Approximation Via Cost-Sharing: A Simple Approximation Algorithm for the

Multicommodity Rent-or-Buy Problem

Anupam Gupta Amit Kumary Martin Pál z Tim Roughgardenz

Abstract an edge is given by a simple concave function: capacity can

be rented, with cost incurred on a per-unit of capacity basis,
We study the multicommodity rent-or-buy problem, a or bought, which allows unlimited use after payment of a
type of network design problem with economies of scale. large fixed cost. Precisely, there are positive parameters 
In this problem, capacity on an edge can be rented, with and M , with the cost of renting capacity equal to  times
cost incurred on a per-unit of capacity basis, or bought, the capacity required (per unit length), and the cost of buy-
which allows unlimited use after payment of a large fixed ing capacity equal to M >  (per unit length). By scaling,
cost. Given a graph and a set of source-sink pairs, we seek a there is no loss of generality in assuming that  = 1.
minimum-cost way of installing sufficient capacity on edges The MRoB problem is a simple model of network design
so that a prescribed amount of flow can be sent simultane- with economies of scale, and is a central special case of the
ously from each source to the corresponding sink. The first more general buy-at-bulk network design problem, where
constant-factor approximation algorithm for this problem the cost of installing capacity can be described by an arbi-
was recently given by Kumar et al. (FOCS ’02); however, trary concave function. In addition, the MRoB problem nat-
this algorithm and its analysis are both quite complicated, urally arises as a subroutine in multicommodity versions of
and its performance guarantee is extremely large. the connected facility location problem and the maybecast
In this paper, we give a conceptually simple 12- problem of Karger and Minkoff [21]; see [24] for further
approximation algorithm for this problem. Our analysis of details on these applications.
this algorithm makes crucial use of cost sharing, the task of The MRoB problem is easily seen to be NP- and MAX
allocating the cost of an object to many users of the object SNP-hard—for example, it contains the Steiner tree prob-
in a “fair” manner. While techniques from approximation lem as a special case [8]—and researchers have therefore
algorithms have recently yielded new progress on cost shar- sought approximation algorithms for the problem. For
ing problems, our work is the first to show the converse— many years, the best known performance guarantee for
that ideas from cost sharing can be fruitfully applied in the MRoB was the O(log n log log n)-approximation algorithm
design and analysis of approximation algorithms. due to Awerbuch and Azar [3] and Bartal [5], where n =
jV j denotes the number of nodes in the network. The first
constant-factor approximation algorithm for the problem
1 Introduction was recently given by Kumar et al. [24]. However, both
We study the multicommodity rent-or-buy (MRoB) prob- the analysis and the primal-dual algorithm of [24] are quite
lem. In this problem, we are given an undirected graph complicated, and the performance guarantee shown for the
G = (V; E ) with non-negative weights e on the edges and algorithm is extremely large.1 The problem of obtaining an
a set D = f(s1 ; t1 ); : : : ; (sk ; tk )g of vertex pairs called de- algorithm with constant performance guarantee for MRoB
mand pairs. We seek a minimum-cost way of installing suf- while using only a transparent algorithm and/or obtaining a
ficient capacity on the edges E so that a prescribed amount reasonable constant has since remained open.
of flow can be sent simultaneously from each source si to In a separate recent development, Gupta et al. [19]
the corresponding sink ti . The cost of installing capacity on showed that extremely simple randomized combinatorial al-
 Department of Computer Science, Carnegie Mellon University, Pitts-
gorithms suffice to achieve best-known performance guar-
burgh PA 15213. Email: [email protected]
antees for several network design problems, including the
y Lucent Bell Labs, 600 Mountain Avenue, Murray Hill NJ 07974. single-sink special case of MRoB in which all sink ver-
Email: [email protected].
z Department of Computer Science, Cornell University, Ithaca NY 1 While this constant was neither optimized nor estimated in [24], the

14853. Supported by ONR grant N00014-98-1-0589. Email: approach of that paper only seems suitable to proving a performance guar-
fmpal,[email protected]. antee of at least several hundred.

1
tices ti are identical. However, the analysis of [19] re- Related Work: As stated above, the only previous
quired that the underlying “buy-only” problem (such as constant-factor approximation algorithm for the MRoB
the Steiner tree problem for the single-sink special case of problem studied in this paper is due to Kumar et al. [24].
MRoB) admit a good greedy approximation algorithm (e.g., Additional papers that considered multicommodity network
the MST heuristic for Steiner tree, with Prim’s MST algo- design with economies of scale are Awerbuch and Azar [3],
rithm). Since the “buy-only” version of the MRoB prob- Bartal [5], and Fakcharoenphol et al. [11], whose work
lem is the Steiner forest problem (see e.g. [1]), for which no gives an O(log n)-approximation for the more general mul-
greedy algorithm is known, it was not clear if the techniques ticommodity buy-at-bulk problem. The special case of
of [19] could yield good algorithms for MRoB. MRoB where all commodities share the same sink, and
the closely related connected facility location problem,
Our Results: We show how a nontrivial extension of the
have been extensively studied [18, 19, 21, 22, 23, 27, 29].
randomized framework of [19] gives a 12-approximation al-
The randomized 3.55-approximation algorithm of Gupta et
gorithm for MRoB. The algorithm is conceptually very sim-
al. [19] is the best known approximation algorithm for the
ple: it picks a random subset of the source-sink pairs, buys a
problem. Swamy and Kumar [29] achieve a performance
set of edges spanning these chosen pairs, and greedily rents
guarantee of 4.55 with a deterministic algorithm. Several
paths for the other source-sink pairs.
more general problems that retain the single-sink assump-
Our analysis of the algorithm is based on a novel connec-
tion have also been intensively studied in recent years, in-
tion between approximation algorithms and cost sharing,
cluding the Access Network Design problem [2, 16, 17, 26],
the task of allocating the cost of an object to many users of
the single-sink buy-at-bulk network design problem [13, 17,
the object in a “fair” manner. This connection, rather than
19, 28, 30], and the still more general problems where the
our specific results, is arguably the most important contri-
capacity cost function can be edge-dependent [10, 25] or
bution of this paper.
unknown to the algorithm [14]. The best known approxi-
Cost sharing has been extensively studied in the eco- mation ratios for these four problems are 68 [26], 73 [19],
nomics literature (see e.g. [31]); more recently, techniques O(log n) [10, 25], and O(log n) [14], respectively.
from approximation algorithms have yielded new progress
Finally, our high-level algorithm of randomly reducing
in this field, see e.g. [20]. We believe the present work to be
the MRoB problem to the Steiner forest problem, followed
the first showing the converse, that ideas from cost sharing
by computing shortest paths, is similar to and partially in-
can lead to better approximation algorithms. A second key
spired by previous work that gave online algorithms with
ingredient for our result is a simple but novel extension of
polylogarithmic competitive ratios for many rent-or-buy-
the primal-dual algorithms of Agrawal et al. [1] and Goe-
type problems [4, 6, 7].
mans and Williamson [15] for the Steiner forest problem.
Our performance guarantee of 12 is almost certainly not
the best achievable for the MRoB problem, but it is far bet- 2 Approximation via Cost Sharing
ter than any other approximation ratio known for a network In this section we show how an appropriate cost-sharing
design problem exhibiting economies of scale, with the ex- scheme gives a good approximation algorithm for MRoB.
ception of the single-sink special case of MRoB (for which We will give such a cost-sharing scheme in the technical
a 3.55-approximation is known [19]). Single-sink buy-at- heart of the paper, Sections 3 and 4.
bulk network design, where all commodities share a com-
In Subsection 2.1, we define our desiderata for cost
mon sink but the cost of installing a given amount of capac- shares. We give the main algorithm in Subsection 2.2, and
ity can essentially be an arbitrary concave function, is only
its analysis in Subsection 2.3.
known to be approximable to within a factor of 73 [19].
Keeping the single-sink assumption and placing further re-
strictions on the function describing the cost of capacity
2.1 Some Definitions
yields the Access Network Design problem of Andrews and We now describe precisely what we mean by a cost-sharing
Zhang [2], for which the best known approximation ratio is method, as well as the additional properties required by our
68 [26]. (There are no known constant-factor approxima- application. Cost-sharing methods can be defined quite gen-
tion algorithms for the multicommodity versions of these erally (see e.g. [31]); here, we take a narrower approach. In
two problems.) The previous-best performance guarantee preparation for our first definition, recall that an instance
for MRoB was still larger, at least several hundred [24]. The of MRoB is defined by a weighted undirected graph G (we
present work is thus the first to suggest that restricting the leave the weight vector implicit) and a set D of demand
capacity cost function could yield a more tractable special pairs. By a Steiner forest for (G; D), we mean a subgraph
case of multicommodity buy-at-bulk network design than F of G so that, for each demand pair (s; t) 2 DP , there is
the popular assumption that all commodities share a com- an s-t path in F . For such a subgraph F , (F ) = e2F e
mon sink. denotes its overall cost. Since we are only interested in so-
lutions with small cost and edge costs are non-negative, we 2.2 The Algorithm SimpleMROB
can always assume that F is a forest. We now state the main algorithm. In employs as a subrou-
The next definition states that a cost-sharing method is tine a Steiner forest algorithm A, which in our implementa-
a way of allocating cost to the demand pairs of an instance
(G; D ), with the total cost bounded above by the cost of an
tion will be a constant-factor approximation algorithm. No
optimal Steiner forest for (G; D).
cost-sharing method is needed for the description or imple-
mentation of the algorithm; cost shares arise only in the
Definition 2.1 A cost-sharing method is a non-negative algorithm’s analysis. We assume for simplicity that each
real-valued function  defined on triples (G; D; (s; t)), for source-sink pair wants to route a single unit of flow. This
a weighted undirected graph G, a set D of demand pairs, assumption is not hard to remove (details are deferred to the
and a single demand pair (s; t) 2 D. Moreover, for every full version).
instance (G; D) admitting a min-cost Steiner forest FD ,
X 1. Mark each pair (si ; ti ) with probability 1=M , and let
(G; D; (s; t))  (FD ): Dmark be the set of marked demands.
(s;t)2D
2. Construct a Steiner forest F on Dmark using algorithm
Definition 2.1 permits some rather uninteresting cost-
sharing methods, such as the function that always assigns
A, and buy all edges in F .
all demand pairs zero cost. The key additional property that 3. For each (si ; ti ) pair outside Dmark , rent edges to
we require of a cost-sharing method is that, intuitively, it connect si and ti in a minimum-cost way (at cost
allocates costs to each demand pair commensurate with its dG=F (s; t)).
distance from the edges needed to connect all of the other
demand pairs. Put differently, no demand pair can be a “free 2.3 Proof of Performance Guarantee
rider”, imposing a large burden in building a Steiner forest,
We now state the main theorem of this section: a constant-
but only receiving a small cost share. We call cost sharing
factor approximation algorithm for Steiner forest with an
methods with this property strict.
accompanying O(1)-strict cost-sharing method yields a
To make this idea precise, we require further notion. Let
dG (
;
) denote the shortest-path distance in G (w.r.t. edge
constant-factor approximation algorithm for MRoB. The
costs in G). Given a subset E 0  E of edges, G=E 0 is
proof is an extension of the techniques of [19], where the
the graph obtained from G by contracting the edges in E 0 .
connection to cost sharing was not explicit and only sim-
pler network design problems were considered.
Note that the cheapest way of connecting vertices s and t
by renting edges, given that edges E 0 have already been Theorem 2.3 Suppose A is an -approximation algorithm
bought, is precisely dG=E 0 (s; t). Our main definition is then for the Steiner forest problem that admits a  -strict cost
the following. sharing method. Then algorithm SimpleMROB, employ-
ing algorithm A as a subroutine in Step (2), is an ( +  )-
Definition 2.2 Let A be a deterministic algorithm that,
given instance (G; D), computes a Steiner forest. A cost-
approximation algorithm for the multicommodity rent-or-
sharing method  is  -strict for A if for all (G, D) and
buy problem.
(s; t) 2 D , the cost (G; D ; (s; t)) assigned to (s; t) by  is Proof: Fix an arbitrary instance of MRoB with an optimal
at least a 1= fraction of dG=F (s; t), where F is the Steiner solution OPT, and let Z  denote the cost of OPT. Let B 
forest returned for (G; D n f(s; t)g) by algorithm A.2 denote the cost of the bought edges Eb in OPT, and R the
It is not clear a priori that strict cost-sharing methods cost of theP rented edges Er . (Note that B  = M (Eb ),
with small  exist: Definition 2.1 states that the aggregate and R = e2Er e xe , where xe is the amount of capacity

costs charged to demand pairs must be reasonable, while that OPT reserves on edge e, or equivalently the number of
Definition 2.2 insists that the cost allocated to each demand demand pairs that use e to route their flow). It suffices to
pair is sufficiently large. show that algorithm SimpleMROB incurs an expected cost
We note that strict cost-sharing methods are somewhat of at most Z  in Step (2) and at most Z  in Step (3). We
reminiscent of some central concepts in cooperative game prove each of these bounds in turn.
theory, such as the core and the nucleolus (see e.g. [31]). For the first bound, it suffices to show that:
However, we are not aware of any precise equivalence be-  
tween strict cost-sharing methods and existing solution con- E
cost of buying a min-cost Steiner forest on
 Z :
cepts in the game theory literature. the (random) set of demand pairs Dmark
2A cost-sharing method (which is defined independent of any Steiner
(2.1)
forest algorithm) can be  -strict for one algorithm and not for another,
as strictness depends on the distance dG=F (s; t), and the edge set F is To prove this, it suffices to exhibit a (random) subgraph of
algorithm-dependent. G that spans the vertices of Dmark that has expected cost at
most Z  =M . To construct this subgraph, include all edges sharing methods for any constant  . On the other hand,
of Eb , and every edge e of Er for which some demand pair in these families of examples, an O(1)-strict cost-sharing
using e in OPT is marked. The cost of the edges in Eb is method can be defined if a few extra edges are bought. (Ex-
deterministically B  =M . For each edge in Er , the expected tra edges make Definition 2.2 easier to satisfy, since the
cost of including it is e  1=M  xe = e xe =M , since each shortest-path distance dG=F relative to the bought edges F
of the xe demand pairs using e in OPT contributes 1=M to decreases.) Our main technical result is that buying a few
the probability of e being included. Summing over all edges extra edges beyond what is advocated by the algorithms
of Er , the expected cost of including edges of Er is R =M ; of [1, 15] always suffices for defining an O(1)-strict cost-
since Z  = B  + R , inequality (2.1) is proved. sharing method, enabling the application of Theorem 2.3.
We now bound the expected cost incurred in Step (3).
We will say that demand pair (s; t) incurs buying cost 3.1 The Algorithm PD and the Cost Shares 
Bi = (G; Dmark ; (s; t)) and renting cost Ri = 0 if In this subsection we show how to extend the algorithms
(s; t) 2 Dmark , and buying cost Bi = 0 and renting cost of [1, 15] to “build a few extra edges” while remaining
Ri = dG=F (s; t) otherwise. constant-factor approximation algorithms for the Steiner
For a demand pair (si ; ti ), let Xi = Ri Bi denote forest problem. We also describe our cost-sharing method.
the random variable equal to the renting cost of this pair Recall that we are given a graph G = (V; E ) and a
minus  times its buying cost. We next condition on the set D of source-sink pairs f(si ; ti )g. Let D be the set of
outcome of all of the coin flips in Step (1) except that for demands— the vertices that are either sources or sinks in
(si ; ti ). Since  is  -strict for A, this conditional expecta- D (without loss of generality, all demands are distinct). It
tion is at most 0. Since this inequality holds for any out- will be convenient to associate a cost share (D; j ) with
come of the coin flips for demand pairs other than (si ; ti ), each demand j 2 D; the cost share (G; D; (s; t)) is then
it also holds unconditionally: E [Xi ℄  0. Summing over just (D; s) + (D; t). Note that we have also dropped the
allP i and applying P linearity of expectations, we find that reference to G; in the sequel, the cost shares are always
E [ i Ri ℄   E [ i Bi ℄. The left-hand side of this inequal- w.r.t. G.
ity is precisely the expected cost incurred P in Step (3) of the Before defining our algorithm, we review the LP relax-
algorithm. By Definition 2.1, the sum i Bi is at most the ation and the corresponding LP dual of the Steiner forest
cost of the min-cost Steiner forest P on the set Dmark of de- problem that was used in [15]:
mands; by (2.1), it follows that E [ i Ri ℄ is at most Z  , P
and the proof is complete. min e e xe (LP)
In Sections 3 and 4, we will give a 6-approximation al- x(Æ(S ))  1 8 valid S (3.2)
gorithm for Steiner forest that admits a 6-strict cost sharing xe  0
method. The main theorem of the paper is then a direct con-
sequence of Theorem 2.3.
P
S yS
Theorem 2.4 There is a 12-approximation algorithm for max (DP)
P
MRoB.
S V :e2Æ(S ) yS  e (3.3)

3 The Steiner Forest Algorithm yS  0;

We first motivate the algorithm. Linear programming du- where a set S is valid if for some i, it contains precisely one
ality is well known to have an economic interpretation, of si ; ti .
and moreover to be useful in cost sharing applications (see We now describe a general way to define primal-dual al-
e.g. [20] for a recent example). It is therefore natural gorithms for the Steiner forest problem. As is standard for
to expect strict cost-sharing methods to fall out of exist- the primal-dual approach, the algorithm will maintain a fea-
ing primal-dual approximation algorithms for the Steiner sible (fractional) dual, initially the all-zero dual, and a pri-
forest problem, such as those by Agrawal et al. [1] and mal integral solution (a set of edges), initially the empty
Goemans and Williamson [15]. In particular, one might set. The algorithm will terminate with a feasible Steiner
hope that taking the subroutine A of algorithm SimpleM- forest, which will be proved approximately optimal with the
ROB to be such a primal-dual algorithm, and defining the dual solution (which is a lower bound on the optimal cost
cost shares  according to the corresponding dual solution, by weak LP duality). The algorithms of [1, 15] arise as a
would be enough to obtain a constant-factor approximation particular instantiation of the following algorithm. Our pre-
for MRoB. sentation is closer to [1], where the “reverse delete step” of
Unfortunately, we have found examples showing that Goemans and Williamson [15] is implicit; this version of
naive implementations of this idea cannot give  -strict cost- the algorithm is more suitable for our analysis.
 Our algorithm has a notion of time, initially 0 and in- current cluster containing it does not contain ti . This
creasing at a uniform rate. At any point in time, some de- implementation of the algorithm is equivalent to the
mands will be active and others inactive. All demands are algorithms of Agrawal et al. [1] and Goemans and
initially active, and eventually become inactive. The vertex Williamson [15].
set is also partitioned into clusters, which can again be ei-
ther active or inactive. In our algorithm, a cluster will be 2. Algorithm Timed(G; D; T ): This algorithm takes as
one or more connected components w.r.t. the currently built an additional input a function T : V ! R0 which
edges. A cluster is defined to be active if it contains some assigns a stopping time to each vertex. (We can also
active demand, and is inactive otherwise. view T as a vector with coordinates indexed by V .) A
Initially, each vertex is a cluster, and the demands are the vertex j is active at time  if j 2 D and   T (j ). (T
active clusters. We will consider different rules by which is defined for vertices not in D for future convenience,
demands become active or inactive. To maintain dual fea- but such values are irrelevant.)
sibility, whenever the constraint (3.3) for some edge e be-
tween two clusters S and S 0 becomes tight (i.e., first holds To get a feeling for Timed(G; D; T ), consider the fol-
with equality), the clusters are merged and replaced by the lowing procedure: run the algorithm GW(G; D) and set
cluster S [ S 0 . We raise dual variables of active clusters TD (j ) to be the time at which vertex j becomes inactive
until there are no more such clusters. during this execution. (If j 2
= D, then TD (j ) is set to zero.)
We have not yet specified how an edge can get built. To- Since the period for which a vertex stays active in the two
ward this end, we define a (time-varying) equivalence rela- algorithms GW(G; D) and Timed(G; D; TD ) is the same,
tion R on the demand set. Initially, all demands lie in their they clearly have identical outputs.
own equivalence class; these classes will only grow with The Timed algorithm gives us a principled way to essen-
time. When two active clusters are merged, we merge the tially force the GW algorithm to build additional edges: run
equivalence classes of all active demands in the two clus- the Timed algorithm with a vector of demand activity times
ters. Since inactive demands cannot become active, this rule larger than what is naturally induced by the GW algorithm.
ensures that all active demands in a cluster are in the same
equivalence class. The Algorithm PD: The central algorithm, Algorithm
We build edges to maintain the following invariant: the PD(G; D) is obtained thus: run GW(G; D) to get the time
demands in the same equivalence class are connected by vector TD ; then run Timed(G; D; TD )—the timed algo-
built edges. This clearly holds at the beginning, since rithm with the GW-induced time vector scaled up by a pa-
the equivalence classes are all singletons. When two ac- rameter  1—to get a forest FD . (We will fix the value
tive clusters meet, the invariant ensures that, in each clus- of later in the analysis.)
ter, all active demands lie in a common connected compo- We claim that the output FD of this algorithm is a feasi-
nent. To maintain the invariant, we join these two com- ble Steiner network for D. Intuitively, this is true because
ponents by adding a path between them. Building such Timed(G; D; TD ) only builds more edges than GW(G; D)
paths without incurring a large cost is simple but some- for  1. We defer a formal proof to the full version. We
what subtle; Agrawal et al. [1] (and implicitly, Goemans and now define the cost shares .
Williamson [15]) show how to accomplish it. We will not
repeat their work here, and instead refer the reader to [1]. The Cost Shares : For a demand j 2 D, the cost share
Remark 3.1 For the reader more familiar with the exposi- (D; j ) is the length of time during the run GW(G; D) in
tion of Goemans and Williamson [15], let us give an (in- which j was the only active vertex in its cluster. Formally,
formal) alternate description of the network output by the let a(j;  ) be the indicator variable for the event that j is the
algorithm given above. Specifically, we grow active clus- only active vertex in its cluster at time  ; then
R
(D; j ) = a(j;  ) d;
ters uniformly, and when any two clusters merge, we build
an edge between them. At the end, we perform a reverse- (3.4)
delete step—when looking at an edge e, if e lies on the path
between some x and y with (x; y ) in the final relation R,
where the integral is over the execution of the algorithm.
then we keep the edge, else we delete it. We assert that the In the sequel, we will prove our main technical result:
network output by this algorithm is the same as that of the Theorem 3.2 PD is a  = 2 -approximation for the
original algorithm. Steiner network problem, and  is a  = 6 =(2 3)-strict
Specifying the rule by which demands are deemed active cost-sharing method for PD.
or inactive now gives us two different algorithms: Setting = 3 then gives us a 6-approximation algorithm
that admits a 6-strict cost-sharing method, as claimed in
1. Algorithm GW(G; D): A demand si is active if the Theorem 2.4.
4 Outline of Proof of Theorem 3.2 4.1 Outline of Proof of Strictness
We first show that algorithm PD is a 2 -approximation al- We first recall some notation we will use often. The algo-
gorithm for Steiner forest. We begin with a monotonicity rithm PD(G; D) first runs GW(G; D) to find a time vec-
lemma stating that the set of edges made tight by algorithm tor TD , and then runs Timed(G; D; TD ) to build a for-
PD is monotone in the parameter . We omit its proof. est FD . Let (s; t) be a new source-sink pair. Define
Lemma 4.1 Let T and T 0 be two time vectors with T (j )  D0 = D [ f(s; t)g, D0 = D [ fs; tg, and TD0 by the time
T 0(j ) for all demands j 2 D. Then at any time  , the vector obtained by running GW(G; D0 ). Our sole remain-
set of tight edges in Timed(G; D; T ) is a subset of those in ing hurdle is the following theorem, asserting the strictness
Timed(G; D; T 0). of our cost-sharing method  for the algorithm PD.
Theorem 4.4 (Strictness) Let (s; t) be a source-sink pair
We can now outline the proof of the claimed approxima-
62 D, and let D0 = D + (s; t) denote the demand set ob-
tained by adding this new pair to D. Then the length of the
tion ratio.
Lemma 4.2 The cost of the Steiner forest FD con- shortest path dG=FD (s; t) between s and t in G=FD is at
(G; D ) is at most
P by algorithm PD for instance
structed
 is an optimal Steiner most  ((D0 ; s) + (D0 ; t)), where  = 6 =(2 3).
2 S yS  2 (F D ) , where F D
forest for (G; D). 4.1.1 Simplifying our goals
Proof: Let fyS g be the dual variables constructed The main difficulty in proving Theorem 4.4 is that the
by GW(G; D). If a( ) is the P numberRof active clusters at addition of the new pair (s; t) may change the behav-
time  during this run, then S yS = a( )d . ior of primal-dual algorithms for Steiner forest in fairly
Similarly, let a0 ( ) be the number of active clusters at unpredictable ways. In particular, it is difficult to ar-
time  during the execution of Timed(G; D; TD ), and gue about the relationship between the two algorithms we
fPyS0 g be the dual solution constructed by it. As above, care about: (1) the algorithm Timed(G; D; TD ) that gives
y 0 = R a0 ( )d . us the forest FD , and (2) the algorithm GW(G; D0 ) =
S S P
First, we claim that the cost of FD is at most 2 S yS0 . Timed(G; D0 ; TD0 ) that gives us the cost-shares. The dif-
This follows from the arguments in Agrawal et al. [1], ficulty of understanding the sensitivity of primal-dual algo-
since our algorithm builds paths between merging clusters rithms to small perturbations of the input is well known, and
as in [1]. has been studied in detail in other contexts by Garg [12] and
P
We next relate S yS0 to the cost of an optimal Steiner Charikar and Guha [9].
forest, via the feasible dual solution fyS g. Toward this end, In this section, we apply some transformations to the in-
we claim that a0 (  )  a( ). Indeed, let C1 ; : : : ; Ck be put data to partially avoid the detailed analyses of [9, 12].
the active clusters in Timed(G; D; TD ) at time  . Each In particular, we will obtain a new graph H from G
active cluster must have an active demand – let these be (with analogous demand sets DH and DH 0 ), as well as
j1 ; : : : ; jk . By the definition of TD , these demands must a new time vector TH so that it now suffices to relate
have been active at time  in GW(G; D), and by Lemma 4.1, (1’) the algorithm Timed(H; DH ; TH ) and (2’) the algo-
no two of them were in the same cluster at this time. With rithm Timed(H; DH 0 ; TH ). In the rest of this section, we
this claim in hand, we can derive will define the new graph and time vector; Section 4.1.2 will
Z Z Z X show that this transition is kosher, and then Sections 4.1.3
a0 ( )d = a0 (  )d  a( )d  yS : and 4.3 will complete the argument.
S
P A simpler time vector T : To begin, let us note that the
Since S yS is a lower bound on the optimal cost (FD ) time vectors TD and TD0 may be very different, even though
(by LP duality), the lemma is proved. the two are obtained from instances that differ only in the
Is is also easy to show that  is a cost-sharing method in presence of the pair (s; t). However, a monotonicity prop-
the sense of Definition 2.1. erty does hold till time TD0 (s) = TD0 (t), as the following
Lemma 4.3 The function  satisfies lemma shows:
X X X Lemma 4.5 The set of tight edges at time   TD0 (s) dur-
(G; D; (s; t)) = (D; j )  yS  (FD ): ing the run of the algorithm GW(G; D) is a subset of the set
(s;t)2D j 2D S of tight edges at the same time in GW(G; D0 ).
P Proof: Since   TD0 (s), both s and t are active at time
Proof: For all  , j a(j;  )  a( ), since each ac-
tive cluster can have at most oneR demand j with  in GW(G; D0 ). Any cluster that has not merged yet with
RP P non-zero
a(j;  ). Thus j a (j;  )d  a ( )d = S yS .
clusters containing s or t has the same behavior in both runs.
A cluster that merges with a cluster containing s or t will
continue to grow. So compared with GW(G; D), only more DH . The set DH0 is just DH [ fs; tg. Each demand
edges can get tight in GW(G; D0 ). j 2 DH0 has a set Cj  D0 of demands that were identi-
fied to form j . A new time-vector TH is defined by setting
Corollary 4.6 Let T be the vector obtained by truncating TH (s) = TH (t) = T (s) = T (t); furthermore, for j 2 DH ,
TD0 at time TD0 (s), i.e., T (j ) = min(TD0 (j ); TD0 (s)). we set TH (j ) = maxx2Cj T (x).
Then for all demands j 2 D, T (j )  TD (j ).
Going from Timed(G; D; TD ) to Timed(H; DH ; TH ):
Proof: If TD (s)  TD (j ), the claim clearly holds. If
0 Note that H was obtained from G by identifying some
TD (j ) < TD (s), then there is a tight path from j to its
0 demands; the edge sets of G and H are exactly the
partner at time TD (j ). But the monotonicity Lemma 4.5 same. We now show that the two instances in some
implies that these edges must be tight in GW(G; D0 ) at time sense behave identically. We denote the execution of
TD (j ) as well, and hence T (j ) = TD (j )  TD (j ).
0 Timed(H; DH ; TH ) by E .
The vector T is now a time vector for which we can
We first require another monotonicity lemma, whose
proof we omit.
say something interesting for both runs. Suppose we
were to now run the algorithm Timed(G; D; T ) as the Lemma 4.8 For all  1, the set of tight edges
second step of PD(G; D) (instead of Timed(G; D; TD ) in Timed(H; DH ; TH ) contains the tight edges
prescribed by the algorithm). By the monotonicity re- in Timed(G; D; T ). Similarly, the tight edges of
sults in Lemma 4.1 and Corollary 4.6, the edges that are Timed(H; DH0 ; TH ) contain those in Timed(G; D0 ; T ).
made tight in Timed(G; D; T ) are a subset of those in Lemma 4.9 The set of tight edges in the two runs
Timed(G; D; TD ). Hence it suffices to show that the Timed(G; D; T ) and E at any time  are the same.
distance between s and t in the forest resulting from
Timed(G; D; T ) is small; this is made precise by the fol- Proof: By Lemma 4.8, we know that the latter edges con-
lowing construction. tain the former; we just have to prove the converse. For a
demand j 2 DH , we claim that each demand x 2 Cj is in
A simpler graph H : Let us look at the equivalence re- some active cluster of Timed(G; D; T ) till time TH (j ).
lation defined by the run of Timed(G; D; T ) over the Suppose not: let x 2 Cj be in an inactive cluster at time
demands, which we shall denote by R. (Recall that for  < TH (j ). By the definition of the equivalence relation
j1 ; j2 2 D, (j1 ; j2 ) 2 R if at some time  during the R, no more demands are added to the equivalence class of
run, they are both active and the clusters containing them x (which is Cj ). But then maxy2Cj T (y)   < TH (j ),
meet. Equivalently, at some time  , both j1 and j2 are ac- a contradiction. Hence all demands in Cj are in active clus-
tive and some path between them becomes tight.) Simi- ters till time TH (j ). Thus, in the run GW(G; D; T ), we
larly, let the equivalence relation RD be obtained by run- had grown regions around these demands till time  , giving
ning Timed(G; D; TD ). Lemma 4.1 and Corollary 4.6 us the same tight edges as in E .
Corollary 4.10 For all  1, an active cluster at time
now imply that the former is a refinement of the latter, i.e.,
R  RD . (Note that this also means that the equiva-  during the run of E contains only one active demand. In
lence classes of RD can be obtained by taking unions of
the equivalence classes of R.)
particular, two active clusters never merge during the entire
run of the algorithm.
Since FD connects all the demands that lie in the same
equivalence class of RD , the fact that R  RD implies that Proof: Suppose j 6= j 0 2 DH are two active demands
it connects up all demands in the same equivalence class in in the same (active) cluster at time  . Let Cj and Cj 0 be
R as well. Hence, to show Theorem 4.4 that there is a short the corresponding sets of demands in D respectively. By
s-t path in G=FD it suffices to show the following result. Lemma 4.9, the set of edges that become tight at time  is
Theorem 4.7 (Strictness restated) Let H be the graph ob- the same as that in Timed(G; D; T ), and hence there must
tained from G by identifying all the vertices that lie in the be demands x 2 Cj ; x0 2 Cj 0 which lie in the same ac-
same equivalence class of R. Then the distance between s tive cluster of Timed(G; D; T ) at time  . By an argument
and t in H is at most  ((D0 ; s) + (D0 ; t)). identical to that in the proof of the previous lemma, there
must be active demands y 2 Cj and y 0 2 Cj 0 in the same
4.1.2 Relating the runs on G and H cluster as well. Hence Cj = Cj0 by the definition of the
We will need some new (but fairly obvious) notation: note equivalence class, contradicting that j 6= j 0 .
that each vertex in H either corresponds to a single ver- Note that this implies that the run E is very simple: each
tex in G, or to a subset of the demands D that formed demand j 2 DH grows a cluster around itself; though this
an equivalence class of R. The vertices of the latter type cluster may merge with inactive clusters, it never merges
are naturally called the demands in H , and denoted by with another active one.
Going from Timed(G; D0 ; TD0 ) to Timed(H; DH 0 ; TH ) : 4.1.3 The path between s and t
0 0
Recall that the cost-shares (D ; s) and (D ; t) are defined Lemma 4.14 The vertices s and t lie in the same tree in the
by running GW(G; D0 ) = Timed(G; D0 ; TD0 ) and using the Steiner forest output by E 0 .
formula (3.4). The following lemma shows that it suffices
to look instead at Timed(H; DH 0 ; TH ) in order to define
0 Proof: By the definition of T , s and t lie in the same
the cost-share. (Let E be short-hand for the execution cluster at time T (s) = T (t). Since TH (s) = TH (t) =
Timed(H; DH0 ; TH ).) T (s), applying Lemma 4.8 implies that s and t lie in the
Lemma 4.11 In E 0 , if s (respectively, t) is the only active same cluster in E 0 as well. Since they are both active at
vertex in its cluster at time  , then a(s;  ) = 1 (resp., the time their clusters merge, they lie in the same connected
a(t;  ) = 1) in the run Timed(G; D0 ; TD0 ). component of the Steiner forest.

Proof: Suppose a(s;  ) = 0 in Timed(G; D0 ; TD0 ), and This simple lemma gives us a path P between s and t
some active demand j 6= s lies in s’s cluster at time  . in H whose length we will argue about. Note that this path
By the definition of T , j and s are both active in the same is already formed at time st , and hence the rest of the ar-
cluster in Timed(G; D0 ; T ) as well. Now by Lemma 4.8, gument can (and will) be done truncating the time vector at
j and s must lie in the same cluster in E 0 too; furthermore, time st instead of at TH (s).
they must both be active (by the definition of TH ). This A useful fact to remember is that all edges in P must be
contradicts the assumption of the theorem. tight in the run of E 0 . The proof of the following theorem,

Corollary 4.12 Let alone(s) be the total time in the run E 0

along with Corollary 4.12, will complete the proof of The-
orem 4.7.
during which s is the only active vertex in its cluster, and
define alone(t) analogously. Then (D0 ; s) + (D0 ; t)  Theorem 4.15 (Strictness restated again) The length of
alone(s) + alone(t). the path P is bounded by  (alone(s) + alone(t)).
A property similar to Corollary 4.10 can also be proved We will prove the theorem with  = 6 =(2 3). Before
about the run E 0 . we proceed, here is some more syntactic sugar:
Given an execution of an algorithm, a layer is a tuple
Lemma 4.13 Let st be the time at which s and t become
(C; I = [1 ; 2 )), where C  V is an active cluster be-
part of the same cluster in E 0 . For  2, any active cluster
tween times 1 and 2 . If I = [1 ; 2 ) is a time interval, its
at time   st in E 0 contains at most one active demand
thickness is 2 1 . A layering L of an execution is a set
from DH , and at most one active demand from the set fs; tg.
of layers such that, for every time  and every active cluster
Proof: Let C be an active cluster in E 0 at time  with C , there is exactly one layer (C; I ) 2 L such that  2 I .
two active demands j; j 0 2 DH . If C contains neither s Given any layering L, another layering L0 can be ob-
nor t, then C is also a cluster (with two active demands) in tained by splitting some layer (C; I ) 2 L—i.e., replac-
E at the same time  , which contradicts the implication of ing it with two layers (C; [1 ; ^)); (C; [^  ; 2 )) for some
Corollary 4.10. ^ 2 (1 ; 2 ). Conversely, two layers with the same set C
Hence C must contain one of s or t (in addition to j; j 0 ); and consecutive time intervals can be merged. It is easy to
it cannot contain both since   st . Suppose it has s; we see that given some layering of an execution, all other lay-
claim that we can prove that j and j 0 must lie in the same erings of the same execution can be obtained by splittings
cluster in E at time 2  . (For this to make sense, we have and mergings.
to assume that  2.) To prove this, consider a tight path We fix layerings of the two executions E and E 0 , which
between s and j in E 0 at time  —since they both lie in the we denote by L and L0 respectively. The only property we
same cluster, such a path must exist. Since s has been active desire of these layerings is this: if (C; I ) 2 L and (C 0 ; I 0 ) 2
for time  , the portion of this path which is tight due to a s’s L0 with 1= I \ I 0 6= ;, then I = I 0 . (I.e., I = [ 1 ; 2 )
cluster has length at most  . Now consider the same path and I 0 = [1 ; 2 ).) It is easy to see that such a condition can
in E at time  — all but a portion of length at most  of be satisfied by making some splittings in L and L0 .
this path must already be tight. (Note that though we have Lemma 4.13 implies that each active layer (C; I ) 2 L0
dropped both s and t, we lose only  , since no part of the is categorized thus:
path could have been made tight by t’s cluster. If this were
to happen, then st <  , which is assumed not to happen.)  lonely: The only active demand in C is either s and t.
Hence the cluster containing j will contain s after another Assign the layer to that demand.
 units of time. The same argument can be made for j 0 . But
this would imply that j and j 0 would lie in the same cluster  shared: C contains one active demand from DH and
in E at time 2 , contradicting Corollary 4.10. one of s and t. The layer is assigned to the active de-
mand from DH .
  unshared: The only active demand in C is from DH . Proof: The proof of the claim is similar to that of
Again, the layer is assigned to that demand. Lemma 4.13; we just sketch the idea again. Consider a tight
path between j and s at time 1 ; at most a 1 portion of it
can be tight due to s. Hence by time 1  21 , the path
Note that the total thickness of the lonely layers is a lower
bound on alone(s)+ alone(t). Furthermore, since P consists
only of tight edges in E 0 , the length of P is exactly the total
must be completely tight.
thickness of the layers of L0 that P crosses. If L, S , and U
denote the total thickness of the alone, shared and unshared 4.2 Finally, the book-keeping
layers that P crosses, then
Note that both s and t both contribute separately to S and A
len(P) = L + S + U (4.5) till time st , and hence

Of course, any layer (C; I ) 2 L0 that crosses P must have S + L = 2 st : (4.6)
I  [0; st ℄, since the path is tight at time st . Hence their
corresponding layers have time intervals that lie in [0; st ℄. Let `0 = (C 0 ; I ) 2 L0 be an unshared layer that P
Note that the total thickness of the layers of L that P crosses, and let its corresponding layer be ` = (C; I ) 2 L.
crosses is a lower bound on its length. This suggests the If v 2 P \ C 0 is a vertex on the path, then v 2 P \ C by
following plan for the rest of the proof: for each shared or Lemma 4.16. If both s and t are outside C , then P crosses
unshared layer in L0 , there is a corresponding distinct layer C twice as well, and we get a contribution of 2 times the
in L that is times thicker. In an ideal world, each crossing thickness of `0 . Suppose not, and s 2 C or t 2 C : then we
of a layer in L0 would also correspond to a crossing of the lose times the thickness of ` for each such infringement.
corresponding layer in L; in this case, (S + U )  len(P), For shared layers `0 (which map to ` 2 L), Lemma 4.17
and hence L  1 len(P). Sadly, we do not live in an ideal implies that we lose only when s and t both lie inside `.
world and the argument is a bit more involved than this, Hence, if W denotes the wastage, the
though not by much.
Mapping shared and unshared layers of L0 : Each such
len(P)  (S + U ) W: (4.7)
layer `0 = (C 0 ; I ) is assigned to a demand j 2 DH \ C 0 .
Since j is active during the interval I in E , there must be
We bound the wastage in two ways: firstly, the wastage only
occurs when s or t lie in layers in L they are “not supposed
a layer ` = (C; I ) 2 L containing j —this is defined to be to”. However, s and t can only lie in one layer at a time,
the layer corresponding to `0 . (The properties of the layer- and hence they can only waste st amount each, which
ings ensure that the two time intervals are just rescalings of by (4.6) gives
each other by a factor .) Furthermore, since each layer in
L contains only one active vertex, the mapping is one-one. W  (S + L): (4.8)
It remains to show that crossings of P by layers is preserved
(approximately) by this correspondence. For another bound, let meet be the earliest time that s
Lemma 4.16 Each unshared layer `0 = (C 0 ; I = [1 ; 2 )) and t lie in the same layer in L (not to be confused with
is crossed by P either zero or two times. Furthermore, if its st , which is the earliest time they lie in the same layer in
corresponding layer is ` = (C; I ), then C 0  C . L0 ). We can be very pessimistic and claim that all the con-
tribution due to the unshared layers (i.e., U ) is lost. A
Proof: Since P begins and ends outside `0 , it must cross shared layer `0 = (C 0 ; I ) (containing s, say) mapped to
`0 an even number of times. Furthermore, P cannot cross ` = (C; I ) can incur loss only when t also lies in C —and
`0 more than twice—if it does so, there must exist vertices hence I  [meet ; st ℄. But now we claim that t must have
j1 , j2 , and j3 visited by P (in order), such that j1 ; j3 2 C 0 , been lonely. Indeed, suppose not, and that t was sharing
but j2 2 = C 0 . However, our algorithms ensure that if j1 layer `0 2 L0 assigned to j , and `0 was assigned to j 0 . But
and j3 lie in the same cluster C 0 , then the edges joining then both these layers would be mapped to `, contradicting
them lie within C 0 as well. This implies that there must be the fact that the mapping was one-one.
two disjoint paths between j1 and j2 , contradicting that the Hence we can charge the loss for each shared layer to a
algorithms construct a forest. lonely layer, and bound the wastage by:
For the second part, note that if C 0 is assigned to j and
does not contain s or t at time 2 , then the cluster containing W  (U + L): (4.9)
j at time 2 , i.e., C must contain C 0 .
Lemma 4.17 If `0 = (C 0 ; I = [1 ; 2 )) is a shared layer
Averaging (4.8) and (4.9), and plugging into (4.7), we get
containing s (resp., t) which assigned to j , and  2, then
its corresponding layer ` = (C; I ) contains s (resp., t).
len(P)  2 (S + U L:
2 ) (4.10)
Subtracting this from 2 times (4.5), we get [8] M. Bern and P. Plassman. The Steiner problem with edge lengths 1
and 2. Info. Proc. Let., 32:171–176, 1989.

alone(s) + alone(t)  L  3 2 len(P); (4.11)

[9] M. Charikar and S. Guha. Improved combinatorial algorithms for
the facility location and k -median problems. In 40th FOCS, pages
378–388, 1999.
proving Theorem 4.15 with  = 3 2 . The calculations of [10] C. Chekuri, S. Khanna, and J. S. Naor. A deterministic algorithm for
the cost-distance problem. In 12th SODA, pages 232–233, 2001.
the next subsection show how to improve this to  = 2 6 3 . [11] J. Fakcharoenphol, S. Rao, and K. Talwar. A tight bound on ap-
proximating arbitrary metrics by tree metrics. In 35th STOC, pages
4.3 Tightening the constants 448–455, 2003.
[12] N. Garg. A 3-approximation for the minimum tree spanning k ver-
One can get a better bound than given in (4.11). Suppose we tices. In 37th FOCS, pages 302–309, 1996.
were split each of the various quantities L; S; U; W in two
(L1 ; L2 etc.) to account for layers of L0 that correspond to
[13] N. Garg, R. Khandekar, G. Konjevod, R. Ravi, F. S. Salman, and
A. Sinha. On the integrality gap of a natural formulation of the
layers before and after time meet . Then, analogous to (4.8), single-sink buy-at-bulk network design formulation. In 8th IPCO,
we have volume 2081 of LNCS, pages 170–184, 2001.
[14] A. Goel and D. Estrin. Simultaneous optimization for concave costs:
W1  (S1 + L1 ) (4.12) Single sink aggregation or single source buy-at-bulk. In 14th SODA,

W2  (S2 + L2 )
pages 499–505, 2003.
(4.13) [15] M. X. Goemans and D. P. Williamson. A general approximation tech-

The reasoning of (4.9) implies W2  (U2 + L2 ). How-

nique for constrained forest problems. SIAM J. Comput., 24(2):296–
317, 1995.
ever, before time meet , we can do the following argument: [16] S. Guha, A. Meyerson, and K. Munagala. Hierarchical placement
each unshared layer gets mapped to a layer that can contain and network design problems. In 41st FOCS, pages 603–612, 2000.
only one of s or t, and hence loses at most one crossing with [17] S. Guha, A. Meyerson, and K. Mungala. A constant factor approxi-
mation for the single sink edge installation problems. In 33rd STOC,
P (while each shared layer loses no crossings at all), giving pages 383–388, 2001.
us that [18] A. Gupta, A. Kumar, J. Kleinberg, R. Rastogi, and B. Yener. Pro-

W1  U1 =2
visioning a Virtual Private Network: A network design problem for
(4.14) multicommodity flow. In 33rd STOC, pages 389–398, 2001.
W2  (U2 + L2) (4.15) [19] A. Gupta, A. Kumar, and T. Roughgarden. Simpler and better ap-
proximation algorithms for network design. In 35th STOC, pages
Multiplying (4.12-4.15) by 1
3
; 32 ; 23 ; 31 respectively, and
365–372, 2003.
[20] K. Jain and V. Vazirani. Applications of approximation algorithms to
adding, we get cooperative games. In 33rd STOC, pages 364–372, 2001.

W1 + W2 
[21] D. R. Karger and M. Minkoff. Building Steiner trees with incomplete
W = S S
[ 1+2 2+ L1 + 2L2 + U1 + U2 ℄: global knowledge. In 41st FOCS, pages 613–623, 2000.
3
[22] S. Khuller and A. Zhu. The general Steiner tree-star problem. Info.
Proc. Let., 84:215–220, 2002.
Now using the fact that S2  L2 , which follows from the [23] T. U. Kim, T. J. Lowe, A. Tamir, and J. E. Ward. On the location of
basic argument behind (4.9), and algebra, we get alone(s)+
a tree-shaped facility. Networks, 28(3):167–175, 1996.

alone(t)  L  2 6 3 len(P), as desired. [24] A. Kumar, A. Gupta, and T. Roughgarden. A constant-factor approx-
imation algorithm for the multicommodity rent-or-buy problem. In
43rd FOCS, pages 333–342, 2002.
References [25] A. Meyerson, K. Munagala, and S. Plotkin. Cost-distance: Two met-
[1] A. Agrawal, P. Klein, and R. Ravi. When trees collide: an approx- ric network design. In 41st FOCS, pages 624–630, 2000.
imation algorithm for the generalized Steiner problem on networks. [26] A. Meyerson, K. Munagala, and S. Plotkin. Designing networks in-
SIAM J. Comput., 24(3):440–456, 1995. crementally. In 42nd FOCS, pages 406–415, 2001.
[2] M. Andrews and L. Zhang. Approximation algorithms for access [27] R. Ravi and F. S. Salman. Approximation algorithms for the traveling
network design. Algorithmica, 34(2):197–215, 2002. purchaser problem and its variants in network design. In ESA ’99,
volume 1643 of LNCS, pages 29–40. 1999.
[3] B. Awerbuch and Y. Azar. Buy-at-bulk network design. In
[28] F. S. Salman, J. Cheriyan, R. Ravi, and S. Subramanian. Approx-
38th FOCS, pages 542–547, 1997.
imating the single-sink link-installation problem in network design.
[4] B. Awerbuch, Y. Azar, and Y. Bartal. On-line generalized Steiner SIAM J. on Opt., 11(3):595–610, 2000.
problem. In 7th SODA, pages 68–74, 1996. [29] C. Swamy and A. Kumar. Primal-dual algorithms for the connected
[5] Y. Bartal. On approximating arbitrary metrics by tree metrics. In facility location problem. In 5th APPROX, volume 2462 of LNCS,
30th STOC, pages 161–168, 1998. pages 256–269, 2002.
[6] Y. Bartal, M. Charikar, and P. Indyk. On-line generalized steiner [30] K. Talwar. Single-sink buy-at-bulk LP has constant integrality gap.
problem. In 5th SODA, pages 43–52, 1997. In 9th IPCO, volume 2337 of LNCS, pages 475–486, 2002.
[31] H. P. Young. Cost allocation. In R. J. Aumann and S. Hart, editors,
[7] Y. Bartal, A. Fiat, and Y. Rabani. Competitive algorithms for dis-
Handbook of Game Theory, volume 2, chapter 34, pages 1193–1235.
tributed data management. J. Comput. System Sci., 51(3):341–358,
North-Holland, 1994.
1995.