∗

An Introduction to Ramanujan Graphs

KRISTAPS J. BALODIS
Department of Mathematics and Statistics
University of Manitoba, Winnipeg MB, CANADA
balodis-k@webmail.uwinnipeg.ca

Abstract

Basic spectral estimates of expanders are developed by applying ideas from Functional
Analysis and Riemannian Manifolds to (finite) graphs. After establishing several foundational
inequalities, the concept of Ramanujan graphs is introduced, and existence results are dis-
cussed. In the final section it is shown that zeta functions of Ramanujan graphs satisfy an
analogue of the Riemann Hypothesis.

MSC: 05C40, 05C50

Keywords: Ramanujan Graphs, Spectral Graph Theory, Expanders, Zeta Functions.

∗
This was paper completed April, 2018 in fulfillment of the requirements for a reading course in Spectral Graph
Theory, directed by S. Kirkland, University of Manitoba.

1
Contents
1 Spectral Graph Theory and Functional Analysis 2

2 Expanders 4

3 Constructions of Ramanujan Graphs 13

4 A Proof of The Riemann Hypothesis! ...(for graphs) 14

1 Spectral Graph Theory and Functional Analysis

In this section we recall various classical results in the spectral theory of graphs, and develop the
basic theory of functional analysis of graphs. Many of the proofs and derivations of classical spectral
graph theory can be found in [1], and references will be given for those which can not.
For a graph Γ with  vertices V Γ = {v1 , ..., vn } the Adjacency Matrix A(Γ) = A of a (simple) graph
1, i 6= j, v ∼ v
i j
is defined by Aij = .
0, otherwise.
We write the eigenvalues µ0 ≥ µ1 ≥ ... ≥ µn−1 . If Γ is regular, then k = µ0 .
Thus the largest eigenvalue of A is in some sense uninteresting for k-regular graphs, for we already
know the information it’s telling us. Recall also that for bipartite regular graphs µn−1 = −k, thus
for regular graphs we will refer to µi 6= ±k as the Non-Trivial
 eigenvalues of Γ.

deg(vi ), i = j



The Laplacian matrix of a graph Γ is given by Lij = −1, vi ∼ vj


0,

otherwise.
with eigenvalues 0 = λ0 ≤ λ1 ≤ ... ≤ λn−1 , where the multiplicity of 0 is equal to the number of
components of Γ, and has corresponding eigenvector ~1 the all 1’s vector. If Γ is k-regular, we have
that L = kI − A, and so λi = k − µi and thus the multiplicity of k is also equal to the number of
connected components.
Fiedler pioneered the study of the smallest non-trivial eigenvalue λ1 which he dubbed the Algebraic
Connectivity having established it’s connection to vertex and edge connectivity in [4]. In light of
Fiedler’s work, plenty of useful results have been established between Laplacian eigenvalues and
different measures of a graph’s connectivity.
As an application of the theory of Rayleigh Quotients for matrices, we have the following expressions
µ1 = max(~v6=~0)⊥~1 hA~ v ,~v i
h~v ,~v i
, and similarly k − µ1 = λ1 = min(~v6=~0)⊥~1 hL~ v ,~v i
h~v ,~v i
.

2
 For an arbitrary orientation of EΓ, we may define the incidence matrix to be

1, x is the tail of e,



Nx,e = −1, x is the head of e,


0,

otherwise.

where N N T = L, and thus for an arbitrary vector ~v , ~v T L~v = − vy )2 where the vx , vy

P
(x,y)∈E (vx
refer to the x and y entries of the vector ~v .

When certain techniques work well for one branch of mathematics, it’s natural to ask if they can
be extended to related fields, and to what extent. It turns out that applying methods of Functional
Analysis and Riemannian Manifolds to Graph Theory is a fairly fruitful endeavor. We outline some
of the basics of these analogies here, and suggest the reader consult the notes [11] and [12] for more
details.
We begin by defining an analogue of Lp -spaces for graphs. For a graph Γ with vertex set V and
edge set E, we give E an arbitrary orientation and define
X X
l2 (V ) = {f : V → C| |f (v)|2 < ∞} and l2 (E) = {f : E → C| |f (e)|2 < ∞}
v∈V e∈E

Both of the above space can also be equipped with respective hermitian products,
X X
hf, gil2 (V ) f¯(v)g(v) and hf, gil2 (E) = f¯(e)g(e)
v∈V e∈E

1
Having Hermitian products naturally gives rise to a concept of norm on our spaces by ||f || = hf, f i 2 .
As edges are simply pairs of vertices one should expect that there are some natural ways in which one
can transfer functions from l2 (V ) to l2 (E) and vice-versa. When working with maps of continuous
spaces, we’re of course interested in linear approximations, such as the differential. Fortunately in
the discrete setting we need not deal with approximations, for we can explicitly compute the finite
differences between a function’s values. We define a coboundary operator d : l2 (V ) → l2 (E) where
for some function f ∈ l2 (V ) we define df (e) = f (e+ ) − f (e− ). We also define the boundary of a set
U ⊆ V (Γ) to be ∂U = {e ∈ E|e is incident with exactly one vertex of U }.
In standard functional analysis we also work with adjoint operators for maps between spaces.
Classically the adjoint of some operator H : W1 → W2 is defined to be an operator H ∗ such that

3
hHf, giW1 = hf, H ∗ giW2 . Define δ : V × E → {−1, 0, 1} by

1,


 x = e+
δ(x, e) = −1, x = e−


0,

otherwise.

Thus we may write df (e) = δ(e+ , e)f (e+ ) + δ(e− , e)f (e− ) = x∈V δ(x, e)f (x) since δ(x, e) is zero
P

for x not incident with e. Now

* + !
X X X XX
f, δ(x, e)g(e) = f¯(x) δ(x, e)g(e) = f¯(x)δ(x, e)g(e)
e∈E x∈V e∈E x∈V e∈E

and since δ is real-valued,

!
X X
= f (x)δ(x, e) g(e) = hdf, gil2 (E)
e∈E x∈V

Therefore d∗ g(x) = e∈E δ(x, e)g(e).

P

By analogy with the so called ”Geometer’s Laplacian”, we define the Laplace operator
∆ = d∗ d : l2 (V ) → l2 (V ). For f ∈ l2 (V ) we get that,
!
X X X
∆f = d∗ (df ) = δ(x, e)df = δ(x, e) δ(x, e)f (x)
e∈E e∈E v∈V

Observe though, by definition δ(x, e) = Nx,e is just the incidence as defined above, and we get that
∆f = N N T f = Lf . Therefore the Laplace operator is just the Laplacian Matrix for the graph! It
may very well be that this fact is the etymological justification for calling the matrix L associated
the the graph the Laplacian Matrix in the first place.

2 Expanders
In this section we introduce the notion of expander graphs, and prove a relationship between the
”quality” of an expander graph, and it’s algebraic connectivity.

Definition 2.1 The Expanding Constant of a graph Γ is defined as

|∂U |
c(Γ) = min
U ⊆V
|V Γ|
(G):|U |≤ 2 |U |

The constant defined above goes by many names in the literature including: Cheeger Constant,

4
Isoperimetric Constant and Magnifying Constant. When we’re considering graphs to be models of
a network, there is some sense in which the expander constant measures the quality of the network.
The ratio |∂U|U |
|
describes how well connected the vertices U are to the rest of the graph. If the
ratio is large there are many more edges joining U to the rest of Γ, and conversely if it’s small,
there’s relatively few connections to the rest of the graph. Thus the minimum ratio over all sets (of
order less than half the graph’s order) measures a worst case scenario. Therefore a relatively large
expanding constant signifies that a graph is efficiently connected.
Note that |∂|UŪ|| = |V|∂U |
−U |
and |U | ≤ |V2 | ⇐⇒ |Ū | ≥ |V2 | , so there is no need to consider the sets
of order larger than half the vertices, for we already get their expansion information ”for free” by
considering the smaller sets. Also ∂V = ∅ so if we were to focus our attention on the larger sets,
we would always get 0 as the minimum value.
Nothing from the definition of the expanding constant clues us in to what explicit values of c(Γ) are
to be considered ”good”. Of course complete graphs are highly connected, though not extremely
efficient networks edge-wise. Since regular graphs enjoy many nice properties practically and theo-
retically , we will make our notion of having a good expansion constant somewhat more precise by
making it relative to some fixed regularity.

Definition 2.2 A family (Γm )m∈Z + of connected k-regular graphs such that limm→∞ |V Γm | = ∞ is
called a family of expanders if ∃ ∈ R+ such that c(Γ) ≥ , ∀m ≥ 1.

As we mentioned in the first section, the Laplacian eigenvalues (λ1 in particular) have been
shown to be closely related to various notions of graph connectivity. One may expect then, that
the Cheeger constant of a graph bears some relation to the spectrum of the Laplacian matrix, and
indeed,

Theorem 2.3 For Γ a k-regular, connected graph

λ1 p
≤ c(Γ) ≤ 2kλ1
2
Proof. The matrix L is both real and symmetric, thus it’s orthogonally diagonalizable as
∆ = diag(0, λ1 , ..., λn−1 ). Define f ∈ l2 (V ) by fixing some subset F ⊆ V and taking

|V − F |, x ∈ F
f (x) =
−|F |, x∈V −F

Thus f takes the value |V − F |, |F | times and −|F |, |V − F | times and thus
P
x∈V f (x) = |V − F ||F | − |F ||V − F | = 0, and

X
||f ||2 = f 2 (x) = |V − F |2 |F | + |F |2 |V − F | = |F ||V − F |(|V − F | + |F |) = |F ||V − F ||V |
x∈V

5
 
±|V |, e ∈ ∂F
Observe that df (e) =
0, otherwise

||df ||2 = hdf |df i = hd∗ df |f i = hLf |f i = h∆f |f i ≥ λ1 ||f ||2

X
and ||df ||2 = df (e)2 = |V |2 |∂F | ≥ (λ1 )|F ||V − F ||V |
e∈E

|∂F | |V − F |
∴ ||df ||2 = |V |2 |∂F | ≥ (λ1 )|F ||V − F ||V | ⇒ ≥ (λ1 )
|F | |V |
|V |
Now in our computation of the Cheeger constant of a graph we want |F | ≤ 2
and since our choice
of F was otherwise arbitrary,

|∂F | λ1 |∂U | λ1
≥ ⇒ c(Γ) = min ≥
|F | 2 |V
U ⊆V (G):|U |≤ 2
Γ| |U | 2

In order to establish the remaining part of the inequality we first let g ∈ l2 (V ) to be an eigen-
function of L (which is also to say an eigenvector) corresponding to λ1 .
Define G+ = {x ∈ V |g(x) > 0}, and assume |G+ | ≤ |V2 | . Otherwise for |G+ | ≥ |V2 | , we could take
the function −g with G− = {x ∈ V |g(x) > 0} so that |G− | ≤ |V2 | . Since −g is just a scalar multiple
of g they’re essentially the same eigenfunction as far as the Laplacian is concerned, and thus our
assumption is justified.
Define h(x) = max{g(x), 0} for x ∈ V and H = e∈E |h2 (e+ ) − h2 (e− )|.
P

Of course our graph is finite, and thus h only takes on finitely many distinct values say
u0 < u1 < ... < ur where we let Ui = {x ∈ V |h(x) ≥ ui }.
Observe that if for some e ∈ E, h(e+ ) = h(e− ) then |h2 (e+ ) − h2 (e− )| = 0, so letting
Eh = {e ∈ E|h(e+ ) 6= h(e− )} (we don’t have to account for negative values since h is non-negative
by definition) we have that H = e∈Eh |h2 (e+ ) − h2 (e− )|. For e ∈ Eh , h takes distinct values ui , uj
P

for some distinct 1 ≤ i, j ≤ r on it’s endpoints. By convention we will write i(e) and j(e) for the
indices of the u’s corresponding to the endpoints of e, chosen such that i(e) > j(e). Thus,
X X
H= (u2i(e) − u2j(e) ) = (u2i(e) − u2i(e)−1 + u2i(e)−1 ∓ ... − u2j(e)+1 + u2j(e)+1 − u2j(e) )
e∈Eh e∈Eh

i(e)
X X
= u2l − u2l−1
e∈Eh l=j(e)+1

For each edge e the left-most sigma runs over, we added a term 0 = u2l − u2l for each value ul which
is (strictly) between the value h takes on the endpoints of e. By the definition of Eh the endpoints

6
differ in value for each edge in the sum, thus the difference u2l − u2l−1 will appear once for each edge
in Ul , so
i(e) r
X X X
2 2
H= ul − ul−1 = |∂Ul |(u2l − u2l−1 )
e∈Eh l=j(e)+1 l=1

Note that the sum begins at l = 1 as U0 = V and thus ∂U0 = ∂V = ∅. Since h = max{g, 0} and
|G+ | ≤ |V2 | , |supp(h)| ≤ |V2 | and thus |Ul | ≤ |V2 | for 1 ≤ l ≤ r. By the minimality of c(Γ), c(Γ) ≤ |∂U l|
|Ul |
.

r
X r
X
∴H= |∂Ul |(u2l − u2l−1 ) ≥ c(Γ) |Ul |(u2l − u2l−1 )
l=1 l=1

= c(Γ) |Ur |u2r + (|Ur−1 | − |Ur |)u2r−1 + ... + (|U1 | − |U2 |)u21
 

r−1
!
X
= c(Γ)(|U1 |(u21 − u20 ) + |U2 |(u22 − u21 ) + ... + |Ur |(u2r − u2r−1 )) = c(Γ) u2r |Ur | + |Ul − Ul+1 |u2l
l=1

where the final equality follows since Ul+1 ⊆ Ul and thus |Ul | − |Ul+1 |.
Note that f (x) = ul for x ∈ Ul − Ul+1 , and in fact all such x belong to Ul − Ul+1 and thus we have
that, !
X r−1 X
X X
H ≥ c(Γ) h2 (x) + h2 (X) = c(Γ) h2 (x) = c(Γ)||h(x)||2
Sr
x∈Ur i=1 x∈Ul x∈ i=1
Sr
since i=1 =V.
Now from the definition, we can write H as
! 21 ! 12
X X X
H= |h(e+ ) + h(e− )||h(e+ ) − h(e− )| ≤ (h(e+ ) + h(e− ))2 (h(e+ ) − h(e− ))2
e∈E e∈E e∈E

For a, b ∈ R where a ≤ b WLOG, (a + b)2 = a2 + 2ab + b2 ≤ a2 + 2a2 + b2 ≤ 2(a2 + b2 ), which we

apply to the first term. The sum in the second set of parenthesis is just the hermitian product of
df with itself, and by the Cauchy-Schwarz inequality | h~u, ~v i | ≤ ||~u||||~v ||,
! 21
√ X
H≤ 2 h2 (e+ ) − h2 (e− ) ||dh||
e∈E

In the remaining sum, since Γ is k-regular h2 (x) will appear k times (once for every edge it’s incident
with) for each x ∈ V so
! 12
√ X √
H≤ 2 k h2 (x) ||dh|| = 2k||h|| · ||dh||
x∈V

7
Recall that G+ = {x ∈ V |g(x) > 0}, and since g is an eigenvector with eigenvalue λ1 6= λ0 which
has eigenvector ~1, g ⊥ ~1 so x∈V g(x) = 0 and since g is not itself trivial, ∃x0 ∈ V such that
P

g(x0 ) 6= 0. Thus we may find some x ∈ G+ and,

X X
(Lh)(x) = kf (x) − (Ah)(x) = kg(x) − Axy h(y) = kg(x) − Axy g(x)
y∈V y∈V +

X
≤ kg(x) − Axy g(y) = (Lg)(x) = λ1 g(x)
y∈V

Recalling our diagonalization ∆ of L,

1 1 1
||dh|| = hdh|dhi 2 = hLh|hi 2 = h∆h|hi 2
! 12 ! 21
X X p
≤ (Lh)(x)g(x) ≤ λ1 g 2 (x) ≤ λ1 ||h||
x∈V + x∈V +

√ √ p
∴ c(Γ)||h||2 ≤ H ≤ 2k||dh|| · ||h|| ≤ 2k λ1 ||h||2
p
⇒ c(Γ) ≤ 2kλ1

It’s worth making sure that we can in fact divide through by ||h||2 . As noted above ∃x ∈ G+ and
thus h(x) 6= 0 so ||h||2 6= 0
Now that our Cheeger constant is bounded by the algebraic connectivity of our graph, we can
talk about the asymptotic behavior of the Cheeger constant(s) for a family of graphs, by exploring
the asymptotic behavior of the algebraic connectivity. That is, if we can bound the algebraic
connectivities of a family of graphs away from zero, the expanding constant will in turn be bounded
away from zero.

Theorem 2.4 For a k-regular finite connected graph Γ such that diam(Γ) ≥ 2b + 2 ≥ 4 for some
b, √
√ 2 k−1−1
µ1 > 2 k − 1 −
b
Proof. For convenience, we will drop the orientation of EΓ, and simply deal with the underlying
graph. Let u, v ∈ V Γ such that d(u, v) ≥ 2b + 2. For each 0 ≤ i ≤ b we define the sets
Ui = {w ∈ V Γ|d(u, w) = i}, Vi = {w ∈ V Γ|d(v, w) = i}. Assume towards a contradiction that
for some distinct pair of sets W, W 0 each from either of the above families ∃w ∈ W ∩ W 0 . Thus
d(u, v) ≤ d(u, w) + d(w, v), and since each of d(u, w), d(w, v) is equal to the index of the set they
belong to, which are both at most b, we have that d(u, v) ≤ 2b a contradiction.
Suppose towards a further contradiction that there exists some 0 ≤ j, k ≤ b such that
∃w ∈ Uj , ∃w0 ∈ Vk such that w ∼ w0 . Thus

8
d(u, v) ≤ d(u, w) + d(w, w0 ) + d(w0 , v) ≤ 2b + 1 by a similar argument to the previous one, which
again contradicts our assumptions.
For i > 0, if w ∈ Ui (resp Vi ) then there exists a path u = u1 , ..., ui−1 , ui = w
(reps. v = v1 , ..., vi−1 , vn = w) which is the shortest possible, and thus
u = u1 , ..., ui−1 (resp v = v1 , ..., vi−1 ) is a shortest u − ui−1 (resp. v − vi−1 ) path. Therefore Ui is at
least as big as Ui−1 .
If i < b, since there are at most k − 1 ways to extend u = u1 , ..., ui = w (resp. v = v1 , ..., vi = w) to
a path of length i + 1. Since these paths may not necessarily be the shortest path between the end
points, we have that |Ui | ≤ (k − 1)|Ui+1 | (resp. |Vi | ≤ (k − 1)|Vi+1 |).
Therefore |Ub | ≤ (k − 1)|Ub−1 | ≤(k − 1)2 |Ub−2 | ≤ ... ≤ (k − 1)b−i |Ub−i | (reps. |Vb | ≤ (k − 1)b−i |Vb−i |).


 α, w=u


β, w=v



i−1
Define f : V Γ → R by f (w) = α(k − 1) 2 , w ∈ Ui , i ≥ 1

β(k − 1) i−1

2 , w ∈ Vi , i ≥ 1





0, otherwise

where α, β ∈ R are chosen so that hf, 1i = 0. Define ∂ + Ui = {(x, y) ∈ ∂Ui |x ∈ Ui , y ∈ / Ui−1 } so
+ +
Sb Sb
∂Ui = ∂ Ui ∪ ∂ Ui−1 (resp. for Vi ’s). By definition supp(f ) = ( i=0 Ui ) ∪ ( i=0 Vi ) and since f is
constant on any Ui (resp. Vi ), (f (x) − f (y))2 = 0 on Ui (resp. Vi ) thus
  ! !
X b
[ b
[
S = supp  (f (x) − f (y))2  = ∂Ui ∪ ∂Vi
(x,y)∈EΓ i=0 i=0

Since ∂U0 = ∂ + U0 , N (u) = U1 and ∂Ui = ∂ + Ui ∪ ∂Ui−1 for 1 ≤ i ≤ b,

b b b
! b
! b
[ [ [ [ [
+ + + +
∂Ui = ∂U0 ∪ (∂ Ui ∪ ∂ Ui−1 ) = U0 ∪ ∂ Ui ∪ ∂ Ui−1 = ∂ + Ui = U
i=0 i=i i=1 i=1 i=0

where the last equality is by definition, and we choose to define V in a similar fashion.
 
1 X X
∴ (f (x) − f (y))2 + (f (x) − f (y))2 
2
(x,y)∈U (x,y)∈V

1 X 1 X
= (f (x) − f (y))2 = (f (x) − f (y))2
2 2
(x,y)∈S (x,y)∈EΓ

1 XX 2 1 XX 2 XX
f (x) + f 2 (y) − 2f (x)f (y) = f (x) + f 2 (y) − f (x)f (y)
2 x∈V Γ y∼x 2 x∈V Γ y∼x x∈V Γ y∼x

9
 !
1 X X X
= kf 2 (x) + f 2 (y) − Af (x)
2 x∈V Γ y∼x x∈V Γ

but for each x ∈ V Γ there’s k vertices y ∼ x so

1 X X
k(f 2 (x) + f 2 (y)) − hAf, f i = k f 2 (x) − hAf, f i
2 x∈V Γ x∈V Γ

= k hf, f i − hAf, f i = hLf, f i

X b−1
X X X
2
Evaluating (f (x) − f (y)) = (f (x) − f (y))2 + (f (x) − f (y))2
(x,y)∈U i=0 (x,y)∈∂ + Ui (x,y)∈∂ + Ub

By the definition of f , and that each element of Ui , 0 < i < b has at most k − 1 neighbours in Ui+1
for i < b, and at least one in Ui−1 . As (f (x) − f (y))2 = (f (y) − f (x))2 we can write the sum over
∂ + U so that x ∈ Ui ,
i−1 i
X
(f (x) − f (y))2 ≤ (k − 1)|Ui |(α(k − 1) 2 − α(k − 1) 2 )2
(x,y)∈∂ + Ui

1−1
X
note that (f (x) − f (y))2 ≤ (k − 1)|U0 |(α − α(k − 1) 2 ) = (k − 1)(α − α)2 = 0
(x,y)∈∂ + U0

and (x, y) ∈ ∂ + Ub ⇒ x ∈ Ub , y ∈
/ Ub−1 , y ∈
/ supp(f ) so
X
(f (x) − f (y))2 ≤ (k − 1)|Ub |α2 (k − 1)1−b
(x,y)∈∂ + Ub

1 X X
thus (f (x) − f (y))2 ≤ (f (x) − f (y))2
2
(x,y)∈U (x,y)∈U

b−1
!
1−i i
X
≤ (k − 1)|Ui |(α(k − 1) 2 − α(k − 1) )2 2 + (k − 1)|Ub |α2 (k − 1)1−b
i=1
√ √
noting that in the last term we can write (k − 1) = ( k − 1 − 1)2 + 2 k − 1 − 1, so the above
becomes
b−1
!
2
√ 2
X
1−i
√ √
α ( k − 1 − 1) (k − 1)|Ui |(k − 1) +(α2 ( k − 1−1)2 |Ub |q 1−b +α2 (2 k − 1−1)|Ub |(k−1)1−b )
i=1

Recall it was demonstrated that |Ub | ≤ (k − 1)b−i |Ub−i | for 1 ≤ i ≤ b and thus
b(α2 |Ub |(k − 1)1−i ) ≤ bi=1 α2 (k − 1)1−i |Ui | which give us that
P

10
 √ √ Pb
α2 (k−1)1−i |Ui |
α2 (2 k − 1 − 1)|Ub |(k − 1)1−b ≤ (2 k − 1 − 1) i=1 b so ultimately

b √ b
1 X 2
√ 2
X
2 1−i 2 k−1−1X 2
(f (x) − f (y)) ≤ ( k − 1 − 1) α (k − 1) |Ui | + α (k − 1)1−i |Ui |
2 i=1
b i=1
(x,y)∈U

√ b
√ 2 k−1−1 2 X 2
≤ (k − 1 − 2 k − 1 + 1 + )(α + α (k − 1)1−i |Ui |)
b i=1

Replace each instance of Ui with Vi in the above, the same argument reveals
√ b
1 X 2
√ 2 k−1−1 2 X 2
(f (x) − f (y)) ≤ (k − 2 k − 1 + (β + β (k − 1)1−i |Ui |))
2 b i=1
(x,y)∈V

X X X
As supp(f ) = U ∪ V, hf, f i = f ( x) = f 2 (x) + f 2 (x)
x∈V Γ x∈U x∈V

b X b X b
! b
X X X X
= f 2 (x) + f 2 (x) = α2 + α(k − 1)1 − i|Ui | + β2 + β 2 (k − 1)1−i |Vi |
i=0 x∈Ui i=0 x∈Vi i=1 i=1

f is constant on each Ui , Vi , so it’s value on the set is simply the size of the set multiplied by the
corresponding constant, that is

b
! b
X X
2 2 1−i 2
α + α (k − 1) |Ui | +β + β 2 (k − 1)1−i |Vi |
i=1 i=1

hLg, gi hLf, f i
∴ k − µ1 = λ1 = min ≤
g6=0,g⊥~1 hg, gi hf, f i
  !−1
1 X X X
= (f (x) − f (y))2 + (f (x) − f (y))2  f 2 (x)
2 x∈V Γ
(x,y)∈U (x,y)∈V

√ b b
!
√
 
2 k−1−1 X X
= k−2 k−1+ (α2 + α2 (k − 1)1−i |Ui |) + β 2 + β 2 (k − 1)1−i |Ui | hf, f i−1
b i=1 i=1
√
√ 2 k−1−1
⇒ k − µ1 ≤ k − 2 k − 1 +
b
√
√ 2 k−1
∴ µ1 ≥ 2 k − 1 −
b

Thus we can obtain even sharper detail about the behavior of the algebraic connectivity if we
can identify the growth of the diameter of a family of regular graphs. For a family of graphs with

11
fixed regularity, and increasing orders, the graphs will be forced to get progressively sparse. To be
precise,

Theorem 2.5 For a family of k-regular graphs {Γn }n∈N0 such that |V Γn | = n,

lim diam(Γn ) = ∞
n→∞

Proof. Let m ∈ N0 , v ∈ V Γm . By definition, every vertex of Γm must belong to a walk of length

dm = diam(Γm ) which starts at v. Inductively building any walk from v, one has k choices for each
successive vertex, and thus there are k dm walks of length dm . Since each walk has at most dm + 1
vertices we must have that |V Γm | ≤ (dm + 1)k dm = dm km d
+ kmd
. By arbitrary selection this holds
for any choice of v and m, thus in general letting dn = diam(Γn ) for n ∈ N0 ,

∞ = lim n = lim |Γn | ≤ lim (dn + 1)k dn

n→∞ n→∞ n→∞

Since k is fixed, this can only hold if limn→∞ dn = ∞.

Therefore, for a a family of k-regular graphs {Γn }n∈N0 such that
diam(Γ) ≥ 2b + 2 ≥ 4, limn→∞ diam(Γ) = ∞ which implies that we can higher and higher values
for b, and thus for µ1 (Γm ) the largest non-trivial eigenvalue of Γm ,
√ √
√ 2 k−1−1 √ 2 k−1−1
lim µ1 (Γm ) ≥ lim 2 k − 1 − ≥ 2 k − 1 − lim
n→∞ n→∞ b n→∞ 2b + 2
√
√ 2 k−1−1 √
≥ 2 k − 1 − lim =2 k−1
n→∞ diam(Γn )

Therefore the µ1 -eigenvalue increase as the order increase, for a family of graphs of fixed regularity.
Thus any graph which beats this bound, is in optimal in an asymptotic sense.

Definition 2.6 A finite k-regular connected graph Γ is said to be Ramanujan if for every non-trivial
√
eigenvalue µ we have |µ| ≤ 2 k − 1

We’ve phrased the definition in terms of the norms of all non-trivial eigenvalues to align with
√
the literature, note though that this is equivalent to demanding that µ1 ≤ 2 k − 1.
√
Suppose {Γn }n∈Z+ is a family of k-regular Ramanujan graphs, that is µ1 (Γn ) ≤ 2 k − 1, ∀n ∈ Z+ .
√ √
Or rephrasing this in terms of Laplacian eigenvalues, k − λ1 ≤ 2 k − 1 ⇒ k − 2 k − 1λ1 . Note
√ √
that 0 < k − 2 k − 1 for k ≥ 2 and by Theorem 2.3 0 < k−2 2 k−1 ≤ c(Γ) and therefore a family of
Ramanujan graphs makes for a family of expanders.

12
3 Constructions of Ramanujan Graphs
As a first example of a Ramanujan graph, the adjacency spectrum (with multiplicity) of everybody’s
√ √
favorite graph (known by many as Petersen’s Graph) is {(−2)(4) , 1(5) , 3} so µ1 = 1 ≤ 2 2 = 2 3 − 1
and since the Petersen graph is 3-regular, it is Ramanujan.
Of course we’re really after infinite families1 of Ramanujan graphs for some fixed regularity. The
most accessible of all such families is simply {Cn }n∈N the family of n-cycles. The spectra of Cn
√
consists of cos 2πj


n
for 0 ≤ j < n which are all bounded above by 2 = 2 2 − 1. Now cycles are
not efficiently connected graphs in general, but they are the only connected 2-regular graphs, and
thus certainly optimally connected in the context of a family of graphs with fixed regularity 2.
With the exception of the above trivial example, families of Ramanujan graphs are generally hard
to come by. While several constructions of families of Ramanujan graphs are known, they all rely
on fairly developed and deep ideas in Number Theory and Algebraic Geometry. Consequently the
rest of this section, and in fact the rest of the paper, will primarily contain references in place of
proofs. Unless otherwise stated, the results of this section may be found in [6].

For a finite non-empty subset S of a group G, we can define the Cayley Graph Γ(G, S) such
that V Γ(G, S) = G, EΓ(G, S) = {(x, y) : ∃s ∈ S where y = xs}.
The real normed division algebra with basis {1, i, j, k} relations given by,

i2 = j 2 = k 2 = ijk = −1

are referred as the Quaternions. We define for some ring R, H(R) = {a + bi + cj + dk|a, b, c, d ∈ R}.
The quaternions were famously discovered by Hamilton who carved the basis relations on the
Brougham Bridge, Dublin. [2] is a great exposition (which is by no means exhaustive) of the beauty
of quaternions.
A classic result of Jacobi states that r1 (n) the number of ways to write n ∈ Z as a sum of 4 squares
is, 
8 P m, n odd
m|n
r1 (n) = P
24
m|n,m odd m, n even

This combined with Legendre’s famous 4-square theorem tells us that every integer get a special
set of distinct representatives Sp of p + 1 elements (a, b, c, d) satisfying a2 + b2 + c2 + d2 = p for an
odd prime p ∈ Z.
Letting H(Fq ) denote the ring on quaternions with elements in the finite field of order q a prime,
take τq : H(Z) → H(Fq ) to be the map (a, b, c, d) 7→ (a, b, c, d) mod q.

1
The more the merrier.

13
Proposition 3.1 For char(Fq ) 6= 2, ∃x, y ∈ Fq such that x2 + y 2 + 1 = 0 and an isomorphism
ψq : H(Fq ) → M2 (Fq ) where M2 (−) denotes the algebra of 2 × 2 matrices with entries in the
argument, and 
a + bx + cy −by + c + dx
ψq (a + bi + cj + dk) =
−by − c + dx a − bx − dy

Letting φ : GL2 (Fq )  P GL2 (Fq ) be the standard projection map, efine Sp,q to be the image of
Sp under φ◦ψq ◦τq . It turns out that the Cayley graphs X p,q = Γ(P GL2 (Fq ), Sp,q ) are (p+1)-regular
Ramanujan graphs.
The fact that these are Ramanujan graphs was originally due to Lubotsky, Phillips and Sarnak [7].
Their proof depends on Eichler and Igusa’s ( [3] and [5]) work on an a conjecture of Ramanujan. If
rm (n) represents the number of integral solutions to
1
a2 + m2 b2 + m2 c2 + m2 d2 = n, then rm (n) = cm d|n d + O(n 2 + ), where cm depends only on the
P

index m.
Eichler’s contribution depends on the famed Weil Conjectures involving an analogue of the Riemann
Hypothesis set over finite fields.
It is the dependence on Ramanujan’s conjecture that lead Lubotsky, Phillips and Sarnak to name
Ramanujan graphs as such. While other constructions of Ramanujan graphs exist, they also rest
on highly non-trivial ideas from Algebraic Geometry and Number Theory. It wasn’t until 2013 that
the existence of Ramanujan graphs of all degrees was settled in [9].

4 A Proof of The Riemann Hypothesis! ...(for graphs)

The Euler product formula for the famed Riemann Zeta Function is given by
Y
ζ(s) = (1 − p−s )− 1
p prime

As the reader is surely aware, one need only show that the non-integer zeros of ζ(s) all have real
part 21 to earn themselves a cool million.
In analogy with the classical Zeta Function, we define for a graph Γ the Ihara Zeta Function.
Our notion of prime numbers in the product formula will be replaced by Prime Walks. We say
that a closed walk has a tail if the last edge traversed is the same as the first edge in the walk,
and call it backtrack-less if at no point an edge is traveled, and then immediately traversed in the
opposite direction. Finally we say that a closed backtrack-less walk with no tail is Prime if it’s not
a shorter backtrack-less, tail-less closed walk traced out multiple times. We say that two closed
back-trackless walks with no tail are equivalent if their vertex orderings are cyclic permutations of
each other, and let [p] denote equivalence class of all such walks. Now the Ihara Zeta Function is

14
given by, Y
ZΓ (s) = (1 − (k − 1)−sv(p) )− 1
[p]

The reader may complain that while we’ve constructed something that looks like the original zeta
function, there should be no reason to expect that ZΓ (s) has any relationship to the original zeta
function, or that it possesses any useful information for that matter.
The inspiration from this definition draws from another analogue of the Riemann Zeta function
which was defined by Selberg [10] for Riemannian manifolds. Selberg’s Zeta Function appears no
less absurd than Ihara’s Zeta function (in fact it appears more so!) and yet it does relate to the
original zeta function, it stores information depending on whether or not it’s roots have real part
1
2
. Selberg’s definition uses lengths special paths on a manifold known as geodesics, as it’s ”prime
numbers”. If one considers the rich analogies which manifolds and functional analysis, the above
definition of the Ihara Zeta function becomes less suspicious.
It turns out that the Ihara Zeta Function has a nice closed form formula given by

ZΓ (s) = (1 − (k − 1)−2s )r−1 det(I − (k − 1)−s A + (k − 1)1−2s I)−1

Where r is the Circuit Rank of the graph Γ, that is r = |EΓ| − |V Γ| + 1.

1
Theorem 4.1 The poles of ZΓ (s) are entirely one the line Re(z) = 2
iff Γ is Ramanujan.

1+(k−1)1−2s
Proof. Let pΓ (z) be the characteristic polynomial of Γ. Writing z = (k−1)−s
, one can see that
for some z0 ∈ C,

 1 + (k − 1)1−2s
 
−|V Γ|s −s −|V Γ|s −s

0 = pΓ (z) = |z0 I − A| ⇒ 0(k − 1) (k − 1) = (k − 1) (k − 1) 
 (k − 1)−s I − A


0 = |(1 + (k − 1)1−2s )I − (k − 1)−s A|

and thus z0 corresponds to a pole of ZΓ (s), and poles of ZΓ (s) will be zeros of pΓ (z).
Now suppose that for a given graph Γ, the poles of ZΓ (s) all have real part 21 , and let s0 = 21 + bi be
1 1−2s0
a such a point. Thus (k −1)|(k −1) 2 +bi |2 = |(k −1)bi | = 1. Let λ = 1+(k−1)
(k−1)−s0
be the corresponding
eigenvalue of A.
(k − 1)−s0 1 + (k − 1)1−2s0
λ =
(k − 1)−s0 (k − 1)−s0 (k − 1)−s0
(k − 1)−s0 + |(k − 1)−s0 |2 (k − 1)1−s0
= = (k − 1)((k − 1)−s0 + (k − 1)−s0 )
|(k − 1)−s0 |2
√
∴ |λ| = (k − 1)|(k − 1)−s0 + (k − 1)−s0 | ≤ 2 k − 1

15
 √
Conversely, suppose that Γ is Ramanujan. Thus any (non-trivial) eigenvalue λ, |λ| ≤ 2 k − 1,
which corresponds to a zero of pΓ /pole of ZΓ (s). Since all eigenvalues of A are real λ2 ≤ 4(k − 1).
1−2s
Since we set z = 1+(k−1)
(k−1)−s
if we take x = (k − 1)−s we get that (k − 1)x2 − λx + 1 = 0 and thus
by the quadratic formula p
λ ± λ2 − 4(k − 1)
x=
2(k − 1)
1
Thus if we have equality λ = 4(k − 1), x = √±1
k−1
⇒ (k − 1)2 = ±(k − 1) 2 and thus Re(s) = 21 .
If λ < 4(k − 1), then x is non-real, and writing λ in place of z0 in the arguments of the forwards
direction, we may write

(k − 1)−s + (k − 1)|(k − 1)−s |2 (k − 1)−s

λ=
|(k − 1)−s |2

but since λ is real, in order for (k − 1)−s + (k − 1)|(k − 1)−s |2 (k − 1)−s to have no imaginary part, it
must be that (k − 1)|(k − 1)−s |2 = 1 and so for s = a + bi, (k − 1)−2a |(k − 1)−bi | = (k − 1)−1 ⇒ a = 12 .
In either case the result holds.

References
[1] Brouwer. Haemers
The Spectra of Graphs. Springer, New York, Dordecht, Heidelberg, London (2012)

[2] John Conway, Derek Smith.

On Quaternions and Octonions: Their Geometry, Arithmetic, and Symmetry.
A.K. Peters, Ltd. Wellesley, MA (2003)

[3] M. Eichler
Quaternare Quadratische Formen und die Riemannsche Vermutung fur die Kangruentz Zeta
Funktion.
Archiv. der. Math. vol V, 355-366 (1954)

[4] Miroslav Fiedler

Algebraic connectivity of graphs
Czechoslovak Mathematical Journal, Vol. 23 (1973), No. 2, 298305

[5] J. Igusa
Fibre Systems of Jacobian Varieties III.
American Jnl. of Math. 81, 453-476 (1959)

16
[6] G. Davidoff, P. Sarnak, A. Valette
Elementary Number Theory, Group Theory, and Ramanujan Graphs.
London Mathematical Society (2003)

[7] Lubotzky, A., Phillips, R. Sarnak, P.

Combinatorica (1988) 8: 261. https://doi.org/10.1007/BF02126799

[8] Maxim Kontsevich,

Formal (non)commutative Symplectic Geometry.
The Gelfand Mathematical Seminars, 19901992, 173187, Birkhauser Boston, Boston, MA, 1993.

[9] Adam W. Marcus, Daniel A. Spielman, Nikhil Srivastava

Interlacing families I: Bipartite Ramanujan graphs of all degrees
Volume 182 (2015) Issue 1, Pages 307-325

[10] Atle Selberg,

Harmonic analysis and discontinuous groups in weakly symmetric Riemannian spaces with ap-
plications to Dirichlet series.
J. Indian Math. Soc. (N.S.), 20: 4787 (1956)

[11] Audrey Terras.

Zeta Functions and Chaos. http://math.ucsd.edu/ aterras/msripaper.pdf

[12] Ch’ing, W. Li, W.

A Survey of Ramanujan Graphs. https://math.byu.edu/cbms/sRgraph.pdf

17