ICCM 2007 · Vol.

II · 1–4

Four proofs for the Cheeger

inequality and graph partition
algorithms
∗†
Fan Chung

Abstract

We will give four proofs of the Cheeger inequality which relates

the eigenvalues of a graph with various isoperimetric variations of the
Cheeger constant. The first is a simplified proof of the classical Cheeger
inequality using eigenvectors. The second is based on a rapid mixing
result for random walks by Lovász and Simonovits. The third uses
PageRank, a quantitative ranking of the vertices introduced by Brin
and Page. The fourth proof is by an improved notion of the heat kernel
pagerank. The four proofs lead to further improvements of graph par-
tition algorithms and in particular the local partition algorithms with
cost proportional to its output instead of in terms of the total size of
the graph.

2000 Mathematics Subject Classification: 05C50, 68R10.

Keywords and Phrases: Cheeger inequalities, Laplacian, random
walks, heat kernel, Dirichlet eigenvalues, PageRank, graph partition
algorithms

1. Introduction
One of the major tools in spectral graph theory is the Cheeger
inequality. It concerns the relationship between two important graph in-
variants, the spectral gap and the Cheeger constant. For an (undirected)
graph G = (V, E), we denote by λG the spectral gap of the (normalized)
Laplacian (see [5]). Let hG denote the Cheeger constant of G, defined
as the minimum value h such that any cut separating a set of volume x
requires at least hx edges if x is less than half of the total volume. (The
∗ University of California, San Diego, USA
† Research supported in part by NSF Grants DMS 0457215 and ITR 0426858
2 F. Chung

detailed definitions will be defined later.) The Cheeger inequality states

that for a connected graph G, we have

h2G
2hG ≥ λG ≥ .
2
The proof of the above (discrete) Cheeger inequality is quite similar to
the proof in the continuous case in spectral geometry due to Jeff Cheeger
[4]. The idea of the proof is to use eigenvectors to guide the search for
good cuts. Indeed, the constructive proof gives the following expanded
version of the Cheeger inequality:

α2G h2
2hG ≥ λG ≥ ≥ G (11)
2 2
where αG is the minimum Cheeger ratio of the size of the edge boundary
and the volume of sets consisting of the vertices associated with the
largest i coordinates of the eigenvector which is associated with λG .
A somewhat simplified proof for (11) will be given in Section 3. In
the search for the optimum cut (evaluated by the Cheeger constant)
among an exponential number of possibilities, we can then focus on
a linear number of choices for the cut using the order determined by
the eigenvector. The above Cheeger inequality guarantees that the cut
resulting from this efficient algorithm has the Cheeger ratio within a
quadratic factor of the optimum. This spectral partition algorithm has
been widely used in numerous applications in the generic divide-and-
conquer approach by reducing problems recursively to smaller and more
manageable sizes [12].
Nowadays we often are dealing with problems involving graphs or
networks of prohibitively large sizes, such as the Web graph, various so-
cial networks, biological networks or many information networks arising
from massive data sets [9]. It is often infeasible to compute eigenvec-
tors for a graph with hundreds of thousands of vertices. For problems
involving such large graphs, the total number of vertices and edges of
the graph are no longer realistic parameters. Alternative approaches are
needed in a way that computation can be carried out “locally” and the
local optimum can be analyzed. As it turns out, one of the key ideas
rests on random walks.
In the early 90’s, Lovász and Simonovits derived an elegant result
on rapid mixing of random walks in their work on approximating the
volume of convex polytopes [13, 14]. Using this mixing result, Spielman
and Teng in 2004 gave a graph partition algorithm which has running
time proportional to the size of the output, and not depending on the
total number of vertices [15]. Although it was not specifically described
in the papers mentioned above, the basic thrust of this approach is a
 The Cheeger inequalities and graph partition algorithms 3

Cheeger inequality,
2
βG h2
2hG ≥ λG ≥ ≥ G (12)
8 8
where βG is the minimum Cheeger ratio of the sets determined by the
largest i values using the vector which is the probability distribution of
a random walk starting at a vertex v after k steps over all vertices v,
i ≤ n = |V (G)| and k ≤ (16 log n)/λ2G . Some local variations will be
further discussed in Section 5.
A third Cheeger inequality relies on the notion of PageRank, which
was first introduced by Brin and Page [3] as a method for quantitatively
ranking Webpages by the Web search engines. The PageRank can be
viewed as a combination of random walks, scaled by a parameter called
the jumping constant. (The detailed definition will be given later.) In
[1, 2], PageRank is used for developing a local partition algorithm which
improves upon the partition algorithm using random walks. PageRank
provides a way to deal with random walks of various lengths simulta-
neously in an organized fashion. The underlying theme here is again a
Cheeger inequality, although it was not mentioned in [1, 2]. In Section
6, we will show that for a subset S in G with vol(S) ≤ vol(G)/2, we have

γS2
hS ≥ λS ≥
4 log(vol(G))
where λS is the Dirichlet eigenvalues on S and γS denotes the mini-
mum of the Cheeger ratios determined by the PageRank involving only
vertices in S with appropriately chosen parameters. As a result, the lo-
cal partition algorithm using PageRank improves both the running time
and performance of the previous algorithms by a factor of powers of
log(vol(S)) (see [1]).
The fourth Cheeger inequality involves a notion of pagerank based
on the heat kernel of a graph. The heat kernel pagerank can be viewed
as an exponential sum of random walks while PageRank is a geometric
sum. The rate of diffusion of the heat kernel pagerank is controlled by a
heat parameter t ≥ 0. The heat kernel has many useful properties with
close relations to the spectrum of the graph. We will use a rapid mixing
inequality of the heat kernel to prove the following Cheeger inequality:
For a subset S of volume s ≤ vol(G)2/3 , we have

κ2S
hS ≥ λS ≥ (13)
8
where κS denote the minimum Cheeger ratio of subsets Si determined
in a sweep of the heat kernel pagerank associated with vertices in S with
vol(Si ) at most 2s for some appropriately chosen parameters for the heat
4 F. Chung

kernel pagerank. The details will be given in Section 7. This again leads
to an improved local partition algorithm.
In this paper, we examine four Cheeger inequalities and their proofs
using different analytic methods. In their entirety, it is of interest to
observe the comparisons as well as pointing out the interconnections.
The methods intertwine, ranging from their roots in differential geometry
through spectral graph theory, random walks, PageRank, heat kernels
to graph partition algorithms. These topics have been rapidly advancing
and emerging as main tools for the age of information and the Internet.

2. Preliminaries
There are basically two ways to give definitions for the heat ker-
nel and eigenvalues of a graph. One way is to use symmetric matrices
throughout which in some situations simplify the arguments (as in [5]).
However, in dealing with random walks, the transition probability ma-
trix W is not necessarily symmetric. Namely,

W = D−1 A

where D is the diagonal degree matrix and A denotes the adjacency

matrix. The normalized Laplacian L is defined by

L = I − D−1/2 AD−1/2 = D1/2 (I − W )D−1/2 .

Since the main results will be stated in the language of random walks,
we will use the unsymmetrical version. The spectral gap λG is the least
nonzero eigenvalue of L. The value λG can be expressed as the infimum
of the Rayleigh quotient:
P
(g(u) − g(v))2
λG = inf R(g) = inf u∼v P 2
u g (u)du
g g

where g ranges over functions defined on the vertices of G satisfying

P
u g(u)du = 0 (see [5]).
For aPset S, the volume of S, denoted by vol(S) is definedPto be
vol(S) = v∈S dv . The volume of G is written as vol(G) = v dv .
Let π denote π = (d1 /vol(G), d2 /vol(G), . . . , dn /vol(G)), indicated as a
row vector. For a connected non-bipartite graph, π is the stationary
distribution for the random walk W on G. In this paper, we consider
only connected graphs.
For a set S, the Cheeger ratio of S is defined to be

|∂S|
hS = ,
min{vol(S), vol(G) − vol(S)}
 The Cheeger inequalities and graph partition algorithms 5

where the boundary of S is denoted ∂S = {{u, v} ∈ E : u ∈ S and v 6∈

S}. Let S̄ denote the complement of S. Then hS = hS̄ . The Cheeger
constant of G is denoted by
hG = min hS .
S

Brin and Page introduced the notion of PageRank which has been
used as a major tool for conducting Web search. The definition of Page-
Rank can be entirely described in graph-theoretical terms since its orig-
inal application is in the Webgraph which has its vertex set consisting of
all webpages and edge set consisting of all hyperlinks. In a graph G, the
PageRank operator can be described as a combination of random walks
on G and a jumping constant c, 0 ≤ c ≤ 1. The PageRank matrix Rc is
defined to be
∞
X
Rc = c (1 − c)k Wk .
k=0
An equivalent definition is that the PageRank matrix satisfies the
following recurrence:
Rc = cI + (1 − c)Rc W. (21)
For a vertex u, let χu denote the (0, 1) indicator vector, i.e., χu (v) = 1
if v = u and 0 otherwise. The personalized PageRank has additional
parameters, the jumping constant and a preference vector. For example,
a typical preference vector associated with a vertex u is χu . All vec-
tors here are taken to be row vectors unless mentioned otherwise. For
a vertex u, the corresponding personalized PageRank, denoted by pru ,
is defined by pru = χu Rc , the matrix product of χu and Rc . The per-
sonalized PageRank can be viewed as quantitative ranking of all vertices
with respect to u. The original definition for PageRank has the prefer-
ence vector (1/n, 1/n, . . . , 1/n). In general, we can consider an arbitrary
starting vector s, summing to 1, in place of χu .
We will consider a new notion of pagerank which is based on the
heat kernel of a graph [8]. The heat kernel pagerank also has two pa-
rameters, the heat t and a seed distribution f . The heat kernel pagerank
ρt,f is the matrix product of f and the heat kernel Ht which is defined
as follows for t ≥ 0.

t2 2 tk
Ht = e−t (I + tW + W + . . . + W k + . . .)
2 k!
= e−t(I−W )
= e−tL
t2 2 tk
= I − tL + L + . . . + (−1)k Lk + . . . .
2 k!
6 F. Chung

Equivalently, Ht can be defined by the following heat equation:

∂
Ht = −(I − W )Ht .
∂t

3. The proof of the Cheeger inequality using

eigenvectors
The Cheeger inequality in (11) is the discrete analog of the classical
Cheeger inequality that appeared in Jeff Cheeger’s thesis in differential
geometry [4, 16]. The proof for the discrete version is in fact quite similar
to that in the continuous version. Several proofs for (11) can be found
in various contexts (see [5, 11]). Here we will give a proof that is simpler
than that in [5].

Theorem 1 In a graph G, the Cheeger constant hG and the spectral gap

λG are related as follows:

α2G h2
2hG ≥ λG ≥ ≥ G
2 2
where αG is the minimum Cheeger ratio of subsets Si consisting of ver-
tices with the largest i values in the eigenvector associated with λG , over
all i.

Proof: Suppose the Cheeger constant hG is achieved by the set S,

i.e., hG = hS = |∂S|/vol(S). By considering

vol(S)
g = χS − 1,
vol(G)

we have
λG ≤ R(g) ≤ 2hG .
Thus, the proofs for the Cheeger inequality mainly concerns a lower
bound for λG in terms of Cheeger ratios.
Now let g denote an eigenvector achieving λG . Namely, λG = R(g)
P
and v g(v)dv = 0. We order the vertices so that

g(v1 ) ≥ g(v2 ) ≥ . . . ≥ g(vn ).

Let Si = {v1 , . . . , vi } and define

αG = min hSi .
i
 The Cheeger inequalities and graph partition algorithms 7

Let r denote the largest integer such that vol(Sr ) ≤ vol(G)/2. Since
P
v g(v)dv = 0,

X X X
g(v)2 dv = min (g(v) − c)2 dv ≤ (g(v) − g(vr ))2 dv .
c
v v v

We define the positive and negative part of g − g(vr ), denoted by g+ and

g− , respectively, as follows:


g(v) − g(vr ) if g(v) ≥ g(vr ),
g+ (v) =
0 otherwise,


|g(v) − g(vr )| if g(v) ≤ g(vr ),
g− (v) =
0 otherwise.

We consider
P
(g(u) − g(v))2
λG = P
u∼v
2
v g(v) dv
P
(g(u) − g(v))2
≥ P u∼v 2
v (g(v) − g(vr )) dv
P

(g+ (u) − g+ (v))2 + (g− (u) − g− (v))2
≥ u∼v P
.
2 2
v g+ (v) + g− (v) du

Without loss of generality, we assume R(g+ ) ≤ R(g− ) and therefore we

have λG ≥ R(g+ ) since

a+b a b
≥ min{ , }.
c+d c d

We here use the notation

˜
vol(S) = min{vol(S), vol(G) − vol(S)}

so that

˜ i ).
|∂(Si )| ≥ αG vol(S
8 F. Chung

Then we have
λG ≥ R(g+ )
P
(g+ (u) − g+ (v))2
= u∼vP 2
u g+ (u)du
P 2

P 2


u∼v (g+ (u) − g+ (v)) u∼v (g+ (u) + g+ (v))
= P 2 P 2
u g+ (u)du u∼v (g+ (u) + g+ (v))
P 2 2

2
u∼v (g+ (u) − g+ (v) )
≥ P 2
2 by the Cauchy-Schwarz inequality,
2 u g+ (u)du
P 2 2

2
i |g+ (vi ) − g+ (vi+1 ) | |∂(Si )|
= P 2
2 by counting,
2 u g + (u)d u
P

|g (v )2
− ˜ i )| 2
g+ (vi+1 )2 |αG |vol(S
i + i
≥ P 2
2 by the def. of αG ,
2 u g+ (u)du
P 2

2
α2G ˜ ˜
i g+ (vi ) (|vol(Si ) − vol(Si+1 )|)
= P 2
2
2
u g+ (u)du
P
2
α2 g+ (vi )2 dvi
= G Pi
2
2 g 2 (u)du
u +
α2G
= .
2
Therefore we have proved the Cheeger inequality in (11):
α2G h2
2hG ≥ λG ≥ ≥ G.
2 2

4. Isoperimetric properties of random walks

In this section, we will describe some useful isoperimetric proper-
ties for random walks, which originated from [13, 14]. Both the second
and third proofs of Cheeger inequalities in (12) and (13) will use these
properties.
For a function f : V → R, we define
 f (u)
du if u is adjacent to v;
f (u, v) =
0 otherwise.
For a set T of pairs of vertices of G, we extend the domain of f to include
all subsets of the edge set as follows.
X
f (T ) = f (u, v).
(u,v)∈T
 The Cheeger inequalities and graph partition algorithms 9

For a set S of vertices of G, we define

Sin = {(u, v) : v ∈ S, u is adjacent to v},
Sout = {(u, v) : u ∈ S, u is adjacent to v}.
Also, we define X
f (S) = f (u).
u∈S
It follows from the definition that
f (S) = f (Sout ).
For any positive value x, we further extend the domain of f to include
the set of reals between 0 and vol(G) by defining:
f (x) = P sup
vc =x
u,v
cu,v f (u, v),

over cu,v with 0 ≤ cu,v ≤ 1. Suppose we order the vertices so that

f (v1 ) f (v2 ) f (vn )
≥ ≥ ... ≥ .
dv1 dv2 dvn
Let Si denote the set consisting of v1 , . . . , vi . Let hf denotes the least
Cheeger ratio hSi over all Si determined by f .
Then the following facts result directly from the above definitions.
Fact 1
P
(i) If x = j≤i dvj + rdvi+1 for 0 ≤ r < 1, then
X
f (x) = f (vi ) + rf (vi+1 )
j≤i
= (1 − r)f (vol(Si )) + rf (vol(Si+1 )).
(ii) f (x) is convex in x.
Let W denote the lazy walk defined by
I +W
W= .
2
The following lemma on lazy walks is in the same spirit as that in Lovász
and Simonovits [13, 14] and also in [1, 6]. For the sake of completeness,
we include a simplified proof here.
Lemma 1 For a subset S ⊆ V , we have
f (Sin ) + f (Sout )
f W(S) =
2
f (vol(S)(1 + hS )) + f (vol(S)(1 − hS ))
≤ .
2
10 F. Chung

Proof: We note that

f (S) + f (Sin )
f W(S) =
2
f (Sout ) + f (Sin )
=
2
f (Sin ∪ Sout ) + f (Sin ∩ Sout )
=
2
f (vol(S) + |∂S|) + f (vol(S) − |∂S|)
≤
2
f (vol(S)(1 + hS )) + f (vol(S)(1 − hS ))
≤ .
2

Thus, from the definitions and the convexity of f , we have

Lemma 2 For a function f : V → R, we have, for 0 ≤ x ≤ vol(G)/2,

f (x(1 + hf )) + f (x(1 − hf ))
f W(x) ≤ . (41)
2

5. A proof of the Cheeger inequality using

random walks
Here we will give a short proof for the following lemma which is
essentially the mixing result of Lovász and Simonovits in [13, 14].

Lemma 3 In a graph G, a subset S of vertices with vol(S) ≤ vol(G)/2

satisfies that for any vertex u,
s
vol(S) β2
|W (u, S) − π(S)| ≤
k
(1 − k )k (51)
du 8

where βk is the minimum Cheeger ratio over sets determined by the i

largest values of the distribution of the lazy random walk starting at u
after k steps. In other words,
0
βk = inf{hf : f = χu Wk for u ∈ V and k 0 ≤ k}.

Proof: For a fixed u, we choose f (v) = fk (v) = Wk (u, v) − π(v) as

in (41). We will prove (51) by induction on k. For the case of k = 0, it
 The Cheeger inequalities and graph partition algorithms 11

is easy to see that (51) holds. By Lemma 1, we have

fk+1 (x) = fk W(x)

fk (x(1 + βk+1 )) + fk (x(1 − βk+1 ))
≤
r p2 p
x βk2 k 1 + βk+1 + 1 − βk+1
≤ (1 − )
du 8 2
r 2
x β
≤ (1 − k+1 )k+1 .
du 8

√ √
Here we use the fact that 1 + y + 1 − y ≤ 2 − y 2 /4 for 0 ≤ y ≤ 1.
To finish the proof of Lemma 3, we consider g(v) = gk (v) = −fk (v). We
can prove in a similar way that
s
vol(S) β2
gk (v) ≤ (1 − k )k .
du 8

The proof can be done by induction on k and we note that it is true for
k = 0 since g0 (x) ≤ 1. This finishes the proof for (51). 
A special case of (51) is the following:

r
dv β2
|W (u, v) − π(v)| ≤
k
(1 − k )k . (52)
du 8

Theorem 2 For a graph G, we have the following Cheeger inequality:

2
βG h2
2hG ≥ λG ≥ ≥ G,
8 8

where
16 log n
βG = min{βt : t ≤ d e}.
λ2G

Let a left eigenvector of I − W associated with λG be denoted by

φ. Note that φ is orthogonal to the right eigenvector 1. On one hand,
we have

λG
t
kφ(Wt − 1∗ π)D−1/2 k = kφWt D−1/2 k = 1 − kφD−1/2 k.
2
12 F. Chung

On the other hand, we have

X 1
2
kφ(Wt − 1∗ π)D−1/2 k2 = φ(Wt − 1∗ π)(v) √
v
dv
X 1 X
2
= φ(u) Wt (u, v) − π(v)
v
dv u
X 1 X 2
≤ |φ(u)| |Wt (u, v) − π(v)|
v
dv u
r
X 1 X βt2 t dv 2
≤ |φ(u)|(1 − )
v
dv u 8 du
βt2 2t X X φ(u)2

≤ (1 − ) n
8 v u
du
βt2 2t 2
= (1 − ) n kφD−1/2 k2 .
8
Together we have
λG β2
1− ≤ (1 − t )n1/t .
2 8
This implies:
βt2 2 log n
2hG ≥ λG ≥ − .
4 t
Therefore for
16 log n
βG = min{βt : t ≤ d e},
λ2G
we have
2
βG h2
2hG ≥ λG ≥ ≥ G
8 8
as desired. 

6. Proving the Cheeger inequality using the

PageRank
The third proof of the Cheeger inequality is based on the notion of
PageRank. The personalized PageRank has two parameters, the jumping
constant c and a preference vector.
For a vertex u, the personalized PageRank pru associated with u
has the preference vector χu and we write pru = χu Rc . Although pru is
defined on the vertex set V of G = (V, E), we can extend the domain of
f = pru to V ∪ 2V ∪ E ∪ 2E ∪ [0, vol(G)] as defined in Section 4.
 The Cheeger inequalities and graph partition algorithms 13

We will use Lemma 1 and (21) by choosing f = pru − π. We have,

for any set S of vertices,
1 1

f (S) = c(1 − π(S)) + (1 − c) f (Sin ∩ Sout ) + f (Sin ∪ Sout ) .
2 2
We will establish the following inequality with a simpler proof than that
in [1]:
Lemma 4 For any positive integer k, a jumping constant c, 0 ≤ c ≤ 1,
a vertex u and a subset S of vertices, the personalized pagerank pru
satisfies
s
 vol(S) γ 2
k 
pru (S) − π(S) ≤ 1 − (1 − c) +
k
1 − u (1 − c)k (1 − π(S)),
du 8

where γu is the minimum Cheeger ratio determined by a sweep of pru .

Proof: It suffices to show that for any real value x, 0 ≤ x ≤ vol(G)/2,
f (x) = (pru (x) − π(x))/(1 − π(x)) satisfies
r
x γ 2
k
f (x) ≤ 1 − (1 − c) +
k
1 − u (1 − c)k . (61)
du 8

First we observe that (61) holds for the case that k = 0 since, for x ≥ du ,
we have f (x) ≤ 1, and for x ≤ du ,
r
x x
f (x) = f (du ) ≤ .
du du

Suppose that (61) holds for some k ≥ 0. We wish to establish the

inequality for k + 1 for x = vol(Si ).

f (x) ≤ c + (1 − c)f W(x)

f (x(1 − γu )) + f (x(1 + γu ))

≤ c + (1 − c) by Lemma 1
2 p p
( x(1 − γu ) + x(1 + γu )) γ2

≤ c + (1 − c) 1 − (1 − c)k + √ (1 − u )k (1 − c)k
2 du 8
by induction
r 2
x γ
≤ 1 − (1 − c)k+1 + (1 − u )k+1 (1 − c)k+1
du 8
√ √
by using the fact that 1 + w + 1 − w ≤ 2(1 − w2 /8) for 0 ≤ w ≤ 1.
We have proved (61). To show that |f (x)| has the same upper bound,
we consider g(x) = −f (x) and apply the same inductive proof. 
We also need the following for later lower bound arguments:
14 F. Chung

Lemma 5 For a subset S of vertices in G with vol(S) ≤ vol(G)/2, there

is a subset T ⊂ S with vol(T ) ≥ vol(S)/2 such that for any u ∈ T , the
personalized pagerank pru satisfies that

(1 − c)hS
pru (S) ≥ 1 − .
c
Proof: We first consider

χS DRc (S) = χS D(cI + (1 − c)DWRc )χ∗S

1−c
= cvol(S) − χS (D − A)Rc χ∗S + (1 − c)χS DRc (S).
2
Therefore we have
1−c
χS DRc (S) = vol(S) − χS (D − A)Rc χ∗S
2c
1−c X
= vol(S) − (χS (u) − χS (v))(χS Rc∗ (u) − χS Rc∗ (v))
2c u∼v
1−c X
= vol(S) − |χu Rc (S) − χv Rc (S)|
2c
{u,v}∈∂S
1−c
≥ vol(S) − |∂S|.
2c
P
Here we use the fact that f (D − A)g ∗ = u∼v (f (u) − f (v))(g(u) − g(v).
This implies that

χS DRc (S̄) = vol(S) − χS DRc (S)

1−c
≤ |∂S|.
2c
We consider a subset T 0 of S defined by
1−c
T 0 = {v ∈ S : χv Rc (S̄) ≥ hS }
c
and we have

χS DRc (S̄) ≥ χT 0 DRc (S̄)

X
= dv χv Rc (S̄)
u∈T 0
X 1−c
≥ dv hS
c
v∈T 0
1−c
≥ vol(T 0 )hS .
c
 The Cheeger inequalities and graph partition algorithms 15

Therefore, we have vol(T 0 ) ≤ vol(S)/2. We define T = S \ T 0 . Thus for

u in T , we have
1−c
pru (S) = 1 − χu Rc (S̄) ≥ 1 − hS
c
as claimed. 
We are now ready to prove the following local version of the Cheeger
inequality using PageRank.

Theorem 3 For a subset S in G with vol(S) ≤ vol(G)/2, there is a

subset T with vol(T ) ≥ vol(S)/2 such that for any u in T , we have

γu2
hS ≥
32 log(vol(S))

where γu denotes the Cheeger ratio determined by the PageRank pru with
jumping constant c = γu2 /(16 log(vol(S))).

Proof: By combining Lemma 4 and Lemma 5, there is a subset T of

S with vol(T ) ≥ vol(S)/2 such that for u ∈ T , we have

(1 − c)hS
1− − π(S)
 c 
vol(S) γu2 k
≤ 1 − (1 − c) +
k
(1 − ) (1 − c) (1 − π(S)).
k
du 8

This implies

hS (1 − c) vol(S) γ2

≥ (1 − c)k 1 − (1 − u )k . (62)
c(1 − π(S)) du 8

We then choose k and c as follows:

16 log(vol(S)) 1
k=d e, c= .
γu2 k

Then (62) implies

c γu2
hS ≥ ≥ ,
2 32 log(vol(S))
as desired. 
The above local Cheeger inequality suggests that by choosing one or
more random vertices near the seed, the PageRank associated with these
vertices can be used to find good cuts with high probability. Of course,
the above inequality does not involve any eigenvector. This correlation
can be added as follows:
16 F. Chung

For a given subset S ⊆ V , A function f : V → R is said to satisfy

the Dirichlet boundary condition if f (v) = 0 for all v 6∈ S. The Dirichlet
eigenvalue, denoted by λS is defined by
λS = inf R(f ) (63)
f

where f ranges over all nontrivial function f which satisfy the Dirichlet
boundary condition for S.
Theorem 4 For a subset S in G with vol(S) ≤ vol(G)/2 and a constant
c ≤ 1/2, we have
γS2
hS ≥ λS ≥
8 log(vol(S))
where γS denotes the minimum of the Cheeger ratios determined by the
PageRank pru for u in S and the jumping constant c ≥ γS2 /(8 log(vol(S))).
Proof: Let S denote the subset achieving the Cheeger constant hG .
From the definition in (63), we have hG = R(χS ) ≥ λS . To prove the
lower bound, we consider WS which has entries WS (u, v) the same as
W(u, v) if u, v are both in S and 0 otherwise. We define
X
Rc0 = c (1 − c)k WSk
k

so that
Rc0 = cIS + (1 − c)Rc0 WS . (64)
Clearly, by Lemma 4, we have
χu Rc0 χS ≤ χu Rc χS
s
vol(S) γ2
≤ 1 − (1 − c)k + (1 − S )k (1 − c)k .
du 8
P
Let ϕ denote the function achieving λS and ϕχS = u∈S ϕ(u) = 1. It
is not hard to see we can choose that ϕ(u) ≥ 0 for all u ∈ S. Then,
X
ϕRc0 χS ≤ ϕ(u)χu Rc0 χS
u∈S
p γ2
≤ 1 − (1 − c)k + vol(S)(1 − S )k (1 − c)k .
8
Note that from (64), we have
1−c

ϕRc0 χS = ϕ IS − RS (IS − WS ) χS
c
(1 − c)λS
= 1− .
c + (1 − c)λS
 The Cheeger inequalities and graph partition algorithms 17

By combining the preceding two inequalities, we get

(1 − c)λS p γ2
(1 − c)k ≤ + vol(S)(1 − S )k (1 − c)k . (65)
c + (1 − c)λS 8
We can now select k so that
p γS2 k
vol(S)(1 − ) ≤ 1/2.
8
Namely, k is chosen to be k = 4γS−2 log(vol(S)).
Then (65) implies that

(1 − c)k (1 − c)λS
≤ .
2 c + (1 − c)λS
Therefore
1 c
ck ≥ log ≥ .
1 − c/((1 − c)λS ) (1 − c)λS
We then have
1
λS ≥
(1 − c)k
γS2
≥ .
8 log(vol(S))
The proof of Theorem 4 is complete. 

7. The Cheeger inequality using the heat ker-

nel
For a vertex u, we define the heat kernel pagerank ρt,u to be

ρt,u (v) = χu Ht χ∗v

where Ht is as defined in Section 2. For a positive value s, we define a s-

local Cheeger ratio of ρt,u to be the minimum Cheeger ratio of cuts that
separate sets Si which consist of the vertices v that have the i largest
values of ρt,u (v)/dv and so that vol(Si ) is at most 2s.
We will use the following rapid mixing result on ρt,u (see [7]): In a
graph G, for a positive value t ≥ 0, a vertex u and a subset S of volume
s ≤ vol(G)/4, we have
r
s −tκ2t,u,s /4
ρt,u (S) − π(S) ≤ e (76)
du
18 F. Chung

where κt,u,s denotes the minimum s-local Cheeger ratio of ρt,u .

We will prove the following local Cheeger inequality that relates
the Cheeger ratio hS of a subset S with the Dirichlet eigenvalue λS and
the local Cheeger ratios.
Theorem 5 In a graph G, for a subset S of volume s ≤ vol(G)2/3 , we
have
κ2
hS ≥ λS ≥ S
8
where κS denotes the minimum s-local Cheeger ratio of ρt,u over all u
in S and t = blog(vol(G)/s)/λS c.
Proof: We consider a left eigenvector ϕ of IP− W which is associated
with the Dirichlet eigenvalue λS and satisfies u∈S ϕ(u) = 1. By using
(76), we have
X
ϕ(Ht − 1∗ π)χ∗S = ϕ(u)ρt,u (S) − π(S)
u∈S
r
s −tκ2t,S /4
≤ e .
du
We consider
∞ k
X t Wk
Ht0 = e−t S
k!
k=0
where WS is as defined in the previous section. It is not difficult to see
that Ht (u, v) ≥ Ht0 (u, v) for all u and v. We have
ϕ(Ht − 1∗ π)χ∗S ≥ ϕ(Ht0 − 1∗ π)χ∗S
s
= e−tλS − .
vol(G)
We choose t satisfying
s
e−tλS ≥ 2 .
vol(G)
By considering t ≤ log(vol(G)/s)/λS , we have
1 −tλS
e ≤ ϕ(Ht − 1∗ π)χ∗S
2 r
s −tκ2t,S /4
≤ e .
du
This implies,
κ2t,S log s
λS ≥ −
4 2t
κ2t,S
≥ − λS
4
 The Cheeger inequalities and graph partition algorithms 19

log s
if t ≥ 2λS . Such t exists if s ≤ vol(G)2/3 . This leads to

κ2t,S
λS ≥
8
for t = blog(vol(G)/s)/λS c. The proof of Theorem 5 is complete. 
For applications, the following modified Cheeger inequality can be
derived along the similar spirit of Theorem 4. Its proof is contained in
[7] and will be omitted.
Theorem 6 In a graph G for a subset of volume s, s ≤ vol(G)/4 and
Cheeger ratio hS ≤ φ2 /4, there is a subset S 0 ⊂ S with vol(S 0 ) ≥ s/2
such that for any u ∈ S 0 , the sweep by using the heat kernel pagerank
−2
√ t = dφ /4e, will find a set T with s-local Cheeger ratio at
ρt,u , with
most φ log s.

8. Graph partition algorithms

The four Cheeger inequalities and their constructive proofs lead
to a number of graph partition algorithms. These algorithms are one-
sweep algorithms so that the output is a subset consisting of vertices
associated with the highest i values determined by some function (which
sometimes is adjusted by the degrees of the vertices). Except for the
spectral partition algorithm, all the other three methods, using random
walks, PageRank and heat kernel pagerank yield local partition algo-
rithms.
Before proceeding to discuss local partition algorithms, we shall
mention the mode of computing. The input of all our algorithms of
course includes the whole graph (which can be very large) which is com-
puted off-line. For example, for search engines, such as Google, Yahoo or
Baidu, the vertices and edges of the Webgraph are meticulously, exhaus-
tively and continuously computed in an off-line mode and then compiled
and stored (in some appropriate data structure which we will not get
into here). When one uses the search engine by inputting a word or key
phrase (as the seed), the search is in on-line mode since one expects the
results of the search in a fraction of a second. The output of an on-line
algorithm is usually quite small in comparison with the whole size of
the data corpus. The local algorithms are all on-line algorithms which
have running time proportional to the sizes of their outputs. In order to
achieve this, the algorithmic steps for a local algorithm usually rely on
local operations such as propagating through the neighbors iteratively.
The spectral partition algorithm.
The one sweep algorithm uses the eigenvector that achieves the spectral
gap as stated in Theorem 1. The Cheeger inequality in (11) guarantees
20 F. Chung
√
that the output has its Cheeger ratio αG which is no more than 2 hG .
The running time of this algorithm is basically the same as the running
time for matrix multiplication, of order O(nω ), ω = 2.376 . . . (see [10])
where n is the number of vertices in G.
Graph partition algorithm using random walks
Spielman and Teng [15] gave a graph partition algorithm which, for a
graph G containing a set of vertices such that hS ≤ Φ and vol(S) ≤
vol(G)/2, can find a subset T with vol(T ) ≥ vol(S)/2 and hT ≤ φ
where Φ = O(φ3 / log2 (vol(G))). The partition algorithm contains quick
approximations of distributions after steps of random walks. The algo-
rithmic analysis uses the ideas of Lovász and Simonovits in [13, 14]. The
running time of their algorithm is O(vol(G) log6 (vol(G))/φ5 ). A main
subroutine NIBBLE in their algorithm has inputs including a seed, a
target volume k and a target Cheeger ratio Φ. The Nibble, which runs
in time O(k log4 (vol(G))/φ5 ), outputs a subset T with hT ≤ φ, where
Φ = O(φ3 / log2 (vol(G))), and the volume is within a factor of 2 of the
target volume.
Graph partition algorithm using PageRank
The local partition algorithm of Andersen, Chung and Lang uses a fast
approximation of PageRank [1]. The algorithm basically follows from the
modified Cheeger inequality in Theorem 3 and improves the previous al-
gorithm using random walks both in its performance and in running
time. For any set S with Cheeger ratio Φ, there are a large number of
starting vertices within S for which a sweep over pan appropriate Page-
Rank vector find a subset with Cheeger ratio O( Φ log vol(S)), in time
O(vol(S) log2 (vol(G))/Φ). The local graph partition algorithm can be
used iteratively to derive a general graph partition
p algorithm which will
find a cut which has Cheeger ratio at most hG log(vol(G)) in time
O(vol(G) log4 (vol(G))/hG ). There is a second paper by the same au-
thors [2] that further simplifies the local partition algorithm by focusing
on cuts which are associated with “sharp” drops in the sorted Page-
Rank functions. The algorithmic analysis is also simpler although the
computational complexity stays of the same order.
Graph partition algorithm using heat kernel pagerank
The Cheeger inequalities in Theorems 5 and 6 suggest a graph parti-
tion algorithm in a straightforward way. This improves the previous
algorithm by removing the factor of log(vol(G)) in the estimate of the
Cheeger ratio, in comparison with the previous algorithm using Page-
Rank. In order to obtain an even more efficient algorithm, we need to
get an efficient approximation of the heat kernel pagerank which the
formulation as an exponential sum can easily yield. For any set S with
Cheeger ratio Φ, there are a large number of starting vertices within S
 The Cheeger inequalities and graph partition algorithms 21

so that the associated heatpkernel pagerank can be use to find a cut with
Cheeger ratio at most O( Φ log(vol(S))) (see [7]). The running time
is basically O(vol(S) log(vol(S))/Φ) for approximating and sorting the
heat kernel pagerank with a support no more than the target volume.

References
[1] R. Andersen, F. Chung and K. Lang, Local graph partitioning using
pagerank vectors, Proceedings of the 47th Annual IEEE Symposium
on Founation of Computer Science (FOCS’2006), 475–486.
[2] R. Andersen, F. Chung and K. Lang, Detecting sharp drops in
PageRank and a simplified local partitioning algorithm, Theory and
Applications of Models of Computation, Proceedings of TAMC 2007,
1–12.
[3] S. Brin and L. Page, The anatomy of a large-scale hypertextual
Web search engine, Computer Networks and ISDN Systems, 30 (1-
7), (1998), 107–117.
[4] J. Cheeger, A lower bound for the smallest eigenvalue of the Lapla-
cian, Problems in Analysis (R. C. Gunning, ed.), Princeton Univ.
Press (1970), 195–199.
[5] F. Chung, Spectral Graph Theory, AMS Publications, 1997.
[6] F. Chung, Random walks and local cuts in graphs, LAA 423 (2007),
22–32.
[7] F. Chung, The heat kernel as the pagerank of a graph, PNAS, to
appear.
[8] F. Chung and S.-T. Yau, Coverings, heat kernels and spanning trees,
Electronic Journal of Combinatorics 6 (1999), #R12.
[9] F. Chung and L. Lu, Complex Graphs and Networks, CBMS Re-
gional Conference Series in Mathematics, 107, AMS Publications,
RI, 2006. viii+264 pp.
[10] D. Coppersmith and S. Winograd, Matrix multiplication via arith-
metic progressions, J. Symbolic Comput. 9 (1990), 251–280.
[11] M. Jerrum and A. J. Sinclair, Approximating the permanent, SIAM
J. Computing 18 (1989), 1149-1178.
[12] R. Kannan and S. Vempala and A. Vetta, On clusterings: Good,
bad and spectral, JACM 51 (2004), 497–515.
[13] L. Lovász and M. Simonovits, The mixing rate of Markov chains,
an isoperimetric inequality, and computing the volume, 31st IEEE
Annual Symposium on Foundations of Computer Science, (1990),
346–354.
[14] L. Lovász and M. Simonovits, Random walks in a convex body and
an improved volume algorithm, Random Structures and Algorithms
4 (1993), 359–412.
22 F. Chung

[15] D. Spielman and S.-H. Teng, Nearly-linear time algorithms for graph
partitioning, graph sparsification, and solving linear systems, Pro-
ceedings of the 36th Annual ACM Symposium on Theory of Com-
puting, (2004), 81–90.
[16] R. M. Schoen and S. T. Yau, Differential Geometry, International
Press, Cambridge, Massachusetts, 1994 .