Bounds of the Relaxed Laplace Matrix

First we treat eigenvalue bounds of the relaxed Laplace matrixL_ρ. The negative adjacency matrix−Aand the Laplace matrixLare special cases ofL_ρand therefore their eigenvalue bounds, too. A negative upper bound of −A is a lower bound of A and a negative lower bound of−A is an upper bound of A. We denote the maximum degree of the underlying graph as ∆, the minimum degree asδ and the deleted row sumsR_i as in definition 3.4 :

R_i :=

n

X

j=1 j6=i

|ω_ij| , 1≤i≤n .

The matrix L_ρ has real spectrum. So Gershgorin’s theorem (3.5) gives us the following theorem:

Theorem 4.21

Given is a relaxed Laplace matrix L_ρ of a graphG= (V, E, ω). Then

1≤i≤nmin(1−ρ)d_i−ω_ii−R_i ≤ λ^Lρ ≤ max

1≤i≤n(1−ρ)d_i−ω_ii+R_i

for all eigenvalues λ^Lρ of L_ρ. Suppose now, that all weights are nonnegative and the underlying graph has no self-loops. Ifρ∈[0,2], the bounds can be written as

−ρ∆ ≤ λ^Lρ ≤ (2−ρ)∆ .

Under the same assumptions holds for the eigenvalues λ^L of the Laplace matrix 0 ≤ λ^L ≤ 2∆

and for the eigenvalues λ^A of the adjacency matrix

−∆ ≤ λ^A ≤ ∆ .

CHAPTER 4. GRAPH RELATED MATRICES 40 Ifρ <0, then

−ρδ ≤ λ^Lρ ≤ (2−ρ)∆ , and if ρ >2, then

−ρ∆ ≤ λ^Lρ ≤ (2−ρ)δ .

These bounds can be sharpened using the extensions of Gershgorin’s theorem:

Theorem 4.22

Given is a relaxed Laplace matrixL_ρ with all weights nonnegative,ρ∈[0,1] and without self-loops. Let ∆₂ denote the second-largest degree of L_ρ. If there are two nodes in the underlying graph with degree ∆, then ∆2 := ∆ . Then holds for all eigenvaluesλ^Lρ ofLρ:

−ρ∆ ≤ 1 2

(1−ρ)(∆ + ∆₂)−p

(1−ρ)²(∆−∆₂)²+ 4∆∆₂

≤ λ^Lρ ≤ 1

2

(1−ρ)(∆ + ∆2) +p

(1−ρ)²(∆−∆2)²+ 4∆∆2

≤ (2−ρ)∆ . Under the same assumptions holds for the eigenvalues λ^L of the Laplace matrix

0 ≤ λ^L ≤ ∆ + ∆2

and for the eigenvalues λ^A of the adjacency matrix

−p

∆∆₂ ≤ λ^A ≤ p

∆∆₂ .

Proof:

Theorem 3.1 yields that all eigenvalues of L_ρ are contained in

n

[

i,j=1 i6=j

K_ij with

K_ij :={x∈R:|x−(1−ρ)d_i| · |x−(1−ρ)d_j| ≤d_id_j} .

CHAPTER 4. GRAPH RELATED MATRICES 41 We want to find the minium and the maximum of the above expression. We define for d_i ≥d_j and that thus x₁ and x₂ are the only zeros. Because of

x→±∞lim f_ij(x) = +∞ ,

max1≤i,j≤nx1 and min1≤i,j≤nx2 are an upper and a lower bound for allKij . To determine the maximum and the minimum of x₁ and x₂ consider again (∗) :

x_1/2 = 1

CHAPTER 4. GRAPH RELATED MATRICES 42 As (k−1)4ab ≥ 0 we get the maximum and minimum of x₁ and x₂ if a and b become maximal, i.e. a = ∆ andb = ∆₂ . For ρ= 0 and ρ= 1 we get the results for the Laplace matrix and the negative adjacency matrix, respectively. As shown in lemma 3.7 these bounds are always as least as good as Gershgorin’s bounds. 2

The bounds of the last theorem become as better as larger the difference between ∆ and

∆2 becomes.

The question is now, how good the upper and lower bounds we statet above approximate the spectrum. We will see, that certain graphs have these bounds as smallest and largest eigenvalue. Thus our bounds are in this sense tight. But first we prove an upper bound for the smallest eigenvalue and a lower bound for the largest one.

Theorem 4.23

Given is a relaxed Laplace matrix L_ρ with ρ∈[0,1]. Then λ₁ ≤ max

1≤i≤n−ρd_i λ_n≥ max

1≤i≤n(1−ρ)d_i−ω_ii .

If the underlying graph has nonnegative degrees and no self-loops, then λ₁ ≤ −ρδ

λ_n≥(1−ρ)∆ .

Proof:

We get the lower bounds by setting A = L, B = −ρD and k = 1 in Weyl’s theorem (3.12). With the Rayleigh-Ritz theorem (2.9) we have

λ_n≥ x^TL_ρx

x^Tx for all x6= 0.

The upper bounds follow byx=e_i, i= 1, .., n . 2

Theorem 4.24

Given is a bipartite regular graphGwith nonnegative edge weights. Thenλis eigenvalue of L_ρ, iff 2(1−ρ)∆−λ is eigenvalue ofL_ρ . In particular

λ^L₁^ρ =−ρ∆ and λ^L_n^ρ = (2−ρ)∆ .

CHAPTER 4. GRAPH RELATED MATRICES 43 Proof:

Bipartite graphs have no self-loops. Thus there exists for a bipartite graphG an isomor-phic graph G⁰ with relaxed Laplace matrix L_ρ of the form

 an eigenvector ofLρ with eigenvalue λ, i.e.

L_ρ

Since 0 is the smallest eigenvalue of the Laplace matrix L, the largest now must be 2∆.

The exact values of the extremal eigenvalues ofL_ρ follow with lemma 4.10. 2

The last theorem can be sharpened for the adjacency matrix using the Perron-Frobenius theorem (4.2). We state a theorem taken from [GvL, p. 178]. On the same page a proof is sketched for unweighted adjacency matrices only. But the proof carries over to the weighted case, if the weights are nonnegative.

CHAPTER 4. GRAPH RELATED MATRICES 44 Theorem 4.25

LetAbe the adjacency matrix of a graphGandr_ρits spectral radius. Then the following is equivalent:

a) G is bipartite.

b) The spectrum of Ais symmetric around the origin, i.e. if λis an eigenvalue, then also −λ is an eigenvalue with the same multiplicity.

c) −r_ρ is an eigenvalue.

The next theorem is an upper bound for the largest eigenvalue of a Laplace matrix with nonnegative weights and without self-loops. The bound is better than all other bounds stated here, but we found no generalization for L_ρ. Theorem and proof is taken from [AM].

Theorem 4.26

Given is a graph G = (V, E, ω) with nonnegative weights, without self-loops and the Laplace matrixL(G). Then:

λ^L_n ≤max{d_i+d_j|(i, j)∈E} .

Proof (sketched):

Given is a graph G = (V, E, ω), |V| = n, |E| = m. We first introduce the vertex-edge incidence matrix E(G) = (e_ik)∈R^n×m of G:

eik =







−√

ω_ij if there is an edge k = (i, j)

√ω_ij if there is an edge k = (j, i)

0 otherwise

, i < j .

So in every rowk of E there are the entries −√

ω_ij and √

ω_ij with k = (i, j) and all other are zero. Now it can be shown that EE^T = L. We define another matrix N := E^TE, N ∈ R^m×m. If Lx = λx with λ 6= 0, then N E^Tx = E^TLx = λE^Tx, so that λ is also an eigenvalue of N. The sum of the absolute values of the row k of N equals d_i +d_j, if k= (i, j). The assertion is the upper gershgorin bound of N. 2

We found forL_ρ,ρ6= 0, no matrix with similar properties as the incidence matrix has for L.

CHAPTER 4. GRAPH RELATED MATRICES 45

An important class of graphs later on are those with nonnegative weights and no self-loops. We can summarize their eigenvalue bounds ofL_ρ, ρ∈[0,1], with

−ρ∆ ≤ ¹₂

(1−ρ)(∆ + ∆₂)−p

(1−ρ)²(∆−∆₂)²+ 4∆∆₂

≤ λ^L₁^ρ ≤

−ρδ ≤ (1−ρ)∆

≤ λ^Ln^ρ ≤

1 2

(1−ρ)(∆ + ∆₂) +p

(1−ρ)²(∆−∆₂)²+ 4∆∆₂

≤ (2−ρ)∆ .

Im Dokument Spectral graph drawing (Seite 40-46)