Linear

Programming
Brings

Marital Bliss
V. J. Sethuraman and C.P. Teo

1. Motivation & Carl, and Debbie & Dan. Observe that Dan likes
Brenda better than Debbie (his current partner), and
By all means marry; if you get a good wife, you'll Brenda likes Dan better than Bob (her current partner);
be happy. If not, you'll become a philosopher. - so, the proposed marriage would "break-up," and Dan
Socrates & Brenda would "elope." We conclude that the
proposed marriage in this case is "unstable". Obviously,
Alan, Bob, Carl and Dan, the only four bachelors in such a situation is undesirable for Marx, and so the least
oddtown, finally contemplate marriage. They approach he should look for is a stable marriage, where no pair of
Marx, the matchmaker, who introduces them to Alice, man and woman will find it beneficial to divorce their
Brenda, Cindy and Debbie. After the meeting, each respective spouses and marry each other. Can Marx
person ranks all of the members of the opposite sex, and always find a stable marriage? How many stable
hands it to Marx. marriages are there? Are some stable marriages better
than the- others? Can people misrepresent their true
Men's preference lists preferences and thereby gain an advantage? We will
Alan Cindy Alice Debbie Brenda answer some of these questions in the next section.
Bob Alice Cindy Debbie Brenda
Carl Debbie Cindy Alice Brenda The stable marriage problem has held a fascination
Dan Cindy Brenda Alice Debbie for computer scientists, mathematicians and economists
ever since its introduction in the pioneering paper of Gale
Women's preference lists and Shapley [62]. Research conducted during the past
Alice Dan Carl Alan Bob thirty-five years has helped us understand and
Brenda Carl Dan Alan Bob appreciate its connections to a variety of problems
Cindy Carl Bob Alan Dan arising in combinatorics, operations research and
Debbie Dan Bob Carl Alan economics. It has been applied to, for example, the
matching of graduating medical students to hopsitals and
college admissions.
For example, Alan likes Cindy best and Brenda least;
Cindy, on the other hand, likes Carl and Bob better than •••••••••••••••••••••••••••••••••••
Alan. Marx's job is to find a match for each man, and
his reputation depends on the number of successful * This article is edited version ofa paper of the same title by
the first author. The paper was submitted for the "Writing
marriages arranged. What should Marx do?
and Humanistic Studies Prize for Scientific and Engineering
Writing" competition organised by Massachusetts Institute
It is easier to answer what Marx should not do. of Technology in 1997; it won a second prize in the
Suppose he matches Alice & Alan, Brenda & Bob, Cindy competition.

Math•maHcat
EDLEY ~
..,_
 2. Helping Marx out and so he has a partner if and only if he is married to
exactly one of these four women . The corresponding
We shall translate Marx' s problem to a problem of decision variables are x~,~ 'xl ,2'xl ,3 'xJ,4respectively,
finding value of boolean variables satisfying linear and so mathematically, this constraint can be written as
constraints, and find all the stable marriage solutions to (1) xl ,l +x1 .2 +xl ,3 -txl ,4 =1.
this problem.
Similarly, we can write down the matching
constraints for each of the other people. The remaining
Formulation
seven constraints are :
The stable marriage problem involves two disjoint
(2) x 2. 1 +x2.2 +x2.3 +x2,4 =1
sets of size n, the men and the women. Each person has
a strictly ordered preference list of all of the members
(3) x3 ,1 + x 3.2 + x 3,3 + x 3, 4 = 1
of the opposite sex. A marriage M is a one-to-one
correspondence between the men and the women. A
(4) x4 ,1 +x4,2 +x4.3 +x4 ,4 =1
man-woman pair (m,w) is said to block the marriage M
if m and ware not married in M, but prefer each other to
their partners under M; such a pair is a blocking pair for
(5) x~,~ +X2.1+x3 ,1 +x4 ,1 =1
the marriage M. A stable marriage M is one with no
(6) X1 .2 + X 2.2 + x3 ,2 + x 4,2 = 1
blocking pairs.
(7) Xu +x2 ,3 +x3 ,3 +x4 ,3 =1
For each man-woman pair (m,w) , we introduce a
boolean variableXm,w :thus, x m.w is either zero or one
(8) x, ,4 + x2, 4 + x 3,4 + x4 ,4 = 1.
and x m,w = 1 if m and w have been paired in M and
x m,w = 0 otherwise. For (mathematical) convenience,
To write down the stability constraints, we use a
we "rename" the players involved. We use 1, 2, 3, 4 for
similar reasoning. Notice that we have sixteen stability
Alan, Bob, Carl and Dan respectively; similarly, 1, 2, 3, 4
constraints- one for each (man, woman) pair. We shall
will also represent Alice, Brenda, Cindy and Debbie
illustrate here how one obtains the stability constraint
respectively. This will cause no confusion because we
corresponding to the pair (Bob, Debbie). The stability
will always use ordered pairs, where the first "compo-
constraint, in English, says that if Bob and Debbie are
nent" denotes a man, and the second denotes a woman.
not married to each other, then at least one of them gets
For example, the pair (Bob, Cindy) would be denoted by
a "better" partner. In our example, Bob prefers Alice
(2, 3). Altogether, we have sixteen decision variables in
and Cindy to Debbie, whereas Debbie prefers Dan to
this problem .
Bob. In order for a marriage to be stable, either Bob and
Debbie are married, or Bob is married to Alice or Cindy,
The key idea is to express all of the constraints of
or Debbie is married to Dan. Another way to express
the problem as linear relations involving these decision
this is that any marriage in which Bob is married to
variables. So, what are the constraints? A little thought
Brenda, and Debbie is married to either Carl or Alan
convinces us that there are two classes of constraints. '
is unstable. A mathematical translation of the last
statement is
• Matching Constraints:
These express the fact that each person has
exactly one partner (of the opposite sex). Let us convince ourselves that equality (9) does
indeed capture the stability condition for the pair (2,4).
• Stability Constraints: When is (9) violated? If x 2 4 =1,all the other decision
variables appearing in (9) a;e forced to be zero by the
If man m and woman w have not been paired up, matching constraints (2) and (8), and so (9) is satisfied.
then at least one of them has been paired with a "better" So, the only way in which (9) is violated is if x 2 4 = 0,
partner. In other words, if m and w are not married to x 2 ,2 = 1, and x 1•4 + x 3.4 = 1. ,
each other, either m is married to somebody he prefers (Since X 1•4 + x 3 •4 ~1 ,by the matching constraint.) Thus
tow, or w is married to someone she prefers tom. If this (9) is violated if and only if Bob is married to Brenda,
is not true, clearly, (m, w) blocks the current marriage. and Debbie is married to Carl or Alan, which is exactly
the stability condition for the pair (Bob, Debbie).
Let us consider the matching constraints first. How do
we ensure that Alan has a partner? Well, Alan could By applying similar argument, one can write down
only be married to Alice, or Brenda, or Cindy, or Debbie, the remaining fifteen constraints.

...... MalllemaNtil
llill:oll EDLEY
 The problem of finding a stable marriage now programming problem, a problem which has been well-
reduces to finding the values of xm,w satisfying the studied in discrete optimization. Typically, the goal is to
above eight equations and the sixteen stability constraints. find an integer vector which optimizes (maximizes or
minimizes, depending on the problem) a given objective
Solution function, subject to 1in ear constraints. Integer program-
ming is a powerful modeling framework that provides
great flexibility to express a wide variety ofprobelms.
Let us now find all possible xm,w satisfying the
On the other hand, this flexibility comes with a price. To
above eight matching constraints and the sixteen
date, there is no known "efficient" algorithm to solve
stability constraints. It turns out that certain variables
general integer programming problems. In fact, it is
are forced to be zero, namely
widely believed that such algorithms are unlikely to
x4,3 =x4,4 =x3,1=x3,2 =xl,4 =x1.2 =x2.2 =x3,2 =x4,1=0;
exist.
while x 4,2 =lAnd the remaining variables are subjected
to the following seven constraints:
A linear programming porblem is syntactically
similar to an integer programming problem, except that
(10) xi ,J +xl ,3 =1
the variables are not restricted to integer values only.
Despite its similarity to integer programming, linear
(11) xi,J + x2.1 = 1 programming is an "easy" problem to solve and has many
efficient algorithms.
(12) xl3, +x2 ,3 +x2 ,4 sl
Can we make use of our expertise in linear
(13) X1,3 + x 2,3 + x 3.3 = 1 programming to solve integer programming problems
efficiently? Sometimes we can!
(14) X2.1 +x2,3 +x2,4 =1
With an integer programming problem (IP), we
(15) x2,4 + x3,4 = 1 associate a linear programming problem (LP), called its
linear programming relaxation. The associated LP is
(16) x3,3 + x3,4 = 1 obtained by relaxing the integer constraints. Suppose we
use an efficient LP algorithm to solve the relaxation. If
We finally arrive at the following three possible
we are incredibly lucky and find a solution x which is
solutions:
integral, we have solved the IP. This follows because
every solution feasible to the IP is also feasible to the
Suppose x 1 3 = 1. Then (1 0) forces xl,l to be zero,
LP.
which in turn fo;ces x 2 1 =1 (by equation (11)).
By (14), x 2,3 =x2,4 =O,a~d so (15) forces x 3,4 =1.
Fortunately for us, we can solve the stable
This specifies a complete solution, which is
marria~e problem using this approach. There are other
M 1 = {(1,3), (2,1), (3,4), (4,2)}. variations of the stable marriage problem. Interested
readers are urged to consult the references.
Suppose X2 , 3 =1. By similar consideration as
above, we obtain a complete solution References
M 2
= {(1,1), (2,3), (3,4), (4,2)}. [/] Gale, D., L. S. Shapley (1962) .
College admissions and the stability of marriage, American
Suppose x 3,3 =1. The corresponding solution is Mathematical Monthly 69 9-15.
[2] Gusfield, D., R. W. Irving (1989).
M 3 = {(1,1), (2,4), (3,3), (4,2)}. The Stable Marriage Problem : Structure and Algorithms, MIT
Press, Massachusetts.
Since exactly one of Xl,3,x2.3'x3,3 is one (by equation [3] Roth, A. E., M. Sotomayor (1991).
(13)), the three cases discussed above are exhaustive. Two-sided matching: a study in game-theoretic modeling and
Thus, we conclude that Marx's problem has exactly three analysis, Cambridge University Press, Cambridge.
stable marriage solutions. [4] Teo, C. P., J. Sethuraman (1997).
LP based approach to optimal stablematchings. Proc. of the
8th ACM- SIAM Symposium on Discrete Algorithms 710- 719.
3. Linear programming and the
stable marriage problem Mr Jay Sethuraman is a graduate student in operations
research at Massachusetts Institute of Technology. His
research interests are in applied probability and optimization.
In the previous section, we formulated Marx's
problem as a problem of finding value of boolean Dr Teo Chung Piaw is a lecturer in the Department of
variables satisfying certain linear constraints. We can Decision Sciences, Faculty of Business administration at
do the same for a general stable marriage problem. The NUS. His main research interests are in optimization, logistics
formulation presented here is an example of an integer and operations research.

M EDLU
olbr~ihtll m