Ordering Trees by Algebraic Connectivity : Graphs and Combinatorics

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Graphs and Combinatorics 6, 229-237 (1990)

Graphs and
Combinatorics
© Springer-Verlag 1990

Ordering Trees by Algebraic Connectivity*


Robert Grone 1 and Russell Merris 2
1 Department of Mathematical Sciences, San Diego State University, San Diego, CA 92182, USA
Department of Mathematics and Computer Science, California State University, Hayward, CA
94542, USA

Abstract. Let G be a graph on n vertices. Denote by L(G) the difference between the diagonal
matrix of vertex degrees and the adjacency matrix. It is not hard to see that L(G) is positive
semidefinite symmetric and that its second smallest eigenvalue, a(G)> 0, if and only if G is
connected. This observation led M. Fiedler to call a(G)the algebraicconnectivityof G. Given two
trees, Ti and T2, the authors explore a graph theoretic interpretation for the difference between
a(T~)and a(T2).

Let G = (V, E) be a graph with vertex set V = {1, 2 .... , n} and edge set E. Denote
by D(G) the diagonal matrix of vertex degrees and by A(G) the adjacency matrix.
Then the corresponding Laplacian matrix is L(G)= D(G)- A(G). In statistical
models of polymer molecules, L(G) is known as a " Z i m m matrix" [ 12] or "Kirchhoff
matrix" [5; 7, p. 152]. Also called a "matrix of admittance" [6; 7], L(G) first appeared
in the Matrix-Tree Theorem: If Lii is the submatrix of L(G) obtained by deleting
its i-th row and j-th column, then ( - 1)i+Jdet(Lii)is the number of spanning trees
in G (see, e.g., [3] or I-6]. More recent work involving L(G) has been done in
[1; 2; 4; 8 - 1 l ; and 13-20]. Note, especially, the survey paper [17].)
It is easy to see (e.g., by the Gerggorin Circle Theorem) that L(G) is positive
semidefinite. Moreover, the column n-tuple e, each of whose entries is 1, is in the
kernel of L(G). Denote the eigenvalues of L(G) by 21 _> 22 > " " > 2, = 0. F r o m the
Matrix-Tree Theorem, one may deduce that 2,_ 1 > 0 if and only if G is connected.
This observation led M. Fiedler to think of 2,-1 as a quantitative measure of
connectivity [10]. Following him, we will set a(G) = 2,_1 and call it the algebraic
connectivity of G.
M a n y approaches to ordering graphs by their spectra are contained in the
literature (see, e.g., [7, p. 70]). It is the main purpose of this article to make some
graph theoretic sense out of the (quasi) ordering of trees by algebraic connectivity,
in a sense to explain why some trees are "more connected" than others.
The m a x i m u m algebraic connectivity of a tree on n _> 3 vertices is 1, and
a(T) = 1 if and only if T = K I , , _ l, the "star". At the other extreme, the minimum

* Research supported by ONR contract 85K0335


230 R. Grone, R. Merris

algebraic connectivity occurs at the path, namely a(P,) = 2 [1 - cos(n/n)]. It follows


from an inequality of M. D o o b [14, Corollary 4.4] that, roughly speaking, a(T)
decreases as diameter increases, and a glance at Table 1 makes this easy to believe.
(Table 2 illustrates the need to "speak roughly" here.)
We begin our development with a complete discussion of the trees T(s, t) of
diameter 3. These trees have exactly 2 n o n p e n d a n t vertices and they are adjacent.
O u r notation reflects the fact that one of these vertices is connected to s pendants
(degree 1 vertices) and the other one to t pendants. In particular, n = s + t + 2.
We will always assume that 1 < s _ t.

Table 1

Type a(T)
I 1.000 Star

"k/' II .466 T(1,4)

II .398 T(2, 3)

I .382 Trivinylmethyl
i °
I .382

II .322

T7 * ; II .296

oIoT: I .268

II .260 3-Hexatrienylmethyl

r,o , I II .225

TH :. : :. - : : - I .198 P7
Ordering Trees by Algebraic Connectivity 231

Table 2

T a(T) Diameter

1. e e e I e ; : .1981 6

3. ~ .1864 6

4. _ e ; ~ : I : .1667 6

Proposition 1. Fix n = s + t + 2. Let T = T(s, t). Then the characteristic polynomial


of L(T)is
x ( x - 1)"-4Ix 3 - (n + 2)x 2 + (2n + st + 1)x - hi.

Proof. The singularity of L ( T ) explains the factor x. Isabel Faria [-9] has shown that
the multiplicity of 1 as an eigenvalue of L(G) is at least the "star degree" of G. In
our case, this is (s - 1) + (t - 1 ) = n - 4. We are left with a cubic factor ps(X)=
x 3 - b l x 2 + bzx - b3. Now, b3 is the sum of all the (n - 1)-by-(n - 1) principal
subdeterminants of L(T). By the Matrix-Tree Theorem, each of these is 1. Since there
are n of them, b3 = n. T h e trace of L ( T ) is Z(n - 1 ) = n - 4 + bl, s o b l = n + 2. T o
evaluate b2, consider the sum of all the 2-by-2 principal subdeterminants of L ( T ) .
On the one hand, it is the coefficient of x "-z in the characteristic polynomial
x ( x - 1)"-4ps(x), namely (n - 4)(n - 5)/2 + (n - 4)ba + b 2 = ( n - 4)(3n - 1)/2 +
b2. O n the other hand, a direct c o m p u t a t i o n from L ( T ) gives 3(s 2 + t 2 + s + t)/2 +
4st. Since n = 2 + s + t, we conclude that b2 = 2s + 2t + st + 5 = 2n + st + 1.
[]

Corollary 2. Let n = s + t + 2 be fixed. Then L ( T ( s , t)) has exactly one positive


eigenvalue less than 1, namely its algebraic connectivity as = a(r(s, t)). Moreover,
as is a strictly decreasing function of s, 1 _< s _< (n - 2)/2.
Proof. If n = 4, there is only one tree of diameter 3, namely the path P4 = T(1, 1).
Its characteristic p o l y n o m i a l is x ( x - 2)(x 2 - 4x + 2). If n = 5, the only tree of
diameter 3 is T(1,2). Its characteristic polynomial is x ( x - 1)(x a - 7x z + 13x - 5).
The roots of the cubic are (approximately) 0.5188, 2.3111, and 4.17. We now
assume n > 6. Write ps(x) = x 3 - (n + 2)x 2 + (2n + st + 1)x - n. Observe that
ps(0) = - n , ps(1) = st > 0, so ps(x)has one or three roots in (0, 1). Since the sum of
the roots is n + 2 > 3, L ( T ( s , t ) ) has exactly one eigenvalue, as e (0, 1).
It remains to show a s < as_ I, 2 _< s _< (n - 2)/2. Observe that
p~(x) - p~_l(x) = (t -- s + 1)x
232 R. Grone, R. Merris

so that ps(as_l) = (t - s + 1)as_ 1 > 0. By our previous observations, this implies


as_ 1 > as. []

In terms of Table 1, we have now explained the relative positions of T2 = T(1, 4)


and Ta = T(2, 3). In our remaining results, we show that the algebraic connectivity
goes down when trees are "reorganized" by breaking off branches and reattaching
them "farther from the center." (In Table 2, the fourth tree can be obtained from
any of the first three by moving a pendant farther from the center.) The "center" in
this discussion is described by means of so-called "characteristic valuations."
Suppose z = (Zl,Z2,...,z,) t is an eigenvector for L(T) corresponding to a(T).
We may use z to label the vertices of T, i.e., vertex i is labeled z i. In this context, z
has been called a characteristic valuation of T.

Example 1. The tree exhibited in Fig. 1 has the somewhat intimidating chemical
name Methylene-l,4,6-Heptatriene. It turns out that a ( T ) - 0.1864 is a simple
eigenvalue. Apart from nonzero multiples, the exhibited characteristic valuation is
unique.

Theorem A [11, Theorem (3, 14)]. Let T be a tree. Suppose z is a characteristic


valuation of L(T). Then two cases can occur.
Case I. I f ~" = {i e Vlzi = 0} ¢ ~ then the graph T = (V,E) induced by T on ~" is
connected and there is exactly one vertex j e V which is adjacent (in T) to a vertex
not belonging to V. Moreover, the values of z along any path starting at j are
increasing, decreasing, or identically zero.
Case II. I f z i ¢ 0 for all i e V, then T contains exactly one edge {j,k} such that
zj and z k have different signs, say zj > 0 and z k < 0. Moreover, the values of z along
any path that starts at j and does not contain k increase while the values of z
along any path that starts at k and does not contain j decrease.

We refer to a tree in which Case I (resp. Case II)occurs as a Type I (resp. Type II)
tree. The Type of T turns out to be independent of which eigenvector of L(T) is
chosen to play the role ofz (so long as it corresponds to a(T)). The vertex described
in Case I is also independent of z and is referred to as the characteristic vertex of T
[16]. For Type II trees, a(r) must be a simple eigenvalue of L(T) and the edge
described in Case II is referred to as the characteristic edge of T. The tree in Figure 1
is a Type II tree; the ends of its characteristic edge are labeled .2109 and -.0266.
Suppose that T is a Type I tree and that vertex 1 is the characteristic vertex of

.2592
?
.3608 ] -.0266 -.4434
0 0 ~- 0 0 0 0
.4434 .2109 -.2592 -.5451
Fig. 1
Ordering Trees by Algebraic Connectivity 233

T. By a branch of T at vertex i we m e a n one of the c o m p o n e n t s of the graph obtained


by deleting vertex i and all adjacent edges from T. It m a y h a p p e n that zj = 0 for
every vertex j of some branch B at vertex 1, no matter h o w the characteristic
valuation is chosen. In that case B is said to be a passive branch. A branch at vertex
1 which is not passive is said to be active. It was seen in [13] that the multiplicity
of a ( T ) a s an eigenvalue of L ( T ) i s one less than the n u m b e r of active branches.

T h e a r e m 3. Let T = (V, E) be a Type I tree whose characteristic vertex has degree d.


Let B1,132 . . . . . Bd be the branches of T at the characteristic vertex. We may assume
BI and B2 are active. Suppose B 3 . . . . . B d contain a total of p vertices altogether. Let
F be any graph with p vertices. Let G be the graph obtained from T by discarding
branches B 3 . . . . , B d and then adjoining the characteristic vertex of T to some or all
of the vertices of F by new edges. Then a(G) <_ a(T).
Proof. We m a y assume that T and G share the same vertex set V. Thus, we m a y
write G = (V, F). Now, if x is any real column n-tuple, the quadratic form
x'L(G)x= E ( x , - x A ~.
{i,j}cF
If e is the vector each of whose c o m p o n e n t s is 1, then

a(G) = min x~L(G)x


x e=o x'x (1)
x4-O
=min 2 (xl-xJ) 2
xte=O {i,j}eF
xtx=l
It was seen in [13] that there exists a characteristic valuation z of L ( T ) which
is positive on B 1, negative on B 2 and, m o r e importantly, zero elsewhere. W i t h o u t
loss of generality, we m a y assume that this z is a unit vector. Denote by i a . . . . , id
the vertices of B a . . . . . Bd that are adjacent (in T) to vertex 1. Then F is obtained
from E by deleting {1, iT}, 3 < r < d, and adjoining {1,jr}, 1 < r _< q, wherej~ . . . . . jq
are the vertices of F joined to vertex 1 (by these new edges) to form G. Then
d q
z t L ( G ) Z ~-- z t L ( r ) Z - 2 ( Z 1 - Zir) 2 "~- Z ( Z 1 - Zj~)2
r=3 r=l
= z'L(r)z

= a(T),
because zl = zi3 = "'" = zld = zjl = "'" = zjq = 0.
But, this means

a ( r ) >_ min 2 ( x i - )9) 2


xte=O {i,j}eF
xtx=l
= a(G). []

Roughly, T h e o r e m 3 says that we can "reorganize" a significant part of a Type


I tree without increasing the algebraic connectivity. This theme wilt be continued
in T h e o r e m s 4 and 5 below.
234 R. Grone, R. Merris

o o

T GI
0 "0 0 0

Fig. 2

Example 2. Let T be the tree in Figure 2. Then T is a Type I tree and the vertex of
degree 5 is the characteristic vertex. The three pendants attached to it are passive
branches and the remaining two are active. The approximate value ofa(T) is 0.3820.
The graphs G~ and G2 in Fig. 2 are two of the possible reorganizations. We have
computed a(G1) - 0.3065 and a(G2) = a(T).

Theorem 4. Let T = (V, E) be a Type I tree with vertex 1 the characteristic vertex.
Let BE, B3,... , B r be r - 1 active branches. Suppose v is the vertex of B v adjacent (in T)
to vertex 1, 2 <_ v < r. Let J2, ..., Jr be (not necessarily different) vertices of passive
branches of T. Form tree T' = (V, E') by deleting edges {1, v} from E and replacing
them with new edges {jv, v}, 2 <_ v <_ r. Then a(T') <_ a(T), with strict inequality if
the total number of active branches of T is at least r.
Proof. Let z be a unit eigenvector of L(T) corresponding to a(T). Then

ztL(T')z = z t L ( T ) z - ~ ( z l - z~)2+ ~, ( z j v - z~)2


v=2 v=2

= ztL(T)z=a(T),
because zl = 0 and z;v = 0, 2 < v < r. It follows as in (1) that a(T') < a(T).
Suppose, now, that B, is an r-th active branch with vertex k of B1 adjacent (in
both T and T') to vertex 1. As in the proof of Theorem 3, we may choose a unit
eigenvector z for a ( T ) t h a t is positive on B1 and negative on B2. If a(T) = a(T'),
then z must be a characteristic valuation of T'. This, however, violates Theorem A
since z 1 = 0 = zj2, but z k ~ 0 ~ z 2 yet {1,k} and {2,j2} ~ E'. []

Since the only requirement is that zyv = 0, 2 _< v < r, for some characteristic
valuation z, the proof actually establishes a little more than claimed.
In Theorem 4, we reorganized T by moving some of its active branches from
the characteristic vertex (the "center") farther away from this "characteristic center"
by attaching them to vertices of passive branches. The next result is a companion
theorem in which a passive branch is moved away from the characteristic center.
Ordering Trees by Algebraic Connectivity 235

Theorem 5. Let T = (V, E) be a Type I tree with vertex 1 the characteristic vertex.
Let B be a passive branch and suppose v is the vertex of B adjacent to vertex 1. Let j
be a vertex of some active branch of T. Form tree T' = (V, E') by deleting edge {l, v}
and replacing it with a new edge {j,v}. Then a(T') < a(T).
Proof. Let z be a unit eigenvector of L ( T ) corresponding to a(T) such that
z v = 0 = z 1 < zj. Break branch B off from vertex 1 and reattach it to vertex j
forming tree T' with the labels of z still attached. Then, as before, we arrive at
f L ( T ' ) z = f L ( T ) z + z 2. Suppose the branch B has a total of p vertices altogether
(including vertex v). Form vector y by adding zj to each of the p coordinates of z
corresponding to vertices of B. Summing ( Y s - Yt)2 over the edges {s, t} of T'
produces y'L(T')y. But, we have contrived y so that y ' L ( T ' ) y = z ' L ( T ) z = a(T).
Since e~y = p%, we let x = y - ( p % / n ) e . Then x~e = 0 and x~L(T')x = ytL(T')y.
It remains only to show that xtx > 1 because then
x'L(T')x a(T)
a(r') < - < a(r).
- - xtx xtx

Observe that

x'x=p\ n '] z,- zj


2 2pzj p(n -- p) z2
=Zz
s~V n s~V n

because z s = 0, s e B. Now, the right hand side of this equation collapses to


1 + p(n - p)zf/n because ztz = 1 and e'z = 0. Since 1 < p < n and zj > 0, xtx > 1.
[]

Example 3. Tree T5 in Table 1 is a Type I tree. The vertex of degree 4 is the


characteristic vertex. The two pendants off the characteristic vertex are the passive
branches and the other two branches are active. We get from Theorem 3 that
a(T4) = a(Ts). From Theorem 4, we see that a(T9) < a(Ts). Finally, Theorem 5
yields both a(T9) < a(Ts) and a(T6) < a(Ts) , depending on how vertexj is chosen.
A better name for passive branches might be "thorns." Dealing with them is
more difficult than the likelihood of their occurrence would justify. The next two
theorems say, roughly, that apart from these thorns, reorganizing a tree by moving
a branch farther from the characteristic center decreases the algebraic connectivity.

Theorem 6. Let T = (V, E) be a Type I tree with characteristic vertex 1. Let B be an


active branch at vertex 1 and suppose i is a vertex of B. Let j be a vertex o f B such
that vertex i is on the unique path from vertex 1 to vertex j. Let K be a branch at
vertex i containing neither vertex j nor vertex 1. Let k be the vertex of K adjacent
(in T) to vertex i. I f T' = (V, E') is the tree obtained from T by deleting {i, k} from E
and adjoining a new edge {j, k}, then a(T) > a(T').

Theorem 7. Let T = (V, E) be a Type II tree with characteristic edge {1, 2}. Let i,j e V
be two vertices such that vertex 1 is on the path from vertex 2 to vertex j, and either
236 R. Grone, R. Merris

i = 1 or vertex i is on the path from vertex 1 to vertex j. Suppose K is a branch at


vertex i containing neither vertex j nor vertex 2. Let k be the vertex of K adjacent
(in T) to vertex i. I f T' = (V, E') is the tree obtained from T by deleting {i, k} from E
and adjoining a new edge {j, k}, then a(T') < a(T).

Proofs of Theorems 6 and 7. The proofs are similar. They depend on the existence
(guaranteed by Theorem A) of a unit eigenvector z for a(T)such that 0 < zi < zj and
zi < Zk. We proceed as before, obtaining ztL(T')z = z t L ( T ) z - (zi - Zk)2 + (Z~ -- Zk)2.
NOW, it might happen that Zk > (Zi + Zj)/2, in which case a(T') < z t L ( T ' ) z <
ztL(T)z = a(T), and the proof is finished. We proceed under the assumption that
_< (zi + zj)/2.
If z k = (zi + zj)/2, then z t L ( T ' ) z = z~L(T)z. But (by Theorem A), z is not a
characteristic valuation of T' because, if it were, the path 1 ~ . . " ~ i ~ . . . ~ j ~ k
would be increasing in z, meaning zj < Zk. This contradicts Zk = (Zi + Z~)/2 < Zj.
Thus, a(T') must be afforded by a strictly smaller Rayleigh quotient.
We have now reached the stage where we may assume that ~ = ( z ~ - Z k ) -
(Zk -- Zi) > 0. Suppose that branch K has a total of p vertices altogether (including
vertex k). Form a new vector y by adding ~ to each of the p coordinates of z
corresponding to a vertex of K and let x = y - (p~/n)e. Then, as in the proof of
Theorem 5, x t L ( T ' ) x = a(T), but # x > 1 because zv > 0 for every vertex v of K.
[]

While we can now explain more of the relationships in Table 1, our results
are not sufficient to deduce a(T3)> a(T4),a(T6)> a(Tv),a(T7)> a(Tg), nor a(Tg) >
a(T9). Apart from Corollary 2, we are missing any results that would allow a branch
to be moved from one side of a characteristic edge to the other. Moreover, we do
not have a purely graph theoretic algorithm for locating the characteristic center.
Indeed, it was suggested in [16] that any characteristic vertex might have to lie on
the path joining the geometric center(s) to the "centroid(s)." That this is a flase hope
can be seen from Fig. 3. Vertex 3 is both the unique center and unique centroid yet
{1, 2} is the characteristic edge.
Ifn _> 5, the diameter 3 tree T(1, n - 3) is the near star; the nearpath is the unique
tree with three pendant vertices, one vertex of degree 3 that is adjacent to 2 pendants,
and n -4 vertices of degree 2. When n = 5, the near star and near path are the
same.

0 0 0
0•0 Fig. 3
2
0 ~0
1
Ordering Trees by Algebraic Connectivity 237

Corollary 8. Let T be a tree on n >__5 vertices. Denote by T1 the near star and by T2
the near path on n vertices. If T ~ KI,,_ ~, then a(T)< a(T1). If T ¢ Pn, then
a(r)>_a(r2)
The p r o o f is clear from Corollary 2 and Theorems 3-7.
Let a, = a(T(1, n - 3)), n _> 5, be the algebraic connectivity of the near star.
Then a, is the smallest root of x 3 - (n + 2)x 2 + (3n - 2)x - n.
It is not difficult to show that if n > 27 then 0.375 < a, < .4. It follows that if T
is a tree on n > 27 vertices which is not a star, then a(T) < .4 and also that a(T) > .4
for only finitely m a n y trees of diameter greater than 2.

References

1. Anderson, W.N. Jr., Morley, T.D.: Eigenvalues of the Laplacian of a graph. Linear and
Multilinear Algebra 18, 141-145 (1985)
2. Bapat, R.B.: A bound for the permanent of the Laplacian matrix. Linear Algebra Appl. 74,
219 223 (1986)
3. Biggs, N.: Algebraic Graph Theory. Cambridge University Press, 1974.
4. Brualdi, R.A., Goldwasser, J.L.: Permanent of the Laplacian matrix of trees and bipartite
graphs, Discrete Math. 48, 1 21 (1984)
5. Constantine, G.M.: Schur convex functions on the spectra of graphs. Discrete Math. 45,
181-188 (1983)
6. Cvetkovi6, D., Doob, M., Sachs, H.: Spectra of Graphs. New York: Academic Press, 1979
7. Cvetkovi6, D., Doob, M., Gutman, I., Torgasev, A.: Recent Results in the Theory of Graph
Spectra. Amsterdam: North-Holland, 1988
8. Dinic, E.A., Kelmans A.K., Zaitsev, M.A.: Nonisomorphic trees with the same T-polynomial.
Inf. Process. Lett. 6, 73-76 (1977)
9. Faria, I.: Permanental roots and the star degree of a graph. Linear Algebra Appl. 64, 255 265
(1985)
10. Fiedler, M.: Algebraic connectivity of graphs. Czech. Math. J. 23(98), 298-305 (1973)
11. Fiedler, M.: A property of eigenvectors of nonnegative symmetric matrices and its application
to graph theory. Czech. Math. J. 25(100), 619-633 (1975)
12. Forsman, W.: Graph theory and the statistics of polymer chains. J. Chem. Physics 65,
4111-4115 (1976)
13. Grone R., Merris R.: Algebraic connectivity oftrees. Czech. Math. J. 37(112), 660-670 (1987)
14. Grone, R., Merris, R., Sunder, V.S.: The Laplacian spectrum of a graph, SIAM J. MATRIX
ANAL. 11 (1990)218-238.
15. Hattori, Y.: Nonisomorphic graphs with the same T-polynomial. Inf. Process. Lett. 22,
133-134 (1986)
16. R. Merris, Characteristic vertices of trees. Linear Multilinear Algebra 22, 115-131 (1987)
17. Mohar, B.: The Laplacian spectrum of graphs. Preprint series dept. math., University E.K. of
Ljubljana 26, 353-384 (1988)
18. Powers, D.L.: Tree eigenvectors, unpublished.
19. Vrba, A.: The permanent of the Laplacian matrix of a bipartite graph. Czech. Math. J. 36(111),
7-17 (1986)
20. Vrba, A.: Principal subpermanents of the Laplacian matrix. Linear Multilinear Algebra 19,
335 346(1986)

Received: September 29, 1988


Revised: November 29, 1989

You might also like