CN9 Spectral properties


Objective

  • how does the topology of a network influence diffusion speed?
  • how fast the stationary distribution can be reached?
  • eigenvalues of transition matrix: captures the influence of a network on diffusion process
  • eigenvectors: define feedback centrality

Questions

  • what is the largest eigenvalue of a transition matrix?
    • 1.
  • what can we learn from the second-largest eigenvalue of a transition matrix?
    • the speed of convergence to stationary distribution
    • the influence of the network topology on the diffusion speed.
  • how is the diffusion speed related to the eigenvalues of the transition matrix?
    • the second largest eigenvalue.
  • what is the minimum time $t_\epsilon$, s.t. $\delta(\pi^{(t)}, \pi^{(0)})\leq \epsilon, \forall\epsilon>0$
    • $t\propto\frac{\log\epsilon}{\log \lambda_2}$
  • what is the spectrum of a transition matrix?
    • the sequence of ascending ordered eigenvalues
  • what is the spectral gap? what does it tell us about the underlying network?
    • 1-second largest eigenvalue.
    • the speed of diffusion
  • what is the spectral gap of a network that is not strongly connected?
    • much less than 1.
  • what is the algebraic connectivity of a network?
    • the second smallest eigenvalue of the Laplacian matrix.
  • what is the Fiedler vector of a network? What does it tell us about the topology?
    • the second smallest eigenvectors.
    • the cluster belonging.

Keywords

diffusion speed, Laplacian spectrum, algebraic connectivity, Fiedler vector

Eigen-decomposition

For aperiodic and irreducible stochastic matrix T

$$
|\lambda_i|<1%2C\forall i\geq 2
$$

  • is the transition matrix always has a eigenvalue of 1?
    • I think so. Even i haven’t proved this.

Rewrite the t-th step distribution

$$
\pi^{(t)}=\pi^{(0)}\cdot T^t
$$

$$
T = U^{-1}DU, T^t = U^{-1}D^tU
$$

$$
\pi^{(t)}=\pi^{(0)}\cdot U^{-1}D^t U=\sum_{i=1}^n\alpha_i\lambda_i^tu_i=\pi,\sum_{i=2}^n\alpha_i\lambda_i^tu_i
$$

we set $\pi^{(0)}U^{-1}=\alpha$

The last equation is the property of eigenvalue 1’s eigenvector.

Total variation distance

Compared to the final stable distribution
$$
\delta(\pi^{(t)},\pi)=\frac{1}{2}\sum_{j=1}^n|\pi_j-\pi^{(t)}_j|
$$

$$
=\frac{1}{2}\sum_{j=1}^n|(\sum_{i=2}^n\alpha_i\lambda_i^tu_i)_j|
$$

To control the distance by epsilon and use approximation,
$$
\delta(\pi, \pi^{(t)})<\epsilon\sim \frac{1}{2}\sum_{j=1}^n|\lambda_2^t \alpha_2(u_2)_j|<\epsilon
$$

$$
t\propto\frac{\log\epsilon}{\log \lambda_2}
$$

The second largest eigenvalue captures the influence of network topology on diffusion speed

Spectrum

Definition

the descending-ordered sequence of the absolute value of eigenvalues of the transition matrix.

Spectral/eigenvalue gap

$$
1-|\lambda_2|
$$

  • the value range is (0,1)
  • the larger the spectral gap, the faster the diffusion, $t$​ is smaller
  • the smaller, the slower, $t$ is larger​​

Allows to study the efficiency of network topologies.

Laplacian matrix

Definition

$$
L=D-A
$$

  • $D$ is a diagonal matrix, with the total degree of node $i$ as the $i$-th diagonal.
  • $A$: the adjacency matrix
  • the row of $L$ is 0.

Continuous time differential model

$$
\frac{dx^{(t)}}{dt}=C(A-D)x^{(t)}
$$

Which is from
$$
\frac{\mathrm{d}x_i^{(t)}}{\mathrm{d}t}=C\sum_{j=1}^nA_{ij}[x^{(t)}_j-x^{(t)}_i]
$$

Written by Laplacian matrix is
$$
\frac{\mathrm{d}x^{(t)}}{\mathrm{d}t}=-CLx^{(t)}
$$
Using the knowledge of ODE, we will get
$$
x^{(t)}=\sum_{i=1}^n \alpha_i e^{-C\lambda_i t}\vec{v}_i
$$

  • $C>0$
  • eigenvalues must be larger than or equal to 0.

Eigenvalues

sort eigenvalues in ascending order

smallest eigenvalue is 0. And its eigenvector is $(1,\cdots,1)$. Because $(1,\cdots,1)L=0L$​ , the sum of each column/row of $L$ is 0

Algebraic connectivity

the second eigenvalue of Laplacian matrix is called algebraic connectivity

  • $\lambda_2=0\Leftrightarrow$ the network is disconnected.

  • number of 0 eigenvalues is the number of components

  • Diffusion speed

if we hold $\lambda t$​ as fixed, the larger $\lambda$​ is, the smaller $t$​ is required, which means faster diffusion.
$$
x^{(t)}=\sum_{i=1}^n \alpha_i e^{-C\lambda_i t}\vec{v}_i
$$

$$
=\alpha_1\vec{v_1}+\sum_{i=2}^n \alpha_i e^{-C\lambda_i t}\vec{v}_i
$$

The second smallest eigenvalue will control the $t$​.

  • $\lambda_2$ captures how well connected the network is.
  • $\lambda_2$ captures the influence of network topology on diffusion speed.

Fiedler vector

$\vec{v_2}$ is called Fiedler vector

the eigenvector of the Laplacian matrix corresponding to the second smallest eigenvalue.

  • sign of this vector allows to detect communities. Bisect the network.
    • It can be applied recursively to detect more than two communities.

Summary

matrix Adjacency A Transition T Laplacian L
meaning simple representation of networks transition probabilities generalization of Laplacian operator
eigenvalue largest: 1, second-largest: diffusion speed, spectral gap smallest: 0, second smallest: algebraic connectivity, connectedness/diffusion speed
eigenvector largest: eigenvector centrality, not necessarily largest :1, stable distribution second-smallest: Fiedler vector
suitable network undirected/directed/weighted directed/undirected/weighted undirected, more complicated for directed

Author: Fululu
Reprint policy: All articles in this blog are used except for special statements CC BY 4.0 reprint policy. If reproduced, please indicate source Fululu !
  TOC