Networks - Directed Reading

With Prof. Simon Board

networks
games
Published

April 14, 2025

Starting Spring 2025, I’m doing some directed reading in network dynamics and games with the help of economic theory Professor Simon Board. I’ve found this line of work of his particularly interesting, so I’m excited to learn more and get more acquainted with theoretical economics. Here I’ll document some of my learning and notes in preparation for (hopefully) future project(s).

import networkx as nx
from matplotlib import pyplot as plt
import random

Materials

Learning in Games:

  • “The Theory of Learning in Games”, Friedberg and Levine
  • Population Games and Evolutionary: Dynamics Sandholm

Networks:

  • “Social and Economic Networks”, Matthew O. Jackson
  • Networks: An Economics Approach”, Sanjeev Goyal
  • “Random Graphs”, Bela Bollobas

Courses:

  • Econ 142: Undergraduate Networks

How to come up with a question topic?

Running list of topic ideas:

  • How can the structure of a network of neighbors affect crime/safety in the neighborhood?
  • A game for career-building/students, how does who you actively choose to become friends with impact your outcomes (job, salary, etc.)

“Social and Economic Networks”, Matthew O. Jackson

Chapter 1. Intro and Motivation for Networks

Example 1: Florentine Marriages

Suppose there is an edge between family \(i\) and \(j\) if there is a marriage between their respective families. Let \(P(i,j)\) be the number of shortest paths from family \(i\) to family \(j\). Let \(P_k(i,j)\) be the number of these that family \(k \neq i,j\) lies on. Then we can develop a “betweenness” measure of a family by:

\[ \text{Betweenness}(k) := \sum_{i,j : i \neq j, k \notin \{i,j\}} \frac{P_k(i,j)/P(i,j)}{{n-1}\choose{2}} \]

where \({n-1}\choose{2}\) is the total number of pairs of families in the graph.

We see that \(\text{Betweenness(Medici)} = 0.522\) in the figure, so they lie on over half of the shortest paths between any two other families. This could be an explanation for why they rose to power, as marriages were key to communication, trade deals, and political decisions.

Example 2: High school Friendships

This following network depicts students organized by race/ethnicity. There are about 52% whites, yet 86% of whites’ friendships are with other whites, and similar for the 38% of blacks, with 85% of friendships being with other blacks.

This is a phenomenon known as homophily, and leads to students being somewhat segregated by race, impacting the diffusion of information, learning, and the speed at which information/disease/etc. may be spread through the network.

image.png

Example 3. Erdos-Renyi Random Networks

Key model:

  1. Fix a set of \(n\) nodes.
  2. An edge between any two nodes, \(i,j\) is given by a probability \(p\), and independent across links.

Ex. Given \(n=3\), a complete \(K_3\) graph forms with probability \(p^3\) (3 edges must form), one with two links \(3(p^2)(1-p)\), etc. (binomial distribution).

The degree distribution of an ER network describes the prob. any given node will have a degree of \(d\). The probability that any node will have \(d\) links is:

\[ {{n-1}\choose{d}}p^d(1-p)^{n-1-d} \]

However, there is still some correlation between node degrees, but as \(n\) becomes large, the correlation of degree between any two nodes vanishes to \(0\), and the degree distribution will approach that above.

We can approximate (for large \(n\), small \(p\)) the binomial expression with Poisson distribution, i.e. \[ \frac{e^{-(n-1)p}((n-1)p)^d}{d!} \] and given this approximation, random graphs with independently formed links with identical probabilities are called Poisson random networks.

Ex. \(n=50, p=0.02\) which gives approx expected degree of a node to be \(1\).

Increasing \(p = 0.078 = \log(50)/50\), we expect that only \(2%\) of the nodes will have degree of 0 (i.e. isolated). Indeed, we see 1/50 of the nodes in this one are isolated; so the rest are in some connected component (which happens to be 1 component).

Using the Poisson approximation, we can see when \(d=0\), the probability of a node having \(0\) links is \(e^{-(n-1)p}\) (for large \(n\)). Thus, if we want to expect just 1 node to be isolated, we can set:

\[ e^{-(n-1)p} = 1/n \] and solving for the average degree, \((n-1)p\) yields: \[ (n-1)p = \log(n) \]

Thus, if the average degree is substantially above \(\log(n)\), \(\mathbb{P}(\text{there is an isolated node}) \rightarrow_{n\rightarrow\infty} 0\), and vice versa, the probability goes to 1. This is also known as the threshold for a phase transition.

Example 4: Symmetric Connections Model (Game-Theory)

Now, we incorporate game-theory, and consider when links are actively chosen by agents in the model (i.e. with the Florentine marriages).

For example:

  • Links could be friendships between players, which offer benefits like favors, information, etc.
  • Indirect links (paths) from \(i\) to \(j\) could offer some (lesser) indirect benefits, so it deteriorates with the “distance” of a friendship.

We formalize with a “reward” parameter \(\delta \in [0,1]\), and let the benefit gained for a connection with \(i,j\) to be \(\delta^{l_{ij}}\), where \(l_{ij}\) is the “distance” (number of links in shortest path) between them.

We include a fixed cost \(c > 0\) for a “maintaining” a direct friendship, i.e. the total cost for \(i\) would be \(c\cdot d_i\) where \(d_i\) is the number of neighbors (degree) that \(i\) has.

Thus, we can classify player \(i\)’s utility in network \(g\) by: \[ u_i(g) = \sum_{j\neq i: i,j \text{path connected}} \delta^{l_{ij}(g)}-d_i(g)\cdot c \]

image.png

Now, we can start asking questions like:

  1. What model \(g^*(\delta, c)\) is best for overall maximizing society utility, i.e. \(g^* = \text{argmax}_g \sum_i u_i(g)\)?
  2. What networks might form (under \(\delta, c\)) when players are self-interested?

For (1), we see if \(c < \delta - \delta^2\) (already getting at most \(\delta^2\) if any existing indirect connection, then the new utility \(\delta-c\) should be larger), then adding a link between any two agents will always increase welfare, and thus, the ideal graph is complete.

When \(c > \delta - \delta^2\), most optimal for society to have 1 player with \(n-1\) links, and all other players linked to her (“star formation”). (Intuition, ensures all pairs are path-connected, and each within 2 links, while minimizing costs). Note adding another link between any two players would be suboptimal then (see previous). image.png

We kind of fall into 3 equilibriums: a complete graph if costs are low, a star if costs high, and empty graph if costs astronomically high.

Now for (2), we introduce an equilibrium concept called pairwise stability, which requires that:

  1. No agent would raise their payoff by deleting some link (deletion only requires will of one individual)
  2. No two agents would both benefit by adding a link between them (i.e. formation requires the consent of both individuals)

Case 1: \(c < \delta - \delta^2\) - unique PS equilibrium is complete graph. Always benefit by adding a link.

Case 2: \(c > \delta - \delta^2\), \(\delta > c\), i.e. star network. Note, the star is both pairwise stable and society-efficient. The peripherals don’t want to delete their links, center wouldn’t either as \(\delta > c\). No formation of a link benefits both players by condition 1.

  • However, pairwise stability is not unique. Others exist, but they don’t maximize society benefit as well.

Case 3: \(\delta > c\), then star cannot form. Any player with just one link benefits by deleting it, so any PS network that has players with links must have at least 2.

Thus,

  • Individual incentives not always aligned with societal whole
  • How can we predict which networks will form?

Exercises

1.1: A Weighted-Betweenness Measure

Let \(\ell_{ij}\) be the length of the shortest path between nodes \(i\) and \(j\) and let \(W_k(ij) = P_k(ij)/(\ell_{ij} - 1)\), (setting \(\ell_{ij} = \infty\) and \(W_k(ij) = 0\) if \(i\) and \(j\) are not connected). Then the weighted betweenness measure for a given node \(k\) be defined by

\[ WB_k = \sum_{i \ne j; i,j \ne k; \ell_{ij} < \infty} \frac{W_k(ij)/P(ij)}{(n-1)(n-2)/2}, \]

where we take the convention that \(\frac{W_k(ij)}{P(ij)} = 0/0 = 0\) if there are no paths connecting \(i\) and \(j\).

Show that:

  • \(WB_k > 0\) if and only if \(k\) has more than one link in a network and some of \(k\)’s neighbors are not linked to each other.

(\(\implies\)) Assume \(WB_k >0\), so \(\exists i,j \neq k\) s.t. \(W_k(ij) > 0\), so \(P_k(ij) > 0\), so namely, \(k\) is on some shortest path from \(i\) to \(j\). In particular, \(\delta(k) \geq 2\), and consider a shortest path \(k\) lies on, \(p = (i, \cdots, k^1, k, k^2, \cdots, j)\). If \(\exists (k^1, k^2) \in E\), this contradicts \(p\) being a shortest path from \(i\) to \(j\).

(\(\impliedby\)) Assume \(\delta(k) \geq 2\). Pick \(2\) such neighbors with no link between them, call them \(i,j\), and consider the shortest paths from \(i\) to \(j\). Clearly, \((i, k, j)\) must is a shortest path, so \(k\) lies on some shortest path for \(i,j \neq k\), and so tracing back through definitions, \(WB_k > 0. \square\)

  • \(WB_k = 1\) for the center node in a star network that includes all nodes (with \(n \ge 3\)).

  • \(WB_k < 1\) unless \(k\) is the center node in a star network that contains all nodes.

  • Calculate this measure for the the network pictured in Figure 1.3 for nodes 5 and 6.

  • Contrast this measure with the betweenness measure in (1.1).

1.3 The Upper Bound for a Star to be Efficient

Find the maximum level of cost in terms of \(\delta\) and \(n\), for which a star is an efficient network in the symmetric connections model.

1.5 Homophily and Balance Across Groups

Consider a society of two groups, where the set \(N_1\) comprises the members of group 1 and the set \(N_2\) comprises the members of group 2, with cardinalities \(n_1\) and \(n_2\), respectively. Suppose that \(n_1 > n_2\). For an individual \(i\), let \(d_i\) be \(i\)’s degree (total number of friends) and let \(s_i\) denote the number of friends that \(i\) has that are within own group. Let \(h_k\) denote a simple homophily index for group \(k\), defined by

\[ h_k = \frac{\sum_{i \in N_k} s_i}{\sum_{i \in N_k} d_i}. \]

Show that if \(h_1\) and \(h_2\) are both above 0 and below 1, and the average degree in group 1 is at least as high as the average degree in group 2, then \(h_1 > h_2\). What are \(h_1\) and \(h_2\) in the case where friendships are formed in percentages that correspond to the relevant populations?

Ch. 2: Representing and Measuring Networks