Price of Anarchy: Theory & Applications

• Price of Anarchy is a metric that compares equilibrium social cost to the optimal cost achievable by a centralized planner in decentralized systems.
• It applies across various domains such as routing, network formation, auctions, and dynamic games, highlighting regime-dependent inefficiencies.
• Recent research reveals that its interpretation varies with different welfare benchmarks, comparability assumptions, and methodological frameworks.

Price of Anarchy (PoA) is a quantitative index of the efficiency loss generated by decentralized strategic behavior. In its standard cost-minimization form, it compares the social cost attained at an equilibrium—typically the worst equilibrium in a specified class—with the minimum social cost achievable by a centralized planner. Across routing, congestion, network formation, auctions, queueing, pricing, differential games, and mean field games, PoA functions as a common comparative statistic for relating equilibrium structure to system-level performance; recent work also shows that its precise meaning depends on the welfare functional and on assumptions about interpersonal comparability of costs or utilities (Cominetti et al., 2019, Shilov et al., 9 Jun 2026).

1. Formal definitions and conventions

The canonical PoA convention is a ratio of system costs,

$$\mathrm{PoA}=\frac{\text{worst equilibrium social cost}}{\text{social optimum}},$$

so that $\mathrm{PoA}\ge 1$. In nonatomic routing with demand $m$, for example, the definition can be written as

$$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$

where $f^*(m)$ is a Wardrop equilibrium, $\lambda(m)$ is the equilibrium path cost, and $\mathrm{OPT}(m)$ is the minimum social cost (Cominetti et al., 2019). In the sum version of the Network Creation Game, the same cost-ratio structure appears as

$$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$

with $C(s)$ the total edge-purchase-plus-distance cost and $\mathcal E$ the set of Nash equilibria (Àlvarez et al., 2019).

The convention is not universal. In decentralized multi-product pricing, PoA is defined as a revenue ratio,

$\mathrm{PoA}\ge 1$0

so larger values indicate smaller efficiency loss (Chen et al., 26 Jan 2026). In two-buyer sequential multiunit auctions, the paper likewise uses equilibrium welfare divided by optimal welfare, so the worst-case PoA can be $\mathrm{PoA}\ge 1$1 rather than $\mathrm{PoA}\ge 1$2 (Ahunbay et al., 2020). A still more substantial reinterpretation appears in crisis distribution systems, where PoA is not a ratio at all but the normalized accumulated frustration induced by equilibrium trade relative to a fairness benchmark encoded by buying rights (Sychrovský et al., 2023).

These differing conventions are not merely notational. They reflect a more basic choice of welfare benchmark. Recent work on invariant PoA argues that if costs or utilities are only meaningful up to agent-specific affine transformations, then the standard utilitarian ratio is representation-dependent, and the admissible welfare aggregator must instead be matched to the assumed comparability structure (Shilov et al., 9 Jun 2026).

2. Routing and congestion games

In nonatomic routing, PoA is most naturally formulated on a directed multigraph with path flows, continuous nondecreasing edge latencies, and Wardrop equilibrium. The Beckmann–McGuire–Winsten potential identifies equilibria as minimizers of

$\mathrm{PoA}\ge 1$3

while the social optimum minimizes

$\mathrm{PoA}\ge 1$4

For $\mathrm{PoA}\ge 1$5 costs with convex marginal social costs, optimal flows are Wardrop equilibria for the marginal latencies $\mathrm{PoA}\ge 1$6. This formulation makes PoA a function of total demand, not only a worst-case constant (Cominetti et al., 2019).

That demand-indexed viewpoint yields a sharp structural picture for affine latencies $\mathrm{PoA}\ge 1$7. Equilibrium flow is piecewise linear in demand, with break points exactly at demand values where the active shortest-path set changes. The number of break points is finite, though potentially exponential in network size. Between consecutive break points, equilibrium and optimal social costs are quadratic in demand, and PoA is either monotone or decreases to a unique interior minimum and then increases. In particular, for affine costs the maximum of $\mathrm{PoA}\ge 1$8 can occur only at break points (Cominetti et al., 2019). This refines the classical sharp worst-case affine bound $\mathrm{PoA}\ge 1$9: the worst-case constant exists, but the actual inefficiency profile of a fixed network is piecewise smooth and regime-dependent.

The same literature also shows that low- and high-demand asymptotics are often benign. For nonatomic congestion games with multiple origin-destination pairs, if the network is tight in the regular-variation sense, then

$m$0

and this includes all polynomial latency functions (Colini-Baldeschi et al., 2017). In highly congested single-commodity nonatomic networks, one bounded path suffices for $m$1, and for parallel networks with asymptotically comparable regularly varying latencies the heavy-traffic limit is again 1; however, simple two-edge counterexamples show that without such asymptotic regularity, PoA may oscillate, stay bounded away from 1, or even have $m$2 (Colini-Baldeschi et al., 2016).

The growth rate of latency functions is decisive in atomic unsplittable congestion games as well. For arbitrary nondecreasing latencies, the paper on superpolynomial costs proves that if resource costs grow faster than every polynomial, then there exist games whose PoA is unbounded and increases with the number of players even when resources remain fixed; by contrast, if latencies are bounded above by a polynomial, PoA is bounded independently of the number of players (Kannan et al., 2013). A complementary random-network result links inefficiency to geometry rather than asymptotic growth: in a directed lattice with randomly placed constant and linearly congestible links, PoA is maximized at the directed percolation threshold $m$3, where a system-spanning congestible path first appears but redundant alternatives are still scarce (Skinner, 2014).

3. Network creation and adversarial network formation

In the Fabrikant et al. Network Creation Game, each player may buy undirected links at cost $m$4 and also pays the sum of graph distances to all other players. PoA therefore measures the ratio between the social cost of the worst equilibrium graph and that of the optimal graph. The asymptotic behavior depends sharply on the link-cost regime. For every fixed non-integral $m$5, the paper “Anarchy is free in network creation” proves

$m$6

whereas for each integral $m$7 it constructs equilibria showing

$m$8

The proof uses a detailed structural analysis of equilibrium graphs, including bounded-diameter and degree-layer arguments; the qualitative conclusion is an arithmetic dichotomy between integral and non-integral edge costs (Graham et al., 2013).

A different line of work studies the high-price regime near the threshold $m$9. Two complementary papers establish that for every fixed $$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$0, PoA is constant whenever $$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$1. One proof shows that every equilibrium of sufficiently large diameter must be a tree, so the known constant PoA bound for tree equilibria applies (Àlvarez et al., 2018). Another proves that in any non-tree Nash equilibrium, every nontrivial biconnected component has size at most a constant $$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$2, which forces bounded diameter and hence bounded inefficiency (Àlvarez et al., 2019). Both results substantially enlarge the known constant-PoA regime but stop short of resolving the tree conjecture for all $$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$3.

Network formation with exogenous failure leads to a different PoA landscape. In the adversary model, after the graph is formed an adversary deletes one edge, and each player’s indirect cost is the expected number of vertices from which it becomes disconnected. Under unilateral link formation, PoA remains constant for both of the paper’s special adversaries: for the simple-minded adversary, which deletes a uniformly random edge, PoA is bounded by $$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$4; for the smart adversary, which deletes an edge maximizing separation, PoA is bounded by $$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$5 (Kliemann, 2012). Under bilateral link formation, the picture changes sharply. For the simple-minded adversary the bound becomes

$$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$6

while for the smart adversary the paper proves

$$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$7

hence $$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$8 when $$\mathrm{PoA}(m)=\frac{\mathrm{SC}(f^*(m))}{\mathrm{OPT}(m)}=\frac{m\,\lambda(m)}{\mathrm{OPT}(m)},$$9 is constant (Kliemann, 2012). This is a structural separation between unilateral and bilateral consent.

4. Market, queueing, auction, and coding applications

Outside routing and network formation, PoA has been adapted to several mechanism-design and queueing environments. In the two-class preemptive-resume $f^*(m)$0 queue with strategic priority choice, equilibrium inefficiency is measured by the ratio of worst-equilibrium average waiting time to the socially optimal waiting time. The paper proves a tight global bound

$f^*(m)$1

with the extremal regime occurring asymptotically as load $f^*(m)$2 and service-time variability $f^*(m)$3 (Chamberlain et al., 2020).

In sequential multiunit auctions with two buyers and complete information, the relevant convention is equilibrium welfare divided by optimal welfare. Under arbitrary non-decreasing valuations, the exact worst-case value is

$f^*(m)$4

for $f^*(m)$5 sequentially auctioned identical items. Under concave valuations, the efficiency guarantee improves to

$f^*(m)$6

and this bound is asymptotically tight as $f^*(m)$7 (Ahunbay et al., 2020). The source of the gap is intertemporal strategic distortion: with unrestricted valuations, an early misallocation may destroy large bundle complementarities.

In decentralized multi-product pricing, product managers set prices independently under a symmetric linear demand system $f^*(m)$8. The centralized benchmark is joint revenue maximization, while the decentralized equilibrium is a Nash equilibrium of product-level revenue maximization. Under strict diagonal dominance,

$f^*(m)$9

the paper derives the tight worst-case bound

$\lambda(m)$0

and refines it to the exact spectral formula

$\lambda(m)$1

where $\lambda(m)$2 is the top eigenvalue of a normalized cross-price interaction matrix (Chen et al., 26 Jan 2026).

Risk aversion changes the mechanism-design interpretation of PoA but does not destroy smoothness-based guarantees in every setting. For many mechanisms whose quasilinear PoA bounds are proved by smoothness, the same qualitative guarantees survive under normalized risk-averse utilities if the smoothness deviations never yield negative quasilinear utility. First-price auctions and second-price auctions under pointwise no-overbidding remain constant-PoA, whereas all-pay auctions admit Bayes–Nash equilibria with unbounded PoA under sufficiently strong downside sensitivity (Kesselheim et al., 2018).

A further application appears in selfish distributed compression with network coding and correlated sources. Terminals choose both multicast flows and source coding rates, and the paper defines PoA as the ratio of the cost of the worst Wardrop equilibrium to the minimum social cost. Its principal conclusion is that, unlike the independent-source case, correlation can induce $\lambda(m)$3; the analysis combines Lagrangian duality with the supermodularity of conditional entropy (0804.1840).

5. Dynamic, informational, and mean-field extensions

The differential-games literature extends PoA from static strategic interaction to continuous-time dynamic systems. For an $\lambda(m)$4-player differential game with information structure $\lambda(m)$5, weighted social cost $\lambda(m)$6, and centralized optimal-control value $\lambda(m)$7, the paper defines

$\lambda(m)$8

The same work introduces the Price of Information, which compares PoAs across open-loop and closed-loop feedback structures, and the Price of Cooperation, which measures the effect of altruistic weighting on an individual’s cost. For scalar linear-quadratic differential games, the paper derives explicit PoA formulas and large-population asymptotics, showing in particular that feedback can be more inefficient than open loop and that PoA often increases with the number of players (Basar et al., 2011).

Mean field games supply an infinite-population analogue. There the selfish benchmark is a mean field game equilibrium $\lambda(m)$9, and the social benchmark is the solution of a McKean–Vlasov control problem. The paper defines

$\mathrm{OPT}(m)$0

For linear-quadratic extended mean field games, interaction occurs through both the law of the state and the law of the control, and explicit Riccati reductions become available. A key identity is

$\mathrm{OPT}(m)$1

which shows that inefficiency is generated entirely by the mismatch between the equilibrium and planner equations for the mean component (Carmona et al., 2018). The same paper gives necessary and sufficient conditions for $\mathrm{OPT}(m)$2, proves $\mathrm{OPT}(m)$3 in several limiting regimes such as $\mathrm{OPT}(m)$4, $\mathrm{OPT}(m)$5, $\mathrm{OPT}(m)$6, and $\mathrm{OPT}(m)$7, and shows $\mathrm{OPT}(m)$8 as $\mathrm{OPT}(m)$9 (Carmona et al., 2018).

These dynamic extensions preserve the core PoA logic—equilibrium versus planner—but make explicit that inefficiency depends not only on payoff geometry but also on information structure, state dynamics, and population coupling.

6. Invariance, comparability, and reinterpretations of welfare loss

A recent foundational development argues that conventional PoA is not always well-posed. In games with von Neumann–Morgenstern utilities, player $$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$0’s utility may be transformed as

$$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$1

without changing best responses or equilibrium sets. Under such player-specific affine transformations, utilitarian welfare and classical smoothness inequalities are generally not invariant. The 2026 paper “Invariant Price of Anarchy and Multiplicative Smoothness” therefore replaces utilitarian aggregation by weighted Nash welfare under Cardinal Non-Comparability (CNC),

$$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$2

and defines an invariant PoA as optimal Nash welfare divided by worst-equilibrium Nash welfare. It also introduces multiplicative smoothness,

$$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$3

which yields pure-Nash bounds $$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$4 and, for $$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$5, coarse-correlated-equilibrium bounds (Shilov et al., 9 Jun 2026).

The same comparability issue appears in traffic control. The 2025 paper on invariant PoA reformulates cost as reservation-adjusted surplus,

$$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$6

and defines

$$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$7

The welfare functional $$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$8 depends on the admissible transformation class: under Cardinal Unit Comparability with $$\mathrm{PoA}=\max_{s\in\mathcal E}\frac{C(s)}{\min_{s\in\mathscr S}C(s)},$$9, one obtains a utilitarian sum of surpluses; under CNC, a weighted Nash product; under extreme inequality aversion, a max-min form. In a toy routing example, a toll of 0 has lower invariant PoA than a toll of 4 under the utilitarian comparability regime, the ordering reverses under CNC, and the max-min regime makes them tie; the Zurich-network case study shows that the toll level judged least efficient likewise depends on the comparability assumption (Shilov et al., 5 Dec 2025).

These results imply that PoA is not a single metric independent of welfare theory. Its numerical value depends on which interpersonal comparisons are admissible. A related but distinct reinterpretation appears in critical-resource distribution: the Crisdis model defines

$C(s)$0

where $C(s)$1 is the normalized shortfall between a buyer’s centrally assigned rights and the amount of goods actually acquired at equilibrium. Here PoA becomes a fairness-centered analogue of the classical concept rather than a welfare or cost ratio (Sychrovský et al., 2023).

Taken together, these developments broaden the scope of the concept while also narrowing its unqualified meaning. PoA remains a unifying measure of decentralization-induced inefficiency, but the object being compared—total latency, waiting cost, revenue, surplus, Nash welfare, or frustration—must be specified explicitly. In that sense, PoA is simultaneously a game-theoretic and a welfare-theoretic notion.