Chapter 5:

Numerical Integration and

Differentiation
– Newton-Cotes Integration Formulas
– Trapezoidal Rule
– Simpson’s Rules
– Integration with Unequal Segments
– Integration of Equations
– Romberg Integration
– Gauss Quadrature
– Numerical Differentiation

1
 Newton-Cotes Integration Formula
 Replacing the complicated integrand function or tabulated data by
an approximating function such as a polynomial.
f(x) f(x)

x x
Replacing by a Replacing by a
straight line parabola
 These formulas are applicable to both closed forms (i.e., data points
at the beginning and end of the integration limits are known) and
open forms (otherwise).
 These formulas are applied for equally spaced intervals. 2
 f(x) f(a)+f(b)
=average height
2

f(a)
I  ( width )  (average height )
f(b)

a b x
width
 All the Newton-Cotes formulas can be formulated in general
form above (only definition of average height changes).

3
Trapezoidal Rule

4
Trapezoidal Rule

5
 Trapezoidal Rule

In Example 1 we deliberately chose an integral whose value can be

computed explicitly so that we can see how accurate the Trapezoidal
and Midpoint Rules are. By the Fundamental Theorem of Calculus,

6
 Trapezoidal Rule

The error in using an approximation is defied to be the amount that

needs to be added to the approximation to make it exact. From the
values in Example 1 we see that the errors in the Trapezoidal and
Midpoint Rule approximations for n − 5 are

7
 Trapezoidal Rule

Comparing this error estimate of 0.006667 with the actual error of about
0.002488, we see that it can happen that the actual error is substantially
less than the upper bound for the error given by (3) 8
 Trapezoidal Rule

Thus n − 41 will ensure the desired accuracy.

For the same accuracy with the Midpoint Rule we choose n so that

9
 Simpson’s Rule

Another rule for approximate integration results from using parabolas

instead of straight line segments to approximate a curve. As before, we
divide [a, b ] into n subintervals of equal length h = △X = , but this time we
assume that n is an even number. Then on each consecutive pair of
intervals we approximate the curve y = f(x) ≥ 0 by a parabola as shown in
Figure 1. If = f( then ( , is the point on the curve lying above . A typical
parabola passes through three consecutive points , , and .

FIGURE 2
FIGURE 1 10
 Simpson’s Rule
To simplify our calculations, we fist consider the case where = -h, = 0, and
= h. (See Figure 2.) We know that the equation of the parabola through , ,
and is of the form y = A + Bx + C and so the area under the parabola from x
= − h to x = h is

11
 Simpson’s Rule
Now by shifting this parabola horizontally we do not change the area under
it. This means that the area under the parabola through , , and from x = to
x = in Figure 1 is still

12
Simpson’s Rule

13
Multiple-application trapezoidal rule:
 We can increase the accuracy of the integration by dividing the
interval into many segments and apply the trapezoidal rule to
each segment.
f(x)
For n+1 equally spaced base
points, there are n segments of
step size:
(b  a )
h
n
x0 x1 x2 x
h
h n 1

I
2 
f ( x0 )  2 
i 1
f ( xi )  f ( x )
n 


Initial point final point

14
 or  n 1

 f ( x0 )  2  f ( xi )  f ( x )
n 

I  (b  a )  i 1 
2n

width average height

 Error for the multiple-application trapezoidal rule is found by

summing individual errors.
b  a
3 n
Et 
12n 3

i 1
f '' ( i )
or
n Convergence:
Ea 
b  a
3
f
''
where ''
f ''
( i ) Error is inversely
12n 2 f  i 1
related to the
n square of n !
15
Multiple-application Simpson’s 1/3 rule
For n segments:
(b  a )
h
n
Total integral can be represented
x2 x4 xn

I  f ( x)dx   f ( x)dx  ...   f ( x)dx

x0 x2 xn 2

Substituting Simpson’s 1/3 formula

n 1 n2
f ( x0 )  4  f (x )  2  f (x )  f (x )
i 1, 3, 5
i
i  2, 4, 6
i n

I  (b  a )
3n

width average height

16
 Error for the multiple (n) segment
Simpon’s 1/3 rule is calculated by
adding error for individual segments.
Hence, we get:

(b  a ) 5 ( 4 )
Et   4
f
180n

 Multiple application Simpson’s rule returns very accurate results

compare to trapezoidal rule. It is the method of choice for most
applications.
 As for all Newton-Cotes formulas, the intervals must be equally
spaced.
 Because of the need for three points for applications, the method
is limited to odd number of points (even number of segments).
17
 Simpson’s Rule
EXAMPLE How large should we take n in order to guarantee
that the Simpson’s Rule approximation for dx is accurate to
within 0.0001?

18
Simpson’s 3/8 rule
 Simpson’s 3/8 rule is used when the odd-number of segments
are encountered.
 A third-order polynomial is used for replacing the function.
b b
I   f ( x)dx   f 3 ( x)dx
a a

Applying a third order Legendre polynomial to four points gives

3h Simpson’s 3/8
I  f ( x0 )  3 f ( x1 )  3 f ( x2 )  f x3  formula.
8

I  (b  a )
 f ( x0 )  3 f ( x1 )  3 f ( x2 )  f x3  h=(b-a)/3
8

average height
width
19
 Truncation error for Simpson’s 3/8 rule:

(b  a ) 5 ( 4 ) Simpson’s 3/8 rule is

Et   f   somewhat more
6480 accurate than 1/3 rule

 In general Simpson’s 1/3 rule is the method of choice because of

obtaining third order accuracy and using three points instead of
four.
 3/8 rule has the utility when the number of segments is odd.
 In practice, use 1/3 rule for all the even number of segments and
use 3/8 rule for the remaining last three segments.

Higher-order Newton-Cotes formulas:

 Higher order (n=4,5) formulas have the same order error as n=1
(Trapezoid) or n=2,3 (Simpson’s) formulas.
 Higher-order formulas are rarely used in engineering practices.
 To increase the accuracy, just increase the number of segments!
20
 Integration with Unequal Segments
 There may be cases where the spacing between data points may
not be even (e.g., experimentally derived data points)
 One way is to use trapezoidal rule for each segment

f ( x0 )  f ( x1 ) f ( x1 )  f ( x2 ) f ( xn 1 )  f ( xn )
I  h1  h2  ..  hn
2 2 2
width of the segments (h) are not constant so the
formula cannot be written in a compact form.

 A computer algorithm can easily be developed to do the

integration for unequal-sized segments.
 The algorithm can be constructed such that, use Simpson’s 1/3
rule wherever two consecutive equal-sized segments are
encountered, and use 3/8 rule wherever three consecutive equal-
sized segments are encountered. When adjacent segments are
unequal-sized , just use trapezoidal rule. 21
 Open Integration Formulas
 There may be cases where
integration limits are beyond the f(x)
range of the data.
 General characteristics and order of
error for open forms of the Newton-
Cotes formulas are similar to the
closed forms.
 Even segment formulas are usually a x
b
the method of choice as they
require fewer points.
 Open forms are not used for definite integration.
 They have utility for analyzing improper integrals (discussed later).
 They have connection to multi-step method for solving ordinary
differential equations.
22
 Integration of Equations
 Previously discussed methods are appropriate for tabulated data.
 If the function to be integrated is available, other modified
methods are available.
> Romberg integration (Richardson extrapolation)
> Gauss quadrature
Newton-Cotes method for functions:
 Use of multiple-application of
trapezoidal and Simpson’s rules are

Percent relative error

Trapezoidal
not convenient for analyzing rule
integration of functions.
> Large number of operations
Simpson’s
required to evaluate the functions. rule
> For large values of n, round of
errors starts to dominate.
n
23
 Romberg Integration
 It uses the trapezoidal rule, but much more efficient results are
obtained through iterative refinement techniques .
Richardson’s extrapolation:
 A sequence acceleration method to improve the rate of
convergence.
 It offers a very practical tool for numerical integration and
differentiation.
 Application of iterative refinement techniques to improve the
error at each iteration. For each iteration
I  I ( h)  E ( h)
exact approximate truncation
integral integral error

 In Richardson extrapolation, two approximate integrals are used

to compute a third more accurate integral. 24
 Assume that two integrals with step sizes of h1 and h2 are
available. Then,
I (h1 )  E (h1 )  I (h2 )  E (h2 )
(b  a )
Error for multiple-equation trapezoidal rule ( n  h
)

E
b  a  2 ''
h f
12
Assuming same f’’ for two step sizes
2
E (h1 ) h 2
 h1 
 1
2 or E (h1 )  E (h2 ) 
E (h2 ) h 2  h2 

25
Inserting the error to the previous equation, and solving for E(h 2)

I (h1 )  I (h2 )
E (h2 ) 
1  h1 / h2 
2

Then,
1
I  I (h2 )  I (h2 )  I (h1 )
(h1 / h2 )  1
2

 It can be shown that the error for above approximation is O(h4)

although the error from use of trapezoidal rule is O(h2).
 For the special case of interval being halved (h2=h1/2), we get

4 1
I  I (h2 )  I (h1 )
3 3
O(h4) O(h2) 26
EX: For the numerical integration of
f ( x)  0.2  25 x  200 x 2  675 x 3  900 x 4  400 x 5
we previously obtained the following, using trapezoidal rule.
n h I t
1 0.8 0.1728 89.5
2 0.4 1.0688 34.9
4 0.2 1.4848 9.5

Use Romberg integration approach to obtain improved approximations.

use n=1 and n=2 for a improved approximation

4 1  t  16.6%
I  (1.0688)  (0.1728)  1.367467
3 3
If use n=2 and n=4
4 1  t  1. 0%
I  (1.4848)  (1.0688)  1.623467
3 3
27
 The method can be generalized, e.g. combine two O(h4)
approximation to obtain an O(h2) approximation, and so on…
 Using two improved O(h4) approximations in the previous
examples, we can use the following formula
16 1 Im: more accurate approx.
I  Im  Il
15 15 Il: less accurate approx.

to obtain an O(h6) approximation.

 Similarly, two O(h6) approximation can be combined to obtain an
O(h8) approximation
64 1 Im: more accurate approx.
I I m  Il
63 63 Il: less accurate approx.

 This sequential improvement convergence technique is named

under the general approach of Richardson's extrapolation.
28
 EX: In the previous example use the two O(h4) approximations to obtain an O(h6)
approximation for the integral.

16 1
I (1.623467)  (1.367467)  1.640533
15 15
True result is 1.640533 (correct answer to 7 S.F.!).

Algorithm:
O(h2) O(h4) O(h6) O(h8)

…
… …
… … …
…
… …
… …
… … … …
… … …
…
29
Trapezoidal rule
 Gauss Quadrature
 In Newton-Cotes formulation, predetermined and evenly spaced
data are used. Points of integration are covered by the data.
 In Gauss quadrature, points of integration are undetermined.

f(x) f(x)

a b x a b x
 For the example above, integral approximation can be improved
by using two wisely chosen interior points.
 Gauss quadrature offer a method to find these points.
30
Method of undetermined coefficients: f(x)
 Assume two points (x0, x1) so that the f(x1)
integral can be written in the form f(x0)

I  c0 f ( x0 )  c1 f ( x1 )

c‘s: constants
x0
 This requires calculation of four unknowns.
-1 x1 1 x
 We choose these four equations according to our preference.
 We can choose a function value that is exact to third order.
1
c0 f ( x0 )  c1 f ( x1 )   1dx  2 In general, integration
1 limits can be made -1..1 for
1
any integral by a simple
c0 f ( x0 )  c1 f ( x1 )   xdx  0 change of variables.
1

31
 1
2
c0 f ( x0 )  c1 f ( x1 )   x 2 dx 
1
3
1
c0 f ( x0 )  c1 f ( x1 )   x 3 dx  0
1

Or, by evaluating the function values

c0  c1  2

c0 x0  c1 x1  0
2
c x c x 
2
0 0
2
1 1
3
c0 x03  c1 x13  0

 By these deliberate choices, we force these two points give a an

integral that is exact to cubic order.
32
Solving for the unknowns
c0  c1  1
1
x0    0.5773503...
3
1
x1   0.5773503...
3
Then,
1 1 Two-point Gauss-
I  f ( ) f ( )
3 3 Legendre formula

33
 To change (normalize) the limits of integration, we do the
following change of variable:
x  a0  a1 xd
For the lower limit (x=a  xd=-1 )
(b  a )
a0 
a  a0  a1 (1) 2
For the upper limit (x=b  xd=1 )
(b  a )
a1 
b  a0  a1 (1) 2

Substitute into the original formula:

(b  a )  (b  a ) xd These equations are
x subsituted into the
2 original integrand to
Differentiate both sides: transfrom into suitable
(b  a ) form for application to
dx  dxd Gauss-Legendre formula
2
34
EX: Use two-point Gauss-Legendre quadrature formula to integrate
f ( x)  0.2  25 x  200 x 2  675 x 3  900 x 4  400 x 5
from x=0 to x=0.8 (remember that exact solution I=1.640533).

First, perform a change of variable to convert the limits -1..1:

x  0.4  0.4 xd dx  0.4dxd
Substitute into the original equation to get:
0.8

      )dx 
2 3 4 5
( 0 . 2 25 x 200 x 675 x 900 x 400 x
0
1
0.2  25(0.4  0.4 xd )  200(0.4  0.4 xd ) 2 
   0.4dxd
 675(0.4  0.4 xd )  900(0.4  0.4 xd )  400(0.4  0.4 xd ) 
3 4 5
1 

RHS is suitable for evaluation using Gauss quadrature. Evaulating for -1/√3
(=0.516741) and for 1/√3 (=1.305837), and gives the following result:
Equivalent to third
I  0.516741  1.305837  1.822578  t  11 .1% order accuracy in
Simpson’s method.
35
 Numerical Differentiation
Types of Differentiation:
Forward expansion of Taylor series:
f '' ( xi ) 2 h=step size
f ( xi 1 )  f ( xi )  f ( xi )h 
'
h  ...
2!
f ( xi 1 )  f ( xi ) first forward
f ( xi ) 
'
 O ( h)
h difference
Backward expansion of Taylor series
''
f ( xi ) 2
f ( xi 1 )  f ( xi )  f ' ( xi )h  h  ...
2!
f ( xi )  f ( xi 1 ) first backward
f ( xi ) 
'
 O ( h) difference
h
Subtract forward expansion from the backward expansion:
f ( xi 1 )  f ( xi 1 ) centered
f ' ( xi )   O(h 2 )
2h difference
36
Higher-accuracy Differentiation Formulas:
 High-accuracy diffentiation formulas can be obtained by adding
more terms in Taylor expansion.
Forward Taylor series expansion:
''
f ( xi ) 2
f ( xi 1 )  f ( xi )  f ( xi )h 
'
h  ...
2
or
f ( x )  f ( x ) f ''
( xi )
f ' ( xi )  i 1 i
 h  O(h 2 )
h 2

Approximate the second derivative using finite difference formula

f ( xi  2 )  2 f ( xi 1 )  f ( xi )
f ( xi ) 
''

h2

37
Then,
f ( xi 1 )  f ( xi ) f ( xi  2 )  2 f ( xi 1 )  f ( xi )
f ' ( xi )   2
h  O ( h 2
)
h 2h
or,
 f ( xi  2 )  4 f ( xi 1 )  3 f ( xi ) forward
f ( xi ) 
'
 O(h 2 ) scheme
2h

 Backward and centered finite difference formulas can be derived in

a similar way:

3 f ( xi )  4 f ( xi 1 )  f ( xi  2 ) backward
f ' ( xi )   O(h 2 ) scheme
2h

 f ( xi  2 )  8 f ( xi 1 )  8 f ( xi 1 )  f ( xi  2 ) centered
f ( xi ) 
'
 O(h 4 ) scheme
12h
38
Richardson Extrapolation:
 As done for the integration, Richardson extrapolation uses two
derivatives of different step sizes to obtain a more accurate
derivative.
 In a similar fashion applied for the integration, use two step sizes
such that h2=h1/2. Richardson extrapolation recursive formula:

4 1
D D(h2 )  D(h1 )
3 3
O(h4) O(h2) [centered difference scheme]

 The approach can be iteratively used by Romberg algorithm to get

higher accuracies.

39
EX: Calculate the approximate derivative of
f ( x)  0.1x 4  0.15 x 3  0.5 x 2  0.25 x  1.2
at x=0.5 using a step size of h=0.25 (true value=-0.9125).

First, evaluate the following data points:

xi-2=0 ; f(xi-2)=1.2 xi+1=0.75 ; f(xi+1)=0.6363281
xi-1=0.25 ; f(xi-2)=1.103516 xi+2=1 ; f(xi+2)=0.2
xi=0.5 ; f(xi)=0.925

Forward difference 0scheme:

.2  4(0.6363281)  3(0.925)
f ' ( xi )   0.859375  t  5.82%
2(0.25)

Backward difference scheme:

3(0.925)  4(1.035156)  1.2
f ' (0.5)   0.878125  t  3.77%
2(0.25)
Centered difference scheme:
 0.2  8(0.6363281)  8(1.035156)  1.2  t  0%
f ' ( xi )   0.9125
12(0.25)
40
EX: Use Richardson extrapolation of step sizes h=0.5 and h=0.25 to calculate the
derivative of the function
f ( x)  0.1x 4  0.15 x 3  0.5 x 2  0.25 x  1.2

at x=0.5.

Centered scheme finite difference approximation for h=0.5:

0.2  1.2
D(0.5)   1.0  t  9.6%
1

Centered scheme finite difference approximation for h=0.25:

0.6363281  1.103516
D(0.5)   0.934375  t  2.4%
0.5
Apply Richardson extrapolation for improved accuracy:

4 1  t  0%
D(0.5)  (0.934375)  (1)  0.9125
3 3

41
Derivatives of Unequally Spaced Data
 In the previous discussion, both finite difference approximations
and Richardson extrapolation reqires evenly distributed data. So,
these methods are more suitable to evaulate functions.
 Emprically derived data, experimental or from field surveys, are
usually not even.
 One way is to fit a second-order Lagrange interpolating polynomial
to each set of three data points. Derivative of the polynomial:
2 x  xi  xi 1 2 x  xi 1  xi 1 2 x  xi 1  xi
f ' ( x)  f ( xi 1 )  f ( xi )  f ( xi 1 )
( xi 1  xi )( xi 1  xi 1 ) ( xi  xi 1 )( xi  xi 1 ) ( xi 1  xi 1 )( xi 1  xi )

 Using above formula, any point in the range of three data points
can be evaluated.
 This equation has the same accuracy of high-accuracy centered
difference approximation even though data is not need to be
equally spaced. 42
Derivatives and Integrals for Data with Error:
 Differentiation process amplifies the error:

y dy/dt

x x
 As a remedy, fit a smoother function (low-order polynomial) to the
uncertain data.
 On the other hand, integration process reduces the error.
(Succusive negative and positive errors cancel out during
integration). No further action is required.

43