Prize-Collecting Steiner Network Problems For Peer Review

Transcription

1 Transactions on Algorithms Prize-Collecting Steiner Network Problems Journal: Transactions on Algorithms Manuscript ID: Draft Manuscript Type: Original Article Date Submitted by the Author: n/a Complete List of Authors: Hajiaghayi, MohammadTaghi; University of Maryland, CS Khandekar, Rohit; IBM T.J. Watson research center Kortsarz, Guy; Rutgers University, Camden Nutov, Zeev; The open university of Israel Computing Classification Systems: F. ANALYSIS OF ALGORITHMS AND PROBLEM COMPLEXITY, F.. Nonnumerical Algorithms and Problems, G.. Graph Theory

2 Page of Transactions on Algorithms 0 0 Prize-Collecting Steiner Network Problems MohammadTaghi Hajiaghayi, Rohit Khandekar, Guy Kortsarz, and Zeev Nutov AT&T Research Lab Research & University of Maryland. hajiagha@cs.umd.edu IBM T.J. Watson Research Center. rohitk@us.ibm.com Rutgers University, Camden. guyk@camden.rutgers.edu The Open University of Israel. nutov@openu.ac.il Abstract. In the Steiner Network problem we are given a graph G with edge-costs and connectivity requirements r uv between node pairs u, v. The goal is to find a minimum-cost subgraph H of G that contains r uv edge-disjoint paths for all u, v V. In Prize-Collecting Steiner Network problems we do not need to satisfy all requirements, but are given a penalty function for violating the connectivity requirements, and the goal is to find a subgraph H that minimizes the cost plus the penalty. The case when r uv {0, } is the classic Prize-Collecting Steiner Forest problem. In this paper we present a novel linear programming relaxation for the Prize-Collecting Steiner Network problem, and by rounding it, obtain the first constant-factor approximation algorithm for submodular and monotone non-decreasing penalty functions. In particular, our setting includes all-or-nothing penalty functions, which charge the penalty even if the connectivity requirement is slightly violated; this resolves an open question posed in [NSW0]. We further generalize our results for elementconnectivity and node-connectivity. Introduction Prize-collecting Steiner problems are well-known network design problems with several applications in expanding telecommunications networks (see for example [JMP00,SCRS00]), cost sharing, and Lagrangian relaxation techniques (see, e.g., [JV0,CRW0]). A general form of these problems is the Prize-Collecting Steiner Forest problem : given a network (graph) G = (V, E), a set of sourcesink pairs P = {{s, t }, {s, t },..., {s k, t k }}, a non-negative cost function c : E R +, and a non-negative penalty function π : P R +, our goal is a minimum-cost way of installing (buying) a set of links (edges) and paying the penalty for those pairs which are not connected via the installed links. Part of this work was done while the authors were meeting at DIMACS. We would like to thank DIMACS for hospitality. A version of this paper appeared in the proceedings of the th conference on Integer Programming and Combinatorial Optimization, pages, 00. Partially supported by NSF Award Grant number 0. In the literature, this problem is also called prize-collecting generalized Steiner tree.

3 Transactions on Algorithms Page of 0 0 When all penalties are, the problem is the classic APX-hard Steiner Forest problem, for which the best known approximation ratio is n (n is the number of nodes of the graph) due to Agrawal, Klein, and Ravi [AKR] (see also [GW] for a more general result and a simpler analysis). The case of Prize-Collecting Steiner Forest problem when all sinks are identical is the classic Prize-Collecting Steiner Tree problem. Bienstock, Goemans, Simchi-Levi, and Williamson [BGSLW] were the first to consider this problem (based on a problem earlier proposed by Balas [Bal]) and gave a -approximation algorithm for it. The current best ratio for this problem is. by Archer, ( Bateni, ) Hajiaghayi, and Karloff [ABHK0], improving upon a primal-dual n -approximation algorithm of Goemans and Williamson [GW]. When in addition all penalties are, the problem is the classic Steiner Tree problem, which is known to be APX-hard [BP] and for which the best approximation ratio is. [RZ0]. Very recently, Byrka et al. [BGRS0] have announced an improved approximation algorithm for the Steiner tree problem. The general form of the Prize-Collecting Steiner Forest problem has been first formulated by Hajiaghayi and Jain [HJ0]. They showed how by a primal-dual algorithm to a novel integer programming formulation of the problem with doublyexponential variables, we can obtain a -approximation algorithm for the problem. In addition, they show that the factor in the analysis of their algorithm is tight. However they show how a direct randomized LP-rounding algorithm with approximation factor. can be obtained for this problem. Their approach has been generalized by Sharma, Swamy, and Williamson [SSW0] for network design problems where violating arbitrary 0- connectivity constraints are allowed in exchange for a general penalty function. The work of Hajiaghayi and Jain has also motivated a game-theoretic version of the problem considered by Gupta et al. [GKL + 0]. In this paper, we consider a more general, high-connectivity version of Prize- Collecting Steiner Forest, called Prize-Collecting Steiner Network, in which we are also given connectivity requirements r uv for pairs of nodes u and v and a penalty function in case we do not satisfy all r uv. Our goal is to find a minimum way of constructing a network (graph) in which we connect u and v with r uv r uv edge-disjoint paths and paying a penalty for all violated connectivity between source-sink pairs. This problem can arise in real-world network design, in which a typical client not only might want to connect to the network but also might want to connect via a few disjoint paths (e.g., to have a higher bandwidth or redundant connections in case of edge failures) and a penalty might be charged if we cannot satisfy its connectivity requirement. When all penalties are, the problem is the classic Steiner Network problem. Improving on a long line of earlier research that applied primal-dual methods, Jain [Jai0] obtained a -approximation algorithm for Steiner Network using the iterative rounding method. This algorithm was generalized to so-called element-connectivity problem by Fleischer, Jain, and Williamson [FJW0] and by Cheriyan, Vempala, and Vetta [CVV0]. Recently, some results were obtained for the node-connectivity version; the currently best known ratios for the node-connectivity case are O(R log n) for gen-

4 Page of Transactions on Algorithms 0 0 eral requirements [CK0] and O(R log R) for rooted requirements [Nut0], where R = max u,v V r uv is the maximum requirement. See also the survey by Kortsarz and Nutov [KN0] for various min-cost connectivity problems. Hajiaghayi and Nasri [HN0] generalize the iterative rounding approach of Jain to Prize-Collecting Steiner Network when there is a separate non-increasing marginal penalty function for each pair u, v whose r uv -connectivity requirement is not satisfied. They obtain an iterative rounding -approximation algorithm for this case. For the special case when penalty functions are linear in the violation of the connectivity requirements, Nagarajan, Sharma, and Williamson [NSW0] using Jain s iterative rounding algorithm as a black box give a.-factor approximation algorithm. They also generalize the 0- requirements of Prize-Collecting Steiner Forest problem introduced by Sharma, Swamy, and Williamson [SSW0] to include general connectivity requirements Assuming the monotone submodular penalty function of Sharma et al. is generalized to a function on multi-sets that can be decomposed into functions of the same type as that of Sharma et al., they and present an O(log R)-approximation algorithm (recall that R is the maximum connectivity requirement). In this algorithm, they assume that we can use each edge possibly many times (without bound). They raise the question whether one can obtain a constant approximation ratio without all these assumptions, when penalty is a submodular multi-set function of the set of disconnected pairs? More importantly they pose as an open problem designing a good approximation algorithm for the all-or-nothing version of penalty functions: i.e., penalty functions which charge the penalty even if the connectivity requirement is slightly violated. In this paper, we answer affirmatively all these open problems by presenting the first constant-factor.-approximation algorithm which is based on a novel LP formulation of the problem. We further generalize our results for element-connectivity and node-connectivity. In fact, for both types of connectivities, we prove a very general result (see Theorem ) stating that if Steiner Network (the version without penalties) admits an LP-based ρ- approximation algorithm, then the corresponding prize-collecting version admits a (ρ + )-approximation algorithm.. Problems we consider In this section, we formally define the terms used in the paper. For a subset S of nodes in a graph H, let λ S H (u, v) denote the S-connectivity between u and v in H, namely, the maximum number of edge-disjoint uv-paths in H so that no two of them have a node in S {u, v} in common. In the Generalized Steiner-Network (GSN) problem we are given a graph G = (V, E) with edgecosts c e 0, a node subset S V, a collection {u, v },..., {u k, v k } of node pairs from V, and S-connectivity requirements r,..., r k. The goal is to find a minimum cost subgraph H of G so that λ S H (u i, v i ) r i for all i. Extensively studied particular cases of GSN are the Steiner Network problem, also called Edge-Connectivity GSN (S = ), Node-Connectivity GSN (S = V ), and Element- Connectivity GSN (S {u i, v i } = for all i). The case of rooted requirements

5 Transactions on Algorithms Page of 0 0 is when there is a root s that belongs to all pairs {u i, v i }. We consider the following prize-collecting version of GSN. All-or-Nothing Prize Collecting Generalized Steiner Network (PC-GSN): Instance: A graph G = (V, E) with edge-costs c e 0, S V, a collection {u, v },..., {u k, v k } of node pairs from V, S-connectivity requirements r,..., r k > 0, and a penalty function π : {,...,k} R +. Objective: Find a subgraph H of G that minimizes the value val(h) = c(h) + π(unsat(h)) of H, where unsat(h) = {i λ S H (u i, v i ) < r i } is the set of requirements not (completely) satisfied by H. We will assume that the penalty function π is given by an value oracle. We will also assume that π is submodular, namely, that π satisfies the decreasing marginal value property, π(a) + π(b) π(a B) + π(a B) for all A, B and that it is monotone non-decreasing, namely, π(a) π(b) for all A, B with A B. As was mentioned, approximating the edge-connectivity variant of PC-GSN was posed as the main open problem by Nagarajan, Sharma, and Williamson [NSW0]. We next define the second problem we consider. We will in fact reduce this problem to the first one. Generalized Steiner Network with Generalized Penalties (GSN-GP): Instance: A graph G = (V, E) with edge-costs c e 0, S V, a collection {u, v },..., {u k, v k } of node pairs from V, and non-increasing penalty functions p,..., p k : {0,,..., n } R +. Objective: Find a subgraph H of G that minimizes the value val (H) = c(h) + k p i (λ S H(u i, v i )). The above problem captures general penalty functions of the S-connectivity λ S (u i, v i ) for given pairs {u i, v i }. It is natural to assume that the penalty functions are non-increasing, i.e., we pay less in the objective function if the achieved connectivity is more. This problem was posed as an open question by Nagarajan et al. [NSW0]. In this paper, we use the convention that p i (n) = 0 for all i. We need some definitions to introduce our results. A pair T = {T, T } of subsets of V is called a setpair (of V ) if T T =. Let K = {,..., k}. Let T = {T, T } be a setpair of V. We denote by δ(t ) the set of edges in E between T and T. For i K we use T (i, S) to denote that T {u i, v i } =, T {u i, v i } = and V \ (T T ) S. While in the case of edge-connectivity i=

6 Page of Transactions on Algorithms 0 0 a cut consists of edges only, in the case of S-connectivity a cut that separates between u and v is mixed, meaning it may contain both edges in the graph and nodes from S. Note that if T (i, S), then δ(t ) (V \ (T T )) is such a mixed cut that separates between u i and v i. Intuitively, Menger s Theorem for S-connectivity (see [KN0]) states that the S-connectivity between u i and v i equals the minimum size of such a mixed cut. Formally, for a node pair u i, v i of a graph H = (V, E) and S V, we have λ S H(u i, v i ) = min ( δ(t ) + V \(T T ) ) = min ( δ(t ) + V ( T + T )). T (i,s) T (i,s) Hence if λ S H (u i, v i ) r i for a graph H = (V, E), then for any setpair T with T (i, S) we must have δ(t ) r i (T ), where r i (T ) = max{r i + T + T V, 0}. Consequently, a standard cut-type LP-relaxation of the GSN problem is as follows (see [KN0]): min c e x e x e r i (T ) T (i, S), i K; x e [0, ] e. () e E. Our results e δ(t ) We introduce a novel LP relaxation of the problem which is shown to be better, in terms of the integrality gap, than a natural LP relaxation considered in [NSW0]. Using our LP relaxation, we prove the following main result. Theorem. Suppose that for an instance defined by any subset K K of node pairs, there exists a polynomial-time algorithm that computes an integral solution to LP () of cost at most ρ times the fractional optimal value of LP () for K. Then PC-GSN, defined on all pairs K, admits a ( e /ρ ) -approximation algorithm, provided that the penalty function π is submodular and monotone non-decreasing. Note that since ρ < e ρ < ρ+ holds for ρ, we have ρ < ( e /ρ ) < ρ +. Let R = max i r i denote the maximum requirement. The best known values of ρ are as follows: for Edge-GSN [Jai0], for Element-GSN [FJW0,CVV0], O(R log V ) for Node-GSN [CK0], and O(R log R) for Node-GSN with rooted requirements [Nut0]. Substituting these values in Theorem, we obtain: Corollary. PC-GSN problems admit the following approximation ratios provided that the penalty function π is submodular and monotone non-decreasing:. for edge- and element-connectivity, O(R log V ) for node-connectivity, and O(R log R) for node-connectivity with rooted requirements. Our results for GSN-GP follow from Corollary.

7 Transactions on Algorithms Page of 0 0 Corollary. GSN-GP problems admit the following approximation ratios:. for edge- and element-connectivity, O(R log V ) for node-connectivity, and O(R log R) for node-connectivity with rooted requirements. Here R = max i k min{λ 0 p i (λ) = 0}. Proof. We present an approximation ratio preserving reduction from the GSN- GP problem to the corresponding PC-GSN problem. Given an instance of the GSN-GP problem, we create an instance of the PC-GSN problem as follows. The PC-GSN instance inherits the graph G, its edge-costs, and the set S. Let (u i, v i ) be a pair in GSN-GP and let R i = min{λ 0 p i (λ) = 0}. We introduce R i copies of this pair, {(u i, v i ),..., (uri i, v Ri i )}, to the set of pairs in the PC-GSN instance. We set the edge-connectivity requirement of a pair (u t i, vt i ) to be t for t R i. We also set the penalty function for singleton sets as follows π({(u t i, vt i )}) = p i(t ) p i (t) for all t R i. Finally, we extend this function π to a set of pairs P by linearity, i.e., π(p ) = p P π({p}). Note that such a function π is clearly submodular and monotone non-decreasing. It is sufficient to show that for any subgraph H of G, its value in the GSN-GP instance equals its value in the PC-GSN instance, i.e., val(h) = val (H); then we can use the algorithm from Corollary to complete the proof. Fix a pair (u i, v i ) in the GSN-GP instance. Let λ S H (u i, v i ) = t i. Thus the contribution of pair (u i, v i ) to the objective function val(h) of the GSN-GP instance is p i (t i ). On the other hand, since π is linear, the total contribution of pairs {(u i, v i ),..., (uri i, v Ri i )} to the objective function val (H) of the PC-GSN instance is R i t=t i+ π({(ut i, vt i )}) = Ri t=t (p i+ i(t ) p i (t)) = p i (t i ). Note that the pairs (u t i, vt i ) for t t i do not incur any penalty. Summing up over all pairs, we conclude that val(h) = val (H), as claimed. A new LP relaxation We use the following LP-relaxation for the PC-GSN problem. We introduce variables x e [0, ] for e E, f i,e [0, ] for i K and e E, and z I [0, ] for I K. In the intended integral solution H, these variables are supposed to take the following values: x e = if e H, f i,e = if i unsat(h) and e appears on a chosen set of r i S-disjoint {u i, v i }-paths in H and z I = if I = unsat(h). Minimize Subject to e E c ex e + I K π(i)z I e δ(t ) f i,e ( I:i I z I)r i (T ) i T (i, S) f i,e I:i I z I i e x e f i,e i e I K z I = x e, f i,e, z I [0, ] i e I We first prove that () is a valid LP-relaxation of the PC-GSN problem. ()

8 Page of Transactions on Algorithms 0 0 Lemma. The optimal value of LP () is at most the value of any optimal solution to the PC-GSN problem. Moreover, if π is monotone non-decreasing, the optimum solution value to the PC-GSN problem is at most the value of the optimum integral solution of LP (). Proof. Given a feasible solution H to the PC-GSN problem define a feasible solution to LP () as follows. Let x e = if e H and x e = 0 otherwise. Let z I = if I = unsat(h) and z I = 0 otherwise. For each i unsat(h) set f i,e = 0 for all e E, while for i / unsat(h) the variables f i,e take values as follows: fix a set of r i pairwise S-disjoint {u i, v i }-paths in H, and let f i,e = if e belongs to one of these paths and f i,e = 0 otherwise. The defined solution is feasible for LP (): the first set of constraints are satisfied by Menger s Theorem for S-connectivity, while the remaining constraints are satisfied by the above definition of variables. It is also easy to see that the above solution has value exactly val(h). If π is monotone non-decreasing, we prove that for any integral solution {x e, f i,e, z I } to (), the graph H with edge-set {e E x e = } has val(h) at most the value of the solution {x e, f i,e, z I }. To see this, first note that there is a unique set I K with z I = since the variables z I are integral and I K z I =. Now consider an index i / I. Since I:i I z I = 0, we have e δ(t ) x e e δ(t ) f i,e r i (T ) for all T (i, S). This implies that i / unsat(h), by Menger s Theorem for S-connectivity. Consequently, unsat(h) I, hence π(unsat(h)) π(i) by the monotonicity of π. Thus val(h) = c(h) + π(unsat(h)) e E c ex e + I K π(i)z I and the lemma follows.. Why does a natural LP relaxation not work? One may be tempted to consider a natural LP without using the flow variables f i,e, namely, the LP obtained from LP () by replacing the the first three sets of constraints by the set of constraints x e ( z I )r i (T ) e δ(t ) I:i I for all i and T (i, S). Here is an example demonstrating that the integrality gap of this LP can be as large as R = max i r i even for edge-connectivity. Let G consist of R edge-disjoint paths between two nodes s and t. All the edges have cost 0. There is only one pair {u, v } = {s, t} that has requirement r = R and penalty π({}) =. Let π( ) = 0. Clearly, π is submodular and monotone non-decreasing. We have S =. No integral solution can satisfy the requirement r, hence an optimal integral solution pays the penalty π({}) and has value. A feasible fractional solution (without the flow variables) sets x e = for all e, and sets z {} = /R, z = /R. The new set of constraints is satisfied since e δ(t ) x e ( /R) R = ( z {} )r (T ) for any {s, t}-cut T. Thus the optimal LP-value is at most /R, giving a gap of at least R. With flow variables, however, we have an upper bound f,e z {}. Since there is an {s, t}-cut T with δ(t ) = R, we cannot satisfy the constraints

9 Transactions on Algorithms Page of 0 0 e δ(t ) f,e ( z {} )r (T ) and f,e z {} simultaneously unless we set z {} =. Thus in this case, our LP () with flow variables has the same value as that of the integral optimum.. Some technical results regarding LP () We will prove the following two statements that together imply Theorem. Lemma. Any basic feasible solution to () has a polynomial number of nonzero variables. Furthermore, an optimal basic solution to () (the non-zero entries) can be computed in polynomial time. Lemma. Suppose that for an instance defined by any subset K K of node pairs, there exists a polynomial-time algorithm that computes an integral solution to LP () of cost at most ρ times the optimal value of LP () for K. Assume also that π is submodular and monotone non-decreasing. Then, there exists a polynomial-time algorithm that given a feasible solution F = {x, f, z } to (), computes as a solution to PC-GSN, a subgraph H of G so that val(h) = c(h) + π(unsat(h)) is at most ( e /ρ ) times the value of F. Before proving these lemmas, we prove some useful results. The following statement can be deduced from a theorem of Edmonds for polymatroids (see [KV0, Chapter.]), as the dual LP d(γ) in the lemma seeks to optimize a linear function over a polymatroid. We provide a direct proof for completeness of exposition. Lemma. Let γ [0, ] k be a vector. Let π be a monotone non-decreasing and submodular function such that π( ) = 0. Consider a primal LP p(γ) := min π(i)z I z I γ i i K, z I 0 I K and its dual LP d(γ) := max I K { i K γ i y i i I I:i I y i π(i) I K, y i 0 i K Let σ be a permutation of K such that γ σ() γ σ()... γ σ(k). Let us also use the notation that γ σ(0) = 0. The optimum solutions to p(γ) and d(γ), respectively, are given by { γσ(i) γ z I = σ(i ), for I = {σ(i),..., σ(k)}, i K, 0, otherwise; and y σ(i) = π({σ(i),..., σ(k)}) π({σ(i + ),..., σ(k)}), for i K. }.

10 Page of Transactions on Algorithms 0 0 Proof. To simplify the notation, we assume without loss of generality that γ γ γ k, i.e., that σ is the identity permutation. We also use the notation γ 0 = 0. We argue that {z I } and {y i } form feasible solutions to the primal and dual LPs, respectively. Note that z I 0 for all I and I:i I z I = i j= (γ j γ j ) = γ i for all i. Since π is monotone non-decreasing, the above-defined y i satisfy y i 0 for all i. Now fix I K. Let I = {i,..., i p } where i < < i p. Therefore, y i = i I p y ij = j= = = p [π({i j, i j +,..., k}) π({i j +,..., k})] j= p [π({i j } {i j +,..., k}) π({i j +,..., k})] j= p [π({i j } {i j+,..., i p }) π({i j+,..., i p })] j= p [π({i j, i j+,..., i p }) π({i j+,..., i p })] j= = π({i,..., i p }) = π(i). The above inequality holds because of the submodularity of π. Next observe that the solutions {z I } and {y i } satisfy π(i)z I = I = k π({i,..., k}) (γ i γ i ) i= k γ i (π({i,..., k}) π({i +,..., k})) = i= k γ i y i. Thus from weak LP duality, they in fact form optimum solutions to primal and dual LPs respectively. Recall that a vector d R k is a subgradient of a convex function g : R k R at a point γ R k if for any γ R k, we have g(γ ) g(γ) d (γ γ). For a differentiable convex function g, the subgradient corresponds to gradient g. The function p(γ) defined in Lemma is essentially Lovasz s continuous extension of the submodular function π. The fact that p is convex and its subgradient can be computed efficiently is given in [Fuj0]. We provide a full proof for completeness of exposition. Lemma. The function p(γ) in Lemma is convex and given γ [0, ] k, both p(γ) and its subgradient p(γ) can be computed in polynomial time. Proof. We first prove that p is convex. Fix γ, γ [0, ] k and α [0, ]. To show that p is convex, we will show p(αγ + ( α)γ ) αp(γ ) + ( α)p(γ ). i=

11 Transactions on Algorithms Page 0 of 0 0 Let {zi } and {z I } be the optimum solutions of the primal LP defining p for γ and γ, respectively. Note that the solution {αzi + ( α)z I } is feasible for this LP for γ = αγ + ( α)γ. Thus, the optimum solution has value not greater than the value of this solution which is αp(γ ) + ( α)p(γ ). From Lemma, it is clear that given γ [0, ] k, the value p(γ) can be computed in polynomial time. Lemma also implies that the optimum dual solution y = (y,..., yk ) Rk + can be computed in polynomial time. We now argue that y is a subgradient of p at γ. Fix any γ R k. First note that, from LP duality, p(γ) = y γ. Thus we have p(γ) + y (γ γ) = y γ + y (γ γ) = y γ p(γ ). The last inequality holds from weak LP duality since y is a feasible solution for the dual LP d(γ ) as well. The lemma follows. Proof of Lemma We now describe how to round an LP () solution to obtain a (ρ+)-approximation for PC-GSN. Later we show how to improve the guarantee to ( e /ρ ). Let {x e, fi,e, z I } be a feasible solution to LP (). Let α (0, ) be a parameter to be fixed later. We partition the requirements into two classes: we call a requirement i K good if I:i I z I α and bad otherwise. Let K g denote the set of good requirements. The following statement shows how to satisfy the good requirements. Lemma. There exists a polynomial-time algorithm that computes a subgraph H of G of cost c(h) ρ α e c ex e that satisfies all good requirements. Proof. Consider the LP-relaxation () of the GSN problem with good requirements only, with K replaced by K g ; namely, we seek a minimum cost subgraph H of G that satisfies the set K g of good requirements. We claim that = min {, x e/( α)} for each e E is a feasible solution to LP (). Thus the optimum value of LP () is at most e E c ex e. Consequently, using the algorithm that computes an integral solution to LP () of cost at most ρ times the optimal value of LP (), we can construct a subgraph H that satisfies all good requirements and has cost at most c(h) ρ e E c ex e ρ α e c ex e, as desired. x e We now show that {x e } is a feasible solution to LP (), namely, that e δ(t ) x e r i (T ) for any i K g and any T (i, S). Let i K g and let ζ i = I:i I z I. Note that ζ i α, by the definition of K g. By the second and the third sets of constraints in LP (), for every e E we have min{ζ i, x e} fi,e. Thus we obtain: x e f i,e ζ i = f i,e P I:i I z I = min {, x e α } = ζ i min { ζ i, ζ i α x e } ζ i min{ζ i, x e}. Consequently, combining with the first set of constraints in LP (), for any T (i, S) we obtain that e δ(t ) x e P e δ(t ) f i,e P I:i I z I r i (T ).

12 Page of Transactions on Algorithms 0 0 Let H be as in Lemma, and recall that unsat(h) denotes the set of requirements not satisfied by H. Clearly each requirement i unsat(h) is bad. The following lemma bounds the total penalty we pay for unsat(h). Lemma. π(unsat(h)) α I π(i)z I. Proof. Define γ [0, ] k as follows: γ i = if i unsat(h) and 0 otherwise. Now consider LP p(γ) defined in Lemma. Since each i unsat(h) is bad, from the definition of bad requirements, it is clear that {zi /α} is a feasible solution to LP p(γ). Furthermore, from Lemma, the solution {z I } defined as z I = if I = unsat(h) and 0 otherwise is the optimum solution to p(γ). The cost of this solution, π(unsat(h)), is therefore at most the cost of the feasible solution {zi /α} which is α I π(i)z I. The lemma thus follows. Combining Lemmas and, we obtain max{ ρ α, α }-approximation. If we substitute α = /(ρ + ), we obtain a (ρ + )-approximation for PC-GSN. Improving the approximation to ( e /ρ ). We pick α uniformly at random from the interval (0, β] where β = e /ρ. From Lemmas and, the expected cost of the solution is at most [ ] ρ E α [c(h) + π(unsat(h))] = E α c e x e + E α [π(unsat(h))]. () α To complete the proof of β -approximation, we now argue that the above expectation is at most β e E (c ex e + I π(i)z I ). Note that E α [ ρ α ] = β β 0 e E ρ α dα = ρ β ( ln( β) + ln ) = ρ β ( ρ ) = β, the first term in () is at most β e E c ex e. Since unsat(h) {i I:i I z I α} and since π is monotone non-decreasing, the second term in () is at most E α [ π ( {i I:i I z I α})]. Lemma bounds this quantity as follows. The ideas used here are also presented in Sharma et al. [SSW0]. Lemma. We have ( E α [π {i )] zi α} β π(i)zi. () I:i I Proof. Let γ i = I:i I z I for all i K. Let us, without loss of generality, order the elements i K such that γ γ γ k. We also use the notation γ 0 = 0. Note that {zi } forms a feasible solution to the primal LP p(γ) given in Lemma. Therefore, from Lemma, its objective value is at least that of the optimum solution: π(i)zi I I k [(γ i γ i ) π({i,..., k})]. () i=

13 Transactions on Algorithms Page of 0 0 We now observe that the LHS of () can be expressed as follows. Since α is picked uniformly at random from (0, β], we have that for all i k, with probability at most γi γi β, the random variable α lies in the interval (γ i, γ i ]. When this event happens, we get that {i I:i I z I α} = {i γ i α} = {i,..., k}. Thus the expectation in LHS of () is at most k [ γi γ i i= β ] π({i,..., k}). () From expressions () and (), the lemma follows. Thus the proof of ( e /ρ ) -approximation is complete. It is worth mentioning that in this section we obtain a solution with a bound on its expected cost. However, the choice of α can be simply derandomized by trying out all the breakpoints where a good demand pair becomes a bad one (plus 0 and β). Proof of Lemma We next show that although LP () has exponential number of variables and constraints, the following lemma holds. Lemma. Any basic feasible solution to LP () has a polynomial number of non-zero variables. Proof. Fix a basic feasible solution {x e, fi,e, z i } to (). For i K, let γ i = min T :T i e δ(t ) f i,e and γ i = max fi,e. r i e Now fix the values of variables {x e, f i,e } to {x e, fi,e } and project the LP () onto variables {z I } as follows. c e x e + min π(i)z I e E I K z I =, γ i z I γ i i K, z I 0 I K. () I K I:i I Since {x e, fi,e, z i } is a basic feasible solution to (), it cannot be written as a convex combination of two distinct feasible solutions to (). Thus, we get that {zi } cannot be written as a convex combination of two distinct feasible solutions to (), and hence it forms a basic feasible solution to (). Since there are + K non-trivial constraints in (), at most + K variables z I can be non-zero in any basic feasible solution of (). The lemma follows.

14 Page of Transactions on Algorithms 0 0 We prove that LP () can be solved in polynomial time. Introduce variables γ [0, ] k and obtain the following program (the function p is as in Lemma ). Minimize Subject to e E c ex e + p(γ) e δ(t ) f i,e ( γ i )r i (T ) i T (i, S) f i,e γ i i e x e f i,e i e x e, f i,e, γ i [0, ] i e It is clear that solving () is enough to solve (). Now note that this is a convex program since p is a convex function. To solve (), we convert its objective function into a constraint e E c ex e + p(γ) opt where opt is the target objective value and thus reduce it to a feasibility problem using standard techniques. Now to find a feasible solution using an ellipsoid algorithm, we need to show a polynomial-time separation oracle. The separation oracle for the constraints that are not in the first set of constraints is trivial. The separation oracle for the first set of constraints can be reduced to the problem of finding a minimum capacity mixed cut (containing both nodes and edges) that separates between two given nodes, in an undirected network with both edge and node capacities. The latter problem can be solved in polynomial time using standard network-flow algorithms. The reduction is as follows. Consider the constraints for a specific i K. Denoting d e = f i,e /( γ i ), the constraints set for i can be rewriten as d e r i (T ) T (i, S). e δ(t ) By the definition of r i (T ), this is equivalent to d e + V \ (T T ) r i T (i, S). () e δ(t ) View the input graph G as a network with source u i, sink v i, edge capacities {d e : e E}, and node capacities d v = if v S\{u i, v i } and d v = otherwise. Then the following statement finishes the reduction. Lemma 0. There exists a setpair T = (T, T ) with T (i, s) for which () is violated if, and only if, in the network above there exists a mixed cut C of capacity strictly less than r i that separates between u i and v i. Proof. Let T = (T, T ) be a setpair with T (i, s) for which () is violated. Let C = δ(t ) (V \ (T T )). Then the capacity of C equals the left-hand side of (), and thus is strictly less than r i, since () is violated for T. Now suppose that a mixed cut C as in the lemma exists. Let T C = (T, T ) be a setpair defined as follows: T is the connected component of G\C containing u i and T = V \(C T ). Then the left-hand side of () is at most the capacity of C. Consequently, the constraint () of the setpair T C is violated, and T C (i, S) holds, since the capacity of C is less than r i and thus is finite. ()

15 Transactions on Algorithms Page of 0 0 The separation oracle for the objective function is as follows. Given a point (x, γ) = {x e, γ i } that satisfies e E c ex e + p(γ) > opt, we compute a subgradient of the function e E c ex e + p(γ) w.r.t. variables {x e, γ i }. The subgradient of e E c ex e w.r.t. x is simply the cost vector c. The subgradient of p(γ) w.r.t. γ is computed using Lemma ; denote it by y R k +. From the definition of subgradient, we have that the subgradient (c, y) to the objective function at point (x, γ) satisfies ( ) ( ) c e x e + p(γ ) c e x e + p(γ) (c, y) ((x, γ ) (x, γ)). e E e E Now fix any feasible solution (x, γ ), i.e., one that satisfies e E c ex e +p(γ ) opt. Substituting (x, γ ) = (x, γ ) in the above equation we get, ( ) ( ) 0 = opt opt > c e x e + p(γ ) c e x e + p(γ) e E e E (c, y) (x, γ ) (c, y) (x, γ). Thus (c, y) defines a separating hyperplane between the point (x, γ) and any point (x, γ ) that satisfies e E c ex e+p(γ ) opt. Hence we have a polynomialtime separation oracle for the objective function as well. Thus we can solve () using the ellipsoid algorithm. The proof of Lemma is hence complete. References [ABHK0] A. Archer, M. Bateni, M. Hajiaghayi, and H. Karloff. Improved approximation algorithms for prize-collecting Steiner tree and TSP. In Proc. of the th annual IEEE Symp. on Foundations of Computer Science, pages, 00. [AKR] A. Agrawal, P. Klein, and R. Ravi. When trees collide: an approximation algorithm for the generalized Steiner problem on networks. SIAM J. Comput., ():,. [Bal] E. Balas. The prize collecting traveling salesman problem. Networks, [BGRS0] ():,. J. Byrka, F. Grandoni, T. Rothvoss, and L. Sanita. An improved LP-based approximation for Steiner tree. In Proc. of the nd annual ACM Symp. on Theory of computing, pages, 00. [BGSLW] D. Bienstock, M. Goemans, D. Simchi-Levi, and D. Williamson. A note on the prize collecting traveling salesman problem. Math. Programming, (, Ser. A): 0,. [BP] [CK0] M. Bern and P. Plassmann. The Steiner problem with edge lengths and. Information Processing Letters, :,. J. Chuzhoy and S. Khanna. An O(k log n)-approximation algorithms for vertex-connectivity network design. In Proc. of the th annual IEEE Symp. on Foundations of Computer Science, pages, 00.

16 Page of Transactions on Algorithms 0 0 [CRW0] [CVV0] F. Chudak, T. Roughgarden, and D. Williamson. Approximate k-msts and k-steiner trees via the primal-dual method and Lagrangean relaxation. In Proc. of the th International Conference on Integer Programming and Combinatorial Optimization, pages 0, 00. J. Cheriyan, S. Vempala, and A. Vetta. Network design via iterative rounding of setpair relaxations. Combinatorica, ():, 00. [FJW0] L. Fleischer, K. Jain, and D. Williamson. An iterative rounding - approximation algorithm for the element connectivity problem. In Proc. of the nd IEEE Symp. on Foundations of Computer Science, pages, 00. [Fuj0] S. Fujishige. Submodular functions and optimization. Elsevier, 00. [GKL + 0] [GW] [HJ0] [HN0] [Jai0] [JMP00] [JV0] [KN0] A. Gupta, J. Könemann, S. Leonardi, R. Ravi, and G. Schäfer. An efficient cost-sharing mechanism for the prize-collecting Steiner forest problem. In Proc. of the th ACM-SIAM Symp. on Discrete algorithms, pages, 00. M. Goemans and D. Williamson. A general approximation technique for constrained forest problems. SIAM J. Comput., ():,. M. Hajiaghayi and K. Jain. The prize-collecting generalized Steiner tree problem via a new approach of primal-dual schema. In Proc. of the th annual ACM-SIAM Symp. on Discrete Algorithms, pages, 00. M. Hajiahayi and A. Nasri. Prize-collecting Steiner networks via iterative rounding. In Proc. of LATIN Theoretical Informatics, pages, 00. K. Jain. A factor approximation algorithm for the generalized Steiner network problem. Combinatorica, ():, 00. D. Johnson, M. Minkoff, and S. Phillips. The prize collecting Steiner tree problem: theory and practice. In Proc. of the th annual ACM-SIAM Symp. on Discrete Algorithms, pages, 000. K. Jain and V. Vazirani. Approximation algorithms for metric facility location and k-median problems using the primal-dual schema and Lagrangian relaxation. J. ACM, ():, 00. G. Kortsarz and Z. Nutov. Approximating minimum cost connectivity problems, Ch. in Approximation Algorithms and Metahueristics, T. F. Gonzales ed.,. CRC Press, Boca Raton, 00. [RZ0] [KV0] B. Korte and J. Vygen. Combinatorial Optimization: Theory and Algorithms. Springer, Berlin, 00. [NSW0] C. Nagarajan, Y. Sharma, and D. Williamson. Approximation algorithms for prize-collecting network design problems with general connectivity requirements. In Proc. of the th International Workshop on Approximation and Online Algorithms, pages, 00. [Nut0] Z. Nutov. Approximating minimum cost connectivity problems via uncrossable bifamilies. Manuscript, 00. G. Robins and A. Zelikovsky. Tighter bounds for graph Steiner tree approximation. SIAM J. on Discrete Mathematics, ():, 00. [SCRS00] F. Salman, J. Cheriyan, R. Ravi, and S. Subramanian. Approximating the single-sink link-installation problem in network design. SIAM J. on Optimization, ():, 000. [SSW0] Y. Sharma, C. Swamy, and D. Williamson. Approximation algorithms for prize collecting forest problems with submodular penalty functions. In Proc. of the th annual ACM-SIAM Symp. on Discrete Algorithms, pages, 00.