MA 214: Introduction To Numerical Analysis: Shripad M. Garge. IIT Bombay (Shripad@math - Iitb.ac - In)

MA 214: Introduction to numerical analysis
Lecture 28
Shripad M. Garge.
IIT Bombay
( shripad@math.iitb.ac.in )
2021-2022
1/2
shripad@math.iitb.ac.in MA214 (2021-2022) L28
Improper integrals
Improper integrals result when either the function is unbounded or

the interval is unbounded.
The latter case is where the interval is pa, 8q, p´8, bq or

p´8, 8q.
In either circumstance, the normal rules of integral approximation

must be modified.
We will first consider the situation when the integrand is

unbounded at the left endpoint of the interval of integration.
In this case we say that f has a singularity at the endpoint a.
In fact, other improper integrals can be reduced to problems of this

form so this is a general case.
2/2
Improper integrals
3/2
Improper integrals
Our calculus courses tell us that the improper integral with a

singularity at the left endpoint
żb
dx
a px ´ 1qp
converges if and only if 0 ă p ă 1, and in this case we define

żb ˇx“b
dx px ´ aq1´p ˇˇ pb ´ aq1´p
“ lim “ .
a px ´ aqp MÑa` 1´p ˇ
x“M 1´p
ż1
dx
For instance, the improper integral ? converges (to 2) but
0 x
ż1
dx
2
diverges.
0 x
4/2
Improper integrals
g pxq
If f pxq “ where 0 ă p ă 1 and g : ra, bs Ñ R is
px ´ aqp
continuous then the improper integral
żb
f pxqdx
a
exists.
We will approximate this integral using the composite Simpson’s

rule, provided that g is 5-times continuously differentiable on
ra, bs, to be able to use the fourth Taylor polynomial, P4 pxq, for g
about a:
g 2 paq g 3 paq g p4q paq
P4 pxq “ g paq ` g 1 paqpx ´ aq ` px ´ aq2 ` px ´ aq3 ` px ´ aq4 .
2 3! 4!
5/2
Improper integrals
We write
żb żb żb
g pxq ´ P4 pxq P4 pxq
f pxqdx “ dx ` dx.
a a px ´ aqp a px ´ aqp
The latter integral can be computed easily, as P4 is a polynomial,

which is
4 ż b pkq 4
ÿ g paq ÿ g pkq paq
px ´ aqk´p dx “ pb ´ aqk`1´p .
k“0 a
k! k“0
k!pk ` 1 ´ pq
This is generally the dominant portion of the approximation,

especially when the P4 pxq agrees closely with g pxq throughout the
interval ra, bs.
6/2
Improper integrals
żb
g pxq ´ P4 pxq
We now need to determine dx. We define
a px ´ aqp
#
g pxq´P4 pxq
pxáqp if a ă x ď b
G pxq “
0 if x “ a.
pkq
Since 0 ă p ă 1 and P4 paq agrees with g pkq paq for k “ 0, . . . , 4,
G is 4-times continuously differentiable on ra, bs.
Therefore composite Simpson’s rule can be applied to approximate

the integral of G on ra, bs.
Adding this approximation to the value obtained in the previous

slide gives an approximation to the improper integral of f on ra, bs,
within the accuracy of the composite Simpson’s rule.
7/2
An example
ż1
ex
Compute ? dx.
0 x
x2 x3 x4
Here P4 pxq “ 1 ` x ` ` ` , so
2 6 24
ż1 ż1ˆ ˙
P4 pxq ´1{2 1{2 1 3{2 1 5{2 1 7{2
? dx “ x `x ` x ` x ` x dx
0 x 0 2 6 24
„ 1
2 1 1 1 9{2
“ lim 2x ` x 3{2 ` x 5{2 ` x 7{2 `
1{2
x
MÑ0` 3 5 21 108 M
2 1 1 1
“ 2` ` ` `
3 5 21 108
« 2.92354497354.
8/2
The example, continued
ż1
Now we need to approximate G pxqdx where
0
#
?1 pe x ´ P4 pxqq if 0 ă x ď 1
G pxq “ x
0 if x “ 0.
Using composite Simpson’s rule, we get

ż1
p0.25q
G pxqdx « rG p0q ` 4G p0.25q ` 2G p0.5q ` 4G p0.75q ` G p1qs
0 3
p0.25q
“ r0 ` 4p0.0000170q ` 2p0.0004013q
3
`4p0.0026026q ` 0.0099485s “ 0.0017691.
Hence
ż1
ex
? dx « 2.9235450 ` 0.0017691 “ 2.9253141.
0 x
9/2
Singularity at b
If the integrand has a singularity at the right endpoint, b, and not

at a, then we could develop a similar technique making appropriate
changes.
Alternatively, we can make the substitution z “ ´x, dz “ ´dx to

get
żb ż á
f pxqdx “ f p´zqdz
a ´b
which has a singularity at the left endpoint now.
Then we can apply the left endpoint singularity technique we have

already developed.
10/2
Improper integrals
11/2
Improper integrals
If the integrand f has singularities at both the end points or if it

has a singularity at c P pa, bq then we can write
żb żc żb
f pxqdx “ f pxqdx ` f pxqdx
a a c
and use the techniques developed above.
The other type of improper integral involves infinite limits of

integration.
ż8
1
The basic integral of this type has the form p
dx for p ą 1.
a x
We use the substitution
t “ x ´1 , dt “ ´x ´2 dx, so dx “ ´x 2 dt “ ´t ´2 dt.
12/2
Improper integrals
Then ż0 ż 1{a
tp
ż8
1 1
dx “ ´ 2 dt “ dt.
a xp 1{a t 0 t 2´p
Similarly, if f is defined on ra, 8q then we use the substitution
t “ x ´1 to obtain
ż8 ż 1{a ˆ ˙
´2 1
f pxqdx “ t f dt.
a 0 t
It can now be approximated by the techniques developed above.
If a “ 0 or if both the endpoints are infinite then we write

żb żc żb
f pxqdx “ f pxqdx ` f pxqdx
a a c
for a suitable a ă c ă b and apply the above substitutions.

13/2
The fourth theme is complete
This completes the fourth theme of our course:
Numerical Differentiation and Integration.
From our next lecture, we will begin the next theme:
Ordinary Differential Equations.
14/2
Lecture 29
Shripad M. Garge.
IIT Bombay
2021-2022
15/2
Ordinary differential equations
Differential equations are used to model problems in science and

engineering that involve the change of some variable with respect
to another.
Most of these problems require the solution of an initial-value

problem, that is, the solution to a differential equation that
satisfies a given initial condition.
In common real-life situations, the differential equation that

models the problem is too complicated to solve exactly, and one of
two approaches is taken to approximate the solution.
The first approach is to modify the problem by simplifying the

differential equation to one that can be solved exactly and then use
the solution of the simplified equation to approximate the solution
to the original problem.
16/2
Ordinary differential equations
The other approach, which we will examine here, uses methods for
approximating the solution of the original problem.
This is the approach that is most commonly taken because the

approximation methods give more accurate results and realistic
error information.
The methods that we consider here do not produce a continuous

approximation to the solution of the initial-value problem.
Rather, approximations are found at certain specified, and often

equally spaced, points.
17/2
Some basic material
A function f pt, y q is said to satisfy a Lipschitz condition in the

variable y on a set D Ă R2 if a constant L ą 0 exists with
|f pt, y1 q ´ f pt, y2 q| ď L|y1 ´ y2 |
whenever pt, y1 q and pt, y2 q are in D. The constant L is called a

Lipschitz constant for f .
Theorem
Let D “ tpt, y q : a ď t ď b, ´8 ă y ă 8u and f pt, y q be
continuous on D. If f satisfies a Lipschitz condition on D in the
variable y , then the initial-value problem
y 1 ptq “ f pt, y q, y paq “ α
has a unique solution y ptq for t P ra, bs.
18/2
Well-posed problems
Initial-value problems obtained by observing physical phenomena

generally only approximate the true situation, so we need to know
whether small changes in the statement of the problem introduce
correspondingly small changes in the solution.
This is also important because of the introduction of round-off

error when numerical methods are used.
We say that an initial-value problem is well-posed if

(i) it has a unique solution, and
(ii) an initial-value problem obtained by small perturbations also
has a unique solution.
We make the above part more precise.
19/2
Well-posed problem
Consider the initial-value problem

dy
“ f pt, y q, a ď t ď b, y paq “ α.
dt
We say that this is a well-posed problem if
‚ a unique solution, y ptq, to the problem exists, and
‚ there are 0 ą 0, k ą 0 such that for any 0 ą ą 0, whenever
δptq is continuous with |δptq| ă over ra, bs and when δ0 ă , the
initial-value problem
dz
“ f pt, zq ` δptq, a ď t ď b, zpaq “ α ` δ0
dt
has a unique solution, zptq, that satisfies
|zptq ´ y ptq| ă k for all t P ra, bs.

20/2
Perturbed problem
The problem specified above, whose solution is zptq, is called a

perturbed problem associated with the original problem.
It assumes the possibility of an error being introduced in the

statement of the differential equation, as well as an error δ0 being
present in the initial condition.
Numerical methods will always be concerned with solving a

perturbed problem because any round-off error introduced in the
representation perturbs the original problem.
Unless the original problem is well-posed, there is little reason to

expect that the numerical solution to a perturbed problem will
accurately approximate the solution to the original problem.
21/2
Condition ensuring well-posed property
The following theorem specifies conditions that ensure that an

initial-value problem is well-posed.
Theorem
Suppose D “ tpt, y q|a ď t ď b, ´8 ă y ă 8u. If f is continuous
and satisfies a Lipschitz condition in the variable y on the set D,
then the initial-value problem
dy
“ f pt, y q, a ď t ď b, y paq “ α
dt
is well-posed.
Now, we study the numerical methods to solve the initial-value

problems.
22/2
Euler’s method
Euler’s method is the most elementary approximation technique for

solving initial-value problems.
Although it is seldom used in practice, the simplicity of its

derivation can be used to illustrate the techniques involved in the
construction of some of the more advanced techniques.
The object of Euler’s method is to obtain approximations to the

well-posed initial-value problem
dy
dt
We will not obtain a continuous solution but we generate

approximations to y at various values, called mesh points, in the
interval ra, bs.
23/2
Euler’s method
Once the approximate solution is obtained at the points, the

approximate solution at other points in the interval can be found
by interpolation.
We select the mesh points ti “ a ` ih, for each i “ 0, 1, . . . , N for

a positive integer N.
The common distance between the points
h “ pb ´ aq{N “ ti`1 ´ ti
is called the step size.
We will use Taylor’s Theorem to derive Euler’s method.
24/2
Euler’s method
Suppose that y ptq, the unique solution to the problem, has two
continuous derivatives on ra, bs, so that for each i “ 0, . . . , N ´ 1,
pti`1 ´ ti q2 2
y pti`1 q “ y pti q ` pti`1 ´ ti qy 1 pti q ` y pξi q
2
h2 2
“ y pti q ` hy 1 pti q ` y pξi q
2
for some ξi P pti , ti`1 q. This then gives
h2 2
y pti`1 q “ y pti q ` h f pti , y pti qq ` y pξi q.
2
Euler’s method constructs wi « y pti q, for each i by deleting the
remainder term.
We take w0 “ α and wi`1 “ wi ` hf pti , wi q for i ě 0.

25/2
An example
We illustrate Euler’s method by means of an example.
y 1 “ y ´ t 2 ` 1, 0 ď t ď 2, y p0q “ 0.5.
Let us suppose that we want to approximate y p2q.
We take h “ 0.5. Here f pt, y q “ y ´ t 2 ` 1. We have w0 “ 0.5,
w1 “ w0 ` p0.5qpw0 ´ p0.0q2 ` 1q “ 0.5 ` 0.5p1.5q “ 1.25,

w2 “ 2.25, w3 “ 3.375, and y p2q « w4 “ 4.4375.
Geometrically, when wi is a close approximation to y pti q, the
well-posed property implies that
f pti , wi q « f pti , y pti qq “ y 1 pti q.
26/2
We now take N “ 10 and compare the results with the actual

values given by
y ptq “ pt ` 1q2 ´ 0.5e t .
We then have h “ 0.2, ti “ p0.2qi, w0 “ 0.5 and
wi`1 “ wi ` hpwi ´ ti2 ` 1q

“ wi ` p0.2qrwi ´ 0.04i 2 ` 1s
“ 1.2wi ´ 0.008i 2 ` 0.2
for 0 ď i ď 9. This gives w1 “ 0.8, w2 “ 1.152 and so on.
We compare the approximate values and the actual values in a

table in the following slide.
27/2
28/2
When we had h “ 0.5, the approximate value for y p2q was 4.4375
whereas for h “ 0.2 it is 4.8657845, much closer to the actual
value.
Note that the error grows slightly as the value of t increases.
This controlled error growth is a consequence of the stability of

Euler’s method, which implies that the error is expected to grow in
no worse than a linear manner.
We will study the error analysis of Euler’s method in our next

lecture.
29/2
Lecture 30
Shripad M. Garge.
IIT Bombay
2021-2022
30/2
Euler’s method
We consider an initial-value problem

dy
dt
and give an approximate solution at equally spaced nodes in ra, bs.
The method begins by choosing a positive integer N, letting

h “ bá
N and taking the nodes, ti “ a ` ih for i “ 0, . . . , N.
We then define w0 “ α and wi`1 “ wi ` hf pti , wi q for i ě 0.
We did an example in the last lecture with two different values of

N.
Now, we study the error pattern in this method.
31/2
Error in Euler’s method
Theorem
Suppose f is continuous and satisfies a Lipschitz condition with
the constant L on D “ ra, bs ˆ R.
Assume that the unique solution, y ptq, to the initial-value problem
dy
dt
satisfies |y 2 ptq| ď M.
If wi for i “ 0, . . . , N, are the Euler approximations to y pti q then
hM ” Lpti áq ı
|y pti q ´ wi | ď e ´1 .
2L
32/2
The result is true if i “ 0. Let us denote y pti q by yi . Then
h2 2
yi`1 ´ wi`1 “ pyi ´ wi q ` hrf pti , yi q ´ f pti , wi qs ` y pξi q
2
and
h2 M
|yi`1 ´ wi`1 | ď p1 ` hLq|yi ´ wi | `
2
h2 M h2 M
ˆ ˙
ď p1 ` hLq p1 ` hLq|yi´1 ´ wi´1 | ` `
2 2
2 2
ˆ ˙
h M h M
ď e pi`1qhL |y0 ´ w0 | ` ´
2hL 2hL
hM ` pti`1 áqL ˘
ď e ´1 .
2L
33/2
The weakness of the above theorem is that we need to know a

bound for y 2 ptq.
Although this condition often prohibits us from obtaining a realistic

error bound, it should be noted that if Bf {Bt and Bf {By both exist,
the chain rule gives
dy 1 df Bf Bf
y 2 ptq “ ptq “ pt, y ptqq “ pt, y ptqq` pt, y ptqq¨f pt, y ptqq.
dt dt Bt By
So it is, at times, possible to obtain an error bound for y 2 ptq

without explicitly knowing y ptq.
The principal importance of the error formula is that the bound

depends linearly on the step size h. Consequently, diminishing the
step size gives greater accuracy to the approximations.
34/2
We did not consider the round-off error while deriving the above
formula.
As h becomes smaller, more calculations are necessary and more

round-off error is expected.
In actuality then, we do not use
w0 “ α, wi`1 “ wi ` hf pti , wi q
but
u0 “ α ` δ 0 , ui`1 “ ui ` hf pti , ui q ` δi`1
where δ0 , . . . are the round-off errors.
We now describe an error bound for Euler’s method taking the

round-off errors into account.
35/2

dy
dt
satisfying the hypothesis of the above theorem.
Let u0 , . . . be the approximations in the Euler’s method with

round-off errors δi and assume that |δi | ă δ for all i.
Then
ˆ ˙
1 hM δ
|y pti q ´ ui | ď ` re Lpti áq ´ 1s ` |δ0 |e Lpti áq
L 2 h
Note that this error bound is not linear in h. In fact, the error will
be expected to be large for small step size.
36/2
Local truncation error
Since the object of a numerical technique is to determine accurate

approximations with minimal effort, we need a means for
comparing the efficiency of various approximation methods.
The first device we consider is called the local truncation error of

the method.
The local truncation error at a specified step measures the amount

by which the exact solution to the differential equation fails to
satisfy the difference equation being used for the approximation at
that step.
We want to know how well the approximations generated by the

method satisfy the differential equation, not the other way around.
However, we don’t know the exact solution so we cannot generally

determine this.
37/2
The local truncation serves quite well to determine not only the
local error of a method but also the actual approximation error.
y 1 “ f pt, y q, a ď t ď b, y paq “ α.
The difference method w0 “ α, wi`1 “ wi ` hφpti , wi q for

i “ 0, . . . , N ´ 1, has local truncation error
yi`1 ´ pyi ` hφpti , yi qq yi`1 ´ yi

τi`1 phq “ “ ´ φpti , yi q.
h h
This error is a local error because it measures the accuracy of the
method at a specific step, assuming that the method was exact at
the previous step.
38/2
Consider Euler’s method, for instance, where the difference

equation is w0 “ α and wi`1 “ wi ` hf pti , wi q.
The local truncation error is

yi`1 ´ yi h
τi`1 phq “ ´ f pti , yi q “ y 2 pξi q.
h 2
If one knows that |y 2 ptq| ď M for some M then the local
hM
truncation error is bounded by .
2
So the local truncation error in Euler’s method is Ophq.
We select difference-equation methods for solving initial-value

problems in such a manner that their local truncation errors are
Ophp q for as large a value of p as possible, while keeping the
number and complexity of calculations within a reasonable bound.
39/2
Higher order Taylor methods
Since Euler method used Taylor’s theorem, we try using Taylor’s

theorem for higher orders.
y 1 “ f pt, y q, a ď t ď b, y paq “ α
with f being n-times continuously differentiable.
By Taylor’s theorem
hn pnq hpn`1q pn`1q

y pti`1 q “ y pti q ` hy 1 pti q ` ¨ ¨ ¨ ` y pti q ` y pξi q
n! pn ` 1q!
for some ξ P pti , ti`1 q.
We can compute the derivatives of y at ti using the function

f pt, y q.
40/2
Higher order Taylor methods
Successive differentiation of y ptq gives
y 1 ptq “ f pt, y ptqq, y 2 ptq “ f 1 pt, y ptqq, ¨ ¨ ¨ , y pn`1q ptq “ f pnq pt, y ptqq.
Putting this in the above formula, we get
hn pn´1q hpn`1q pnq

yi`1 “ yi ` hf pti , yi q ` ¨ ¨ ¨ ` f pti , yi q ` f pξi , y pξi qq.
n! pn ` 1q!
The Taylor method of order n is then w0 “ α and

hn pn´1q
wi`1 “ wi ` hf pti , wi q ` ¨ ¨ ¨ ` f pti , wi q.
n!
Euler’s method is Taylor method of order one.
41/2
Lecture 31
Shripad M. Garge.
IIT Bombay
2021-2022
42/2
Taylor method of order n

dy
dt
The Taylor method of order n gives an approximate solution to the
above problem at equally spaced nodes ti “ a ` ih where h “ bá
N .
These solutions are w0 “ α and wi`1 “ wi ` hT pti , wi q where
h hn´1 pn´1q
T pti , wi q “ f pti , wi q ` f 1 pti , wi q ` ¨ ¨ ¨ ` f pti , wi q.
2 n!
Let us now do some examples.
43/2
An example
Let us apply Taylor’s method of order two with N “ 10 to solve
y 1 “ y ´ t 2 ` 1, 0 ď t ď 2, y p0q “ 0.5.
We will need to compute y 2 “ f 1 pt, y q:
d
f 1 pt, y ptqq “ py ´t 2 `1q “ y 1 ´2t “ py ´t 2 `1q´2t “ y ´t 2 ´2t`1.
dt
Then
h
T pti , wi q “ f pti , wi q ` f 1 pti , wi q
2
h
“ wi ´ ti ` 1 ` pwi ´ ti2 ´ 2ti ` 1q
2
ˆ ˙ 2
h
“ 1` pwi ´ ti2 ` 1q ´ hti .
2
44/2
Since N “ 10, we get that h “ 0.2 and ti “ 0.2i for i “ 0, . . . , 10.

Then the method gives w0 “ 0.5 and
„ˆ ˙ 
h
wi`1 “ wi ` h 1 ` pwi ´ ti2 ` 1q ´ hti
2
„ˆ ˙ 
0.2 2
“ wi ` p0.2q 1 ` pwi ´ 0.04i ` 1q ´ 0.04i
2
“ 1.22wi ´ 0.0088i 2 ´ 0.008i ` 0.22.
We have
y p0.2q « 0.83 and y p0.4q « 1.2158.
We now list the values and the errors in the following table.
45/2
Order 2 and N “ 10.
46/2
Another example
We now apply Taylor’s method of order 4 with N “ 10 to the same

problem
y 1 “ y ´ t 2 ` 1, 0 ď t ď 2, y p0q “ 0.5.
We compute
f 1 pt, y ptqq “ y ´ t 2 ´ 2t ` 1,
d
f 2 pt, y ptqq “ py ´ t 2 ´ 2t ` 1q “ y 1 ´ 2t ´ 2
dt
“ py ´ t 2 ` 1q ´ 2t ´ 2 “ y ´ t 2 ´ 2t ´ 1,
d
f 3 pt, y ptqq “ py ´ t 2 ´ 2t ´ 1q “ y 1 ´ 2t ´ 2
dt
“ y ´ t 2 ´ 2t ´ 1.
47/2
Then
h h2 h3
T pti , wi q “ f pti , wi q ` f 1 pti , wi q ` f 2 pti , wi q ` f 3 pti , wi q
2 6 24
2 h 2
“ pwi ´ ti ` 1q ` pwi ´ ti ´ 2ti ` 1q
2
h2 h3
` pwi ´ ti2 ´ 2ti ´ 1q ` pwi ´ ti2 ´ 2ti ´ 1q
ˆ6 24
h h2 h3 h h2
˙ ˆ ˙
2
“ 1` ` ` pwi ´ ti q ´ 1 ` ` phti q
2 6 24 3 12
h h2 h3
`1 ` ´ ´ .
2 6 24
48/2
The method then gives w0 “ 0.5 and
wi`1 “ wi ` hT pti , wi q
h h2 h3
„ˆ ˙
“ wi ` h 1 ` ` ` pwi ´ ti2 q
2 6 24
h h2 h h2 h3
ˆ ˙ 
´ 1` ` phti q ` 1 ` ´ ´
3 12 2 6 24
With N “ 10 and h “ 0.2, this becomes
wi`1 “ 1.2214 wi ´ 0.008856 i 2 ´ 0.00856 i ` 0.2186.
y p0.2q « 0.8293 and y p0.4q « 1.214091.

We note the results in the next table.
49/2
Order 4 and N “ 10.
50/2
Order 2 Order 4
51/2
Suppose now that with this data, we want to approximate the

value of y at some other point, say at y “ 1.25.
We then use interpolation.
If we use linear interpolation on the nodes 1.2 and 1.4 then we get
ˆ ˙ ˆ ˙
1.25 ´ 1.4 1.25 ´ 1.2
y p1.25q « 3.1799640 ` 3.7324321
1.2 ´ 1.4 1.4 ´ 1.2
“ 3.3180810.
The true value is y p1.25q “ 3.3173285, so the error of this

approximation is 0.0007525, about 30 times the average of the
errors at 1.2 and 1.4.
We can improve upon this by using cubic Hermite interpolation on

the nodes 1.2 and 1.4.
52/2
For that we will need to compute the values of y 1 at these two

nodes:
y 1 p1.2q “ y p1.2q ´ p1.2q2 ` 1 « 2.7399640
and
y 1 p1.4q “ y p1.4q ´ p1.4q2 ` 1 « 2.7724321.
The corresponding cubic Hermite polynomial is
3.1799640 ` pt ´ 1.2qp2.7399640q ` pt ´ 1.2q2 p0.1118825q
`pt ´ 1.2q2 pt ´ 1.4qp´0.3071225q

Then
y p1.25q « 3.1799640 ` 0.1369982 ` 0.0002797 ` 0.0001152

“ 3.3173571.
53/2
This value is accurate to within 0.0000286, which is about the

average of the errors at 1.2 and at 1.4, and only 4% of the error
obtained using linear interpolation.
This improvement in the accuracy certainly justifies the added

computations required for the Hermite method.
54/2
Error in Taylor’s method
Theorem
If Taylor’s method of order n is used to approximate the solution to
y 1 “ f pt, y q, a ď t ď b, y paq “ α
with step size h then the local truncation error is Ophn q.
There are, of course, limitations and draw backs to this method.
The number of calculations of various derivatives of f pt, y ptqq is a

bit of a concern, even if the error is Ophn q.
We would like to reduce the number of calculations while keeping

the error to the similar order.
55/2
Lecture 32
Shripad M. Garge.
IIT Bombay
2021-2022
56/2
Taylor method of order n
The Taylor methods outlined in the previous lectures have the

desirable property of high-order local truncation error.
But they also have the disadvantage of requiring the computation

and evaluation of the derivatives of f pt, y q.
This is a complicated and time-consuming procedure for most

problems, so the Taylor methods are seldom used in practice.
We would like to learn approximation methods to solve the

initial-value problems which do not require as high computations
and evaluations but have the error to the similar order.
We will study the Runge-Kutta method but before that we will

need the statement of Taylor’s theorem in two variables.
57/2
Taylor’s theorem in two variables
Theorem
Suppose that f pt, y q and all its partial derivatives of order less
than or equal to n ` 1 are continuous on
D “ tpt, y q : a ď t ď b, c ď y ď du,
and let pt0 , y0 q P D. For every pt, y q P D, there exists ξ between t

and t0 and µ between y and y0 with
f pt, y q “ Pn pt, y q ` Rn pt, y q
where Pn is called the n-th Taylor polynomial in two variables for

the function f about pt0 , y0 q, and Rn pt, y q is the remainder term
associated with Pn pt, y q.
The error term Rn pt, y q involves ξ and µ.

58/2
Taylor’s theorem in two variables
„ 
Bf Bf
Pn pt, y q “ f pt0 , y0 q ` pt ´ t0 q pt0 , y0 q ` py ´ y0 q pt0 , y0 q
Bt By
2 2 B2f
„
pt ´ t0 q B f
` pt 0 , y 0 q ` pt ´ t0 qpy ´ y0 q pt0 , y0 q`
2 Bt 2 BtBy
py ´ y0 q2 B 2 f

pt ,
0 0 y q ` ¨¨¨
2 By 2
« ff
n ˆ ˙ nf
1 ÿ n B
` pt ´ t0 qn´j py ´ y0 qj n´j j pt0 , y0 q
n! j“0 j Bt By
and
n`1
1 ÿ ˆn ` 1˙ B n`1 f
Rn pt, y q “ pt´t0 qn`1´j py ý0 qj n`1´j j pξ, µq.
pn ` 1q! j“0 j Bt By
59/2
Runge-Kutta Methods of Order Two
The first step in deriving a Runge-Kutta method is to determine

values for a1 , α1 and β1 with the property that a1 f pt ` α1 , y ` β1 q
approximates
h
T pt, y q “ f pt, y q ` f 1 pt, y q
2
with error no greater than Oph2 q. Since
df Bf Bf
f 1 pt, y q “ pt, y q “ pt, y q ` pt, y q ¨ y 1 ptq
dt Bt By
Bf Bf
“ pt, y q ` pt, y q ¨ f pt, y q
Bt By
we get
h Bf h Bf
T pt, y q “ f pt, y q ` pt, y q ` pt, y q ¨ f pt, y q.
2 Bt 2 By
60/2
Using Taylor’s theorem of order 1 we get
a1 f pt ` α1 , y ` β1 q “ a1 f pt, y q
Bf Bf
à1 α1 pt, y q ` a1 β1 pt, y q
Bt By
à1 R1 pt ` α1 , y ` β1 q
where
α12 B 2 f B2f β12 B 2 f
R1 pt ` α1 , y ` β1 q “ pξ, µq ` α 1 β1 pξ, µq ` pξ, µq
2 Bt 2 BtBy 2 By 2
for some ξ between t and t ` α1 and µ between y and y ` β1 .
We now match the equations for T pt, y q and a1 f pt ` α1 , y ` β1 q

and get some equations.
61/2
These equations are

Bf h Bf h
f pt, y q : a1 “ 1, pt, y q : a1 α1 “ , pt, y q : a1 β1 “ f pt, y q.
Bt 2 By 2
We then conclude
h h
a1 “ 1, α1 “ , and β1 “ f pt, y q
2 2
and get
ˆ ˙ ˆ ˙
h h h h
T pt, y q “ f t ` , y ` f pt, y q ´ R1 t ` , y ` f pt, y q
2 2 2 2
where the error R1 can be written as follows:
62/2
h2 B 2 f
ˆ ˙
h h
R1 t ` , y ` f pt, y q “ pξ, µq
2 2 8 Bt 2
h2 B2f
` f pt, y q pξ, µq
4 BtBy
h2 B 2 f
` pξ, µq
8 By 2
If all the second-order partial derivatives of f are bounded, then
ˆ ˙
h h
R1 t ` , y ` f pt, y q
2 2
is of Oph2 q. As a consequence, the order of error for this method is

the same as that of the Taylor method of order two.
63/2
The difference-equation method resulting from replacing T pt, y q in

Taylor’s method of order two by
ˆ ˙
h h
f t ` , y ` f pt, y q
2 2
is a specific Runge-Kutta method known as the Midpoint-method.
It gives w0 “ α and
ˆ ˙
h h
wi`1 “ wi ` hf ti ` , wi ` f pti , wi q
2 2
for i “ 0, . . . , N ´ 1.
We do an example now.
64/2
An example
Use the Midpoint method with N “ 10, h “ 0.2, ti “ 0.2i, and

w0 “ 0.5 to approximate the solution to our usual example
y 1 “ y ´ t 2 ` 1, 0 ď t ď 2, y p0q “ 0.5.
The difference formula is
wi`1 “ 1.22wi ´ 0.0088i 2 ´ 0.008i ` 0.218.
for each i “ 0, . . . , 9.
The first two steps of these methods give
w1 “ 0.828, and w2 “ 1.21136.
The table is given on the next slide.
65/2
66/2

MA 214: Introduction To Numerical Analysis: Shripad M. Garge. IIT Bombay (Shripad@math - Iitb.ac - In)

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MA 214: Introduction To Numerical Analysis: Shripad M. Garge. IIT Bombay (Shripad@math - Iitb.ac - In)

Uploaded by

Copyright:

Available Formats

MA 214: Introduction to numerical analysis

Improper integrals result when either the function is unbounded or

The latter case is where the interval is pa, 8q, p´8, bq or

In either circumstance, the normal rules of integral approximation

We will first consider the situation when the integrand is

In this case we say that f has a singularity at the endpoint a.

In fact, other improper integrals can be reduced to problems of this

Our calculus courses tell us that the improper integral with a

converges if and only if 0 ă p ă 1, and in this case we define

We will approximate this integral using the composite Simpson’s

The latter integral can be computed easily, as P4 is a polynomial,

This is generally the dominant portion of the approximation,

Therefore composite Simpson’s rule can be applied to approximate

Adding this approximation to the value obtained in the previous

Using composite Simpson’s rule, we get

If the integrand has a singularity at the right endpoint, b, and not

Alternatively, we can make the substitution z “ ´x, dz “ ´dx to

Then we can apply the left endpoint singularity technique we have

If the integrand f has singularities at both the end points or if it

and use the techniques developed above.

The other type of improper integral involves infinite limits of

We use the substitution

It can now be approximated by the techniques developed above.

If a “ 0 or if both the endpoints are infinite then we write

for a suitable a ă c ă b and apply the above substitutions.

This completes the fourth theme of our course:

Numerical Differentiation and Integration.

From our next lecture, we will begin the next theme:

Ordinary Differential Equations.

Differential equations are used to model problems in science and

Most of these problems require the solution of an initial-value

In common real-life situations, the differential equation that

The first approach is to modify the problem by simplifying the

This is the approach that is most commonly taken because the

The methods that we consider here do not produce a continuous

Rather, approximations are found at certain specified, and often

A function f pt, y q is said to satisfy a Lipschitz condition in the

|f pt, y1 q ´ f pt, y2 q| ď L|y1 ´ y2 |

whenever pt, y1 q and pt, y2 q are in D. The constant L is called a

y 1 ptq “ f pt, y q, y paq “ α

has a unique solution y ptq for t P ra, bs.

Initial-value problems obtained by observing physical phenomena

This is also important because of the introduction of round-off

We say that an initial-value problem is well-posed if

We make the above part more precise.

Consider the initial-value problem

|zptq ´ y ptq| ă k for all t P ra, bs.

The problem specified above, whose solution is zptq, is called a

It assumes the possibility of an error being introduced in the

Numerical methods will always be concerned with solving a

Unless the original problem is well-posed, there is little reason to

The following theorem specifies conditions that ensure that an

Now, we study the numerical methods to solve the initial-value

Euler’s method is the most elementary approximation technique for

Although it is seldom used in practice, the simplicity of its

The object of Euler’s method is to obtain approximations to the

We will not obtain a continuous solution but we generate

Once the approximate solution is obtained at the points, the

We select the mesh points ti “ a ` ih, for each i “ 0, 1, . . . , N for

The common distance between the points

is called the step size.

We will use Taylor’s Theorem to derive Euler’s method.

|zptq ´ y ptq| ă k for all t P ra, bs.