Vous êtes sur la page 1sur 8

A Solvable Routing Problem

E. N. Gilbert A T&T Bell Laboratories, Murray Hill, New Jersey 07974

Stations 1,2, . . . ,n are interconnected; b,, channels join stations i and j . Channels may be grouped together in cables to reduce cost. A routing problem requires a channel layout (or network of cables) that has minimum cost. The cost of a cable is taken to be independent of its length but a functionf(k) of the cable size k. That kind of cost is unusual in practice, but might be appropriate if the cable is actually a satellite link. Requiringf(k) to be concave gives a discount for large cables. That imposes a number of special properties on the minimizing network. An extra assumption b,, = b, a constant for all i, j , produces a solvable problem. Depending on the shape off#), the solution network is either a complete graph or a particular kind of tree.

1. INTRODUCTION Routing problems arise in determining a customers bill for renting a large communications network. The bill is not based on the actual circuits supplied; these may be roundabout and may change from time to time as other customers change their demands. Instead, the bill is obtained by applying some simple formula to an ideal network, that would provide all the required facilities at minimum cost. Finding this minimizing network is a routing problem. It can be so difficult that only computer searches or heuristic methods apply [4]. For instance, a simplified kind of network a billing leads to the Steiner minimal tree problem [2,3,5], problem known to be NPcomplete [ 11. One simplified billing problem, requiring a minimal spanning tree, can be solved by an easy algorithm [6]. This paper gives another routing problem that also has a direct solution. The new routing problem is an oversimplification of current billing procedures. However it may supply useful insight because it retains a common feature of most billing formulas, a discount for grouping many channels together into a common cable. 2. THE PROBLEM There are n stations. A prescribed number b, of two-way channels join each pair ( i , j) of stations. The b, channels need not go directly from i to j. When routed
NETWORKS, Vol. 19 (1989) 587-594 0 1989 John Wiley & Sons, Inc.

CCC 0028-3045/89/050587-08$04.00

588

GILBERT

b.

C.

e.

f.

Im

>(
D

C
h.

0.

FIG. 1. Eight networks for 4 stations, one channel for each pair of stations.

through other stations, these channels combine with other channels into large cables. Ordinarily the cost of a cable containing k channels is less than k times the cost of a single channel. Then routing channels indirectly into large cables may be economical. Each way of combining channels into cables determines a distinct network. Figure 1 shows some networks with n = 4 and bij = 1 for all (i, j ) . Each line represents a cable. The number labeling a line is the size of the cable, i.e., the number of channels it contains. Some channels in a cable entering point P may not terminate at P but may leave P via other cables. Figure 1 does not specify how channels of different cables interconnect at each point but it should be clear that cables of the sizes shown can provide the required channels. Thus, in Figure l b the channel ( B , D )can go through point C while the channel (A,C) goes through D. Steiner points (S, S', in Figs. If and h) are extra points (not stations) where cables meet and exchange channels, none terminating. In some routing problems Steiner points can help to reduce cost [2,3,5]. Ordinarily the cost of a cable would depend both on its size k and on its length. Here the cable cost will be a functionf(k) of size only. This cable cost might be appropriate if terminal equipment at the ends of the cable costs much more than the cable itself (e.g.. in satellite systems). If Steiner points are allowed, and have different terminal equipment from stations, it may not be appropriate to assume that cable cost depends only on cable size. Under that assumption, Theorem 2 will show that Steiner points are uneconomical. Ignoring cable length as a cost factor is the most unrealistic assumption that will be made. Other assumptions will be

SOLVABLE ROUTING PROBLEM (iii) f ( k ) is an increasing (or at least non-decreasing) function of k, (iv) f ( k ) is concave.

589

Concavity off@) means that the marginal costf(k + 1) - f ( k ) of adding an extra channel to a cable is a non-increasing function of the cable size k. Thus (iv) allows a discount for using large cables. A typical concave function might bef(k) = a (3k for k > 0. There OL is a positive first cost of buying a right of way and digging a trench; p is a small positive constant incremental cost for each channel installed. If cables came only in one fixed size y, f ( k ) might increase stepwise at multiples of y; but that f ( k ) would not be concave. In a given network, let q ( k ) denote the number of cables of size k. The cost of the network will be the sum of the costs of its cables,

The problem is to find a network with minimum C.The assumptions made above will ensure that the minimizing network has some special properties (Theorems 1, 2, 3, and 4 below) and a further assumption about the form of bij will ensure a simple solution (Theorems 5 and 6). Some of the networks in Figure 1 cannot possibly be solutions. Network f costs more than network c, having two extra cables of size 3. Network h has a cable of size 3 in addition to those of b or g . The increasing property of@) eliminates network c because, compared with d, c replaces a cable of size 3 by one of size 4. Comparisons between the remaining networks a, b, d, e, and g must use concavity or specific numerical details about f(k).
3. BASIC PROPERTIES

The assumptions in Section 2 impose special properties on the solution.

Theorem 1. There is a minimizing network in which no two cables share the same pair of endpoints.
Proof. If two cables, of sizes r and s, have the same endpoints, combine them s) 2 0, into a single cable of size r + s. The net saving in cost isf(r) + f(s) - f(r rn the inequality following from (iv).

Theorem 2. A minimizing network has no Steiner points.


Proof. If S is a Steiner point, let P be one of the points connected to S by a cable. Without breaking any channel, move S to coincide with P.This motion changes cable lengths but not their costs. But now the cable PS, of length zero, can be removed to save its cost.

For example, moving S to A and then S to A improves Figure If to Figure Id. ' Removing a Steiner point may produce a network with cables in parallel (as in moving S to A in Fig. lh), in which case Theorem 1 will give a further saving (to Fig. Id). Next consider all the channels that pass through a given pair of stations (ij).These channels include the b, channels that connect i and j . Other channels may visit i and

590

GILBERT

ih FIG. 2. Network w t split routing.


j without terminating. Do all these channels follow the same route between i and j? If not, the network will be said to use split routing between i and j. Figure 2 illustrates split routing between A and C. Unlike Figure 1, Figure 2 shows each channel as a

line in order to show the channel terminations explicitly. The channels that visit both
A and C are AC (routed via B) and AB (routed via D and C ) .

Theorem 3. There is a minimizing network that does not employ split routing.
The proof appears essentially in [3]. If some channels visiting both points i. j follow different routes U and V, two different modifications are possible. Either reroute the channels of U to take route V or vice versa. The proof uses concavity of f ( k ) to demonstrate that the average of the changes in network cost from these two modifications is not positive; then both changes cannot increase the cost. As a corollary to Theorem 3, the b, channels joining i and j can all follow the same route; otherwise split routing would occur between i and j . Also if there is a cable between i and j then the b, channels (i, j ) may be sent directly over this cable to avoid split routing. In particular, if b, > 0 , there need be no cable i j containing only channels that do not terminate at i. If b, = 0 , a cable i j with no channel destined for i or for j may actually be needed. Figure 3 illustrates this by an example with n = 8, bM = bAc = bm = 100, bH = bH = bH = lo(), b E = bF = b G = b E = bcF = b G E F G B B B C C = b E = b F = bm = 1, and bij = 0 otherwise. Supposef(k) = CY D D k with CY so large that the minimizing network must use the fewest cables (a tree). Large cables will be needed from A to B,C,D and from H to E,F,G. The remaining 9 channels can be routed together via cable of size 9 connecting one of A,B,C,D with one of E,F.G,H (Fig. 3a). The cheapest network (Fig. 3b) places the cable betweenA and H (bM = 0 ) ; but none of the channels in this cable terminate at A or at H . Although costs often decrease as channels are grouped into larger cables, there is a point of diminishing return.

Theorem 4. Let Bj denote the total number of channels that terminate atj, B j = Zibij. There is a minimizing network in which, for all j, no cable to station j has size more than Bj.
Proof. Consider any cable ij, where Bj S Bi. If this cable has size B > Bj, remove it. Move the endpoint of every other cable at j from j to i (at no change in cost). That

SOLVABLE ROUTING PROBLEM

591

a.

103

103

H, ,
"

"O

103"

9
103

403

103

4. SOLVABLE PROBLEM
Further results may be obtained when bij has special form. In this section each point i contains a population of pi > 0 individuals and each pair of individuals requires a separate channel; then b , = pip,. Regarding stations and cables as vertices and lines of a graph, and ignoring cable sizes, one may speak of networks that are trees (Figs. lc,d,f), complete graphs (Fig. la), etc. One may also distinguish between end vertices (stations having one cable) and infernal vertices (stations with two or more cables). The next theorem applies to trees only.

592

GILBERT

Theorem 5. Let b,, = ppJ for all (i, j ) and let n 5 3. Among trees, one with least cost has only one internal vertex, a point I with largest population. The cost of the minimizing tree is Z.f(B,), summed for i f 1.
Proof. A tree with 2 or more internal vertices contains a pair i, jjoined by a cable. If this cable were broken, the tree would separate into two parts, say with populations PI and PJ. Then the cable ij contains B = PQ, channels. Suppose P, is the smaller pn is the of PI andP,, i.e., P, 6 P/2, whereP = PI + P, = pI + p2 + * . total population. Now

B, = p,(P - p J ) C P,(P - PI) = B follows becausep, < P, and the function p(P - p) is strictly increasing in 0 6 p < P/2. Then Theorem 4 shows how to save cost by making j an end vertex. Moreover, this modification again leaves a tree. Continuing in this way, one eventually obtains only one internal vertex I, which is connected by cables of sizes B, to the other vertices i. The cost of this tree is the sum Zf(B,) for i f I , which is minimized by taking B, as largeaspossible. BecauseBI - B, = (p, - p,)(P - pI - p,), astationIwithlargest w B, also has largest population pI. The minimizing t e is star-shaped, with all lines radiating from I. Of course other re shapes (Fig. 3) are possible if b, is not of the form plpJ. While Theorem 5 gives a least-cost tree, networks that are not trees may be even cheaper. The next theorem gives a complete solution when the populations all have the same size, i.e., b,, = b, a constant for all (i,j). Because of Theorem 3 the b channels ( i j ) can all take the same route. Each of the (d2)groups of b channels can be regarded as a kind of superchannel and the problem reduces to finding a network giving one superchannel to each pair of stations. Then it suffices to solve the case b = 1.

Theorem 6. Let b,, = b = 1 for all i j . If f(n - l)/f(l) G n/2, a tree of cost ( n - l)f(n - 1) is a minimizing network. Otherwise a complete graph of cost (n/ 2)f(l) is a minimizing network.
Proof. Let 4 ( k ) denote the number of cables of size k as in (1). Define the length of the network to be

L = Z4(k) k.
k

(2)

In ( 1 ) and (2) the summation extends over 1 G k G n - 1 because Theorem 4 makes cables of size n or more unnecessary. In that range, k can be expressed as an average of 1 and n - 1 with weights ( n - 1 - k ) / ( n - 2) and ( k - l)/(n - 2)
k = { ( n - 1 - k)l

+ (k -

1) ( n - l)}/(n - 2 )

Then, because f(k) is concave,


(n -

2)f(k)

3 (n -

1 - kM1)

( k - l)f(n - 1).

This inequality converts (1) into


( n - 2) C 3 {(n - 1)Al) - A n - 1)) &(k)

+ Mn

- 1)

A1))L.

(3)

SOLVABLE ROUTING PROBLEM


In (3) the sum, which extends over k = 1 , cables

593

. . . ,n

- 1, is the total number of

(q(0) is the number of station pairs not joined by a cable). Moreover each pair ( j j ) with no cable ij has a channel routed via two or more cables. Hence

Now (3) simplifies to

c3

(;)

f l l ) - (nf(1) - 2fln - l)}q(O)/(n - 2).

(4)

U f ( n - l)/f (1) 3 n/2 the coefficient of q(0) in (4) is negative or zero; the righthand side of (4) is minimized by taking q(0) = 0. The bound (4) is achieved by a complete graph, providing each pair (ij)with a direct channel in a cable of size 1 at minimum cost C = (n/2)f(l). Iff(n - l)/f(l) d n/2 the right-hand side of (4) is minimized by taking q(0) large, i.e., using a graph with as few lines as possible. The connected graphs with fewest lines are the trees, having n - 1 lines and hence
q(0) =

(;)

-n

+ 1 = ( n - l)(n - 2)/2.
1)fln - 1).
(5)

For trees (4) becomes

c 3 (n -

The bound ( 5 ) is achieved if the channels are routed into n - 1 cables of size n - 1 to a single vertex, as in Theorem 5 . With more general p i the minimizing network may be neither a complete graph nor
a tree. One such example has n = 6 stations with populations 1, 1, 1, 3, 3, and 3. Supposef(k) = k fork = 1,2, . . . , 11 andf(k) = 11 for k 3 11. The costs of the

complete graph and cheapest tree (Theorem 5 ) are 57 and 55. However, the reader will easily find other cheaper networks.

REFERENCES
[l] M. R. Garey, R. L. Graham,and D. S. Johnson, The complexity of computing Steiner minimal trees. SIAM J. Appl. Math. 32 (1977) 835-859. [2] E. N. Gilbert, Minimum cost communications networks. BSTJ 46 (1967) 2209-2227. [3] E. N. Gilbert and H.0. Pollak, Steiner minimal trees. SIAM J. Appl. Math 16 (1968) 129. [4] S. Lin, Heuristic programming as an aid to network design. Networks 5 (1975) 33-43.

594

GILBERT

[S] 2. A. Melzak, On the problem of Steiner. C a d . M a r k Bull. 4 (1961) 143-148. [a] R. C. Prim, Shortest connecting networks and some generalizations. BSTJ 36 (1957) 13891401.

Received August 1987 Accepted August 1988

Vous aimerez peut-être aussi