1

Advanced Operations Research

Prof. G. Srinivasan
Department of Management Studies
Indian Institute of Technology, Madras
Lecture - 11
Complexity of Simplex Algorithm (Continued)
Integer Programming
(Refer Slide Time: 00:17)

In the last class we looked at the Klee and Minty problems, and said that the Klee and Minty
problems would take 2 to the power of n minus 1 iterations if there are n decision variables in the
problem. We also saw that there could be indirect ways of overcoming this difficulty particularly
for the Klee and Minty problems. Those ways of overcoming the difficulty could be, to have a
different starting solution or it could be a different entering variable rule, or scaling the variables
could also get us out of the Klee and Minty difficulty.
We should keep in mind that the idea of the Klee and Minty problem was not to solve the Klee
and Minty problem, but to expose the fact that simplex algorithm particularly when you use the
maximum or the largest coefficient rule: the C
j
minus Z
j
rule, can turn out to be exponential for
certain types of problems.
2
The algorithm is called a polynomial algorithm, if the time taken to solve is a polynomial
function of the problem parameters. It becomes non-polynomial; it becomes exponential when
time taken can be shown to be an exponential function. At this point, we are not going to go
deeper into complexity, but based on the Klee and Minty example; understand that simplex can
take an exponential number of iterations, particularly when you are looking at Klee and Minty.
Very few examples of Klee and Minty have been published. There could be many more which
have not been published yet. But this was enough to show that simplex is a worst case
exponential algorithm and it is not the worst case polynomial algorithm.
(Refer Slide Time: 02:32)

The fact that it could be exponential was already evident by the algebraic method which will take
superscript nC
m
iterations or it evaluates all the superscript nC
m
possible solutions. One could go
back and perhaps argue that there are no Klee and Minty examples for the largest product rule or
the largest increase rule. But it would not be too difficult to create an equivalent Klee and Minty
example, because they have been created for the largest coefficient. The main understanding is
the fact that, simplex is an exponential algorithm, when it comes to looking at it, in terms of
complexity.
The next question that people started asking was can we have polynomially bounded algorithms
to solve linear programming problems? In this course, we are not going to go very deep into the
3
understanding of the polynomial algorithm. We will just give an outline or an overview of at
least one of them; also try to explain some of the issues related to this.
When this question was raised after the Klee and Minty examples were published, then people
also started arguing that, simplex however was an excellent algorithm when it came to average
case performance. It is only for the very small amount of Klee and Minty and related problems it
behaves in an exponential manner. Whereas when you test it on normal problems, problems that
are taken from real life, problems for which a linearly programming solution is applicable,
meaningful, practical, can be implemented and so on.
In all those cases, simplex was only taking about 3 m by 2 iterations on an average and very
rarely going up to 3 m. So simplex, in average case was a very good algorithm. However,
average case performance of algorithms is meant for people who practice and in worst case,
meant for people who developed. So theory still felt that, simplex is an exponential algorithm.
Now the question was, are there polynomials algorithms and how do we develop that. When this
thing was going on, there was also another issue that was being discussed which is called solving
inequalities. Remember that, even though all the linear programming constraints are inequality,
there could be equations, but predominantly all of them are inequalities. We never tried solving
them directly. We converted the inequalities to equations and then we solve them. That is
because we know or we have learnt how to solve equations, whereas we have not learnt how to
solve the inequalities.
Solving inequalities simply means, given a set of inequalities can we find a solution?
For example, if we give you say four or five inequalities of two variables. For example: X
1
plus
X
2
less than and equal to something; 3X
1
plus 2X
2
greater than or equal to something. So, you
define a certain number of variables and certain number of inequalities. Now solving this
inequalities simply means, finding an X
1
, X
2
, which satisfies all the inequalities. Or if there are n
variables, finding even one X
1
, X2, X
3
, X
n
which satisfies all the constraints is called solving
inequalities.
When you represent the solving inequalities problem and solving equations problem, one may be
tempted to think that solving inequalities is actually easier than solving equations. If you take a 2
4
by 2 set and plot the inequalities on the graph, but if you end up getting a feasible region you can
actually close your eyes and put your finger at any point in the region that will satisfy all the
constraints, that is the solution to inequalities whereas you cannot do that to equations.
Most of the time there is only one point which satisfies all the equations. You could have other
cases like, no solution, infinite solution, and so on. In practice, actually speaking, solving
inequalities is not easy; people have not developed an algorithm that could solve inequalities, a
polynomial algorithm that could solve inequalities, till 1979 or so.
When this issue is being discussed, the ellipsoid algorithm was developed in 1979 to solve
inequalities. It was a polynomially bonded algorithm to solve the inequalities. We will not go
into too much of detail on the ellipsoid algorithm, but we will give you some basics of the
ellipsoid algorithm. If you make an assumption that there exists a feasible region in certain
number of dimensions. If you initially make an assumption, that the given set of inequality has a
solution, which means there exists a feasible region for this. It is only a matter of finding a point
which is inside that feasible region.
The basic idea of the ellipsoid algorithm is to try and imagine, a very large the biggest possible
ellipsoid of the same number of dimensions, as the number of variables that you have. For
example, if your inequalities have 10 variables, irrespective of the number of constraints, you
should imagine a very huge ellipsoid which is of 10 dimensions. Then create such an ellipsoid
and locate its centre. When that center satisfies all these constraints that you have, then you have
solved it. If you have not, then the underlying assumption is, if the existing feasible region is the
centre of the ellipsoid does not satisfy the existing feasible region, and if such a region exists, it
is always possible to put that region in a smaller ellipsoid, which is proportionately reduced in
terms of volume. There are lots of governing equations and mappings and all that which we will
not get into.
5
(Refer Slide Time: 08:57)

The underlying assumption is, to just to put it on the black board. For example: if this is your
region and if this is your huge ellipsoid, it need not be a circle. If it is a huge ellipsoid and this is
the centre and it does not satisfy the feasible region, then you can always make an assumption
that, we can put something called a half ellipsoid, which will be like this, whose center in this
case satisfies. If the feasible region is here, then the half ellipsoid center also will not satisfy.
How do you get this half ellipsoid from the bigger ellipsoid is a matter of a lot rigour; we mean,
so much of mapping, so much of equations, you cannot arbitrarily choose half ellipsoid. It has to
be proportionately reduced. All those, there is a lot of theory, which we are not going to look at
today. We are only going to see the principle in which the algorithm was developed. You try to
put it in a half ellipsoid and then try to take the center of the half ellipsoid. If the center of the
half ellipsoid again it satisfies this, then you have solved it. If it does not put another half
ellipsoid from the given half ellipsoid and keep doing it, till it satisfies. This is the basic idea.
There are a whole lot of issues associated with it. The most important thing being, if there is a
solution, then you have to prove that you will reach that point in polynomial number of steps and
that is the most important thing. All those things were done and we are not going to see how they
were done, but all those things were done, so that, this algorithm was developed, to solve
inequalities.
6
This algorithm is called the Khachiyans algorithm. It came in 1979 and Khachiyan proved that,
there is a polynomially bounded algorithm to solve inequalities. This was a major development,
particularly when people were thinking about whether we could develop a polynomial algorithm
for simplex. The moment this was available, of course, there were other issues. For example, this
is under the assumption that the given set of inequalities has a feasible region. There are other
issues like, what if it does not have, what if it is not solvable. All those are, in some sense, a little
more involved than even trying to understand the algorithm which will solve the case where a
solution exists.
We will right now leave that; we will only take the learning from this part. The principle is you
start with a largest possible ellipsoid, get it centered. Center does not satisfy you, assume that
there are going to be two half ellipsoids and it will be in one of them. Then try to get that. It is
not that you made 2 into 4, 4 into 8 and so on it, becomes exponential, but then you choose the
one and gradually reduce the volumes of ellipsoids, identify the centres, and then keep
proceeding till you are able to get it.
When this algorithm was developed, people said, we can use this algorithm to solve linear
programming problems. How we do that?
(Refer Slide Time: 12:03)

7
The linear programming problem is actually an objective function plus constraints. If we hide
this, then this becomes solving inequalities, but linear programming is something more. Here,
this algorithm takes care of finding just one point among the many that may be available which
satisfies this. But on the other hand, what you want is among the infinite or many available, you
want the best. So you cannot directly apply this algorithm to this. What people said is, let us do a
small trick and get it. The trick goes like this.
(Refer Slide Time: 12:50)

If we have a linear programming problem, which is maximise CX subject to AX less than or
equal to b, X greater than or equal to 0. Now its dual will be, to minimise Yb, subject to say AY
greater than or equal to C, Y greater than or equal to 0. You need not worry so much about the
matrix notation, consistency of the matrices and so on. We will just take it crudely, that this is
how the dual will look like.
What will happen is, your duality theorems have told you, that if there is an optimum, at the
optimum, the CX star should be equal to Y star b, because the values of the objective function
will be equal for both the primal and the dual at the optimum. Your weak duality theorem also
tells you, that for every feasible solution to the primal and to the dual; this Yb is greater than CX.
All you need to do is, you need to put Yb either equal to CX, or more importantly less than or
equal to CX. So the optimum to both, if you solve this set of inequalities, you have actually got
8
the optimum to both the primal and the dual. To repeat again, from the weak duality theorem, Yb
is greater than or equal to CX, but at the optimum Yb is equal to CX. You change the constraint
to, Yb less than or equal to CX. So the only point that will satisfy, the only XY that will satisfy
this, is the optimum. Because you are dealing with inequalities you do not want to put an
equation because you may have to again convert the equation into two inequalities.
You just take it as the opposite. This is the opposite of the weak duality theorem result, but this is
consistent with what we are trying to do. So add this one, so when you add this one, you get a set
of inequality. The objective function does not come any more, so you get a set of inequalities. So
you are trying to solve a set of inequalities, apply this algorithm and try to solve it.
People said if this can solve inequalities in polynomial time, now this can solve linear
programming problems in polynomial time. But then there were another issue. The issue is very
simple. If this system has a solution, then there is no feasible region for this. There is only one
point, there is no feasible region. This algorithm had to be modified to look at a situation, when
the set of inequalities has only one solution, which is unique, if it exists. Whereas, in the general
case of trying to check whether the centre satisfies the region, now it is more like defining the
ellipsoids in such a way that ultimately the centre will coincide with that unique point; so that
had to be looked at.
Obviously, this algorithm was modified and then there were perturbation schemes introduced.
You try to put a small area around this point and perturb, and then check that you define a small
feasible region around this optimum point. Then if it comes into that, you say it is, you have to
define the epsilons, and then you have to define the number of iterations in which it will
converge into that. All those things happened and then people were kind of happy and convinced
that the ellipsoid algorithm is capable of solving the linear programming problems in polynomial
time.
This was ascertained and the ellipsoid algorithm was the first algorithm in principle, to be
developed, not necessarily or not with the primary objective of solving LP, but then it could be
adapted to solving LPs, in providing an algorithm that solves it in polynomial time. Let us try to
solve some more problems in this. Now there is a theoretical proof that ellipsoid algorithm can
solve it in polynomial time. Let us try out some problems.
9
If you go back and try out the ellipsoid algorithm, for a simple problem like: maximise 5X
1
plus
6X
2
, subjected to 2 sets of constraint and all that. Ellipsoid algorithm will take at least about
1000 iterations, for it to converge whereas simplex will do it in 2 iterations.
(Refer Slide Time: 17:13)

For example, if your optimum is X
1
equal to 5, X
2
equal to 4, say what will happen is ellipsoid
will kind of alternate. It will start alternating between say, 5.5 and 3.5 on one hand then 4.5 and
4.5 on the other. It will look at 5.5, 3.5 here; it will look at 4.5, 4.5., it will move a little bit closer
and then it will slowly try to converge towards this 5 and 4.
What you have to do is you need to put an epsilon and say that, the moment it comes within a
certain range, you terminate. But you do not know what that epsilon is. You run into those kinds
of difficulties. What was evident from such a study, not by a single person but many people was
that this has a very poor average performance. In terms of average performance, simplex will just
beat the ellipsoid algorithm hands down for the kind of problems that people were solving.
People are not the researchers, people were everybody, we mean in linear programming
terminology, problems even with 500 constraints or 1000 constraints are considered to be small.
It becomes large only when you are looking at a million constraints, a million variables and
things like that. If you are looking at 50,000, 100,000, you are somewhere in the medium, today.
10
That is a kind of problem size that is associated with the definition of it being small, large,
medium, and so on.
We are not looking at the 2 by 2s and 3 by 3s that we are familiar with. If you want look at a
real life application that comes, say communication, or comes from scheduling in airline, or
comes from forecasting, demand in an airline industry, or in an FMCG firm or whatever be the
example, which can be modeled as linear programming related problem, then you are running
into tens and thousands or may be hundreds and thousands of variables. Even when it comes to
solving medium sized problems, the simplex was beating the ellipsoid algorithm, and ellipsoid
algorithm had a kind of a poor average scale performance. But it did have a very good worst case
performance which is polynomial.
(Refer Slide Time: 19:40)

When this was going on then came a most important break-through which is called the
Karmarkars algorithm, which was also a kind of an interior point algorithm. If you go back to the
simplex, simplex is not an interior point algorithm. Simplex is basically a corner point algorithm.
It start with one corner point and keeps moving along the corner points in a certain way so that
you reach the optimum because you have a result that, the optimum solution is always a corner
point solution.
11
(Refer Slide Time: 20:30)

For example, if this were a feasible region. In a 2 by 2 case what simplex would do is, it would
start here, it either moves here or here, here and then check whichever is optimum (20:38). It
keeps moving along the corner points. For example, nothing prevents you from starting with the
point here, knowing fully well that there it is not a corner point, provided you can move in the
right direction from this. Not necessarily in one jump, go to this corner point or this corner point
and so on, but keep jumping in a certain way, till you finally reach the corner point. This was the
basic idea. Now such algorithms are called interior point algorithms.
In a 2 by 2, you cannot actually appreciate the whole thing. If you can think of a three-
dimensional representation then what you can do is you can be in an intermediate point then you
try to move either in x or in y or in z up to a certain limit and then keep moving that way. It is not
in one jump you can get here, you still have to move progressively, you have to make multiple
jumps and then slowly move towards the corner point that is optimum. That was the basic idea of
the Karmarkars algorithms, wherein there are mapped problems. You map all the xs into the ys
and then you try put something into a huge sphere of higher dimensions, instead of ellipse and
ellipsoids, use circles and spheres higher dimensions and then you solve the y problem and then
get the optimum solution to the y in a certain way, and then map it back to x.
12
Your x moves from certain place from inside the feasible region to some other point inside the
feasible region. Keep doing this back and forth, till you reach the corner point. Reaching the
corner point is not exactly sitting on the corner point, it will be like you are within some 2 to the
power minus n of the corner point and then you will terminate it. It is not that, if the answer is (4,
5), it will beautifully sit at (4, 5) and tell you that this is the optimum, with simplex would do. It
would always converge a 3.998 and 5.001, if you are giving 3 digit accuracy or whatever. If you
want 25 digit accuracy, then it will be some 2 to the power something error and so on. That was
the basic idea of the Karmarkars algorithm.
There were lots of similarities and differences between the ellipsoid algorithm and the
Karmarkars algorithm. There are lots of research papers which talk about, may be the basic idea
is from the ellipsoid, some people would say, it is not. There are other related issues about the
origin. Both these algorithms came from scientists who worked in different countries which
compete all the time. So there were issues of origin and so on. But that is beyond our discussion.
Our discussion is only to look at algorithms that help solving linear programming problems.
Today Karmarkars algorithm is accepted which is exclusively developed to solve linear
programming problems unlike the ellipsoid algorithm; perhaps a little faster. Both of them have a
worst case polynomial nature of the algorithm. The important thing is that, the Karmarkars
algorithm should start with an interior point, which means the interior point implies that the point
is inside the feasible region. What Karmarkars algorithm would do is, it will solve particular
linear programming problem for which the value of the objective function is 0 at the optimum.
The point 1 by n, 1 by n, 1 by n, 1 by n is feasible and so on.
There are certain conditions associated with the Karmarkar problem. The problem that he used as
a base problem to develop the polynomial algorithm, would satisfy certain things, at least some
of which are the value of the objective function is 0 at the optimum, and the point 1 by n, 1 by n,
1 by n is feasible. He also devised a mechanism by which, any problem that has an optimum can
be converted into the Karmarkars form, by suitably writing the dual, by suitably scaling some
variables. It is a slightly complicated way to convert a given problem into a Karmarkars problem.
For example, if you take a simple 2 by 2 that you are very familiar with. If you convert it into the
Karmarkars form, you will end up getting some 10 constraints and some 10 variables. That does
13
not matter; it is still theoretically a polynomially bounded algorithm. Once again, you have to go
back and look at in terms of average case performance and complexity; simplex was still better
than Karmarkars algorithm for small and medium sized linear programming problems. Only for
very large problems, which run into millions of constraints and tens of millions of variables and
so on the Karmarkars algorithm could be faster than the simplex. This is where the theory is,
with respect developing efficient algorithms for linear programming.
As far as this course is concerned, simplex will be the algorithm that you will use to solve linear
programming, as well as integer and the other kind of programming problems. As people who
learned operations research, one should be aware of the limitations of the simplex algorithm and
the availability of algorithms that are theoretically very superior. Depending on the size of the
problem that is being solved, may be faster than the simplex, even for average case, if the
problem size is exceptionally large. For the kind of small and medium problems, that people
encounter in practice, simplex is still the best, when it comes to average case performance and
general average running performance. With this we will wind up the discussion on how fast or
how efficient the simplex is.
We have not looked at, for example, there are obviously results available, if the given problem
with hundred thousand constraints is infeasible or if the given problem is hundred thousand
constraints may be unbounded, one does not know. But then there are results available. Both for
the ellipsoid as well as the Karmarkars algorithms there are efficient versions, the moment the
Karmarkars algorithms came in 1985, nearly for ten years there were papers with extended
Karmarkars algorithms to various situations discussing implementation of the Karmarkars
algorithms, discussed their own experience in solving the large real life problems and so on.
There are plenty of researches available and people should still be working in several places to
develop polynomially bounded algorithm better than Karmarkars in terms of performance.
As far as we are concerned and this course is concerned, we would look at simplex and keep
simplex as the algorithm to solve linear programming problems. We will wind up the discussion
on Klee and Minty problems, as well as the goodness of the simplex algorithm.
We will move on to the next topic today which is called the integer programming.
14
(Refer Slide Time: 28:04)

We will start integer programming; we will also look at a few things here. I will give you an
introduction to integer programming and explain why it is different from linear programming.
We will also look at some formulations. We will also try to see that certain nice things can be
modeled using IP. Then we will go to solving integer programming problems. We will first look
at this 0 1 case and then we look at the Gomory cutting plane algorithm and then some all integer
algorithms. That will be the topic of integer programming.
What is an integer programming called?
It is a linear programming problem plus the condition, that X
j
are integers. Moment you add an
integer restriction it becomes an integer programming problem.
The question is can we define integer programming problem as one which has a linear objective
function, a set of linear constraints, non-negative restriction plus an additional condition that
variables are integers. The answer is no. It is actually beyond that.
An integer programming problem can be a linear integer programming problem, it can be a
nonlinear integer programming problem; it can be both. In terms of objective function, it can be
linear integer; it can be non linear integer. But common terminology would mean that, if you say
it is an integer programming problem, it means, it is linear integer; unless otherwise stated,
15
integer programming means linear integer, but it could be nonlinear integer. You could have
nonlinear objective functions.
(Refer Slide Time: 30:11)

What do we mean by an integer variable?
Integer variable, for example, if we have a feasible region to an LP like this, now it simply
means, like for example, if this is 0 0 and say this is 5 0, this is 7 by 2, 7 by 2 and say this is 0 4.
Now what does integer mean? Integer would mean that, for example, a point here, say (2, 2),
which is an integer point, is feasible, whereas, this point itself is not feasible to the integer
programming problem, because it is not integer. In terms of the integer restriction of the
variables there are two categories.
One category is called an all integer problem; where all variables are forced to take integer
values. You have zero-one problems, where the variable can take either a 0 or a 1 or binary. It
cannot take value 2 or 3 or whatever. In the assignment problem that we have seen in the earlier
course, this is zero-one problem; X
ij
is 0 or 1. But then we solved it as a linear programming due
to other reasons but the problem as such is a zero-one problem where variables take either 0 or 1.
Then you could have mixed integer programming problems. Normally when you say mixed
integer it means you have some variables that are greater than or equal to 0 and some variables
16
that are integer. When you say a variable is integer, it means you are looking at this category. You
have to explicitly say that it is a binary variable, if it is a binary variable. When you say the
variable is integer, it does not include binary. So variables can be of three types: continuous,
integer and binary.
Continuous is greater than or equal to 0, which is in linear programming. Integer means, it takes
integer values 0, 1, 2, 3, 4, any integer. Zero-one means it takes either a 0 or 1 or binary. When
you say the problem is mixed, it could be any two out of the three or all the three. Once again,
normally when you say mixed, it means, it is continuous and integer. The zero-one, unless
otherwise stated does not figure in a mixed problem. But in reality, a problem may have all three
types of variables; continuous, integer, and binary, it could be anything.
Typically you can have something like six or seven different types of IPs. You can have a linear
integer program, all integer linear, linear objective function and all integers; you can have a
linear and a mixed integer; you could have a linear and a zero-one; you could have a nonlinear
integer, nonlinear zero-one, nonlinear mixed and so on. The moment you get into the nonlinear
segment, it becomes tougher and within that if you get into mixed, it is much tougher. The only
solvable thing in the nonlinear is a nonlinear binary. Nonlinear binary is reasonably easy because
if the variable X is binary any power of X is also binary. It is only 0 or 1. Therefore it becomes
simple. Only difficulty is, when you have a product, you get into trouble there. As long as you
have an X that is binary, any X to the power n is also 0 or 1. So we will look at that as we
proceed.
Typically a linear programming problem has all these things. Let us look at the first one, the
simplest one, with a linear and an all integer. This would simply mean a linear programming
problem with the additional restriction that variables are integers. In a 2 by 2 case, you would go
back and say, this is my LP solution. Now we know that this point is not feasible to the integer
problem. All we want to do is to find out that integer point, which has the best value of the
objective function. Assume maximization, say we want to find that integer point, which
maximises my objective.
17
(Refer Slide Time: 34:44)

What are the issues?
The first and most important issue is, if this happens to be a corner point, if we leave the integer
restriction, relax it into an LP and solve it, and if we say that this is the corner point solution. The
normal feeling that one would have is, that the integer point that is closest to this corner point is
likely to be optimum. A feasible integer point which is nearest to the corner point is likely to be
the optimum, is a very common assumption that we can make. That assumption need not be
correct all the time. You can have a problem where this could be optimum. You can have an
objective function where, a point, a different corner point; a different integer corner point could
be optimum. It is all a matter of how the constrains are defined and how the objective function is
defined.
It is not absolutely necessary that the feasible integer point closest to the corner point is
optimum. That hypothesis need not be correct all the time. You could get a different integer
corner point or a different integer point which is away from the existing LP. For example, this
could be optimum, if this is 5 0, then 5 0 could be optimum. So that is number one.
Now, if we compromise and say that, we are finding it difficult, we know that the nearest integer
point to the infeasible corner point need not be optimum to the IP. However, let me try and find
an approximation, after all it cannot be a bad solution because the LP optimum is here, we just
18
put an integer restriction, it cannot be a very bad solution. It can be a descent solution, so we
want to compromise and we find out that point.
Now we are compromising, we are not looking at the exact optimum, but we are compromising
somewhere. We start doing that, then what we have to do, is to round off this. So, rounding off, is
most of the time seen as a very efficient strategy to solve IP. When you round it off, what will
happens is, if you have n variables, then each variable is rounded off to its upper limit and the
lower limit. You can have 2 to the power n rounded solution that are possible.
2 to the power n is not a polynomial number in n, as n increases, 2 to the power n is exponential
and therefore rounding off is not a polynomial way of solving it. But as you are always interested
in algorithm that is polynomial, you will not take an algorithm the moment it is established, that
it is exponential. Even if you do an exponential search and get these 2 to the power n and get the
best, it is still not optimum; there is no guarantee. Therefore rounding off is not a very good
strategy, even though it is followed by many equations.
You will also know that out of these 2 to the power n, there could be a very large number that
would be infeasible, because when you start rounding the mass on the higher side, for example,
you will get points like this. This you will have four rounded points and in this case, it can turn
out three of them is infeasible and only one of them is feasible. You can also have a situation
where none of the rounded points are feasible, but there is an optimum somewhere else.
19
(Refer Slide Time: 38:39 min)

For example, you can create a feasible region like this. This is the LP corner point and let us say
these are the four rounded points; none of them will be feasible to the IP. You could have an
integer point somewhere here, which is optimum. These are all extreme examples. However,
they are relevant, because we are trying to understand the principles involved. There are three
aspects to it; one - rounding is exponential and therefore difficult to do; two - the best feasible
solution among the rounded, need not be optimum; three - you can encounter situations where all
the rounded solutions are infeasible, but there is an optimum somewhere. Therefore one should
not do this. One should try and formulate the problem separately and solve the problem
separately. What are the other implications? If we have an algorithm which will directly solve
integer programming without using ideas from linear programming, it is another issue.
20
(Refer Slide Time: 39:58)

If you look at it, this discussion was based on the fact that you relax the integer assumption.
Solve the resultant LP and then try to get the integer optimum from the linear programming
optimum, that was the hypothesis or that was the assumption which led us to the discussion.
Let us go back and see: let this be the feasible region,, what we actually want to do? If we can
identify all the integer points in this feasible region and then we substitute all the integer points
into the objective function we can get the best, which is optimum. It is only a question of trying
or to be able to find out all the integer points which are there. For example, your integer points
are, the moment you start looking at this way, these are your integer points and so on. Your
feasible region for the IP, is not continuous; it is a collection of discrete points. You see, this
point is infeasible, this point is infeasible. Whereas, in linear programming, your feasible region
is convex and the entire results of linear programming is built around the convexity property of
your feasible region.
You cannot directly apply results from linear programming and say we can do this, look at the
corner point. You need to still be within the linear programming territory and do things in such
way that the convexity of the feasible region is not lost. Then somehow you try to get corner
point to an LP, which is integer. Most integer programming algorithms are developed with that
thing in mind.
21
We will always be solving LPs, because we do not want to lose the convexity property of the
feasible regions, and somehow do something such that the LP optimum becomes integer. Please
remember, that in all IPs if you relax the integer assumption and solve it as an LP and if the LP
optimum happens to be integer valued we have solved the LP. That is the best thing that you can
hope for whenever you are solving IP. You solve it as an LP and you hope that the LP optimum is
integer valued so that straight away you have to solve the IP. If it is not, then you still do things
which are under the purview of LP, keep solving LPs in successive iterations and somehow get
the optimum LP to be integer valued so that it is also optimum IP. Algorithms to solve integer
programming problems are developed that way. After this, let us go back and try to formulate
some practical situations which are integer programming problems.
(Refer Slide Time: 43:11)

We start with one simple example called the grouping problem. Let us define the grouping
problems and see whether we can complete the formulations.
22
(Refer Slide Time: 43:27)

Let us assume that we have five points which are like this. If you look at it normally and if we
ask anyone of you to come up with two groups, you will quickly go back and say, these two will
form a group, and these three will form a group, very obvious. This is under the assumption of
minimising the distance among the groups.
How do we first formulate this problem and then try and solve them, to get this, how can we
formulate this?
One of the ways of formulating it is this, even before we define the decision variables let us look
at this.
23
(Refer Slide Time: 44:13)

If we know the number of groups and if we want to solve or identify the groups, for a known
number of groups, say two in this case. What we can do is we arbitrarily define two seed points
called, S
1
and S
2
. The moment you defined two seed points S
1
and S
2
, we will go back and try to
find out the given points 1, 2, 3, 4, 5; we will try to assign them or cluster them around the seed
points. That is a very simple thing to do. These two will go to this seed point; these two will go
this seed point.
Now what is the difficulty there?
The difficulty is not in assigning or clustering around the seed point, but the difficulty is in
creating the seed points. For example, if we had arbitrarily created, my S
1
to be here and my S
2

to be somewhere here, then you will see that all the points will cluster around S
1
and you will
have a null cluster around S
2
.
When we say that we want a group, in that case the first group will have all the 5 and the second
will have 0, which means you have only one group. When we say that we want two groups, a
group is characterised by the presence of at least one element. Therefore, what we do is, we try to
not identify seed points that are different from the given points. We will simply go back and
assume 2 out of the 5 given points as seed points, so that we know for sure that we will not get
null groups.
24
(Refer Slide Time: 45:50)

I could assume this one to be S
1
and this one to be S
2
. So we are guaranteed that, this 1 will form
a group; 5 will be in another group, and the rest of them will get mapped into different groups. If
we are going to form a group based on the distance criterion, then we assume that we know the
seed points, we can quickly do a distance criterion and form the groups. What we want to do is, a
good way by which we pick 2 out of the 5, or in general if we want p groups, we fix p out of the
n. as seeds. The moment we fix p out of n as seeds, because of my objective function which is to
minimise distance, we will quickly cluster the rest of the points around the p seed points. These p
seeds points are called medians or nucleus points, nucleus, nuclei and so on. We will call them
medians.
25
(Refer Slide Time: 47:03)

The formulation will look like this: we have i points, i equal to 1 to n. We will first put this
condition: X
jj
equal to 1, if point j is a median. X
ij
equal to 1, if i goes to median j; equal to 0
otherwise. The formulation looks like this. This is under the assumption that we know that p, the
number of groups that we want.
(Refer Slide Time: 47:53)

26
What we will do is, my first constraint will be, sigma X
jj
equal to p. 2 out of the 5, will take 1
and the remaining 3 will take 0. Whichever j is a median will take that X
jj
will be 1. If it is not a
median, then it will be 0. X
ij
should be less than or equal to X
jj
, because if we are attaching for
example, if point 1 goes to median to point 2, then point 2 should be a median. If 2 is not a
median, then 2 cannot build clusters around itself.
You can build clusters only around the median. So X
ij
less than equal to X
jj
would imply, that any
point that is not a median, can only be clustered around a median. Sigma X
ij
equal to 1 implies,
that any i that is not a median can go to only one cluster. Then your X
ij
is 0 or 1 and then you
minimise double sigma d
ij
X
ij
and your d
ij
is 0, the distance between the point and itself is 0.
You assume that for the given 5 points, you have a 5 by 5 distance matrix and if there are points
here, they will be symmetric, so you can define it right. You will get a formulation like this. This
is one example of how integer programming can be used to formulate a reasonably complex
problem. You will know that the grouping problem is actually a very difficult problem. This is
called the p median problem or a k median problem, depending on whether you call this as p or
k. This is an integer programming formulation of a grouping problem under known number of
groups. You cannot solve this if the p is not known, the only thing you can do is you can solve
for p equal to 2, 3, 4, 5, up to n minus 1 and then do something. But this is under the assumption
that you know the number of groups that you want and having given the number of groups you
want to get the best grouping that is possible out of this.
We will look at other types of formulation in the subsequent lectures.