Economic Models of Network Formation Networked Life CSE 112 Spring 2006 Prof. Michael Kearns Background and Motivation • First half of course: – identification/quantification of common or “universal” structural properties of “natural” networks • small diameter, high clustering, heavy-tailed degree distributions,… – development of statistical models of network formation • Watt’s “Caveman/Solaria”, alpha model, pref. att., Kleinberg’s model… – analyzed/criticized simple “transmission” dynamics • disease/fad spread, forest fires, PageRank,… • Second half of course: – examination of “rational” dynamics • interdependent security games, exchange economies,… – interaction of rational dynamics with statistical formation models • e.g. when network is formed via pref. att., what will price variation be? • This lecture: let network formation be “rational” as well • A very recent topic – thx to Eyal Even-Dar & Sid Suri A Shortest Paths Game • • • • • • Let’s consider a simply stated network formation game Have N players, consider them vertices in the network Each player has to decide which edges to build or buy Assume a fixed cost c to build an edge Player’s goal: be as “central” in the network as possible Cost to player i: – Cost(i) = S {j <> i} Distance(i,j) + c x (# edges bought by i) – Distance (i,j) = shortest-path distance between i and j in the network jointly formed by all the players – Players want to minimize their cost – So need to balance edge costs vs. centrality Comments and Clarifications • Are formalizing as a one-shot game – contrast with “gradual” or incremental statistical formation models – could imagine multi-round or stage game; more complex • Each player has a huge choice of actions – action for player i: any subset S_i of all the N-1 edges i could buy – number of choices for S_i = 2^(N-1) – cost of choosing S_i = c|S_i| • Are assuming that if i buys edge to j, j (and all others) can “use” or benefit from this edge • Joint action for all N players: – choice of edge sets for all players: S_1, S_2, …, S_N • Let G = G(S_1,S_2,…,S_N) be the result overall graph/NW From Incentives to Networks • Q: How can we view this game a NW formation model? • A: View the NWs generated as being the Nash equilibria • More precisely: say that G can be “formed” by the game if: – G = G(S_1,S_2,…,S_N) for some choices for the S_i – S_1,S_2,…,S_N form a Nash equilibrium of the shortest-paths game – so, no player i can improve Cost(i) by unilaterally: • dropping an edge they bought and saving the cost c • adding an edge they didn’t buy and paying the cost c Properties of Equilibria • Questions we might ask in NW Life: – – – – What’s the diameter of the equilibria graphs? What do their degree distributions look like? What are their clustering coefficients? Etc. • Not much known precisely, but we’ll make some inferences • Another measure of interest: the Price of Anarchy: – for a given G = G(S_1,S_2,…,S_N), consider sum of all player costs: • Cost(G) = S i Cost(i) • Let Cost* = minimum possible Cost(G) (social optimum) • Price of Anarchy = Cost(G)/Cost* for G a Nash eq. • Which Nash equilibrium? Pick worst (largest Cost(G)) • Inefficiency or cost of “capitalism” over “socialism” What Happens? • Note that for a single player, sum of distances is between – a small constant independent of N (e.g. constant diameter graphs) – ~ N^2 (e.g. a cycle or a line graph) • Price of Anarchy: – edge cost c < sqrt(N): P.O.A. < some constant (independent of N) – edge cost c > N log(N): P.O.A. < 1.5 – in between: unknown whether P.O.A. is bounded • Structural properties: very little is known, but seems – Nash equilibria very sparse • often trees, but not always! – Nash equilibria very “regular” or “structured” • e.g. “star” or “hub” graph – Small diameter? Sometimes. – Heavy-tailed degree distribution? Don’t know. – High clustering? Seems unlikely. Kleinberg’s Model • Similar in spirit to the a-model • Start with an n by n grid of vertices (so N = n^2) – add local connections: all vertices within grid distance p (e.g. 2) – add distant connections: • q additional connections • probability of connection at distance d: ~ 1/d^r – so full model given by choice of p, q and r – large r: heavy bias towards “more local” long-distance connections – small r: approach uniformly random • Kleinberg’s question: – what value of r permits effective search? • Assume parties know only: – grid address of target – addresses of their own direct links • Algorithm: pass message to neighbor closest to target An Economic Variation on Kleinberg • Again have N players/vertices, but arrange them in a grid • Grid connections provide free connectivity • Instead of variable probabilities for long-distance edges, introduce variable costs: – – – – Let cost to i to purchase edge to j = g(i,j)^a g(i,j) = grid or “Manhattan” distance from i to j a = some constant value so cost grows with distance on grid, at a rate determined by value a • So now just have another network formation game • Another striking “tipping point”: – for any a <= 2, all Nash equilibria have constant diameter • i.e. diameter does not grow with N! – for any a > 2, all Nash equilibria have unbounded diameter • i.e. diameter grows with N – again, Nash equilibria seem to be “regular”, but we don’t know much at this point… Economic Formation + Economic Dynamics • Recall our simple 2-good exchange economy model: – – – – – – start with a bipartite network between “buyers” and “sellers” buyers start with $1 but value only wheat sellers start with 1 unit wheat but value only dollars prices = proposed rates of exchange price p means party is willing to exchange their $1/1u for p of other equilibrium prices: prices for each party such that • all parties behave “rationally” = trade only with “best price” neighbor(s) • everyone is able to trade away their initial endowment – at equilibrium, party charging p only trades with parties charging 1/p – equilibrium prices = equilibrium wealths • Before we examined wealth distribution for given networks Price Variation vs. a and n n=1 n = 250, scatter plot n=2 Exponential decrease with a; rapid decrease with n (Statistical) Structure and Outcome • Wealth distribution at equilibrium: • Price variation (max/min) at equilibrium: • Random graphs result in “socialist” outcomes • Price variation in arbitrary networks: – Power law (heavy-tailed) in networks generated by preferential attachment – Sharply peaked (Poisson) in random graphs – Grows as a root of n in preferential attachment – None in random graphs – Despite lack of centralized formation process – – – – Characterized by presence/absence of a perfect matching Alternately: an expansion property Theory of random walks Economic vs. geographic isolation Economic Formation + Economic Dynamics • Now imagine that network is not given, but must be bought • Each player i (buyer or seller) chooses a set S_i of trading partners on the opposing side (sellers or buyers) – as before, assume each edge costs c to purchase, where c can be in dollars (buyers) or wheat (sellers) – edge purchased by one party can be used “for free” by other party – once all parties have decided what edges to buy, have some graph G – at price equilibrium of G, player i receives wealth W_i – overall payoff to player i: • W_i – c x (#edges purchased by i) • Can again view this as a network formation model: – possible networks = Nash equilibria of the above game – that is, a choice of edge sets S_i bought for each player i such that no party can unilaterally improve their overall payoff by dropping or adding an edge Price Variation? • • • • • • • Can price/wealth variation still be present? How much? What do the equilibrium networks look like? Suppose G = G(S_1,S_2,…,S_N) is a Nash equil. of this game Let W_min < 1 be smallest wealth (w/o edge costs) Let c be the cost of an edge Then must have W_min > 1 – c (same as c > 1 – W_min) This inequality is tight – can construct networks where W_min = 1-c • So rational NW formation eradicates inequality… – …up to the cost to buy an edge Network Structure • If G is some Nash equil. of this game, then – G equals its “exchange subgraph” --- no “unused” edges – G consists of (possibly multiple) connected components • each component has uniform prices p, 1/p – don’t know much yet about structure within components • some components may have a range of degrees