Improved approximation algorithms for the facility location problems with linear/submodular penalty

Yu Li∗1,2 , Donglei Du†2 , Naihua Xiu‡1 , and Dachuan Xu§3
1

Department of Mathematics, School of Science, Beijing Jiaotong University, 3 Shangyuancun, Haidian District, Beijing 100044, P.R. China.

2

Faculty of Business Administration, University of New Brunswick, Fredericton, NB Canada E3B 5A3.

3

Department of Applied Mathematics, Beijing University of Technology, 100 Pingleyuan, Chaoyang District, Beijing 100124, P.R. China. February 4, 2012

Abstract We consider the facility location problem with submodular penalty (FLPSP) and the facility location problem with linear penalty (FLPLP), two extensions of the classical facility location problem (FLP). First, we introduce a general algorithmic framework for a class of covering problems with submodular penalty, extending the recent result of Geunes et al. [12] with linear penalty. This framework leverages existing approximation results for the original covering problems to obtain corresponding results for their submodular penalty counterparts. Specifically, any LP-based α-approximation for the original ( )−1 -approximation algorithm for the councovering problem can be leveraged to obtain an 1 − e−1/α terpart with submodular penalty. Consequently, any LP-based approximation algorithm for the classical FLP (as a covering problem) can yield, via this framework, an approximation algorithm for the submodular penalty counterpart. Second, by exploiting some special properties of FLP, we present an LP rounding algorithm which has the currently best 2-approximation ratio over the previously best 2.50 by Hayrapetyan et al. [14] for the FLPSP and another LP rounding algorithm which has the currently best 1.5148-approximation ratio over the previously best 1.853 by Xu and Xu [29] for the FLPLP, respectively.
∗ Email: † Email:

liyubjut@gmail.com. ddu@unb.ca. ‡ Email: nhxiu@center.njtu.edu.cn. § Corresponding author. Email: xudc@bjut.edu.cn.

24. 31. [5] and Hayrapetyan et al. 2.463. 26]. but is less robust. four basic schemes have been emerged in the design of these approximation algorithms. The FLPLP can be formally defined as follows. the currently best approximation ratio of 1. 20. a penalty cost pj . we refer to [1. and inventory management and supply chain management in particular. [14]. 10. [32]). These four competitive and complementary schemes possess different features. The problem is to open a subset of facilities such that each client j ∈ D is either assigned to an open facility or rejected with the objective to minimize the total cost. and local search [4. there was a long list of work on designing improved approximation algorithms for this problem over the years. first introduced.g. unless N P ⊆ DT IM E(nlog log n ). including the open. and usually offers better approximation ratio than a typical primal-dual. 22].488 by Li [19] is obtained by combining LP-rounding and dual-fitting algorithm. Among the latter three. On the impossibility of approximation. the FLPLP does not require that every client must be served by some open facility. 19. This problem is NP-hard and therefore much attention has been focused on designing approximation algorithms with good performance. respectively. However. Instead. we study two facility location problems with penalty. 8. Introduction The classical facility location problem (FLP) is one of the most important models in combinatorial optimization with applications in operations research and computer science in general. 23. 21. We assume that the connection cost between clients and facilities 2 . Local search is more powerful on the hard capacitated version of the FLP (cf. [24]. and a connection cost cij . As a result. For other variants of the FLP. In this work.Keywords: Approximation algorithm · Facility location problem · LP rounding · Submodular function Mathematics Subject Classification (2000) Primary 68W25.. primal-dual [17]. hence offering faster approximation algorithms than LP rounding. connection and linear penalty costs. Following the first constantfactor approximation algorithm by Shmoys et al. 32] and the references therein. while the other three are combinatorial in nature. there is a nonnegative opening cost fi . later strengthened to unless P = N P by Sviridenko ([27]). 30. 18]. Guha and Khuller [13] show that the approximation ratio for FLP is at least 1. 13. namely the facility location problem with linear penalty (FLPLP) and the more general facility location problem with submodular penalty (FLPSP). Unlike the regular FLP. the first scheme usually allows us to design algorithms with better approximation ratios compared to the other three—e. Secondary 90C27 1. For every facility i ∈ F and every client j ∈ D. namely LP-rounding [3. by Charikar et al. 16. LP-rounding is non-combinatorial in nature mainly because of the resolving of the underlying LP-relaxation. as important extensions of the classical FLP. 6. primal-dual can be adapted to solve variants of the classical FLP. dual-fitting [15. Consider a set F of facilities and a set D of clients. Dual-fitting is essentially one special type of primal-dual methods. a client j can be either served by an open facility or rejected for service with penalty cost pj . 25.

Recently. along with the exploitation of the special properties of the submodular penalty function. cij ≤ cij ′ + ci′ j ′ + ci′ j .853 [29]. for the FLPLP. In general. Xu and Xu [28.736-approximation algorithm.e. To overcome 3 .is metric.5148-approximation algorithm which exploits the special properties of the linear penalty function. the best known combinatorial approximation ratio is 3 [9] and the best known non-combinatorial ratio is 2. Du et al. is not directly applicable to the FLPSP. We assume that P (∅) = 0. [14] gave a simple LP-rounding based 2. i. (iii) Note that for the FLPLP. four constant-factor approximation algorithms were known in the literature. 16]. The FLPSP is similar to the FLPLP except that the linear penalty cost is replaced by a general monotonically increasing submodular function. and a combinatorial 1. the existing combinatorial ratio is better than the existing non-combinatorial ratio. For the FLPSP. In summary. The main contributions of this work are summarized as follows. A set function P : 2D → R+ is submodular if P (X ∩ Y ) + P (X ∪ Y ) ≤ P (X) + P (Y ) for any subsets X. we provide an improved 2approximation algorithm for the FLPSP over the previously best approximation ratio 2. [5] gave a primal-dual 3-approximation algorithm. In particular. This phenomena is generally “abnormal” and our third contribution corrects this “abnormality” by offering the currently best LP-rounding 1. (ii) Combining a novel LP-rounding technique with the JMS algorithm of [15. Hayrapetyan et al. Geunes et al.056 [12]. There are intrinsic differences between the two LP-based algorithms for the FLPSP in (ii) and the FLPLP in (iii) due to the essential difference between linear and submodular functions. such as that by Byrka and Aardal [3]. Chudak and Nagano [7] gave a faster (2. Charikar et al. (i) We extend Geunes et al.056-approximation algorithm. [12] gave an improved LProunding based 2.50-approximation algorithm. [12] presented an algorithmic ( )−1 framework which can convert any LP-based α-approximation for the classical FLP to an 1 − e−1/α approximation algorithm for the counterpart with linear penalty.50 [14].50+ϵ)-approximation algorithm by solving a convex relaxation rather than an LP relaxation of the FLPSP. For the FLPLP. the best known approximation ratio is 1. [9] presented a primal-dual 3-approximation algorithm.. Our algorithm and analysis indicate that the latter is substantially harder to approximate than the former because the techniques for the FLPLP. 29] presented an LP-rounding based 2 + e−1 ≈ 2. Very recently.853-approximation algorithm by integrating primal-dual with local search technique. and the best known noncombinatorial ratio is 2. linear penalty functions possess important properties not applicable to submodular penalty functions. three approximation algorithms were proposed in the literature. Y ∈ 2D . [12]’s algorithmic framework for linear penalty to submodular penalty by ( )−1 showing that our framework can leverage any LP-based α-approximation to construct an 1 − e−1/α approximation algorithm for the counterpart with submodular penalty.50 [14]. Geunes et al. For the FLPSP.

zS ≥ 0. respectively. Algorithm 2. S ⊆ D.5148-approximation algorithm for the FLPLP. Solve the LP relaxation (2. we first show that any LP-based α( )−1 approximation algorithm for the classical FLP can be leveraged to an 1 − e−1/α -approximation algorithm for the submodular penalty counterpart. we give an LP-rounding based approximation algorithm similar to that by Geunes et al. (2. we present the 2-approximation algorithm for the FLPSP. and pay the penalty cost P (S). And this framework also extends Geunes et al. yi . Now. where P (S) is a nondecreasing submodular function and P (∅) = 0. y ∗ . δ).1 An LP-rounding approximation algorithm for the FLPSP The following LP relaxation for the FLPSP first appeared in Hayrapetyan et al. min s. j ∈ D. 2. ij STEP 3. ∀i ∈ F. We use the following notations throughout the paper: nf = |F| and nc = |D|. z ∗ ). Select parameter β uniformly at random from the interval [0.1) to obtain an optimal fractional solution (x∗ . xij . This result will serve as a concrete example in deriving a general framework for a class of covering problems with submodular penalty. [14]. and the 1.this difficulty. ∑∑ i∈F j∈D cij xij + ∑ ∑ i∈F fi yi + ∑ S⊆D P (S)zS ∑ xij + zS ≥ 1. STEP 2.1 STEP 1. which then will be extended to a more general class of covering problems. 2 Algorithmic scheme for problems with submodular penalty We will use the FLPSP as an example to show our algorithmic framework.t. [12]. Reject the subset S := {j|zj ≥ β}. ∀i ∈ F. j ∈ D. our algorithm for the latter will exploit the special structure of the optimal LP relaxation solution of the FLPSP. The rest of this paper is organized as follows.1) i∈F S⊆D:j∈S xij ≤ yi . In Section 2. In Sections 3 and 4. 4 . ∀j ∈ D. Construct a new variable z such that zj := 1 − ∑ i∈F x∗ . [12]’s technique for the linear penalty model to the submodular case. STEP 4.

. We recall the S⊆D:j∈S zS .. j = 1. where αj = P ({zj . .. The following claim is immediate. .1 is no more than δ −1 ∑ S⊆D ∗ P (S)zS . j∈S αj ≤ P (S) for any ∑nc S ⊆ D. We estimate the rejecting probability qSi for any subset Si . Proof : ∑ ∑ Note that (x∗ . From these properties. we assume that z1 ≤ z2 · · · ≤ znc .STEP 5. znc }). y ∗ . Property 3. Property 1. ∑ ∑ i∈S j∈D αj zj αj ≤ P (S). From Lemma 3. . Then run the 1. zj+1 . . . . .1 in [11]. P ′ (I(S)) = P (S) and P ′ (0) = P (∅). ∑ ∑ Claim 2. . zj+2 . ∀S ⊆ D.. When Si is rejected. αnc ).. ℓ). P ′ (z) ≤ P ′  S⊆D S⊆D S⊆D (2. where nc = |D|. we can construct an optimal solution α = (α1 . . Lemma 2. P ′ (·) is an monotonically increasing convex function. zj+1 .3 The optimal solution α satisfies j∈Si αj = P (Si ) for i = 1 · · · ℓ. implying that qSi ≤ δ −1 (zki − zki−1 ) since β is from the interval [0.. nc .2 The expected penalty cost of the solution generated by Algorithm 2. Property 2.488-approximation algorithm [19] for the instance and assign the clients in D \ S to the closest open facilities. Let zk0 = 0.. it is easy to show that   ∑ ∑ ∑ ∗ ∗ ∗ zS I(S) ≤ zS P ′ (I(S)) = P (S)zS . . and j ∈ D \ S).3) Without loss of generality.t. for any j ≤ nc . Let Si = {j ∈ D|zj ≥ zki } (i = 1 . δ). 5 . For any number a ≥ 0 and any vector w ≥ 0. znc }) − P ({zj+1 . Note further z ≤ definition of Lov´sz extension function P ′ (·) of any given submodular function P (·) along with its properties a for later usage. ∀j ∈ D. where I(S) is the incidence vector of subset S. Then these sets satisfy that Sℓ ⊂ · · · ⊂ S1 and are all the ones that could be rejected by our algorithm. the set of clients D \ S and the connection cost cij (i ∈ F.. z ∗ ) is an optimal fractional solution. zki ]. Construct an instance of the classical FLP with the set of facilities F. implying that zj = 1 − i∈F x∗ ≤ ij ∑ ∗ ∗ S⊆D zS I(S). the random variable β must fall in the interval (zki−1 . ...2) αj ≥ 0. znc }). and moreover there are exactly ℓ nonzero and distinct numbers zk1 < zk2 · · · < zkℓ in {zj }. (2. P ′ (z) = max s. P ′ (aw) = aP ′ (w). and j αj = P ({zj .

and P be the open. E[P ] = l ∑ i=1 P (Si )qSi =   kj ∑ i=1 l ∑ i=1   ∑  αj  qSi = ≤ where the last inequality is from (2.2 implies that E[P ] ≤ 1 ∑ ∗ P (S)zS .488 ∑ ∑ ∗ cij x∗ + fi yi  . ij δ 1−δ δ i∈F j∈D i∈F = 6 .4 Setting δ = 1 − e−1/1.488-approximation algorithm [19] called in Step 5 of Algorithm 2. respectively. ∑ j∈D ′  j∈Si qSi  αj = 1∑ z j αj δ j∈D P (z) 1 ∑ ∗ ≤ P (S)zS .1 is an LP-based algorithm.488 ∑ ∑ ∗ dβ cij x∗ + fi yi  ij 0 δ 1−β i∈F j∈D i∈F   ( ) ∑∑ ∑ 1 1.1.488 ∗  ln cij x∗ + fi yi  . F +C ≤ ij 1−β i∈F j∈D i∈F ∗ ∗ implying that ∫ E[F + C] ≤   ∑ 1 1.1 for the FLPSP is no ( )−1 more than 1 − e−1/1. δ δ S⊆D 2 Theorem 2. ij i∈F implying that for any j ∈ D \ S.Now we can bound the expected penalty cost as follows.3). Proof : Let F .044. connection and penalty costs of the solution obtained by Algorithm 2. the approximation ratio of Algorithm 2. y /(1 − β)) is a feasible solution to the LP relaxation of the classical FLP. So   ∑ 1. the set of unrejected clients is { } ∑ D \ S = j ∈ D|1 − x∗ < β . For the penalty cost. ∑ i∈F x∗ > 1 − β. δ S⊆D For the open and connection costs.488 ≤ 2. given any β.488 . Recall that the 1. C. Lemma 2. ij Therefore (x /(1 − β).

488 i∈F j∈D i∈F S⊆D   ∑∑ ∑ ∑ ∗ ∗ ≤ 2.1 the solution (x∗ . S ⊆ D.6) S⊆D:j∈S wj .t. ln δ δ ( 1 1−δ )}   ∑∑ i∈F j∈D cij x∗ + ij ∑ i∈F ∗ fi yi + ∑ S⊆D  ∗ P (S)zS  . ∀j ∈ D. αj ≥ 0. (2. min s. zS ∈ {0.1 by just solving the convex relaxation (2.5) to obtain an new algorithm with the same approximation ratio except the extra term ε.5). ∑∑ i∈F j⊆D cij xij + ∑ i∈F fi yi + P ′ (z) (2. y ∗ .488 .5) i∈F xij ≤ yi . To avoid solving (2. ∑ ∑ j∈S j∈D αj zj αj ≤ P (S). zj ≥ 0.t. ∀i ∈ F. ∀S ⊆ D. ∀j ∈ D. 2. xij . Choosing δ = 1 − e−1/1. j ∈ D. 7 ∀j ∈ D. P ′ (z) = max s. yi .2 General rounding framework ∑ ∑ S⊆D We can extend our rounding technique in the previous section to the following more general model: min s. ∀j ∈ D.Finally we have { E[F + C + P ] ≤ max 1 1. z) constructed in Step 2 is actually an optimal solution to (2.488 .4) ∑ xij + zj ≥ 1. j ∈ D. we get E[F + C + P ] ≤ 1  ∗ ∗  cij x∗ + fi yi + P (S)zS  ij 1 − e−1/1. ∀i ∈ F. in which. (2. .t. 1}. we can replace the first two steps in Algorithm 2. ij i∈F j∈D i∈F S⊆D ∑∑ ∑ ∑  2 We remark that there is a convex relaxation for the FLPSP presented by Chudak and Nagano [7]. φ(w) + wj + P (S)zS zS ≥ 1.1) with exponential number of variables.044  cij x∗ + fi yi + P (S)zS  . It is not hard to see that in Algorithm 2.

for example.5 There exists a function φ: [0. where α ≥ 1. z ∗ ). Assumption 2. we can efficiently find a solution to ϕ(w) of cost at most αφ(w).g. in which the clients can receive service or be rejected. ∀ 0 ≤ β < 1. ¯ ¯ 2) for any fixed w ∈ {0. ∀j ∈ D. This new problem embeds a subproblem. 1}nc .5 below. (2. Algorithm 2. ∗ STEP 3. i. can be solved efficiently. φ(w) ≤ φ(w). Assumption 2. for the FLPSP. Construct variable w ∈ {0. for all w ∈ [0. 1]nc . 1}nc by setting w := I(D \ S). We need one more assumption called scaling property [12]. φ(w) + ¯ wj + ∑ ∑ S⊆D P (S)zS zS = 1. ¯ 1−β ∀ w∗ ∈ [0. Find a solution to the subproblem ϕ(w) and serve all the unrejected clients D \ S. and pay the penalty cost P (S). Select the parameter β uniformly at random from the interval [0. The binary vector w indicates whether the client j receives service or not. such that ¯ 1) φ is a lower bound on φ.7 The function φ satisfies the scaling property if ¯ φ(w) ≤ ¯ 1 φ(w∗ ).This model captures a class of covering problems.t. STEP 4.7) to obtain an optimal fractional solution (w∗ .. ϕ(w) is the classical FLP. The following algorithm is an extension of Algorithm 2. 8 . 1]nc . Solve the relaxation problem (2.1) in the case of the FLPSP. Reject the subset S := {j|1 − wj ≥ β}. S ⊆ D. The above assumption will make it possible for us to design a constant-factor approximation algorithm for the general model. Note that (2. δ). denoted as ϕ(w).7) S⊆D:j∈S wj .e. ¯ 3) the optimization problem min s. In general the subproblem φ(w) is NP-hard—e.1. 1]nc −→ R+ . ∀j ∈ D. 1]. and becomes (2. and the subset of clients Srej rejected will incur a penalty cost P (Srej ).6). The term φ in the objective function satisfies Assumption 2. STEP 2. zS ∈ [0.7) is a relaxation of (2.6 STEP 1.

1) and convert its solution into (x∗ . y ∗ . the algorithm and its analysis ignored the possibility where unrejected clients may have paid fractional cost for penalty.4) STEP 2. z ∗ ). y ∗ . Construct an instance of classical FLP with the set of facilities F. For example. 2 STEP 3.4 lead to the following result: Theorem 2. z ∗ ) for (2.7. y ∗ /(1 − β)) is a feasible solution to the relaxed problem.1 offers an improved approximation ratio. we know the total cost of the optimal fractional solution (x∗ . 2 We will show that Algorithm 3.1. we presented a general algorithmic frame for a class of covering problems with submodular penalty.1 STEP 1. Moreover.Note that this assumption indeed holds in the FLPSP.6 for (2. (2) P (Sr ) ≤ 2 ∑ j∈Sr ∗ ∗ αj zj .8 Setting δ = 1 − e−1/α . the approximation ratio of Algorithm 2. and pay the penalty cost P (S). Algorithm 3.5 and 2. 16] for the instance and assign the clients in D \ Sr to the closest open facilities. Or solve the LP relaxation (2. the set of clients D\Sr and the connection cost cij .4) of the problem to obtain an optimal fractional solution (x∗ . With Assumptions 2. z ∗ ) for (2. y ∗ . After the Step 1 of Algorithm 3. similar proofs to those in Lemma 2. Then run the JMS algorithm [15.1. In this section.6) is no more than ( )−1 1 − e−1/α . but this frame can be too crude to be applied to certain specific problems and hence sometimes we need to refine the general frame to yield improved approximation ratios. i ∈ F.2 and Theorem 2.1) is ∑∑ i∈F j⊆D cij x∗ + ij ∑ i∈F ∗ fi yi + P ′ (z ∗ ). 3 Improved 2-approximation algorithm for the FLPSP In the previous section. j ∈ D\Sr . 9 .2 (1) ∑ j∈D ∗ ∗ αj zj = P ′ (z ∗ ). we will incorporate this ignored cost into the design and analysis of our algorithms to obtain improved approximation ratio for the FLPSP. when we applied it to the FLPSP earlier in Section 2. because (x∗ /(1 − β). we have Lemma 3. Reject the subset Sr := {j|zj ≥ 1 }. Solve the convex relaxation (2.

Suppose on the contrary.e. the total cost of the solution generated by Algorithm 3. we extend the LP 10 . Lemma 3.1 is a 2-approximation algorithm for the FLPSP.3 directly.∗ ∗ (3) For any j ∈ Sr . Theorem 3.1 is no more than   ∑∑ ∑ ∗ 2 cij x∗ + fi yi + P ′ (z ∗ ) . where ϵ > 0 is sufficiently small to maintain the order of {z ∗ }.2 implies that the open and connection costs of the new fractional solution are x ¯ respectively given by ∑ i∈F 2 fi yi = 2 ¯ ∑ i∈F ∗ fi yi and ∑ ∑ i∈F j∈D\Sr cij xij ≤ ¯ ∑ ∑ i∈F j∈D\Sr cij x∗ + ij ∑ j∈D\Sr ∗ ∗ αj z j . but αj < cij . a contradiction to the optimality of (x∗ . y ∗ ) as follows: scale all / the y ∗ ’s by 2. z1 ≤ z2 · · · zj−1 ≤ zj + ϵ ≤ zj+1 · · · ≤ znc . In this section. i. 2). z ∗ ). So we focus on (3). Then {α∗ } is still the optimal solution of (2.5148-approximation algorithm for the FLPLP Recently. there ∗ ∗ exists a client j ∈ Sr satisfying zj > 0 and x∗ > 0 for some facility i. if zj > 0 and x∗ > 0. y ).2. ij i∈F j∈D i∈F 2 4 Improved 1.2) and we get a better solution to the convex relaxation (2. z) = (x∗ − ϵ. y ∗ . We now prove the approximation ratio of the above algorithm. the cost for the unrejected clients is no more than   ∑ ∑ ∑ ∑ ∗ ∗ ∗ αj z j  . y ∗ . then αj ≥ cij . we can construct a feasible fractional solution from (x∗ . y. Proof : For all j ∈ Sr . Li [19] offered an improved analysis of the LP rounding algorithm by Byrka and Aardal [3] to obtain the currently best approximation ratio for the classical to FLP. 2 cij x∗ + fi yi + ij i∈F i∈F j∈D\Sr j∈D\Sr Combining with Lemma 3.3 Algorithm 3.4): (x.4). Since the JMS algorithm [15. Then we construct a / ij new feasible solution to convex relaxation (2. z ∗ + ϵ). / ij Proof : (1) and (2) follow from Claim 2. and reassign the clients j ∈ Sr to its closest fractional open facilities to get a new fractional / solution (¯. 16] has a bi-factor of (1.

The total optimal fractional cost includes three parts: the opening cost F ∗ = i∈F fi yi . y ∗ . y ∗ . yi . and binary variable zj indicates whether client j is penalized or not. ij We have the following lemma to capture the property of the optimal fractional solution (x∗ . the new z-variable must be zero. y ∗ . 11 . z ∗ ). for each client j.1 For any j ∈ D. we say that a facility i is one of its close facilities if it fractionally serves client j in (¯. but ij ∗ pj < cij . { Dγ ¯ Dγ = j ∈ D| } . z ∗ ) and the corresponding optimal dual ∑ ∗ solution as (α∗ . a ij contradiction to the optimality of (x∗ . If xij = 0.8) i∈F xij ≤ yi . β ∗ ). there exists no i ∈ F and j ∈ D such that 0 < xij < yi . γy ∗ . y ) be the resultant x ¯ complete solution (i. z) = (0. then we say that i is a x ¯ ¯ distant facility of client j. For a client j. We partition the set of clients D into two subsets. xij .rounding algorithm by Byrka and Aardal and the analysis by Li to obtain an improved algorithm for the FLPLP. y ). Then we can generate a better feasible solution to the LP relaxation: (x. z ∗ ). j ∈ D. then pj ≥ cij . ∀i ∈ F. y ∗ . ij ∗ Proof : Suppose on the contrary. making room for modification of x∗ and z ∗ in order to reduce the total cost. we first randomly scale y ∗ by a number γ > 1 to obtain a suboptimal fractional solution (x∗ . j ∈ D. ∑∑ i∈F j∈D cij xij + ∑ i∈F fi yi + ∑ j∈D pj zj ∑ xij + zj ≥ 1. y. In our algorithm. binary variable yi indicates whether facility i is open or not. (4. resulting in a new feasible solution.1 implies that for any client j ∈ Dγ . we can omit z-variables and let (¯. However. if zj > 0 and x∗ > 0. Then. z ∗ ). otherwise we may split ¯ ¯ the facilities to obtain an equivalent instance with a complete solution—refer to Lemma 1 in [26] for a more detailed argument). So for j ∈ Dγ . ∀i ∈ F. Denote the optimal solution of the LP relaxation as (x∗ . zj + x∗ ). but facility i was serving client j in solution (x∗ .t. z ∗ ). and the penalty cost P ∗ = j∈D pj zj = j∈D Pj∗ . we modify the corresponding x∗ and z ∗ to reduce his fractional connection cost and his fractional penalty cost. ∑ ∑ ∑ ∑ ∑ ∗ ∗ the connection cost C ∗ = i∈F j∈D cij x∗ = j∈D Cj . 2 ∑ i∈F x∗ ij 1 ≥ γ = D \ Dγ . there exists a client j satisfying zj > 0 and x∗ > 0 for some facility i. ∗ Lemma 4. zj ∈ {0. ∀j ∈ D. 1}. Lemma 4. which has an natural integer programming formulation: min s. where the binary variable xij indicates whether client j is connected to facility i or not. y ∗ .e.

. where each client j ∈ Dγ is directly connected to his close facilities. we choose γ between (1. .55 probability. We let γ = 1. to which client j is connected to fractionally in (x∗ .3360. Clients connected to the same x ¯ facility in G are called neighbors. arranged in the nondecreasing order of distances. where t is minimum ∑t ∑m ∗ ∗ number such that s=1 yis ≥ p. Note that for every client j ∈ D.For any client j ∈ D. . let i1 . We are now ready to present our algorithm. z ∗ ).45 and with the remaining 0.3360 with probability 0. Consider the bipartite graph G obtained from the solution (¯. . The main component in the algorithm is to cluster this graph recursively in a greedy fashion (Step 4). the following facts hold: • The sum of average connection and penalty costs equals Dav = C • The average connection cost to a close facility equals Dav = γ ∫1 0 hj (p)dp = Fj∗ + Pj∗ . 12 . im be the facilities. Then we define hj (p) = cit . γ γ−1 ∫ 1/γ 0 D • The average connection cost to a distant or rejected facility equals Dav = ∫1 1/γ hj (p)dp. C • The maximal distance to a close facility is at most the average distance to a distant facility. i2 . y ).1. Dmax = hj (1/γ). or hj (p) = pj when p > s=1 yis .j . similarly to that used by Byrka and Aardal [3].9860] uniformly. y ∗ . We denote the distribution as µ(γ). hj (p)dp.

For every cluster center j. STEP 3. connect it to an open facility or pay penalty cost when the minimal connection cost to an open facility is larger than pj . Scale up the value of the facility opening variables y ∗ by a random number γ obeying µ(γ). y ). j ∈ Dγ . which is different from the LP rounding algorithm in Byrka and Aardal [3] and Li [19]. We have ( ) 1 γ 0 ( ∫ 1 ( )) ∫ 1 1 hj (p)dp + (3 − γ)hj . If necessary. we will use the following lemma from [19] which holds also for our algorithm. j ∈ Dγ . 2 Now we present our main result for the FLPLP which follows the approach of Li [19]. The lemma is concluded by Lemma 3 in [19]. γ Proof : Let us consider the following two possibilities. During the analysis of the above algorithm. Solve the LP relaxation (4. We remark that only clients in Dγ can be chosen as cluster centers. we have ( ∫ ∫ 1 E[Cj + Pj ] ≤ hj (p)e−γp γdp + e−γ γ 0 0 1 hj (p)dp + (3 − γ)hj ( )) 1 .Algorithm 4. x ¯ STEP 4. open it independently with probability yi . For each facility i that is not a close facility of any cluster center.8) to obtain an optimal fractional solution (x∗ . STEP 5. z ∗ -variables so that each client is connected to its closest fractionally open facilities. ¯ Case 2. STEP 2. 13 . y ∗ . We only give the sketch of the proof.2 STEP 1. Construct a greedy clustering from solution (¯. Lemma 4.3 For any client j. z ∗ ). ¯ STEP 7. For any client j ∈ D. y ) by choosing recursively as cluster centers the x ¯ C C unclustered clients in Dγ with minimal Dav (j) + Dmax (j). ¯ STEP 6. open one of its close facilities randomly with probabilities xij . ≤ hj (p)e−γp γdp + e−γ γ γ 0 0 ∫ 1 E[Cj + Pj ] ≤ hj (p)e−γp γdp + e−γ hj where the last inequality holds since γ ≤ 2 in our distribution µ(γ). Case 1. Then modify the value of the x∗ . split facilities to obtain a complete solution (¯.

Proof : The scaling factor for the facility cost is λf = 0.2246 e−1.2 is an 1.5146(C ∗ + P ∗ ) ≤ 1.3360q − e−1.3309. 0 ≤ q < 1/1.3360. 2 To obtain the worst approximation factor λQ for the connection cost and penalty cost.4 Algorithm 4.8462(e−1.45 (1 − q)q 1−q 1−q 0.0019 − (2 − 1/q)e−1/q ).5148(F ∗ + C ∗ + P ∗ ). achieved at q = 0. the maximum value of λQ (q) is 1. 1/1. 0.Theorem 4.3360q + + 0.1138 e−1.5148. 1/1.9860. (1 − q)q 1−q 1−q λQ (q) = For all three cases.3360 + 0.9860 = 1. implying that Algorithm 4.3360q + + 0.3360q − e−1. which is a convex combination of step-functions.45 × 1. for p < q. 0.5148-approximation algorithm.9860q ) −0.3360q + + 0. 2). λQ (q) = 0.45 + 0. 1−q Case 3.9860 ≤ q ≤ 1/1. The research of the second author is supported by the Natural Sciences and Engineering 14 . and the objective is a linear function):   0.5148F ∗ + 1.3360 + 1.8462(e−1.5148(F ∗ + C ∗ + P ∗ ).9860q ) 0. University of New Brunswick and supported in part by NSERC grants 283103.45 + 0. Therefore we get a solution with value no more than than 1. hq (p) =  1 .3360 < q < 1.8462(e−1.5146.2 produces a solution with expected cost E[F + C + P ] ≤ 1. 2 Acknowledgements This work was partially done while the first author was a visiting doctorate student at the Faculty of Business Administration.7400 (See Fig.3309 + (0.8462 +0. (1 − q)q 1−q 1−q λQ (q) = Case 2. 1−q Now we can explicitly calculate λQ (q) in three cases: Case 1.1138 e−1.3360q − e−1.55 1.3309. for p ≥ q. we only need to consider step-functions as follows (since h(p) is a nondecreasing function.9860q ) 0.

Zhang. 1102001). K. and Scientific Research Common Program of Beijing Municipal Commission of Education (No. J. Aardal. Y..8 0.I.2 0.: Algorithms for facility location problems with outliers. S. M. A. Chen. Beijing Natural Science Foundation (No. 378-388 (1999) [5] Charikar. pp.I. SIAM Journal on Computing 39..3 0.Research Council of Canada (NSERC) grant 283103.7 0. K.6 0. pp.. The third author’s research is supported by the National Basic Research Program of China (No. In: Proceedings of SODA. Shmoys. B. G..: Improved combinatorial algorithms for facility location and k-median problems.. X. 642-651 (2001) [6] Chen.B. Algorithmica 53.M. Mount. D. Chudak. M..A. Khuller. S.. References [1] Aardal. KM201210005033). Ye..: Improved combinatorial approximation algorithms for the k-level facility location problem. 2010CB732501). 161-167 (1999) [2] Ageev.9 1 15 . 263-297 (2007) 0.: Approximation algorithms for soft-capacitated facility location in capacitated network design.: A 3-approximation algorithm for the k-level uncapacitated facility location problem..5 0. J. 11071268). 2212-2231 (2010) [4] Charikar. D. The fourth author’s research is supported by NSF of China (No.. SIAM Journal on Discrete Mathematics 18. Naraasimban.: An optimal bi-factor approximation algorithm for the metric uncapacitated facility location problem. Information Processing Letters 72. 207-217 (2003) [3] Byrka. Guha. F. In: Proceedings of FOCS.4 0.

S.: Approximation algorithms for integrated distribution network design problems..B.. S. Wang.. E.. Mahdian.: Improved approximation algorithms for the uncapacitated facility location problem. Xu.. Submitted (2010) 16 . X. M. V. Saberi. Mathematical Programming 130. H. vol. V.: Submodular Functions and Optimization (Second Edition).: A 1. Plaxton. Shmoys..[7] Chudak. K. 1-10 (1998) [19] Li.: Analysis of a local search heuristic for facility location problems. 191-200 (2012) [10] Du. Mahdian. Wang.. S. Journal of the ACM 48.A.. J. 795-824 (2003) [16] Jain. Xu.. 228-248 (1999) [14] Hayrapetyan.A. 79-88 (2007) [8] Chudak... Zhang.. 361-368 (2010) [11] Fujishige. N. R. Levi... In: Proceedings of SODA. X. Annals of Discrete Mathematics. pp. Tard¨s. J.: A primal-dual approximation algorithm for the facility location problem with submodular penalty. A. K. pp. 274-296 (2001) [18] Korupolu.. D. C.: Greedy strike back: improved facility location algorithms. SIAM Journal on Computing 33. Part II.: Greedy facility location algorithms analyzed using dual fitting with factor-revealing LP. 58. D. F... a pp. D.: An approximation algorithm for the k-level capacitated facility location problem. Xiu.: Approximation algorithms for supply chain planning and logistics problems with market choice. A. Vazirani.G.B. Journal of Combinatorial Optimization 20. pp. E.V.: Approximation algorithms for metric facility location and k-median problems using the primal-dual schema and Lagrangian relaxation. A new greedy approach for facility location problems.. D. Lu. A. Elsevier (2005) [12] Geunes.. F. In: Proceedings of SODA. Journal of Algorithms 31.. Saberi. S...: Network design for information networks.: Efficient solutions to relaxations of combinatorial problems with submodular penalty via the Lov´sz extension and non-smooth convex optimization.. Shu. In: Proceedings of STOC. Markakis. Xu. K.V.R. Khuller. In: Proceedings of the ICALP. Journal of the ACM 50. 77-88 (2011) [20] Li. 1-25 (2003) [9] Du. 85-106 (2011) [13] Guha.488 Approximation algorithm for the uncapacitated facility location problem. K. D.. C. Vazirani.. Nagano. pp. D. Romeijn. J. M. M. Shmoys.. In: Proceedings of o SODA. 933-942 (2005) [15] Jain. Rajaraman. Y. Swamy. R.E.. 731-740 (2002) [17] Jain. Algorithmica 63. D. R.

Aardal. Report No.: Facility location and the analysis of algorithms through factor-revealing programs.de/vygen/fl.: Improved approximation algorithms for metric facility location problems. Zhang. Kluwer Academic Publishers. M. Information Processing Letters 94.. Ye.. K. C. 265-274 (1997) [25] Shu. Tard¨s.. 05950-OR. G. Z.. In: Proceedings of IPCO. 411-432 (2006) [23] Shmoys. In: o Proceedings of STOC.. M.or. J.pdf (2005) [28] Xu. J. G. Research Institute for Discrete Mathematics. Shen. D. University of Bonn.. Journal of the ACM 53. Operations Research 53. Ph.B.: An LP rounding algorithm for approximating uncapacitated facility location problem with penalty. Chen. pp. 978-1012 (2006) [24] Shmoys. Max: Stochastic transportation-inventory network design problem. J. 48-60 (2005) [26] Sviridenko. Zhang.: An improved approximation algorithm for the metric uncapacitated facility location problem. Y.B.: A multiexchange local search algorithm for the capacitated facility location problem.. Journal of Combinatorial Optimization 17.I.unibonn. D..[21] Mahdian. MIT. Teo. 424-436 (2008) [30] Ye. MA (2004) [22] Mahdian. Y.: Approximating the two-level facility location problem via a quasi-greedy approach. Xu. http://www.: An improved approximation algorithm for uncapacitated facility location problem with penalty..: An approximation scheme for stochastic linear programming and its application to stochastic integer programs. Y. Ye. J. M. Combinatorial Optimization in Communication Networks.P. Xu. 159-176 (2006) [32] Zhang. 389-403 (2005) 17 . J. Cambridge.J. Swamy. 119-123 (2005) [29] Xu. 623-637 (2005) [31] Zhang. Mathematics of Operations Research 30.: Approximation algorithms for facility location problems. thesis... SIAM Journal on Computing 36. pp.. B.: An approximation algorithm for the dynamic facility location problem. C. Mathematical Programming 108. D. E. 240-257 (2002) [27] Vygen. J.: Approximation algorithms for facility location problems (Lecture Notes). pp. J. J.

Sign up to vote on this title
UsefulNot useful