On the irreducible core and the equal remaining obligations rule of minimum cost spanning extension problems

by Vincent Feltkampi
Maastricht School of Management, POBox 1203, 6201BE Maastricht, The Netherlands

Stef Tijs
CentER and Econometrics Department, Tilburg University, P.O. Box 90153, 5000 LE Tilburg, The Netherlands,

and Shigeo Mutoii

Dept of Value and Decision Science, Graduate School of Decision Science and Technology, Tokyo Institute of Technology, 2-12-1 Oh Okayama, Meguro-ku, Tokyo 152-8552, Japan.

95-96?

Abstract Minimum cost spanning extension problems are generalizations of minimum cost spanning tree problems in which an existing network has to be extended to connect users to a source. This paper generalizes the denition of irreducible core to minimum cost spanning extension problems and introduces an algorithm generating all elements of the irreducible core. Moreover, the equal remaining obligations rule, a one-point renement of the irreducible core is presented. Finally, the paper characterizes these solutions axiomatically. The classical Bird tree allocation of minimum cost spanning tree problems is obtained as a particular case in our algorithm for the irreducible core.

Introduction

Consider a group of villages, each of which needs te be connected directly or via other villages to a source. Such a connection needs costly links. Each village could connect itself directly to the source, but by cooperating costs might be reduced. This cost minimization problem is an old problem in Operations Research, and Boruvka (1926) came up with algorithms to construct a tree connecting everybody to the source with minimal total cost. Later, Kruskal (1956), Prim (1957) and Dijkstra (1959) found similar algorithms. A historic overview of this minimum cost spanning tree (henceforth mcst) problem can be found in Graham and Hell (1985). In this paper, we analize a more general problem, in which a partial network of links exists already and has to be extended to a network connecting every player to the source. However, nding a minimal cost spanning extension is only part of the problem: if the cost of this extension has to be borne by the villages, then a cost allocation problem has to be addressed as well. Claus and Kleitman (1973) introduced this cost allocation problem for the original minimum cost spanning tree setting, whereupon Bird (1976) treated this problem with game-theoretic methods and proposed a cost allocation closely related to the Prim-Dijkstra algorithm. Granot and Huberman (1981) proved that this allocation is an extremal point of the core of the associated minimum cost spanning tree game. This game is dened as follows: the players are the villages and the worth of a coalition is the minimal cost of connecting this coalition to the source via links between members of this coalition. Aarts (1992) found other extreme points of the core in case the mcst problem has an mcst that is a chain, i.e. a tree with only two leaves. Kuipers (1993) investigated the core of information games. These are games arising from mcst problems in which the costs of a connection are either one or zero.
i Email: vincent.feltkamp@bigfoot.com. This author was sponsored during the major part of this research by the Foundation for the Promotion of Research in Economic Sciences, which is part of the Nehterlands Organization for Scientic Research (NWO). ii This author wishes to acknowledge the Canon Foundation in Europe Visiting Research Fellowship which made his visit to CentER possible.

The outline of this paper is as follows. Section 2 presents a formal model of the minimum cost spanning extension (mcse) problem in which an existing network has to be extended to a spanning network, i.e. a network connecting every village to the source. An algorithm to nd a minimum cost spanning extension and an associated set of cost allocations is presented. In contrast to earlier work (see Feltkamp, Tijs and Muto (1994a)), this extension and set of allocations are not generated by an algorithm ` a la Prim-Dijkstra, but by an algorithm similar to Kruskals (1956) algorithm. It is proved that the set of allocations generated is a subset of the core of the associated mcse game and that is is independent of the extension that is constructed. Section 3 generalizes the denition of the irreducible core, proposed in Bird (1976) for minimum cost spanning tree problems, to minimum cost spanning extension problems and proves that the set of allocations generated by the algorithm in section 2 coincides with the irreducible core. A corollary is that Birds tree allocations (see Bird (1976)) for minimum cost spanning tree problems are also generated by our algorithm. Section 4 introduces the equal remaining obligations (ERO) rule, a one-point renement of the irreducible core. It is obtained as a special case of the algorithm for the irreducible core presented in section 2. Like the irreducible core, the ERO rule is independent of the extension constructed. This in contrast with Birds tree allocation rule, which is dependent on the tree constructed. Section 5 axiomatically characterizes the irreducible core and the equal remaining obligations value. Among others, axioms we use are eciency, consistency and converse consistency. Finally, section 6 concludes with some remarks and suggestions for further research. The proofs of the main theorems of section 2 are provided in an appendix. Preliminaries and notations We recall some standard denitions from graph theory which can be found in any elementary textbook on graph theory to show the notational conventions. A graph G =< V, E > consists of a set V of vertices and a set E of edges. An edge e incident with two vertices i and j is identied with {i, j }1 . For a graph G =< V, E > and a set W V , E (W ) := {e E | e W } is the set of edges linking two vertices in W and for a subset E of E , V (E ) := {v V | there exists an edge e E with v e} is the set of vertices incident with E . The complete graph on a vertex set V is the graph KV =< V, EV >, where EV := {{v, w} | v, w V and v = w}. A path from a vertex i to a vertex j in a graph < V, E > is a sequence (i = i0 , i1 , . . . , ik = j ) of vertices such that for all l k , the edge {il1 , il } lies in E . A cycle is an edge-disjoint path whose begin and end points coincide. Two vertices i, j V are connected in a graph < V, E > if there is a path from i to j in < V, E >. A subset W of V is connected in < V, E > if every two vertices i, j W are connected in the subgraph < W, E (W ) >. A connected set W is a connected component of the graph < V, E > if no superset of W is connected. We will usually say component when we mean connected component. If W V , the set of connected components of the graph < W, E (W ) > is denoted W/E . A connected graph is a graph < V, E > with V connected in < V, E >. A tree is a connected graph that contains no cycles. A connected component of a graph will be denoted by the letter C , and for a vertex v of the graph, the connected component containing v is denoted Cv . The economic situations in the sequel involve a set N of users of a source . For a coalition S N , we denote S {} by S . For two vectors x I RS and y I RT , where S and T are two disjoint coalitions, we denote (x, y ) the vector with coordinates { xk if k S (x, y )k = yk if k T. Furthermore, for two coalitions S T and a vector x I RT , we denote xS the restriction of x to S . The symbol N 1S is used to denote the vector in I R with coordinates { 1 if k S 1S,k = 0 if k N \ S. For any coalition S , the simplex S is dened by S := {x I RS + |
1 This

iS

xi = 1}.

can be done because we do not consider multigraphs: two vertices are connected by at most one edge.

With many economic situations in which costs have to be divided one can associate a cost game (N, c) consisting of a nite set N = {1, . . . , n} of players, and a characteristic function c : 2N I R, with c() = 0. Here c(S ) represents the maximal cost for coalition S if it secedes, i.e. if people of S cooperate and do not count upon help from people outside S . The core Core(c) of a cost game (N, c), is dened by Core(c) = {x I RN | xi = c(N ) and xi c(S ) for all S N }.
iN iS

The cardinality of a set S will be denoted by |S |.

Mcse problems: a solution

In this section we formally present minimum cost spanning extension problems, mcse games and an algorithm which for any mcse problem computes a minimum cost spanning extension and an associated set of allocations, which appears to be contained in the core of the mcse game. A minimum cost spanning extension problem consists of a set N of users who have to extend an existing network in order to be connected to a source, denoted . The links are costly, and the users have to pay for the extension. Such a problem is represented by a complete graph < N , EN > on the set N containing all users and the source, together with a set E EN of already constructed links and a weight function w : EN I R+ . The cost of constructing an edge e is given by the positive weight w(e) > 0 of this edge. Because the graph of possible edges is always the complete graph, we denote a mcse problem with set of users N , source , weight function w and existing edge set E by < N, , w, E >. If the set of existing edges E is empty, the mcse problem becomes the classical minimum cost spanning tree problem, and instead of writing < N, , w, >, we will write < N, , w >. Mcse problems can be split up into two subproblems, an Operations Research problem of connecting all users to the source in a extended graph < N , E E > such that the cost of the extension E is minimal and a cost allocation problem of allocating this cost to the users in a reasonable way. In the special case of mcst problems, an mcst is constructed by Kruskals algorithm, which is dened as follows: Algorithm 1 (Kruskal 1956) input : an mcst construction problem < N, , w > output : an edge set E of a minimum cost spanning tree 1. start with the empty set E = . 2. Find an edge e of minimum cost such that the graph < N , E {e} > does not contain a cycle. 3. Join this edge to the set E (E := E {e}). 4. If the graph < N , E > is not connected, go back to step 2. 5. E is the edge set we were seeking. For mcse problems, a generalization of Kruskals algorithm is demonstrated in example 2. Example 2 Let N = {1, 2, 3, 4, 5}, and let the weights of the edges and the graph which is already constructed be as in gure 1. The costs of edges that are not indicated are $200. First construct the edge {1, 2}, it is the cheapest one which does not introduce a cycle. The same reasoning picks {2, 3} as second edge, as third edge {3, 4} and nally as last edge, {2, } is constructed. The algorithm demonstrated is the following: Algorithm 3 (Kruskal generalized to mcse problems) input : an mcse construction problem < N, , w, E > output : an mcse 1. Given M < N, , w, E >, dene t E0 = 0 the initial stage, = |N /E | 1 the number of stages, = the initial extension.

1 u A A A $10 A

5 u H $70 HH HH Hu4 $40 u 3

1 u

5 u H HH HH Hu4

A Au 2

$20

u 3

$100 u The weights of links. u {4, 5} is already constructed.

Figure 1: A simple mcse problem 2. t := t + 1. 3. At stage t, given E t1 , choose a cheapest edge et E E t1 such that the graph < N , E E t1 {et } > does not contain more cycles than the graph < N , E E t1 >. 4. Dene E t := E t1 {et }. 5. If t < , go to step 2. 6. E is the extension we were seeking. Denote the sequence of edges constructed by E = (e1 , . . . , e ). In lemma 35 it is proved that the extension constructed is indeed a mcse. Note this algorithm constructs the edges of a minimum cost spanning extension in the order of non-decreasing costs. Because the costs of the edges in any minimum cost spanning extension are the same, it follows that any sequence E = (e1 , . . . , e ) which consists of the edges of a minimum cost spanning extension ordered by non-decreasing costs can be constructed with algorithm 3 by a suitable choice of the edges chosen in step 3. Bird (1976) associates a tree allocation with every minimum cost spanning tree in an mcst problem. In Feltkamp, Tijs and Muto (1994a), we prove that this tree allocation can be associated with the sequence of edges which Prim and Dijkstras algorithm generates when generating this mcst. This suggests looking for allocations associated with the sequences generated by Kruskals algorithm. For an mcse problem M < N, , w, E >, the minimal cost spanning extensions E have one edge less than the number of components |N /E | (if more edges were built, a cycle would be introduced, which cannot be minimal in cost, as the weights of the edges are positive). Hence, dening := |N /E | 1, we associate an allocation with every sequence E = (e1 , . . . , e ) of edges which does not introduce new cycles in the mcse problem M. Note that any such sequence connects all players to the source. The idea behind the allocation is that at each successive stage t, the cost of the edge et which is constructed at stage t is shared among the players in N according to a fraction vector f t N . Three rules will be observed when allocating the cost of et : At stage t, the edge et connects two components of the graph < N , E {e1 , . . . , et1 } > creating a component C t of the graph < N , E {e1 , . . . , et } >. Only players in C t contribute to the cost of et .
t 1 The component C of the source in the graph < N , E {e1 , . . . , et1 } > constructed before stage t does not contribute to the cost of et .

Furthermore, summing over all edges in the sequence, every component of the original graph < N , E > which does not contain the source pays fractions of edges to a total of one. Hence, dene the set V E (M) of sequences of fraction vectors valid for the sequence E in M by

5 E V (M) := (f 1 , . . . , f ) [0, 1] |N | .

k C t kC s=1

t fk

= and

for all t

s fk

t fk

1 C N /E with C and t 1 = 0 for all t, k s.t. k C =

For a sequence F (f 1 , . . . , f ) valid for E in M, dene the allocation xE ,F (M) := and dene the set DE (M) by
t=1

f t w(et ) I RN

(2.1)

DE (M) := {xE ,F | F V E }.

(2.2)

If no confusion can occur, we drop the argument M. The intuition behind the rst rule is that at the moment et is constructed, we do not know yet who is going to be connected through et to the source. It can be either one of the components being connected, depending on how the later edges are chosen. We also dont know yet which player outside C t will need et . Because we want to allocate its cost immediately, it is allocated to the players in C t . The intuition behind the two other rules is obvious. Lemma 4 For all mcse problems M, for all sequences E = (e1 , . . . , e ) such that E := {e1 , . . . , e } is an mcse of M and all F valid for E , the allocation x := xE ,F (M) is ecient: iN xi = eE w(e). Proof : Validity of F implies every edge et E is paid for by the component C t it constructs. Note that because the set of valid sequences of fraction vectors V E is convex and the map xE : F xE ,F is linear, the set DE is also convex, for any sequence E . Instead of rst constructing the edges and later allocating their cost, one could allocate the cost of the edge et immediately, because the validity of a sequence of fraction vectors can be checked stage by stage: a sequence f 1 , . . . , f is valid for e1 , . . . , e in M if and only if at every stage t it satises the component C t constructed at stage t pays the cost of the edge et , for every component in the original graph < N , E >, the total of the fractions paid up to stage t does not exceed 1, the people outside C t pay nothing,
t 1 the people in the component C of the source pay nothing.

In formula, this gives

t fk k C t t s fk kC s=1 t fk t fk

= = =

1, 1 0 0 for all C N /E, if k C t , t 1 if k C (2.3)

for every stage t. Example 5 For an mcst problem T < N, , w >, Prim and Dijkstras algorithm (cf Prim (1957) or Dijkstra (1959)) constructs a sequence E = (e1 , . . . , e|N | ) of edges leading to an mcst < N , T > as follows: at every stage t, et is an edge which connects a player with the component of the source in the graph < N , {e1 , . . . , et1 } > and which has minimal cost among all such edges. Without loss of generality, we number the players in N such that for every t, the edge et connects player t with the component of the source,

6 {, 1, . . . , t 1}. Hence, the edge e1 connects player 1 to the source and using the system 2.3, we see player 1 has to pay the cost of e1 and the other players do not contribute. In the second stage, player 2 is connected to the component of the source, which now equals {, 1}. The rst equation in system 2.3 implies players 1 and 2 pay the cost of edge e2 . The fourth implies that player 1, who is in the component of the source, does not contribute, hence, player 2 is assigned the cost of edge e2 . The third equation implies the other players do not contribute. The inequality is satised, because up to now, every component in the original graph (i.e. every player) paid either one edge or no edges. By induction, we see that at every stage, the component C t t 1 consists of the component C and the newly connected player t. Because the component of the source does not contribute to the cost of et , the unique valid allocation of the cost of this edge is to allocate it completely to player t. Hence, DE (T ) consists of one allocation, in which each player i is allocated the cost of the edge incident to i on the unique path in the tree from i to the source. This allocation is precisely Birds tree allocation associated with the mcst < N , T >, denoted T (cf Bird (1976)).

Example 6 Computing the extreme points of the set of valid fraction vectors for the sequence E constructed in example 2 shows that in this case DE (M) is the convex hull of the vectors (10, 20, 40, 100, 0), (10, 40, 20, 100, 0), (20, 10, 40, 100, 0), (40, 10, 20, 100, 0), (10, 20, 40, 0, 100), (10, 40, 20, 0, 100), (20, 10, 40, 0, 100), (40, 10, 20, 0, 100), (10, 20, 100, 40, 0), (10, 100, 20, 40, 0), (20, 10, 100, 40, 0), (100, 10, 20, 40, 0), (10, 20, 100, 0, 40), (10, 100, 20, 0, 40), (20, 10, 100, 0, 40), (100, 10, 20, 0, 40).

Inspired by Bird (1976), we associate a minimum cost spanning extension game (N, cM ) with an mcse problem M < N, , w, E > as follows. Each coalition S N , if it cannot count on the players in its complement, has to solve a problem similar to the problem of the grand coalition, namely, extending the existing graph to a graph connecting all users in S to the source. The cost of this extension is the worth cM (S ) of coalition S in the mcse game. When computing the cost of a coalition S , several questions arise. Can the coalition use all or some of the edges which are already present? Is it allowed to use vertices outside S ? We opt for the following answers: a coalition S is allowed to use all edges which are initially present, but can only use those vertices which lie in a component of < N , E > which contains members of S or the source. Let us consider an example to clarify what we mean. 2 u HH $1 HH 1 3 H u Hu J
$4 J
$3 J
$3 $1 J
J
Ju
Figure 2: {1, 2} is allowed to use the edge {2, 3}, but {1} is not. Example 7 In the problem depicted in gure 2, edge {2, 3} is already constructed. Coalition {1, 2} is allowed to use the edge {2, 3} and can connect itself by building the edges {1, 2} and {3, }, so c({1, 2}) = 1 + 1 = 2. Coalition {1} is not allowed to use the edge {2, 3} because the component {2, 3} does not have any vertices in common with {1} or the source, hence c({1}) = 3. The other worths are c({2}) = c({3}) = 1 (connect player 2 via player 3), and nally c({1, 3}) = c(N ) = 2. In general, the formula becomes { } E S C and E contains only edges between cM (S ) := min w(e) | components containing members of S
e E

7
E for all S N , where C is the component of the source in the graph < N , E E >. The next theorem states an allocation associated with a sequence of edges generated by algorithm 3 is a core element of the mcse game.

Theorem 8 For any mcse problem M, for any sequence of choices E = (e1 , . . . , e ) in the algorithm 3 applied to M and any sequence of fractions F valid for E the allocation xE ,F , as dened in equation 2.1, is a core-allocation of the mcse game (N, cM ) associated with M. The proof of this theorem is lengthy and technical, and can be found in the appendix. An immediate consequence is Corollary 9 For any sequence E of edges ordered according to non-decreasing costs leading to a minimum cost spanning extension for an mcse problem M with associated mcse game (N, cM ), DE (M) Core(N, cM ). Proof : For any x DE (M), there is a sequence F which is valid for E with x = xE ,F (M) Core(N, c). 2

A question which arises is, how does the set DE depend on the sequence E ? It is answered in the next proposition. Proposition 10 For any mcse problem M, for any E and E constructed by the algorithm 3 applied to M, DE (M) = DE (M). A proof can be found in the appendix. Because DE is independent of the sequence E of edges, as long as this sequence is constructed by the algorithm 3, we dene for an mcse problem M: DGK (M) := DE (M) for any sequence E obtained by the algorithm 3 applied to M. (The superscript GK stands for generalized Kruskal).

The irreducible core of mcse problems

In this section we generalize the concept of irreducible core, known from mcst problems, to mcse problems and prove that our set of allocations DGK coincides with this irreducible core. Denition 11 Given an mcse problem < N, , w, E >, we dene the associated mcst problem < NE , E , wE > as follows: NE consists of the components of < N , E > which do not contain the source , the new source E is the component of < N , E > which contains the original source and wE is dened by wE (C, D) := min{w(i, j ) | i C, j D} for all components C and D of the graph < N , E >. Furthermore, for an edge e = {i, j } EN , dene eE := {Ci , Cj } and for a set of edges F EN , dene FE := {{Ci , Cj } | {i, j } F }, where Ci and Cj are the components of < N , E > containing players i and j , respectively. (3.2) (3.1)

The intuitive idea is to shrink each component not containing the source into a single player, and to shrink the component of the source into a new source. It is easy to see that if F is an mcse of the mcse problem < N, , w, E >, then the tree < NE , FE > is an mcst of the associated mcst problem < NE , E , wE >. Conversely, if < NE , T > is an mcst of the associated mcst problem, then there exists an mcse F with FE = T . This correspondence, though possibly not one to one, transfers the well-known structure of the collection of mcs trees of an mcst problem onto the set of mcs extensions of an mcse problem. More about this structure will be said in the appendix, in the proof of proposition 10. We now dene the irreducible core of an mcse problem. It is a straightforward generalization of the denition of irreducible core of an mcst game provided in Bird (1976). Apparently, it depends on an mcse. However, this is not the case, as we will prove later. Denition 12 Given an mcse problem M = < N, , w, E > and an mcse E , dene the irreducible core IC(M, E ) of M with respect to E as follows: consider the set Var(M, E ) of all mcse problems obtained from M by varying the weight w(e) of edges e E E , that still have E as mcse. Now IC(M, E ) is the intersection of the cores of all mcse games associated with an mcse problem in Var(M, E ), i.e. } { IC(M, E ) := Core(N, cM ) | M Var(M, E ) . If the set E of initially present edges is empty, the present denition coincides with the denition of irreducible core of an mcst problem in Bird (1976). For mcst problems, it is already known that the irreducible core is independent of the mcst used to dene it. Equivalently, one could dene the irreducible core as follows: given an mcse problem M = < N, , w, E > and an mcse E , for any two players i, j N , let Pij be a path in the graph < N , E E > from i to j . It is possible that this path is not unique, but the part of the path in E is. Dene a new weight function { max{w(e) | e Pij E } if Ci = Cj (3.3) w(i, j ) := w({i, j }) otherwise where Ci and Cj are the components of i and j , respectively, in the graph < N , E >. Then the following holds: Lemma 13 The irreducible core IC(M, E ) of a mcse problem M = < N, , w, E > coincides with the core of the game (N, c) = (N, c<N,,w,E> ). be a Proof : First, E is an mcse of the problem < N, , w, E >. This can be seen as follows. Let E spanning extension, which is cheaper than E . Order the edges in E by increasing weight and take the rst \ E . Adding e to < N , E E > creates a cycle formed by e and the edges in Pij . Hence, edge {i, j } = e E , such that substituting e with e in E yields another spanning extension there exists an edge e Pij E \ E of the problem < N, , w, E >. Now e connects two components of < N , E >; hence w(e) w(e ) for all edges e Pij E . This spanning extension is thus an mcse, which, moreover, has one more edge in common with E . Repeat this argument enough times to see that E is indeed an mcse of the problem < N, , w, E >, which implies that < N, , w, E > Var(M, E ). Hence, the irreducible core of M is included in the core of the game (N, c). On the other hand, note that for an edge e = {i, j } E connecting two components Ci and Cj , max{w(e ) | e Pij E } is a lower bound on the weight that can be assigned to this edge e in an mcse problem M Var(M, E ): if a lower weight is assigned to e, then there is an edge e Pe E with higher weight than e and replacing e by e would yield a spanning extension E \ {e } {e} which has lower cost than E . Hence, w(e) w (e) for all edges e EN , for every problem < N, , w , E > Var(M, E ). This implies that c(S ) cM (S ) for all coalitions S and for all mcse problems M in Var(M, E ). Because E is an mcse of M (N ). Hence, Core(c) Core(cM ) for all < N, , w, E >, it also follows that c(N ) = e E w ( e ) = c M Var(M, E ), which implies that the core of c is included in the irreducible core of M. Together with the rst part of the proof, this proves the theorem. 2 In order to analyze the structure of the irreducible core of an mcse problem, we relate its structure to the structure of the irreducible core of the associated mcst problem of denition 11, using the concept of marionettes. Zumsteg (1992) dened two players i, j in a game (N, c) to be marionettes if c(S {i}) = c(S {j }) = c(S {i, j }) for all S N . Considering players to be marionettes of themselves turns being marionettes into an equivalence relation, we denote it by . For any player i, the set of marionettes of i is denoted by Si .

9 Denition 14 For a game (N, game (N , c ) is the game in which N = {Si | i N }, c), the marionette-reduced and which satises c (C ) = c( S C S ) for all C N . Hence a player in the marionette-reduced game consists of all marionettes of one player in the original game. Equivalently, one could obtain the marionette-reduced game as a subgame of the original game: for each player U N , take one representative player jU U and dene T = {jU | U N }. Dene the subgame (T, cT ) by cT (U ) = c(U ) for all U T . For every player i in T , there is a unique player Si in N satisfying jSi = i and for every player U in N , there is a unique player jU in T satisfying SjU = U . Furthermore this bijection between the players of T and N turns out to be an isomorphism between the games (N , c ) and (T, cT ): c (C ) = c( S ) = c({jS | S C }) = cT ({jS | S C })
S C

c (U )

c(U )

c(

iU

Si )

c ({Si | i U })

for all coalitions C N and U T . Lemma 15 If (N, c) is a game, and (N , c ) is its marionette-reduced game, their cores are related as follows: 1. if x Core(N, c), then y Core(N , c ), where y I RN is dened by yS = iS xi for all S N . 2. if y Core(N , c ) I RN RN + , then x Core(N, c), for all x I + satisfying yS = Moreover, such an x exists.

iS

xi for all S N .

Proof : The proof of part 1 is trivial. To prove part 2, take y Core(N , c ). We rst prove an x that satises the requirements exists. For all S N , choose a representative player iS S and assign { yS if i = iS for an S N , xi := 0 otherwise. Because y is non-negative, so is x. Now x(N ) = x({iS | S N }) = y (N ) = c (N ) = c(N ) and for any coalition T N , there exists a subset T of T , such that iT Si = iT Si , where the right-hand side is a disjoint union. Hence, c(T ) = c( Si ) = c( Si ) = c ({Si | i T }) iT iT ySi = x(Si ) = x( Si )
iT iT iT

x(T ),

which implies x satises the requirements. Now take any x I RN + satisfying yS = iS xi for all S N . Then x(N ) =
S N

x(S ) =

S N

yS = y (N ) = c (N ) = c(N )

and for any coalition T , take again a subset T such that iT Si = iT Si , where the right-hand side is a disjoint union. Then because x is non-negative and y Core(N , c ), x(T ) xj = x(Si ) = ySi c ({Si | i T }) = Hence, x Core(N, c).
iT j Si iT

c(T )

iT

c(T ) 2

Lemma 16 For a given mcse problem M = < N, , w, E > with associated minimum-cost spanning tree problem T = < NE , E , wE >, the mcst game (NE , cT ) associated with T coincides with the marionette-reduction (N , c ) of the mcse game (N, cM ).

10 Proof : We rst prove that the sets NE and N coincide. Two players that are in the same component of < N , E > are marionettes in the mcse game: if either one is connected to the source, so is the other, so that the cost of connecting one is the cost of connecting both. Hence, a player in NE , being a component of < N , E >, is a coalition of marionettes of the mcse game. On the other hand, if two players are marionettes in the mcse game, it means that connecting one of them to the source is as costly as connecting the other player or connecting both players. Because the cost of all edges is positive, it follows that both players must lie in the same component. Hence the set of players NE in the mcst game coincides with the set of players N in the marionette-reduction of the mcse game. Consider a coalition C N . By denition of the associated mcst problem, cT (C ) = cM ( S ).
S C

As for all S C , all players in S are marionettes in the mcse game, it follows that cM ( S ) = c (C ).
S C

Hence cT (C ) = c (C ), which concludes the proof.

The next proposition states the relation between the irreducible core of an mcse problem and the associated mcst problem. Proposition 17 For an mcse problem M = < N, , w, E > and an mcse E , the irreducible core IC(M, E ) satises IC(M, E ) = {x I RN xi S NE } + | y IC(< NE , E , wE >) where yS :=
iS

Proof : This follows easily from lemmata 15 and 16 and the fact that an mcse is transformed into an mcst by the transition from mcse problem to associated mcst problem. 2 Corollary 18 The irreducible core of an mcse problem is independent of the mcse used to dene it. Accordingly, we will denote the irreducible core of an mcse problem M by IC(M). Having given some properties of the irreducible core, we now proceed to prove that it coincides with the set of allocations generated by algorithm 3. Lemma 19 Let M < N, , w, E > be an mcse problem. Then DGK (M) IC(M). Proof : In section 2 we stated that for an mcse problem M < N, , w, E >, the set DGK (M) of allocations generated by the algorithm 3 applied to M are core allocations of the associated game (N, cM ). Since to prove this in section 7, we only use the weights of edges in the mcse E , which is an mcse in all mcse problems M Var(M, E ), it holds that DGK (M) is a subset of the core of any mcse game associated with an mcse problem in the set Var(M, E ). Hence, DGK (M) IC(M). (3.4) 2 In order to prove the reverse inclusion, we need the following lemma. Lemma 20 Let T = < N, , w > be an mcst problem and let < N , T > be an mcst for T . Then Birds tree allocation T lies in the set DGK (T ). Proof : The number of edges in T equals n := |N |. Consider any sequence E = (e1 , . . . , en ) of edges obtained by ordering the edges of T by non-decreasing cost. Dene F = (f 1 , . . . , f n ) by { 1 if et is the rst edge on the path in < N , T > from i to the source, fit := 0 otherwise.

11 It follows that F V E (T ) and that T = xE ,F DE = DGK (T ). 2 Theorem 21 Let T be an mcst problem. Then DGK (T ) = IC(T ). Proof : It follows from lemma 19 that we only have to prove IC(T ) DGK (T ). Bird (1976) proved IC(T ) is the convex hull of the set of all Bird allocations of the mcst problem T , with reduced weight function dened by equation 3.3. By proposition 20, these Bird allocations lie in DGK (T ). Now DGK (T ) equals DGK (T ), because the set DGK is obtained by only considering the weights of edges in an mcst, and these edges have the same weight in T as in T (cf. Aarts and Driessen (1993)). Moreover DGK (T ) is convex, hence IC(T ) = conv hull{ T | T is an mcst of T } DGK (T ) = DGK (T ). 2 Corollary 22 Let M be an mcse problem. Then DGK (M) = IC(M). Proof : Let x IC(M), and let T =< NE , E , wE > be the mcst problem associated with M. We know by proposition 17 that the vector y I RNE , dened by yS = iS xi for all S NE , lies in IC(T )(= DGK (T )). Hence, there exists a sequence E = (e1 , . . . , e ) of edges leading to an mcst of T and a sequence F = (f 1 , . . . , f ) of fraction vectors valid for E , such that y = xE ,F . Now for each edge et = {Ci , Cj }, there exists an edge e t with same weight in the weighted graph < N , EN , w >, which connects the components Ci and Cj . Hence, 1 , . . . , f ) by E = ( e1 , . . . , e ) is a sequence leading to an mcse of M. Dene F = (f { t = f i
t xi fC i yCi 0

if yCi > 0 if yCi = 0.

for all t and all i N , where Ci is the component containing i. Then F is valid for E . Moreover, for any ,F player i, xE = 0 if yCi = 0, but then also xi = 0, and if yCi = 0, then i
,F xE = i t=1 t fC i xi xi t yC = f = xi i = xi . yCi yCi t=1 Ci yCi

Hence, x = xE ,F DGK (M), which completes the proof.

The equal remaining obligations value

In most cases, the irreducible core of an mcse problem M contains a continuum of allocations. If the objective is to choose a division of the cost, one might be better o with a one-point solution. The equal remaining obligations value (henceforth ERO value), suggested by Jos Potters, is a one-point renement of the irreducible core and is constructed by the following extension of algorithm 3. Algorithm 23 (Equal remaining obligations solution) input : an mcse problem < N, , w, E > output : an mcse and the ERO value 1. Given M < N, , w, E >, dene t E0 2. t := t + 1. = 0 the initial stage, = |N /E | 1 the number of stages, = the initial extension.

12 3. At stage t, given E t1 , choose a cheapest edge et E E t1 such that the graph < N , E E t1 {et } > does not contain more cycles than the graph < N , E E t1 >.
t1 t 1 t1 4. If C t = Ci Cj is the connected component just formed by connecting the components Ci and t 1 t1 t t t Cj of the graph < N , E > with the edge e = {i, j }, dene the vector f = (fk )kN of fractions by

t fk =

1 t1 |Ci | 1 t1 | |Cj 1 t1 |Ci | 1 t1 |Cj |

1 |C t | 1 |C t |

t1 if k Ci and C t , t1 if k Cj and C t , t1 t1 if k Ci and Cj , t1 t1 if k Cj and Ci ,

otherwise.

5. Dene E t := E t1 {et }. 6. If t < , go to step 2. 7. E is the mcse we sought. As before, denote E = (e1 , . . . , e ) the sequence of edges constructed. 8. Dene EROE (M) :=
t=1

f t w ( et ) .

Applied to the mcse problem of example 2, this algorithm generates successively edge {1, 2}, of which players 1 1 1 2 1 1 1 and 2 each pay 1 2 =2 , edge {2, 3}, of which player 3 pays 1 1 3 = 3 and players 1 and 2 each pay 2 3 = 6 , 1 1 2 1 1 3 edge {3, 4}, of which players 1, 2 and 3 each pay 3 5 = 15 while players 4 and 5 pay 2 5 = 10 and nally 1 edge {2, }, to which each player contributes 5 . This yields the allocation ERO(M) = 1 (101, 101, 116, 96, 96). 3

Generically, the choice of edge in step 3 is unique but even in the case that the sequence E is not uniquely dened, this algorithm yields only one allocation, independently of the choice of edges made. This in contrast with Birds tree allocation rule, that may associate a dierent allocation with each mcst of an mcst problem. Proposition 24 For any two sequences of edges E and E chosen by the algorithm 23 applied to an mcse problem M, EROE (M) = EROE (M). The proof is similar to the proof of proposition 10, so we do not give it. This proposition allows us to dene ERO(M) := EROE (M) for any sequence E constructed by algorithm 23. Clearly, the fractions constructed are valid for the edges constructed; the ERO solution is thus a renement of the irreducible core. To see that the ERO value deserves its name, dene for an mcse problem M the initial obligation oi of a player i by { 1 if Ci |C i | oi := (4.1) 0 if Ci where Ci is the component of < N , E > containing player i. For a sequence F = (f 1 , . . . , f ) of fraction vectors, after a stage t a player i N has paid st fis , while the initial obligation was oi . Hence player is remaining obligation ot i satises 1 ot fis = ot fit . (4.2) i = oi i
st

13

Theorem 25 The algorithm 23 has the property that after each stage t, in each component C of the graph < N , E E t >, every player k in the component C has the same remaining obligation { 1 if C, t |C | ok = (4.3) 0 if C. Proof : The proof goes by induction on the stage t. 1. After stage zero, for k in a component C of < N , E >, { 1 t |C | ok = ok 0 = 0

if C, if C.

t 1 t1 2. Suppose equation 4.3 holds after stage t 1. Let C t = Ci Cj be the connected component formed t1 t1 of the graph < N , E E t1 > with the edge and Cj at stage t by connecting the components Ci t 1 t et = {i, j }. Let Ck and Ck be the components of player k in < N , E E t > and < N , E E t1 >. Then

ot k

t 1 t = ok fk {

= =

00

1 t 1 |Ck |

t fk

t1 if k C t1 if k C t 1 if k Ci and C t t 1 if k Cj and C t t 1 t 1 if k Ci and Cj

1 t 1 |Ci | 1 t 1 |Cj | 1 t 1 |Ci |

( |C t1 1 |
i

1 |C t | ) 1 |C t | )

( |C t1 1 |
j

1 t1 |Ci |

t 1 t 1 1 1 if k Cj and Ci t 1 t1 |Cj | |Cj | 1 t 1 if k C t C t 1 |Ck | t 1 0 if k C 1 if k C t |C t | t 1 t 1 0 if k Ci and Cj t 1 t 1 0 if k Cj and Ci = 1 t 1 if k C t C t 1 |Ck | 0 t 1 if Ck 1 t |C t | if Ck k = t 0 if Ck . Hence equation 4.3 holds after stage t as well. This completes the proof. 2

Axiomatic characterizations

In sections 2 and 4 we introduced the irreducible core and the equal remaining obligations value for mcse problems. We axiomatically characterize these rules in this section. In contrast with the characterizations in Feltkamp, Tijs and Muto (1994a) and Feltkamp, Tijs and Muto (1994b), here, a solution consists only of a set of allocations. This is possible because both the irreducible core and the ERO rule are independent of the set of edges constructed. M (N ). In An allocation of an mcse problem < N, , w, E > is a vector x I RN + which satises iN xi c eect, an allocation is a vector that allocates at least the cost of a minimum cost spanning extension to the players. Properties that an allocation x of an mcse problem < N, , w, E > can satisfy are

14

Denition 26 E x is ecient if
iN

xi = cM (N ).

MC x has the minimal contribution property if every component that does not contain the source contributes at least the cost of a minimum cost edge that connects two components. In formula: for each component C N /E that does not contain the source, xi min{w(e) | e connects two components of < N , E >}.
iC

FSC x has the free for source component property if xi = 0 for all i in the component of the source in the graph < N , E >. The minimal contribution property and the free for source component are motivated as follows: every component that has to be connected to the source has to contribute at least the cost of an edge, and the component of the source should not contribute, because it is already connected. A solution of mcse problems is a map assigning to every mcse problem a set of allocations. Denition 27 NE A solution is said to be non-empty if (M) = for all M. We will say a solution is ecient, satises the minimal contribution property or the free for source component property if for all M all elements of the solution (M) satisfy the corresponding property.

Denition 28 Given an mcse problem M < N, , w, E > and an edge e = {i, j } E that connects two components of < N , E >, dene the edge-reduced mcse problem Me =< N, , w, E {e} > . Note that the edge-reduced problem is indeed a simpler problem than the original problem: less edges have to be constructed. However, it has the same number of players as the original problem. The next three properties use edge-reduced mcse problems, with as extra edge an edge which constructs a new component if it is adjoined to the graph < N , E >, and which has minimum cost among all edges with this property. Denition 29 Loc is local if for all M, for all x (M), for all minimum cost edge e that when added to < N , E > constructs a new component C , there exists an x I RC such that ( x, xN \C ) (Me ). In eect, this axiom requires that when a minimum cost edge is added, this has no inuence on the allocation to players that are not in the component constructed by adding this edge. ECons is minimum cost edge consistent if for all M < N, , w, E >, for all x (M), for each minimum cost edge e that when added to < N , E > constructs a new component C , for each C satisfying w(e) xC , it holds that x xe, (Me ), where xe, := (w(e), 0N \C ). This axiom means that when a minimum cost edge is added, the savings obtained by not having to construct this edge can be allocated arbitrarily over the players that are in the component constructed by adding this edge. Obviously, edge consistency implies locality.

15 CECons is converse minimum cost edge consistent if for all M < N, , w, E >, for every minimum cost edge e that when added to < N , E > constructs a new component C , for every x I RN that satises a MC, b FSC, c w(e) xC implies x xe, (Me ) for all C , it holds that x (M). This axiom requires that if adding an allocation to the solution does not destroy the MC, FSC and ECons properties, then it should be part of the solution. In eect, it requires the solution to be the largest solution that satises the other properties.

Lemma 30 The irreducible core satises the properties NE, MC, E, FSC, ECons and CECons. Proof : Because of the coincidence of the irreducible core with the set DGK , the irreducible core is non-empty for any mcse problem: there are always valid sequences of fraction vectors for any sequence of edges constructed by the algorithm 3. It satises the minimum contribution property because every component that does not contain the source has to pay for fractions of edges that total one, so it contributes at least the minimum cost of an edge that connects two components. That it is ecient is proved in lemma 4. By construction, it is clear that DGK satises FSC. To prove edge consistency, take an mcse problem M and suppose x IC(M) = DGK (M). For each minimum cost edge e that when added to < N , E > constructs a new component C , there exists a sequence E = (e1 , e2 , . . . , e ) starting with the edge e1 = e, that is constructed by the algorithm 3. Because the set DGK (M) is independent of the sequence of edges constructed, there exists a sequence F = (f 1 , . . . , f ) V E 1 := fi1 for all i C . Then such that x = xE ,F . For any C satisfying w(e) xC , dene 1 C by i 1 i i = 0, iC 1 1 ( ) w ( e ) f t w(et ), t=2 f t w(et ) 0 for t 2. Hence there exist vectors 2 , . . . , I RC satisfying t i fit t = 0 i 1 1 ( )w(e ) =
iC t=2 t

for t 2 and for all i C, for t 2, (5.1)

t

w(e ).

E.g. t the projection of f t on the hyperplane with coordinates zero, along the line through the points take s s 1 1 s=2 f w (e ) and ( )w (e ), i.e. t C f i i =f ( f s w(es ) ( 1 )w(e1 )) s s iC s=2 fi w (e ) s=2
t t

for all t 2. Rewriting the last equation of system (5.1), we obtain w(e) =
t=1

t w ( et ) .

It follows that x xe, = t=2 (f t t )w(et ). The rst two equations of system (5.1) imply that the sequence (f 2 2 , . . . , f ) is valid for (e2 , . . . , e ). Hence, x xe, Me . Because edge consistency implies locality, the irreducible core satises locality as well. To prove that the irreducible core satises CECons, take an mcse problem M, take a minimum cost edge e that when added to < N , E > constructs a new component C , take an x I RN that satises

16

a MC, b FSC, c w(e) xC implies x xe, IC(Me ) for all C . We have to prove that x IC(M). Denote by C1 and C2 the two components that are joined by e. The allocation x satises FSC; hence if one of these components (say C1 ) contains the source, xi = 0 for i C1 and an C with xC satises i = 0 for i C1 . For such an (which exists), there exists a sequence (e2 , . . . , e ) constructed by the algorithm 3 2 2 2 applied to the problem Me and a sequence (f 2 , . . . , f ) V (e ,...,e ) such that x xe, = x(e ,...,e ),(f ,...,f ) . So with f dened by { k if k C fk := 0 otherwise for all k N , it holds that (f, f 2 , . . . , f ) is valid for the sequence (e, e2 , . . . , e ) and x = x(e,e
2

,...,e ),(f,f 2 ,...,f )

(M).

If neither of the two components contain the source, then by the minimal contribution property, both components contribute at least w(e) in the allocation x. On the other hand, both together contribute x(C ), so there exists an a1 [0, 1], such that { x(C1 ) = a1 w(e) + (1 a1 )(x(C ) w(e)) x(C2 ) = (1 a1 )w(e) + a1 (x(C ) w(e)) Dene C , by i = { a1 xi /x(C1 ) (1 a1 )xi /x(C2 ) if i C1 , if i C2 .
2

C Then, w(e) x Hence, there exists a mcse {e2 , . . . , e } and a sequence (f 2 , . . . , f ) V (e ,...,e ) ( M) such . t 1 t 2 t 1 t = a by a , . . . , a and a f = (1 a ) by a , . . . , a that x xe, = t=2 f t w(et ). Dene a2 2 2 2 1 1 1 iC f i . iC i Then t for all t 2 fit at 1 + a2 = i C 1 1 = 1 a1 + fit at 1 = a + (1 a ) t=2 t=2 iC 1 t 1 1 1 a + a = 1 a + a fit = 1 2 t=2 t=2 iC

Dening g = (, 0

N \C

) and
t gi

{ =

at 1 xi /x(C1 ) at 2 xi /x(C2 )

if i C1 if i C2

we see that (g 1 , . . . , g ) V (e,e

,...,e )

(M). Furthermore, (1 a1 )(x(C ) w(e)) = (1 a1 ) fit w(et )

t=2 iC

x(C1 ) a1 w(e) =

= This implies that for i C1 , xi = = =

2

t=2

t at 1 w (e ).

xi x(C1 ) x(C1 )
xi t (a1 w(e) + at 1 w (e )) x(C1 ) t=2 1 gi w ( e)
1

t=2

t gi w(et ).

Similarly for i C2 . Hence, x = x(e,e ,...,e ),(g ,...,g irreducible core satises converse edge consistency.

is in the irreducible core of M. This implies that the 2

17

Lemma 31 If a solution of mcse problems satises MC, FSC and ECons, and a solution of mcse problems satises NE, FSC and CECons, then (M) (M) for each mcse problem M. Proof : We prove the lemma for any mcse problem M by induction on the number of components of the graph < N , E >. First, consider an mcse problem M < N, , w, E >, where the graph < N , E > is connected. Then by FSC, x = 0 for any x (M) and for any x (M). By NE, there has to be an x (M), hence (M) {0} = (M). Now suppose the lemma holds for every mcse problem M with p 1 components in the graph < N , E >. Take an mcse problem M such that < N , E > has p components and take x (M). By ECons of , for any minimum cost edge that constructs a new component C when added to < N , E >, for each C satisfying w(e) xC , it holds that x xe, (Me ) (Me ), where the inclusion holds by the induction hypothesis. Because x satises MC, it holds that x (M) by CECons of . Hence (M) (M). 2 Theorem 32 The unique solution of mcse problems that satises NE, MC, FSC, ECons and CECons is the irreducible core. Proof : By proposition 30, the irreducible core has these properties. By lemma 31, if two solutions have these properties, they contain each other and hence they coincide. 2 To characterise the ERO value, we need some other properties that a solution can have. Denition 33 ET a solution satises equal treatment if for every mcse problem M, for all x (M), for each component C of the original graph < N , E >, and for all pairs of players i and j C , xi = xj . IPCons A solution is inversely proportional consistent if for every mcse problem M < N, , w, E >, for every minimum cost edge e that when added to < N , E > connects two components C1 and C2 , neither of which contains the source, for all x (M), there exists an x (Me ) such that |C1 | (xi x i ) = |C2 | (xi x i ).
iC1 iC 2

Equal treatment is motivated by the idea that players in the same component need the same connections, hence should be allocated the same amount. Inversely proportional consistency means that if two components are connected by a minimum-cost edge, they contribute amounts to the cost of this edge which are inversely proportional to their number of elements. This is motivated by the idea that a bigger component has apparently already constructed more edges in E , so should pay less for the edge under consideration. Proposition 34 The unique solution of mcse problems that satises NE, FSC, Loc, E, ET and IPCons is the ERO value. Here, the ERO value is identied with the solution that assigns the singleton {ERO(M)} to every M. Proof : First we prove that the ERO value satises the required properties. That the ERO value satises the properties NE, MC, FSC, Loc and E is a consequence of its being a renement of the irreducible core. That is satises equal treatment is also easy to see. To prove it satises IPCons, take an mcse problem M, and a minimum cost edge connecting the components C1 and C2 , neither of which contains the source, into a component C . Then there exists a sequence of edges E = (e = e1 , . . . , e ) a sequence of fractions F = (f 1 , . . . , f ) constructed by the algorithm 23 such that ERO(M) = xE ,F . Moreover, by denition of the 2 2 algorithm, x(e ,...,e ),(f ,...,f ) = ERO(Me ). Because e connects two components that do not contain the source, 1 1 |C w(e)( |C | ) if k C1 , 1| (e2 ,...,e ),(f 2 ,...,f ) ,F 1 1 |C w(e)( |C xE xk = k | ) if k C2 , 2| 0 if k C.

18

Then

kC1

(ERO(M) ERO(Me )) = 1
kC2

|C1 | |C2 | = |C | |C | |C1 | . |C |

and similarly,

(ERO(M) ERO(Me )) =

Hence |C1 |

kC1

(EROk (M) EROk (Me )) =

|C1 ||C2 | = |C2 | (EROk (M) EROk (Me )). |C |

kC2

To prove uniqueness, suppose a solution satises these six properties. We prove (M) = {ERO(M)} by induction on the number of components of the graph < N , E >. Let < N , E > have one component. By FSC, xi (M) = 0 = EROi (M) for all i N and all x (M). Suppose (M) = {ERO(M)} for all mcse problems M such that < N , E > has less than p components. Consider an mcse problem M such that < N , E > has p components. Take a minimum cost edge e that connects two components C1 and C2 into a new component C in < N , E >. By ET of applied to M and Me , for all x (M), for all x (Me ), for all i, j C1 we have xj x j = xi x i =: 1 (x, x ) and for all i, j C2 we have xj x j = xi x i =: 2 (x, x ).

Moreover, by FSC of , if C1 contains the source, then 1 (x, x ) = 0 and by locality and eciency, there exists w ( e) e an x (M ) such that 2 (x, x ) = |C2 | . If C2 contains the source, then similarly one proves that 1 (x, x ) and 2 (x, x ) are uniquely determined. If neither of C1 and C2 contain the source, then by IPCons, there exists an x (Me ) such that 2 (x, x ), 1 (x, x ) = |C2 | |C1 |
kC1 k C 2

and by locality and eciency,

iC1

1 (x, x ) +

iC2

2 (x, x ) =

(xi x i ) = w(e).
iC

Hence, 1 (x, x ) = w ( e) |C1 |2 and 2 (x, x ) = w(e) . |C2 |2

So whether C1 or C2 contain the source or not, the numbers 1 (x, x ) and 2 (x, x ) are uniquely determined and independent of x and x . Now the ERO value also satises the six properties and so has these same numbers 1 (x, x ), 2 (x, x ) characterizing the dierence between ERO(M) and ERO(Me ). The induction hypothesis then implies x 1 (x, x )1C1 2 (x, x )1C2 = x = ERO(Me ) = ERO(M) 1 (x, x )1C1 2 (x, x )1C2 and so x = ERO(M). 2

Concluding remarks and suggestions for further research

New in this paper is the integrated approach : we solve the problem of constructing an optimal network and at the same time allocate the costs. Because of this integrated view of the problem, the dierent algorithms to construct a minimum cost spanning tree suggested several closely related algorithms for solutions to the cost allocation problem. For instance, Prim and Dijkstras algorithm appeared to be closely linked to Birds tree allocations. This suggested to look for allocation rules that are related to Kruskals (1956) algorithm for constructing an mcst. In the present paper we relate the irreducible core to Kruskals algorithm, and moreover provide a one-point renement, the equal remaining obligations solution. A third algorithm for constructing an mcst is known, and in Feltkamp, Tijs and Muto (1994b) we associate an allocation rule with this algorithm. Second, instead of looking only at the extreme case where no edges are present at the beginning and a spanning network has to be constructed, we consider problems where some network can be present already, construct a minimum cost spanning extension and prove that the cost of this extension can be allocated in a a stable way. This has two advantages. The mathematical advantage is that a half-solved problem is again

19

in the same class of problems, which allows for a recursive solution, the advantage from an applied point of view is that not only problems in which all edges have yet to be constructed are treated, but also extensions of networks can be solved. If the original problem was suggested by among others electrication of Moravia at the beginning of the century, by now the problem is rather how to extend an already present network and allocate the cost of the extension. Third, we provide axiomatic characterizations of the irreducible core and the equal remaining obligations solution. These axiomatic characterizations enable one to select an allocation rule, based on the properties the rule should have. Another way of approaching mcse problems is to dene the cost of a coalition S as the minimal cost of an extension connecting S to the source, without any restriction on the vertices of the extension. This approach yields a monotonic game, with the same cost for the grand coalition, but smaller costs for the other coalitions, because these now have more opportunities to save costs. Hence, in general the core of this variant is contained in the core of the mcse game we dened. However, such a monotonic game associated with an mcse problem < N, , w, E > can be considered as an mcse game according to our denition associated to the mcse problem < N, , w , E > where the weights of links have been reduced to satisfy the triangle inequality w ({i, j }) w ({i, k }) + w ({k, j }) for all i, j, k N .

So this does not introduce new games. Moreover, core elements of the monotonic game can be computed by applying algorithm 3 to the problem with reduced weights. A second possible variation is not to allow a coalition to use any players in its complement when connecting to the source. This would yield a game with the same cost for the grand coalition, but larger costs for the other coalitions. The core of this variant contains the core of our mcse game and so all algorithms presented in this paper yield core elements of the variant. Because the set DGK and the ERO value are not dened in function of a game but rather from the mcse problem itself, they are independent of the game which one chooses. Moreover, for both variants, the irreducible core coincides with the set DGK . All allocations introduced here lie in the irreducible core of the mcse problems. In order to get the whole core of the corresponding games, one needs to use weights of edges that are not used in any minimum cost spanning extension. In general, it is still an open problem to compute the whole core of an mcse game directly from the weights of the edges, even if the attention is restricted to mcst games. Faigle, Kern, Fekete, and Hochst attler (1996) proved that it is an N P -hard problem to decide whether a given vector is a member of the core. Because of the polynomial equivalence of membership testing and optimizing linear functions relative to a polyhedron (see Gr otsel, Lovasz, and Schrijver (1988)), it is therefore doubtful whether there exists ecient ways to obtain the core of a mcst game. Finally, it would be interesting to nd non-cooperative games of which equilibria sustain the cooperative solutions presented here. Acknowledgements : Inspiring discussions with Harry Aarts, Peter Borm, Ruud Jeurissen, Michael Maschler, Gert-Jan Otten, Jos Potters and Hans Reynierse about the subject of the paper are gratefully acknowledged.

Appendix

In this appendix, we prove theorem 8 and proposition 10. First, we need a few lemmas. Lemma 35 For all sequences E chosen by algorithm 3 the constructed extension E is an mcse for the problem M < N, , w, E >. Proof : There are + 1 = |N /E | components in < N , E >, at each stage two components are connected, so after stage , the resulting graph is connected and no new cycles have been introduced. Assume that the with |E | = , such that extension E constructed is not minimal in cost, i.e. there exists a set of edges E > is a connected graph, and < N , E E w(e) < w(e). (7.1)
e E e E

ordered by non-decreasing weight. Equation 7.1 Let the sequence E = ( e1 , . . . , e ) consist of the edges in E t t implies there exists a smallest t such that e = e for 1 t < t and et = e t . Because et is a minimum t1 1 weight edge that does not introduce a cycle in < N , E E > =< N , E { e ,...,e t1 } > and e t does not

20 introduce a cycle in < N , E E t1 >, it follows that w(et ) w( et ). Consider the end points i and j of et . >, hence there exists a path from i to j in < N , E E >. But They have to be connected in < N , E E not all edges in this path can be present in the graph < N , E E t1 >, otherwise et would introduce a cycle. \ E t1 in this path that comes later in E than e Hence there is an edge e E t , so e costs at least w( et ), which t t is at least w(e ). Now E := (E \ {e}) {e } is a spanning extension of < N , E > such that E does not cost , and E has one edge more in common with E . Repeating this process enough times shows that more than E . This is a contradiction, hence the assumption that the algorithm 3 does not E does not cost more than E lead to an mcse is wrong. 2 In order to prove that DE is a subset of the core of the associated mcse game if E is constructed by algorithm 3, we need to compare the outcome of the algorithm 3 applied to related mcse problems. Suppose we have an mcse problem M < N, , w, E > and an edge e = {i, j } connecting the component C := E {e}. Consider of the source with the component Cj of some player j in the graph < N , E >. Dene E the mcse problem M = < N, , w, E >. Distinguish the graphs, components, edges and allocations used in algorithm 3 applied to the problems M and M by giving those in the latter problem a tilde. With this setup, we prove two lemmata and a corollary, which we need to prove theorem 8. Lemma 36 For every sequence of choices E = (e1 , . . . , e ) in the algorithm 3 applied to M, one can nd an s such that the sequence E ( e1 , . . . , e 1 ) := (e1 , . . . , es1 , es+1 , . . . , e ), obtained by deleting the edge es from E , is a sequence of edges that can be obtained by algorithm 3 applied to M and that satises
t t t )) \ {C t } and C t C t = C t for all t < s, 1. (N /(E E t )) \ {Ci , Cj } = (N /(E E i i j i E t > have the same components, exthat is, as long as t < s, the graphs < N , E E t > and < N , E t E t >. cept for the components of i and j in < N , E E >, which are connected to each other in < N , E

t1 ) for all t {s, . . . , }, 2. N /(E E t ) = N /(E E t 1 > that is, after stage s, the components of < N , E E t > coincide with the components of < N , E E at the previous stage. E 0 = E = E {e} = E E 0 {e}, hence Proof : We prove the statements by induction on t. For t = 0: E t t t t t t t t = C C and (N /(E E )) \ {C , C } = (N /(E E )) \ {C }. Hence case 1 holds. C i i j i j i t t1 . Two cases can occur. If case 1 holds at stage t 1, look at the eect of adding e to E E t1 >. Now et a et = e and adding the edge et does not introduce a cycle in the graph < N , E t1 is a cheapest edge which does not introduce a cycle in < N , E E > and any edge which does not E t1 >, does not introduce a cycle in < N , E E t1 >. Hence et is also introduce a cycle in < N , E E t1 >. This means et is a legitimate choice a cheapest edge that does not introduce a cycle in < N , E t for e . Consequently, case 1 still holds at stage t. E t1 >. This means b et = e or adding the edge et does introduce a cycle in the graph < N , E t 1 t1 t1 t1 t1 t t1 and e connects the components Ci and Ci of < N , E E >. Then Ci = Ci Cj =C i t 1 t 1 t the other components are unchanged, so (N /(E E t )) \ {Ci } = (N /(E E t1 )) \ {Ci , Cj } = t1 t1 t t 1 (N /(E E )) \ {Ci }. Hence N /(E E ) = N /(E E ), and case 2 holds for stage t.
t

Suppose case 2 holds for stage t 1. Then the edge et = {k, l} is a legitimate choice for e t1 (it does t 1 t not introduce a cycle, and has minimal cost among the edges satisfying this). Hence, Ck = Ck Clt1 = t2 C t2 = C t1 , which implies N /(E E t ) = N /(E E t1 ), and case 2 holds for stage t as well. C k l k Hence, if the rst stage at which case 2 holds is called s, we see that case 1 holds for t < s and case 2 holds for t s. E t1 ), hence case 2 holds at stage , so s . Note that N /(E E ) = {N } = N /(E 2 Lemma 37 Let M and M be as above, let E be a sequence of choices made by algorithm 3 applied to M, and k )kN , let F be valid for E . Let the sequence E and the stage s be as dened in lemma 36. Then there exist (t 1 , . . . , f 1 ) dened by such that the sequence F = (f t k , s} fk if t < min{t t k < s fk if t = t t t = tk fk := for all t {1, . . . , 1} and all k N t+1 fk if tk t s k 0 if t > t

21 is valid for E . In formula: F V E (M). Proof : For all k N , dene E t >}, k := min{t | there exists a path from k to in < N , E t t is the edge set resulting in stage t in the algorithm 3 applied to M. Note that t k = 0 for all k Cj , where E t the component of < N , E > connected to C by the edge e. Hence fk = 0 for all stages t if k Cj . This means the players in Cj do not contribute to any edge. We now prove the lemma in three steps. t = C t1 C t1 be the component in the graph < N , E E t > 1. For t 1, let e t = {k, l} and let C k l t t formed by the addition of e . Then mC t fm = 1. To prove this, we distinguish several cases:
t1 E t >. Then t m = t k > t for all Suppose e t is not incident to C , the component of in < N , E t m C , hence t = f t for m C t . Moreover, C t = C t , the component in the graph < N , E E t > if t < s then f m m t t formed by the addition of e = e . Hence, t t = f fm =1 m t m C m C t

by the assumption on F . t = f t+1 for m C t . Moreover, e t = C t+1 , the component in the if t > s then f t = et+1 and C m k t+1 t+1 graph < N , E E > formed by the addition of e . Hence, t+1 t = f fm =1 m
t m C mC t+1

by the assumption on F .
t 1 t1 m = t l = t . This means t . Then one of k and l, say k , lies in C Suppose that e t is incident to C t 1 t 1 for all m C and t < t for all m C . Hence, m l k t = 0 for m C t = f t for m C t1 and f t1 . This implies if t < s, then f m m t =t m l k t t = f f m m t m C t1 m C l

t fm t 1 t =1

t1 t =t m C l

t fm

t fm ).

t1 t =1 m C l

t1 = C t1 is a union of a number, say p, of components C1 , . . . , Cp of the graph < N , E >. Now C l l Remember that for all q {1, . . . , p}:
mCq t =1 t fm = 1,

hence

t fm =

p q =1 mCq t =1

t fm =p

(7.2)

p and as Clt1 = q=1 Cq contains exactly those players that contributed to the p 1 edges in t 1 {e1 , . . . , et1 } that connect the components (Cq )p , q =1 into Cl
t1 t fm = p 1.

t1 t =1 m C l

(7.3)

t1 t =1 m C l

Equations 7.2 and 7.3 imply

t m C

t = p (p 1) = 1. f m

22 t = f t+1 , e t = C t+1 , the component in the graph < N , E t+1 > if t s, then f t = et+1 and C m m t+1 formed by the addition of e . Hence, t t+1 fm =1 fm =
t m C mC t+1

by the assumption on F . ) that does not contain : C is also a component of the graph < N , E > 2. For each component C (N /E k = t l for all (because < N , E > and < N , E > dier only in the component of the source). Moreover, t k, l C and k < s for all k C , then if t
1 kC t=1

t f k

= = = =

k t

t f k

kC t=1 k 1 t

k t + f t ( f k k ) (
t fk + k t=t t fk )

kC t=1 k 1 t kC t=1 kC t=1

t fk

= 1. t = 0 for t > t k , the third equality because of the denition of The rst equality follows because f k tk fk , and the fth by the assumptions on F . E s1 > = < N , E E s >. Hence, k s for k C , then C is not connected to in < N , E if t s s s . This implies fk =0 according to F , nobody in C contributes to e , which is an edge incident to C for all k C . But then k t t = t f f
k k kC t=1

= =

kC t=1 s1

t fk

k t t=s

t+1 fk )

kC t=1 k +1 t kC t=1

t fk

= 1.
t 1 t = 0 E t1 >. Hence, f k t if t 1 and k C 3. Furthermore, t , the component of in < N , E k by denition.

Steps 1, 2 and 3 imply F V E (M).

Corollary 38 Let M and M be as above, let E be a sequence of choices made by algorithm 3 applied to M, and let F be valid for E . Then
,F ,F xE (M) xE (M) k k

for all k N,

where E is as dened in lemma 36 and F in lemma 37.

,F Proof : xE (M) = k

t k
t=1

t w( f et ). k
k t t=1 ,F t w( f et ) = 0 x E (M) k k

k = 0 then If t

23 k < s (with s as dened in lemma 36) then If 0 < t

k t t=1

t w( f et ) = k = = =

k 1 t t=1 k 1 t t=1 k 1 t t=1

k t w( t f et ) + f e tk ) k k w (

t fk w(et ) + (

t fk )w(etk )

t fk w(et ) +

k t= t k t= t

t fk w(et )

t=1

t fk w(et )

,F xE (M). k

The second equation follows from the denition of F and the inequality holds because E is ordered by non-decreasing weights.
s1 s s k s then k is not contained in C If t = C , so k is not allowed to contribute to es , i.e. fk = 0. Hence, k t t=1

t w( f et ) k

= = =

s1 t=1

t fk w(et )

k t

t+1 fk w(et+1 ) t fk w(et )

s1 t=1 t=1

t fk w(et ) + t w(et ) fk

t=s t=s+1

,F = xE (M). k

2 This now enables us to prove theorem 8. Theorem 8 For any mcse problem M, for any sequence of choices E = (e1 , . . . , e ) in the algorithm 3 applied to M and any sequence of fractions F that is valid for E the allocation xE ,F , as dened in equation 2.1, is a core-allocation of the mcse game (N, cM ) associated with M. Proof : Take any coalition S N . We have to prove iS xi c(S ). Construct for coalition S a minimum cost extension E containing only edges between components of < N , E > containing members of S , such that S is connected to the source in < N , E E >. Let p be one less than the number of components of < N , E > containing members of S . Then |E | = p and the only dierence between < N , E > and < N , E E > is that the component C of in < N , E E > is a union of the component C of and other components of < N , E >. Construct the nested sequence E = E0 Ep = E E such that Eq \ Eq1 consists of exactly one edge which connects the component of the source in < N , Eq1 > to another component of < N , Eq1 >. Consider the mcse problems Mq =< N , , w, Eq >, where q varies from 0 to p, and note that for any q > 0, lemmata 36 and 37 and corollary 38 are applicable to the pair Mq1 and Mq = Mq1 . Dene E 0 = E , F 0 = F and for 1 q p, dene E q = E q1 and F q = F q1 recursively given E q1 and F q1 , as in lemmata 36 and 37. Then by corollary 38, E ,F E ,F xk q q (Mq ) xk q1 q1 (Mq1 ) for all k N and all 1 q p. Summing over q {1, . . . , p}, we obtain E ,F E ,F E 0 ,F 0 (M ) = xE ,F (M) xk p p (M ) = xk p p (Mp ) xk 0 k for all k N . Summing over k N \ S yields E ,F E ,F xk p p (M ) xk (M).
kN \S kN \S

(7.4)

p p Denoting (e1 ) := E p , we see that the graph < N , E E {e1 } > is a spanning extension p , . . . , ep p , . . . , ep for M. On the other hand, < N , E {e1 , . . . , e } > is a minimum cost spanning extension for M, hence E ,F w(e) w(e) = c(N ) = xk (M). (7.5)
p eE {e1 } p ,...,ep

e{e1 ,...,e }

k N

24 Now by denition of E ,

e E

w(e) = c(S )

(7.6)

and by eciency of xE p ,F p (M ) (cf. lemma 4),

p e{e1 } p ,...,ep

w(e) =

kN \S

E ,F xk p p (M ).

(7.7)

Plugging equations 7.6 and 7.7 into inequality 7.5 and using inequality 7.4, we obtain c(S ) +
k N \ S

E ,F xk p p (M ) =

E ,F xk (M)
k N

E ,F E ,F xk (M) + xk (M). E ,F E ,F xk (M) + xk p p (M ),

kN \S kN \S

k S

(7.8)

k S

which is equivalent to c(S )

E ,F xk (M).
k S

As we proved in lemma 4 that xE ,F (M) is ecient, it is a core element of (N, cM ). We now give a proof of proposition 10. Proposition 10 For any mcse problem M, for any E and E constructed by the algorithm 3 applied to M, DE (M) = DE (M).

Proof : First we prove that DE is independent of the order of E . Suppose that E and E are two sequences constructed in algorithm 3 applied to M, both leading to the same mcs extension E , i.e. E and E dier only in their order. Because in algorithm 3, the edges in E and E are ordered by non-decreasing cost, this means that E equals E except for edges of the same cost that are swapped. If more than two edges are swapped, it is possible to construct a series E = E 0 , . . . , E p = E of sequences all leading to the same edge set E , with for any q p, E q equal to E q1 except for exactly two subsequent edges with the same cost that are swapped. So it suces to prove DE = DE for E = (e1 , . . . , et , et+1 , . . . , e ) and E = (e1 , . . . , et+1 , et , . . . , e ), for some t < , with w(et ) = w(et+1 ). Two cases have to be distinguished: 1. the components C t and C t+1 formed by adjoining et and et+1 to < N , E E t > and < N , E E t+1 >, 1 , . . . , f ) by respectively are disjoint. For F V E , dene F = (f s f t f t+1 f = fs = f t+1 , = f t. if s {t, t + 1},

Obviously, F V E and xE ,F (M) = xE ,F (M). 2. the components C t and C t+1 are not disjoint. Then we are in the situation drawn in gure 3: C t consists of two components C1 and C2 , connected by et , and C t+1 is formed by connecting C3 to C t via et+1 . Without loss of generality, we suppose et+1 is incident to C2 . '$ '$ '$ C1 et C2 et+1 C3

&% &% &% Figure 3: et links C1 to C2 and et+1 links C1 C2 to C3 .

25 s ) by Now let F (f 1 , . . . , f ) V E be a valid sequence of fraction vectors, and dene F (f s=1 s f k t f k

s = fk 0 = f t+1 + gk k f t+1 t t+1 fk + fk t = f gk k 0

if s {t, t + 1} or k C1 C2 C3 , if k C1 , if k C2 , if k C3 , if k C1 , if k C2 , if k C3 ,

t+1 f k

where g I RC2 is a vector such that

t+1 gk = fk
k C 2 t fk kC1

(7.9)

gk 0 for all k C2 .

The system of equations (7.9) are feasible, because F V E implies t+1 t t fk =1 fk fk = gk .

kC2 k C 1 k C 1 kC2

One easily sees that F is valid for E and that ,F ,F t t w( xE = f et ) = fk w(et ) = xE . k k k

t=1 t=1

Hence DE is independent of the order of E , so we dene DE := DE

for any sequence E constructed by the algorithm 3 that leads to E . To prove that DE = DE for all minimum of M, we have to know more about the structure of the set of all mcses cost spanning extensions E and E of M. Now constructing an mcse for the mcse problem M is equivalent to constructing an mcst on the associated mcst problem < NE , E , wE > (cf. denition 11). It is well known that for any two mcsts < N , T > > of an mcst problem < N, , w >, for every edge e T \ T , there exists an edge e \ T such and < N , T T that < N , T {e } \ {e} > is again a minimum cost spanning tree. > are two mcses for M. Then with E and E E as dened Suppose that < N , E E > and < N , E E E in equation 3.2, note that < NE , EE > and < NE , EE > are mcsts of < NE , E , wE >. and e dened as in equation 3.1, it holds that either e E \E E or e E E E . Now for every e E \E E E E E E E \ E such that In the former case, there exists an edge {C, D} E E
{{C, D}} \ {e < NE , EE E} >

Hence, to prove that DE is independent of the mcse E , it suces to prove that DE = DE for two mcses of M with |E \ E | = 1. E and E are minimum cost extensions, the edge e E \ E and the edge e \ E have to have Because E and E E 1 s by non-decreasing cost into sequences E = (e , . . . , e 1 , e , es+1 , . . . , e ) the same cost. Now order E and E and E = (e1 , . . . , es1 , e , es+1 , . . . , e ) where s equals the number of edges in E with cost not greater than w(e ) = w( e). Then w(et ) > w(e ) for all t > s, and moreover E and E are two sequences that can be constructed by algorithm 3 applied to M. Consider the graph < N , E {e1 , . . . , es1 , e } >. As the next edge es+1 has greater cost than e , it has to be the case that adding e would introduce a cycle. But adding e to < N , E {e1 , . . . , es1 } > does not introduce a cycle. This means that e and e connect the same two components of < N , E {e1 , . . . , es1 } > t > are the same for all t, (see gure 4). Hence the components of the graphs < N , E E t > and < N , E E which implies that V E = V E . Together with w(e ) = w( e), this implies DE = DE = DE = DE . 2

E there exists an edge e \ E with e is also a minimum cost spanning tree. By denition of E E E = {C, D} \ E with e and w( e) = wE (e) = wE (eE ) = w(e ). In the latter case, there exists an edge e E E = e E , so e and E are mcses, w( connects the same components as e . Because both E e) = w(e ). So in both cases, we obtain that E E {e E } \ {e E} than E does. is a minimum cost spanning extension of M, which diers one edge less from E

26 '$ e C1 e &% Figure 4: e and e both link C1 to C2 . '$ C2 &%

References
Aarts H (1992) Marginal allocations in the core of minimal chain games, working paper no. 1103, Dept. of Applied Math, University of Twente, Enschede, The Netherlands. Aarts H and Driessen Th (1993) The irreducible core of a minimum cost spanning tree game, Zeitschrift f ur Operations Research 38 : 163174. Bird CG (1976) On cost allocation for a spanning tree : A game theoretic approach. Networks 6 : 335350. Boruvka O (1926a) O jist em probl emu minim aln m. Pr ace Mor. P r rodov ed. Spol. v Brn e (Acta Societ. Scient Natur. Moravicae) 3 : 3758. Boruvka O (1926b) P r sp evek k re sen ot azky ekonomick e stavby elektrovodn ch s t . Elektrotechnick y obzor 15 : 153154. Claus A and Kleitman DJ (1973) Cost allocation for a spanning tree. Networks 3 : 289304. Dijkstra EW (1959) A note on two problems in connection with graphs, Numerische Mathematik 1 : 269271. Faigle, U, Kern, W, Fekete, SP, and Hochst attler, W (1996) On the complexity of testing membership in the core of min-cost spanning tree games, to appear in Int. J. of Game Th. Feltkamp V, Tijs S, and Muto S (1994a) Birds tree allocations revisited, CentER DP 9435, University of Tilburg, The Netherlands. Feltkamp V, Tijs S, and Muto S (1994b) Minimum cost spanning extension problems : the proportional rule and the decentralized rule, mimeo, Tilburg University, The Netherlands. Graham RL and Hell P (1985) On the history of the minimum spanning tree problem. Annals of the History of Computing 7 : 4357. Granot D and Huberman G (1977) The relationship between convex games and minimal cost spanning tree games : A case for permutationally convex games, SIAM J. Alg Discr. Meth. 3 : 288292. Granot D and Huberman G (1981) On minimum cost spanning tree games, Mathematical programming 21 : 1 18. Granot D and Huberman G (1984) On the core and nucleolus of minimum cost spanning tree games, Mathematical programming 29 : 323347. Gr otsel, M, Lov asz, L, and Schrijver, A (1988) Geometric algorithms and combinatorial optimization, Springer Verlag, Berlin. Kuipers J (1993) On the core of information graph games. Int. J. of Game Th. 21 : 339350. Kruskal JB (1956) On the shortest spanning subtree of a graph and the traveling salesman problem, Proc. American Math. Soc. 7 : 4850. Prim RC (1957) Shortest connection networks and some generalizations, Bell Systems Techn. J. 36 : 13891401. Zumsteg S (1992) Marionette, Mimeo, Eidgenossische Technische Hochshule, Zurich, Switzerland.