Introduction to Spectral Theory

Laplacian
Defined as:

All column and row sums are zero:

T r[f
T
L = D − A;

L =

Semidefinite positve (f is an eigenfunction):

T r[L]

Lf ] =
1

⎢⎥
⎡

⎣
−a n1

∀i : ∑ L ij = 0

T r[f

n
d1

−a 21

∑ ∑ a ij (f i − f j )
i=1

This quantity is referred to as Dirichlet energy

If Lf = f − D
−1
j=1

Af ,
D = diag(d 1 , … , d n )

−a 12

d2

−a n2

then
and

Eigenfunctions of L (λ and f is the Fiedler eigenvalue/vector respectively):

Properties of the eigenvectors/values:

Lf 1 =

Lf 2 =

Lf n =

1
2
⋮
0

λ2 f2

λn fn

⊥ f 1 , f 3 ⊥ {f 1 f 2 }, ⋯

< λ2 ≤ ⋯ ≤ λn

The Fiedler eigenvalue minimizes the cut between two sets of nodes:
→
…

Lf ] ≥ 0, ∀f ∈ R

2
=

f (i) =
−a 1n

−a 2n

is ∝ to the variability imposed by the structure of the graph (Fiedler's Theorem):

dn

∀j : ∑ L ij

λ1 = 0
1
⎤

∑ (f i − f j )
i∼j

di
∑ a ij f (i)
j
2
 λ 2 = n ⋅ min {

Spectral decomposition of the Laplacian:

L = ΦΛΦ
T

at time t = 0, each node has one unit of heat

Dirichlet Energies:

Expanders and Sparsifiers:

E G (x) = x
T

∀x ∈ R : (1 − ϵ)x
n

Φ = [f 1 , f 2 , ⋯ , f n ],

Diffusion through heat kernels

i=1

vol(G)

LG x =
∑

T
n

⎢⎥
∑ i=1 ∑ j=1 a ij (x i − x j )

j=1

the Heat Kernel Signature (HKS) of a node is its heat trace over time
Heat H
time t.

Commute time
t
(i, j)

∑
= ∑

(u,v)∈E

LG x ≤ x
(x i − x j ) 2

Λ = diag(λ 1 , λ 2 , ⋯ , λ n )

Heat equation that describes information flow across edges of graph with time:

By solving it, we get the heat kernel:

K t = exp(−tL) = Φexp(−tΛ)Φ
T
∂h t

= Φ
∂t

⎡
e
= −Lh t

−tλ 1

0
⋮

i=2
e

λi
0

−tλ 2

0
⋮
2

: x ≠ c ⋅

is the probability that a lazy random walk starting at node i hits node j at

Defined as: The time needed by a random walk to hit j (hitting time) and return. Formally:

It has a direct relationship with:

Eigenvectors
CT (i, j) = H (i, j) + H (j, i)

CT uv
n
⋯

⋯ e

(f i (u) − f i (v))

(x u − x v )

T
2
= T r [X

L G ′ x ≤ (1 + ϵ)x
0

⋮
𝟙, c ∈ R}

−tλ n ⎦

T
⎤

L G X]

T
LG x
T
n

= ∑ e
i=1
n

L = ∑ λi fi fi
i=1

−tλ i
T

fi fi
T
 Transductive Rewiring > Cheeger constant
Curvature:

1 2(p u + p v )
p u := 1 − ∑ R uv κ uv :=
2 R uv
v∈N (u)

The commute time can be robust to deletion of individual edges.

Smooth eigenfunctions will contribute less, but the contribution of high variance
eigenfunctions is reduced by their large inverse λ . The equation expresses a weighted (by
i

effectively the pseudo-inverse of the Laplacian) Euclidian distance between the

This gives us the embedding of commute time:

−1/2 T
Θ = √ vol(G)Λ Φ

Resistance distance
Also, a relevant metric is the resistance distance between two vertices:

CT (i, j)
R(i, j) =
vol(G)