Chapter 4

Concepts of Water Resources System Analysis

The concept of systems analysis is indeed very wide, but in this section, we shall
confine our discussion in to a typical water resources system.

4.1 Definition and Types of Systems

The basic concept of a system is that it relates two or more things. Out of the several
definitions of a system, the simplest one states that it is a scheme that connects one
or more inputs and generates one or more outputs. A comprehensive definition of a
system is given by Dooge (1973) as “any structure, device, scheme, or procedure,
real or abstract, that interrelates in a given time reference, an input, cause, or
stimulus, of matter, energy, or information, and an output, effect or response, of
information, energy or matter.” The input and output referred to in mechanics is
synonymous with cause and effect in physics and philosophy, and stimulus and
response in biological sciences. Typical examples of a system are a university with its
various departments, a central government with its regional governments, a river
basin with all its tributaries, and so on.

4.1.1 Types of systems

Some commonly understood types of systems are discussed below.

1. Mindless and multi-minded Systems

Physical processes such as rainfall-runoff is called mindless system where as systems
such as social system are called multi-minded system.

2. Simple and Complex Systems

A simple system is one in which there is a direct relation between the input and the
output of the system. A complex system is a combination of several subsystems each
of which is a simple system. Therefore, a complex system may be subdivided into a
number of simple systems. Each subsystem has a distinct relation between input and
output. For example, a river basin system is a complex system comprising several
subsystems, each corresponding to a tributary.

39
3. Linear and Non Linear Systems
A 1inear system is one in which the output is a constant ratio of the input. In a
linear system the output due to a combination of inputs is equal to the sum of the
outputs from each of the inputs individually, i.e. the principle of superposition is valid.
For example, a system (watershed) in which the input x (rainfall) and the output y
(runoff) are related by y = mx, in which m is a constant, is a linear system. The unit
hydrograph in hydrology is a linear system (as the hydrograph ordinate of the direct
runoff hydrograph is proportional to the rainfall excess), On the other hand, a system
in which the input x and the output y are related by the linear equation y = mx + c, in
which m and c are constants, is not a linear system (why?). A nonlinear system is one
in which the input-output relation is such that the principle of superposition is not
valid. In reality, a watershed is a nonlinear system, as the runoff from the watershed
due a storm is a nonlinear function of the (storm) rainfall over its area.

4. Time Variant and Invariant Systems

In a time invariant system, the input-output relationship does not depend on the time
of application of the input, i.e. the output is the same for the same input at all times.
The unit hydrograph in hydrology is a linear time invariant system.

5. Continuous, Discrete and Quantized Systems

In a continuous system, the changes in the system take place continuously. E.g.
weather; whereas in a discrete system, the state of the system changes at discrete
intervals of time. A variable, input or an output, is said to be quantized when it
changes only at certain discrete intervals of time and holds a constant value between
intervals e.g. rainfall records.

6. Lumped Parameter and Distributed Parameter Systems

A lumped parameter is one whose variation in space is either nonexistent or
ignored (e.g. average rainfall over a watershed). A parameter is said to be a distributed
one if its variation in one or more spatial dimensions is taken into account. The
parameters of the system, the input or the output may be lumped. A lumped
parameter system is governed by ordinary differential equations (with time as an
independent variable), whereas a distributed parameter system is governed by partial
differential equations (with spatial coordinates as independent variables). For

40
example, a homogeneous isotropic aquifer is analyzed as a lumped parameter system.
Instead, if the spatial variation of the transmissivity in modelling a water table
aquifer is to be taken into account, the aquifer has to be modelled as a distributed
parameter system.

7. Deterministic and Probabilistic Systems

In a deterministic system, if the input remains the same, the output remains the
same, i.e. the same input will always produce the same output. The input itself
may be deterministic or stochastic. In a probabilistic system, the input-output
relationship is probabilistic rather than deterministic. The output corresponding to a
given input will have a probability associated with it.

8. Stable Systems
A stable system is one in which the output is bounded if the input is bounded.
Virtually all systems in hydrology and water resources are stable systems.

4.2 System Approach and System analysis

4.2.1 Systems approach

The input-output relationship of a system is controlled by the nature,
parameters of the system and the physical laws governing the system. In many of
the systems in practice, the nature and the principal laws are very complex, and
systems’ modelling in such cases uses simplifying assumptions and transformation
functions, which convert the input to the corresponding output, ignoring the
mechanics of the physical processes involved in the transformation. This requires
conceptualization of the system and its configuration to be able to construct a
mathematical model of it in which the input-output relationships are established
through operating the system in a defined fashion. The specification of the system
operation is what we refer to as the operating policy.

4.2.2 Systems analysis

Systems analysis may be said to be a formalization of the operation of the total system
with all of its subsystems together. Systems analysis is usually understood as a set

41
of mathematical planning and design techniques, which includes some formal
optimization procedure. When scarce resources must be used effectively, systems
analysis techniques stand particularly promising (for example, in optimal crop water
allocation to several competing crops, under conditions of limited water supply). It
must be clearly understood that systems analysis is not merely an exercise in
mathematical modelling but spans much farther into processes such as design and
decision. The techniques may use both descriptive as well as prescriptive models. The
descriptive models deal with the way the system works, whereas the prescriptive ones
are aimed at deciding how the system should be operated to best achieve the specified
objectives.

Basic Problems in Systems Analysis

Basically there are two types of problems: analysis and synthesis. The first one is
essentially a problem of prediction or a direct problem, whereas the second is a
problem of identification or an inverse problem. In a prediction problem, we are
required to determine the output knowing the input and the system operation. In an
inverse problem, we are required to find the system (parameters), given the input and
the output. In synthesis, we devise a model (system) that will convert a known input to
its corresponding known output. Here, we have to keep in mind the nature of the
system and its solution.

Example Problems
Prediction: In surface water hydrology, the problem is to predict the storm runoff
(output), knowing the rainfall excess (input) and the unit hydrograph (system). In
ground water hydrology, the problem to determine the response (output) of a given
aquifer (system), for given rainfall and irrigation application (input). In a reservoir
(system) the problem is to determine irrigation allocations (output) for given inflow and
storage (input), based on known or given operating policy.

Identification: In surface water hydrology, the problem is to derive the unit

hydrograph (system) given precipitation (input) and runoff data (output) for the
concurrent period. It is assumed that all the complexities of the watershed including
the geometry and physical processes in the runoff conversion are described by the unit

42
hydrograph. In ground water hydrology, the problem is to determine the aquifer
parameters (system), given the aquifer response (output) for known rainfall and
irrigation application (input).
In a reservoir system, the problem is to determine the reservoir release policy (system
operation) for a specified objective (output) for given inflows a (input).

Synthesis: The problem of synthesis is even more complex than the inverse problem
mentioned earlier. Here no record of input and output are available. An example is the
derivation of Snyder’s synthetic unit hydrograph using watershed characteristics to
convert known values of rainfall excess to runoff.

4.2.3 Techniques of Water Resources Systems Analysis

The basic techniques used in water resources systems analysis are optimization and
simulation. Whereas optimization techniques are meant to give global optimum
solutions, simulation is a trial and error approach leading to the identification of the
best solution possible. The simulation technique cannot guarantee global optimum
solution; however, solutions, which are very close to the optimum, can be arrived at
using simulation and sensitivity analysis.
Optimization models are embodied in the general theory of mathematical
programming. They are characterized by a mathematical statement of the objective
function, and a formal search procedure to determine the values of decision variables,
for optimizing the objective function. The principal optimization techniques are:

1. Linear Programming
The objective function and the constraints are all linear. It is probably the single-most
applied optimization technique all over the world. In integer programming, which is a
variant of linear programming, the decision variables take on integer values. In mixed
integer programming, only some of the variables are integers.

2. Nonlinear Programming
The objective function and/or (any of) the constraints involve nonlinear terms. General
solution procedures do not exist. Special purpose solutions, such as quadratic
programming, are available for limited applications. However, linear programming may

43
still be used in some engineering applications, if a nonlinear function can be either
transformed to a linear function, or approximated by piece-wise linear segments.

3. Dynamic Programming
Offers a solution procedure for linear or nonlinear problems, in which multistage
decision-making is involved. The choice of technique for a given problem depends on
the configuration of the system being analyzed, the nature of the objective function
and the constraints, the availability and reliability of data, and the depth of detail
needed in the investigation. Linear programming (LP) and dynamic programming (DP)
are the most common mathematical programming models used in water resources
systems analysis. Simulation, by itself, or in combination with LP, DP, or both LP and
DP is used to analyze complex water resources systems.

4.3 Systems Techniques in Water Resources

In most engineering problems, decisions need to be made to optimize (i.e. minimize or
maximize) an appropriate physical or economic measure, For example, we may want to
design a reservoir at a site with known inflows for meeting known water demands,
such that the reservoir capacity is minimum, or we may be interested in locating wells
in a region such that the aquifer draw-down is minimum for a given pumping pattern.
Most engineering decision making problems may be posed as optimization
problems. In general, an optimization problem consists of:
(a) An objective function, which is a mathematical function of decision variables,
that needs to be optimized, and
(b) A set of constraints that represents some physical (or other) conditions to be met.
The decision variables are the variables for which decisions are required such that the
objective function is optimized subject to the constraints. In the first example
mentioned, the reservoir capacity is one of the decision variables and the reservoir
mass balance defines one set of constraints. In the second example, the locations in
terms of coordinates of the wells are the decision variables, and again, the mass
balance forms a set of constraints.

A general optimization problem may be expressed mathematically as

Maximize f(X)

44
 Subjected to
gj(X)≤O , j=l,2....m
Where, X is a vector of decision variables, X = [x1, x2 ... xn.]

In this general problem there are n decision variables (viz. x1, x2, x3…xj,) and m
constraints. The complexity of the problem varies depending on the nature of the
function f(X), the constraint functions gj(X) and the number of variables and
constraints.

Simulation is a technique by which we imitate the behavior of a system.

Typically, we use simulation to answer ‘what if’ type of questions. As against
optimization, where we are typically looking for the ‘best possible’ solution, in
simulation we simply look at the behavior of the system for given sets of inputs.
Simulation is a very powerful technique in analyzing most complex water resource
systems in detail for performance evaluation, while optimization models yield
results helpful in planning and management of large systems. In some
situations, optimization models cannot be even applied due to computational
limitations. In many situations, however, decision-makers would be interested in
examining a number of scenarios rather than just looking at one single solution that is
optimal. Typical examples where simulation is used in water resources include:
(a) Analysis of river basin development alternatives,
(b) Multi-reservoir operation problems,
(c) Generating trade-offs of water allocations among various uses such as hydropower,
irrigation, industrial and municipal use, etc., and
(d) Conjunctive use of surface and ground water resources,

It may be noted that by repeatedly simulating the system with various sets of inputs it
is possible to obtain near-optimal solutions.

4.3.1 Linear programming

Liner programming may be classified as the most popular optimization technique used
ever in water resources systems planning. Its popularity is partly because of the
readily available software packages for problem solution, apart from its ability to
screen large-scale water resources systems in identifying potential smaller systems for

45
detailed modelling and analysis. Systems analysts find this tool extremely useful as a
screening model for very large systems, and as a planning model to determine the
design and operating parameters for a detailed operational study of a given system.

A. Graphical Method:
Linear programming (LP) is a scheme of solving an optimization problem in which both
the objective function and the constraints are linear functions of decision variables.
There are several ways of expressing a linear programming formulation, which lend
themselves to solutions, with appropriate modifications to the original problem.
We shall illustrate here maximization problem in LP in its classical form first, and
discuss variations later in the section.

Where, z is the objective function, x1, x2…xn are decision variables,

c1, c2…cn are coefficients of x1, x2…xn, respectively, in the objective function,
a11…a1n, a21…a2n… am1…amn are coefficients in the constraints,
b1, b2 … bn are non-negative right hand side values.
Each of the constraints can be converted to an equation by adding a slack variable
to the left hand side. The coefficient of this slack variable in the objective function will
be zero.

Standard Form (Equality Constraints)

There are many standard forms in which an LP problem is expressed, and we shall
follow the standard form with equality constraints, as given here, throughout the
section:

Maximize z  c1 x1  c2 x2  ...  cn xn  cn1 xn1  ...cn m xn m

46
Where the variables xn+1, xn+2 ... xn+m are called slack variables. The objective
function is written including the slack variables with coefficients cn+1, cn+2
…cn+m=0

In this standard form, we have a total of n+m variables (n decision variables+m

slack variables) and a constraint set of only m equations. These equations can
be solved uniquely for any set of n variables if the remaining n variables are set
to zero.

For example, in the simplex method (an iterative method, discussed later in the
section), the starting solution is chosen to be the one in which the decision
variables x1, x2… xn, are assumed zero, so that the slack variable in each
equality constraint equals the right hand side of the equation, i.e., xn+1 = b1,
xn+2 = b2, ..., Xn+m = bm, in the starting solution in the simplex method.
Obviously, the objective function value for this starting solution is z = 0.
Iterations are performed in the simplex method on this starting solution for
better values of the objective function till optimality is reached.

Before discussing the simplex method, the graphical method to a LP problem in

two variables is illustrated in the following example to gain some insight into
the method. It may be noted that if the problem has more than two decision
variables, the graphical method in two dimensions illustrated below cannot be
used.

Example 4.1
Two crops are grown on a land of 200 ha. The cost of raising crop 1 is 3 unit/ha, while
for crop 2 it is 1 unit/ha. The benefit from crop 1 is 5 unit/ha and from crop 2, it is 2

47
unit/ha. A total of 300 units of money is available for raising both crops. What should
be the cropping plan (how much area for crop 1 and how much for crop 2) in order to
maximize the total net benefits?

Solution:
The net benefit of raising crop 1 = 5 - 3 = 2 unit/ha
The net benefit of raising crop 2 = 2 - 1 = 1 unit/ha
Let x1 be the area of crop 1 in hectares and x2 be that of crop 2, and z, the’ total net
benefit.
Then the net benefit of raising both crops is 2x 1 + x2. However, there are two
constraints. One limits the total cost of raising the two crops to 300, and the other
limits the total area of the two crops to 200 ha. These two are the resource
constraints. Thus the complete formulation of the problem is:
Maximize z = 2x1+ x2 4.1
Subject to
3x1 + x2  300
x1 + x2  200
x 1, x 2  0 4.2
Equation (4.1) is the objective function and Eqs (4.2) are the constraints. The non-
negative constraints for x1 and x2 indicate that neither x1 nor x2 can physically be
negative (area cannot be negative).

First, the feasibility region for the constraint set should be mapped. To do this, plot
the lines 3x1+x2 = 300 and x1+x2 = 200, along with x1= 0 and x2 = 0 as in Fig. 4.1. The
region bounded by the non-negativity constraints is the first quadrant in which x1≥0
and x2 ≥0. The region bounded by the constraint 3x1+x2 ≤300 is the region OCD (it is
easily seen that since the origin x1 = 0, x2 = 0 satisfies this constraint, the region to the

Figure 4.1 Graphical method

48
left of the line CD in which the origin lies is the feasible region for this constraint).
Similarly, the region OAB is the feasible region for the constraint X1+X2 ≤ 200.
Thus, the feasible region for the problem taking all constraints into account is OAPD,
where P is the point of intersection of the lines AB and CD. Any point within or on the
boundary of the region, OAPD, is a feasible solution to the problem. The optimal
solution, however, is that point which gives the maximum value of the objective
function, z, within or on the boundary of the region OAPD.
Next, consider a line for objective function, z=2x 1+x2=c, for an arbitrary value c. The
line shown in the figure is drawn for c= 40 and the arrows show the direction in which
lines parallel to it will have higher value of c, i.e. if the objective function line is plotted
for two different values of c, the line with a higher value of c plots farther from the
origin than the one with a lower value (of c). We need to determine that value of c
corresponding to a line parallel to 2x1+x2=c, farthest point from the origin and at the
same time passing through a point lying within or on the boundary of the feasible
region. If the z line is moved parallel to itself away from the origin, the farthest point
on the feasible region that it touches is the point P (50, 150). This can be easily seen
by an examination of the slopes of the z line and the constraint lines.
Since the slope of the z line is -2 which lies between -3 (slope of the line
3x1+2x2=300) and -1 (slope of the line x1+x2=200), the farthest point, in the feasible
region away from the origin, lying on a line parallel to the z line. Thus, the point
P(x1=50, x2=150) presents the optimal solution to the problem. The maximized net
benefit z= Rs 250.
The graphical method can be used only with a two variable problem. For a general
LP problem, the most common method used is the simplex method.

Terminology:
Solution : A set of values assigned to the variables in a given problem is referred to as
a solution. A solution, in general, may or may not satisfy any or all of the constraints.

Basis and basic variables: The basis is the set of basic variables. The number of basic
variables is equal to the number of equality constraints. The variables in the basis
only can be non-negative. The non-basic variables are zeros. In the optimal solution of
the example mentioned earlier, out of the total of four variables x 1, x2, x3 and x4, the

49
variables x1 and x2 are in the basis, and x3 and x4 are out of the basis, i.e. x1 and x2 are
basic variables and x3 and x4 are non-basic variables in the optimal solution.

Nonbasic variables: variables which are outside (or not in) the basic are non-basic
variables. x3 and x4 in the optimal solution of the example are non-basic variables.

Feasible solution: Any solution (set of values associated with each variable) that
satisfies all the constraints is a feasible solution.

Infeasible solution: A solution which violates at least one of the constraints is

infeasible solution.

Basic solution: Assume there are a total number of n+m variables (n decision
variables and m slack variables) and total number of m equality constraints. Then a
basic solution is one which has m number of basic variables and n number of non-
basic variables. All non-basic variables and at least n zero valued variables.

Basic feasible solution: A basic solution which is also feasible is a basic feasible
solution.

Initial basic feasible solution: The basic feasible solution used as an initial solution
in the simplex method is called an initial basic feasible solution.

B. Simplex Method
Prelude to Simplex Method
Let us look upon the solution as one resulting from the following set of simultaneous
equations from Example 4.1:

3x1 + x2 + x3  300 4.3

x1 + x2 + x4  200 4.4
x1, x2, x3, x4  0

Where, x3 and x4 are the slack variables in the respective constraints, Eq. (4.3) and Eq.
(4.4), introduced to facilitate equality of the left and right hand sides.

50
 Equations 4.3 and 4.4 have four variables in two equations. These can be
uniquely solved when two of the variables x1, x2, x3 and x4 assume zero values. Our
aim is to look for such a combination of x1, x2, x3 and x4 satisfying Eqs. 4.3 and 4.4,
which make the objective function z=2x1+x2+ 0.x3+0.x4 a maximum. If we assume two
of these variables to be zero, then the remaining two can be solved from the two
equations. In general, if there are m number of equality constraints and n+m total
number of values (including slack variables), we can solve for any m of the n+m
variables if we assign zero value to each of the remaining n variables. In the starting
solution for the simplex method, we assign zero values to the n decision variables and
the remaining m variables are solved from m simultaneous equations. In the example,
we now need a search procedure to determine the optimal combinations of the four
variables x1, x2, x3 and x4 that maximizes the value of the objective function, z. This is
done by iterating the starting solution to move to that adjacent corner point solution,
which results in the best value of z, in the simplex method. In any corner point
solution, it may be noted that there can be at most m number of nonnegative variables
and at least n number of zero valued variables. This is the second important feature
implicit in the simplex method.
The example 4.1 will now be solved using the simplex method
Maximize z = 2x1+ x2
Subject to
3x1 + x2  300
x1 + x2  200
x 1, x 2  0
First introduce slack variables (non-negative) and convert the constraints into equality
constraints.
Maximize z=2x1+x2+ 0.x3+0.x4
Subject to
3x1+x2+x3  300
x1+x2+x4= 200
x1, x2, x3, x4  0
Here, the total number of variables, m=4, and the number of constraints, n=2.
Therefore two of these four variables have to be necessarily set to zero to enable us to
solve the two equations to determine the remaining two variables. For example,

51
consider the solution (x1, x2, x3 and x4)=(25,25,200,150). This solution is feasible
solution, but not a basic solution (verify). On the other hand, the solution (100, 25,
0,0) is a basic feasible solution (because it has at least two zero valued variables and
the solution satisfies both constraints).

The procedure is explained by means of the simplex tableau, Table 4.1. For
convenience, the objective function is also written as an equality constraint as follows:

Table 4.1 shows the basic variables under the column ‘Basis’. The last column ‘RHS’
in each row gives the value of the basic variable in the current solution. The elements
in each row connecting these two columns are the coefficients of the variables in the
constraint represented by that row. Two important features of this table are:

52
Table 4.1 Starting solution

53
54
55
Example 4.2 (Multiple solution)

Table 4.2 shows the iterations required to arrive at the final simplex table,
giving the optimal solution using the simplex method.
Table 4.2 Starting solution

56
4.4 Dynamic Programming
4.4.1 Introduction

A single stage decision problem may be represented as in Fig. 2.9. In this

representation, S is the input, X is the decision and T is the output. Because of
decision S and X. The transformation of input S to output T is called the state
transformation. In a serial multistage decision problem, the output T from one stage
forms the input S to the next stage, and a decision is associated with each stage.

Figure 4.2 Single stage decision problem.

4.4.2 Example application of dynamic programming

Shortest route problem
Consider the problem of determining the shortest route for a pipeline from among
various possible routes available from destination to sources. Figure 4.3 shows the
network of possible routes connecting several nodes, along with the distance between
two nodes, shown along the arrows, on a pipe route. Starting from the source node A,
the pipeline must run through one of the nodes B1, B2 or B3, one of the nodes, C1, C2,
or C3, and one of the nodes, D1, D2 or D3, before reaching the destination node E. The
problem is to determine the shortest route between node A and node E. We may pose
this problem as a multi stage, sequential decision problem and solve the problem with
dynamic progrogramming. To do this, we define stages as shown in the figure. Starting

57
from the source node A, We can reach one of the nodes B1, B2 or B3, in stage 1 and one
of the nodes C1, C2, or C3 in stage 2 and one of the nodes D1, D2 or D3 in stage 3, and
reach the destination node E in stage 4. Using forward recursion, we define the state
of the system, Sn, at a stage n (n=1,2,3,4) as the node reached in that stage , and the
decision variable as the node xn in the previous stage, n-1, from which the nodes Sn is
reached . we look for the node xn, out of all possible xn, which results in the shortes
distance from the source to the node Sn, We denote this node as xn*. In stage n , we
answer the question , “If we are in node Sn, then which node in the previous stage n-1
we must have come from so that the total distance up to the node Sn, starting from the
source node A is minimum?”

Figure 4.2 Network of routes in the shortest route problem

58
59
4.5 Optimization using calculus
Some basic concepts and rules of optimization of a function of a single variable and a
function of multiple variables are presented in this section.

A) Function of a Single Variable

Let f(x) be a function of a single variable x, defined in the range a<x<b (fig 4.3)

Fig. 4.3 function of a Single Variable

Local Maximum:
The function f(x) is said to have a local maximum at x1 and x4, where it has a value
higher than that at any other value of x in the neighbourhood of x 1 and x4. The
function is a local maximum at x1, if
f ( x1  x1 )  f ( x1 )  f ( x1  x1 )

60
Local Minimum
The function f(x) is said to have a local minimum at x 2 and x5, where it has a value
lower than that at any other value of x in the neighbourhood of x 2 and x5.
The function is a local minimum at x2, if
f ( x2  x2 )  f ( x2 )  f ( x2  x2 )

Saddle Point
The function has a saddle point at x, where the value of the function is lower on one
side of x3 and higher on the other, compared to the value at x3. The slope of the
function at x3 is zero.
f ( x3  x3 )  f ( x3 )  f ( x3  x3 ) ; slope of f(x) at x = x3 is zero.

Global Maximum
The function f(x) is a global maximum at x 4 in the range a <x < b, where the value of
the function is higher than that at any other value of x in the defined range.

Global Minimum
The function f(x) is a global minimum at x 2 in the range a <x < b, where the value of
the function is lower than that at any other value of x in the defined range.

Convexity
A function is said to be strictly convex, if a straight line connecting any two points
on the function lies completely above the function. Consider the function f(x) in
Fig. 4.4.

Fig. 4.4 Convex function

61
f(x) is said to be convex if the line AB is completely above the function (curve AB). Note
that the value of x for any point n between A and B can be expressed as ax1 + (1 - a)
x2, for some value of a, such that 0 ≤ a≤1. Therefore, f(x) is said to be strictly convex if,
f[ax1+(l - a)x2]< af(x1) + (l - a)f(x2); where 0 ≤ a≤ 1

1. If the inequality sign < is replaced by ≤ sign, then f(x) is said to be convex, but not
strictly convex.
2. If the inequality sign < is replaced by = sign, f(x) is a straight line and satisfies the
condition for convexity mentioned in 1 above. Therefore, a straight line is a convex
function.

3. If a function is strictly convex its slope increases continuously, or

d2 f/d2x> 0. For a convex function, however, d2 f/d2x ≥ 0.

Concavity
A function is said to be strictly concave if a straight line connecting any two points
on the function lies completely below the function. Consider the function(x), in
Fig. 4.5.

Fig. 4.5 Concave function

A function f(x) is strictly concave, if the line AB connecting any two points A and B on
the function is completely below the function (Curve AB).

f[ax1 + (1 - a)x2] > af(x1) + (1 - a)f(x,)

1. If the inequality > is replaced by ≥, then f(x) is said to be concave, but not strictly
concave.

62
2. If the inequality > is replaced by = sign, then f(x) is a straight line still satisfying the
condition for concavity. Therefore a straight line is a concave function.
3. If a function is strictly concave, its slope decreases continuously, or d 2f /d2x<0. For
a concave function, however d2f /d2x<0.

It may be noted that a straight line is both convex and concave, and is neither
strictly convex nor strictly concave.
A local minimum of a convex function is also its global minimum.
A local maximum of a concave function is also its global maximum.
The sum of (strictly) convex functions is (strictly) convex.
The sum of (strictly) concave functions is (strictly) concave.
If f(x) is a convex function, -f(x) is a concave function.
If f(x) is a concave function, -f(x) is a convex function.
In general, if f(x) is a convex function, and a is a constant, af(x) is convex, if a> 0 and
af(x) is concave if a<0.

B) Optimization of a Function of a Single Variable

The point at which a function will have a maximum or minimum is called a
stationary point. A stationary point is a value of the independent variable at which
the slope of the function is zero.
x=x0 is a stationary point if df/dx|xo = 0. This is a necessary condition for f(x) to be a
maximum or minimum at x0.
Sufficient condition is examined as follows:
1. If d2f/dx2>0 for all x, f(x) is convex and the stationary point is a global minimum.
2. If d2f/dx2<0 for all x, f(x) is concave and the stationary point is a global maximum.
3. If d2f/dx2=0, we should investigate further.
In case of 3, find the first nonzero higher order derivative. Let this be the derivative of
nth order. Thus, at the stationary point, x =xo
dnf/dxn<,=,>0

1. If n is even, x0 is a local minimum or a local maximum.

If dnf/dxn |xo> 0, xo is a local minimum
If dnf/dxn |xo< 0, xo is a local maximum

63
2. If n is odd, xo is a saddle point.

C) Function of Multiple Variables

Let f(X) is a function of n variables represented by the vector X = (x 1, x2, xn). Before
coming to the criteria for convexity and concavity of a function of multiple
variables, we should know the Hessian matrix (H-matrix) of the function. The
Hessian matrix, H [f(X)], of the function, f(X), is defined as:

The convexity and concavity of a function of multiple variables is determined by

an examination of the Eigen values of its Hessian matrix. The Eigen values of H
[f(X)] are given by the roots of the characteristic equation,
|λI – H [f(X)]| = 0

Where: I is an identity matrix, and λ is the vector of Eigen values. The function
f(X) is said to be positive definite if all its Eigen values are positive, i.e. all the values of
λ should be positive. Similarly, the function f(X) is said to be negative definite if all its
Eigen values are negative, i.e. all the values of λ should be negative.

Convexity and Concavity:

If all Eigen values of the Hessian matrix are positive, the function is strictly
convex. If all the Eigen values of the Hessian matrix are negative, the function
is strictly concave. If some Eigen values are positive and some negative, or if some
are zero, then the function is neither strictly convex nor strictly concave.

D) Optimization of a Function of Multiple Variables

Unconstrained Optimization
Let f(X) be a function of multiple variables, X = (x1, x2, x3,...,xn).
A necessary condition for a stationary point X = Xo is that each first partial derivative
of f(X) should equal zero.

64
Whether the function is a minimum or maximum at X = X1 depends on the nature of
the Eigen values of its Hessian matrix evaluated at X0.
1. If all Eigen values are positive at X0, X0 is a local minimum. If all Eigen values are
positive for all possible values of X, then X0 is a global minimum.
2. If all Eigen values are negative at X0, X0 is a local maximum. If all Eigen values are
negative for all possible values of X, then X0 is a global maximum.
3. If some Eigen values are positive and some negative or some are zero, then X 0 is
neither a local minimum nor a local maximum.

Example 4.2: Examine the following functions for convexity, concavity and then
determine their values at the extreme points.
1. f(X)= x12+x22-4x1-2x2+5
Solution:
First determine the Hessian Matrix.
df/dx1= 2x1 - 4 =0→x1=2
df/dx2 = 2x2 - 2 =0→x2=1
Check f(x) 2, 1= 22+12-8-2+5=0
d2f/dx12 = 2, d2f/dx1dx2= 0 , d2f/dx2dx1= 0, d2f/dx22= 2

Therefore, H f(x) =

Eigen values of H are obtained by:

The Eigen values are λ1= 2, λ2 = 2. As both the Eigen values are positive, the function
is a convex function (strictly convex). Also, as the Eigen values do not depend on the
value of x1 or x2, the function is strictly convex.
The stationary points are given by solving

65
Therefore the function f(X) has a global minimum at X = (2, 1).

Example 4.3: Examine the following functions for convexity, concavity and then
determine their values at the extreme points.

Solution:

The Eigen values are λ1= -2, λ2 = -2.

Both Eigen values of H Matrix are negative and are independent of the value of x 1 and
x2. Therefore f(X) is a strictly concave function.

That is X = (-2, 0) and is a global maximum. The function f(X) has a global maximum
at X = (-2, 0) equal to -4.

Example 4.4: Examine the following functions for convexity, concavity and then
determine their values at the extreme points.

66
Eigen values are given by the equation

Therefore λ1 =6x1 and λ2 = 6x2. That is, if both x1 and x2 are positive, then both Eigen
values are positive, and f(X) is convex; or if both x 1 and x2 are negative, then both
Eigen values are negative, and f(X) is concave.

Stationary points:

Therefore,
(i) f(X) has a local minimum at (x1, x2) = (1, 2), equal to13+ 23-3 (1) - 12(2) + 20 =
2. f min(x) = 2 at X (1, 2).
(ii) f(X) has a local maximum at (x1, x2) = (-1, -2) equal to (-l)3 + (-2)3-3 (-1) -12 (-2) + 20
= 38. fmax(X) = 38 at X = (-1, -2). At the points (1, -2) and (-1, 2), the function is neither
convex nor concave. They are saddle points.

Constrained Optimization
We shall discuss in this section the conditions under which a function of multiple
variables will have a local maximum or a local minimum, and those under which its
local optimum also happens to be its global optimum. Let us first consider a function
with equality constraints.

67
1. Function f(X) of n-Variables with a Single Equality Constraint
Maximize or Minimize f(X)
Subject to g(X) = 0
Note that f(X) and g(X) may or may not be linear.
We shall write down the Lagrangean of the function f(X) denoted by Lf(X, λ), and apply
the Lagrangean multiplier method.
Lf(X) =f(X) - λ g(X), where λ is a Lagrangean multiplier.
When g (K) = 0, optimizing Lf(X) is the same as optimizing f(X). The original problem of
constrained optimization is now transformed into an unconstrained optimization
problem (through the introduction of an additional variable, the Lagrangean
multiplier).

If more than one equality constraint is present in the problem,

Maximize or Minimize f(X)
Subject to gp(X) = 0, p = 1, 2…n, the Lagrangean function of f(X) in this case is
Lf(X, λ) =f(X) - λ1 g1(X) - λ2 g2(X) - ..., gm(X)
Where: λ = (λ1, λ2... λm)
A necessary condition for the function to have a maximum or minimum is that
the first partial derivatives of the function L should be equal to zero,

The (n + m) simultaneous equations are solved to get a solution, (X0, λ0). Let the
second partial derivatives be denoted by:

The sufficiency condition is specified below.

Consider the determinant D, denoted as |D|, given by:

68
This is a polynomial in p of order (n - m) where n is the number of variables and m is
the number of equality constraints. If each root of p in the equation |D| = 0 is
negative, the solution X0 is a local maximum. If each root is positive, then X0 is a local
minimum. If some roots are positive and some negative, X0 is neither a local maximum
nor a local minimum. Also, if all the roots are negative and independent of X, then X 0
is the global maximum. If all the roots are positive and independent of X, then X 0 is
the global minimum.

Example 4.5:

These equations yield x1 = x2= 2, λ=-4.

Now we shall determine if this is a maximum.

69
Or 2μ +4 = 0 giving μ = -2.
As the only root is negative, the stationary point x= (2, 2) is a local maximum of
f(X) and fmax(X) = -8.

2. Function f(X) with Inequality Constraints

An inequality constraint can be converted to an equality constraint by introducing an
additional variable on the left-hand side of the constraint. Thus a constraint g(X) ≤ 0 is
converted as g(X) + s2 = 0, where: s2 is a nonnegative variable (being square of s).
Similarly, a constraint g (X) ≥ 0 is converted as g(X) - s2 = 0.
The solution is found by the Lagrangean multiplier method, as indicated, treating s as
an additional variable in each inequality constraint.
When the Lagrangean of f(X) is formed with either type of constraint, equating the
partial derivative with respect to (w.r.t.) s gives,
λ.s = 0, meaning either λ = 0 or s = 0.

1. If λ > 0, s = 0. This means that the corresponding constraint is an equality

constraint (binding constraint or active constraint).
2. If s2 > 0, λ = 0. This means that the corresponding constraint is inactive or
redundant.

Kuhn-Tucker Conditions
The conditions mentioned above lead to the statement of Kuhn-Tucker conditions.
These conditions are necessary for a function f(X) to be a local maximum or a local
minimum. The conditions for a maximization problem are given below.
Maximize f(X)

70
Subject to gj(X) ≤ 0, j= 1...m.
The conditions are as follows:

In addition if f(X) is concave, and the constraints form a convex set, these conditions
are sufficient for a global maximum.

General Problem
The necessary and sufficient conditions for optimization of a function of multiple
variables subject to a set of constraints are discussed below.
A general problem may be one of maximization or minimization with equality
constraints, and inequality constraints of both ≥ and ≤ type.
Consider the problem:

Introduce variables s into the inequality constraints to make them equality constraints
or equations. Let S denote the vector with elements sj.
The Lagrangean is

Where, λ is the Lagrangean multiplier associated with constraint i.

Necessary Conditions for a Maximum or Minimum The first partial derivatives of

L(X, S, λ) with respect to each variable in X, S and λ should be equal to zero. The
solution for a stationary point (X0, S0, λ0) is obtained by solving these simultaneous
equations. This is a necessary condition. Sufficiency is checked by the following
conditions.

71
Sufficiency Conditions for a Maximum
f(X) should be a concave function.
gi (X) should be concave; λi>= 0, i= 1, 2…. j
gi(X) should be linear; λi<=0, i=j+1…k
gi(x) is linear, λi unrestricted, i = k + 1… m.

Similarly, Sufficiency Conditions for a Minimum

f(X) should be a convex function.
gi(X) should be a convex function; λi ≥ 0, i = 1… j.
gi(X) should be a concave; λi ≤ 0, i = j + 1… k.
gi(X) should be Linear; λi unrestricted, i = k + 1… m.

Note: For a maximum or a minimum, the feasible space or the solution space should
be a convex region. A constraint set g,(X) ≤ 0 defines a convex region, if gi(X) is a
convex function for all i. Similarly, a region defined by a constraint set gi(X) ≥ 0 is a
convex region, if gi(X) is a concave function for all i.
It is practically better to stick to one set of criteria, i.e. either for maximization or
minimization. We shall follow the criteria for maximization in the following examples
while testing the sufficiency criterion. For this purpose, we shall convert the given
problem to the following form:
Maximize f(X)
Subjected to gi(X) ≤ 0
We shall reiterate here that a linear function is both convex and concave.

Example 4.6:

72
λ1= 2; λ2 = 2, both being positive.
Thus f(X) is a convex function (strictly convex). Therefore the function
-f(X) is concave and can be maximized. First convert the problem to a form
Maximize f(X)
Subject to g(X) ≤0
The original problem is rewritten as:

(i) Assuming λ2=O, sl =O; x1=8/5, x2=6/5 and λ1= 4/5>0, s22=3/5>0
Here the conditions for a maximum are satisfied. No violations.
(ii) Assume λ1 = 0 and λ2 = 0.
Then the simultaneous equations give
x1= x2 = 2; s12= -2 (not possible)
s22 = -l (not possible).
This is not a solution to the problem. Similarly,
(iii) Assume. λ1 = 0 and s2 = 0. The equations to be solved are:

73
Note: In a clear case like this, when f(X) is strictly convex or -f(X) is strictly concave
and the solution set is convex (i.e. the constraint set is a convex region being bounded
by linear functions), there is a unique solution.
That is, only a particular combination of λ and s yields the optimum solution.
Thus, in a given trial in a problem such as Example 4.5, with two constraints:
If λ1 and λ2 are assumed to be zero, then s12 and s22 should both be positive,
If λ1 and s2 are assumed to be zero, then λ2 and s12 should both be positive,
If λ2 and s1 are assumed to be zero, then λ1 and s22 should both be positive,
If s1 and s2 are assumed to be zero, then λ1 and λ2 should both be positive.
The first trial, which satisfies these conditions, will be the optimal solution to the
problem, and the computations can stop there.

Problem 1

74
75