I'm trying to understand the Newman Modularity (doi:10.1073/pnas.0601602103) by investigating its calculation on the Wikipedia example . My question is why are probabilities of self-loops included in the Q formula while the adjacency matrix A has no self-loops? i.e., given the standard definition of modularity Q...

$Q=\frac{1}{2m}\Sigma_{ij} (A-P_{ij})\delta(s_{i},s_{j})$

where A is the adjacency matrix and $P_{ij}$ is probability of a link between i and j, and $\delta()=1$ if entries are in the same community. While $A_{ii}=0$, this is not true for P, where $P_{ii}=\frac{k_{i}k_{i}}{2m} \neq0$, $k_i = \Sigma_{j}a_{ij}$ is the degree of i.

It would seem to me that probabilities of self loops should NOT be included. A-P leads to negative entries for the community of i. Here is R code to show how self-loop probabilities need to be included for agreement with both wikipedia example and the same calculation in the 'igraph' package function 'modularity'

# modularity function, input matrix A and membership vector mem
Q <- function(A,mem){ diag(A) <- rep(0,ncol(A)) 
m <- sum(A)/2 # number of links
k <- colSums(A) #degree
kk <- outer(k,k,"*") 
memmem <- outer(mem,mem,"==")*1 # diag is 1!
Pij <- kk/(2*m) # probability of link

# Make a graph from Wikipedia entry on Modularity (Q=0.4895)
Aw <- matrix(c(0,1,1,0,0,0,0,0,0,1,1,0,1,0,0,0,0,0,0,0,1,1,0,0,0,0,
membership <- c(1,1,1,2,2,2,3,3,3,1)
Q(Aw, membership) # = 0.4895833

library(igraph) # check in igraph package
modularity(graph.adjacency(Aw),membership) # = 0.4895833

I think that the entries $P_{ii}$ are needed to fulfill the constraint $\Sigma_{j} p_{ij} = \Sigma_{j} a_{ij} = k_i$ , i.e, the expectation of i's degree is its degree, but if $p_{ii}\neq 0$, than the probabilities are not distributed across the other nodes $i\neq j$.


2 Answers 2


I will answer my own question, after stumbling upon Squartini and Garlaschelli 2011 article "Analytical maximum-likelihood method to detect patterns in real networks", in ArXiv. Yes, the expected probability of a link between two nodes in a binary graph, as calculated in the classic Modularity score, includes self-loops, and yes this is not ideal. However, it is known to be an approximation, $\langle p_{ij} \rangle = \frac{k_i \cdot k_j}{2m}$. Squartini and Garleschelli also state that the approximation becomes particularly probablematic in scale-free networks that have high density.


Self-loops always have positive contribution to modularity. However if the difference between the self-loops and the other edges is high, the detection algorithm would not be able to detect big communities and instead detects lots of singular communities.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.