Inequalities for doubly nonnegative functions

Alexander Sidorenko
Rényi Institute
Budapest, Hungary
[email protected]

Submitted: Aug 16, 2019; Accepted: Jan 31, 2021; Published: Feb 12, 2021
© The author. Released under the CC BY-ND license (International 4.0).

Abstract
LetRg be a bounded symmetric measurable nonnegative function on [0, 1]2 , and
∥g∥ = [0,1]2 g(x, y)dxdy. For a graph G with vertices {v1 , v2 , . . . , vn } and edge set
E(G), we define
Z Y
t(G, g) = g(xi , xj ) dx1 dx2 · · · dxn .
[0,1]n {v ,v }∈E(G)
i j

We conjecture that t(G, g) ⩾ ∥g∥|E(G)| holds for any graph G and any function g with
nonnegative spectrum. We prove this conjecture for various graphs G, including
complete graphs, unicyclic and bicyclic graphs, as well as graphs with 5 vertices or
less.
Mathematics Subject Classifications: 05C35, 05C22, 26D20

1 Introduction
Let µ be the Lebesgue measure on [0, 1]. Let H denote the space of bounded measurable
real functions on [0, 1]2 , and G ⊂ H denote the subspace of symmetric functions. Let H+
and G+ denote the subsets of nonnegative functions in H and G, respectively.
Let G be a simple graph with vertices {v1 , v2 , . . . , vn } and edge set E(G). We would
like to know what conditions on G and g ∈ G+ guarantee that
Z Y Z |E(G)|
def n 2
t(G, g) = g(xi , xj ) dµ ⩾ g dµ . (1.1)
[0,1]n {v ,v }∈E(G) [0,1]2
i j

One approach is to ask what graphs G satisfy (1.1) for every function g ∈ G+ . It is easy
to show that such graphs can not have odd cycles, so only graphs with chromatic number
2 are suitable candidates. It led to

the electronic journal of combinatorics 28(1) (2021), #P1.32 https://doi.org/10.37236/8947

Conjecture 1.1 ([21, 22]). Let H be a bipartite graph with two vertex sets V = {v1 , v2 ,
. . . , vn }, W = {w1 , w2 , . . . , wm } and edge set E(H) ⊆ V × W . Then for any function
h ∈ H+ (not necessarily symmetric)
Z Y Z |E(H)|
def n+m 2
t(H, h) = h(xi , yj ) dµ ⩾ h dµ . (1.2)
[0,1]n+m (v ,w )∈E(H) [0,1]2
i j

We discuss Conjecture 1.1 in Section 3.

For a (simple or bipartite) graph G, let E(G) denote its edge set, and e(G) = |E(G)|.
For a simple graph G, let V (G) denote its vertex set, and v(G) = |V (G)|.
The 1-subdivision of a simple graph G is a bipartite graph H = Sub(G) with vertex
sets V (G) and E(G), where v ∈ V (G) and e ∈ E(G) form an edge in H if v ∈ e in G.
We call a bipartite graph H symmetric if it has an automorphism ϕ which switches
its vertex-sets V and W : ϕ(V ) = W , ϕ(W ) = V .
We callR a function g ∈ G+ doubly nonnegative if there is a function h ∈ H such that
g(x, y) = [0,1] h(x, z)h(y, z)dµ(z). Equivalently, a doubly nonnegative function is a non-
negative symmetric function with nonnegative spectrum. We callRa function g ∈ G+ com-
pletely positive if there is a function h ∈ H+ such that g(x, y) = [0,1] h(x, z)h(y, z)dµ(z).
The terms “doubly nonnegative” and “completely positive” come from matrix theory;
there exist functions which are doubly nonnegative but not completely positive (see Sec-
tion 2).
In this article, we study two problems: (a) what functions g ∈ G+ satisfy t(G, g) ⩾
∥g∥e(G) for all simple graphs G (we call such functions nice); and (b) what graphs G satisfy
the same inequality for any doubly nonnegative function g (we call such graphs good).
If G is good, then Conjecture 1.1 holds for H = Sub(G). We show in Section 3 that
for a fixed G, inequality (1.1) holds for any completely positive function g if and only if
Conjecture 1.1 holds for H = Sub(G). Thus, it is reasonable to expect that all completely
positive functions are nice.
Conjecture 1.2. All doubly nonnegative functions are nice. All simple graphs are good.
Our Theorem 2.1 demonstrates that there are nice functions which are not doubly
nonnegative.
If chromatic number χ(G) = 2, then goodness of G should follow from Conjecture 1.1.
In Sections 4-7, we give examples of good graphs G with χ(G) ⩾ 3. In particular, we
prove that complete graphs, graphs whose complements consist of disjoint edges, unicyclic
and bicyclic graphs, generalized theta graphs, and graphs with ⩽ 5 vertices are all good.
In Section 6, we consider a strengthened variant of inequality (1.1) which in many
instances is easier to prove than the original one. We say that a simple graph G is extra-
good if for any doubly nonnegative function g and any bounded measurable nonnegative
functions f1 , . . . , fv(G) on [0, 1],
Z v(G) Z e(G)
Y Y
v(G) 2
g(xi , xj ) fi (xi ) dµ ⩾ f (x)g(x, y)f (y) dµ , (1.3)
[0,1]v(G) {v ,v }∈E(G) i=1 [0,1]2
i j

the electronic journal of combinatorics 28(1) (2021), #P1.32 2

 Q 1/(2e(G))
v(G)
where f (x) = fi (x)
i=1 . Obviously, if G is extra-good, then G is good. For
each graph that we proved to be good, we also were able to prove that it is extra-good.
It is possible that every good graph is extra-good.
A connection to Kohayakawa–Nagle–Rödl-Schacht conjecture is discussed in Section 8.
In Section 9, we discuss generalizations of inequality (1.1) for bounded measurable
nonnegative symmetric functions of r ⩾ 3 variables.

2 Doubly nonnegative and completely positive matrices

A doubly nonnegative matrix is a real positive semidefinite square matrix with nonnega-
tive entries. A completely positive matrix is a doubly nonnegative matrix which can be
factorized as A = BB T where B is a nonnegative (not necessarily square) matrix. It is
well known (see [3]) that for any k ⩾ 5 there exist doubly nonnegative k × k matrices
which are not completely positive.
For a k × k matrix A = [aij ], we define a function gA on [0, 1]2 as g(x, y) = aij for
(i − 1)/k < x ⩽ i/k, (j − 1)/k < y ⩽ j/k, and g(x, y) = 0 if xy = 0. Obviously, gA is a
doubly nonnegative (completely positive) function if and only if A is a doubly nonnegative
(completely positive) matrix.
If A(H) is the adjacency matrix of a k-vertex simple graph H, then the number of
homomorphisms from G to H is equal to t(G, gA(H) )k v(G) .
Notice that if a nonzero k × k matrix A has zero diagonal, then gA is not nice, since
t(G, gA ) = 0 for any graph G with chromatic number χ(G) > k. We are going to
demonstrate that presence of a single positive diagonal entry can be sufficient to make gA
nice.
Theorem 2.1. Let P be a symmetric permutation matrix of order k with a ⩾ 1 diagonal
entries equal to 1, and b ⩾ 1 pairs of off-diagonal entries equal to 1 (a + 2b = k). Then
gP , while not being positive semidefinite, is a nice function.
Proof. P has eigenvalues 1 with multiplicity a+b, and −1 with multiplicity b ⩾ 1. There-
fore, P is not positive semidefinite. If graph G has connected components G1 , G2 , . . . , Gm
then t(G, g) = m
Q
i=1 t(G i , g). Hence, to prove (1.1) for g = gP it is sufficient to consider
connected graphs G. If G is a tree, then validity of (1.1) follows from (1.2) (Conjecture 1.1
has been proved for trees by various authors; a short proof can be found in [11]). Hence,
we may assume that G is not a tree. If n = v(G), then e(G) ⩾ n. R As P has2 a ⩾ 1 diagonal
n e(G)
entries equal to 1, we get t(G, gP ) ⩾ a (1/k) ⩾ (1/k) and [0,1]2 gP dµ = 1/k.

3 More on Conjecture 1.1

The earliest known works where inequalities of type (1.1) and (1.2) appear are [18] and [2].
In 1959, Mulholland and Smith [18] proved that for any symmetric nonnegative matrix A
and any nonnegative vector z,
(zTAk z) · (zT z)k−1 ⩾ (zTAz)k , (3.1)

the electronic journal of combinatorics 28(1) (2021), #P1.32 3

where equality takes place if and only if z is an eigenvector of A or a zero vector. Note
that (3.1) is a particular case of (1.1) where H is the k-edge path Pk .
Almost at the same time, Atkinson, Watterson and Moran [2] proved that nm ·
s(AATA) ⩾ s(A)3 , where A is an (asymmetric) nonnegative (n × m)-matrix, and s(A)
is the sum of entries of A. They presented their inequality in both matrix and integral
form, and conjectured validity of (1.2) for H = Pk with k ⩾ 3.
In 1965, Blakley and Roy [4], being unaware of the article [18], rediscovered (3.1).
Lately, Conjecture 1.1 has been proved for various bipartite graphs (see [6, 7, 8, 9, 12,
15, 17, 19, 21, 22, 23, 24, 25]), among them: trees, complete bipartite graphs, and graphs
with 9 vertices or less. Some of the authors restricted (1.2) to symmetric functions h.
Nevertheless, the proofs of their results can be extended to asymmetric h as well. Let
S be the class of bipartite graphs that satisfy Conjecture 1.1, and S∗ be the class of
bipartite graphs H that satisfy (1.2) for all h ∈ G+ . Obviously, S ⊆ S∗ . It would be nice
to prove S = S∗ .

Theorem 3.1. If H ∈ S∗ is symmetric, then H ∈ S.

In the proof of Theorem 3.1, we will use the so called “tensor-trick” lemma.

Lemma 3.2 ([21, 22]). If there exists a constant c = cH > 0 such that for any h ∈ H+ ,
R e(H)
t(H, h) ⩾ c · [0,1]2 hdµ2 , then H ∈ S.

Proof of Theorem 3.1. It is sufficient to consider the case when H is connected. Denote
by n the size of each vertex set of H, so the total number of vertices is 2n. Let h ∈ H+ .
Define its “transpose” hT as hT (x, y) = h(y, x). As H is symmetric, t(H, h) = t(H, hT ).
Define symmetric function h̃ ∈ G+ as follows:


0 if 0 ⩽ x, y < 1/2;

h(2x, 2y − 1) if 0 ⩽ x < 1/2 ⩽ y ⩽ 1;
h̃(x, y) =


h(2y, 2x − 1) if 0 ⩽ y < 1/2 ⩽ x ⩽ 1;

0 if 1/2 ⩽ x, y ⩽ 1.
R R
Notice that [0,1]2
h̃dµ2 = (1/2) h[0,1]2 dµ2 . As H is connected,

t(H, h̃) = 2−2n (t(H, h) + t(H, hT )) = 21−2n t(H, h).

R e(H)
Since H ∈ S∗ , we get t(H, h̃) ⩾ [0,1]2
h̃dµ2 . Hence,

Z e(H)
2n−1−e(H) 2
t(H, h) ⩾ 2 hdµ ,
[0,1]2

and by Lemma 3.2, H ∈ S.

the electronic journal of combinatorics 28(1) (2021), #P1.32 4

Remark 3.3. It is a classical fact that there exists a measure preserving bijection between
any two atomless measure spaces with total measure 1. In particular, if µ1 and µ2 are
atomless measures on [0, 1], and a bipartite graph H ∈ S has vertex sets of sizes n and
m, then for any bounded non-negative function h on [0, 1]2 , measurable with respect to
µ1 ⊗ µ2 ,
Z Y Z e(H)
n m
h(xi , yj ) dµ1 dµ2 ⩾ h dµ1 dµ2 .
[0,1]n+m (v ,w )∈E(H) [0,1]2
i j

Theorem 3.4. If Conjecture 1.1 holds for a bipartite graph H, and all vertices from
the first vertex set of H have the same degree a, then t(H, h) ⩾ t(K1,a , h)e(H)/a for any
h ∈ H+ . If all vertices from the second vertex set of H have the same degree b, then
t(H, h) ⩾ t(Kb,1 , h)e(H)/b for any h ∈ H+ .
Proof. We will prove the first part of the statement (the proof of the second part is
similar). Notice that e(H) = na. It is sufficient to consider R functions h ∈ H+ that
are separated from zero: inf [0,1]2 h > 0. Denote φ(x) = [0,1] h(x, y)dµ(y). Then c =
R
[0,1]
φ(x)a dµ > 0, and f (x) = φ(x)a /c is positive and bounded on [0, 1]. Consider a mea-
h(x, y) = h(x, y)f (x)−1/a .
sure µ∗ on [0, 1] defined by dµ∗ = f dµ, so µ∗ ([0, 1]) = 1. Denote b
Clearly, b h is bounded and measurable with respect to µ∗ ⊗ µ. By Remark 3.3,
 1/n
Z Y
t(H, h)1/n =  h(xi , yj )dµn dµm 
[0,1]n+m (vi ,wj )∈E(H)
 1/n
Z Y
= h(xi , yj )dµn∗ dµm 
b
[0,1]n+m (v ,w )∈E(H)
i j
Z a
⩾ h(x, y)dµ∗ (x)dµ(y)
b
[0,1]2
Z a
−1/a
= h(x, y)f (x) dµ∗ (x)dµ(y)
[0,1]2
Z a
1−(1/a) 2
= h(x, y)f (x) dµ
[0,1]2
Z a
1−(1/a)
= φ(x)f (x) dµ
[0,1]
Z a Z
1−a a
=c φ(x) dµ = φ(x)a dµ = t (K1,a , h) .
[0,1] [0,1]

Theorem 3.5. For a fixed graph G, inequality (1.1) holds for any completely positive
function g if and only if Conjecture 1.1 holds for H = Sub(G).
Proof. Suppose that Conjecture 1.1 holds for H =R Sub(G), and a function g is completely
positive. There exists h ∈ H+ such that g(x, y) = [0,1] h(x, z)h(y, z)dµ(z). Then t(G, g) =

the electronic journal of combinatorics 28(1) (2021), #P1.32 5

t(H, h). Every vertex in the second vertex set of H has degree 2. ByR Theorem 3.4, we
have t(H, h) ⩾ t(K2,1 , h)e(H)/2 . As e(G) = e(H)/2 and t(K2,1 , h) = [0,1]2 gdµ2 , we get
(1.1).
Now suppose R that (1.1) holds for any completely positive functionRg. Let h ∈ H+ .
Set g(x, y) = [0,1] h(x, z)h(y, z)dµ(z). Then t(H, h) = t(G, g) ⩾ ( [0,1]2 gdµ2 )e(G) ⩾
R R
( [0,1]2 hdµ2 )2e(G) = ( [0,1]2 hdµ2 )e(H) .

4 Subdivisions that are norming

We say that a bipartite graph H with vertex sets V = {v1 , v2 , . . . , vn } and W = {w1 , w2 ,
. . . , wm } has the Hölder property if for any assignment f : E(H) → H,
 e(H)
Z Y Y
 fe (xi , yj ) dµn+m  ⩽ t(H, fe ). (4.1)
[0,1]n+m e=(v ,w )∈E(H) e∈E(H)
i j

It is known (see [10, 16]) that every graph H with the Hölder property (except a star
with even number of edges) is a norming graph: t(H, h)1/e(H) is a norm on H. Conversely,
every norming graph has the Hölder property.
Theorem 4.1. If Sub(G) has the Hölder property then G is good.
R
Proof. Let H = Sub(G). If g(x, y) = [0,1] h(x, z)h(y, z)dµ(z) then t(G, g) = t(H, h).
Select a pair of edges e′ , e′′ in H which subdivide the same edge of G. Assign fe′ = fe′′ = h
R 2e(G)
and fe = 1 for e ̸= e′ , e′′ . Then the left hand side of (4.1) is [0,1]2 gdµ2 , and the
right hand side is t(G, g)2 .
The 1-subdivision of cycle Cn is an even cycle C2n which is a norming graph. The
1-subdivision of the octahedron K2,2,2 is norming (see [8, Example 4.15]). Hence, Cn and
K2,2,2 are good graphs.
In a norming graph, the degrees of vertices are even (see Observation 2.5 in [10]).
Hence, Sub(K2r ) is not norming. While Sub(K3 ) = C6 is norming, it is not known
whether Sub(K2r+1 ) with r ⩾ 2 is norming. We will prove in the next section that all
complete graphs are good.

5 Complete graphs are good

Theorem 5.1. If graph G is good, and graph G1 is obtained from G by adding a new
vertex adjacent to all vertices of G, then G1 is good.
Corollary 5.2. The complete graphs are good.
Theorem 5.3. If graph G is good, and graph G2 is obtained from G by adding two vertices
adjacent to all vertices of G but not to each other, then G2 is good.

the electronic journal of combinatorics 28(1) (2021), #P1.32 6

Corollary 5.4. Any graph whose complement is a set of independent edges is good.
To prove Theorems
R 5.1 and 5.3 we need a couple of auxiliary results. For g ∈ G+ , we
2
denote ∥g∥ = [0,1]2 gdµ .

Lemma 5.5. If function g is doubly nonnegative, then t(K3 , g) ⩾ t(K1,2 , g)3/2 and
t(K4 − e, g) ⩾ t(K2,2 , g)5/4 .
R
Proof. As g is doubly nonnegative, g(x, y) = [0,1] h(x, z)h(y, z)dµ(z). Then t(K3 , g) =
t(C6 , h), t(K1,2 , g) = t(P4 , h), t(K2,2 , g) = t(C8 , h), where P4 denotes the 4-edge path.
As C6 is norming, and P4 is a subgraph of C6 , we have t(C6 , h)1/6 ⩾ t(P4 , h)1/4 .
By the Cauchy–Schwarz inequality, t(K4 − e, g) ∥g∥ ⩾ t(K3 , g)2 = t(C6 , h)2. As
t(C2k , h)1/2k is the (2k)-th Schatten norm of h, we get t(C6 , h)1/6 ⩾ t(C8 , h)1/8 . Hence,
t(K4 − e, g) ⩾ t(K2,2 , g)3/2 ∥g∥−1 ⩾ t(K2,2 , g)5/4 .
For n > m ⩾ 1, let Kn − Km denote the complement of Km in Kn . It follows from
Lemma 5.5 that for m = 1, 2 and any doubly nonnegative function g,
t(Km+2 − Km , g) ⩾ t(K2,m , g) · ∥g∥ . (5.1)
Thus, Theorems 5.1 and 5.3 follow from the next proposition.
Proposition 5.6. Let integer m be such that (5.1) holds for all doubly nonnegative func-
tions. If graph G is good, and graph Gm is obtained from G by adding a group of m
independent vertices that are adjacent to all vertices of G, then Gm is good.
Proof. We may assume v(G) ⩾ 2. It is sufficient to consider functions Q g that are sepa-
rated from zero: inf [0,1]2 g > 0. Then function φ(x1 , . . . , xm ) = [0,1] m
R
i=1 g(xi , y)dµ(y) is
m m
positive and bounded on [0, 1] . For each x = (x1 , . .Q . , xm ) ∈ [0, 1] , consider measure
µx on [0, 1], defined by dµx = fx dµ, where fx (y) = m −1
i=1 g(xi , y)φ(x1 , . . . , xm ) . It is
easy to see that µx ([0, 1]) = 1, and g is bounded and measurable with respect to µx ⊗ µx .
By Remark 3.3,
Z Y Z e(G)
def v(G) 2
t(G, g, µx ) = g(yi , yj ) dµx ⩾ g dµx .
[0,1]v(G) {v ,v }∈E(G) [0,1]2
i j

Hence,
Z
t(Gm , g) = t(G, g, µx1 ,...,xm )φ(x1 , . . . , xm )v(G) dµ(x1 ) · · · dµ(xm )
[0,1]m
Z Z e(G)
⩾ g dµ2x1 ,...,xm φ(x1 , . . . , xm )v(G) dµ(x1 ) · · · dµ(xm )
[0,1]m [0,1]2
Z Z m
!e(G)
Y
= g(y, z) (g(xi , y)g(xi , z)) dµ(y)dµ(z)
[0,1]m [0,1]2 i=1

× φ(x1 , . . . , xm )v(G)−2e(G) dµ(x1 ) · · · dµ(xm ).

the electronic journal of combinatorics 28(1) (2021), #P1.32 7

As 1 + e(G) − v(G)/2 ⩽ e(G), by using the Hölder inequality, we get

t(Gm , g)·t(K2,m , g)e(G)−v(G)/2

Z e(G)−v(G)/2
2
= t(Gm , g) · φ (x1 , . . . , xm ) dµ(x1 ) · · · dµ(xm )
[0,1]m
Z m
!e(G)
Y
⩾ g(y, z) (g(xi , y)g(xi , z)) dµ(y)dµ(z)dµ(x1 ) · · · dµ(xm )
[0,1]m+2 i=1
e(G)
= t(Km+2 − Km , g) .

As Conjecture 1.1 holds for complete bipartite graphs (see [21]), K2,m is good. By using
(5.1), we get

t(Gm , g) ⩾ t(K2,m , g)−e(G)+v(G)/2 t(K2,m , g)e(G) ∥g∥e(G)

= t(K2,m , g)v(G)/2 ∥g∥e(G) ⩾ ∥g∥mv(G) ∥g∥e(G) = ∥g∥e(Gm ) .

6 Extra-good graphs
Theorem 6.1. If G is vertex-transitive and good, then G is extra-good.
1/(2e(G))
Proof. We denote n = v(G), f (x) = ( ni=1 fi (x))
Q
and g̃(x, y) = f (x)g(x, y)f (y).
Notice that g̃ is doubly nonnegative. Since G is vertex-transitive, any permutation of the
functions f1 , f2 , . . . , fn does not change the value of the integral on the left hand side of
(1.3). By applying the Hölder inequality to the geometric mean of all n! possible integrals,
we get
Z Y n
Y Z Y n
Y
n
g(xi , xj ) fi (xi ) dµ ⩾ g(xi , xj ) f (xi )2e(G)/n dµn
[0,1]n {vi ,vj }∈E(G) i=1 [0,1]n {vi ,vj }∈E(G) i=1
Z Y
= g̃(xi , xj ) dµn .
[0,1]n {v ,v }∈E(G)
i j

Since G is good,
Z Y Z e(G)
n 2
g̃(xi , xj ) dµ ⩾ g̃(x, y) dµ .
[0,1]n {v ,v }∈E(G) [0,1]2
i j

Corollary 6.2. Cycles and complete graphs are extra-good.

Theorem 6.3. If vertex v is a leaf in graph G, and G − v is extra-good, then G is

extra-good too.

the electronic journal of combinatorics 28(1) (2021), #P1.32 8

Proof. Let v(G) = n + 1, V (G) = {v1 , . . . , vn , vn+1 }, and v = vn+1 is a leaf. For
a set Rof functions f1 , . . . , fn , fn+1 , define f̃i = fi for i = 1, . . . , n − 1, and f˜n (x) =
fn (x) [0,1] g(x, y) fn+1 (y) dµ(y). Then
Z Y n+1
Y Z Y n
Y
g(xi , xj ) fi (xi )dµ n+1
= g(xi , xj ) f˜i (xi )dµn
[0,1]n+1 {v ,v }∈E(G) i=1 [0,1]n {v ,v }∈E(G−v) i=1
i j i j
Z e(G)−1
⩾ f˜(x)g(x, y)f˜(y) dµ2
[0,1]2
e(G)−1
= I0 ,
Q 1/(2e(G)−2)
˜ n ˜
R
where f (x) = i=1 fi (x) . Set φ(x) = [0,1] g(x, y)fn+1 (y)dµ(y),
Z Z
−1
I1 = φ(x) g(x, y)fn+1 (y)dµ , 2
I2 = fn+1 (x)g(x, y)φ(y)−1 dµ2 .
[0,1]2 [0,1]2

φ(x)−1 φ(x)dµ = 1, and similarly, I2 = 1. By the Hölder

R
It is easy to see that I1 = [0,1]
inequality,
Z e(G)
e(G)−1 e(G)−1 1/2 1/2 2
I0 ·1·1 = I0 · I1 · I2 ⩾ f (x) g(x, y) f (x) dµ ,
[0,1]2

where
n
Y n−1
Y n+1
Y
f 2e(G) = (f˜)2e(G)−2 φ−1 fn+1 = f˜i · φ−1 fn+1 = fi · (fn φ) · φ−1 fn+1 = fi .
i=1 i=1 i=1

A connected graph G is called unicyclic if e(G) = v(G), and bicyclic if e(G) = v(G)+1.
Corollary 6.4. Trees and unicyclic graphs are extra-good.
Let P (k1 , k2 , . . . , kr ) denote a graph consisting of two vertices joined by r internally
disjoint paths of length k1 , k2 , . . . , kr . Such a graph is called theta graph (see [5]). If r = 2,
then P (k1 , k2 ) is simply a cycle of length k1 + k2 .
Theorem 6.5. P (k1 , k2 , . . . , kr ) is extra-good.
Let C(k1 , k2 , m) denote a graph which consists of two cycles of length k1 and k2
connected by a path of length m ⩾ 0 (when m = 0, the cycles share a vertex). Notice
that P (k1 , k2 , k3 ) and C(k1 , k2 , m) are the only bicyclic graphs without leaves. In view
of Theorems 6.3 and 6.5, in order to prove that all bicyclic graphs are extra-good, it is
sufficient to show that C(k1 , k2 , m) is extra-good. This will follow from the next result.
Theorem 6.6. Let graph G0 be formed by attaching a cycle to one of the vertices of graph
G. If G is extra-good, then G0 is extra-good.

the electronic journal of combinatorics 28(1) (2021), #P1.32 9

Corollary 6.7. Bicyclic graphs are extra-good.

The proofs of Theorems 6.5 and 6.6 are given in the appendix.
Consider a tree T whose vertices are arbitrarily colored in black and white so that
at least one vertex is black. Take r disjoint copies of T and glue together “sister” black
vertices from different copies. We call the resulting graph a multitree. For example
P (k, k, . . . , k) is a multitree. The case of even cycle in Theorem 6.6 is a particular case
of the following statement.

Theorem 6.8. Let graph G0 be formed by gluing a black vertex of a multitree to one of
the vertices of graph G. If G is extra-good, then G0 is extra-good.

Theorem 6.9. If graph G is extra-good, and graph G1 is obtained from G by adding a

new vertex adjacent to all vertices of G, then G1 is extra-good.

Theorem 6.10. If graph G is extra-good, and graph G2 is obtained from G by adding two
vertices adjacent to all vertices of G but not to each other, then G2 is extra-good.

Theorems 6.8, 6.9, and 6.10 are not used in the rest of the article. We omit their
proofs as they are very similar to the proofs of Theorems 6.6, 5.1, and 5.3.

7 Graphs with small number of vertices

Theorem 7.1. All graphs with 5 vertices or less are good.

Proof. If graph G has connected components G1 , G2 , . . . , Gm then t(G, g) = m

Q
i=1 t(Gi , g).
Hence, it is sufficient to consider connected graphs G only. As Conjecture 1.1 has been
proved for bipartite graphs with 9 vertices or less, it is sufficient to consider connected
simple graphs with at least one odd cycle. The results of Sections 5 and 6 cover all such
graphs with ⩽ 5 vertices. A table of all 5-vertex graphs that do not have isolated vertices
can be found in [1, Figure 6].
Some 6-vertex graphs are not covered by the results of Sections 5 and 6, but for almost
all of them we were able to prove that they are good by using the Cauchy–Schwarz and
Hölder inequalities. The only graph with 6 vertices that we were unable to prove to be
good is the complement of the 5-edge path.

8 Locally dense graphs

A simple graph H is called (ε, d)-dense if every subset X ⊆ V (H) of size |X| ⩾ ε|V (H)|
spans at least d2 |X|2 edges.

Conjecture 8.1 ([13]). For any graph G and δ, d ∈ (0, 1), there exists ε = ε(δ, d, G) such
that there are at least (de(G) − δ)v(H)v(G) homomorphisms of G into any sufficiently large
(ε, d)-dense graph H.

the electronic journal of combinatorics 28(1) (2021), #P1.32 10

 When χ(G) = 2, Conjecture 8.1 follows from Conjecture 1.1. It is known that Con-
jecture 8.1 holds for complete graphs and multipartite complete graphs (see[13, 20]).
Christian Reiher [20] proved Conjecture 8.1 for odd cycles. Joonkyung Lee [14] proved
that adding an edge to a cycle or a tree produces graphs that satisfy the conjecture. He
also proved that Conjecture 8.1 holds for a class of graphs obtained by gluing complete
multipartite graphs (or odd cycles) in a tree-like way. All graphs with 5 vertices or less
satisfy the conjecture.
For a nonnegative symmetric k×k matrix A, we define its density d(A) as the minimum
of xT Ax over all nonegative k-dimensional vectors x with the sum of entries equal to 1.
Clearly, d(A) ⩽ ∥gA ∥. We call a graph G density-friendly if t(G, gA ) ⩾ d(A)e(G) for any A.
It is easy to see that any density-friendly graph satisfies Conjecture 8.1, but the converse
is not obvious.
While Conjectures 1.2 and 8.1 looks very different, the sets of graphs which are known
to satisfy them are surprisingly similar.
One can try to build a bridge between these two topics by defining

c(A) = inf t(G, gA )1/e(G) .

G

Then Conjecture 1.2 claims that c(A) = ∥gA ∥ for any doubly nonnegative matrix A, and
Conjecture 8.1 claims that c(A) ⩾ d(A) for any nonnegative symmetric matrix A.

9 Functions of r variables
Let G be an r-uniform hypergraph with vertex set V (G) = {v1 , v2 , . . . , vn } and edge set
E(G) (edges are r-element subsets of the vertex set). Let g be Ra bounded symmetric
measurable nonnegative function defined on [0, 1]r . Denote ∥g∥ = [0,1]r gdµr and
Z Y
t(G, g) = g(xi1 , xi2 , . . . , xir ) dµn .
[0,1]n {vi1 ,vi2 ,...,vir }∈E(G)

Problem 9.1. Characterize functions g such that

t(G, g) ⩾ ∥g∥|E(G)| (9.1)

holds for every r-uniform hypergraph G.

When r = 1, it is obvious that (9.1) holds for any nonnegative function h on [0, 1].
The incidence graph of an r-uniform hypergraph G is a bipartite graph Inc(G) with
vertex sets V (G) and E(G), where v ∈ V (G) and e ∈ E(G) form an edge {v, e} in Inc(G)
if and only if v ∈ e in G.
If there is a function h ∈ H such that
Z r
Y
g(x1 , x2 , . . . , xr ) = h(xi , y) dµ(y) , (9.2)
[0,1] i=1

the electronic journal of combinatorics 28(1) (2021), #P1.32 11

then t(G, g) = t(Inc(G), h).
Similarly to Theorem 3.5, if Inc(G) satisfies Conjecture 1.1, then (9.1) holds for func-
tions g that have representation (9.2) with nonnegative h. Similarly to Theorem 4.1, if
Inc(G) is norming (it requires r to be even), then (9.1) holds for functions g that have
representation (9.2), where h ∈ H can take negative values.

Acknowledgements
The author would like to thank Joonkyung Lee for his valuable comments and suggestions.

References
[1] P. Adams, D. Bryant, and M. Buchanan. A survey on the existence of G-designs. J.
Combin. Designs, 16(5):373–410, 2008. https://doi.org/10.1002/jcd.20170.
[2] F. V. Atkinson, G. A. Watterson, and P. A. D. Moran. A matrix inequality. Quarterly
J. of Math., 11(42):137–140, 1960. https://doi.org/10.1093/qmath/11.1.137.
[3] A. Berman and N. Shaked-Monderer. Completely Positive Matrices, World Scientific,
2003. https://doi.org/10.1142/5273.
[4] G. R. Blakley and P. Roy. Hölder type inequality for symmetric matri-
ces with nonnegative entries. Proc. Amer. Math. Soc., 16(6):1244–1245, 1965.
https://doi.org/10.1090/S0002-9939-1965-0184950-9.
[5] A. Blinco. Theta graphs, graph decompositions and related graph la-
belling techniques. Bull. Austral. Math. Soc., 69(1):173–175, 2004.
https://doi.org/10.1017/S0004972700034377.
[6] D. Conlon, J. Fox, and B. Sudakov. An approximate version of
Sidorenko’s conjecture. Geom. Funct. Anal., 20(6):1354–1366, 2010.
https://doi.org/10.1007/s00039-010-0097-0.
[7] D. Conlon, J. H. Kim, C. Lee, and J. Lee. Some advances on Sidorenko’s conjecture. J.
London Math. Soc., 98(3):593–608, 2018. https://doi.org/10.1112/jlms.12142.
[8] D. Conlon and J. Lee. Finite reflection groups and graph norms. Adv. Math., 315:130–
165, 2017. https://doi.org/10.1016/j.aim.2017.05.009.
[9] D. Conlon and J. Lee. Sidorenko’s conjecture for blow-ups. Discrete Analysis, to
appear. arXiv:1809.01259, 2018.
[10] H. Hatami. Graph norms and Sidorenko’s conjecture. Israel J. Math., 175(1):125–
150, 2010. https://doi.org/10.1007/s11856-010-0005-1.
[11] C. Jagger, P. Šťoviček, and A. Thomason. Multiplicities of subgraphs. Combinatorica,
16(1):123–141, 1996. https://doi.org/10.1007/BF01300130.
[12] J. H. Kim, C. Lee, and J. Lee. Two approaches to Sidorenko’s conjecture. Trans.
Amer. Math. Soc., 368(7):5057–5074, 2016. https://doi.org/10.1090/tran/6487.

the electronic journal of combinatorics 28(1) (2021), #P1.32 12

[13] Y. Kohayakawa, B. Nagle, V. Rödl, and M. Schacht. Weak hypergraph regu-
larity and linear hypergraphs. J. Combin. Theory Ser. B, 100(2):151–160, 2010.
https://doi.org/10.1016/j.jctb.2009.05.005.
[14] J. Lee. On some graph densities in locally dense graphs. Random Struct Alg., 58:322–
344, 2021. https://doi.org/10.1002/rsa.20974.
[15] J. L. X. Li and B. Szegedy. On the logarithmic calculus and Sidorenko’s conjecture.
arXiv:1107.1153, 2011.
[16] L. Lovász. Large networks and graph limits, volume 60 of Colloquium Publications.
AMS, 2012. https://doi.org/10.1090/coll/060.
[17] L. Lovász. Subgraph densities in signed graphons and the local Sidorenko conjecture.
Electr. J. Combin., 18(1) #P127, 2011.
[18] H. P. Mulholland and C. A. B. Smith. An inequality arising in genetical theory.
Amer. Math. Monthly, 66(8):673–683, 1959. https://doi.org/10.2307/2309342.
[19] O. Parczyk. On Sidorenko’s conjecture, Master’s thesis, Freie Universität Berlin,
2014. http://www.uni-frankfurt.de/58522166.
[20] C. Reiher. Counting odd cycles in locally dense graphs. J. Combin. Theory Ser. B,
105:1–5, 2014. https://doi.org/10.1016/j.jctb.2013.12.002.
[21] A. Sidorenko. Inequalities for functionals generated by bipartite graphs. Discrete
Math. Appl., 2(5):489–504, 1992. https://doi.org/10.1515/dma.1992.2.5.489.
[22] A. Sidorenko. A correlation inequality for bipartite graphs. Graphs and Combina-
torics, 9(2):201–204, 1993. https://doi.org/10.1007/BF02988307.
[23] A. Sidorenko. An analytic approach to extremal problems for graphs and hyper-
graphs. In Extremal Problems for Finite Sets (Visegrád, 1991), volume 3 of Bolyai
Soc. Math. Stud., pages 423–455. János Bolyai Math. Soc., Budapest, 1994.
[24] B. Szegedy. An information theoretic approach to Sidorenko’s conjecture.
arXiv:1406.6738, 2014.
[25] B. Szegedy. Sparse graph limits, entropy maximization and transitive graphs.
arXiv:1504.00858, 2015.

A Proofs of Theorems 6.5 and 6.6

Proof of Theorem 6.5. The proof would be easy if all ki were even: then we could
use the Cauchy–Schwarz inequality and the fact that the tree formed by paths of length
k1 /2, k2 /2, . . . , kr /2 connected at one endpoint is an extra-good graph. To deal with odd
values of ki , we are going to subdivide the middle edge of the path. We assume that
k1 , . . . , ks are odd, and ks+1 , . . . , kr are even (0 ⩽ s ⩽ r). We denote the vertices of
the ith path by vi0 , vi1 , . . . , viki . Gluing together vertices vi0 for all i produces vertex
v0 . Gluing P together vertices viki produces vertex v∞ . P The total number of vertices is
n = 2 + ri=1 (ki − 1), and the number of edges is m = ri=1 ki .

the electronic journal of combinatorics 28(1) (2021), #P1.32 13

 R
Let g be a doubly negative function: g(x, y) = [0,1] h(x, z)h(y, z)dµ(z). Let f0 , f∞ , fij
(1 ⩽ i ⩽ r, 1 ⩽ j ⩽ ki − 1) be bounded measurable nonnegative functions on [0, 1]. Let
!1/(2m)
Yr kYi −1

f = f0 · f∞ · fij
i=1 j=1

For i = 1, 2, . . . , r, denote
i −1 i −1
kY
! kY
Pi = g(x0 , xi1 ) g(xi,j−1 , xi,j ) g(xi,ki −1 , x∞ ) fi,j (xi,j ) .
j=2 j=1
m
We need to prove I ⩾ J , where
Z r
Y Z
I= f0 (x0 )f∞ (x∞ ) Pi dµn , J= f (x)g(x, y)f (y) dµ2 .
[0,1]n i=1 [0,1]2

We can assume that f0 = f∞ , and fi,j = fi,ki −j for all i and j. Indeed, we could swap f0
with f∞ , and fi,j with fi,ki −j for all i, j (this would not change the value of I), and then
apply the Cauchy–Schwarz inequality to the geometric mean of both expressions for I.
For i = 1, 2, . . . , s (that is when ki is odd), we can add a variable zi and replace
g(xi,(ki −1)/2 , xi,(ki +1)/2 ) in the expression for Pi with the product of h(xi,(ki −1)/2 , zi ) and
h(xi,(ki +1)/2 , zi ) :
 
(ki −1)/2
Y
Pi = g(x0 , xi,1 )  g(xi,j−1 , xi,j ) h(xi,(ki −1)/2 , zi ) h(xi,(ki +1)/2 , zi )
j=2
 
i −1
kY i −1
kY
× g(xi,j−1 , xi,j ) g(xi,ki −1 , x∞ ) fi,j (xi,j ) .
j=(ki +3)/2 j=1

Now in the expression for I we have to integrate over n + s variables. Define for i ⩽ s,
(ki −1)/2 (ki −1)/2
Y Y
Qi = g(x0 , xi1 ) g(xi,j−1 , xi,j ) h(xi,(ki −1)/2 , zi ) fi,j (xi,j ) ,
j=2 j=1

and for i > s,

ki /2 ki /2−1
Y Y
1/2
Qi = g(x0 , xi,1 ) g(xi,j−1 , xi,j ) fi,j (xi,j ) fi,ki /2 (xi,ki /2 ) .
j=2 j=1
Qr
Let A(z1 , . . . , zs , xs+1,ks+1 /2 , . . . , xr,kr /2 ) denote the integral of f0 (x0 ) i=1 Qi over all vari-
ables except z1 , . . . , zs , xs+1,ks+1 /2 , . . . , xr,kr /2 . Set

B(y1 , . . . , ys , xs+1,ks+1 /2 , . . . , xr,kr /2 )

Z s
Y
= A(z1 , . . . , zs , xs+1,ks+1 /2 , . . . , xr,kr /2 ) h(yi , zi ) dµ(z1 ) · · · dµ(zs ) .
[0,1]s i=1

the electronic journal of combinatorics 28(1) (2021), #P1.32 14

Then Z
I = A(z1 , . . . , zs , xs+1,ks+1 /2 , . . . , xr,kr /2 )2 dµr ,
[0,1]r
and
Z s Z
Y 2
s
I ·J = I· f (yi ) h(yi , zi ) dµ(yi ) dµ(z1 ) · · · dµ(zs )
[0,1]s i=1 [0,1]
Z 2
r
⩾ B(y1 , . . . , ys , xs+1,ks+1 /2 , . . . , xr,kr /2 ) dµ .
[0,1]r
R
Notice that [0,1]r Bdµr is the left hand side of (1.3) for the tree G formed by paths of length
(k1 + 1)/2, . . . , (ks +R1)/2, ks+1 /2, . . . , kr /2 connected at one endpoint. By Corollary 6.4,
G is extra-good, so [0,1]r Bdµr ⩾ J e(G) . As e(G) = (m + s)/2, we get
Z 2
−s r
I ⩾ J Bdµ ⩾ Jm .
[0,1]r

Proof of Theorem 6.6. Let G0 be formed by attaching a k-edge cycle (v0 , v1 , . . . , vk−1 ,
vk = v0 ) to a vertex v of graph G, so v = v0 =Rvk . Let g be a doubly nonnegative func-
tion. There exists h ∈ H such that g(x, y) = [0,1] h(x, z)h(y, z)dµ(z). Assign bounded
measurable nonnegative functions on [0, 1] to all vertices of G, and denote by F (x) the
integral on the left hand side of (1.3) taken over all variables except the one that corre-
sponds to v. Assign bounded measurable nonnegative functions f1 , f2 , . . . , fk−1 to vertices
v1 , v2 , . . . , vk−1 . For functions γ1 , γ2 , . . . , γr , denote
Z Yr
Ir (x0 , xr ; γ1 , γ2 , . . . , γr ) = (g(xi−1 , xi ) γi (xi )) dµ(x1 ) · · · dµ(xr−1 ).
[0,1]r−1 i=1

Let IG0 be the value of the integral on the left hand side of (1.3) for G0 . Then
Z
IG0 = F (x) Ik (x, x; f1 , f2 , . . . , fk−1 , 1) dµ(x) .
[0,1]

Let f0 be the product of Rall v(G0 ) functions assigned to vertices of G0 . We need to prove
IG0 ⩾ J e(G0 ) , where J = [0,1]2 f (x)g(x, y)f (y)dµ2 and f = (f0 )1/(2e(G0 )) .
p
Set γi = fi fk−i , so γk−i = γi . By the Cauchy–Schwarz inequality,
Ik (x, x; f1 , f2 , . . . , fk−1 , 1)2 = Ik (x, x; f1 , f2 , . . . , fk−1 , 1) · Ik (x, x; fk−1 , fk−2 , . . . , f1 , 1)
⩾ Ik (x, x; γ1 , γ2 , . . . , γk−1 , 1)2 .
If k = 2a, then
√
Z
I2a (x, x; γ1 , . . . , γ2a−1 , 1) = Ia (x, y; γ1 , . . . , γa−1 , γa )2 dµ(y)
[0,1]
Z 2
√
⩾ Ia (x, y; γ1 , . . . , γa−1 , γa ) dµ(y) .
[0,1]

the electronic journal of combinatorics 28(1) (2021), #P1.32 15

Construct graph Ga from graph G by attaching two disjoint a-edge paths to vertex v. By
√
Theorem 6.3, Ga is extra-good. Assign functions γ1 , . . . , γa−1 , γa to vertices of each of
the two paths. Let IGa be the value of the integral on the left hand side of (1.3) for Ga .
Then
Z 2
√
Z
IGa = F (x) Ia (x, y; γ1 , . . . , γa−1 , γa ) dµ(y) dµ(x) ⩽ IG0 .
[0,1] [0,1]

The right hand side of (1.3) is the same for both G0 and Ga , so we are done with the case
k = 2a. If k = 2a + 1, we have

I2a+1 (x, x; γ1 , . . . , γ2a , 1)

Z
= Ia (x, y ′ ; γ1 , . . . , γa ) g(y ′ , y ′′ ) Ia (x, y ′′ ; γ2a , . . . , γa+1 ) dµ(y ′ )dµ(y ′′ )
[0,1]2
Z
= (Ia (x, y; γ1 , . . . , γa ) h(y, z) dµ(y))2 dµ(z) ,
[0,1]

and
Z Z Z 2
′ ′ ′′ ′′ 2
J = f (s ) g(s , s ) f (s ) dµ = h(s, z) f (s) dµ(s) dµ(z) .
[0,1]2 [0,1] [0,1]

Hence, by the Cauchy–Schwarz inequality,

I2a+1 (x, x; γ1 , . . . , γ2a , 1) · J ⩾

Z 2
⩾ Ia (x, y; γ1 , . . . , γa ) h(y, z) h(s, z) f (s) dµ(y)dµ(z)dµ(s)
[0,1]3
Z 2
= Ia+1 (x, s; γ1 , . . . , γa , f ) dµ(s) .
[0,1]

Construct Ga+1 from G by attaching two disjoint (a + 1)-edge paths to vertex v. By

Theorem 6.3, Ga+1 is extra-good. Assign functions γ1 , . . . , γa , f to vertices of each of the
two paths. Let IGa+1 be the value of the integral on the left hand side of (1.3) for Ga+1 .
Then
Z Z 2
IGa+1 = F (x) Ia+1 (x, s; γ1 , . . . , γa , f ) dµ(s) dµ(x) ⩽ IG0 · J .
[0,1] [0,1]

The right hand side of (1.3) is equal to J e(G0 ) for G0 , and J e(Ga+1 ) = J · J e(G0 ) for Ga+1 .
As (1.3) holds for Ga+1 , it holds for G0 , too.

the electronic journal of combinatorics 28(1) (2021), #P1.32 16