Sp ectral Partitioni ng Works:
Planar graphs and nite element meshes
Preliminary Draft
y
Daniel A. Spielman Shang-Hua Teng
February 13, 1996
Abstract
Sp ectral partitioning metho ds use the Fiedler vector|the eigenvector of the second-
smallest eigenvalue of the Laplacian matrix|to nd a small separator of a graph. These
metho ds are imp ortant comp onents of many scienti c numerical algorithms and have b een
demonstrated by exp erimenttowork extremely well. In this pap er, we show that sp ectral
partitioning metho ds work well on b ounded-degree planar graphs and nite element meshes|
the classes of graphs to which they are usually applied. While naive sp ectral bisection do es
not necessarily work, we prove that sp ectral partitioning techniques can b e used to pro duce
p
n for b ounded-degree planar separators whose ratio of vertices removed to edges cut is O
1=d
graphs and two-dimensional meshes and O n for well-shap ed d-dimensional meshes. The
heart of our analysis is an upp er b ound on the second-smallest eigenvalues of the Laplacian
matrices of these graphs.
1. Intro duction
Sp ectral partitioning has b ecome one of the most successful heuristics for partitioning graphs
and matrices. It is used in many scienti c numerical applications, such as mapping nite el-
ement calculations on parallel machines [Sim91, Wil90], solving sparse linear systems [PSW92 ],
and partitioning for domain decomp osition [CR87, CS93 ]. It is also used in VLSI circuit de-
sign and simulation [CSZ93 , HK92, AK95]. Substantial exp erimental work has demonstrated
that sp ectral metho ds nd go o d partitions of the graphs and matrices that arise in many ap-
plications [BS92, HL92 , HL93 , PSL90 , Sim91, Wil90]. However, the quality of the partition that
Computer Science Division, U.C. Berkeley, CA 94720, [email protected]. Supp orted in part byan
NSF p ostdo c. After July 1996, Department of Mathematics, M.I.T.
y
Department of Computer Science, University of Minnesota, Minneap olis, MN 55455, [email protected]. Sup-
p orted in part by an NSF CAREER award CCR-9502540 and an Intel research grant. Part of the work was done
while visiting XeroxPARC. 1
these metho ds should pro duce has so far eluded precise analysis. In this pap er, we will prove that
sp ectral partitioning metho ds give go o d separators for the graphs to which they are usually applied.
The size of the separator pro duced by sp ectral metho ds can b e related to the Fiedler value|the
second smallest eigenvalue of the Laplacian|of the adjacency structure to which they are applied.
2=d
,we By showing that well-shap ed meshes in d dimensions have Fiedler value at most O 1=n
1 1=d
. show that sp ectral metho ds can b e used to nd bisectors of these graphs of size at most O n
While a small Fiedler value do es not immediately imply that there is a cut along the Fiedler vector
that is a balanced separator, it do es mean that there is a cut whose ratio of vertices separated to
1=d
. By removing the vertices separated by this cut, computing a Fiedler vector of edges cut is O n
1 1=d
edges. In particular, the new graph, and iterating as necessary, one can nd a bisector of O n
we prove that b ounded-degree planar graphs have Fiedler value at most O 1=n, which implies that
p
sp ectral techniques can b e used to nd bisectors of size at most O n in these graphs. These
b ounds are the b est p ossible for well-shap ed meshes and planar graphs.
1.1. History
The sp ectral metho d of graph partitioning was b orn in the works of Donath and Ho man [DH72,
DH73 ] who rst suggested using the eigenvectors of adjacency matrices of graphs to nd partitions.
Fiedler [Fie73, Fie75a , Fie75b] asso ciated the second-smallest eigenvalue of the Laplacian of a graph
with its connectivity and suggested partitioning by splitting vertices according to their value in the
corresp onding eigenvector. Thus, we call this eigenvalue the Fied ler value and a corresp onding
vector a Fied ler vector.
A few years later, Barnes and Ho man [Bar82 , BH84] used linear programming in combination
with an examination of the eigenvectors of the adjacency matrix of a graph. In a similar vein, Bop-
pana [Bop87 ] analyzed eigenvector techniques in conjunction with convex programming. However,
the use of linear and convex programming made these techniques impractical for most applications.
By recognizing a relation b etween the Fiedler value and the Cheeger constant [Che70] of con-
tinuous manifolds, Alon [Alo86] and Sinclair and Jerrum [SJ89 ] demonstrated that if the Fiedler
value of a graph is small, then directly partitioning the graph according to the values of vertices
in the eigenvector will pro duce a cut with a go o d ratio of cut edges to separated vertices see
also [AM85, Fil91, DS91 , Mih89, Moh89 ]. Around the same time, improvements in algorithms
for approximately computing eigenvectors, such as the Lanczos algorithm, made the computation
of eigenvectors practical [PSS82 , Sim91]. In the next few years, a wealth of exp erimental work
demonstrated that sp ectral partitioning metho ds work well on graphs that usually arise in prac-
tice [BS92, HL92 , PSL90 , Sim91, Wil90]. Sp ectral partitioning b ecame a standard to ol for mesh
partitioning in many areas [HL93 ]. Still, researchers were unable to prove that sp ectral partition-
ing techniques would work well on the graphs encountered in practice. This failure is partially
explained by results of Guattery and Miller [GM95] demonstrating that naive applications of sp ec-
tral partitioning, such as sp ectral bisection, will fail miserably on some graphs that could con-
ceivably arise in practice. By b ounding the Fiedler values of the graphs of interest in scienti c
applications|b ounded-degree planar graphs and well-shap ed meshes|we are able to show that 2
sp ectral partitioning metho ds will successfully nd go o d partitions of these graphs.
In a related line of research, algorithms were develop ed along with pro ofs that they will always
nd small separators in various families of graphs. The seminal work in this area was that of
Lipton and Tarjan [LT79 ], who constructed a linear-time algorithm that pro duces a 1=3-separator
p
of 8n no des in any n-no de planar graph. Their result improved a theorem of Ungar [Ung51 ] which
p
demonstrated that every planar graph has a separator of size O n log n . Gilb ert, Hutchinson, and
Tarjan [GHT84] extended these results to show that every graph of genus at most g has a separator
p
of size O gn . Another generalization was obtained by Alon, Seymour, and Thomas [AST90],
p
3=2
who showed that graphs that do not haveanh-clique minor have separators of O h n no des.
3=2
Plotkin, Rao, and Smith [PRS94] reduced the dep endency on h from h to h. Using geometric
techniques, Miller, Teng, Thurston, and Vavasis [MT90, MTTV96a , MTTV96b , MTV91 , MV91,
Ten91 ] extended the planar separator theorem to graphs emb edded in higher dimensions and showed
d 1 1=d
has a 1=d + 2-separator of size O n . Using multicommo dity that every well-shap ed mesh in R
ow, Leighton and Rao [LR88 ] designed a partitioning metho d guaranteed to return a cut whose
ratio of cut size to vertices separated is within logarithmic factors of optimal. While sp ectral
metho ds have b een favored in practice, they lacked a pro of of e ectiveness.
1.2. Outline of pap er
In Section 2, weintro duce the concept of a graph partition, review some facts from linear algebra
that we require, and describ e the class of sp ectral partitioning metho ds.
In Section 3, we prove the embedding lemma, which relates the quality of geometric emb eddings
of a graph with its Fiedler value. We then show using the main result of Section 4 that every
planar graph has a \nice" emb edding as a collection of spherical caps on the surface of a unit
sphere in three dimensions. By applying the emb edding lemma to this emb edding, we prove that
the Fiedler value of every b ounded-degree planar graph is O 1=n.
In Section 4, we show that, for almost every arrangement of spherical caps on the unit sphere in
d
R , there is a sphere-preserving map that transforms the caps so that the center of the sphere is the
centroid of their centers. It is this fact that enables us to nd nice emb eddings of planar graphs.
In Sections 5 and 6, we extend our sp ectral planar separator theorem to the class of overlap
graphs of k -ply neighb orho o d systems emb edded in any xed dimension. This extension enables us
1=d
to show that the sp ectral metho d nds cuts of ratio O 1=n for k -nearest neighb or graphs and
well-shap ed nite element meshes.
In Section 7, we present an elementary pro of that from anyvector p erp endicular to the all-ones
vector with small Rayleigh quotient, one can obtain a cut of small ratio.
In Section 8, we extend the results of Guattery and Miller to show that our results are essentially
the b est p ossible given currentcharacterizations of well-shap ed meshes. We present natural families
of graphs for which Fiedler vectors can b e used to nd cuts of go o d ratio, but not go o d balance.
We discuss why these graphs exist and why they might not app ear in practice. 3
2. Intro duction to Sp ectral Partitioning
In this section, we de ne the sp ectral partitioning metho d and intro duce the terminology that we
will use throughout the pap er.
2.1. Graph Partitioning
Throughout this pap er, G =V; E will b e a connected, undirected graph on n vertices.
A partition of a graph G is a division of its vertices into two disjoint subsets, A and A. Without
loss of generality,we can assume that jAj A . Let E A; A b e the set of edges with one endp oint
in A and the other in A. The cut size of the partition A; A is simply jE A; Aj. The cut ratio,or
simply the ratio of the cut, denoted A; A, is equal to the ratio of the size of the cut to the size
of A, namely,
jEA; Aj
A; A= :
minjAj; jAj
The isop erimetric numb er of a graph, which measures how go o d a ratio cut one can hop e to nd,
is de ned to b e
E A; A
G= min :
jAjn=2
jAj
In Section 7, we describ e a relation b etween the isop erimetric numb er of a graph and its Fiedler
value.
A partition is a bisection of G if A and A di er in size by at most 1. For in the range
0 < 1=2, a partition is called a -separator if minjAj; jAj n.We use the word cut to refer
to a partition separating anynumber of vertices and reserve the word separator for partitions that
are -separators for some >0. Given an algorithm that can nd cuts of ratio in G and its
subgraphs, we can nd a bisector of G of size O n see Lemma 22.
2.2. Laplacians and Fiedler Vectors
The adjacency matrix, AG, of a graph G is the n n matrix whose i; j -th entry is 1 if i; j 2 E
and 0 otherwise. The diagonal entries are de ned to b e 0. Let D b e the n n diagonal matrix with
entries D = d , where d is the degree of the ith vertex of G. The Laplacian, LG, of the graph
i;i i i
G is de ned to b e LG=D A.
Let M be an n n matrix. An n-dimensional vector ~x is an eigenvector of M if there is a scalar
such that M~x = ~x. is the eigenvalue of M corresp onding to the eigenvector ~x.IfMis a real
symmetric matrix, then all of its n eigenvalues are real. The only matrices we consider in this pap er
will b e the Laplacians of graphs. Notice that the all-ones vector is an eigenvector of any Laplacian
matrix and that its asso ciated eigenvalue is 0. Because Laplacian matrices are p ositive semide nite,
all the other eigenvalues must b e non-negative. We will fo cus on the second smallest eigenvalue,
, of the Laplacian and an asso ciated eigenvector ~u. Fiedler called this eigenvalue the \algebraic
2
connectivity of a graph", so we will call it the Fied ler value and an asso ciated eigenvector a Fied ler
vector. 4
The following prop erties of Fiedler values and vectors play an imp ortant role in this pap er:
The Fiedler value of a graph is greater than zero if and only if the graph is connected.
A Fiedler vector ~u =u ; :::; u satis es
1 n
n
X
u =0;
i
i=1
b ecause all-ones vector is an eigenvector of the Laplacian and the eigenvectors of a symmetric
matrix are orthogonal.
The Fiedler value, ,of G satis es
2
T
~x LG~x
; = min
2
T
~x?1;1;:::;1
~x ~x
with the minimum o ccurring only when ~x is a Fiedler vector.
n
For anyvector ~x 2 R ,wehave
X
T 2
~x LG~x= x x :
i j
i;j 2E
Let M b e a symmetric n n matrix and ~x be an n-dimensional vector. Then, the Rayleigh
quotient of ~x with resp ect to M is
T
~x M~x
:
T
~x ~x
For pro ofs of these statements and many other fascinating facts ab out the eigenvalues and
eigenvectors of graphs consult one of [Str88, Moh88 , CDS90 , Big93].
The Fiedler value, , of a graph is closely linked to its isop erimetric numb er. If G is a graph
2
on more than three no des, then one can show [AM85, Moh89 , SJ88 ]
q
2
G 2d :
2 2
2
In Section 7, we will fo cus on the second inequality.
2.3. Sp ectral Partitioning Metho ds
Let ~u =u; :::; u b e a Fiedler vector of the Laplacian of a graph G. The idea of sp ectral
1 n
partitioning is to nd a splitting value s and partition the vertices of G into the set of i such that
u >sand the set such that u s. We call such a partition a Fied ler cut. There are several
i i
p opular choices for the splitting value s:
bisection: s is the median of fu ; :::; u g.
1 n 5
ratio cut: s is the value that gives the b est ratio cut.
sign cut: s is equal to 0.
gap cut: s is a value in the largest gap in the sorted list of Fiedler vector comp onents.
Other variations have b een prop osed.
In this pap er, we will analyze the sp ectral metho d that uses the splitting value that achieves
the b est ratio cut. We will show that, for b ounded-degree planar graphs and well-shap ed meshes, it
always nds a go o d ratio cut. In fact, it is not necessary to use a Fiedler vector; an approximation
will suce. If a vector ~x that is orthorgonal to the all-ones vector has a small Rayleigh quotient
with resp ect to the Laplacian of G, then ~x can b e used to nd a go o d ratio cut of G.
Guattery and Miller [GM95] showed that there exist b ounded-degree planar graphs on n vertices
with constant-size separators for which sp ectral bisection and sp ectral sign cuts give separators that
cut n=3 edges.
p
We will show that planar graphs have Fiedler cuts of ratio O 1= n. By Lemma 22, our result
p
implies that a bisector of size O n can b e found by rep eatly nding Fiedler cuts. In Section 8,
we extend the results of Guattery and Miller to show that this rep eated application of Fiedler cuts
is necessary,even for some quite natural graphs. We will show that, for any constant in the range
0 < 1=2, there are natural families of well-shap ed two-dimensional meshes that have no Fiedler
cut of small ratio that is also a -separator. We discuss why these graphs exist as well as why they
might fail to arise in practice. Our explanation is not entirely satisfactory and the problem remains
of nding a b etter characterization of the graphs that do arise in practice as well as those for which
p
there is a Fiedler cut that pro duces a -separator of size O n.
3. The eigenvalues of planar graphs
In this section, we will prove that the Fiedler value of every b ounded-degree planar graph is O 1=n.
Our pro of establishes and exploits a connection b etween the Fiedler value and geometric emb eddings
of graphs. We obtain the eigenvalue b ound by demonstrating that every planar graph has a \nice"
emb edding in Euclidean space.
p
A b ound of O 1= n can b e placed on the Fiedler value of any planar graph by combining the
planar separator theorem of Lipton and Tarjan [LT79 ] with the fact that =2 G. Bounds of
2
O 1=n on the Fiedler values of planar graphs were previously known for graphs such as regular grids
[PSL90 ], quasi-uniform graphs [GK95 ], and b ounded-degree trees. Bounds on the Fiedler values of
regular grids and quasi-uniform graphs essentially follow from the fact that the diameters of these
graphs are large see [Chu89]. Bounds on trees can b e obtained from the fact that every b ounded-
degree tree has a -separator of size 1 for some constant in the range 0 < <1=2 that dep ends
only on the degree. However, in order to estimate the Fiedler value of general b ounded-degree
planar graphs and well-shap ed meshes, we need di erent techniques.
p
T
x x.We relate We denote the standard l norm of a vector ~x in Euclidean space by k~xk =
2
the qualityofanemb edding of a graph in Euclidean space with its Fiedler value by the following
lemma: 6
Lemma 1 emb edding lemma. Let G =V; E beagraph. Then , the Fied ler value of G,is
2
given by
P
2
k~v ~v k
i j
i;j 2E
= min ;
P
2
n
2
k~v k
i
i=1
n
where the minimum is taken over vectors f~v ;:::;~v gR such that
1 n
n
X
~
~v = 0;
i
i=1
~
where 0 denotes the al l-zeroes vector.
n m
Remark 2. While we state this lemma for vectors in R , it applies equal ly wel l for vectors in R
for any m 1.
Pro of: Because the all-ones vector is the eigenvector of LG corresp onding to the eigenvalue 0,
can b e characterized by
2
P
2
x x
i j
i;j 2E
; = min
P
2
n
2
x
i
i=1
P
n
where the minimum is taken over real x 's such that x = 0. The minimum is achieved precisely
i i
i=1
when x ;:::;x is an eigenvector.
1 n
The emb edding lemmanow follows from a comp onent-wise application of this fact. Write ~v as
i
P
n
~
v ;:::;v . Then, for all f~v ;:::;~v g such that ~v = 0, wehave
i;1 i;n 1 n i
i=1
P P P
n
2 2
k~v ~v k v v
i j i;k j;k
i;j 2E i;j 2E
k =1
=
P P P
n n n
2
2
k~v k v
i
i=1 i=1 k =1 i;k
P P
n
2
v v
i;k j;k
i;j 2E
k =1
= :
P P
n n
2
v
k =1 i=1 i;k
But, for each k ,
P
2
v v
i;k j;k
i;j 2E
;
P
2
n
2
v
i=1 i;k
so
P P
n
2
v v
i;k j;k
i;j 2E
k =1
P P
2
n n
2
v
k =1 i=1 i;k
P P
this follows from the fact that x = y min x =y , for x ;y > 0. 2
i i i i i i i
i i
Our metho d of nding a go o d geometric emb edding of a planar graph is similar to the way
in which Miller, Teng, Thurston, and Vavasis [MTTV96a ] directly nd go o d separators of planar
graphs.
We rst nd an emb edding of the graph on the plane by using the \kissing disk" emb edding of
Ko eb e, Andreev, and Thurston [Ko e36, And70a , And70b , Thu88 ]: 7
Theorem 3 Ko eb e-Andreev-Thurston. Let G be a planar graph with vertex set V =
f1;:::;ng and edge set E . Then, there exists a set of disks fD ;:::;D g in the plane with disjoint
1 n
interiors such that D touches D if and only if i; j 2 E .
i j
Suchanemb edding is called a kissing disk emb edding of G.
The analogue of a disk on the sphere is a cap.Acap is given by the intersection of a half-space
with the sphere, and its b oundary is a circle. We de ne kissing caps analogously with kissing disks.
Following [MTTV96a ], we use stereographic pro jection to map the kissing disk emb edding of the
graph on the plane to a kissing cap emb edding on the sphere See Section 4 for more information
on stereographic pro jection. In Theorem 9, we will show that we can nd a sphere preserving map
map that sends the centroid also known as the center of gravity or center of mass of the centers
of the caps to the center of the sphere. Using this theorem, we can b ound the eigenvalues of planar
graphs:
Theorem 4. Let G be a planar graph on n nodes of degree at most d. Then, the Fied ler value of
G is at most
8d
:
n
p
Accordingly, G has a Fied ler cut of ratio O 1= n, and one can iterate Fied ler cuts to nd a
p
bisector of size O n.
Pro of: By Theorem 3 and Theorem 9, there is a representation of G by kissing caps on the unit
sphere so that the centroid of the centers of the caps is the center of the sphere. Let ~v ;:::;~v be
1 n
P
n
~
the centers of these caps. Make the center of the sphere the origin, so that ~v = 0.
i
i=1
Let r ;:::;r b e the radii of the caps. If cap i kisses cap j , then the edge from ~v to ~v will have
1 n i j
2 2 2
length at most r + r . As this is at most 2r + r , we can divide the contribution of this edge
i j
i j
between the two caps. That is, we write
n
X X
2 2
k~v ~v k 2d r :
i j
i
i=1
i;j 2E
But, b ecause the caps do not overlap,
n
X
2
r 4:
i
i=1
Moreover, k~v k = 1 b ecause the vectors are on the unit sphere.
i
Applying the emb edding lemma, we nd that the Fiedler value of G is at most
P
2
k~v ~v k
8d
i j
i;j 2E
P :
n
2
n kv~ k
i
i=1
Given the b ound on the Fiedler value, the ratio achievable by a Fiedler cut follows immediately
from Theorem 21 and the corresp onding bisector size follows Lemma 22. 2 8
Remark 5. One can prove a slightly weaker result by using a result of Mil ler, Teng, Thurston,
and Vavasis [MTTV96a] to nd a circle-preserving map that makes the center of the spherea
centerpoint [Ede87] of the images of particular points in the caps. If the center of the sphereisa
centerpoint, then the centroid must be far away from at leastaconstant fraction of the centers of
the caps. Thus, the numerator of the Rayleigh quotient wil l be the same as in Theorem 4, and the
denominator wil l be n.
Remark 6. The embedding lemma can be viewed as a semi-de nite relaxation of an integer program
for minimum balanced cut. The integer program would be
X
2
x x min
i j
i;j 2E
n
X
s:t: x =0; and
i
i=1
x 2f1g:
i
However, unlike the semi-de nite relaxation of MaxCut usedbyGoemans and Wil liamson [GW94],
we do not know if our relaxation provides a constant factor approximation of the optimum.
4. Sphere-preserving maps
P
n
d 2 d
Let B b e the unit ball in d dimensions: fx ;:::;x j x 1g. Let S denote the sphere
1 d
i=1 i
d d d
de ning the surface of B . This section is concerned with sphere-preserving maps from S to S .
d d
A sphere-preserving map from S to S is a continuous function that sends every sphere of lower
d d d
dimension contained in S to a sphere in S and such that every sphere in S has a pre-image
under the map that is also a sphere. Familiar sphere-preserving maps include rotations and the
map that sends each p oint to its antip o de.
We will make use of a slightly larger family of sphere-preserving maps. We obtain this family
d
by rst considering sphere-preserving maps b etween the sphere and the plane. Let H b e the
d d d
hyp erplane tangenttoS at 1; 0;:::;0. One can map H to S by stereographic projection:
d d
: H ! S by
d
z = the intersection of S with the line connecting z to 1; 0;:::;0: