Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

AGGREGATION ERROR BOUNDS FOR A CLASS OF LOCATION

MODELS
R. L. FRANCIS
Department of Industrial and Systems Engineering, University of Florida, Gainesville, Florida 32611; francis@ise.ufl.edu

T. J. LOWE
College of Business Administration, University of Iowa, Iowa City, Iowa 52242, timothy-lowe@uiowa.edu

ARIE TAMIR
Department of Statistics & Operations Research, School of Mathematical Sciences, Tel-Aviv University, Tel-Aviv 69978, Israel, atamir@math.tau.ac.il
(Received February 1997; revisions received January 1998, November 1998; accepted December 1998)

Many location models involve distances and demand points in their objective function. In urban contexts, there can be millions of
demand points. This leads to demand point aggregation, which produces error. We identify a general model structure that includes
most such location models, and present a means of obtaining error bounds for all models with this structure. The error bounds
suggest how to do the demand point aggregation so as to keep the error small.

W hen solving real-world, large-scale location prob-


lems, it is quite common to aggregate demand
points. This aggregation amounts to intentionally reducing
this scale if customers are modeled at the “street address”
level. This very large number of customers makes the loca-
tion problem difficult to deal with, and for computational
the accuracy of the model to make it more computation- purposes it is quite common in such situations to aggregate
ally tractable. It is the purpose of this paper to develop the customers. For example, if there are 100 homes in a
worst-case error estimates for such aggregation, which in city block, they might be aggregated into (replaced by, for
turn suggest means of doing demand point aggregation so modeling purposes) a single imaginary home at the middle
as to help control the resulting error. We believe that our of the block. Such an aggregation approach simplifies the
approach can facilitate an analysis of the trade-off between location problem computationally but introduces error, be-
model accuracy and computational tractability. cause the “true” model has been replaced by an approxi-
Location problems occur whenever some private or pub- mating model.
lic sector organization is faced with a logistics problem There are other reasons besides computational ones for
involving locating facilities to provide services to custom- doing aggregation of customer data. Francis et al. (1999)
ers. Providing good service is essential, and the conve-
point out that information on individuals requiring a service
nience, which is usually measured by either travel distance
is sometimes regarded as confidential. Privacy considerations
or travel time between the facilities and the customers, is
lead individuals to expect information that would identify
an important measure of good service. The field of loca-
them not be released and made publicly available. In fact, it is
tion theory, as a research area, has been very active for a
for this reason that the National Census releases to the public
number of years. As an example of the diversity of location
only information on aggregations of individuals.
models studied, see Mirchandani and Francis (1990).
Many types of location models have been formulated and Also we note that in many instances there could be a
analyzed. One reason for the widespread study of these high degree of uncertainty (as measured by the coefficient
problems is that they arise frequently in practice. A reason of variation) regarding a single given customer’s demand
they have drawn interest from the research community is for services. However, when many such customers are ag-
that they are often extremely difficult to solve to optimality. gregated, the problem of predicting total demand becomes
In fact, many of the location problems are NP-hard, so re- much easier. For example, it is much easier to predict with
search effort has been expended toward the development of a given level of confidence that over an aggregated set of
heuristic solution methods for general models as well as tai- customers the total demand for services will be “D units”
lored solution approaches for special cases of the models. per time period than it is to predict the demand rate from
In many applied location contexts, customers can num- any single customer. This advantage of aggregation is often
ber in millions. For example, modeling the public sector referred to as “risk pooling,” and has found extensive ap-
decision of determining the location of emergency services plications in inventory and distribution settings (see, for
in a large metropolitan area could involve a problem of example, Lee et al. 1993, Kumar et al. 1995).

Subject classifications: Facilities Location: demand point aggregation error.


Area of review: SERVICES & MILITARY.

Operations Research, 䉷 2000 INFORMS 0030-364X/00/4802-0294 $05.00


Vol. 48, No. 2, March–April 2000, pp. 294 –307 294 1526-5463 electronic ISSN
FRANCIS, LOWE, AND TAMIR / 295
Although aggregation/approximation is pervasive in ap- are the n-median and the n-center models, introduced by
plied location work when there are many customers, aggre- Hakimi (1964, 1965). Francis et al. (1996) consider the
gation is often done on a rather ad-hoc basis, at least from n-median problem where customers are distributed over
the standpoint of minimizing the error associated with the the planar location space, and the travel distance is mea-
aggregation process. We note that pre-existing aggrega- sured with the rectilinear norm. Rayco et al. (1997) con-
tions of geographical data are readily available. For exam- sider the n-center problem, again in the context of
ple, one can obtain household data on an aggregated basis customers distributed in the plane, and the rectilinear
in terms of city blocks, postal zip codes, tracts, etc. In fact, norm. In both of these papers, an upper bound (in the
aggregation of customer facilities by replacement with cen- form of an error bound function) is developed using an
troids of groups of customers is quite common. Such ag- approach first reported in Francis and Lowe (1992) and
gregation schemes may seem reasonable to the layperson, then used to drive an aggregation scheme. That is, the
but they ignore the issue of aggregation error. method of doing aggregation is driven by the desire to
Rushton (1989), in an insightful discussion of applica- minimize the error associated with aggregating customer
tions of location models, discusses aggregation. He states: data.
“The effect of employing discrete spatial structures to rep- In this paper, we generalize and extend many of the
resent data that is distributed continuously is largely un- results in Francis et al. (1996) and Rayco et al. (1997). Our
known, though recent research has shown that the effects purpose herein is to report error bounds for an entire class
on the validity of results from location-allocation models of location models. Note that our focus is not on solving
can be considerable. In most cases, the decision to use a the original (reported) location problem, but instead is to
particular data structure is a matter of convenience and develop an error bound (function) that can potentially be
most analysts do not discuss the potential consequences of used to develop good data aggregation schemes for the
the data system they use or the alternative they reject. . . . problem.
Clearly, optimum locations must be sensitive, to an impor-
tant degree, at some level of spatial aggregation of
data. . . . As for capturing geographical complexity, the his- 1. MODEL INTRODUCTION
tory of applications show little development: decisions to
use crude distance functions and spatially aggregated data We now develop notation that we shall use in the remain-
units continue to be the rule and the lack of any recog- der of our paper. We let P ⫽ ( p1, . . . , pm) represent the
nized method for evaluating the consequences of these vector of m customer locations (with pi the location of
choices on the quality of the results has encouraged ana- customer i), and X ⫽ ( x1, . . . , xn) represent the set of n
lysts to simply dismiss the problem as intractable.” locations (to be determined) of the new facilities (with xj
We now briefly discuss some other selected aggregation the location of facility j). Whenever a fixed number of new
literature. Goodchild (1979) discusses aggregation effects facilities is to be located, we assume that the number of
on both median and minimax type models. Murray and such facilities is n. Note that the location context is quite
Gottsegen (1997) study the effects of aggregation on loca- general, i.e., we are not restricted to the case where cus-
tion model solutions. Mirchandani and Reilly (1986) exam- tomers are located (and the new facilities are to be lo-
ine the effect of using point nodes to approximate “spatial cated) in the plane or on a transport network, etc.
nodes.” Daskin et al. (1989) study aggregation effects in For location problems, the notion of distance is para-
covering problems. Hillsman and Rhoda (1978) identify mount. Given two locations, say u and v, we denote by
three sources of aggregation error, and error measurement d(u, v) the distance between the two points. We assume
is studied by Casillas (1987). Techniques for correcting the distance has the following (metric space) properties:
different types of error are proposed by Current and (1) d(u, v) ⫽ d(v, u) for all u and v (symmetry);
Schilling (1987, 1990) and Hodgson and Neuman (1993). (2) d( x, y) 聿 d( x, z) ⫹ d( z, y) for all x, y, z (triangle
For a more detailed accounting of published work on inequality); (3) d(u, v) 肁 0 for all u and v, with d(u, v) ⫽
n-median model aggregation, see Erkut and Bozkaya 0 meaning u ⫽ v (positivity). As is usually the case, if the
(1999). location problem is posed on an undirected transport net-
Many of the aggregation studies have been experimental work, then d⵺ is the shortest path distance. If the location
in nature. Aggregations have been simulated over gener- problem is posed in the plane, then d⵺ is some relevant
ated data sets, and estimates have been evaluated after the distance defined by a norm.
aggregation has been implemented. With few exceptions, Given a location u and a vector (or a finite set) X of
e.g., Current and Schilling (1987, 1990) and Plastria locations, we denote by D(X, u) the distance from u to the
(1996), aggregation procedures are generally not devel- closest component (member) of X. Often, we will be deal-
oped with the objective of minimizing aggregation error. ing with objective functions that involve sums and/or max-
However, aggregation schemes that attempt to reduce an ima of (nonnegative) weighted distances. The weights
upper bound on aggregation error have recently been de- reflect the relative importance of various terms in the lo-
veloped for two specific location problem settings. Perhaps cation problem objective function. We will reserve the
the two most important location models in the literature symbol wij for the weight associated with distance between
296 / FRANCIS, LOWE, AND TAMIR
new facility j and customer i, and vjk will be used to weight That is, it is guaranteed that regardless of the locations X
the distance between new facilities j and k. of the new facilities, the absolute difference of the objec-
We let f(X:P) represent an original location model, i.e., tive function values of the original problem and the ap-
the underlying location model of interest. The modifier proximate problem will not exceed eb. As we will see in
“original” is important because, via the process of aggrega- Table 1 the exact form of eb depends upon the structure of
tion, we derive an approximate location model. The ap- the original problem, as well as the vector P⬘ of aggregate
proximate model arises as follows. Via aggregation, the locations. Thus, one can readily see from (1) that a good
vector P is replaced by an (aggregate) vector P⬘, whose aggregation scheme can be derived by paying close atten-
elements are taken from a collection Q of distinct aggre- tion to the error bound expression when doing the aggre-
gate demand points. The number of elements, q, in Q is gation. In fact, in both Francis et al. (1996) and Rayco et
significantly less than m, the number of elements in P. The al. (1997), aggregations are derived for the problems stud-
approximate location model is denoted by f(X:P⬘). ied by first deriving the appropriate eb function, and then
The following discussion is taken from Francis et al. heuristically minimizing (by appropriately choosing P⬘) the
(1999). eb function.
We now demonstrate the usefulness of error bounds by
If p⬘i replaces pi, we customarily say that pi is aggregated into
showing how to use these error terms to bound other re-
p⬘i. Often the rule used is to replace pi by a closest member of
Q to pi. Another possible rule derives from using some par- lated important quantities. Specifically, we wish to bound
tition of the demand points into q disjoint and exhaustive the additional cost (error) due to implementing the opti-
sets. Then an aggregate demand point for each set is chosen mal solution to the aggregate model and not the optimal
and it replaces every demand point in the set. For example, solution to the original problem: the “cost” of approxima-
demand points might be partitioned according to the postal tion. We will also show how we can use these bounds to
code area (PCA) they are in; every demand point in each
PCA might be replaced by the centroid of the PCA (Domich
control the error by adjusting the number of aggregate
et al. 1991). Thus, to obtain an aggregation we must decide demand points. As above, let f(X:P) represent the location
on three things: model with respect to the original customer vector P, and
(D1): q, the number of aggregate demand points, let f(X:P⬘) represent the model with respect to the aggre-
(D2): the locations of the aggregate demand points, gate vector P⬘. Let X*(X⬘) be the vector of optimal facility
(D3): the replacement rule.
locations for model f(X:P)( f(X:P⬘)). (Occasionally in what
In principle, one could imagine an approach which deter-
mines (D2) and (D3) simultaneously instead of sequentially, follows, we assume that we have the optimal solution to
but we do not know of one. The freedom of choice in the the aggregated location problem available to us.) Consider
three decisions, D1, D2, D3 allows numerous aggregation the following two differences:
schemes. There are many ways in geography of partitioning
demand points, including tiling, space filling curves, and 兩f共X*:P兲 ⫺ f共X⬘:P⬘兲兩,
voter districting methods (Williams 1995). Each of these par-
兩f共X*:P兲 ⫺ f共X⬘:P兲兩.
titioning methods can result in an aggregation method. A
customary way of choosing each aggregate demand point lo- The first is the difference between the optimal objective
cation, given the partition, is to use the centroid of the de-
mand points in each part of the partition. We also note that
function value (OFV) using the optimal locations for the
medians can also be used, provided they are well-defined . . . . new facilities when serving the original demand point set
When the replacement rule (decision D3) is to replace and the OFV using the optimal locations for the new facil-
each demand point by a closest aggregate point, we note that ities when serving the aggregate demand point set. The
the three decisions, D1, D2 and D3, are identical to ones made second is the difference between the OFV using the opti-
in solving a location problem; to choose the number of new
mal locations for the new facilities when serving the origi-
facilities, the locations of the new facilities, and the assign-
ment of each demand point to a closest new facility. Thus it nal demand point set and the OFV using the optimal new
is not surprising that previous work in location theory can be facility locations for the aggregated problem when these
helpful in doing aggregation. locations serve the original demand point set.
It is our observation that there seems to be no agreed- Let eb denote the error bound defined in (1). (See Table
upon way to measure aggregation error. Thus, not surpris- 1 for appropriate expressions for a variety of location mod-
els.) Geoffrion’s (1977) theorems 1 and 5 can be used to
ingly there is no agreed-upon way to compare two different
show the following:
aggregations. This also means there is as yet little point in
developing a list of test problems to compare two different 兩f共X*:P兲 ⫺ f共X⬘:P⬘兲兩 ⭐ eb, (2)
aggregation approaches. In spite of this lack of agreement,
兩f共X⬘:P兲 ⫺ f共X*:P兲兩 ⭐ 2eb. (3)
we note that there are at least two fundamental measures
of error, incorrect locations of new facilities, and an incor- The difference 兩f(X⬘:P) ⫺ f(X*:P)兩 in (3) is the additional
rect objective function value. We focus on this latter con- forgone savings due to implementing an approximate solu-
cept of error. By an error bound, we mean a number tion (based on a solution to the aggregate model) and not
(hereafter denoted by eb) such that an optimal solution. Note that the bound 2eb in (3) can be
computed for any aggregate vector P⬘ before we solve the
兩f共X:P兲 ⫺ f共X:P⬘兲兩 ⭐ eb, for all X. (1) aggregate model to optimality and obtain X⬘.
FRANCIS, LOWE, AND TAMIR / 297
Following Geoffrion (1977), a bound that is stronger Equality holds if and only if some x⬘[i] in X that is closest to
than 2eb can be obtained when X⬘ is known. First rewrite p⬘i is also closest to pi and satisfies d( x⬘[i], pi) ⫽ d( x⬘[i], p⬘i) ⫹
inequality (2) as d( p⬘i, pi).
⫺ f共X*:P兲 ⭐ eb ⫺ f共X⬘:P⬘兲. We now discuss briefly the models in the table, referring
Adding the term f(X⬘:P) results in to models by their table numbers. In §3 we discuss the
methodology for obtaining the error bounds and give error
f共X⬘:P兲 ⫺ f共X*:P兲 ⭐ eb ⫹ 关 f共X⬘:P兲 ⫺ f共X⬘:P⬘兲兴. (4) bounds for some models not in the table.
The optimality of X* for P implies that the left-hand side The conditional n-median problem, #1 (see Minieka
of (4) is bounded below by zero. The term in brackets on 1980, Berman and Simchi-Levi 1990, and Drezner 1995), is
the right-hand side is bounded above by eb. Thus the right to locate n new facilities; the travel distance for each de-
side of (4) is at most 2eb. The result is a post- mand point can be either to a closest new facility location
computational error bound (a posterior bound) that re- in X or to a closest existing service facility in E, whichever
quires X⬘. This bound will be strictly better than 2eb if of the two is closest. The nonnegative “weight” wi associ-
f(X:P⬘) undershoots f(X:P) at X⬘ by less than the greatest ated with each demand point is typically a measure of
possible amount (which is eb). relative frequency of travel. Thus f(X:P) is a total travel
We now give an overview of the remainder of the paper. distance to closest facilities (new or existing). The condi-
In §2, we discuss several location models that have ap- tional n-center model, #6 (Minieka 1980 and Drezner
peared in the literature. For each discussed model, we 1989), has a similar interpretation but occurs in a worst-
provide an eb function. The main results of §2 are cap- case context; thus it is the maximum nearest distance that
tured in Table 1. Also, we discuss the problem of finding a is of interest. We note that the error bounds for these two
minimal number of aggregate demand points to guarantee conditional models are the same as those for the conven-
the error bounds will not exceed any given positive con- tional n-median and n-center models (Francis and Lowe
stant. Section 3 contains our theoretical results. We de- 1992), that is, the “unconditional” models.
velop a general model structure and obtain an error bound The multifacility minisum (MFMS) model, #2 (see
for it. Also we identify the general structure of location Francis et al. 1992), adds weighted distances between new
models for which our error bound analysis is possible, and facilities and demand points to weighted distances between
show that many known models have such structure. Then pairs of new facilities. The analogous multifacility minimax
we illustrate the results by deriving the error bound for (MFMM) model, #7 (Francis et al. 1992), considers the
one of the models studied. Also, we give enough guidance maximum of such weighted distances.
so that the reader can easily derive the error bounds for The quadratic assignment problem, #3, abbreviated as
the other problems in Table 1. Finally, in §4 we discuss QAP (see Burkhard 1990 and Pardalos and Wolkowicz
possible future research issues. 1994) is a form of the MFMS problem with a discrete
possible choice of new facility locations. We denote the
location of each new facility j by a[ j] and choose a[ j] from
2. PROBLEMS STUDIED AND ERROR BOUNDS
a finite set.
Table 1 shows a partial list of the location models for Goldman’s (1971) two-stage processing model, #4, ad-
which we have error bounds. Each model includes (near- dresses the processing of mail. Mail originating at a de-
est) distances between facilities explicitly in its objective mand point goes through processing at two facilities and
function, and each has given demand points as well as new then goes on to some other demand point. The model is
facility locations. Each distance can be multiplied by a set up so as to consider a “cheapest” choice of the two
nonnegative number called a weight. The operators involv- facilities for each demand point pair. For demand points h
ing the (weighted) distances are the maximum, summation, and i, the weights ahi and chi are associated with the dis-
and minimum operators. The last column of the table tances between demand points and new facilities, while the
shows the error bounds for the models. Since the error weight bhi is associated with the distance between the new
bounds hold for all X, we leave unstated the sets of feasi- facilities.
ble solutions for the models. The supporting median model, #5 (Mirchandani and
As a preliminary motivation for the error bounds, we Odoni 1979), considers travel between demand point loca-
state the following known result (Zemel 1985, Francis and tions ( pi) and airports (vt). Travel can either be directly as
Lowe 1992), which is based on the triangle inequality. The measured by a metric d2, or by means of a cheapest sup-
lemma amounts to a sort of triangle inequality for nearest porting median as measured by a different metric d1. The
distances, and it is essential for deriving the aggregation supporting medians could be public transit stops, such as
error bounds. subway stations. Note that there is no need to aggregate
LEMMA 1. Suppose we are given any X. For i ⫽ 1, . . . , m, the airport locations.
we have the following: The round-trip model, #9 (Elzinga and Hearn 1972),
considers responding to an emergency at some demand
D共X, p i 兲 ⭐ D共X, p⬘i 兲 ⫹ d共 p⬘i , p i 兲. point pi from a closest center in time wiD(X, pi). From pi
298 / FRANCIS, LOWE, AND TAMIR
there is then travel to some given location qi (perhaps a herit” these operators from the original model structure,
hospital) in a time of ␻id( pi, qi). The result is a “round- with the one exception of the obnoxious location model.
trip” time of wiD(X, pi) ⫹ ␻id( pi, qi). The weights wi and We consider all these properties quite reasonable. Indeed,
␻i reflect the possibility of different travel velocities de- it can be shown that each of the error bounds defines a
pending upon the “leg” of the trip. The time to return distance between P and P⬘, given the reasonable assump-
from qi to the starting point of the round trip is omitted tion that if the error bound value is zero then every d( p⬘i,
from the model as being unimportant compared with the pi) ⫽ 0. Thus attempts to make the lower bound small can
other times. Because the problem occurs in a worst-case be viewed as attempts to make P⬘ “close” to P.
context, it is the maximum time that is relevant. The error bound expressions lead in a natural way to the
The following scenario motivates the multistop model, definition of “second-order” location problems. Such
#8 (Handler and Mirchandani 1979). A breakdown repair second-order problems have previously been noted by
service is to be located at some point x. With probability Francis and Lowe (1992) when the original model is either
␣, a service unit is dispatched directly to some breakdown an n-median or n-center model.
point pi. With probability 1 ⫺ ␣ the unit is first dispatched To see how such second-order problems are obtained,
to some point p0 to collect existing equipment and then to suppose Q is a collection of aggregate demand points, with
the point pi. The maximum expected response time to a 兩Q兩 fixed. Suppose we choose as p⬘i a closest point in Q to
random call is of interest. Note the similarity in model pi, so that d( pi, p⬘i) ⫽ D(Q, pi). (Picking as p⬘i a point in Q
structure to the round-trip problem. other than a closest one cannot give a smaller error bound
The cent-dian model, #10 (Halpern 1978), also called value.) If we make the replacement of each d( pi, p⬘i) by
the medi-center model (Handler 1985), considers what D(Q, pi) in the error bound expression we get an expres-
amounts to a weighted average of an n-median and a sion, say eb(Q:P). We can consider the problem of mini-
n-center objective. The goal is to find a balance, using an mizing eb(Q:P) over Q, subject to 兩Q兩 fixed (or an
adjustable “balancing” parameter ␤, of efficiency (least- adjustable parameter), as the second-order location model.
cost) and equity (worst-case). The error bound reflects the For models 1– 4 in the table, the second-order model is
model structure and combines n-median and n-center er- a 兩Q兩-median model. For models 6 – 8 and model 12 in the
ror bounds. table, the second-order model is a 兩Q兩-center model. The
The one-facility ᐉp model, #11 (Shier and Dearing second-order models for models 5 and 9 are new, as best
1983), captures several different objective functions, de- as we can tell. The second-order model for model 10 is a
pending upon the value of p (1 聿 p 聿 ⬁). When p ⫽ 1, cent-dian/medi-center model, while for model 11 it is a
the median problem results, whereas p ⫽ ⬁ corresponds to sort of ᐉp兩Q兩-median model. We believe that the predomi-
the center problem. Note that the error bound inherits the nance of resulting secondary models that are 兩Q兩-center or
structure of the model. 兩Q兩-median models reinforces the central role these two
The obnoxious location model, #12 (Erkut and Neuman models play in location theory.
1989), considers the minimum of the distances between The error bounds of Table 1 can be used to drive de-
each demand point and a closest obnoxious new facility. mand point aggregation heuristic algorithms. When the
This minimum is to be maximized. This model is the only original model is the n-median or n-center model, Francis
maximization model in the table. If we also include inside et al. (1996) and Rayco et al. (1997), respectively, have
the min-operator the weighted minimum of the distances studied the use of low computational order “row-column”
between the obnoxious new facilities, the error bound re- heuristics for minimizing the resulting 兩Q兩-median or 兩Q兩-
mains valid. Note that the error bound does not reflect the center models to obtain a small error bound. Their point
“min” structure of the original model. All the models in of departure is what Francis and Lowe (1992) call the
the table but this one fit into the general class of models “paradox of aggregation.”
we discuss. We provide separate but related theory for this Essentially, the paradox states that the same reasons
model. that require aggregation in the original model also limit
In §3 we show that if each weight in the table that what can be done computationally in minimizing the error
multiplies a distance is replaced by a concave and nonde- bound model exactly. Certainly the paradox applies to all
creasing function of the distance, then the error bounds the models in the table. There are, however, several things
are still valid. It is relevant to note that such a function can that can be done. First, when the original distances are
model economies of scale in distances. network shortest-path distances, the distances in the error
Certain error bound properties are evident from the bound model can be taken to be simpler planar distances,
error bound column. The error bound expressions all in- such as rectilinear distances. Second, 兩Q兩 can be treated as
clude the weights involving the demand points and omit an adjustable model parameter.
the other weights. The error bounds all involve the dis- This second recourse, adjusting 兩Q兩, is particularly im-
tances between the pi and the p⬘i and are nondecreasing in portant to note because, for many location problems, the
these distances. The error bounds are zero if and only if number of new facilities is relatively small and can seldom
pi ⫽ p⬘i for every positively weighted pi. The error bounds be varied within a large range. In this sense the second-
all involve max-operators and/or sum-operators; they “in- order model is fundamentally different from the first-order
FRANCIS, LOWE, AND TAMIR / 299
Table 1 Aggregation error bounds for a family of location models.
No. Location Model f(X:P) Model Name Error Bound
1. ¥ {wi min{D(E, pi), D(X, pi)兩i} Conditional n-median ¥ {wid( p⬘i, pi)兩i}
2. ¥ {wijd( xj, pi)兩i, j} ⫹ ¥ {vjkd( xj, xk)兩j ⬍ k} MFMS ¥ {wijd( p⬘i, pi)兩i, j}
3. ¥ ¥ {wijd(a[ j], pi)兩i, j} ⫹ ¥ {vjkd(a[ j], a[k])兩j ⬍ k} QAP ¥ {wijd( p⬘i, pi)兩i, j}
4. ¥ ¥ {min{ahid( ph, xj) ⫹ bhid( xj, xk) ⫹ chid( xk, pi)兩j, Goldman’s 2-stage ¥ ¥ {ahid( ph, p⬘h) ⫹ chid( pi, p⬘i)兩h, i}
k}兩h, i}
5. ¥ ¥ {wit min{d2( pi, vt), min{d1( pi, xj) ⫹ d1( xj, vt)兩j}兩i, t} Supporting median ¥ ¥ {wij max{d1( pi, p⬘i), d2( pi, p⬘i)}兩i, j}
6. max{wi min{D(E, pi), D(X, pi)}兩i} Conditional n-center max{wid( p⬘i, pi)兩i}
7. max{max{wijd( xj, pi)兩i, j}, max{vjkd( xj, xk)兩j ⬍ k}} MFMM max{wijd( p⬘i, pi)兩i, j}
8. max{␣d( x, pi) ⫹ (1 ⫺ ␣)[d( x, p0) ⫹ d( p0, pi)]兩i} multistop max{d( pi, p⬘i)兩i}
␣ is a probability
9. max{wiD(X, pi) ⫹ ␻id( pi, qi)兩i} Round-trip max{wid( p⬘i, pi) ⫹ ␻i兩d( pi, qi) ⫺
d( p⬘i, qi)兩兩i}
10. ¥ {wiD(X, pi)兩i} ⫹ ␤ max{wiD(X, pi)兩i} Cent-dian/medi-center ¥ {wid( p⬘i, pi)兩i} ⫹ ␤ max{wid( p⬘i, pi)兩i}
␤ is a positive constant
11. [¥ {d( x, pi)p兩i}]1/p One-facility lp model [¥ {d( p⬘i, pi)p}]1/p
12. min{wiD(X, pi)兩i} Obnoxious max{wid( p⬘i, pi)兩i}

model. Indeed, this difference suggests that asymptotic gate demand points can be approximated within reason-
analysis, as done by Zemel (1985) and by Rayco and Fran- able accuracy in polynomial time. For example, if qmin
cis (1996), may be a fruitful approach to solving the sec- denotes the minimum number of aggregate demand points
ondary models. needed and the space is the Euclidean plane, Hochbaum
For more detail on using error bounds to suggest aggre- and Maas (1985) provide efficient algorithms that will find
gation methods, using simpler distances, and adjusting 兩Q兩, q⬘ centers (aggregate points) ensuring the desired bound,
see Francis et al. (1996), Rayco et al. (1997), and Francis and q⬘ is arbitrarily close to qmin. Specifically, for any fixed
et al. (1999). In the first two listed papers the authors find positive ⑀ their scheme computes in time which is polyno-
that the error bound for the n-median problem was usually mial in m ⫽ 兩P兩 and ⑀, a set of q⬘ aggregate demand points,
not tight but worked well as a surrogate error measure, such that the error bound is at most B, and q⬘ 聿 [(1 ⫹
and that experimentally, the error bound for the (un- ⑀)2]qmin. They give similar results for d-dimensional Eu-
weighted) n-center problem was usually found to be tight. clidean and rectilinear spaces. For general metric spaces
In the latter paper, the authors observed that, based on or networks the best known approximation of qmin is not as
experimental results, error is decreasing in 兩Q兩/兩X兩. good. The best values of q⬘ that can be computed efficiently
It is also possible, as discussed by Andersson et al. are of the order [log兩P兩]qmin. Such bounds are guaranteed if
(1998), to view the use of approximating planar distances we use, for example, a greedy scheme to solve the above
in the secondary problems as being “strategic” planning. covering problem on the points in P (see Chvatal 1979).
One can then obtain a family of smaller “tactical” sub- The above approach to bound and control the number
problems to solve, where the distances can be the original of aggregate demand points can also be adapted to apply
distances, but using the aggregate demand points. They to the n-median problem. Again, suppose that we want to
found that this approach worked quite well for n-median find the minimum number of aggregate demand points
aggregation, whereas results were mixed for n-center ag- qmin needed to ensure an upper bound of B on the error
gregation. The approach may be viewed as a heuristic de- term 2eb for the n-median problem. Unlike the above re-
composition method. sults for the n-center problem, we are unaware of any
Consider now how the error bound expressions can be polynomial time algorithm that can approximate qmin by a
useful in controlling or reducing the error by adjusting q, constant factor and at the same time ensure the bound B
the number of distinct aggregate demand points in P⬘, i.e., on the error term 2eb. However, the results of Lin and
the number of points in Q. As an illustration, consider first Vitter (1992) can be used to achieve the following result.
the (unweighted) n-center model. Suppose we wish to ag- For any fixed positive ⑀ we can compute, in time that is
gregate the original vector of demand points P and ensure polynomial in m ⫽ 兩P兩 and ⑀, a set of q⬘ aggregate demand
that the error term, 2eb (from (3)), is at most B. The points, such that the error bound is at most 2(1 ⫹ ⑀) B and
relevant question is how big should the set Q be, i.e., what q⬘ 聿 (1 ⫹ 1/⑀)qmin. For example, if ⑀ ⫽ 1, an error bound
is the minimum number of aggregate demand points of 4B can be guaranteed by computing in polynomial time
needed to ensure that 2eb 聿 B. Of course, this problem by q⬘ aggregate demand points, where q⬘ 聿 2qmin.
itself is NP-hard for general metric spaces or networks. It To conclude, we have demonstrated that for several im-
is equivalent to the problem of finding the minimum num- portant location models, the error bounds we will derive in
ber of centers (i.e., aggregate demand points) needed to §3 can be used quite efficiently to assess and control the
cover each one of the original demand points within a “cost of approximation” by selecting the set of aggregate
distance of B/ 2. However, the minimum number of aggre- demand points of the appropriate size.
300 / FRANCIS, LOWE, AND TAMIR
Figure 1. Construction process for SAND location models.

3. SAND LOCATION MODELS capturing transportation economies of scale) results in a


In this section we present a general and unifying approach SAND location model. Knowing that certain compositions
to derive error bounds for several of the most common of SAND functions give a SAND function is useful in
location models, including all the models in Table 1. We Theorem 5, which shows how to combine two such loca-
employ composition function notation concepts introduced tion models to obtain another one.
for location modeling by Hooker et al. (1991). A SAND costing function gk considered in §3.2 permits
Figure 1 illustrates a fundamental feature of all the lo- capturing a family of location models in one model; the
cation models we consider. k-centrum model of Slater (1978) and Andreatta and Ma-
All the models we study are constructed by beginning son (1985a, 1985b). If T(X:P) ⫽ (w1D(X, p1), . . . , wmD(X,
with a vector of new facility locations X ⫽ ( x1, . . . , xn) and pm)), then f(X:P) ⫽ gk(T(X:P)) is the sum of the k-largest
a vector of demand point locations P ⫽ ( p1, . . . , pm). entries in T(X:P). Taking k ⫽ 1 gives the n-center model;
Then various distances between elements in X and P are taking k ⫽ m gives the n-median model. The fact that the
computed, resulting in a “distance-like” vector we denote error bound for this model, gk(w1d( p1, p⬘1), . . . , wmd( pm,
by T(X:P). Finally a “costing function” g is applied to p⬘m)), reflects the structure of the costing function is a
T(X:P) to give the location model f(X:P) ⫽ g(T(X:P)). strong indication that it is important to consider carefully
For example, consider T(X:P) ⫽ (D(X, p1), . . . , D(X, the structure of the location model when doing demand
pm)). For any nonnegative vector U ⫽ (u1, . . . , um), if the point aggregation. Similarly, we discuss the ᐉp-norm model
costing function is g(U) ⫽ w1u1 ⫹ . . . ⫹ wmum, then f(X: of Shier and Dearing (1983), listed in Table 1. Again, this
P) ⫽ g(T(X:P)) is the n-median model. If g(U) ⫽ model captures a family of location models in one.
max{w1u1, . . . , wmum}, then f(X:P) ⫽ g(T(X:P)) is the Subsection 3.3 focuses on the general SAND location
n-center model. model, which includes all but the obnoxious model in Ta-
It is the choice of a costing function that gives the model ble 1 as a special case.
much of its structure. Two properties of the costing func- In §3.4 we derive the error bound for the general model
tion are essential to our approach: (1) the nondecreasing via Theorem 3, and then illustrate its application to the
(ND) property, and (2) subadditivity (SA). We refer to a
supporting-median location model, one of the most diffi-
costing function with both these properties as a SAND
cult models to obtain the error bound for. The general
function. We call a location model constructed using the
location model is defined using the idea of a metric space,
process the figure illustrates a SAND location model. The
but no insight is lost by supposing all distances to be net-
“distance-like” vector must also satisfy a type of triangle
work shortest-path distances, or planar distances. Also in
inequality. Namely, if P⬘ denotes the vector of aggregate
this section we show how to conclude the resulting model
demand points, there is a nonnegative vector S(P:P⬘) ⫽
is still SAND, if terms such as fixed costs (which are not
S(P⬘:P) so that T(X:P) 聿 T(X:P⬘) ⫹ S(P⬘:P) and T(X:
P⬘) 聿 T(X:P) ⫹ S(P:P⬘) for all X. Lemma 1 illustrates dependent on the demand point location vector P) are
these inequalities when T(X:P) ⫽ (D(X, p1), . . . , D(X, built into the location model. Thus, for example, if we add
pm)) and S(P⬘:P) ⫽ (d( p1, p⬘1), . . . , d( pm, p⬘m)). We define fixed costs to the n-median model, there is no change in
these properties formally in later subsections. the error bound. Similarly, having distances between new
In §3.1 we define SAND functions and list some of their facilities in the model, as in the multifacility and QAP
properties. These properties provide basic “building models in Table 1, has no effect on the error bound.
blocks” for constructing the SAND location models we In §3.5 we show how to combine two SAND location
study. For example, because the SAND property is pre- models to obtain a SAND location model, and we illus-
served under the addition and maximization operations, trate this approach with the cent-dian model of the table.
both the n-median and n-center models are SAND mod- The fact that a combination of SAND models can be a
els. Because a certain type of concave function is a SAND SAND model expands the class of such models and pro-
function, replacing the weights of any model in Table 1 by vides a way to analyze more complicated models by break-
concave, nondecreasing functions of the distances (thus ing them into smaller models.
FRANCIS, LOWE, AND TAMIR / 301
The obnoxious location model shown in Table 1 is not a S-6. If g(u) is a concave, isotone function of a single
SAND model because it does not have the subadditivity real variable u, satisfying g(0) 肁 0, then g(u) is a SAND
property. Nevertheless, in §3.6, we obtain an error bound function.
for it and other related models using theory related to the
foregoing theory. 3.2. Examples of SAND Functions
Finally, in §3.7, we indicate how to generalize our error
In this section we show that several unifying location mod-
bound approach when the distance symmetry property,
els, each of which includes both the n-center and n-median
d( x, y) ⫽ d( y, x) for all x and y, fails, but the triangle
models as special cases, are SAND models.
inequality still holds. Reasons such as one-way streets and
Subadditive functions play a key role in convexity analy-
rush-hour traffic can result in such a failure when the dis-
sis (see Rockafellar 1970, §15). Norm functions are subad-
tance has units of time.
ditive. Specifically, a real-valued function g(u) on Rr is a
norm if:
3.1. Properties of SAND Functions
1. g(u) ⬎ 0, for all u ⫽ 0.
In this subsection we list several of the most useful prop- 2. g(u) is subadditive.
erties of SAND costing functions. These properties derive 3. g(u) is positively homogeneous (of degree 1), i.e.,
from quite similar properties of subadditive functions. See g(␭u) ⫽ ␭g(u), for all u 僆 Rr and ␭ 僆 R⫹.
Rosenbaum (1950) for the proofs of these results on 4. g(⫺u) ⫽ g(u), for all u 僆 Rr.
subadditive functions and for additional results on subad-
ditive functions. Many of the properties are obtained from As a result of the above, we note that every norm func-
r
ones of Rosenbaum simply by replacing “subadditive” by tion that is isotone on R⫹ is a SAND function. In particu-
“SAND.” We have already illustrated the use of most of lar, the ᐉp norm, p 肁 1, defined by g(u1, . . . , ur) ⫽ (up1
the properties in our overview of §3. ⫹ . . . ⫹ urp)1/p, is a SAND function. The use of this costing
Unless otherwise specified, the domain of definition of function g was introduced by Shier and Dearing (1983) as
each real-valued costing function g will be R⫹ r
, the non- a unifying model for the n-median and n-center models
r
negative orthant of R , for any positive integer r. For each mentioned above. Note that the median model corre-
such costing function g we make the following two as- sponds to the case p ⫽ 1, while the center model corre-
sumptions: sponds to p ⫽ ⬁.
Rockafellar (1970, Theorem 4.7) states that a positively
ISOTONICITY/NONDECREASING ASSUMPTION. For each pair homogenous function on Rr is convex if and only if it is
r
of vectors u ⫽ (u1, . . . , ur), v ⫽ (v1, . . . , vr) 僆 R⫹ , u 聿 v, subadditive. Thus we conclude that every isotone, posi-
implies g(u) 聿 g(v). tively homogenous, and convex function is a SAND func-
tion. We use the latter result to prove that certain order
SUBADDITIVITY ASSUMPTION. For each pair of vectors u ⫽ functions are SAND functions.
r
(u1, . . . , ur), v ⫽ (v1, . . . , vr) 僆 R⫹ , g(u ⫹ v) 聿 g(u) ⫹ We have already noted that for each vector u ⫽
g(v). Thus a SAND function is both isotone (nondecreas- (u1, . . . , ur) in Rr, the functions ¥j⫽1
r
uj and max{uj兩j ⫽
ing) and subadditive. In particular, if g(u) is a SAND func- 1, . . . , r} are both SAND functions. (These functions are
tion it is nonnegative. used to derive error bounds for the n-median and n-center
models, respectively.) In an effort to unify these two loca-
From Rosenbaum’s properties of subadditive functions tion models and obtain a more general framework, we
we conclude the following: consider certain order functions and show that they are
S-1. If g1(u) and g2(u) are SAND functions, and c1 and SAND functions as well. For the use of these order func-
c2 are positive constants, then the function c1g1(u) ⫹ tions in k-centrum tree network location models, see An-
c2g2(u) is a SAND function. dreatta and Mason (1985a, 1985b) and Slater (1978). Let
S-2. If g1(u) is a nonnegative and subadditive function, T(X:P) ⫽ (w1D(X, p1), . . . , wmD(X, pm)). The k-centrum
g2(u) is nonnegative and ⫺g2(u) is isotone, then the func- location model is as follows: f(X:P) ⫽ gk(T(X:P)), where
tion g1(u) g2(u) is a SAND function if it is isotone. gk(T(X:P)) is the sum of the k-largest entries in T(X:P).
r
S-3. If g1(u) is a SAND function over R⫹ , and g2(u) is a
1
SAND function on R⫹, then the composite function Order Functions. For any vector u ⫽ (u1, . . . , ur) in Rr,
g2( g1(u)) is a SAND function of u on R⫹ r
. let u[1] 肁u[2] 肁 . . . 肁 u[r] denote the components of u in
S-4. If { gi(u)} is a set of SAND functions, all bounded decreasing order. For each j ⫽ 1, . . . , r, hj(u) ⫽ u[ j] is
above at each point of R⫹ r
, then the function defined at called the jth largest component of u, and gj(u) ⫽ ¥i⫽1 j

r
each point u of R⫹ as the maximum (or sup) of the values hi(u) is called the jth largest component sum of u.
of { gi(u)} at that point is a SAND function.
r
S-5. If g(u) is a SAND function over R⫹ , then for any THEOREM 1. For each j ⫽ 1, . . . , r, hj(u) is an isotone
fixed u ⫽ (u1, . . . , ur), the function g(␭u1, . . . , ␭ur) is a function defined on Rr, and gj(u) is a SAND function de-
SAND function of the single nonnegative variable ␭. fined on Rr.
302 / FRANCIS, LOWE, AND TAMIR
r
PROOF. To prove the isotonicity of the function hj, con- tion g(u1, . . . , ur) defined over R⫹ , the nonnegative or-
r
sider a pair of vectors, u, v, satisfying ui 肁 vi, for i ⫽ thant of R . Specifically,
1, . . . , r. Without loss of generality, suppose that v1
f共X:P兲 ⫽ g共T 1共X:P兲, . . . , T r共X:P兲兲 ⫽ g共T共X:P兲兲.
肁 . . . 肁 v r.
The function hj(u) is the jth largest element of the set At the beginning of §3 we discussed two special cases of
{u1, . . . , uj, . . . , ur}. Therefore, it is greater than or equal this model: the classical n-median and n-center models.
to the jth largest element of the subset {u1, . . . , uj}, which Every model in the table, except the last, is a special case
is equal to min{u1, . . . , uj}. Using the isotonicity of the of this model.
minimum operator we obtain We now continue by formally defining the distance-like
assumptions on the function T(X:P).
h j 共v兲 ⫽ v j ⫽ min兵v 1 , . . . , v j 其 ⭐ min兵u 1 , . . . , u j 其
Associated with T(X:P) we assume that there is a second
⭐ h j 共u兲. mapping. Given any pair of vectors P ⫽ ( p1, . . . , pm) and
We conclude that hj is isotone. P⬘ ⫽ ( p⬘1, . . . , p⬘m), this second mapping is defined by
The function gj is the sum of j isotone functions, and S(P:P⬘) ⫽ (S1(P:P⬘), . . . , Sr(P:P⬘)), from Ym ⫻ Ym into
r
therefore it is isotone. Because gj is positively homoge- R⫹ , where, for j ⫽ 1, . . . , r, Sj(P:P⬘) is a nonnegative real
neous, it follows from Rockafellar (1970, Theorem 4.7), number. The two mappings satisfy the following assump-
mentioned above, that gj is subadditive if and only if it is tions.
convex. Thus, it remains to show that gj is convex. DISTANCE-LIKE ASSUMPTION. For each triplet of vectors of
Let u, v be a pair of vectors in Rr and let t be any real points, X ⫽ ( x1, . . . , xn), P ⫽ ( p1, . . . , pm), and P⬘ ⫽
number satisfying 0 聿 t 聿 1. We have gj(tu ⫹ (1 ⫺ t)v) ⫽ ( p⬘1, . . . , p⬘m), in the metric space Y,
¥ {tuk ⫹ (1 ⫺ t)vk兩k 僆 I} for some I 傺 {1, . . . , r}, 兩I兩 ⫽ j.
Thus S共P:P⬘兲 ⫽ S共P⬘:P兲,

g j共tu ⫹ 共1 ⫺ t兲v兲 ⫽ t 冘 兵u 兩k 僆 I其
k
T共X:P兲 ⭐ T共X:P⬘兲 ⫹ S共P⬘:P兲.

⫹ 共1 ⫺ t兲 冘 兵v 兩k 僆 I其 ⭐ tg 共u兲 ⫹ 共1 ⫺ t兲 g 共v兲
k
j j The above assumption generalizes the triangle inequal-
ity property of the distance (metric) function. For example,
We now conclude gj is a SAND function. □ consider again the n-median model, where Ti(X:P) ⫽ D(X,
pi). For i ⫽ 1, . . . , m, define Si(P:P⬘) ⫽ d( pi, p⬘i). It is easy
THEOREM 2. Let a1 肁 . . .肁 ar be a sequence of nonnega- to verify that the distance-like property is implied by the
r
tive reals. The function g(u) ⫽ ¥j⫽1 ajhj(u) is a SAND symmetry of distance, the triangle inequality, and Lemma
r
function defined on R . 1.
Our last supposition of the general location model in-
PROOF. Using summation by parts, the function g(u) can
volves the objective function f(X:P) ⫽ g(T1(X:P), . . . ,
be represented as
Tr(X:P)). The real function g(u1, . . . , ur), defined on R⫹ r
,

冘 共a ⫺ a
r⫺1
is assumed to be SAND.
g共u兲 ⫽ a r g r共u兲 ⫹ j j⫹1 兲 g
j
共u兲.
j⫽1
3.4. Derivation of the Error-Bound Expression
Using this representation, the result follows directly from We now use the general model presented above to derive
property S-1 and the previous theorem, because gj(u) is a explicit expressions for eb, which are independent of X but
SAND function, and aj ⫺ aj⫹1 肁 0 for j ⫽ 1, . . . , r ⫺ 1. □ depend on the original set of demand points P and the set
of aggregate demand points P⬘ to be selected.
3.3. The General SAND Location Model
In a typical facility location model there is a vector of THEOREM 3. Consider the location model f(X:P) ⫽ g(T(X:
demand points P ⫽ ( p1, . . . , pm), embedded in some met- P)), where T(X:P) satisfies the distance-like assumption,
ric space Y (see Goldberg 1976). The objective is to select and g is a SAND function. Then,
a vector X ⫽ ( x1, . . . , xn) of facilities (servers) in Y to
兩f共X:P兲 ⫺ f共X:P⬘兲兩 ⭐ g共S共P:P⬘兲兲 for all X.
optimize some function of the distances {d( xj, pi)兩j ⫽
1, . . . , n; i ⫽ 1, . . . , m} between the m demand points PROOF. The distance-like property implies
and the n servers and the distances {d( xj, xt)兩j, t ⫽ 1, . . . ,
T共X:P兲 ⭐ T共X:P⬘兲 ⫹ S共P⬘:P兲.
n} among the n servers themselves.
To formulate a model, let T(X:P) denote a “distance- Next, from the isotonicity of the function g, we have
like” vector (to be defined shortly) of nonnegative reals,
g共T共X:P兲兲 ⭐ g共T共X:P⬘兲 ⫹ S共P⬘:P兲兲.
involving the above distances. Formally, T(X:P) has r real
components, say T1(X:P), . . . , T r(X:P), then T j(X:P), j ⫽ Applying the subadditivity of g, we obtain
1, . . . , r, is viewed as a mapping from the space Yn ⫻ Ym
g共T共X:P兲兲 ⭐ g共T共X:P⬘兲兲 ⫹ g共S共P⬘:P兲兲.
into the real line R. The objective function of the location
model f(X:P) is modeled by some real-valued costing func- Due to symmetry, we also have
FRANCIS, LOWE, AND TAMIR / 303
g共T共X:P⬘兲兲 ⭐ g共T共X:P兲兲 ⫹ g共S共P:P⬘兲兲. min兵d 2 共 p i , v t 兲, min 兵d 1 共 p i , x j 兲 ⫹ d 1 共 x j , v t 兲其其,
j⫽1,. . .,n
Finally, combining the above inequalities and using the ⭐ min兵d 2 共 p⬘i , v t 兲, min 兵d 1 共 p⬘i , x j 兲 ⫹ d 1 共 x j , v t 兲其其
symmetry assumption S(P:P⬘) ⫽ S(P⬘:P), it follows that j⫽1,. . .,n

兩f共X:P兲 ⫺ f共X:P⬘兲兩 ⫽ 兩g共T共X:P兲兲 ⫺ g共T共X:P⬘兲兲兩 ⫹ A i.

⭐ g共S共P:P⬘兲兲. □ Using the above definitions, we get

To demonstrate the above result we apply it to the sup- T i,t共共X, V兲:P兲 ⭐ T i,t共共X, V兲:P⬘兲 ⫹ S i,t共P:P⬘兲.
porting median model discussed in Mirchandani and Thus, we conclude that an error bound for the supporting
Odoni (1979). The input consists of a set of demand points median model is given by
P ⫽ ( p1, . . . , pm), a set of existing facilities (e.g., airports)
冘冘w
m r
V ⫽ (v1, . . . , vr), and nonnegative weights {wi,t兩i ⫽ 1, . . . , i,t
i,t S 共P:P⬘兲
m; t ⫽ 1, . . . , r}. The objective is to select a set of new i⫽1 t⫽1
facilities to optimize the objective,
冘冘w
m r
⫽ max兵d 1 共 p i , p⬘i 兲, d 2 共 p i , p⬘i 兲其.
冘冘
m r i,t
i⫽1 t⫽1
w i,t min兵d 2 共 p i , v t 兲, min 兵d 1 共 p i , x j 兲
i⫽1 t⫽1 j⫽1,. . .,n
Now consider the n-median model with concave SAND
⫹ d 1 共 x j , v t 兲其其. functions of nearest distances. From Property S-6, §3.1
(d1 and d2 are two different metrics defined on Y, repre- (see Rosenbaum 1950, Theorem 1.4.3) we know that a
senting travel by a supporting median and direct travel, concave, isotone function g(u) of a single real variable u,
respectively.) satisfying g(0) 肁 0, is a SAND function. As an example
We now define the functions T((X, V):P), S(P:P⬘), and consider the following extension of the classical n-median
g, corresponding to the supporting median model. For model, where we build in economies of scale in distance
each pair of indices i, t; i ⫽ 1, . . . , m; t ⫽ 1, . . . , r, set into the model.
Suppose that each customer pi, i ⫽ 1, . . . , m, is associ-
T i,t共共X, V兲:P兲 ⫽ min兵d 2 共 p i , v t 兲, min 兵d 1 共 p i , x j 兲 ated with a transportation cost function ci of the distance
j⫽1,. . .,n
traveled to the respective server, D(X, pi). Specifically, ci is
⫹ d 1 共 x j , v t 兲其其, a concave, isotone function of a single variable, satisfying
i,t
S 共P:P⬘兲 ⫽ max兵d 1 共 p i , p⬘i 兲, d 2 共 p i , p⬘i 兲其. ci(0) 肁 0. The location model is defined by the objective

冘 c 共D共X, p 兲兲.
Finally, set m
f共X:P兲 ⫽ i i

冘冘w
m r i⫽1
g共u兲 ⫽ i,t u i,t .
i⫽1 t⫽1 To obtain an error bound for this extended model, we
apply Theorem 3. Define the vectors T(X:P) ⫽ (T⬘(X:
To apply Theorem 3 we demonstrate that all the assump-
P), . . . , Tm(X:P)) and (S⬘(P:P⬘), . . . , Sm(P:P⬘)) as follows:
tions are satisfied. It is clear that g is a SAND function,
and S(P:P⬘) is symmetric and nonnegative. We only need T i共X:P兲 ⫽ c i 共D共X, p i 兲兲, S i共P:P⬘兲 ⫽ c i 共d共 p i , p⬘i 兲兲,
to satisfy the distance-like assumption,
for all i ⫽ 1, . . . , m. Using the concavity of the function ci,
T i,t共共X, V兲:P兲 ⭐ T i,t共共X, V兲:P⬘兲 ⫹ S i,t共P:P⬘兲. we note that ci is a SAND function and therefore,
Indeed, for any fixed i, t, and j ⫽ 1, . . . , n, c i 共D共X, p i 兲兲 ⭐ c i 共D共X, p⬘i 兲兲 ⫹ c i 共d共 p⬘i , p i 兲兲.
d 1 共 p i , x j 兲 ⫹ d 1 共 x j , v t 兲 ⭐ d 1 共 p⬘i , x j 兲 ⫹ d 1 共 x j , v t 兲 Thus,
⫹ d 1 共 p i , p⬘i 兲. T i共X:P兲 ⭐ T i共X:P⬘兲 ⫹ S i共P⬘:P兲.
Also m
Setting g(u1, . . . , um) ⫽ ¥i⫽1 ui and applying Theorem 3,
d 2 共 p i , v t 兲 ⭐ d 2 共 p⬘i , v t 兲 ⫹ d 2 共 p i , p⬘i 兲. we conclude that an error bound for this model is given by
Let
冘 c 共d共 p , p⬘ 兲兲.
m
i i i
A i ⫽ max兵d 1 共 p i , p⬘i 兲, d 2 共 p i , p⬘i 兲其. i⫽1

Then from the above inequalities, Finally we point out that the results of Theorem 3 ex-
d 1 共 p i , x j 兲 ⫹ d 1 共 x j , v t 兲 ⭐ d 1 共 p⬘i , x j 兲 ⫹ d 1 共 x j , v t 兲 ⫹ A i, tend easily to models that include fixed or setup costs for
the new facilities, as well as other terms that are indepen-
d 2 共 p i , v t 兲 ⭐ d 2 共 p⬘i , v t 兲 ⫹ A i,
dent of the set of customers P. To formulate this exten-
for a fixed pair i, t and all j ⫽ 1, . . . , n. Taking the sion, let H(X) be a mapping from Yn to the space Rl.
minimum operator over both sides of the above n ⫹ 1 Specifically, if H(X) has l real components, H1(X), . . . ,
inequalities implies that H1(X), then Hj(X), j ⫽ 1, . . . , l, is viewed as a mapping
304 / FRANCIS, LOWE, AND TAMIR
from Yn into the real line R. For example, H(X) can be a g1(T(X:P)) is the median objective, f2(X:P) ⫽ g2(T(X:P))
vector of setup costs of some or all of the facilities in X or is the center objective, and g0(v1, v2) ⫽ v1 ⫹ v2.
the vector of distances between some pairs of points in X. We note that the above results for the hyper-SAND
We have the following generalization of Theorem 3. (For model apply to any composition of any number of SAND
brevity we omit the proof because it is very similar to the models from the table in §2.
proof of Theorem 3.)
3.6. Obnoxious Location Models
THEOREM 4. Consider the location model f(X:P) ⫽ g(T(X: The crucial assumption in the general model requiring that
P), H(X)), where T(X:P) satisfies the distance-like assump- the objective is a SAND function does not hold for prob-
tion, and g is a SAND function. Then, lems dealing with the location of obnoxious facilities,
兩f共X:P兲 ⫺ f共X:P⬘兲兩 ⭐ g共S共P:P⬘兲, 0兲. where the goal is to maximize some isotone function of the
distances between the customers and the new facilities.
As an example of Theorem 4, f(X:P) might be the Therefore, the above approach for deriving error bounds is
n-median model including a collection of fixed costs, as- not applicable to obnoxious models.
sembled into a vector H(X), and added to the sum of We present a weaker model that is useful to obtain error
weighted closest distances. The error bound for this model bounds for some obnoxious facility location problems. We
would be the same as for the model without the fixed costs. replace the subadditivity property by the following assump-
tion: Let e 僆 Rr denote the vector, all of whose compo-
3.5. Hyper-SAND Models
nents are equal to 1. A function g(u) is e-subadditive on
The above general model can be further generalized to R⫹r r
, if for any u 僆 R⫹ , and ␭ 僆 R⫹1
,
obtain error bounds for complex location models, defined
by compositions of simpler location models. For example, g共u ⫹ ␭ e兲 ⭐ g共u兲 ⫹ g共 ␭ e兲.
consider the n-cent-dian model, where the objective is de- To motivate this definition, note that the function
fined as a sum (or convex combination) of the objective min{uj兩j ⫽ 1, . . . , r} is e-subadditive but not subadditive.
functions of the n-median and the n-center models. In this (Indeed, it is superadditive.)
case, it can be shown directly that the sum of the error Consider now a location model f(X:P) ⫽ g(T(X:P)),
bounds for the n-median and the n-center models is an where T(X:P) satisfies the distance-like assumption, and g
error bound for the n-cent-dian model. We will show that is isotone and e-subadditive. From the distance-like as-
this result can be extended to any composition of models sumption we have
as long as the composition function itself can be expressed
in terms of a SAND function. T j共X:P兲 ⭐ T j共X:P⬘兲 ⫹ S j共P:P⬘兲,
Consider a collection of s location models, { f1(X: for all j ⫽ 1, . . . , r. Let ␭(P:P⬘) ⫽ max{Sj(P:P⬘)兩j ⫽
P), . . . , fs(X:P)}. Suppose that for each j ⫽ 1, . . . , s, 1, . . . , r}. Then,
fj(X:P) ⫽ gj(T(X:P)), where gj is a SAND function. Let g0
be a SAND function, and define a composite model by T j共X:P兲 ⭐ T j共X:P⬘兲 ⫹ ␭ 共P:P⬘兲,

f共X:P兲 ⫽ g 0 共 f 1 共X:P兲, . . . , f s 共X:P兲兲. for all j ⫽ 1, . . . , r. Applying the isotonicity of g, we


obtain
An error bound for the composite model is obtained by a
composition of the error bounds of the individual models. g共T共X:P兲兲 ⭐ g共T共X:P⬘兲 ⫹ ␭ 共P:P⬘兲e兲.
We omit the proof, as it is quite similar to that of Theorem Finally, from the e-subadditivity we conclude
3.
g共T共X:P兲兲 ⭐ g共T共X:P⬘兲兲 ⫹ g共 ␭ 共P:P⬘兲e兲.
THEOREM 5. Let g1(u), . . . , gs(u) be a collection of s
r From symmetry we also have
SAND functions, defined on R⫹ . Let g0(v) be a SAND
s
function defined on R⫹. Consider the location model g共T共X:P⬘兲兲 ⭐ g共T共X:P兲兲 ⫹ g共 ␭ 共P:P⬘兲e兲.
f(X:P) defined by
Therefore, we obtain the following error bound:
f共X:P兲 ⫽ g 0 共 g 1 共T共X:P兲兲, . . . , g s 共T共X:P兲兲兲.
兩f共X:P兲 ⫺ f共X:P⬘兲兩 ⭐ g共 ␭ 共P:P⬘兲e兲.
Then,
As an example consider the obnoxious facility location
兩f共X:P兲 ⫺ f共X:P⬘兲兩 model, where the objective is to maximize the minimum
distance between the customers and their respective near-
⭐ g 0 共 g 1 共S共P:P⬘兲兲, . . . , g s 共S共P:P⬘兲兲兲.
est servers. In this model,
Notice that with the above notation, each function gj can
f共X:P兲 ⫽ min兵D共X, p i 兲兩i ⫽ 1, . . . , m其.
depend on a different component of T and a different
component of S. Set r ⫽ m, Ti(X:P) ⫽ D(X, pi), and Si(P:P⬘) ⫽ d( pi, p⬘i) for
The example of the n-cent-dian model presented above i ⫽ 1, . . . , r.
is a special case of the theorem, where s ⫽ 2, f1(X:P) ⫽ Therefore, an error bound is given by
FRANCIS, LOWE, AND TAMIR / 305
␭ 共P:P⬘兲 ⫽ max兵d共 p i , p⬘i 兲兩i ⫽ 1, . . . , r其. work is a tree. An examination of our error bound
derivations shows we assume very little about X. In fact, X
It is worth mentioning that the same error bound holds for
can be any one of these structures and the error bound
a more general model defined by
results still go through.
f共X:P兲 ⫽ min{min兵D共X, p i 兲兩i ⫽ 1, . . . , m其, Also, models have appeared in the literature, e.g., Ber-
man et al. (1995), where a demand “facility” is a path, and
min兵d共 x j , x t 兲兩j, t ⫽ 1, . . . , n; j ⫽ t其}.
the path is served by a closest new facility. For our pur-
3.7. Asymmetric Distances and Error Bounds poses, we assume that the distance from a serving facility
In the above discussion we have assumed that the distance to a demand path is the distance to the closest point in the
function d( , ) is symmetric with respect to its two argu- path. For median and center versions of these problems, it
ments. This assumption was made as part of the distance- is possible to derive error bounds using our approach.
like assumption, where we required that S(P:P⬘) ⫽ S(P⬘:P) Covering models have no distances in their objectives,
for all pairs of subsets P and P⬘. and we have no error bounds for such models. However,
The above symmetry assumption is certainly not justified their well-known close connection to center problems (e.g.,
in cases of one-way streets, up- and down-hill travel, or see Iyer and Ratliff 1990) suggests it may be possible to
where traffic is more congested one way than the other. To obtain error bounds for them.
overcome this difficulty we note that the above models can Because our work is on objective function approxima-
be easily modified as follows when S(P:P⬘) and S(P⬘:P) are tion, the following comment is in order. In exact optimiza-
not identical. The only modification needed is their re- tion, an optimal solution is unaffected by applying any
placement by the vector S*(P:P⬘) whose jth entry is the monotone transformation to the original objective func-
maximum of the jth entry in S(P:P⬘) and S(P⬘:P) for all j. tion, e.g., squaring a nonnegative objective function has no
It is easy to check that, with this modification, all the above effect on the optimal set. This is not the case whenever
results are now valid for this asymmetric case as well. optimization is not feasible or tractable for the original
model and one must use approximations. Approximate so-
lutions, as well as error bounds, are crucially dependent on
4. CONCLUSIONS the particular representation of the objective function. In
In this section we summarize our work and speculate particular, in our context, if we apply a monotone transfor-
about possible extensions and generalizations. mation that does not preserve subadditivity, then uniform
In §1 we introduced the problem of demand point ag- error bounds might not be achievable.
gregation for location models and discussed related litera- The process we have followed in this paper is to begin
ture, including the idea of an aggregation error bound, an with a location model and then obtain an error bound. It
upper bound on the absolute error, 兩f(X:P) ⫺ f(X:P⬘)兩, might be interesting to consider reversing the process. Per-
which holds for all X. In §2 we presented error bounds in haps we could state an error bound model, some function
Table 1 and discussed the use of the bounds for a class of of P and P⬘, and then determine if there exists a location
location models; we also outlined some generalizations, model with this error bound. Such results might be helpful
including replacing weights by concave, nondecreasing in trying to further classify models with common error
functions of the distances. In §3 we presented our theory bounds. Ideally, it would be desirable to have necessary
for deriving error bounds, including error bounds for mod- and sufficient conditions for a model to have a given error
els not listed in Table 1. We exploited the ideas of the bound.
triangle inequality with the SAND properties, subadditiv-
ity, and monotonicity to develop a general form of the
location model and its associated error bound. We also ACKNOWLEDGMENTS
showed that the idea of order functions allowed us to cap- This research is sponsored in part by the National Science
ture a class of location models in a single form. In addi- Foundation, Grant DMI-9522882.
tion, we showed how to obtain error bounds for certain We are happy to acknowledge also the constructive
combinations of location models, which we call hyper- comments of the referees. In particular, we appreciate the
SAND models. We also pointed out that the error bound suggestion that risk-pooling is another positive reason to
methodology is easily generalized to allow for directed net- do aggregation.
works where we do not have d( x, y) ⫽ d( y, x) for all x and
y.
We note there are a number of network location models REFERENCES
that involve locating “structures.” Hakimi et al. (1993)
Andersson, G., R. L. Francis, T. Normark, M. B. Rayco. 1998.
consider a class of such models with “center” or “median” “Aggregation method experimentation for large-scale
objectives where the structure can be one or more of any network location problems,” Location Sci., 6 25–39.
of the following: points in the network, partial arcs, full Andreatta, G., F. M. Mason. 1985a. K-eccentricity and abso-
arcs, paths, or subtrees. Tamir and Lowe (1992) also con- lute k-centrum of a probabilistic tree. European J. Oper.
sider several of these problems when the underlying net- Res. 19 114 –117.
306 / FRANCIS, LOWE, AND TAMIR
——, ——. 1985b. Properties of the k-centra in a tree net- Goodchild, M. F. 1979. The aggregation problem in location-
work. Networks 15 21–25. allocation. Geographical Analysis 11 240 –254.
Berman, O., D. Simchi-Levi. 1990. Conditional location prob- Halpern, J. 1978. Finding minimal center-median convex combi-
lems on networks. Transp. Sci. 24 77–78. nation (cent-dian) of a graph. Management Sci. 24 535–544.
——, J. Hodgson, D. Krass. 1995. Flow interception problems. Hakimi, S. L. 1964. Optimum locations of switching centers
Chapter 17 in Facility Location: a Survey of Applications and the absolute centers and medians of a graph. Oper.
and Methods. Z. Drezner (ed.). Springer Verlag, New Res. 12 450 – 459.
York, 389 – 426. ——. 1965. Optimum distribution of switching centers in a
Burkard, R. 1990. Locations with spatial interactions: the qua- communications network and some related graph-
dratic problem. Chapter 9 in Discrete Location Theory. theoretic problems. Oper. Res. 13 462– 475.
P. B. Mirchandani and R. L. Francis (eds.). J. Wiley & ——, E. F. Schmeichel, M. Labbé. 1993. On locating path- or
Sons, New York, 387– 438. tree-shaped facilities on networks. Networks 23 543–555.
Casillas, P. A. 1987. Data aggregation and the p-median prob- Handler, G. Y. 1985. Medi-centers of a tree. Transp. Sci. 19
lem in continuous space. In Spatial Analysis and Location 246 –260.
Allocation Models, A. Ghosh and G. Rushton (eds.), Van ——, P. B. Mirchandani. 1979. Location on Networks: Theory
Nostrand Reinhold Publishers, New York, 227–244. and Algorithms, The MIT Press, Cambridge, MA.
Chvatal, V. 1979. A greedy heuristic for the set covering prob- Hillsman, E. L., R. Rhoda. 1978. Errors in measuring dis-
lem. Math. Oper. Res. 4 233–235. tances from populations to service centres. Ann. Regional
Current, J. R., D. A. Schilling. 1987. Elimination of source A Sci. 12 74 – 88.
and B errors in p-median location problems. Geographi- Hochbaum, D. S., W. Maas. 1985. Approximation schemes for
cal Analysis 19 95–110. covering and packing problems in image processing and
——, ——. 1990. Analysis of errors due to demand data ag- VLSI. JACM 32 130 –136.
gregation in the set covering and maximal covering loca- Hodgson, M. J., S. Neuman. 1993. A GIS approach to elimi-
tion problems. Geographical Analysis 22 116 –126. nating source C aggregation error in p-median models.
Daskin, M. S., A. E. Haghani, M. Khanal, C. Malandraki. Location Sci. 1 155–170.
1989. Aggregation effects in maximum covering models. Hooker, J. N., R. S. Garfinkel, C. K. Chen. 1991. Finite dom-
Ann. Oper. Res. 18 115–140. inating sets for network location problems. Oper. Res. 39
Domich, P. D., K. L. Hoffman, R. H. F. Jackson, M. A. Mc- 100 –118.
Clain. 1991. Locating tax facilities: a graphics-based mi- Iyer, A. V., H. D. Ratliff. 1990. Accumulation point location
crocomputer optimization model. Management Sci. 37 on tree networks for guaranteed time distribution. Man-
960 –979. agement Sci. 36 958 –969.
Drezner, Z. 1989. On the conditional p-center problem. Kumar, A., L. Schwartz, J. Ward. 1995. Risk-pooling along a
Transp. Sci. 23 51–53. fixed delivery route using a dynamic inventory-allocation
——. 1995. On the conditional p-median problem. Comput. policy. Management Sci. 41 344 –362.
Oper. Res. 22 525–530. Lee, H., C. Billington, B. Carter. 1993. Hewlett-Packard gains
Elzinga, J., D. W. Hearn. 1972. Geometrical solutions for control of inventory and services through design for lo-
some minimax location problems. Transp. Sci. 6 379 –394. calization. Interfaces 23 1–11.
Erkut, E., B. Bozkaya. 1999. Analysis of aggregation errors for Lin, J.-H., J. S. Vitter. 1992. Approximation algorithms for
the p-median problem. Computers and Operations Re- geometric median problems. Information Processing Let-
search (Special Issue on Aggregation & Disaggregation ters 44 245–249.
for Operations Research) 26 1075–1096. Minieka, E. 1980. Conditional centers and medians on a
——, S. Neuman. 1989. Analytical models for locating unde- graph. Networks 10 262–272.
sirable facilities. European J. Oper. Res. 40 275–291. Mirchandani, P. B., A. R. Odoni. 1979. Locating new passen-
Francis, R. L., T. J. Lowe. 1992. On worst-case aggregation ger facilities on a transportation network. Transp. Res.
analysis for network location problems. Ann. Oper. Res. 13B 113–122.
40 229 –246. ——, R. L. Francis (ed.). 1990. Discrete Location Theory. J.
——, ——, M. B. Rayco. 1996. Row-column aggregation for Wiley & Sons, New York.
rectilinear distance p-median problems. Transp. Sci. 30 ——, J. M. Reilly. 1986. Spatial nodes in discrete location
160 –174. problems. Ann. Oper. Res. 10 329 –350.
——, ——, G. Rushton, M. B. Rayco. 1999. “A synthesis of Murray, A. T., J. M. Gottsegen. 1997. The influence of data
aggregation methods for multi-facility location problems: aggregation on the stability of p-median location model
strategies for containing error,” Geographical Anal., 31 solutions. Geographical Analysis 29 200 –213.
67– 87. Pardalos, P. M., H. Wolkowicz (eds.). 1994. Quadratic Assign-
——, L. F. McGinnis, J. A. White. 1992. Facility Layout and ment and Related Problems. DIMACS Series in Discrete
Location: an Analytical Approach. Prentice-Hall, Inc., Mathematics and Theoretical Computer Science, 16,
Englewood Cliffs, NJ. American Mathematical Society: Chapter 1, “The Qua-
Geoffrion, A. 1977. Objective function approximations in mathe- dratic Assignment Problem,” pp. 1– 42.
matical programming. Mathematical Programming 13 23–37. Plastria, F. 1996. How to choose aggregation points for con-
Goldberg, R. R. 1976. Methods of Real Analysis. Second Edi- tinuous p-median problems: a case for the gravity centre.
tion, J. Wiley and Sons, Inc., New York. Report BEIF/88. Vrije Universiteit Brussel, Belgium.
Goldman, A. J. 1971. Optimum location for centers in a net- Rayco, M. B., R. L. Francis. 1996. Asymptotically optimal
work. Transp. Sci. 5 212–221. aggregation for some unweighted p-center problems with
FRANCIS, LOWE, AND TAMIR / 307
rectilinear distances. Studies in Locational Analysis 10 Shier, D. R., P. M. Dearing. 1983. Optimal locations for a
25–36. class of nonlinear, single-facility location problems on a
——, ——, T. J. Lowe. 1997. Error-bound driven demand network. Oper. Res. 31 292–303.
point aggregation for the rectilinear distance p-center Slater, P. J. 1978. Centers to centroids in a graph. J. Graph
model. Location Sci. 4 213–235. Theory 2 209 –222.
Rockafellar, R. T. 1970. Convex Analysis. Princeton University Tamir, A., T. J. Lowe. 1992. The generalized p-forest problem
Press, Princeton, NJ. on a tree network. Networks 22 217–230.
Rosenbaum, R. A. 1950. Sub-additive functions. Duke Math. J. Williams, J. C., Jr. 1995. Political redistricting: a review. Pa-
17 227–247. pers in Regional Sci. 74 13– 40.
Rushton, G. 1989. Applications of location models. Ann. Zemel, E. 1985. Probabilistic analysis of geometric location
Oper. Res. 18 25– 42. problems. SIAM J. Alg. Disc. Meth. 6 189 –200.

You might also like