Abstract
The adjacency and Laplacian matrices of complex networks with two species of nodes are studied and the spectral density is evaluated by using the replica method in statistical physics. The network nodes are classified into two species (A and B) and connections are made only between the nodes of different species. A static model of such bipartite networks with power law degree distributions is introduced by applying Goh, Kahng and Kim's method to construct scale-free networks. As a result, the spectral density is shown to obey a power law in the limit of large mean degree.
Export citation and abstract BibTeX RIS
1. Introduction
The theory of complex networks, which has been dramatically developed since the end of the last century, is based on the observation that there are universal features in real biological and social networks [1]. One such feature is the scale-free property, meaning that the degree (the number of nodes directly connected to each node) distribution function P(Δ) obeys a power law P(Δ)∝Δ−λ for large Δ. Barabási and Albert explained the origin of this scale-free property by focusing on the network growing process [2]. Goh, Kahng and Kim formulated a static network model which exhibits the scale-free property [3].
The connection pattern of a network is mathematically described by the adjacency matrix. When the network has the scale-free property, the spectral (eigenvalue) density ρ(μ) of the adjacency matrix is also expected to obey a power law ρ(μ)∝μ−γ for large μ. Dorogovtsev et al presented an analytic evidence of this power law behaviour [4, 5]. Moreover, a relation γ = 2λ − 1 was found between the exponents of the power laws. Rodgers et al analysed Goh, Kahng and Kim's static model and confirmed the power law behaviour of ρ(μ) [6].
In this paper, we shall study scale-free networks with two species (A and B) of nodes. Connections are made only between the nodes of different species. We introduce a static model of such bipartite scale-free networks by applying Goh, Kahng and Kim's method, and observe that each species has its own degree distribution function obeying a power law. Suppose that the exponent of the degree distribution function is λA for the species A and λB for the species B. Using the replica method in statistical physics, we are able to analytically evaluate the spectral density ρ(μ) in the limit of large mean degree [6–8]. As a result, we find that ρ(μ) also obeys a power law and the exponent γ is associated with the exponents λA and λB. In addition, the spectral density of the Laplacian matrix is similarly analysed and the power law behaviour is confirmed.
Bipartite networks find applications in the analysis of human sexual contacts [9], and of the connections between collaborators and collaboration acts [10], such as actors and movies, scientists and papers. The adjacency matrices of bipartite networks are also interesting from the viewpoint of random matrix theory, since the Gaussian matrix model with the same structure is called the chiral Gaussian ensemble and applied to physics, such as the QCD gauge theory [11].
The outline of this paper is as follows. In section 2, a static model of bipartite scale-free networks with two species of nodes is introduced, and the adjacency and Laplacian matrices are defined. In section 3, in order to evaluate the spectral density, we apply the replica method to the network model. In the limit of large mean degree, the power law behaviour of the spectral density is analytically derived. In section 4, the effective medium approximation is briefly discussed as an attempt to treat the case with a finite mean degree.
2. Complex networks with two species of nodes
Let us suppose that there are N nodes of type A and M nodes of type B (N ⩾ M). We are interested in the asymptotic behaviour of bipartite networks with two species of nodes A and B in the limit
We introduce a static model of such networks with power law degree distributions by applying Goh, Kahng and Kim's method. Each node of type A is assigned a probability Pj normalized as
while each node of type B has a probability Qk with
The nodes of type A and B are connected according to the following procedure. In each step, we choose a node j of type A and a node k of type B with probabilities Pj and Qk, respectively. Then the nodes j and k are connected, unless they are already connected. After repeating such a step pN times, a node j of type A and a node k of type B are connected with the probability
Let us consider an N × M matrix C (N ⩾ M), where Cjk = 1 if the node j of type A is directly connected to the node k of type B, and Cjk = 0 otherwise. This random matrix C describes the connection pattern of the network with two species of nodes. Each matrix element Cjk is independently distributed with the probability density function (p.d.f.)
We assume that Pj and Qk are given by
and
There are thus two parameters α and β controlling the p.d.f. of the matrix C.
We define the degree dj of the type-A node j as the number of directly connected type-B nodes:
Then the type-A node degree distribution function is given by
where the brackets denote the average over the p.d.f. (2.5) and δ(x) is Dirac's delta function. We can similarly introduce the degree ek of the type-B node k:
and the type-B node degree distribution function
In appendix A, a useful asymptotic relation
is derived in the limit (2.1). Here tjk, which depends neither on N nor on M, is in the neighbourhood of the origin so that .
As special cases, we can readily derive asymptotic relations for
as
Then we can readily see that
so that the mean degree m(A) of the type-A node is
while the mean degree m(B) of the type-B node is
It can be seen from (2.6), (2.9) and (2.14) that the type-A node degree distribution function can be written as
Then in the limit Δ → ∞, we find
and similarly obtain
Thus we have seen that the network has the scale-free property, as the node degree distribution functions obey power laws. The exponents of the power laws defined as
are found to be λA = (1/α) + 1 and λB = (1/β) + 1.
In this paper, we study the adjacency and Laplacian matrices of this scale-free network. The adjacency matrix of this network is defined as
where CT is the transpose of C and On is an n × n matrix with zero elements. The Laplacian matrix is an (N + M) × (N + M) symmetric matrix with
3. Spectral density
Let us define that J is the adjacency matrix or the Laplacian matrix . The spectral density of J is defined as
where μj, j = 1, 2, ..., N + M, are the eigenvalues of J. In order to calculate ρ(μ), we introduce the partition function
Using the partition function Z, we can write the spectral density as
where is an infinitesimal positive number and I is an (N + M) × (N + M) identity matrix. Then we can utilize the relation
to obtain
Therefore, it is necessary to evaluate the average 〈Zn〉.
The replica method explained in appendix B is known to be a powerful tool for that purpose. It follows in the limit (2.1) that
Here,
and
with
The functional integrations are taken over the auxiliary functions and satisfying
In the limit (2.1), the functional integrations over and are dominated by the stationary point satisfying
where θj and ωk are the Lagrange multipliers. It follows from this equation that
where Θj and Ωk are normalization constants.
In the limit of large mean degree p → ∞, the variational equations (3.13) are satisfied by the Gaussian ansatz
as shown in appendix C. Here, Im σj ⩽ 0 and Im τk ⩽ 0. This property simplifies the problem and enables us to evaluate the asymptotic spectral density.
Let us consider the limit p → ∞ with a scaling variable . In appendix C, we find the asymptotic spectral density of the adjacency matrix as
in the tail region E → ∞. The exponent γ of the spectral density defined as
is (2/α) + 1 if α ⩾ β, and is (2/β) + 1 if β ⩾ α. Thus γ is associated with λA = (1/α) + 1 and λB = (1/β) + 1 as γ = 2min (λA, λB) − 1.
It is also explained in appendix C that the asymptotic spectral density of the Laplacian matrix is given by
in the region μ = O(p) with p → ∞. Here, H(x) is defined as
The exponent γL of the spectral density (μ → ∞) is (1/α) + 1 if α ⩾ β, and is (1/β) + 1 if β ⩾ α. Thus γL is associated with λA and λB as γL = min (λA, λB).
4. Effective medium approximation
In the previous section, we have dealt with the spectral density in the limit p → ∞. The calculation of the spectral density with a finite mean degree p is a much more involved problem, for which sophisticated numerical schemes have been proposed [12–15]. In this section, we briefly discuss a simple approximation method (effective medium approximation) for that problem [8, 16–20]. In this approximation, we put the Gaussian ansatz (3.14) into formulas (3.7), (3.8) and (3.9), and solve the stationary point equations
and
In the case of the adjacency matrix , the above procedure results in the effective medium approximation (EMA) equations
As for scale-free networks with a single species of nodes, Nagao and Rodgers calculated the 1/p expansion of the spectral density by using the corresponding EMA equation [20]. A similar analytical treatment could also be possible in the present case. Here, however only results of numerical iterations of (4.3) are shown in figure 1 as the EMA spectral densities. They are compared with the spectral densities of positive eigenvalues calculated by numerical diagonalizations of numerically generated adjacency matrices (averaged over 100 samples). The EMA gives a better fit for a larger p, as expected from the fact that the variational equations (3.13) are satisfied by the Gaussian ansatz (3.14) in the limit p → ∞. When p = 1, the agreement significantly breaks down around the origin, although it is still fairly good in the tail region with large μ.
In the limit α, β → 0, we obtain the adjacency matrix of a classical random graph with two species, where the connections are made only between the nodes of different species. In that case, σj and τk can be written as σ and τ, respectively, because they depend neither on j nor on k. The EMA equations become a cubic equation for σ
and
These equations are equivalent to Nagao and Tanaka's SEMA (symmetric EMA) equations concerning the spectral density of sparse correlation matrices [19], and can be analysed in the same way.
We can similarly derive the EMA equations for the Laplacian matrix as
In the limit α, β → 0, σj and τk can again be reduced to σ and τ, respectively. Then we find a cubic equation for σ
and
Acknowledgments
The author thanks Professor G J Rodgers and Professor Toshiyuki Tanaka for valuable discussions. This work was partially supported by the Japan Society for the Promotion of Science (KAKENHI 20540372).
Appendix A
In this appendix, we derive an asymptotic relation
where tjk is a parameter which is independent of N and M. We moreover assume that tjk is in the neighbourhood of the origin so that |Sjk| < 1 holds for . A similar argument for Goh, Kahng and Kim's model is found in [21].
The Taylor expansion of the logarithmic function gives
We show
in two steps.
Step 1
Let us first prove that
We define
and
Then we see that
A monotonously decreasing continuous function F(x) satisfies
so that
with
Then one can again use (A.8) to obtain
where
Using the notations
we see that
Then, using the inequality
we find
where
In the case min (α, β) > 1/2, we similarly employ
to obtain
where
Using inequality (A.15), we can similarly derive the estimates
If min (α, β) > 1/2, then we utilize (A.18) to find
Moreover, one can readily see from the inequality G1(x) ⩽ x (x ⩾ 0) that
It follows from (A.16), (A.19), (A.21), (A.22) and (A.23) that
which yields (A.4).
Step 2
We next prove
Using
we see that
We can again employ (A.8) to obtain
where
Making use of the identity
we find
where
In the case α + β > 1, by means of
we obtain
Here the symbol N(α, β) is defined in (A.20). Inequality (A.33) similarly gives the estimates
Moreover, it is evident from the inequality G0(x) ⩽ 1 (x ⩾ 0) that
Now we can easily see from (A.31), (A.34), (A.35) and (A.36) that
for any ℓ ⩾ 2. This relation results in the asymptotic estimate (A.25).
Appendix B
Let us first discuss the spectral density of the adjacency matrix . The eigenvalues μj, j = 1, 2, ..., N + M of consist of M pairs ±νj, j = 1, 2, ..., M and N − M zeros. Note that ν2j > 0 are identified with the eigenvalues of the M × M correlation matrix V with
Using the notations
and
we can rewrite the partition function Z defined in (3.2) as
Then we introduce the replica variables
and
to obtain
Now we can see from (2.12) that
It should be noted that this asymptotic relation holds if is in the neighbourhood of the origin. This condition is justified in the limit of large mean degree p → ∞, since and are scaled as O(p−1/2) or O(p−1) (see equations (C.4) and (C.23)).
Using the notation
we obtain
so that we find
Here, S1 and S2 are defined in (3.8) and (3.9). The auxiliary functions and satisfy (3.11). If J is the Laplacian matrix , we can similarly derive the same formula (B.11) for 〈Zn〉, except the change of S2 according to (3.10).
It can be readily seen that
where
In the limit N → ∞, the dominant contribution comes from the stationary point satisfying
which means
Therefore, we find an asymptotic estimate
One can similarly derive another estimate
in the limit M → ∞. Then we arrive at
where S0 is defined in (3.7).
Appendix C
Putting the Gaussian ansatz (3.14) into (3.13), we see in the limit n → 0 that
where
and
Let us first consider the adjacency matrix . We are in a position to take the limit p → ∞ with the scalings
Then, we obtain
The variational equations (3.13) are satisfied by the Gaussian ansatz (3.14), if σj and τk are determined by these equations.
In order to analytically treat (C.5), we define the scaling variables
Then it is straightforward to find
and
in the limit (2.1). Using the notations
we obtain
In order to evaluate the behaviour of S and T in the tail region E → ∞, we write
with real s(R), s(I), t(R) and t(I). Then it can be seen that
Let us employ an asymptotic formula [6]
and obtain an estimate
so that
One can similarly derive
so that
It follows from (C.15) and (C.17) that
Now we can evaluate the asymptotic behaviour of the spectral density ρ(μ) in the tail region E → ∞. Equations (3.5) and (3.6) can be utilized as
in the limit (2.1). Here,
and we can similarly obtain
Then it can be seen from (C.15), (C.17) and (C.18) that
This gives the asymptotic spectral density of the adjacency matrix in the tail region E → ∞.
We next compute the spectral density of the Laplacian matrix . Using the scalings
and taking the limit p → ∞, we find
so that
where is an infinitesimal positive number. Then it follows in the limit (2.1) that
and
where H(x) is defined in (3.18). Therefore, we arrive at
This gives the asymptotic spectral density of the Laplacian matrix in the region μ = O(p).