Power System Operations and Control

Prof. S. N. Singh
Department of Electrical Engineering
Indian Institute of Technology, Kanpur

Module - 4
Lecture - 3

So, welcome to lecture number three. In this lecture I will discuss about the various
optimization methods, techniques to solve the power system problems.

(Refer Slide Time: 00:32)

To begin with, just I will try to introduce the optimization what is the optimization every
person everybody in this life always try to optimize even though in terms of your daily
routine works. For example, if you are traveling from one point to another point, always
you find the minimum distance, and then you try to cover means you try to optimize your
travel subjected to your various constraints. Similarly, suppose you are having some
money in your pocket always you try to optimize that money so that you can utilize that
resource very efficiently and optimized way. So, in the power system also in always we
try to use the optimization method to achieve our various objectives.

So, there are various optimization methods existing and they can be categorized based on
again several ways based on whether they are linear or non-linear problems. So, we can
linearize that the linear programming optimization methods or we can go for your non-
linear programming methods. So, in these linear programming methods our objective
function where we are going to optimize it must be linear and linear and your constraints
that are constraints must also be linear. So, once you have formulation your objective as
well as the constraints if they are there. So, it must be linear then we can say our
optimization is your linear programming approach, we have to apply and that is called
LP method sometime very powerful and it is very fast.

Another is your non-linear of programming in which at least any of either constraint or

objectives or both will be non-linear function then it is called your NLP method, and
then we have to use the NLP method for optimization of this method. Most of non-linear
programming method use techniques on the basis of information obtained by the gradient
to reach the optimal point and these methods are termed as the gradient methods. So,
gradient methods are very very popular and they use some sort of information to move
and that information basically based on the gradient to reach the optimal point.

For example, in this figure you can say if you are starting here with some x0 again you
must also know for getting the optimal solution, always we start with some initial guess,
and normally that guess should be very close to your optimal value. Otherwise, there
may be possibility that your optimization method may fail to get this optimal solution it
may give some local solution it may give some optimal solution and so on so forth. So,
here thus we use the gradient method, we start with the some initial point x0, it moves to
another point your x1 here. Then, it calculates another gradient, and then it will move to
x2 and then finally, it is coming to your x star that is your optimal point.

So, it is based on your gradient method and it is moving like this and reaching your
optimal value here. Now, to again to understand this, to go in the detail in the non-linear,
let us suppose we are having a function. Now, question, why we start with the initial
guess? That is very close to that in any optimization methods there may be some various
optimal points. Some may be your global, some may be your local, for example, let us
suppose this is your function, now you can see this is one maximum this is another
maxima and this is the highest one, so there may be the various local optimal points.

So, here we can say the local optima and another is your global means that is one highest
among other let us suppose again here we are having them, so what is happening? We
have the several peaks, so these are the maxima value, these are your the minima values.
So, in any this is function let our function of x, let us go for one variable, this function is
going in this way, so what happens here? We are having only one global optima, this
maxima and only one global here in minima. So, if your objective is to minimize this
function, then we are optimization that we have to give global minima here.

If your objective is to optimize for the maxima, then your global maxima will be here,
there are several local minima’s and maxima’s and therefore, there may be possibility
you may stuck at some global local point optimization problem. That is why we start
very close for example, if you are starting from here let us suppose your x0 is here, you
are moving with the some direction here, and here there may be possibility you can land
up here, and you will get here the solution. You are not coming here and this is giving
your local optimal solution, so for this we have to start somewhere here very close to
your global point and then we can reach here at this optimal point.

So, this selection of initial guess is very very important and it also depends upon method
to methods some methods give some direction some ideas to check it, but mostly if you
are giving here your starting point very close to your optimal value. Let us say very easy
that you will get a minimum time and you can get your global optimization. So, this is
important of this x0, another requirement for the optimization is that we should have the
function of the continuous because most of the classical methods I am talking here about
the non-linear programming approaches. Most of the classical methods the method that
which are conventional methods, I can say Newton’s method, I will discuss later.

They require some derivative information, if your function is not differentiable at any
point, means it has some discontinuities. Then, we cannot apply the NLP the
conventional NLP, then we have to go for some heuristic or we can go for some non
conventional application of AI technique. We can apply like genetic algorithms some
related dynamic say other programming and so on so forth. So, it is your let us see now
gradient conventional method.
(Refer Slide Time: 07:31)

I am just discussing at optimum seeking gradient based search methods using the
following quantity for the new update of the state variable. If your x is your value here,
the bar is showing the vector because your function may be of the several variable it is
not only x1, it is our x, it may be x1, x2 xn. So, we can write your update value of x that is
the states that we are obtaining so that we can get our function f x here that should have
the optimal value here. Again, it is maximum minima, it depends again because again the
optima the maximization and minimization are reciprocal. It means if you are just for
example, you are going to say minimization of x function 𝑓(𝑥̅ ), it will be similar to
maximization of function −𝑓(𝑥̅ ).

It means if you are multiplied by minus one your function transfers from minima to
maxima. So, it is immaterial means whether you are going to minimize or maximize only
simple here minus sign is changed and you can solve by using any conventional
methods. So, most of the method basically go for the minimum suppose you want to
maximize then you multiplies that function with the minus one and then you can
minimize and you can apply to the any method. How It is true?, for example, I will show
you, let us suppose a function here, we have here let us suppose minima, so we have this
minimum value this is your 𝑓(𝑥̅ ), this is your x for a single variable I am talking. The
point here it is your minima and we can get here the minimum value x0.
Now, if you want to maximize, then you have to multiply it by (- 1) and you can say
what you are getting here the function that is −𝑓(𝑥̅ ). You are getting this function and
this is your x, so this curve it is just inverted, so it is (-x) I am talking of here, so this
value will be like this and you can get it. So, this means minima and maxima should not
confuse, normally we that is why we call the optimization, we never say minimization or
maximization because both are can say complementary to each other. So, here I was
talking about the updated value of the state variable that is the x I am talking x is a vector
that is (k + 1) means at k at plus 1 iteration, k is a iteration count we are in any iteration
(k + 1).

Then, we can use the value of x in the previous iteration that is the kth iteration plus
minus some α ̅̅̅̅̅̅̅̅
𝑥 (𝑘+1) = 𝑥̅ 𝑘 ± 𝛼𝑀∇𝑓(𝑥̅ 𝑘 ) multiplied by m into your change here del f at
x k what is this let us say first basically this alpha is a scalar quantity, M here is of
matrix. That is your n crossn matrix and this value here is the differentiation, this value x
k is nothing but this differentiation of this function 𝑓(𝑥̅ ) this is vector, so it is a here x at
your 𝑥̅ 𝑘 th iteration. So, this is your partial derivative of function f with respect to x at
variable. So, we can write it in this general form one most of method, they differ only in
̅̅̅𝑘 .
terms of selection of m and alpha at 𝑥0

Now, we have the two type of methods based on this the first type of method follows the
direction of the steepest descent that is the ascent we call and a closely as possible during
the search. We use the direction of the steepest descent, we will see what is the steepest
descent, and it will follow that path means this given the direction here in the second
type of method it use the gradient to guide the direction of search. The search direction is
not necessarily along with the direction of steepest steepest descent, means it will give
the gradient will give the guidance of direction, but it is not the steepest steepest descent.
This method basically utilizes the conjugate gradient technique, and we will see later on,
now let us sees what this is? Your steepest descent method.
(Refer Slide Time: 12:02)

It is also called the first order gradient method in this method, let us we have again the
function 𝑓(𝑥̅ ) here we want to optimize this function; now this function we can expand
using the let us say expansion term. That we can write this ̅̅̅̅̅̅̅̅̅̅̅̅
𝑓(𝑥 (𝑘+1) ) = 𝑓(𝑥̅ ) +
𝜕𝑓(𝑥̅ )
|x= 𝑥̅ . So, based on the tailor series expansion any fractional function at the point
𝑑𝑥̅

(x + k) this function at k bar we can write here in this way. Now, we can ignore the
higher order terms here and we can simplify this, so we can take this term this side we
can move. So, we can write here that we can get this differentiation term we can write
this one this 𝜕𝑥̅ 𝑘 = 𝑥̅ 𝑘+1 − 𝑥̅ 𝑘 .

This is your partial derivative at x is equal to (x + k) sometimes called the gradient. So,
you can say now you are getting this change in f here we are getting this function now if
your objective is to find the minimum value. The left hand side of this expression here
side will be negative we want to minimum value, what will happen? This value will be
less than this value, means we are keep on minimizing you are moving in that direction.
For example, here this is your function, so you are here at 𝑥̅ 𝑘 , now here this is function is
̅̅̅𝑘 ) here, now you are moving here in this direction your ̅̅̅̅̅̅̅̅̅̅̅̅
your 𝑓(𝑥 𝑓(𝑥 (𝑘+1) ) will be less
than your 𝑥̅ 𝑘 . Now, you can say this whole quantity will be negative means this will be
lesser than this value.

It means whole this will be negative which shows that either for the negative this either
we should have this negative or we should have this negative two possibilities. Either we
should have this negative or this negative to have this negative because we want to
minimize this thus the moment is in the negative direction here 𝑥̅ 𝑘+1 if you are
incrementing keep on moving, so this value will be in the negative direction.

(Refer Slide Time: 14:48)

Now, I can say here your ̅̅̅̅̅̅̅̅

𝑥 (𝑘+1) = 𝑥̅ 𝑘 − 𝛼𝑘 ∇𝑓(𝑥̅ 𝑘 ), but it is shown it is showing that
this value is nothing but your 𝜕𝑥̅ 𝑘 = −𝛼𝑘 ∇𝑓. So, if this is negative, you can see it is
here we are going to write 𝛼𝑘 some magnitude we are adding. It means your iterations
that is we are keep on adding with the some constant, and then we are getting this
another important feature that here we are achieving this 𝛼𝑘 is the step length and it is a
positive scalar value at the k th expression. At the same time, the direction of here 𝑓(𝑥̅ 𝑘 ),
will be perpendicular to the contour of the function 𝑓(𝑥̅ ) as shown here you can say this
grad of f this is a contour this is your function let us suppose.

So, the contour of this one this is moving like this, it is in two variable function x1 and
x2, it is moving and then it is your like this value is achieving means this function is
nothing but I can say 𝑓 (𝑥̅ ) = 𝑥12 + 𝑥22 . So, here you can say this function is a 𝜋𝑟 2 and it
is a circle the equation.

Here, the contour will be the circles you can say this is the circle we are having. So, at
any point this grade of f will be perpendicular to this value as I said. So, the objective
function contour this is called objective function contour or constant objective function
curve here means for fix values here this is a equation of circle. So, this is the contour
this is a function x y the variables two variables are there and it will be the perpendicular
of that one. The value of 𝛼𝑘 this one, sorry we are talking the value of this 𝛼𝑘 here is very
critical, and with small value of 𝛼𝑘 the solution time required will be larger where as a
large value of 𝛼𝑘 may result in divergence of solution.

Hence, the optimal value of 𝛼𝑘 is required, so here 𝛼𝑘 we have to take the optimal value
and that we have to decide it optimally, so the optimal choice of 𝛼𝑘 come into the
picture.

(Refer Slide Time: 17:39)

So, let us see the optimal choice of 𝛼𝑘 , and then we have to min minimize this 𝛼𝑘 such
that the condition is that such that here your minimum ̅̅̅̅̅̅̅̅̅̅̅̅
𝑓(𝑥 (𝑘+1) ) here this is the vector
or we can say the minimum here. Now, I can replace this with your minimum of your f
that I can write here 𝑥̅ 𝑘 − 𝛼𝑘 ∇𝑓(𝑥̅ 𝑘 ) here x I can say k this is vector and this is your
form. So, we have to minimize this subjected that we can get the minimum of this value
because we are going to minimize this.

So, here let us take a quadratic function here f, now I can take 𝑓(𝑥̅ ) it is your quadratic
and I am adding this function is your 𝑓(𝑥̅ ) = 𝑥 𝑇 𝑄𝑥 + 𝑥 𝑇 𝑏. So, this is basically nothing
but your quadratic function thus we want to calculate what will be the optimal value of
alpha for this objective function bar term denotes. Again, the x is having the various
variables d is the vector curve representing the linear term here it was b and this is your
quadratic term and Q is your basically the coefficient matrix for this quadratic curve.
Now, I want to calculate the grad this grad, I want to calculate this f x here it will be
nothing but I can write this ∇𝑓(𝑥̅ 𝑘 ) = 𝑄𝑥 + 𝑏 this is again it is also vector.

So, we are getting the ∇𝑓 this, so this is nothing but let us write here this 𝑔𝑘 it is nothing
but I can write the ∇𝑓(𝑥̅ 𝑘 ) at x is equal to x here k bar that is we are trying to write.
Now, our objective is to minimize this function here or you can say minimize this
function here and we can use this value here. So, I can write now this function f here that
is your 𝑓 (𝑥̅ 𝑘+1 ). I can write using the Taylor series expansion it is nothing, but simply I
am putting that value here in terms of this function we are putting. So, we are getting this
1
your simply here 𝑥̅ 𝑘 and now I can write it is (𝑥̅ 𝑘 − 𝛼𝑘 𝑔𝑘 )𝑇 𝑄(𝑥̅ 𝑘 − 𝛼𝑘 𝑔𝑘 )+ (𝑥̅ 𝑘 −
2

𝛼𝑘 𝑔𝑘 )𝑇 𝑏.

Here, your again (𝑥̅ 𝑘 − 𝛼𝑘 𝑔𝑘 ) here this is a quadratic term plus here I can write (𝑥̅ 𝑘 −
𝛼𝑘 𝑔𝑘 )𝑇 𝑏. Now, what I did just I replaced this term here, you can say this was your α
grad just value I wrote here is the𝑔𝑘 at k value. Then, we can write in this fashion now to
𝜕𝑓(𝑥̅ 𝑘+1)
have minimum of we know this = 0. For the sample expression, we know to be
𝜕𝛼𝑘

the minima always the function is here is for the minima the necessary condition that this
should be zero already we have this you know it very well. Again, if you are going for
the double differentiation here if it is a negative, then that value will be the maxima value
at which value we are getting negative.

It means that we will give maxima if it is a positive, it will give your minima at 0 again
we cannot say anything and we can go for further. So, here you can differentiate this and
then we are getting here that is I can say minus 𝑔𝑘 because this is 0. Here, we are going
for −𝑔𝑘 𝑇 𝑄(𝑥̅ 𝑘 − 𝛼𝑘 𝑔𝑘 ) − 𝑔𝑘 𝑇 𝑏̅ = 0

Or if you will solve this what we are going to get if you are going to replace this xk
etcetera you will get here that is−𝑔𝑘 𝑇 𝑄 (𝑥̅ 𝑘 − 𝛼𝑘 𝑔𝑘 ) − 𝑔𝑘 𝑇 𝑏̅ = 0. Or I can say 𝛼𝑘 is
𝑔𝑘 𝑇 𝑔𝑘
nothing but your = 𝛼𝑘 only which I skip from this expression to this expression
𝑔𝑘 𝑇 𝑄𝑔𝑘

vary directly what we can do if you will expand this. While expansion, you will find
some of the term will become 0, for example here I just want to explain this term this
you can remember this 𝑔𝑘 𝑇 𝑔𝑘 , this multiplied by this and so on, I am writing here.
(Refer Slide Time: 23:34)

We can write here that −𝑔𝑘 𝑇 𝑄𝑥̅ 𝑘 + 𝑔𝑘 𝑇 𝑄𝛼𝑘 𝑔𝑘 − 𝑔𝑘 𝑇 𝑏̅ = 0 Now, you can see from this
value from this value and this value you combine together, now we can write from here
we are getting −𝑔𝑘 𝑇 (𝑄𝑥̅ 𝑘 + 𝑏̅) + 𝑔𝑘 𝑇 𝑄𝛼𝑘 𝑔𝑘 = 0. This value what is this value this
value is nothing but your g k if you remember here you can see here this value is the 𝑔𝑘 .
So, I can write simply here this is if this was basically this is negative value.

If this was negative and this was negative, so it is negative and this will be positive. So,
this was your 𝑔𝑘 and then based on that I wrote this expression here you can see this one.

Then, I can say 𝛼𝑘 here, now this update of state vector using the steepest descent
technique for the quadratic function will be given by this expression that is your𝑥̅ 𝑘+1 .
Here, sometimes I am writing 𝑥̅ 𝑘 at lower or upper basically 𝑥̅ 𝑘 is the x variable at
𝑔 𝑇𝑔
(k+1)th iteration, here I can write here 𝑥̅ 𝑘+1 = 𝑥̅ 𝑘 − (𝑔 𝑘𝑇𝑄𝑔𝑘 )𝑔𝑘 . So, this is for your
𝑘 𝑘

quadratic function as I have explained in the previous case, we can use in this fashion the
main disadvantage of steepest descent method is that it is very slow this method is very
slow.

It gives your value and then we can get this value is you can get the optimal value. So,
here the 𝛼𝑘 is keep on changing means every time we are calculating the ∇𝑓 and based
that α is calculated that is the step length. So, another method that is called the second
order and I can write here second method that is second method I am discussing it is your
Newton method it is called and it is called second order gradient method. The Newton
Raphson method proposes the property of quadratic convergence as we saw in the load
flow we know that in load flow that is Newton Raphson method is very fast because it is
giving quadratic convergence.

The optimality of function is obtained in finite number of steps the main condition of the
method is to have the initial guess very close to the optimal point. Otherwise, there will
be possibility of diverge solution already I explained that point that we should have the
initial guess very close to your optimal value. Otherwise, you will get the diverge
solution the main condition of this method to have initial guess very close as I said since
the second derivative of the function is required, therefore it must exist must exist that
𝜕𝑓
one. It means it is using the second order gradient second order. So, it means here is
𝜕𝑥
𝜕 2𝑓
existing we should also have upon this function should also exist if it is not existing
𝜕𝑥 2

then this method will not be applied now to see this method again I can just write the
function 𝑓(𝑥̅ ) here.

(Refer Slide Time: 27:35)

We can write at any point at any k th iteration I can write it here 𝑓(𝑥̅ )value here plus we
can differentiate here del x or I can write the grad.
𝜕𝑓 1
𝑓 (𝑥̅ 𝑘 ) = 𝜕𝑥 (𝑥̅ − 𝑥̅ 𝑘 ) + 2 ∇2 𝑓 (𝑥̅ 𝑘 )(𝑥̅ − 𝑥̅ 𝑘 )2 + ⋯Here, it is your this value is your x

minus x k here plus half of your either say double differentiation of this function here
and here it is x k I am talking and here your x minus x k here that is square plus other
high order term. So, here we are going for higher order term normally here in the
Newton Raphson method, we ignore the order that is more than two here it is a square.
So, we are going up to this portion in the first order we took up to here and remaining we
ignored.

Here, we are talking this term as well we saw this the first order necessary condition for
𝜕𝑓
optimality that I said here this = 0, you already know it. Now, based on this the
𝜕𝑥

equation here, and the equation this equation and this equation now we can get here this
value as I said 0. So, what we are going to get we can define this ∇𝑓(𝑥̅ 𝑘 ) +
∇2 𝑓 (𝑥̅ 𝑘 )(𝑥̅ − 𝑥̅ 𝑘 ) = 0, using this expression we can get this one very well what happen
now you are having this function ignore that one.

Now, you differentiate this function and put it 0 means from here you can differentiate
𝜕𝑓
this = 0 plus here no you have to differentiate this function and then here you can say
𝜕𝑥

∇𝑓 (𝑥̅ 𝑘 ) here.

Then, you are differentiating this function that is ∇𝑓(𝑥̅ − 𝑥̅ 𝑘 ) = 0. So, you are getting
this expression that is this expression because here to the double your differentiating,
now from here this is 0 or from this equation. We can write your𝑥̅ will be your 𝑥̅ 𝑘 minus,
this is your term this [∇2 𝑓(𝑥̅ 𝑘 )]−1 ∇𝑓(𝑥̅ 𝑘 ). From here, what we are getting we can
expand this. So, x we can get in terms of this value and we are getting this now you can
see with the previous this is the general expression which I wrote I say this is your α.
Then, we are using some grad, so in this case the α is your double differentiation of your
function with at the value of 𝑥̅ 𝑘 and the inverse value is existing.

So, here we are going for, so in this condition, we must have the value here that that is
why I said the double differentiation must exist if it is not existing this is say 0 inverse of
this will be infinite and you cannot solve it. So, this is Newton method and here alpha is
also keep on changing this value is kept on changing in every iteration. So, I can say
simply here for any iteration (x + 1) we can write this expression. This method is very
fast because it provides the quadratic convergence, but it has some disadvantages that it
requires additional memory to keep this is a matrix Hessian matrix inverse it is existing if
it is a vector, so it will be matrix.
The additional CPU time it has some more for every time you are inverting this and
inversion of Hessian matrix is every time must exist. This is also called Hessian matrix
another method is called your conjugate gradient method conjugate gradient method. So,
this method is superior to both previous method means your first order and second order.
In this method, the search is made along the certain sets of direction to ensure the
optimality of the function is achieved in certain number of iteration Fletcher’s power
method is most popular minimization of unconstraint non-linear functions. Another
method is gradient projection method, these are the various method and I am not going to
detail about that method.

Normally, they use some sort of the similar type of technique, now this is the case when
your optimization problem is not having constraints means you are having simple
objective function f here that is your x and having the different value. It means you are
having f that is your x1 to xn the function of n variable. You are going to minimize for
one variable it is very well clear that we can differentiate, and then we can go for double
differentiation, and then we can get it, but if you are going for the multiple variable.

This this function its value is of course, true, but we have to go for these methods and we
can calculate and it is iterative process. So, for the constraints optimization then we have
to use another one and normally this Lagrange’s multiplier method is very popular that
we normally use.

(Refer Slide Time: 33:37)

So, if your objective function here is 𝑓(𝑥̅ ) is there and you have some function here that
is the equality constraints 0 means here you are going to minimize this function that is
objective function subject to your constraints. We are talking about the constraints here
on that is the equality constraints, we can have the constraints on the limiting value of x
value here, but we are not having any inequality. The function inequality we are not
using functional means we can have here let us suppose your g x here and that should be
less than your 0. So, this is a function of x, so we are not including here then here is your
your objective function this is your equality constraint and then what we do normally we
go for Lagrange’s function method and we use another function called is called L.

We use this 𝑓(𝑥̅ ) here, and then we add some λ ℎ(𝑥̅ ), so this is also called augmented
function or sometimes called another Lagrange’s function and this gamma is called
Lagrange’s multiplier. This is known as the Lagrange’s function equation and also
sometimes called the augmented objective function because this is your objective
function is augmented by your equality constant.

This gamma is called your Lagrange’s multiplier Lagrange’s multiplier now here then
for objective for this your minimum value again, thus we are going to have this as per
𝜕𝐿 𝜕𝐿
this optimization. This = 0 and also this your this = 0, now here this is a x bar I
𝜕𝑥 𝜕𝜆

am writing, so this objective function if it is having x of n variable. So, we are going to

have here for every basically I for means every xi it will be suppose you are having your
function x1 to xn, then this will be your number of (n + 1), so you are going to have (n +
1) equation.

Then, you can have variables now how much x to xn here and plus λ then you can solve it
very easily. So, this is basically nothing, but I can say it is your x is 0, you can
differentiate this is your 0. We are getting this differentiation with this differentiation
equal to 0 means you are getting this 0 because this is independent of gamma. This is
partial derivative, so this is nothing but again you are equality constant equality
constraints. Now, we can solve to see one example just I will show you one example
here let us take a function that is your function objective function is here I can say f x2
variable I am talking.

Let us suppose your 𝑓 (𝑥̅ ) = 𝑥12 + 4𝑥22 and your equality here at here I can say ℎ(𝑥̅ ) =
5 − 𝑥1 − 𝑥2 = 0 and I want to solve this I want to here maximize minimize this function
here. So, the minimization of this function at I am trying to minimize subjected to this
constraint no doubt if this constraint is not there minimum will be certainly you can see
the x1 and x2 will be 0. Due to this, if you are putting x1 =0 and x2 =0, this is not five is
equal to not 0. So, it is not satisfying, so we have to get the minimum value subjected to
these equality constraints here. To solve this, we have to again use this procedure and we
have to use these condition condition number one condition number two that is for
objective function.

So, it is your augmented objective function must be differentiated with the state
variables. We will get here the number of equation is equal to number of state variables
plus another equation we are getting here. So, now we are having n number (n + 1)
equations and (n + 1) unknown, means n here plus λ another one and we can solve
uniquely you will get the another solution correct. So, just let us see how to minimize the
function if you are having these equality constraints. We have to apply the first order
necessary condition and that condition. As I said here this your differentiation of L with
𝜕𝐿 𝜕𝐿
respect to your 𝜕𝑥 , 𝜕𝑥 because we are having two state variable x1 and x2..
1 2

(Refer Slide Time: 38:15)

It means your function is a of two variable function, so it must be 0 here, it must be 0.

So, if you are differentiating this function with respect to x one you can see here you are
getting 2x for the 𝑓(𝑥̅ ) here and I can say here we are getting that will be equal to your
2x1. Now, the second term here if you are going for this will be 0, now another term we
are just differentiating with the λ means here what I am going to do means your L is
nothing, but your 𝑓(𝑥̅ ). I can say here 𝑓 (𝑥̅ ) + λℎ(𝑥̅ ) let us take here I am just going to
take plus let us take plus now because I am defined with this. Here, it is your lambda h x
means we can have this 𝑥12 + 4𝑥22 + λ(5 − 𝑥1 − 𝑥2 ) .

𝜕𝐿
So, we are differentiating this function with the x 1, so this was your𝜕𝑥 = 0, 2x1, this
1

was 0 now here with the λ as it is. So, I can say plus λ this is 0 and I can say x one is
differentiated. So, we are getting minus here and that is equal to 0 if you are
differentiating with this two, so this is 0, this term here it is 8. Now, I can write 8 x 2
minus here the lambda here the x 2 and I can say here the 0 another term we are getting
here that is I can say del l upon del lambda and this is nothing but this is whole 0. We are
getting 5 − 𝑥1 − 𝑥2 = 0.

Now, we are having three unknowns that is your x1 x2 and the λ and we are having three
equations, we can solve it uniquely. It means you can see here I can say x one from here
I can say x1 will be your λ/2 here your x2 = λ/8 and we can put here. So, we can set
5 − λ/2 − λ/8 = 0. If you will solve here, you will get λ = 8 value you can see λis
equal to 8, it is unity therefore, and then it is equal to 0. So, λ you are getting 8 from this
value you can say. Now, your 𝑥1 = 4 and your 𝑥2 = 1, so this is your optimal value and
then we can solve I, so this is using your Lagrange’s multiplier method.

In this method, what we are doing you can see very well that we are not considering the
inequality constraints means first the methods we saw with the help of without
constraints now we are solving if you are having equality constraints. Then, we can use
Lagrange’s multiplier methods here very easily another method that I am going to
discuss that is if you are having some inequality constraint even though to this generalize
means you are having several constraints what will happen in this function. So, it is your
L it is nothing, but your 𝑓(𝑥̅ ) + λ1 ℎ1 (𝑥̅ ) + +λ2 ℎ2 (𝑥̅ ) + ⋯ …. So, in general form I can
say 𝑓(𝑥̅ ) + λ𝑖 𝑇 ℎ𝑖 (𝑥̅ ) and then we can get this.

So, here we can have the several equality constraints, and then we can form in this
fashion, and then we can again apply the same procedure that we can go for the first
order necessary condition. Then, we can go for this one, and then we can solve and we
can get the optimal value here for this one now we can also. So, this method basically is
also called the classical method, now let us include some inequality constraints.
(Refer Slide Time: 42:54)

It means we are having equality as well as inequality means our objective function here
is your 𝑓 (𝑥̅ ), then it is we are going to minimize here, then we are having some equality
constant ℎ(𝑥̅ ) = 0. We have some inequality constant that is here equal to zero again the
function of several variables. So, these are the constraints here it is your equality this is
your inequality now for this the very there are so many methods. In the various methods,
one is called the Kuhn Tucker condition it is called Kuhn Tucker condition Kuhn Tucker
condition method. These methods that we can use your equality as well as inequality
constraints, it can handle very well.

So, this approach is similar to Lagrange’s method and here what we can do we use the
Lagrange’s augmented function. Then, we use this 𝑓 (𝑥̅ ) here plus here the summation of
here several, let us suppose we have several function of this and several functions of this.
So, I can say here L= 𝑓 (𝑥̅ ) + ∑𝑚 𝐾
𝑖=1 𝜆𝑖 ℎ𝑖 (𝑥̅ ) + ∑𝑘=1 𝜇𝑖 𝑔𝑖 (𝑥̅ )lambda I into hi x you can

add all the function plus here you are going for the 𝜇𝑖 and the 𝑔𝑖 (𝑥̅ ) means we are having
let us suppose this is your m equality and we are having the k inequality. So, here it is up
to k here up to m means we have to add all this equality with your Lagrange’s multiplier
lambda. Then, we use another variable that is mu related to your inequality constraint,
and this is your augmented objective function.

So, in this kuhn tucker condition is that first condition is that we have the several
𝜕𝐿
condition for kuhn tucker first one is that your 𝜕𝑥 = 0,.
1
It means for all the state variables this is the partial derivative of this function augmented
function objective function must be 0. Here, it is for all xi is equal to your x1 to xn second
𝜕𝐿
condition is that your again that is called = 0, and this shows that your ℎ𝑖 (𝑥̅ ) = 0,
𝜕𝜆1

this is nothing but your equality constraint previously your classical method. So, these
two condition are as it is with your Lagrange’s multiplier method or you can say classical
method, the third condition here that is added here it is called your 𝑔𝑖 (𝑥̅ ) ≤ 0. The fourth
condition is called that is your 𝜇𝑖 𝑔𝑖 (𝑥̅ ) = 0 and where 𝜇𝑖 ≥ 0.

So, these are the four condition based on that we go for, basically you can see this is your
first condition is primary condition second one is your equality constraints as it is this
condition is your equality constraints as it is we are adding one extra here condition. This
condition is called your complementary slackness condition, it is called complementary
complementary slackness condition in this condition what is happening it is said that this
here 𝜇𝑖 𝑔𝑖 (𝑥̅ ) = 0. It means here the possibility is that either 𝜇𝑖 = 0or your 𝑔𝑖 (𝑥̅ ) = 0 or
both are 0 means possibility that this is equal to 0 or this is equal to 0 or here we are
having both are equal to 0, then we can get this condition.

So, if mu is 0 means your 𝑔𝑖 (𝑥̅ ) ≠ 0, and then it will call 𝑔𝑖 is free to binding means if
𝜇𝑖 = 0, then this indicate that 𝑔𝑖 (𝑥̅ ) is free to be not binding is free to be non binding. It
means this constraint is non binding means this is 0, so what happens? The whole
function is 0, so this function sis even though there is no need to consider means it is a
non binding constraint. Now, if it is not, so if your 𝜇𝑖 ≥ 0 because we have the condition
too either it will be equal to 0 or more than 0 if this is more than 0, then what is
happened then 𝑔𝑖 in this condition 𝑔𝑖 (𝑥̅ ) = 0.

Then 𝑔𝑖 (𝑥̅ ) = 0 and then it is your binding constraint and it is just like it is binding and it
is sometimes called active constraints to see this, let us take same example with some
extra addition of that one.
(Refer Slide Time: 48:26)

It means again our objective function here is your f x here are two variable same function
I am using 𝑓 (𝑥̅ ) = 𝑥12 + 4𝑥22 your ℎ(𝑥̅ ) = 5 − 𝑥1 − 𝑥2 = 0. We are having your
𝑔(𝑥̅ ) = 4𝑥1 + 𝑥2 − 14 ≤ 0, so we are having this function is 𝑔(𝑥̅ ). It means your 𝑔(𝑥̅ )
is this value means this value is your 𝑔(𝑥̅ ) this 14, now we can again solve this 𝐿 =
𝑥12 + 4𝑥22 + λ(5 − 𝑥1 − 𝑥2 ) + μ(4𝑥1 + 𝑥2 − 14), here we are using this function.

So, this is your augmented Lagrange’s function or augmented objective function now the
𝜕𝐿 𝜕𝐿
condition first that it is your = 0, and = 0,, two variables will be considered. So,
𝜕𝑥1 𝜕𝑥2

𝜕𝐿 𝜕𝐿
now for this we can get it ()𝜕𝑥 = 2𝑥1 − 𝜆 + 4μ1, this is 0, here it is = 8𝑥2 − 𝜆 + μ,
1 𝜕𝑥2

here we are getting plus four mu for this other will be 0. For this, we are getting your
𝜕𝐿
= 8𝑥2 − 𝜆 + μ for this case and here we are getting plus your mu now second
𝜕𝑥2

condition is your h x that we are getting this 5 − 𝑥1 − 𝑥2 = 0.

So, this is your equation number one this is your equation number 2, this is your equation
number 3, your third condition that your 4𝑥1 + 𝑥2 − 14 ≤ 0 equation 4.

Another equation just we are going to have your µ that is multiplied by your 4𝑥1 + 𝑥2 −
14 = 0 where μ ≥ 0, that is your fifth equation. Now, what we can do now we can go
for the solving this first to see this first let us take μ = 0 if μ is 0, what happens? Now,
this will be not 0, this μ is 0, so we can again solve this value. It means here there is no
mu we cannot we do not know this. So, this is a equality which we can solve and this is
nothing but if mu is 0, it is our previous case when there was no inequality.

So, we again we can get your 𝑥1 = 4 and your 𝑥2 = 1 𝜆 = 8 if you put this value here
for if I am using this if you are putting here let us say what we are getting. So, we are
getting this 4 into 4 plus this 1 minus 14, how much this less than 0 16 plus 17, then it is
a positive means the 17 is less than 0, which know seventeen minus here 14 it is 3 is not
less than, so it is not correct. So, mu cannot be 0 because we saw it here now it means
μ ≥ 0, then in that case it is your 4𝑥1 + 𝑥2 − 14 = 0 because the 𝑔(𝑥̅ ) = 0. Now, what
now we have now we have equation number one equation number two equation number
three and now we have another equation here that is a I can say 6, so we have this four
equations.

Now, our unknowns are known because mu value you have to calculate here. So, we
have x 1 we have 2 x 2 we have your 𝜆 and we have μ, so four equation four unknown.
Then, we can solve it and we will find the different value, so in this case when we
consider that suppose this μ = 0we found that 3 is less than 0, which is not feasible
means this μ ≠ 0 It means our 𝑔𝑖 is binding and binding means it is the just equality
constraint because equality constraints are always binding than that one. So, just I put
μ ≠ 0 means we have to have this 𝑔(𝑥̅ ) = 0 means we have this condition here. It means
we have to now solve that is value using this, so now we have the four equations, now I
can solve now I can write here.
(Refer Slide Time: 53:50)

This four4𝑥1 + 𝑥2 − 14 = 0 that is equality constraint, we had your 2𝑥1 − 𝜆 + 4μ = 0

We have this 8𝑥2 − 𝜆 + μ = 0 and we have this 5 − 𝑥1 − 𝑥2 = 0. So, this is your
equality constraint this is your inequality constraint and these are your the first order
necessary condition from here if you will saw am not going to solve completely. So, we
will get here 𝑥1 = 3 and your 𝑥2 = 2 μ = 10/3 𝜆 = 56/3. So, these are your optimal
values, now what will be the optimal function value, then you can use in this value this
simply 𝑓 (𝑥̅ ) here basically these are called the star value means optimal value.

So, 𝑓 (𝑥̅ ∗ )is means we have to use this 𝑥12 + 4𝑥22 you can put here the value. So, it is a
three square plus four into 2 squares, so how much getting this 9 plus 16 we are getting
25. Now, you can see in the previous example when equality inequality was not there we
got your 𝑥1 = 4 and your 𝑥2 = 1, now you can see in these conditions sorry x 2 is equal
to 1. Here this 𝑓 (𝑥̅ )here optimal value was your 4 square plus 4 into your one square and
we are getting 16 plus 4, but we are getting twenty what does it mean you can say this
minimum value here, now it is moving from here because of binding constant.

You are not getting minimum inequality was not there means the inequality is binding
and this minimum value is shifted from here 20 to this 25 when there is no equality then
what was that function. That function this simply let us suppose this x one this 𝑓(𝑥̅ )
without here the 𝑥12 + 4𝑥22 this the optimal value here will be you𝑥1 = 0 and your
𝑥2 = 0 and here the 𝑓 (𝑥̅ ) = 0. This means you can say the minimum value here is 0
when we use some constraints here the value will be either zero or it will be more than
that no doubt about it means if the constraints are binding.

These equality constants are all binding, so it is a binding, so it moves from 0 to 20,
similarly when we put another inequality constraints it may be binding it may not be
binding, but it was the binding and we saw that is we are getting the 25. Similarly, we
can put the constraints on the state variable that is the x1 and x2 we can put the certain
values that should not be less than that and then we can using that optimizing program
and then we can solve it accordingly. So, this is your Kuhn tucker condition to get the
optimal value and this is very good method that we can get here no doubt the number of
equations if you are your number of variables is more.

Your number of constraints is more your number of equations is keeping on increasing

and it is not possible to solve this equation by simple elimination method. So, you have
to go for the some sort of technique that may be iterative techniques. You can use the
gauss siedel method or any method to arrive the values here from here. So, numbers of
equations are more increasing then you have to use some sort of techniques of to solve
these values and then we can get the optimal value of this objective function and this
state variable accordingly.

So, this is your Kuhn Tucker technique, so we saw the non-linear of course, we are
starting with the without constants then we saw the constraints with equality then we saw
equality as well as inequality. Then, we discuss the Kuhn Tucker method, another
method is called the slack variable formation and that I will discuss in the next lecture.

Thank you.