Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
Source. Alon, Theorem 1.1, statement on paper p. 1 and proof on
pp. 3-4 (PDF pp. 1, 3-4).
Statement
For a graph G, let b(G) be the maximum number of edges in a bipartite
subgraph, and define
F(e)=∣E(G)∣=eminb(G).
There are constants c>0 and n0 such that, for every even integer
n>n0 and e=n2/2,
F(e)≥2e+8e+ce1/4.
Equivalently, for every sufficiently large positive integer m, every graph
with 2m2 edges has a bipartite subgraph with at least
m2+m/2+c′m edges, for another absolute constant c′>0.
Dependencies and notation
The proof uses
Lemma 2.1 and the
Edwards lower bound
[[extremal_graph_theory/edwards_1973_extremal_properties_bipartite_subgraphs/theorem_12|Theorem
12]]. For disjoint vertex sets A,B, write e(A) for the number of edges
inside A and e(A,B) for the number between them. A maximum bipartite
subgraph may be taken to be a cut: the two classes of any bipartite subgraph
define a cut of the ambient graph containing all its edges.
Rewritten proof
Fix a sufficiently small constant ε>0, say
ε=1/100. Let n be a sufficiently large even integer and let
G have e=n2/2 edges.
Case 1: the chromatic number is below the threshold
Suppose G is 2s-colorable for an integer s satisfying
Now suppose χ(G)≥n−εn. Write
χ(G)=n−k. In a proper coloring with the minimum number of colors,
every pair of color classes has an edge between them, so
(2χ(G))≤e<(2n+1). Hence
0≤k≤εn. Take a vertex-critical
(n−k)-chromatic subgraph H. Every vertex of H has degree at least
n−k−1. Therefore, writing h=∣V(H)∣,
h(n−k−1)≤2∣E(H)∣≤n2,
and hence h≤n+2εn once n is large.
Color H properly with n−k colors. If a color classes are
singletons, then
h≥a+2(n−k−a)=2(n−k)−a,
so a≥n−4εn. Any two classes in a coloring using the
minimum number of colors have an edge between them; otherwise they could be
merged. The singleton classes consequently form a clique. Thus G contains
a clique U on n−r vertices for some
0≤r≤4εn. Put W=V(G)∖U and
q=n−r. Direct calculation gives
e(U)=(2q)=2n2−2(2r+1)n+2r(r+1)
and
D:=e(W)+e(U,W)=r(n−r)+2n+2r2−r.(5)
The last expression is rq+A, where A=(n+r2−r)/2. With
ε=1/100 and n large, ∣A−n/2∣≤n/100 and
q=n−r≥99n/100. Thus both A and q−A exceed n/4, so the distance
from A to every multiple of q is at least n/4. The same is true of
D; in particular, D≥n/4.
Subcase 2a: e(W)≥n/32
Apply the Edwards bound inside W to obtain a partition W=W1⊔W2
with
e(W1,W2)≥2e(W)+8e(W)+O(1)≥2e(W)+16n+O(1).
Balance the complete graph on U into U1⊔U2. Its cut has
⌊q2/4⌋ edges, and therefore
There are two ways to align the cuts of U and W. The numbers of
U-W edges crossing in the two alignments sum to e(U,W), so one
alignment keeps at least half of them. For that alignment,
By (5), e(U,W)=D−e(W) has distance at least n/5 from every multiple
of q. List U as v1,…,vq so that their numbers of neighbors in
W satisfy d1≤⋯≤dq, and put
U1={v1,…,v⌊q/2⌋},U2=U∖U1.
Let dˉ=e(U,W)/q and δ=n/(5q). The distance from dˉ to
every integer is at least δ; also dˉ≥δ. If
d⌊q/2⌋≥dˉ, integrality gives
d⌊q/2⌋≥dˉ+δ. Every vertex of U2 has at
least that many neighbors in W, and hence
e(U2,W)≥2e(U,W)+10n.
Otherwise every vertex of U1 has at most
dˉ−δ neighbors in W. Since dˉ≥δ, the inequality
$\lfloor q/2\rfloor(\bar d-\delta)\leq
q(\bar d-\delta)/2$ gives the same conclusion after subtracting
e(U1,W) from e(U,W). This also handles odd q.
Use the cut (U1∪W,U2). Since the complete graph on U contributes
⌊q2/4⌋ edges,
This is stronger than the claimed ce1/4 improvement once n is large.
The two cases complete the proof.
Consequence for Problem 127
For e=n2/2,
88e+1−1=8e+O(1).
The theorem therefore makes the integral correction above the exact Edwards
baseline at least ce1/4−O(1) along the infinite sequence of even n.
It tends to infinity.