Semilinear Representations for Series-Parallel Atomic Congestion Games

We consider atomic congestion games on series-parallel networks, and study the structure of the sets of Nash equilibria and social local optima on a given network when the number of players varies. We establish that these sets are definable in Presburger arithmetic and that they admit semilinear representations whose all period vectors have a common direction. As an application, we prove that the prices of anarchy and stability converge to 1 as the number of players goes to infinity, and show how to exploit these semilinear representations to compute these ratios precisely for a given network and number of players


Introduction
Network congestion games are used to model situations in which agents share resources such as routes or bandwidth [31], and have applications in communication networks (e.g.[1]).We consider the atomic variant of these games, where each player controls one unit of flow and must assign it to a path in the network.All players using an edge then incur a cost (a.k.a.latency) that is a nondecreasing function of the number of players using the same edge.Since all players try to minimize their own cost, this yields a noncooperative multiplayer game.It is well known that Nash equilibria exist in these games [31] but that they can be inefficient, that is, a global measure such as the total cost, or the makespan may not be minimized by Nash equilibria [30].
To quantify this inefficiency, [27] introduced the notion of price of anarchy (PoA), which is the ratio of the cost of the worst Nash equilibrium and the social optimum.Here, social optimum refers to the sum of the individual costs.A tight bound of 5  2 on this ratio was given in [3,9].Various works have studied bounds on the PoA for restricted classes of graphs or types of cost functions; see [28].While the price of anarchy is interesting to understand behaviors that emerge in a system from a worst-case perspective, the best-case is also interesting if, for instance, the network designer is able to select a Nash equilibrium.

32:2 Semilinear Representations for Series-Parallel Atomic Congestion Games
The price of stability (PoS) is thus the ratio between the cost of the best Nash equilibrium and the social cost, and was studied in [2]; a bound of 1 + √ 3/3 was given in the atomic case; see [7,8,2].
These bounds have been studied for restrictions of this problem such as particular classes of graphs.One such class is series-parallel networks which are built using single edges, or parallel and serial composition of smaller series-parallel networks (see e.g.[34]).On these networks with affine cost functions, [24] reports that PoA is between 27  19 and 2, where a previous known lower bound due to [19] was 15  11 .[19] also proves an upper bound on PoA for extension-parallel networks which are a strict subclass of series-parallel networks.
While tight bounds are known for atomic network congestion games and even for particular subclasses, this does not help one to evaluate the price of anarchy of a given specific game with a given number of players.In fact, the upper bounds mentioned above are obtained by building particular networks, and these are shown to be tight by exhibiting families of instances in which both the networks and the number of players vary.One of our objectives is to provide tools to analyze a given network congestion game, by computing both ratios precisely for varying numbers of players.We are interested both in the case of a given number of players, and in the case of the limit behavior.Note that we are considering a hard problem since computing extreme Nash equilibria, that is, best and worst ones, is NP-hard in networks with only three and two players respectively [33].
In this paper, we consider series-parallel networks with linear cost functions and establish interesting properties of their Nash equilibria and social optima.We start with the observation that Nash equilibria and locally social optima can be expressed in Presbuger arithmetic; it follows that these sets admit semilinear representations [22].Our main result is that that these semilinear sets have a particular structure.In fact, the flows (i.e.edge loads) induced by Nash equilibria and local social optima admit semilinear representations with a common direction for all period vectors, which is moreover efficiently computable.We call this direction the characteristic vector of the network.Intuitively, this vector determines how the flow evolves in Nash equilibria and social local optima as the number of players increases.
We believe that the form of these representations and the characteristic vector have an interest on their own.We give one application of these representations here, namely, that the PoA and the PoS both tend to 1 as the number of players goes to infinity in series-parallel networks with linear cost functions with positive coefficients.This result was proven recently [37]; we provide here new techniques and thus an alternative proof.Observe that a similar result holds in nonatomic congestion games [11,10] (see Section 6 for a discussion).
We also illustrate how these semilinear representations allow one to study the evolution of PoA and PoS for a given network.The computation of these representations is an expensive step, but once this is done, thanks to the particular form of the representations, for any n, one can easily query the exact value for PoA and PoS in the network instantiated with n players.One can thus analyze both ratios precisely and specifically for a given network as a function of n, while the limits will always be 1.

Illustrating Example
Let us illustrate our results on a simple example.Consider the network with two parallel links in Figure 1a.There are n players who would like to go from src to tgt, by taking either the bottom edge (b) with the cost function x → x, or the top edge (t) with cost function x → 4x.The cost function determines the cost each player pays when taking a given edge, and it is a function of the total number of players using the same edge.For instance, for n = 4, assume that 3 players use b, and 1 uses t.Then, each player using b pays a cost of 3, while the only player using t pays 4. Here, a strategy profile can be seen as a pair (k, n − k) determining how many players take t and how many take b.(b) Nash equilibria in the network on the left for varying total numbers of players.There is a point at coordinates (x, y) if a strategy profile assigning x players to b, and y players to t is a Nash equilibrium.Intuitively, in Nash equilibria, the number of players taking b should be roughly four times the number of them taking t; so 4  5 of them should take b, and 1 5 of them should take t.This would make sure that both edges have identical cost, and make profitable deviations impossible.Although this intuition holds in the nonatomic case, players cannot always be split with this proportion in the atomic case, as in the case of n = 4 above; and there are indeed equilibria that do not match this proportion exactly.Figure 1b shows the Nash equilibria in this game, while the line with direction (4, 1) shows the ideal distribution (as in the nonatomic case).Not all Nash equilibria are on this line, but one can notice that they do form a tube around this line that go in the same direction.Formally, our results determine that the Nash equilibria form the semilinear set B NE + ⃗ δ • N, where ⃗ δ = (4, 1), and B NE = {(0, 0), (1, 0), (2, 0), (3, 0), (4, 0), (3, 1)}.In other terms, it is the union of the integer points of six lines with the same direction vector ⃗ δ.
Similarly, we describe the set of locally social optimal profiles and show that it admits a semilinear representation.Locally optimal profiles are those in which the social cost cannot be decreased by changing the strategy of a single player; the formal definition is given in Section 3.1.It turns out that these have a structure very similar to that of Nash equilibria.In our example, local optima are given by B SO + ⃗ δ • N, thus with the same vector ⃗ δ as above, and with B SO = {(0, 0), (1, 0), (2, 0), (2, 1)}.
The particular structures of these semilinear sets we obtain allow us to compute the prices of anarchy and stability for any given number of players.In fact, given n, one can easily find, in a set B + ⃗ δ • N, all strategy profiles with n players.So one can compute the worst and the

32:4
Semilinear Representations for Series-Parallel Atomic Congestion Games best Nash equilibria, as well as the social optimum for any given n.The plot in Fig. 2 shows how PoA and PoS evolve as n increases, and was calculated from the previous representations.This plot illustrates our objective of analyzing the inefficiency of these games precisely for varying n.Even in this simple example, PoA and PoS can significantly vary depending on n, and they are far from the known tight bounds for the whole class of networks.One can notice in Fig. 2 that both PoA and PoS seem to converge to 1 as n increases.This is indeed the case and is a consequence of our results (Theorem 14).We believe that this approach can allow one to better understand the specific network under analysis.Section 5 contains more examples.
Paper Overview.We provide formal definitions in Section 2. Section 3 characterizes the form of semilinear representations of local social optima; and Section 4 proves that of Nash equilibria.Section 5 shows how to use these representations to compute PoA and PoS.We provide more discussion on related work in Section 6, and present conclusions in Section 7.

Network Congestion Games and Series-Parallel Arenas
A network is a weighted graph A = ⟨V, E, orig, dest, wgt, src, tgt⟩, where V is a finite set of vertices, E is a finite set of edges, orig : E → V and dest : E → V indicate the origin and destination of each edge, wgt : E → N >0 is a weight function assigning positive weights to edges, and src and tgt are, respectively, a source and a target states.Let In(v) and Out(v) denote, respectively, the set of incoming and outgoing edges of v.We restrict to acylic networks in which all vertices are reachable from src and tgt is reachable from all vertices.A path π of A is a sequence e 1 e 2 . . .e n of edges with dest(e i ) = orig(e i+1 ) for all 1 ≤ i ≤ n − 1.For an edge e and a path π = e 1 e 2 . . .e n , we write e ∈ π if e = e i for some 1 ≤ i ≤ |π|.We will sometimes see paths as sets of edges and apply set operations such as intersection and set difference.Let Paths A (s, t) denote the set of simple paths from s to t in A, and let Paths A = Paths A (src, tgt).
In this work, we consider series-parallel networks [34].These are built inductively from single edges using serial and parallel composition.Two networks are composed in series to a new network denoted by A 1 ; A 2 = ⟨V, E, orig, dest, wgt, src, tgt⟩ obtained by taking the disjoint union of A 1 and A 2 , and merging the vertices tgt 1 and src 2 , and setting src = src 1 , tgt = tgt 2 .Two networks ⟨V, E, orig, dest, wgt, src, tgt⟩ obtained by taking the disjoint union of A 1 and A 2 , and merging src 1 and src 2 , then merging tgt 1 and tgt 2 , and setting src = src 1 , tgt = tgt 1 .A network is said to be series-parallel if it is either a single edge, or it is a serial or parallel composition of series-parallel graphs.
A network congestion game (NCG) is a pair G = ⟨A, n⟩ where A = ⟨V, E, orig, dest, wgt, src, tgt⟩ is a network, and n ∈ N is the number of players in the game.We consider the symmetric case where all players start at src and want to reach tgt.A strategy of a player is a path in Paths A .The setting can be seen as a one-shot game in which each player selects a strategy simultaneously.In our study, we do not need to identify players, we thus represent strategy profiles by counting how many players choose each strategy.That is, a strategy profile is a tuple ⃗ p = (p π ) π∈Paths A where p π is the number of players taking path π.In this case, the number of players is given by ∥⃗ p∥ = π∈Paths p π ; thus ⃗ p is a strategy profile in the game ⟨A, ∥⃗ p∥⟩.Let S(A) denote the set of all strategy profiles, and S n (A) the set of strategy profiles with n players, that is, {⃗ p ∈ S(A) | n = ∥⃗ p∥}.For π ∈ Paths A , let ⃗ p + π (resp.⃗ p − π) denote the strategy profile obtained by incrementing (resp.decrementing) p π by one.
Another useful notion we use is the flow of a strategy profile, which consists in the projection of a strategy profile to edges.Formally, given a strategy profile ⃗ p, flow(⃗ p) = (q e ) e∈E where q e = π∈Paths A :e∈π p π , that is, q e is the number of players that use the edge e in the profile ⃗ p.This vector satisfies the following flow equations: We refer to a vector ⃗ q = (q e ) e∈E with nonnegative coefficients satisfying (1) as a flow; and denote by F(A) the set of all flows.Observe that F(A) is the image of S(A) by flow.For a flow ⃗ q, let ∥⃗ q∥ = e∈E:orig(e)=src q e , which is the number of players.Let F n (A) define the set of flows with n players as follows: Observe that this corresponds to a flow of size n, and that several strategy profiles can project to the same flow.
For a strategy profile ⃗ p and π ∈ Paths A , each player using the path π incurs a cost equal to cost π (⃗ p) = Σ e∈π wgt(e) • flow e (⃗ p), where flow e (⃗ p) is the number of players using edge e in the strategy profile ⃗ p.The social cost of a strategy profile ⃗ p is the sum of the costs for all players, i.e., soccost Observe that the cost of a path in a strategy profile, and the social cost of a strategy profile, are determined by the flow of that profile.Thus, we define the social cost of a flow ⃗ q as soccost(⃗ q) = e∈E q 2 e • wgt(e) (in fact, q e players use strategies that include e, and each of them pays q e wgt(e) for crossing this edge).A flow ⃗ q ∈ F n is socially optimal if soccost(⃗ q) = opt(G).
A strategy profile ⃗ p is a Nash equilibrium if no player can reduce their cost by unilaterally changing strategy, i.e., if where ⃗ p ′ is defined by , and p ′ τ = p τ for all other paths τ .In fact, cost π (⃗ p) is the cost of a player playing π in the profile ⃗ p, while cost π ′ ( ⃗ p ′ ) is their cost in the new profile ⃗ p ′ obtained by switching from π to π ′ .Let NE(A) denote the set of strategy profiles satisfying (2), that is, the set of Nash equilibria, and let NE n (A) denote the set of Nash equilibria for n players.The price of anarchy is the ratio of the social cost of the worst Nash equilibrium, and the social optimum: PoA(⟨A, n⟩) = max ⃗ p∈NEn soccost(⃗ p)/opt(⟨A, n⟩).The price of stability is the ratio of the social cost of the best Nash equilibrium, and the social optimum: PoS(⟨A, n⟩) = min ⃗ p∈NEn soccost(⃗ p)/opt(⟨A, n⟩).

Presburger Arithmetic and Semilinear Sets
We recall the definition and some basic properties of semilinear sets; see e.g.[23] for more details.A set S ⊆ N m is called linear if there is a base vector ⃗ b ∈ N m and a finite set of Such a linear set S will be denoted as L( ⃗ b, P ).

32:6 Semilinear Representations for Series-Parallel Atomic Congestion Games
A set S ⊆ N m is said to be semi-linear if it is a finite union of linear sets.Therefore, a semi-linear set S can be written in the form S = ∪ i∈I L( ⃗ b i , P i ) where I is a finite set, P i 's are finite sets of period vectors, and the ⃗ b i are the base vectors of the same dimension.Note that in a linear set L( ⃗ b, P ), P can be empty, which corresponds to a singleton set.Thus, finite sets are semilinear; and the union of any semilinear set with a finite set is semilinear.Furthermore, each semilinear set admits a non-ambiguous representation in the sense that S = ∪ i∈I L( ⃗ b i , P i ) such that each P i is linearly independent and L( ⃗ b i , P i ) ∩ L( ⃗ b j , P j ) = ∅ for all i ̸ = j ∈ I [16,25].
Presburger arithmetic is the first-order theory of integers without multiplication.It is well-known that any set expressible in Presburger arithmetic is semilinear [22].So, in order to show that a set is semilinear, one can either exhibit its semilinear representation, or show that it is expressible in Presburger arithmetic.

Local Social Optima
In this section, our goal is to obtain a representation of social optima in a given network congestion game as a function of the number n of players.Characterizing the social optimum directly by a formula brings two difficulties.First, expressing that a flow ⃗ q is optimal would require to quantify over all flows ⃗ q ′ and writing that soccost(⃗ q) ≤ soccost( ⃗ q ′ ), so such a formula contains universal quantifiers.Second, the formula is quadratic since soccost(⃗ q) = e∈E q 2 e • wgt(e), so this cannot be represented by a semilinear set.Here, we introduce the notion of local optimality which allows us to circumvent both difficulties, providing semilinear representations which, moreover, allow us to compute the global optimum.

Locally-Optimal Profiles
Let us fix a series-parallel network A. Intuitively, a strategy profile is locally-optimal if the social cost cannot be reduced by exchanging one path for another.Formally, ⃗ p ∈ S n (A) is locally-optimal if for all π, π ′ ∈ Paths A with p π > 0, soccost(⃗ p) ≤ soccost(⃗ p − π + π ′ ).By extension, a flow ⃗ q ∈ F n (A) is locally-optimal if it is the image of a locally-optimal strategy profile.Observe that the (global) social optimum is locally-optimal.
In the following lemma, we see paths π, π ′ as sets of edges.
▶ Lemma 1.In a network congestion game ⟨A, n⟩, a flow ⃗ q is locally-optimal if, and only if, for all π, π ′ ∈ Paths A such that ∀e ∈ π, q e > 0, We define LocOpt(A) to be the set of locally-optimal flows, and LocOpt n (A) those with n players.It follows from Lemma 1 that LocOpt(A) and LocOpt n (A) are expressible in Presburger arithmetic, and are thus semilinear.We will now characterize the form of the semilinear set describing LocOpt(A) by proving that it admits a single and computable period vector ⃗ δ, that is, LocOpt(A) can be written as B ∪ i∈I L( ⃗ b i , ⃗ δ), where B is a finite set of flows, I is a finite set of indices, and ( ⃗ b i ) i∈I are the base vectors.

Large Numbers of Players
To simplify the proof of the characterization of the period vector ⃗ δ, we would like to consider instances in which (3) holds for all paths π, π ′ ∈ Paths A , that is, we would like to get rid of the assumption on π in Lemma 1.It turns out that the assumption that q e > 0 for all edges e of π holds whenever the number of players is large enough.Moreover, we do not lose generality by focusing on these instances; in fact, as we will see, a semilinear representation with the same period vector ⃗ δ for all locally-optimal profiles can be derived once this representation is established for instances with large numbers of players.
The next lemma shows that all edges are used in locally-optimal profiles whenever the number of players is sufficiently large.This property is specific to series-parallel graphs and may not hold in a more general network, as the following example shows.▶ Example 2. Consider the network of Fig. 3.We claim that the edge from v to v ′ cannot be used in any locally-optimal profile.Write a, b and c for the number of players taking paths Observe (by contradiction) that if there are at least two players, then paths π a and π c will be taken by at least one player.Writing Eq. ( 3) for π a and π c yields −1 ≤ a − c ≤ 1.Assuming b > 0, we can also apply Eq. ( 3) to π b and π a , and get −2a + 8b + 6c − 5 ≤ 0. It follows 8b + 4c − 7 ≤ 0. This cannot be preserved when the number of players grows.Hence π b cannot be used in any locally-optimal profile.◀ ▶ Lemma 3. In all series-parallel networks A, there exists n 0 ∈ N such that for all n ≥ n 0 , all flows ⃗ q ∈ LocOpt n (A) are such that q e > 0 for all e ∈ E.
Proof.We inductively define a value n 0 (A) for each series-parallel network A, such that n 0 (A) ≥ 4 (for technical reasons) and satisfying the following property: The lemma follows by taking k = 1.
If A is a single edge, we define n 0 (A) = 4, and the property trivially holds.If A = A 1 ; A 2 , then we define n 0 (A) = max(n 0 (A 1 ), n 0 (A 2 )).Observe that if ⃗ q is locally-optimal in A, then each ⃗ q| Ai is locally-optimal in A i , and they have the same number of players.So the property holds by induction.
The non-trivial case is when ), and n 0 (A) = 2|E|W m 2 , where E is set of edges of A and W is the largest weight of A (notice that W ≥ 1).Then m ≥ 4 and n 0 (A) ≥ 4. Take n ≥ kn 0 (A), and ⃗ q ∈ LocOpt n .
We show that for all i ∈ {1, 2}, we have ∥⃗ q| Ai ∥ ≥ kn 0 (A i ).Towards a contradiction, assume for example that ∥⃗ q| A1 ∥ < kn 0 (A 1 ).Then For all pair (π 1 , π 2 ) ∈ Paths A1 × Paths A2 , by (3) we have that e∈π2 wgt(e)(2q e − 1) ≤ e∈π1 wgt(e)(2q e + 1).Take an arbitrary edge e 0 of π 2 ; then e 0 does not appear in π 1 , and we get 2q e0 ≤ 2W Now, take k ′ = 2k|E|W (m−1).Then k ′ ≥ 1, and ∥⃗ q| A2 ∥ > k ′ m ≥ kn 0 (A 2 ).By induction hypothesis applied to A 2 , we have q e0 ≥ k ′ , which implies . By the previous lemma, all ⃗ q ∈ LocOpt ≥n0 (A) satisfy ( 3) for all pairs of paths π, π ′ .Observe that LocOpt(A) and LocOpt ≥n0 (A) have the same period vectors.In fact, LocOpt ≥n0 (A) differs from LocOpt(A) only by a finite set; so, given a semilinear representation i∈I L( ⃗ b i , P i ) for the former, one can obtain a representation for the latter as B ∪ i∈I L( ⃗ b i , P i ) where B is the finite difference between the two.Therefore, establishing that LocOpt ≥n0 (A) has a single period vector suffices to prove the same result for LocOpt(A).Note also that I is non-empty here since the set LocOpt ≥n0 (A) is infinite.
The following lemma shows that the period vectors of LocOpt ≥n0 (A) assign the same cost to all paths.Intuitively, this is because if the cost along two paths were different in the period vector, then by adding a large number of copies of the period vector to its base vector, one could amplify this difference and obtain a vector that is not locally-optimal.
▶ Lemma 4. In a series-parallel network A, for all period vectors ⃗ d ∈ N E of a semilinear representation of LocOpt ≥n0 (A), there exists κ ≥ 0 such that for all π ∈ Paths A , we have e∈π wgt(e) • d e = κ.

A Unique Period Vector: The Characteristic Vector
We now establish that LocOpt(A) admits a unique period vector.For any κ ∈ R, we study the following system E(κ) of equations with unknowns {q e } e∈E : ∀v ∈ V \ {src, tgt}, Note that all period vectors of LocOpt ≥n0 (A) satisfy the above.In fact, (4) comes from Lemma 4, and ( 5) is the set of flow equations (1).
The following lemma states that E(κ) has a unique solution whenever we fix κ.
▶ Lemma 5.For a series-parallel network A, for each κ ∈ R, the system E(κ) admits a unique solution.
The system E(κ) is actually the characterization of the flows of Nash equilibria in the non-atomic congestion games, and the unicity of the solution of this equation system is known (see [12]).We represent the system E(κ) in the matrix form as M A • X = κ ⃗ b, where X = (q e ) e∈E is the vector of unknowns, and ⃗ b a {0, 1}-column vector.Lemma 5 means that M A admits a left-inverse M −1 A .It follows that all period vectors can be written as κM ▶ Definition 6 (Characteristic Vector).Let κ 0 denote the least rational number such that Since period vectors of LocOpt ≥n0 (A) have natural number coefficients, any κ corresponding to a period vector is also a natural.We have that all period vectors are integer multiples of ⃗ δ.
⃗ δ, and κ κ0 is an integer since otherwise its denominator would divide ⃗ δ, which would contradict the minimality of κ 0 .▶ Corollary 7. Consider a series-parallel network A, and its characteristic vector ⃗ δ.There exist finite sets of vectors B and Proof.Since each linear set can be assumed to have linearly-independent period vectors (Section 2.2) all linear sets included in LocOpt ≥n0 (A) can be written in the form of ⃗ b + m ⃗ δN, We show that m = 1.By Lemma 3, ⃗ b has only positive coefficients, so it satisfies (3) for all pairs of paths.We show that ⃗ b + k ⃗ δ also satisfies (3) for all k ≥ 0, which proves that L( ⃗ b, ⃗ δ) is included in LocOpt ≥n0 (A).For all paths π, π ′ , consider (3) by adding an identical term to both sides: ▶ Remark 8.Note that we chose here to study the semilinear representations of locallyoptimal flows, rather than locally-optimal strategy profiles.The latter are also expressible in Presburger arithmetic, thus also admit a semilinear representation.However, the set of locally-optimal strategy profiles admit, in general, several linearly independent period vectors, so their representation is more complex, and more difficult to use, for instance, to compute the global optimum for given number n of players.
To see this, consider the example of Fig. 4. Consider the strategy profile ⃗ p with p ac = p bd = 1 and p ad = p bc = 0; and ⃗ p ′ such that p ′ ac = p ′ bd = 0 and p ′ ad = p ′ bc = 1.For all k ≥ 0, both k⃗ p and k ⃗ p ′ are socially optimal, but they are linearly independent.In larger networks, there can be a larger number of period vectors due to similar phenomena.Notice however that flow(⃗ p) = flow( ⃗ p ′ ), that is, the projections of these period vectors to their flows are identical, and are, in fact, equal to the characteristic vector.

Nash Equilibria
In this section, we show how to compute a semilinear representation of the flows of Nash equilibria, which will allow us to compute the costs of the best and the worst equilibria.

F S T T C S 2 0 2 2 32:10 Semilinear Representations for Series-Parallel Atomic Congestion Games
The following lemma is a characterization of Nash equilibria that follows from (2).
It follows from the previous lemma that Nash equilibria, but also their flows, are definable in Presburger arithmetic.
The study of the semilinear representation is similar to what was done for locally-optimal profiles.We prove that the period vectors of flow(NE(A)) are colinear to ⃗ δ, thus establishing the form of their semilinear representation.In particular, we use Lemma 5 and show that the period vectors are multiples of the characteristic vector.Here, we consider flow(NE(A)) rather than NE(A) due to Remark 8 which also holds for Nash equilibria.
The following lemma shows that a semilinear representation of flow(NE(A)) can be deduced from that of NE(A) and follows by the definition of flow.In particular, all period vectors of the former are projections of the period vectors of the latter by flow.

▶ Lemma 11. For any semilinear representation
As in the case of locally-optimal strategy profiles, we establish that the period vectors of flow(NE(A)) for series-parallel networks have constant cost along all paths.▶ Lemma 12.For a series-parallel network A, and any period vector ⃗ q of flow(NE(A)), there exists κ, such that for all π ∈ Paths A , e∈π wgt(e) • q e = κ.
Proof.We use structural induction on the series-parallel network.The base case is when the network is a single edge, which trivially satisfies the property.
Consider a network A = ⟨V, E, orig, dest, wgt, s, t⟩ with Let ⃗ q be a period vector of flow(NE(A)).By Lemma 11, there exists a period vector ⃗ p of NE(A), with ⃗ q = flow(⃗ p).Let us show that each ⃗ q| Ai is a period vector of flow(NE(A i )).It suffices to show that prj Ai (⃗ p) is a period vector of NE(A i ): since ⃗ q| Ai = flow(prj Ai (⃗ p)), we can then conclude by Lemma 11.
For some base vector ⃗ b, we have that ⃗ b + k⃗ p ∈ NE(A) for all k ≥ 0. By Lemma 16, Let us call ⃗ q i = ⃗ q| Ai .By the induction hypothesis, there exist constants κ 1 , κ 2 such that ∀π ∈ Paths Ai , e∈π wgt(e) • q i e = κ i .

It follows that
where π 1 π 2 denotes the decomposition of π such that π i ∈ Paths Ai .

Consider now the case
Let ⃗ q be a period vector of flow(NE(A)).By Lemma 11, there exists a period vector ⃗ p of NE(A), with ⃗ q = flow(⃗ p).Let us show that each ⃗ q| Ai is a period vector of flow(NE(A i )).The argument is identical to the first case, using Lemma 17 in place of Lemma 16.It suffices to show that ⃗ p| Ai is a period vector of NE(A i ) since we can then conclude by Lemma 11.For some base vector ⃗ b, we have that ⃗ b+k⃗ p ∈ NE(A) for all k ≥ 0. By Lemma 17, ( ⃗ b + k⃗ p)| Ai ∈ NE(A i ), and since ( ⃗ b + k⃗ p) = ⃗ b| Ai + k⃗ p| Ai , ⃗ p| Ai is indeed a period vector.
We need to prove that κ 1 = κ 2 .Assume otherwise; for instance, κ 1 < κ 2 .Then, for paths π 1 , π 2 ∈ Paths A , we have cost π1 (⃗ q) < cost π2 (⃗ q).Consider a period vector ⃗ p of NE(A) with flow(⃗ p) = ⃗ q.Observe that ⃗ p and its multiples are also a Nash equilibria (by Lemma 9).Then, there must exist π 2 ∈ Paths A2 such that p π2 > 0; if not, paths in Paths A2 would become profitable deviations for k⃗ p for large k.Take such a π 2 ∈ Paths A2 .By ( 6), for all π 1 ∈ Paths A1 , and all k ≥ 0, we have We do not know whether the m i in the above corollary can be different from 1.We did not encounter such a case in our experiments, and we conjecture that ⃗ δ is the only period vector of flow(NE(A)).
An immediate consequence of Corollaries 7 and 13 is that the prices of anarchy and stability converge to 1 for a fixed series-parallel network, when the number of players goes to infinity.This is intuitively due to the fact that both LocOpt(A) and flow(NE(A)) have the same direction ⃗ δ.This result already appeared recently in [37]; our setting provides an alternative proof.As we already argued, the social cost only depends on the flow, so worst n and best n can be computed by maximizing or minimizing the social cost among flow(NE(A)) restricted to n players.The optimum can be computed by minimizing over LocOpt n (A) since the global optimum is also locally optimal.

Let flow(NE(
so that all Nash equilibria with n players belong to some L( ⃗ b i , m i ⃗ δ), and similarly, all local optima with n players are in some We have that Similarly, Consider (i 0 , i ′ 0 ) ∈ I × I ′ such that worst n is maximized for i 0 ∈ I, and opt n is minimized for i ′ 0 ∈ I ′ for infinitely many n.Let (α k ) k≥0 denote the increasing sequence of indices n such that this is the case.We are going to show that the limit of the sequence (worst α k /opt α k ) k∈N is 1 (independently of i 0 and i ′ 0 ), which yields the result.We have We have worst A similar expression can be obtained for opt α k since it has the same form as worst α k .
In particular, the first term is again A. We can obtain that opt It follows that lim n→∞ PoA(⟨A, n⟩) = 1.This also implies that lim n→∞ PoS(⟨A, n⟩) = 1.◀

Semilinear Representations
Let us show how the semilinear representation for LocOpt(A) can be computed.First, the characteristic vector ⃗ δ can be computed by solving the homogeneous equation system E(0) using a symbolic solver (so as to obtain a rational solution), and multiplying the unique solution by the gcd of its coefficients.Next, one can incrementally construct the semilinear representation using an integer-arithmetic solver to find the linear sets and the finite set B. This can be done with quantifier-free formulas only.Although integer linear programming is already NP-hard [6,35], available solvers are efficient for small instances.
At a given iteration, assume that the current subset of LocOpt(A) is B ∪ i∈I L( ⃗ b i , ⃗ δ).We write a linear quantifier-free formula ϕ(⃗ q) with free variables a flow ⃗ q which requires that ⃗ q is locally optimal (by Lemma 1), and that ⃗ q is not included in the current set.We already saw that the former is a Presburger formula.The latter constraints can be written as but the existentially quantified k can be determined from the number of players of ⃗ q, ⃗ b i , ⃗ δ, so this can be simplified as follows: where ∥ ⃗ δ∥ and ∥ ⃗ b i ∥ are fixed numbers.
If ϕ(⃗ q) is not satisfiable, then the current representation is complete.Otherwise, a model satisfying the above formula gives a new vector ⃗ q that is locally optimal.To determine whether ⃗ q should belong to B, or whether L(⃗ q, ⃗ δ) is to be added to our set, we simply check if ⃗ q + ⃗ δ satisfies the condition of Lemma 1: if this is the case, then we keep the linear set, otherwise we add ⃗ q to B. In fact, for all k ≥ 1, ⃗ q + k ⃗ δ has the same set of paths π satisfying ∀e ∈ π, (q e + kδ e ) > 0, so this check is sufficient.Since the set admits a finite semilinear representation, this procedure terminates.
Let us explain the computation of flow(NE(A)).Let NE(⃗ p) denote the linear constraints of (6).Assume that we currently have a subset of flow(NE A ) in the form The following formula ψ(⃗ p, ⃗ q) with free variables ⃗ p, ⃗ q is satisfiable iff some Nash equilibrium ⃗ p (with flow(⃗ p) = ⃗ q) is not in the set: Assume that this is satisfiable, and let ⃗ p, ⃗ q be a model.We need to check whether ⃗ q is to be added to B, or whether L(⃗ q, m i ⃗ δ), for some m i , is to be included.We use the following properties of the period vectors of NE(A).Let us first define S ⃗ δ = {π ∈ Paths A | e∈π wgt(e)δ e ≤ min π ′ ∈Paths A e∈π ′ wgt(e)δ e }.Intuitively, S ⃗ δ is the set of paths in Paths A with minimum cost in the profile ⃗ δ.
Then, for all π such that p π > 0 or p ′ π > 0, we have (q e + q ′ e )wgt(e) ≤ e∈π ′ \π (q e + q ′ e + 1)wgt(e).(q e + 1)wgt(e), where the equivalence holds by Lemma 4. This already holds for π ∈ Paths(A) such that p π > 0. Thus any path π such that p ′ π > 0 must satisfy π ∈ S ⃗ p , which is equivalent to the above.So any period vector ⃗ p ′ for the base ⃗ p satisfies π̸ ∈S ⃗ p p ′ π = 0. Conversely, for any such vector ⃗ p ′ , ⃗ p + k ⃗ p ′ is a Nash equilibrium for all k ≥ 0. ◀ Given the pair ⃗ p, ⃗ q, we write another query to guess m, ⃗ p ′ such that If this is satisfiable, then we query again the solver to find the smallest such m, and keep the set L(⃗ q, m ⃗ δ).Otherwise ⃗ q is added to B.

Price of Anarchy and Stability
Given a semilinear representation B ∪ i∈I L( ⃗ b i , ⃗ δ) of LocOpt(A), observe that there is only a finite number of vectors with a given n number of players.So in order to compute the global social optimum with n players, we iterate over all vectors with n players in this representation, and keep the minimal social cost.We first iterate over { ⃗ b ∈ B | ∥ ⃗ b∥ = n} and consider the vector with the least social cost among those (this set can be empty).Second, for each linear is an integer, we compute the vector ⃗ b i + ∥⃗ q∥−∥ ⃗ b i ∥ ∥ ⃗ δ∥ ⃗ δ, compute its social cost, and keep it if it is less than the previous value.
We compute the social costs of the best and the worst Nash equilibria similarly on the semilinear representation of flow(NE A ).
Figure 5 shows the plots of the PoA and PoS computed for four examples.We used the Python sympy package for solving linear equations, and the Z3 SMT solver for integer arithmetic queries.The characteristic vector was always easy to compute since it is computable in polynomial time in the size of the linear equation system.However, the number of base vectors can be large and depends on the weights used in the network.Our prototype is currently limited in scalability but it allows us to explore small yet non-trivial networks.

More on Related Works
Inefficiency in congestion networks were mentioned in [30], and equilibria were first mathematically studied in [36].Network congestion games are mainly studied in two settings which have different mathematical properties: the nonatomic case, where one considers a large number of players, each of which contributes an infinitesimal amount to congestion; and the atomic case, as we do, where there is a discrete number of players involved.Existence and properties of Nash equilibria in the nonatomic case were established in [4].The price of anarchy of the nonatomic case was studied in [32] which gives a tight bound of 4 3 for networks with affine cost functions.[28,Chapter 18] presents a survey of these results.It is shown in [17] that Nash equilibria in atomic network congestion games can be found in polynomial time, by reduction to maximum flow in the symmetric case, that is, when all players share the same source and target vertices.The problem in the non-symmetric case is however complete for the class PLS, Polynomially Local Search, and is believed to be intractable [26,29].In extension-parallel networks, best-response procedures are shown to converge in linear time in [19]; but this does not extend to general series-parallel graphs.
The complexity of finding extremal Nash equilibria, that is, the best and the worst ones is however higher.Finding such equilibria is NP-hard for the makespan objective with varying sizes [20].For the makespan objective and unit sizes, finding a Nash equilibrium minimizing the makespan in series-parallel networks with linear cost functions is strongly NP-hard when the number of players is part of the input, while a polynomial-time greedy algorithm allows one to find a worst Nash equilibrium [21].For the total cost objective (as in this paper), NP-hardness holds for both best and worst equilibria for three and two players respectively [33].Note that in our work, we are interested in computing such equilibria (or their costs) for arbitrarily large numbers of players.In the more general case of network congestion games with dynamic strategies which allow players to choose each move according to the current state of the game, doubly exponential-time algorithms were given for computing such equilibria in [5] when the number of players is encoded in binary.
It is possible to efficiently compute the social optimum in atomic congestion games by transforming the cost function, and reducing the problem to the computation of Nash equilibria [15].In our case, this transformation would yield affine cost functions.This direction could be exploited to compute the costs of socially optimal profiles using semilinear representations for Nash equilibria, if these could be extended to affine costs.The behaviors of PoA for large numbers of players have been studied before.[18] considers congestion games with large numbers of players, and shows that the PoA of atomic congestion games converges to the PoA of the nonatomic game; the result holds for games with affine cost functions and positive coefficients (as in our case).A consequence is that, although the PoA for the atomic case is often larger than that in the nonatomic case, this difference vanishes in the limit, and thus the upper bound of 4  3 holds for the atomic case in the limit.In [13], the limit of atomic congestion games is considered in a setting where either players participate in the game with given probabilities that tend to 0, or they have weights that tend to 0; in both cases, the limit of the PoA for mixed equilibria is equal to that in a corresponding nonatomic game.Asymptotic PoA bounds (of ≈ 1.35188) are provided in [14] for symmetric atomic congestion games with affine cost functions, by restricting to specific strategies called k-uniform.In the nonatomic case, the works [11,10] establish that the limit of the PoA is 1.Paper [12] considers nonatomic congestion games as a function of the demand and studies continuous derivability properties of the PoA function.

Conclusion
Our results provide theoretical tools that allow us to have a better understanding of the structure of Nash equilibria and social optima in atomic congestion games over series-parallel networks.An immediate question is how the semilinear representations would change if we allow affine cost functions rather than linear ones.However, extending further our approach to nonlinear cost functions would not be immediate since these sets would no longer be definable in Presburger arithmetic.
Although the characteristic vector is easy to compute, the exact computation of the whole semilinear representations is costly and currently does not scale to large networks.One might investigate how this computation can be rendered more efficient in practice.Another possible direction is to explore more efficient approximation algorithms using just the characteristic vector, and perhaps a subset of the base vectors.▶ Lemma 1.In a network congestion game ⟨A, n⟩, a flow ⃗ q is locally-optimal if, and only if, for all π, π ′ ∈ Paths A such that ∀e ∈ π, q e > 0, Proof.Observe that flow ⃗ q is locally optimal iff for all paths π, π ′ such that ∀e ∈ π, q e > 0, the vector ⃗ q ′ defined by satisfies soccost(⃗ q) ≤ soccost( ⃗ q ′ ).Given such paths π, π ′ , let us thus write this inequality as wgt(e) • (q e + 1), (6) where ⃗ q = flow(⃗ p).
Proof.The lemma is a reformulation of (2): for π, π ′ ∈ Paths with p π > 0, cost π (⃗ p) ≤ cost π ′ ( ⃗ p ′ ) can be written as e∈π wgt(e) • q e ≤ e∈π ′ wgt(e) • q ′ e , where ⃗ q and ⃗ q ′ are the respective flows of ⃗ p and ⃗ p ′ .This is equivalent to Here we use the fact that Paths is finite so that the above is a well-defined formula.Now, existentially quantifying {q e } e∈E in ϕ yields a formula describing NE(A).Existentially quantifying {p π } π∈Paths(A) in ϕ yields flow(NE(A)).◀ Our next results prove that the "projections" of the Nash equilibria of series-parallel networks onto their constituent subnets still are Nash equilibria.We first formally define those projections.
For a network A = A 1 ; A 2 and ⃗ p ∈ S(A), for i ∈ {1, 2}, let us define prj Ai (⃗ p) ∈ S(A i ) where for each π i ∈ Paths Ai , prj Ai (⃗ p) πi = π3−i∈Paths A 3−i p π1π2 .Thus, prj Ai (⃗ p) πi is the number of players that cross the path π i in the profile ⃗ p.For a network A = A 1 ∥ A 2 , and a vector ⃗ p ∈ S(A), for i ∈ {1, 2}, let us denote ⃗ p| Ai ∈ S(A i ) obtained by restricting ⃗ p to Paths Ai .Similarly, for a vector ⃗ q ∈ F(A), let ⃗ q| Ai the restriction of ⃗ q to the edges of A i .

4 (
a) A network with two parallel links where each player can choose either the top edge (t) with cost x → 4x, or the bottom edge (b) with cost x → x.

Figure 1
Figure 1 Analysis of a simple network congestion game.

Figure 2
Figure 2 Prices of anarchy and stability as a function of n.

Figure 3
Figure3A network with an edge v → v ′ that is never used in any locally-optimal profile.

− 1 A 1 A
⃗ b for some κ.Since M A and ⃗ b have integer coefficients, M −⃗ b is a vector with rational coefficients.

1 Figure 4
Figure 4 Network whose set of local optima admit several period vectors.

▶ Corollary 13 .
e∈π2\π1 wgt(e)kq e ≤ e∈π1\π2 wgt(e)(kq e + 1).Since π 1 and π 2 are disjoint, we get e∈π2 wgt(e)kq e ≤ e∈π1 wgt(e)(kq e + 1), hence k(κ 2 − κ 1 ) ≤ e∈π1 wgt(e), which is a contradiction for large k.This concludes the proof.◀ It follows from Lemma 5 that all period vectors of flow(NE(A)) are multiples of ⃗ δ, the characteristic vector of A. For a series-parallel network A, the set flow(NE(A)) admits a semilinear representation in the form B ∪ i∈I L( ⃗ b i , m i ⃗ δ), for a finite set B, and finitely-many base vectors ⃗ b i and natural numbers m i .

Figure 5
Figure 5 Plots for PoA and PoS on four series-parallel networks.Unlabeled edges have weight 1.