I am reading "Using gossips to spread information: Theory and evidence from two Randomised Control Trials" by Banerjee, Chandrashekhar, Duflo and Jackson, where they discuss the efficacy of gossips in spreading information regarding policies/ business offers.

In this paper, in page 21, they have introduced a generalised version of the network parameter "Diffusion Centrality", which the same authors had defined earlier in a 2013 paper.

If $w$ is a directed and weighted adjacency matrix for a social network, then they define a 'hearing matrix' for $w$ and time duration $T \in N$ as:

$$H(w, T) = \sum_{t=1}^T (w)^t$$

The $ij$-th entry of $H$, $H(w, T)_{ij}$ , is the expected number of times, within $T$ periods, that $j$ hears about a piece of information originating from $i$.

Then Diffusion Centrality is defined as:

$$DC(w, T) = H(w, T) \cdot 1$$

The paper says, "$DC(w, T)_i$ is the expected total number of times that some piece of information that originates from $i$ is heard by any of the members of the society during a $T$-period time interval".

I am not very familiar with network economics. I would like to understand why do $H$ and $DC$ match the verbal descriptions of them, that is, why is the $ij$-th entry of $H$ the expected number of times $j$ hears something from $i$ and similarly for $DC$. A proof directly written in the answer, or a reference to any good material to study this will help.


Let me use capital letter $W$ instead of small letter $w$ (as $W$ is actually a matrix).

So, let $W$ be the matrix that contains in row $i$ and column $j$ the probability that $j$ hears from $i$. Then average number of pieces of information that $j$ gets from $i$ in one period is given by $W(i,j)$.

The average number of information that $j$ gets from $i$ that takes 2 periods to travel is the information that goes first from $i$ to some intermediate person $k$ and then from $k$ to $j$, via the path $i \to k \to j$. The amount of information transmitted through this path is $W(i,k) \cdot W(k,j)$. To see this, notice that $W(i,k)$ is the amount of info send to $k$ and of this a fraction $W(k,j)$ is sent forward to $j$. Now, this has to be summed over all intermediate nodes $k$ (i.e. all paths of length 2). so we get: $$ \sum_{k} W(i,k) \cdot W(k,j) = (W \times W)_{i,j} = (W^2)_{i,j} $$ Here $\times$ is matrix multiplication.

Let's now look at the information that takes 3 periods. This is the information that goes from $i$ to some intermediate $k$ then from $k$ to some intermediate $\ell$ and finally to $j$, so the path $i \to k \to \ell \to j$. The amount of information on this path is $W(i,k) \cdot W(k,\ell) \cdot W(\ell, j)$. If we sum over all these paths of length 3, we get: $$ \sum_k \sum_\ell W(i,k) \cdot W(k,\ell) \cdot W(\ell, j) = (W \times W \times W)_{i,j} = (W^3)_{i,j}. $$ If we continue, then we can generalize this to any path of length $T$. The information sent from $i$ to $j$ through all such paths is given by: $$ (W \times W \times \ldots W)_{i,j} = (W^T)_{i,j}. $$

Now if info goes from $i$ to $j$ either takes 1 period, 2 periods, 3 periods, ... $T$ periods to travel. As such, to compute the total expected pieces of info that $j$ receives from $i$ in $T$ periods is given by the sum over all these periods: $$ H(W, T) = \sum_{t = 1}^T (W^t)_{i,j}. $$ Now the total amount of info sent from $i$ and heard by someone can be obtained by taking the sum over all recipients $j$. $$ DC(W,T) = \sum_{j} (H(W,T))_{i,j} = (H(W,T)\times 1)_{i,j} $$ This amounts to adding up the rows of the matrix $H(W,T)$.

An example

Let's take the example of three individuals $1,2$ and $3$. Assume that $W$ is given by: $$ W = \begin{bmatrix} 0 & 0.3 & 0.2\\ 0.2 & 0 & 0.6\\ 0 & 0.4 & 0\end{bmatrix} $$ This implies that (for example) the probability that 2 hears from 1 in one period equals 0.3.

Then $$ W^2 = \begin{bmatrix} 0.06 & 0.08 & 0.18\\ 0 & 0.3 & 0.04\\ 0.08 & 0 & 0.24\end{bmatrix} $$ Here, for example, the amount received by 2 from 1 in two periods is $0.08 = 0.2 \times 0.4$

Next, multiplying $W^2$ once more by $W$ gives: $$ W^3 = \begin{bmatrix} 0.016 & 0.09 & 0.06\\ 0.06 & 0.016 & 0.18\\ 0 & 0.12 & 0.016\end{bmatrix} $$ Then: $$ H(W, 3) = W + W^2 + W^3 = \begin{bmatrix} 0.076 & 0.47 & 0.44\\ 0.26 & 0.316 & 0.82\\ 0.08 & 0.52 & 0.256\end{bmatrix} $$ So in three periods $2$ receives on average $0.47$ pieces of info from $1$

The total information sent from $1$ to someone is computed by taking the sum of the elements in row $1$, which gives: $$ 0.076 + 0.47 + 0.44 = 0.986 $$ In general: $$ DC(W, 3) = \begin{bmatrix} 0.986\\1.396\\0.856\end{bmatrix} $$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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