that is,
(4−3y)x<2−2y<(4−3y)/2, and so x<1/2, a contradiction.
This proves (2).
(3) If v1,v2,v3∈C and N(v1)∩N(v2)=∅ then N(v3) is disjoint from exactly one of
N(v1),N(v2).
If N(v3) is disjoint from both N(v1),N(v2), then every two of N(v1),N(v2),N(v3) have union of
cardinality more than (1−x)∣B∣, and so every vertex in A belongs to at least two of NA2(vi)(i=1,2,3).
Consequently one of NA2(vi)(i=1,2,3) has cardinality at least a∣A∣/3, a contradiction.
Now suppose that N(u3) has nonempty intersection with both N(v1),N(v2). Thus ∣NA2(vi)∩NA2(v3)∣≥x∣A∣
for i=1,2, and since ∣NA2(v1)∩NA2(v2)∣<∣A∣/3, it follows that
∣NA2(v3)∣≥(2x−1/3)∣A∣≥2∣A∣/3 since x≥1/2, a contradiction. This proves (3).
Let H be the bipartite graph G[B∪C]. From (2) and (3), H has exactly two components H1 and H2 say.
Let Ci=V(Hi)∪C and Bi=V(Hi)∩B for i=1,2. Then from (3), every two vertices
in Ci have a common neighbour in Bi, for i=1,2.
Let ci=∣Ci∣/∣C∣, for i=1,2. Thus c1+c2=1.
We may assume that b1≥1/2.
Choose v1∈C1. Since ∣A∖NA2(v1)∣>∣A∣/3, and every vertex in A∖NA2(v1) has at least y∣C∣
second neighbours in C1, some vertex v2∈C1 has at least (y/(3c1))∣A∣ second neighbours in
A∖NA2(v1). But since v1,v2∈C1, they have a common neighbour in B1; therefore
∣NA2(v2)∣≥(y/(3c1)+x)∣A∣, and so y/(3c1)+x<2/3. Now c1≤1−y, so 3x+y/(1−y)<2, contrary to the
hypothesis.
This proves 7.3.
If A⊆V(G) and X⊆V(G)∖A, NA(X) denotes the set of vertices in A with a neighbour in X.
The next result is a useful lemma which says, roughly speaking, the larger X is, the larger NA(X) is. In this section we only use it for k=1, but we will
use it with k=2 in the section discussing when ψ(x,y)≥3/4.
7.4
Let x,y,z∈(0,1], and suppose that G is (x,y)-biconstrained via (A,B,C), and ∣NA2(w)∣<z∣A∣ for all
w∈C.
Then for all integers k≥1, if B′⊆B with
[TABLE]
then ∣NA(B′)∣>(x+k(1−z))∣A∣.
Proof. We proceed by induction on k, and so we assume that either k=1 or the result holds for k−1.
Let A′=NA(B′). Every vertex in B∖B′ has at least y∣C∣ neighbours in C, so there exists v∈C
with at least y∣B∖B′∣ neighbours in B∖B′.
By hypothesis,
∣B′∣/∣B∣>1−x/(1−y),
that is, x∣B∣+y∣B∖B′∣>∣B∖B′∣.
But every vertex in A∖A′ has at least x∣B∣ neighbours in B∖B′, and therefore has one such
neighbour adjacent to v; and so A∖A′⊆NA2(v).
Let B′′=B′∩NB(v), and A′′=NA(B′′).
(1) ∣A′′∣≥(x+(k−1)(1−z))∣A∣.
Since
∣NB(v)∣≥y∣B∣, it follows that
∣B′′∣≥y∣B∣−(∣B∣−∣B′∣)=∣B′∣−(1−y)∣B∣. If k=1, then ∣B′∣/∣B∣>1−y by hypothesis,
and therefore B′′=∅, and so ∣A′′∣≥x∣A∣ as claimed. If k≥2, then since
∣B′∣/∣B∣>(k−1)(1−y)+max(1−y,1−x/(1−y)), it follows that
[TABLE]
and so ∣A′′∣≥(x+(k−1)(1−z))∣A∣ from the inductive hypothesis. This proves (1).
Since A′′⊆A′, and A′′∪(A∖A′)⊆NA2(v),
it follows that
Let x,y∈(0,1], such that y>1/2, x+3y>2 and x>2(1−y)2/(2−y). Then ψ(x,y)≥2/3.
Proof. Let G be (x,y)-biconstrained, via (A,B,C), and suppose for a contradiction that ∣NA2(v)∣<2∣A∣/3
for each v∈C.
(1) If B′⊆B with ∣B′∣/∣B∣>max(1−y,1−x/(1−y)), then there are at least (x+1/3)∣A∣ vertices in A
with a neighbour in B′. In particular, if ∣B′∣≥(x+2y−1)∣B∣ then the same conclusion holds.
The first statement follows from 7.4 with k=1.
For the second, x+2y−1>1−y since x+3y≥2;
and x+2y−1>1−x/(1−y) since x>2(1−y)2/(2−y). Consequently x+2y−1>max(1−y,1−x/(1−y)).
This proves the second statement and so proves (1).
Say w1,w2∈C are close if ∣NB(w1)∪NB(w2)∣≤(1−x)∣B∣.
(2) There exists w1∈C such that the set of vertices in C that are close to w1 has cardinality at least ∣C∣/2.
This is trivial if every two vertices in C are close; so we assume there exist w1,w2∈C that are not close. Consequently
every vertex in A has a neighbour in NB(w1)∪NB(w2), and so NA2(w1)∪NA2(w2)=A. If there exists w∈C
that is not close to either of w1,w2 then similarly NA2(w1)∪NA2(w)=A and NA2(w2)∪NA2(w)=A;
and so every vertex in A
belongs to at least two of NA2(w1),NA2(w2),NA2(w), and therefore one of these three sets has cardinality at least 2∣A∣/3,
a contradiction. Thus, exchanging w1,w2 if necessary, we may assume that at least half of all vertices in C are close to
w1. This proves (2).
Let C1 be the set of vertices in C that are close to w1; thus ∣C1∣≥∣C∣/2. Let B1=NB(w1) and A1=NA2(w1).
Since ∣B1∣≥y∣B∣ and every vertex in A∖A1 has at least x∣B∣ neighbours in B∖B1, there exists
v2∈B∖B1 with at least 1−yx∣A∖A1∣ neighbours in A∖A1. Since y>1/2 and ∣C1∣≥∣C∣/2,
v2 has a neighbour w2∈C1. Since w2 is close to w1 it follows that ∣NB(w1)∪NB(w2)∣≤(1−x)∣B∣, and so
∣NB(w1)∩NB(w2)∣≥(x+2y−1)∣B∣. From (1), there are at least (x+1/3)∣A∣ vertices in A
with a neighbour in NB(w1)∩NB(w2). These vertices all belong to A1, and so
[TABLE]
Consequently 3(1−y)x+x+1/3<2/3, that is, 1−yx+3x<1. By hypothesis, x>2(1−y)2/(2−y), and so
substitution for x yields
[TABLE]
which simplifies to (3y−2)(2y−3)<0, contrary
to 1.2. This proves 7.5.
7.6
Let x,y∈(0,1]. If either
•
x≤11/17* and y≤1/7 and x+3y≤1; or*
•
x<1/8* and y≤11/17 and 3x+y≤1*
then ψ(x,y)<2/3.
Proof. Take the graph consisting of seven disjoint copies of a three-vertex path, numbered ai,bi,ci in order
(1≤i≤7).
For 1≤i≤3 and 4≤j≤7, make ai adjacent to bj and make aj adjacent to bi, forming G.
Let A={ai:1≤i≤7} and define B,C similarly.
For the first statement, we may assume by increasing x,y that x+3y=1. It follows that x≥4/7
(because y≤1/7) and similarly y≥2/17, and 4x=4−12y≥1+9y.
For 1≤i≤3, let w(ai)=3/17, w(bi)=(4x−1)/9, and w(ci)=1/7.
For 4≤i≤7, let w(ai)=2/17, w(bi)=(1−x)/3, and w(ci)=1/7.
Then this weighted graph is (x,y)-biconstrained via (A,B,C) and shows that ψ(x,y)<2/3. This proves the first statement.
For the second statement, we may assume that 3x+y=1, and so x≥2/17 and y>5/8, and
so 3y=3−9x≤1+8x.
Let us take the same graph and redefine w, as follows.
For 1≤i≤3, let w(ai)=(1−y)/2 and w(bi)=w(ci)=(1−4x)/3.
For 4≤i≤7, let w(ai)=(3y−1)/8 and w(bi)=w(ci)=x.
Then this weighted graph is (x,y)-biconstrained via (C,B,A) and shows that ψ(x,y)<2/3. This proves the second statement,
and hence proves 7.6.
7.7
Let x′,y′,z′∈(0,1] such that ψ(x′,y′)≤z′<1/2; and let x,y∈(0,1] satisfy
x≤2−x′1, x<1−3(1−z′)1−x′, y≤1+y′y′ and x+y′1−x′y≤1.
Then ψ(x,y)<2/3. Consequently:
•
ψ(x,y)<2/3* if x≤3/5 and y≤1/4 and x+2y≤1;*
•
ψ(x,y)<2/3* if x≤5/8 and y≤1/6 and x+3y≤1.*
Proof. The first statement follows from 3.4 taking z slightly less than 2/3.
The two statements in bullets follow by setting x′=y′=z′=1/3, and
then x′=z′=2/5 and y′=1/5. This proves 7.7.
7.8
Let x′,y′,z′∈(0,1] such that ψ(x′,y′)≤z′<1/2; and let x,y∈(0,1] satisfy
x<32x′, y≤2−y′1, and x′1−y′x+y≤1.
Then ψ(x,y)<2/3. Consequently:
•
ψ(x,y)<2/3* if y≤3/5 and 2x+y≤1; and*
•
ψ(x,y)<2/3* if y≥3/5 and x+3y≤2, and x+3y<2 if x or y is irrational.*
Proof. The first statement follows from 3.5, taking z slightly less than 2/3.
To prove the first bullet, let 2x+y≤1, and so x≤1/2. If also y≤1/2 then ψ(x,y)≤1/2
by 1.4; so we may assume that y>1/2.
We claim there is an integer k≥1 with
[TABLE]
To see this, if y>5/9 we can take k=1 (because we are given that y≤3/5), and if y≤5/9, then
[TABLE]
and so again k exists. Thus
y≤4k+12k+1, and
x≤21−y<6k+32k.
Let x′=z′=k/(2k+1), and y′=1/(2k+1). Then the claim follows from the first statement.
For the second bullet, let x+3y≤2 with y≥3/5, with x+3y<2 if x or y is irrational.
Consequently, we may assume that x,y are rational, by increasing them slightly if necessary.
Let x′=2/y−3, and y′=2−1/y; it follows that x′+2y′≤1 and
2x′+y′≤1, and x′,y′ are rational,
and so ψ(x′,y′)<1/2 by 3.3. The result follows from the first statement.
This proves 7.8.
For the mono-constrained question, we have:
7.9
For x,y∈(0,1], if y≤1/2 and x>(1−y)2/(1−2y2) then ϕ(x,y)≥2/3.
Proof. Let G be (x,y)-constrained via (A,B,C).
If x+y>1 the result follows from 5.2, so we may assume that x+y≤1. Since
x>(1−y)2/(1−2y2), we may also assume that
•
x,y are rational; and
•
every vertex in A has strictly more than x∣B∣ neighbours in B
by reducing x and y a little if necessary while retaining the
property that x>(1−y)2/(1−2y2).
Let p=(1−x−y)/(1−2y). Thus p is rational, so we may assume (by multiplying vertices) that p∣B∣ is an integer.
Also p≤y, since x>(1−y)2/(1−2y2).
Let s=(x−(1−y)2)/(y(1−y)). It follows that 0≤s≤1, since x>(1−y)2/(1−2y2) and x+y≤1.
Choose v1∈C with at
least y∣B∣ neighbours in B, and let
B1⊆N(v1) with ∣B1∣=y∣B∣.
Choose v2∈C such that
sb0+b2≥y(sy+(1−y)),
where b0∣B∣=∣N(v2)∩B1∣ and b2∣B∣=∣N(v2)∖B1∣.
(Such a vertex exists by averaging.)
We claim that b0+b2≥p; for from the definition of s,
[TABLE]
and p=(1−x−y)/(1−2y)≤xy/(1−y) since x>(1−y)2/(1−2y2).
Also we claim that b2≥1−x−y; for from the definition of s,
[TABLE]
Consequently ∣N(v1)∪N(v2)∣≥(1−x)∣B∣, and there exist P1⊆N(v1) and P2⊆N(v2),
both of cardinality p∣B∣. Choose v3∈C with at least y(1−2p)∣B∣ neighbours in B∖(P1∪P2).
Then for i=1,2,
[TABLE]
Since every vertex in A has strictly more than x∣B∣ neighbours in B, it follows that every vertex in A belongs to at least
two of the sets NA2(vi)(i=1,2,3); and so one of these sets has cardinality at least 2∣A∣/3. This proves 7.9.
7.10
For all x,y∈(0,1] with y≤1/2,
if 1−xx+1−2yy≤2, then ϕ(x,y)<2/3.
Proof. We may assume that x>1/2, or else the result is true since ϕ(1/2,1/2)=1/2.
The claim follows from 3.7 taking z slightly less than 2/3.
This proves 7.10.
8 The 1/3 level
Next we do the same for ψ(x,y)≥1/3 and ϕ(x,y)≥1/3. The figure summarizes our results.
8.1
Let x,y∈(0,1] with y>51 and 3x+3(1−y)y≥1. Then ψ(x,y)≥31.
Proof. We may assume that y≤1/3, by 1.2, and so y/(3(1−y))≤1/6. Consequently x≥5/18, and in particular
x>2y/3 (we will need this later). Also, since 1/5≤y≤1/3, it follows that 3y−y/(3(1−y))>1/2; and so
[TABLE]
and consequently x+y>1/2.
Let G be (x,y)-biconstrained via (A,B,C), and suppose that ∣NA2(v)∣<∣A∣/3 for each v∈C.
Let H be the subgraph induced on B∪C, and let H1,…,Hk
be its components. Let Bi=V(Hi)∩B and Ci=V(Hi)∩C, and
bi=∣Bi∣/∣B∣, ci=∣Ci∣/∣C∣, for 1≤i≤k. Since y>0, Bi,Ci are both nonempty
and so bi,ci≥y for 1≤i≤k.
For 1≤i≤k, let Ai be the set of vertices in A with a neighbour in Bi, and let Ai∗ be the set of vertices in A
such that N(v)⊆Bi.
(1) k≥2.
Suppose that k=1, and let H′ be the graph with vertex set B in which u,u′ are adjacent if u,u′ have
a common neighbour in H. Then every stable set of H′ has cardinality at most 4. By 6.2
there is a vertex u1∈B with H′-distance at most
four to every other
vertex in B; and so the H-distance from u1 to each vertex in B is at most eight. Let v1∈C be adjacent to u1.
Let A′=A∖NA2(v1) and B′=B∖N(v1). Hence ∣A′∣>2∣A∣/3.
Since every vertex in
A′ has at least x∣B∣ neighbours in B′, and ∣B′∣≤(1−y)∣B∣, some vertex u∈B′ has at least
[TABLE]
neighbours in A′. Choose a path of H between u1 and u of length at most eight, and let its vertices
be u1-v2-u2-⋯-vt-ut=u
say, in order. Thus t≤5,
and so there exists i with 1≤i≤t−1 such that there are at least ∣NA′(u)∣/4 vertices that belong to
NA(ui+1)∖NA(ui). Since ∣NA(ui)∣≥x∣A∣, it follows that
[TABLE]
a contradiction, since x≥2y/3 and so x+x/(6(1−y))≥x+y/(9(1−y))≥1/3 . This proves (1).
(2) bi≤1−x−y<1/2* for 1≤i≤k, and so k≥3.*
Suppose that b1>1−x−y say.
Thus, if u∈A∖A1, then u∈NA2(v) for every v∈C∖C1; and so ∣A∖A1∣<∣A∣/3, and
so ∣A1∣>2∣A∣/3.
Let H′ be the graph with vertex set C1 in which v,v′ are adjacent if they have a common H1-neighbour in B1.
Thus H′ has stability number at most three (by (1)) and so has radius at most three, by 6.2. Choose
v1∈C1 such that every vertex in C1 has H1-distance at most six from v1. Let A′=A1∖NA2(v1);
thus ∣A′∣>∣A∣/3. Since every vertex in
A′ has a neighbour in B1 and hence has at least y∣C∣ second neighbours in C1, there exists v∈C1
such that
[TABLE]
since ∣C1∣≤(1−y)∣C∣. Choose a path of H1 between v1,v of length at most six, with vertices
v1-u1-v2-⋯-ut−1-vt=v
say where t≤4. Then for some i with 1≤i≤t−1,
[TABLE]
and hence
[TABLE]
since all vertices of NA(ui) belong to NA2(vi+1) and do not belong to NA′2(vi+1)∖NA′2(vi).
But 3x+y/(3(1−y))≥1, a contradiction. This proves (2).
By (2), k≥3; and k≤4 since y>1/5. We may assume that ∣B1∣,∣B2∣≥∣Bi∣ for i≥3; let B0=⋃3≤i≤kBi,
and C0=⋃3≤i≤kCi. Hence ∣B0∣≤∣B∣/2 since k≤4. Let b0=∣B0∣/∣B∣ and c0=∣C0∣/∣C∣; let
A0 be the set of vertices in A with a neighbour in B0, and let A0∗ be the set of vertices in A
such that N(v)⊆B0. For 0≤i<j≤2, choose Aij=Aji⊆Ai∩Aj such that the sets
A12,A13,A23,A0∗,A1∗,A2∗ are pairwise disjoint and have union A.
For 0≤i≤2 let ai=∣Ai∣/∣A∣ and ai∗=∣Ai∗∣/∣A∣, and for 0≤i,j≤2 with i=j let aij=∣Aij∣/∣A∣.
Since b1,b2≤1−x−y<x+y and b0≤1/2<x+y, we have bi<x+y
for i=0,1,2. Let 0≤i≤2, and choose v∈Ci uniformly at random. Then Ai∗⊆NA2(v) because bi<x+y, and the
expected value of ∣NA2(v)∩(Ai∖Ai∗)∣ is at least (y/ci)∣Ai∖Ai∗∣; so the expected value of
∣NA2(v)∣ is at least
[TABLE]
Since ∣NA2(v)∣<∣A∣/3, it follows that
ai∗+(y/ci)(ai−ai∗)<1/3. Now A0∗,A01,A02 are pairwise disjoint subsets of A0, so
a01+a02≤a0−a0∗; and hence
[TABLE]
Similarly we have a1∗+(y/c1)(a01+a12)<1/3 and a2∗+(y/c2)(a02+a12)<1/3; and by summing these three
inequalities and using the equation
[TABLE]
we obtain
[TABLE]
Consequently there exist distinct i,j∈{0,1,2} with y/ci+y/cj−1<0. But
1/ci+1/cj≥4/(ci+cj), and ci+cj≤1−y,
and so 4y/(1−y)<1, a contradiction.
This proves 8.1.
For ϕ, we need the following.
8.2
Let x,y∈(0,1]. Let G be (x,2/3)-constrained via (A,B,C), such that every three vertices in B have a
common neighbour in C, and every vertex w∈C
satisfies ∣NA2(w)∣<(1−y)∣A∣. If v1∈B has a∣A∣ neighbours in A then
[TABLE]
Proof. Let v1∈B have a∣A∣ neighbours in A. Define A1=NA(v1) and choose C1⊆NC(v1) with ∣C1∣=2∣C∣/3
(we may assume this is an integer). Let A1′=A∖A1
and C1′=C∖C1.
(1) If some vertex v2∈B has a set A2 of t∣B∣ neighbours in A1′, then
•
every v∈B has at most (1−y−a−t)∣A∣ neighbours in A∖(A1∪A2); and
•
the sum over all v∈B of the number of neighbours of v in A∖(A1∪A2)
is (1−a−t)x∣A∣∣B∣.
The first claim follows since v1,v2,v have a common neighbour in C. The second holds since every vertex in
A∖(A1∪A2) has x∣B∣ neighbours. The third follows. This proves (1).
The sum over u∈A1′, of ∣NC12(u)∣, is at most 32(1−y−a)∣A∣∣C∣; and for each u,
∣NC12(u)∣≥maxv∈NB(u)∣NC1(v)∣. But the latter is at least
[TABLE]
It follows that
[TABLE]
Consequently
[TABLE]
Moreover, each vertex in C1′ has at least x∣B∣ nonneighbours in B, and so there are at most (1−x)/3∣B∣∣C∣
edges between B and C1′. Hence there are at least (1+x)/3∣B∣∣C∣ edges between B and C1.
For each v∈B, let p(v)=∣NA1′(v)∣/∣A∣.
Thus ∑v∈Bp(v)=x(1−a)∣B∣. By setting q(v)=3∣NC1(v)∣/∣C∣−1 we deduce:
for each v∈B there exists q(v) such that
•
for each v∈B, 1/3≤q(v)/3+1/3≤2/3, that is, 0≤q(v)≤1;
•
∑v∈B(q(v)/3+1/3)≥(1+x)/3∣B∣, that is, ∑v∈Bq(v)≥x∣B∣;
•
∑v∈Bp(v)(q(v)/3+1/3)≤32x(1−y−a)∣B∣, that is, ∑v∈Bp(v)q(v)≤(x−2xy−xa)∣B∣.
Let Q⊆B be the x∣B∣ vertices in B (we may assume this is an integer) with p(v) smallest.
Then the expression in the last bullet above is minimized by setting q(v)=1 for v∈Q, and q(v)=0 for v∈B∖Q.
Consequently ∑v∈Qp(v)≤(x−xa−2xy)∣B∣.
Choose v2∈B∖Q with ∣NA1′(v2)∣ maximum; A2 say, where ∣A2∣=t∣A∣. By (1),
every v∈B has at most (1−y−a−t)∣A∣ neighbours in A1′∖A2, and the sum over all v∈B of the number
of neighbours of v in A1′∖A2
is (1−a−t)x∣A∣∣B∣. So the number of edges between A1′∖A2 and B∖Q is at most
(1−y−a−t)(1−x)∣A∣∣B∣; and the number between A1′∖A2 and Q is at most (x−xa−2xy)∣A∣∣B∣, since
∑v∈Qp(v)≤(x−xa−2xy)∣B∣. Hence
the number between A1′∖A2 and B is at most ((1−y−a−t)(1−x)+x−xa−2xy)∣A∣∣B∣, and since this number
equals (1−a−t)x∣A∣∣B∣, it follows that
[TABLE]
that is,
[TABLE]
Consequently t≤(1−x−y−a+xa−xy)/(1−2x).
Now t(∣B∣−∣Q∣)∣A∣+(x−xa−2xy)∣A∣∣B∣≥x(1−a)∣A∣∣B∣, so
Let x,y∈(0,1] with (1−x)2>(2x+1)(1+2x−5x2)y
and x≥1/4 and 4x2y2≥(1−y)(x−y)2. Then ϕ(x,y)≥1/3.
Consequently if x>0.28231 then ϕ(x,x)≥1/3.
Proof. Let ϕ(x,y)=z and suppose that z<1/3. Then there is a graph G that is (x,1−z)-constrained via A,B,C, such that
∣NA2(w)∣≤(1−y)∣A∣ for each w∈C, by 2.5
and 2.3. As in 5.7, there exists w∈C such that there are at least x(1−z)∣A∣⋅∣B∣ edges between
NB(w) and NA2(w). Define B1=NB(w) and B2=B∖B1; and let B2=t∣B∣. For each u∈A,
let u have d(u)∣B∣ neighbours in B1.
Let A1=NA2(w) and A2=A∖A1; and let ∣A2∣=s∣A∣. Thus d(u)=0 for each u∈A2.
(1) t≥x, and s≥y, and 0≤d(u)≤min(x,1−t) for each u∈A. Also
[TABLE]
We may assume that every vertex in A has degree exactly x∣B∣;
so 0≤d(u)≤min(x,1−t) for each u∈A. Since ∣A1∣≤(1−y)∣A∣, it follows that s≥y. In particular, A2=∅, and so some vertex in A2
has x∣B∣ neighbours in B2, and so t≥x. Since there are at least x(1−z)∣A∣⋅∣B∣ edges between
NB(w) and NA2(w), it follows that ∑u∈A1d(v)≥x(1−z)∣A∣. Since ∣A1∣≤(1−y)∣A∣
and every vertex in A1 has degree exactly x∣B∣, it follows that the number of edges between A1 and B1
is at most x(1−y)∣A∣⋅∣B∣, and so x(1−y)∣A∣≥∑u∈A1d(v).
This proves (1).
(2) 1−t≥(1−x)2−(1+2x−5x2)y2x(1−x)2/3. Consequently x<1−3t/2 and so x<1−t.
There are at least 2x∣A∣⋅∣B∣/3 edges between B1 and A, since z<1/3. But each vertex in B1 has at most
[TABLE]
neighbours in A, by 8.2, and the first claim follows.
To show that x<1−3t/2, suppose not; then
[TABLE]
and so
[TABLE]
that is,
[TABLE]
contrary to the hypothesis.
This proves (2).
Let us choose v1,v1′∈A1 uniformly and independently at random, and choose v2∈A2 uniformly at random. Then
for u∈A, the probability that all of v1,v1′,v2 are nonadjacent to u is
[TABLE]
Since 1−y>2/3 and so v1,v1′,v2 have a common neighbour in C, say w′, and ∣NA2(w′)∣≤(1−y)∣A∣, it follows
that
[TABLE]
that is,
[TABLE]
This can be rewritten as:
[TABLE]
where f(r) is the polynomial (r+t−x)(r+t−1)2.
We therefore need to investigate the maximum value of ∑u∈Af(d(v)) (which we call “the objective function”)
over all choices of the numbers d(u)(u∈A) satisfying the various constraints, and verify that this maximum is
less than t(1−t)2(1−y)∣A∣.
The derivative of f(r) is zero when 3r2+2(3t−x−2)r+(3t−2x−1)(t−1)=0, which has roots
r=1−t and r=(2x+1)/3−t. Let us define r0=(2x+1)/3−t. Since r0<1−t, the function f(r) increases for r<r0
and for r>1−t, and decreases for r0<r<1−t.
The second derivative of f(r) is zero when 3r+3t−x−2=0, that is, when r=r1 where r1=2/3+x/3−t.
By (2), x<r1, and we are only concerned f(r) for r in with the range 0≤r≤x; so in particular
all such r are less than r1. The function f(r) is concave through the range 0≤r≤r1, since its second derivative
is at most zero.
Let us choose real numbers d(v)(v∈A) satisfying the constraints
•
0≤d(u)≤x for each u∈A;
•
d(u)=0 for at least y∣A∣ vertices u∈A;
•
x(1−y)∣A∣≥∑u∈A1d(v)≥2x∣A∣/3
to maximize the function ∑u∈Af(d(v)). From the concavity of f, it follows that
there exists r∗ with 0<r∗≤x such that d(v)∈{0,r∗} for all v (because if there were u,v with d(u),d(v)
distinct and nonzero, replacing them both by (d(u)+d(v))/2 would still satisfy the constraints and increase the
objective function). Similarly, if there were more than y∣A∣ vertices v with d(v)=0, then choose some one of them, v
say, and choose some u with d(u)>0; then again replacing them both by (d(u)+d(v))/2 would still satisfy the constraints and increase the
objective function. We deduce that there are exactly y∣A∣ vertices v with d(v)=0.
Now the problem breaks into three cases, depending which of the constraints x(1−y)∣A∣≥∑u∈A1d(v)≥2x∣A∣/3
hold with equality.
Suppose first that neither holds with equality. Then from the optimality of the objective function, it follows that
r∗=r0, and since
[TABLE]
it follows that
[TABLE]
that is,
[TABLE]
Thus, if t satisfies
[TABLE]
then there is a possible
optimal solution where the objective function has value
[TABLE]
Now f(0)=(t−x)(t−1)2, and
[TABLE]
We must therefore check that for t in the given range,
[TABLE]
This simplifies to:
[TABLE]
Now the function 27(1−t)2(t−2ty+xy) has no local minimum at t with t<1, and so is minimized at one of
the ends of the range. Since t≥x, we might as well replace the lower extreme of the range by t≥x (because
it makes the arithmetic easier); so to check the lower extreme, we need to check that
[TABLE]
that is,
4(1−x)<27x, which is true by hypothesis.
For the upper extreme, t≤(2x+1)/3−2x/(3−3y)<1/3; so it suffices to check that
[TABLE]
when t=1/3, that is, to check
(1−x)3(1−y)<(1−2y+3xy). But (1−x)3<1/2 by hypothesis, and (1−2y+3xy)/(1−y)≥(1−2y)/(1−y)≥1/2 since y<1/3.
This finished the first of the three cases.
Now let us assume that x(1−y)∣A∣=∑u∈A1d(v). It follows from the optimality of the objective function that
r∗≤r0. Moreover, since x(1−y)∣A∣=∑u∈A1d(v), it follows that x(1−y)∣A∣=(1−y)∣A∣r∗, so
r∗=x. This is only possible if x≤r0, that is, t<(1−x)/3; and this is impossible since t≥x≥1/4.
This finishes the second case.
Finally, we assume that ∑u∈A1d(v)=2x∣A∣/3. It follows from the optimality of the objective function that
r∗≥r0. Moreover, since ∑u∈A1d(v)=2x∣A∣/3, it follows that (1−y)∣A∣r∗=2x∣A∣/3, that is,
r∗=2x/(3−3y).
We must check that
[TABLE]
This is cubic in t, and, collecting the various powers of t, it becomes:
[TABLE]
[TABLE]
[TABLE]
This simplifies to:
[TABLE]
The derivative of the left side with respect to t is
[TABLE]
which can be rewritten as
[TABLE]
Since by hypothesis, −(x−y)2/(3y)+4x2y/(3−3y)≥0, the derivative is nonnegative, at every value of t.
Thus we only need verify the inequality for the maximum value of t that lies in the range.
By (2), t≤2(1−x)/3; so it is enough to verify that
[TABLE]
holds when t=2(1−x)/3. Thus we need to check that
[TABLE]
[TABLE]
According to WolframAlpha, this is true for all x,y with 1/4<x<1/3 and 0<y≤x. This proves 8.3.
8.4
Let x,y∈(0,1] with x≤41 and y<31 and y<3−12x+16x2(1−2x)2; then ϕ(x,y)<31.
In this section we investigate when ψ(x,y)≥3/4 and ϕ(x,y)≥3/4. The results are shown in figure 8.
9.1
Let x,y∈(0,1], such that y>1/3, x≥1/2, and 2y−2y2>1−x. Then ψ(x,y)≥3/4.
Proof. Let G be a graph that is (x,y)-biconstrained via (A,B,C).
We can assume that x,y are rational, and by multiplying vertices if necessary that y∣B∣∈Z. Let v1∈C, and
let B1⊆N(v1) be such that ∣B1∣=y∣B∣. Choose v2∈C with at least y∣B∖B1∣=y(1−y)∣B∣ neighbours
in B∖B1, and choose B2⊆N(v2) with ∣B2∣=y∣B∣. Choose v3∈C with at least y(1−2y) neighbours in
B∖(B1∪B2). Thus ∣N(v1)∪N(v2)∣≥y+y(1−y)>1−x, and for i=1,2, ∣N(vi)∪N(v3)∣≥y+y(1−2y)>1−x.
For 1≤i≤3, let Ai=NA2(vi). Since y>1/3, it follows that there exist i,j with 1≤i<j≤3
such that N(vi)∩N(vj)=∅, and so ∣Ai∩Aj∣≥x∣A∣≥∣A∣/2.
But Ai∪Aj=A since ∣N(vi)∪N(vj)∣>1−x, and
so ∣Ai∣+∣Aj∣≥3∣A∣/2, and therefore one of ∣Ai∣,∣Aj∣≥3∣A∣/4. This proves 9.1.
9.2
Let x,y∈(0,1], such that x>2/3, x+2y>1, and either 4(1−x)(1−y)≤1 or x>1−2y+2y2.
Then ψ(x,y)≥3/4.
Proof. If x+y>1 the result follows from (theorem 4.1 of the paper); so we may assume that y≤1−x<1/3.
Let G be (x,y)-biconstrained via (A,B,C), and suppose that ∣NA2(v)∣<3∣A∣/4 for each v∈C.
Let H be the graph with V(H)=V(C), where distinct u,v are adjacent in H if and only if u and v have distance two in G
(that is, in G they have a common neighbour in B).
(1) For all w1,w2,w3∈C, if N(w1)∩N(w2)=∅ and N(w2)∩N(w3)=∅, then
N(w1)∩N(w3)=∅. Consequently each component of H is a complete graph.
Suppose that w1,w2,w3∈C, and N(w1)∩N(w3)=∅, and vi∈N(wi)∩N(w2) for i=1,3.
Let Ai=NA2(wi) for i=1,2,3. Since x+2y>1 we have A1∪A3=A. Consequently ∣A1∩A3∣<∣A∣/2, and since
NA(v1)∩NA(v3)⊆A1∩A3, it follows that ∣NA(v1)∩NA(v3)∣<∣A∣/2. Thus
[TABLE]
a contradiction. This proves (1).
Let α be the size of the largest stable set in H, that is, the number of components of H.
Let the vertex sets of the components of H be C1,…,Cα, and for 1≤i≤α let Bi be the set of
vertices in B with a
neighbour in Ci. The sets B1,…,Bα have union B, and from the definition of H, they are pairwise disjoint.
For 1≤i≤α, let wi∈Ci and let Ai=NA2(wi).
(2) For 1≤i<j≤α, Ai∪Aj=A, and so ∣Ai∩Aj∣<∣A∣/2. Consequently α≤3.
Since wi,wj have no common neighbour in B, it follows that ∣N(wi)∪N(wj)∣≥2y∣B∣>(1−x)∣B∣, and so
Ai∪Aj=A. Since ∣Ai∣,∣Aj∣<3∣A∣/4, it follows that ∣Ai∩Aj∣<∣A∣/2. This proves the first assertion.
Suppose that α≥4. By the first assertion, every vertex in A belongs to at least three
of A1,…,A4. Consequently some Ai has cardinality at least 3∣A∣/4, a contradiction. This proves (2).
(3) α=1.
Suppose that α=1. Every vertex in A∖A1 has at least x∣B∣ neighbours in B∖N(w1), so we may
choose v∈B with at least
[TABLE]
neighbours in A∖A1. Let w∈C be a neighbour of v. Since
w1,w have a common neighbour, it follows that
[TABLE]
a contradiction. This proves (3).
(4) For 1≤i≤α, if ∣Bi∣>(1−x)∣B∣ then ∣Ci∣>3−4xy∣C∣.
Suppose that ∣Bi∣>(1−x)∣B∣ say, and let ∣Ci∣=c∣C∣. Since x>2/3, every vertex u∈A∖Ai has a neighbour in
Bi, and so ∣NC2(u)∩Ci∣≥y∣C∣=(y/c)∣Ci∣. Hence there exists w∈Ci such that
[TABLE]
But w,wi have a common neighbour, and so
∣NA2(w)∩Ai∣≥x∣A∣, and therefore
[TABLE]
Consequently 3/4>x+y/(4c), and so c>y/(3−4x). This proves (4).
(5) x>1−2y+2y2* and α=2.*
Since α≤3, we may assume without loss of generality that ∣B1∣≥∣B∣/3>(1−x)∣B∣. Since each ∣Ci∣≥y∣C∣,
it follows that ∣C1∣≤(1−(α−1)y)∣C∣. By (4), 1−(α−1)y>y/(3−4x), and since α≥2, it follows that
1−y>y/(3−4x), that is, 4(1−x)(1−y)>1. From the hypothesis it follows that x>1−2y+2y2. This proves the first claim.
Suppose that α>2; then
[TABLE]
which simplifies to (1−y)(1−4y)2<0, a contradiction. This proves (5).
We may assume without loss of generality that ∣C1∣≤∣C∣/2. By (4),
∣B1∣≤(1−x)∣B∣<∣B∣/2, and so ∣B2∣≥∣B∣/2.
Every vertex in B2∖N(w2) is adjacent to at least a fraction y/(1−y) of the vertices
of C2, and hence there exists w∈C2 with
[TABLE]
Thus
[TABLE]
Since ∣B2∣≥x∣B∣ and ∣N(w2)∣≥y∣B∣, it follows that
[TABLE]
(because x>1−2y+2y2 by (5)).
Thus, NA2(w)∪NA2(w2)=A, and since w and w2 have a common neighbour in B2 it follows that
[TABLE]
and so one of ∣NA2(w)∣,∣NA2(w2)∣≥3∣A∣/4, a contradiction.
This proves 9.2.
9.3
Let x,y∈(0,1], with x>1/3, y>1/2, x+3y>2, and either x≥(5−6y)/(11−12y) or x≥(3−4y)/(4−4y).
Then ψ(x,y)≥3/4.
Proof. Let G be (x,y)-biconstrained via (A,B,C), and suppose that NA2(v)<3∣A∣/4
for each v∈C. By (theorem 4.1 of the paper) it follows that x+y≤1.
Let H be the graph with V(H)=V(C), in which distinct u,v
are adjacent if and only if ∣N(u)∪N(v)∣≤(1−x)∣B∣.
(1) For all u,v∈C, if u,v are nonadjacent in H then NA2(u)∪NA2(v)=A. If u,v
are adjacent in H then ∣NA2(u)∩NA2(v)∣>(x+1/4)∣A∣.
If u,v are nonadjacent in H, then ∣N(u)∪N(v)∣>(1−x)∣B∣, and so every vertex in A has a neighbour
in N(u)∪N(v), that is, NA2(u)∪NA2(v)=A. Now we assume that u,v are adjacent in H. Consequently
[TABLE]
by hypothesis. Moreover, x>1/3 and y>1/2 imply that
[TABLE]
Thus, 7.4 implies that ∣NA2(u)∩NA2(v)∣>(x+1/4)∣A∣. This proves (1).
If there exist w1,…,w4∈C, pairwise nonadjacent in H, then by (1) each pair of the sets
of the sets NA2(wi)(1≤i≤4) has union A, and so each vertex in A belongs to at least three of the four sets; and
so one of the four sets has cardinality at least 3∣A∣/4, a contradiction. Thus we may choose w1,w2,w3∈C
such that every other vertex in C is adjacent in H to at least one of w1,w2,w3. Choose a partition C=C1∪C2∪C3
such that for 1≤i≤3, every vertex in Ci is equal or adjacent in H to wi.
(2) ∣Ci∣≤(1−y)∣C∣* for 1≤i≤3.*
Suppose that ∣C1∣>(1−y)∣C∣. Define B1=N(w) and A1=NA2(w). Choose v∈B∖N(w) with at least
x∣A∖A1∣/(1−y)>x∣A∣/(4−4y) neighbours in A∖A1.
Since ∣C1∣>(1−y)∣C∣, there exists
w∈C1 adjacent to v. Then
[TABLE]
since x>1/3 and y>1/2, a contradiction. This proves (2).
(3) Every vertex in B has neighbours in exactly two of C1,C2,C3.
Since each ∣Ci∣≤(1−y)∣C∣<y∣C∣ by (2), it follows that every vertex in B has neighbours in at least two of C1,C2,C3.
Suppose that v∈B has a neighbour wi′∈Ci for i=1,2,3.
Let Ai=NA2(wi′) for i=1,2,3. For 1≤i<j≤3, Ai∪Aj=A by (1), and so every vertex of A belongs to
at least two of A1,A2,A3, and NA(u) is a subset of all three of A1,A2,A3. Consequently
[TABLE]
and so some ∣Ai∣≥3∣A∣/4, a contradiction. This proves (3).
From (3) we may partition B into B1,B2,B3 such that
every vertex
in B1 has neighbours in C2 and in C3 but not in C1, and similarly for B2,B3.
Without loss of generality,
we may assume that
∣B1∣≤1/3. Let A1=NA2(w1).
(4) ∣NA2(w)∖A1∣<(2−4x)∣A∖A1∣* for each w∈C1.*
By (1) ∣NA2(w1)∩NA2(w)∣≥(x+1/4)∣A∣, and since ∣NA2(w)∣<3∣A∣/4,
it follows that
[TABLE]
This proves (4).
This has two consequences. The first is that x<(3−4y)/(4−4y). To see this, by (4) we may choose u∈A∖A1 such that ∣NC2(u)∩C1∣<(2−4x)∣C1∣.
Since ∣B1∣≤∣B∣/3, u has a neighbour v∈B2∪B3, and we may assume that v∈B2 from the symmetry.
So at least y∣C∣ neighbours of v belong to C1∪C3, and therefore at least
(y−(1−y))∣C∣ neighbours belong to C1, since ∣C2∣≤(1−y)∣C∣.
So (2y−1)∣C∣<(2−4x)∣C1∣≤(2−4x)(1−y)∣C∣, and hence x<(3−4y)/(4−4y) as claimed.
The second consequence is that x<(5−6y)/(11−12y). To see this, let S=B∖(B1∪N(w1)). By (4)
and since each vertex in B2∪B3 has a neighbour in C1, it follows that each vertex v∈B2∪B3
has fewer than (2−4x)∣A∖A1∣ neighbours in A∖A1. Since
S⊆B2∪B3, it follows that some vertex u∈A∖A1
has fewer than (2−4x)∣S∣ neighbours in S. But u has no neighbours in
N(w1), and only at most r∣B∣ neighbours in B1; and since it has at least x∣B∣ neighbours in total, we deduce that
[TABLE]
(since B1∩N(w)=∅ and ∣N(w1)∣≥y∣B∣ and therefore ∣S∣≤(1−r−y)∣B∣). Consequently
[TABLE]
and so x<(5−6y)/(11−12y).
We have shown then that x<(3−4y)/(4−4y) and x<(5−6y)/(11−12y); but this contradicts the hypothesis.
This proves 9.3.
9.4
Let x,y∈(0,1] with 1/2<y<2/3 and x>6y2−8y+3. Then ψ(x,y)≥3/4.
Proof. We may assume that y∈Q, by decreasing y if necessary. Let G be (x,y)-biconstrained via (A,B,C), and suppose that
∣NA2(w)∣<3∣A∣/4 for each w∈C. By multiplying vertices if necessary, we may assume that y∣B∣∈Z.
Since x>6y2−8y+3=6(y−2/3)2+1/3, it follows that x>1/3.
Let H be the graph with V(H)=C in which distinct u,v∈C are adjacent if and only if ∣N(u)∪N(v)∣≤(1−x)∣B∣.
It follows that if u,v are nonadjacent in H, then NA2(u)∪NA2(v)=A.
As in the proof of 9.3, there do not w1,…,w4∈C, pairwise nonadjacent in H; and so we may choose w1,w2,w3∈C
and a partition C=C1∪C2∪C3 such that for 1≤i≤3, every vertex in Ci is equal or adjacent in H to wi.
Let ∣Ci∣=ci∣C∣, and choose Bi⊆N(wi) with ∣Bi∣=y∣B∣ for 1≤i≤3.
Let F be the set of all edges
vw of G with v∈B and w∈C, such that for 1≤i≤3, not both v∈Bi and w∈Ci.
(1) ∣B∣∣C∣∣F∣≤(1−x−y)<(2−3y)(2y−1).
Let w∈C, with w∈Ci say; then since w,wi are adjacent in H, it follows that
w has at most (1−x−y)∣B∣ neighbours in B∖Bi. Thus ∣F∣≤(1−x−y)∣B∣⋅∣C∣. But 1−x−y<(2−3y)(2y−1)
since x>6y2−8y+3. This proves (1).
Let p1=∣B1∖(B2∪B3)∣/∣B∣, and define p2,p3 similarly. Let
q1=∣(B2∪B3)∖B1∣/∣B∣, and define q2,q3 similarly. Let p0=∣B∖(B1∪B2∪B3)∣/∣B∣, and
q0=∣B1∪B2∪B3∣/∣B∣. Let q=q1+q2+q3. Thus
[TABLE]
By subtracting the last three of these from the first, we obtain
[TABLE]
and so p0=2q0+q−3y+1.
Every vertex in B∖(B1∪B2∪B3) is incident with at least y∣C∣ edges in F,
every vertex in B1∖(B2∪B3) is incident with at least (y−c1)∣C∣ edges in F, and every vertex
in (B2∪B3)∖B1 is incident with at least max(y−c2−c3,0)=max(y+c1−1,0) edges in F (and similar
statements hold for c2,c3). Summing, we deduce that
[TABLE]
Since pi=y−q0−q+qi for i=1,2,3, and c1+c2+c3=1, it follows that
[TABLE]
Also, p0=2q0+q−3y+1, and so
[TABLE]
This simplifies to
[TABLE]
where I is the set of i∈{1,2,3} such that ci<1−y. From (1) we deduce that
[TABLE]
In particular it follows that (1−y)q0<(2−3y)(2y−1)≤(1−y)(2y−1), and so q0<2y−1. Moreover, since ∣B2∪B3∣≤∣B∣,
it follows that ∣B2∩B3∣≥(2y−1)∣B∣, and so q1≥2y−1−q0, and the same holds for q2,q3.
Consequently
[TABLE]
and so
[TABLE]
since 2y−1−q0>0. This simplifies to
(2y−1)q0<0, a contradiction. This proves 9.4.
9.5
Let x,y∈(0,1] with x≥1/4 and y>2/3. Then ψ(x,y)≥3/4.
Proof. Suppose that G is (x,y)-biconstrained via (A,B,C), and ∣NA2(w)∣<3∣A∣/4 for each w∈C.
Since x≥1/4 and y>2/3, it follows that x>2(1−y)2, that is, y>(1−y)+1−x/(1−y) ; and also y>2−2y.
Let w∈C. By 7.4 with k=2 and B′=N(w), it follows that
Let x,y∈(0,1] with y>2/3, x+4y>3 and x>3(1−y)2/(2−y).
Then ψ(x,y)≥3/4.
Proof. Suppose that G is (x,y)-biconstrained via (A,B,C), and ∣NA2(w)∣<3∣A∣/4 for each w∈C. Consequently y≤3/4,
and so x≥3/20 since x>3(1−y)2/(2−y).
From the hypotheses it follows
that
[TABLE]
If w1,w2∈C with ∣N(w1)∪N(w2)∣≤(1−x)∣B∣ then ∣N(w1)∩N(w2)∣≥(2y+x−1)∣B∣. Thus, 7.4
applied with k=2 tells us that, for all such w1,w2∈C, more than (x+1/2)∣A∣ vertices in A have a neighbour in
N(v1)∩N(v2). Let H be the graph with vertex set C, in which w1,w2 are adjacent if ∣N(w1)∪N(w2)∣≤(1−x)∣B∣.
As in the proof of 9.3, there is no stable set of size at least four in H.
It follows that there exist w1,w2,w3∈C and a partition C=C1∪C2∪C3 such that for 1≤i≤3, every
vertex in Ci is equal to or adjacent in H to wi. We assume without loss of generality that ∣C1∣≥1/3. Since y>2/3,
every vertex in B has a neighbour in C1. Let B1=N(w1) and A1=NA2(w1), and choose v∈B∖B1 with
more than x∣A∣/(4−4y) neighbours in A∖A1.
Since y>2/3,
there exists w∈C1 adjacent to v. Then
[TABLE]
since x≥3/20≥1/7 and y≥2/3, a contradiction.
This proves 9.6.
9.7
Let x,y∈(0,1] with y>1/2 and x2y≥(3/4−x)2. Then ϕ(x,y)≥3/4.
Let x,y∈(0,1] with y<1/3 and x>1+y−3y21−y. Then ϕ(x,y)≥3/4.
Proof. Suppose that G is (x,y)-constrained via (A,B,C), and ∣NA2(w)∣<3∣A∣/4 for each w∈C. Consequently x+y≤1.
By reducing x or y
if necessary, we may assume that every vertex in A has strictly more than x∣B∣ neighbours in B, and that
x,y are rational. Let
[TABLE]
By multiplying vertices, we may also assume that y∣B∣ and p∣B∣ are integers. Note that the hypotheses imply
that p<xy.
(1) There exists s∈[0,1] such that for all b,c, if 0≤a≤y and sa+b≥y(sy+1−y),
then a+b≥p and b≥1−x−y.
We claim first that
[TABLE]
To see this, we need to check that 0≤y−y22y−y2+x−1, and y2p−y(1−y)≤1, and
y2p−y(1−y)≤y−y22y−y2+x−1. The first is true since
[TABLE]
The second is true since p≤xy≤y. The third simplifies to p/y≤x/(1−y), and this is true since p≤xy.
This proves the claim, and so there exists s such that
[TABLE]
We will show that s satisfies (1). Suppose that 0≤a≤y and sa+b≥y(sy+1−y). Then
[TABLE]
and
[TABLE]
(and therefore b≥1−x−y). This proves (1).
(2) There exists t∈[0,1] such that for all a,b, if 0≤a≤2p and ta+b≥y(1−2p(1−t)) then a+b≥p and p+b≥1−x.
We claim first that
[TABLE]
To see this we must check that
0≤2p(1−y)x+y+p−2py−1, and 2py2py+p−y≤1, and
2py2py+p−y≤2p(1−y)x+y+p−2py−1.
The first is true since
[TABLE]
The second is true since p≤xy≤y; and the third simplifies to
p≤xy and so is true. This proves the claim, and so there exists t with
[TABLE]
We will show that t satisfies (2). Let a,b satisfy 0≤a≤2y and ta+b≥y(1−2p(1−t)). Then
[TABLE]
and
[TABLE]
(and so b≥1−x−p). This proves (2).
Choose w1∈C with at least y∣B∣ neighbours in B.
(3) There exists w2∈C such that ∣N(w2)∣≥p∣B∣ and ∣N(w1)∪N(w2)∣≥(1−x)∣B∣.
Choose B1⊆N(w1) with
∣B1∣=y∣B∣. Choose s as in (1).
Then
[TABLE]
Consequently we may choose w2∈C such that
[TABLE]
Since
[TABLE]
the choice of s implies that ∣N(w2)∣≥p∣B∣ and ∣N(w2)∖B1∣≥(1−x−y)∣B∣, and so ∣N(w1)∪N(w2)∣≥(1−x)∣B∣.
This proves (3).
(4) There exists w3∈C such that ∣N(w3)∣≥p∣B∣, and ∣N(wi)∪N(w3)∣≥(1−x)∣B∣ for i=1,2.
Since ∣N(w1)∣,∣N(w2)∣≥p∣B∣ and p∣B∣ is an integer, we may choose R⊆B with ∣R∣=2p∣B∣
such that ∣N(w1)∩R∣,∣N(w2)∩R∣≥p∣B∣. Choose t as in (2). As in the proof of (3), there exists w3∈C with
[TABLE]
From the choice of t, it follows that ∣N(w3)∣/∣B∣≥p, and ∣N(w3)∖R∣≥1−x−p, and consequently
∣N(w1)∪N(w3)∣,∣N(w2)∪N(w3)∣≥(1−x)∣B∣. This proves (4).
For 1≤i≤3, choose Bi⊆N(wi) with ∣Bi∣=p∣B∣.
Since ∣B1∪B2∪B3∣≤3p, we may choose w4∈C with at least y(1−3p)∣B∣ neighbours in
B∖(B1∪B2∪B3). Then for all 1≤i≤3 we have:
[TABLE]
by the definition of p. It follows that for 1≤i<j≤4, ∣N(wi)∪N(wj)∣≥(1−x)∣B∣, and so (since every vertex in A
has strictly more than x∣B∣ neighbours in B) it follows that NA2(wi)∪NA2(wj)=A. Thus every vertex in A belongs to
at least three of the four sets NA2(wi)(1≤i≤4), and so one of them has cardinality at least 3∣A∣/4, a contradiction.
This proves 9.8.
9.9
Let x,y∈(0,1]. Then ψ(x,y)<3/4 if either:
•
x≤1/6* and y≤5/7 and 2x+y≤1; or*
•
x≤3/20* and x+4y≤3, and x+4y<3 if x is irrational; or*
•
x≤17/23* and y≤1/8 and x+3y≤1; or*
•
x≤5/7* and y≤1/6 and x+2y≤1.*
Proof. If x′,y′ with 2x′+y′≤1 and
y′≤3/5, then ψ(x′,y′)<2/3 by the first bullet of 7.8. Given x,y as in the first bullet,
the hypotheses imply that there is a choice of x′,y′ with 2x′+y′≤1 and
y′≤3/5, and which also satisfy the hypotheses of 3.5
with z′=ψ(x′,y′) and z slightly less than 3/4 (checking this needs some lengthy calculation, which we omit); and
so the first statement follows from 3.5.
The second statement follows similarly from 3.5 and the second bullet of 7.8. The third statement
follows
from 3.4 and the first bullet of 7.6; and the fourth follows by applying 3.4 with
z=max(2/7,x), taking x′=3/5, y′=1/5 and z′=3/5. This proves 9.9.
9.10
If x,y∈(0,1], with y≤1/3 and
[TABLE]
then ϕ(x,y)<3/4.
Proof. Apply 3.7 with z slightly less than 3/4 to 7.10. This proves 9.10.
10 The 2/5 Level
Next, we analyze when ψ,ϕ≥2/5. The results are shown in figure 9.
10.1
Let x,y∈(0,1] with x≥1/5, y>1/3, and 3y−2y2>1−x; then ψ(x,y)≥2/5.
Proof. Suppose that G is (x,y)-biconstrained via (A,B,C), and ∣NA2(w)∣<(2/5)∣A∣
for each w∈C.
Let w1∈C, and let A1=NA2(w1).
By averaging, there exists w2∈C such that
[TABLE]
where A2=NA2(w2).
Since
∣A2∣<2∣A∣/5, it follows that
[TABLE]
and so N(w1)∩N(w2)=∅.
Let B′=N(w1)∪N(w2); thus ∣B′∣≥2y∣B∣. By averaging, there exists w3∈C such that
∣N(w3)∖B′∣≥y∣B∖B′∣,
and so
[TABLE]
Hence, setting A3=NA2(w3), it follows that A1∪A2∪A3=A.
Since y>1/3, some pair of N(w1),N(w2),N(w3) have nonempty intersection, and so some pair of A1,A2,A3 have
intersection of cardinality at least x∣A∣≥∣A∣/5.
But then
[TABLE]
which is impossible since ∣Ai∣<(2/5)∣A∣ for 1≤i≤3.
This proves 10.1.
10.2
Let x,y∈(0,1] with x>1/3, x+3y>1, and either y≥1/4 or x+y/(10(1−2y))≥2/5; then ψ(x,y)≥2/5.
Proof. Suppose that G is (x,y)-biconstrained via (A,B,C), and ∣NA2(w)∣<(2/5)∣A∣ for each w∈C. Choose w1,…,wα∈C
with α
maximum such that N(w1),…,N(wα) are pairwise disjoint.
(1) α=3.
Suppose that α≤2. Let A′ be the union of the sets NA2(wi) for 1≤i≤α−1, and let B′ be the union of the
sets N(wi) for 1≤i≤α−1. So
∣A′∣<(2α/5)∣A∣≤(4/5)∣A∣, and ∣B′∣≥y∣B∣. By averaging, there exists v∈B∖B′ such that
[TABLE]
let w∈C be adjacent to v. Since
N(w) has nonempty intersection with N(wi) for some i<α, it follows that ∣NA2(w)∩A′∣≥x∣A∣.
Adding, we deduce that
[TABLE]
a contradiction.
Thus α≥3; suppose that α≥4.
Since x+3y>1 it follows that every vertex in A belongs to at least two of the sets NA2(wi)(1≤i≤4), and so one of these four sets has cardinality at least ∣A∣/2≥(2/5)∣A∣, a contradiction. This proves (1).
(2) If w∈C then N(w)∩N(wi) is nonempty for exactly one value of i∈{1,2,3}.
Since α=3, it follows that N(w)∩N(wi) is nonempty for at least one such i; suppose that
N(w) has nonempty intersection with both N(w1),N(w2) say. Let Ai=NA2(wi) for i=1,2,3, and A0=NA2(w).
Since A1∪A2∪A3=A, and each ∣Ai∣<(2/5)∣A∣, there are fewer than ∣A∣/5 vertices in A that belong to more than
one of A1,A2,A3, and in particular ∣A1∩A2∣<∣A∣/5. But ∣A0∩Ai∣≥x∣A∣ for i=1,2, and so
∣A0∣≥(2x−1/5)∣A∣≥(2/5)∣A∣, a contradiction. This proves (2).
From (2), we can partition B=B1∪B2∪B3, and partition C=C1∪C2∪C3, such that all six of these sets are
nonempty,
and for all distinct i,j∈{1,2,3} there is no edge between Bi and Cj, and for all i∈{1,2,3} and all w,w′∈Ci,
N(w)∩N(w′)=∅. Let ∣Bi∣=bi∣B∣ and Ci∣=ci∣C∣
for i=1,2,3.
Without loss of generality we may assume that
b3≤1/3<x, and so
every vertex in A has a neighbour in B1∪B2.
(3) x+y/(10(1−2y))<2/5* and so y≥1/4.*
Suppose that x+y/(10(1−2y))≥2/5.
Without loss of generality, we may assume that at least ∣A∣/2 vertices in A have a neighbour in B1. Choose w∈C1.
Since
∣NA2(w)∣<(2/5)∣A∣, there are at least ∣A∣/10 vertices u∈A∖NA2(w) that have a neighbour in B1.
For each such u, ∣NC2(u)∩C1∣≥y∣C∣, and since ∣C1∣≤(1−2y)∣C∣ (because ∣C2∣,∣C3∣≥y∣C∣), it follows that
∣NC2(u)∩C1∣≥(y/(1−2y))∣C1∣. Consequently there exists w′∈C1 such that NA2(w′) contains at least
(y/(10(1−2y)))∣A∣ vertices in A∖NA2(w). Since w,w′ have a common neighbour, it follows that ∣NA2(w)∩NA2(w′)∣≥x∣A∣, and so
[TABLE]
a contradiction. Thus x+y/(10(1−2y))<2/5, and so y≥1/4 from the hypothesis. This proves (3).
Since
(b1−y)+(b2−y)+(b3−y)=1−3y<x,
it follows that for every vertex u∈A, there exists i∈{1,2,3} such that ∣N(u)∩Bi∣≥(bi−y)∣B∣;
and consequently there is a partition A=A1∪A2∪A3 such that
for i=1,2,3, every vertex in Ai has more than (bi−y)∣B∣ neighbours in Bi. It follows that
Ai⊆NA2(w) for each w∈Ci.
Let ∣Ai∣=ai∣A∣ for i=1,2,3.
For i=1,2, let Di be the set of vertices in A3 with a neighbour in Bi, and
let di=∣Di∣/∣A∣. For i=1,2, if u∈Di then
[TABLE]
and so there exists w∈Ci such that
[TABLE]
and since Ai⊆NA2(w), it follows that
(y/(1−2y))di+ai<2/5. Since d1+d2≥a3, and a1+a2=1−a3, summing for i=1,2 yields
that
[TABLE]
that is, (1−3y)a3/(1−2y)>1/5; and since a3<2/5, this implies that y<1/4, a contradiction.
This proves 10.2.
10.3
If x,y∈(0,1] and 12x2y≥5(1−x−y)2, then ϕ(x,y)≥2/5.
Proof. Suppose not. Then ϕ(x,y)=1−(3/5+ϵ) for some ϵ>0, so by rotating we have ϕ(x,3/5+ϵ)≤1−y.
But 5.7 gives ϕ(x,3/5+ϵ)>1−y, a contradiction. This proves 10.3.
10.4
If x,y∈(0,1] and x>1/3 and y≥(5−3)/11, then ϕ(x,y)≥2/5.
Proof. Suppose not. Then ϕ(x,y)=1−(3/5+ϵ) for some ϵ>0, so by rotating we have
ϕ(3/5+ϵ,y)≤1−x<2/3. But 3/5≥(1−y)2/(1−2y2), since y≥(5−3)/11; and so 7.9
gives that
ϕ(3/5+ϵ,y)≥2/3, a contradiction. This proves 10.4.
10.5
If x,y∈(0,1], and either 5x/2+y≤1 and 2y≤x, or x+5y/2≤1 and 2x≤y, then ψ(x,y)<2/5.
If x,y∈(0,1] with y≤1/3 and x/(1−2x)+y/(1−3y)≤2, then ϕ(x,y)<2/5.
Proof. Apply 3.1 (with z slightly less than 2/3) to 7.10. This proves 10.6.
11 The 3/5 Level
Next, we analyze when ψ≥3/5, and similarly for ϕ. The results are shown in figure 10.
11.1
If x,y∈(0,1] with y>1/2, x≥1/5, and
[TABLE]
then ψ(x,y)≥3/5.
Proof. Suppose not, and let G be (x,y)-biconstrained via (A,B,C), such that ∣NA2(w)∣<3∣A∣/5 for all w∈C. We can
assume that
x,y are rational, and by multiplying vertices if necessary, we can assume that both x∣B∣ and ∣C∣/5 are integers. By averaging,
there exists u∈A such that ∣NC2(u)∣<3∣C∣/5. Choose B′⊆N(u) with ∣B′∣=x∣B∣, and choose
C′⊆C with NC2(u)⊆C′ and ∣C′∣=3∣C∣/5.
(1) There exist w1∈C′ and w2∈C∖C′ such that ∣N(w1)∪N(w2)∣>(1−x)∣B∣.
Choose w1∈C′ and w2∈C∖C′ uniformly and independently at random, and let
S=N(w1)∪N(w2). We will compute the expectation of ∣S∣.
Let t=2/5. For each v∈B1, v has at least y∣C∣ neighbours in C′, so the probability it is in S is at least
y/(1−t). For each v∈B∖B′, define d(v)=∣N(v)∩(C∖C1)∣/∣C∣. Then the probability that v
is a neighbour of w2 is d(v)/t, and the probability that v is a neighbour of w1 and not a neighbour of w2
is at least
[TABLE]
Thus the
expectation of ∣S∣ is at least
[TABLE]
Choose q with
[TABLE]
Each vertex w∈C∖C′ has at least y∣B∣ neighbours in B∖B′, so it follows that
q≥y. Thus, the expectation of ∣S∣ is at least
[TABLE]
Since ∑v∈B∖B′d(v)=qt∣B∣ and ∣B∖B′∣=(1−x)∣B∣, it follows by Cauchy-Schwarz that
[TABLE]
Thus the expectation of ∣S∣ is at least
[TABLE]
To prove (1), it suffices to show that the expectation of ∣S∣ is more than (1−x)∣B∣, and so it suffices to show
that
[TABLE]
Remembering that t=2/5, this is
[TABLE]
and the derivative of the left-hand side with respect to q is
[TABLE]
for q≥y. It follows that the left-hand side is minimized when q=y, so it suffices to show
that
[TABLE]
which is equivalent to the hypothesis. This proves (1).
Let w1,w2 be as in (1), and let Ai=NA2(wi) for i=1,2. Then A=A1∪A2, and since y>1/2
we have N(v1)∩N(v2)=∅, and consequently ∣A1∩A2∣≥x∣A∣≥∣A∣/5. Then
[TABLE]
and so we have ∣Ai∣≥3∣A∣/5 for some i, a contradiction. This proves 11.1.
11.2
If x,y∈(0,1] with x>1/2 and x+2y>1, then ψ(x,y)≥3/5.
Proof. Let G be (x,y)-biconstrained via (A,B,C), and suppose that ∣NA2(w)∣<(3/5)∣A∣ for each w∈C.
Choose α maximum such that there exist w1,…,wα∈C where N(wi)∩N(wj)=∅ for 1≤i<j≤α.
(1) α=2.
Since x+2y>1 it follows that NA2(wi)∪NA2(wj)=A for all distinct i,j∈{1,…,α, and so if α≥3
then every vertex in A belongs to at least two of the sets NA2(w1),NA2(w2),NA2(w3), which is impossible since they each
have cardinality less than (3/5)∣A∣. So α≤2.
Suppose that α=1; then
every w2∈C satisfies N(w1)∩N(w2)=∅. Let B1=N(w1) and A1=NA2(w1); thus ∣A1∣<(3/5)∣A∣ and ∣B1∣≥y∣B∣. Choose v∈B∖B1 with v at least x∣A∖A1∣/(1−y)>2x∣A∣/(5(1−y)) neighbours in
A∖A1, and let w2∈C be a neighbour of v. Then
[TABLE]
since x>1/2, a contradiction. This proves (1).
Since α=2, every vertex w∈C shares a neighbour with at least one of w1,w2. Let Ai=NA2(wi)
for i=1,2. Since x+2y>1, we have A=A1∪A2, and so ∣A1∩A2∣<∣A∣/5 because ∣A1∣,∣A2∣<3∣A∣/5.
Then, if some w∈C shares a neighbour with w1 and shares a neighbour with w2, it follows that
∣NA2(w)∣>2x∣A∣−∣A∣/5>4∣A∣/5, a contradiction.
Thus, every vertex in C shares a neighbour with exactly one of w1 and w2. Let H be the bipartite graph G[B∪C].
It follows that there are exactly two components of H, say H1,H2, where wi∈V(Hi) for i=1,2.
Let Bi=B∩Hi and Ci=C∩Hi for i=1,2. Without loss of generality we may assume that ∣B1∣≥∣B∣/2.
It follows that for each u∈A, u has a neighbour in B1
and consequently
[TABLE]
because ∣C2∣≥y∣C∣, and thus ∣C1∣≤(1−y)∣C∣.
Since ∣A∖A1∣>(2/5)∣A∣, there exists w∈C1 with more than 2∣A∣y/(5(1−y)) neighbours in A∖A1.
But w and w1 share a common neighbour, so
[TABLE]
since the last inequality is equivalent to 5y2−5y+1≥0, which is true because y≤1/4 (since x+2y>1, and x>1/2).
This proves 11.2.
11.3
If x,y∈(0,1] with y>1/2 and 40x2y≥(3−5x)2, then ϕ(x,y)≥3/5.
If x,y∈(0,1] with x≤4/7 and y≤1/2 and x+3y≤1, then ψ(x,y)<3/5.
Proof. We may assume that x>1/2 since ψ(1/2,1/2)<3/5; and so y<1/6 since x+3y≤1.
The claim follows from applying 3.4 with z slightly less than 3/5 and x′=y′=z′=1/4.
This proves 11.4.
11.5
If x,y∈(0,1], such that 3x+y≤1, and x+5y≤3, with strict inequality in both if x or y is irrational, then
ψ(x,y)<3/5.
Proof. By increasing x,y if necessary, we may assume that x,y are rational. Suppose that ψ(x,y)≥3/5. We claim first that:
(1) x<1/6, and y>1/2, and 5xy+15y<9, and x<(1−y)/(5y), and x≤3(1−y)2/(1+5y).
Since 3x+y≤1, it follows that x<1/3, and y>1/2 since ψ(1/2,1/2)=1/2<3/5. Thus x<1/6, since
3x+y≤1. This proves the first two statements.
Since x+5y≤3, it follows that y<3/5, and so 5xy+15y<3x+15y≤9. This proves the third statement.
For the fourth, 5x<3x/y (since y<3/5), and 3x≤1−y, and so 5x<(1−y)/y.
Finally, for the fifth statement, if y≤4/7, then 1+5y≤9−9y, and so
[TABLE]
and if y≥4/7, then
[TABLE]
and so
[TABLE]
This proves (1).
Since x<1/6, it follows that x/(1−x)<(5x)/3 and (1−y)/(3y)<1/3.
The hypotheses (via (1)) imply that
[TABLE]
and
[TABLE]
Consequently there exists a rational x′ with x/(1−x)<5x/3<x′, and
[TABLE]
Thus
[TABLE]
choose a rational y′ between them. Then x′+3y′≤1 and 3x′+y′≤1, and so ψ(x′,y′)<1/3, by (theorem 3.3 of the paper).
Let ψ(x′,y′)=z′<1/3, and choose z<3/5 with (1−z)/z≤1−z′, and (1−z)/(1−x)≤1−z′, and z≥x/x′.
Then from 3.5, ψ(x,y)≤z<3/5, a contradiction. This proves 11.5.
11.6
If x,y∈(0,1] with y<1/3 and 2−3x2x−1+1−3yy≤1, and with strict inequality if x or y is irrational,
then ϕ(x,y)<3/5.
Proof. We may assume that x,y are rational. Let x′=2−1/x and y′=y/(1−y); it follows that x′,y′<1/2, and
1−2x′x′+1−2y′y′≤1. By 3.2 for k=2, it follows that
ϕ(x′,y′)<1/3. Then applying 3.7 with z slightly less than 3/5 gives the result.
This proves 11.6.
12 Peaceful coexistence
We have not been able to evaluate ϕ(x,y) in general, but here is an easier question (that we also cannot do, but
it seems to be less far out of reach).
It is always true that ϕ(x,y)≥x, by 1.2, but if y is sufficiently small then equality may hold.
For fixed x, what is the largest y such that ϕ(x,y)=x?
Let (G,w) be a weighted graph. We say it is x-regular via a bipartition (A,B)
if
•
∣A∣=∣B∣, and w(v)>0 for each v∈V(G);
•
the 0,1-adjacent matrix between A and B is nonsingular;
•
∑u∈Aw(u)=∑v∈Bw(v)=1; and
•
for each u∈V(G), ∑v∈N(u)w(v)=x.
(Note that the fourth bullet is required to hold both for u∈A and for u∈B.)
Its order is ∣A∣, and its min-weight is minv∈Bw(v).
We will show:
12.1
For x,y∈(0,1], ϕ(x,y)=x if and only if there is an x-regular bipartite weighted graph with order at
most 1/y.
Proof. If there is a such a weighted graph (G,w), via (A,B), where ∣A∣=∣B∣=n say, let C be a set of n new vertices,
and add a perfect matching between B and C. Extend w to C by defining w(v)=1/n for each v∈C.
The weighted graph just made is (x,1/n)-constrained, and shows that ϕ(x,1/n)≤x, and consequently ϕ(x,y)≤x
(and so ϕ(x,y)=x).
For the converse, suppose that G is (x,y)-constrained via (A,B,C), and ∣NA2(v)∣≤x∣A∣ for each v∈C.
(1) Each vertex in A has exactly x∣B∣ neighbours in B, and each vertex in B has exactly
x∣A∣ neighbours in A.
Each vertex u∈B has at most x∣A∣ neighbours in A, since u has a neighbour v∈C and
∣NA2(v)∣≤x∣A∣. Since each vertex in A has at least x∣B∣ neighbours in B, averaging shows that equality holds
throughout. That proves (1).
Say two vertices in A are twins if they have the same neighbour set in B, and two vertices in B are
twins if they have the same neighbour set in A. This defines equivalence relations of A and B, and we call
the equivalence classes twin classes.
(2) For each vertex v∈C, all its neighbours in B are twins, and so N(v) is a subset of a twin class of B.
By (1) each vertex in N(v) has x∣A∣ neighbours in A, and all these vertices belong to NA2(v);
and since ∣NA2(v)∣=x∣A∣, equality holds, and in particular, all vertices in N(v) are twins.
This proves (2).
Let T be the set of all twins classes of B.
For each T∈T, let C(T) be the set of all v∈C with N(v)⊆T. Thus the sets C(T)(T∈T)
are nonempty, pairwise disjoint and have union C. There is one of cardinality at most ∣C∣/∣T∣, say
C(T); and then each vertex in T has only at most ∣C∣/∣T∣ neighbours in C, and so y≤1/∣T∣.
Choose one vertex from each twin class of A and of B, and let H be the subgraph induced on this set.
For each vertex v of H, let w(v)=∣T∣/∣B∣ if v∈T for some twin class T of B, and w(v)=∣T∣/∣A∣
if v∈T for some twin class T of A. Then we have:
•
(H,w) is a bipartite graph, with bipartition (A0,B0) say;
•
∑u∈A0w(u)=∑v∈B0w(v)=1;
•
for each u∈V(H), ∑v∈N(u)w(v)=x; and
•
∣B0∣≤1/y.
Let us choose a weighted graph (H,w) and bipartition with these properties, with ∣V(H)∣ minimum.
If there is a function f:A→R such that ∑u∈N(v)f(u)=0 for each v∈B, not identically zero,
then by adding a suitable multiple of f to the restriction of w to A, we can arrange that w(u)=0 for some u∈A,
and then u can be deleted, contrary to the minimality of ∣V(H)∣. Thus there is no such f, and similarly there
is no f:B→R such that ∑v∈N(u)f(v)=0 for each u∈A, not identically zero. Consequently
∣A0∣=∣B0∣=n say, and the adjacency matrix between A0 and B0 is nonsingular. Moreover w(v)>0
for each v∈V(H), from the minimality of V(H). This proves 12.1.
By 2.3, ϕ(x,y)=x if and only ϕ(y,x)=x, so this also answers the analogous question for ϕ(y,x).
If x is irrational, there is no x-regular bipartite weighted graph, and so ϕ(x,y)>x for all y>0.
If x∈(0,1] is rational,
let us define the order of x∈(0,1] to be the minimum order of x-regular bipartite weighted graphs. If x=p/q
say where p,q>0 are integers, then the order of x is at most q,
because one can construct an appropriate cyclic
shift graph.
But the order of x can be strictly less than q. For instance, the top part of the graph of figure
1 is 13/27-regular (take as vertex-weights the numbers given, divided by 27), and so the order
of 13/27
is at most seven.
Figure 11 gives a smaller example, showing that the order of 2/5 is at most four.
We can prove that the order is also bounded below by a function of q that goes to infinity with q.
More exactly, if G is p/q-regular (in lowest terms) and has order n, then q is at most (n+1)(n+1)/2.
This follows from a theorem of Hadamard [2], that
every n×n0,1-matrix has determinant at most (n+1)(n+1)/22−n.
We do not know whether
there are weighted bipartite graphs with order n that are p/q-regular (in lowest terms), where q is exponentially large in n.
(Hadamard n×n0,1-matrices have determinant that achieve Hadamard’s bound, and they exist when n+1 is a power of two,
but they give
weighted bipartite graphs that are vertex-transitive, and which therefore are p/q-regular with q=n.)
One could ask the same question for the biconstrained problem: given x, for which values of y is it true that
ψ(x,y)=x? A similar analysis (we omit the details) shows:
12.2
For x,y∈(0,1], the following are equivalent:
•
ψ(x,y)=x;
•
ψ(y,x)=x; and
•
there is an x-regular bipartite weighted graph with min-weight at
least y.
Bibliography5
The reference list from the paper itself. Each links out to its DOI / PubMed record.
1[1] P. Erdős, M. Saks, and V. Sós, “Maximum induced trees in graphs”, J. Combinatorial Theory, Ser. B , 41 (1986), 61–79.
2[2] J. Hadamard, “Résolution d’une question relative aux déterminants”, Bulletin des Sciences Mathématiques , 17 (1893), 240–246.
3[3] P. Hompe, The Girth of Digraphs and Concatenating Bipartite Graphs , Senior Thesis, Princeton, 2019.
4[4] P. Hompe, “Some results on concatenating bipartite graphs”, ar Xiv:1908.07453 .
5[5] M. Kneser, “Abschätzungen der asymptotischen Dichte von Summenmengen”, Math. Z. (in German), 58 (1953), 459–484.