Numerical Methods &

Statistical Analysis

BSDS-304
 M.Sc. (IT) Final Year
MIT - 12

NUMERICAL METHODS
AND STATISTICAL ANALYSIS

MADHYA PRADESH BHOJ (OPEN) UNIVERSITY - BHOPAL

 Reviewer Committee
1. Dr. Sharad Gangele 3. Dr. K. Mani Kandan Nair
Professor Department of Computer Science
R.K.D.F. University, Bhopal (M.P.) Makhanlal Chaturvedi National
2. Dr. Romsha Sharma University of Journalism and
Professor Communication, Bhopal (M.P.)
Sri Sathya Sai College for Women,
Bhopal (M.P.)

Advisory Committee
1. Dr. Jayant Sonwalkar 4. Dr. Sharad Gangele
Hon’ble Vice Chancellor Professor
Madhya Pradesh Bhoj (Open) University, R.K.D.F. University Bhopal (M.P.)
Bhopal (M.P.)
2. Dr. L.S. Solanki 5. Dr. Romsha Sharma
Registrar Professor
Madhya Pradesh Bhoj (Open) University, Sri Sathya Sai College for Women,
Bhopal (M.P.) Bhopal (M.P.)
3. Dr. Kishor John 6. Dr. K. Mani Kandan Nair
Director Department of Computer Science
Madhya Pradesh Bhoj (Open) University, Makhanlal Chaturvedi National University of
Bhopal (M.P.) Journalism and Communication, Bhopal (M.P.)

COURSE WRITERS
Dr. N. Dutta, Professor (Mathematics), Head, Department of Basic Sciences & Humanities, Heritage Institute of
Technology, Kolkata
Units (1, 2.0-2.3, 2.6-2.10, 3)
Manisha Pant, Former Lecturer, Department of Mathematics, H.N.B. Garhwal University, Srinagar
(A Central University), U.K.
Units (2.4, 2.5.2, 5.5.3, 4.2, 4.3, 4.6.5, 5.2, 5.6)
C. R. Kothari, Ex-Associate Professor, Department of Economic Administration & Financial Management, University of
Rajasthan
Units (2.5-2.5.1, 4.7-4.7.5)
J. S. Chandan, Retd. Professor, Medgar Evers College, City University of New York
Units (4.0-4.1, 4.3.1-4.6.4, 4.8-4.12, 5.0-5.1, 5.3-5.5, 5.7-5.11)
Copyright © Reserved, Madhya Pradesh Bhoj (Open) University, Bhopal

All rights reserved. No part of this publication which is material protected by this copyright notice
may be reproduced or transmitted or utilized or stored in any form or by any means now known or
hereinafter invented, electronic, digital or mechanical, including photocopying, scanning, recording
or by any information storage or retrieval system, without prior written permission from the Registrar,
Madhya Pradesh Bhoj (Open) University, Bhopal

Information contained in this book has been published by VIKAS® Publishing House Pvt. Ltd. and has
been obtained by its Authors from sources believed to be reliable and are correct to the best of their
knowledge. However, the Madhya Pradesh Bhoj (Open) University, Bhopal, Publisher and its Authors
shall in no event be liable for any errors, omissions or damages arising out of use of this information
and specifically disclaim any implied warranties or merchantability or fitness for any particular use.
Published by Registrar, MP Bhoj (open) University, Bhopal in 2020

Vikas® is the registered trademark of Vikas® Publishing House Pvt. Ltd.

VIKAS® PUBLISHING HOUSE PVT. LTD.
E-28, Sector-8, Noida - 201301 (UP)
Phone: 0120-4078900  Fax: 0120-4078999
Regd. Office: A-27, 2nd Floor, Mohan Co-operative Industrial Estate, New Delhi 1100 44
 Website: www.vikaspublishing.com  Email: [email protected]
 SYLLABI-BOOK MAPPING TABLE
Numerical Methods and Statistical Analysis
Syllabi Mapping in Book

UNIT - I
Introduction, Limitation of Number Representation, Arithmetic rules for Unit-1: Representation of Numbers
Floating Point Numbers, Errors in Numbers, Measurement of Errors, (Pages 3-38)
Solving Equations, Introduction, Bisection Method, Regula Falsi Method,
Secant Method, Convergence of the iterative methods.

UNIT - II
Interpolation, Introduction, Lagrange Interpolation, Finite Differences, Unit-2: Interpolation and
Truncation Error in Interpolation, Curve Fitting, Introduction, Linear Curve Fitting
Regression, Polynomial Regression, Fitting Exponential and Trigonometric (Pages 39-113)
Functions

UNIT - III
Numerical Differentiation and Integration, Introduction, Numerical Unit-3: Numerical Differentiation
Differentiation Formulae, Numerical Integration Formulae, Simpson's Rule, and Integration
Errors in Integration Formulae, Gaussian Quadrature Formulae, Comparison (Pages 115-184)
of Integration Formulae, Solving Numerical Differential Equations,
Introduction, Euler's Method, Taylor Series Method, Runge-Kutta Method,
Higher Order Differential Equations.

UNIT - IV
Introduction to Statistical Computation, History of Statistics, Meaning and Unit-4: Statistical Computation and
scope of Statistics, Various measures of Average, Median, Mode, Geometric Probability Distributiona
Mean, Harmonic Mean, Measures of Dispersion, Range, Standard (Pages 185-290)
Deviation, Probability Distributions, Introduction, Counting Techniques,
Probability, Axiomatic or Modern Approach to Probability, Theorems on
Probability, Probability Distribution of a Random Variable, Mean and
Variance of a Random Variable, Standard Probability Distributions, Binomial
Distribution, Hyper geometric Distribution Geometrical Distribution,
Uniform Distribution (Discrete Random Variable), Poisson Distribution,
Exponential Distribution, Uniform Distribution (Continuous Variable),
Normal Distribution

UNIT - V
Estimation, Sampling Theory, Parameter and Statistic, Sampling Distribution Unit-5: Estimation and
of Sample Mean, Sampling Distribution of the Number of Successes, The Hypothesis Testing
Student's Distribution, Theory of Estimation, Point Estimation, Interval (Pages 291-328)
Estimation, Hypothesis Testing, Test of Hypothesis, Test of Hypothesis
Concerning Mean, Test of Hypothesis Concerning Proportion, Test of
Hypothesis Concerning Standard Deviation
 CONTENTS
INTRODUCTION

UNIT 1 REPRESENTATION OF NUMBERS 3-38

1.0 Introduction
1.1 Objectives
1.2 Introduction to Numerical Computing
1.3 Limitations of Number Representations
1.3.1 Arithmetic Rules for Floating Point Numbers
1.4 Errors in Numbers and Measurement of Errors
1.4.1 Generation and Propagation of Round-Off Error
1.4.2 Round-Off Errors in Arithmetic Operations
1.4.3 Errors in Evaluation of Functions
1.4.4 Characteristics of Numerical Computation
1.4.5 Computational Algorithms
1.5 Solving Equation
1.5.1 Bisection Method and Convergence of the Iterative Method
1.5.2 Newton-Raphson Method
1.5.3 Secant Method
1.5.4 Regula-Falsi Method
1.5.5 Descarte’s Rule
1.6 Answers to ‘Check Your Progress’
1.7 Summary
1.8 Key Terms
1.9 Self-Assessment Questions and Exercises
1.10 Further Reading

UNIT 2 INTERPOLATION AND CURVE FITTING 39-113

2.0 Introduction
2.1 Objectives
2.2 Interpolation
2.2.1 Iterative Linear Interpolation
2.2.2 Lagrange’s Interpolation
2.2.3 Finite Difference for Interpolation
2.2.4 Symbolic Operators
2.2.5 Shift Operator
2.2.6 Central Difference Operator
2.2.7 Differences of a Polynomial
2.2.8 Newton’s Forward Difference Interpolation Formula
2.2.9 Newton’s Backward Difference Interpolation Formula
2.2.10 Extrapolation
2.2.11 Inverse Interpolation
2.2.12 Truncation Error in Interpolation
2.3 Curve Fitting
2.3.1 Method of Least Squares
2.4 Trigonometric Functions
2.5 Regression
2.5.1 Linear Regression
2.5.2 Polynomial Regression
2.5.3 Fitting Exponential
2.6 Answers to ‘Check Your Progress’
 2.7 Summary
2.8 Key Terms
2.9 Self-Assessment Questions and Exercises
2.10 Further Reading

UNIT 3 NUMERICAL DIFFERENTIATION AND INTEGRATION 115-184

3.0 Introduction
3.1 Objectives
3.2 Numerical Differentiation Formula
3.2.1 Differentiation Using Newton’s Forward Difference Interpolation Formula
3.2.2 Differentiation Using Newton’s Backward Difference Interpolation Formula
3.3 Numerical Integration Formule
3.3.1 Simposon's One-Third Rule
3.3.2 Weddle’s Formula
3.3.3 Errors in Itegration Formulae
3.3.4 Gaussian Quadrature
3.4 Solving Numerical
3.4.1 Taylor Series Method
3.4.2 Euler’s Method
3.4.3 Runge-Kutta Methods
3.4.4 Higher Order Differential Equations
3.5 Answers to ‘Check Your Progress’
3.6 Summary
3.7 Key Terms
3.8 Self-Assessment Questions and Exercises
3.9 Further Reading

UNIT 4 STATISTICAL COMPUTATION AND PROBABILITY DISTRIBUTIONA 185-290

4.0 Introduction
4.1 Objectives
4.2 History and Meaning of Statistics
4.2.1 Scope of Statistics
4.3 Various Measures of statistical computations
4.3.1 Average
4.3.2 Mean
4.3.3 Median
4.3.4 Mode
4.3.5 Geometric Mean
4.3.6 Harmonic Mean
4.3.7 Quartiles, Percentiles and Deciles
4.3.8 Box Plot
4.4 Measures of Dispersion
4.4.1 Range
4.4.2 Quartile Deviation
4.4.3 Mean Deviation
4.5 Standard Deviation
4.5.1 Calculation of Standard Deviation by Short-cut Method
4.5.2 Combining Standard Deviations of Two Distributions
4.5.3 Comparison of Various Measures of Dispersion
4.6 Probability
4.6.1 Probability Distribution of a Random Variable
4.6.2 Axiomatic or Modern Approach to Probability
4.6.3 Theorems on Probability
 4.6.4 Counting Techniques
4.6.5 Mean and Variance of Random Variables
4.7 Standard Probability Distribution
4.7.1 Binomial Distribution
4.7.2 Poisson Distribution
4.7.3 Exponential Distribution
4.7.4 Normal Distribution
4.7.5 Uniform Distribution (Discrete Random and Continous Variable)
4.8 Answers to ‘Check Your Progress’
4.9 Summary
4.10 Key Terms
4.11 Self-Assessment Questions and Exercises
4.12 Further Reading

UNIT 5 ESTIMATION AND HYPOTHESIS TESTING 291-328

5.0 Introduction
5.1 Objectives
5.2 Sampling Theory
5.2.1 Parameter and Statistic
5.2.2 Sampling Distribution of Sample Mean
5.3 Sampling Distribution of the Number of Successes
5.4 The Student’s Distribution
5.5 Theory of Estimation
5.5.1 Point Estimation
5.5.2 Interval Estimation
5.6 Hypothesis Testing
5.6.1 Test of Hypothesis Concerning Mean and Proportion
5.6.2 Test of Hypothesis Conerning Standard Deviation
5.7 Answers to ‘Check Your Progress’
5.8 Summary
5.9 Key Terms
5.10 Self-Assessment Questions and Exercises
5.11 Further Reading
 Introduction
INTRODUCTION

Numerical method and Statistical analysis is the study of algorithms to find solutions
for problems of continuous mathematics or considered as a mathematical science NOTES
pertaining to the collection, analysis, interpretation or explanation, and presentation
of data and can be categorized as Inferential Statistics and Descriptive Statistics.
Numerical method helps in obtaining approximate solutions while maintaining
reasonable bounds on errors. Although numerical analysis has applications in all
fields of engineering and the physical sciences, yet in the 21st century life sciences
and both the arts have adopted elements of scientific computations. Ordinary
differential equations are used for calculating the movement of heavenly bodies,
i.e., planets, stars and galaxies. Besides, it evaluates optimization occurring in
portfolio management and also computes stochastic differential equations to solve
problems related to medicine and biology. Airlines use sophisticated optimization
algorithms to finalize ticket prices, airplane and crew assignments and fuel needs.
Insurance companies too use numerical programs for actuarial analysis. The basic
aim of numerical analysis is to design and analyse techniques to compute
approximate and accurate solutions to unique problems. In numerical analysis,
two methods are involved, namely direct and iterative methods. Direct methods
compute the solution to a problem in a finite number of steps whereas iterative
methods start from an initial guess to form successive approximations that converge
to the exact solution only in the limit. Iterative methods are more common than
direct methods in numerical analysis. The study of errors is an important part of
numerical analysis. There are different methods to detect and fix errors that occur
in the solution of any problem. Round-off errors occur because it is not possible to
represent all real numbers exactly on a machine with finite memory. Truncation
errors are assigned when an iterative method is terminated or a mathematical
procedure is approximated and the approximate solution differs from the exact
solution.
Statistical analysis is very important for taking decisions and is widely used
by academic institutions, natural and social sciences departments, governments
and business organizations. The word ‘Statistics’ is derived from the Latin word
‘Status’ which means a political state or government. It was originally applied in
connection with kings and monarchs collecting data on their citizenry that pertained
to state wealth, collection of taxes, study of population, and so on. In the beginning
of the Indian, Greek and Egyptian civilizations, data was collected for the purpose
of planning and organizing civilian and military projects. Proper records of such
vital events as births and deaths have been kept since the Middle Ages. By the end
of the 19th century, the field of statistics extended from simple data collection and
record keeping to interpretation of data and drawing useful conclusions from it.
Statistics can be called a science that deals with numbers or figures describing the
state of affairs of various situations with which we are generally and specifically
concerned. To a layman, it often means columns of figures, or perhaps tables,
graphs and charts relating to population, national income, expenditures, production,
consumption, supply, demand, sales, imports, exports, births, deaths and accidents. Self - Learning
Similarly, statistical records kept at universities may reflect the number of students, Material 1
Introduction the percentage of female and male students, the number of divisions and courses
in each division, the number of professors, the tuition received, the expenditures
incurred, and so on. Hence, the subject of statistics deals primarily with numerical
data gathered from surveys or collected using various statistical methods.
NOTES
This book is divided into five units. The topics discussed is designed to be
a comprehensive and easily accessible book covering the limitations of number
representation, measurement of errors, solving equations, Regula Falsi method,
secant method, interpolation, Lagrange interpolation, curve fitting, regression,
numerical differentiation, Simpson’s rule, Gaussian quadrature formulae, solving
numerical differential equations, Euler’s method, Taylor series method, Runge-
Kutta method, history of statistics, various measures of statistical computation,
probability distribution, standard probability distribution, sampling theory, point
estimation and test of hypothesis.
The book follows the Self-Instructional Mode (SIM) wherein each unit
begins with an ‘Introduction’ to the topic. The ‘Objectives’ are then outlined before
going on to the presentation of the detailed content in a simple and structured
format. ‘Check Your Progress’ questions are provided at regular intervals to test
the student’s understanding of the subject. ‘Answers to Check Your Progress
Questions’, a ‘Summary’, a list of ‘Key Terms’, and a set of ‘Self-Assessment
Questions and Exercises’ are provided at the end of each unit for effective
recapitulation.

Self - Learning
2 Material
 Representation of Numbers

UNIT 1 REPRESENTATION
OF NUMBERS
NOTES
Structure
1.0 Introduction
1.1 Objectives
1.2 Introduction to Numerical Computing
1.3 Limitations of Number Representations
1.3.1 Arithmetic Rules for Floating Point Numbers
1.4 Errors in Numbers and Measurement of Errors
1.4.1 Generation and Propagation of Round-Off Error
1.4.2 Round-Off Errors in Arithmetic Operations
1.4.3 Errors in Evaluation of Functions
1.4.4 Characteristics of Numerical Computation
1.4.5 Computational Algorithms
1.5 Solving Equation
1.5.1 Bisection Method and Convergence of the Iterative Method
1.5.2 Newton-Raphson Method
1.5.3 Secant Method
1.5.4 Regula-Falsi Method
1.5.5 Descarte’s Rule
1.6 Answers to ‘Check Your Progress’
1.7 Summary
1.8 Key Terms
1.9 Self-Assessment Questions and Exercises
1.10 Further Reading

1.0 INTRODUCTION
The use of computers to solve problems involving real numbers is referred to as
‘Numerical Calculations.’A finite string of digits can represent a large number of
real numbers. Most scientific computers limit the amount of digits that can be used
to represent a single number to a set number. Numerical error is the combined
effect of two kinds of error in a calculation. The first is caused by the finite precision
of computations involving floating point or integer values. The second usually called
truncation error is the difference between the exact mathematical solution and the
approximate solution obtained when simplifications are made to the mathematical
equations to make them more amenable to calculation. The number of significant
figures in a measurement, such as 2.531, is equal to the number of digits that are
known with some degree of confidence (2, 5 and 3) plus the last digit (1), which is
an estimate or approximation. Zeros within a number are always significant. Zeros
that do nothing but set the decimal point are not significant. Trailing zeros that are
not needed to hold the decimal point are significant. A round-off error, also called
rounding error, is the difference between the calculated approximation of a number
and its exact mathematical value. Numerical analysis specifically tries to estimate
this error when using approximation equations and/or algorithms, especially when
using finitely many digits to represent real numbers. Self - Learning
Material 3
Representation of Numbers In root finding and curve fitting, a root-finding algorithm is a numerical
method, or algorithm, for finding a value x such that f(x) = 0, for a given function f.
Such an x is called a root of the function f. Generally speaking, algorithms for solving
problems numerically can be divided into two main groups: direct methods and
NOTES iterative methods. Direct methods are those which can be completed in a
predetermined unite number of steps. Iterative methods are methods which
converge to the solution over time. These algorithms run until some convergence
criterion is met. When choosing which method to use one important consideration
is how quickly the algorithm converges to the solution or the method’s convergence
rate.
In this unit, you will learn about the limitations of number representation,
arithmetic rules for floating point numbers, errors in numbers and measurement of
errors, solving equation, bisection method and convergence of the iterative method,
secant method and Regula-Falsi method.

1.1 OBJECTIVES
After going through this unit, you will be able to:
 Understand the basic concept of limitations of number representation
 Explain about the arithmetic rules for floating point numbers
 Analysis the errors in numbers and measurement of errors
 Define solving equation
 Discuss about the bisection method and convergence of the iterative method
 Elaborate on the secant method
 Explain Regula-Falsi method

1.2 INTRODUCTION TO NUMERICAL

COMPUTING
Numerical methods are useful in almost all fields of Science and Engineering,
especially when analytical solutions are either not available or are very complicated.
There are several problems for which numerical solutions are simpler than analytical
solutions. The development of computers and the advancement in software
engineering have spurred further research in numerical analysis. Well-defined
algorithms lead to faster computation, improved storage capacity better accuracy
and stability.
The methods employed in numerical analysis are at times approximate and the
data used in computation are of finite decimal representation. Thus in most of the
cases, the results obtained by numerical methods have some errors. Before defining
numerical computing we must be aware of sources of errors in a numerical solution
and accordingly handle the case. Numerical analysis basically deals with the
development of suitable methods for obtaining applicable numerical solutions for
mathematical problems along with an indication of the accuracy of the solution.
Self - Learning
4 Material
 Numerical methods have been specifically developed for finding accurate Representation of Numbers

numerical solutions using a computer. While performing arithmetic operations on

real numbers, using a computer, we use fixed number of decimal digits. Most
numbers usually have infinite decimal representation, but for machine computation
the numbers are given and stored with a finite number of digits. NOTES

1.3 LIMITATIONS OF NUMBER

REPRESENTATIONS
Numerical methods are methods used for solving problems through numerical
calculations providing a table of numbers and/or graphical representations or figures.
Numerical methods emphasize that how the algorithms are implemented. Thus,
the objective of numerical methods is to provide systematic methods for solving
problems in a numerical form. Often the numerical data and the methods used are
approximate ones. Hence, the error in a computed result may be caused by the
errors in the data or the errors in the method or both. Generally, the numbers are
represented in decimal (base 10) form, while in computers the numbers are
represented using the binary (base 2) and also the hexadecimal (base 16) forms.
To perform a numerical calculation, approximate them first by a representation
involving a finite number of significant digits. If the numbers to be represented
are very large or very small, then they are written in floating point notation. The
Institute of Electrical and Electronics Engineers (IEEE) has published a standard
for binary floating point arithmetic. This standard, known as the IEEE Standard
754, has been widely adopted. The standard specifies formats for single precision
and double precision numbers. The simplest way of reducing the number of
significant digits in the representation of a number is simply to ignore the unwanted
digits known as chopping. All these topics are discussed in the following section.
Significant Figures
In approximate representation of numbers, the number is represented with a finite
number of digits. All the digits in the usual decimal representation may not be
significant while considering the accuracy of the number. Consider the following
numbers:
1514, 15.14, 1.324, 1524
Each of them has four significant digits and all the digits in them are significant.
Now consider the following numbers,
0.00215, 0.0215, 0.000215, 0.0000125
The leading zeroes after the decimal point in each of the above numbers are
not significant. Each number has only three significant digits, even though they
have different number of digits after the decimal point.
1.3.1 Arithmetic Rules for Floating Point Numbers
Every real number is usually represented by a finite or infinite sequence of decimal
digits. This is called decimal system representation. For example, we can represent
Self - Learning
Material 5
Representation of Numbers
1 1 1
as 0.25, but as 0.333... Thus is represented by two significant digits only,,
4 3 4
1
while is represented by an infinite number of digits. Most computers have two
NOTES 3
forms of storing numbers for performing computations. They are fixed-point and
floating point. In a fixed-point system, all numbers are given with a fixed number of
decimal places. For example, 35.123, 0.014, 2.001. However, fixed-point
representation is not of practical importance in scientific computation, since it cannot
deal with very large or very small numbers.
In a floating-point representation, a number is represented with a finite
number of significant digits having a floating decimal point. We can express the
floating decimal number as follows:
623.8 as 0.6238 × 103, 0.0001714 as 0.1714 × 10–3
A very large number can also be representated with floating-point
representation, keeping the first few significant digits such as 0.14263218 × 1039.
Similarly, a very small number can be written with only the significant digits, leaving
the leading zeros such as 0.32192516 × 10–19.
In the decimal system, very large and very small numbers are expressed in
scientific notation as follows: 4  69  1023 and 1  601 10 19 . Binary numbers can
also be expressed by the floating point representation. The floating point
representation of a number consists of two parts: the first part represents a signed,
fixed point number called the mantissa (m); the second part designates the position
of the decimal (or binary) point and is called the exponent (e). The fixed point
mantissa may be a fraction or an integer. The number of bits required to express
the exponent and mantissa is determined by the accuracy desired from the
computing system as well as its capability to handle such numbers. For example,
the decimal number + 6132.789 is represented in floating point as follows:
sign sign
0 0  6132789 0 04
    
mantissa exponent

The mantissa has a 0 in the leftmost position to denote a plus. Here, the
mantissa is considered to be a fixed point fraction. This representation is equivalent
to the number expressed as a fraction 10 times by an exponent, that is 0.6132789
× 10+04. Because of this analogy, the mantissa is sometimes called the fraction
part.
Consider, for example, a computer that assumes integer representation for the
mantissa and radix 8 for the numbers. The octal number + 36.754 = 36754 × 8–3 in its
floating point representation will look like this:
sign sign
0 36754 1 03
 

mantissa exponent

Self - Learning
6 Material
 When this number is represented in a register in its binary-coded form, the Representation of Numbers

actual value of the register becomes 0 011 110 111 101 100 and 1 000 011.
Most computers and all electronic calculators have a built-in capacity to
perform floating-point arithmetic operations.
NOTES
Example 1.1: Determine the number of bits required to represent in floating point
notation the exponent for decimal numbers in the range of 10 86 .
Solution: Let n be the required number of bits to represent the number 10 86.

2n  1086
n log 2  86
86 86
n   285.7
log 2 0.3010
Therefore, 1086  2285.7

The exponent ±285 can be represented by a 10-bit binary word. It has a

range of exponents (+511 to –512).

1.4 ERRORS IN NUMBERS AND

MEASUREMENT OF ERRORS
The errors in a numerical solution are basically of two types. They are truncation
error and computational error. The error which is inherent in the numerical method
employed for finding numerical solution is called the truncation error. The
computational error arises while doing arithmetic computation due to representation
of numbers with a finite number of decimal digits.
The truncation error arises due to the replacement of an infinite process
such as summation or integration by a finite one. For example, in computation of a
transcendental function we use Taylor series/Maclaurin series expansion but retain
only a finite number of terms. Similarly, a definite integral is numerically evaluated
using a finite sum with a few function values of the integral. Thus, we express the
error in the solution obtained by numerical method.
Inherent errors are errors in the data which are obtained by physical
measurement and are due to limitations of the measuring instrument. The analysis
of errors in the computed result due to the inherent errors in data is similar to that
of round-off errors.
1.4.1 Generation and Propagation of Round-Off Error
During numerical computation on a computer, a round-off error is generated by
taking an infinite decimal representation of a real, rational number such as 1/3,
4/7, etc., by a finite size decimal form. In each arithmetic operation with such
approximate rounded-off numbers there arises a round-off error. Also round-off
errors present in the data will propagate in the result. Consider two approximate
floating point numbers rounded-off to four significant digits.
x = 0.2234 × 103 and y = 0.1112 × 102 Self - Learning
Material 7
Representation of Numbers The sum x + y = 0.23452 × 103 is rounded-off to 0.23456 × 103 with an
absolute error, 2 × 10–2. This is the new round-off error generated in the result.
Besides this error, the result will have an error propagated from the round-off
errors in x and y.
NOTES
1.4.2 Round-Off Errors in Arithmetic Operations
To get an insight into the propagation of round-off errors, let us consider them for
the four basic operations of addition, subtraction, multiplication and division. Let
xT and yT be two real numbers whose round-off errors in their approximate
representations x and y are 1 and 2 respectively, so that
xT  x  1 and yT  y   2
Their addition gives, ( xT  yT )  ( x  y )  1   2
Hence, the propagated round-off error is given by,
( xT  yT )  ( x  y )  1   2

Thus the propagated round-off error is the sum of two approximate numbers
(having round-off errors) equal to the sum of the round-off errors in the individual
numbers.
The multiplication of two approximate numbers has the propagated round-
off error given by,
xT  yT  xy  1 y   2 x  1 2
Since the product 12 is a small quantity of higher order, then 1or 2 may
take the propagated round-off error as 1 x1   2 y1 and the relative propagated
error is given by,
1 x   2 y 1  2
 
xy x y
This is equal to the sum of the relative errors in the numbers x and y.
Similarly, for division we get the relative propagated error as,
xT x

yT y 1  2
 
x x y
y

Thus, the relative error in division is equal to the difference of the relative
errors in the numbers.
1.4.3 Errors in Evaluation of Functions
The propagated error in the evaluation of a function f (x) of a single variable x
having a round-off error  is given by,
f ( x   )  f ( x)   f '( x)
In the evaluation of a function of several variables x1, x2, …, xn, the
n
f
propagated round-off error is given by  1 , where 1 ,  2 ,...,  n are the round-
i 1 xi
Self - Learning
8 Material off errors in x1, x2,..., xn, respectively.
Significance Errors Representation of Numbers

During arithmetic computations of approximate numbers having fixed precision,

there may be loss of significant digits in some cases. The error due to loss of
significant digits is termed as significance error. Significance error is more serious
than round-off errors, since it affects the accuracy of the result. NOTES
There are two situations when loss of significant digits occur. These are,
(i) Subtraction of two nearly equal numbers.
(ii) Division by a very small divisor compared to the dividend.
For example, consider the subtraction of the nearly equal numbers
x = 0.12454657 and y = 0.12452413, each having eight significant digits. The
result x – y = 0.22440000 × 10–4, is correct to four significant figures only. This
result when used in further computations leads to serious error in the result.
Consider the problem of computing the roots of the quadratic equation,
ax2 + bx + c = 0
The roots of this equation are,
b  b2  4ac b  b 2  4ac
and
2a 2a

If b2 >> 4ac, then the evaluation of  b  b 2  4ac leads to subtraction of

nearly equal numbers. One can avoid this by rewriting the expression,
b  b 2  4ac
2a
It can be written as,

 b  b 2  4ac )(b  b 2  4ac  2c


2a  b  b  4ac 2
 b  b 2  4ac

Let the quadratic equation be,

x2 + 100.0001x + 0.01 = 0
Using the first formula, we get the smaller root = 0.10050000 × 10–3, whereas
exact root is 0.10000 0000 × 10–3. But using the last expression we get the smaller
root as 0.10000000 × 10–3 which does not has the effect of significance error.
Consider an example where loss of significant digits occur due to division
by a small number.
1  cos x
Computation of f ( x)  , for small values of x would have loss of
x2
significant digits.
The Table 1.1 shows the computed values of f (x) upto six decimal places
along with the correct values and error.

Self - Learning
Material 9
Representation of Numbers Table 1.1 Computed Value of f(x) upto Six Decimal Places

x Computed f (x) Correct f (x) Error

0.1 0.499584 0.499583 – 0.000001
NOTES 0.01 0.50008 0.499996 – 0.000012
0.001 0.506639 0.500000 – 0.006639
0.0001 0.500000 0.745058 0.245058

Table 1.1 shows that the error in the computed value becomes more serious
for smaller value of x. It may be noted that the correct values of f (x) can be
computed by avoiding the divisions by small number by rewriting f (x) as given
below.
1  cos x 1  cos x
f ( x)  
x2 1  cos x

sin 2 x
i.e., f ( x) 
x 2 (1  cos x)

1.4.4 Characteristics of Numerical Computation

A numerical solution can never be exact but attempts are made to know the accuracy
of the approximate solution. Thus one attempts to get an approximate solution
which differs from the exact solution by less than a specified tolerance limit.
Some numerical methods find the solution by a direct method but many others
are of repetitive nature. The first step in the solution procedure is to take an
approximate solution. Then the numerical method is applied repeatedly to get
better results till the solution is obtained up to a desired accuracy. This process is
known as iteration.
To get a numerical solution on a computer, one has to write an algorithm. An
algorithm is a sequence of unambiguous steps used to solve a given problem. In
the design of such computer programs one considers the input data required to
implement the numerical method and writes the computer program in a suitable
programming language. The output of the program should give the solution with
the desired accuracy.
It may be noted that the iterative method gives rise to a sequence of results.
The convergence of this sequence to get the output upto a desired accuracy is
dependent on the initial data. Hence, one has to suitably choose the input data.
Thus, if for some input data the sequence is not convergent for certain pre-assigned
number of iterations then the input data is changed. It is for this reason that one has
to limit the number of iterations to be employed while designing the computer
program.
While computing a solution with the help of an algorithm, one has to check
the correctness of the solution obtained. To do so, one has to have some test data
whose solution is known.
Example 1.2: The numbers 28.483 and 27.984 are both approximate and are
correct up to the last digits shown. Compute their difference. Indicate how many
Self - Learning
significant digits are present in the result and comment.
10 Material
 Solution: We have 28.483 – 27.984 = 00.499. The result has only three significant Representation of Numbers

digits. This is due to the loss of significant digits during subtraction of nearly equal
numbers.
Example 1.3: Round the number x = 2.2554 to three significant figures. Find the
absolute error and the relative error. NOTES
Solution:The rounded-off number is 2.25.
The absolute error is 0.0054.

The relative error is ~ 0.0054  0.0024

2.25
The percentage error is 0.24 per cent.
22
Example 1.4: If  = 3.14 instead of , find the relative error..
7

 22 22
Solution:Relative error =  7  3.14   0.00090.
  7

Example 1.5: Determine the number of correct digits in x = 0.2217, if it has a

relative error  r  0.2  101.
Solution:Absolute error = 0.2 × 10–1 × 0.2217 = 0.004434
Hence, x has only one correct digit x ~ 0.2.
Example 1.6: Round-off the number 4.5126 to four significant figures and find
the relative percentage error.
Solution:The number 4.5126 rounded-off to four significant figures is 4.153.

Relative error   0.0004  100   0.0088 per cent

4.5126

5 xy 2
Example 1.7: Given f (x, y, z) = , find the relative maximum error in the
z2
evaluation of f (x, y, z) at x = y = z = 1, if x, y, z have absolute errors
x  y  z  0.1
Solution:The value of f (x, y, z) at x = y = z = 1 is 5. The maximum absolute error
in the evaluation of f (x, y, z) is,
f f f
(f ) max  x  y  z
x y z
5y2 10 xy  10 xy 2
 2
x  2
y  z
z z z3
At, x = y = z = 1, the maximum relative error is,

25  0.1
( ER ) max   0.5
5
Example 1.8: Find the relative propagated error in the evaluation of x + y where
x = 13.24 and y = 14.32 have round-off errors 1  0.004 and  2  0.002 respectively.
Self - Learning
0.004 and 0.002 respectively. Material 11
Representation of Numbers
Solution:Here, x  y  27.56 and 1   2  0.006 .
0.006
Thus, the required relative error =  0.0002177 .
27.56
NOTES Example 1.9: Find the relative percentage error in the evaluation of u = xy with
x = 5.43, y = 3.82 having round-off errors 0.01 in both x and y.
Solution:Now, xy = 5.43 × 3.82 ~ 20.74
0.01
The relative error in x is  0.0018.
5.43
0.01
The relative error in y is  0.0026.
3.82
Thus, the relative propagated error in x and y = 0.0044.
The percentage relative error = 0.44 per cent.
Example 1.10: Given u = xy + yz + zx, find the estimate of relative percentage
error in the evaluation of u for x = 2.104, y = 1.935 and z = 0.845. What are the
approximate values correct to the last digit?
Solution:Here, u = x (y + z) + yz = 2.104 (1.935 + 0.845) + 1.935 × 0.845
= 5.849 + 1.635 = 7.484
Error, u  ( y  z )x  ( z  x)y  ( x  y )z
 0.0005  2( x  y  z )  x  y  z  0.0005
 2  4.884  0.0005  0.0049

0.0049
Hence, the relative percentage error =  100  0.062 per cent.
7.884
Example 1.11: The diameter of a circle measured to within 1 mm is d = 0.842 m.
Compute the area of the circle and give the estimated relative error in the computed
result.
d 2
Solution: The area of the circle A is given by the formula, A  .
4

Thus, A  3.1416  (0.842) 2 m2 = 0.5568 m2.

4
Here the value of  is taken upto 4th decimal palce since the data of d has
accuracy upto the 3rd decimal place. Now the relative percentage error in the
above computation is,
2 d 4  d 2 d 2  0.01
Ep    100   100   0.24 per cent
4 d 2
d 0.842
Example 1.12: The length a and the width b of a plate is measured accurate up
to 1cm as a = 5.43 m and b = 3.82 m. Compute the area of the plate and indicate
its error.
Solution: The area of the plate is given by,
Self - Learning
12 Material A = ab = 3.82 × 5.43 sq. m. = 20.74 m2.
The estimate of error in the computed value of A is given by, Representation of Numbers

A  a .b  b . a
 0.01  3.82  0.01  5.43, since a  b  0.01
NOTES
 0.0925  10 m 2

1.4.5 Computational Algorithms

For solving problems with the help of a computer, one should first analyse the
mathematical formulation of the problem and consider a suitable numerical method
for solving it. The next step is to write an algorithm for implementing the method.
An algorithm is defined as a finite sequence of unambiguous steps to be followed
for solving a given problem. Finally, one has to write a computer program in a
suitable programming language. A computer program is a sequence of computer
instructions for solving a problem.
It is possible to write more than one algorithm to solve a specific problem.
But one should analyse them before writing a computer program. The analysis
involves checking their correctness, robustness, efficiency and other characteristics.
The analysis is helpful for solving the problem on a computer. The analysis of
correctness of an algorithm ensures that the algorithm gives a correct solution of
the problem. The analysis of robustness is required to ascertain if the algorithm is
capable of tackling the problem for possible cases or for all possible variations of
the parameters of the problem. The efficiency is concerned with the computational
complexities and the total time required to solve the problem.
Computer oriented numerical methods must deal with algorithms for
implementation of numerical methods on a computer. The following algorithms of
some simple problems will make the concept clear.
Consider the problem of solving a pair of linear equations in two unknowns
given by,
a1 x  b1 y  c1
a2 x  b2 y  c2

where a1, b1, c1, a2, b2, c2 are real constants. The solution of the equations are
given by cross multiplication as,
b2 c1  b1c2 c2 a1  c1a2
x , y
a1b2  a2 b1 a1b2  a2 b1
It may be noted that if a1 b2 – a2 b1 = 0, then the solution does not exist. This
aspect has to be kept in mind while writing the algorithm as given below.

Algorithm: Solution of a pair of equations a1 x  b1 y  c1 , a2 x  b2 y  c2

Step 1: Read a1, b1, c1, a2, b2, c2
Step 2: Compute d = a1 b2 – a2 b1
Step 3: Check if d = 0, then go to Step 8 else
go to next step
Self - Learning
Step 4: Compute x Material 13
Representation of Numbers Step 5: Compute y
Step 6: Write ‘x =’, x, ‘y =’, y
Step 7: Go to Step 9
NOTES Step 8: Write ‘No Solution’
Step 9: Stop
Example 1.13: Write an algorithm to compute the roots of a quadratic equation,
ax2 + bx + c = 0.
Solution: We know that the roots of the quadratic equation are given by,
 b  b 2  4ac
x
2a

Further, if b 2  4 ac, the roots are real, otherwise they are complex conjugates.
This aspect is to be considered while writing an algorithm.
Algorithm: Computation of roots of a quadratic equation.
Step 1: Read a, b, c
Step 2: Compute d = b2 – 4ac
Step 3: Check if d  0, go to Step 4 else go to Step 8

Step 4: Compute x1  (b  d ) (2a)

Step 5: Compute x 2  (b  d ) (2a)

Step 6: Write ‘Roots are real’, x1, x2
Step 7: Go to Step 11

Step 8: Compute xi   d ( 2a)

Step 9: Compute x r  b (2a)
Step 10: Write ‘Roots are complex’, ‘Real part =’ xr ‘Imaginary part =’, xi
Step 11: Stop

Check Your Progress

1. What are the two parts of floating point representation?
2. Define truncation and computational errors.
3. How will you define the inherent errors?
4. What is propagated round-off error?
5. What are significance errors?
6. Write the situations when loss of significant digits occur.
7. Why we write an algorithm?
8. Define features and purpose of computational algorithms.

Self - Learning
14 Material
 Representation of Numbers
1.5 SOLVING EQUATION
In this section, we consider numerical methods for computing the roots of an
equation of the form, NOTES
f (x) = 0 (1.1)
where f (x) is a reasonably well-behaved function of a real variable x. The function
may be in algebraic form or polynomial form given by,
f ( x)  a n x n  a n1 x n 1  ...  a1 x  a0 (1.2)
It may also be an expression containing transcendental functions such as cos
x ,
sin x, e , etc. First, we would discuss methods to find the isolated real roots of a
x

single equation. Later, we would discuss methods to find the isolated roots of a
system of equations, particularly of two real variables x and y, given by
f (x, y) = 0 , g (x, y) = 0 (1.3)
A root of an equation is usually computed in two stages. First, we find the
location of a root in the form of a crude approximation of the root. Next we use
an iterative technique for computing a better value of the root to a desired accuracy
in successive approximations/computations. This is done by using an iterative
function.

Methods for Finding Location of Real Roots

The location or crude approximation of a real root is determined by the use of any
one of the two methods, (a) Graphical and (b) Tabulation.
Graphical Method: In the graphical method, we draw the graph of the function
y = f (x), for a certain range of values of x. The abscissae of the points where the
graph intersects the x-axis are crude approximations for the roots of the Equation
(1.1). For example, consider the equation,
f (x) = x2 + 2x – 1 = 0
From the graph of the function y = f (x) shown in Figure 1.1, we find that it
cuts the x-axis between 0 and 1. We may take any point in [0, 1] as the crude
approximation for one root. Thus, we may take 0.5 as the location of a root. The
other root lies between – 2 and – 3. We can take – 2.5 as the crude approximation
of the other root.

Self - Learning
Fig. 1.1 Graph of y  x 2  2 x  1 Material 15
Representation of Numbers In some cases, where it is complicated to draw the graph of y = f (x), we may
rewrite the equation f (x) = 0, as f1(x) = f2(x), where the graphs of y = f1 (x) and
y = f2(x) are standard curves. Then we find the x-coordinate(s) of the point(s) of
intersection of the curves y = f1(x) and y = f2(x), which is the crude approximation
NOTES of the root (s).
For example, consider the equation
x 3  15.2 x  13.2  0
This can be rewritten as,
x 3  15.2 x  13.2
where it is easy to draw the graphs of y = x3 and y = 15.2 x + 13.2. Then, the
abscissa of the point(s) of intersection can be taken as the crude approximation(s)
of the root(s).

20 y = 15.2 x + 13.2

10 y = x3

Fig. 1.2 Graph of y = x3 and y = 15.2x + 13.2

Example 1.14: Find the location of the root of the equation x log10 x  1.
1
Solution: The equation can be rewritten as log10 x  .
x
1
Now the curves y  log10 x and y  , can be easily drawn and are shown
x
in Figure below.
Y
1
y= x

y = log10 x

1 2 3 X
O

1
Graph of y  and y  log10 x
x
Self - Learning
16 Material
 The point of intersection of the curves has its x-coordinates value 2.5 Representation of Numbers
approximately. Thus, the location of the root is 2.5.
Tabulation Method: In the tabulation method, a table of values of f (x) is made
for values of x in a particular range. Then, we look for the change in sign in the NOTES
values of f (x) for two consecutive values of x. We conclude that a real root lies
between these values of x. This is true if we make use of the following theorem on
continuous functions.
Theorem 1.1: If f (x) is continuous in an interval (a, b), and f (a) and f(b) are of
opposite signs, then there exists at least one real root of f (x) = 0, between a and
b.
Consider for example, the equation f (x) = x3 – 8x + 5 = 0.
Constructing the following table of x and f (x),
x  4  3  2 1 0 1 2 3
f ( x)  27 2 13 12 5  2  3 8

we observe that there is a change in sign of f (x) in each of the sub-intervals (–3,
–4), (0, 1) and (2, 3). Thus we can take the crude approximation for the three real
roots as – 3.2, 0.2 and 2.2.

1.5.1 Bisection Method and Convergence of the

Iterative Method
Bisection Method: The bisection method is a root finding method which
repeatedly bisects an interval and then selects a subinterval in which a root must lie
for further processing. It is an extremely simple and robust method, but it is relatively
slow. It is normally used for obtaining a rough approximation to a solution which is
then used as a starting point for more rapidly converging methods. When an interval
contains more than one root, the bisection method can find one of them. When an
interval contains a singularity, the bisection method converges to that singularity.
The notion of the bisection method is based on the fact that a function will change
sign when it passes through zero. By evaluating the function at the middle of an
interval and replacing whichever limit has the same sign, the bisection method can
halve the size of the interval in each iteration to find the root.
Thus, the bisection method is the simplest method for finding a root to an
equation. It needs two initial estimates xa and xb which bracket the root. Let
fa = f(xa) and fb = f(xb) such that fa fb  0. Evidently, if fa fb = 0 then one or both
of xa and xb must be a root of f(x) = 0. As shown in Figure 1.3 is a graphical
representation of the bisection method showing two initial guesses xa and xb
bracketing the root.

Self - Learning
Material 17
Representation of Numbers

NOTES

Fig. 1.3 Graph of the Bisection Method showing Two Initial Estimates
xa and xb Bracketing the Root

The method is applicable when we wish to solve the equation f(x) = 0 for the real
variable x, where f is a continuous function defined on an interval [a, b] and f(a)
and f(b) have opposite signs.
The bisection method involves successive reduction of the interval in which
an isolated root of an equation lies. This method is based upon an important theorem
on continuous functions as stated below.
Theorem 1.2: If a function f (x) is continuous in the closed interval [a, b], and
f (a) and f (b) are of opposite signs, i.e., f (a) f (b) < 0, then there exists at least
one real root of f (x) = 0 between a and b.
The bisection method starts with two guess values x0 and x1. Then this interval
1
[x0, x1] is bisected by a point x2  ( x0  x1 ), where f(x0) . f(x1) < 0. We compute
2
f(x2). If f(x2) = 0, then x2 is a root. Otherwise, we check whether f(x0) . f(x2) < 0
or f(x1) . f(x2) < 0. If f (x2)/f (x0) < 0, then the root lies in the interval (x2, x0).
Otherwise, if f(x0) . f(x1) < 0, then the root lies in the interval (x2, x1).
The sub-interval in which the root lies is again bisected and the above process
is repeated until the length of the sub-interval is less than the desired accuracy.
The bisection method is also termed as bracketing method, since the method
successively reduces the gap between the two ends of an interval surrounding the
real root, i.e., brackets the real root.
The algorithm given below clearly shows the steps to be followed in finding a
real root of an equation, by bisection method to the desired accuracy.
Algorithm: Finding root using bisection method.
Step 1: Define the equation, f (x) = 0
Step 2: Read epsilon, the desired accuracy
Setp 3: Read two initial values x0 and x1 which bracket the desired root
Step 4: Compute y0 = f (x0)
Step 5: Compute y1 = f (x1)
Step 6: Check if y0 y1 < 0, then go to Step 6
else go to Step 2
Self - Learning
18 Material
Step 7: Compute x2 = (x0 + x1)/2
 Step 8: Compute y2 = f (x2) Representation of Numbers

Step 9: Check if y0 y2 > 0, then set x0 = x2

else set x1 = x2
Step 10: Check if | ( x1  x0 ) / x1 | > epsilon, then go to Step 3 NOTES
Step 11: Write x2, y2
Step 12: End
Next, we give the flowchart representation of the above algorithm to get a
better understanding of the method. The flowchart also helps in easy implementation
of the method in a computer program.

Flowchart for Bisection Algorithm

Begin

Define f (x)

Read epsilon

Read x0, x1

Compute y0 = f (x0)

Compute y1 = f (x1)

Is
No y0y1 > 0

Yes

Compute x2 = ( x0 + x1)/2

Compute y2 = f (x2)

Is Yes x0 = x2
y0y2 > 1

No
x1 = x2

Is
No Yes
|(x1 – x0) / x0|
> epsilon

print ‘root’ = x2

Self - Learning
End
Material 19
Representation of Numbers Example 1.15: Find the location of the smallest positive root of the equation
x3 – 9x + 1 = 0 and compute it by bisection method, correct to two decimal
places.
Solution: To find the location of the smallest positive root we tabulate the function
NOTES
f (x) = x3 – 9x + 1 below.

x 0 1 2 3
f ( x) 1  2  9 1

We observe that the smallest positive root lies in the interval [0, 1]. The
computed values for the successive steps of the bisection method are given in the
table.

n x0 x1 x2 f ( x2 )
1 0 1 0 .5  3.37
2 0 0.5 0.25  1.23
3 0 0.25 0.125  0.123
4 0 0.125 0.0625 0.437
5 0.0625 0.125 0.09375 0.155
6 0.09375 0.125 0.109375 0.016933
7 0.109375 0.125 0.11718  0.053

From the above results, we conclude that the smallest root correct to two
decimal places is 0.11.
Simple Iteration Method: A root of an equation f (x) = 0, is determined using
the method of simple iteration by successively computing better and better
approximation of the root, by first rewriting the equation in the form,
x = g(x) (1.4)
Then, we form the sequence {xn} starting from the guess value x0 of the root
and computing successively,
x1  g ( x0 ), x2  g ( x1 ),.., xn  g ( xn 1 )
In general, the above sequence may converge to the root  as n   or it
may diverge. If the sequence diverges, we shall discard it and consider another
form x = h(x), by rewriting f (x) = 0. It is always possible to get a convergent
sequence since there are different ways of rewriting f (x) = 0 in the form x = g(x).
However, instead of starting computation of the sequence, we shall first test whether
the form of g(x) can give a convergent sequence or not. We give below a theorem
which can be used to test for convergence.
Theorem 1.3: If the function g(x) is continuous in the interval [a, b] which contains
a root of the equation f (x) = 0, and is rewritten as x = g(x), and | g ( x) |  l  1 in
this interval, then for any choice of x0  [a, b] , the sequence {xn} determined by
the iterations,
xk 1  g ( xk ), for k  0, 1, 2,... (1.5)
Self - Learning This converges to the root of f (x) = 0.
20 Material
Proof: Since x = , is a root of the equation x = g(x), we have Representation of Numbers

  g () (1.6)
The first iteration gives x1 = g(x0) (1.7)
Subtracting Equation (1.7) from Equation (1.6), we get NOTES

  x1  g ()  g ( x0 )
Applying mean value theorem, we can write
  x1  (  x0 ) g  ( s0 ), x0  s0   (1.8)
Similarly, we can derive
  x2  (  x1 ) g  ( s1 ), x1  s1   (1.9)
....
  xn 1  (  xn ) g ( sn ), xn  sn   (1.10)

From Equations (1.8), (1.9) and (1.10), we get

  xn 1  (  x0 ) g  ( s0 ) g  ( s1 )... g  ( sn ) (1.11)
Since | g ( xi ) |  l for each xi, the above Equation (1.11) becomes,
|   xn 1 |  l n 1 |   x0 | (1.12)
Evidently, since l < l, l n 1  0, as n  , the right hand side tends to zero
and thus it follows that the sequence {xn}converges to the root  if   ()  1. This
completes the proof.
Order of Convergence: The order of convergence of an iterative process is
determined in terms of the errors en and en+1 in successive iterations. An iterative
e
process is said to have kth order of convergence if lim n k1  M , where M is a
n  e
n
finite number.
Roughly speaking, the error in any iteration is proportional to the kth power of
the error in the previous iteration.
Evidently, the simple iteration discussed in this section has its order of
convergence 1.
The above iteration is also termed as fixed point iteration since it determines
the root as the fixed point of the mapping defined by x = g(x).
Algorithm: Computation of a root of f (x) = 0 by linear iteration.
Step 1: Define g(x), where f (x) = 0 is rewritten as x = g(x)
Step 2: Input x0, epsilon, maxit, where x0 is the initial guess of root, epsilon is
accuracy desired, maxit is the maximum number of iterations allowed.
Step 3: Set i = 0
Step 4: Set x1 = g (x0)
Step 5: Set i = i + 1
Step 6: Check, if |(x1 – x0)/ x1| < epsilon, then print ‘Root is’, x1
Self - Learning
else go to Step 6 Material 21
Representation of Numbers Step 7: Check, if i < n, then set x0 = x1 and go to Step 3
Step 8: Write ‘No convergence after’, n, ‘Iterations’
Step 9: End
NOTES Example 1.16: In order to compute a real root of the equation x3 – x – 1 = 0,
near x = 1, by iteration, determine which of the following iterative functions can be
used to give a convergent sequence.
x 1 x 1
(i) x = x3 – 1 (ii) x  (iii) x 
x2 x
Solution:
(i) For the form x  x3  1, g(x) = x3 – 1 and g  ( x )  3x 2 . Hence,
| g  ( x ) |  1, for x near 1. So, this form would not give a convergent
sequence of iterations.
x 1 x 1 1 2
(ii) For the form x  , g ( x )  2 . Thus, g  ( x )   2  3 and
x 2
x x x
| g  (1) |  3  1. Hence, this form also would not give a convergent
sequence of iterations.
1

x 1 1  x 1 2  1 
(iii) For the form, g ( x)  , g ( x)       2 .
x 2 x   x 

1 x 1
 | g (1) |   1. Hence, the form x  would give a convergent
2 2 x
sequence of iterations.
Example 1.17: Compute the real root of the equation x3 + x2 – 1 = 0, correct to
five significant digits, by iteration method.
Solution: The equation has a real root between 0 and 1 since f (x) = x3 + x2 – 1
has opposite signs at 0 and 1. For using iteration, we first rewrite the equation in
the following different forms:
1 1 1
(i) x 1 (ii) x 1 (iii) x
x2 x x 1
1 2
For the form (i), g ( x)  1  2
, g ( x)   and for x in (0, 1), | g  ( x) | 1 .
x x3
So, this form is not suitable.

1 1  1 
For the form (ii) g ( x)  .   2  1 and | g ( x) |  1 for all x in
2 1  x 
1
x
(0, 1).
1 1
Finally, for the form (iii) g  ( x)   . 3
and g  ( x)  1 for x in (0, 1).
2
( x  1) 2

Self - Learning
22 Material
 Thus this form can be used to form a convergent sequence for finding the root. Representation of Numbers

1
We start the iteration x  with x0 = 1. The results of suecessive iterations
1 x
are, NOTES
x1  0.70711 x2  0.76537 x3  0.75236 x4  0.75541
x5  0.75476 x6  0.75490 x7  0.75488 x8  0.75488

Thus, the root is 0.75488, correct to five significant digits.

Example 1.18: Compute the root of the equation x2 – x – 0.1 = 0, which lies in
(1, 2), correct to five significant figures.
Solution: The equation is rewritten in the following form for computing the root
by iteration:
1
x  x  0.1. Here, g ( x)  and | g ( x) |  1, for x in (1, 2).
2 x  0.1
The results for successive iterations, taking x0 = 1, are
x1 = 1.0488 x2 = 1.0718 x3 = 1.0825
x4 = 1.0874 x5 = 1.0897.
Thus, the root is 1.09, correct to three significant figures.
Example 1.19: Solve the following equation for the root lying in (2, 4) by using
the method of linear iteration: x3 – 9x + 1 = 0. Show that there are various ways
of rewriting the equation in the form, x = g (x) and choose the one which gives a
convergent sequence for the root.
Solution: We can rewrite the equation in the following different forms:

1 3 1 1
(i) x ( x  1) (ii) x  9/ x 2 (iii) x  9
9 x x

12
In case of (i), g ( x)  x and for x in [2, 4], | g  ( x) | 1. Hence it will not give
3
rise to a convergent sequence.
9 2
In case of (ii) g  ( x )  2 x  2
 3 and for x in [2, 4], | g  ( x) | 1
x x
1

In case of (iii) g  ( x)   9  1  1
2
and| g  ( x) | 1
 x 2x2
Thus, the forms (ii) and (iii) would give convergent sequences for finding the
root in [2, 3].
We start the iterations taking x0 = 2 in the iteration scheme (iii). The result for
successive iterations are,
x0 = 2.0 x1 = 2.91548 x4 = 2.94282
x2 = 2.94228 x3 = 2.94281
Self - Learning
Thus, the root can be taken as 2.94281, correct to four decimal places. Material 23
Representation of Numbers
1.5.2 Newton-Raphson Method
Newton-Raphson method is a widely used numerical method for finding a root of
an equation f (x) = 0, to the desired accuracy. It is an iterative method which has
a faster rate of convergence and is very useful when the expression for the derivative
NOTES
f (x) is not complicated. Newton-Raphson method, also called the Newton’s
method, is a root finding algorithm that uses the first few terms of the Taylor series
of a function f(x) in the neighborhood of a suspected root. In the Newton-Raphson
method, to find the root start with an initial guess x1 at the root, the next guess x2 is
the intersection of the tangent from the point [x1, f(x1)] to the x-axis. The next
guess x3 is the intersection of the tangent from the point [x2, f(x2)] to the x-axis as
shown in Figure 1.4.

f(x)
f(x1) B

x3 x2 x1

Fig. 1.4 Graph of the Newton-Raphson Method

The Newton-Raphson method can be derived from the definition of a slope as
follows:
f ( x1 )  0 f ( x1 )
f (x1) =  x2 = x1 –
x1  x2 f '( x1 )
As a general rule, from the point [xn, f(xn)], the next guess is calculated as
follows:
f ( xn )
xn+1 = xn –
f '( xn )
The derivative or slope f(xn) can be approximated numerically as follows:
f ( xn  x )  f ( xn )
f (xn) =
x

To derive the formula for this method, we consider a Taylor’s series expansion of
f (x0 + h), x0 being an initial guess of a root of f (x) = 0 and h a small correction to
the root.

h2
Self - Learning f ( x0  h)  f ( x0 )  h f  ( x0 )  f" ( x0 )  ...
24 Material 2 !
 Assuming h to be small, we equate f (x0 + h) to 0 by neglecting square and Representation of Numbers

higher powers of h.
 f ( x0 )  h f  ( x0 )  0 2
NOTES
f ( x0 )
or, h
f  ( x0 )
Thus, we can write an improved value of the root as,

x1  x0  h
f ( x0 )
i.e., x1  x0 
f  ( x0 )

Successive approximations x2 , x3 ,..., xn 1 can thus be written as,

f ( x1 )
x2  x1 
f ( x1 )
f ( x2 )
x3  x 2 
f ( x2 )
... ... ...
f ( xn )
xn 1  xn 
f ( xn ) (1.13)

If the sequence {xn } converges, we get the root.

Algorithm: Computation of a root of f (x) = 0 by Newton-Raphson method.
Step 1: Define f (x), f (x)
Step 2: Input x0, epsilon, maxit
[x0 is the initial guess of root, epsilon is the desired accuracy of the
root and maxit is the maximum number of iterations allowed]
Step 3: Set i = 0
Step 4: Set f0 = f (x0)
Step 5: Compute df0 = f  (x0)
Step 6: Set x1 = x0 – f0/df0
Step 7: Set i = i + 1
Step 8: Check if |(x1 – x0) |x1| < epsilon, then print ‘Root is’, x1 and stop
else if i < n, then set x0 = x1 and go to Step 3
Step 9: Write ‘Iterations do not converge’
Step 10: End
Example 1.20: Use Newton-Raphson method to compute the positive root of
the equation x3 – 8x – 4 = 0, correct to five significant digits.
Solution: Newton-Raphson iterative scheme is given by,
f ( xn )
xn 1  xn  , for n  0, 1, 2, ... Self - Learning
f  ( xn ) Material 25
Representation of Numbers For the given equation, f (x) = x3 – 8x – 4.
First we find the location of the root by the method of tabulation. The table for f
(x) is,

NOTES x 0 1 2 3 4
f ( x)  4  13  12  1 28

Evidently, the positive root is near x = 3. We take x0 = 3 in Newton-Raphson

iterative scheme.
xn3  8 xn  4
xn 1  xn 
3xn2  8
27  24  4
We get, x1  3   3.0526
27  8
Similarly, x2 = 3.05138 and x3 = 3.05138.
Thus, the positive root is 3.0514, correct to five significant digits.
Example 1.21: Find a real root of the equation x3  7 x 2  9  0, correct to five
significant digits.
Solution: First we find the location of the real root by tabulation. We observe that
the real root is negative and since f (–7) = 9 > 0 and f (–8) = – 55 < 0, a root lies
between –7 and – 8.
For computing the root to the desired accuracy, we take x0 = –8 and use
Newton-Raphson iterative formula,
xn3  7 xn2  9
xn 1  xn  , for n  0, 1, 2, ...
3 xn2  14 xn
The successive iterations give,
x1 = –7.3125
x2 = –7.17966
x3 = –7.17484
x4 = –7.17483
Hence, the desired root is –7.1748, correct to five significant digits.
1 a 
Example 1.22: For evaluating a , deduce the iterative formula xn 1   xn  ,
2 xn 

by using Newton-Raphson scheme of iteration. Hence, evaluate 2 using this,

correct to four significant digits.
Solution: We observe that, a is the solution of the equation x2 – a = 0.
Now, using f (x) = x2 – a in the Newton-Raphson iterative scheme,
xn2  a
xn 1  xn 
2 xn

xn2  a
We have, xn 1  xn 
2 xn

Self - Learning xn2  a

26 Material xn 1  xn 
2 xn
 Representation of Numbers
1 a
i.e., xn 1   xn   , for n  0, 1, 2,...
2 xn 

Now, for computing 2 , we assume x0 = 1.4. The successive iterations give,

NOTES
1 2  3.96
x1  1.4    1.414
2 1.4  2.8
1 2 
x2  1.414    1.41421
2 1.414 

Hence, the value of 2 is 1.414 correct to four significant digits.

Example 1.23: Prove that k a can be computed by the iterative scheme,
1 a 
x n 1  (k  1) x n  k 1 . Hence evaluate 3
2 , correct to five significant digits.
k  xn 

Solution: The value k a is the positive root of xk – a = 0. Thus, the iterative

scheme for evaluating k a is,
xnk  a
xn 1  xn  1
kxnk

1 a 
or, xn 1  (k  1) xn  k 1  , for n  0, 1, 2,...
k xn 
Now, for evaluating 3 2 , we take x0 = 1.25 and use the iterative formula,
1 2
xn 1   2 xn  2  .
3 xn 

1 2 
We have, x1  1.25  2    1.26
3 (1.25)2 

x2  1.259921, x3  1.259921

Hence, 3 2 = 1.2599, correct to five significant digits.

Example 1.24: Find by Newton-Raphson method, the real root of 3x – cos x –
1 = 0, correct to three significant figures.
Solution: The location of the real root of f (x) = 3x – cos x – 1 = 0, is [0, 1] since
f (0) = – 2 and f (1) > 0.
We choose x0 = 0 and use Newton-Raphson scheme of iteration.
3 xn  cos xn  1
xn 1  xn  , n  0, 1, 2,...
3  sin xn
The results for successive iterations are,
x1 = 0.667, x2 = 0.6075, x3 = 0.6071
Thus, the root is 0.607 correct to three significant figures.
Example 1.25: Find a real root of the equation xx + 2x – 6 = 0, correct to four
Self - Learning
significant digits. Material 27
Representation of Numbers Solution: Taking f (x) = xx + 2x – 6, we have f (1) = –3 < 0 and f (2) = 2 > 0.
Thus, a root lies in [1, 2]. Choosing x0 = 2, we use Newton-Raphson iterative
scheme given by,
x
NOTES x n  2 xn  6
xn 1  xn  xn n , for n  0, 1, 2,...
xn (log e xn  1)  2
The computed results for successive iterations are,
446
x1  2   1.72238
4  (log e 2 x 2  1)  2
x2  1.72321
x3  1.72308
Hence, the root is 1.723 correct to four significant figures.
Order of Convergence: We consider the order of convergence of the Newton-
Raphson method given by the formula,
f ( xn )
xn 1  xn 
f  ( xn )
Let us assume that the sequence of iterations {xn} converge to the root 
Then, expanding by Taylor’s series about xn, the relation f () = 0, gives
1
f ( xn )  (  xn ) f  ( xn )  (  xn ) 2 f  ( xn )  ...  0
2

f ( xn ) 1 f  ( xn )
     xn  (  xn ) 2 .  ...
f  ( xn ) 2 f '( xn )
1 f  ( xn )
 xn 1    (  xn ) 2 .
2 f  ( xn )

Taking n as the error in the nth iteration and writing n = n – , we have,

1 2 f  ()
 n 1  n (1.14)
2 f  ( )

Thus,  n 1  k  2 n, where k is a constant.

This shows that the order of convergence of Newton-Raphson method is 2.
In other words, the Newton-Raphson method has a quadratic rate of convergence.
The condition for convergence of Newton-Raphson method can easily be
derived by rewriting the Newton-Raphson iterative scheme as xn 1   ( xn ) with
f ( x)
 ( x)  x  .
f  ( x)
Hence, using the condition for convergence of the linear iteration method, we
f ( x) f  ( x)
can write   ( x)  .
[ f  ( x)]2
Self - Learning
28 Material
 Thus, the sufficient condition for the convergence of Newton-Raphson method Representation of Numbers
is,

f ( x) f  ( x)
 1, in the interval near the root.
[ f  ( x)]2 NOTES

i.e., | f ( x) f ( x) |  | f  ( x ) |2 (1.15)

1.5.3 Secant Method

Secant method can be considered as a discretized form of Newton-Raphson
method. The iterative formula for this method is obtained from formula of Newton-
Raphson method on replacing the derivative f ( x0 ) by the gradient of the chord
joining two neighbouring points x0 and x1 on the curve y = f (x).
Thus, we have
f ( x1 )  f ( x0 )
f  ( x0 ) 
x1  x0
The iterative formula is given by,
f ( x1 )
x2  x1  ( x1  x0 )
f ( x1 )  f ( x0 )
This can be rewritten as,
x0 f ( x1 )  x1 f ( x0 )
x2 
f ( x1 )  f ( x0 )
Secant formula in general form is,
xn = xn–1 – f(xn–1)
xn –1  xn –1
f ( xn –1 )  f ( xn – 2 )
The iterative formula is equivalent to the one for Regula-Falsi method. The
distinction between secant method and Regula-Falsi method lies in the fact that
unlike in Regula-Falsi method, the two initial guess values do not bracket a root
and the bracketing of the root is not checked during successive iterations, in secant
method. Thus, secant method may not always give rise to a convergent sequence
to find the root. The geometrical interpretation of the method is shown in
Figure 1.5.

(Line AB meets x-axis alone)

Fig. 1.5 Secant Method Self - Learning
Material 29
Representation of Numbers Algorithm: To find a root of f (x) = 0, by secant method.
Step 1: Define f (x).
Step 2: Input x0, x1, error, maxit [x0, x1, are initial guess values, error is the
NOTES prescribed precision and maxit is the maximum number of iterations
allowed].
Step 3: Set i = 1
Step 4: Compute f0 = f (x0)
Step 5: Compute f1 = f (x1)
Step 6: Compute x2 = (x0 f1 – x1 f0)/(f1 – f0)
Step 7: Set i = i + 1
Step 8: Compute accy = |x2 – x1| / |x1|
Step 9: Check if accy < error, then go to Step 14
Step 10: Check if i  maxit then go to Step 16
Step 11: Set x0 = x1
Step 12: Set x1 = x2
Step 13: Go to Step 6
Step 14: Print ‘Root =’, x2
Step 15: Go to Step 17
Step 16: Print ‘Iterations do not converge’
Step 17: Stop

1.5.4 Regula-Falsi Method

Regula-Falsi method is also a bracketing method. As in bisection method, we
start the computation by first finding an interval (a, b) within which a real root lies.
Writing a = x0 and b = x1, we compute f (x0) and f (x1), and check if f (x0) and f
(x1) are of opposite signs. For determining the approximate root x2, we find the
point of intersection of the chord joining the points (x0, f (x0)) and (x1, f (x1)) with
the x-axis, i.e., the curve y = f (x0) is replaced by the chord given by,
f ( x1 )  f ( x0 )
y  f ( x0 )  ( x  x0 ) (1.16)
x1  x0

Thus, by putting y = 0 and x = x2 in Equation (1.16), we get

f ( x0 )
x 2  x0  ( x1  x0 ) (1.17)
f ( x1 )  f ( x0 )

Next, we compute f (x2) and determine the interval in which the root lies in the
following manner. If (a) f (x2) and f (x1) are of opposite signs, then the root lies in
(x2, x1). Otherwise if (b) f (x0) and f (x2) are of opposite signs, then the root lies in
(x0, x2). The next approximate root is determined by changing x0 by x2 in the first
case and x1 by x2 in the second case.
The aforesaid process is repeated until the root is computed to the desired
accuracy , i.e., the condition
Self - Learning
30 Material
 Representation of Numbers
( xk 1  xk ) / xk  , should be satisfied.
Regula-Falsi method can be geometrically interpreted by the following
Figure 1.6.
NOTES
Y

x1, f (x1)

X
O
x2, f (x2)

x0, f (x2)

Fig. 1.6 Regula-Falsi Method

Algorithm: Computing root of an equation by Regula-Falsi method.
Step 1: Define f (x)
Step 2: Read epsilon, the desired accuracy
Step 3: Read maxit, the maximum number of iterations
Step 4: Read x0, x1, two initial guess values of root
Step 5: Compute f0 = f (x0)
Step 6: Compute f1 = f (x1)
Step 7: Check if f0 f1 < 0, then go to the next step
else go to Step 4
Step 8: Compute x2 = (x0 f1 – x1 f0) / (f1 – f0)
Step 9: Compute f2 = f (x2)
Step 10: Check if |f2| < epsilon, then go to Step 18
Step 11: Check if f2 f0 < 0 then go to the next step
else go to Step 15
Step 12: Set x1 = x2
Step 13: Set f1 = f2
Step 14: Go to Step 7
Step 15: Set x0 = x2
Step 16: Set f0 = f2
Step 17: Go to Step 7
Step 18: Write ‘Root =’ , x2, f3
Step 19: End
Example 1.26: Use Regula-Falsi method to compute the positive root of x3 – 3x
– 5 = 0, correct to four significant figures.
Solution: First we find the interval in which the root lies. We observe that f (2)
= –3 and f (3) = 13. Thus, the root lies in [2, 3]. For using the Regula-Falsi
Self - Learning
method, we use the formula, Material 31
Representation of Numbers f ( x0 )
x 2  x0  ( x1  x0 )
f ( x1 )  f ( x0 )
with x0 = 2, and x1 = 3, we have
NOTES 3
x2  2  (3  2)  2.1875
13  3
Again, since f (x2) = f (2.1875) = –1.095, we consider the interval [2.1875,
3]. The next approximation is x3 = 2.2461. Also, f (x3) = – 0.4128. Hence, the root
lies in [2.2461, 3].
Repeating the iterations, we get
x4 = 2.2684, f (x4) = – 0.1328
x5 = 2.2748, f (x5) = – 0.0529
x6 = 2.2773, f (x6) = – 0.0316
x7 = 2.2788, f (x7) = – 0.0028
x8 = 2.2792, f (x8) = – 0.0022
The root correct to four significant figures is 2.279.

Roots of Polynomial Equations

Polynomial equations with real coefficients have some important characteristics
regarding their roots. A polynomial equation of degree n is of the form pn(x) =
anxn + an–1xn–1 + an–2xn–2 + ... + a2x2 + a1x + a0 = 0.
(i) A polynomial equation of degree n has exactly n roots.
(ii) Complex roots occur in pairs, i.e., if   i  is a root of pn(x) = 0, then
  i  is also a root.
(iii) Descarte’s rule of signs can be used to determine the number of possible
real roots (positive or negative).
(iv) If x1, x2,..., xn are all real roots of the polynomial equation, then we can
express pn(x) uniquely as,
pn ( x )  an ( x  x1 )( x  x2 )...( x  xn )
(v) pn(x) has a quadratic factor for each pair of complex conjugate roots.
Let,   i  and   i  be the roots, then {x 2  2 x  ( 2   2 )} is the
quadratic factor.
(vi) There is a special method, known as Horner’s method of synthetic
substitution, for evaluating the values of a polynomial and its derivatives
for a given x.

1.5.5 Descarte’s Rule

The number of positive real roots of a polynomial equation is equal to the number
of changes of sign in pn(x), written with descending powers of x, or less by an
even number.
Consider for example, the polynomial equation,
3x5  2 x 4  x3  2 x 2  x  2  0
Self - Learning
32 Material
 Clearly there are three changes of sign and hence the number of positive real Representation of Numbers

roots is three or one. Thus, it must have a real root. In fact, every polynomial
equation of odd degree has a real root.
We can also use Descarte’s rule to determine the number of negative roots by
NOTES
finding the number of changes of signs in pn(–x). For the above equation,
pn ( x)  3 x 5  2 x 4  x3  2 x 2  x  2  0 and it has two changes of sign. Thus,
it has either two negative real roots or none.

Check Your Progress

9. The roots of an equation are computed in how many stages?
10. Define tabulation method.
11. State the procedure of bisection method.
12. How is the order of convergence of an iterative process determined?
13. State a property of Newton-Raphson method.
14. Define secant method.
15. Give the procedure of Regula-Falsi method.

1.6 ANSWERS TO ‘CHECK YOUR PROGRESS’

1. The floating point representation of a number consists of mantissa and
exponent.
2. The errors in a numerical solution are basically of two types. They are
truncation error and computational error. The error which is inherent in the
numerical method employed for finding numerical solution is called the
truncation error. The computational error arises while doing arithmetic
computation due to representation of numbers with a finite number of decimal
digits.
3. Inherent errors are errors in the data which are obtained by physical
measurement and are due to limitations of the measuring instrument. The
analysis of errors in the computed result due to the inherent errors in data is
similar to that of round-off errors.
4. Propagated round-off error is the sum of two approximate numbers (having
round-off errors) equal to the sum of the round-off errors in the individual
numbers.
5. During arithmetic computations of approximate numbers having fixed
precision, there may be loss of significant digits in some cases. The error
due to loss of significant digits is termed as significance error.
6. There are two situations when loss of significant digits occur. These are,
(i) Subtraction of two nearly equal numbers
(ii) Division by a very small divisor compared to the dividend
7. To get a numerical solution on a computer, one has to write an algorithm.
Self - Learning
Material 33
Representation of Numbers 8. For solving problems with the help of a computer, one should first analyse
the mathematical formulation of the problem and consider a suitable numerical
method for solving it. The next step is to write an algorithm for implementing
the method.
NOTES
9. A root of an equation is usually computed in two stages. First, we find the
location of a root in the form of a crude approximation of the root. Next we
use an iterative technique for computing a better value of the root to a
desired accuracy in successive approximations/computations.
10. In the tabulation method, a table of values of f (x) is made for values of x in
a particular range. Then, we look for the change in sign in the values of f (x)
for two consecutive values of x. We conclude that a real root lies between
these values of x.
11. The bisection method involves successive reduction of the interval in which
an isolated root of an equation lies. The sub-interval in which the root lies is
again bisected and the above process is repeated until the length of the sub-
interval is less than the desired accuracy.
12. The order of convergence of an iterative process is determined in terms of
the errors en and en + 1 in successive iterations.
13. Newton-Raphson method is a widely used numerical method for finding a
root of an equation f (x) = 0, to the desired accuracy. It is an iterative
method which has a faster rate of convergence and is very useful when the
expression for the derivative f (x) is not complicated.
14. Secant method can be considered as a discretized form of Newton-Raphson
method. The iterative formula for this method is obtained from formula of
Newton-Raphson method on replacing the derivative by the gradient of
the chord joining two neighbouring points x0 and x1 on the curve y = f (x).
15. Regula-Falsi method is also a bracketing method. As in bisection method,
we start the computation by first finding an interval (a, b) within which a real
root lies. Writing a = x0 and b = x1, we compute f(x0) and f(x1) and check
if f(x0) and f(x1) are of opposite signs. For determining the approximate
root x2, we find the point of intersection of the chord joining the points (x0,
f(x0)) and (x1, f(x1)) with the x-axis, i.e., the curve y = f(x0) is replaced by
the chord given by,
f ( x1 )  f ( x0 )
y  f ( x0 )  ( x  x0 )
x1  x0

1.7 SUMMARY
 Numerical methods are methods used for solving problems through numerical
calculations providing a table of numbers and/or graphical representations
or figures. Numerical methods emphasize that how the algorithms are
implemented.
 To perform a numerical calculation, approximate them first by a
representation involving a finite number of significant digits. If the numbers
to be represented are very large or very small, then they are written in
Self - Learning floating point notation.
34 Material
  The Institute of Electrical and Electronics Engineers (IEEE) has published a Representation of Numbers

standard for binary floating point arithmetic.

 In approximate representation of numbers, the number is represented with
a finite number of digits. All the digits in the usual decimal representation
NOTES
may not be significant while considering the accuracy of the number.
 In a floating representation, a number is represented with a finite number of
significant digits having a floating decimal point.
 Floating point representation of a number consists of mantissa and exponent.
 The errors in a numerical solution are basically of two types termed as
truncation error and computational error.
 The error which is inherent in the numerical method employed for finding
numerical solution is called the truncation error.
 The truncation error arises due to the replacement of an infinite process
such as summation or integration by a finite one.
 Inherent errors are errors in the data which are obtained by physical
measurement and are due to limitations of the measuring instrument.
 Numerical methods can be employed for computing the roots of an equation
of the form, f(x) = 0, where f(x) is a reasonably well-behaved function of a
real variable x.
 The location or crude approximation of a real root is determined by the use
of any one of the following methods, (a) Graphical and (b) Tabulation.
 In general, the roots of an equation can be computed using bisection and
simple iteration methods.
 The bisection method is also termed as bracketing method.
 Newton-Raphson method is a widely used numerical method for finding a
root of an equation f (x) = 0, to the desired accuracy.
 Secant method can be considered as a discretized form of Newton-Raphson
method. The iterative formula for this method is obtained from formula of
Newton-Raphson method on replacing the derivative by the gradient of
the chord joining two neighbouring points x0 and x1 on the curve y = f (x).
 Regula-Falsi method is also a bracketing method.

1.8 KEY TERMS

 Truncation error: This error is inherent in the numerical method employed
for finding numerical solution. It occurs due to the replacement of an infinite
process such as summation or integration by a finite one.
 Computational error: This error occurs during arithmetic computation
due to representation of numbers having a finite number of decimal digits.
 Inherent error: This error occurs in the data type which is obtained using
physical measurement and also due to limitations of the measuring instruments.
 Significance error: This error occurs due to loss of significant digits. Self - Learning
Material 35
Representation of Numbers  Tabulation method: In the tabulation method, a table of values of f (x) is
made for values of x in a particular range.
 Bisection method: It involves successive reduction of the interval in which
an isolated root of an equation lies.
NOTES

1.9 SELF-ASSESSMENT QUESTIONS AND

EXERCISES

Short-Answer Questions
1. What are floating point numbers?

2. Find the percentage error in approximating 5 by 0.8333 correct upto four

6
significant figures.
3. Write the characteristics of numerical computation.
4. Find the relative error in the computation of x – y for x = 12.05 and y =
8.02 having absolute errors x  0.005 and y  0.001.

5. Find the percentage error in computing y  3x 2  6 x at x = 1, if the error

in x is 0.05.
6. Given a = 1.135 and b = 1.075 having absolute errors a = 0.011 and
b = 0.12. Estimate the relative percentage error in the computation of a – b.
7. What are isolated roots?
8. What is crude approximation in graphical method?
9. Why is bisection method also termed as bracketing method?
10. What is the order of convergence of the Newton-Raphson method?
11. State the similarity between secant method and Regula-Falsi method.
Long-Answer Questions
1. Round-off the following numbers to three decimal places:
(i) 0.230582 (ii) 0.00221118 (iii) 2.3645 (iv) 1.3455
2. Round-off the following numbers to four significant figures:
(i) 49.3628 (ii) 0.80022 (iii) 8.9325 (iv) 0.032588
(v) 0.0029417 (vi) 0.00010211 (vii) 410.99
3. Round-off each of the following numbers to three significant figures and
indicate the absolute error in each.
(i) 49.3628 (ii) 0.9002 (iii) 8.325 (iv) 0.0039417
4. Find the sum of the following approximate numbers, correct to the last
digits.
0.348, 0.1834, 345.4, 235.2, 11.75, 0.0849, 0.0214, 0.0002435
5. Find the number of correct significant digits in the approximate number
Self - Learning
36 Material 11.2461. Given is its absolute error = 0.25 × 10–2.
 6. Given are the following approximate numbers with their relative errors. Representation of Numbers

Determine the absolute errors.

(i) x A  12165,  R  0.1% (ii) x A  3.23,  R  0.6%
(iii) x A  0.798,  R  10% (iv) x A  67.84,  R  1% NOTES

7. Round-off the following numbers to four significant digits.

(i) 450.92 (ii) 48.3668 (iii) 9.3265 (iv) 8.4155
(v) 0.80012 (vi) 0.042514 (vii) 0.0049125 (viii) 0.00020215
8. Write the following numbers in floating-point form rounded to four significant
digits.
(i) 100000 (ii) – 0.0022136 (iii) – 35.666
9. Determine the number of correct digits in the number x in each of the
following (the relative errors are given).
(i) x  0.2217,  R  0.2  101 (ii) x  32.541,  R  0.1
(iii) x  0.12432,  R  10% (iv) x  0.58632,  R  1%

10. Find the percentage error in computing z  x for x = 4.44, if x is correct

to its last digit only.
11. Let u = 4x6 + 3x – 9. Find the relative percentage error in computing u at
x = 1.1, if the error in x is 0.05.
12. Use graphical method to find the location of a real root of the equation
x3 + 10x – 15 = 0.
13. Draw the graphs of the function f (x) = cos x – x, in the range [0, /2) and
find the location of the root of the equation f (x) = 0.
14. Compute the root of the equation x3 – 9x + 1 = 0 which lies between 2 and
3 correct upto three significant digits using bisection method.
15. Compute the root of the equation x3 + x2 – 1 = 0, near 1, by the iterative
method correct upto two significant digits.
16. Compute using Newton-Raphson method the root of the equation ex = 4x,
near 2, correct upto four significant digits.
17. Find the real root of x log10x – 1.2 = 0 correct upto four decimal places
using Regula-Falsi method.

1.10 FURTHER READING

Chance, William A. 1969. Statistical Methods for Decision Making. Illinois:
Richard D Irwin.
Chandan, J.S., Jagjit Singh and K.K. Khanna. 1995. Business Statistics. New
Delhi: Vikas Publishing House.
Elhance, D.N. 2006. Fundamental of Statistics. Allahabad: Kitab Mahal.
Self - Learning
Material 37
Representation of Numbers Freud, J.E., and F.J. William. 1997. Elementary Business Statistics – The
Modern Approach. New Jersey: Prentice-Hall International.
Goon, A.M., M.K. Gupta, and B. Das Gupta. 1983. Fundamentals of Statistics.
Vols. I & II, Kolkata: The World Press Pvt. Ltd.
NOTES
Gupta, S.C. 2008. Fundamentals of Business Statistics. Mumbai: Himalaya
Publishing House.
Kothari, C.R. 1984. Quantitative Techniques. New Delhi: Vikas Publishing
House.
Levin, Richard. I., and David. S. Rubin. 1997. Statistics for Management. New
Jersey: Prentice-Hall International.
Meyer, Paul L. 1970. Introductory Probability and Statistical Applications.
Massachusetts: Addison-Wesley.
Gupta, C.B. and Vijay Gupta. 2004. An Introduction to Statistical Methods,
23rd Edition. New Delhi: Vikas Publishing House Pvt. Ltd.
Hooda, R. P. 2013. Statistics for Business and Economics, 5th Edition. New
Delhi: Vikas Publishing House Pvt. Ltd.
Anderson, David R., Dennis J. Sweeney and Thomas A. Williams. Essentials of
Statistics for Business and Economics. Mumbai: Thomson Learning,
2007.
S.P. Gupta. 2021. Statistical Methods. Delhi: Sultan Chand and Sons.

Self - Learning
38 Material
 Interpolation and

UNIT 2 INTERPOLATION Curve Fitting

AND CURVE FITTING

NOTES
Structure
2.0 Introduction
2.1 Objectives
2.2 Interpolation
2.2.1 Iterative Linear Interpolation
2.2.2 Lagrange’s Interpolation
2.2.3 Finite Difference for Interpolation
2.2.4 Symbolic Operators
2.2.5 Shift Operator
2.2.6 Central Difference Operator
2.2.7 Differences of a Polynomial
2.2.8 Newton’s Forward Difference Interpolation Formula
2.2.9 Newton’s Backward Difference Interpolation Formula
2.2.10 Extrapolation
2.2.11 Inverse Interpolation
2.2.12 Truncation Error in Interpolation
2.3 Curve Fitting
2.3.1 Method of Least Squares
2.4 Trigonometric Functions
2.5 Regression
2.5.1 Linear Regression
2.5.2 Polynomial Regression
2.5.3 Fitting Exponential
2.6 Answers to ‘Check Your Progress’
2.7 Summary
2.8 Key Terms
2.9 Self-Assessment Questions and Exercises
2.10 Further Reading

2.0 INTRODUCTION
Interpolation is the process of defining a function that takes on specified values at
specified points. Polynomial interpolation is the most known one-dimensional
interpolation method. Its advantages lies in its simplicity of realization and the good
quality of interpolants obtained from it. You will learn about the various interpolation
methods, namely Lagrange’s interpolation, Newton’s forward and backward
difference interpolation formulae, iterative linear interpolation and inverse
interpolation.
Curve fitting is the process of constructing a curve, or mathematical function,
which has the best fit to a series of data points, possibly subject to constraints.
In mathematics, the trigonometric functions, also called the circular functions,
are functions of an angle. They relate the angles of a triangle to the lengths of its
sides. The most familiar trigonometric functions are the sine, cosine and tangent. In
the context of the standard unit circle (a circle with radius 1 unit), where a triangle
Self - Learning
is formed by a ray originating at the origin and making some angle with the x-axis, Material 39
Interpolation and the sine of the angle gives the length of the y-component (the opposite to the angle
Curve Fitting
or the rise) of the triangle, the cosine gives the length of the x-component (the
adjacent of the angle or the run), and the tangent function gives the slope (y-
component divided by the x-component). Trigonometric functions are commonly
NOTES defined as ratios of two sides of a right triangle containing the angle, and can
equivalently be defined as the lengths of various line segments from a unit circle.
Regression analysis, is the mathematical process of using observations to
find the line of best fit through the data in order to make estimates and predictions
about the behaviour of variables. This technique is used to determine the statistical
relationship between two or more variables and to make prediction of one variable
on the basis of one or more other variables.
In this unit, you will learn about the interpolation, curve fitting, trigonometric
function and regression.

2.1 OBJECTIVES
After going through this unit, you will be able to:
 Describe the method of iterative linear interpolation
 Understand polynomial interpolation
 Explain the importance of Lagrange’s interpolation
 Perform interpolation of equally spaced tabular values
 Explain finite, forward and backward differences
 Evaluate interpolation using symbolic, shift and central difference operators
 Know differences of polynomials
 Define Newton’s forward and backward interpolation formulae
 Explain extrapolation and inverse interpolation
 Understand the concept of curve fitting
 Explain the various trigonometric functions
 Discuss regression analysis in detail

2.2 INTERPOLATION
The problem of interpolation is very fundamental problem in numerical analysis.
The term interpolation literally means reading between the lines. In numerical analysis,
interpolation means computing the value of a function f (x) in between values of x
in a table of values. It can be stated explicitly as ‘given a set of (n + 1) values y0,
y1, y2,..., yn for x = x0, x1, x2, ..., xn respectively. The problem of interpolation is
to compute the value of the function y = f (x) for some non-tabular value of x.’
The computation is often made by finding a polynomial called interpolating
polynomial of degree less than or equal to n such that the value of the polynomial
is equal to the value of the function at each of the tabulated points. Thus if,
Self - Learning
40 Material  ( x )  a0  a1 x  a2 x 2    an x n (2.1)
is the interpolating polynomial of degree  n , then Interpolation and
Curve Fitting
 ( xi )  yi , for i  0, 1, 2, ..., n (2.2)
It is true that, in general, it is difficult to guess the type of function to
approximate f (x). In case of periodic functions, the approximation can be made NOTES
by a finite series of trigonometric functions. Polynomial interpolation is a very useful
method for functional approximation. The interpolating polynomial is also useful as
a basis to develop methods for other problems such as numerical differentiation,
numerical integration and solution of initial and boundary value problems associated
with differential equations.
The following theorem, developed by Weierstrass, gives the justification for
approximation of the unknown function by a polynomial.
Theorem 2.1: Every function which is continuous in an interval (a, b) can be
represented in that interval by a polynomial to any desired accuracy. In other
words, it is possible to determine a polynomial P(x) such that f ( x)  P( x)  ,
for every x in the interval (a, b) where  is any prescribed small quantity.
Geometrically, it may be interpreted that the graph of the polynomial y = P(x) is
confined to the region bounded by the curves y  f ( x )   and y  f ( x )   for
all values of x within (a, b), however small  may be.

Fig. 2.1 Interpolation

The following theorem is regarding the uniqueness of the interpolating

polynomial.
Theorem 2.2: For a real-valued function f (x) defined at (n + 1) distinct points
x0, x1, ..., xn, there exists exactly one polynomial of degree  n which interpolates
f (x) at x0, x1, ..., xn.
We know that a polynomial P(x) which has (n + 1) distinct roots x0, x1, ...,
xn can be written as,
P(x) = (x – x0) (x – x1) .....(x – xn) q (x)
where q(x) is a polynomial whose degree is either 0 or (n + 1) which is less than
the degree of P(x).
Suppose that two polynomials  ( x ) and  ( x ) are of degree  n and that
both interpolate f(x). Here P ( x )  ( x )   ( x ) at x  x0 , x1 ,..., xn . Then P(x)
vanishes at the n +1 points x0 , x1 ,..., xn . Thus P(x) = 0 and ( x )   ( x ). Self - Learning
Material 41
Interpolation and
Curve Fitting
2.2.1 Iterative Linear Interpolation
In this method, we successively generate interpolating polynomials, of any degree,
by iteratively using linear interpolating functions.
NOTES Let p01(x) denote the linear interpolating polynomial for the tabulated values at
x0 and x1. Thus, we can write as,
( x1  x) f 0  ( x0  x) f1
p01 ( x) 
x1  x0

This can be written with determinant notation as,

f0 x0  x
f1 x1  x (2.3)
p01 ( x) 
x1  x0

This form of p01(x) is easy to visualize and is convenient for desk computation.
Thus, the linear interpolating polynomial through the pair of points (x0, f0) and
( x j , f j ) can be easily written as,

1 f0 x0  x
p0 j ( x)  , for j  1, 2, ..., n (2.4)
x j  x0 f j xj  x

Now, consider the polynomial denoted by p01j (x) and defined by,

1 p01 ( x) x1  x
p01 j ( x)  , for j  2, 3, ..., n (2.5)
x j  x1 p0 j ( x) x j  x

The polynomial p01j(x) interpolates f(x) at the points x0, x1, xj (j > 1) and is a
polynomial of degree 2, which can be easily verified that,
p0ij ( x0 )  f 0 , p0ij ( xi )  f i and p0ij ( x j )  f j because p01 ( x0 )  f 0  p0ij ( x0 ), etc.

Similarly, the polynomial p012 j ( x) can be constructed by replacing p01(x) by

p012 (x) and p0j (x) by p01j (x).
Thus,

1 p012 ( x) x2  x
p012 j ( x)  , for j  3, 4, ..., n (2.6)
x j  x2 p01 j ( x) x j  x

Evidently, p012j (x) is a polynomial of degree 3 and it interpolates the function

at x0, x1, x2 and xj.
i.e., p012 j ( x0 )  f 0 ; p012 j ( xi )  f1 ; p012 j ( x2 )  f 2 and p012 j ( x j )  f j
This process can be continued to generate higher and higher degree
interpolating polynomials.
The results of the iterated linear interpolation can be conveniently represented
as given in the following table.

Self - Learning
42 Material
 Interpolation and
xk fk p0 j p01 j ... x j  x Curve Fitting
x0 f0 x0  x
x1 f1 p01 x1  x
x2 f2 p02 p012 x2  x
NOTES
x3 f3 p03 p013 x3  x
... ... ... ... ... ...
xj fj p0 j p01 j xj  x
... ... ... ... ... ...
xn fn p0n p01n xn  x

The successive columns of interpolation results can be conveniently filled by

computing the values of the determinants written using the previous column and
the corresponding entries in the last column xj – x. Thus, for computing p01j’s for
j = 2, 3, ..., n, we evaluate the determinant whose elements are the boldface
quantities and divide the determinant’s value by the difference ( x j  x)  ( x1  x) .
Example 2.1: Find s(2.12) using the following table by iterative linear interpolation:

x 2.0 2.1 2.2 2.3

s( x) 0.7909 0.7875 0.7796 0.7673

Solution: Here, x = 2.12. The following table gives the successive iterative linear
interpolation results. The details of the calculations are shown below in the table.

xj s( x j ) p0 j p01 j p012 j xj  x
2.0 0.7909  0.12
2.1 0.7875 0.78682  0.02
2.2 0.7796 0.78412 0.78628 0.08
2.3 0.7673 0.78146 0.78628 0.78628 0.18

1 0.7909 0.12
p01   0.78682
2.1  2.0 0.7875 0.02
1 0.7909 0.12
p02   0.78412
2.2  2.0 0.7796 0.08
1 0.7909 0.12
p03   0.78146
2.3  2.0 0.7673 0.18
1 0.78682 0.02
p012   0.78628
2.2  2.1 0.78412 0.08
1 0.78682 0.02
p013   0.78628
2.3  2.1 0.78146 0.18
1 0.78628 0.08
p012   0.78628
2.3  2.2 0.78628 0.18
The boldfaced results in the table give the value of the interpolation at x =
2.12. The result 0.78682 is the value obtained by linear interpolation. The result
Self - Learning
0.78628 is obtained by quadratic as well as by cubic interpolation. We conclude Material 43
Interpolation and that there is no improvement in the third degree polynomial over that of the second
Curve Fitting
degree.
Notes 1. Unlike Lagrange’s methods, it is not necessary to find the degree of the
NOTES interpolating polynomial to be used.
2. The approximation by a higher degree interpolating polynomial may
not always lead to a better result. In fact it may be even worse in some
cases.
Consider, the function f(x) = 4.
We form the finite difference table with values for x = 0 to 4.

x f ( x ) f ( x) 2 f ( x) 3 f ( x) 4 f ( x)
0 1
3
1 4 9
12 27
2 16 36 81
48 108
3 64 144
192
4 256

Newton’s forward difference interpolating polynomial is given below by taking

x0  0,

x  x0 9 27 81
u  x,  ( x )  1  3x  x( x  1)  x( x  1)( x  2)  x( x  1)( x  2)( x  3)
h 2 6 24
Now, consider values of  ( x) at x = 0.5 by taking successively higher and
higher degree polynomials.
Thus,

1 (0.5)  1  0.5  3  2.5, by linear interpolation

0.5  ( 0.5)
 2 (0.5)  2.5   9  1.375, by quadratic interpolation
2
0.5  ( 0.5)  (1.5)
 3 (0.5)  1.375   27  3.0625, by cubic interpolation
6
(0.5)(0.5)(1.5)( 2.5)
 4 (0.5)  3.0625   81   0.10156, by quartic interpolation
24
We note that the actual value 40.5 = 2 is not obtainable by interpolation. The
results for higher degree interpolating polynomials become worse.
Note: Lagrange’s interpolation formula and iterative linear interpolation can easily
be implemented for computations by a digital computer.
Example 2.2: Determine the interpolating polynomial for the following table of
data:
x 1 2 3 4
Self - Learning y 1 1 1 5
44 Material
Solution: The data is equally spaced. We thus form the finite difference table. Interpolation and
Curve Fitting

x y y 2 y
1 1
0 NOTES
2 1 2
2
3 1 2
4
4 5

Since the differences of second order are constant, the interpolating polynomial
is of degree two. Using Newton’s forward difference interpolation, we get

u (u  1) 2
y  y0  u y0   y0 ,
2!
Here, x0  1, u  x  1.
( x  1)( x  2)
Thus, y  1  ( x  1)  0   2  x 2  3x  1.
2

Example 2.3: Compute the value of f(7.5) by using suitable interpolation on the
following table of data.

x 3 4 5 6 7 8
f ( x) 28 65 126 217 344 513

Solution: The data is equally spaced. Thus for computing f(7.5), we use Newton’s
backward difference interpolation. For this, we first form the finite difference table
as shown below.

x f ( x) f ( x) 2 f ( x) 3 f ( x)
3 28
37
4 65 24
61 6
5 126 30
91 6
6 217 36
127 6
7 344 42
169
8 513

The differences of order three are constant and hence we use Newton’s
backward difference interpolating polynomial of degree three.
v(v  1) 2 v(v  1)(v  2) 3
f ( x)  yn  v yn   yn   yn ,
2 ! 3 !

x  xn
v , for x  7.5, xn  8 Self - Learning
h Material 45
Interpolation and 7.5  8
Curve Fitting  v  0.5
1
( 0.5) ( 0.5  1) 0.5  0.5  1.5
f (7.5)  513  0.5  169   42  6
2 6
NOTES
 513  84.5  5.25  0.375
 422.875
Example 2.4: Determine the interpolating polynomial for the following data:

x 2 4 6 8 10
f ( x) 5 10 17 29 50

Solution: The data is equally spaced. We construct the Newton’s forward

difference interpolating polynomial. The finite difference table is,
x f ( x) f ( x) 2 f ( x) 3 f ( x) 4 f ( x)
2 5
5
4 10 2
7 3
6 17 5 1
12 4
8 29 9
21
10 50

Here, x0 = 2, u = (x–x0)/h = (x–2)/2.

The interpolating polynomial is,
u (u  1) 2
f ( x )  f ( x 0 )  u f ( x 0 )   f ( x0 )  ...
2!
x2 x  2  x  2  2 x  2  x  2  x  2 3
 5 5    1    1  2
2 2  2  2 ! 2  2  2  3!
x  2  x  2  x  2  x  2 1
   1  2   3
2  2  2  2  4!
1
 ( x 4  4 x 3  52 x 2  1040 x)
384
Example 2.5: Find the interpolating polynomial which takes the following values:
y(0) = 1, y(0.1) = 0.9975, y(0.2) = 0.9900, y(0.3) = 0.9980. Hence compute
y (0.05).
Solution: The data values of x are equally spaced we form the finite difference
table,
x y y 2 y 3 y
0.0 1.0000
 25
0.1 0.9975  50
 75 25
0.2 0.9900  25
 100
Self - Learning 0.3 0.9800
46 Material
 Interpolation and
x Curve Fitting
Here, h = 0.1. Choosing x0 = 0.0, we have s  10 x. Newton’s forward
0.1
difference interpolation formula is,
s ( s  1) 2 s ( s  1)(s  2) 3 NOTES
y  y 0  s y 0   y0   y0
2! 3!
10 x(10 x  1) 10 x(10 x  1)(10 x  2)
 1  10 x(0.0025)  (0.0050)   0.0025
2! 6
2 2.5 3 300 2 0.025
 1.0  0.25 x  0.25 x  0.25 x  x   0.0025x  x
6 4 6
2 3
 1.0  0.004 x  0.375 x  0.421x
y (0.05)  1.0002
Example 2.6: Compute f(0.23) and f(0.29) by using suitable interpolation formula
with the table of data given below.
x 0.20 0.22 0.24 0.26 0.28 0.30
f ( x) 1.6596 1.6698 1.6804 1.6912 1.7024 1.7139

Solution: The data being equally spaced, we use Newton’s forward difference
interpolation for computing f(0.23), and for computing f(0.29), we use Newton’s
backward difference interpolation. We first form the finite difference table,

x f ( x ) f ( x ) 2 f ( x)
0.20 1.6596
102
0.22 1.6698 4
106
0.24 1.6804 2
108
0.26 1.6912 4
112
0.28 1.7024 3
115
0.30 1.7139

We observe that differences of order higher than two would be irregular. Hence,
we use second degree interpolating polynomial. For computing f(0.23), we take
x  x0 0.23  0.22
x0 = 0.22 so that u    0.5.
h 0.02
Using Newton’s forward difference interpolation, we compute
(0.5)(0.5  1.0)
f (0.23)  1.6698  0.5  0.0106   0.0002
2
 1.6698  0.0053  0.000025
 1.675075
 1.6751
Again for computing f (0.29), we take xn = 0.30,
x  xn 0.29  0.30
so that v     0.5 Self - Learning
n 0.02 Material 47
Interpolation and Using Newton’s backward difference interpolation we evaluate,
Curve Fitting
( 0.5)(0.5  1.0)
f (0.29)  1.7139  0.5  .0115   0.0003
2
NOTES  1.7139  0.00575  0.00004
 1.70811
 1.7081
Example 2.7: Compute values of ex at x = 0.02 and at x = 0.38 using suitable
interpolation formula on the table of data given below.
x 0.0 0.1 0.2 0.3 0.4
e x 1.0000 1.1052 1.2214 1.3499 1.4918

Solution: The data is equally spaced. We have to use Newton’s forward difference
interpolation formula for computing ex at x = 0.02, and for computing ex at
x = 0.38, we have to use Newton’s backward difference interpolation formula.
We first form the finite difference table.
x y  ex y 2 y 3 y 4 y
0.0 1.0000
1052
0.1 1.1052 110
1162 13
0.2 1.2214 123 2
1285 11
0.3 1.3499 134
1419
0.4 1.4918

For computing e0.02 , we take x  0 0

x  x0 0.02  0.0
 u   0 .2
h 0.1
By Newton’s forward difference interpolation formula, we have

0.2(0.2  1) 0.2(0.2  1)(0.2  2)

e0.02  1.0  0.2  0.1052   0.0110   0.0013
2 6
0.2 (0.2  1)(0.2  2)(0.2  3)
  0.0002
24
 1.0  .02104  0.00088  0.00006  0.00001
 1.02023  1.0202

For computing e0.38 we take xn = 0.4. Thus, v  0.38  0.4  0.2

0.1
By Newton’s backward difference interpolation formula, we have

( 0.2)(0.2  1)
e0.38  1.4918  (0.2)  0.1419   0.0134
2
( 0.2)( 0.2  1)(0.2  2) 0.2( 0.2  1)(0.2  2)(0.2  3)
  0.0011   ( 0.0002)
6 24
 1.4918  0.02838  0.00107  0.00005  0.00001
 1.49287  0.02844
 1.46443  1.4644
Self - Learning
48 Material
2.2.2 Lagrange’s Interpolation Interpolation and
Curve Fitting
Lagrange’s interpolation is useful for unequally spaced tabulated values. Let y = f
(x) be a real valued function defined in an interval (a, b) and let y0, y1,..., yn be the
(n + 1) known values of y at x0, x1,...,xn, respectively. The polynomial (x), NOTES
which interpolates f (x), is of degree less than or equal to n. Thus,
 ( xi )  yi , for i  0,1, 2, ..., n
(2.7)
The polynomial  (x) is assumed to be of the form,
n
 ( x)   li ( x) yi
i0

(2.8)
where each li(x) is a polynomial of degree  n in x and is called Lagrangian
function.
Now,  (x ) satisfies Equation (2.7) if each li(x) satisfies,
li ( x j )  0 when i  j
1 when i  j
(2.9)
Equation (2.9) suggests that li(x) vanishes at the (n+1) points x0, x1, ... xi–1,
xi+1,..., xn. Thus, we can write,
li(x) = ci (x – x0) (x – x1) ... (x – xi–1) (x – xi+1)...(x – xn)
where ci is a constant given by li (xi) =1,
i.e., ci ( xi  x0 ) ( xi  x1 )...( xi  xi 1 ) ( xi  xi 1 )... ( xi  xn )  1
( x  x0 )( x  x1 )...( x  xi 1 )( x  xi 1 )...( x  xn )
Thus, li ( x)  for i  0, 1, 2, ..., n
( xi  x0 )( xi  x1 )...( xi  xi 1 )( xi  xi 1 )...( xi  xn )
(2.10)
Equations (2.8) and (2.10) together give Lagrange’s interpolating polynomial.
Algorithm: To compute f (x) by Lagrange’s interpolation.
Step 1: Read n [n being the number of values]
Step 2: Read values of xi, fi for i = 1, 2,..., n.
Step 3: Set sum = 0, i = 1
Step 4: Read x [x being the interpolating point]
Step 5: Set j = 1, product = 1
Step 6: Check if j  i, product = product × (x – xj)/(xi – xj) else go to Step
7
Step 7: Set j = j + 1
Step 8: Check if j > n, then go to Step 9 else go to Step 6
Step 9: Compute sum = sum + product × fi
Step 10: Set i = i + 1 Self - Learning
Material 49
Interpolation and Step 11: Check if i > n, then go to Step 12
Curve Fitting
else go to Step 5
Step 12: Write x, sum
NOTES Example 2.8: Compute f (0.4) for the table below by Lagrange’s interpolation.
x 0.3 0.5 0.6
f ( x) 0.61 0.69 0.72

Solution: The Lagrange’s interpolation formula gives,

(0.4  0.5)(0.4  0.6) (0.4  0.3)(0.4  0.6) (0.4  0.3)(0.4  0.5)
f (0.4)   0.61   0.69   0.72
(0.3  0.5)(0.3  0.6) (0.5  0.3)(0.5  0.6) (0.6  0.3)(0.6  0.5)
 0.203  0.69  0.24  0.653 ~  0.65

Thus, f (0.4) = 0.65.

Example 2.9: Using Lagrange’s formula, find the value of f (0) from the table
given below.
x 1  2 2 4
f ( x)  1  9 11 69

Solution: Using Lagrange’s interpolation formula, we find

 (0  2)(0  2)(0  4)   (0  1)(0  2)(0  4) 
f (0)    (1)    (9)
 (1  2) (1  2)(1  4)   (2  1)(2  2)(2  4) 
 (0  1) (0  2)(0  4)   (0  1)(0  2)(0  2) 
 11    69
 (2  1)(2  2)(2  4)   (4  1)(4  2)(4  2) 
16 9 11 69 20 85
     
15 3 3 15 3 15
20 17
  1
3 3
Example 2.10: Determine the interpolating polynomial of degree three for the
table given below.
x 1 0 1 2
f ( x) 1 1 1  3
Solution: We have Lagrange’s third degree interpolating polynomial as,
3
f ( x)   l ( x) f ( x )
i 0
i i

where
( x  0)( x  1)( x  2) 1
l0 ( x)    x( x  1)( x  2)
(1  0)(1  1)(1  2) 6
( x  1)( x  1)( x  2) 1
l1 ( x )   ( x  1)( x  1)( x  2)
(0  1)(0  1)(0  2) 2
( x  1)( x  0)( x  2) 1
l2 ( x)    ( x  1) x( x  2)
(1  1)(1  0)(1  2) 2
( x  1)( x  0)( x  1) 1
l3 ( x )   ( x  1) x ( x  2)
Self - Learning (2  1)(2  0)(2  1) 6
50 Material
 1 1 1 1 Interpolation and
f ( x)   x( x  1)( x  2)  1  ( x  1)( x  1)( x  2)  1  ( x  1) x ( x  2)  1  ( x  1) x( x  2)  (3) Curve Fitting
6 2 2 6
1 3
  (4 x  4 x  6)
6
1 NOTES
 (2 x 3  2 x  3)
3
Example 2.11: Evaluate the values of f (2) and f (6.3) using Lagrange’s interpolation
formula for the table of values given below.

x 1.2 2.5 4 5.1 6 6.5

f ( x) 6.84 14.25 27 39.21 51 58.25

Solution: It is not advisable to use a higher degree interpolating polynomial. For

evaluation of f (2) we take a second degree polynomial using the values of f (x) at
the points x0 = 1.2, x1 = 2.5 and x2 = 4.
Thus,
f (2) = l0(2) × 6.84 + l1(2) × 14.25 + l2(2) × 27
where
(2  2.5)(2  4)
l0 (2)   0.275
(1.2  2.5)(1.2  4)
(2  1.2)(2  4)
l1 (2)   0.821
(2.5  1.2)(2.5  4)
(2  1.2)(2  2.5)
l 2 (2)   0.095
(4  1.2)(4  2.5)
 f (2) = 0.275 × 6.84 + 0.821 × 14.25 – 0 .095 × 27 = 11.015 ~
 11.02
For evaluation of f (6.3), we consider the values of f (x) at x0 = 5.1, x1 = 6.0,
x2 = 6.5.
Thus, f (6.3) = l0(6.3) × 39.21 + l1(6.3) × 51 + l2(6.3) × 58.25
where
(6.3  6.0)(6.3  6.5)
l0 (6.3)   0.048
(5.1  6.0)(5.1  6.5)
(6.3  5.1)(6.3  6.5)
l1 (6.3)   0.533
(6  5.1)(6.0  6.5)
(6.3  5.1)(6.3  6.0)
l 2 (6.3)   0.514
(6.5  5.1)(6.5  6.0)

 f (6.3)   0.048  39.21  0.533  51  0.514  58.25

 55.241  55.24

Since, the computed result cannot be more accurate than the data, the final
result is rounded-off to the same number of decimals as the data. In some cases,
a higher degree interpolating polynomial may not lead to better results.
2.2.3 Finite Difference for Interpolation
For interpolation of an unknown function when the tabular values of the argument
x are equally spaced, we have two important interpolation formulae, viz.,
Self - Learning
Material 51
Interpolation and (i) Newton’s forward difference interpolation formula
Curve Fitting
(ii) Newton’s backward difference interpolation formula
We will first discuss the finite differences which are used in evaluating the
NOTES above two formulae.
Finite Differences
Let us assume that values of a function y = f (x) are known for a set of equally
spaced values of x given by {x0, x1,..., xn}, such that the spacing between any
two consecutive values is equal. Thus, x1 = x0 + h, x2 = x1 + h,..., xn = xn–1 + h,
so that xi = x0 + ih for i = 1, 2, ...,n. We consider two types of differences known
as forward differences and backward differences of various orders. These
differences can be tabulated in a finite difference table as explained in the subsequent
sections.
Forward Differences
Let y0, y1,..., yn be the values of a function y = f (x) at the equally spaced values of
x = x0, x1, ..., xn. The differences between two consecutive y given by y1 – y0, y2
– y1,..., yn – yn–1 are called the first order forward differences of the function y = f
(x) at the points x0, x1,..., xn–1. These differences are denoted by,
y0  y1  y0 , y1  y2  y1 , ..., yn1  yn  yn1
(2.11)
where  is termed as the forward difference operator defined by,,
f ( x )  f ( x  h )  f ( x )
(2.12)
Thus,  yi = yi+1 – yi, for i = 0, 1, 2, ..., n – 1, are the first order forward
differences at xi.
The differences of these first order forward differences are called the second
order forward differences.
Thus, 2
 yi   (yi )
 yi 1  yi , for i  0, 1, 2, ..., n  2
(2.13)
Evidently,
2 y0  y1  y0  y 2  y1  ( y1  y0 )  y2  2 y1  y0

And, 2 yi  yi  2  yi 1  ( yi 1  yi )

i.e.,  2 yi  yi  2  2 yi 1  yi , for i  0, 1, 2, ..., n  2

(2.14)
Similarly, the third order forward differences are given by,
 3 yi   2 yi 1   2 yi , for i  0, 1, 2, ..., n  3

i.e.,  3 y i  y i  3  3 y i  2  3 y i 1  y i
(2.15)
Self - Learning
52 Material
 Finally, we can define the nth order forward difference by, Interpolation and
Curve Fitting
n( n  1)
n y 0  y n  ny n 1  y n  2  ...  ( 1) n y0
2!
(2.16) NOTES
The coefficients in above equations are the coefficients of the binomial expansion
(1 – x)n.
The forward differences of various orders for a table of values of a function
y = f (x), are usually computed and represented in a diagonal difference table. A
diagonal difference table for a table of values of y = f (x), for six points x0, x1, x2,
x3, x4, x5 is shown here.
Diagonal difference Table for y = f(x):
i xi yi yi 2 yi 3 yi 4 yi 5 yi
0 x0 y0
y 0
1 x1 y1 2 y0
y1 3 y0
2
2 x2 y2  y1 4 y 0
y 2 3 y1 5 y0
2 4
3 x3 y3  y2  y1
y 3 3 y 2
4 x4 y4 2 y 3
y 4
5 x5 y5

The entries in any column of the differences are computed as the differences
of the entries of the previous column and one placed in between them. The upper
data in a column is subtracted from the lower data to compute the forward
differences. We notice that the forward differences of various orders with respect
to yi are along the forward diagonal through it. Thus y0, 2y0, 3y0, 4y0 and
5y0 lie along the top forward diagonal through y0. Consider the following example.
Example 2.12: Given the table of values of y = f (x),
x 1 3 5 7 9
y 8 12 21 36 62

form the diagonal difference table and find the values of f (5), 2 f (3), 3 f (1) .
Solution: The diagonal difference table is,
i xi yi y i 2 yi 3 yi 4 yi
0 1 8
4
1 3 12 5
9 1
2 5 21 6 4
15 5
3 7 36 11
26 Self - Learning
4 9 62 Material 53
Interpolation and
Curve Fitting
From the table, we find that f (5)  15, the entry along the diagonal through
the entry 21 of f (5).
Similarly, 2 f (3)  6, the entry along the diagonal through f (3). Finally,,
NOTES
3 f (1)  1.

Backward Differences
The backward differences of various orders for a table of values of a function y =
f (x) are defined in a manner similar to the forward differences. The backward
difference operator  (inverted triangle) is defined by f ( x)  f ( x)  f ( x  h).
Thus, yk  yk  yk 1 , for k  1, 2, ..., n
i.e., y1  y1  y0 , y 2  y 2  y1 ,..., y n  y n  y n 1
(2.17)
The backward differences of second order are defined by,

 2 yk  yk  yk 1  yk  2 yk 1  yk  2
Hence,
 2 y2  y2  2 y1  y0 , and  2 yn  yn  2 yn 1  yn 2
(2.18)
Higher order backward differences can be defined in a similar manner.
Thus,  3 yn  yn  3 yn 1  3 yn 2  yn 3 , etc.
(2.19)
Finally,
n( n  1)
 n yn  yn  nyn 1  yn 2 – ...  ( 1) n y0
2 i
(2.20)
The backward differences of various orders can be computed and placed in a
diagonal difference table. The backward differences at a point are then found
along the backward diagonal through the point. The following table shows the
backward differences entries.
Diagonal difference Table of backward differences:

i xi yi y i  2 yi  3 yi  4 yi  5 yi
0 x0 y0
y1
1 x1 y1  2 y2
y 2  3 y3
2
2 x2 y2  y3  4 y4
y 3  3 y4
2
3 x3 y3  y4  4 y5
3
y 4  y5
4 x4 y4  2 y5
 y5
Self - Learning 5 x5 y5
54 Material
 The entries along a column in the table are computed (as discussed in previous Interpolation and
Curve Fitting
example) as the differences of the entries in the previous column and are placed in
between. We notice that the backward differences of various orders with respect
to yi are along the backward diagonal through it. Thus, y5 ,  2 y5 ,  3 y5 ,  4 y5 and
NOTES
 5 y 5 are along the lowest backward diagonal through y5.

We may note that the data entries of the backward difference table in any
column are the same as those of the forward difference table, but the differences
are for different reference points.
Specifically, if we compare the columns of first order differences we can see
that,
y0  y1 , y1  y 2 , ..., y n 1  y n

Hence, yi  yi 1 , for i  0, 1, 2, ..., n  1

Similarly, 2 y 0   2 y 2 , 2 y1   2 y3 ,..., 2 y n  2   2 y n

Thus,  2 yi  2 yi  2 , for i  1, 2, ..., n  2

In general, k yi   k yi  k .

Conversely,  k yi  k yi  k .
Example 2.13: Given the following table of values of y = f (x):

x 1 3 5 7 9
y 8 12 21 36 62

Find the values of  y (7) ,  2 y (9) ,  3 y (9) .

Solution: We form the diagonal difference table,

xi yi y i  2 yi  3 yi  4 yi
1 8
4
3 12 5
9 1
5 21 6 4
15 5
7 36 11
26
9 62

From the table, we can easily find y(7) 15,  y(9) 11,  y(9)  5.
2 3

2.2.4 Symbolic Operators

We consider the finite differences of an equally spaced tabular data for developing
numerical methods. Let a function y = f (x) has a set of values y0, y1, y2,..., Self - Learning
Material 55
Interpolation and corresponding to points x0, x1, x2,..., where x1 = x0 + h, x2 = x0 + 2h,...., are
Curve Fitting
equally spaced with spacing h. We define different types of finite differences such
as forward differences, backward differences and central differences, and express
them in terms of operators.
NOTES
The forward difference of a function f (x) is defined by the operator , called
the forward difference operator given by,
f ( x )  f ( x  h )  f ( x )
(2.21)
At a tabulated point xi , we have
f ( xi )  f ( xi  h)  f ( xi )
(2.22)
We also denote f ( xi ) by yi , given by
yi  yi 1  yi , for i  0, 1, 2, ...
(2.23)
We also define an operator E, called the shift operator which is given by,
E f(x) = f(x + h)
(2.24)
 f ( x)  Ef ( x)  f ( x )
Thus,   E  1 is an operator relation. (2.25)
While Equation (2.21) defines the first order forward difference, we can define
second order forward difference by,
 2 yi   (yi )   ( yi 1  yi )

  2 yi  yi 1  yi (2.26)

2.2.5 Shift Operator

The shift operator is denoted by E and is defined by E f (x) = f (x + h). Thus,
Eyk = yk+1
Higher order shift operators can be defined by, E2f (x) = Ef (x + h) = f (x +
2h).
E2yk = E(Eyk) = E(yk + 1) = yk + 2
In general, Emf (x) = f (x + mh)
Emyk = yk+m

Relation between forward difference operator and shift operator

From the definition of forward difference operator, we have

y ( x )  y ( x  h )  y ( x )
 Ey ( x)  y ( x)
Self - Learning  ( E  1) y ( x)
56 Material
This leads to the operator relation, Interpolation and
Curve Fitting
  E 1
or, E  1  (2.27)
NOTES
Similarly, for the second order forward difference, we have

 2 y ( x)  y ( x  h)  y ( x)
 y ( x  2 h)  2 y ( x  h)  y ( x )
 E 2 y ( x )  2 Ey ( x )  y ( x )
 ( E 2  2 E  1) y ( x )

This gives the operator relation, 2  ( E  1) 2 .

Finally, we have  m  ( E  1)m , for m  1, 2, ... (2.28)

Relation between the backward difference operator with shift operator
From the definition of backward difference operator, we have
 f ( x )  f ( x )  f ( x  h)
 f ( x)  E 1 f ( x)  (1  E 1 ) f ( x)

This leads to the operator relation,   1  E 1 (2.29)

Similarly, the second order backward difference is defined by,

 2 f ( x )  f ( x )  f ( x  h )
 f ( x )  f ( x  h)  f ( x  h)  f ( x  2h)
 f ( x )  2 f ( x  h)  f ( x  2h)
 f ( x)  E 1 f ( x)  E  2 f ( x)
 (1  E 1  E  2 ) f ( x)
 (1  E 1 ) 2 f ( x)

This gives the operator relation,  2  (1  E 1 ) 2 and in general,

 m  (1  E 1 ) m (2.30)
Relations between the operators E, D and 
We have by Taylor’s theorem,

h2
f ( x  h)  f ( x )  hf  ( x)  f  ( x)  ...
2!
h2 D2 d
Thus, Ef ( x)  f ( x)  hDf ( x)  f ( x)  ..., where D 
2! dx
 h2 D2 
or, (1   ) f ( x )  1  hD   ...  f ( x)
 2! 
Self - Learning
 e f ( x)
hD
Material 57
Interpolation and
Curve Fitting
Thus, e hD  1    E
(2.31)
Also, hD  log (1   )

NOTES 2 3 4
or, hD       ...
2 3 4
1   2 3  4 
 D  
h  2

3

4
 ... 


2.2.6 Central Difference Operator

The central difference operator denoted by  is defined by,
 h  h
y ( x )  y  x    y  x  
 2  2
Thus,
y ( x)  ( E1/ 2  E 1/ 2 ) y ( x)
Giving the operator relation,   E1/ 2  E 1/ 2 or  E1/ 2  E  1
Also,
 yn  ( E1/ 2  E 1/ 2 ) y ( xn )  E1/ 2 yn  E 1/ 2 yn
i.e.,  yn  yn 1/ 2  yn 1/ 2
Further,
 2 yn   ( yn )   yn 1/ 2   yn 1/ 2


 E1/ 2  E 1/ 2   y   E
n 1/ 2
1/ 2
 E 1/ 2  y 
n 1/ 2

E 1/ 2
 yn1/ 2  yn1/ 2   E 1/ 2
 yn 1/ 2  yn1/ 2 
 yn 1  yn  yn  yn 1
 yn 1  2 yn  yn 1  ( E1/ 2  E 1/ 2 ) 2 yn    2 yn 1   2 yn 1 
 ( E  E 1  2) yn

  2  E  E 1  2 (2.32)
Even though the central difference operator uses fractional arguments, still it is
widely used. This is related to the averaging operator and is defined by,
1 1/ 2
 ( E  E 1/ 2 ) (2.33)
2
1 1 1
Squaring,  2  ( E  2  E 1 )  ( 2  2  2)  1   2
4 4 4
1
 2  1  2 (2.34)
4
It may be noted that,  y1/ 2  y1  y0  y1
Self - Learning
58 Material
 Also,  E1/ 2 y1   y 1  y2  y1  y1 Interpolation and
1 Curve Fitting
2

  E1/ 2    E  1 (2.35)
Further, NOTES

 
 3 yn  ( 2 yn )    y 1   y 1 
 n 2 n 
2

 2 y 1  2 y 1   ( yn 1  2 yn  yn 1 )
n n
2 2

Example 2.14: Prove the following operator relations:

(i)   E (ii) (1   ) (1  )  1
Solution:
(i) Since, f (x)  f (x  h)  f (x)  Ef (x)  f (x) ,   E  1 (1)

and since f ( x)  f ( x)  f ( x  h)  (1  E 1 ) f ( x) ,   1  E 1 (2)

E 1
Thus,   or E  E  1  
E
Hence proved.
(ii) From Equation (1), we have E    1 (3)
and from Equation (2) we get E 1  1   (4)
Combining Equations (3) and (4), we get (1   )(1  )  1.
Example 2.15: If fi is the value of f (x) at xi where xi = x0 + ih, for i = 1,2,...,
prove that,
i
i
fi  E i fo     i f0
j 0  j 

Solution: We can write Ef (x) = f (x + h)

Using Taylor series expansion, we have

h2
Ef ( x)  f ( x)  hf ( x)  f ( x)  ...
2!
h2 d
 f ( x)  hDf ( x)  D2 f ( x)  ..., where D 
2! dx
 hD2 2

 (1  ) f ( x)  1  hD   ... f ( x), since E  1  
 2! 

 ehD . f ( x)

 1    ehD

Hence, eihD  (1  )i

Self - Learning
Material 59
Interpolation and
Curve Fitting Now, f i  f ( xi )  f ( x0  ih)  E i f ( x0 )

 f i  (1   )i f ( x0 ), since E  1  

NOTES i
 i
f i      i f 0 , using binomial expansion.
j 0  j

Hence proved.
Example 2.16: Compute the following differences:
(i) n e x (ii) n x n
Solution:
(i) We have,  e x  e x  h  e x  e x (e h  1)

Again, 2 e x  (e x )  (e h  1)e x  (e h  1) 2 e x

Thus by induction, n e x  (e h  1) n e x .
(ii) We have,

 ( x n )  ( x  h) n  x n
n(n  1) 2 n  2
 n h x n 1  h x  ....  h n
2!

Thus, ( x n ) is a polynomial of degree (n – 1)

Also, (h n )  0. Hence, we can say that 2 ( x n ) is a polynomial of degree

(n – 2) with the leading term n(n  1)h 2 x n 2 .
Proceeding n times, we get
 n ( x n )  n(n  1)... 1h n  n!h n

Example 2.17: Prove that,

 f ( x )  g ( x ) f ( x )  f ( x ) g ( x )
(i)   
 g ( x)  g ( x ) g ( x  h)

  f ( x) 
(ii) {log f ( x)}  log 1  
 f ( x) 

Solution:
(i) We have,
 f ( x )  f ( x  h) f ( x )
  
 g ( x )  g ( x  h) g ( x )
f ( x  h) g ( x )  f ( x ) g ( x  h)

g ( x  h) g ( x )

Self - Learning
60 Material
 f ( x  h) g ( x )  f ( x ) g ( x )  f ( x ) g ( x )  f ( x ) g ( x  h) Interpolation and
 Curve Fitting
g ( x  h) g ( x )
g ( x){ f ( x  h)  f ( x)}  f ( x){g ( x  h)  g ( x)}

g ( x ) g ( x  h)
g ( x) f ( x)  f ( x) g ( x)
NOTES

g ( x ) g ( x  h)
(ii) We have,
 {log f ( x)}  log{ f ( x  h)}  log{ f ( x)}
f ( x  h)  f ( x  h)  f ( x )  f ( x ) 
 log  log  
f ( x)  f ( x) 
 f ( x ) 
 log   1
 f ( x) 

2.2.7 Differences of a Polynomial

We now look at the differences of various orders of a polynomial of degree n,
given by
y  f ( x)  an x n  an1 x n 1  an  2 x n 2  ...  a1 x  a0
The first order forward difference is defined by,
f ( x)  f ( x  h)  f ( x) and is given by,,

y  an {( x  h) n  x n }  an 1{( x  h) n 1  x n 1}  ...  a1 ( x  h  x)

n(n  1) 2 n  2
 an {n h x n 1  h x  ...}  an 1{( n  1) h x n  2  ...}
2 !
 bn 1 x n 1  bn  2 x n  2  ...  b1 x  b0

where the coefficients of various powers of x are collected separately.

Thus, the first order difference of a polynomial of degree n is a polynomial of
degree n – 1, with bn 1  an . nh
Proceeding as above, we can state that the second order forward difference
of a polynomial of degree n is a polynomial of degree n – 2, with coefficient of xn
–2
as n(n  1)h 2 a0 .
Continuing successively, we finally get  n y  a0 n!h n , a constant.
We can conclude that for polynomial of degree n, all other differences having
order higher than n are zero.
It may be noted that the converse of the above result is partially true and
suggests that if the tabulated values of a function are found to be such that the
differences of the kth order are approximately constant, then the highest degree of
the interpolating polynomial that should be used is k. Since the tabulated data may
have round-off errors, the acutal function may not be a polynomial.

Self - Learning
Material 61
Interpolation and Example 2.18: Compute the horizontal difference table for the following data
Curve Fitting
and hence, write down the values of f (4),  2 f (3) and  3 f (5).

x 1 2 3 4 5
NOTES f ( x) 3 18 83 258 627

Solution: The horizontal difference table for the given data is as follows:
x f ( x ) f ( x )  2 f ( x )  3 f ( x )  4 f ( x )
1 3    
2 18 15   
3 83 65 50  
4 258 175 110 60 
5 627 369 194 84 24

From the table we read the required values and get the following result:
f (4)  175,  2 f (3)  50,  3 f (5)  84

Example 2.19: Form the difference table of f (x) on the basis of the following
table and show that the third differences are constant. Hence, conclude about the
degree of the interpolating polynomial.

x 0 1 2 3 4
f ( x) 5 6 13 32 69

Solution: The difference table is given below

x f ( x) f ( x) 2 f ( x) 3 f ( x)
0 5
1
1 6 6
7 6
2 13 12
19 6
3 32 18
37
4 69

It is clear from the above table that the third differences are constant and
hence, the degree of the interpolating polynomial is three.

2.2.8 Newton’s Forward Difference Interpolation

Formula
Newton’s forward difference interpolation formula is a polynomial of degree less
than or equal to n. This is used to find the value of the tabulated function at a non-
tabular point. Consider a function y = f (x) whose values y0, y1,..., yn at a set of
Self - Learning
equidistant points x0 , x1 ,..., xn are known.
62 Material
 Interpolation and
Let  ( x ) be the interpolating polynomial, such that Curve Fitting
 ( xi )  f ( xi )  yi
xi  x0  ih, for i  0, 1, 2, ..., n (2.36)
NOTES
We assume the polynomial  ( x ) to be of the form,
 ( x)  a0  a1 ( x  x0 )  a2 ( x  x0 )( x  x1 )  a3 ( x  x0 )( x  x1 )( x  x2 ) 
...  an ( x  x0 )( x  x1 )...( x  xn 1 ) (2.37)
The coefficients ai’s in Equation (2.37) are determined by satisfying the
conditions in Equation (2.36) successively for i = 0, 1, 2,...,n.
Thus, we get
y0   ( x0 )  a0 , gives a0  y0
y1  y0
y1   ( x1 )  a0  a1 ( x1  x0 ), gives a1 
h
y0
 a1 
h
y2   ( x2 )  a0  a1 ( x2  x0 )  a2 ( x2  x0 )( x2  x1 )
or,
y0
y2  y0   a2 (2h) h
h
y  2 y1  y0  2 y0
a2  2 
2h 2 2 ! h2
Proceeding further, we get successively,
 3 y0  n y0
a3  ,..., an 
3 ! h3 n ! hn
Using these values of the coefficients, we get Newton’s forward difference
interpolation in the form,
x  x0 ( x  x0 )( x  x1 ) 2 ( x  x0 ) ( x  x1 ) ( x  x2 )  3 y0
 ( x)  y0   y0   y0   ...
h 2 ! h2 h h h 3!
( x  x0 ) ( x  x1 ) ( x  xn 1 )  n y0
...  ...
h h h n!
x  x0
This formula can be expressed in a more convenient form by taking u  as
h
shown here.
We have,
x  x1 x  ( x0  h) x  x0
  1  u 1
h h h
x  x2 x  ( x0  2h) x  x0
  2 u2
h h h
x  xn 1 x  {x0  (n  1)h} x  x0
   (n  1)  u  n  1 Self - Learning
h h h
Material 63
Interpolation and Thus, the interpolating polynomial reduces to:
Curve Fitting
u (u  1) 2 u (u  1)(u  2) 3
 (u )  y0  u y0   y0   y0
2 ! 3 !
NOTES
u (u  1)(u  2)...(u  n  1) n
 ...   y0 (2.38)
n !
This formula is generally used for interpolating near the beginning of the table.
For a given x, we choose a tabulated point as x0 for which the following condition
is satisfied.
For better results, we should have
x  x0
| u |  0.5
h

The degree of the interpolating polynomial to be used is less than or equal to

n and is determined by the order of the differences when they are nearly same so
that the differences of higher orders are irregular due to the propagated round-off
error in the data.

2.2.9 Newton’s Backward Difference Interpolation

Formula
Newton’s forward difference interpolation formula cannot be used for interpolating
at a point near the end of a table, since we do not have the required forward
differences for interpolating at such points. However, we can use a separate formula
known as Newton’s backward difference interpolation formula. Let a table of
values {xi, yi}, for i = 0, 1, 2, ..., n for equally spaced values of xi be given. Thus,
xi = x0 + ih, yi = f(xi), for i = 0, 1, 2, ..., n are known.
We construct an interpolating polynomial of degree n of the form,
y   ( x)  b0  b1 ( x  xn )  b2 ( x  xn )( x  xn 1 )  ...  bn ( x  xn )( x  xn 1 )...( x  x1 )
(2.39)
We have to determine the coefficients b0, b1, ..., bn by satisfying the relations,
 (xi )  yi , for i  n, n  1, n  2, ..., 1, 0
(2.40)
Thus,  (xn )  yn , gives b0  yn
(2.41)
Similarly,  (xn 1 )  yn 1 , gives yn 1  b0  b1 ( xn 1  xn )

yn  yn 1 yn
or, b1   (2.42)
h h
Again
 ( xn  2 )  yn  2 , gives yn  2  b0  b1 ( xn  2  xn )  b2 ( xn  2  xn )( xn 1  xn )

yn  yn 1
or, yn  2  yn  ( 2h)  b2 (2h)(h)
Self - Learning h
64 Material
 Interpolation and
y  2 yn 1  yn  2  2 yn
 b2  n  (2.43) Curve Fitting
2h 2 2 ! h2
By induction or by proceeding as mentioned earlier, we have
NOTES
3 yn  4 yn  n yn
b3  , b4  , ..., bn  (2.44)
3 ! h3 4 ! h4 n ! hn
Substituting the expressions for bi in Equation (2.39), we get
yn  2 yn  n yn
 ( x )  yn  ( x  xn )  ( x  x n ) ( x  xn 1 )  ...  ( x  xn )( x  xn 1 )...( x  x1 )
h 2 ! h2 n ! hn
(2.45)
This formula is known as Newton’s backward difference interpolation formula.
It uses the backward differences along the backward diagonal in the difference
table.
x  xn
Introducing a new variable v  ,
h
x  xn 1 x  ( xn  h)
we have,   v 1.
h h
x  xn  2 x  x1
Similarly,  v  2,...,  v  n  1.
h h
Thus, the interpolating polynomial in Equation (2.45) may be rewritten as,

v(v  1) 2 v(v  1)(v  2) 3 v (v  1)(v  2)...(v  n  1) n

 ( x)  yn  vyn   yn   yn  ...   yn
2 ! 3 ! n !

(2.46)
This formula is generally used for interpolation at a point near the end of a
table.
The error in the given interpolation formula may be written as,
E ( x)  f ( x)  ( x)
( n  1)
( x  x n )( x  x n 1 )...( x  x1 )( x  x 0 ) f ( )
 , where x0    x n
( n  1) !
y ( n  1) (  ) n 1
 v ( v  1)( v  2)...( v  n ) h
( n  1) !

2.2.10 Extrapolation
The interpolating polynomials are usually used for finding values of the tabulated
function y = f(x) for a value of x within the table. But, they can also be used in
some cases for finding values of f(x) for values of x near to the end points x0 or xn
outside the interval [x0, xn]. This process of finding values of f(x) at points beyond
the interval is termed as extrapolation. We can use Newton’s forward difference
interpolation for points near the beginning value x0. Similarly, for points near the
end value xn, we use Newton’s backward difference interpolation formula.
Example 2.20: With the help of appropriate interpolation formula, find from the
Self - Learning
following data the weight of a baby at the age of one year and of ten years: Material 65
Interpolation and
Curve Fitting Age  x 3 5 7 9
Weight  y (kg ) 5 8 12 17

NOTES Solution: Since the values of x are equidistant, we form the finite difference table
for using Newton’s forward difference interpolation formula to compute weight of
the baby at the age of required years.
x y y 2 y
3 5
3
5 8 1
4
7 12 1
5
9 17

x  x0
Taking x = 2, u   0.5.
h
Newton’s forward difference interpolation gives,
(0.5)(1.5)
y at x  1, y (1)  5  0.5  3  1
2
 5  1.5  0.38  3.88  3.9 kg.
Similarly, for computing weight of the baby at the age of ten years, we use
Newton’s backward difference interpolation given by,
x  xn 10  9
v   0.5
h 2
0.5  1.5
y at x  10, y (10)  17  0.5  5  1
2
 17  2.5  0.38  19.88

2.2.11 Inverse Interpolation

The problem of inverse interpolation in a table of values of y = f (x) is to find the
value of x for a given y. We know that the inverse function x = g (y) exists and is

unique, if y = f (x) is a single valued function of x and dy exists and does not
dx
vanish in the neighbourhood of the point where inverse interpolation is desired.
When the values of x are unequally spaced, we can apply Lagrange’s
interpolation or iterative linear interpolation simply by interchanging the roles of x
and y. Thus Lagrange’s formula for inverse interpolation can be written as,
n
x   li ( y) xi
i 0
n
where li ( y )   [( y  y j ) /( yi  y j )]
j 0
j i

Self - Learning When x values are equally spaced, we can apply the method of successive
66 Material approximation as described below.
 Consider Newton’s formula for forward difference interpolation given by, Interpolation and
Curve Fitting
u (u  1) 2 u (u  1)(u  2) 3
y  y 0  u y 0   y0   y0  ...
2! 3!

Retaining only two terms on the RHS, we can write the first approximation, NOTES

1
u (1)  ( y  y0 )
 y0

The second approximation can be written as,

1  u (1) (u (1)  1) 2 
u ( 2)  ( y  y 0 )   y0 
y 0  2 

on replacing u by u(1) in the coefficient of 2 y0 .

Similarly, the third approximation can be written as,

1  u ( 2) (u ( 2)  1) 2 u ( 2) (u ( 2)  1)(u ( 2)  2) 3 
u (3)   y  y 0   y 0   y0 
yo  2 6 

The process can be continued until two successive approximations have a

reasonable accuracy. Then x is obtained by the relation,
x = x0 + uh
Example 2.21: Using inverse interpolation, find the value of x for y = 5, from the
given table.

x 1 3 4
y 3 12 19

Solution: Applying inverse interpolation,

2
x  l ( y). x
i 0
i i

Thus, for y = 5, we have

(5  12)(5  19) (5  3)(5  19) (5  3)(5  12)
x 1  3 4
(3  12)(3  19) (12  3)(12  19) (19  3)(19  12)
7  14 2  14 2  7
  3 4
9  16 9  7 16  7
 0.6805  1.3333  0.5000
 1.5138
 1.514 correct upto four significant figures.
Example 2.22: Given the following tabular values of cosh x, find x for which
cosh x = 1.285.

x 0.738 0.739 0.740 0.741 0.742

cos h x 1.2849085 1.2857159 1.2865247 1.2873348 1.2881461
Self - Learning
Material 67
Interpolation and Solution: Since finding x for an equally spaced table of cosh x is a problem of
Curve Fitting
inverse interpolation, we employ the method of successive approximation using
Newton’s formula of inverse interpolation. We first form the finite difference table.
x f ( x)  cos h x f ( x)  2 f ( x)  3 f ( x)
NOTES
0.738 1.2849085
8074
0.739 1.2857159 14
8088 1
0.740 1.2865247 13
8101 1
0.741 1.2873348 12
8113
0.742 1.2881461

Using Newton’s forward difference interpolation formula for the first

( x  x0 )
approximation u  we get,
h
1
u (1)  ( y  y0 )
f ( x0 )
For, y  1.285, we take x0  0.739.

1
u (1)   (1.285  1.2857159)  0.8851384, then x  0.739885
0.0008088
For a second approximation,
1 u (1) (u (1)  1) 2
u (2)  u (1)   f ( x0 )
f ( x0 ) 2
1
 0.8851384   ( 0.8851384)  (1.8851384)  0.0000013
0.0008088  2
 0.8851384  0.0013409  0.8864793  x  0.7398864
Similarly,
u (2) (u (2)  1)  2 f 0 1 (2) (2) 3 f 0
u (3)  u (1)   u (u  1)(u (2)  2)
2 f 0 6 f 0
 0.8851384  0.0013430  0.000073600
 0.886555  x  0.7398865
Example 2.23: Find the divided difference interpolation for the following table of
values:
x 4 7 9
f ( x)  43 83 327

Solution: We first form the Divided Difference (DD) table as given below.
x f ( x) 1st DD 2nd DD
4  43
42
7 83 16
122
Self - Learning
68 Material 9 327
 Newton’s divided difference interpolation formula is, Interpolation and
Curve Fitting

f ( x)  f ( x0 )  ( x  x0 ) f ( x0 , x1 )  ( x  x0 ) ( x  x1 ) f  x0 , x1 , x2 
 f ( x)  43  ( x  4) 42  ( x  4) ( x  7)  16
 16 x 2  134 x  237
NOTES

Example 2.24: Given the following table of values of the function

y  log e x, construct the Newton’s forward difference interpolating polynomial.
Comment on the degree of the polynomial and find loge1001.

x 1000 1010 1020 1030 1040

log e x 3.00000 3.00432 3.00860 3.01284 3.01703

Solution: We form the difference table as given below:

x y y 2 y
1000 3.00000
432
1010 3.00432 4
428
1020 3.00860 4
424
1030 3.01284 5
419
1040 3.01703

We observe that, the differences of second order are nearly constant. Thus,
the degree of the interpolating polynomial is 2 and is given by,
u (u  1) 2 x  x0
y  y 0  u y 0   y0 , where u 
2 h
For x = 1001, we take x0 = 1000.

1001  1000
u   0.1
10
0.1  0.9
log e 1001  3.00000  0.1  0.00432   (0.00004)
2
 3.000430  3.00043

Example 2.25: Determine the interpolating polynomial for the following data table
using both forward and backward difference interpolating formulae. Comment on
the result.

x 0 1 2 3 4
f ( x) 1.0 8.5 36.0 95.5 199.0

Self - Learning
Material 69
Interpolation and Solution: Since the data points are equally spaced, we construct the Newton’s
Curve Fitting
forward difference interpolating polynomial for which we first form the finite
difference table as given below:

NOTES x f ( x) f ( x) 2 f ( x) 3 f ( x)
0 1 .0
7.5
1.0 8 .5 20.0
27.5 12.0
2 .0 36.0 32.0
59.5 12.0
3 .0 95.5 44.0
103.5
4.0 199.0

Since the differences of order 3 are constant, we construct the third degree
Newton’s forward difference interpolating polynomial given by,
x ( x  1) x ( x  1) ( x  2)
f ( x)  1.0  x  0.75   20   12
2 6

Since x0  0, h  1.0

x  x0
 u x
h

i.e., f ( x)  1.0  1.5x  4 x 2  2 x3 , on simplification.

Taking x n  4, we also construct the backward difference interpolating

polynomial given by,

( x  4) ( x  3)
f ( x)  199  ( x  4)  103.5   44
2
( x  4) ( x  3) ( x  3)
  12
6
 1.0  1.5 x  4 x 2  2 x 3 , on simplification.

This is the same as the forward difference interpolating polynomial, because

the difference of third order is constant.
Example 2.26: Use Newton’s divided difference interpolation to evaluate f(18)
and f (12) for the following data:

x 4 5 7 10 11 13
f ( x) 48 100 294 900 1210 2028

Self - Learning
70 Material
Solution: We first form the divided difference table as given below. Interpolation and
Curve Fitting

x f ( x) 1st DD 2nd DD 3rd DD

4 48
52 NOTES
5 100 15
97 1
7 294 21
202 1
10 900 27
310 1
11 1210 33
409
13 2028

Since 3rd order divided differences are same, higher order divided differences
vanish. We have the Newton’s divided difference interpolation given by,

f ( x )  f 0  ( x  x0 ) f  x , x1   ( x  x0 )( x  x1 ) f  x0 , x1 , x2 
 ( x  x0 )( x  x1 )( x  x 2 ) f  x0 , x1 , x2 , x3 

For x = 8, we take x0 = 4,
f (8)  48  (8  4)52  (8  4)(8  5)  15  (8  4)(8  5)(8  7)  1
 48  208  180  12  448

For x = 12, we take x0 = 13,

f (12)  2028  (12  13)  409  (12  13) (12  11)  33
 (12  13) (12  11) (12  10)  1
 f (12)  2028  409  33  2  1584

Example 2.27: Using inverse interpolation, find the zero of f (x) given by the
following tabular values.

x 0.3 0.4 0.6 0.7

y  f ( x) 0.14 0.06  0.04  0.06

Solution: Using Lagrange’s form of inverse interpolation, we calculate the formula

using y = 0. 14, 0.06, – 0.04 and – 0.06, as given below:
( y  0.06)( y  0.04)( y  0.06)
P3 ( y )   0.3
(0.14  0.06)(0.14  0.04)(0.14  0.06)
( y  0.14)( y  0.04)( y  0.06)
  0.4
(0.06  0.14)(0.06  0.04)(0.06  0.06)
( y  0.14)( y  0.06)( y  0.06)
  0.6
(0.04  0.14)(0.04  0.06)( 0.04  0.06)
( y  0.14)( y  0.06)( y  0.04)
  0.7
(0.06  0.14)(0.06  0.06)(0.06  0.04)

Self - Learning
Material 71
Interpolation and 0.06  0.04  0.06  0.3 0.14  0.04  0.06  0.4
Curve Fitting Thus, P3 (0)  
0.08  0.18  0.20 0.08  0.1  0.12
0.14  0.06  0.06  0.6 0.14  0.06  0.04  0.7
 
0.18  0.1  0.02 0.2  0.12  0.02
NOTES  0.015  0.14  0.84  0.49  0.475
Thus, the zero of f (x) is 0.475 which is approximately equal to 0.48, since the
accuracy depends on the accuracy of the data which is the significant digits.

2.2.12 Truncation Error in Interpolation

Refer Unit 1 (Level Head 1.4)

Check Your Progress

1. What do we generate in iterative linear interpolation?
2. Define interpolation.
3. How is Lagrange's interpolation useful?
4. Which interpolation will you use for equally spaced tabular values?
5. Define the shift operator.
6. What is the Newton forward difference interpolation formula used?
7. Define extrapolation.
8. Define the problem of inverse interpolation.

2.3 CURVE FITTING

In this section, we consider the problem of approximating an unknown function
whose values, at a set of points, are generally known only empirically and are, thus
subject to inherent errors, which may sometimes be appreciably larger in many
engineering and scientific problems. In these cases, it is required to derive a
functional relationship using certain experimentally observed data. Here the
observed data may have inherent or round-off errors, which are serious, making
polynomial interpolation for approximating the function inappropriate. In polynomial
interpolation the truncation error in the approximation is considered to be important.
But when the data contains round-off errors or inherent errors, interpolation is not
appropriate.
The subject of this section is curve fitting by least square approximation. Here
we consider a technique by which noisy function values are used to generate a
smooth approximation. This smooth approximation can then be used to approximate
the derivative more accurately than with exact polynomial interpolation.
There are situations where interpolation for approximating function may not
be efficacious procedure. Errors will arise when the function values f (xi), i = 1, 2,
…, n are observed data and not exact. In this case, if we use the polynomial
interpolation, then it would reproduce all the errors of observation. In such situations
one may take a large number of observed data, so that statistical laws in effect
Self - Learning
72 Material
cancel the errors introduced by inaccuracies in the measuring equipment. The
approximating function is then derived, such that the sum of the squared deviation Interpolation and
Curve Fitting
between the observed values and the estimated values are made as small as
possible.
Mathematically, the problem of curve fitting or function approximation may be NOTES
stated as follows:
To find a functional relationship y = g(x), that relates the set of observed data
values Pi(xi, yi), i = 1, 2,..., n as closely as possible, so that the graph of y = g(x)
goes near the data points Pi’s though not necessarily through all of them.
The first task in curve fitting is to select a proper form of an approximating
function g(x), containing some parameters, which are then determined by minimizing
the total squared deviation.
For example, g(x) may be a polynomial of some degree or an exponential or
logarithmic function. Thus g (x) may be any of the following:
(i) g ( x)     x (ii) g ( x)     x   x 2

(iii) g ( x )   e x (iv) g ( x)   e x

(v) g ( x)   log( x )
Here , ,  are parameters which are to be evaluated so that the curve
y = g(x), fits the data well. A measure of how well the curve fits is called the
goodness of fit.
In the case of least square fit, the parameters are evaluated by solving a system
of normal equations, derived from the conditions to be satisfied so that the sum of
the squared deviations of the estimated values from the observed values, is minimum.

2.3.1 Method of Least Squares

Let (x1, f1), (x2, f2),..., (xn, fn) be a set of observed values and g(x) be the
approximating function. We form the sums of the squares of the deviations of the
observed values fi from the estimated values g (xi),
n
S    f i  g ( xi )
2
i.e.,
i 1

(2.47)
The function g(x) may have some parameters, , , . In order to determine
these parameters we have to form the necessary conditions for S to be minimum,
which are

S S S
 0,  0, 0 (2.48)
  
These equations are called normal equations, solving which we get the
parameters for the best approximate function g(x).

Curve Fitting by a Straight Line: Let g ( x)     x, be the straight line which

fits a set of observed data points (xi, yi), i = 1, 2, ..., n.
Self - Learning
Material 73
Interpolation and Let S be the sum of the squares of the deviations g(xi) – yi, i = 1, 2,..., n, given
Curve Fitting
by,
n
S       xi  yi 
2
(2.49)
NOTES i 1

We now employ the method of least squares to determine  and  so that S

will be minimum. The normal equations are,
S n
 0, i.e.,  (   xi  yi )  0 (2.50)
 i 1

S n
and,  0, i.e.,  xi (   xi  yi )  0 (2.51)
 i 1

These conditions give,

n  S1  S 01  0
S1  S 2  S11  0
n n n n
S1   xi , S01   yi , S 2   xi , S11   xi yi
2
where,
i 1 i 1 i 1 i 1

Solving,
  1 S S
  . Also   01   1 .
 S1S11  S1S2 nS11  S1S01 nS2  S12
n n
Algorithm: Fitting a straight line y = a + bx.
Step 1: Read n [n being the number of data points]
Step 2: Initialize : sum x = 0, sum x2 = 0, sum y = 0, sum xy = 0
Step 3: For j = 1 to n compute
Begin
Read data xj, yj
Compute sum x = sum x + xj
Compute sum x2 + xj × xj
Compute sum y = sum y + yi × yj
Compute sum xy = sum xy + xj × yj
End
Step 4: Compute b = (n × sum xy – sum x × sum y)/ (n × sum x2 – (sum x)2)
Step 5: Compute x bar = sum x / n
Step 6: Compute y bar = sum y / n
Step 8: Compute a = y bar – b × x bar
Step 9: Write a, b
Step 10: For j = 1 to n
Begin
Compute y estimate = a + b × x
write xj, yj, y estimate
Self - Learning End
74 Material
 Step 11: Stop Interpolation and
Curve Fitting
Curve Fitting by a Quadratic (A Parabola): Let g(x) = a + bx + cx2, be the
approximating quadratic to fit a set of data (xi, yi), i = 1, 2, ..., n. Here the
parameters are to be determined by the method of least squares, i.e., by minimizing
the sum of the squares of the deviations given by, NOTES
n
S   ( a  bxi  cxi2  yi ) 2
i 1
(2.52)
S S S
Thus the normal equations,  0,  0,  0, are as follows:
a b c
(2.53)
n

 (a  bx  cx
i 1
i
2
i  yi )  0
n

 x (a  bx
i 1
i i  cxi2  yi )  0
n

x
i 1
2
i (a  bxi  cxi2  yi )  0. (2.54)

These equations can be rewritten as,

na  s1b  s2 c  s01  0
s1a  s2b  s3c  s11  0
s2 a  s3b  s4 c  s21  0 (2.55)

n n n n

where, s1   xi , s2   xi2 , s3   xi3 , s4   xi4

i 1 i 1 i 1 i 1
n n n
s01   yi , s11   xi yi , s21   xi2 yi (2.56)
i 1 i 1 i 1

It is clear that the normal equations form a system of linear equations in the
unknown parameters a, b, c. The computation of the coefficients of the normal
equations can be made in a tabular form for desk computations as shown below.

x xi yi xi2 xi3 xi4 xi yi xi2 yi

1 x1 y1 x12 x13 x 41 x1 y1 x12 y1
2 x2 y2 x22 x23 x24 x2 y2 x22 y2
... ... ... ... ... ... ... ...
n xn yn xn2 xn3 xn4 xn yn xn2 yn
Sum s1 s01 s2 s3 s4 s11 s21

The system of linear equations can be solved by Gaussian elimination method.

It may be noted that number of normal equations is equal to the number of unknown Self - Learning
parameters. Material 75
Interpolation and Example 2.28: Find the straight line fitting the following data:
Curve Fitting

xi 4 6 8 10 12
y1 13.72 12.90 12.01 11.14 10.31
NOTES
Solution: Let y = a + bx, be the straight line which fits the data. We have the
S S
normal equations  0,  0 for determining a and b, where
a b
5
S  ( y  a  bx )
i 1
i i
2
.

5 5

Thus, y
i 1
i  na  b xi  0
i 1

5 5 5

x y  a  xi  b xi  0
2
and , i i
i 1 I 1 i 1

The coefficients are computed in the table below.

xi yi xi2 xi yi
4 13.72 16 54.88
6 12.90 36 77.40
8 12.01 64 96.08
10 11.14 100 111.40
12 10.31 144 123.72
Sum 40 60.08 360 463.48
Thus the normal equations are,
5a  40b  60.08  0
40a  360b  463.48  0
Solving these two equations we obtain,
a = 15.448, b = 0.429
Thus, y = g(x) = 15.448 – 0.429 x, is the straight line fitting the data.
Example 2.29: Use the method of least square approximation to fit a straight line
to the following observed data:

xi 60 61 62 63 64
yi 40 40 48 52 55

Solution: Let the straight line fitting the data be y = a + bx. The data values being
large, we can use a change in variable by substituting u = x – 62 and v = y – 48.
Let v = A + B u, be a straight line fitting the transformed data, where the
normal equations for A and B are,
5 5


i 1
vi  5 A  B u
i 1
i

5 5 5

76
Self - Learning
Material 
i 1
u i vi  A 
i 1
ui  B u
i 1
2
i
 The computation of the various sums are given in the table below, Interpolation and
Curve Fitting

xi yi ui vi ui vi ui2
60 40 –2 –8 16 4
61 42 1 6 6 1 NOTES
62 48 0 0 0 0
63 52 1 4 4 1
64 55 2 7 14 4
Sum 0 3 40 10

Thus the normal equations are,

3  5 A and 40  10 B
3
A  and B  4
5

This gives the line, v  3/ 5  4u

or, 20u  5v  3  0.
Transforming we get the line,
20 (x – 62) – 5 (y – 48) – 3 = 0
or, 20 x – 5y – 1003 = 0
Curve Fitting with an Exponential Curve: We consider a two parameter
exponential curve as,
y  ae  bx (2.57)
For determining the parameters, we can apply the principle of least squares
by first using a transformation,
z = log y, so that Equation (2.57) is rewritten as, (2.58)
z = log a – bx (2.59)
Thus, we have to fit a linear curve of the form z     x, with z – x variables
and then get the parameters a and b as,
a  e , b   (2.60)
Thus proceeding as in linear curve fitting,
n n n
n  xi log yi   xi  log yi
 i 1 i 1 i 1
2
n
  n
(2.61)
n  xi2    xi 
i 1  i 1 

 n   n 
and,   z  px , where x    xi  n , z    log yi  n (2.62)
 i 1   i 1 

After computing  and , we can determine a and b given by Equation (2.59).

Finally, the exponential curve fitting data set is given by Equation (2.57). Self - Learning
Material 77
Interpolation and Algorithm: To fit a straight line for a given set of data points by least square error
Curve Fitting
method.
Step 1: Read the number of data points, i.e., n
NOTES Step 2: Read values of data-points, i.e., Read (xi, yi) for i = 1, 2,..., n
Step 3: Initialize the sums to be computed for the normal equations,
i.e., sx = 0, sx2 = 0, sy = 0, syx = 0
Step 4: Compute the sums, i.e., For i = 1 to n do
Begin
sx  sx  xi
sx 2  sx 2  xi2
sy  sy  yi
syx  syx  xi yi
End
Step 5: Solve the normal equations, i.e., solve for a and b of the line y = a +
bx
Compute d  n  sx 2  sx  sx
b  (n  sxy  sy  sx) / d
xbar  sx / n
ybar  sy / n
a  ybar  b  x bar
Step 6: Print values of a and b
Step 7: Print a table of values of xi , yi , y pi  a  bxi for i  1, 2, ..., n
Step 8: Stop
Algorithm: To fit a parabola y  a  bx  cx 2 , for a given set of data points by least
square error method.
Step 1: Read n, the number of data points
Step 2: Read (xi, yi) for i = 1, 2,..., n, the values of data points
Step 3: Initialize the sum to be computed for the normal equations,
i.e., sx = 0, sx2 = 0, sx3 = 0, sx4 = 0, sy = 0, sxy = 0.
Step 4: Compute the sums, i.e., For i = 1 to n do
Begin
sx  sx  xi
x 2  xi  xi
sx 2  sx 2  x 2
sx 3  sx 3  xi  x 2
sx 4  sx 4  x 2  x 2
sy  sy  yi
sxy  sxy  xi  yi
Self - Learning
78 Material sx 2 y  sx 2 y  x 2  yi
 End Interpolation and
Curve Fitting
Step 5: Form the coefficients {aij } matrix of the normal equations, i.e.,

a11  n, a21  sx, a31  sx 2 NOTES

a12  sx, a22  sx 2 , a32  sx3
a13  sx 2 , a23  sx 3 , a33  sx 4

Step 6: Form the constant vector of the normal equations.

b1 = sy, b2 = sxy, b3 = sx2y
Step 7: Solve the normal equation by Gauss-Jordan method
a12  a12 / a11 , a13  a13 / a11 , b1  b1 / a11
a 22  a 22  a 21a12 , a23  a 23  a21a13
b2  b2  b1a21

a32  a32  a31a12

a33  a33  a31a13
b3  b3  b1a31
a 23  a23 / a 22
b2  b2 / a 22
a33  a33  a23 a32
b3  b3  a32 b2
c  b3 / a33
b  b2  c a23
a  b1  b a12  c a13

Step 8: Print values of a, b, c (the coefficients of the parabola)

Step 9: Print the table of values of xk , yk and y pk where y pk  a  bxk  cx 2 k ,

i.e., print xk , yk , y pk for k  1, 2,..., n.

Step 10: Stop.

2.4 TRIGONOMETRIC FUNCTIONS

Let a revolving line OP start from OX in the anticlockwise direction and trace out
an angle XOP. From P draw PM  OX (Produce OX, if necessary). Let XOP
= .

Then (1) MP is called sine of angle  and is written as sin .

OP
(2) OM is called cosine of angle and is written as cos .
OP
(3) MP is called tangent of angle  and is written as tan .
OM
Self - Learning
Material 79
Interpolation and
Curve Fitting (4) OM is called cotangent of angle  and is written as cot .
MP
OP
(5) is called secant of angle  and is written as sec .
OM
NOTES (6) OP is called cosecant of angle and is written as cosec .
MP
These ratios are called Trigonometrical Ratios of the angle .
P


X
O M

Fig. 2.2

Remarks:
1. It follows from the definition that
1 1 1
sec  = , cosec = , cot  = ,
cos  sin  tan 
sin  cos 
tan  = , cot  = .
cos  sin 
2. Trigonometrical ratios are same for the same angle. For, let P be any
point on the revolving line OP. Draw PM  OX. Then triangles OPM
MP M  P
and OPM are similar, so = i.e., each of these ratios is sin .
OP OP

Therefore, whatever be the triangle of reference (i.e., OPM or OPM)

might be, we find that sin remains the same for a particular angle .
It can be similarly shown that no trigonometrical ratio depends on the size
of triangle of reference.
3. (sin )n is written as sinn , where n is positive. Similar notation holds
good for other trigonometrical ratios.
4. sin–1  denotes that angle whose sine is . Note that sin–1  does not
1
stand for . Similar notation holds good for other trigonometrical ratios.
sin 

P'


X
O M M'

Fig. 2.3

Self - Learning
80 Material
For Any Angle  Interpolation and
Curve Fitting
1. sin2  + cos2  = 1
2. sec2  = 1 + tan2 
3. cosec2  = 1 + cot2  NOTES
Proof. Let the revolving line OP start from OX and trace out an angle  in the
anti-clockwise direction. From P draw PM  OX. (Produce OX, if necessary.)
(Refer Figure 2.2).
Then XOP = 
MP OM
(1) sin  = , cos  =
OP OP
( MP ) 2  (OM ) 2 (OP ) 2
Then sin2  + cos2  = = = 1.
( OP) 2 (OP ) 2
OP MP
(2) sec  = , tan  =
OM OM
( MP) 2 (OM ) 2  ( MP) 2
Then 1 + tan2  = 1  =
(OM ) 2 (OM ) 2

=
(OP ) 2
=
FG OP IJ 2
= (sec )2 = sec2 
(OM ) 2 H OM K
OM OP
(3) cot  = , cosec  = .
MP MP

Then 1 + cot2  = 1 
FG OM IJ 2
=
( MP ) 2  (OM ) 2
H MP K ( MP ) 2

=
(OP ) 2
=
FG OP IJ 2
= (cosec )2 = cosec2 
( MP) 2 H MP K
Signs of Trigonometrical Ratios
Consider four lines OX, OX, OY, OY at right angles to each other. Let a revolving
line OP start from OX in the anticlockwise direction. From P draw PM  OX or
OX. We have the following convention of signs regarding the sides of OPM.
1. OM is positive, if it is along OX.
2. OM is negative, if it is along OX.
3. MP is negative, if it is along OY.
4. MP is positive, if it is along OY.
5. OP is regarded always positive.
Y
P
P

M O M
X' X
M M

P P
Y'
Self - Learning
Fig. 2.4 Material 81
Interpolation and First Quadrant: If the revolving line OP is in the first quadrant, then all the sides
Curve Fitting
of the triangle OPM are positive. Therefore, all the trigonometrical ratios are positive
in the first quadrant.
Second Quadrant: If the revolving line OP is in the second quadrant, then OM is
NOTES
negative and the other two sides of OPM are positive. Therefore, ratios involving
OM will be negative. So, cosine, secant, tangent, cotangent of an angle in the
second quadrant are negative while sine and cosecant of anlge in the second quadrant
are positive.
Third Quadrant: If the revolving line is in the third quadrant, then sides OM and
MP both are negative. Since OP is always positive, therefore, ratios involving
each one of OM and MP alone will be negative. So, sine, cosine, cosecant and
secant of an angle in the third quadrant are negative. Since tangent or cotangent of
any angle involve both OM and MP, therefore, these will be positive. So, tangent
and cotangent of an angle in the third quadrant are positive.
Fourth Quadrant: If the revolving line OP is in the fourth quadrant, then MP is
negative and the other two sides of OPM are positive. Therefore, ratios involving
MP will be negative and others positive. So, sine, cosecant, tangent and cotangent
of an angle in the fourth quadrant are negative while cosine and secant of an angle
in the fourth quadrant are positive.
Limits to the Value of Trigonometrical Ratios
We know that sin2  + cos2  = 1 for any angle . sin2  and cos2  being perfect
squares, will be positive. Again neither of them can be greater than 1 because then
the other will have to be negative.
Thus sin2  1, cos2 1.
 sin  and cos  cannot be numerically greater than 1.
1 1
Similarly, cosec  = and sec  = cannot be numerically less
sin  cos 
than 1.
There is no restriction on tan  and cot . They can have any value.
Example 2.30: Prove that sin6  + cos6  = 1 – 3 sin2  cos2 .
Solution: Here LHS = sin6  + cos6 
= (sin2 )3 + (cos2 )3
= (sin2  + cos2 )(sin4  – sin2  cos2  + cos4 )
= 1 . (sin4 – sin2  cos2 + cos4 )
= [(sin2  + cos2 )2 – 3 sin2  cos2 ]
= 1 – 3 sin2  cos2  = RHS.
1  cos 
Example 2.31: Prove that = cosec  + cot . Provided cos  1.
1  cos 
1  cos 
Solution. LHS =
1  cos 
(1  cos )(1  cos ) 1  cos 
Self - Learning = = 
82 Material (1  cos )(1  cos ) 1  cos2 
 1  cos  1 cos  Interpolation and
= =  = cosec + cot  Curve Fitting
sin  sin  sin 
Example 2.32: Prove that (1 + cot  – cosec )(1 + tan  + sec ) = 2.
Solution: LHS = (1 + cot  – cosec )(1 + tan  + sec ) NOTES
F cos   1 IJ FG1  sin   1 IJ
= G1 
H sin  sin  K H cos  cos  K
(sin   cos   1)(cos   sin   1)
LHS =
sin  cos 
(sin   cos ) 2  1
=
sin  cos 
sin 2   cos2   2 sin  cos   1
=
sin  cos 

1  2 sin  cos   1 2 sin  cos 

= = = 2 = RHS.
sin  cos  sin  cos 
tan  cot 
Example 2.33: Prove that  = 1 + cosec  sec , if cot  1,
1  cot  1  tan 
0 and tan  1, 0.
tan  cot 
Solution: LHS = 
1  cot  1  tan 
1
tan  tan 
= 
1 1  tan 
1
tan 
tan 2  1
= 
tan   1 tan (1  tan )
tan 2  1
= 
tan   1 tan  (tan   1)
tan 3   1
=
tan  (tan   1)
(tan   1)(tan 2   tan   1)
=
tan  (tan   1)
tan 2   tan   1
= since tan  1
tan 
sec2   tan 
=
tan 

sec 2 
=  1 = sec  cosec  + 1 = RHS.
tan 
Example 2.34: Which of the six trigonometrical ratios are positive for (i) 960º
(ii) – 560º?
Solution: (i) 960º = 720º + 240º.
Therefore, the revolving line starting from OX will make two complete
Self - Learning
revolutions in the anticlockwise direction and further trace out an angle of 240º in Material 83
Interpolation and the same direction. Thus, it will be in the third quadrant. So, the tangent and
Curve Fitting
cotangent are positive and rest of trigonometrical ratios will be negative.
(ii) – 560º = – 360º – 200º.
NOTES Therefore, the revolving line after making one complete revolution in the
clockwise direciton, will trace out further an angle of 200º in the same direction.
Thus, it will be in the second quadrant. So, only sine and cosecant are positive.
7
Example 2.35: In what quadrants can lie if sec  = ?
6
Solution: As sec  is negative in second and third quadrants,  can lie in second
or third quadrant only.
12
Example 2.36: If sin  = , determine other trigonometrical ratios of .
13
Solution. cos2  = 1 – sin2 
144 169  144 25
= 1 = = .
169 169 169
5
 cos  =  .
13
sin 
So tan  = = ± 12
cos  5
13 13 5
cosec  = , sec  =  , cot  = ± .
12 5 12
Example 2.37: Express all the trigonometrical ratios of  in terms of the sin .
Solution: Let sin  = k.
Then, cos2  = 1 – sin2  = 1 – k2  cos  =  1  k 2
sin  k
tan  = = 
cos  1 k2

cos  1 k2
cot  = = 
sin  k
1 1
sec  = = 
cos  1k2
1 1
cosec  = = .
sin  k
1
Example 2.38: Prove that sin  = a  is impossible, if a is real.
a
1 a2  1
Solution: sin  = a   sin  =
a a
 a2 – a sin  + 1 = 0

sin   sin 2   4
 a=
2
For a to be real, the expression under the radical sign, must be positive or
Self - Learning
zero.
84 Material i.e., sin2  – 4  0
or sin2   4  sin  is numerically greater than or Interpolation and
Curve Fitting
equal to 2 which is impossible.
1
Thus, if a is real, sin  = a  is impossible.
a NOTES
Example 2.39: Determine the quadrant in which  must lie if cot  is positive and
cosec  is negative.
Solution: cot  is positive  lies in first or third quadrant.
cosec  is negative  lies in third or fourth quadrant.
In order that cot is positive and cosec  is negatie, we see that  must lie in
third quadrant.
Example 2.40: Prove that
1 1 1 1
 = 
cosec   cot  sin  sin  cosec   cot 

1 1
Solution: LHS = 
cosec   cot  sin 

sin  1
= 
1  cos  sin 

sin 2   (1  cos )
=
(1  cos ) sin 

 (1  sin 2 )  cos 
=
(1  cos ) sin 

 cos2   cos 
=
(1  cos ) sin 

 cos  (1  cos )
= = – cot 
(1  cos ) sin 

1 1
RHS = 
sin  cosec   cot 

1 sin 
= 
sin  1  cos 

1  cos   sin 2 
=
sin  (1  cos )

cos 2   cos 
=
sin  (1  cos )
 cos  (1  cos )
= = – cot 
sin  (1  cos )
Therefore, LHS = RHS.
Example 2.41: Prove that,
sin (1 + tan ) + cos  (1 + cot ) = sec  + cosec .
Self - Learning
Material 85
Interpolation and Solution: LHS = sin  (1 + tan ) + cos  (1 + cot )
Curve Fitting
FG
= sin  1 
sin  IJ FG
 cos  1 
cos  IJ
H cos  K Hsin  K
NOTES sin 2  cos2 
= sin    cos  
cos  sin 

sin 2  cos   sin 3   cos 2  sin   cos3 

=
sin  cos 

sin 2  (sin   cos )  cos2  (sin   cos )

=
sin  cos 

(sin 2   cos2 ) (sin   cos )

=
sin  cos 
sin   cos 
=
sin  cos 
1 1
=  = sec  + cosec  = RHS.
cos  sin 
4
Example 2.42: If tan  = , find the value of
3
2 sin   3 cos 
.
4 cos   3 sin 
8
3 23
2 sin   3 cos  2 tan   3
Solution: = = 5 12 = .
4 cos   3 sin  4  3 tan  4
32
5
Example 2.43: State giving the reason whether the following equation is possible.
2 sin2  – 3 cos  – 6 = 0
Solution: 2 sin2  – 3 cos  – 6 = 0
 2(1 – cos2 ) – 3 cos  – 6 = 0
 – 2 cos2 – 3 cos – 4 = 0
 2 cos2  + 3 cos  + 4 = 0
 3  9  32  3   23
 cos  = =
4 4
 cos  is imaginary, which is not true.
Example 2.44: Prove that
1  sin  1  sin 
 = 2 cos  (cot  + cosec2 )
1  sec  1  sec 

(1  sin ) cos  (1  sin ) cos 

Solution: LHS = 
1  cos  cos   1

= cos 
LM (1  sin )  (1  sin ) OP
N (1  cos ) (1  cos ) Q
Self - Learning
86 Material
 Interpolation and
 (1  sin ) (1  cos )  Curve Fitting
  (1  sin ) (1  cos ) 
= cos   
 1  cos 2  
 
 
NOTES
LM1  sin   cos  sin  cos OP
 1  sin   cos   sin  cos 
= cos  M PP
MM sin  2

N QP
= cos 
LM 2  2 sin  cos  OP
MN sin  PQ
2

= 2 cos [cosec2  + cos ] = RHS.

sin   cos 
Example 2.45: If tan x = where  and x are both positive and
sin   cos 
acute angles, prove that
1
sin x = ( sin   cos )
2
sin 2   cos2   2 sin  cos 
Solution: 1 + tan2 x = 1
sin 2   cos2   2 sin  cos 
(1  2 sin  cos ) 2
= 1 =
(1  2 sin  cos ) 1  2 sin  cos 
2
Therefore, sec2 x =
1  2 sin  cos 
1  2 sin  cos 
 cos2 x =
2
2  (1  2 sin  cos )
 1 – cos2 x =
2
1  2 sin  cos  (sin   cos ) 2
= =
2 2
(sin   cos )2
 sin2 x =
2
(sin   cos )
 sin x =  .
2
Since  is acute and tan x  0, sin   cos 
 sin – cos   0
Also x is acute sin x  0
(sin   cos )
 sin x =  .
2
Example 2.46: Exhibit (sin  – 3)(sin  – 1)(sin + 1)(sin  + 3) + 16
as a perfect square and examine if there is any suitable value of  for which the
above expression can vanish.
Self - Learning
Material 87
Interpolation and Solution: Now (sin  – 3)(sin  – 1)(sin + 1)(sin  + 3) + 16
Curve Fitting
= (sin2  – 1)(sin2  – 9) + 16
= sin4  – 10 sin2 + 25
NOTES = (sin2  – 5)2.
This is 0 only when sin2  – 5 = 0 i.e., only when sin2  = 5
which is not possible as the maximum value of sin2  is 1.
Thus, there is no value of for which the given expression can vanish.
Example 2.47: Find the value in terms of p and q of
p cos   q sin  p
where cot  = .
p cos   q sin  q

Here use of any figure not allowed.

cos 
p q
p cos   q sin  sin 
Solution: = cos 
p cos   q sin  p q
sin 
p2
q
p cot   q q p2  q 2
= = 2 = 2 2.
p cot   q p p q
q
q
Example 2.48: Show that
tan  tan 
 = 2 cosec .
sec   1 sec   1
tan  tan 
Solution: LHS = 
sec   1 sec   1

= tan 
LM 1  1 OP
N sec   1 sec   1Q
= tan  M
L 2 sec  OP
NM sec   1QP
2

L 2 sec  OP
= tan  M
MN tan  PQ
2

2 sec  2
= = = 2 cosec  = RHS.
tan  sin 

2.5 REGRESSION
The term ‘Regression’ was first used in 1877 by Sir Francis Galton who made a
study that showed that the height of children born to tall parents will tend to move
back or ‘regress’ toward the mean height of the population. He designated the
word regression as the name of the process of predicting one variable from another
variable. He coined the term multiple regression to describe the process by which
Self - Learning several variables are used to predict another. Thus, when there is a well-established
88 Material
relationship between variables, it is possible to make use of this relationship in Interpolation and
Curve Fitting
making estimates and to forecast the value of one variable (the unknown or the
dependent variable) on the basis of the other variable/s (the known or the
independent variable/s). A banker, for example, could predict deposits on the
basis of per capita income in the trading area of bank. A marketing manager, may NOTES
plan his advertising expenditures on the basis of the expected effect on total sales
revenue of a change in the level of advertising expenditure. Similarly, a hospital
superintendent could project his need for beds on the basis of total population.
Such predictions may be made by using regression analysis. An investigator may
employ regression analysis to test his theory having the cause and effect relationship.
All these explain that regression analysis is an extremely useful tool especially in
problems of business and industry involving predictions.
Assumptions in regression analysis
While making use of the regression techniques for making predictions, the following
are always assumed:
(i) There is an actual relationship between the dependent and independent
variables.
(ii) The values of the dependent variable are random but the values of the
independent variable are fixed quantities without error and are chosen by
the experimentor.
(iii) There is a clear indication of direction of the relationship. This means that
dependent variable is a function of independent variable. (For example,
when we say that advertising has an effect on sales, then we are saying that
sales has an effect on advertising).
(iv) The conditions (that existed when the relationship between the dependent
and independent variable was estimated by the regression) are the same
when the regression model is being used. In other words, it simply means
that the relationship has not changed since the regression equation was
computed.
(v) The analysis is to be used to predict values within the range (and not for
values outside the range) for which it is valid.
2.5.1 Linear Regression
In case of simple linear regression analysis, a single variable is used to predict
another variable on the assumption of linear relationship (i.e., relationship of the
type defined by Y = a + bX) between the given variables. The variable to be
predicted is called the dependent variable and the variable on which the prediction
is based is called the independent variable.
Simple linear regression model3 (or the Regression Line) is stated as,
Yi = a + bXi + ei
Where, Yi = The dependent variable
Xi = The independent variable
ei = Unpredictable random element (usually called
Self - Learning
residual or error term) Material 89
Interpolation and (i) a represents the Y-intercept, i.e., the intercept specifies the value of the
Curve Fitting
dependent variable when the independent variable has a value of zero.
(However, this term has practical meaning only if a zero value for the
independent variable is possible).
NOTES
(ii) b is a constant, indicating the slope of the regression line. Slope of the line
indicates the amount of change in the value of the dependent variable for a
unit change in the independent variable.
If the two constants (viz., a and b) are known, the accuracy of our prediction of Y
(denoted by Ŷ and read as Y--hat) depends on the magnitude of the values of ei. If
in the model, all the ei tend to have very large values then the estimates will not be
very good, but if these values are relatively small, then the predicted values ( Ŷ )
will tend to be close to the true values (Yi).
Estimating the intercept and slope of the regression model (or estimating
the regression equation)
The two constants or the parameters viz., ‘a’ and ‘b’ in the regression model for
the entire population or universe are generally unknown and as such are estimated
from sample information. The following are the two methods used for estimation:
(i) Scatter diagram method
(ii) Least squares method
1. Scatter diagram method
This method makes use of the Scatter diagram also known as Dot diagram. Scatter
diagram is a diagram representing two series with the known variable, i.e.,
independent variable plotted on the X-axis and the variable to be estimated, i.e.,
dependent variable to be plotted on the Y-axis on a graph paper (Refer Figure
2.5) to get the following information illustrated in Table 2.1:
Table 2.1 Table Derived from Scatter Diagram

Income Consumption Expenditure

X Y
(Hundreds of Rupees) (Hundreds of Rupees)
41 44
65 60
50 39
57 51
96 80
94 68
110 84
30 34
79 55
65 48

Self - Learning
90 Material
 y-axis Interpolation and
Curve Fitting

Consumption Expenditure (’00 Rs)

120
100
NOTES
80
60
40
20
x-axis
0 20 40 60 80 100 120

Fig. 2.5 Scatter Diagram

The scatter diagram by itself is not sufficient for predicting values of the dependent
variable. Some formal expression of the relationship between the two variables is
necessary for predictive purposes. For the purpose, one may simply take a ruler
and draw a straight line through the points in the scatter diagram and this way can
determine the intercept and the slope of the said line and then the line can be
defined as Yˆ  a  bX i , with the help of which we can predict Y for a given value of
X. However, there are shortcomings in this approach. For example, if five different
persons draw such a straight line in the same scatter diagram, it is possible that
there may be five different estimates of a and b, especially when the dots are more
dispersed in the diagram. Hence, the estimates cannot be worked out only through
this approach. A more systematic and statistical method is required to estimate the
constants of the predictive equation. The least squares method is used to draw the
best fit line.
2. Least square method
The least squares method of fitting a line (the line of best fit or the regression line)
through the scatter diagram is a method which minimizes the sum of the squared
vertical deviations from the fitted line. In other words, the line to be fitted will pass
through the points of the scatter diagram in such a way that the sum of the squares
of the vertical deviations of these points from the line will be a minimum.
The meaning of the least squares criterion can be easily understood through
as shown in Figure 2.6, where the earlier as shown in Figure 2.5 in scatter diagram
has been reproduced along with a line which represents the least squares line to fit
the data.

Fig. 2.6 Scatter Diagram, Regression Line and Short Vertical Lines Self - Learning
Material 91
Representing ‘e’
Interpolation and As shown in Figure 2.6, the vertical deviations of the individual points from the line
Curve Fitting
are shown as the short vertical lines joining the points to the least squares line.
These deviations will be denoted by the symbol ‘e’. The value of ‘e’ varies from
one point to another. In some cases it is positive, while in others it is negative. If
NOTES the line drawn happens to be the least squares line, then the values of  ei is the
least possible. It is because of this feature the method is known as Least Squares
Method.
Why we insist on minimizing the sum of squared deviations is a question that
needs explanation. If we denote the deviations from the actual value Y to the
n
estimated value Ŷ as (Y – Yˆ ) or ei, it is logical that we want the (Y – Yˆ ) or  ei , to
i 1
n
be as small as possible. However, mere examining (Y – Yˆ ) or  ei , is
i 1
inappropriate, since any ei can be positive or negative. Large positive values and
large negative values could cancel one another. However, large values of ei
regardless of their sign, indicate a poor prediction. Even if we ignore the signs
n
while working out | ei | , the difficulties may continue. Hence, the standard
i 1

procedure is to eliminate the effect of signs by squaring each observation.

Squaring each term accomplishes two purposes, viz., (i) It magnifies (or penalizes)
the larger errors, and (ii) It cancels the effect of the positive and negative values
(since a negative error when squared becomes positive). The choice of minimizing
the squared sum of errors rather than the sum of the absolute values implies that
there are many small errors rather than a few large errors. Hence, in obtaining
the regression line, we follow the approach that the sum of the squared deviations
be minimum and on this basis work out the values of its constants viz., ‘a’ and
‘b’ also known as the intercept and the slope of the line. This is done with the
help of the following two normal equations:
Y = na + bX
XY = aX + bX2
In these two equations, ‘a’ and ‘b’ are unknowns and all other values viz., X,
Y, X2, XY, are the sum of the products and cross products to be calculated
from the sample data, and ‘n’ means the number of observations in the sample.
Example 2.49 explains the Least squares method.
Example 2.49: Fit a regression line Yˆ  a  bX i by the method of Least squares
to the following sample information.
Observations 1 2 3 4 5 6 7 8 9 10
Income (X) (00 ) 41 65 50 57 96 94 110 30 79 65
Consumption
Expenditure (Y) (00 ) 44 60 39 51 80 68 84 34 55 48

Self - Learning
92 Material
Solution: Interpolation and
Curve Fitting
We are to fit a regression line Yˆ  a  bX i to the given data by the method of
Least squares. Accordingly, we work out the ‘a’ and ‘b’ values with the help
of the normal equations as stated above and also for the purpose, work out NOTES
X, Y, XY, X2 values from the given sample information table on
summations for regression equation.
Summations for Regression Equation

Observations Income Consumption XY X2 Y2

X Expenditure
Y
(00 ) (00 )

1 41 44 1804 1681 1936

2 65 60 3900 4225 3600
3 50 39 1950 2500 1521
4 57 51 2907 3249 2601
5 96 80 7680 9216 6400
6 94 68 6392 8836 4624
7 110 84 9240 12100 7056
8 30 34 1020 900 1156
9 79 55 4345 6241 3025
10 65 48 3120 4225 2304
n = 10 X = 687 Y =563 XY = 42358X2= 53173 Y2 = 34223

Putting the values in the required normal equations we have,

563 = 10a + 687b
42358 = 687a + 53173b
Solving these two equations for a and b we obtain,
a = 14.000 and b = 0.616
Hence, the equation for the required regression line is,
Ŷ = a + bXi
or, Ŷ = 14.000 + 0.616Xi
This equation is known as the regression equation of Y on X from which Y
values can be estimated for given values of X variable.

Checking the accuracy of equation

After finding the regression line, one can check its accuracy also. The method to
be used for the purpose follows from the mathematical property of a line fitted by
the method of least squares, viz., the individual positive and negative errors must
sum to zero. In other words, using the estimating equation one must find out whether
the term  Y  Yˆ  is zero and if this is so, then one can reasonably be sure that he
has not committed any mistake in determining the estimating equation.
Self - Learning
Material 93
Interpolation and The problem of prediction
Curve Fitting
When we talk about prediction or estimation, we usually imply that if the relationship
Yi = a + bXi + ei exists, then the regression equation, Yˆ  a  bX i provides a base
NOTES for making estimates of the value of Y which will be associated with particular
values of X. In Example 2.49, we worked out the regression equation for the
income and consumption data as,
Ŷ = 14.000 + 0.616Xi
On the basis of this equation, we can make a point estimate of Y for any
given value of X. Suppose we wish to estimate the consumption expenditure of
individuals with income of 10,000. We substitute X = 100 for the same in our
equation and get an estimate of consumption expenditure as,
Yˆ  14.000  0.616 100  75.60
Thus, the regression relationship indicates that individuals with 10,000 of
income may be expected to spend approximately 7,560 on consumption.
However, this is only an expected or an estimated value and it is possible that
actual consumption expenditure of same individual with that income may deviate
from this amount and if so, then our estimate will be an error, the likelihood of
which will be high if the estimate is applied to any one individual. The interval
estimate method is considered better and it states an interval in which the expected
consumption expenditure may fall. Remember that the wider the interval, the greater
the level of confidence we can have, but the width of the interval (or what is
technically known as the precision of the estimate) is associated with a specified
level of confidence and is dependent on the variability (consumption expenditure
in our case) found in the sample. This variability is measured by the standard
deviation of the error term, ‘e’, and is popularly known as the standard error of
the estimate.
Standard error of the estimate
Standard error of estimate is a measure developed by statisticians for measuring
the reliability of the estimating equation. Like the standard deviation, the Standard
Error (S.E.) of Ŷ measures the variability or scatter of the observed values of Y
around the regression line. Standard Error of Estimate (S.E. of Ŷ ) is worked out
as,

S.E. of Yˆ (or Se ) 
 (Y  Yˆ )2 
 e2
n2 n2
where, S.E. of Ŷ (or Se) = Standard error of the estimate
Y = Observed value of Y
Ŷ = Estimated value of Y
e = The error term = (Y– Ŷ )
n = Number of observations in the sample
Self - Learning
94 Material
Note: In the above formula, n – 2 is used instead of n because of the fact that two degrees of Interpolation and
freedom are lost in basing the estimate on the variability of the sample observations about Curve Fitting
the line with two constants viz., ‘a’ and ‘b’ whose position is determined by those same
sample observations.
The square of the Se, also known as the variance of the error term, is the NOTES
basic measure of reliability. The larger the variance, the more significant are the
magnitudes of the e’s and the less reliable is the regression analysis in predicting
the data.
Interpreting the standard error of estimate and finding the confidence
limits for the estimate in large and small samples
The larger the S.E. of estimate (SEe), the greater happens to be the dispersion,
or scattering, of given observations around the regression line. However, if the
S.E. of estimate happens to be zero, then the estimating equation is a ‘Perfect’
estimator (i.e., cent per cent correct estimator) of the dependent variable.
(i) In case of large samples, i.e., where n > 30 in a sample, it is assumed that
the observed points are normally distributed around the regression line and we
may find that,
 68 per cent of all points lie within Yˆ  1 SEe limits.
 95.5 per cent of all points lie within Yˆ  2 SEe limits.
 99.7 per cent of all points lie within Yˆ  3 SEe limits.
This can be stated as,
a. The observed values of Y are normally distributed around each estimated
value of Ŷ and;
b. The variance of the distributions around each possible value of Ŷ is the
same.
(ii) In case of small samples, i.e., where n  30 in a sample the ‘t’ distribution
is used for finding the two limits more appropriately.
This is done as follows:
Upper limit = Ŷ + ‘t’ (SEe)
Lower limit = Ŷ – ‘t’ (SEe)
Where, Ŷ = The estimated value of Y for a given value of X.
SEe = The standard error of estimate.
‘t’ = Table value of ‘t’ for given degrees of freedom for a
specified confidence level.

Some other details concerning simple regression

Sometimes the estimating equation of Y also known as the regression equation of
Y on X, is written as,
Y
Ŷ  Y  = r X X
X i

Self - Learning
Material 95
Interpolation and

Curve Fitting
or, 
r Y Xi  X  Y
Ŷ =  X 
Where, r = Coefficient of simple correlation between X and
NOTES Y
Y = Standard deviation of Y
X = Standard deviation of X
_
X = Mean of X
_
Y = Mean of Y
Ŷ = Value of Y to be estimated
Xi = Any given value of X for which Y is to be estimated
This is based on the formula we have used, i.e., Yˆ  a  bX i . The coefficient of
Xi is defined as,
Y
Coefficient of Xi = b = r 
X

(Also known as regression coefficient of Y on X or slope of the regression line

of Y on X) or bYX.

2
 XY  nX Y   Y 2  nY
= 2 2
 Y 2  nY 2  X 2  n X  X 2  nX
XY  n X Y
= 2
X 2  n X

Y 
and a = r  X  Y
X

 Y 
= Y  bX  since b  r  
X

Similarly, the estimating equation of X, also known as the regression equation of X

on Y, can be stated as,

 X̂  X  = r  Y  Y 
X

Y

or
X

X̂ = r  Y  Y  X
Y

and the
 X  XY  n X Y
Regression coefficient of X on Y (or bXY)  r 
Y  Y 2  nY
2

If we are given the two regression equations as stated above, along with the values
of ‘a’ and ‘b’ constants to solve the same for finding the value of X and Y, then the
values of X and Y so obtained, are the mean values of X (i.e., X ) and the mean
value of Y (i.e., Y ).
Self - Learning
96 Material
 If we are given the two regression coefficients (viz., bXY and bYX), then we Interpolation and
Curve Fitting
can work out the value of coefficient of correlation by just taking the square root
of the product of the regression coefficients as shown,
r = bYX .bXY
NOTES
Y  X
= r .r
 X Y
= r.r =r
The (±) sign of r will be determined on the basis of the sign of the given regression
coefficients. If regression coefficients have minus sign then r will be taken with
minus (–) sign and if regression coefficients have plus sign then r will be taken with
plus (+) sign, (Remember that both regression coefficients will necessarily have
the same sign, whether it is minus or plus, for their sign is governed by the sign of
coefficient of correlation.) To understand it better, Refer Examples 2.50 and 2.51.
Example 2.50: Given is the following information:
X Y
Mean 39.5 47.5
Standard Deviation 10.8 17.8
Simple correlation coefficient between X and Y is = + 0.42.
Find the estimating equation of Y and X.
Solution:
Estimating equation of Y can be worked out as,

 Ŷ  Y  = r   Xi  X 
Y


or Ŷ = r
Y

X  X Y
X i

17.8
= 0.42  X i  39.5   47.5
10.8
= 0.69 X i  27.25  47.5
= 0.69Xi + 20.25
Similarly, the estimating equation of X can be worked out as
X
  X̂  X  = r 
Y Y
Y i

X
or X̂ = r
Y i

Y Y  X 
10.8
or = 0.42 Yi  47.5   39.5
17.8
= 0.26Yi – 12.35 + 39.5
= 0.26Yi + 27.15
Self - Learning
Material 97
Interpolation and Example 2.51: The following is the given data:
Curve Fitting
Variance of X = 9
Regression equations:
NOTES 4X – 5Y + 33 = 0
20X – 9Y – 107 = 0
Find: (i) Mean values of X and Y
(ii) Coefficient of Correlation between X and Y
(iii) Standard deviation of Y
Solution:
(i) For finding the mean values of X and Y, we solve the two given regression
equations for the values of X and Y as follows:
4X – 5Y + 33 = 0 ...(1)
20X – 9Y –107 = 0 ...(2)
If we multiply Equation (1) by 5, we have the following equations:
20X – 25Y = –165 ...(3)
20X – 9Y = 107 ...(2)
– + –
– 16Y = –272
Subtracting Equations (2) from (3)
or Y = 17
Putting this value of Y in Equation (1) we have,
4X = – 33 + 5(17)
33  85 52
or X =   13
4 4
_
Hence, X = 13 and Y = 17
(ii) For finding the coefficient of correlation, first of all we presume one of the
two given regression equations as the estimating equation of X. Let equation
4X – 5Y + 33 = 0 be the estimating equation of X, then we have,
5Y 33
Xˆ  i 
4 4
and
5
From this we can write bXY  .
4
The other given equation is then taken as the estimating equation of Y and can
be written as,
20 X i 107
Yˆ  
9 9
20
and from this we can write bYX  .
Self - Learning
9
98 Material
If the above equations are correct then r must be equal to, Interpolation and
Curve Fitting
r = 5 / 4  20 / 9  25 / 9 = 5/3 = 1.6
which is an impossible equation, since r can in no case be greater than 1.
Hence, we change our supposition about the estimating equations and by reversing NOTES
it, we re-write the estimating equations as,
9Y 107
Xˆ  i 
20 20
4 X i 33
and Yˆ  
5 5
Hence, r = 9 / 20  4 / 5
= 9 / 25
= 3/5
= 0.6
Since, regression coefficients have plus signs, we take r = + 0.6
(iii) Standard deviation of Y can be calculated,
 Variance of X = 9  Standard deviation of X = 3
Y 4 
 bYX  r =  0.6 Y  0.2 Y
X 5 3

Hence, Y = 4
Alternatively, we can work it out as,
X 9  1.8

Y
bXY  r = 20  0.6 3  
Y Y

Hence, Y = 4

2.5.2 Polynomial Regression

In statistics, polynomial regression is a form of regression analysis in which the
relationship between the independent variable x and the dependent variable y is
modelled as an nth degree polynomial in x. Polynomial regression fits a nonlinear
relationship between the value of x and the corresponding conditional mean of y,
denoted E(y|x). Although polynomial regression fits a nonlinear model to the data,
as a statistical estimation problem it is linear, in the sense that the regression function
E(y|x) is linear in the unknown parameters that are estimated from the data. For
this reason, polynomial regression is considered to be a special case of multiple
linear regression.
The explanatory (independent) variables resulting from the polynomial
expansion of the ‘Baseline’ variables are known as higher-degree terms. Such
variables are also used in classification settings.
Ancient Times of Polynomial regression
Polynomial regression models are usually fit using the method of least squares. The
least-squares method minimizes the variance of the unbiased estimators of the
Self - Learning
Material 99
Interpolation and coefficients, under the conditions of the Gauss–Markov theorem. The least-squares
Curve Fitting
method was published in 1805 by Legendre and in 1809 by Gauss. The first design
of an experiment for polynomial regression appeared in an 1815 paper of
Gergonne. In the twentieth century, polynomial regression played an important role
NOTES in the development of regression analysis, with a greater emphasis on issues of
design and inference. More recently, the use of polynomial models has been
complemented by other methods, with non-polynomial models having advantages
for some classes of problems.
Definition and Example of Polynomial Regression
The goal of regression analysis is to model the expected value of a dependent
variable y in terms of the value of an independent variable (or vector of independent
variables) x. In simple linear regression, the model

is used, where ε is an unobserved random error with mean zero conditioned

on a scalar variable x. In this model, for each unit increase in the value of x, the
conditional expectation of y increases by β1 units.
In many settings, such a linear relationship may not hold. For example, if we
are modeling the yield of a chemical synthesis in terms of the temperature at which
the synthesis takes place, we may find that the yield improves by increasing amounts
for each unit increase in temperature. In this case, we might propose a quadratic
model of the form

In this model, when the temperature is increased from x to x + 1 units, the

expected yield changes by

(This can be seen by replacing x in this equation with x+1 and subtracting
the equation in x from the equation in x+1.) For infinitesimal changes in x, the effect
on y is given by the total derivative with respect to x:
The fact that the change in yield depends on x is what makes the relationship
between x and y nonlinear even though the model is linear in the parameters to be
estimated.
In general, we can model the expected value of y as an nth degree polynomial,
yielding the general polynomial regression model

Conveniently, these models are all linear from the point of view of estimation,
since the regression function is linear in terms of the unknown parameters β0, β1,
.... therefore, for least squares analysis, the computational and inferential problems
of polynomial regression can be completely addressed using the techniques of
multiple regression. This is done by treating x, x2, ... as being distinct independent
Self - Learning variables in a multiple regression model.
100 Material
Matrix form and Calculation of Estimates Interpolation and
Curve Fitting
The polynomial regression model

NOTES
can be expressed in matrix form in terms of a design matrix X is a
Vandermonde matrix, the invertibility condition is guaranteed to hold if all the xi
values are distinct. This is the unique least-squares solution.
Explanation of Polynomial Regression
Although polynomial regression is technically a special case of multiple linear
regression, the interpretation of a fitted polynomial regression model requires a
somewhat different perspective. It is often difficult to interpret the individual
coefficients in a polynomial regression fit, since the underlying monomials can be
highly correlated. For example, x and x2 have correlation around 0.97 when x is
uniformly distributed on the interval (0, 1). Although the correlation can be reduced
by using orthogonal polynomials, it is generally more informative to consider the
fitted regression function as a whole. Point-wise or simultaneous confidence bands
can then be used to provide a sense of the uncertainty in the estimate of the
regression function.
Alternative Approaches of Polynomial Regression
Polynomial regression is one example of regression analysis using basis functions
to model a functional relationship between two quantities. More specifically, it
replaces in linear regression with polynomial basis
A drawback of polynomial bases is that the basis
functions are ‘Non-Local’, meaning that the fitted value of y at a given value x = x0
depends strongly on data values with x far from x0. In modern statistics, polynomial
basis-functions are used along with new basis functions, such as splines, radial basis
functions, and wavelets. These families of basis functions offer a more parsimonious
fit for many types of data.
The goal of polynomial regression is to model a non-linear relationship
between the independent and dependent variables (technically, between the
independent variable and the conditional mean of the dependent variable). This is
similar to the goal of nonparametric regression, which aims to capture non-linear
regression relationships. Therefore, non-parametric regression approaches such
as smoothing can be useful alternatives to polynomial regression. Some of these
methods make use of a localized form of classical polynomial regression. An
advantage of traditional polynomial regression is that the inferential framework of
multiple regression can be used (this also holds when using other families of basis
functions such as splines). A final alternative is to use kernelized models such as
support vector regression with a polynomial kernel. If residuals have unequal
variance, a weighted least squares estimator may be used to account for that.
2.5.3 Fitting Exponential
In this section, we consider the problem of approximating an unknown function
whose values, at a set of points, are generally known only empirically and are, thus
subject to inherent errors, which may sometimes be appreciably larger in many Self - Learning
Material 101
Interpolation and engineering and scientific problems. In these cases, it is required to derive a
Curve Fitting
functional relationship using certain experimentally observed data. Here the
observed data may have inherent or round-off errors, which are serious, making
polynomial interpolation for approximating the function inappropriate. In polynomial
NOTES interpolation the truncation error in the approximation is considered to be important.
But when the data contains round-off errors or inherent errors, interpolation is not
appropriate.
The subject of this section is curve fitting by least square approximation. Here
we consider a technique by which noisy function values are used to generate a
smooth approximation. This smooth approximation can then be used to approximate
the derivative more accurately than with exact polynomial interpolation.
There are situations where interpolation for approximating function may not
be efficacious procedure. Errors will arise when the function values f (xi), i = 1, 2,
…, n are observed data and not exact. In this case, if we use the polynomial
interpolation, then it would reproduce all the errors of observation. In such situations
one may take a large number of observed data, so that statistical laws in effect
cancel the errors introduced by inaccuracies in the measuring equipment. The
approximating function is then derived, such that the sum of the squared deviation
between the observed values and the estimated values are made as small as
possible.
Mathematically, the problem of curve fitting or function approximation may be
stated as follows:
To find a functional relationship y = g(x), that relates the set of observed data
values Pi(xi, yi), i = 1, 2,...,n as closely as possible, so that the graph of y = g(x)
goes near the data points Pi’s though not necessarily through all of them.
The first task in curve fitting is to select a proper form of an approximating
function g(x), containing some parameters, which are then determined by minimizing
the total squared deviation.
For example, g (x) may be a polynomial of some degree or an exponential or
logarithmic function. Thus g (x) may be any of the following:
(i) g ( x)     x (ii) g ( x)     x   x 2

(iii) g ( x)   e  x (iv) g ( x)   e   x
(v) g ( x)   log(  x)
Here  ,  ,  are parameters which are to be evaluated so that the curve
y = g(x), fits the data well. A measure of how well the curve fits is called the
goodness of fit.
In the case of least square fit, the parameters are evaluated by solving a system
of normal equations, derived from the conditions to be satisfied so that the sum of
the squared deviations of the estimated values from the observed values, is minimum.

Method of Least Squares

Let (x1, f1), (x2, f2),..., (xn, fn) be a set of observed values and g(x) be the
Self - Learning approximating function. We form the sums of the squares of the deviations of the
102 Material observed values fi from the estimated values g (xi),
 n
Interpolation and
S    f i  g ( xi ) Curve Fitting
2
i.e.,
i 1

(2.63)
The function g(x) may have some parameters,  ,  ,  . In order to determine NOTES
these parameters we have to form the necessary conditions for S to be minimum,
which are
S S S
 0,  0, 0 (2.64)
  

These equations are called normal equations, solving which we get the
parameters for the best approximate function g(x).
Curve Fitting by a Straight Line: Let g ( x)     x, be the straight line which
fits a set of observed data points (xi, yi), i = 1, 2, ..., n.
Let S be the sum of the squares of the deviations g(xi) – yi, i = 1, 2,...,n; given
by,
n
S     x  y 
i 1
i i
2
(2.65)

We now employ the method of least squares to determine  and  , so that S

will be minimum. The normal equations are,
S n


 0, i.e.,  (  
i 1
xi  yi )  0 (2.66)

S n

and, 
 0, i.e.,  x (  
i 1
i xi  yi )  0 (2.67)

These conditions give,

n  S1   S01  0
S1  S 2   S11  0

n n n n
S1   xi , S01   yi , S 2   xi , S11   xi yi
2
where,
i 1 i 1 i 1 i 1

Solving,
  1 S S
  . Also   01   1 .
 S1 S11  S1 S2 nS11  S1S01 nS2  S12
n n
Algorithm. Fitting a straight line y = a + bx.
Step 1. Read n [n being the number of data points]
Step 2. Initialize : sum x = 0, sum x2 = 0, sum y = 0, sum xy = 0
Step 3. For j = 1 to n compute
Begin
Read data xj, yj
Self - Learning
Compute sum x = sum x + xj Material 103
Interpolation and Compute sum x2 + xj × xj
Curve Fitting
Compute sum y = sum y + yi × yj
Compute sum xy = sum xy + xj × yj
NOTES End
Step 4. Compute b = (n × sum xy – sum x × sum y)/ (n × sum x2 – (sum x)2)
Step 5. Compute x bar = sum x / n
Step 6. Compute y bar = sum y / n
Step 8. Compute a = y bar – b × x bar
Step 9. Write a, b
Step 10. For j = 1 to n
Begin
Compute y estimate = a + b × x
write xj, yj, y estimate
End
Step 11. Stop
Curve Fitting by a Quadratic (A Parabola): Let g(x) = a + bx + cx2, be the
approximating quadratic to fit a set of data (xi, yi), i = 1, 2, ..., n. Here the
parameters are to be determined by the method of least squares, i.e., by minimizing
the sum of the squares of the deviations given by,
n
S  (a  bx  cx
i 1
i
2
i  yi ) 2

(2.68)
Thus the normal equations,  S  0,  S  0,  S  0, are as follows:
a b c
(2.69)
n

 (a  bx  cx
i 1
i
2
i
 yi )  0

 x (a  bx  cx
i 1
i i
2
i  yi )  0

x
i 1
2
i (a  bxi  cxi2  yi )  0. (2.70)

These equations can be rewritten as,

na  s1b  s2 c  s01  0
s1a  s2b  s3 c  s11  0
s2 a  s3b  s4 c  s21  0 (2.71)
n n n n
Where, s1  x ,
i 1
i
s2  x
i 1
2
i
, s3  x ,
i 1
3
i
s4  x
i 1
4
i

n n n

Self - Learning
s01  y ,
i 1
i
s11  x
i 1
i
yi , s 21  x
i 1
2
y
i i
(2.72)
104 Material
 It is clear that the normal equations form a system of linear equations in the Interpolation and
Curve Fitting
unknown parameters a, b, c. The computation of the coefficients of the normal
equations can be made in a tabular form for desk computations as shown below,

2 3 4 2 NOTES
x xi yi xi xi xi xi y i xi y i
2 3 4 2
1 x1 y1 x1 x1 x1 x1 y1 x1 y1
2 3 4 2
2 x2 y2 x2 x2 x2 x2 y 2 x2 y 2
... ... ... ... ... ... ... ...
2 3 4 2
n xn yn xn xn xn xn y n xn y n
sum s1 s01 s2 s3 s4 s11 s 21

The system of linear equations can be solved by Gaussian elimination method.

It may be noted that number of normal equations is equal to the number of unknown
parameters.
Example 2.52. Find the straight line fitting the following data:

xi 4 6 8 10 12
y1 13.72 12.90 12.01 11.14 10.31

Solution: Let y = a + bx, be the straight line which fits the data. We have the

normal equations  S  0,  S  0 for determining a and b, where

a b

5
S  ( y  a  bx )
i 1
i i
2

5 5

Thus, y
i 1
i  na  b xi  0
i 1

5 5 5

x y  a  xi  b xi  0
2
and i i
i 1 I 1 i 1

The coefficients are computed in the table below,

xi yi xi2 xi yi
4 13.72 16 54.88
6 12.90 36 77.40
8 12.01 64 96.08
10 11.14 100 111.40
12 10.31 144 123.72
Sum 40 60.08 360 463.48

Thus the normal equations are,

5a  40b  60.08  0
40a  360b  463.48  0
Solving these two equations we obtain,
Self - Learning
a = 15.448, b = 0.429 Material 105
Interpolation and Thus y = g(x) = 15.448 – 0.429 x, is the straight line fitting the data.
Curve Fitting
Example 2.53. Use the method of least square approximation to fit a straight line
to the following observed data.
NOTES xi 60 61 62 63 64
yi 40 40 48 52 55

Solution: Let the straight line fitting the data be y = a + bx. The data values being
large, we can use a change in variable by substituting u = x – 62, and v = y – 48.
Let v = A + B u, be a straight line fitting the transformed data, where the
normal equations for A and B are,
5 5


i 1
vi  5 A  B u
i 1
i

5 5 5


i 1
u i vi  A i 1
ui  B u
i 1
2
i

The computation of the various sums are given in the table below,

xi yi ui vi ui vi ui2
60 40 –2 –8 16 4
61 42 1 6 6 1
62 48 0 0 0 0
63 52 1 4 4 1
64 55 2 7 14 4
Sum 0 3 40 10

Thus the normal equation are,

3  5 A and 40  10 B
3
 A   , and B  4
5
This gives the line, v  3 / 5  4u
or, 20u  5v  3  0.
Transforming we get the line,
20 (x – 62) – 5 (y – 48) – 3 = 0
or, 20 x – 5y – 1003 = 0
Curve Fitting with an Exponential Curve: We consider a two parameter
exponential curve as
y  ae bx (2.73)
For determining the parameters, we can apply the principle of least square by
first using a transformation,
z = log y, so that Equation (2.73) is rewritten as, (2.74)
Self - Learning
i.e., z = log a – bx (2.75)
106 Material
 Interpolation and
Thus we have to fit a linear curve of the form z     x, with z – x variables Curve Fitting
and then get the parameters a and b as,
a  e , b    (2.76)
Thus proceeding as in linear curve fitting, NOTES

n n n
n 
i 1
xi log yi    log y
i 1
xi
i 1
i
 2
n  n  (6.77)
n 
i 1
xi2  
 xi 
 i 1 


 n
  n

and,   z  px , where x    xi  n , z    log yi  n (6.78)
 i 1   i 1 
After computing  and  , we can determine a and b given by equation
(2.75). Finally, the exponential curve fitting the data set is given by equation (2.73).
Algorithm. To fit a straight line for a given set of data points by least square
error method.
Step 1. Read the number of data points, i.e., n
Step 2. Read values of data-points, i.e., Read (xi, yi) for i = 1, 2,..., n
Step 3 Initialize the sums to be computed for the normal equations,
i.e., sx = 0, sx2 = 0, sy = 0, syx = 0
Step 4. Compute the sums, i.e., For i = 1 to n do
Begin
sx  sx  xi
sx 2  sx 2  xi2
sy  sy  yi
syx  syx  xi yi
End
Step 5. Solve the normal equations, i.e., solve for a and b of the line y = a +
bx
Compute d  n  sx 2  sx  sx
b  (n  sxy  sy  sx) / d
xbar  sx / n
ybar  sy / n
a  ybar  b  x bar
Step 6. Print values of a and b
Step 7. Print a table of values of xi , yi , y pi  a  bxi for i  1, 2, ..., n
Step 8. Stop
Algorithm. To fit a parabola y  a  bx  cx 2 , for a given set of data points by
Self - Learning
least square error method. Material 107
Interpolation and Step 1. Read n, the number of data points
Curve Fitting
Step 2. Read (xi, yi) for i = 1, 2,..., n; the values of data points
Step 3. Initialize the sum to be computed for the normal equations,
i.e., sx = 0, sx2 = 0, sx3 = 0, sx4 = 0, sy = 0, sxy = 0.
NOTES
Step 4. Compute the sums, i.e., For i = 1 to n do
Begin
sx  sx  xi
x 2  xi  xi
sx 2  sx 2  x 2
sx 3  sx 3  xi  x 2
sx 4  sx 4  x 2  x 2
sy  sy  yi
sxy  sxy  xi  yi
sx 2 y  sx 2 y  x 2  yi
End
Step 5. Form the coefficients {aij } matrix of the normal equations, i.e.,

a11  n, a21  sx, a31  sx 2

a12  sx, a22  sx 2 , a32  sx3
a13  sx 2 , a23  sx 3 , a33  sx 4

Step 6. Form the constant vector of the normal equations.

b1 = sy, b2 = sxy, b3 = sx2y
Step 7. Solve the normal equation by Gauss-Jordan method
a12  a12 / a11 , a13  a13 / a11 , b1  b1 / a11
a 22  a 22  a 21a12 , a23  a 23  a21a13
b2  b2  b1a21

a32  a32  a31a12

a33  a33  a31a13
b3  b3  b1a31
a 23  a23 / a 22
b2  b2 / a 22
a33  a33  a23 a32
b3  b3  a32 b2
c  b3 / a33
b  b2  c a23
a  b1  b a12  c a13
Step 8. Print values of a, b, c (the coefficients of the parabola)
Step 9. Print the table of values of xk , yk and y pk where y pk  a  bxk  cx 2 k ,
i.e., print xk , y k , y pk for k  1, 2,..., n.
Self - Learning
Step 10. Stop.
108 Material
 Interpolation and
Curve Fitting
Check Your Progress
9. When does an error arise in function interpolation?
10. How is approximating function found in the method of least squares? NOTES
11. Define the term first quadrant.
12. List the two methods used for estimation.
13. What are the reasons behind squaring each terms in the least square
method?

2.6 ANSWERS TO ‘CHECK YOUR PROGRESS’

1. In this method, we successively generate interpolating polynomials of any
degree by iteratively using linear interpolating functions.
2. It can be stated explicitly as ‘given a set of (n + 1) values y0, y1, y2,..., yn for
x = x0, x1, x2, ..., xn respectively. The problem of interpolation is to compute
the value of the function y = f (x) for some non-tabular value of x.’
3. Lagrange’s interpolation is useful for unequally spaced tabulated values.
4. For interpolation of an unknown function when the tabular values of the
argument x are equally spaced, we have two important interpolation
formulae, viz.,
(a) Newton’s forward difference interpolation formula
(b) Newton’s backward difference interpolation formula
5. The shift operator is denoted by E and is defined by E f (x) = f (x + h).
6. The Newton’s forward difference interpolation formula is a polynomial of
degree less than or equal to n.
7. The interpolating polynomials are usually used for finding values of the
tabulated function y = f(x) for a value of x within the table. But they can also
be used in some cases for finding values of f(x) for values of x near to the
end points x0 or xn outside the interval [x0, xn]. This process of finding
values of f(x) at points beyond the interval is termed as extrapolation.
8. The problem of inverse interpolation in a table of values of y = f (x) is to find
the value of x for a given y.
9. There are situations where interpolation for approximating function may not
be efficacious procedure. Errors will arise when the function values
f(xi), i = 1, 2, …, n are observed data and not exact.
10. Let (x1, f1), (x2, f2), ..., (xn, fn) be a set of observed values and g(x) be the
approximating function. We form the sums of the squares of the deviations
of the observed values fi from the estimated values g(xi),
n
S    fi  g ( xi )
2
i.e.,
i 1

11. If the revolving line OP is in the first quadrant, then all the sides of the
triangle OPM are positive. Therefore, all the trigonometrical ratios are positive Self - Learning
in the first quadrant. Material 109
Interpolation and 12. The following are the two methods used for estimation:
Curve Fitting
(i) Scatter diagram method
(ii) Least squares method
NOTES 13. Squaring each term accomplishes two purposes, viz., (i) It magnifies
(or penalizes) the larger errors, and (ii) It cancels the effect of the positive
and negative values (since a negative error when squared becomes positive).

2.7 SUMMARY
 The problem of interpolation is very fundamental problem in numerical
analysis.
 In numerical analysis, interpolation means computing the value of a function
f (x) in between values of x in a table of values.
 Lagrange’s interpolation is useful for unequally spaced tabulated values.
 For interpolation of an unknown function when the tabular values of the
argument x are equally spaced, we have two important interpolation
formulae, viz., Newton’s forward difference interpolation formula and
Newton’s backward difference interpolation formula.
 The forward difference operator is defined by, f ( x)  f ( x  h)  f ( x) .
 The backward difference operator is defined by, f ( x)  f ( x  h)  f ( x) .
 We define different types of finite differences such as forward differences,
backward differences and central differences, and express them in terms of
operators.
 The shift operator is denoted by E and is defined by E f (x) = f (x + h).
 The first order difference of a polynomial of degree n is a polynomial of
degree n–1. For polynomial of degree n, all other differences having order
higher than n are zero.
 Newton’s forward difference interpolation formula is generally used for
interpolating near the beginning of the table while Newton’s backward
difference formula is used for interpolating at a point near the end of a table.
 In iterative linear interpolation, we successively generate interpolating
polynomials, of any degree, by iteratively using linear interpolating functions.
 The process of finding values of a function at points beyond the interval is
termed as extrapolation.
 Horner’s method of synthetic substitution is used for evaluating the values
of a polynomial and its derivatives for a given x.
 Descarte’s rule is used to determine the number of negative roots by finding
the number of changes of signs in pn(–x).
 By using the method of least squares, noisy function values are used to
generate a smooth approximation. This smooth approximation can then be
used to approximate the derivative more accurately than with exact
Self - Learning
110 Material
polynomial interpolation.
  The term ‘Regression’ was first used in 1877 by Sir Francis Galton who Interpolation and
Curve Fitting
made a study that showed that the height of children born to tall parents will
tend to move back or ‘Regress’ toward the mean height of the population.

NOTES
2.8 KEY TERMS
 Interpolation: Interpolation means computing the value of a function f(x)
in between values of x in a table of values.
 Extrapolation: The process of finding values of a function at points beyond
the interval is termed as extrapolation.
 Newton-Raphson method: Newton-Raphson method is a widely used
numerical method for finding a root of an equation f (x) = 0, to the desired
accuracy.
 First quadrant: If the revolving line OP is in the first quadrant, then all the
sides of the triangle OPM are positive. Therefore, all the trigonometrical
ratios are positive in the first quadrant.
 Scatter diagram: A diagram representing two series with the known
variable, i.e., independent variable plotted on the X-axis and the variable to
be estimated, i.e., dependent variable to be plotted on the Y-axis on a
graph paper.

2.9 SELF-ASSESSMENT QUESTIONS AND

EXERCISES

Short-Answer Questions
1. What is the significance of polynomial interpolation?
2. Define the symbolic operators E and D.
3. What is the degree of the first order forward difference of a polynomial of
degree n?
4. What is the degree of the nth order forward difference of a polynomial of
degree n?
5. Write newton’s forward and backward difference formulae.
6. State an application of iterative linear interpolation.
7. What is the advantage of extrapolation?
8. State Lagrange’s formula for inverse interpolation.
9. How many roots are there in a polynomial equation of degree n?
10. How many positive real roots are there in a polynomial equation?
11. Define the term first quadrant.
12. List the basic precautions and limitations of regression and correlation
analyses.
13. Differentiate between scatter diagram and least square method. Self - Learning
Material 111
Interpolation and Long-Answer Questions
Curve Fitting
1. Use Lagrange’s interpolation formula to find the polynomials of least degree
which attain the following tabular values:
NOTES x 2 1 2
(a) y 25 8 15

x 0 1 2 5
(b) y 2 3 12 147

x 1 2 3 4
(c) y 1 1 1 5

2. Form the finite difference table for the given tabular values and find the
values of:
(a)  f (2)
(b)  f 2(1)
(c)  f 3(0)
(d)  f 4(1)
(e) f (5)
(f) f (3)
x 0 1 2 3 4 5 6
f ( x) 3 4 13 36 79 148 249

3. How are the forward and backward differences in a table related? Prove
the following:
(a) yi  yi 1
(b)  2 yi   2 yi  2
(c)  n yi   n yi  n
4. Describe Newton’s forward and backward difference formulae using
illustrations.
5. Explain iterative linear interpolation with the help of examples.
6. Illustrate inverse interpolation procedure.
7. Use the method of least squares to fit a straight line for the following data
points:

x 1 0 1 2 3 4 5 6
y 10 9 7 5 4 3 0  1

8. Discuss about the trigonometric function with the help of giving examples.
9. What is regression analysis? What are the assumptions in it?
10. Explain scatter diagram and the least square method in detail. Also, mention
how scatter diagram helps in studying correlation between two variables.

Self - Learning
112 Material
 Interpolation and
2.10 FURTHER READING Curve Fitting

Chance, William A. 1969. Statistical Methods for Decision Making. Illinois:

Richard D Irwin. NOTES
Chandan, J.S., Jagjit Singh and K.K. Khanna. 1995. Business Statistics. New
Delhi: Vikas Publishing House.
Elhance, D.N. 2006. Fundamental of Statistics. Allahabad: Kitab Mahal.
Freud, J.E., and F.J. William. 1997. Elementary Business Statistics – The
Modern Approach. New Jersey: Prentice-Hall International.
Goon, A.M., M.K. Gupta, and B. Das Gupta. 1983. Fundamentals of Statistics.
Vols. I & II, Kolkata: The World Press Pvt. Ltd.
Gupta, S.C. 2008. Fundamentals of Business Statistics. Mumbai: Himalaya
Publishing House.
Kothari, C.R. 1984. Quantitative Techniques. New Delhi: Vikas Publishing
House.
Levin, Richard. I., and David. S. Rubin. 1997. Statistics for Management. New
Jersey: Prentice-Hall International.
Meyer, Paul L. 1970. Introductory Probability and Statistical Applications.
Massachusetts: Addison-Wesley.
Gupta, C.B. and Vijay Gupta. 2004. An Introduction to Statistical Methods,
23rd Edition. New Delhi: Vikas Publishing House Pvt. Ltd.
Hooda, R. P. 2013. Statistics for Business and Economics, 5th Edition. New
Delhi: Vikas Publishing House Pvt. Ltd.
Anderson, David R., Dennis J. Sweeney and Thomas A. Williams. Essentials of
Statistics for Business and Economics. Mumbai: Thomson Learning, 2007.
S.P. Gupta. 2021. Statistical Methods. Delhi: Sultan Chand and Sons.

Self - Learning
Material 113
 Numerical Differentiation

UNIT 3 NUMERICAL and Integration

DIFFERENTIATION
AND INTEGRATION NOTES

Structure
3.0 Introduction
3.1 Objectives
3.2 Numerical Differentiation Formulae
3.2.1 Differentiation Using Newton’s Forward Difference Interpolation Formula
3.2.2 Differentiation Using Newton’s Backward Difference Interpolation
Formula
3.3 Numerical Integration Formule
3.3.1 Simpson’s One-Third Rule
3.3.2 Weddle’s Formula
3.3.3 Errors in Itegration Formulae
3.3.4 Gaussian Quadrature
3.4 Solving Numerical
3.4.1 Taylor Series Method
3.4.2 Euler’s Method
3.4.3 Runge-Kutta Methods
3.4.4 Higher Order Differential Equations
3.5 Answers to ‘Check Your Progress’
3.6 Summary
3.7 Key Terms
3.8 Self-Assessment Questions and Exercises
3.9 Further Reading

3.0 INTRODUCTION
In numerical analysis, numerical differentiation is the process of finding the numerical
value of a derivative of a given function at a given point. It is the process of
computing the derivatives of a function f(x) when the function is not explicitly
known, but the values of the function are known only at a given set of arguments
x = x0, x1, x2,..., xn. For finding the derivatives, a suitable interpolating polynomial
is used and then its derivatives are used as the formulae for the derivatives of the
function. Thus, for computing the derivatives at a point near the beginning of an
equally spaced table, Newton’s forward difference interpolation formula is used,
whereas Newton’s backward difference interpolation formula is used for computing
the derivatives at a point near the end of the table.
Numerical integration constitutes a broad family of algorithms for calculating
the numerical value of a definite integral. The numerical computation of an integral
is sometimes called quadrature. The most straightforward numerical integration
technique uses the Newton-Cotes formulas, which approximate a function tabulated
at a sequence of regularly spaced intervals by various degree polynomials. If the
functions are known analytically instead of being tabulated at equally spaced
intervals, the best numerical method of integration is called Gaussian quadrature. Self - Learning
Material 115
Numerical Differentiation The basic problem considered by numerical integration is to compute an
and Integration
b
approximate solution to a definite integral a f ( x) dx. If f(x) is a smooth well performed
function integrated over a small number of dimensions and the limits of integration
NOTES are bounded then there are many methods of approximating the integral with
arbitrary precision. Numerical integration methods can generally be described as
combining evaluations of the integrand to get an approximation to the integral. The
integrand is evaluated at a finite set of points called integration points and a weighted
sum of these values is used to approximate the integral. The integration points and
weights depend on the specific method used and the accuracy required from the
approximation. Modern numerical integrations methods based on information theory
have been developed to simulate information systems such as computer controlled
systems, communication systems, and control systems.
An ordinary differential equation is a relation that contains functions of only
one independent variable and one or more of their derivatives with respect to that
variable. Ordinary differential equations are distinguished from partial differential
equations, which involve partial derivatives of functions of several variables.
Ordinary differential equations arise in many different contexts including geometry,
mechanics, astronomy and population modelling. The Picard—Lindelöf theorem,
Picard’s existence theorem or Cauchy–Lipschitz theorem is an important theorem
on existence and uniqueness of solutions to first-order equations with given initial
conditions. The Picard method is a way of approximating solutions of ordinary
differential equations. Originally, it was a way of proving the existence of solutions.
It is only by advanced symbolic computing that it has become a practical way of
approximating solutions. Euler’s method is a first-order numerical procedure for
solving ordinary differential equations with a given initial value. It is the most basic
kind of explicit method for numerical integration of ordinary differential equations
and is the simplest kind of Runge-Kutta method.
In this unit, you will learn about the numerical differentiation formulae,
Simpson’s rule, errors in integration formulae, Gaussian quadrature formulae,
solving numerical differential equation, Euler’s method, Taylor series method,
Runge-Kutta method and higher order differential equation.

3.1 OBJECTIVES
After going through this unit, you will be able to:
 Describe numerical differentiation
 Differentiate using Newton’s forward difference interpolation formula
 Differentiate using Newton’s backward difference interpolation formula
 Describe numerical integration
 Identify the numerical methods for evaluating a definite integral
 Know Newton-Cotes general quadrature
 Understand Simpson’s one-third and three-eighth rule
Self - Learning  Explain interval halving technique
116 Material
  Numerically evaluate double integrals solution of non-linear equation Numerical Differentiation
and Integration
 Define Picard’s method of successive approximation
 Describe Euler’s method and Taylor series method
 Explain Runge-Kutta and multistep methods NOTES
 Understand predictor-corrector methods
 Find numerical solution of boundary value problems
 Define higher order differential equations

3.2 NUMERICAL DIFFERENTIATION

FORMULAE
Numerical differentiation is the process of computing the derivatives of a function
f(x) when the function is not explicitly known, but the values of the function are
known only at a given set of arguments x = x0, x1, x2,..., xn. For finding the
derivatives, we use a suitable interpolating polynomial and then its derivatives are
used as the formulae for the derivatives of the function. Thus, for computing the
derivatives at a point near the beginning of an equally spaced table, Newton’s
forward difference interpolation formula is used, whereas Newton’s backward
difference interpolation formula is used for computing the derivatives at a point
near the end of the table. Again, for computing the derivatives at a point near the
middle of the table, the derivatives of the central difference interpolation formula is
used. If, however, the arguments of the table are unequally spaced, the derivatives
of the Lagrange’s interpolating polynomial are used for computing the derivatives
of the function.

3.2.1 Differentiation Using Newton’s Forward

Difference Interpolation Formula
Let the values of an unknown function y = f(x) be known for a set of equally
spaced values x0, x1, …, xn of x, where xr = x0 + rh. Newton’s forward differ-
ence interpolation formula is,
u (u  1) 2 u (u  1)(u  2) 3 u (u  1)(u  2)...(u  n  1) n
 (u )  y0  u  y0   y0   y0  ...   y0
2 ! 3 ! n !
x  x0
where u 
h
dy
The derivative can be evaluated as,
dx
dy d d du 1 d 
 { (u )}  . 
dx dx du dx h du
1 2u  1 2 3u 2  6u  2 3 2u 3  9u 2  11u  3 4 
Thus, y  ( x)   y0   y0   y0   y0  ...
h 2 6 12 
(3.1)
1
Similarly, y  ( x)    (u )
h2
Self - Learning
Material 117
Numerical Differentiation
and Integration  2 1 6u 2  18u  11 4 
Or, y ( x) 
  y0  (u  1)  3
y0   y0  ... (3.2)
 h2 12 
For a value of x near the beginning of a table, u = (x – xo)/h is computed first
NOTES and then Equations (3.1) and (3.2) can be used to compute f ( x) and f ( x). At
the tabulated point x0, the value of u is zero and the formulae for the derivatives
are given by,
1 1 1 1 1 
y ( x0 )  y 0  2 y 0  3 y0  4 y0  5 y0  ... (3.3)
h  2 3 4 5 
1  2 11 5 
y ( x0 )  2 
 y0  3 y0  4 y 0  5 y0  ... (3.4)
h  12 6 

3.2.2 Differentiation Using Newton’s Backward

Difference Interpolation Formula
For an equally spaced table of a function, Newton’s backward difference
interpolation formula is,
v(v  1) 2 v (v  1)(v  2) 3 v(v  1)(v  2)(v  3) 4
 ( v )  yn  v y n   yn   yn   yn  ...
2 ! 3 ! 4 !
v(v  1)...(v  n  1) n
  yn
n !
x  xn
where v 
h
dy d2y
The derivatives and 2 , obtained by differentiating the above formula
dx dx
are given by,
dy 1  2v  1 2 3v 2  6v  2 3 2v 3  9v 2  11v  3 4 
 y n   yn   yn   y n  ...
dx h  2 6 12 
(3.5)
d2y 1  2 3 6v 2  18v  11 4 
  y n  ( v  1)  y n   y n  ... (3.6)
dx 2 2
h  12 

dy d2y
For a given x near the end of the table, the values of and are com-
dx dx 2
puted by first computing v = (x – xn)/h and using the above formulae. At the
tabulated point xn, the derivatives are given by,
1 1 1 1 
y ( xn )  y n   2 y n   3 y n   4 y n  ...
h  2 3 4 
(3.7)
1  2 11 5 
y ( xn )  2 
 y n   3 y n   4 y n   5 y n  ...
h  12 6 
(3.8)
Example 3.1: Compute the values of f (2.1), f (2.1), f (2.0) and f (2.0) when
f (x) is not known explicitly, but the following table of values is given:
Self - Learning
118 Material
 x f(x) Numerical Differentiation
and Integration
2.0 0.69315
2.2 0.78846
2.4 0.87547
Solution: Since the points are equally spaced, we form the finite difference table. NOTES

x f ( x) f ( x) 2 f ( x)
2.0 0.69315
9531
2.2 0.78846  83
8701
2.4 0.87547

For computing the derivatives at x = 2.1, we have

1 2u  1 2 1 2
 [f 0 
f  ( x)   f 0 ] and f  ( x)  2  f0
h 2 h
x  x0 2.1  2.0
u   0.5
h 0.2
1  2  0.5  1 2 
 f  (2.1)   0.09531   f 0   0.4765
0.2  2 
1
f  (2.1)   ( 0.00083)  0.21
(0.2) 2

The value of f (2.0) is given by,,

1  1 
f (2.0)   f 0   2 f 0 
0.2  2 
1  1 
  0.09531   0.00083
0.2  2 
0.09572
  0.4786
0.2
1
f (2.0)   ( 0.0083)
(0.2) 2
 0.21
Example 3.2: For the function f(x) whose values are given in the table below
compute values of f (1), f (1), f (5.0), f (5.0).

x 1 2 3 4 5 6
f ( x) 7.4036 7.7815 8.1291 8.4510 8.7506 9.0309

Solution: Since f(x) is known at equally spaced points, we form the finite differ-
ence table to be used in the differentiation formulae based on Newton’s interpo-
lating polynomial.

Self - Learning
Material 119
Numerical Differentiation 2 3 4 5
and Integration x f ( x) f ( x )  f ( x)  f ( x)  f ( x)  f ( x)
1 7.4036
0.3779
NOTES 2 7.7815  303
0.3476 46
3 8.1291  257  12
0.3219 34 8
4 8.4510  223 4
0.2996 30
5 8.7506  193
0.2803
6 9.0309

To calculate f (1) and f (1), we use the derivative formulae based on Newton’ss
forward difference interpolation at the tabulated point given by,

1 1 1 1 1 
f ( x0 )   f 0   2 f 0   3 f 0   4 f 0   5 f 0 
h 2 3 4 5 
 2
1 11 4 5 5 
f ( x0 )    f 0   f 0  12  f 0  6  f 0 
3

h2
1 1 1 1 1 
 f (1)  0.3779   (0.0303)   0.0046   (0.0012)   0.0008
1 2 3 4 5 
 0.39507
 11 5 
f (1)   0.0303  0.0046   (0.0012)   0.0008
 12 6 
 0.0367

Similarly, for evaluating f (5.0) and f (5.0), we use the following formulae
1 1 1 1 1 
f ( xn )   f n   2 f n   3 f n   4 f n   5 f n 
h 2 3 4 5 
1  2 11 5 
f ( xn )  2 
 f n  3 f n   4 f n  5 f n 
h  12 6 
 1 1 1 
f (5)  0.2996  (0.0223)   0.0034  (0.0012)
 2 3 4 
 0.2893
11
f (5)  [0.0223  0.0034   0.0012]
12
 0.0178

Example 3.3: Compute the values of y (0), y (0.0), y (0.02) and y (0.02) for the
function y = f(x) given by the following tabular values:
x 0.0 0.05 0.10 0.15 0.20 0.25
y 0.00000 0.10017 0.20134 0.30452 0.41075 0.52110

Solution: Since the values of x for which the derivatives are to be computed lie
near the beginning of the equally spaced table, we use the differentiation formulae
based on Newton’s forward difference interpolation formula. We first form the
Self - Learning
120 Material
finite difference table.
 Numerical Differentiation
x y y 2 y 3 y 4 y
and Integration
0. 0 0.00000
0.10017
0.05 0.10017 100
0.10117 101
NOTES
0.10 0.20134 201 3
0.10318 104
0.15 0.30452 305 3
0.10623 107
0.20 0.41075 412
0.11035
0.25 0.52110

For evaluating y  (0,0), we use the formula

1 1 1 1 
y ( x0 )   y0   2 y0   3 y0   4 y0 
h 2 3 4 
1  1 1 1 
 y(0.0)   0.10017   0.00100   0.00101   0.00003 
0.05  2 3 4 
 2.00000
For evaluating y  (0,0), we use the formula
1 2 11 4 
y ( x0 )    y0   y0   y0 
3

h2  12 
1  11 
  0.00100  0.00101   0.00003 
(0.05) 2  12 
 0.007

For evaluating y  (0.02) and y  (0.02), we use the following formulae, with
0.02  0.00
u  0.4
0.05
1 2u  1 2 3u 2  6u  2 3 2u 3  9u 2  11u  3 4 
y (0.02)   y0   y0   y0   y0 
h 2 6 12 
1  2 6(u  1) 3 6u 2  18u  11 4 
y (0.02)    y0   y0    y0 
h2  6 12 
1  2  0.4  1 3  (0.4) 2  6  0.4  2
 y (0.02)  0.10017   0.00100   0.00101
0.05  2 6
2  0.43  9  0.42  11  0.4  3 
  0.00003
12 
 4.00028
1  6  0.16  18  0.4  11 
y (0.02)  0.00100  0.00101  (0.6)   0.00003
(0.05)2  12 
 0.800

Example 3.4: Compute f (6.0) and f (6.3) by numerical differentiation formulae

for the function f(x) given in the following table.

x 6.0 6.1 6.2 6.3 6.4

f ( x)  0.1750  0.1998  0.2223  0.2422  0.2596

Self - Learning
Material 121
Numerical Differentiation Solution: We first form the finite difference table,
and Integration
x f ( x) f ( x) 2 f ( x) 3 f ( x)
6.0  0.1750
 248
NOTES
6.1  0.1998 23
 225 3
6.2  0.2223 26
 199 1
6.3  0.2422 25
 174
6.4  0.2596

For evaluating f (6.0) , we use the formula derived by differentiating Newton’ss

forward difference interpolation formula.

1 1 1 
f ( x0 )   f 0   2 f 0   3 f 0 
h 2 3 
1  1 1 
 f (6.0)   0.0248   0.0023   0.0003
0.1  2 3 
 10[0.0248  0.00115  0.0001]
 0.2585
For evaluating f (6.3), we use the formula obtained by differentiating Newton’ss
backward difference interpolation formula. It is given by,
1 2
f ( xn )  2
[ f n  3 f n ]
h
 1
f (6.3)  [0.0026  0.0003]  0.29
(0.1) 2
Example 3.5: Compute the values of y (1.00) and y (1.00) using suitable numeri-
cal differentiation formulae on the following table of values of x and y:

x 1.00 1.05 1.10 1.15 1.20

y 1.0000 1.02470 1.04881 1.07238 1.09544

Solution: For computing the derivatives, we use the formulae derived on differ-
entiating Newton’s forward difference interpolation formula, given by
1 1 1 1 
f ( x0 )  y0  2 y0  3 y0  4 y0  ...
h  2 3 4 
1  2 11 
f ( x0 )  2 
 y0  3 y0  4 y 0  ...
h  12 
Now, we form the finite difference table.

Self - Learning
122 Material
 Numerical Differentiation
x y y 2 y 3 y 4 y and Integration
1.00 1.00000
2470
1.05 1.02470  59
NOTES
2411 5
1.10 1.04881  54 2
2357 3
1.15 1.07238  51
2306
1.20 1.09544

Thus with x0 = 1.00, we have

1  1 1 1 
y (1.00)   0.02470   0.00059   0.00005   0.00002 
0.05  2 3 4 
 0.502

1 11 
y (1.00)    0.00059  0.00005   0.00002 
2
(0.05)  12 
 0.26
Example 3.6: Using the following table of values, find a polynomial representa-
tion of f (x) and then compute f (0.5).

x 0 1 2 3
f ( x) 1 3 15 40

Solution: Since the values of x are equally spaced we use Newton’s forward
difference interpolating polynomial for finding f ( x) and f (0.5). We first form the
finite difference table as given below:
x f ( x ) f ( x) 2 f ( x) 3 f ( x )
0 1
2
1 3 10
12 3
2 15 13
25
3 40

x  x0
Taking x0  0, we have u   x. Thus the Newton’s forward difference
h
interpolation gives,
u (u  1) 2 u (u  1) (u  2) 3
f  f 0  uf 0   f0   f0
2! 3!
x ( x  1) x ( x  1) ( x  2)
i.e., f ( x)  1  2 x   10  3
2 6
or, 13 2 1 3
f ( x)  1  3 x  x  x
2 2
 3 2
f ( x )  3  13 x  x
2
3
and, f (0.5)  3  13  0.5   (0.5) 2  3.12 Self - Learning
2
Material 123
Numerical Differentiation Example 3.7: The population of a city is given in the following table. Find the rate
and Integration
of growth in population in the year 2001 and in 1995.
Year x 1961 1971 1981 1991 2001
NOTES Population y 40.62 60.80 79.95 103.56 132.65

dy
Solution: Since the rate of growth of the population is , we have to compute
dx
dy
at x = 2001 and at x = 1995. For this we consider the formula for the deriva-
dx
tive on approximating y by the Newton’s backward difference interpolation given
by,

dy 1  2u  1 2 3u 2  6u  2 3 2u 3  9u 2  11u  3 4 
 y n   yn   yn   y n  ...
dx h  2 6 12 

x  xn
Where u 
h
For this we construct the finite difference table as given below:
x y y 2 y 3 y 4 y
1961 40.62
20.18
1971 60.80  1.03
19.15 5.49
1981 79.95 4.46  4.47
23.61 1.02
1991 103.56 5.48
29.09
2001 132.65

x  xn
For x = 2001, u  0
h

 dy  1  1 1 1 
    29.09   5.48   1.02   (4.47) 
  dx 2001 10  2 3 4 
 3.105

1995  1991
For x = 1995, u   0.4
10

 dy  1  1.8 3  0.16  6  0.4  2 

    23.61   4.46   5.49 
 dx 1995 10  2 6 
 3.21

3.3 NUMERICAL INTEGRATION FORMULE

The evaluation of a definite integral cannot be carried out when the integrand f (x)
is not integrable, as well as when the function is not explicitly known but only the
Self - Learning
124 Material
function values are known at a finite number of values of x. However, the value of
the integral can be determined numerically by applying numerical methods. There Numerical Differentiation
and Integration
are two types of numerical methods for evaluating a definite integral based on the
following formula.
b
NOTES
 f ( x) dx
a
(3.9)

They are termed as Newton-Cotes quadrature and Gaussian quadrature. We

first confine our attention to Newton-Cotes quadrature which is based on inte-
grating polynomial interpolation formulae. This quadrature requires a table of val-
ues of the integrand at equally spaced values of the independent variable x.

3.3.1 Newton-Cotes General Quadrature

We start with Newton’s forward difference interpolation formula which uses a
table of values of f (x) at equally spaced points in the interval [a, b]. Let the
interval [a, b] be divided into n equal sub-intervals such that,
a = x0, xi = xo+ ih, for i = 1, 2, ..., n – 1, xn = b (3.10)
so that, nh = b–a
Newton’s forward difference interpolation formula is,
s ( s  1) 2 s ( s  1)( s  2) 3 s ( s  1)( s  2)...( s  n  1) n
 ( s )  f 0  s f 0   f0   f 0  ...   f0
2! 3! n!

(3.11)
x  x0
where, s
h
Replacing f (x) by (s) in Equation (3.9), we get
xn n
 s ( s  1) 2 
 f ( x) dx  h  f
x0 0
0  s f 0 
2!
 f 0  ... ds

since when x = x0, s = 0 and x = xn, s = n and dx = h du.
Performing the integration on the RHS we have,
xn
 n2 1  n3 n 2  1  n4 n3 n2 

x0
f ( x)dx  h nf 0 
 2
f 0    2 f 0  
2  3 2  6  4
 3  2 3 f 0
3 2 

1  n 5 3n 4 11n 3   (3.12)
    3n 2 4 f 0  ...
24  5 2 3 
 
We can derive different integration formulae by taking particular values of
n = 1, 2, 3, .... Again, on replacing the differences, the Newton-Cotes formula can
be expressed in terms of the function values at x0, x1,..., xn, as
xn n

 f ( x)dx  h c
k 0
k f ( xk ) (3.13)
x0

The error in the Newton-Cotes formula is given by,

n
hn 2
En  f n 1 ()   s ( s  1)  ( s  n) ds (3.14)
(n  1)! 0
Self - Learning
Material 125
Numerical Differentiation Trapezoidal Formula of Numerical Integration
and Integration
Taking n = 1 in Equation (3.12), we get the trapezoidal formula given by,
x1
 1 
NOTES  f ( x)dx  h f
x0
0
 f
2 0 

since all other differences of higher order are absent.

Replacing f 0 by f1 – f0, we have
x1
h
 f ( x)dx  2 [ f
x0
0  f1 ] (3.15)

This is termed as trapezoidal formula of numerical integration.

This formula can be geometrically interpreted as the definite integral of the
function f (x) between the limits x0 to x1, as is approximated by the area of the
trapezoidal region bounded by the chord joining the points (x0, f0) and (x1, f1), the
x-axis and the ordinates at x = x0 and at x = x1. This is represented by the shaded
area as shown in the Figure 3.1.
Y

(x1, f1)

f(x)
(x0, f0)

O X
x0 x1

Fig. 3.1 Trapezoidal Region

Thus, the area under the curve y = f (x) is replaced by the area under the
chord joining the points.
The error in the trapezoidal formula is given by,
1
h3 h3
ET  f  ()   s ( s  1) ds   f  (), where x0    x1 (3.16)
2 0
12

Trapezoidal Rule
xn
For evaluating the integral  f ( x)dx, we have to sum the integrals for each of the
x0

sub-intervals (x0, x1), (x1, x2),..., (xn–1, xn). Thus,

xn
h
 f ( x)dx  2 [( f
x0
0  f1 )  ( f1  f 2 )  ...  ( f n 1  f n )]

xn
h (3.17)
Self - Learning
or  f ( x)dx  2 [ f
x0
0  2( f1  f 2  ...  f n 1 )  f n ]
126 Material
 This is known as trapezoidal rule of numerical integration. Numerical Differentiation
and Integration
The error in the trapezoidal rule is,
xn
h
E   f ( x)dx  2 [ f  2( f1  f 2  ...  f n 1 )  f n ]
n
T 0
x0 NOTES
h 3
 [ f (1 )  f  ( 2 )  ...  f  ( n )]
12

where x0  1  x1 , x1   2  x2 ,..., xn 1   n  xn
Thus, we can write

h3
ETn   [nf  ()], f  () being the mean of f  (1 ), f ( 2 ),..., f ( n )
12
h2
  nh f ()
12
h2
where ETn   (b  a) f (), since nh  b  a
12
or, x0    xn
b
Algorithm: Evaluation of  f ( x)dx by trapezoidal rule.
a

Step 1: Define function f (x)

Step 2: Initialize a, b, n
Step 3: Compute h = (b–a)/n
Step 4: Set x = a, S = 0
Step 5: Compute x = x + h
Step 6: Compute S = S + f (x)
Step 7: Check if x < b, then go to Step 4 else go to the next step
Step 8: Compute I = h (S + (f (a) + f (b))/2)
Step 9: Output I, n

Simpson’s One-Third Formula

Taking n = 2 in the Newton-Cotes formula in Equation (3.12), we get Simpson’s
one-third formula of numerical integration given by,
x2
 22 1 
 f ( x)dx 
x0
h  2 f 0  f 0  (2  23  3  22 ) 2 f 0 
 2 12 
 1 
 h  2 f 0  2 ( f1  f 0 )  ( f 2  2 f1  f0 )  (3.18)
 3 
x2
h
  f ( x)dx  [ f 0  4 f1  f 2 ]
x0
3

This is known as Simpson’s one-third formula of numerical integration.

The error in Simpson’s one-third formula is defined as,
x2
h
ES   f ( x) dx  3 ( f
x0
0  4 f1  f 2 ) Self - Learning
Material 127
Numerical Differentiation
and Integration
Assuming F ( x)  f ( x), we obtain:
h
E S  F ( x 2 )  F ( x0 )  ( f  4 f1  f 2 )
3 0

NOTES Expanding F(x2) = F(x0+2h), f1 = f (x0+h) and f2 = f (x0+2h) in powers of

h, we have:
(2h) 2 (2h)3
ES  2hF  ( x0 )  F  ( x) ( x0 )  F ( x0 )  ...
2! 3!
h  h2  (2h) 2 
  f 0  4  f 0  hf 0  f  (0)  ...  f 0  2hf 0  f  (0)  ...
3  2!  2! 
4 2 4 iv
 2hf 0  2h 2 f 0  h3 f  (0)  h 4 f  (0)  h5 y0 ()
3 3 15
h
 [6 f 0  6hf 0  4h 2 f  (0)  2h3 f  (0)...]
3
5
h iv
ES   f (), on simplification, where x0    x2
90
(3.19)
Geometrical interpretation of Simpson’s one-third formula is that the integral
represented by the area under the curve is approximated by the area under the
parabola through the points (x0, f0), (x1, f1) and (x2, f2) shown in Figure 3.2.
Y

Parabola y = f (x)

X
0 x0 x1 x2

Fig. 3.2 Simpson’s One-Third Integration

3.3.1 Simpson’s One-Third Rule

On dividing the interval [a, b] into 2m sub-intervals by points x0 = a, x1 = a + h,
x2 = a + 2h, ..., x2m = a+2mh, where b = x2m and h = (b–a)/(2m), and using
Simpson’s one-third formula in each pair of consecutive sub-intervals, we have
b x2 x4 x2 m


a
f ( x )dx  
x0

f ( x) dx  f ( x )dx  ... 
x2
 f ( x)dx
x2 m 2


h
3

( f 0  4 f1  f 2 )  ( f 2  4 f 3  f 4 )  ( f 4  4 f 5  f 6 )  ...  ( f 2 m  2  4 f 2 m 1  f 2 m ) 
b

 f ( x)dx  3  f 
h
0  4 ( f1  f 3  f 5  ...  f 2 m 1 )  2 ( f 2  f 4  f 6  ...  f 2 m  2 )  f 2 m .
a

This is known as Simpson’s one-third rule of numerical integration.

The error in this formula is given by the sum of the errors in each pair of
intervals as,
h5 iv
ES2 m   [ f (1 )  f iv ( 2 )  ...  f iv ( m )]
Self - Learning 90
128 Material
 Which can be rewritten as, Numerical Differentiation
and Integration
h5
ES2 m  
m f iv (), f iv () being the mean of f iv (1 ), f iv ( 2 ),..., f iv (m )
90
Since 2mh  b  a, we have NOTES
h4
ES2 m   (b  a ) f iv ( ), where a    b.
180
(3.20)
b

Algorithm: Evaluation of  f ( x)dx by Simpson’s one-third rule.

a

Step 1: Define f (x)

Step 2: Input a, b, n (even)
Step 3: Compute h = (b–a)/n
Step 4: Compute S1 = f (a) + f (b)
Step 5: Set S2 = 0, x = a
Step 6: Compute x = x + 2h
Step 7: Compute S2 = S2+ f (x)
Step 8: Check If x < b then go to Step 5 else go to next step
Step 9: Compute x = a + h
Step 10: Compute S4 = S4+ f (x)
Step 11: Compute x = x + 2h
Step 12: Check If x > b go to next Step else go to Step 9
Step 13: Compute I = (S1+ 4S4+ 2S2)h/3
Step 14: Write I, n

Simpson’s Three-Eighth Formula

Taking n = 3, Newton-Cotes formula can be written as,
x3 3
u (u  1) 2 u (u  1)(u  2) 3
 f ( x)dx  h ( f
x0 0
0  u f 0 
2!
 f0 
3!
 f 0 )du

3
 u
2
1u
3
u  2
2
1u
4
2 3

f 0     f0    u  u  f 0 
3
 h uf 0 
 2 
2 3 2  
6 4   0

 9 9 2 3 3 
 h 3 y0  y 0   y0   y0 
 2 4 8 
 9 9 3 
 h 3 y0  ( y1  y0 )  ( y 2  2 y1  y0 )  ( y3  3 y 2  3 y1  y 0 )
 2 4 8 
x3
3h
 f ( x) dx 
x0
( y  3 y1  y3 )
8 0
Self - Learning
Material 129
Numerical Differentiation (3.21)
and Integration
3h5 iv
The truncation error in this formula is  f (), x0    x3 .
80
NOTES This formula is known as Simpson’s three-eighth formula of numerical inte-
gration.
As in the case of Simpson’s one-third rule, we can write Simpson’s three-eighth
rule of numerical integration as,
b
3h
 f ( x) dx 
a
[ y  3 y1  3 y 2  2 y3  3 y 4  3 y5  2 y 6  ...  2 y3m 3  3 y3m  2  3 y3m 1  y3m ]
8 0

(3.22)
where h = (b–a)/(3m); for m = 1, 2,...
i.e., the interval (b–a) is divided into 3m number of sub-intervals.
The rule in Equation (3.22) can be rewritten as,
b
3h
 f ( x) dx 
a
8
[ y0  y3m  3 ( y1  y 2  y 4  y5  ...  y3m  2  y3m 1 )  2 ( y3  y6  ...  y3m 3 )]

(3.23)
The truncation error in Simpson’s three-eighth rule is
3h4
(b  a) f iv (), x0    xg m
240

3.3.2 Weddle’s Formula

In Newton-Cotes formula with n = 6 some minor modifications give the Weddle’s
formula. Newton-Cotes formula with n = 6, gives
x6
 123 5 41 6 
 ydx  h6 y
x0
0
 18 y0  272 y0  24 y3 y 0 
10
 y0 
140
 y0 


41 6
This formula takes a very simple form if the last term  y0 is replaced by
140
42 6 3
 y0   6 y0 . Then the error in the formula will have an additional term
140 10
1 6
 y0 . The above formula then becomes,
140
x6
 123 5 3 
 ydx  h 6 y
x0
0  18y0  27 2 y0  243 y0 
10
 y0   6 y0 
10 
 x6
3h
 ydx  10
x0
[ y0  5 y1  y2  6 y3  y4  5 y5  y6 ]

(3.24)
On replacing the differences in terms of yi’s, this formula is known as Weddle’s
Self - Learning formula.
130 Material
 Numerical Differentiation
1 7 ( vi ) and Integration
The error Weddle’s formula is  h  y ( ) (3.25)
140
Weddle’s rule is a composite Weddle’s formula, when the number of sub-
intervals is a multiple of 6. One can use a Weddle’s rule of numerical integration by NOTES
sub-dividing the interval (b – a) into 6m number of sub-intervals, m being a posi-
tive integer. The Weddle’s rule is,
b
3h
 f ( x)dx  10 [y +5y +y +6y +y +5y +2y +5y +y +6y +y
a
0 1 2 3 4 5 6 7 8 9 10
+5y11+...

+2y6m–6+5y6m–5+y6m–4+6y6m–3+y6m–2+5y6m–1+y6m] (3.26)
where, b–a = 6mh
b
3h
i.e.,  f ( x) dx  10 [ y
a
0
 y6 m  5 ( y1  y5  y7  y11  ...  y6 m 5  y6 m 1 )  y 2  y 4  y8  y10  ...

 y6 m  4  y6 m 2  6 ( y3  y9  ...  y6 m 3 )  2 ( y6  y12  ...  y  6 )]

1 6
The error in Weddle’s rule is given by  h (b  a ) y ( vi ) ()
840
(3.27)
2
Example 3.8: Compute the approximate value of  x 4 dx by taking four sub-inter-
0

vals and compare it with the exact value.

2 1
Solution: For four sub-intervals of [0, 2], we have h = = = 0.6. We tabu-
4 2
late f ( x)  x 4 .
x 0 0.5 1.0 1.5 2.0
f ( x) 0 0.0625 1.0 5.062 16.0

By trapezoidal rule, we get

2
0 .5
 x dx 
4
[0  2  (0.0625  1.0  5.062)  16.0]
2
0
1 28.2690
 [12.2690  16.0]   7.0672
4 4
By Simpson’s one-third rule, we get
2
05
 x dx 
4
[0  4  (0.0625  5.062)  2  1.0  16.0]
3
0
1 38.5380
 [4  5.135  18.0]   6.4230
6 6
2 5 32
Exact value    6 .4
5 5
Error in the result by trapezoidal rule = 6.4 – 7.0672 = – 0.6672
Error in the result by Simpson’s one third rule = 6.4 – 6.4230 = – 0.0230
Self - Learning
Material 131
Numerical Differentiation Example 3.9: Evaluate the following integral:
and Integration
1

 (4 x  3x
2
)dx by taking n = 10 and using the following rules:
0
NOTES
(i) Trapezoidal rule and (ii) Simpson’s one-third rule. (iii) Also compare them
with the exact value and find the error in each case.
Solution: We tabulate f (x) = 4x–3x2, for x = 0, 0.1, 0.2, ..., 1.0.
x 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
f ( x) 0.0 0.37 0.68 0.93 1.12 1.25 1.32 1.33 1.28 1.17 1.0

(i) Using trapezoidal rule, we have

1
0.1
 (4 x  3 x
2
) dx  0  2 (0.37  0.68  0.93  1.12  1.25  1.32  1.33  1.28  1.17)  1.0
2
0
0.1
  (18.90  1.0)  0.995
2
(ii) Using Simpson’s one-third rule, we have
1
0.1
 (4 x  3x
2
) dx  [0  4 (0.37  0.93  1.25  1.33  1.17)  2(0.68  1.12  1.32  1.28)  1.0]
3
0
0.1
 [4  5.05  2  4.40  1.0]
3
0.1
  [30.0]  1.00
3

(iii) Exact value = 1.0

Error in the result by trapezoidal rule is 0.005 and there is no error in the result
by Simpson’s one-third rule.
1
Example 3.10: Evaluate  e  x dx, using (i) Simpson’s one-third rule with 10 sub-
2

0
intervals and (ii) Trapezoidal rule.
2
Solution: (i) We tabulate values of e  x for the 11 points x = 0, 0.1, 0.2, 0.3, ....,
1.0 as given below.
2
x e x
0.0 1.00000
0.1 0.990050
0.2 0.960789
0.3 0.913931
0.4 0.852144
0.5 0.778801
0.6 0.697676
0.7 0.612626
0.8 0.527292
0.9 0.444854
1.0 0.367879
Self - Learning
132 Material 1.367879 3.740262 3.037901
 Hence, by Simpson’s one-third rule we have, Numerical Differentiation
and Integration
1
h
e
 x2
dx  [ f 0  f10  4 ( f1  f 3  f5  f 7  f 9 )  2 ( f 2  f 4  f 6  f8 )]
0 3
0.1
NOTES
 [1.367879  4  3.740262  2  3.037901]
3
0.1
 [1.367879  14.961048  6.075802]
3
2.2404729
  0.7468243  0.746824
3
(ii) Using trapezoidal rule, we get
1
h
e
 x2
dx  [ f 0  f10  2 ( f1  f 2  ...  f9 )]
0
2
0.1
 [1.367879  6.778163]
2
 0.4073021

4
Example 3.11: Compute the integral I   ( x 3  2 x 2  1)dx, using Simpson’s one-
0

third rule taking h = 1 and show that the computed value agrees with the exact
value. Give reasons for this.
Solution: The values of f (x) = x3–2x2+1 are tabulated for x = 0, 1, 2, 3, 4 as

x 0 1 2 3 4
f ( x) 1 0 1 10 33

The value of the integral by Simpson’s one-third rule is,

1 1
I [1  4  0  2  1  4  10  33]  25
3 3

44 43 1
The exact value   2   1  4  25
4 3 3
Thus, the computed value by Simpson’s one-third rule is equal to the exact
value. This is because the error in Simpson’s one-third rule contains the fourth
order derivative and so this rule gives the exact result when the integrand is a
polynomial of degree less than or equal to three.
0.5

Example 3.12: Compute  e dx by (i) Trapezoidal rule and (ii) Simpson’s one-
x

0.1

third rule and compare the results with the exact value, by taking h = 0.1.

Solution: We tabulate the values of f (x) = ex for x = 0.1 to 0.5 with spacing
h = 0.1.
Self - Learning
Material 133
Numerical Differentiation
and Integration x 0.1 0.2 0.3 0.4 0.5
x
f ( x)  e 1.1052 1.2214 1.3498 1.4918 1.6847

(i) The value of the integral by trapezoidal rule is,

NOTES
0.1
IT  [1.1052  2 (1.2214  1.3498  1.4918)  1.6487]
2
0.1
 [2.7539  2  4.0630]  0.5439
2
(ii) The value computed by Simpson’s one-third rule is,

0.1
IS  [1.1052  4 (1.2214  1.4918)  2  1.3498  1.6487]
3
0.1 0.1
 [2.7539  4  2.7132  2.6996]  [16.3063]  0.5435
3 3

Exact value = e0.5– e0.1 = 1.6487–1.1052 = 0.5435

The trapezoidal rule gives the value of the intergal with an error – 0.0004 but
Simpson’s one-third rule gives the exact value.
1
dx
Example 3.13: Compute  1  x using (i) Trapezoidal rule (ii) Simpson’s one-
0

third rule taking 10 sub-intervals. Hence, (iii) Find log e and compare it with the 2

exact value up to six decimal places.

1
Solution: We tabulate the values of f(x) = for x = 0, 0.1, 0.2,..., 1.0 as given
1 x
below:
1
x y f ( x) 
1 x
0.0 y0 1.000000
0.1 y1 0.9090909
0.2 y2 0.8333333
0.3 y3 0.1692307
0.4 y4 0.7142857
0.5 y5 0.6666667
0.6 y6 0.6250000
0.7 y7 0.5882352
0.8 y8 0.5555556
0.9 y9 0.5263157
1.0 y10 0.500000
1.500000 3.4595391 2.7281746
(i) Using trapezoidal rule, we have
1
dx h
 1 x  2 [ f
0
0  f10  2 ( f1  f 2  f 3  f 4  ...  f 0 )]

0.1
 [1.500000  2  (3.4595391  2.7281745)]
2
0.1
Self - Learning  [1.500000  12.3754272]  0.6437714.
134 Material 2
(ii) Using Simpson’s one-third rule, we get Numerical Differentiation
and Integration
1
dx h
 1 x  3 [ f
0
0  f10  4 ( f1  f 3  ...  f 9 )  2 ( f 2  f 4  ...  f 8 )]

0.1
NOTES
 [1.500000  4  3.4595391  2  2.7281745]
3
0.1 0.1
 [1.5  13.838156  5.456349]   20.794505  0.6931501
3 3
(iii) Exact value:
1
dx 0.1
1 x
0
 log e2 
3
[1.500000  4  3.4595391  2  2.7281745]

 0.6931472

The trapezoidal rule gives the value of the integral having an error 0.693147 –
0.6437714 = 0.0493758, while the error in the value by Simpson’s one-third rule
is – 0. 000029.

2
Example 3.14: Compute 0
cos d , by (i) Simpson’s rule and (ii) Weddle’ss

formula taking six sub-intervals.


Solution: Sub-division of [0, ] into six sub-intervals will have
2
 1
h .  15  0  26179. For applying the integration rules we tabulate cos .
2 6
 0 15 30 45 60 75 90
cos  1 0.98281 0.93061 0.84089 0.70711 0.50874 0
(i) The value of the integral by Simpson’s one-third rule is given by,
0.26179
IS  [1  4  (0.98281  0.84089  0.50874)  2  (0.093061  0.070711)  0)]
3
0.26179
 [1  4  2.33244  2  1.63772]
3
0.26179
  13.6052  1.18723
3
(ii) The value of the integral by Weddle’s formula is,
3
IW   0.26179 [1.05  7.45775  5.04534  0.93061  0.070711]
10
 3  0.026179 [14.554411]  1.143059  1.14306

2
Example 3.15: Evaluate the integral  1  0.162sin 2  d  by Weddle’s formula.
0

Solution: On dividing the interval into six sub-intervals, the length of each sub-
1 
interval will be h    0  26179  15. For computing the integral by Weddle’ss
6 2
Self - Learning
formula, we tabulate f ()  1  0 162sin 2  . Material 135
Numerical Differentiation
and Integration  0 15 30 45 60 75 90
f ( ) 1.0 0.99455 0.97954 0.95864 0.93728 0.92133 0.91542

The value of the integral by Weddle’s formula is given by,

NOTES
3  0.26179
IW  [1.0  5 (0.99455  0.92133)  0.97954  6  0.95864  0.93728  0.91542]
10
 0.078537  19.16348  1.50504

3.3.3 Errors in Itegration Formulae

For evaluating a definite integral correct to a desired accuracy, one has to make a
suitable choice of the value of h, the length of sub-interval to be used in the formula.
There are two ways of determining h, by considering the truncation error in the
formula to be used for numerical integration or by successive evaluation of the
integral by the technique of interval halving and comparing the results.
Truncation Error Estimation Method
In the truncation error estimation method, the value of h to be used is determined
by considering the truncation error in the formula for numerical integration. Let E
be the error tolerance for the integral to be evaluated. Then h is chosen by using
the condition,
R  /2
2
dx
As an illustration, consider the evaluation of x
1
using Simpson’s one-third

rule accurate up to the third decimal place. We may take   103.

If we wish to use Simpson’s one-third rule, then the truncation error is R,
h4
R (2  1) f iv ( ); 1   2
180
Then h is determined by satisfying the condition,
h4
| f iv () | 0  5  103
180
1 iv 2  3 4
For the given problem, f (x)  , thus f ( x )  . Hence,
x x5

max f iv ( x)  24
[1, 2 ]

1  24
4
Thus, h   0  5  10 3 or h < 0.102
180
But h has to be so chosen such so that the interval [1, 2] is divided into an
even number of sub-intervals. Hence we may take h = 0.1 < 0.102, for which n =
10, i.e., there will be 10 sub-intervals.
The value of the integral is,
2
dx 0.1   1 1 1 1 1   1 1 1 1  1

1
x

3 
1.0  4     
 1.1 1.3 1.5 1.7 1.9 
 2    
 1.2 1.4 1.6 1.8  2 
0.1
 [1.5  4  3.4595  2  2.7282]
3
Self - Learning 0.1
  2.0749  0.06931 which agrees with the exact value of log e2 .
136 Material 3
Interval Halving Technique Numerical Differentiation
and Integration
When the estimation of the truncation error is cumbersome, the method of interval
halving is used to compute an integral to the desired accuracy.
In the interval halving technique, an integral is first computed for some moderate NOTES
h
value of h. Then, it is evaluated again for spacing , i.e., with double the number
2
of subdivisions. This requires the evaluation of the integrand at the new points of
subdivision only and the previous function values with spacing h are also used.
Ih
Now the difference between the integral Ih and is used to check the accuracy
2

of the computed integral. If I h  I h / 2  , where  is the permissible error, then

Ih/2 is to be taken as the computed value of the integral to the desired accuracy. If
the above accuracy is not achieved, i.e., I h  I h / 2  , then the computation of the
h
integral is made again with spacing and the accuracy condition is tested again.
4
The equation of I h / 4 will require the evaluation of the integrand at the new points of
sub-division only.
Notes:
1. The initial choice of h is sometimes taken as m  where m = 2 for trapezoidal
rule and m = 4 for Simpson’s one-third rule.
2. The method of interval halving is widely used for computer evaluation since it
enables a general choice of h together with a check on the computations.
3. The truncation error R can be estimated by using Runge’s principle given by,
1
R Ih  Ih/2 for trapezoidal rule and R  1 I h  I h / 2 for Simpson’s one-
3 15
third rule.
Algorithm: Evaluation of an integral by Simpson’s one-third rule with interval
halving.
Step 1: Set/initialize a, b, 
[a, b are limits of integration, is error tolerance]
ba
Step 2: Set h 
2
Step 3: Compute S1 = f (a) + f (b)
Step 4: Compute S4 = f (a + h)
Step 5: Set S2 = 0, I1 = 0
( S1  4S 4  S 2 )  h
Step 6: Compute I 2 
3

Step 7: If ( I 2  I1 )  , go to Step 17 else go to the next step Self - Learning

Material 137
Numerical Differentiation
and Integration h
Step 8: Set h  , I1 = I2
2
Step 9: Compute S2 = S2+ S4
NOTES Step 10: Set S4 = 0
Step 11: Set x = a + h
Step 12: Compute S4 = S4+ f (x)
Step 13: Set x = x + h
Step 14: If x < b, go to Step 12 else go to the next step
( S1  2S 2  4S 4 )  h
Step 15: Compute I 2 
3

Step 16: Go to step 7

Step 17: Write I2, h, 
Step 18: End

Algorithm: Evaluation of an integral by trapezoidal rule with interval halving.

Step 1: Initialize/set a, b,  [a, b are limits of integration, is error
tolerance]
Step 2: Set h = b–a
f ( a )  f (b )
Step 3: Compute S1 
2
Step 4: Compute I1 = S1×h
h
Step 5: Compute x = a +
2

Step 6: Compute I 2  ( S1  f ( x))  h

Step 7: If I 2  I1  , go to Step 13 else go to the next step

h
Step 8: Set h 
2
Step 9: Set x = a + h
Step 10: Set I2 = I2+ h × f (x)
Step 11: If x < b, go to Step 9 else go to next step
Step 12: Go to Step 7
Step 13: Write I 2 , h, 
Step 14: End

Numerical Evaluation of Double Integrals

We consider the evaluation of a double integral,

138
Self - Learning
Material
I  f ( x, y)dx dy
R
(3.28)
 where R is the rectangular region a  x  b, c  y  d. The double integral Numerical Differentiation
and Integration
can be transformed into a repeated integral in the following form,
b d 
 
a
dx  f ( x, y )dy 
c



(3.29) NOTES
d
Writing F ( x)   f ( x, y )dy considered as a function of x, we have (3.30)
c


I  F ( x)dx
a
(3.31)

Now for numerical integration, we can divide the interval [a, b] into n sub-
intervals with spacing h and then use a suitable rule of numerical integration.
Trapezoidal Rule for Double Integral
By trapezoidal rule, we can write the integral Equation (3.31) as,
b
h
 F ( x) dx  2 [ F
a
0
 Fn  2 ( F1  F2  F3  ...  Fn 1 )] (3.32)

ba
where x0 = a, xn = b, h  and
n
1
Fi  F ( xi )   f ( x , y) dy, x
0
i i
 a  ih (3.33)

for i = 0, 1, 2,..., n.
Each Fi can be evaluated by trapezoidal rule. For this, the interval [c, d] may

be divided into m sub-intervals each of length k  c  d . Thus we can write,

m
k
Fi  [ f ( xi , y0 )  f ( xi , y m )  2{ f ( xi , y1 )  f ( xi , y 2  ...  f ( xi , y m 1 )}] (3.34)
2
y0 = c, ym = d, yi = c+ik; i = 0, 1,..., m.
This Equation (3.34) can be written in a compact form,
k
Fi  [ f  f im  2 ( f i1  f i 2  ...  f im 1 )]. (3.35)
2 i0
The relation Equations (3.32) and (3.35) together form the trapezoidal rule
for evaluation of double integrals.

Simpson’s One-Third Rule for Double Integrals

1
For the evaluation of double integrals we can write Simpson’s rule. Thus we
3
have,
b
h

I  F ( x) dx 
a
[ F  Fn  2 ( F2  F4  ...  Fn  2 )  4 ( F1  F3  ...  Fn 1 )]
3 0 (3.36)

ba
where h  , n is even and Self - Learning
n Material 139
Numerical Differentiation d
and Integration
Fi = F(xi)   f ( xi , y )dy, xi = a + ih, for i = 0, 1, 2,..., n
c

(3.37)
NOTES and, x0 = a and xn = b
For evaluating I, we have to evaluate each of the (n + 1) integrals given in
Equation (3.37). For evaluation of Fi, we can use Simpson’s one-third rule by
dividing [c, d] into m sub-intervals. Fi can be written as,
k
Fi  [ f ( xi , y0 )  f ( xi , y m )  2 f ( xi , y 2 )  f ( xi , y4 )  ...  f ( xi , y m 2 )  4{ f ( xi , y1 )  f ( xi , y3 )
3
 ...  f ( xi , y m1 )}]
...(3.38)
Equation (3.38) can be written in a compact notation as,
k
Fi  [ f  f im  2 ( f i 2  f i 4  ...  f in  2 )  4 ( f i1 , f i 3  ...  f im 1 )]
3 i0
where fij = f (xi, yj), j = 0, 1, 2,...,m.

 ( x
2
 y 2 )dx dy
Example 3.16: Evaluate the following double integral where R
R
is the rectangular region 1  x  3, 1  y  2, by Simpson’s one-third rule taking
h = k = 0.5.
Solution: We write the integral in the form of a repeated integral,
3 2 

1
 
I  dx  ( x 2  y 2 )dy 
1




2
Taking n = 4 sub-intervals along x, so that h  = 0.5
4

y=2
y=1

x=1 x=3
3
0 .5

 I  F ( x)dx 
1
3 [F0 + F4 + 2F2 + 4(F1 + F3)]
2

where F(x)   ( x  y )dy

2 2

 Fi  F ( xi )   ( xi  y )dy; xi = 1+ 0.5i, where i = 0, 1, 2, 3, 4.

2 2

1
For evaluating Fi’s, we take k   0.5 and get,
2
2
0.5 0 .5
 (1  y
2
F0  ) dy  [1  12  4 {1  (1.5) 2 }  1  2 2 ]   20
3 3
1
2
Self - Learning 0 .5 0.5
 (1.5
2
140 Material F1   y 2 ) dy  [(1.5) 2  12  4{1.5) 2  (1.5) 2 }  (1.5) 2  2 2 ]   27.50
3 3
1
 2
0.5 0.5 Numerical Differentiation
F2   (22  y 2 ) dy  [22  12  4 (22  1.5) 2 }  22  22 ]   38 and Integration
1
3 3
2
0.5 0.5
F3   ((2.5) 2  y 2 )dy  [(2.5)2  12  4 {(2.5)2  (1.5)2 }  (2.5) 2  22 ]   51.50
1 3 3
2
0.5 2 2 0.5 NOTES
F4   (32  y 2 ) dy  [3  1  4{32  (1.5)2 }  32  22 ]   68
1
3 3
0.25
I  [20  68  2  38  4 (27.50  51.50)]
9
0.25
  480  13.333
9

 ( x
2
Example 3.17: Compute  y 2 )dx dy by trapezoidal rule with h = 0.5.
R

y=2

y=1

x=1 x=3
3
0.5
Solution: I T   F ( x)dx  [F0+F4+2 (F1+F2+F3)]
2
1
2


where Fi = F(xi)  ( xi2  y 2 )dy, xi = 1 + 0.5i, i = 0, 1, 2, 3, 4.
1
2
0.5 2 2
Thus, F0   (1  y 2 )dy  [1  1  2{12  (1.5)2 }  12  22 ]
1
2
0.5
  13.50  3.375
2
2
0.5
F1   [(1.5)2  y 2 ]dy  [(1.5)2  12  2 {(1.5)2  (1.5)2 }  (1.5) 2  22 ]
1
2
0.5
  18.50  4.625
2
2
0.5
F2   [22  y 2 ]dy  [22  12  2{22  (1.5) 2 }  22  22 ]
1
2
0.5
  25.50  6.375
2
2
0.5
F3   [(2.5) 2  y 2 ]dy  [(2.5)2  12  2{(2.5) 2  (1.5) 2 }  (2.5)2  22 ]
1
2
0.5
  34.50  8.625
2
2
0.5 2 2
F4   [32  y 2 ]dy  [3  1  2{32  (1.5) 2 }  32  22 ]
1 2
0.5
  45.50  11.375
2
0.5
 IT   [3.375  11.375  2(4.625  6.375  8.625)]
2
1
 [14.750  2  19.625]
4
1 1 Self - Learning
 14.750  39.250   54  13.5 Material 141
4 4
Numerical Differentiation Example 3.18: Evaluate the following double integral using trapezoidal rule with
and Integration
2 2
dx dy
length of sub-intervals h = k = 0.5,   x y .
1 1
NOTES 1
Solution: Let f ( x, y ) 
x y
y

1.5

x
1 1.5 2

By trapezoidal rule with h = 0.5, the integral

2 2
I   dx dy f ( x, y ) is computed as,
1 1

0.5  0.5
I [ f (1, 1)  f (2, 1)  f (1, 2)  f (2, 2)  2{ f (1.5, 1)  f (1, 1.5)
4
 f (2, 1.5)  f (1.5, 2)}  4 f (1.5, 1.5)]

1 1 1 1 1 2 2 2 2 1
      2       4 
16  2 3 3 4 5 5 7 7 3
1 4  12 4 
  0.666667  0.75  2   
16  35 3
1
 5.492857
16
 0.343304.

2 2
dxdy
Example 3.19: Evaluate  x  y by Simpson’s one-third rule. Take sub-inter--
1 1

vals of length h = k = 0.5.

2 2
Solution: The value of the integral I    f ( x, y)dx dy by Simpson’s one-third
1 1

rule with h = k = 0.5 is,

0.5  0.5
I [ f (1, 1)  f (2, 1)  f (1, 2)  f (2, 2)  4{ f (1, 1.5)  f (1.5, 1)
33
 f (2, 1.5)  f (1.5, 2)}  16 f (1.5, 1.5)]
1 1 1 1 1  2 2 2 2 1
     4       16  
36  2 3 3 4  5 5 7 7 3
1  4  12 16 
 0.666667  0.75  4   
36  35 3
1
Self - Learning  [12.235714]  0.339880
142 Material 36
3.3.4 Gaussian Quadrature Numerical Differentiation
and Integration
We have seen that Newton-Cotes formula of numerical integration is of the form,
b n

 f ( x)dx   c f (x )
i 0
i i (3.39) NOTES
a

ba
where xi = a+ih, i = 0, 1, 2, ..., n; h 
n
This formula uses function values at equally spaced points and gives the exact
result for f (x) being a polynomial of degree less than or equal to n. Gaussian
quadrature formula is similar to Equation (3.39) given by,
1 n

 F (u )du   w F (u )
i 1
i i (3.40)
1

where wi’s and ui’s called weights and abscissae, respectively are derived such
that above Equation (3.40) gives the exact result for F(u) being a polynomial of
degree less than or equal to 2n–1.
In Newton-Cotes Equation (3.39), the coefficients ci and the abscissae xi are
rational numbers but the weights wi and the abscissae ui are usually irrational
numbers. Even though Gaussian quadrature formula gives the integration of F(u)
between the limits –1 to +1, we can use it to find the integral of f (x) from a to b
by a simple transformation given by,
ba ab
x u (3.41)
2 2
Evidently, then limits for u become –1 to 1 corresponding to x = a to b and
writing,
b  a a b
f ( x)  f  u  F (u )
 2 2 

b 1
ba
we have, a
f ( x)dx 
2  F (u)du
1

(3.42)
It can be shown that the ui are the zeros of the Legendre polynomial Pn(u) of
degree n. These roots are real but irrational and the weights are also irrational.
Given below is a simple formulation of the relevant equations to determine ui
and wi. Let F(u) be a polynomial of the form,
2 n 1
F (u )  a u
k 0
k
k
(3.43)

Then, we can write

1 1 2 n 1
 
 
k
F (u )du   a k u  du (3.44)
1  k 0
1   Self - Learning
Material 143
Numerical Differentiation
1
and Integration 2 2 2
or,  F (u )du  2a0  a2  a4  ...  a
3 5 2n  2 2 n  2
1

(3.45)
NOTES
Equation (3.40) gives,
1 n  2 n 1 
 F (u )du   
i 1
wi 
 k 0
ak uik 

1

 w a 
n
 i 0  a1ui  a2 ui2  ...  a2 n 1ui2 n 1
i 1

(3.46)
The Equations (3.45) and (3.46) are assumed to be identical for all polynomi-
als of degree less than or equal to 2n–1 and hence equating the coefficients of ak
on either side we obtain the following 2n equations for the 2n unknowns w1,
w2,...,wn and u1, u2,...,un.

n n n n
2

i 1
wi  2, i 1
wi ui  0, 
i 1
wi ui2  
,... wi ui2 n 1  0
3 i 1

(3.47)
The solution of Equation (3.47) is quite complicated. However, use of Legendre
polynominals makes the labour unnecessary. It can be shown that the abscissae ui
are the zeros of the Legendre polynomial Pn(x) of degree n. The weights wi can
then be easily determined by solving the first n Equations of Equations (3.47). As
an illustration, we take n = 2. The four equations for u1, u2, w1 and w2 are,
w1+w 2 = 2
w1u1+w2u2 = 0
2
w1u12  w2 u 22 
3
w1u13  w2u 23  0
Eliminating w1, w2, we get
w1 u u3
  2   23
w2 u1 u1

or, u13u 2  u1u 23  0 or u1u 2 (u12  u 22 )  0

Since, u1  u 2  0 , we have u1 = – u2.

2 1 1
Also, w1 = w2 = 1. The third equation gives, 2u12   u1  , u2  
3 3 3
Hence, two point Gauss-Legendre quadrature formula is,
1
 1   1 
Self - Learning  F (u )du  F 
1
 F
3 

3
144 Material
 The Table 3.1 gives the abscissae and weights of the Gauss-Legendre Numerical Differentiation
and Integration
quadraturen for values of n from 2 to 6.
Table 3.1 Values of Weights and Abscissae for Gauss-Legendre Quadrature

n Weights Abscissae NOTES

2 1.0  0.57735027
3 0.88888889 0.0
0.55555556  0.77459667
4 0.65214515  0.33998104
0.34785485  0.86113631
5 0.56888889 0.0
0.47862867  0.53846931
0.23692689  0.90617985
6 0.46791393  0.23861919
0.36076157  0.66120939
0.17132449  0.93246951

It is seen that the abscissae are symmetrical with respect to the origin and the
weights are equal for equidistant points.
2
Example 3.20: Compute  (1  x)dx, by Gauss two point quadrature formula.
0
2
Solution: Substituting x = u + 1, the given integral  (1  x)dx
0
reduces to
1


I  (u  2)du .
1
Using a two point Gauss quadrature formula, we have
I = (0.57735027+2) + (– 0.57735027+2) = 4.0.
As expected, the result is equal to the exact value of the integral.
Example 3.21: Show that Gauss two-point quadrature formula for evaluating
b b N

 f ( x)dx can be written in the composite form as  f ( x)dx  h  [ f (r )  f (s )]

i 0
i i
a a

1
where ri = xi + hp, si = xi + (1 – p)h, p  (3  3 ).
6
Solution: We subdivide the interval [a, b] into N sub-intervals, each of length h,
given by h  b  a .
N
xi1

Consider the integral Ii over the interval (xi, xi+1), i.e., I x 

i  f ( x)dx .
xi

We transform the integral Ii by putting x  h u   xi  h , so that x = xi gives

2  2
1
u = –1 and x = xi+1 gives u = 1. Thus, I i  h  f  h u  xi  h du .
2 2 2 Self - Learning
1
Material 145
Numerical Differentiation The Gauss two point quadrature gives,
and Integration

h h 1 h  h h 
Ii   f    xi    f    xi  
2   2 3 2  2 3 2 
NOTES h
 [ f (ri )  f ( si )]
2

1
where ri = xi + ph, si = xi + (1 – p)h, p  (3  3 )
6
b N 1 N 1
h
Hence,  f ( x)dx  
i 0
Ii 
2  [ f ( r )  f ( s )]
i 0
i i
a

Note: Instead of considering Gauss integration formula for more and more num-
ber of points for better accuracy, one can use a two point composite formula for
larger number of sub-intervals.
Example 3.22: Evaluate the following integral by Gauss three point quadrature
formula:
1
dx
I 
0
1 x
Solution: We first transform the interval [0, 1] to the interval (–1, 1) by substitut-
i n g
1 1
dx dt
t = 2x – 1, so that   .
0
1 x 1 t  3
Now by Gauss three point quadrature we have,
1 1
I [8 F (0)  5 F (3  0.77459667)  5F (3.77459667)] with F (t ) 
9 t 3
 I  0.693122

1
dx
The exact value of  1  x  ln 2  0.693147
0

Error = 0.000025

Romberg’s Procedure
This procedure is used to find a better estimate of an integral using the evaluation
of the integral for two values of the width of the sub-intervals.
b

Let I1 and I2 be the values of an integral I   f ( x) dx, with two different num-
a
ber of sub-intervals of width h1 and h2 respectively using the trapezoidal rule. Let
E1 and E2 be the corresponding truncation errors. Since the errors in trapezoidal
rule is of order of h2, we can write,

Self - Learning
146 Material
 Numerical Differentiation
I  I1  Kh12 and I  I 2  Kh22 , where K is approximately same. and Integration

 I1  Kh12  I 2  Kh22
I1  I 2
 K NOTES
h22  h12

I1  I 2 I1h22  I 2 h12
Thus, I  I1  .h 2 
2 1
h22  h1 h22  h12

h1
In Romberg procedure, we take h2  and we then have,
2
2
h 
I1  1   I 2 h12
4I  I
I   2
2
 2 1
 h1  3
  h
2

2
I I 
or, I  I2   2 1 
 3 
This is known as Romberg’s formula for trapezoidal integration.
The use of Romberg procedure gives a better estimate of the integral without
any more function evaluation. Further, the evaluation of I2 with h/2 uses the func-
tion values required in evaluation of I1.
1
dx
Example 3.23: Evaluate I  1 x
0
2 by trapezoidal rule with h1  0.5and h2  0.25

and then use Romberg procedure for a better estimate of I. Compare the result
with exact value.
1
Solution: We tabulate the value of x and y  with h = 0.25.
1 x2
x 0 0.25 0.5 0.75 1.0
y 1 0.9412 0.80 0.64 0.5

Thus using trapezoidal rule, with h1 = 0.5, we have

0.5
I1   (1  0.5  2  0.8)  0.516
3
Similarly, with h2 = 0.25,
0.25
I2  [1  0.5  2 (0.8  0.9412  0.64)]
3
 0.5218

The evaluation of I2 uses the function values for evaluation of I1.

By Romberg formula,
Self - Learning
Material 147
Numerical Differentiation
1
and Integration I  I2  ( I 2  I1 )
3
1
 0.5218  (0.5218  0.516) 
3
NOTES  0.5218  0.0019
 0.5237
1 
The exact integral  tan 1 x 0   0.5237.
4
Thus we can take the result correct to four places of decimals.
2
dx
Example 3.24: Evaluate I   by trapezoidal rule with two and four sub-
x
1

intervals and then use Romberg procedure to get a better estimate of I.

1 1
Solution: We form a table of value of y  with spacing h   0.25.
x 4

x 1 1.25 1.5 1.75 2.0

y 1 0.8 0.6667 0.5714 0.5

0.5
I1  [1  0.5  2  0.6667]  0.7084
2
0.25
I2  [1  0.5  2 (0.8  0.6667  0.5714)]  0.6970
2
By Romberg proecedure,
I 2  I1 1
I  I2   0.6970  (0.0114)
3 3
 0.6970  0.0038  0.6932

1
dx
Example 3.25: Compute the value of  1 x ,
0
(i) By Gauss two point and

(ii) By Gauss three point formulas.

ba 1
Solution: We first transform the integral by substituting x  t  (b  a)
2 2

1 1 1
dx 1 1 1 2

0

1 x 2 
11 
1 1
 t

2 3t
1
dt 
2 2
1
 1   1 
(i) By Gauss two point quadrature  F (t ) dt  F    F    we get ,
1  3  3 

 
1  
1 1 1
 dt      0.6923
3t  1 1 
Self - Learning 1  3 3 
148 Material  3 3 
(ii) By Gauss three point quadrature, Numerical Differentiation
and Integration
1
dt 1 0.55555556 
 3t
1
dt    0.888888 
3 3  0.77459667 
 0.443478 NOTES
2
Example 3.26: Compute  e x dx by Gauss three point quadrature.
1

Solution: We first transform the integral by substituting

6a 1 1 3
x t  (b  a )  t 
2 2 2 2
2 1 t 3 3 1 t
1  1
  
x
 e dx  e 2 2 dt  e 2 e 2 dt
2 2
1 1 1

1 2   1 
3 1
e 0.88888889  e  0.55555556  e 2  0.77459667  e 2  0.77459667
0

2   

 4.67077

Check Your Progress

1. Define the process of numerical differentiation.
2. Write Newton's forward difference interpolation formula.
3. Write Newton's backward difference interpolation formula.
4. How will you evaluate a definite integral?
5. Write the trapezoidal formula for numerical integration.
6. What is Simpson's one-third formula of numerical integration?
7. Define Simpson's three-eighth rule of numerical integration.
8. State Weddle's rule.
9. Why is Romberg's procedure used?

3.4 SOLVING NUMERICAL

Even though there are many methods to find an analytical solution of ordinary
differential equations, for many differential equations solutions in closed form cannot
be obtained. There are many methods available for finding a numerical solution for
differential equations. We consider the solution of an initial value problem associated
with a first order differential equation given by,
dy
 f ( x, y )
dx
(3.48)
With y (x0) = y0 (3.49)
In general, the solution of the differential equation may not always exist. For
the existence of a unique solution of the differential equation (3.48), the following Self - Learning
conditions, known as Lipshitz conditions must be satisfied, Material 149
Numerical Differentiation (i) The function f(x, y) is defined and continuous in the strip
and Integration
R : x0  x  b,   y  
(ii) There exists a constant L such that for any x in (x0, b) and any two num-
bers y and y1
NOTES
|f(x, y) – f(x, y1)|  L|y – y1|
(3.50)
The numerical solution of initial value problems consists of finding the ap-
proximate numerical solution of y at successive steps x1, x2,..., xn of x. A number
of good methods are available for computing the numerical solution of differential
equations.
3.4.1 Taylor Series Method
Consider the solution of the first order differential equation,
dy
 f ( x, y ) with y ( x0 )  y 0 (3.51)
dx
where f (x, y) is sufficiently differentiable with respect to x and y. The solution y
(x) of the problem can be expanded about the point x0 by a Taylor series in the
form,
h2 y k ( x0 ) k h k 1
y ( x0  h)  y ( x0 )  hy  ( x0 )  y ( x0 )  ...  h  ()
2! k! (k  1)!
The derivatives in the above expansion can be determined as follows,
y ( x0 )  f ( x0 , y 0 )
y  ( x0 )  f x ( x0 , y0 )  f y ( x0 , y 0 ) y ( x0 )

y  ( x0 )  f xx ( x0 , y0 )  2 f xy ( x0 , y 0 ) y ( x0 )  f yy ( x0 , y0 ) { y  ( x0 )}2  f y ( x, y ) y  ( x0 )

where a suffix x or y denotes partial derivative with respect to x or y.

Thus the value of y1 = y (x0+h), can be computed by taking the Taylor series
expansion shown above. Usually, because of difficulties in obtaining higher order
derivatives, commonly a fourth order method is used. The solution at x2 = x1+h,
can be found by evaluating the derivatives at (x1, y1) and using the expansion;
otherwise, writing x2 = x0+2h, we can use the same expansion. This process can
be continued for determining yn+1 with known values xn, yn.
Note: If we take k = 1, we get the Euler’s method, y1 = y0+h f (x0, y0).
Thus, Euler’s method is a particular case of Taylor series method.
Example 3.27: Form the Taylor series solution of the initial value problem,
dy
 xy  1, y (0)  1 up to five terms and hence compute y (0.1) and y (0.2), correct
dx
to four decimal places.
Solution: We have, y   xy  1, y (0)  1
Differentiating successively we get,
y ( x)  xy   y ,  y (0)  1
y ( x)  xy   2 y ,  y (0)  2
y (iv ) ( x)  xy   3 y ,  y (iv ) (0)  3
Self - Learning
150 Material y ( v ) ( x)  xy (iv )  3 y ,  y ( v ) (0)  6
 Hence, the Taylor series solution y (x) is given by, Numerical Differentiation
and Integration

x2 x3 x 4 ( iv ) x 5 (v)
y ( x)  y (0)  xy (0)  y (0)  y (0)  y (0)  y (0)
2 3! 4! 5! NOTES
x 2 x3 x4 x5 x 2 x3 x 4 x5
 1 x   2 3  6  1 x    
2 6 24 120 2 3 8 20
0.01 0.001 0.0001 0.00001
 y (0.1)  1  0.1      1.1053
2 3 8 20

Similarly, y (0.2)  1  0.2  0.04  0.008  0.0016  0.00032  1.04274

2 3 8 20
Example 3.28: Find first two non-vanishing terms in the Taylor series solution of
the initial value problem y   x 2  y 2 , y (0)  0. Hence, compute y (0.1), y (0.2), y
(0.3) and comment on the accuracy of the solution.
Solution: We have, y   x 2  y 2 , y (0)  0
Differentiating successively we have,

y   2 x  2 yy ,  y (0)  0
y   2  2[ yy   ( y ) 2 , y (0)  2
(iv ) ( iv )
y  2 ( yy   3 y y ), y (0 )  0
(v) (iv ) 2
y  2[ yy  4 y y   3 ( y ) ],  y (v)  0
y (vi )  2[ yy ( v )  5 y y (iv )  10 y y ],  y ( vi )  0
y ( vii )  2[ yy ( vi )  6 y  y ( v )  15 y  y (iv )  10 ( y ) 2 ]  y ( vi ) (0)  80

x3 x7 80 1 x7
The Taylor series up to two terms is y ( x)   2    x3 
6 7 7! 3 63
Example 3.29: Given x y  = x – y2, y(2) = 1, evaluate y(2.1), y(2.2) and y(2.3)
correct to four decimal places using Taylor series method.
Solution: Given y   x  y 2 , i.e., y   1  y 2 / x and y  1 for x  2. To compute
y(2.1) by Taylor series method, we first find the derivatives of y at x = 2.
1
y  1  y 2 / x  y (2)  1   0.5
2
xy   y   1  2 yy 
1 1 1 2 1
2 y (2)   1  2.  y (2)     0.25
2 2 4 2 2
xy   2 y   2 y 2  2 yy 
2
 1 1  1
 2 y (2)  2     2    2   
 4  2  4
1 1
2 y (2)   y (2)   0.25
Or, 2 4
xy  3 y   4 y y   2 y y   2 yy 
( iv )

Self - Learning
Material 151
Numerical Differentiation
1 1  1 1
and Integration 2 y  (2)  3   6   2
4 2  4 4
 3 3 1 1
y  (2)       0.25
 4 4 2 2
NOTES
(0.1) 2 (0.1)3 (0.1)4
y (2.1)  y (2)  0.1 y  (2)  y  (2)  y (2)  y  (2)
2 3! 4!
0.01 0.001 0.0001
 1  0.1  0.5   (0.25)   0.25   (0.25)
2 6 24
 1  0.05  0.00125  0.00004  0.000001
 1.0488

0.04 0.008 0.0016

y (2.2)  1  0.2  0.5   (0.25)   0.25   (0.5)
2 6 24
 1  0.1  0.005  0.00032  0.00003
 1.0954
0.09 0.009 0.0081
y (2.3)  1  0.3  0.5  ( 0.25)   0.25   (0.5)
2 2 24
 1  0.15  0.01125  0.001125  0.000168
 1.005043

Picard’s Method of Successive Approximations

Consider the solution of the initial value problem,
dy
 f ( x, y ) with y(x0) = y0
dx
Taking y = y (x) as a function of x, we can integrate the differential equation
with respect to x from x = x0 to x, in the form
x
y  y0   f ( x, y( x)) dx
x0
(3.52)

The integral contains the unknown function y (x) and it is not possible to
integrate it directly. In Picard’s method, the first approximate solution y (1) ( x) is
obtained by replacing y (x) by y0.
x

Thus, y (1) ( x )  y0   f ( x, y )dx

x0
0

(3.53)
The second approximate solution is derived on replacing y by y (x). Thus, (1)

 f ( x, y
( 2) (1)
y ( x)  y0  ( x)) dx
(3.54)
x0

The process can be continued, so that we have the general approximate solu-
tion given by,
Self - Learning
152 Material
 x Numerical Differentiation
and Integration
 f ( x, y
(n) ( n 1)
y ( x)  y 0  ( x))dx,
for n = 2, 3... (3.55)
x0

This iteration formula is known as Picard’s iteration for finding solution of a

NOTES
first order differential equation, when an initial condition is given. The iterations are
continued until two successive approximate solutions y(k) and y(k+1) give
approximately the same result for the desired values of x up to a desired accuracy.
Note: Due to practical difficulties in evaluating the necessary integration, this method
cannot be always used. However, if f (x, y) is a polynomial in x and y, the succes-
sive approximate solutions will be obtained as a power series of x.
Example 3.30: Find four successive approximate solutions for the following ini-
tial value problem: y   x  y, with y (0) = 1, by Picard’s method. Hence compute
y (0.1) and y (0.2) correct to five significant digits.
Solution: We have, y   x  y , with y (0) = 1.
The first approximation by Picard’s method is,
x

y (1) ( x)  y (0)   [ x  y (0)] dx

0
x
x2
 y (1) ( x)  1   ( x  1) dx  1  x 
0
2
The second approximation is,
x
x2 x3

( 2) 2
y ( x)  1  ( x  1  x  )dx  1  x  x 
2 6
0

Similarly, the third approximation is,

x
x3
y (3) ( x)  1   (1  2 x  x 2  )dx
0
6
x3 x 4
 y (3) ( x)  1  x  x 2  
3 24
The fourth approximation is,
x
x3 x4
y (4) x  1   (1  2x  x 2   )dx
0
3 24
x3 x 4 x5
 y (4) ( x)  1  x  x 2   
3 12 120

It is clear that successive approximations are easily determined as power se-

ries of x having one degree more than the previous one. The value of y (0.1) is
given by,
(0.1)3 (0.1) 4
y (0.1)  1  0.1  (0.1) 2    ...  1.1103, correct to five significant
3 4
digits.
3 4 5
Similarly, y (0.2)  1  0.2  (0.2)2  (0.2)  (0.2)  (0.2)  1.2431. Self - Learning
3 4 120 Material 153
Numerical Differentiation Example 3.31: Find the successive approximate solution of the initial value prob-
and Integration
lem, y   xy  1, with y (0) = 1, by Picard’s method.
Solution: The first approximate solution is given by,
NOTES x
x2

(1)
y ( x)  1  ( x  1)dx  1  x 
2
0

The second and third approximate solutions are,

x
x2 x2 x3 x 4

y ( 2) ( x )  1  [ x(1  x 
0
2
)  1]dx  1  x 
2

3

8
x
x 2 x3 x4 x 2 x3 x 4 x5 x6

y (3) ( x )  1  [ x(1  x 
0
2

3
 )  1]dx  1  x 
4 2

3
  
8 15 48

Example 3.32: Compute y (0.25) and y (0.5) correct to three decimal places by
solving the following initial value problem by Picard’s method:
dy x2
 , y (0)  0
dx 1  y 2

dy x2
Solution: We have dx  1  y 2 , y (0)  0

By Picard’s method, the first approximation is,

x
x2 x3
y (1) ( x)  0   1  0 dx 
0
3

The second approximate solution is,

x
x2
y ( 2) ( x )   1[y
0
(1)
( x)]
2
dx

x
x2 x3
  x6
dx  tan 1
3
0 1
9

(0.25) 2
For x  0.25, y (1) (0.25)   0.0052
3
(0.25) 2
y (2) (0.25)  tan 1  0.0052
3
 y (0.25)  0.005, Correct to three decimal place.

(0.5)2
Again, for x = 0.5, y (1) (0.5)   0.083333
3

(0.5) 3
y ( 2) (0.5)  tan 1  0.0416
3
Self - Learning
154 Material
Thus, correct to three decimal places, y (0.5) = 0.042.
Note: For this problem we observe that, the integral for getting the third and Numerical Differentiation
and Integration
higher approximate solution is either difficult or impossible to evalute, since
x
x2
y (3) ( x)   2 is not integrable.
0  x3  NOTES
1   tan 1 
 3
Example 3.33: Use Picard’s method to find two successive approximate solu-
tions of the initial value problem,
dy y  x
 , y (0 )  1
dx y  x

Solution: The first approximate solution by Picard’s method is given by,

x
y (1) ( x )  y0   f ( x, y0 )dx
0
x x
1 x 2  (1  x)
 y (1) ( x )  1   dx  1   dx
0 1 x 0 1 x
 y (1) ( x )  1  2log e |1  x |  x

The second approximate solution is given by,

x

 f ( x, y
( 2) (1)
y ( x)  y 0  ( x))dx
0
x x
x  2 x  2 log e | 1  x | x
 1 
0
1  2 log e | 1  x |
dx  1  x  2  1  2 log
0 e |1 x |
dx

We observe that, it is not possible to obtain the integral for getting y(2)(x). Thus
Picard’s method is not applicable to get successive approximate solutions.

3.4.2 Euler’s Method

This is a crude but simple method of solving a first order initial value problem:
dy
 f ( x, y ), y ( x0 )  y0
dx
This is derived by integrating f (x0, y0) instead of f (x, y) for a small interval,
x0  h x0  h

 
x0
dy  
x0
f ( x0 , y0 )dx

 y ( x0  h)  y ( x0 )  hf ( x0 , y0 )

Writing y1 = y (x0+ h), we have

y1 = y0+h f (x0, y0) (3.56)
Similarly, we can write
y2 = y (x1+ h) = y1+ h f (x1, y1) (3.57)
where x1 = x0+ h.
Self - Learning
Material 155
Numerical Differentiation Proceeding successively, we can get the solution at any xn = x0+ nh, as
and Integration
yn = yn–1+ h f (xn–1, yn–1) (3.58)
This method, known as Euler’s method, can be geometrically interpreted, as
NOTES shown in Figure 3.3.
Y

X
0 x0 x1 x2

Fig. 3.3 Euler’s Method

For small step size h, the solution curve y = y (x), is approximated by the
tangential line.
The local error at any xk , i.e., the truncation error of the Euler’s method is
given by,
ek = y(xk+1) – yk+1
where yk+1 is the solution by Euler’s method.

 ek  y ( xk  h)  { yk  hf ( xk , yk )}
h2
 yk  hy ( xk )  y( xk  h)  yk  hy( xk ), 0    1
2
h2
ek  y ( xk  h), 0    1
 2

Note: The Euler’s method finds a sequence of values {yk} of y for the sequence
of values {xk}of x, step by step. But to get the solution up to a desired accuracy,
we have to take the step size h to be very small. Again, the method should not be
used for a larger range of x about x0, since the propagated error grows as integra-
tion proceeds.
Example 3.34: Solve the following differential equation by Euler’s method for
dy
x = 0.1, 0.2, 0.3; taking h = 0.1;  x 2  y, y (0)  1. Compare the results with
dx
exact solution.
dy
Solution: Given  x 2  y, with y (0) = 1.
dx
In Euler’s method one computes in successive steps, values of y1, y2, y3,... at
x1 = x0+ h, x2 = x0 + 2h, x3 = x0 + 3h, using the formula,
yn 1  yn  hf ( xn , yn ), for n  0, 1, 2,...
 yn 1  yn  h ( x 2 n  yn )

With h = 0.1 and starting with x0 = 0, y0 = 1, we present the successive

Self - Learning computations in the table given below.
156 Material
 2 Numerical Differentiation
n xn yn f ( xn , y n )  x n  y n y n 1  y n  hf ( xn , y n ) and Integration
0 0.0 1.000  1.000 0.9000
1 0.1 0.900  0.8900 0.8110
2 0.2 0.8110  0.7710 0.7339 NOTES
3 0.3 0.7339  0.6439 0.6695

dy
The analytical solution of the differential equation written as  y  x 2 , is
dx
ye x   x 2 e x dx  c
or, ye x  x 2 e x  2 xe x  2e x  c.
Since, y  1 for x  0, c  1.
 y  x 2  2 x  2  e x .
The following table compares the exact solution with the approximate solution
by Euler’s method.
n xn Approximate Solution Exact Solution % Error
1 0.1 0.9000 0.9052 0.57
2 0.2 0.8110 0.8213 1.25
3 0.3 0.7339 0.7492 2.04
Example 3.35: Compute the solution of the following initial value problem by
Euler’s method, for x = 0.1 correct to four decimal places, taking h = 0.02,
dy y  x
 , y (0)  1 .
dx y  x

Solution: Euler’s method for solving an initial value problem,

dy
 f ( x, y ), y ( x0 )  y0 , is yn 1  yn  h f ( xn , yn ), for n  0, 1, 2,...
dx
Taking h = 0.02, we have x1 = 0.02, x2 = 0.04, x3 = 0.06, x4 = 0.08, x5 =
0.1.
Using Euler’s method, we have, since y(0) = 1

1 0
y (0.02)  y1  y 0  h f ( x0 , y0 )  1  0.02   1.0200
1 0
1.0200  0.02
y (0.04)  y 2  y1  h f ( x1 , y1 )  1.0200  0.02   1.0392
1.0200  0.02

1.0392  0.04
y (0.06)  y3  y 2  h f ( x2 , y 2 )  1.0392  0.02   1.0577
1.0392  0.04
1.0577  0.06
y (0.08)  y 4  y3  h f ( x3 , y3 )  1.0577  0.02   1.0756
1.0577  0.06
1.0756  0.08
y (0.1)  y5  y 4  h f ( x4 , y 4 )  1.0756  0.02   1.0928
1.0756  0.08
Hence, y (0.1) = 1.0928.

Self - Learning
Material 157
Numerical Differentiation Modified Euler’s method
and Integration
In order to get somewhat moderate accuracy, Euler’s method is modified by
computing the derivative y   f ( x, y ), at a point xn as the mean of f (xn, yn) and f
(xn+1, y(0)n+1), where,
NOTES
y (0) n 1  yn  h f ( xn , yn )
h
y (1) n 1  yn  [ f ( xn , yn )  f ( xn 1 , y (0) n 1 )]
2
(3.59)
This modified method is known as Euler-Cauchy method. The local truncation
error of the modified Euler’s method is of the order O(h3).
Note: Modified Euler’s method can be used to compute the solution up to a
desired accuracy by applying it in an iterative scheme as stated below.
Compute y ( k ) n 1  yn  h f ( xn , yn )
h
Compute yn( k11)  yn   f ( xn , y n )  f ( xn 1 , yn( k)1  , for k  0, 1, 2,...
2 
(3.60)
The iterations are continued until two successive approximations yn( k)1 and yn( k11)
coincide to the desired accuarcy. As a rule, the iterations converge rapidly for a
sufficiently small h. If, however, after three or four iteration the iterations still do
not give the necessary accuracy in the solution, the spacing h is decreased and
iterations are performed again.
Example 3.36: Use modified Euler’s method to compute y (0.02) for the initial
value problem, dy  x 2  y, with y (0) = 1, taking h = 0.01. Compare the result
dx
with the exact solution.
Solution: Modified Euler’s method consists of obtaining the solution at successive
points, x1 = x0 + h, x2 = x0 + 2h,..., xn = x0 + nh, by the two stage computations
given by,
y n(0)1  yn  hf ( xn , yn )

y n(1)1  y n 
h
2
 
f ( xn , y n )  f ( x n 1 , y n( 0)1 ) .

For the given problem, f (x, y) = x2 + y and h = 0.01

y1(0)  y0  h[ x02  y0 ]  1  0.01  1  1.01
0.01
y1(1)  1  [1.0  1.01  (0.01)2 ]  1.01005
2
i.e., y1  y (0.01)  1.01005

Next, y 2(0)  y1  h [ x12  y1 ]

 1.01005  0.01[(0.1) 2  1.01005]
 1.01005  0.010102  1.02015

Self - Learning
158 Material
 Numerical Differentiation
0.01 and Integration
y (1)
2  1.01005  [(0.01)2  1.01005  (0.01)2  1.02015]
2
0.01
 1.01005   (2.02140)
2 NOTES
 1.01005  0.10107
 1.11112
 y2  y (0.02)  1.11112

Euler’s Method for a Pair of Differential Equations

Consider an initial value problem associated with a pair of first order differential
equation given by,
dy dz
 f ( x, y, z ),  g ( x, y , z ) (3.61)
dx dx
with y (x0) = y0, z (x0) = z0 (3.62)
Euler’s method can be extended to compute approximate values yi and zi of
y (xi) and z (xi) respectively given by,
yi+1 = yi+h f (xi, yi, zi)
zi+1 = zi+h g (xi, yi, zi) (3.63)
starting with i = 0 and continuing step by step for i = 1, 2, 3,... Evidently, we can
also extend Euler’s method for an initial value problem associated with a second
order differential equation by rewriting it as a pair of first order equations.
Consider the initial value problem,
2
d y  dy 
 g  x, y,  , with y(x ) = y , y ( x0 )  y0
2
dx  dx  0 0

dy dz
We write  z, so that  g ( x, y , z ) with y (x0) = y0 and z (x0) = y0 .
dx dx
Example 3.37: Compute y(1.1) and y(1.2) by solving the initial value problem,
y
y    y  0, with y (1) = 0.77, y  (1) = –0.44
x
z
Solution: We can rewrite the problem as y   z, z     y; with y, (1) = 0.77 and
x
z (1.1) = –0.44.
Taking h = 0.1, we use Euler’s method for the problem in the form,
yi 1  yi  hzi
 z 
z i 1  z i  h  1  yi , i  0, 1, 2,...
 xi 
Thus, y1 = y (1.1) and z1 = z (1.1) are given by,
y1  y0  hz 0  0.77  0.1 (0.44)  0.726
 z 
z1  z 0  h  0  y 0   0.44  0.1 (0.44  0.77)
 x0 
 0.44  0.33  0.473 Self - Learning
Material 159
Numerical Differentiation Similarly, y2 = y(1.2) = y1 + hz1 = 0.726 – 0.1 (–0.473) =
and Integration
0.679

 z 
NOTES z2  z (1.2)  z1  h   1  y1 
 x1 
 0.473 
 0.473  0.1    0.726
 1.1 
 0.473  0.1  0.296  0.503
Thus, y (1.1)  0.726 and y (12)  0.679.

Example 3.38: Using Euler’s method, compute y (0.1) and y (0.2) for the initial
value problem,
y   y  0, y (0)  0, y (0)  1

Solution: We rewrite the initial value problem as y   z, z    y, with y (0) = 0,

z (0) = 1.
Taking h = 0.1, we have by Euler’s method,
y1  y (0.1)  y0  hz0  0  0.1  1  0.1
z1  z (0.1)  z0  h( y0 )  1  0.1  0  1.0
y2  y (0.2)  y1  hz1  0.1  0.1  1.0  0.2
z2  z (0.2)  z1  hy1  1.0  0.1  0.1  0.99

Example 3.39: For the initial value problem y   xy   y  0 , y(0) = 0, y  (0) = 1.

Compute the values of y for 0.05, 0.10, 0.15 and 0.20, having accuracy not
exceeding 0.510–4.
Solution: We form Taylor series expansion using y (0), y  (0) = 1 and from the
differential equation,
y   xy   y  0, we get y (0)  0
y ( x)   xy   2 y   y (0)  2

y (iv ) ( x)   xy   3 y   y iv (0)  0

y v ( x)   xy (iv )  4 y   y v (0)  8

And in general, y(2n)(0) = 0, y ( 2n 1) (0)  2ny ( 2n 1) (0)  (1) n 2 n.2!

x3 x5 2n n ! x 2n 1
Thus, y ( x)  x  3  15  ...  (1) (2n  1)!  ...
n

This is an alternating series whose terms decrease. Using this, we form the
solution for y up to 0.2 as given below:

x 0 0.05 0.10 0.15 0.20

y ( x) 0 0.0500 0.0997 0.1489 0.1973

Self - Learning
160 Material
3.4.3 Runge-Kutta Methods Numerical Differentiation
and Integration
Runge-Kutta method can be of different orders. They are very useful when the
method of Taylor series is not easy to apply because of the complexity of finding
higher order derivatives. Runge-Kutta methods attempt to get better accuracy
NOTES
and at the same time obviate the need for computing higher order derivatives.
These methods, however, require the evaluation of the first order derivatives at
several off-step points.
Here we consider the derivation of Runge-Kutta method of order 2.
The solution of the (n + 1)th step is assumed in the form,
yn+1 = yn+ ak1+ bk2 (3.64)
where k1 = h f (xn, yn) and
k2 = h f(xn+ h, yn+ k1), for n = 0, 1, 2,... (3.65)
The unknown parameters a, b, , and  are determined by expanding in
Taylor series and forming equations by equating coefficients of like powers of h.
We have,
h2 h3
y n 1  y ( xn  h)  y n  h y ( xn )  y ( xn )  y ( xn )  0 (h 4 )
2 6
h2 h3
 y n  h f ( xn , y n )  [ f x  ff y ]n  [ f xx  2 ff yy  f yy f 2  f x f y  f y2 f ]n  0(h 4 )
2 6
(3.66)
The subscript n indicates that the functions within brackets are to be evaluated
at (xn, yn).
Again, expanding k2 by Taylor series with two variables, we have
 2 2  2 k12
k 2  h[ f n  ah ( f x ) n  k1 ( f y ) n  ( f xx ) n  hk1 ( f xy ) n  ( f yy ) n  0(h3 )]
2 2
(3.66) Thus on substituting the expansion of k2, we get from Equation (3.66)
 2 2 2 
yn 1  yn  ( a  b) h f n  bh 2 (f x  ff y ) n  bh3  f xx  ff xx  f f yy   0(h 4 )
 2 2 
On comparing with the expansion of yn+1 and equating coefficients of h and h2
we get the relations,
1
a  b  1, b  b 
2
There are three equations for the determination of four unknown parameters.
Thus, there are many solutions. However, usually a symmetric solution is taken by
1
setting a  b  , then     1
2
Thus we can write a Runge-Kutta method of order 2 in the form,
h
yn 1  yn  [ f ( xn , yn )  f ( xn  h, yn  h f ( xn , yn ))], for n  0, 1, 2,...
2
(3.67)
Proceeding as in second order method, Runge-Kutta method of order 4 can
be formulated. Omitting the derivation, we give below the commonly used Runge-
Kutta method of order 4. Self - Learning
Material 161
Numerical Differentiation
and Integration 1 5
y n 1  y n  (k1  2k 2  2k3  k 4 )  0 (h )
6
k1  h f ( xn , y n )

NOTES  h k 
k 2  h f  xn  , y n  1 
 2 2
 h k 
k 3  h f  xn  , y n  2 
 2 2 
k 4  h f ( x n  h, y n  k 3 ) (3.68)
Runge-Kutta method of order 4 requires the evaluation of the first order
derivative f (x, y), at four points. The method is self-starting. The error estimate
with this method can be roughly given by,
y n*  y n
|y (xn) – yn|  (3.69)
15
h
where yn* and yn are the approximate values computed with and h respectively
2
as step size and y (xn) is the exact solution.
Note: In particular, for the special form of differential equation y   F (x), a function
of x alone, the Runge-Kutta method reduces to the Simpson’s one-third formula
of numerical integration from xn to xn+1. Then,
xn1

yn+1 = yn+  F ( x)dx

xn

h h
or, yn+1 = yn+ [F(xn) + 4F(xn+ ) + F(xn+h)]
6 2
Runge-Kutta methods are widely used particularly for finding starting values
at steps x1, x2, x3,..., since it does not require evaluation of higher order derivatives.
It is also easy to implement the method in a computer program.
Example 3.40: Compute values of y (0.1) and y (0.2) by 4th order Runge-Kutta
method, correct to five significant figures for the initial value problem,
dy
 x  y , y ( 0)  1
dx
dy
Solution: We have  x  y , y ( 0)  1
dx

 f ( x, y )  x  y, h  0.1, x0  0, y0  1
By Runge-Kutta method,
1
y (0.1)  y (0)  ( k  2k 2  2k 3  k 4 )
6 1
where, k1  h f ( x0 , y0 )  0.1  (0  1)  0.1
 h k 
k2  h f  x0  , y0  2   0.1  (0.05  1.05)  0.11
 2 2
 h k 
k3  h f  x0  , y0  2   0.1  (0.05  1.055)  0.1105
 2 2
k4  h f ( x0  h, y0  k3 )  0.1  (0.1  1.1105)  0.12105
Self - Learning 1
162 Material  y (0.1)  1  [0.1  2  (0.11  0.1105  0.12105]  1.130516
6
 Thus, x1  0.1, y1  1.130516 Numerical Differentiation
and Integration
1
y (0.2)  y (0.1)  (k1  2k2  2k3  k4 )
where, 6
k1  h f ( x1 , y1 )  0.1  (0.1  1.11034)  0.121034
NOTES
 h k
k2  h f  x1  , y1  1   0.1 (0.15  1.17086)  0.132086
 2 2
 h k 
k3  h f  x1  , y1  2   0.1 (0.15  1.17638)  0.132638
 2 2
k4  h f ( x1  h, y1  k3 )  0.1 (0.2  1.24298)  0.144298
1
y2  y (0.2)  1.11034  [0.121034  2 (0.132086  0.132638)  0.144298]  1.2428
6
Example 3.41: Use Runge-Kutta method of order 4 to evaluate y (1.1) and y
(1.2), by taking step length h = 0.1 for the initial value problem,
dy
 x 2  y 2 , y (1)  0
dx
Solution: For the initial value problem,
dy
 f ( x, y ), y ( x0 )  y0 , the Runge-Kutta method of order 4 is given as,
dx
1
y n 1  y n  ( k1  2k 2  2k 3  k 4 )
6

k1  h f ( xn , yn )
 h k
k2  h f  xn  , yn  1 
 2 2
 h k2 
where k3  h f  xn  , yn  
 2 2
k4  h f ( xn  h, yn  k3 ), for n  0, 1, 2,...

For the given problem, f (x, y) = x2 + y2, x0 = 1, y0 = 0, h = 0.1.

Thus,

k1  h f ( x0 , y0 )  0.1  (12  02 )  0.1

 h k
k2  h f  x0  , y0  1   0.1  [(1.05)2  (0.5) 2 ]  0.13525
 2 2
 h k 
k3  h f  x0  , y0  2   0.1  [(1.05) 2  (0.05525)2 ]  0.13555
 2 2

k4  h f ( x0  h, y0  k3 )  0.1  [(1.1) 2  (0.13555)2 ]  0.12283

1
 y1  y0  (k1  2k2  2k3  k4 )
6
1 1
 (0.1  0.2705  0.2711  0.12283)   0.76443
6 6
 0.127405

Self - Learning
Material 163
Numerical Differentiation For y (1.2):
and Integration

k1  0.1[(1.1)2  (0.11072) 2 ]  0.12226

k2  0.1[(1.15) 2  (0.17183)2 ]  0.135203
NOTES k3  0.1[(1.15) 2  (0.17832) 2 ]  0.135430
k4  0.1[(1.2) 2  (0.24615) 2 ]  0.150059.
1
 y2  y (1.2)  0.11072  (0.12226  0.270406  0.270860  0.150069)
6
 0.24631

Algorithm: Solution of first order differential equation by Runge-Kutta method

of order 2: y   f (x) with y (x0) = y0.
Step 1: Define f (x, y)
Step 2: Read x0, y0, h, xf [h is step size, xf is final x]
Step 3: Repeat Steps 4 to 11 until x1 > xf
Step 4: Compute k1 = f (x0, y0)
Step 5: Compute y1 = y0+ hk1
Step 6: Compute x1 = x0+ h
Step 7: Compute k2 = f (x1, y1)
Step 8: Compute y1  y0  h  (k1  k 2 ) / 2
Step 9: Write x1, y1
Step 10: Set x0 = x1
Step 11: Set y0 = y1
Step 12: Stop
Algorithm: Solution of y1  f ( x, y ), y ( x0 )  y0 by Runge-Kutta method of
order 4.
Step 1: Define f (x, y)
Step 2: Read x0, y0, h, xf
Step 3: Repeat Step 4 to Step 16 until x1 > xf
Step 4: Compute k1 = h f (x0, y0)
h
Step 5: Compute x  x0 
2

k1
Step 6: Compute y  y0 
2
Step 7: Compute k2 = h f (x, y)
k2
Step 8: Compute y  y0 
2
Step 9: Compute k3 = h f(x, y)
Step 10: Compute x1 = x0+ h
Self - Learning
164 Material Step 11: Compute y = y0+ k3
 Step 12: Compute k4 = h f (x1, y) Numerical Differentiation
and Integration
Step 13: Compute y1 = y0+ (k1+ 2 (k2+ k3) + k4)/6
Step 14: Write x1, y1
Step 15: Set x0 = x1 NOTES
Step 16: Set y0 = y1
Step 17: Stop

Runge-Kutta Method for a Pair of Equations

Consider an initial value problem associated with a system of two first order ordinary
differential equations in the form,

dy dz
 f ( x, y, z ),  g ( x, y , z )
dx dx
with y ( x0 )  y0 and z ( x0 )  z0

The Runge-Kutta method of order 4 can be easily extended in the following

form,

1
yi 1  yi  (k1  2k2  2k3  k4 )
6
1 (3.70)
zi 1  zi  (l1  2l2  2l3  l4 ) for i  0, 1, 2,...
6

where k1  hf ( xi , yi , z i ), l1  hg ( xi , yi , z i )
 h k l   h k l 
k 2  hf  xi  , yi  1 , z i  1 , l 2  hg  xi  , yi  1 , z1  1 
 2 2 2  2 2 2
 h k l   h k l 
k 3  hf  xi  , yi  2 , z i  2 , l3  hg  xi  , yi  2 , zi  2 
 2 2 2  2 2 2
k 4  hf ( xi  h, y1  k 3 , z i  l3 ), l 4  hf ( xi  h, yi  k 3 , z i  l3 )

yi  y ( xi ), z i  z ( xi ), i  0, 1, 2,...

The solutions for y(x) and z(x) are determined at successive step points x1 = x0+h,
x2 = x1+h = x0+2h,..., xN = x0+Nh.

Runge-Kutta Method for a Second Order Differential Equation

Consider the initial value problem associated with a second order differential
equation,
d2y
 g ( x, y, y )
dx 2
with y (x0) = y0 and y  (x0) =  0
On substituting z  y , the above problem is reduced to the problem,
dy dz
 z,  g ( x, y , z )
dx dx

with y (x0) = y0 and z (x0) = y  (x0) =  0

Self - Learning
Material 165
Numerical Differentiation which is an initial value problem associated with a system of two first order differential
and Integration
equations. Thus we can write the Runge-Kutta method for a second order
differential equation as,

NOTES 1
yi 1  yi  (k1  2k 2  2k3  k4 ),
6
1
zi 1  yi1  zi  (l1  2l2  2l3  l4 ) for i  0, 1, 2,...
6
(3.71)

where k1  h( zi ), l1  hg ( xi , yi , zi )
 l   h k l 
k2  h  zi  1  , l2  hg  xi  , yi  1 , zi  1 
 2  2 2 2
 l   h k l 
k3  h  zi  2  , l3  hg  xi  , yi  2 , zi  2 
 2  2 2 2
k4  h( zi  l3 ), l4  hg ( xi  h, yi  k3 , zi  l3 )

Multistep Methods
We have seen that for finding the solution at each step, the Taylor series method
and Runge-Kutta methods requires evaluation of several derivatives. We shall
now develop the multistep method which require only one derivative evaluation
per step; but unlike the self starting Taylor series or Runge-Kutta methods, the
multistep methods make use of the solution at more than one previous step points.
Let the values of y and y1 already have been evaluated by self-starting methods
at a number of equally spaced points x0, x1,..., xn. We now integrate the differential
equation,

dy
 f ( x, y ), from xn to xn 1
dx
xn1 xn1

i.e.,  dy   f ( x, y ) dx
xn xn
xn 1

 yn 1  yn  
xn
f ( x, y ( x )) dx

To evaluate the integral on the right hand side, we consider f (x, y) as a function
of x and replace it by an interpolating polynomial, i.e., a Newton’s backward
difference interpolation using the (m + 1) points xn, xn+1, xn–2,..., xn–m,
m
x  xn
pm ( x )   (1)k (k s ) k f n  k , where s 
k 0 h
1
 
s
k   s ( s  1)( s  2)...( s  k  1) 
k!

Substituting pm(x) in place of f (x, y), we obtain

Self - Learning
166 Material
 1 m Numerical Differentiation
yn 1  yn  h  (1) k
 
s
k
k
f n  k ds and Integration
0 k 0

 yn  h [ f n   1f n 1   2  2 f n  2  ...   m f n  m ]
1
NOTES
where  k  (1)k    ds
s
k
0

The coefficients k can be easily computed to give,

1 5 3 251
 0  1,  1  ,  2  ,  3  ,  4  , etc.
2 12 8 720
Taking m = 3, the above formula gives,
 1 5 3 
y n 1  y n  h  f n  f n 1  2 f n  2  3 f n 3 
 2 12 8 

Substituting the expression of the differences in terms of function values given

by,
f n 1  f n  f n 1 , 2 f n  2  f n  2 f n 1  f n  2
3 f n 3  f n  3 f n 1  3 f n  2  f n 3
We get on arranging,
h
y n 1  y n  [55 f n  59 f n 1  37 f n  2  9 f n 3 ] (3.72)
24
This is known as Adams-Bashforth formula of order 4. The local error of
this formula is,
 s  3
1

E  h f (  ) 
5 iv
ds (3.73)
0
 4 

By using mean value theorem of integral calculus,

 s  3
1

E  h f ( ) 
5 iv
ds
0
 4 
or, 251 (3.74)
E  h5 f iv ( ).
720
The fourth order Adams-Bashforth formula requires four starting values, i.e.,
the derivaties, f3, f2, f1 and f0. This is a multistep method.

Predictor-Correction Methods
These methods use a pair of multistep numerical integration. The first is the Predictor
formula, which is an open-type explicit formula derived by using, in the integral, an
interpolation formula which interpolates at the points xn, xn–1,..., xn–m. The second
is the Corrector formula which is obtained by using interpolation formula that
interpolates at the points xn+1, xn, ..., xn–p in the integral.

Self - Learning
Material 167
Numerical Differentiation Euler’s Predictor-Corrector Formula
and Integration
The simplest formula of the type is a pair of formula given by,
y n( p1)  y n  h f ( xn , yn ) (3.75)
NOTES
(c) h ( p)
y n 1  y n  [ f ( xn , y n )  f ( xn 1 , y n 1 )] (3.76)
2
In order to determine the solution of the problem upto a desired accuracy, the
corrector formula can be employed in an iterative manner as shown below:
Step 1: Compute yn(0)1 , using Equation (3.75)

i.e., yn(0)1 = yn+ h f (xn, yn)

Step 2: Compute yn( k)1 using Equation (3.76)

h
i.e., yn 1  yn 
(k )
[ f ( xn , yn )  f ( xn 1 , yn( k11) )], for K  1, 2, 3,...,
2
The computation is continued till the condition given below is satisfied,
yn( k1)  yn( k11)
 (3.77)
yn( k1)
where  is the prescribed accuracy.
It may be noted that the accuracy achieved will depend on step size h and on
the local error. The local error in the predictor and corrector formula are,
h2 h3
y  ( 1 ) and  y  ( 2 ), respectively..
2 12

Milne’s Predictor-Corrector Formula

A commonly used Predictor-Corrector system is the fourth order Milne’s
Predictor-Corrector formula. It uses the following as Predictor and Corrector.
( p) 4h *
y n 1  y n 3  (2 f n  f n 1  2 f n  2 )
3
h
y n( c)1  y n 1  [ f n 1  4 f n  f n 1 ( xn 1 , y n( p1) )] (3.78)
3
The local errors in these formulae are respectively,
14 5 ( v ) 1
h y (1 ) and  h5 y ( v )  2 ) (3.79)
45 90
Example 3.42: Compute the Taylor series solution of the problem
dy
 xy  1, y (0)  1, up to x5 terms and hence compute values of y(0.1), y (0.2)
dx
and y (0.3). Use Milne’s Predictor-Corrector method to compute y (0.4) and y
(0.5).
Solution: We have y   xy  1, with y (0) = 1,  y (0)  1
Differentiating successively, we get
y ( x)  xy   y  y (0) 1
y ( x)  xy   y   y (0) 2
(iv ) (iv )
Self - Learning y ( x)  xy   3 y  y (0)  3
(v) (iv ) (iv )
168 Material y ( x)  xy  4 y   y (0)  8
 Thus the Taylor series solution is given by, Numerical Differentiation
and Integration
x2 x3 x 4 iv x5 ( v )
y ( x)  y (0)  xy (0)  y (0)  y (0)  y (0)  y (0)
2 3! 4! 5!
x 2 x3 x4 x5 NOTES
 1 x   2 3 8
2 3! 4! 5!
2 3 4 5
x x x x
 y ( x)  1  x    
2 3 8 15
0.01 0.001 0.0001 0.00001
 y (0.1)  1  0.1    
2 3 8 15
 1.1053
0.04 0.008 0.0016 0.00032
y (0.2)  1  0.2    
2 3 8 15
 1.22288
0.09 0.027 0.0081 0.00243
y (0.3)  1  0.3    
2 3 8 15
 1.35526

For application of Milne’s Predictor-Corrector method, we compute y  (0.1),

y  (0.2) and y  (0.3).

y (0.1)  0.1  1.1053  1  1.11053

y (0.2)  0.2  1.22288  1  1.244576
y (0.3)  0.3  1.35526  1  1.40658

4h
The Predictor formula gives, y4 = y(0.4) = y0+ (2 y1  y 2  2 y3 ) .
3

4  0.1
 y4(0)  1  (2  1.11053  1.24458  2  1.40658)
3
 1.50528  y4  1  0.4  1.50528  1.602112
(1) h
The Corrector formula gives, y 4  y 2  ( y 2  4 y3  y 4 ) .
3
0.1
y (0.4)  1.22288  (1.24458  4  1.40658  1.60211)
3
 1.22288  0.28243
 1.50531

3.4.4 Higher Order Differential Equations

We consider the solution of ordinary differential equation of order 2 or more,
when value of the dependent variable is given at more than one point, usually at the
two ends of an interval in which the solution is required. For example, the simplest
boundary value problem associated with a second order differential equation is,
y +p (x) y  +q (x)y = r (x) (3.80)
with boundary conditions, y (a) = A, y (b) = B. (3.81)
The following two methods reduce the boundary value problem into initial
value problems which are then solved by any of the methods for solving such Self - Learning
Material 169
problems.
Numerical Differentiation Reduction to a Pair of Initial Value Problem
and Integration
This method is applicable to linear differential equations only. In this method, the
solution is assumed to be a linear combination of two solutions in the form,
NOTES y(x) = u(x) +  (x) (3.82)
where  is a suitable constant determined by using the boundary condition
and u(x) and (x) are the solutions of the following two initial value problems:
(i) u  + p(x) u  + q(x)u = r(x)
u(a) = A, u  (a) = 1, (say). (3.83)
(ii)  + P(X)  + Q(X)  = R(X)
 (a) = 0 and (a) =  2 , (say) (3.84)
where 1 and 2 are arbitrarily assumed constants. After solving the two initial
value problems, the constant  is determined by satisfying the boundary condition
at x = b. Thus,
B  u (b)   (b)
or, B   (b) (3.85)
 , provided  (b)  0
 (b)
Evidently, y(a) = A, is already satisfied.
If (b) = 0, then we solve the initial value problem for  again by choosing
(a) = 3, for some other value for which (b) will be non-zero.
Another method which is commonly used for solving boundary problems is
the finite difference method discussed below.
Finite Difference Method
In this method of solving boundary value problem, the derivatives appearing in the
differential equation and boundary conditions, if necessary, are replaced by
appropriate difference gradients.
Consider the differential equation, y +p(x) y  + q(x)y = r(x) (3.86)
with the boundary conditions, y (a) =  and y (b) = 
(3.87)
The interval [a, b] is divided into N equal parts each of width h, so that
h = (b–a)/N, and the end points are x0 = a and xn = b. The interior mesh points xi
at which solution values y(xi) are to be determined are,
xn = x0+ nh, n = 1, 2, ..., N – 1
(3.88)
The values of y at the mesh points is denoted by yn given by,
yn = y (x0+ nh), n = 0, 1, 2, ..., N
(3.89)
The following central difference approximations are usually used in finite
difference method of solving boundary value problem,
y n 1  y n 1
y ( xn ) 
Self - Learning
2h
170 Material (3.90)
 Numerical Differentiation
y n 1  2 y n  y n 1 and Integration
y ( xn )  2
h
(3.91)
Substituting these in the differential equation, we have NOTES
2(yn+1–2yn+ yn–1) + pn h(yn+1– yn–1) + 2h2gnyn = 2rnh2,
where pn = p(xn), qn = q(xn), rn = r(xn) (3.92)
Rewriting the equation by regrouping we get,
(2–hpn)yn–1+(–4+2h2qn)yn+(2+h2qn)yn+1 = 2rnh2
(3.93)
This equation is to be considered at each of the interior points, i.e., it is true for
n = 1, 2, ..., N–1.
The boundary conditions of the problem are given by,
y0   , yn   (3.94)
Introducing these conditions in the relevant equations and arranging them, we
have the following system of linear equations in (N–1) unknowns y1, y2, ..., yn–1.

(4  2h 2 q1 ) y1  (2  hp1 ) y2  2r1h 2  (2  hp1 )

(2  hp2 ) y1  (4  2h 2 q2 ) y2  (2  hp2 ) y3  2r2 h 2
(2  hp3 ) y2  ( 4  2h 2 q3 ) y3  (2  hp3 ) y4  2r3 h 2
... ... ... ... ...
(2  hpN  2 )  ( 4  2h 2 qN  2 ) y N  2 (2  hpN  2 ) yN 1  2rN  2 h 2
(2  hpN 1 ) yN  2  ( 4  2h 2 q N 1 ) yN 1  2rN 1h 2  (2  hpN 1 ) (3.95)
The above system of N–1 equations can be expressed in matrix notation in the
form
Ay  b
(3.96)
where the coefficient matrix A is a tridiagonal one, of the form

 B1 C1 0 0... 0 0 0 
A B2 C2 0... 0 0 0 
 2 
0 A3 B3 C3 ... 0 0 0 
A  (3.97)
 ... ... ... ... ... ... ... 
0 0 0 0... AN  2 BN 2 C N 2 
 
 0 0 0 0... 0 AN 1 B N 1 

Where, Bi  4  2h 2 qi , i  1, 2,..., N  1
Ci  2  hpi , i  1, 2,..., N  2
Ai  2  hpi , i  2, 3,..., N  1

(3.98)
Self - Learning
Material 171
Numerical Differentiation The vector b has components,
and Integration

b1  2  1h 2  (2  hp1 )
bi  2  i h 2 , for i  2, 3,..., N  2
NOTES bN 1  2  N 1  h 2  (2  hlp N 1 )
(3.99)
The system of linear equations can be directly solved using suitable methods.
Example 3.43: Compute values of y (1.1) and y (1.2) on solving the following
initial value problem, using Runge-Kutta method of order 4:
y
y    y  0 , with y(1) = 0.77, y  (1) = –0.44
x
Solution: We first rewrite the initial value problem in the form of pair of first order
equations.
z
y   z, z   y
x
with y (1) = 0.77 and z (1) = –0.44.
We now employ Runge-Kutta method of order 4 with h = 0.1,
1
y (1.1)  y (1)  (k1  2k2  2k3  k4 )
6
1
y (1.1)  z (1.1)  1  (l1  2l2  2l3  l4 )
6
k1  0.44  0.1  0.044
 0.44 
l1  0.1    0.77   0.033
 1 
 0.033 
k2  0.1   0.44    0.04565
 2 
 0.4565 
l2  0.1    0.748   0.031323809
 1.05 

 0.03123809 
k3  0.1   0.44    0.0455661904
 2 
 0.0455661904 
l3  0.1    0.747175   0.031321128
 1.05 
k4  0.1  (0.47132112)  0.047132112
 0.047132112 
l4  0.1    0.72443381  0.068158643
 1.1 
1
 y (1.1)  0.77  [0.044  2  (0.045661904)  0.029596005]  0.727328602
6
1
y (1.1)  0.44  [0.033  2(0.031323809)  2(0.031321128)  0.029596005]
6
1
 0.44  [0.33  0.062647618  0.062642256  0.029596005]
6
Self - Learning  0.526322021
172 Material
Example 3.44: Compute the solution of the following initial value problem for Numerical Differentiation
and Integration
x = 0.2, using Taylor series solution method of order 4: n.l.
d2 y dy
2
 y  x , y(0)  1, y(0)  0
dx dx NOTES
Solution: Given y   y  xy , we put z  y , so that
z   y  xz , y   z and y (0)  1, z (0)  0.
We solve for y and z by Taylor series method of order 4. For this we first
compute y (0), y (0), y iv (0),...
We have, y (0)  y(0)  0  y (0)  1, z (0)  1
y (0)  z (0)  y (0)  z (0)  0.z (0)  0
y iv (0)  z (0)  y (0)  2 z (0)  0.z (0)  3
z iv (0)  4 z (0)  0.z (0)  0
By Taylor series of order 4, we have
x2 x3 x 4 iv
y (0  x)  y (0)  xy (0)  y (0)  y (0)  y (0)
2! 3! 4!
x2 x4
or, y ( x)  1   3
2! 4!
(0.2) 2 (0.2) 4
 y (0.2)  1    1.0202
2! 8
(0.2)3
Similarly, y  (0.2)  z (0.2)  0.2   3  0.204
4!
Example 3.45: Compute the solution of the following initial value problem for
2
d y
x = 0.2 by fourth order Runge -Kutta method: n.l.  xy, y (0)  1, y (0)  1
dx 2
Solution: Given y   xy, we put y   z and the simultaneous first order problem,
y   z  f ( x, y , z ), say z   xy  g ( x, y , z ), say with y (0)  1 and z (0)  1
We use Runge-Kutta 4th order formulae, with h = 0.2, to compute y (0.2)
and y (0.2), given below..
k1  h f ( x0 , y0 , z0 )  0.2  1  0.2
l1  h g ( x0 , y0 , z0 )  0.2  0  0
 h k l
k2  h f  x0  , y0  1 , z0  1   0.2  (1  0)  0.2
 2 2 2
 h k l 0.2  0.2 
l2  h g  x0  , y0  1 , z0  1   0.2  1    0.022
 2 2 2 2  2 
 h k l 
k3  h f  x0  , y0  2 , z0  2   0.2  1.011  0.2022
 2 2 2
 h k l 
l3  h g  x0  , y0  2 , z0  2   0.2  0.1  1.1  0.022
 2 2 2
k4  h f ( x0  h, y0  k3 , z0  l3 )  0.2  1.022  0.2044
l4  h g ( x0  h, y0  k3 , z0  l3 )  0.2  0.2  1.2022  0.048088
1
y (0.2)  1  (0.2  2(0.2  0.2022)  0.2044)  1.2015
6
1 Self - Learning
y (0.2)  1  (0  2 (0.022  0.022)  0.048088)  1.02268
6 Material 173
Numerical Differentiation
and Integration
Check Your Progress
10. How are Euler's method and Taylor's method related?
NOTES 11. Define Picard's method of successive approximation.
12. Why should we not use Euler's method for a larger range of x?
13. When are Runge-Kutta methods applied?
14. What is a predictor formula?
15. What are local errors in Milne's predictor-corrector formulae?
16. Where can the method of reduction to a pair of initial value problem be
applied?

3.5 ANSWERS TO ‘CHECK YOUR PROGRESS’

1. Numerical differentiation is the process of computing the derivatives of a
function f(x) when the function is not explicitly known, but the values of the
function are known for a given set of arguments x = x0, x1, x2, ..., xn. To
find the derivatives, we use a suitable interpolating polynomial and then its
derivatives are used as the formulae for the derivatives of the function.
2. Newton’s forward difference interpolation formula is,
u (u  1) 2 u (u  1)(u  2) 3 u (u  1)(u  2)...(u  n  1) n
 (u )  y0  u  y0   y0   y0  ...   y0
2 ! 3 ! n !

where u  x  x0
h
3. Newton’s backward difference interpolation formula is,
v(v  1) 2 v(v  1)(v  2) 3 v (v  1)(v  2)(v  3) 4
(v )  yn  v yn   yn   yn   yn  ...
2 ! 3 ! 4 !
v(v  1)...(v  n  1) n
  yn
n !
x  xn
Where v 
h
4. The evaluation of a definite integral cannot be carried out when the integrand
f(x) is not integrable, as well as when the function is not explicitly known
but only the function values are known at a finite number of values of x.
There are two types of numerical methods for evaluating a definite integral
based on the following formula:
b

 f ( x) dx
a

x1
h
5. The formula is,  f ( x)dx  2 [ f
x0
0  f1 ] .

x2
h
6. The formula is,  f ( x)dx  3 [ f
x0
0  4 f1  f 2 ] .

Self - Learning
174 Material
 Numerical Differentiation
b
3h and Integration
7. Simpson’s three-eighth rule of numerical integration is,  f ( x) dx  [y
a
8 0
+ 3y1 + 3y2 + 2y3 + 3y4 + 3y5 + 2y6 + …+ 2y3m – 3 + 3y3m – 2 + 3y3m – 1 + y3m]
where h = (b–a)/(3m); for m = 1, 2, ... NOTES
b
3h
8. The Weddle’s rule is,  f ( x)dx  10
a
[y0 + 5y1 + y2 + 6y3 + y4 + 5y5 + 2y6 +
5y7 + y8 + 6y9 + y10 + 5y11 + ... + 2y6m – 6 + 5y 6m – 5 + y6m – 4 + 6y6m – 3 + y6m
–2
+ 5y6m – 1 + y6m], where b – a = 6mh.
9. This procedure is used to find a better estimate of an integral using the
evaluation of the integral for two values of the width of the sub-intervals.
10. If we take k = 1, we get the Euler’s method, y1 = y0 + h f(x0, y0).
11. In Picard’s method the first approximate solution y (1) ( x ) is obtained by
x
replacing y(x) by y 0. Thus, y (1) ( x )  y0   f ( x, y0 )dx . The second
x0

approximate solution is derived on replacing y by y(1) (x). Thus,

x

 f ( x, y
( 2) (1)
y ( x)  y0  ( x)) dx
.
x0

This iteration formula is known as Picard’s iteration for finding solution of a

first order differential equation, when an initial condition is given. The iterations
are continued until two successive approximate solutions yk and yk + 1 give
approximately the same result for the desired values of x up to a desired
accuracy.
12. The method should not be used for a larger range of x about x0, since the
propagated error grows as integration proceeds.
13. Runge-Kutta methods are very useful when the method of Taylor series is
not easy to apply because of the complexity of finding higher order
derivatives.
14. A predictor formula is an open-type explicit formula derived by using, in the
integral, an interpolation formula which interpolates at the points xn, xn – 1,
..., xn – m.
14 5 (v ) 1
15. The local errors in these formulae are h y (1 ) and  h 5 y (v )  2 ) .
45 90
16. This method is applicable to linear differential equations only.

3.6 SUMMARY
 Numerical differentiation is the process of computing the derivatives of a
function f(x) when the function is not explicitly known, but the values of the
function are known only at a given set of arguments x = x0, x1, x2,..., xn. Self - Learning
Material 175
Numerical Differentiation  For finding the derivatives, we use a suitable interpolating polynomial and
and Integration
then its derivatives are used as the formulae for the derivatives of the function.
 For computing the derivatives at a point near the beginning of an equally
spaced table, Newton’s forward difference interpolation formula is used,
NOTES
whereas Newton’s backward difference interpolation formula is used for
computing the derivatives at a point near the end of the table.
 Let the values of an unknown function y = f(x) be known for a set of
equally spaced values x0, x1, …, xn of x, where xr = x0 + rh. Newton’s
forward difference interpolation formula is,

u (u  1) 2 u (u  1)(u  2) 3 u (u  1)(u  2)...(u  n  1) n

 (u )  y0  u  y0   y0   y0  ...   y0
2 ! 3 ! n !

x  x0
where u  .
h
 At the tabulated point x0, the value of u is zero and the formulae for the
derivatives are given by,

1 1 1 1 1 
y ( x0 )   y0  2 y0  3 y0  4 y0  5 y0  ...
h 2 3 4 5 

1  2 11 5 
y ( x0 )  2 
 y 0  3 y 0  4 y 0  5 y 0  ...
h  12 6 

dy d2y
 For a given x near the end of the table, the values of and 2 are
dx dx
computed by first computing v = (x – xn)/h and using the above formulae.
At the tabulated point xn, the derivatives are given by,

1 1 1 1 
y ( x n )  y n   2 y n   3 y n   4 y n  ...
h 2 3 4 

1  2 11 5 
y ( xn )  2 
 y n   3 y n   4 y n   5 y n  ...
h  12 6 
 For computing the derivatives at a point near the middle of the table, the
derivatives of the central difference interpolation formula is used.
 If the arguments of the table are unequally spaced, then the derivatives of
the Lagrange’s interpolating polynomial are used for computing the derivatives
of the function.
 Numerical differentiation is the process of computing the derivatives of a
function f(x) when the function is not explicitly known, but the values of the
function are known only at a given set of arguments x = x0, x1, x2, ..., xn.
 For computing the derivatives at a point near the beginning of an equally
spaced table, Newton’s forward difference interpolation formula is used,
Self - Learning
176 Material
 whereas Newton’s backward difference interpolation formula is used for Numerical Differentiation
and Integration
computing the derivatives at a point near the end of the table.
 Numerical methods can be applied to determine the value of the integral
when the integrand is not integrable as well as when the function is not
NOTES
explicitly known but only the function values are known.
 The two types of numerical methods for evaluating a definite integral are
Newton-Cotes quadrature and Gaussian quadrature.
 Taking n = 2 in the Newton-Cotes formula, we get Simpson’s one-third
formula of numerical integration while taking n = 3, we get Simpson’s three-
eighth formula of numerical integration.
 In Newton-Cotes formula with n = 6 some minor modifications give the
Weddle’s formula.
 For evaluating a definite integral correct to a desired accuracy, one has to
make a suitable choice of the value of h, the length of sub-interval to be
used in the formula.
 There are two ways of determining h, by considering the truncation error in
the formula to be used for numerical integration or by successive evaluation
of the integral by the technique of interval halving and comparing the results.
 In the truncation error estimation method, the value of h to be used is
determined by considering the truncation error in the formula for numerical
integration.
 When the estimation of the truncation error is cumbersome, the method of
interval halving is used to compute an integral to the desired accuracy.
 Numerical evaluation of double integrals is done by applying trapezoidal
rule and Simpson’s one-third rule.
 This procedure is used to find a better estimate of an integral using the
evaluation of the integral for two values of the width of the sub-intervals.
 There are many methods available for finding a numerical solution for
differential equations.
 Picard’s iteration is a method of finding solutions of a first order differential
equation when an initial condition is given.
 Euler’s method is a crude but simple method for solving a first order initial
value problem.
 Euler’s method is a particular case of Taylor’s series method.
 Runge-Kutta methods are useful when the method of Taylor series is not
easy to apply because of the complexity of finding higher order derivatives.
 For finding the solution at each step, the Taylor series method and Runge-
Kutta methods require evaluation of several derivatives.
 The multistep method requires only one derivative evaluation per step; but
unlike the self starting Taylor series on Runge-Kutta methods, the multistep
methods make use of the solution at more than one previous step points.
 These methods use a pair of multistep numerical integration. The first is the
predictor formula, which is an open-type explicit formula derived by using,
in the integral, an interpolation formula which interpolates at the points
Self - Learning
Material 177
Numerical Differentiation xn, xn – 1, ..., xn – m. The second is the corrector formula which is obtained by
and Integration
using interpolation formula that interpolates at the points xn + 1, xn, ..., xn – p in
the integral.
 The solution of ordinary differential equation of order 2 or more, when
NOTES values of the dependent variable is given at more than one point, usually at
the two ends of an interval in which the solution is required.
 The methods used to reduce the boundary value problem into initial value
problems are reduction to a pair of initial value problem and finite difference
method.

3.7 KEY TERMS

 Numerical differentiation: It is the process of computing the derivatives
of a function f(x) when the function is not explicitly known, but the values of
the function are known for a given set of arguments x = x0, x1, x2, ..., xn.
 Newton’s forward difference interpolation formula: The Newton’s
forward difference interpolation formula is used for computing the derivatives
at a point near the beginning of an equally spaced table.
 Newton’s backward difference interpolation formula: Newton’s
backward difference interpolation formula is used for computing the
derivatives at a point near the end of the table.
 Central difference interpolation formula: For computing the derivatives
at a point near the middle of the table, the derivatives of the central difference
interpolation formula is used.
 Newton-Cotes quadrature: This is based on integrating polynomial
interpolation formulae and requires a table of values of the integrand at
equally spaced values of the independent variable x.
 Trapezoidal formula: The trapezoidal formula of numerical integration is
defined using the definite integral of the function f (x) between the limits x0
to x1, as it is approximated by the area of the trapezoidal region bounded
by the chord joining the points (x0, f0) and (x1, f1), the x-axis and the
ordinates at x = x0 and at x = x1.
 Romberg’s procedure: This procedure is used to find a better estimate of
an integral using the evaluation of the integral for two values of the width of
the sub-intervals.
 Weddle’s rule: It is a composite Weddle’s formula and is used when the
number of sub-intervals is multiple of 6.
 Predictor formula: It is an open-type explicit formula derived by using, in
the integral, an interpolation formula which interpolates at the points xn,
xn – 1, ..., xn – m.
 Corrector formula: It is obtained by using interpolation formula that
interpolates at the points xn + 1, xn, ..., xn – p in the integral.

Self - Learning
178 Material
 Numerical Differentiation
3.8 SELF-ASSESSMENT QUESTIONS AND and Integration

EXERCISES
NOTES
Short-Answer Questions
1. Define the term numerical differentiation.
2. Give the differentiation formula for Newton’s forward difference interpolation.
dy
3. How the derivative can be evaluated?
dx
4. Give the formulae for the derivatives at the tabulated point x0 where the
value of u is zero.
5. Give the differentiation formula for Newton’s backward difference
interpolation.
6. Give the Newton’s backward difference interpolation formula for an equally
spaced table of a function.
7. State Newton-Cotes formula.
8. State the trapezoidal rule.
9. What is the difference between Simpson’s one-third formula and one-third
rule?
10. What is the error in Weddle’s rule?
11. Give the truncation error in Simpson’s one-third rule.
12. Where is interval halving technique used?
13. Name the methods used for numerical evaluation of double integrals.
14. State the Gauss quadrature formula.
15. State an application of Romberg’s procedure.
16. What are ordinary differential equations?
17. Name the methods for computing the numerical solution of differential
equations.
18. What is the significance of Runge-Kutta methods of different orders?
19. When is multistep method used?
20. Name the predictor-corrector methods.
21. How will you find the numerical solution of boundary value problems?
Long-Answer Questions
1. Discuss numerical differentiation using Newton’s forward difference
interpolation formula and Newton’s backward difference interpolation
formula.

Self - Learning
Material 179
Numerical Differentiation
3
and Integration
2. Use the following table of values to compute  f ( x ) dx :
0

NOTES x 0 1 2 3
f ( x) 1.6 3.8 8.2 15.4

3. Use suitable formulae to compute y (1.4) and y (1.4) for the function y =
f(x), given by the following tabular values:
x 1.4 1.8 2.2 2.6 3.0
y 0.9854 0.9738 0.8085 0.5155 0.1411

dy d2y
4. Compute and for x =1 where the function y = f(x) is given by the
dx dx 2
following table:
x 1 2 3 4 5 6
y 1 8 27 64 125 216

5. A rod is rotating in a plane about one of its ends. The following table gives
the angle  (in radians) through which the rod has turned for different values
d
of time t seconds. Find its angular velocity and angular acceleration
dt
d 2
at t = 1.0.
dt 2

t secs 0.0 0.2 0.4 0.6 0.8 1.0

 radians 0.0 0.12 0.48 1.10 2.00 3.20

dy d2y
6. Find and at x = 1 and at x = 3 for the function y = f(x), whose
dx dx 2
values in [1, 6] are given in the following table:
x 1 2 3 4 5 6
y 2.7183 3.3210 4.0552 4.9530 6.0496 7.3891

dy d2y
7. Find and at x = 0.96 and at x = 1.04 for the function y = f(x)
dx dx 2
given in the following table:
x 0.96 0.98 1.0 1.02 1.04
y 0.7825 0.7739 0.7651 0.7563 0.7473

8. Use suitable formulae to compute y(1.4) and y(1.4) for the function y = f
(x), given by the following tabular values.
x 1.4 1.8 2.2 2.6 3.0
y 0.9854 0.9738 0.8085 0.5155 0.1411

Self - Learning
180 Material
 dy d2y Numerical Differentiation
9. Compute dx and dx 2
for x = 1 where the function y = f (x) is given by the and Integration

following table:
x 1 2 3 4 5 6
NOTES
y 1 8 27 64 125 216

20
10. Compute  f ( x) dx by Simpson’s one-third rule, where:
0

x 0 5 10 15 20
f ( x) 1.0 1.6 3.8 8.2 15.4

4
11. Compute  x 3 dx by Simpson’s one-third formula and comment on the result:
0

x 0 2 4
x3 0 8 64

12. Compute  x 3 dx by Simpson’s one-third formula and comment on the result:

0

2
13. Compute  e x dx by Simpson’s one-third formula and compare with the exact
0
value, where e0 = 1, e1 = 2.72, e2 = 7.39.
1
dx
14. Compute an approximate value of  , by integrating
 1 x
0
2
, by Simpson’ss

one-third formula.
15. A rod is rotating in a plane about one of its ends. The following table gives
the angle  (in radians) through which the rod has turned for different values
d
of time t seconds. Find its angular velocity and angular acceleration
dt
d 2
at t = 1.0.
dt 2

t secs 0.0 0.2 0.4 0.6 0.8 1.0

θ radius 0.0 0.12 0.48 1.10 2.00 3.20

dy d2y
16. Find and at x = 1 and at x = 3 for the function y = f (x), whose
dx dx 2
values are given in the following table:
x 1 2 3 4 5 6
y 2.7183 3.3210 4.0552 4.9530 6.0496 7.3891

Self - Learning
Material 181
Numerical Differentiation
and Integration dy d2y
17. Find and at x = 0.96 and at x = 1.04 for the function y = f (x)
dx dx 2
given in the following table:
NOTES x 0.96 0.98 1.0 1.02 1.04
y 0.7825 0.7739 0.7651 0.7563 0.7473

18. Compute  ( x  1)dx, by trapezoidal rule by taking four sub-intervals and

0

comment on the result by comparing it with the exact value.

1.4

19. Compute  ( x3  2)dx, by Simpson’s one-third rule by taking four sub-intervals

1

and find the error in the result.

1

20. Evaluate  cos x dx, correct to three significant figures taking five equal sub-
0

intervals.
1
xdx
21. Compute the value of the integral  correct to three significant figures
1 x 0

by Simpson’s one-thrid rule with six sub-intervals.

1
dx
22. Compute the integral  , by Simpson’s one-third rule taking four sub-
1 x 0

intervals and use it to compute the approximate value of .

4

23. Compute  e x dx by Simpson’s rule correct to four significant digits taking

0

four sub-intervals and compare it with the exact value.

2
dx
24. Compute the approximate value of  by Simpson’s one-third rule with
1 x
four sub-intervals.
1

25. Evaluate  1  x3 dx by trapezoidal rule, taking four sub-intervals; give the

0

result upto four decimal places.

26. Compute the following integral by Simpson’s one-third rule taking h = 0.05
1.3
correct to five significant digits  x dx. ,
1


2
27. Compute the integral  sin x dx by (a) Trapezoidal rule and (b) Simpson’ss
0

one-third rule taking six sub-intervals and compare the results with the exact
value.
28. Evaluate the following integrals by Weddle’s rule,
1
dx
(a)  1  x 2 , taking n = 12
0

1
x2  1
Self - Learning (b)  x 2  1 dx, taking n = 12
182 Material 0
 1 Numerical Differentiation
29. Compute  xe  x dx by Gauss-Legendre two point and three point formula, and Integration
0

and compare with the exact value.

30. Evaluate the following integrals by Gauss-Legendre three point formula:
NOTES

2
(a)  sin x dx
0
1

(b)  cos
2
xe x dx
0
1
dx
(c) 
1  x2
0

31. Illustrate Romberg’s procedure.

32. Use Picard’s method to compute values of y(0.1), y(0.2) and y(0.3) correct
to four decimal places, for the problem, y = x + y, y(0) = 1.
33. Compute values of y at x = 0.02, by Euler’s method taking h = 0.01, given
dy
y is the solution of the following initial value problem: dx
= x3 + y, y(0) = 1.
34. Evaluate y(0.02) by modified Euler’s method, given y = x2 + y, y(0) = 1,
correct to four decimal places.
1
35. Given y = , y(4) = 4, compute y(4.2) by Taylor series method,
x2  y
taking h = 0.1.
36. Using Runge-Kutta method of order 4, compute y(0.1) for each of the
following problems:
dy
(a)  x  y, y (0)  1
dx
dy
(b)  x  y 2 , y (0)  1
dx
37. Compute solution of the following initial value problem by Runge-Kutta
method of order 4 taking h = 0.2 upto x = 1; y = x – y, y(0) = 1.5.
dy 1
38. Given  (1  x 2 ) y 2 , and y(0) = 1, y(0.1) = 1.06, y(0.2) = 1.12, y(0.3)
dx 2
= 1.21. Compute y(0.4) by Milne’s predictor-corrector method.

3.9 FURTHER READING

Chance, William A. 1969. Statistical Methods for Decision Making. Illinois:
Richard D Irwin.
Chandan, J.S., Jagjit Singh and K.K. Khanna. 1995. Business Statistics. New
Delhi: Vikas Publishing House.
Elhance, D.N. 2006. Fundamental of Statistics. Allahabad: Kitab Mahal.
Freud, J.E., and F.J. William. 1997. Elementary Business Statistics – The
Modern Approach. New Jersey: Prentice-Hall International. Self - Learning
Material 183
Numerical Differentiation Goon, A.M., M.K. Gupta, and B. Das Gupta. 1983. Fundamentals of Statistics.
and Integration
Vols. I & II, Kolkata: The World Press Pvt. Ltd.
Gupta, S.C. 2008. Fundamentals of Business Statistics. Mumbai: Himalaya
Publishing House.
NOTES
Kothari, C.R. 1984. Quantitative Techniques. New Delhi: Vikas Publishing
House.
Levin, Richard. I., and David. S. Rubin. 1997. Statistics for Management. New
Jersey: Prentice-Hall International.
Meyer, Paul L. 1970. Introductory Probability and Statistical Applications.
Massachusetts: Addison-Wesley.
Gupta, C.B. and Vijay Gupta. 2004. An Introduction to Statistical Methods,
23rd Edition. New Delhi: Vikas Publishing House Pvt. Ltd.
Hooda, R. P. 2013. Statistics for Business and Economics, 5th Edition. New
Delhi: Vikas Publishing House Pvt. Ltd.
Anderson, David R., Dennis J. Sweeney and Thomas A. Williams. Essentials of
Statistics for Business and Economics. Mumbai: Thomson Learning,
2007.
S.P. Gupta. 2021. Statistical Methods. Delhi: Sultan Chand and Sons.

Self - Learning
184 Material
 Statistical Computation and

UNIT 4 STATISTICAL COMPUTATION Probability Distributiona

AND PROBABILITY
DISTRIBUTION NOTES

Structure
4.0 Introduction
4.1 Objectives
4.2 History and Meaning of Statistics
4.2.1 Scope of Statistics
4.3 Various Measures of statistical computations
4.3.1 Average
4.3.2 Mean
4.3.3 Median
4.2.4 Mode
4.3.5 Geometric Mean
4.3.6 Harmonic Mean
4.3.7 Quartiles, Percentiles and Deciles
4.3.8 Box Plot
4.4 Measures of Dispersion
4.4.1 Range
4.4.2 Quartile Deviation
4.4.3 Mean Deviation
4.5 Standard Deviation
4.5.1 Calculation of Standard Deviation by Short-cut Method
4.5.2 Combining Standard Deviations of Two Distributions
4.5.3 Comparison of Various Measures of Dispersion
4.6 Probability
4.6.1 Probability Distribution of a Random Variable
4.6.2 Axiomatic or Modern Approach to Probability
4.6.3 Theorems on Probability
4.6.4 Counting Techniques
4.6.5 Mean and Variance of Random Variables
4.7 Standard Probability Distribution
4.7.1 Binomial Distribution
4.7.2 Poisson Distribution
4.7.3 Exponential Distribution
4.7.4 Normal Distribution
4.7.5 Uniform Distribution (Discrete Random and Continous Variable)
4.8 Answers to ‘Check Your Progress’
4.9 Summary
4.10 Key Terms
4.11 Self-Assessment Questions and Exercises
4.12 Further Reading

4.0 INTRODUCTION
Statistics is the discipline that concerns the collection, organization, analysis,
interpretation, and presentation of data. Every day we are confronted with some
form of statistical information through different sources. All raw data cannot be Self - Learning
Material 185
Statistical Computation and termed as statistics. Similarly, single or isolated facts or figures cannot be called
Probability Distribution
statistics as these cannot be compared or related to other figures within the same
framework. Hence, any quantitative and numerical data can be identified as statistics
when it possesses certain identifiable characteristics according to the norms of
NOTES statistics.
In statistics, the term statistical computation specifies the method through
which the quantitative data have a tendency to cluster approximately about some
value. A measure of statistical computation is any precise method of specifying this
‘Central Value’. In the simplest form, the measure of statistical computation is an
average of a set of measurements, where the word average refers to as mean,
median, mode or other measures of location. Typically the most commonly used
measures are arithmetic mean, mode and median. The measures of dispersion,
which in itself is a very important property of a distribution and needs to be measured
by appropriate statistics. Hence, this unit has taken into consideration several aspects
of dispersion. It describes absolute and relative measures of dispersion. It deals
with range, the crudest measure of dispersion. It also explains quartile deviation,
mean deviation and standard deviation. The standard deviation is the most useful
measure of dispersion.
The subject of probability in itself is a cumbersome one, hence only the
basic concepts will be discussed in this unit. The word probability or chance is
very commonly used in day-to-day conversation, and terms such as possible or
probable or likely, all have similar meanings. Probability can be defined as a measure
of the likelihood that a particular event will occur. It is a numerical measure with a
value between 0 and 1 of such likelihood where the probability of zero indicates
that the given event cannot occur and the probability of one assures certainty of
such an occurrence. The probability theory helps a decision-maker to analyse a
situation and decide accordingly. We study why all these uncertainties require
knowledge of probability so that calculated risks can be taken. Since the outcomes
of most decisions cannot be accurately predicted because of the impact of many
uncontrollable and unpredictable variables, it is necessary that all the known risks
be scientifically evaluated. Probability theory, sometimes referred to as the science
of uncertainty, is very helpful in such evaluations. It helps the decision-maker with
only limited information to analyse the risks and select the strategy of minimum
risk.
The probability distribution of a discrete random variable is a list of
probabilities associated with each of its possible values. It is also sometimes called
the probability function or the probability mass function. The probability density
function of a continuous random variable is a function which can be integrated to
obtain the probability that the random variable takes a value in a given interval.
The binomial distribution is used in finite sampling problems where each observation
is one of two possible outcomes (‘Success’ or ‘Failure’). The Poisson distribution
is used for modelling rates of occurrence. The exponential distribution is used to
describe units that have a constant failure rate. The term ‘Normal Distribution’
refers to a particular way in which observations will tend to pile up around a
particular value rather than be spread evenly across a range of values, i.e., the
Self - Learning central limit theorem.
186 Material
 In this unit, you will learn about the history and meaning of statistics, scope Statistical Computation and
Probability Distributiona
of statistics, various measures of statistical computations, measures of dispersion,
standard deviation, probability and standard probability distribution.

NOTES
4.1 OBJECTIVES
After going through this unit, you will be able to:
 Examine the functions and meaning of statistics
 Understand the various measures of statistical data
 Analysis the absolute and relative measures of distribution
 Discuss the meaning, uses and merits of range in statistical presentation
 Define standard deviation
 Understand the basic concept of probability
 Do random experiment
 Explain about the concepts of probability distribution
 Describe the Poisson distribution
 Analyse Poisson distribution as an approximation of binomial distribution
 Understand exponential distribution
 Learn about the uniform distribution (discrete random and continuous
variable)

4.2 HISTORY AND MEANING OF STATISTICS

The term statistics is used to mean either statistical data or statistical method.
When it is used in the sense of statistical data it refers to quantitative aspects of
things, and is a numerical description. Thus, the distribution of family incomes is a
quantitative description, as also the annual production figures of various industries.
These quantities are numerical to begin with. But there are also some quantities
which are not in themselves numerical, but can be made so by counting. The sex
of a baby is not a number, but by counting the number of boys, we can associate
a numerical description to the sex of all new-born babies, for example, when
saying that 54 per cent of all live-born babies are boys. This information, then,
comes within the realm of statistics. Likewise, the statistics of students of a college
include count of, the number of students, and separate counts of numbers of various
kinds, as males and females, married and unmarried, of postgraduates and
undergraduates. They may also include such measurements as their heights and
weights. In addition, there may also be numbers computed on the basis of these
measurements or counts, e.g., the proportion of female students; their average
height or average weight. An example of statistical data is given in Table 4.1.

Self - Learning
Material 187
Statistical Computation and Table 4.1 Statistics of Students of a College where Total Number of Students is 1,000
Probability Distribution

Sex-wise distribution Class-wise distribution

NOTES Sex Number Percentage Course Number Percentage

Male 900 90 Undergraduate 800 80

Female 100 10 Postgraduate 200 20

Total 1,000 100 Total 1,000 100

Distribution According to Height

Height Number

From 160 cm to 170 600

From 170 cm to 180 390
From 180 cm to 190 10

Total 1,000

The other aspect of statistics is as a body of theories and techniques employed

in analysing the numerical information and using it to make informed decisions.
It is a branch of the scientific method used in dealing with those phenomena which
can be described numerically, either by measurements or by counts. For example,
if a preliminary test using a new vaccine shows that in a sample of 100 cases the
incidence of the disease is reduced to 8 while that in the unvaccinated population
it is 10, can we say if the vaccine is effective or not? Surely, we can select 100
individuals in the unvaccinated population in which the incidence is 8 or even less.
Is the observed reduction due to chance causes or does it show the effect of the
vaccine? The statistical method provides with theories and techniques for checking
this out. In this text, we will be primarily concerned with the statistical method, its
theories and its techniques.
Consider the following situations typical of what are faced by decision-
makers. In each of these the method of arriving at policy decision consists of first
understanding the parameters of the problem, for which statistical data is called
for.
 A large technical university experiences a fall in the number of persons
seeking admission. Is this fall due to factors which are peculiar to that
particular university, or is it a country-wide trend? The exact course of
policy action will depend upon the answer to this question.
 At what age should one retire airline pilots? Does the increase in age affect
the safety record ? How is it balanced by the increasing experience, if at
all?
 In certain circles it is taken as axiomatic that more the people travel in the
country, the more emotional integration is achieved. Recently there has been
some doubt cast on this premise. The implications of this are wide ranging,
Self - Learning
affecting the government attitude towards promotion of tourism. How does
188 Material one verify the truth of the contention?
  More than 50 per cent of heart-transplant patient die within a year. Is heart- Statistical Computation and
Probability Distributiona
transplant surgery beneficial? Would not a heart-patient be better off without
a transplant than with it? Would he live longer?
 Is advertising on television more cost effective than advertising in print? Are
street-corner hoardings effective at all? NOTES
Anybody would see immediately that we need data to answer any of these
questions. But mere data would not help. The data will have to be systematically
collected and analysed so that our answers are not affected by other factors. For
example, how does one separate the effect of age and experience when one knows
that experience increases with age? Also in each of these there are a whole lot of
chance factors interacting with the outcome. Therefore, we have to isolate the
effect that we want to study, and for this specialized statistical methods are called
for.
The statistical method, when used properly, helps in understanding phenomena
using numerical evidence. As a further example, suppose we want to understand
the factors that affect the yield of farms. We may note that various factors such as
rainfall, soil fertility, quality of seed, soil nutrients used, method of cultivation, etc.
are all more or less important. One can never for sure predict the influence of one
parameter, because we cannot control all of these independently. But it is possible
to design experiments and collect data so that one is able to, more or less, isolate
each effect to a predetermined level of certainty. The procedures for doing so are
provided by statistical method.
As a further example, let us suppose we are interested in studying the level
of income of the people living in a certain village.
For this purpose the following procedure may be adopted with advantages:
(i) Collect data: Information should be collected regarding
 the number of persons living in the village
 the number of persons who are getting income
 the daily income of each earning member
(ii) Organize the data obtained above so as to show the number of persons
within different income groups, and in that way reduce bulk.
(iii) Present this information by means of diagrams or other visual aids.
(iv) Analyse the data to determine the ‘Average’ income of the people
and the extent of disparities that exist.
(v) On the basis of the above it would be possible to have an understanding
of the phenomenon (income of people), and one would know (a) The
average income of people’, and (b) The extent of disparity in the level
of incomes.
(vi) All this may lead to a policy decision for improvement of the existing
situation.
4.2.1 Scope of Statistics
The proper function of statistics is to enlarge our knowledge of complex phenomena,
and to lend precision to our ideas that would otherwise remain vague and
indeterminate. Our knowledge of such things as ‘National Income,’ ‘Population,’ Self - Learning
Material 189
Statistical Computation and ‘National Resources’, etc., would not have been so definite and precise, if there
Probability Distribution
were no reliable statistics pertaining to each one of these. To say that the per
capita income in India is low is a vague statement. The term ‘Low’ may mean one
thing to one individual while to another it might mean something altogether different.
NOTES One may take it to be near about 100 while someone else may think it to be in
the neighborhood of 5,000. But the moment we say that our per capita income
is 750 we make a statement which is precise and convincing. Again a statement,
viz., the per capita income in agricultural sector is lower than in the industrial sector,
is vague and indefinite. But if the per capita incomes for both these sectors are
ascertained; the comparison would be easier and even a layman would be able to
appreciate the difference in the productivity of these two sectors. It can thus be
said that ‘Statistics increases the field of mental vision, as an opera glass or telescope
increases the field of physical vision. Statistics is able to widen our knowledge
because of the following services that it renders.
It presents facts in a definite form. It is the quality of definiteness which is
responsible for the growing universal application of statistical methods. The
conclusions stated numerically are definite and hence more convincing than
conclusions stated qualitatively. This fact can be readily understood by a simple
example. In an advertisement, statements expressed numerically have greater
attraction and are more appealing than those expressed in a qualitative manner.
The caption, ‘We have sold more cars this year’ is certainly less attractive than
‘Record sale of 10,000 cars in 1985 as compared to 6,000 in 1984’. The latter
statement emphasizes in a much better manner the growing popularity of the
advertiser’s cars.
Statistics simplifies unwieldy and complex mass of data and presents them
in such a manner that they at once become intelligible: The complex data
may be reduced to totals; averages, percentages, etc., and presented either
graphically or diagrammatically. These devices help us to understand quickly the
significant characteristics of the numerical data, and consequently save us from a
lot of mental strain. Single figures in the form of averages and percentages can be
grasped more easily than a mass of statistical data comprising thousands of facts.
Similarly, diagrams and graphs, because of their greater appeal to the eye and
imagination tender valuable assistance in the proper understanding of numerical
data. Time and energy of business executives are thus economized, if the statistician
supplies them with the results of production, sales and finances in a condensed
form.
Statistics classifies numerical facts: The procedure of classification brings into
relief the salient features of the variable that is under investigation. This can be
clearly illustrated by an example. If we are given the marks in mathematics of each
individual student of a class and if it is desired to judge the performance of the
class on the basis of these data it will not be an easy matter. Human mind has its
limitations and cannot easily grasp a multitude of figures. But if the students are
classified i.e., if we put into one group all those boys who get more than second
division marks, in still another group those who get third division marks, and have
a separate group of those who fait to get pass marks, it will be easier for us to
Self - Learning form a more precise idea about the performance of the class.
190 Material
Statistics furnishes a technique of comparison: The facts, having been Statistical Computation and
Probability Distributiona
classified, are now in a shape when they can be used for purposes of comparisons
and contrasts. Certain facts, by themselves, may be meaningless unless they are
capable of being compared with similar facts at other places or at other periods in
time. We estimate the national income of India not essentially for the value of that NOTES
fact itself, but’ mainly in order that we may compare the income of today with that
of the past and thus draw conclusions as to whether the standard of living’ of the
people is on the increase, decrease or is stationary. Statistics affords suitable
technique for comparison. It is with the help of statistics that the cost accountant is
able to compare the actual accomplishment (in terms of cost) with programmes
laid out (in terms of standard cost). Some of the modes of comparison provided
by statistics include totals, ratios, averages or measure of central tendencies, graphs
and diagrams, and coefficients. Statistics thus ‘Serves as a scale in which facts in
various combinations are weighed and valued,
Statistics endeavours to interpret conditions: Like an artist statistics renders
useful service in presenting an attractive picture of the phenomenon under
investigation- But it frequently does far more than this by enabling the interpretation
of condition, by developing possible causes for the results described. If the
production manager discovers that a certain machine is turning out some articles
which are not up to the standard specifications, he will be able to find statistically
if this condition is due to some defect in the machine or whether such a condition
is normal.
Statistical Method
Statistical approach to a problem may broadly be summarized as: (i) Collection of
facts: (ii) Organization of facts; (iii) Analysis of facts; and (iv) Interpretation of
facts.
A detailed discussion of the various methods of collection, presentation,
analysis and interpretation of facts is given later in the unit. Here the intention is to
give only a bird’s eye-view of the entire statistical procedure,
(i) Collection of facts is the first step in the statistical treatment of a problem.
Numerical facts are the raw materials upon which the statistician is to work
and just as in a manufacturing concern the quality of a finished product
depends, inter alia, upon the quality of the raw material, in the same manner,
the validity of statistical conclusions will be governed, among other
considerations, by the quality of data used. Assembling of the facts is thus a
very important process and no pains should be spared to see that the data
collected are accurate, reliable and thorough. One thing that should be noted
here is that the work of collecting facts should be undertaken in a planned
manner. Without proper planning the facts collected may not be suitable for
the purpose and a lot of time and money may be wasted.
(ii) The data so collected will more often than not be a huge mass of facts
running into hundreds and thousands of figures. Human mind has its
limitations. No One can appreciate at a glance or even after a careful study
hold in mind the information contained in a hundred or a thousand schedules.
For a proper understanding of the data their irregularities must be brushed Self - Learning
Material 191
Statistical Computation and off and their bulk be reduced, i.e., some process of condensation must
Probability Distribution
take place. Condensation implies the organization, classification, tabulation
and presentation of the data in a suitable form.
(iii) The process of statistical analysis is a method of abstracting significant facts
NOTES
from the collected mass of numerical data. This process includes such things
as ‘measures of central tendency’—the determination, of Mean, Median
and Mode—’measures of dispersion’ and the determination of trends and
tendencies, etc. This is more or less a mechanical process involving the use
of elementary mathematics.
(iv) The interpretation of the various statistical constants obtained through a
process of statistical analysis is the final phase or the finishing process of the
statistical technique. It involves those methods by which judgments are
formed and inferences obtained. To make estimates of the population
parameters on the basis of sample statistics in an example of the problem of
interpretation. For the interpretation of results a knowledge of advanced
mathematics is essential.
Characteristics of Statistical Data
Even a casual look at Table 4.2 would lead us to the conclusion that statistical data
always denotes ‘Figures’, i.e., numerical descriptions. Whereas this is true, it must
be remembered that all numerical descriptions are not statistical data. In order
that numerical descriptions may be called statistics they must possess the following
characteristics:
(i) They must be in aggregates.
(ii) They must be affected to a marked extent by a multiplicity of causes.
(iii) They must be enumerated or estimated according to reason able standard
of accuracy,
(iv) They must have been collected in a systematic manner for a predetermined
purpose.
(v) They must be placed in relation to each other.
Let us explain these characteristics:
Statistics are aggregates of facts: This means that statistics are a ‘number
of facts.’A single fact,’ even though numerically stated, cannot be called statistics.
‘A single death, an accident, a sale, a shipment does not constitute statistics. Yet
numbers of deaths, accidents; sales and shipments are statistics.’ Observe carefully
Table 4.2 containing information about the population of India. Column (a) states
the population only for one year whereas Column (b) gives population figures for
seven different years. The data given in Column (b) are statistics whereas the
figure given in Column (a) is not so, for the simple reason that it is a single solitary
figure.

Self - Learning
192 Material
 Table 4.2 Characteristics of Statistical Data Statistical Computation and
Probability Distributiona
Column (a) Column (b)
Population Population
Year (in lakhs) Year (in lakhs)
NOTES
1951 3,569 1911 2,490
1921 2,481
1931 2,755
1941 3,128*
1951 3,569
1961 4,390
1971 5,470
*After deducting estimated amount of inflation of returns in West Bengal and Punjab
(20 lakhs).
They must be affected to a marked extent by a multiplicity of causes:
The term statistical data can be used only when we cannot predict exactly the
values of the various physical quantities. This means that the numerical value of
any quantity at any particular moment is the result of the action and interaction of
a number of forces, differing amongst themselves and it is not possible to say as to
how much of it is due to any one particular cause. Thus, the volume of wheat
production is attributable to a number of factors, viz., rainfall, soil, fertility, quality
of seed, methods of cultivation, etc. All these factors acting jointly determine the
amount of the yield and it is not possible for any one to assess the individual
contribution of any one of these factors.
Statistics must be enumerated or estimated according to reasonable
standards of accuracy: This means that if aggregates of numerical facts are to be
called’ statistics’ they must be reasonably accurate. This is necessary because
statistical data are to serve as a basis for statistical investigations. If the basis
happens to be incorrect the results are bound to be misleading. It must, however,
be clearly stated that it is not ‘mathematical accuracy, but only reasonable accuracy’
that is necessary in statistical work. What standard of accuracy is to be regarded
as reasonable will depend upon the aims and objects of inquiry. Where precision
is required’ accuracy is necessary; where generel impressions are sufficient,
appreciable errors may be tolerated. Again, whatever standard of accuracy is
once adopted, it should be uniformly maintained throughout the inquiry.
Statistics are collected in a systematic manner for a predetermined
purpose: Numerical data can be called statistics only if they have been compiled
in a properly planned manner and for a purpose about which the enumerator had
a definite idea. So long as the compiler is not clear about the object for which facts
are to be collected, he will not be able to distinguish between facts that are relevant
and those that are unnecessary; and as such the data collected will, in all probability,
be a heterogeneous mass of unconnected facts. Again, the procedure of data
collection must be properly planned, i.e., it must be decided beforehand as to
what kind of information is to be collected and the method that is to be applied in
obtaining it. This involves decisions on matters like ‘statistical unit,’ ‘standard of
accuracy,’ ‘list of questions,’ etc. Facts collected in an unsystematic manner, and
without a complete awareness of the object, will be confusing and cannot be
made the basis of valid conclusions. Self - Learning
Material 193
Statistical Computation and Statistics should be placed in relation to each other. Numerical facts
Probability Distribution
may be placed in relation to each other either in point of time, space or condition.
The phrase ‘placed in relation to each other’ suggests that the facts should be
comparable. Facts are comparable in point of time when we have measurements
NOTES of the same object, obtained in an identical manner, for different periods. They are
said to be related in point of space or condition when we have the measurements
of the same phenomenon at different places or in different conditions, but at the
same time. Numerical facts will be comparable, if they pertain to the same inquiry
and have been compiled in a systematic manner for a predetermined purpose.
Putting all these characteristics together, Secrets has defined statistics
(numerical descriptions) as: ‘Aggregates of facts, affected to a marked extent by
multiplicity of causes, numerically expressed, enumerated or estimated, according
to reasonable standard of accuracy, collected in a systematic manner, for a
predetermined purpose, and placed in relation to each other.’
Some Other Definitions of Statistics
As numerical data
Waster has defined statistics as ‘Classified facts respecting the condition of the
people in a state especially those facts which can be stated in numbers or in tables
or in any other tabular or classified arrangement.’ No doubt, this definition was
correct at a time when statistics were collected only for purposes of internal
administration or for knowing, for purposes of war, the wealth of the State. The
scope of statistics is now considerably wider and it has almost a universal application.
Obviously, therefore, the definition is inadequate.
Bowley defines statistics as ‘numerical statements of facts in any department
of inquiry placed in relation to each other.’ This is somewhat more accurate. It
means that if numerical facts do not pertain to a department of inquiry or if such
facts are not related to each other they cannot be called statistics. The leads us to
the conclusion that ‘all statistics are numerical facts but all numerical facts are not
statistics.’ This definition is certainly better than the previous one But it is not
comprehensive enough in as much as it does not give any importance either to the
nature of facts or the standard of accuracy.
As Statistical Methods
Bowley has called it ‘The science of measurement of the social organism, regarded
as a whole, in all its manifestations.’ This definition is too narrow as it confines the
scope of statistics only to human activities. Statistics in fact has a much wider
application and is not confined only to the social organism. Besides, statistics is
not only the technique of measuring but also of analysing and interpreting. Again,
statistics, strictly speaking, is not a science but a scientific method. It is a device of
inferring knowledge and not knowledge itself.
Bowley has also called statistics ‘the science of counting,’ and ‘the science
of average.’ These definitions are again incomplete in the sense that they pertain to
only a limited field. True, statistical work includes counting and averaging, but it
also includes many other processes of treating quantitative data. In fact, while
Self - Learning dealing with large numbers, actual count becomes illusory and only estimates are
194 Material
made. Thus these definitions can also be discarded on the ground of inadequacy.
Origin of Statistics Statistical Computation and
Probability Distributiona
Statistics originated from two quite dissimilar fields, viz., games of chance and
political states. These two different fields are also termed as two distinct
disciplines—one primarily analytical and the other essentially descriptive. The NOTES
former is associated with the concept of chance and probability and the latter is
concerned with the collection of data.
The theoretical development of the subject has its origin in the mid-17 century
and many mathematicians and gamblers of France, Germany and England are
credited for its development. Notable amongst them are Pascal (1623–1662),
who investigated the properties of the coefficients of binomial expansion and James
Bernoulli (1654–1705). who wrote the first treatise on the theory of probability.
As regards the descriptive side of statistics it may be stated that statistics is
as old as statecraft. Since time immemorial men must have been compiling
information about wealth and manpower for purpose of peace and war. This activity
considerably expanded at each upsurge of social and political development and
received added impetus in periods of war.
The development of statistics can be divided into the following three stages:
The empirical stage (1600): During this, the primitive stage of the subject,
numerical facts were utilized by the rulers, principally as an aid in the administration
of Government. Information was gathered about the number of people and the
amount of property held by them—the former serving the ruler as an index of
human fighting strength and the latter as an indication of actual and potential taxes.
The comparative stage (1600–1800): During this period statisticians
frequently made comparisons between nations with a view to judging their relative
strength and prosperity. In some countries enquiries were instituted to judge the
economic and social conditions of their people. Colbert introduced in France a
‘mercantile’ theory of government whose basis was essentially statistical in
character. In 1719, Frederick William I began gathering information about
population occupation, house-taxes, city finance, etc., which helped to study the
condition of the people.
The modern stage (1800 up to date): During this period statistics is viewed
as a way of handling numerical facts rather than a mere device of collecting numerical
data. Besides, there has been a considerable extension of the field of its applicability.
It has now become a useful tool and statistical methods of analysis are now being
increasingly used in biology, psychology, education, economics and business.

4.3 VARIOUS MEASURES OF STATISTICAL

COMPUTATIONS
For quantitative data, the mean, median, mode, percentiles, range, variance, and
standard deviation are the most widely used numerical measurements. The mean,
also known as the average, is calculated by summing all of a variable’s data values
and dividing the total by the number of data values.
Self - Learning
Material 195
Statistical Computation and
Probability Distribution
4.3.1 Average
In statistics, the term central tendency specifies the method through which the
quantitative data have a tendency to cluster approximately about some value. A
NOTES measure of central tendency is any precise method of specifying this ‘Central Value’.
In the simplest form, the measure of central tendency is an average of a set of
measurements, where the word average refers to as mean, median, mode or
other measures of location. Typically the most commonly used measures are
arithmetic mean, mode and median. These values are very useful not only in
presenting the overall picture of the entire data but also for the purpose of making
comparisons among two or more sets of data. As an example, questions like
‘How hot is the month of June in Delhi?’ can be answered, generally by a single
figure of the average for that month. Similarly, suppose we want to find out if boys
and girls at age 10 years differ in height for the purpose of making comparisons.
Then, by taking the average height of boys of that age and average height of girls
of the same age, we can compare and record the differences.
While arithmetic mean is the most commonly used measure of central location,
mode and median are more suitable measures under certain set of conditions and
for certain types of data. However, each measure of central tendency should meet
the following requisites.
1. It should be easy to calculate and understand.
2. It should be rigidly defined. It should have only one interpretation so that
the personal prejudice or bias of the investigator does not affect its usefulness.
3. It should be representative of the data. If it is calculated from a sample, then
the sample should be random enough to be accurately representing the
population.
4. It should have sampling stability. It should not be affected by sampling
fluctuations. This means that if we pick 10 different groups of college students
at random and compute the average of each group, then we should expect
to get approximately the same value from each of these groups.
5. It should not be affected much by extreme values. If few very small or very
large items are present in the data, they will unduly influence the value of the
average by shifting it to one side or other, so that the average would not be
really typical of the entire series. Hence, the average chosen should be such
that it is not unduly affected by such extreme values.
All these measures of central tendency are discussed in this section.
4.3.2 Mean
There are several commonly used measures, such as arithmetic mean, mode and
median. These values are very useful not only in presenting the overall picture of
the entire data, but also for the purpose of making comparisons among two or
more sets of data.
As an example, questions like ‘How hot is the month of June in Delhi?’ can
be answered generally by a single figure of the average for that month. Similarly,
Self - Learning suppose we want to find out if boys and girls of age 10 years differ in height for the
196 Material
purpose of making comparisons. Then, by taking the average height of boys of Statistical Computation and
Probability Distributiona
that age and the average height of girls of the same age, we can compare and
record the differences.
While arithmetic mean is the most commonly used measure of central
NOTES
tendency, mode and median are more suitable measures under certain set of
conditions and for certain types of data. However, each measure of central tendency
should meet the following requisites:
(i) It should be easy to calculate and understand.
(ii) It should be rigidly defined. It should have only one interpretation so
that the personal prejudice or the bias of the investigator does not
affect its usefulness.
(iii) It should be representative of the data. If it is calculated from a sample,
the sample should be random enough to be accurately representing
the population.
(iv) It should have a sampling stability. It should not be affected by sampling
fluctuations. This means that if we pick ten different groups of college
students at random and compute the average of each group, then we
should expect to get approximately the same value from each of these
groups.
(v) It should not be affected much by extreme values. If few, very small
or very large items are present in the data, they will unduly influence
the value of the average by shifting it to one side or other, so that the
average would not be really typical of the entire series. Hence, the
average chosen should be such that it is not unduly affected by such
extreme values.
Arithmetic mean is also commonly known as the mean. Even though average,
in general, means measure of central tendency, when we use the word average in
our daily routine, we always mean the arithmetic average. The term is widely used
by almost everyone in daily communication. We speak of an individual being an
average student or of average intelligence. We always talk about average family
size or average family income or Grade Point Average (GPA) for students, and
so on.
For discussion purposes, let us assume a variable X which stands for some
value such as the ages of students. Let the ages of 5 students be 19, 20, 22, 22
and 17 years. Then variable X would represent these ages as,
X: 19, 20, 22, 22, 17
Placing the Greek symbol (Sigma) before X would indicate a command
that all values of X are to be added together. Thus,
X = 19 + 20 + 22 + 22 + 17
The mean is computed by adding all the data values and dividing it by the
number of such values. The symbol used for sample average is X , so that,
19  20  22  22  17
X
5
Self - Learning
Material 197
Statistical Computation and In general, if there are n values in the sample, then,
Probability Distribution
X1  X 2  .........  X n
X
n
NOTES In other words,
n
 Xi
i 1
X , i  1, 2 ... n
n
According to this formula, mean can be obtained by adding all values of Xi,
where the value of i starts at 1 and ends at n with unit increments so that i = 1, 2,
3, ... n.
If instead of taking a sample, we take the entire population in our calculations
of the mean, then the symbol for the mean of the population is (mu) and the size
of the population is N, so that,
N
 Xi
i 1
 , i  1, 2 ...N
N
If we have the data in grouped discrete form with frequencies, then the
sample mean is given by,
f ( X )
X
f
Here, f = Summation of all frequencies
= n
f(X) = Summation of each value of X multiplied by its
corresponding frequency ( f )
Example 4.1: Let us take the ages of 10 students as follows:
19, 20, 22, 22, 17, 22, 20, 23, 17, 18
Solution: This data can be arranged in a frequency distribution as follows:
(X) (f) f(X)
17 2 34
18 1 18
19 1 19
20 2 40
22 3 66
23 1 23
Total = 10 200
In this case, we have f = 10 and f(X) = 200, so that,
f ( X )
X = f
= 200/10 = 20

Self - Learning
198 Material
Characteristics of the Mean Statistical Computation and
Probability Distributiona
The arithmetic mean has some interesting properties. These are as follows:
(i) The sum of the deviations of individual values of X from the mean will always
add up to zero. This means that if we subtract all the individual values from NOTES
their mean, then some values will be negative and some will be positive, but
if all these differences are added together then the sum will be zero. In other
words, the positive deviations must balance the negative deviations. Or
symbolically,
n
 (Xi  X ) = 0, i = 1, 2, ... n
i 1

(ii) The second important characteristic of mean is that it is very sensitive to

extreme values. Since the computation of mean is based upon inclusion of
all values in the data, an extreme value in the data would shift the mean
towards it, thus making the mean unrepresentative of the data.
(iii) The third property of mean is that the sum of squares of the deviations
about the mean is minimum. This means that if we take the differences
between individual values and the mean and square these differences
individually and then add these squared differences, then the final figure will
be less than the sum of the squared deviations around any other number
other than the mean. Symbolically, it means that
n
 ( X i  X )2 = Minimum, i = 1, 2, ... n
i 1

(iv) The product of the arithmetic mean and the number of values on which the
mean is based is equal to the sum of all given values. In other words, if we
replace each item in series by the mean, then the sum of these substitutions
will equal the sum of individual items. Thus, if we take random figures as an
example like 3, 5, 7, 9, and if we substitute the mean for each item 6, 6, 6,
6 then the total is 24, both in the original series and in the substitution series.
This can be shown like,
X
Since, X =
N
 N X = X
For example, if we have a series of values 3, 5, 7, 9, the mean is 6. The
squared deviations will be:
X X  X  X X 2
3 3  6  3 9
5 5  6  1 1
7 7  6 1 1
9 963 9
2
 X   20

This property provides a test to check if the computed value is the correct
arithmetic mean.
Self - Learning
Material 199
Statistical Computation and Example 4.2: The mean age of a group of 100 persons (grouped in intervals
Probability Distribution
10–, 12–,..., etc.) was found to be 32.02. Later, it was discovered that age 57
was misread as 27. Find the corrected mean.

NOTES Solution: Let the mean be denoted by X. So, putting the given values in the
formula of arithmetic mean, we have,
 X , i.e.,
32.02 =
100
X = 3202

Correct X = 3202 – 27 + 57 = 3232

3232
 Correct AM = = 32.32
100
Example 4.3: The mean monthly salary paid to all employees in a company is
500. The monthly salaries paid to male and female employees average 520 and
420, respectively. Determine the percentage of males and females employed by
the company.
Solution: Let N1 be the number of males and N2 be the number of females
employed by the company. Also, let x1 and x2 be the monthly average salaries
paid to male and female employees and x be the mean monthly salary paid to all
the employees.
N1 x1  N 2 x2
x =
N1  N 2
520 N 1  420 N 2
or 500 = or 20N1= 80N2
N1  N 2
N1 80 4
or = 
N2 20 1
Hence, the males and females are in the ratio of 4 : 1 or 80 per cent are
males and 20 per cent are females in those employed by the company.
Short-Cut Methods for Calculating Mean
We can simplify the calculations of mean by noticing that if we subtract a constant
amount A from each item X to define a new variable X' = X – A, the mean X  of
X' differs from X by A. This generally simplifies the calculations and we can then
add back the constant A, termed as the assumed mean such as,
f ( X )
X = A X A 
f

Table 4.3 illustrates the procedure of calculation by short-cut method using

the data given in Example 4.1.
The choice of A is made in such a manner as to simplify calculation the most,
and is generally in the region of concentration of data.

Self - Learning
200 Material
 Table 4.3 Short-Cut Method of Calculating Mean Statistical Computation and
Probability Distributiona
X (f) Deviation from f(X')
Assumed Mean (13) X'

9 1 –4 –4 NOTES
10 2 –3 –6
11 3 –2 –6
12 6 –1 –6

13 10 0 –22
14 11 +1 +11
15 7 +2 +14
16 3 +3 +9
17 2 +4 +8
18 1 +5 +5

+47
–22

f = 46 fX  = 25

The mean,
 f ( X ) 25
X = A  13  = 4.54
f 46
This mean is the same as calculated in Example 4.1.
In the case of grouped frequency data, the variable X is replaced by midvalue
m, and in the short-cut technique, we subtract a constant value A from each m, so
that the formula becomes:
 f ( m  A)
X = A
f
In cases where the class intervals are equal, we may further simplify calculation
by taking the factor i from the variable m–A defining,
m A
X' =
i

where i is the class width. It can be verified that when X' is defined, then the mean
of the distribution is given by
f ( X )
X = A  i
f

Example 4.2 will illustrate the use of short-cut method.

Example 4.4: The ages of twenty husbands and wives are given in the following
table. Form frequency tables showing the relationship between the ages of husbands
and wives with class intervals 20 – 24; 25 – 29; etc.

Self - Learning
Material 201
Statistical Computation and Calculate the arithmetic mean of the two groups after the classification.
Probability Distribution
S.No. Age of Husband Age of Wife
1 28 23
2 37 30
NOTES 3 42 40
4 25 26
5 29 25
6 47 41
7 37 35
8 35 25
9 23 21
10 41 38
11 27 24
12 39 34
13 23 20
14 33 31
15 36 29
16 32 35
17 22 23
18 29 27
19 38 34
20 48 47

Solution:
Calculation of Arithmetic Mean of Husbands’ Age

m  37
Class Intervals Midvalues Husband x1' = f1x1'
5
m Frequency ( f1)
20–24 22 3 –3 –9
25–29 27 5 –2 –10
30–34 32 2 –1 –2
 21
35–39 37 6 0 0
40–44 42 2 1 2
45–49 47 2 2 4
6
f1 = 20 f1x1' = –15

Arithmetic mean of husbands age,

 f 1 x1 15
x = ×i  A  × 5  37 = 33.25
N 20
Calculation of Arithmetic Mean of Wives’ Age
Wife
m  37
Class Intervals Midvalues Frequency x2' = f2x2'
5
m ( f2)

20–24 22 5 –3 –15
25–29 27 5 –2 –10
30–34 32 4 –1 –4
35–39 37 3 0 0
40–44 42 2 1 2
45–49 47 1 2 2
f2 = 20 f2x2' = –25
Self - Learning
202 Material
 Arithmetic mean of wife’s age, Statistical Computation and
Probability Distributiona
f x 25
x =  2 2 i A  5  37 = 30.75
N 20
The Weighted Arithmetic Mean NOTES
In the computation of arithmetic mean we had given equal importance to each
observation in the series. This equal importance may be misleading if the individual
values constituting the series have different importance as in Example 4.5.
Example 4.5: The Raja Toy shop sells
Toy Cars at 3 each
Toy Locomotives at 5 each
Toy Aeroplanes at 7 each
Toy Double Decker at 9 each
What will be the average price of the toys sold, if the shop sells 4 toys, one of each
kind?
Solution:
x 24
Mean Price, i.e., x = = = 6
4 4
In this case, the importance of each observation (price quotation) is equal in
as much as one toy of each variety has been sold. In the computation of the
arithmetic mean, this fact has been taken care of by including ‘once only’ the price
of each toy.
If , however the shop sells 100 toys: 50 cars, 25 locomotives, 15 aeroplanes
and 10 double deckers, the importance of the four price quotations to the dealer
is not equal as a source of earning revenue. In fact, their respective importance is
equal to the number of units of each toy sold, i.e.,
The importance of Toy Car 50
The importance of Locomotive 25
The importance of Aeroplane 15
The importance of Double Decker 10
It may be noted that 50, 25, 15, 10 are the quantities of the various classes
of toys sold. It is for these quantities that the term ‘weights’ is used in statistical
language. Weight is represented by symbol ‘w’, and w represents the sum of
weights.
While determining the ‘Average price of toy sold’, these weights are of
great importance and are taken into account in the manner as shown,
w1 x1  w2 x2  w3 x3  w4 x4
x = =  wx
w1  w2  w3  w4 w

When w1, w2, w3, w4 are the respective weights of x1, x2, x3, x4 which in turn
represent the price of four varieties of toys, viz., car, locomotive, aeroplane and
double decker, respectively.
(50  3)  (25  5)  (15  7)  (10  9) Self - Learning
x = Material 203
50  25  15  10
Statistical Computation and (150)  (125)  (105)  (90) 470
Probability Distribution = = = 4.70
100 100
The following table summarizes the steps taken in the computation of the
NOTES weighted arithmetic mean.
Weighted Arithmetic Mean of Toys Sold by Raja Toy Shop

Toys Price per Toy Number Sold Price × Weight

x w xw
Car 3 50 150
Locomotive 5 25 125
Aeroplane 7 15 105
Double Decker 9 10 90

w = 100 xw = 470

w = 100; wx = 470

x =  wx = 470 = 4.70
w 100
The weighted arithmetic mean is particularly useful where we have to compute
the mean of means. If we are given two arithmetic means, one for each of two
different series, in respect of the same variable, and are required to find the arithmetic
mean of the combined series, the weighted arithmetic mean is the only suitable
method of its determination (Refer Example 4.6).
Example 4.6: The arithmetic mean of daily wages of two manufacturing concerns
A Ltd. and B Ltd. is 5 and 7, respectively. Determine the average daily wages
of both concerns if the number of workers employed were 2,000 and 4,000,
respectively.
Solution:
(i) Multiply each average (viz., 5 and 7), by the number of workers in the
concern it represents.
(ii) Add up the two products obtained in (i).
(iii) Divide the total obtained in (ii) by the total number of workers.
Weighted Mean of Mean Wages of A Ltd. and B Ltd.

Manufacturing Mean Wages Workers Mean Wages ×

Concern x Employed Workers Employed
w wx

A Ltd. 5 2,000 10,000

B Ltd. 7 4,000 28,000

 w = 6,000  wx = 38,000

 wx
x =
w
38,000
=
6, 000
= 6.33

Self - Learning
204 Material
 These examples explain that ‘Arithmetic Means and Percentage’ are not Statistical Computation and
Probability Distributiona
original data. They are derived figures and their importance is relative to the original
data from which they are obtained. This relative importance must be taken into
account by weighting while averaging them (means and percentage).
NOTES
Advantages of Mean
(i) Its concept is familiar to most people and is intuitively clear.
(ii) Every data set has a mean, which is unique and describes the entire data to
some degree. For example, when we say that the average salary of a
professor is 25,000 per month, it gives us a reasonable idea about the
salaries of professors.
(iii) It is a measure that can be easily calculated.
(iv) It includes all values of the data set in its calculation.
(v) Its value varies very little from sample to sample taken from the same
population.
(vi) It is useful for performing statistical procedures, such as computing and
comparing the means of several data sets.
Disadvantages of Mean
(i) It is affected by extreme values, and hence are not very reliable when the
data set has extreme values especially when these extreme values are on
one side of the ordered data. Thus, a mean of such data is not truly a
representative of such data. For example, the average age of three persons
of ages 4, 6 and 80 years gives us an average of 30.
(ii) It is tedious to compute for a large data set as every point in the data set is
to be used in computations.
(iii) We are unable to compute the mean for a data set that has open-ended
classes either at the high or at the low-end of the scale.
(iv) The mean cannot be calculated for qualitative characteristics, such as beauty
or intelligence, unless these can be converted into quantitative figures such
as intelligence into IQs.
4.3.3 Median
The second measure of central tendency that has a wide usage in statistical works
is the median. Median is that value of a variable which divides the series in such a
manner that the number of items below it is equal to the number of items above it.
Half the total number of observations lie below the median, and half above it. The
median is thus a positional average.
The median of ungrouped data is found easily if the items are first arranged
in order of the magnitude. The median may then be located simply by counting,
and its value can be obtained by reading the value of the middle observations. If
we have five observations whose values are 8, 10, 1, 3 and 5, the values are first
arrayed: 1, 3, 5, 8 and 10. It is now apparent that the value of the median is 5,
since two observations are below that value and two observations are above it.
Self - Learning
Material 205
Statistical Computation and When there is an even number of cases, there is no actual middle item and the
Probability Distribution
median is taken to be the average of the values of the items lying on either side of
(N + 1)/2, where N is the total number of items. Thus, if the values of six items of
a series are 1, 2, 3, 5, 8 and 10, then the median is the value of item number (6 +
NOTES 1)/2 = 3.5, which is approximated as the average of the third and the fourth items,
i.e., (3+5)/2 = 4.
Thus, the steps required for obtaining median are as follows:
(i) Arrange the data as an array of increasing magnitude.
(ii) Obtain the value of the (N+ l)/2th item.
Even in the case of grouped data, the procedure for obtaining median is
straightforward as long as the variable is discrete or non-continuous as is clear
from Example 4.7.
Example 4.7: Obtain the median size of shoes sold from the following data:
Number of Shoes Sold by Size in One Year

Size Number of Pairs Cumulative Total

5 30 30
5 21 40 70
6 50 120
6 21 150 270
7 300 570
7 21 600 1170
8 950 2120
8 21 820 2940
9 750 3690
9 21 440 4130
10 250 4380
10 21 150 4530
11 40 4570
11 21 39 4609
Total 4609

( N  1) 4609 + 1
Solution: Median is the value of th = th = 2305th item. Since the
2 2
items are already arranged in ascending order (size-wise), the size of 2305th item
is easily determined by constructing the cumulative frequency. Thus, the median
size of shoes sold is 8½, the size of 2305th item.
In the case of grouped data with continuous variable, the determination of
median is a bit more involved. Consider the following table where the data relating
to the distribution of male workers by average monthly earnings is given. Clearly
the median of 6291 is the earnings of (6291 + l)/2 = 3l46th worker arranged in
ascending order of earnings.
From the cumulative frequency, it is clear that this worker has his income in
Self - Learning the class interval 67.5 – 72.5. However, it is impossible to determine his exact
206 Material
income. We therefore, resort to approximation by assuming that the 795 workers Statistical Computation and
Probability Distributiona
of this class are distributed uniformly across the interval 67.5 – 72.5. The median
worker is (3146–2713) = 433rd of these 795, and hence, the value corresponding
to him can be approximated as,
NOTES
433
67.5  × ( 72.5  67.5) = 67.5 + 2.73 = 70.23
795
Distribution of Male Workers by Average Monthly Earnings

Group No. Monthly No. of Cumulative No.

Earnings ( ) Workers of Workers

1 27.5–32.5 120 120

2 32.5–37.5 152 272
3 37.5–42.5 170 442
4 42.5–47.5 214 656
5 47.5–52.5 410 1066
6 52.5–57.5 429 1495
7 57.5–62.5 568 2063
8 62.5–67.5 650 2713
9 67.5–72.5 795 3508
10 72.5–77.5 915 4423
11 77.5–82.5 745 5168
12 82.5–87.5 530 5698
13 87.5–92.5 259 5957
14 92.5–97.5 152 6109
15 97.5–102.5 107 6216
16 102.5–107.5 50 6266
17 107.5–112.5 25 6291

Total 6291

The value of the median can thus be put in the form of the formula,
N 1
C
Me = l  2 ×i
f

Where l is the lower limit of the median class, i its width, f its frequency, C the
cumulative frequency upto (but not including) the median class, and N is the total
number of cases.
Finding Median by Graphical Analysis
The median can quite conveniently be determined by reference to the ogive which
plots the cumulative frequency against the variable. The value of the item below
which half the items lie, can easily be read from the ogive as is shown in
Example 4.8.
Self - Learning
Material 207
Statistical Computation and Example 4.8: Obtain the median of data given in the following table:
Probability Distribution

Monthly Earnings Frequency Less Than More Than

27.5 __ 0 6291
NOTES 32.5 120 120 6171
37.5 152 272 6019
42.5 170 442 5849
47.5 214 656 5635
52.5 410 1066 5225
57.5 429 1495 4796
62.5 568 2063 4228
67.5 650 2713 3578
72.5 795 3508 2783
77.5 915 4423 1868
82.5 745 5168 1123
87.5 530 5698 593
92.5 259 5957 334
97.5 152 6109 182
102.5 107 6216 75
107.5 50 6266 25
112.5 25 6291 0

Solution: It is clear that this is grouped data. The first class is 27.5 – 32.5, whose
frequency is 120, and the last class is 107.5 – 112.5 whose frequency is 25.
Figure 4.1 shows the ogive of less than cumulative frequency. The median is the
value below which N/2 items lie, is 6291/2 = 3145.5 items lie, which is read of
from Figure 4.2 as about 70. More accuracy than this is unobtainable because of
the space limitation on the earning scale.

6291
6000

5000
MORE THAN LESS THAN

4000
Number of Workers

3000

2000

1000

MEDIAN
0
27.5
32.5
37.5
42.5
47.5
52.5
57.5

67.5
72.5
77.5
82.5
87.5
92.5
97.5
102.5
107.5
112.5
62.5

Monthly Earnings in Rupees

Fig. 4.1 Median Determination by Plotting Less than and More than
Cumulative Frequency

Self - Learning
208 Material
 The median can also be determined by plotting both ‘Less Than’ and ‘More Statistical Computation and
Probability Distributiona
than’ cumulative frequency as shown in Figure 4.1. It should be obvious that the
two curves should intersect at the median of the data.

NOTES
6000

5000

4000
Number of Workers

3000

2000
MEDIAN
1000

0
27.5
32.5
37.5
42.5
47.5
52.5
57.5
62.5
67.5
72.5
77.5
82.5
87.5
92.5
97.5
102.5
107.5
112.5

Monthly Earnings in Rupees

Fig. 4.2 Median

Advantages of Median
(i) Median is a positional average and hence the extreme values in the data set
do not affect it as much as they do to the mean.
(ii) Median is easy to understand and can be calculated from any kind of data,
even from grouped data with open-ended classes.
(iii) We can find the median even when our data set is qualitative and can be
arranged in the ascending or the descending order, such as average beauty
or average intelligence.
(iv) Similar to mean, median is also unique, meaning that, there is only one median
in a given set of data.
(v) Median can be located visually when the data is in the form of ordered
data.
(vi) The sum of absolute differences of all values in the data set from the median
value is minimum. This means that, it is less than any other value of central
tendency in the data set, which makes it more central in certain situations.
Disadvantages of Median
(i) The data must be arranged in order to find the median. This can be very
time consuming for a large number of elements in the data set.
(ii) The value of the median is affected more by sampling variations. Different
samples from the same population may give significantly different values of
the median.
(iii) The calculation of median in case of grouped data is based on the assumption
that the values of observations are evenly spaced over the entire class interval
Self - Learning
and this is usually not so. Material 209
Statistical Computation and (iv) Median is comparatively less stable than mean, particularly for small samples,
Probability Distribution
due to fluctuations in sampling.
(v) Median is not suitable for further mathematical treatment. For example, we
cannot compute the median of the combined group from the median values
NOTES
of different groups.
4.2.4 Mode
Mode is that value of the variable which occurs or repeats itself the greatest
number of times. The mode is the most ‘Fashionable’ size in the sense that it is
the most common and typical, and is defined by Zizek as ‘the value occurring
most frequently in a series (or group of items) and around which the other items
are distributed most densely’.
The mode of a distribution is the value at the point around which the items
tend to be most heavily concentrated. It is the most frequent or the most common
value, provided that a sufficiently large number of items are available, to give a
smooth distribution. It will correspond to the value of the maximum point (ordinate),
of a frequency distribution if it is an ‘ideal’ or smooth distribution. It may be regarded
as the most typical of a series of values. The modal wage, for example, is the wage
received by more individuals than any other wage. The modal ‘hat’ size is that,
which is worn by more persons than any other single size.
It may be noted that the occurrence of one or a few extremely high or low
values has no effect upon the mode. If a series of data are unclassified, not have
been either arrayed or put into a frequency distribution, the mode cannot be readily
located.
Taking first an extremely simple example, if seven men receive daily wages
of 5, 6, 7, 7, 7, 8 and 10, it is clear that the modal wage is 7 per day. If we
have a series such as 2, 3, 5, 6, 7, 10 and 11, it is apparent that there is no mode.
There are several methods of estimating the value of the mode. However, it
is seldom that the different methods of ascertaining the mode give us identical
results. Consequently, it becomes necessary to decide as to which method would
be most suitable for the purpose in hand. In order that a choice of the method may
be made, we should understand each of the methods and the differences that exist
among them.
The four important methods of estimating mode of a series are: (i) Locating
the most frequently repeated value in the array; (ii) Estimating the mode by
interpolation; (iii) Locating the mode by graphic method; and (iv) Estimating the
mode from the mean and the median. Only the last three methods are discussed in
this unit.
Estimating the Mode by Interpolation: In the case of continuous
frequency distributions, the problem of determining the value of the mode is not so
simple as it might have appeared from the foregoing description. Having located
the modal class of the data, the next problem in the case of continuous series is to
interpolate the value of the mode within this ‘modal’ class.

Self - Learning
210 Material
 The interpolation is made by the use of any one of the following formulae: Statistical Computation and
Probability Distributiona
f2 f0
(i) Mo = l1  × i; (ii) Mo = l2  ×i
f0  f2 f0  f2

f1  f 0 NOTES
(iii) Mo = l1  ×i
( f1  f 0 )  ( f1  f 2 )

Where l1 is the lower limit of the modal class, l2 is the upper limit of the modal
class, f0 equals the frequency of the preceding class in value, f1 equals the frequency
of the modal class in value, f2 equals the frequency of the following class (class
next to modal class) in value, and i equals the interval of the modal class. Example
4.9 explains the method of estimating mode.
Example 4.9: Determine the mode for the data given in the following table:
Wage Group Frequency (f)
14 — 18 6
18 — 22 18
22 — 26 19
26 — 30 12
30 — 34 5
34 — 38 4
38 — 42 3
42 — 46 2
46 — 50 1
50 — 54 0
54 — 58 1

Solution:
In the given data, 22 – 26 is the modal class since it has the largest frequency. The
lower limit of the modal class is 22, its upper limit is 26, its frequency is 19, the
frequency of the preceding class is 18, and of the following class is 12. The class
interval is 4. Using the various methods of determining mode, we have,
12 18
(i) Mo = 22  4 (ii) Mo = 26 – 18  12  4
18  12

8 12
= 22  = 26 –
5 5
= 23.6 = 23.6
19  18 4
(iii) Mo = 22   4 = 22  = 22.5
(19  18)  ( 19  12) 8
In formulae (i) and (ii), the frequency of the classes adjoining the modal
class is used to pull the estimate of the mode away from the midpoint towards
either the upper or lower class limit. In this particular case, the frequency of the
class preceding the modal class is more than the frequency of the class following
and therefore, the estimated mode is less than the midvalue of the modal class.
This seems quite logical. If the frequencies are more on one side of the modal class
than on the other it can be reasonably concluded that the items in the modal class
are concentrated more towards the class limit of the adjoining class with the larger
frequency.
Self - Learning
Material 211
Statistical Computation and Formula (iii) is also based on a logic similar to that of formulae (i) and (ii). In
Probability Distribution
this case, to interpolate the value of the mode within the modal class, the differences
between the frequency of the modal class, and the respective frequencies of the
classes adjoining it are used. This formula usually gives results better than the
NOTES values obtained by the other and exactly equals results obtained by graphic method.
The formulae (i) and (ii) give values which are different from the value obtained by
formula (iii) and are more close to the central point of modal class. If the frequencies
of the class adjoining the modal are equal, the mode is expected to be located at
the midvalue of the modal class, but if the frequency on one of the sides is greater,
the mode will be pulled away from the central point. It will be pulled more and
more if the difference between the frequencies of the classes adjoining the modal
class is higher and higher. In given example in this book, the frequency of the
modal class is 19 and that of the preceding class is 18. So, the mode should be
quite close to the lower limit of the modal class. The midpoint of the modal class is
24 and the lower limit of the modal class is 22.
Locating the Mode by the Graphic Method: The method of graphic
interpolation is shown in Figure 4.3. The upper corners of the rectangle over the
modal class have been joined by straight lines to those of the adjoining rectangles
as shown in Figure 4.3; the right corner to the corresponding one of the adjoining
rectangle on the left, etc. If a perpendicular is drawn from the point of intersection
of these lines, we have a value for the mode indicated on the base line. The
graphic approach is, in principle, similar to the arithmetic interpolation explained
earlier.
The mode may also be determined graphically from an ogive or cumulative
frequency curve. It is found by drawing a perpendicular to the base from that
point on the curve where the curve is most nearly vertical, i.e., steepest (in other
words, where it passes through the greatest distance vertically and smallest distance
horizontally). The point where it cuts the base gives us the value of the mode. How
accurately this method determines the mode is governed by: (i) The shape of the
ogive, (ii) The scale on which the curve is drawn.

Self - Learning Fig. 4.3 Method of Mode Determination by Graphic Interpolation

212 Material
 Estimating the Mode from the Mean and the Median. There usually Statistical Computation and
Probability Distributiona
exists a relationship among the mean, median and mode for moderately asymmetrical
distributions. If the distribution is symmetrical, the mean, median and mode will
have identical values, but if the distribution is skewed (moderately) the mean,
median and mode will pull apart. If the distribution tails off towards higher values, NOTES
the mean and the median will be greater than the mode. If it tails off towards lower
values, the mode will be greater than either of the other two measures. In either
case, the median will be about one-third as far away from the mean as the mode
is. This means that,
Mode = Mean – 3 (Mean – Median)
= 3 Median – 2 Mean
Consider Example 4.10 to better understand the calculation of mode.
Example 4.10: Consider the mean to be 68.53 and the median to be 70.2 Calculate
the mode using the formula discussed.
Solution:
In the case of the average monthly earnings, the mean is 68.53 and the median is
70.2. If these values are substituted in the above formula, we get,
Mode = 68.5 – 3(68.5 –70.2)
= 68.5 + 5.1 = 73.6
According to the formula used earlier,
f2
Mode = l1  ×i
f0  f2
745
= 72.5  ×5
795  745
= 72.5 + 2.4 = 74.9
or
1 f  f
0
Mode = l1  2 f  f  f  i
1 0 2
915  795
= 72.5  ×5
2 × 915  795  745

= 72.5  120 × 5 = 74.57

290
The difference between the two estimates is due to the fact that the assumption
of relationship between the mean, median and mode may not always be true which
is obviously not valid in this case.
Example 4.11: (i) In a moderately symmetrical distribution, the mode and mean
are 32.1 and 35.4 respectively. Calculate the median.
(ii) If the mode and median of moderately asymmetrical series are respectively
16'' and 15.7'', what would be its most probable median?
(iii) In a moderately skewed distribution, the mean and the median are
respectively 25.6 and 26.1 inches. What is the mode of the distribution?
Self - Learning
Material 213
Statistical Computation and Solution:
Probability Distribution
(i) We know,
Mean – Mode = 3 (Mean – Median)
NOTES or 3 Median = Mode + 2 Mean
32.1  2  35.4
or Median =
3
102.9
=
3
= 34.3
(ii) 2 Mean = 3 Median – Mode
1 31.1
or Mean = ( 3 × 15. 7  16. 0)  = 15.55
2 2
(iii) Mode = 3 Median – 2 Mean
= 3 × 26.1 – 2 × 25.6 = 78.3 – 51.2 = 27.1
Advantages of Mode
(i) Similar to median, the mode is not affected by extreme values in the data.
(ii) Its value can be obtained in open-ended distributions without ascertaining
the class limits.
(iii) It can be easily used to describe qualitative phenomenon. For example, if
most people prefer a certain brand of tea, then this will become the modal
point.
(iv) Mode is easy to calculate and understand. In some cases, it can be located
simply by observation or inspection.
Disadvantages of Mode
(i) Quite often, there is no modal value.
(ii) It can be bi-modal or multi-modal, or it can have all modal values making
its significance more difficult to measure.
(iii) If there is more than one modal value, the data is difficult to interpret.
(iv) A mode is not suitable for algebraic manipulations.
(v) Since the mode is the value of maximum frequency in the data set, it cannot
be rigidly defined if such frequency occurs at the beginning or at the end of
the distribution.
(vi) It does not include all observations in the data set, and hence, less reliable
in most of the situations.
4.3.5 Geometric Mean
If , ,  are in GP, then is called a geometric mean between  and , written
as GM.
If a1, a2, .....,an are in GP, then a2, ..., an–1 are called geometric means
between a1 and an.
Thus, 3, 9, 27 are three geometric means between 1 and 81.
Self - Learning
214 Material
 Let G1, G3, ... Gn be n geometric means between a and b. Thus, a, G1, Statistical Computation and
Probability Distributiona
G2, ... Gn, b is a GP, b being (n + 2)th term = arn+1, where r is the common ratio
of GP
1
n+1  b  n 1 NOTES
Thus, b = ar r=  
a
1
1
= a   n 1
b
So, G1 = ar =  a n b  n 1
a
2
1
= a   n 1
b
G2 = ar 2
=  a n 1b 2  n1
a
... .... ... ... ... ... .... ... ... ... ....

n 1
1
= a   n 1
b
Gn = ar n–1
=  a 2b n1  n1
a
Example 4.12: Find 7 GM’s between 1 and 256.
Solution: Let G1, G2, ... G7, be 7 GM’s between 1 and 256
Then, 256= 9th term of GP,
= 1. r8, where r is the common ratio of the GP
This gives that, r8 = 256  r = 2
Thus, G1 = ar = 1.2 = 2
G2 = ar2 = 1.4 = 4
G3 = ar3 = 1.8 = 8
G4 = ar4 = 1.16 = 16
G5 = ar5 = 1.32 = 32
G6 = ar6 = 1.64 = 64
G7 = ar7 = 1.128 = 128
Hence, required GM’s are 2, 4, 8, 16, 32, 64, 128.
Example 4.13: Sum the series 1 + 3x + 5x2 + 7x3 + ... up to n terms, x  1.
Solution: Note that nth term of this series = (2n – 1) xn – 1
Let Sn = 1 + 3x + 5x2 + ... + (2n – 1) xn – 1
Then, xSn = x + 3x2 + ... + (2n – 3) xn – 1 + (2n – 1) xn
Subtracing, we get
Sn(1 – x) = 1 + 2x + 2x2 + ... + 2xn – 1 – (2n – 1) xn
 1  xn  1 
= 1 + 2x  1  x  – (2n – 1) xn

1  x  2 x  2 x n  (2n  1) x n (1  x)
=
1 x

1  x  2 x n  (2n  1) x n  (2n  1) xn  1
= 1 x
Self - Learning
Material 215
Statistical Computation and
Probability Distribution 1  x  (2n  1) x n  (2n  1) x n  1
=
1 x

1  x  (2n  1) x n  (2n  1) x n  1
NOTES Hence, S=
(1  x)2

Example 4.14: Sum the series 5 + 55 + 555 + ... up to n terms.

Solution: Let Sn = 5 + 55 + 555 + . . . .
Sn = 5 (1 + 11 + 111 + . . . . )
5
= (9 + 99 + 999 + . . . )
9
5
= [(10 – 1) + (100 – 1) + (1000 – 1) + ...]
9
5
= [(10 + 102 + 103 + ... + 10n )
9
– (1 + 1 + . . . .n terms)]
5 2 3 n
= [(10 + 10 + 10 + ... + 10 ) – n]
9
5 10(1  10n ) 
=   n
9  1  10 
 n
5 10(10  1) 
=   n
9  9 
50 5n
= (10n  1) 
81 9
124
Example 4.15: Three numbers are in GP. Their product is 64 and sum is .
5
Find them.
a
Solution: Let the numbers be , a, ar
r
a 124 a
Since, + a + ar = and × a × ar = 64,
r 5 r
we have, a3 = 64  a = 4
4 124
This gives, + 4 + 4r =
r 5
1 31
 +1+r=
r 5
2
r 1 26
 =
r 5
 5r2 + 5 = 26r
 5r2 – 26r + 5 = 0
2
 5r – 25r – r + 5 = 0
 5r (r – 5) –1 (r – 5) = 0
 (r – 5) (5r – 1) = 0
1
 r= or 5
5
4
Self - Learning
In either case, numbers are , 4 and 20.
5
216 Material
Example 4.16: Sum to n terms the series Statistical Computation and
Probability Distributiona
0.7 + 0.77 + 0.777 + . . .
Solution: Given series,
= 7 + 0.77 + 0.777 + . . . up to n terms NOTES
= 7 (0.1 + 0.11 + 0.111 + ... up to n terms)
7
= (0.9 + 0.99 + 0.999 + ... up to n terms)
9
7  1  1   1  
=  1    1  2   1  3   ...
9  10 10 10 
7  1 1 
=  n  10  2  ... up to n terms 
9  10 
 
1 (1  1/10n )
7 
n  10
= 9  1 

1
 10 
7  1 1 
=  n  1  n  
9  9 10 
7 1 1 
=  n  1  n  
9 9 10 
Example 4.17: A manufacturer reckons that the value of a machine which costs
him 18750 will depreciate each year by 20%. Find the estimated value at the
end of 5 years.
Solution: At the end of first year the value of machine is
80
= 18750 ×
100
4
= (18750)
5
2
At the end of 2nd year it is equal to   (18750); proceeding in this manner,,
4
 5
5
the estimated value of machine at the end of 5 years is   (18750)
4
 5

64  16
=  18750
125  25
1024
=  750
125
= 1024 × 6
= 6144 rupees
Example 4.18: Show that a given sum of money accumulated at 20 % per annum,
more than doubles itself in 4 years at compound interest.
6a
Solution: Let the given sum be a rupees. After 1 year it becomes (it is increased
5
a Self - Learning
by ). Material 217
5
Statistical Computation and 2
Probability Distribution 6  6a   6 
At the end of two years it becomes     a
5  5   5
Proceeding in this manner, we get that at the end of 4th year, the amount will
4
NOTES be   a =
6 1296
a
 5 625
1296 46
Now, a  2a  a, since a is a + ve quantity, so the amount after 4
625 625
years is more than double of the original amount.
4.3.6 Harmonic Mean
If a, b, c are in HP, then b is called a Harmonic Mean between a and c, written as
HM
Let H1, H2, H3, ..., Hn be the required Harmonic Means. Then
a, H1, H2, ..., Hn, b are in HP
1 1 1 1 1
i.e., , ,
a H1 H 2
, ..., ,
Hn b are in AP

1
Then, = (n + 2)th term of an AP
b
1
= + (n + 1)d
a
Where d is the common difference of AP.
ab
This gives, d=
(n  1)ab
1 1 1 ab
Now, = d  
H1 a a (n  1) ab
nb  b  a  b a  nb
= 
(n  1) ab (n  1) ab
1 a  nb
So, =
H1 (n  1) ab
(n  1) ab
 H1 =
a  nb
1 1 1 2 ( a  b)
Again, =  2d  
H2 a a (n  1) ab
nb  b  2a  2b 2a  b  nb
= 
(n  1) ab (n  1) ab
(n  1) ab
 H2 =
2a  b  nb
1 1 3a  2b  nb
Similarly, =  3d 
H3 a ( n  1) ab
(n  1) ab
 H3 = and so on,
3a  2b  nb
1 1 1 n ( a  b)
=  nd  
Hn a a (n  1) ab
nb  b  na  nb
=
(n  1) ab
Self - Learning
na  b (n  1) ab
218 Material =  Hn =
(n  1) ab na  b
 Statistical Computation and
Example 4.19: Find the 5th term of 2, 2 12 , 3 13 , ... ... Probability Distributiona
1 1 2 3
Solution: Let 5th term be x. Then, is 5th term of corresponding AP , , , ... ...
x 2 5 10

1 1 2 1 1  1 
NOTES
Then, =  4     4 
x 2 5 2 2  10 

1 1 2 1
 =    x = 10
x 2 5 10
1 4
Example 4.20: Insert two harmonic means between and .
2 17
1 4
Solution: Let H1, H2 be two harmonic means between and
2 17
1 1 17
Thus, 2, , , are in AP Let d be their common difference
H1 H 2 4
17
Then, = 2 + 3d
4
9 3
 3d =  d=
4 4
1 3 11 4
Thus, =2+   H1 =
H1 4 4 11
1 3 7 2
=2+2×   H2 =
H2 4 2 7
4 2
Required harmonic means are , .
11 7

4.3.7 Quartiles, Percentiles and Deciles

Some measures, other than the measures of central tendency, are often employed
when summarizing or describing a set of data where it is necessary to divide the
data into equal parts. These are positional measures and are called quantiles and
consist of quartiles, deciles and percentiles. The quartiles divide the data into four
equal parts. The deciles divide the total ordered data into ten equal parts and the
percentiles divide the data into 100 equal parts. Consequently, there are three
quartiles, nine deciles and 99 percentiles. The quartiles are denoted by the symbol
Q, which can be fractioned as Q1, Q2, Q3, ..., and so on. Here, Q1 will be such
point in the ordered data which has 25 per cent of the data below and Q2 will
represent 75 per cent of the data above it. In other words, Q1 is the value
 n  1
corresponding to   th ordered observation. Similarly, Q2 divides the data in
4 
the middle, and is also equal to the median and its value, Q2 is given by:
 n 1
Q2 = The value of 2   th ordered observation in the data.
 4 
Similarly, we can calculate the values of various deciles. For instance,
 n 1
D1 =   th observaton in the ordered data, and
 10  Self - Learning
Material 219
Statistical Computation and
Probability Distribution  n 1
D7 = 7   th observation in the ordered data.
 10 
Percentiles are generally used in the research area of education where people
NOTES are given standard tests and it is desirable to compare the relative position of the
subject’s performance on the test. Percentiles are similarly calculated as,
 n 1
P7 = 7  th observation in the ordered data.
 100 
and,
 n 1
P69 = 69  th observation in the ordered data.
 100 
Quartiles
The formula for calculating the values of quartiles for grouped data is given as,
Q = L + (j/f)C
Where,
Q = The quartile under consideration.
L = Lower limit of the class interval which contains the value of Q.
j = The number of units we lack from the class interval which contains
the value of Q, in reaching the value of Q.
f = Frequency of the class interval containing Q.
C = Size of the class interval.
Let us assume, we took the data of the ages of 100 students and a frequency
distribution for this data has been constructed as shown in Table 4.4.
Table 4.4 The Frequency Distribution Ages of 100 Students

Ages (CI) Mid-point (X) (f) f(X) f(X)2

16 and upto 17 16.5 4 66 1089.0
17 and upto 18 17.5 14 245 4287.5
18 and upto 19 18.5 18 333 6160.5
19 and upto 20 19.5 28 546 10647.0
20 and upto 21 20.5 20 410 8405.0
21 and upto 22 21.5 12 258 5547.0
22 and upto 23 22.5 4 90 2025.0
Total = 100 1948 38161
In our case, in order to find Q1, where Q1 is the cut-off point so that 25 per
cent of the data is below this point and 75 per cent of the data is above, we see
that the first group has 4 students and the second group has 14 students, making
a total of 18 students. Since Q1 cuts off at 25 students, it is the third class interval
which contains Q1. This means that the value of L in our formula is 18.
Since we already have 18 students in the first two groups, we need 7 more
students from the third group to make it a total of 25 students, which is the value
of Q1. Hence, the value of (j) is 7. Also, since the frequency of this third class
interval which contains Q1 is 18, the value of (f) in our formula is 18. The size of the
Self - Learning class interval C is given as 1. Substituting these values in the formula for Q, we get,
220 Material
 Q1 = 18 + (7/18)1 Statistical Computation and
Probability Distributiona
= 18 + 0.38 = 18.38
This means that 25 per cent of the students are below 18.38 years of age
and 75 per cent are above this age. NOTES
Similarly, we can calculate the value of Q2, using the same formula. Hence,
Q2 = L + (j/f)C
= 19 + (14/28)1
= 19.5
This also happens to be the median.
By using the same formula and the same logic we can calculate the values of all
deciles as well as percentiles.
We have defined the median as the value of the item which is located at the
centre of the array. We can define other measures which are located at other
specified points. Thus, the Nth percentile of an array is the value of the item such
that N per cent items lie below it. Clearly then, the Nth percentile Pn of grouped
data is given by,
nN
C
Pn = l  100 ×i
f
Here, l is the lower limit of the class in which nN/100th item lies, i its width,
f its frequency, C the cumulative frequency upto (but not including) this class, and
N is the total number of items.
We can similarly define the Nth decile as the value of the item below which
(nN/10) items of the array lie. Clearly,
nN
C
Dn = P10n = l  10 ×i
f
where the symbols have the obvious meanings.
The other most commonly referred to measures of location are the quartiles.
Thus, nth quartile is the value of the item which lies at the n(N/5)th item. Clearly,
Q2, the second quartile, is the median for grouped data.
nN
C
Qn = P25n l 4 ×i
f

4.3.8 Box Plot

In descriptive statistics, a box plot or boxplot, also called as box-and-whisker
diagram or plot, is the most appropriate method to depict clusters of numerical
data graphically by means of the following:
1. The Smallest Observation (Sample Minimum).
2. The Lower Quartile (Q1).
3. The Median (Q2).
4. The Upper Quartile (Q3).
Self - Learning
5. The Largest Observation (Sample Maximum). Material 221
Statistical Computation and A box plot also specifies that which observations can be considered as
Probability Distribution
outliers. Box plots are non-parametric and can be drawn either horizontally or
vertically.
The box plots are formed as follows:
NOTES
 Vertical Axis: Response Variable.
 Horizontal Axis: The Factor of Interest.
To draw a box plot, perform the following steps:
Step 1: Calculate the median and the quartiles. The lower quartile is the 25th
percentile and the upper quartile is the 75th percentile.
Step 2: Plot a symbol at the median or draw a line and then draw a box between
the lower and upper quartiles. This box represents the middle 50% of the data
which is the body of the data.
Step 3: Draw first line from the lower quartile to the minimum point and another
line from the upper quartile to the maximum point. Typically, a symbol is drawn at
these minimum and maximum points, although this is optional.
Step 4: Calculate the interquartile range, i.e., the difference between the upper
and lower quartile, called IQ.
Step 5: Now calculate the following points:
L1 = Lower Quartile  1.5*IQ
L2 = Lower Quartile  3.0*IQ
U1 = Upper Quartile + 1.5*IQ
U2 = Upper Quartile + 3.0*IQ

Step 6: The line from the lower quartile to the minimum can be drawn from the
lower quartile to the smallest point that is greater than L1. Similarly, the line from
the upper quartile to the maximum can be drawn to the largest point smaller than
U1.
Step 7: Points between L1 and L2 or between U1 and U2 can be drawn as small
circles. Points less than L2 or greater than U2 can be drawn as large circles
Thus the box plot identifies the middle 50% of the data, the median and the extreme
points. A single box plot can be drawn for one set of data with no distinct groups.
Alternatively, multiple box plots can be drawn together to compare multiple data
sets or to compare groups in a single data set. For a single box plot, the width of
the box is arbitrary. For multiple box plots, the width of the box plot can be set
proportional to the number of points in the given group or sample.

Self - Learning
222 Material
 Statistical Computation and
Probability Distributiona
Check Your Progress
1. Define statistics.
2. How does statistics classify numerical facts? NOTES
3. What is the first step in the statistical treatment of a problem?
4. What is central tendency in statistics?
5. Define the term arithmetic mean.
6. When is weighted arithmetic mean used?
7. Define the term median.
8. What is mode?
9. What are the four important methods of estimating mode of a series?
10. What are positional measures?

4.4 MEASURES OF DISPERSION

A measure of dispersion, or simply dispersion may be defined as statistics signifying
the extent of the scatteredness of items around a measure of central tendency.
A measure of dispersion may be expressed in an ‘Absolute form’, or in a
‘Relative form’. It is said to be in an absolute form when it states the actual amount
by which the value of an item on an average deviates from a measure of central
tendency. Absolute measures are expressed in concrete units, i.e., units in terms of
which the data have been expressed, e.g., rupees, centimetres, kilograms, etc.,
and are used to describe frequency distribution.
A relative measure of dispersion computed is a quotient by dividing the
absolute measures by a quantity in respect to which absolute deviation has been
computed. It is as such a pure number and is usually expressed in a percentage
form. Relative measures are used for making comparisons between two or more
distributions.
A measure of dispersion should possess all those characteristics which are
considered essential for a measure of central tendency, viz.
 It should be based on all observations.
 It should be readily comprehensible.
 It should be fairly easily calculated.
 It should be affected as little as possible by fluctuations of sampling.
 It should be amenable to algebraic treatment.
The following are some common measures of dispersion:
(i) The range, (ii) the Semi-interquartile range or the quartile deviation, (iii) The
mean deviation, and (iv) The standard deviation. Of these, the standard deviation
is the best measure. We describe these measures in the following sections.

Self - Learning
Material 223
Statistical Computation and
Probability Distribution
4.4.1 Range
The crudest measure of dispersion is the range of the distribution. The range of
any series is the difference between the highest and the lowest values in the series.
NOTES If the marks received in an examination taken by 248 students are arranged in
ascending order, then the range will be equal to the difference between the highest
and the lowest marks.
In a frequency distribution, the range is taken to be the difference between
the lower limit of the class at the lower extreme of the distribution and the upper
limit of the class at the upper extreme.
Table 4.5 Weekly Earnings of Labourers in Four Workshops of the Same Type

No. of workers
Weekly earnings
Workshop A Workshop B Workshop C Workshop D
15–16 ... ... 2 ...
17–18 ... 2 4 ...
19–20 ... 4 4 4
21–22 10 10 10 14
23–24 22 14 16 16
25–26 20 18 14 16
27–28 14 16 12 12
29–30 14 10 6 12
31–32 ... 6 6 4
33–34 ... ... 2 2
35–36 ... ... ... ...
37–38 ... ... 4 ...

Total 80 80 80 80
Mean 25.5 25.5 25.5 25.5

Consider the data on weekly earning of worker on four workshops given in

the Table 4.5. We note the following:
Workshop Range
A 9
B 15
C 23
D 15
From these figures, it is clear that the greater the range, the greater is the
variation of the values in the group.
The range is a measure of absolute dispersion and as such cannot be usefully
employed for comparing the variability of two distributions expressed in different
units. The amount of dispersion measured, say, in pounds, is not comparable with
dispersion measured in inches. So the need of measuring relative dispersion arises.
An absolute measure can be converted into a relative measure if we divide it by
some other value regarded as standard for the purpose. We may use the mean of
the distribution or any other positional average as the standard.

Self - Learning
224 Material
For Table 4.5, the relative dispersion would be: Statistical Computation and
Probability Distributiona
9 23
Workshop A = Workshop C =
25.5 25.5

Workshop B =
15
Workshop D =
15 NOTES
25.5 25.5

An alternate method of converting an absolute variation into a relative one would

be to use the total of the extremes as the standard. This will be equal to dividing
the difference of the extreme items by the total of the extreme items. Thus,
Difference of extreme items, i.e, Range
Relative Dispersion =
Sum of extreme items
The relative dispersion of the series is called the coefficient or ratio of dispersion. In
our example of weekly earnings of workers considered earlier, the coefficients
would be:
9 9 15 15
Workshop A =  Workshop B = 
21  30 51 17  32 49

23 23 15 15
Workshop C =  Workshop D = 
15  38 53 19  34 53

Merits and limitations of range

Merits
Of the various characteristics that a good measure of dispersion should possess,
the range has only two, viz (i) It is easy to understand, and (ii) Its computation is
simple.
Limitations
Besides the aforesaid two qualities, the range does not satisfy the other test of a
good measure and hence it is often termed as a crude measure of dispersion.
The following are the limitations that are inherent in the range as a concept
of variability:
(i) Since it is based upon two extreme cases in the entire distribution, the
range may be considerably changed if either of the extreme cases
happens to drop out, while the removal of any other case would not
affect it at all.
(ii) It does not tell anything about the distribution of values in the series
relative to a measure of central tendency.
(iii) It cannot be computed when distribution has open-end classes.
(iv) It does not take into account the entire data. These can be illustrated
by the following illustration. Consider the data given in Table 4.6.

Self - Learning
Material 225
Statistical Computation and Table 4.6 Distribution with the Same Number of Cases,
Probability Distribution but Different Variability
No. of students
Class
Section Section Section
NOTES A B C
0–10 ... ... ...
10–20 1 ... ...
20–30 12 12 19
30–40 17 20 18
40–50 29 35 16
50–60 18 25 18
60–70 16 10 18
70–80 6 8 21
80–90 11 ... ...
90–100 ... ... ...
Total 110 110 110
Range 80 60 60

The table is designed to illustrate three distributions with the same number of cases
but different variability. The removal of two extreme students from section A would
make its range equal to that of B or C.
The greater range of A is not a description of the entire group of 110 students, but
of the two most extreme students only. Further, though sections B and C have the
same range, the students in section B cluster more closely around the central
tendency of the group than they do in section C. Thus, the range fails to reveal the
greater homogeneity of B or the greater dispersion of C. Due to this defect, it is
seldom used as a measure of dispersion.
Specific uses of range
In spite of the numerous limitations of the range as a measure of dispersion, there
are the following circumstances when it is the most appropriate one:
(i) In situations where the extremes involve some hazard for which preparation
should be made, it may be more important to know the most extreme cases
to be encountered than to know anything else about the distribution. For
example, an explorer, would like to know the lowest and the highest
temperatures on record in the region he is about to enter; or an engineer
would like to know the maximum rainfall during 24 hours for the construction
of a storem water drain.
(ii) In the study of prices of securities, range has a special field of activity. Thus
to highlight fluctuations in the prices of shares or bullion it is a common
practice to indicate the range over which the prices have moved during a
certain period of time. This information, besides being of use to the operators,
gives an indication of the stability of the bullion market, or that of the
investment climate.
(iii) In statistical quality control the range is used as a measure of variation. We,
e.g., determine the range over which variations in quality are due to random
causes, which is made the basis for the fixation of control limits.
Self - Learning
226 Material
4.4.2 Quartile Deviation Statistical Computation and
Probability Distributiona
Another measure of dispersion, much better than the range, is the semi-interquartile
range, usually termed as ‘Quartile Deviation’. As stated in the previous unit, quartiles
are the points which divide the array in four equal parts. More precisely, Q1 gives
NOTES
the value of the item 1/4th the way up the distribution and Q3 the value of the item
3/4th the way up the distribution. Between Q1 and Q3 are included half the total
number of items. The difference between Q1 and Q3 includes only the central
items but excludes the extremes. Since under most circumstances, the central half
of the series tends to be fairly typical of all the items, the interquartile range
(Q3– Q1) affords a convenient and often a good indicator of the absolute variability.
The larger the interquartile range, the larger the variability.
Usually, one-half of the difference between Q3 and Q1 is used and to it is given the
name of quartile deviation or semi-interquartile range. The interquartile range is
divided by two for the reason that half of the interquartile range will, in a normal
distribution, be equal to the difference between the median and any quartile. This
means that 50 per cent items of a normal distribution will lie within the interval
defined by the median plus and minus the semi-interquartile range.
Symbolically:
Q3  Q1
Q.D. = ...(4.1)
2
Let us find quartile deviations for the weekly earnings of labour in the four workshop
whose data is given in Table 4.5. The computations are as shown in Table 4.7.
As shown in the table, Q.D. of workshop A is 2.12 and median value in 25.3.
This means that if the distribution is symmetrical the number of workers whose
wages vary between (25.3–2.1) = 23.2 and (25.3 + 2.1) = 27.4, shall be just
half of the total cases. The other half of the workers will be more than 2.1
removed from the median wage. As this distribution is not symmetrical, the distance
between Q1 and the median Q2 is not the same as between Q3 and the median.
Hence the interval defined by median plus and minus semi inter-quartile range will
not be exactly the same as given by the value of the two quartiles. Under such
conditions the range between 23.2 and 27.4 will not include precisely 50 per
cent of the workers.
If quartile deviation is to be used for comparing the variability of any two series, it
is necessary to convert the absolute measure to a coefficient of quartile deviation.
To do this the absolute measure is divided by the average size of the two quartile.
Symbolically:
Q3  Q1
Coefficient of quartile deviation = ...(4.2)
Q3  Q1

Applying this to our illustration of four workshops, the coefficients of Q.D. are as
given below.

Self - Learning
Material 227
Statistical Computation and Table 4.7 Calculation of Quartile Deviation
Probability Distribution
Workshop Workshop Workshop Workshop
A B C D

N 80 80 80 80
NOTES Location of Q2
2 2
 40
2
 40
2
 40
2
 40

40  30 40  30 40  30 40  30
Q2 24.5 + 2 24.5   2 24.5 +  2 24.5 + 2
22 18 16 16
= 24.5 + 0.9 = 24.5 + 1.1 = 24.5 + 0.75 = 24.5 + 0.75
= 25.4 = 25.61 = 25.25 = 25.25
N 80 80 80 80
Location of Q1  20  20  20  20
4 4 4 4 4

20  10 20  16 20  10 20  18
Q1 22.5  2 22.5   2 20.5   2 22.5  2
22 14 10 16
= 22.5 + .91 = 22.5 + .57 = 20.5 + 2 = 22.5 + .25
= 23.41 = 23.07 = 22.5 = 22.75

3N 80
Location of Q3 3  60 60 60 60
4 4

60  52 60  48 60  50
Q3 26.5  2 26.5  2 26.5  2
14 16 12
60  50
26.5  2
12
= 26.5 + 1.14 = 26.5 + 1.5= 26.5 + 1.67 = 26.5 + 1.67
= 27.64 = 28.0 = 28.17 = 28.17

Q3  Q1 27.64  23.41 28  23.07 28.17  22.5 28.17  22. 75

Quartile Deviation
2 2 2 2 2
4.23 4.93 5.67 5.42
= = 2.12 = = 2.46 = = 2.83= = . 2.71
2 2 2 2

Coefficient of quartile
Q3  Q1 27. 64  23. 41 28  23. 07 28.17  22.5 28.17  22. 75
deviation Q  Q =
3 1 27. 64  23. 41 28  23.07 28.17  22.5 28.17  22. 75
= 0.083 = 0.097 = 0.112 = 0.106

Characteristics of quartile deviation

 The size of the quartile deviation gives an indication about the uniformity or
otherwise of the size of the items of a distribution. If the quartile deviation is
small it denotes large uniformity. Thus, a coefficient of quartile deviation
may be used for comparing uniformity or variation in different distributions.
 Quartile deviation is not a measure of dispersion in the sense that it does
not show the scatter around an average, but only a distance on scale.
Consequently, quartile deviation is regarded as a measure of partition.
 It can be computed when the distribution has open-end classes.
Limitations of quartile deviation
Except for the fact that its computation is simple and it is easy to understand, a
quartile deviation does not satisfy any other test of a good measure of variation.
Self - Learning
228 Material
4.4.3 Mean Deviation Statistical Computation and
Probability Distributiona
A weakness of the measures of dispersion discussed earlier, based upon the range
or a portion thereof, is that the precise size of most of the variants has no effect on
the result. As an illustration, the quartile deviation will be the same whether the
NOTES
variates between Q1 and Q3 are concentrated just above Q1 or they are spread
uniformly from Q1 to Q3. This is an important defect from the viewpoint of measuring
the divergence of the distribution from its typical value. The mean deviation is
employed to answer the objection.
Mean deviation also called average deviation, of a frequency distribution is
the mean of the absolute values of the deviation from some measure of central
tendency. In other words, mean deviation is the arithmetic average of the variations
(deviations) of the individual items of the series from a measure of their central
tendency.
We can measure the deviations from any measure of central tendency, but
the most commonly employed ones are the median and the mean. The median is
preferred because it has the important property that the average deviation from it
is the least.
Calculation of the mean deviation then involves the following steps:
(a) Calculate the median (or the mean) Me (or X ).
(b) Record the deviations | d | = | x – Me | of each of the items, ignoring the
sign.
(c) Find the average value of deviations.

Mean Deviation =  |d | ...(4.3)

N
Example 4.21: Calculate the mean deviation from the following data giving marks
obtained by 11 students in a class test.
14, 15, 23, 20, 10, 30, 19, 18, 16, 25, 12.

Solution: Median = Size of 11 + 1 th item

2

= size of 6th item = 18.

Serial No. Marks | x – Median |
|d|
1 10 8
2 12 6
3 14 4
4 15 3
5 16 2
6 18 0
7 19 1
8 20 2
9 23 5
10 25 7
11 30 12
Self - Learning
 |d | = 50 Material 229
Statistical Computation and
=
|d |
Probability Distribution Mean deviation from median
N
50
= = 4.5 marks.
11
NOTES
For grouped data, it is easy to see that the mean deviation is given by
 f |d |
Mean deviation, M.D. = ...(1)
f

where | d | = | x – median | for grouped discrete data, and | d | = M – median | for

grouped continuous data with M as the mid-value of a particular group. The following
examples illustrate the use of this formula.
Example 4.22: Calculate the mean deviation from the following data
Size of item 6 7 8 9 10 11 12
Frequency 3 6 9 13 8 5 4
Solution:

Size Frequency Cumulative Deviations f|d|

f frequency from median
(9)
|d|

6 3 3 3 9
7 6 9 2 12
8 9 18 1 9
9 13 31 0 0
10 8 39 1 8
11 5 44 2 10
12 4 48 3 12

48 60

48  1
Median = the size of = 24.5th item which is 9.
2
Therefore, deviations d are calculated from 9, i.e., | d | = | x – 9 |.
 f |d | 60
Mean deviation =  f = = 1.25
48
Example 4.23: Calculate the mean deviation from the following data:

x 0–10 10–20 20–30 30–40 40–50 50–60 60–70 70–80

f 18 16 15 12 10 5 2 2

Solution:
This is a frequency distribution with continuous variable. Thus, deviations are
calculated from mid-values.
Self - Learning
230 Material
 Statistical Computation and
x Mid-value f Less than Deviation f| d |
Probability Distributiona
c.f. from median
|d|
0–10 5 18 18 19 342
10–20 15 16 34 9 144 NOTES
20–30 25 15 49 1 15
30–40 35 12 61 11 132
40–50 45 10 71 21 210
50–60 55 5 76 31 155
60–70 65 2 78 41 82
70–80 75 2 80 51 102

80 1182

80
Median = the size of th item
2
6
= 20  × 10 = 24
15
 f |d |
and then, mean deviation = f

1182
= = 14.775.
80

Merits and demerits of the mean deviation

Merits
 It is easy to understand.
 As compared to standard deviation (discussed later), its computation is
simple.
 As compared to standard deviation, it is less affected by extreme values.
 Since it is based on all values in the distribution, it is better than range or
quartile deviation.
Demerits
 It lacks those algebraic properties which would facilitate its computation
and establish its relation to other measures.
 Due to this, it is not suitable for further mathematical processing.
Coefficient of mean deviation
The coefficient or relative dispersion is found by dividing the mean deviations
recorded. Thus,
Mean Deviation
Coefficient of M.D. = ...(4.4)
Mean
(when deviations were recorded from the mean)
M.D.
= ...(4.5)
Median
(when deviations were recorded from the median)
Self - Learning
Applying the above formula to Example 4.23. Material 231
Statistical Computation and
14.775
Probability Distribution Coefficient of Mean deviation =
24

= 0.616
NOTES
4.5 STANDARD DEVIATION
By far the most universally used and the most useful measure of dispersion is the
standard deviation or root mean square deviation about the mean. We have seen
that all the methods of measuring dispersion so far discussed are not universally
adopted for want of adequacy and accuracy. The range is not satisfactory as its
magnitude is determined by most extreme cases in the entire group. Further, the
range is notable because it is dependent on the item whose size is largely matter of
chance. Mean deviation method is also an unsatisfactory measure of scatter, as it
ignores the algebraic signs of deviation. We desire a measure of scatter which is
free from these shortcomings. To some extent standard deviation is one such
measure.
The calculation of standard deviation differs in the following respects from
that of mean deviation. First, in calculating standard deviation, the deviations are
squared. This is done so as to get rid of negative signs without committing algebraic
violence. Further, the squaring of deviations provides added weight to the extreme
items, a desirable feature for certain types of series.
Secondly, the deviations are always recorded from the arithmetic mean,
because although the sum of deviations is the minimum from the median, the sum
of squares of deviations is minimum when deviations are measured from the
arithmetic average. The deviation from x is represented by d.
Thus, standard deviation,  (sigma) is defined as the square root of the
mean of the squares of the deviations of individual items from their arithmetic
mean.
2
(x  x)
 = ...(4.6)
N
For grouped data (discrete variables)
2
 f (x  x )
 = ...(4.7)
f
and, for grouped data (continuous variables)
 f (M  x)
 = ...(4.8)
f

where M is the mid-value of the group.

The use of these formulae is illustrated by the following examples.
Example 4.24: Compute the standard deviation for the following data:
11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21.
Self - Learning
232 Material
Solution: Statistical Computation and
Probability Distributiona
Here Equation (4.6) is appropriate. We first calculate the mean as x =  x/ N =
176/11= 16, and then calculate the deviation as follows:
x (x – x ) (x – x )2
NOTES
11 –5 25
12 –4 16
13 –3 9
14 –2 4
15 –1 1
16 0 0
17 +1 1
18 +2 4
19 +3 9
20 +4 16
21 +5 25

176 10

Thus by Equation (4.6).

110
=  10 = 3.16
11

Example 4.25: Find the standard deviation of the data in the following distributions:

x 12 13 14 15 16 17 18 20
f 4 11 32 21 15 8 6 4

Solution:
For this discrete variable grouped data, we use formula 8. Since for calculation of
x , we need  fx and then for  we need  f ( x  x ) 2 , the calculations are
conveniently made in the following format.

x f fx d=x– x d2 fd2
12 4 48 –3 9 36
13 11 143 –2 4 44
14 32 448 –1 1 32
15 21 315 0 0 0
16 15 240 1 1 15
17 8 136 2 4 32
18 5 90 3 9 45
20 4 80 5 25 100

100 1500 304

Here x =  fx /  f = 1500/100 = 15

 fd 2
and  =
f

304
= = 3. 04 = 1.74 Self - Learning
100
Material 233
Statistical Computation and Example 4.26: Calculate the standard deviation of the following data.
Probability Distribution

Class 1–3 3–5 5–7 7–19 9–11 11–13 13–15

frequency 1 9 25 35 17 10 3
NOTES
Solution: This is an example of continuous frequency series and formula 9 seems
appropriate.

Class Mid- Frequency Deviation Squared Squared

point of mid- deviation deviation
point x times
x f fx from d2 frequency
mean (8) d2

1–3 2 1 2 –6 36 36
3–5 4 9 36 –4 16 144
5–7 6 25 150 –2 4 100
7–9 8 35 280 0 0 0
9–11 10 17 170 2 4 68
11–13 12 10 120 4 16 160
13–15 14 3 42 6 36 108
100 800 616

First the mean is calculated as

x =  fx / x = 800/100 = 8.0
Then the deviations are obtained from 8.0. The standard deviation

 f ( M  x )2
=
f

 fd 2 616
= 
f 100
= 2.48
4.5.1 Calculation of Standard Deviation by Short-cut
Method
The three examples worked out above have one common simplifying feature,
namely x in each, turned out to be an integer, thus, simplifying calculations. In
most cases, it is very unlikely that it will turn out to be so. In such cases, the
calculation of d and d2 becomes quite time-consuming. Short-cut methods have
consequently been developed. These are on the same lines as those for calculation
of mean itself.
In the short-cut method, we calculate deviations x' from an assumed mean A.
Then,
for ungrouped data

 =  x 2 FG
 x IJ 2
...(4.9)
N

N H K
Self - Learning
234 Material
and for grouped data Statistical Computation and
Probability Distributiona

 =  fx  2

FG IJ
fx 
2
...(4.10)
f f H K
NOTES
This formula is valid for both discrete and continuous variables. In case of continuous
variables, x in the equation x' = x – A stands for the mid-value of the class in
question.
Note that the second term in each of the formulae is a correction term because of
the difference in the values of A and x . When A is taken as x itself, this correction
is automatically reduced to zero.
Example 4.27: Compute the standard deviation by the short-cut method for the
following data:
11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21
Solution: Let us assume that A = 15.

x' = (x – 15) x'2

11 –4 16
12 –3 9
13 –2 4
14 –1 1
15 0 0
16 1 1
17 2 4
18 3 9
19 4 16
20 5 25
21 6 36

N = 11  x  = 11  x  2 = 121

=  x 2 FG
 x IJ 2

N

H
N K
= 121  11 FG IJ 2

11 H 11K
= 11  1
= 10
= 3.16.
Another method
If we assumed A as zero, then the deviation of each item from the assumed mean
is the same as the value of item itself. Thus, 11 deviates from the assumed mean of
zero by 11, 12 deviates by 12, and so on. As such, we work with deviations
without having to compute them, and the formula takes the following shape:
Self - Learning
Material 235
Statistical Computation and
x x2
Probability Distribution
11 121
12 144
13 169
NOTES 14 196
15 225
16 256
17 289
18 324
19 361
20 400
21 441

176 2,926

 =  x2 xFG IJ 2

N

N H K
= 926  176 FG IJ 2
= 266  256 = 3.16
11 H 11 K
Example 4.28: Calculate the standard deviation of the following data by short
method.
Person 1 2 3 4 5 6 7
Monthly income
(Rupees) 300 400 420 440 460 480 580

Solution: In this data, the values of the variable are very large making calculations
cumbersome. It is advantageous to take a common factor out. Thus, we use
x A
x' = . The standard deviation is calculated using x' and then the true value of
20
 is obtained by multiplying back by 20. The effective formula then is

=  x2 FG
 x IJ 2
C×
N

HN K
where C represents the common factor.
Using x' = (x – 420)/20.

x Deviation from x' x'2

Assumed mean
x = (x – 420)

300 –120 –6 36
400 –20 –1 1
420 0 0 0
–7
440 20 1 1
460 40 2 4
480 60 3 9
580 160 8 64
+ 14
Self - Learning N=7 7 115
236 Material
 Statistical Computation and
=  x2  x FG IJ 2
Probability Distributiona
20 ×
N

N H K
= 20 115  7 FG IJ 2
NOTES
7 H 7K
= 78.56
Example 4.29: Calculate the standard deviation from the following data:
Size 6 9 12 15 18
Frequency 7 12 19 10 2
Solution:
x Frequency Deviation Deviation x' times x'2 times
f from divided frequency frequency
assumed by common fx' fx'2
mean 12 factor 3
x'
6 7 –6 –2 –14 28
9 12 –3 –1 –12 12
12 19 0 0 0 0
15 10 3 1 10 10
18 2 6 2 4 8

N = 50  fx  fx 2
= –12 = 58

Since deviations have been divided by a common factor, we use

 =C  fx  2

FG
 fx  IJ 2

N N H K
= 3 58  12 FG IJ 2

50 H 50 K
= 3 1.1600  . 0576 = 3 × 1.05 = 3.15.
Example 4.30: Obtain the mean and standard deviation of the first N natural
numbers, i.e., of 1, 2, 3, ..., N – 1, N.
Solution: Let x denote the variable which assumes the values of the first N natural
numbers.
Then
N N ( N  1)
x
1 2 N 1
x =  
N N 2
N
because  x = 1 + 2 + 3 + ... + (N – 1) + N
1

N ( N  1)
=
2 Self - Learning
Material 237
Statistical Computation and To calculate the standard deviation , we use 0 as the assumed mean A. Then
Probability Distribution

=  x2 x FG IJ 2

N

N H K
NOTES N ( N  1) ( 2 N  1)
But  x2 = 12 + 22 + 32 + ... (N – 1)2 + N2 =
6
Therefore
N ( N  1) ( 2 N  1) N 2 ( N  1) 2
 = 
6N 4N 2

=
LM
( N  1) 2 N  1 N  1

OP = ( N  1) ( N  1)
2 N3 2 Q 12
Thus for first 11 natural numbers
11  1
x = 6
2
(11  1) (11  1)
and  = = 10 = 3.16
12
Example 4.31:
Mid- Frequency Deviation Deviation Squared
point f from class time deviation
x of assumed frequency times
mean fx' frequency
x' fx'2
0–10 5 18 –2 –36 72
10–20 15 16 –1 –16 16
–52
20–30 25 15 0 0 0
30–40 35 12 1 12 12
40–50 45 10 2 20 40
50–60 55 5 3 15 45
60–70 65 2 4 8 32
70–80 75 1 5 5 25
– 60

79 60 242
–52
 fx = 8

Solution: Since the deviations are from assumed mean and expressed in terms of
class-interval units,

 =  x2

FG IJ
 fx 
2
i×
N H K
N

242 F 8 I
2
= 10 × G J
79 H 79 K

= 10 × 1.75 = 17.5.

Self - Learning
238 Material
4.5.2 Combining Standard Deviations of Two Statistical Computation and
Probability Distributiona
Distributions
If we were given two sets of data of N1 and N2 items with means x1 and x 2 and
standard deviations 1 and 2 respectively, we can obtain the mean and standard NOTES
deviation x and  of the combined distribution by the following formulae:
N 1 x1  N 2 x 2
x = ...(4.11)
N1  N 2

N 1 12  N 2  22  N 1 ( x  x1 ) 2  N 2 ( x  x 2 ) 2
and  = ...(4.12)
N1  N 2

Example 4.32: The mean and standard deviations of two distributions of 100
and 150 items are 50, 5 and 40, 6 respectively. Find the standard deviation of all
taken together.
Solution: Combined mean
N 1 x1  N 2 x 2 100 × 50  150 × 40
x = =
N1  N 2 100  150

= 44
Combined standard deviation

N112  N 2 22  N1 ( x  x1 )2  N 2 ( x  x2 ) 2
 =
N1  N 2

100 × (5) 2  150 ( 6) 2  100 ( 44  50 ) 2  150 ( 44  40 ) 2

=
100  150

= 7.46.
Example 4.33: A distribution consists of three components with 200, 250, 300
items having mean 25, 10 and 15 and standard deviation 3, 4 and 5, respectively.
Find the standard deviation of the combined distribution.
Solution: In the usual notations, we are given here
N 1 = 200, N2= 250, N3 = 300
x1 = 25, x 2 = 10, x 3 = 15
The Equations (4.11) and (4.12) can easily be extended for combination of three
series as
N 1 x1  N 2 x 2  N 3 x 3
x =
N1  N 2  N 3
200 × 25  250 × 10  300 × 15
=
200  250  300

= 12000 = 16
750

Self - Learning
Material 239
Statistical Computation and and
Probability Distribution

N112  N 2 22  N3 32  N1 ( x  x1 )2
 N 2 ( x  x2 ) 2  N3 ( x  x3 )2
NOTES  =
N1  N 2  N 3

200 × 9  250 × 16  300 × 25  200 × 81  250 × 36  300 × 1

=
200  250  300
= 51.73  7.19.
4.5.3 Comparison of Various Measures of Dispersion
The range is the easiest to calculate the measure of dispersion, but since it depends
on extreme values, it is extremely sensitive to the size of the sample, and to the
sample variability. In fact, as the sample size increases the range increases
dramatically, because the more the items one considers, the more likely it is that
some item will turn up which is larger than the previous maximum or smaller than
the previous minimum. So, it is, in general, impossible to interpret properly the
significance of a given range unless the sample size is constant. It is for this reason
that there appears to be only one valid application of the range, namely in statistical
quality control where the same sample size is repeatedly used, so that comparison
of ranges are not distorted by differences in sample size.
The quartile deviations and other such positional measures of dispersions
are also easy to calculate but suffer from the disadvantage that they are not amenable
to algebraic treatment. Similarly, the mean deviation is not suitable because we
cannot obtain the mean deviation of a combined series from the deviations of
component series. However, it is easy to interpret and easier to calculate than the
standard deviation.
The standard deviation of a set of data, on the other hand, is one of the most
important statistics describing it. It lends itself to rigorous algebraic treatment, is
rigidly defined and is based on all observations. It is, therefore, quite insensitive to
sample size (provided the size is ‘Large Enough’) and is least affected by
sampling variations.
It is used extensively in testing of hypothesis about population parameters
based on sampling statistics.
In fact, the standard deviations has such stable mathematical properties that
it is used as a standard scale for measuring deviations from the mean. If we are
told that the performance of an individual is 10 points better than the mean, it really
does not tell us enough, for 10 points may or may not be a large enough
difference to be of significance. But if we know that the s for the score is only 4
points, so that on this scale, the performance is 2.5s better than the mean, the
statement becomes meaningful. This indicates an extremely good performance.
This sigma scale is a very commonly used scale for measuring and specifying
deviations which immediately suggest the significance of the deviation.
The only disadvantages of the standard deviation lies in the amount of work
involved in its calculation, and the large weight it attaches to extreme values because
of the process of squaring involved in its calculations.

Self - Learning
240 Material
 Statistical Computation and
4.6 PROBABILITY Probability Distributiona

Probability can be defined as a measure of the likelihood that a particular

event will occur. It is a numerical measure with a value between 0 and 1 of such NOTES
a likelihood where the probability of zero indicates that the given event cannot
occur and the probability of one assures certainty of such an occurrence. For
example, if a radio weather report indicates a near-zero probability of rain,
it can be interpreted as no chance of rain and if a 90 per cent probability of
rain is reported, then our understanding is, that the rain is most likely to occur.
A 50 per cent, probability or chance of rain indicates that rain is just as likely
to occur as not. This likelihood can be shown as follows:
0 0.5 1

Event does Event as likely Event definitely

not occur to occur as not occurs

Increasing likelihood of occurrence

Probability theory provides us with a mechanism for measuring and analysing

uncertainties associated with future events. Probability can be subjective or
objective. Subjective probability is purely individualistic so that an individual can
assign a probability to the outcome of a particular event based upon whatever
information regarding this event is available to him along with his personal
feelings, experience, judgement and expectations. Two different individuals may
assign two different probabilities for the outcome of the same event.
The objective probability of an event, on the other hand, can be defined
as the relative frequency of its occurrence in the long run. In other words, the
probability of an outcome in which we are interested, known as favourable
outcome or successful outcome can be calculated as the number of favourable
outcomes divided by the total number of outcomes. For example, if (s) defines
the number of successful outcomes and (n) is the total number of outcomes,
then the probability of a successful outcome is given as (s/n).
Experiment: An experiment is any activity that generates data. For example,
tossing of a fair coin is considered as a statistical experiment. An experiment is
identified by two properties.
(i) Each experiment has several possible outcomes and all these outcomes
are known in advance.
(ii) None of these outcomes can be predicted with certainty.
For example, while tossing a fair coin, we are not certain whether the
outcome will be a head or a tail. Some of the experiments and their possible
outcomes can be given as:
Experiment Possible Outcomes
1. Tossing of a fair coin Head, tail
2. Rolling a die 1, 2, 3, 4, 5, 6
3. Selecting an item from a production lot Good, bad
4. Introducing a new product Success, failure Self - Learning
Material 241
Statistical Computation and
Probability Distribution
4.6.1 Probability Distribution of a Random Variable
An experiment is said to be random if we cannot predict the outcome before the
experiment is carried out. A random experiment is one which can be repeated,
NOTES practically or theoretically, any number of times. We can toss a coin or roll a die
any number of times and study the outcomes. The result of any outcome may or
may not influence that of succeeding outcomes. Thus any throw of a coin or dice
is independent of all earlier throws. But if a card is drawn from a deck of cards
and not replaced, the experiment made for a second draw will be influenced by
the result of the first.
The consideration of the random character of a phenomenon is inevitable
because of the complicated characteristics of the laws of nature, ignorance about
the relevant laws and the theoretical or practical difficulty allowing for the effect of
a large number of factors under consideration or disturbing factors or other
shortcomings of our tools of work.
A random variable is a probability variable. A specified degree of probability
is attached to every value of a random variable. What is meant by this statement
is that when we conduct an experiment, say by tossing a coin, we cannot definitely
say what will come up. Random forces will decide the outcome, although each
outcome may have an equal chance of probability of showing up. A random variable
is a quantity which in different observations can assume different values.
4.6.2 Axiomatic or Modern Approach to Probability
Probability theory is also called the theory of chance and can be mathematically
derived using the standard formulas. A probability is expressed as a real number,
p  [0, 1] and the probability number is expressed as a percentage (0 per cent
to 100 per cent) and not as a decimal. For example, a probability of 0.55 is
expressed as 55 per cent. When we say that the probability is 100 per cent, it
means that the event is certain while the 0 per cent probability means that the event
is impossible. We can also express probability of an outcome in the ratio format.
For example, we have two probabilities, i.e., ‘Chance of winning’ (1/4) and ‘Chance
of not winning’ (3/4), then using the mathematical formula of odds, we can say,
‘Chance of Winning’ : ‘Chance of not Winning’ = 1/4 : 3/4 = 1 : 3 or 1/3
We are using the probability in vague terms when we predict something for future.
For example, we might say it will probably rain tomorrow or it will probably a
holiday the day after. This is subjective probability to the person predicting, but
implies that the person believes the probability is greater than 50 per cent.
Different types of probability theories are:
(i) Classical Theory of Probability
(ii) Axiomatic Probability Theory
(iii) Empirical Probability Theory
Classical Theory of Probability
The classical theory of probability is the theory based on the number of favourable
Self - Learning outcomes and the number of total outcomes. The probability is expressed as a
242 Material
ratio of these two numbers. The term ‘favorable’ is not the subjective value given Statistical Computation and
Probability Distributiona
to the outcomes, but is rather the classical terminology used to indicate that an
outcome belongs to a given event of interest.
Classical Definition of Probability: If the number of outcomes belonging to an
NOTES
event E is NE , and the total number of outcomes is N, then the probability of event
NE
E is defined as pE 
N
For example, a standard pack of cards (without jokers) has 52 cards. If we
randomly draw a card from the pack, we can imagine about each card as a possible
outcome. Therefore, there are 52 total outcomes. Calculating all the outcome
events and their probabilities, we have the following possibilities:
• Out of the 52 cards, there are 13 clubs. Therefore, if the event of interest is
drawing a club, there are 13 favourable outcomes, and the probability of this
13 1
event becomes: 
52 4
• There are 4 kings (one of each suit). The probability of drawing a king is:
4 1

52 13
• What is the probability of drawing a king or a club? This example is slightly
more complicated. We cannot simply add together the number of outcomes for
each event separately (4 + 13 = 17) as this inadvertently counts one of the
16
outcomes twice (the king of clubs). The correct answer is: from
52
13 4 1
 
52 52 52
We have this from the probability equation, p(club) + p(king) – p(king of clubs).
• Classical probability has limitations, because this definition of probability implicitly
defines all outcomes to be equiprobable and this can be only used for conditions
such as drawing cards, rolling dice, or pulling balls from urns. We cannot calculate
the probability where the outcomes are unequal probabilities.
It is not that the classical theory of probability is not useful because of the above
described limitations. We can use this as an important guiding factor to calculate
the probability of uncertain situations as mentioned above and to calculate the
axiomatic approach to probability.
Frequency of Occurrence
This approach to probability is widely used to a wide range of scientific disciplines.
It is based on the idea that the underlying probability of an event can be measured
by repeated trials.
Probability as a Measure of Frequency: Let nA be the number of times event
A occurs after n trials. We define the probability of event A as,
Self - Learning
Material 243
Statistical Computation and
Probability Distribution nA
PA  Lim
n n

It is not possible to conduct an infinite number of trials. However, it usually suffices

NOTES to conduct a large number of trials, where the standard of large depends on the
probability being measured and how accurate a measurement we need.
nA
Definition of Probability: The sequence in the limit that will converge to the
n
same result every time, or that it will converge at all. To understand this, let us
consider an experiment consisting of flipping a coin an infinite number of times. We
want that the probability of heads must come up. The result may appear as the
following sequence:
HTHHTTHHHHTTTTHHHHHHHHTTTTTTTTHHHHHHHHHHHHH
HHHTTTTTTTTTTTTTTTT...
This shows that each run of k heads and k tails are being followed by another run
nA 1
of the same probability. For this example, the sequence oscillates between,
n 3
2
and which does not converge. These sequences may be unlikely, and can be
3
right. The definition given above does not express convergence in the required
way, but it shows some kind of convergence in probability. The problem of
formulating exactly can be considered using axiomatic probability theory.
Axiomatic Probability Theory
The axiomatic probability theory is the most general approach to probability, and
is used for more difficult problems in probability. We start with a set of axioms,
which serve to define a probability space. These axioms are not immediately intuitive
and are developed using the classical probability theory.
Empirical Probability Theory
The empirical approach to determining probabilities relies on data from actual
experiments to determine approximate probabilities instead of the assumption of
equal likeliness. Probabilities in these experiments are defined as the ratio of the
frequency of the possibility of an event, f(E), to the number of trials in the experiment,
n, written symbolically as P(E) = f(E)/n. For example, while flipping a coin, the
empirical probability of heads is the number of heads divided by the total number
of flips.
The relationship between these empirical probabilities and the theoretical
probabilities is suggested by the Law of Large Numbers. The law states that as
the number of trials of an experiment increases, the empirical probability approaches
the theoretical probability. Hence, if we roll a die a number of times, each number
would come up approximately1/6 of the time. The study of empirical probabilities
is known as statistics.

Self - Learning
244 Material
4.6.3 Theorems on Probability Statistical Computation and
Probability Distributiona
When two events are mutually exclusive, then the probability that either of the
events will occur is the sum of their separate probabilities. For example, if you
roll a single die then the probability that it will come up with a face 5 or face
NOTES
6, where event A refers to face 5 and event B refers to face 6, both events
being mutually exclusive events, is given by,
P[A or B] = P[A] + P[B]
or P[5 or 6] = P[5] + P[6]
= 1/6 +1/6
= 2/6 = 1/3
P [A or B] is written as P[A  B] and is known as P [A union B].
However, if events A and B are not mutually exclusive, then the probability
of occurrence of either event A or event B or both is equal to the probability
that event A occurs plus the probability that event B occurs minus the probability
that events common to both A and B occur.
Symbolically, it can be written as,
P[A  B] = P[A] + P[B] – P[A and B]
P[A and B] can also be written as P[A  B], known as P [A intersection
B] or simply P[AB].
Events [A and B] consist of all those events which are contained in both
A and B simultaneously. For example, in an experiment of taking cards out of
a pack of 52 playing cards, assume that:
Event A = An ace is drawn
Event B = A spade is drawn
Event [AB] = An ace of spade is drawn
Hence, P[A  B]  P[A] + P[B] – P[AB]
= 4/52 + 13/52 – 1/52
= 16/52 = 4/13
This is because there are 4 aces, 13 cards of spades, including 1 ace of
spades out of a total of 52 cards in the pack. The logic behind subtracting
P[AB] is that the ace of spades is counted twice—once in event A (4 aces) and
once again in event B (13 cards of spade including the ace).
Another example for P[A  B], where event A and event B are not
mutually exclusive is as follows:
Suppose a survey of 100 persons revealed that 50 persons read India
Today and 30 persons read Time magazine and 10 of these 100 persons read
both India Today and Time. Then:
Event [A] = 50
Event [B] = 30
Event [AB] = 10
Self - Learning
Material 245
Statistical Computation and Since event [AB] of 10 is included twice, both in event A as well as in
Probability Distribution
event B, event [AB] must be subtracted once in order to determine the event
[A  B] which means that a person reads India Today or Time or both.
Hence,
NOTES
P[A  B] = P [A] + P [B] – P [AB]
= 50/100 + 30/100 –10/100
= 70/100 = 0.7

Multiplication Rule
Multiplication rule is applied when it is necessary to compute the probability if
both events A and B will occur at the same time. The multiplication rule is
different if the two events are independent as against the two events being not
independent.
If events A and B are independent events, then the probability that they
both will occur is the product of their separate probabilities. This is a strict
condition so that events A and B are independent if, and only if,
P [AB] = P[A] × P[B] or
= P[A]P[B]
For example, if we toss a coin twice, then the probability that the first toss
results in a head and the second toss results in a tail is given by,
P [HT] = P[H] × P[T]
= 1/2 × 1/2 = 1/4
However, if events A and B are not independent, meaning that the probability
of occurrence of an event is dependent or conditional upon the occurrence or
non-occurrence of the other event, then the probability that they will both occur
is given by,
P[AB] = P[A] × P[B/given the outcome of A]
This relationship is written as:
P[AB] = P[A] × P[B/A] = P[A] P[B/A]
where P[B/A] means the probability of event B on the condition that event A
has occurred. As an example, assume that a bowl has 6 black balls and 4 white
balls. A ball is drawn at random from the bowl. Then a second ball is drawn
without replacement of the first ball back in the bowl. The probability of the
second ball being black or white would depend upon the result of the first draw
as to whether the first ball was black or white. The probability that both these
balls are black is given by,
P [two black balls] = P [black on 1st draw] × P [black on 2nd draw/black
on 1st draw]
= 6/10 × 5/9 = 30/90 = 1/3
This is so because, first there are 6 black balls out of a total of 10, but
if the first ball drawn is black then we are left with 5 black balls out of a total
of 9 balls.
Self - Learning
246 Material
4.6.4 Counting Techniques Statistical Computation and
Probability Distributiona
Reverend Thomas Bayes (1702–1761) introduced his theorem on probability
which is concerned with a method for estimating the probability of causes which
are responsible for the outcome of an observed effect. Being a religious preacher NOTES
himself as well as a mathematician, his motivation for the theorem came from his
desire to prove the existence of God by looking at the evidence of the world that
God created. He was interested in drawing conclusions about the causes by
observing the consequences. The theorem contributes to the statistical decision
theory in revising prior probabilities of outcomes of events based upon the
observation and analysis of additional information.
Bayes’ theorem makes use of conditional probability formula where the
condition can be described in terms of the additional information which would
result in the revised probability of the outcome of an event.
Suppose that there are 50 students in our statistics class out of which 20 are
male students and 30 are female students. Out of the 30 females, 20 are Indian
students and 10 are foreign students. Out of the 20 male students, 15 are Indians
and 5 are foreigners, so that out of all the 50 students, 35 are Indians and 15 are
foreigners. This data can be presented in a tabular form as follows:
Indian Foreigner Total
Male 15 5 20
Female 20 10 30
Total 35 15 50

Based upon this information, the probability that a student picked up at

random will be female is 30/50 or 0.6, since there are 30 females in the total class
of 50 students. Now, suppose that we are given additional information that the
person picked up at random is Indian, then what is the probability that this person
is a female? This additional information will result in revised probability or posterior
probability in the sense that it is assigned to the outcome of the event after this
additional information is made available.
Since we are interested in the revised probability of picking a female student
at random provided that we know that the student is Indian. Let A1 be the event
female, A2 be the event male and B the event Indian. Then based upon our
knowledge of conditional probability, Bayes’ theorem can be stated as follows,
P( A1 ) P( B / A1 )
P( A1 / B) =
P( A1 ) P( B / A1 ) + P( A2 )( P( B / A2 )

In the example discussed here, there are 2 basic events which are A1 (female)
and A2 (male). However, if there are n basic events, A1, A2, .....An, then Bayes’
theorem can be generalized as,

P ( A1 ) P( B / A1 )
P( A1 / B) 
P( A1 ) P ( B / A1 )  P( A2 )(P (B / A2 )  ...  P ( An ) P ( B / An )
Self - Learning
Material 247
Statistical Computation and Solving the case of 2 events we have,
Probability Distribution
( 30 / 50)( 20 / 30)
P ( A1 / B )   20 / 35  4 / 7  0.57
( 30 / 50)( 20 / 30)  ( 20 / 50)(15 / 20)
NOTES
This example shows that while the prior probability of picking up a female
student is 0.6, the posterior probability becomes 0.57 after the additional
information that the student is an American is incorporated in the problem.
Another example of application of Bayes’ theorem is as follows:
Example 4.34: A businessman wants to construct a hotel in New Delhi. He
generally builds three types of hotels. These are 50 rooms, 100 rooms and 150
rooms hotels, depending upon the demand for the rooms, which is a function of
the area in which the hotel is located, and the traffic flow. The demand can be
categorized as low, medium or high. Depending upon these various demands, the
businessman has made some preliminary assessment of his net profits and possible
losses (in thousands of dollars) for these various types of hotels. These pay-offs
are shown in the following table.
States of Nature
Demand for Rooms
Low (A1) Medium (A2) High (A3)

0.2 0.5 0.3 D e m a n d

Probability
Number of Rooms R1 = (50) 25 35 50
R2 = (100) –10 40 70
R3 = (150) –30 20 100

Solution: The businessman has also assigned ‘prior probabilities’ to the demand
structure or rooms. These probabilities reflect the initial judgement of the
businessman based upon his intuition and his degree of belief regarding the outcomes
of the states of nature.
Demand for rooms Probability of Demand
Low (A1) 0.2
Medium (A2) 0.5
High (A3) 0.3
Based upon these values, the expected pay-offs for various rooms can be
computed as follows,
EV (50) = ( 25 × 0.2) + (35 × 0.5) + (50 × 0.3) = 37.50
EV (100) = (–10 × 0.2) + (40 × 0.5) + (70 × 0.3) = 39.00
EV (150) = (–30 × 0.2) + (20 × 0.5) + (100 × 0.3) = 34.00
This gives us the maximum pay-off of $39,000 for building a 100 rooms
hotel.
Now the hotelier must decide whether to gather additional information
regarding the states of nature, so that these states can be predicted more accurately
Self - Learning
248 Material
than the preliminary assessment. The basis of such a decision would be the cost of
obtaining additional information. If this cost is less than the increase in maximum Statistical Computation and
Probability Distributiona
expected profit, then such additional information is justified.
Suppose that the businessman asks a consultant to study the market and
predict the states of nature more accurately. This study is going to cost the
NOTES
businessman $10,000. This cost would be justified if the maximum expected profit
with the new states of nature is at least $10,000 more than the expected pay-off
with the prior probabilities. The consultant made some studies and came up with
the estimates of low demand (X1), medium demand (X2), and high demand (X3)
with a degree of reliability in these estimates. This degree of reliability is expressed
as conditional probability which is the probability that the consultant’s estimate of
low demand will be correct and the demand will be actually low. Similarly, there
will be a conditional probability of the consultant’s estimate of medium demand,
when the demand is actually low, and so on. These conditional probabilities are
expressed in Table 4.8.
Table 4.8 Conditional Probabilities

X1 X2 X3
States of (A1) 0.5 0.3 0.2
Nature (A2) 0.2 0.6 0.2
(Demand) (A3) 51 0.3 0.6
The values in the preceding table are conditional probabilities and are
interpreted as follows:
The upper north-west value of 0.5 is the probability that the consultant’s
prediction will be for low demand (X1) when the demand is actually low. Similarly,
the probability is 0.3 that the consultant’s estimate will be for medium demand
(X2) when in fact the demand is low, and so on. In other words, P(X1/ A1) = 0.5
and P(X2/ A1) = 0.3. Similarly, P(X1 / A2) = 0.2 and P(X2 / A2) = 0.6, and so on.
Our objective is to obtain posteriors which are computed by taking the
additional information into consideration. One way to reach this objective is to
first compute the joint probability which is the product of prior probability and
conditional probability for each state of nature. Joint probabilities as computed is
given as,
State Prior Joint Probabilities

of Nature Probability P(A1X1) P(A1X2) P(A1X3)

A1 0.2 0.2 × 0.5 = 0.1 0.2 × 0.3 = 0.06 0.2 × 0.2 = 0.04
A2 0.5 0.5 × 0.2 = 0.1 0.5 × 0.6 = 0.3 0.5 × 0.2 = 0.1
A3 0.3 0.3 × 0.1 = 0.03 0.3 × 0.3 = 0.09 0.3 × 0.6 = 0.18
Total marginal probabilities = 0.23 = 0.45 = 0.32

Now, the posterior probabilities for each state of nature Ai are calculated as
follows:
Joint probability of Ai and Xj
P (Ai / Xj ) 
Marginal probability of Xj
Self - Learning
Material 249
Statistical Computation and By using this formula, the joint probabilities are converted into posterior
Probability Distribution
probabilities and the computed table for these posterior probabilities is given as,
States of Nature Posterior Probabilities
NOTES P(A1/X1) P(A1/X2) P(A1/X3)

A1 0.1/.023 = 0.435 0.06/0.45 = 0.133 0.04/0.32 = 0.125

A2 0.1/.023 = 0.435 0.30/0.45 = 0.667 0.1/0.32 = 0.312
A3 0.03/.023 = 0.130 0.09/0.45 = 0.200 0.18/0.32 = 0.563

Total = 1.0 = 1.0 = 1.0

Now, we have to compute the expected pay-offs for each course of action
with the new posterior probabilities assigned to each state of nature. The net profits
for each course of action for a given state of nature is the same as before and is
restated as follows. These net profits are expressed in thousands of dollars.
Low (A1) Medium (A2) High (A3)
Number of Rooms (R1) 25 35 50
(R2) –10 40 70
(R3) –30 20 100
Let Oij be the monetary outcome of course of action (i) when (j) is the
corresponding state of nature, so that in the above case Oi1 will be the outcome of
course of action R1 and state of nature A1, which in our case is $25,000. Similarly,
Oi2 will be the outcome of action R2 and state of nature A2, which in our case is –
$10,000, and so on. The expected value EV (in thousands of dollars) is calculated
on the basis of actual state of nature that prevails as well as the estimate of the
state of nature as provided by the consultant. These expected values are calculated
as follows,
Course of action = Ri
Estimate of consultant = Xi
Actual state of nature = Ai
where, i = 1, 2, 3
Then
(A) Course of action = R1 = Build 50 rooms hotel

R  A 
EV  1  = P  i  Oi1
 X1   X1 
= 0.435(25) + 0.435 (–10) + 0.130 (–30)
= 10.875 – 4.35 – 3.9 = 2.625

R   A 
EV  1  = P  i  Oi1
 X2   X2 
= 0.133(25) + 0.667 (–10) + 0.200 (–30)
Self - Learning
250 Material
 = 3.325 – 6.67 – 6.0 = –9.345 Statistical Computation and
Probability Distributiona

R   Ai 
EV  1  = P   Oi1
 X3   X3 
NOTES
= 0.125(25) + 0.312(–10) + 0.563(–30)
= 3.125 – 3.12 – 16.89
= –16.885
(B) Course of action = R2 = Build 100 rooms hotel

R   A 
EV  2  = P  i  Oi 2
 X1   X1 
= 0.435(35) + 0.435 (40) + 0.130 (20)
= 15.225 + 17.4 + 2.6 = 35.225

R   A 
EV  2  = P  i  Oi 2
 X2   X1 
= 0.133(35) + 0.667 (40) + 0.200 (20)
= 4.655 + 26.68 + 4.0 = 35.335

R   A 
EV  2  = P  i  Oi 2
 X3   X3 
= 0.125(35) + 0.312(40) + 0.563(20)
= 4.375 + 12.48 + 11.26 = 28.115
(C) Course of action = R3 = Build 150 rooms hotel

R   A 
EV  3  = P  i  Oi 3
 X1   X1 
= 0.435(50) + 0.435(70) + 0.130 (100)
= 21.75 + 30.45 + 13 = 65.2

R   A 
EV  3  = P  i  Oi 3
 X2   X2 

= 0.133(50) + 0.667 (70) + 0.200 (100)

= 6.65 + 46.69 + 20 = 73.34

R   A 
EV  3  = P  i  Oi 3
 X3   X3 
= 0.125(50) + 0.312(70) + 0.563(100)
= 6.25 + 21.84 + 56.3 = 84.39 Self - Learning
Material 251
Statistical Computation and The calculated expected values in thousands of dollars, are presented in a
Probability Distribution
tabular form.
Expected Posterior Pay-offs
NOTES
Outcome EV (R1/Xi) EV (R2/Xi) EV (R3/Xi)

X1 2.625 35.225 65.2

X2 –9.345 35.335 73.34
X3 –16.885 28.115 84.39
This table can now be analysed in the following manner.
If the outcome is X1, it is desirable to build 150 rooms hotel, since the
expected pay-off for this course of action is maximum of $65,200. Similarly, if the
outcome is X2, the course of action should again be R3 since the maximum pay-off
is $73,34. Finally, if the outcome is X3, the maximum pay-off is $84,390 for course
of action R3.
Accordingly, given these conditions and the pay-off, it would be advisable
to build a hotel which has 150 rooms.
4.6.5 Mean and Variance of Random Variables
Mean
A discrete random variable’s mean xi represents a weighted average of the random
variable’s possible values. The mean of a random variable, unlike the sample mean
of a series of observations, which gives each observation equal weight, weights
each result pi according to its likelihood, . The mean (also known as the expected
value of X) is denoted by the symbol W, which is technically defined by

The mean of a random variable is the variable’s long-run average, or the

expected average outcome over a large number of observations.
The mean of a continuous random variable is determined by the distribution’s
density curve. The mean of a symmetric density curve, such as the normal density,
is in the curve’s centre.
The rule of large numbers asserts that as the number of observations of a
random variable grows larger, the observed random mean will always approach
the distribution mean .
That is, as the number of observations grows larger, the mean of these data
approaches the true mean of the random variable. This isn’t to say that short-term
averages will always reflect the mean.
The total of the means of two random variables, X and Y, is the mean:

Self - Learning
252 Material
Variance Statistical Computation and
Probability Distributiona
The variance of a discrete random variable X is defined by and measures the spread,
or variability, of the distribution:
NOTES

The square root of the variance is the standard deviation .

Properties of Variance
The variance is modified as follows when a random variable X is adjusted by
multiplying by the value b and adding the value a:

The value a is ignored since adding or removing a constant has no effect on

the spread of the distribution. Because the variance is a sum of squared terms, any
multiplier value b used to alter the variance must also be squared.
The variance of their total or difference is the sum of their variances for
independent random variables X and Y:

Because the variation in each variable adds to the variation in each case,
variances are added for both the sum and difference of two independent random
variables. Variability in one variable is connected to variability in the other if the
variables are not independent. As a result, applying the preceding formula to
compute the variance of their total or difference may not be possible.
Assume that variable X represents the amount of money (in dollars) spent
on lunch by a group of people, and variable Y represents the amount of money
spent on supper by the same group of people. Because X and Y are not considered
independent variables, the variance of the sum X + Y cannot be calculated as the
sum of the variances.

4.7 STANDARD PROBABILITY DISTRIBUTION

Once the random variable of interest is defined and the probabilities are assigned
to all its values, it is called a probability distribution. Table 4.9 showns the
probability distribution for various sales levels (sales level being the random
variable represented as X) for a new product as stated by the sales manager:

Self - Learning
Material 253
Statistical Computation and Table 4.9 Probability Distribution for Various Sales Levels
Probability Distribution
Sales (in units) Probability
Xi pr. (Xi)

NOTES X1 50 0.10
X2 100 0.30
X3 150 0.30
X4 200 0.15
X5 250 0.10
X6 300 0.05
Total 1.00

Sometimes, the probability distribution may be presented in the form called a

cumulative probability distribution. The probability distribution given in Table 4.9
can also be presented in the form of cumulative probability distribution as in
Table 4.10.
Table 4.10 Cumulative Probability Distribution

Sales (in units) Probability Cumulative Probabilities

(Xi)  pr (Xi) pr (Xi)
X1 50 0.10 0.10
X2 100 0.30 0.40
X3 150 0.30 0.70
X4 200 0.15 0.85
X5 250 0.10 0.95
X6 300 0.05 1.00
The meaning of probability distribution can be made more clear if you remember
the following:
 An observed frequency distribution (often called simply as a frequency
distribution) is a listing of the observed frequencies of all the outcomes of an
experiment that actually occurred while performing the experiment.
 A probability distribution is a listing of the probabilities of all the possible
outcomes that could result if the experiment is performed. The assignment
of probabilities may be based either on theoretical considerations or it may
be a subjective assessment or may be based on experience.
 A theoretical frequency distribution is a probability distribution that
describes how outcomes are expected to vary. In other words, it enlists the
expected values (i.e., observed values multiplied by corresponding
probabilities) of all the outcomes.
Types of Probability Distributions
Probability distributions can be classified as either discrete or continuous. In a
discrete probability distribution, the variable under consideration is allowed to
take only a limited number of discrete values along with corresponding
probabilities. The two important discrete probability distributions are: the binomial
Self - Learning probability distribution and the Poisson probability distribution. In a continuous
254 Material probability distribution, the variable under consideration is allowed to take on
any value within a given range. Important continuous probability distributions are Statistical Computation and
Probability Distributiona
exponential probability distribution and normal probability distribution.
Important disrete and continuous probability distributions are discussed later in
this unit.
NOTES
Probability Functions
In probability distribution, it is not always necessary to calculate probabilities for
each and every outcome in the sample space. There exist many mathematical
formulae for many commonly encountered problems which can assign probabilities
to the values of random variables. Such formulae are generally termed as
probability functions. In fact, a probability function is a mathematical way of
describing a given probability distribution. To select a suitable probability function
that best fits in the given situation, you should work out the values of its
parameters. Once you have worked out the values of the parameters, you can
then assign the probabilities, if required, using the appropriate probability function
to the values of random variables. Various probability functions will be explained
shortly while describing the various probability distributions.

Techniques of Assigning Probabilities

You can assign probability values to the random variables. Since the assignment
of probabilities is not an easy task, you should observe certain rules in this
context as follows:
(i) A probability cannot be less than zero or greater than one, i.e. 0,  pr  1,
where pr represents probability.
(ii) The sum of all the probabilities assigned to each value of the random
variable must be exactly one.
There are three techniques of assignment of probabilities to the values of the
random variable:
(i) Subjective probability assignment: It is the technique of assigning
probabilities on the basis of personal judgement. Such assignment may
differ from individual to individual and depends upon the expertise of the
person assigning the probabilities. It cannot be termed as a rational way
of assigning probabilities but is used when the objective methods cannot
be used for one reason or the other.
(ii) A priori probability assignment: It is the technique under which the
probability is assigned by calculating the ratio of the number of ways in
which a given outcome can occur to the total number of possible outcomes.
The basic underlying assumption in using this procedure is that every
possible outcome is likely to occur equally. But at times the use of this
technique gives ridiculous conclusions. For example, you have to assign
probability to the event that a person of age 35 will live up to age 36.
There are two possible outcomes, he lives or he dies. If the probability
assigned in accordance with a priori probability assignment is half, then
the same may not represent reality. In such a situation, probability can be
Self - Learning
assigned by some other techniques. Material 255
Statistical Computation and (iii) Empirical probability assignment: It is an objective method of assigning
Probability Distribution
probabilities and is used by the decision makers. Using this technique, the
probability is assigned by calculating the relative frequency of occurrence
of a given event over an infinite number of occurrences. However, in
NOTES
practice, only a finite (perhaps very large) number of cases are observed
and relative frequency of the event is calculated. The probability assignment
through this technique may as well be unrealistic, if future conditions do not
happen to be a reflection of the past.
Thus, what constitutes the ‘best’ method of probability assignment can
only be judged in the light of what seems best to depict reality. It depends
upon the nature of the problem and also on the circumstances under which
the problem is being studied.

4.7.1 Binomial Distribution

Binomial distribution (or the binomial probability distribution) is a widely used
probability distribution concerned with a discrete random variable and as such
is an example of a discrete probability distribution. The binomial distribution
describes discrete data resulting from what is often called as the Bernoulli
process. The tossing of a fair coin a fixed number of times is a Bernoulli process
and the outcome of such tosses can be represented by the binomial distribution.
The name of Swiss mathematician Jacob Bernoulli is associated with this
distribution. This distribution applies in situations where there are repeated trials
of any experiment for which only one of the two mutually exclusive outcomes
(often denoted as ‘Success’ and ‘Failure’) can result on each trial.

The Bernoulli process

Binomial distribution is considered appropriate in a Bernoulli process which has
the following characteristics:
(a) Dichotomy: This means that each trial has only two mutually exclusive
possible outcomes, e.g., ‘Success’ or ‘Failure’, ‘Yes’ or ‘No’, ‘Heads’ or
‘Tails’ and the like.
(b) Stability: This means that the probability of the outcome of any trial is
known (or given) and remains fixed over time, i.e., remains the same for
all the trials.
(c) Independence: This means that the trials are statistically independent, i.e.
to say the happening of an outcome or the event in any particular trial is
independent of its happening in any other trial or trials.

Probability Function of Binomial Distribution

The random variable, say X, in the binomial distribution is the number of
‘Successes’ in n trials. The probability function of the binomial distribution is
written as follows:
f (X = r) = nCr prqn–r
r = 0, 1, 2, …, n
Self - Learning
256 Material where n = Numbers of trials
 p = Probability of success in a single trial Statistical Computation and
Probability Distributiona
q = (1 – p) = Probability of ‘Failure’ in a single trial
r = Number of successes in ‘n’ trials
NOTES
Parameters of Binomial Distribution
Binomial distribution depends upon the values of p and n which in fact are its
parameters. Knowledge of p truly defines the probability of X since n is known
by the definition of the problem. The probability of the happening of exactly r
events in n trials can be found out using the previously stated binomial function.
The value of p also determines the general appearance of the binomial distribution,
if shown graphically. In this context, the usual generalizations are as follows:
(i) When p is small (say 0.1), the binomial distribution is skewed to the right,
i.e., the graph takes the form as shown in Figure 4.4.
Probability

0 1 2 3 4 5 6 7 8
No. of Successes
Fig. 4.4 Bonimial Distribution Skewed to the Right
(ii) When p is equal to 0.5, the binomial distribution is symmetrical and the
graph takes the form as shown in Figure 4.5.
Probability

0 1 2 3 4 5 6 7 8
No. of Successes
Fig. 4.5 Symmetrical Binomial Distribution
(iii) When p is larger than 0.5, the binomial distribution is skewed to the left
and the graph takes the form as shown in Figure 4.6.
Probability

0 1 2 3 4 5 6 7 8
No. of Successes
Self - Learning
Fig. 4.6 Bonimial Distribution Skewed to the Left Material 257
Statistical Computation and But if ‘p’ stays constant and ‘n’ increases, then as ‘n’ increases, the vertical
Probability Distribution
lines become not only numerous but also tend to bunch up together to form a
bell shape, i.e. the binomial distribution tends to become symmetrical and the
graph takes the shape as shown in Figure 4.7.
NOTES

Probability

0, 1, 2, ..........
No. of Successes
Fig. 4.7 The Bell-Shaped Binomial Distribution

Important measures of binomial distribution

The expected
_ value of random variable [i.e. E(X)] or mean of random variable
(i.e. X ) of the binomial distribution is equal to np and the variance of random
variable is equal to npq or np (1 – p). Accordingly, the standard deviation of
binomial distribution is equal to npq. The other important measures relating to
binomial distribution are as under:
1 2 p
Skewness =
npq

1  6 p  6q 2
Kurtosis = 3 
npq

When to use binomial distribution

The use of binomial distribution is most appropriate in situations fulfilling the
previously stated conditions. Two such situations, for example, can be described
as follows:
(i) When you have to find the probability of 6 heads in 10 throws of a fair
coin.
(ii) When you have to find the probability that 3 out of 10 items produced by
a machine, which produces 8 per cent defective items on an average, will
be defective.
Example 4.35: A fair coin is thrown 10 times. The random variable X is the
number of head(s) coming upwards. Using the binomial probability function,
find the probabilities of all possible values
_ which X can take and then verify that
binomial distribution has a mean: X = np and variance: 2 = npq.
Solution: Since the coin is fair and so, when thrown, it can come either with
1 1
head upwards or tail upwards. Hence, p (head)  and q (no head)  . The
2 2
Self - Learning required probability function is:
258 Material
 f(X = r) = nCr prqn–r Statistical Computation and
Probability Distributiona
r = 0, 1, 2, ..., 10
The following table of binomial probability distribution is constructed using this
function. NOTES
_ _ _
Xi (Number Probability pri Xi pri (Xi – X) (Xi – X)2 (Xi – X )2.pi
of Heads)
0 10C q10 0 1/10246
0p = 0/1024 –5 25 25/1024
1 10C p1 q9 = 10/1024 10/1024 –4 16 160/1024
1
2 10C p2 q8 = 45/1024 90/1024 –3 9 405/1024
2
3 10C p3 q7 = 120/1024 360/1024 –2 4 480/1024
3
4 10C p4 q6 = 210/1024 840/1024 –1 1 210/1024
4
5 10C p5 q5 = 252/1024 1260/1024 0 0 0/1024
5
6 10C p6 q4 = 210/1024 1260/1024 1 1 210/1024
6
7 10C p7 q3 = 120/1024 840/1024 2 4 480/1024
7
8 10C p8 q2 = 45/1024 360/1024 3 9 405/1024
8
9 10C p9 q1 = 10/1024 90/1024 4 16 160/1024
9
10 10C p10 q0 = 1/1024 10/1024 5 25 25/1024
10
_
_X = 5120/1024 _ = 2 =
Variance
X = 5  (Xi– X )2pri =
2560/1024 = 2.5

The mean of the binomial distribution is given by np = 10 × 1 = 5 and the

2
1 1
variance of this distribution is equal to npq = 10 × × = 2.5
2 2
These values are exactly the same as we have found them in the preceding
table.
Hence, these values stand verified with the calculated values of the two measures
as shown in the table.

Fitting a binomial distribution

When a binomial distribution is to be fitted to the given data, the following
procedure is adopted:
_
(i) Determine the values of ‘p’ and ‘q’ keeping in view that X = np and
q = (1  p).
(ii) Find the probabilities for all possible values of the given random variable
applying the binomial probability function, namely
f(Xi = r) = nCrprqn–r
r = 0, 1, 2, …, n
(iii) Work out the expected frequencies for all values of random variable by
multiplying N (the total frequency) with the corresponding probability as
worked out in case (ii).
The expected frequencies so calculated constitute the fitted binomial
distribution to the given data.
4.7.2 Poisson Distribution
Poisson distribution is also a discrete probability distribution with which is Self - Learning
associated the name of a Frenchman, Simeon Denis Poisson, who developed Material 259
Statistical Computation and this distribution. It is frequently used in the context of operations research, and,
Probability Distribution
for this reason, has a great significance for management people. It plays an
important role in queuing theory, inventory control problems and risk models.
Unlike binomial distribution, Poisson distribution cannot be deducted on purely
NOTES
theoretical grounds based on the conditions of the experiment. In fact, it must
be based on experience, i.e. on the empirical results of past experiments relating
to the problem under study. Poisson distribution is appropriate, especially when
probability of happening of an event is very small [so that q or (1– p) is almost
equal to unity] and n is very large such that the average of series (namely np)
is a finite number. Experience has shown that this distribution is good for calculating
the probabilities associated with X occurrences in a given time period or specified
area.
The random variable of interest in Poisson distribution is the number of
occurrences of a given event during a given interval (interval may be time,
distance, area, etc.). You use capital X to represent the discrete random variable
and lower case x to represent a specific value that capital X can take. The
probability function of this distribution is generally written as under:

 i e 
f ( X i  x) 
x
x = 0, 1, 2,…
where  = Average number of occurrences per specified interval. In other
words, it is the mean of the distribution.
e = 2.7183 being the basis of natural logarithms.
x = Number of occurrences of a given event.

The Poisson process

The poisson distribution applies in the case of the Poisson process which has
the following characteristics:
 Concerning a given random variable, the mean relating to a given interval
can be estimated on the basis of past data concerning the variable under
study.
 If you divide the given interval into very very small intervals you will find the
following:
(a) The probability that exactly one event will happen during the very
very small interval is a very small number and is constant for every
other very small interval.
(b) The probability that two or more events will happen with in a very
small interval is so small that you can assign it a zero value.
(c) The event that happens in a given very small interval is independent,
when the very small interval falls during a given interval.
(d) The number of events in any small interval is not dependent on the
number of events in any other small interval.
Self - Learning
260 Material
Parameter and important measures of poisson distribution Statistical Computation and
Probability Distributiona
Poisson distribution depends upon the value of , the average number of
occurrences per specified interval which is its only parameter. The probability
of exactly x occurrences can be found out using Poisson probability function
stated above. The expected value or the mean of Poisson random variable is NOTES
 and its variance is also . The standard deviation of Poisson distribution
is, λ .
Underlying the Poisson model is the assumption that if there are on the average
 occurrences per interval t, then there are on the average k occurrences
per interval kt. For example, if the number of arrivals at a service counted in
a given hour has a Poisson distribution with  = 4, then y, the number of arrivals
at a service counter in a given 6 hour day, has the Poisson distribution =24,
i.e., 6×4.

When to use Poisson distribution

The use of Poisson distribution is resorted to in cases when you do not know
the value of ‘n’ or when ‘n’ cannot be estimated with any degree of accuracy.
In fact, in certain cases it does not make any sense in asking the value of ‘n’.
For example, if the goals scored by one team in a football match are given, it
cannot be stated how many goals could not be scored. Similarly, if you watch
carefully, you may find out how many times the lightning flashed but it is not
possible to state how many times it did not flash. It is in such cases you use
Poisson distribution. The number of deaths per day in a district in one year due
to a disease, the number of scooters passing through a road per minute during
a certain part of the day for a few months, the number of printing mistakes per
page in a book containing many pages, etc. are a few other examples where
Poisson probability distribution is generally used.
Example 4.36: Suppose that a manufactured product has 2 defects per unit of
product inspected. Use Poisson distribution and calculate the probabilities of
finding a product without any defect, with 3 defects and with 4 defects.
Solution: If the product has 2 defects per unit of product inspected. Hence,
 = 2.
Poisson probability function is as follows:
 x e 
f  Xi  x  
x
x = 0, 1, 2,…
Using this probability function, you will find the required probabilities as follows:
20 e2
P(without any defects, i.e., x = 0) =
0
1. 0.13534
=  0.13534
1

23 e 2 2  2  2  0.13534
P(with 3 defects, i.e., x = 3) = =
3 3  2 1 Self - Learning
Material 261
Statistical Computation and
0.54136
Probability Distribution =  0.18045
3

24 e 2 2  2  2  2  0.13534
P(with 4 defects, i.e., x = 4) = =
NOTES 4 4  3  2 1
0.27068
=  0.09023
3

Fitting a Poisson distribution

When a Poisson distribution is to be fitted to the given data, then the following
procedure is adopted:
(i) Determine the value of , the mean of the distribution.
(ii) Find the probabilities for all possible values of the given random variable
using the Poisson probability function, namely
 x .e 
f  X i  x 
x
x = 0, 1, 2,…
(iii) Work out the expected frequencies as follows:
np(Xi = x)
The result of case (iii) is the fitted Poisson distribution to the given data.

Poisson distribution as an approximation of binomial distribution

Under certain circumstances Poisson distribution can be considered as a
reasonable approximation of binomial distribution and can be used accordingly.
The circumstances which permit all this are when ‘n’ is large, approaching
infinity, and p is small, approaching zero (n = number of trials, p = probability
of ‘success’). Statisticians usually take the meaning of large n, for this purpose,
when n  20 and by small ‘p’ they mean when p 0.05. In the cases where
these two conditions are fulfilled, you can use mean of the binomial distribution
(namely np) in place of the mean of Poisson distribution (namely) so that the
probability function of Poisson distribution becomes as follows:

f  Xi  x 
 np  x e  np 
x
You can explain Poisson distribution as an approximation of the binomial
distribution with the help of following example.
Example 4.37: Given are the following information:
(a) There are 20 machines in a certain factory, i.e. n = 20.
(b) The probability of machine going out of order during any day is 0.02.
What is the probability that exactly three machines will be out of order on the
same day? Calculate the required probability using both binomial and Poissons
distributions and state whether Poisson distribution is a good approximation of
the binomial distribution in this case.
Self - Learning
262 Material
Solution: Probability as per Poisson probability function (using np in place of ) Statistical Computation and
Probability Distributiona
(since n  20 and p  0.05)

f  Xi  x 
 np  x e  np
NOTES
x
Where x means number of machines becoming out of order on the same day.

20  0.02  e 
3  200.02 
P(Xi = 3) =

3

=
 0.4 . 0.67032 (0.064)(0.67032)
3

3 2 1 6
= 0.00715
Probability as per binomial probability function,
f(Xi = r) = nCr prqn–r
where n = 20, r = 3, p = 0.02 and hence q = 0.98
 f(Xi = 3) = 20C
3(0.02)
3 (0.98)17
= 0.00650
The difference between the probability of three machines becoming out of order
on the same day calculated using probability function and binomial probability
function is just 0.00065. The difference being very very small, you can state that
in the given case Poisson distribution appears to be a good approximation of
binomial distribution.
Example 4.38: How would you use a Poisson distribution to find approximately
the probability of exactly 5 successes in 100 trials the probability of success in
each trial being p = 0.1?
Solution: Given:
n = 100 and p = 0.1
  = n.p = 100 × 0.1 = 10
To find the required probability, the Poisson probability function can be used as
an approximation to the binomial probability function as follows:

f  Xi  x 
 x e

 np  x e  np 
x x

105 e 10 100000  0.00005 5.00000

or P(5)7 =  
5 5  4  3  2 1 5  4  3  2 1
1
= = 0.042
24

4.7.3 Exponential Distribution

Exponential probability distribution is the probability distribution of time (say t),
between events and as such it is continuous probability distribution concerned Self - Learning
with the continuous random variable that takes on any value between zero and Material 263
Statistical Computation and positive infinity. In the exponential distribution, you often ask the question: What
Probability Distribution
is the probability that it will take x trials before the first occurrence? This
distribution plays an important role in describing a large class of phenomena,
particularly in the area of reliability theory and in queuing models.
NOTES The probability function of the exponential distribution is as follows:
F(x) = e–x x  0
where µ = The average length of the interval between two occurrences10
e = 2.7183 being the basis of natural logarithms
The only parameter of the exponential distribution is µ.
The expected value or mean of the exponential distribution is 1/µ and its variance
is 1/µ.11
The cumulative distribution (less than type) of the exponential is
F(x) = P(X  x) = 1 – e–x, x  0
= 0, elsewhere
Thus, for instance, the probability that x  2 is 1 – e–2µ.
Example 4.39: In an industrial complex, the average number of fatal accidents
per month is one-half. The number of accidents per month is adequately described
by a Poisson distribution. What is the probability that four months will pass
without a fatal accident?
Solution: You have been given that the average number of fatal accidents per
month is one-half and the number of accidents per month is well described by
a Poisson distribution.
Hence,   0.5
1 1
The average length of the time interval between two accidents   2
 0.5
months, assuming exponential distribution.
Now, by using the cumulative distribution of the exponential, we can find
the required probability that four months will pass without a fatal accident
(i.e. x > 4) as follows:
 F(x) = P(X  x) = 1 – e–x
 P(X  x) = e–x
 P(X  x) = e–(4) = e–8 = 0.00034
Thus, 0.00034 is the required probability that 4 months will pass without a fatal
accident.

4.7.4 Normal Distribution

Among all the probability distributions, the normal probability distribution is by far
the most important and frequently used continuous probability distribution. This is
so because this distribution well fits in many types of problems. This distribution is
of special significance in inferential statistics since it describes probabilistically the
link between a statistic and a parameter (i.e., between the sample results and the
population from which the sample is drawn). The name of Karl Gauss, the
Self - Learning
264 Material eighteenth century mathematician-astronomer, is associated with this distribution
and in honour of his contribution, this distribution is often known as the Gaussian Statistical Computation and
Probability Distributiona
distribution.
The normal distribution can be theoretically derived as the limiting form of many
discrete distributions. For instance, if in the binomial expansion of (p + q)n, the
NOTES
1
value of ‘n’is infinity and p = q = , then a perfectly smooth symmetrical curve
2
would be obtained. Even if the values of p and q are not equal but if the value of
the exponent ‘n’ happens to be very very large, you get a curve normal probability
smooth and symmetrical. Such curves are called normal probability curves (or at
times known as normal curves of error) and represent the normal distributions.
The probability function in the case of normal probability distribution13 is given as
follows:
1  x  2
1   
f  x  e 2  
 2
where  = The mean of the distribution
 2 = Variance of the distribution
The normal distribution is thus defined by two parameters, namely  and 2. This
distribution can be represented graphically (Refer Figure 4.8).

Fig. 4.8 Curve Representing Normal Distribution

Characteristics of Normal Distribution

The characteristics of the normal distribution or those of a normal curve are as
follows:
(i) It is symmetric distribution.
(ii) The mean  defines where the peak of the curve occurs. In other words,
the ordinate at the mean is the highest ordinate. The height of the ordinate at
a distance of one standard deviation from mean is 60.653 per cent of the
height of the mean ordinate, and similarly, the height of other ordinates at
various standard deviations (s) from mean happens to be a fixed relationship
with the height of the mean ordinate.
(iii) The curve is asymptotic to the baseline which means that it continues to
approach but never touches the horizontal axis.
(iv) The variance (2) defines the spread of the curve.
(v) Area enclosed between mean ordinate and an ordinate at a distance of one
Self - Learning
standard deviation from the mean is always 34.134 per cent of the total Material 265
Statistical Computation and area of the curve. It means that the area enclosed between two ordinates at
Probability Distribution
one Sigma Distance (SD) from the mean on either side would always be
68.268 per cent of the total area. This is shown in Figure 4.9.

NOTES (34.134% + 34.134%) = 68.268%

Area of the total
curve between  ± 1()

X or X
–3 –2 –   +2 +3

Fig. 4.9 An Area Enclosed between Two Ordinates at One SD

Similarly, the other area relationships are as given in Table 4.11.

Table 4.11 Area Relationships
Between Area Covered to Total Area of the
Normal Curve 15
µ±1 SD 68.27%
µ±2 SD 95.45%
µ±3 SD 99.73%
µ ± 1.96 SD 95%
µ ± 2.578 SD 99%
µ ± 0.6745 SD 50%

(vi) The normal distribution has only one mode since the curve has a single
peak. In other words, it is always a unimodal distribution.
(vii) The maximum ordinate divides the graph of normal curve into two equal
parts.
(viii) In addition to all the above stated characteristics the curve has the following
properties:
_
(a) µ = x
(b) µ2=2 = Variance
(c) µ4=34
(d) Moment coefficient of Kurtosis = 3
Family of normal distributions or curves
You can have several normal probability distributions but each particular normal
distribution is being defined by its two parameters, namely the mean (µ) and the
standard deviation (). There is, thus, not a single normal curve but rather a family
of normal curves. Figures 4.10–4.12 exhibit some of these normal curves:
Curve having small standard
deviation say ( = 1)
Curve having large standard
deviation say ( = 5)
Curve having very large standard
deviation say ( = 10)
 in a normal
distribution

Self - Learning Fig. 4.10 Normal Curves with Identical Means

266 Material but Different Standard Deviations
 Statistical Computation and
Probability Distributiona

NOTES

Fig. 4.11 Normal Curves with Identical Standard Deviation

but Each with Different Means

Fig. 4.12 Normal Curves Each with Different Standard

Deviations and Different Means

How to measure the area under the normal curve?

You have learned about the the area relationships involving certain intervals of
standard deviations (plus and minus) from the means that are true in case of a
normal curve. But what should be done in all other cases? You can make use of
the statistical tables constructed by mathematicians for the purpose. Using these
tables, you can find the area (or probability, taking the entire area of the curve as
equal to 1) that the normally distributed random variable will lie within certain
distances from the mean. These distances are defined in terms of standard deviations.
While using the tables showing the area under the normal curve, you talk in terms
of standard variate (symbolically Z ) which really means standard deviations without
units of measurement and this ‘Z’ is worked out as under:
X 
Z

where Z = The standard variate (or number of standard deviations from X to the
mean of the distribution)
X = Value of the random variable under consideration
µ = Mean of the distribution of the random variable
 = Standard deviation of the distribution
The table showing the area under the normal curve (often termed as the standard
normal probability distribution table) is organized in terms of standard variate
(or Z) values. It gives the values for only half the area under the normal curve,
beginning with Z = 0 at the mean. Since the normal distribution is perfectly
symmetrical, the values true for one half of the curve are also true for the other
half.
Self - Learning
Material 267
Statistical Computation and Example 4.40: A banker claims that the life of a regular saving account opened
Probability Distribution
with his bank averages 18 months with a standard deviation of 6.45 months. Answer
the following questions:
(a) What is the probability that there will still be money in 22 months in a savings
NOTES
account opened with the said bank by a depositor?
(b) What is the probability that the account will have been closed before two
years?
Solution: (a) For finding the required probability, you are interested in the area of
the portion of the normal curve as shaded and shown in the following diagram:

σ = 6.45

μ = 18
z=0 X = 22

Calculate Z as under:
X  22  18
Z=   0.62
 6.45

The value from the table showing the area under the normal curve for Z = 0.62 is
0.2324. This means that the area of the curve between µ = 18 and X = 22 is
0.2324. Hence, the area of the shaded portion of the curve is (0.5) – (0.2324) =
0.2676 since the area of the entire right hand portion of the curve always happens
to be 0.5. Thus, the probability that there will still be money in 22 months in a
savings account is 0.2676.
(b) For finding the required probability, you are interested in the area of the portion
of the normal curve as shaded and shown in the following figure:

σ = 6.45

μ = 18 X = 24
z=0

Calculate Z as under:
24  18
Z  0.93
6.45

The value from the concerning table, when Z = 0.93, is 0.3238 which refers to the
area of the curve between µ = 18 and X = 24. The area of the entire left hand
Self - Learning portion of the curve is 0.5 as usual.
268 Material
Hence, the area of the shaded portion is (0.5) + (0.3238) = 0.8238 which is the Statistical Computation and
Probability Distributiona
required probability that the account will have been closed before two years, i.e.
before 24 months.
Example 4.41: Regarding a certain normal distribution concerning the income of
NOTES
the individuals, you are given that mean = 500 rupees and standard deviation
=100 rupees. Find the probability that an individual selected at random will belong
to income group:
(a) Rs 550 to Rs 650 (b) Rs 420 to 570
Solution: (a) For finding the required probability, you are interested in the area of
the portion of the normal curve as shaded and shown in the following figure:

 = 100

 =500
= 500 X = 650
z = 0 X = 550
For finding the area of the curve between X = 550 to 650, do the following
calculations:
550  500 50
Z   0.50
100 100
Corresponding to which the area between µ = 500 and X = 550 in the curve as
per table is equal to 0.1915 and
650  500 150
Z   1.5
100 100
Corresponding to which the area between µ = 500 and X = 650 in the curve as
per table is equal to 0.4332
Hence, the area of the curve that lies between X = 550 and X = 650 is
(0.4332) – (0.1915) = 0.2417
This is the required probability that an individual selected at random will belong to
income group of Rs 550 to Rs 650.
(b) For finding the required probability, you are interested in the area of the portion
of the normal curve as shaded and shown in the following figure:
To find the area of the shaded portion we make the following calculations:

 = 100

 = 100
X = 420z = 0X = 570
z=0 Self - Learning
Material 269
Statistical Computation and
Probability Distribution 570  500
Z  0.70
100
Corresponding to which the area between µ = 500 and X = 570 in the curve as
NOTES per table is equal to 0.2580.
420  500
and Z  0.80
100
Corresponding to which the area between µ = 500 and X = 420 in the curve as
per table is equal to 0.2881.
Hence, the required area in the curve between X = 420 and X = 570 is:
(0.2580) + (0.2881) = 0.5461
This is the required probability that an individual selected at random will belong to
income group of Rs 420 to Rs 570.
1''
Example 4.42: A certain company manufactures 1 all-purpose rope made from
2
imported hemp. The manager of the company knows that the average load-bearing
capacity of the rope is 200 lbs. Assuming that normal distribution applies, find the
1''
standard deviation of load-bearing capacity for the 1 rope if it is given that the
2
rope has a 0.1210 probability of breaking with 68 lbs or less pull.
Solution: Given information can be depicted in a normal curve as shown in the
following figure:
Probability of this
area (0.5) – (0.1210) = 0.3790

σ = ? (to be found out)

Probability of this area

(68 lbs. or less)
as given is 0.1210

μ = 200
X = 68 z=0
If the probability of the area falling within µ = 200 and X = 68 is 0.3790 as stated
above, the corresponding value of Z as per the table showing the area of the
normal curve is – 1.17 (minus sign indicates that we are in the left portion of the
curve)
Now to find , you can write:
X 
Z

68  200
or 1.17 

or 1.17  132
Self - Learning
270 Material or   = 112.8 lbs approx.
Thus, the required standard deviation is 112.8 lbs approximately. Statistical Computation and
Probability Distributiona
Example 4.43: In a normal distribution,
_ 31 per cent items are below 45 and 8 per
cent are above 64. Find the X and  of this distribution.
Solution: You can depict the given information in a normal curve as follows: NOTES
X X

X
X

If the probability of the area falling within µ and X = 45 is 0.19 as stated above, the
corresponding value of Z from the table showing the area of the normal curve is
– 0.50. Since, you are in the left portion of the curve so we can express this as
under,
45  
0.50  (1)

Similarly, if the probability of the area falling within µ and X = 64 is 0.42 as stated
above, the corresponding value of Z from the area table is +1.41. Since, you are
in the right portion of the curve, so you can express this as under:
64  
1.41  (2)

If you solve Equations (1) and (2) above to obtain the value of µ or X , you have:
 0.5  = 45 – µ (3)
1.41  = 64 – µ (4)
By subtracting the Equation (4) from (3), you have:
 1.91  = –19
  = 10
Putting  = 10 in Equation (3), you have:
 5 = 45 – 
  = 50
_
Hence, X (or µ)=50 and  =10 for the concerning normal distribution.

4.7.5 Uniform Distribution (Discrete Random and

Continous Variable)
When a random variable x takes discrete values x1, x2,...., xn with probabilities
p1, p2,...,pn, we have a discrete probability distribution of X.
The function p(x) for which X = x1, x2,..., xn takes values p1, p2,....,pn, is the
probability function of X.
Self - Learning
Material 271
Statistical Computation and The variable is discrete because it does not assume all values. Its properties
Probability Distribution
are:
p(xi) = Probability that X assumes the value x
NOTES = Prob (x = xi) = pi
p(x)  0, p(x) = 1
For example, four coins are tossed and the number of heads X noted. X can take
values 0, 1, 2, 3, 4 heads.
4

p(X = 0) =   
1 1
 2  16
3
11 4
p(X = 1) = C1     
4

 2   2  16
2 2
p(X = 2) = C2  1   1   6
4

 2   2  16
3

p(X = 3) = C3     
41 1 4
 2   2  16
4 0
p(X = 4) = 4 C4     
1 1 1
 2   2  16

6
16

5
16

4
16

3
16

2
16

1
16

O
0 1 2 3 4

4
1 4 6 4 1
 p( x)  16  16  16  16  16  1
x 2

This is a discrete probability distribution (Refer Example 4.44).

Example 4.44: If a discrete variable X has the following probability function, then
find (i) a (ii) p(X  3) (iii) p(X  3).
Self - Learning
272 Material
Solution: Statistical Computation and
Probability Distributiona
x1 p(xi)
0 0
1 a NOTES
2 2a
3 2a2
4 4a2
5 2a
Since p(x) = 1 , 0 + a + 2a + 2a2 + 4a2 + 2a = 1
 6a2 + 5a – 1 = 0, so that (6a – 1) (a + 1) = 0
1
a= or a = –1 (not admissible)
6

1 5
For a = , p(X  3) = 0 + a + 2a + 2a2 = 2a2 + 3a =
6 9

4
p(X  3) = 4a2 + 2a =
9

Discrete Distributions
There are several discrete distributions. Some other discrete distributions are
described as follows:
(i) Uniform or Rectangular Distribution
Each possible value of the random variable x has the same probability in the uniform
distribution. If x takes vaues x1, x2....,xk, then,
1
p(xi, k) =
k
The numbers on a die follow the uniform distribution,
1
p(xi, 6) = (Here, x = 1, 2, 3, 4, 5, 6)
6

Bernoulli Trials
In a Bernoulli experiment, an even E either happens or does not happen (E).
Examples are, getting a head on tossing a coin, getting a six on rolling a die, and so
on.
The Bernoulli random variable is written,
X = 1 if E occurs
= 0 if E occurs
Since there are two possible values it is a case of a discrete variable
where, Self - Learning
Material 273
Statistical Computation and Probability of success = p = p(E)
Probability Distribution
Profitability of failure = 1 – p = q = p(E)
We can write,
NOTES For k = 1, f(k) = p
For k = 0, f(k) = q
For k = 0 or 1, f(k) = pkq1–k
Negative Binomial
In this distribution, the variance is larger than the mean.
Suppose, the probability of success p in a series of independent Bernoulli
trials remains constant.
Suppose the rth success occurs after x failures in x + r trials, then
(i) The probability of the success of the last trial is p.
(ii) The number of remaining trials is x + r – 1 in which there should be
r – 1 successes. The probability of r – 1 successes is given by,
x  r –1
Cr –1 p r –1 q x
The combined pobability of cases (i) and (ii) happening together is,
x  r –1
p(x) = px Cr –1 p r –1 q x x = 0, 1, 2,....
This is the Negative Binomial distribution. We can write it in an alternative
form,
p(x) = –r
Cx p r (q) x x = 0, 1, 2,....
This can be summed up as,
In an infinite series of Bernoulli trials, the probability that x + r trials will be
required to get r successes is the negative binomial,
x  r –1
p(x) = Cr –1 p r –1 q x r0
If r = 1, it becomes the geometric distribution.
If p  0,  , rp = m a constant, then the negative binomial tends to be the
Poisson distribution.
(ii) Geometric Distribution
Suppose the probability of success p in a series of independent trials remains
constant.
Suppose, the first success occurs after x failures, i.e., there are x failures
preceding the first success. The probability of this event will be given by p(x) =
qxp (x = 0, 1, 2,.....)
This is the geometric distribution and can be derived from the negative
binomial. If we put r = 1 in the negative binomial distribution, then
x  r –1
Self - Learning
p(x) = Cr –1 p r –1 q x
274 Material
 We get the geometric distribution, Statistical Computation and
Probability Distributiona
p(x) = x C0 p1 q x  pq x
p
p
p(x) = q
n0
x
p
1 q
1 NOTES

p
E(x) = Mean =
q
p
Variance =
q2
x
1
Mode =  
 2
Refer Example 4.45 to understand it better.
Example 4.45: Find the expectation of the number of failures preceding the first
success in an infinite series of independent trials with constant probability p of
success.
Solution:
The probability of success in,
1st trial = p (Success at once)
2nd trial = qp (One failure, then success, and so on)
3rd trial = q2p (Two failures, then success, and so
on)
The expected number of failures preceding the success,
E(x) = 0 . p + 1. pq + 2p2p + ............
= pq(1 + 2q + 3q2 + .........)
1 1 q
= pq (1 – q )2  qp p 2  p

Since p = 1 – q.
(iii) Hypergeometic Distribution
From a finite population of size N, a sample of size n is drawn without replacement.
Let there be N1 successes out of N.
The number of failures is N2 = N – N1.
The disribution of the random variable X, which is the number of successes
obtained in the discussed case, is called the hypergeometic distribution.
N1
C xN Cn  x
p(x) = N (X = 0, 1, 2, ...., n)
Cn

Here, x is the number of successes in the sample and n – x is the number of

failures in the sample. Self - Learning
Material 275
Statistical Computation and It can be shown that,
Probability Distribution
N1
Mean : E(X) = n
N
NOTES
N  n  nN1 nN12 
Variance : Var (X) = N  1  N – N 
 
Example 4.46: There are 20 lottery tickets with three prizes. Find the probability
that out of 5 tickets purchased exactly two prizes are won.
Solution:
We have N1 = 3, N2 = N – N1 = 17, x = 2, n = 5.
3
C2 17C3
p(2) = 20
C5
3
C0 17C5
The probability of no prize p(0) = 20
C5
3
C1 17C4
The probability of exactly 1 prize p(1) = 20
C5
Example 4.47: Examine the nature of the distibution if balls are drawn, one at a
time without replacement, from a bag containing m white and n black balls.
Solution:
It is the hypergeometric distribution. It corresponds to the probability that x balls
will be white out of r balls so drawn and is given by,
x
C x n Cr  x
p(x) = m n
Cr
(iv) Multinomial
There are k possible outcomes of trials, viz., x1, x2, ..., xk with probabilities p1, p2,
..., pk, n independent trials are performed. The multinomial distibution gives the
probability that out of these n trials, x1 occurs n1 times, x2 occurs n2 times, and so
n! n1 n2 n
on. This is given by n !n !....n ! p1 p2 .... pk
1 2 k

Where, n
i 1
i n

Characteristic Features of the Binomial Distribution

The following are the characteristics of binomial distribution:
(i) It is a discrete distribution.
(ii) It gives the probability of x successes and n – x failures in a specific order.
(iii) The experiment consists of n repeated trials.
Self - Learning
276 Material
 (iv) Each trial results in a success or a failure. Statistical Computation and
Probability Distributiona
(v) The probability of success remains constant from trial to trial.
(vi) The trials are independent.
(vii) The success probability p of any outcome remains constant over time. This NOTES
condition is usually not fully satisfied in situations involving management and
economics, e.g., the probability of response from successive informants is
not the same. However, it may be assumed that the condition is reasonably
well satisfied in many cases and that the outcome of one trial does not
depend on the outcome of another. This condition too, may not be fully
satisfied in many cases. An investigator may not approach a second informant
with the same mind set as used for the first informant.
(viii) The binomial distribution depends on two parameters, n and p. Each set of
different values of n, p has a different binomial distribution.
(ix) If p = 0.5, the distribution is symmetrical. For a symmetrical distribution, in
n
Prob. (X = 0) = Prob (X = n)
i.e., the probabilities of 0 or n successes in n trials will be the same. Similarly,
Prob (X = 1) = Prob(X = n – 1), and so on.
If p > 0.5, the distribution is not symmetrical. The probabilities on the right
are larger than those on the left. The reverse case is when p < 0.5.
When n becomes large the distribution becomes bell shaped. Even when n
is not very large but p  0.5, it is fairly bell shaped.
(x) The binomial distribution can be approximated by the normal. As n becomes
large and p is close to 0.5, the approximation becomes better.
Through the following examples you can understand multinomial better.
Example 4.48: Explain the concept of a discrete probability distribution.
Solution:
If a random variable x assumes n discrete values x1, x2, ........xn, with respective
probabilities p1, p2,...........pn(p1 + p2 + .......+ pn = 1), then the distribution of
values xi with probabilities pi (= 1, 2,.....n), is called the discrete probability
distribution of x.
The frequency function or frequency distribution of x is defined by p(x)
which for different values x1, x2, ........xn of x, gives the corresponding probabilities:
p(xi) = pi where, p(x)  0 p(x) = 1
Example 4.49: For the following probability distribution, find p(x > 4) and
p(x  4):

x 0 1 2 3 4 5
p ( x) 0 a a/2 a/2 a/4 a/4

Self - Learning
Material 277
Statistical Computation and Solution:
Probability Distribution
a a a a
Since, p(x) = 1,0  a    1
2 2 4 4
NOTES 5 2
 a =1 or a=
2 5
9 1
p(x > 4) = p(x = 5) = 
4 10
a a a 9a 9
p(x  4) = 0 + a +    
2 2 4 4 10
Example 4.50: A fair coin is tossed 400 times. Find the mean number of heads
and the corresponding standard deviation.
Solution:
1
This is a case of binomial distribution with p = q = , n = 400.
2
1
The mean number of heads is given by  = np = 400  = 200.
2

1 1
and S. D.  = npq  400    10
2 2
Example 4.51: A manager has thought of 4 planning strategies each of which has
an equal chance of being successful. What is the probability that at least one of his
1 3
strategies will work if he tries them in 4 situations? Here p = ,q= .
4 4
Solution:
The probability that none of the strategies will work is given by,
0 4 4
p(0) = C0       
4 1 3 3
 4 4 4
4
 3  175
The probability that at least one will work is given by 1     .
4 256

Example 4.52: For the Poisson distribution, write the probabilities of 0, 1, 2, ....
successes.
Solution:

mx
x p ( x)  e – m
x!
0 p (0)  e m0 / 0!
–m

m
1 p (1)  e – m  p (0).m
Self - Learning 1!
278 Material
 Statistical Computation and
2 Probability Distributiona
m m
2 e– m  p (2)  p (1).
2! 2
3
m m
3 e– m  p (3)  p (2). NOTES
3! 3


and so on.
Total of all probabilities p(x) = 1.
Example 4.53: What are the raw moments of Poisson distribution?
Solution:
First raw moment 1 = m
Second raw moment 2 = m2 + m
Third raw moment 3 = m3 + 3m2 + m
(v) Continuous probability distributions
When a random variate can take any value in the given interval a  x  b, it is a
continuous variate and its distribution is a continuous probability distribution.
Theoretical distributions are often continuous. They are useful in practice
because they are convenient to handle mathematically. They can serve as good
approximations to discrete distributions.
The range of the variate may be finite or infinite.
A continuous random variable can take all values in a given interval. A
continuous probability distribution is represented by a smooth curve.
The total area under the curve for a probability distribution is necessarily
unity. The curve is always above the x axis because the area under the curve for
any interval represents probability and probabilities cannot be negative.
If X is a continous variable, the probability of X falling in an interval with end
points z1, z2 may be written p(z1  X  z2).
This probability corresponds to the shaded area under the curve in Figure
4.13.

z1 z2

Fig. 4.13 Continuous Probability Distribution

Self - Learning
Material 279
Statistical Computation and A function is a probability density function if,
Probability Distribution

–
p ( x ) dx  1, p ( x )  0, –   x   , i.e., the area under the curve p(x) is

NOTES 1 and the probability of x lying between two values a, b, i.e., p(a < x < b) is
positive. The most prominent example of a continuous probability function is the
normal distribution.
Cumulative Probability Function (CPF)
The Cumulative Probability Function (CPF) shows the probability that x takes a
value less than or equal to, say, z and corresponds to the area under the curve up
to z:
z
p( x  z )   p ( x )dx
–

This is denoted by F(x).

Check Your Progress

11. What is absolute measure of dispersion?
12. What is relative measure of dispersion?
13. Define range.
14. Calculate standard deviation for the series 1, 2, 3, 5, 7.
15. Define the terms 'Simple Probability' and 'Joint Probability'.
16. What are the types of probability distributions?
17. Write the probability function of binomial distribution.
18. What are the different parameters of binomial distribution?
19. Under what circumstances would you use binomial distribution?
20. What is Poisson distribution?
21. What is the use of exponential distribution?
22. Define a normal distribution.
23. When is a distribution said to be symmetrical?

4.8 ANSWERS TO ‘CHECK YOUR PROGRESS’

1. The term statistics is used to mean either statistical data or statistical method.
When it is used in the sense of statistical data it refers to quantitative aspects
of things, and is a numerical description.
2. The procedure of classification brings into relief the salient features of the
variable that is under investigation. This can be clearly illustrated by an
example. If we are given the marks in mathematics of each individual student
of a class and if it is desired to judge the performance of the class on the
basis of these data it will not be an easy matter. Human mind has its limitations
Self - Learning and cannot easily grasp a multitude of figures. But if the students are classified
280 Material
 i.e., if we put into one group all those boys who get more than second Statistical Computation and
Probability Distributiona
division marks, in still another group those who get third division marks,
and have a separate group of those who fait to get pass marks, it will be
easier for us to form a more precise idea about the performance of the
class. NOTES
3. Collection of facts is the first step in the statistical treatment of a problem.
Numerical facts are the raw materials upon which the statistician is to work
and just as in a manufacturing concern the quality of a finished product
depends, inter alia, upon the quality of the raw material, in the same manner,
the validity of statistical conclusions will be governed, among other
considerations, by the quality of data used.
4. In statistics, the term central tendency specifies the method through which
the quantitative data have a tendency to cluster approximately about some
value. A measure of central tendency is any precise method of specifying
this ‘central value’.
5. Arithmetic mean is also commonly known as the mean. Even though average,
in general, means measure of central tendency, when we use the word
average in our daily routine, we always mean the arithmetic average. The
term is widely used by almost everyone in daily communication.
6. The weighted arithmetic mean is particularly useful where we have to compute
the mean of means. If we are given two arithmetic means, one for each of
two different series, in respect of the same variable, and are required to find
the arithmetic mean of the combined series, the weighted arithmetic mean is
the only suitable method of its determination.
7. Median is that value of a variable which divides the series in such a manner
that the number of items below it is equal to the number of items above it.
Half the total number of observations lie below the median, and half above
it. The median is thus a positional average.
8. Mode is that value of the variable which occurs or repeats itself the greatest
number of times. The mode is the most ‘Fashionable’ size in the sense that
it is the most common and typical, and is defined by Zizek as ‘The value
occurring most frequently in a series (or group of items) and around which
the other items are distributed most densely’.
9. The four important methods of estimating mode of a series are: (i) Locating
the most frequently repeated value in the array; (ii) Estimating the mode by
interpolation; (iii) Locating the mode by graphic method; and (iv) Estimating
the mode from the mean and the median.
10. Some measures, other than the measures of central tendency, are often
employed when summarizing or describing a set of data where it is necessary
to divide the data into equal parts. These are positional measures and are
called quantiles and consist of quartiles, deciles and percentiles. The quartiles
divide the data into four equal parts. The deciles divide the total ordered
data into ten equal parts and the percentiles divide the data into 100 equal
parts. Consequently, there are three quartiles, nine deciles and 99 percentiles.
Self - Learning
Material 281
Statistical Computation and 11. Absolute measure of dispersion states the actual amount by which an item
Probability Distribution
on an average deviates from a measure of central tendency.
12. Relative measure of dispersion is a quotient computed by dividing the absolute
measures by a quantity in respect to which absolute deviation has been
NOTES
computed.
13. The range of a set of numbers is the difference between the maximum and
minimum values. It indicates the limits within which the values fall.
14. 2.15
15. Simple Probability: The term simple probability refers to a phenomenon
where only a simple or an elementary event occurs. For example, assume
that event (E), the drawing of a diamond card from a pack of 52 cards, is a
simple event. Since there are 13 diamond cards in the pack and each card
is equally likely to be drawn, the probability of event (E) or P[E] = 13/52
or 1/4.
Joint Probability: The term joint probability refers to the phenomenon of
occurrence of two or more simple events. For example, assume that event
(E) is a joint event (or compound event) of drawing a black ace from a
pack of cards. There are two simple events involved in the compound event,
which are: the card being black and the card being an ace. Hence, P[Black
ace] or P[E] = 2/52 since there are two black aces in the pack.
16. There are two types of probability distributions, discrete and continuous
probability distributions. In discrete probability distribution, the variable under
consideration is allowed to take only a limited number of discrete values
along with corresponding probabilities. On the other hand, in a continuous
probability distribution, the variable under consideration is allowed to take
on any value within a given range.
17. The probability function of binomial distribution is written as follows:
f (X = r) = nC r n–r
rp q
r= 0, 1, 2, …, n
where n= Numbers of trials
p= Probability of success in a single trial
q = (1 – p) = Probability of failure in a single trial
r= Number of successes in n trials
18. The parameters of binomial distribution are p and n, where p specifies the
probability of success in a single trial and n specifies the number of trials.
19. The use of binomial distribution is needed under the following circumstances:
(a) When we have to find the probability of heads in 10 throws of a fair
coin.
(b) When we have to find the probability that 3 out of 10 items produced
by a machine, which produces 8 per cent defective items on average,
will be defective.
20. Poisson distribution is a discrete probability distribution that is frequently
Self - Learning used in the context of operations research. Unlike binomial distribution,
282 Material
 Poisson distribution cannot be deduced on purely theoretical grounds based Statistical Computation and
Probability Distributiona
on the conditions of the experiment. In fact, it must be based on the
experience, i.e. on the empirical results of past experiments relating to the
problem under study.
NOTES
21. Exponential distribution is used for describing a large class of phenomena,
particularly in the area of reliability theory and in queuing models.
22. Normal distribution is the most important and frequently used continuous
probability distribution among all the probability distributions. This is so
because this distribution fits well in many types of problems. This distribution
is of special significance in inferential statistics since it describes
probabilistically the link between a statistic and a parameter.
23. If p = 0.5, the distribution is symmetrical.

4.9 SUMMARY
 Statistics influence the operations of business and management in many
dimensions.
 Statistical applications include the area of production, marketing, promotion
of product, financing, distribution, accounting, marketing research, manpower
planning, forecasting, research and development, and so on.
 In statistics, the term central tendency specifies the method through which
the quantitative data have a tendency to cluster approximately about some
value.
 A measure of central tendency is any precise method of specifying this
‘central value’. In the simplest form, the measure of central tendency is an
average of a set of measurements, where the word average refers to as
mean, median, mode or other measures of location. Typically the most
commonly used measures are arithmetic mean, mode and median.
 While arithmetic mean is the most commonly used measure of central location,
mode and median are more suitable measures under certain set of conditions
and for certain types of data.
 There are several commonly used measures, such as arithmetic mean, mode
and median. These values are very useful not only in presenting the overall
picture of the entire data, but also for the purpose of making comparisons
among two or more sets of data.
 Arithmetic mean is also commonly known as the mean. Even though average,
in general, means measure of central tendency, when we use the word
average in our daily routine, we always mean the arithmetic average. The
term is widely used by almost everyone in daily communication.
 The weighted arithmetic mean is particularly useful where we have to compute
the mean of means. If we are given two arithmetic means, one for each of
two different series, in respect of the same variable, and are required to find
the arithmetic mean of the combined series, the weighted arithmetic mean is
the only suitable method of its determination. Self - Learning
Material 283
Statistical Computation and  Median is that value of a variable which divides the series in such a manner
Probability Distribution
that the number of items below it is equal to the number of items above it.
Half the total number of observations lie below the median, and half above
it. The median is thus a positional average.
NOTES
 The median of ungrouped data is found easily if the items are first arranged
in order of the magnitude. The median may then be located simply by
counting, and its value can be obtained by reading the value of the middle
observations.
 The median can quite conveniently be determined by reference to the ogive
which plots the cumulative frequency against the variable. The value of the
item below which half the items lie, can easily be read from the ogive.
 Median is a positional average and hence the extreme values in the data set
do not affect it as much as they do to the mean.
 The mode of a distribution is the value at the point around which the items
tend to be most heavily concentrated. It is the most fre-quent or the most
common value, provided that a sufficiently large number of items are available,
to give a smooth distribution.
 The measures of dispersion bring out this inequality. In engineering problems
too the variability is an important concern.
 The amount of variability in dimensions of nominally identical components
is critical in determining whether or not the components of a mass-produced
item will be really interchangeable.
 Probability can be defined as a measure of the likelihood that a particular
event will occur. It is a numerical measure with a value between 0 and 1 of
such a likelihood where the probability of zero indicates that the given event
cannot occur and the probability of one assures certainty of such an
occurrence.
 Probability theory provides us with a mechanism for measuring and analysing
uncertainties associated with future events. Probability can be subjective or
objective.
 The objective probability of an event, on the other hand, can be defined as
the relative frequency of its occurrence in the long run.
 Binomial distribution is probably the best known of discrete distributions.
The normal distribution, or Z-distribution, is often used to approximate the
binomial distribution.
 If the sample size is very large, the Poisson distribution is a philosophically
more correct alternative to binomial distribution than normal distribution.
 One of the main differences between the Poisson distribution and the binomial
distribution is that in using the binomial distribution all eligible phenomena
are studied, whereas in the Poisson, only the cases with a particular outcome
are studied.
 Exponential distribution is a very commonly used distribution in reliability
Self - Learning
engineering. The reason for its widespread use lies in its simplicity, so much
284 Material that it has even been employed in cases to which it does not apply directly.
  Amongst all types of distributions, the normal probability distribution is by Statistical Computation and
Probability Distributiona
far the most important and frequently used distribution because it fits well in
many types of problems.

NOTES
4.10 KEY TERMS
 Statistics: Numerical statements of facts in any department of inquiry placed
in relation to each other.
 Median: Measure of central tendency and it appears in the centre of an
ordered data.
 Mode: A form of average that can be defined as the most frequently
occurring value in the data.
 The weighted arithmetic mean: The weighted arithmetic mean is
particularly useful where we have to compute the mean of means.
 Mean deviation: The mean deviation of a series of values is the arithmetic
mean of their absolute deviations.
 Standard deviation: The square root of the average of the squared
deviations from their mean of a set of observations.
 Range: The difference between the maximum and minimum values of a set
of number. It indicates the limits within which the values fall.
 Classical theory of probability: It is the theory of probability based on
the number of favourable outcomes and the number of total outcomes.
 Binomial distribution: It is also called the Bernoulli process and is used to
describe a discrete random variable.
 Poisson distribution: It is used to describe the empirical results of past
experiments relating to the problem and plays an important role in queuing
theory, inventory control problems and risk models.
 Exponential distribution: It is a continuous probability distribution and is
used to describe the probability distribution of time between two events.
 Normal distribution: It is referred to as most important and frequently
used continuous probability distribution as it fits well in many types of
problems.

4.11 SELF-ASSESSMENT QUESTIONS AND

EXERCISES

Short-Answer Questions
1. State the significance of statistical methods.
2. How does statistics aid in interpreting conditions?
3. List the various characteristics of statistical data.
4. Write a short note on the origin of statistics. Self - Learning
Material 285
Statistical Computation and 5. Define the term arithmetic mean.
Probability Distribution
6. Differentiate between a mean and a mode.
7. Write three characteristics of mean.
NOTES 8. What is the importance of arithmetic mean in statistics?
9. Define the term median with the help of an example.
10. Differentiate between geometric and harmonic mean.
11. Define the terms quartiles, percentiles and deciles.
12. Write the definition and formula of quartile deviation.
13. How will you calculate the mean deviation of a given data?
14. What is standard deviation? Why is it used in statistical evaluation of data?
15. Define the concept of probability.
16. What are the different theories of probability? Explain briefly.
17. Define probability distribution and probability functions.
18. What do you mean by the binomial distribution and its measures?
19. How can a binomial distribution be fitted to given data?
20. How will define the Poisson distribution and its important measures?
21. Poisson distribution can be an approximation of binomial distribution. Define.
22. When is the Poisson distribution used?
23. What is exponential distribution?
24. Define any six characteristics of normal distribution.
25. Write the formula for measuring the area under the curve.
26. How will you define the circumstances when the normal probability
distribution can be used?
27. What is CPF?
Long-Answer Questions
1. Give a detailed description on the various functions of statistics.
2. Describe the various features of the statistical procedure.
3. According to Bowley statistics is ‘The science of counting’. Do you agree?
Give reasons.
4. An elevator is designed to carry a maximum load of 3200 pounds. If 18
passengers are riding in the elevator with an average weight of 154 pounds,
is there any danger that the elevator might be overloaded?
5. In a car assembly plant, the cars were diagnostically checked after assembly
and before shipping them to the dealers. All such cars with any defect were
returned for extra work. The number of such defective cars returned in one
day of a 16-days period is given as follows:
30, 34, 10, 16, 28, 9, 22, 2, 6, 23, 25, 10, 15, 10, 8, 24

Self - Learning
286 Material
 (i) Find the average number of defective cars returned for extra work per Statistical Computation and
Probability Distributiona
day.
(ii) Find the median for defective cars per day.
(iii) Find the mode for defective cars per day.
NOTES
(iv) Find Q2.
(v) Find D2.
(vi) Find P70.
6. Calculate mean deviation and its coefficient about median, arithmetic mean
and mode for the following figures, and show that the mean deviation about
the median is least.
(103, 50, 68, 110, 108, 105, 174, 103, 150, 200, 225, 350, 103)
7. A group has  = 10, N = 60, 2 = 4. A subgroup of this has X 1 = 11, N1 =
4 0 ,
1 = 2.25. Find the mean and the standard deviation of the other subgroup.
2

8. The following are some of the particulars of the distribution of weights of

boys and girls in a class
Boys Girls
Number 100 50
Mean weight 60 kg 45 kg
Variance 9 4

(i) Find the standard deviation of the combined data.

(ii) Which of the two distributions is more variable?
9. Find the Q.D. and coefficient of Q.D. for the following data:
Marks No. of Students
35 – 40 4
40 – 45 8
45 – 50 12
50 – 55 7
55 – 60 2

10. Find the Q.D. from the mean for the series 5, 7, 10, 12, 6.
11. (i) Calculate the mean deviation from the mean for the following data.
What light does it throw on the social conditions of the community?
Data showing differences in ages of husbands and wives.
Difference in years Frequency
0–5 499
5 – 10 705
10 – 15 507
15– 20 281
20 – 25 109
25 – 30 52
30 – 35 164
Self - Learning
Material 287
Statistical Computation and (ii) The age distribution of 100 life insurance policy-holders is as follows:
Probability Distribution
Age No. of policy holders
17 – 19 9
20 – 25 16
NOTES
26 – 35 12
36 – 40 26
41 – 50 14
51 – 55 12
56 – 60 5
61 – 70 5

12. Calculate the mean deviation from the mean and the median and their
coefficients for the following data.
Size of shoes: 3 6 11 2 4 10 5 7 8 9
No. of pairs sold: 10 15 25 6 4 3 2 8 9 4
13. Discuss briefly about the measures of dispersion with the help of giving
examples and characteristics.
14. Explain briefly about the standard deviation. Give appropriate examples.
15. A family plans to have two children. What is the probability that both children
will be boys? (List all the possibilities and then select the one which would
be two boys.)
16. A card is selected at random from an ordinary well-shuffled pack of 52
cards. What is the probability of getting,
(i) A king (ii) A spade
(iii) A king or an ace (iv) A picture card
17. A wheel of fortune has numbers 1 to 40 painted on it, each number being at
equal distance from the other so that when the wheel is rotated, there is the
same chance that the pointer will point at any of these numbers. Tickets
have been issued to contestants numbering 1 to 40. The number at which
the wheel stops after being rotated would be the winning number. What is
the probability that,
(i) Ticket number 29 wins.
(ii) One person who bought 5 tickets numbered 18 to 22 (inclusive), wins
the prize.
18. (a) Explain the meaning of the Bernoulli process pointing out its main
characteristics.
(b) Give a few examples narrating some situations wherein binomial pr
distribution can be used.
19. State the distinctive features of the binomial, Poisson and normal probability
distributions. When does a binomial distribution tend to become a normal
and a Poisson distribution? Explain.

Self - Learning
288 Material
 20. Explain the circumstances when the following probability distributions are Statistical Computation and
Probability Distributiona
used:
(a) Binomial distribution
(b) Poisson distribution
NOTES
(c) Exponential distribution
(d) Normal distribution
21. Certain articles have been produced of which 0.5 per cent are defective
and the articles are packed in cartons each containing 130 articles. What
proportion of cartons are free from defective articles? What proportion of
cartons contain two or more defective articles?
(Given e–0.5=0.6065).
22. The following mistakes per page were observed in a book:
No. of Mistakes No. of Times the Mistake
Per Page Occurred
0 211
1 90
2 19
3 5
4 0
Total 345

Fit a Poisson distribution to the given data and test the goodness of fit.
23. In a distribution exactly normal, 7 per cent of the items are under 35 and
89 per cent are under 63. What are the mean and standard deviation of the
distribution?
24. Assume the mean height of soldiers to be 68.22 inches with a variance of
10.8 inches. How many soldiers in a regiment of 1000 would you expect to
be over six feet tall?
25. Fit a normal distribution to the following data:
Height in inches Frequency
60–62 5
63–65 18
66–68 42
69–71 27
72–74 8
26. Analyse the types of discrete distributions with the help of giving examples.

4.12 FURTHER READING

Chance, William A. 1969. Statistical Methods for Decision Making. Illinois:
Richard D Irwin.
Chandan, J.S., Jagjit Singh and K.K. Khanna. 1995. Business Statistics. New
Delhi: Vikas Publishing House. Self - Learning
Material 289
Statistical Computation and Elhance, D.N. 2006. Fundamental of Statistics. Allahabad: Kitab Mahal.
Probability Distribution
Freud, J.E., and F.J. William. 1997. Elementary Business Statistics – The
Modern Approach. New Jersey: Prentice-Hall International.
NOTES Goon, A.M., M.K. Gupta, and B. Das Gupta. 1983. Fundamentals of Statistics.
Vols. I & II, Kolkata: The World Press Pvt. Ltd.
Gupta, S.C. 2008. Fundamentals of Business Statistics. Mumbai: Himalaya
Publishing House.
Kothari, C.R. 1984. Quantitative Techniques. New Delhi: Vikas Publishing
House.
Levin, Richard. I., and David. S. Rubin. 1997. Statistics for Management. New
Jersey: Prentice-Hall International.
Meyer, Paul L. 1970. Introductory Probability and Statistical Applications.
Massachusetts: Addison-Wesley.
Gupta, C.B. and Vijay Gupta. 2004. An Introduction to Statistical Methods,
23rd Edition. New Delhi: Vikas Publishing House Pvt. Ltd.
Hooda, R. P. 2013. Statistics for Business and Economics, 5th Edition. New
Delhi: Vikas Publishing House Pvt. Ltd.
Anderson, David R., Dennis J. Sweeney and Thomas A. Williams. Essentials of
Statistics for Business and Economics. Mumbai: Thomson Learning,
2007.
S.P. Gupta. 2021. Statistical Methods. Delhi: Sultan Chand and Sons.

Self - Learning
290 Material
 Estimation and

UNIT 5 ESTIMATION AND Hypothesis Testing

HYPOTHESIS TESTING
NOTES
Structure
5.0 Introduction
5.1 Objectives
5.2 Sampling Theory
5.2.1 Parameter and Statistic
5.2.2 Sampling Distribution of Sample Mean
5.3 Sampling Distribution of the Number of Successes
5.4 The Student’s Distribution
5.5 Theory of Estimation
5.5.1 Point Estimation
5.5.2 Interval Estimation
5.6 Hypothesis Testing
5.6.1 Test of Hypothesis Concerning Mean and Proportion
5.6.2 Test of Hypothesis Conerning Standard Deviation
5.7 Answers to ‘Check Your Progress’
5.8 Summary
5.9 Key Terms
5.10 Self-Assessment Questions and Exercises
5.11 Further Reading

5.0 INTRODUCTION
In statistics, quality assurance, and survey methodology, sampling is the selection
of a subset (a statistical sample) of individuals from within a statistical population
to estimate characteristics of the whole population. Statisticians attempt to collect
samples that are representative of the population in question. Sampling has lower
costs and faster data collection than measuring the entire population and can provide
insights in cases where it is infeasible to sample an entire population. Each
observation measures one or more properties (such as weight, location, colour) of
independent objects or individuals. In survey sampling, weights can be applied to
the data to adjust for the sample design, particularly in stratified sampling. Results
from probability theory and statistical theory are employed to guide the practice.
In business and medical research, sampling is widely used for gathering information
about a population. Acceptance sampling is used to determine if a production lot
of material meets the governing specifications.
Single or isolated facts or figures cannot be called statistics as these cannot
be compared or related to other figures within the same framework. Hence, any
quantitative and numerical data can be identified as statistics when it possesses
certain identifiable characteristics as per the norms of statistics. The area of statistics
can be split up into two identifiable sub-areas. These sub-areas constitute descriptive
statistics and inferential statistics. This unit will describe some of the terms used
extensively in the field of statistics for scientific measurement. Statistical investigation
is a comprehensive process and requires systematic collection of data about some
group of people or objects, describing and organizing the data, analysing the data
with the help of various statistical methods, summarizing the analysis and using the Self - Learning
results for making judgements, decisions and predictions. Material 291
Estimation and A sampling distribution or finite-sample distribution is the probability
Hypothesis Testing
distribution of a given random-sample-based statistic. If an arbitrarily large number
of samples, each involving multiple observations (data points), were separately
used in order to compute one value of a statistic (such as, for example, the sample
NOTES mean or sample variance) for each sample, then the sampling distribution is the
probability distribution of the values that the statistic takes on. In many contexts,
only one sample is observed, but the sampling distribution can be found theoretically.
Sampling distributions are important in statistics because they provide a major
simplification en route to statistical inference. More specifically, they allow analytical
considerations to be based on the probability distribution of a statistic, rather than
on the joint probability distribution of all the individual sample values.
Estimation (or estimating) is the process of finding an estimate, or
approximation, which is a value that is usable for some purpose even if input data
may be incomplete, uncertain, or unstable. The value is nonetheless usable because
it is derived from the best information available. Typically, estimation involves ‘Using
the value of a statistic derived from a sample to estimate the value of a corresponding
population parameter’. The sample provides information that can be projected,
through various formal or informal processes, to determine a range most likely to
describe the missing information. An estimate that turns out to be incorrect will be
an overestimate if the estimate exceeded the actual result, and an underestimate if
the estimate fell short of the actual result.
Statistical hypothesis test is a method of statistical inference used to determine
a possible conclusion from two different, and likely conflicting, hypotheses. In a
statistical hypothesis test, a null hypothesis and an alternative hypothesis is proposed
for the probability distribution of the data. If the sample obtained has a probability
of occurrence less than the pre-specified threshold probability, the significance
level, given the null hypothesis is true, the difference between the sample and the
null hypothesis in deemed statistically significant. The hypothesis test may then
lead to the rejection of null hypothesis and acceptance of alternate hypothesis.
The process of distinguishing between the null hypothesis and the alternative
hypothesis is aided by considering Type I error and Type II error which are
controlled by the pre-specified significance level. Hypothesis tests based on
statistical significance are another way of expressing confidence intervals (more
precisely, confidence sets). In other words, every hypothesis test based on
significance can be obtained via a confidence interval, and every confidence interval
can be obtained via a hypothesis test based on significance.
In this unit, you will learn about the sampling theory, sampling distribution of
the number of successes, the student’s distribution theory of estimation, theory of
estimation and hypothesis testing.

5.1 OBJECTIVES
After going through this unit, you will be able to:
 Explain about the sampling theory
 Discuss the methods of sampling
 Learn about parameter and statistics
Self - Learning
292 Material
  Explain the concept of population in statistics Estimation and
Hypothesis Testing
 Know about estimation
 Understand hypothesis distribution and test of significance
NOTES
5.2 SAMPLING THEORY
A universe is the complete group of items about which knowledge is sought. The
universe may be finite or infinite. Finite universe is one which has a definite and
certain number of items but when the number of items is uncertain and infinite, the
universe is said to be an infinite universe. Similarly the universe may be hypothetical
or existent. In the former case the universe in fact does not exist and we can only
imagine the items constituting it. Tossing of a coin or throwing of a dice are examples
of hypothetical universes. Existent universe is a universe of concrete objects, i.e.,
the universe where the items constituting it really exist. On the other hand, the term
sample refers to that part of the universe which is selected for the purpose of
investigation. The theory of sampling studies the relationships that exist between
the universe and the sample or samples drawn from it.
5.2.1 Parameter and Statistic
It would be appropriate to explain the meaning of two terms viz., parameter and
statistic. All the statistical measures based on all items of the universe are termed
as parameters whereas statistical measures worked out on the basis of sample
studies are termed as sample statistics. Thus, a sample mean or a sample standard
deviation is an example of statistic whereas the universe mean or universe standard
deviation is an example of a parameter.
The main problem of sampling theory is the problem of relationship between
a parameter and a statistic. The theory of sampling is concerned with estimating
the properties of the population from those of the sample and also with gauging the
precision of the estimate. This sort of movement from particular Sample towards
general Universe is what is known as statistical induction or statistical inference. In
more clear terms, ‘From the sample we attempt to draw inferences concerning the
universe. In order to be able to follow this inductive method, we first follow a
deductive argument which is that we imagine a population or universe (finite or
infinite) and investigate the behaviour of the samples drawn from this universe
applying the laws of probability.’ The methodology dealing with all this is known as
Sampling Theory.
Objects of Sampling Theory
Sampling theory is to attain one or more of the following objectives:
(a) Statistical Estimation: Sampling theory helps in estimating unknown
population quantities or what are called parameters from a knowledge of
statistical measures based on sample studies often called as ‘Statistic’. In
other words, to obtain the estimate of parameter from statistic is the main
objective of the sampling theory. The estimate can either be a point estimate
or it may be an interval estimate. Point estimate is a single estimate
Self - Learning
Material 293
Estimation and expressed in the form of a single figure but interval estimate has two limits,
Hypothesis Testing
the upper and lower limits. Interval estimates are often used in statistical
induction.
(b) Tests of Hypotheses or Tests of Significance: The second objective of
NOTES
sampling theory is to enable us to decide whether to accept or reject
hypotheses or to determine whether observed samples differ significantly
from expected results. The sampling theory helps in determining whether
observed differences are actually due to chance or whether they are really
significant. Tests of significance are important in the theory of decisions.
(c) Statistical Inference: Sampling theory helps in making generalization about
the universe from the studies based on samples drawn from it. It also helps
in determining the accuracy of such generalizations.
5.2.2 Sampling Distribution of Sample Mean
In sampling theory we are concerned with what is known as the sampling
distribution. For this purpose we can take certain number of samples and for each
sample we can compute various statistical measures such as mean, standard
deviation etc. It is to be noted that each sample will give its own value for the
statistic under consideration. All these values of the statistic together with their
relative frequencies with which they occur, constitute the sampling distribution.
We can have sampling distribution of means or the sampling distribution of standard
deviations or the sampling distribution of any other statistical measure. The sampling
distribution tends quite closer to the normal distribution if the number of samples is
large. The significance of sampling distribution follows from the fact that the
mean of a sampling distribution is the same as the mean of the universe.
Thus, the mean of the sampling distribution can be taken as the mean of the
universe.
The Concept of Standard Error (or S.E.)
The standard deviation of sampling distribution of a statistic is known as its standard
error and is considered the key to sampling theory. The utility of the concept of
standard error in statistical induction arises on account of the following reasons:
(a) The standard error helps in testing whether the difference between observed
and expected frequencies could arise due to chance. The criterion usually
adopted is that if a difference is upto 3 times the S.E. then the difference is
supposed to exist as a matter of chance and if the difference is more than 3
times the S.E., chance fails to account for it, and we conclude the difference
as significant difference. This criterion is based on the fact that at x ± 3(S.E.),
the normal curve covers an area of 99.73 per cent. The product of the
critical value at certain level of significance and the S. E. is often described
as the Sampling Error at that particular level of significance. We can test the
difference at certain other levels of significance as well depending upon our
requirement.

Self - Learning
294 Material
 (b) The standard error gives an idea about the reliability and precision of a Estimation and
Hypothesis Testing
sample. If the relationship between the standard deviation and the sample
size is kept in view, one would find that the standard error is smaller than
the standard deviation. The smaller the S.E. the greater the uniformity of
the sampling distribution and hence greater is the reliability of sample. NOTES
Conversely, the greater the S.E., the greater the difference between
observed and expected frequencies and in such a situation the unreliability
of the sample is greater. The size of S.E. depends upon the sample size;
the greater the number of items included in the sample the smaller the
error to be expected and vice versa.
(c) The standard error enables us to specify the limits, maximum and minimum,
within which the parameters of the population are expected to lie with a
specified degree of confidence. Such an interval is usually known as
confidence interval. The degree of confidence with which it can be asserted
that a particular value of the population lies within certain limits is known
as the level of confidence.

Procedure of Significance Testing

The following sequential steps constitute, in general, the procedure of significance
testing:
(a) Statement of the Problem: First, the problem has to be stated in clear
terms. It should be quite clear as to in respect of what the statistical decision
has to be taken. The problem may be, Whether the hypothesis is to be
rejected or accepted? Is the difference between a parameter and a statistic
significant? or the like ones.
(b) Defining the Hypothesis: Usually, we start with the null hypothesis
according to which it is presumed that there is no difference between a
parameter and a statistic. If we are take a decision whether the students
have been benefited from the extra coaching and if we start with the
supposition that they have not been benefited then this supposition would
be termed as null hypothesis which in symbolic form is denoted by H0.
As against null hypothesis, the researcher may as well start with some
alternative hypothesis, (symbolically H1) which specifies those values that
the researcher believes to hold true and then may test such hypothesis on
the basis of sample data. Only one alternative hypothesis can be tested
at one time against the null hypothesis.
(c) Selecting the Level of Significance: The hypothesis is examined on a
pre-determined level of significance. Generally, either 5 per cent level or
1 per cent level of significance is adopted for the purpose. However, it
can be stated here that the level of significance must be adequate keeping
in view the purpose and nature of enquiry.
(d) Computation of the Standard Error. After determining the level of
significance the standard error of the concerning statistic (mean, standard
deviation or any other measure) is computed. There are different formulae
for computing the standard errors of different statistic. For example, the

Self - Learning
Material 295
Estimation and
Hypothesis Testing Standard Deviation
Standard Error of Mean  , the standard error of
n
Standard Deviation
NOTES Standard Deviation = , the standard error of Karl
2n
1 r2
Pearson’s Coefficient of Correlation = and so on. (A detailed
n
description of important standard errors formulae has been given on the
pages that follow).
(e) Calculation of the Significance Ratio: The significance ratio, symbolically
described as z, t, f etc. depending on the test we use, is often calculated
by diving the difference between a parameter and a statistic by the standard
error concerned. Thus, in context of mean, of small sample when population
x
variance is not known, in context of t  and in context of difference
SE x
X1  X 2
between two sample means t  . (All this has been fully
SE diff x  x
1 2

explained while explaining sampling theory in respect of small samples of

variables later in this unit).
(f) Deriving the Inference: The significance ratio is then compared with the
predetermined critical value. If the ratio exceeds the critical value then the
difference is taken as significant but if the ratio is less than the critical value,
the difference is considered insignificant. For example, the critical value at
5 per cent level of significance is 1.96. If the computed value exceeds 1.96
then the inference would be that the difference at 5 per cent level is
significant and this difference is not the result of sampling fluctuations but
the difference is a real one and should be understood as such.

5.3 SAMPLING DISTRIBUTION OF THE

NUMBER OF SUCCESSES

Population
A population, in statistical terms, is the totality of things under consideration. It is
the collection of all values of the variable that is under study. For instance, if we
are interested in knowing as to how much on an average an American bachelor
spends on his clothes per year, then all American bachelors would constitute the
population. Similarly, if we want to know the percentage of adult American travellers
who go to Europe, then only those adult Americans who travel are considered as
population.
The amount paid by parents in one year for an average Class I day-boarding
public school student can be evaluated by calculating the fee structure, such as
admission fee, day-boarding fees, tuition fees and annual charges. Thus, Class I
Self - Learning
296 Material day-boarding students would constitute the specific population group.
 Another example we can consider about is the population of coal mines Estimation and
Hypothesis Testing
workers, who are suffering from pneumoconiosis throughout the country. To
evaluate this, we collect information on all cases of pneumoconiosis in different
coal mines in the country.
NOTES
A summary measure that describes any given characteristic of the population
is known as a parameter. For example, the measure, the average income of
American professors, would be considered as a parameter since it describes the
characteristic of income of the population of American professors.
Sample
A sample is a portion of the total population that is considered for study and
analysis. For instance, if we want to study the income pattern of professors at City
University of New York and there are 10,000 professors, then we may take a
random sample of only 1,000 professors out of this entire population of 10,000
for the purpose of our study. Then this number of 1,000 professors constitutes a
sample. The summary measure that describes a characteristic such as average
income of this sample is known as a statistic.
Sampling is the process of selecting a sample from the population. It is
technically and economically not feasible to take the entire population for analysis.
So we must take a representative sample out of this population for the purpose of
such analysis. A sample is part of the whole, selected in such a manner as to be
representing the whole.
Random Sample
A random sample is a collection of items selected from the population in such a
manner that each item in the population has exactly the same chance of being
selected, so that the sample taken from the population would be truly representative
of the population. The degree of randomness of selection would depend upon the
process of selecting the items from the sample. A true random sample would be
free from all biases whatsoever. For example, if we want to take a random sample
of five students from a class of twenty-five students, then each one of these twenty-
five students should have the same chance of being selected into the sample. One
way to do this would be writing the names of all students on separate but small
pieces of paper, folding each piece of this paper in a similar manner, putting each
folded piece into a container, mixing them thoroughly and drawing out five pieces
of paper from this container.
Sampling without Replacement
The sample as taken in the above example is known as sampling without replacement,
as each person can only be selected once. This is because once a piece of paper is
taken out of the container, it is kept aside so that the person whose name appears on
this piece of paper has no chance of being selected again.
Sampling with Replacement
There are certain situations in which a piece of paper once selected and taken into
consideration is put back into the container in such a manner that the same person
has the same chance of being selected again as any other person. For example, if
we are randomly selecting five persons for award of prizes so that each person is
Self - Learning
eligible for any and all prizes, then once the slip of paper is drawn out of the Material 297
Estimation and container and the prize is awarded to the person whose name appears on the
Hypothesis Testing
paper, the same piece of paper is put back into the container and the same person
has the same chance of winning the second prize as anybody else.

NOTES Random Number Tables

For a sample to be truly representative of the population, it must truly be random.
To make the random selection easier, we can make use of tables of random numbers
which are generated by computers. A perfect random number table would be one
in which every digit has been entered randomly. This means that no matter where
you start within the table and no matter in which direction you move, the probability
of encountering any one of the ten digits (0, 1, 2,...9) would be the same. This
means that the chance of any one of these digits being at any place in the table is
exactly one out of ten. Similarly, if these digits are grouped in pairs (00, 01, 02,...99),
then each of these pairs has the same chance of occurring at any place so that
each pair would have a chance of occurring of one out of a hundred.
Table 5.1 illustrates a random number table.
Table 5.1 A Random Number Table
Column Number
1 2 3 4 5
1 81625 42372 07090 23422 10742
2 20891 27833 93079 16274 92818
3 62882 48722 39630 96434 09895
4 59882 84713 82521 29026 08591
Row Number

5 17932 14360 42933 89380 68191

6 67732 36772 09281 26898 30919
7 58198 87824 47958 04701 17369
8 57041 47778 02361 86939 61463
9 05264 49678 02067 58121 61822

10 84935 60407 16547 21359 58913

As an example of use of random number tables, let us assume that we have

to select a random sample from a finite population. The population cannot be
infinite due to the limitation as to how far the random numbers can go. Let there be
100 students in the population from which we have to draw a sample of five
students. Now we assign a two-digit number to each member of the population so
that each member is known as 00, 01, 02 ... 99. For selecting five students at
random from this population, we go to the random number table with groups of
two digits each and starting at any point and moving in any direction we pick the
five groups of numbers. Suppose that the numbers picked up are 07, 22, 23, 58
and 78. Then those members of the population to whom these numbers are assigned
constitute the random sample. In case we want to use a random number table in
which groups of five digits are arranged, as in Table 5.1, then we can use only the
first two digits or any two digits out of the five and reach the same conclusion of
Self - Learning
298 Material
randomness. In Table 5.1, suppose we pick row 5 and go across and pick up the
first two digits from each group of five, we get the following numbers: 17, 14, 42, Estimation and
Hypothesis Testing
89 and 68. Thus, those five members of the population to whom these numbers
are assigned constitute the random sample.
Sample Selection NOTES
Selecting an adequate sample is one of the steps in the primary data collection
process. It is necessary to take a representative sample from the population, since
it is extremely costly, time-consuming and cumbersome to do a complete census.
Then, depending upon the conclusions drawn from the study of the characteristics
of such a sample, we can draw inferences about the similar characteristics of the
population. If the sample is truly representative of the population, then the
characteristics of the sample can be considered to be the same as those of the
entire population. For example, the taste of soup in the entire pot of soup can be
determined by tasting one spoonful from the pot if the soup is well stirred. Similarly,
a small amount of blood sample taken from a patient can determine whether the
patient’s sugar level is normal or not. This is so because the small sample of blood
is truly representative of the entire blood supply in the body.
There are many reasons behind sampling. First, as discussed earlier, it is not
technically or economically feasible to take the entire population into consideration.
Second, due to dynamic changes in business, industrial and social environment, it
is necessary to make quick decisions based upon the analysis of information.
Managers seldom have the time to collect and process data for the entire population.
Thus, a sample is necessary to save time. The time element has further importance
in that if the data collection takes a long time, then the values of some characteristics
may change over the period of time so that data may no longer be up to date, thus
defeating the very purpose of data analysis. Third, samples, if representative, may
yield more accurate results than the total census. This is due to the fact that samples
can be more accurately supervised and data can be more carefully selected.
Additionally, because of the smaller size of the samples, the routine errors that are
introduced in the sampling process can be kept at a minimum. Fourth, the quality
of some products must be tested by destroying the products. For example, in
testing cars for their ability to withstand accidents at various speeds, the environment
of accidents must be simulated. Thus, a sample of cars must be selected and
subjected to accidents by remote control. Naturally, the entire population of cars
cannot be subjected to these accident tests and hence, a sample must be selected.
One important aspect to be considered is the size of the sample. The sampling
size—which is the number of sampling units selected from the population for
investigation—must be optimum. If the sample size is too small, it may not
appropriately represent the population or the universe as it is known, thus leading
to incorrect inferences. Too large a sample would be costly in terms of time and
money. The optimum sample size should fulfil the requirements of efficiency,
representativeness, reliability and flexibility. What is an optimum sample size is
also open to question. Some experts have suggested that 5 per cent of the
population properly selected would constitute an adequate sample, while others
have suggested as high as 10 per cent depending upon the size of the population
under study. However, proper selection and representation of the sample is more
Self - Learning
important than size itself. The following considerations may be taken into account Material 299
in deciding the sample size:
Estimation and (i) The larger the size of the population, the larger should be the sample size.
Hypothesis Testing
(ii) If the resources available do not put a heavy constraint on the sample size,
a larger sample would be desirable.
NOTES (iii) If the samples are selected by scientific methods, a larger sample size would
ensure greater degree of accuracy in conclusions.
(iv) A smaller sample could adequately represent the population, if the population
consists of mostly homogeneous units. A heterogeneous universe would
require a larger sample.
Census vs Sampling
Under the census or complete enumeration survey method, data is collected for
all units (e.g., person, consumer, employee, household, organization) of the
population or universe which are the complete set of entities and which are of
interest in any particular situation. In spite of the benefits of such an all-inclusive
approach, it is infeasible in most situations. Besides, the time and resource
constraints of the researcher, infinite or huge population, the incidental destruction
of the population unit during the evaluation process (as in the case of bullets,
explosives, etc.) and cases of data obsolescence (by the time census ends) do not
permit this mode of data collection.
Sampling is simply a process of learning about the population on the basis
of a sample drawn from it. Thus, in any sampling technique, instead of every unit
of the universe, only a part of the universe is studied and the conclusions are
drawn on that basis for the entire population. The process of sampling involves
selection of a sample based on a set of rules, collection of information and making
an inference about the population. It should be clear to the researcher that a sample
is studied not for its own sake, but the basic objective of its study is to draw
inference about the population. In other words, sampling is a tool which helps us
know the characteristics of the universe or the population by examining only a
small part of it. The values obtained from the study of a sample, such as the
average and dispersion are known as ‘statistics’ and the corresponding such values
for the population are called ‘parameters’.
Although diversity is a universal quality of mass data, every population has
characteristic properties with limited variation. The following two laws of statistics
are very important in this regard:
(i) The law of statistical regularity states that a moderately large number
of items chosen at random from a large group are almost sure on the
average to possess the characteristics of the large group. By random
selection, we mean a selection where each item of the population has
an equal chance of being selected.
(ii) The law of inertia of large numbers states that, other things being equal,
larger the size of the sample, more accurate the results are likely to
be.
Hence, a sound sampling procedure should result in a representative,
adequate and homogeneous sample while ensuring that the selection of items should
occur independently of one another.
Self - Learning
300 Material
Methods of Sampling Estimation and
Hypothesis Testing
The various methods of sampling can be grouped under two broad categories—
probability (or random) sampling and non-probability (or non-random) sampling.
Probability sampling methods are those in which every item in the universe NOTES
has a known chance, or probability of being chosen for the sample. Thus, the
sample selection process is objective (independent of the person making the study)
and hence, random. It is worth noting that randomness is a property of the sampling
procedure instead of an individual sample. As such, randomness can enter
processed sampling in a number of ways and hence, random samples may be of
many types. These methods include: (i) Simple random sampling, (ii) Stratified
random sampling, (iii) Systematic sampling, and (iv) Cluster sampling.
Non-probability sampling methods do not provide every item in the universe
with a known chance of being included in the sample. The selection process is, at
least, partially subjective (dependent on the person making the study). The most
important difference between random and non-random sampling is that whereas
the pattern of sampling variability can be ascertained in case of random sampling,
there is no way of knowing the pattern of variability in non-random sampling process.
The non-probability methods include: (i) Judgement sampling, (ii) Quota sampling,
and (iii) Convenience sampling.
As shown in Figure 5.1 depicts the broad classification and sub-classification
of various methods of sampling.
Sampling Methods

Non-probability Probability
Samples Samples

Judgement Quota Convenience

Sampling Sampling Sampling

Simple Random Stratified Systematic Cluster

Sampling Sampling Sampling Sampling

Fig. 5.1 Methods of Sampling

Non-Probability Sampling Methods

The following are the non-probability sampling methods:
(i) Judgement Sampling
In judgement sampling, the choice of sample items depends exclusively on the
judgement of the investigator. The sample here is based on the opinion of the
researcher, whose discretion will clinch the sample. Though the principles of
sampling theory are not applicable to judgement sampling, it is sometimes found to
be useful. When we want to study some unknown traits of a population, some of
whose characteristics are known, we may then stratify the population according
to these known properties and select sampling units from each stratum on the
basis of judgement. Naturally, the success of this method depends upon the Self - Learning
excellence in judgement. Material 301
Estimation and (ii) Convenience Sampling
Hypothesis Testing
A convenience sample is obtained by selecting convenient population units. It is
also called a chunk, which refers to that fraction of the population being investigated
NOTES which is selected neither by probability nor by judgement, but by convenience. A
sample obtained from readily available lists, such as telephone directories is a
convenience sample and not a random sample, even if the sample is drawn at
random from such lists. In spite of the biased nature of such a procedure,
convenience sampling is often used for pilot studies.
(iii) Quota Sampling
Quota sampling is a type of judgement sampling and is perhaps the most commonly
used sampling technique in non-probability category. In a quota sample, quotas
(or minimum targets) are set up according to some specified characteristics, such
as age, income group, religious or political affiliations, and so on. Within the quota,
the selection of the sample items depends on personal judgement. Because of the
risk of personal prejudice entering the sample selection process, quota sampling is
not widely used in practical works.
It is worth noting that similarity between quota sampling and stratified random
sampling is confined to dividing the population into different strata. The process
of selecting items from each of these strata in the case of stratified random sampling
is random, while it is not so in the case of quota sampling. Quota sampling is often
used in public opinion studies.
Probability Sampling Methods
The following are the probability sampling methods:
(i) Simple Random Sampling
In simple random sampling each unit of the population has an equal chance of
being selected in the sample. One should not mistake the term ‘Arbitrary’ for
‘Random’. To ensure randomness, one may adopt either the lottery method or
consult the table of random numbers, preferably the latter. Being a random method,
it is independent of personal bias creeping into the analysis besides enhancing the
representativeness of the sample. Furthermore, it is easy to assess the accuracy
of the sampling estimates because sampling errors follow the principles of chance.
However, a completely catalogued universe is a prerequisite for this method. The
sample size requirements would be usually larger under random sampling than
under stratified random sampling, to ensure statistical reliability. It may escalate
the cost of collecting data as the cases selected by random sampling tend to be
too widely dispersed geographically.
(ii) Stratified Random Sampling
In stratified random sampling, the universe to be sampled is subdivided (Stratified)
into groups which are mutually exclusive, but collectively exhaustive based on a
variable known to be correlated with the variable of interest. Then, a simple random
sample is chosen independently from each group. This method differs from simple
random sampling in that, in the latter the sample items are chosen at random from
the entire universe. In stratified random sampling, the sampling is designed in such
Self - Learning
302 Material a way that a designated number of items is chosen from each stratum. If the ratio
of items between various strata in the population matches with the ratio of Estimation and
Hypothesis Testing
corresponding items between various strata in the sample, it is called proportionate
stratified sampling; otherwise, it is known as disproportionate stratified sampling.
Ideally, we should assign greater representation to a stratum with a larger dispersion
and smaller representation to one with small variation. Hence, it results in a more NOTES
representative sample than simple random sampling.
(iii) Systematic Sampling
Systematic sampling is also known as quasi-random sampling method because
once the initial starting point is determined, the remainder of the items selected for
the sample are predetermined by the sampling interval. A systematic sample is
formed by selecting one unit at random and then selecting additional units at evenly
spaced interval until the sample has been formed. This method is popularly used in
cases where a complete list of the population from which sample is to be drawn is
available. The list may be prepared in alphabetical, geographical, numerical or
some other order. The items are serially numbered. The first item is selected at
random generally by following the lottery method. The subsequent items are selected
by taking every Kth item from the list where ‘K’ stands for the sampling interval or
the sampling ratio, i.e., the ratio of the population size to the size of the sample.
Symbolically,
K = N / n , where K = Sampling interval; N = Universe size; n = Sample size.
In case K is a fractional value, it is rounded off to the nearest integer.

(iv) Multistage or Cluster Sampling

In multistage or cluster sampling, the primary, intermediate and final (or the ultimate)
units are randomly selected from a given population or stratum. There are several
stages in which the sampling process is carried out. At first, the stage units are
sampled by some suitable method, such as simple random sampling. Then, a
sample of second stage units is selected from each of the selected first stage units,
by applying some suitable method which may or may not be the same method
employed for the first stage units. For example, in a survey of 10,000 households
in AP, we may choose a few districts in the first stage, a few towns/villages/mandals
in the second stage and select a number of households from each town/village/
mandal selected in the previous stage. This method is quite flexible and is
particularly useful in surveys of underdeveloped areas, where no frame is generally
sufficiently detailed and accurate for subdivision of the material into reasonably
small sampling units. However, a multistage sample is, in general, less accurate
than a sample containing the same number of final stage units which have been
selected by some suitable single stage process.
Sampling and Non-Sampling Errors
The basic objective of a sample is to draw inferences about the population from
which such sample is drawn. This means that sampling is a technique which helps
us in understanding the parameters or the characteristics of the universe or the
population by examining only a small part of it. Therefore, it is necessary that the
sampling technique be a reliable one. The randomness of the sample is especially
important because of the principle of statistical regularity, which states that a sample Self - Learning
taken at random from a population is likely to possess almost the same characteristics Material 303
Estimation and as those of the population. However, in the total process of statistical analysis,
Hypothesis Testing
some errors are bound to be introduced. These errors may be the sampling errors
or the non-sampling errors. The sampling errors arise due to drawing faulty
inferences about the population based upon the results of the samples. In other
NOTES words, it is the difference between the results that are obtained by the sample
study and the results that would have been obtained if the entire population was
taken for such a study, provided that the same methodology and manner was
applied in studying both the sample as well as the population. For example, if a
sample study indicates that 25 per cent of the adult population of a city does not
smoke and the study of the entire adult population of the city indicates that 30 per
cent are non-smokers, then this difference would be considered as the sampling
error. This sampling error would be smallest if the sample size is large relative to
the population, and vice versa.
Non-sampling errors, on the other hand, are introduced due to technically
faulty observations during the processing of data. These errors could also arise
due to defective methods of data collection and incomplete coverage of the
population, because some units of the population are not available for study,
inaccurate information provided by the participants in the sample, and errors
occurring during editing, tabulating and mathematical manipulation of data. These
errors can arise even when the entire population is taken under study.
Both the sampling as well as the non-sampling errors must be reduced to a
minimum in order to get as representative a sample of the population as possible.
Parameter and Statistics
Parameter
A parameter is a numeric quantity that describes a certain population characteristic.
Parameters are in general represented by Greek letters. The most common
parameters are the population mean and variance, represented by the Greek letters
 and 2, respectively. For example, the population mean is a parameter that is
often used to indicate the average value of a quantity.
Parameters are often estimated, in view of the fact that their value is generally
unknown, especially when the population is large enough that it is impossible or
impractical to obtain measurements for all people.
Statistics
A statistic is a quantity, calculated from a sample of data, used to estimate a
parameter. For example, the average of the data in a sample is used to give
information about the overall average in the population from which that sample
was drawn. Statistics is usually represented by Latin letters with other symbols.
The sample mean and variance, two of the most common statistics derived from
samples, are denoted by the symbols X n and 2, respectively..
It is possible to draw more than one sample from the same population, and
each sample will have its own value for any statistic used to estimate a particular
parameter. For example, the mean of the data in a sample is used to provide
information about the overall mean in the population from which that sample was
Self - Learning
304 Material
drawn. However, the sample means for two independent samples, drawn from
the same population, will not necessarily be equal. Each sample mean is still an Estimation and
Hypothesis Testing
estimate of the underlying population mean.

Check Your Progress

NOTES
1. What would you use to test the validity of hypothesis?
2. What is one very important aspect of sampling theory?
3. Define a sample.
4. What are the two types of sampling?
5. Define non-probability sampling.
6. What is probability sampling?
7. Name the probability sampling method.

5.4 THE STUDENT’S DISTRIBUTION

One of the major objectives of statistical analysis is to know the ‘True’ values of
different parameters of the population. Since it is not possible due to time, cost
and other constraints to take the entire population for consideration, random samples
are taken from the population. These samples are analysed properly and they lead
to generalizations that are valid for the entire population. The process of relating
the sample results to population is referred to as, ‘Statistical Inference’ or ‘Inferential
Statistics’.
In general, a single sample is taken and its mean X is considered to represent
the population mean. However, in order to use the sample mean to estimate the
population mean, we should examine every possible sample (and its mean, etc.)
that could have occurred, because a single sample may not be representative
enough. If it was possible to take all the possible samples of the same size, then the
distribution of the results of these samples would be referred to as, ‘Sampling
Distribution’. The distribution of the means of these samples would be referred to
as, ‘Sampling Distribution; of the Means’.
The relationship between the sample means and the population mean can
best be illustrated by Example 5.1.
Example 5.1: Suppose a babysitter has 5 children under her supervision with
average age of 6 years. However, individually, the age of each child be as follows:
X1 = 2
X2 = 4
X3 = 6
X4 = 8
X 5 = 10
Now these 5 children would constitute our entire population, so that N = 5.
Solution:
X Self - Learning
The population mean μ 
N Material 305
Estimation and
2  4  6  8  10
Hypothesis Testing =  30 / 5  6
5
and the standard deviation is given by the formula:
NOTES  (   μ) 2
σ
N

Now, let us calculate the standard deviation.

X  (X–)2
2 6 16
4 6 4
6 6 0
8 6 4
10 6 16

Total = ( X – )2 = 40

Then,

40
σ  8  2.83
5
Now, let us assume the sample size, n = 2, and take all the possible samples of
size 2, from this population. There are 10 such possible samples. These are as
follows, along with their means.
X1, X 2 (2, 4) X1 = 3
X1, X 3 (2, 6) X2 = 4
X1, X 4 (2, 8) X3 = 5
X1, X 5 (2,10) X4 = 6
X2, X 3 (4, 6) X5 = 5
X2, X 4 (4, 8) X6 = 6
X2, X 5 (4, 10) X7 = 7
X3, X 4 (6, 8) X8 = 7
X3, X 5 (6, 10) X9 = 8
X4, X 5 (8, 10) X 10 = 9
Now, if only the first sample was taken, the average of the sample would be
3. Similarly, the average of the last sample would be 9. Both of these samples are
totally unrepresentative of the population. However, if a grand mean X of the
distribution of these sample means is taken, then,
10

X i
X  i 1

10
3 4 5 6567 7 89
Self - Learning  60 / 10  6
306 Material 10
 This grand mean has the same value as the mean of the population. Let us Estimation and
Hypothesis Testing
organize this distribution of sample means into a frequency distribution and
probability distribution.
Sample mean Freq. Rel.freq. Prob.
NOTES
3 1 1/10 .1
4 1 1/10 .1
5 2 2/10 .2
6 2 2/10 .2
7 2 2/10 .2
8 1 1/10 .1
9 1 1/10 .1
1.00
This probability distribution of the sample means is referred to as ‘sampling
distribution of the mean.’
Sampling Distribution of the Mean
The sampling distribution of the mean can thus be defined as, ‘A probability
distribution of all possible sample means of a given size, selected from a population’.
Accordingly, the sampling distribution of the means of the ages of children as
tabulated in Example 5.1, has 3 predictable patterns. These are as follows:
(i) The mean of the sampling distribution and the mean of the population are
equal. This can be shown as follows:
Sample mean ( X ) Prob. P( X )
3 .1
4 .1
5 .2
6 .2
7 .2
8 .1
9 .1
1.00
Then,
   XP ( X ) = (3 × .l) + (4 × .l) + (5×.2) + (6 × .2) + (7 × .2) + (8 × .l)
+ 9 ×.l) = 6
This value is the same as the mean of the original population.
(ii) The spread of the sample means in the distribution is smaller than in the
population values. For example, the spread in the distribution of sample
means above is from 3 to 9, while the spread in the population was from 2
to 10.
(iii) The shape of the sampling distribution of the means tends to be, ‘Bell-
shaped’ and approximates the normal probability distribution, even when
the population is not normally distributed. This last property leads us to
Self - Learning
the ‘Central Limit Theorem’. Material 307
Estimation and Central Limit Theorem
Hypothesis Testing
Central Limit Theorem states that, ‘Regardless of the shape of the population, the
distribution of the sample means approaches the normal probability distribution as
NOTES the sample size increases.’
The question now is how large should the sample size be in order for the
distribution of sample means to approximate the normal distribution for any type
of population. In practice, the sample sizes of 30 or larger are considered adequate
for this purpose. This should be noted however, that the sampling : distribution
would be normally distributed, if the original population is normally distributed, no
matter what the sample size.
As we can see from our sampling distribution of the means, the grand mean
X of the sample means or – × equals , the population mean. However,,
realistically speaking, it is not possible to take all the possible samples of size n
from the population. In practice only one sample is taken, but the discussion on
the sampling distribution is concerned with the proximity of ‘a’ sample mean to the
population mean.
It can be seen that the possible values of sample means tend towards the
population mean, and according to Central Limit Theorem, the distribution of
sample means tend to be normal for a sample size of n being larger than 30.
Hence, we can draw conclusions based upon our knowledge about the
characteristics of the normal distribution.
For example, in the case of sampling distribution of the means, if we know
the grand mean –×of this distribution, which is equal to p, and the standard
deviation of this distribution, known as ‘Standard error of free mean’ and denoted
by –× , then we know from the normal distribution that there is a 68.26 per cent
chance that a sample selected at random from a population, will have a mean that
lies within one standard error of the mean of the population mean. Similarly, this
chance increases to 95.44 per cent, that the sample mean will lie within two standard
errors of the mean (–× ) of the population mean. Hence, knowing the properties
of the sampling distribution tells us as to how close the sample mean will be to the
true population mean.
Standard Error
Standard error of the mean (–×)
Standard error of the mean (–×)is a measure of dispersion of the distribution of
sample means and is similar to the standard deviation in a frequency distribution
and it measures the likely deviation of a sample mean from the grand mean of the
sampling distribution.
If all sample means are given, then (–×)can be calculated as follows:

( –  ) where N = Number of sample means

 
N
Thus we can calculate –× for Example 5.1 of the sampling distribution of
the ages of 5 children as follows:
Self - Learning
308 Material
 Estimation and
X (μ  ) ( X  μ  )2 Hypothesis Testing
3 6 9
4 6 4
5 6 1 NOTES
6 6 0
7 6 1
8 6 4
9 6 9
 ( X  μ  ) 2 = 28
Then,
 ( –   )
 
N

28

7
 42
However, since it is not possible to take all possible samples from the
population, we must use alternate methods to compute –× .
The standard error of the mean can be computed from the following formula,
if the population is finite and we know the population mean. Hence,

σ ( N  n)
σ 
n ( N  1)

Where,
 = population standard deviation
N = population size
n = sample size
This formula can be made simpler to use by the fact that we generally deal
with very large populations, which can be considered infinite, so that if the population
size A’ is very large and sample size n is small, as for example in the case of items
tested from assembly line operations, then,
( N – n)
would approach 1.
( N –1)

Hence,

 
n
( N – n)
The factor ( N – n)
is also known as the ‘finite correction factor’, and should be
used when the population size is finite.
As this formula suggests, –× decreases as the sample size (w) increases,
meaning that the general dispersion among the sample means decreases, meaning Self - Learning
Material 309
Estimation and further that any single sample mean will become closer to the population mean, as
Hypothesis Testing
the value of (–×) decreases. Additionally, since according to the property of the
normal curve, there is a 68.26 per cent chance of the population mean being
within one –× of the sample mean, a smaller value of –× will make this range
NOTES shorter; thus making the population mean closer to the sample mean
(Refer Example 5.2).
Example 5.2: The IQ scores of college students are normally distributed with the
mean of 120 and standard deviation of 10.
(a) What is the probability that the IQ score of any one student chosen at
random is between 120 and 125?
(b) If a random sample of 25 students is taken, what is the probability that the
mean of this sample will be between 120 and 125.
Solution:
(a) Using the standardized normal distribution formula,

125
 = 120
 = 10

( X – )
Z


125 –120
Z  5 / 10  .5
10
The area for Z = .5 is 19.15.
This means that there is a 19.15 per cent chance that a student picked up at
random will have an IQ score between 120 and 125.
(b) With the sample of 25 students, it is expected that the sample mean will be
much closer to the population mean, hence it is highly likely that the sample
mean would be between 120 and 125.
The formula to be used in the case of standardized normal distribution for
sampling distribution of the means is given by,
X –
Z

where,

 
Self - Learning n
310 Material
Hence, Estimation and
Hypothesis Testing

NOTES

125
 =120
 =10
X –
Z

 10 10
where,     2
n 25 5
Then,
125 –120
Z  5 / 2  2.5
2
The area for Z = 2.5 is 49.38.
This shows that there is a chance of 49.38 per cent that the sample mean
will be between 120 and 125. As the sample size increases further, this chance will
also increase. It can be noted that the probability of a sample mean being between
120 and 125 is much higher than the probability of an individual student having an
IQ between 120 and 125.

5.5 THEORY OF ESTIMATION

The best estimator should be highly reliable and have such desirable properties as
unbiasedness, consistency, efficiency and sufficiency. These criteria are described
as follows:
(i) Unbiasedness: An estimator is a random variable since it is always a
function of the sample values. For example, the value of the sample average
would depend upon the values of the sample and may differ from sample to
sample. The expected value of the sample average is considered to be an
unbiased estimator if it equals the population mean, which is being estimated.
This means that:
E (X)  
(Since sampling distribution is a probability distribution, we refer to the
average, as expected value instead of simply the average).
(ii) Consistency: This refers to the effect of the sample size on the accuracy of
the estimator. A statistic is said to be consistent estimator of the population
parameter, if it approaches the parameter as the sample size increases, so
that in the case of the mean:
X   as n  N Self - Learning
Material 311
Estimation and (iii) Efficiency: An estimator is considered to be efficient if its value remains
Hypothesis Testing
stable from sample to sample. The best estimator would be the one which
would have the least variance from sample to sample taken randomly from
the same population. From the three point estimators of central tendency,
NOTES namely the mean, the mode and the median, the mean is considered to be
the least variant and hence, a better estimator.
(iv) Sufficiency: An estimator is said to be sufficient if it uses all the information
about the population parameter contained in the sample. For example, the
statistic mean uses all the sample values in its computation, while the mode
and the median do not. Hence, the mean is a better estimator in this sense.
Some of the parameters of the population and their estimators are as follows:

 =X  
X
n

 s
(X  X ) 2

n 1
X
p  ps    , where ps is the sample proportion.
n

5.5.1 Point Estimation

The theory of estimation is a very commonly used and popular statistical method
and is used to calculate the mathematical model for the data to be considered.
This method was introduced by the statistician Sir R. A. Fisher, between 1912
and 1922. This method can be used in the following:
 Finding linear models and generalized linear models.
 Exploratory and confirmatory factor analysis.
 Structural equation modelling.
 Calculating Time-Delay of Arrival (TDOA) in acoustic or electromagnetic
detection.
 Data modelling in nuclear and particle physics.
 Finding the result for hypothesis testing.
The method of estimation is used with known mean and variance. The sample
mean becomes the maximum likelihood estimator of the population mean, and the
sample variance becomes the close approximation to the maximum likelihood
estimator of the population variance.
A point estimate uses a single sample value to estimate the desired population
parameter. For example, a sample mean X is considered as a point estimate of the
population mean . Similarly, a sample standard deviation s, is a point estimate of
population standard deviation . For instance, if we want to know the Grade Point
Average (GPA) of seniors majoring in Business Administration at Medgar Evers
College, then we take a random sample of business major seniors and calculate the
sample mean X of the sample. Then, the value of this X would be considered as a
point estimate of  which is the grade point average of the entire population of students
Self - Learning
312 Material
majoring in business administration. Similarly, the sample variance s2 is the point Estimation and
Hypothesis Testing
estimate of the population variance 2.
In point estimate, we seek the sample statistic, such as X , computed from
sample observations, which is the best estimate of the corresponding population
parameter, such as . But, how do we know that the sample statistic that we computed NOTES
from sample observations is the best estimator of the population parameter? By
best we mean that the value of the sample statistic should be as close to the population
parameter as possible. For example, if the sample mean grade point average for
business students is calculated as 3.5 out of 4, then the population average grade
point average should also be 3.5 or very close to, it in order for sample average to be
a good estimator of population average. Since the population parameter is always
inferred from sample statistic, it is necessary and important that such sample statistic
should be as highly reliable as an estimator for population parameter, as possible.
For example, there are three measures of central tendency, namely mean, mode and
median for a sample that can be used as point estimators for the population average.
It is important to know as to which one of these measures best represents the
population mean. As an illustration, suppose that we want to find out the average
time that a salesman of a company spends with the customer. Suppose further that
we took a sample and found out that on an average, a salesman spent 60 minutes
with a customer (mean). However, most salesmen spent 45 minutes (the mode) and
the median was 65 minutes. The question now is to establish as to which of these
measures would best describe the population parameter as to how much time on an
average a salesman spends with the customer?

5.5.2 Interval Estimation

Point estimator, though simplistic in nature, has some drawbacks. First, a point
estimator from the sample may not exactly locate the population parameter resulting
in some margin of uncertainty. The average of a sample for example, may or may
not be equal or close to the average of the population. If the sample average is
different from the population average, the point estimator does not indicate the
extent of the possible error, even though this error can be reduced by increasing
the sample size. Second, a point estimate does not specify as to how confident we
can be that the estimate is close to the parameter it is estimating.
To reasonably overcome these drawbacks, statisticians use another type of
estimation known as interval estimation. In this method, we first find a point estimate.
Then we use this estimate to construct an interval on both sides of the point estimate,
within which we can be reasonably confident that the true parameter will lie. For
example, suppose that we wanted to find out the average salary of full professors
at a university who had served at least five years at that rank. Suppose further, that
a random sample was taken and the average of the sample was computed to be
$55,000. It is quite possible that the actual average salary of all university professors
is $55,000. However, it is equally possible that the sample was not true
representative of the population and the average of the population is quite far off
the average of the sample. Accordingly, it is much more likely that the average
salary of all the professors lies somewhere, let us say, between $50,000 and $60,000
than exactly at $55,000. Of course, the greater the range of interval around the
sample mean, the more likely it is that the population mean lies in that range. This Self - Learning
degree of likelihood is known as the confidence level and the range around the Material 313
Estimation and sample mean is known as the confidence interval at a given confidence level. (It is,
Hypothesis Testing
of course, assumed that the sample is large enough so that the Central Limit Theorem
holds.)

NOTES Interval estimate of the population mean (Population variance known)

Since the sample means are normally distributed, with a mean of  and a standard
deviation of  X , it follows that sample means follow normal distribution characteristics.
Transforming the sampling distribution of sample means into the standard normal
distribution, we get:

X 
Z
X
or X   ZX
or   X  ZX

Since falls within a range of values equidistant from X ,

  X  ZX

This relationship is shown in the following illustration:

Zx Zx

X1 X X2


This means that the population mean is expected to lie between the values
of X1 and X2 which are both equidistant from X and this distance depends upon
the value of Z which is a function of confidence level.
Suppose that we wanted to find out a confidence interval around the
sample mean within which the population mean is expected to lie 95 per cent of
the time. (We can never be sure that the population mean will lie in any given
interval 100 per cent of the time). This confidence interval is shown in the following
illustration:
95%
47.5% 47.5%

2.5% 2.5%

X1 X X2

Self - Learning
314 Material
 The points X1 and X2 above define the range of the confidence interval as Estimation and
Hypothesis Testing
follows:

X1  X  ZX
and X2  X  ZX NOTES

Looking at the table of Z scores, (given in the Appendix) we find that the
value of Z score for area 10.4750 (half of 95 per cent) is 1.96. This illustration can
be interpreted as follows:
(i) If all possible samples of size n were taken, then on the average 95 per cent
of these samples would include the population mean within the interval around
their sample means bounded by X1 and X2.
(ii) If we took a random sample of size n from a given population, the probability
is 0.95 that the population mean would lie between the interval X1 and X2
around the sample mean, as shown.
(iii) If a random sample of size n was taken from a given population, we can be
95 per cent confident in our assertion that the population mean will lie around
the sample mean in the interval bounded by values of X1 and X2 as shown.
(It is also known as 95 per cent confidence interval.) At 95 per cent
confidence interval, the value of Z score as taken from the Z score table is
1.96. The value of Z score can be found for any given level of confidence,
but generally speaking, a confidence level of 90 per cent, 95 per cent or 99
per cent is taken into consideration for which the Z score values are 1.68,
1.96 and 2.58, respectively.
Refer Examples 5.3 and 5.4 to understand internal estimation better.
Example 5.3: The sponsor of a television programme targeted at the children’s
market (age 4-10 years) wants to find out the average amount of time children
spend watching television. A random sample of 100 children indicated the average
time spent by these children watching television per week to be 27.2 hours. From
previous experience, the population standard deviation of the weekly extent of
television watched () is known to be 8 hours. A confidence level of 95 per cent is
considered to be adequate.
Solution:

1.96X 1.96X

X1 X = 27.2 X2
= 8
The confidence interval is given by,

X  ZX or X  ZX  X  ZX


where  X 
n Self - Learning
Material 315
Estimation and
Hypothesis Testing
Accordingly, we need only four values, namely X , Z,  and n. In our case:

X  27.2
Z  1.96
NOTES
 8
n  100

 8 8
Hence X     0.8
n 100 10
Then,
X1  X  Z X
 27.2  (1.96  0.8)  27.2 1.568
 25.632

and
X 2  X  Z X
 27.2  (1.96  0.8)  27.2  1.568
 28.768

This means that we can conclude with 95 per cent confidence that a child on an
average spends between 25.632 and 28.768 hours per week watching television. (It
should be understood that 5 per cent of the time our conclusion would still be wrong.
This means that because of the symmetry of distribution, we will be wrong 2.5 per
cent of the times because the children on an average would be watching television
more than 28.768 hours and another 2.5 per cent of the time we will be wrong in our
conclusion, because on an average, the children will be watching television less than
25.632 hours per week.)
Example 5.4: Calculate the confidence interval in the previous problem, if we
want to increase our confidence level from 95 per cent to 99 per cent. Other values
remain the same.
Solution:

.495 .495

.005 .005
2.58X 2.58X

X1 X = 27.2 X2
= 8
If we increase our confidence level to 99 per cent, then it would be natural to
assume that the range of the confidence interval would be wider, because we would
want to include more values which may be greater than 28.768 or smaller than
25.632 within the confidence interval range. Accordingly, in this new situation,
Z  2.58
Self - Learning
316 Material X  0.8
Then Estimation and
Hypothesis Testing
X1  X  Z X
 27.2 – (2.58  0.8)  27.2  2.064
 25.136
NOTES
and
X 2  X  Z X
 27.2  2.064
 29.264
(The value of Z is established from the table of Z scores against the area of 0.495 or
a figure closest to it. The table shows that the area close to 0.495 is 0.4949 for
which the Z score is 2.57 or 0.4951 for which the Z score is 2.58. In practice, the Z
score of 2.58 is taken into consideration when calculating 99 per cent confidence
interval.)

5.6 HYPOTHESIS TESTING

A hypothesis is an approximate assumption that a researcher wants to test for its
logical or empirical consequences. Hypothesis refers to a provisional idea whose
merit needs evaluation, but has no specific meaning, though it is often referred as a
convenient mathematical approach for simplifying cumbersome calculation. Setting
up and testing hypothesis is an integral art of statistical inference. Hypotheses are
often statements about population parameters like variance and expected value.
During the course of hypothesis testing, some inference about population like the
mean and proportion are made. Any useful hypothesis will enable predictions by
reasoning including deductive reasoning. According to Karl Popper, a hypothesis
must be falsifiable and that a proposition or theory cannot be called scientific if it
does not admit the possibility of being shown false. Hypothesis might predict the
outcome of an experiment in a lab, setting the observation of a phenomenon in
nature. Thus, hypothesis is an explanation of a phenomenon proposal suggesting a
possible correlation between multiple phenomena.
The characteristics of hypothesis are as follows:
 Clear and accurate: Hypothesis should be clear and accurate so as to
draw a consistent conclusion.
 Statement of relationship between variables: If a hypothesis is relational,
it should state the relationship between different variables.
 Testability: A hypothesis should be open to testing so that other deductions
can be made from it and can be confirmed or disproved by observation.
The researcher should do some prior study to make the hypothesis a testable
one.
 Specific with limited scope: A hypothesis, which is specific, with limited
scope, is easily testable than a hypothesis with limitless scope. Therefore, a
researcher should pay more time to do research on such kind of hypothesis.
 Simplicity: A hypothesis should be stated in the most simple and clear Self - Learning
terms to make it understandable. Material 317
Estimation and  Consistency: A hypothesis should be reliable and consistent with
Hypothesis Testing
established and known facts.
 Time limit: A hypothesis should be capable of being tested within a
reasonable time. In other words, it can be said that the excellence of a
NOTES
hypothesis is judged by the time taken to collect the data needed for the
test.
 Empirical reference: A hypothesis should explain or support all the
sufficient facts needed to understand what the problem is all about.
A hypothesis is a statement or assumption concerning a population. For the
purpose of decision-making, a hypothesis has to be verified and then accepted or
rejected. This is done with the help of observations. We test a sample and make a
decision on the basis of the result obtained. Decision-making plays a significant
role in different areas such as marketing, industry and management.
Statistical Decision-Making
Testing a statistical hypothesis on the basis of a sample enables us to decide whether
the hypothesis should be accepted or rejected. The sample data enables us to
accept or reject the hypothesis. Since the sample data gives incomplete information
about the population, the result of the test need not be considered to be final or
unchallengeable. The procedure, on the basis of which sample results, enables to
decide whether a hypothesis is to be accepted or rejected. This is called Hypothesis
Testing or Test of Significance.
Note 1: A test provides evidence, if any, against a hypothesis, usually called a null
hypothesis. The test cannot prove the hypothesis to be correct. It can give some
evidence against it.
The test of hypothesis is a procedure to decide whether to accept or reject a hypothesis.
Note 2: The acceptance of a hypotheses implies, if there is no evidence from the sample that
we should believe otherwise.
The rejection of a hypothesis leads us to conclude that it is false. This way
of putting the problem is convenient because of the uncertainty inherent in the
problem. In view of this, we must always briefly state a hypothesis that we hope to
reject.
A hypothesis stated in the hope of being rejected is called a null hypothesis
and is denoted by H0.
If H0 is rejected, it may lead to the acceptance of an alternative hypothesis
denoted by H1.
For example, a new fragrance soap is introduced in the market. The null
hypothesis H0, which may be rejected, is that the new soap is not better than the
existing soap.
Similarly, a dice is suspected to be rolled. Roll the dice a number of times to
test.
The null hypothesis H0: p = 1/6 for showing six.
The alternative hypothesis H1: p  1/6.
Self - Learning
318 Material
For example, skulls found at an ancient site may all belong to race X or race Y on Estimation and
Hypothesis Testing
the basis of their diameters. We may test the hypothesis, that the mean is  of the
population from which the present skulls came. We have the hypotheses.
H0 :  =  x, H1 :  =  y NOTES
Here, we should not insist on calling either hypothesis null and the other alternative
since the reverse could also be true.
Committing Errors: Type I and type II
Types of Errors
There are two types of errors in statistical hypothesis, which are as follows:
 Type I error: In this type of error, you may reject a null hypothesis when it
is true. It means rejection of a hypothesis, which should have been accepted.
It is denoted by  (alpha) and is also known alpha error.
 Type II error: In this type of error, you are supposed to accept a null
hypothesis when it is not true. It means accepting a hypothesis, which should
have been rejected. It is denoted by  (beta) and is also known as beta
error.
Type I error can be controlled by fixing it at a lower level. For example, if
you fix it at 2 per cent, then the maximum probability to commit Type I error
is 0.02. However, reducing Type I error has a disadvantage when the sample
size is fixed, as it increases the chances of Type II error. In other words, it
can be said that both types of errors cannot be reduced simultaneously. The
only solution of this problem is to set an appropriate level by considering
the costs and penalties attached to them or to strike a proper balance between
both types of errors.
In a hypothesis test, a Type I error occurs when the null hypothesis is rejected
when it is in fact true; that is, H0 is wrongly rejected. For example, in a clinical trial
of a new drug, the null hypothesis might be that the new drug is no better, on
average, than the current drug; that is H0: there is no difference between the two
drugs on average. A Type I error would occur if we concluded that the two drugs
produced different effects, when in fact there was no difference between them.
In a hypothesis test, a Type II error occurs when the null hypothesis H0 is
not rejected, when it is in fact false. For example, in a clinical trial of a new drug,
the null hypothesis might be that the new drug is no better, on average, than the
current drug; that is H0: there is no difference between the two drugs on average.
A Type II error would occur if it were concluded that the two drugs produced the
same effect, that is, there is no difference between the two drugs on average,
when in fact they produced different ones.
In how many ways can we commit errors?
We reject a hypothesis when it may be true. This is Type I Error.
We accept a hypothesis when it may be false. This is Type II Error.
The other true situations are desirable: We accept a hypothesis when it is true. We
reject a hypothesis when it is false. Self - Learning
Material 319
Estimation and
Accept H0 Reject H0
Hypothesis Testing
H0 Accept True H0 Reject True H0
Desirable Type I Error
True
NOTES H1 Accept False H0 Reject False H0
Type II Error Desirable
False

The level of significance implies the probability of Type I error. A 5 per cent level
implies that the probability of committing a Type I error is 0.05. A 1 per cent level
implies 0.01 probability of committing Type I error.
Lowering the significance level and hence the probability of Type I error is good
but unfortunately, it would lead to the undesirable situation of committing Type II
error.
To Sum Up:
 Type I Error: Rejecting H0 when H0 is true.
 Type II Error: Accepting H0 when H0 is false.
Note: The probability of making a Type I error is the level of significance of a statistical test.
It is denoted by 

Where,  = Prob. (Rejecting H0 / H0 true)

1– = Prob. (Accepting H0 / H0 true)
The probability of making a Type II error is denoted by .
Where,  = Prob. (Accepting H0 / H0 false)
1– = Prob. (Rejecting H0 / H0 false) = Prob. (The test correctly rejects H0
when H0 is false)
1– is called the power of the test. It depends on the level of significance , sample size
n and the parameter value.

5.6.1 Test of Hypothesis Concerning Mean and

Proportion
Test of Significance

Tests for a sample mean X

We have to test the null hypothesis that the population mean has a specified value
µ, i.e., H0: X = µ. For large n, if H0 is true then,

X 
z is approximately nominal. The theoretical region for z depending on
SE ( X )
the desired level of significance can be calculated.
For example, a factory produces items, each weighing 5 kg with variance 4. Can
a random sample of size 900 with mean weight 4.45 kg be justified as having been
Self - Learning
taken from this factory?
320 Material
 n = 900 Estimation and
Hypothesis Testing
X = 4.45
µ=5
= NOTES
4 =2
X  X   4.45  5
z = = = 8.25
SE ( X )  / n 2 / 30

We have z > 3. The null hypothesis is rejected. The sample may not be regarded
as originally from the factory at 0.27 per cent level of significance (corresponding
to 99.73 per cent acceptance region).
Test for equality of two proportions
If P1, P2 are proportions of some characteristic of two samples of sizes n1, n2,
drawn from populations with proportions P1, P2, then we have H0: P1 = P2 vs
H1:P1 ¹ P2 .
 Case (I): If H0 is true, then let P1 = P2 = p
Where, p can be found from the data,
n1 P1  n2 P2
p
n1  n2
q  1 p

p is the mean of the two proportions.

1 1
SE ( P1  P2 )  pq   
 n1 n2 
P1  P2
z , P is approximately normal (0,1)
SE ( P1  P2 )

We write z ~ N(0, 1)
The usual rules for rejection or acceptance are applicable here.
 Case (II): If it is assumed that the proportion under question is not the same in
the two populations from which the samples are drawn and that P1, P2 are the true
proportions, we write,

 Pq P q 
SE ( P1  P2 )   1 1  2 2 
 n1 n2 

We can also write the confidence interval for P1 – P2.

For two independent samples of sizes n1, n2 selected from two binomial
populations, the 100 (1 – a) per cent confidence limits for P1 – P2 are,

 Pq P q 
( P1  P2 )  z / 2  1 1  2 2 
 n1 n2 
The 90% confidence limits would be [with a = 0.1, 100 (1 – a) = 0.90] Self - Learning
Material 321
Estimation and
Hypothesis Testing  Pq P q 
( P1  P2 )  1.645  1 1  2 2 
 n1 n2 

NOTES Consider Example 5.5 to further understand the test for equality.
Example 5.5: Out of 5000 interviewees, 2400 are in favour of a proposal, and
out of another set of 2000 interviewees, 1200 are in favour. Is the difference
significant?

2400 1200
Where, P1   0.48 P2   0.6
5000 2000
Solution:

2400 1200
Given, P1   0.48 P2   0.6
5000 2000
n1 = 5000 n2 = 2000

 0.48  0.52 0.6  0.4 

SE     = 0.013 (using Case (II))
 5000 2000 

P1  P2 0.12
z   9.2  3
SE 0.013

The difference is highly significant at 0.27 per cent level.

5.6.2 Test of Hypothesis Conerning Standard Deviation

Large sample test for equality of two means X 1 , X 2

Suppose two samples of sizes n1 and n2 are drawn from populations having means
1, 2 and standard deviations 1, 2.
To test the equality of means X1 , X 2 we write,

H 0 : 1   2
H1 : 1   2

If we assume H0 is true, then

X1  X 2
z
12  22 , approximately normally distributed with mean 0, and S.D. = 1.

n1 n2
We write z ~ N (0, 1)
As usual, if | z | > 2 we reject H0 at 4.55% level of significance, and so on
(Refer Example 5.6).
Example 5.6: Two groups of sizes 121 and 81 are subjected to tests. Their
means are found to be 84 and 81 and standard deviations 10 and 12. Test for the
Self - Learning
322 Material
significance of difference between the groups.
Solution: Estimation and
Hypothesis Testing
X 1 = 84 X 2 = 81 n1 = 121 n2 = 81
1 = 10 2 = 12
NOTES
X1  X 2 84  81
z , z    = 1.86 < 1.96
12  22 

n1 n2 121 81
The difference is not significant at the 5 per cent level of significance.
Small sample tests of significance
The sampling distribution of many statistics for large samples is approximately
normal. For small samples with n < 30, the normal distribution, as shown in Example
5.4, can be used only if the sample is from a normal population with known .
If  is not known, we can use student’s t distribution instead of the normal. We
then replace  by sample standard deviation  with some modification as shown.
Let x1, x2, ..., xn be a random sample of size n drawn from a normal population
with mean  and S.D. . Then,
x  .
t
s / n 1
Here, t follows the student’s t distribution with n – 1 degrees of freedom.
Note: For small samples of n < 30, the term n  1 , in SE = s / n  1 , corrects the bias,
resulting from the use of sample standard deviation as an estimator of 
Also,
s2 n  1 n 1
2
 or s  S
S n n
Procedure: Small samples
To test the null hypothesis H 0 :     , against the alternative hypothesis H1 :    
X 
Calculate t  and compare it with the table value with n – 1 degrees of
SE ( X )
freedom (d.f.) at level of significance 1 per cent.
If this value > table value, reject H0
If this value < table value, accept H0
(Significance level idea same as for large samples)
We can also find the 95% (or any other) confidence limits for .
For the two-tailed test (use the same rules as for large samples; substitute t for z)
the 95% confidence limits are,
X  t s / n  1   

Self - Learning
Material 323
Estimation and Rejection region
Hypothesis Testing
At a per cent level for two-tailed test if | t | > ta/2 reject.
For one-tailed test, (right) if t > ta reject
NOTES (left) if t > –ta reject
At 5 per cent level the three cases are,
If | t | > t0.025 reject two-tailed
If t > t0.05 reject one-tailed right
If t £ t0.05 reject one-tailed left
For proportions, the same procedure is to be followed.
Example 5.7: A firm produces tubes of diameter 2 cm. A sample of 10 tubes is
found to have a diameter of 2.01 cm and variance 0.004. Is the difference significant?
Given t0.05,9= 2.26.
Solution:
X 
t
s / n 1
2.01 2

0.004/ 10 1
0.01

0.021
 0.48
Since, |t| < 2.26, the difference is not significant at 5 per cent level.

Check Your Progress

8. What are the objects of sample distribution theory?
9. Define the central limit theorem.
10. Give the name of some properties of estimation.
11. What is a hypothesis?
12. How will you define the characteristics of hypothesis?
13. How many types of statistical errors?

5.7 ANSWERS TO ‘CHECK YOUR PROGRESS’

1. To test the validity of the hypothesis we would make use of sample
observations and statistics.
2. A very important aspect of sampling theory is the study of test of significance
which gives us a ground to decide the deviation between the observed
sample statistic and the hypothetical parameter value or the deviation
between two independent sample statistics.
Self - Learning 3. A sample is a portion of the total population that is considered for study and
324 Material analysis.
 4. The various methods of sampling can be grouped under two broad Estimation and
Hypothesis Testing
categories—probability (or random) sampling and non-probability (or non-
random) sampling.
5. Non-probability sampling methods do not provide every item in the universe
NOTES
with a known chance of being included in the sample.
6. Probability sampling methods are those in which every item in the universe
has a known chance, or probability of being chosen for the sample.
7. The following are the probability sampling method:
 Simple random sampling
 Stratified random sampling
 Systematic sampling
 Multistage or cluster sampling
8. One of the major objectives of statistical analysis is to know the ‘true’ values
of different parameters of the population. Since it is not possible due to
time, cost and other constraints to take the entire population for
consideration, random samples are taken from the population.
9. Central Limit Theorem states that, ‘Regardless of the shape of the population,
the distribution of the sample means approaches the normal probability
distribution as the sample size increases.’
10. The best estimator should be highly reliable and have such desirable properties
as unbiasedness, consistency, efficiency and sufficiency.
11. A hypothesis is an approximate assumption that a researcher wants to test
for its logical or empirical consequences. Hypothesis refers to a provisional
idea whose merit needs evaluation, but as no specific meaning.
12. The characteristics of hypothesis are as follows:
 Clear and accurate: Hypothesis should be clear and accurate so as to
draw a consistent conclusion.
 Statement of relationship between variables: If a hypothesis is relational,
it should state the relationship between different variables.
 Testability: A hypothesis should be open to testing so that other deductions
can be made from it and can be confirmed or disproved by observation.
The researcher should do some prior study to make the hypothesis a
testable one.
 Specific with limited scope: A hypothesis, which is specific, with limited
scope, is easily testable than a hypothesis with limitless scope. Therefore,
a researcher should pay more time to do research on such kind of
hypothesis.
13. There are two types of errors in statistical hypothesis, which are as follows:
 Type I error: In this type of error, you may reject a null hypothesis when
it is true. It means rejection of a hypothesis, which should have been
accepted. It is denoted by a (alpha) and is also known alpha error.
Self - Learning
Material 325
Estimation and  Type II error: In this type of error, you are supposed to accept a null
Hypothesis Testing
hypothesis when it is not true. It means accepting a hypothesis, which
should have been rejected. It is denoted by b (beta) and is also known
as beta error.
NOTES
5.8 SUMMARY
 In statistics, population does not mean human population alone. A complete
set of objects under study, living or non-living is called population or universe;
for example, graduates in Pondy, Bajaj tube lights, Ceat car tyres etc.
 Each individual or object is called unit or member or element of that
population. If there are a finite number of elements then it is called finite
population. If the number of elements is infinite then it is called infinite
population.
 A population, in statistical terms, is the totality of things under consideration.
It is the collection of all values of the variable that is under study.
 A sample is a portion of the total population that is considered for study and
analysis.
 A random sample is a collection of items selected from the population in
such a manner that each item in the population has exactly the same chance
of being selected, so that the sample taken from the population would be
truly representative of the population.
 Selecting an adequate sample is one of the steps in the primary data collection
process.
 The various methods of sampling can be grouped under two broad
categories—probability (or random) sampling and non-probability (or non-
random) sampling.
 A parameter is a numeric quantity that describes a certain population
characteristic.
 The best estimator should be highly reliable and have such desirable properties
as unbiasedness, consistency, efficiency and sufficiency.
 A hypothesis is an approximate assumption that a researcher wants to test
for its logical or empirical consequences. Hypothesis refers to a provisional
idea whose merit needs evaluation, but having no specific meaning.
 A hypothesis should be stated in the most simple and clear terms to make it
understandable.
 A hypothesis should be reliable and consistent with established and known
facts.
 A hypothesis should be capable of being tested within a reasonable time. In
other words, it can be said that the excellence of a hypothesis is judged by
the time taken to collect the data needed for the test.
 In type I error, you may reject a null hypothesis when it is true. It means
Self - Learning rejection of a hypothesis, which should have been accepted. It is denoted
326 Material
by a (alpha) and is also known alpha error.
  In type II error, you are supposed to accept a null hypothesis when it is not Estimation and
Hypothesis Testing
true. It means accepting a hypothesis, which should have been rejected. It
is denoted by b (beta) and is also known as beta error.

NOTES
5.9 KEY TERMS
 Population (in statistics): It is a complete set of objects under study,
living or non-living.
 Standard error of mean: Measures the likely deviation of a sample mean
from the grand mean of the sampling distribution.
 Efficiency: An estimator is considered to be efficient if its value remains
stable from sample to sample. The best estimator would be the one which
would have the least variance from sample to sample taken randomly from
the same population. From the three point estimators of central tendency,
namely the mean, the mode and the median, the mean is considered to be
the least variant and hence, a better estimator.
 Null hypothesis: A hypothesis stated in the hope of being rejected.

5.10 SELF-ASSESSMENT QUESTIONS AND

EXERCISES

Short-Answer Questions
1. What is meant by statistical estimation?
2. What do you understand by computation of the standard error?
3. Describe the terms sample and sampling.
4. What are the two laws of statistics?
5. How do sampling and non-sampling errors arise?
6. What is estimation?
7. What are the characteristics of a hypothesis?
Long-Answer Questions
1. A company claims that 5% of its products are defective. In a sample of 400
items 320 are good. Test whether the claim is valid.
2. Discuss why sampling is necessary with the help of giving examples.
3. Write an explanatory note on census and sampling.
4. Explain the various non-probability sampling methods.
5. Discuss the various types of probability sampling methods.
6. Briefly explain about the estimation with the help of giving examples.
7. Explain the two types of errors in statistical hypothesis.

Self - Learning
Material 327
Estimation and
Hypothesis Testing 5.11 FURTHER READING
Chance, William A. 1969. Statistical Methods for Decision Making. Illinois:
NOTES Richard D Irwin.
Chandan, J.S., Jagjit Singh and K.K. Khanna. 1995. Business Statistics. New
Delhi: Vikas Publishing House.
Elhance, D.N. 2006. Fundamental of Statistics. Allahabad: Kitab Mahal.
Freud, J.E., and F.J. William. 1997. Elementary Business Statistics – The
Modern Approach. New Jersey: Prentice-Hall International.
Goon, A.M., M.K. Gupta, and B. Das Gupta. 1983. Fundamentals of Statistics.
Vols. I & II, Kolkata: The World Press Pvt. Ltd.
Gupta, S.C. 2008. Fundamentals of Business Statistics. Mumbai: Himalaya
Publishing House.
Kothari, C.R. 1984. Quantitative Techniques. New Delhi: Vikas Publishing
House.
Levin, Richard. I., and David. S. Rubin. 1997. Statistics for Management. New
Jersey: Prentice-Hall International.
Meyer, Paul L. 1970. Introductory Probability and Statistical Applications.
Massachusetts: Addison-Wesley.
Gupta, C.B. and Vijay Gupta. 2004. An Introduction to Statistical Methods,
23rd Edition. New Delhi: Vikas Publishing House Pvt. Ltd.
Hooda, R. P. 2013. Statistics for Business and Economics, 5th Edition. New
Delhi: Vikas Publishing House Pvt. Ltd.
Anderson, David R., Dennis J. Sweeney and Thomas A. Williams. Essentials of
Statistics for Business and Economics. Mumbai: Thomson Learning,
2007.
S.P. Gupta. 2021. Statistical Methods. Delhi: Sultan Chand and Sons.

Self - Learning
328 Material