4.

INTERPOLATION

Given that

y  f ( x), x0  x  xn

and assuming that f ( x ) is single-valued, continuous and explicitly known, then the value of
f ( x ) corresponding to x0 , x1 ,..., xn can easily be computed and tabulated. Suppose we want to

reverse this process, that is, given the set of tabular values ( x0 , y0 ), ( x1 , y1 ),..., ( xn , yn ) satisfying

the relation y  f ( x), where the explicit nature of f ( x ) is not known, we want to find a
simpler function, say  ( x), such that f ( x ) and  ( x ) agree at the set of tabulated points. Such
a process is called interpolation. If  ( x ) is a polynomial, then the process is called polynomial
interpolation. One reason for approximating unknown function by means of a polynomial is
outlined in the following theorem:

Theorem 4.1.1 (Weierstrass Approximation Theorem)

Suppose that f is defined and continuous on [a, b]. For each   0, there exists a polynomial
p( x), with the property that

f ( x)  p( x)   , for all x [a, b].

Theorem 4.1.1 shows that given any function, defined and continuous on a closed and bounded
interval, there exists a polynomial that is as “close” to the given function as desired.

y y  f ( x)  

y  p ( x)

y  f ( x)

y  f ( x)  

0 a b x

The fact that derivatives and indefinite integrals for polynomials can easily be determined is
another reason for using polynomials to approximate unknown function.

1
Taylor polynomial may not give a correct approximation because it is concentrated at a point
x0 . Thus, moving away from x0 gives inaccurate approximations.

1
Exercise: Check that using the Taylor polynomial for f ( x )  expanded about x  1 to
x
1
approximate f (3)  gives inaccurate value.
3

4. 1. LAGRANGE INTERPOLATING POLYNOMIAL

Suppose we have two distinct points ( x0 , y0 ) and ( x1 , y1 ), and would like to approximate f

for which f ( x0 )  y0 and f ( x1 )  y1. Define the functions

x  x1 x  x0
L0 ( x)  and L1 ( x)  .
x0  x1 x1  x0

The linear Lagrange interpolating polynomial through ( x0 , y0 ) and ( x1 , y1 ) is

x  x1 x  x0
p ( x)  L0 ( x) f ( x0 )  L1 ( x) f ( x1 )  f ( x0 )  f ( x1 )
x0  x1 x1  x0

Example 4.1.1

Determine the linear interpolating polynomial that passes through the points (2, 4) and (5,1).

Solution:

x  x1 x  5 1
L0 ( x)     ( x  5)
x0  x1 2  5 3
x  x0 x  2 1
L1 ( x)    ( x  2)
x1  x0 5  2 3
 1  1 
 p( x)  L0 ( x) f ( x0 )  L1 ( x) f ( x1 )    ( x  5)  (4)   ( x  2)  (1)
 3  3 
 x  6

Questions: 1. What happens when you “swap” points?

2. What do the values p( x0 ) and p ( x1 ) show?

NOTE: Two points give a linear function (polynomial of degree 1). It follows that n 1
points give a polynomial of degree at most n.

2
Suppose that we have n 1 points ( x0 , f ( x0 )), ( x1 , f ( x1 )),..., ( xn , f ( xn )). In this case, we first

construct Ln ,k ( x), for k  0,1, 2,..., n with the property that

1, if i  k
Ln ,k ( xi )   .
0, if i  k

For example, from Example 4.1.1,

1
L1,0 ( x0 )  L1,0 (2)   (2  5)  1
3
1
L1,0 ( x1 )  L1,0 (5)   (5  5)  0
3
1
L1,1 ( x0 )  L1,1 (2)  (2  2)  0
3
1
L1,1 ( x1 )  L1,1 (5)  (5  2)  1.
3

To satisfy Ln,k ( xi )  0, i  k , requires that the numerator of Ln,k ( x) contain the term

( x  x0 )( x  x1 )...( x  xk 1 )( x  xk 1 )...( x  xn ),

and Ln ,k ( xi )  1, i  k , requires that the denominator be evaluated at x  xk . Thus,

( x  x0 )( x  x1 )...( x  xk 1 )( x  xk 1 )...( x  xn )
Ln ,k ( x)  ,
( xk  x0 )( xk  x1 )...( xk  xn )

for example, when n  2,

( x  x1 )( x  x2 )
L2,0 ( x) 
( x0  x1 )( x0  x2 )
( x  x0 )( x  x2 )
L2,1 ( x) 
( x1  x0 )( x1  x2 )
( x  x0 )( x  x1 )
L2,2 ( x)  .
( x2  x0 )( x2  x1 )

The nth Lagrange interpolating polynomial is defined in the following theorem:

Theorem 4.1.2

If x0 , x1 ,..., xn are n 1 distinct numbers and f is a function whose values are given at these

numbers, then a unique polynomial p( x) at most n exists with

3
 f ( xk )  p( xk ), for each k  0,1, 2,..., n.

This polynomial is given by

p ( x)  f ( x0 ) Ln ,0 ( x)  f ( x1 ) Ln ,1 ( x)  ...  f ( xn 1 ) Ln ,n 1 ( x)  f ( xn ) Ln ,n ( x)
n
  f ( xk ) Ln ,k ( x),
k 0

where for each k  0,1, 2,..., n,

( x  x0 )( x  x1 )...( x  xk 1 )( x  xk 1 )...( x  xn )
Ln ,k ( x) 
( xk  x0 )( xk  x0 )...( xk  xn )
n
( x  xi )
 .
i 0 ( xk  xi )
ik

The numbers x0 , x1 ,..., xn are called nodes.

Example 4.1.2

Use the nodes x0  2, x1  2.75 and x2  4 to find the second Lagrange interpolating

1 1
polynomial for f ( x )  and use it to approximate f (3)  .
x 3

Solution:

( x  x1 )( x  x2 ) ( x  2.75)( x  4) 2
L2,0 ( x)    ( x  2.75)( x  4)
( x0  x1 )( x0  x2 ) (2  2.75)(2  4) 3
( x  x0 )( x  x2 ) ( x  2)( x  4) 16
L2,1 ( x)     ( x  2)( x  4)
( x1  x0 )( x1  x2 ) (2.75  2)(2.75  4) 15
( x  x0 )( x  x1 ) ( x  2)( x  2.75) 2
L2,2 ( x)    ( x  2)( x  2.75)
( x2  x0 )( x2  x1 ) (4  2)(4  2.75) 5
 p( x)  f ( x0 ) L2,0 ( x)  f ( x1 ) L2,1 ( x)  f ( x2 ) L2,2 ( x)
1 2 4 16 1 2
 . ( x  2.75)( x  4)  .  ( x  2)( x  4)  . ( x  2)( x  2.75)
2 3 11 15 4 5
1 64 1
 ( x  2.75)( x  4)  ( x  2)( x  4)  ( x  2)( x  2.75)
3 165 10
1 2 35 49
 x  x .
22 88 44
At x  3,
1 35 49 29
f (3)  (32 )  (3)    0.32955.
22 88 44 88

4
Theorem 4.1.3

Suppose the nodes in the interval [ a, b] are distinct and f  C n1[a, b]. Then, for each x in
[a, b], a number  ( x ) (generally unknown) between the nodes, and hence in (a, b) exists
with

f ( n 1)  ( x) 
f ( x)  p( x)  ( x  x0 )( x  x1 )...( x  xn ),
(n  1)!

where p( x) is the nth Lagrange interpolating polynomial.

NOTE: The error formula given in Theorem 4.1.3 is similar to that given by the Taylor
Theorem, only that this error uses information at the distinct numbers x0 , x1 ,..., xn .

Example 4.1.3

Find the error formula for the second Lagrange interpolating polynomial found in Example
4.1.2 and find the maximum error when this polynomial is used to approximate f ( x ) for x
in [2, 4].

Solution:
1 1 2 6
f ( x)   f ( x)   2 , f ( x)  3 , f ( x)   4 .
x x x x
Thus,

f (3)  ( x) 
f ( x)  p ( x)  ( x  x0 )( x  x1 )( x  x2 ),
(2  1)!
where

f (3)  ( x)  6
( x  x0 )( x  x1 )( x  x2 )   ( x  2)( x  2.75)( x  4)
6  ( x) 
4
3!

   ( x)  ( x  2)( x  2.75)( x  4), for  ( x) in (2,4).

4

1
The maximum value of  ( x)  on (2,4) is 24 
4
. We now determine the maximum value
16
35 49
of the absolute value of the polynomial g ( x)  ( x  2)( x  2.75)( x  4)  x 3  x 2  x  22
4 2
35 49 7 7
g ( x)  0  3x 2  x   0  x  or x 
2 2 3 2
25 9
 g  73   or g  72    .
108 16

f (3)  ( x)  1 9 9
Hence, the maximum error is g ( x)     0.0352.
3! 16 16 256

5
4.2. DIVIDED DIFFERENCES

Suppose that pn ( x) is the nth Lagrange interpolating polynomial that agrees with the function

f at distinct numbers x0 , x1 ,..., xn . The divided differences of f with respect to x0 , x1 ,..., xn

are used to express pn ( x) in the form

pn ( x)  a0  a1 ( x  x0 )  a2 ( x  x0 )( x  x1 )  ...  an ( x  x0 )( x  x1 )...( x  xn1 )

for some constants a0 , a1 ,..., an . Note that

pn ( x0 )  a0  f ( x0 )
f ( x1 )  f ( x0 )
pn ( x1 )  a0  a1 ( x1  x0 )  f ( x1 )  a1  .
x1  x0

The zeroth divided difference of f with respect to xi , denoted f [ xi ], is f [ xi ]  f ( xi ).

The first divided difference of f with respect to xi and xi 1 , denoted f [ xi , xi 1 ], is

f [ xi 1 ]  f [ xi ]
f [ xi , xi 1 ]  .
xi 1  xi

The second divided difference is

f [ xi 1 , xi  2 ]  f [ xi , xi 1 ]
f [ xi , xi 1 , xi  2 ]  .
xi  2  xi

Similarly, the k th divided difference relative to xi , xi 1 , xi  2 ,..., xi  k , is

f [ xi 1 , xi  2 ,..., xi  k ]  f [ xi , xi 1 , xi  2 ,..., xi  k 1 ]
f [ xi , xi 1 , xi  2 ,..., xi  k ]  .
xi  k  xi

This process ends with a single nth divided difference

f [ x1 , x2 ,..., xn ]  f [ x0 , x1 ,..., xn 1 ]
f [ x0 , x1 ,..., xn ]  .
xn  x0

Thus, pn ( x) can be written as

pn ( x)  f [ x0 ]  f [ x0 , x1 ]( x  x0 )  ...  f [ x0 , x1 ,..., xn ]( x  x0 )( x  x1 )...( x  xn1 )

n
 f [ x0 ]   f [ x0 , x1 ,..., xk ]( x  x0 )( x  x1 )...( x  xk 1 ) (4.1)
k 1

6
Equation (4.1) is known as Newton’s Divided-Difference formula. The table below outlines
how divided differences from tabulated data can be determined:

x f ( x) 1st 2nd 3rd

x0 f  x0 
f [ x0 , x1 ]
x1 f  x1  f [ x0 , x1 , x2 ]
f [ x1 , x2 ] f [ x0 , x1 , x2 , x3 ]
x2 f  x2  f [ x1 , x2 , x3 ]
f [ x2 , x3 ]
x3 f  x3 

Example 4.2.1

Suppose we have the data (1, 2), (2, 4), (4,3), (5, 0). Then,

x0  1, x1  2, x2  4, x3  5 and f [ x0 ]  2, f [ x1 ]  4, f [ x2 ]  3, f [ x3 ]  0

x f ( x) 1st 2nd 3rd

1 f  x0   2
42
f [ x0 , x1 ]  2
2 1
 12  2 5
2 f  x1   4 f [ x0 , x1 , x2 ]  
4 1 6
3 4 1  56    56 
f [ x1 , x2 ]   f [ x0 , x1 , x2 , x3 ]  0
42 2 5 1
3    12  5
4 f  x2   3 f [ x1 , x2 , x3 ]  
52 6
03
f [ x2 , x3 ]   3
54
5 f  x3   0

7
p3 ( x)  f [ x0 ]  f [ x0 , x1 ]( x  x0 )  f [ x0 , x1 , x2 ]( x  x0 )( x  x1 )  f [ x0 , x1 , x2 , x3 ]( x  x0 )( x  x1 )( x  x2 )
5
 2  2( x  1)  ( x  1)( x  2)  0( x  1)( x  2)( x  4)
6
5 9 5
   x  x2 .
3 2 6
5 9 5
 p3 (3)    (3)  (32 )  5.16.
3 2 6

5 y  p3 ( x)

0 1 2 3 4 5 x

Example 4.2.2

Write the divided difference table for the given values of x and f ( x). Hence, find p4 ( x).

x f ( x)

1.0 0.7651977
1.3 0.6200860
1.6 0.4554022
1.9 0.2818186
2.2 0.1103623

Solution:

Here x0  1.0, x1  1.3, x2  1.6, x3  1.9, x4  2.2

f [ x0 ]  0.7651977, f [ x1 ]  0.6200860, f [ x2 ]  0.4554022, f [ x3 ]  0.2818186, f [ x4 ]  0.1103623

8
 f [ x1 ]  f [ x0 ] 0.6200860  0.7651977
f [ x0 , x1 ]    0.4837057
x1  x0 1.3  1.0
f [ x2 ]  f [ x1 ] 0.4554022  0.6200860
f [ x1 , x2 ]    0.5489460
x2  x1 1.6  1.3
f [ x3 ]  f [ x2 ] 0.2818186  0.4554022
f [ x2 , x3 ]    0.5786120
x3  x2 1.9  1.6
f [ x4 ]  f [ x3 ] 0.1103623  0.2818186
f [ x3 , x4 ]    0.571521
x4  x3 2.2  1.9
f [ x1 , x2 ]  f [ x0 , x1 ] 0.5489460  (0.4837057)
f [ x0 , x1 , x2 ]    0.1087338
x2  x0 1.6  1.0
f [ x2 , x3 ]  f [ x1 , x2 ] 0.5786120  (0.5489460)
f [ x1 , x2 , x3 ]    0.0494433
x3  x1 1.9  1.3
f [ x3 , x4 ]  f [ x2 , x3 ] 0.571521  (0.5786120)
f [ x2 , x3 , x4 ]    0.0118183
x4  x2 2.2  1.6
f [ x1 , x2 , x3 ]  f [ x0 , x1 , x2 ] 0.0494433  (0.1087338)
f [ x0 , x1 , x2 , x3 ]    0.0658783
x3  x0 1.9  1.0
f [ x2 , x3 , x4 ]  f [ x1 , x2 , x3 ] 0.0118183  (0.0494433)
f [ x1 , x2 , x3 , x4 ]    0.0680684
x4  x1 2.2  1.3
f [ x1 , x2 , x3 , x4 ]  f [ x0 , x1 , x2 , x3 ] 0.0680684  0.0658783
f [ x0 , x1 , x2 , x3 , x4 ]    0.0018251
x4  x0 2.2  1.0

x f ( x) 1st 2nd 3rd 4th

1.0 f  x0   0.7651977
f [ x0 , x1 ]  0.4837057
1.3 f  x1   0.6200860 f [ x0 , x1 , x2 ]  0.1087338
f [ x1 , x2 ]  0.5489460 f [ x0 , x1 , x2 , x3 ]  0.0658783
1.6 f  x2   0.4554022 f [ x1 , x2 , x3 ]  0.0494433 f [ x0 , x1 , x2 , x3 ]  0.0018251
f [ x2 , x3 ]  0.5786120 f [ x1 , x2 , x3 , x4 ]  0.0680684
1.9 f  x3   0.2818186 f [ x2 , x3 , x4 ]  0.0118183
f [ x3 , x4 ]  0.5715210
2.2 f  x4   0.1103623

 p4 ( x)  0.7651977  0.4837057( x  1.0)  0.1087338( x  1.0)( x  1.3)

 0.0658783( x  1.0)( x  1.3)( x  1.6)  0.0018251( x  1.0)( x  1.3)( x  1.6)( x  1.9)

9
4.2.1. Newton Forward- and Backward-Difference

Let us try to write equation (4.1) in a simplified form when the nodes are arranged
consecutively with equal spacing. Let h  xi 1  xi and x  x0  sh. Since the nodes are equally

spaced, we have that xi  x0  ih which implies that the difference x  xi is

x  xi  x0  sh  ( x0  ih)  (s  i )h.

Thus, equation (4.1) becomes

pn ( x)  pn ( x0  sh)  f [ x0 ]  shf [ x0 , x1 ]  s(s 1)h2 f [ x0 , x1, x2 ]  ...  s(s 1)...(s  n  1)hn f [ x0 , x1,..., xn ].

s 
Since s ( s  1)...( s  k  1)  k !  , we have that
k 
n
s 
pn ( x)  pn ( x0  sh)  f [ x0 ]     k !h k f [ x0 , x1 ,..., xk ] (4.2)
k 1  k 

Equation (4.2) is known as Newton Forward-Difference formula.

Newton Backward-Difference formula is obtained when the nodes are reordered from last to
first. In this case, (4.1) can be written in the form

pn ( x)  f [ xn ]  f [ xn , xn 1 ]( x  xn )  f [ xn , xn 1 , xn  2 ]( x  xn )( x  xn 1 )
 ...  f [ xn , xn 1 ,..., x0 ]( x  xn )( x  xn 1 )...( x  x1 ).

As observed above, since the nodes are equally spaced, letting x  xn  sh and

x  xi  ( s  n  i )h, given

pn ( x)  pn ( xn  sh)  f [ xn ]  shf [ xn , xn 1 ]  s ( s  1)h 2 f [ xn , xn 1 , xn  2 ]

 ...  s( s  1)...( s  n  1)h n f [ xn , xn 1 ,..., x0 ].

If we use

  s   s( s  1)...( s  k  1) s ( s  1)...(s  k  1)

   (1) k ,
k  k! k!

then the Newton Backward-Difference formula becomes

n
 s 
pn ( x)  pn ( x0  sh)  f [ xn ]     (1) k k !h k f [ xk , xk 1 ,..., x0 ] (4.3)
k 1  k 

10
NOTE: The Newton Forward- Difference formula is useful for interpolation near the beginning
of a set of tabular values and the Newton Backward- Difference formula for interpolation near
the end of a set of tabular values.

Example 4.2.3

Use Example 4.2.2 to approximate f (1.1) and f (2.0).

Solution:

Since x  1.1 is near the first value x0  1.0, we use The Newton Forward- Difference formula.

Here, h  1.3 1.0  1.6 1.3  1.9 1.6  2.2 1.9  0.3.

1
x  x0  sh  1.1  1.0  0.3s  s  .
3

 f (1.1)  p4 (1.1)  p4 (1.0  13 (0.3))

1 11 
 0.7651977  (0.3)(0.4837057)    1 (0.3)2 (0.1087338)
3 33 
1  1  1  1  1  1  1 
   1  2  (0.3)3 (0.0658783)    1  2   3  (0.3) 4 (0.0018251)
3  3  3  3  3  3  3 
 0.719646.

To approximate f (2.0), we use Newton Backward- Difference formula since x  2.0 is near
the last value x  2.2.

2
x  xn  sh  x4  sh  2.0  2.2  0.3s  s  
3

 f (2.0)  p4 (2.0)  p4 (2.2  23 (0.3))

2 2 2 
 0.1103623  (0.3)(0.5715210)     1 (0.3) 2 (0.0118183)
3 3 3 
2  2  2  2  2  2  2 
    1   2  (0.3)3 (0.0680684)     1   2    3  (0.3) 4 (0.0018251)
3  3  3  3  3  3  3 
 0.2238853.

11
4.3. SPLINE INTERPOLATION

One disadvantage of using a single polynomial to approximate an arbitrary function on a closed

interval is that high-degree polynomials can oscillate erratically, that is, a minor fluctuation
over a small portion of the interval can induce large fluctuations over the entire range. An
alternative approach is to divide the approximation interval into a collection of subintervals
and construct a (generally) different approximating polynomial on each subinterval. This is
called piecewise-polynomial or spline approximation. The splines, which will be denoted by
sn ( x), can be classified depending on the order of the polynomials. We will discuss linear,

quadratic and cubic splines. Other high-order splines can be derived in a similar way with a
greater increase in computational difficulty.

4.3.1. Linear Splines

The simplest spline is a linear spline which consists of joining a set of data points

( x0 , f ( x0 )), ( x1 , f ( x1 )),..., ( xn , f ( xn )), where x0  x1  ...  xn ,

by a series of straight lines such that

  x  x1   x  x0 
 s1 ( x)  f ( x0 )    f ( x1 )   , x  [ x0 , x1 ]
  x0  x1   x1  x0 
    
 s2 ( x)  f ( x1 )  x  x2   f ( x2 )  x  x1  , x  [ x1 , x2 ]
S ( x)    x1  x2   x2  x1 


  x  xn   x  xn 1 
 sn ( x)  f ( xn 1 )    f ( xn )   , x  [ xn 1 , xn ]
  xn 1  xn   xn  xn 1 

x0 x1 x2 x3 x4 xn

12
Example 4.3.1
Construct the linear spline interpolating the data below:

x 1 0 1
y 0 1 3

Solution:

 s1 ( x), x  [1, 0]
S ( x)  
 s2 ( x), x  [0,1]

where

 x  x1   x  x0   x0   x  (1) 
s1 ( x)  f ( x0 )    f ( x1 )    f (1)    f (0)  
 x0  x1   x1  x0   1  0   0  (1) 
x x 1
 (0)  (1)  x 1
1 1
 x  x2   x  x1   x 1   x0
s2 ( x)  f ( x1 )    f ( x2 )    f (0)    f (1)  
 x1  x2   x2  x1   0 1   1 0 
x 1 x
 (1)  (3)  1  2 x
1 1

1  x, x  [1, 0]
 S ( x)  
1  2 x, x  [0,1]

NOTE:

A drawback for linear splines is that S ( x) is generally discontinuous at each interior nodes xi .

For example, the derivative of S ( x) in Example 4.3.1 is discontinuous at x  0.

4.3.2. Quadratic Splines

Suppose we want to join the set of data

( x0 , f ( x0 )), ( x1 , f ( x1 )),..., ( xn , f ( xn )), where a  x0  x1  ...  xn  b,

by piecewise-quadratic polynomials. Then,

13
  s1 ( x)  a0  b0 x  c0 x 2 , x  [ x0 , x1 ]

 s ( x)  a1  b1 x  c1 x , x  [ x1 , x2 ]
2
S ( x)   2

 s ( x)  a  b x  c x 2 , x  [ x , x ]
 n n 1 n 1 n 1 n 1 n

To ensure that S ( x) interpolates the data, we set

si ( xi 1 )  f ( xi 1 ), i  1, 2,..., n
si ( xi )  f ( xi ).

To ensure that S ( x) is continuous and has continuous first-order derivative everywhere in

[a, b], we set

si ( xi )  si 1 ( xi ), i  1, 2,..., n  1.

Since each si has 3 unknown constants, S ( x) will give a total of 3n unknown coefficients.

The condition si ( xi )  si 1 ( xi ) imposes n 1 linear conditions and interpolation imposes 2n

linear conditions giving a total of 3n 1 imposed linear conditions. This means that the
conditions will give a system of 3n 1 linear equations in 3n unknown coefficients, which
may give infinite-many solutions. To get a unique solution, we need to impose another
condition. The problem is to determine what additional condition to impose to make the
solution unique, for example, f ( x0 )  S ( x0 ) or f ( xn )  S ( xn ). One way to avoid this

problem is to set the second derivative of s1 to zero, i.e. c0  0.

Example 4.3.2

Construct a quadratic spline interpolating (1, 0), (0,1) and (1,3).

Solution:
Since the interval [1,1] has been divided into three intervals, we must have

 s1 ( x)  a0  b0 x, x  [1, 0]
S ( x)  
 s2 ( x)  a1  b1 x  c1 x , x  [0,1]
2

To ensure that S ( x) interpolates the data, we have that

s1 ( x0 )  f ( x0 ), s1 ( x1 )  s2 ( x1 )  f ( x1 ) and s2 ( x2 )  f ( x2 )
 s1 (1)  f (1)  a0  b0 (1)  0  a0  b0

14
 s1 (0)  s2 (0)  a0  b0 (0)  a1  b1 (0)  c1 (02 )  a0  a1 and s1 (0)  f (0)  a0  a1  1
s2 (1)  f (1)  a1  b1 (1)  c1 (12 )  3  b1  c1  2

To ensure that S ( x) is continuous and has continuous first-order derivative everywhere in

[1,1], we have that

s1 ( x1 )  s2 ( x1 )  b0  b1  c1 x1  b0  b1  c1 (0)

Since b0  1, we get b0  b1  1 and b1  c1  2  c1  1.

1  x, x  [1, 0]
 S ( x)  
1  x  x , x  [0,1]
2

Check that S ( x) interpolates the data and that S  C1[1,1].

4.3.3. Cubic Splines

We consider a piecewise-polynomial approximation that uses cubic polynomials between

each successive pair of nodes.

Definition 4.3.1

Given a function f defined on [ a, b] and a set of nodes a  x0  x1  ...  xn  b, a cubic

spline interpolant S for f is a function that satisfies the following conditions:

1. S ( x) is a cubic polynomial, denoted S j ( x), on the interval [ x j , x j 1 ] for each

j  0,1, 2,..., n  1
2. S j ( x j )  f ( x j ) and S j ( x j 1 )  f ( x j 1 ) for each j  0,1, 2,..., n  1.
3. S j 1 ( x j 1 )  S j ( x j 1 ), for each j  0,1, 2,..., n  2
4. S j 1 ( x j 1 )  S j ( x j 1 ), for each j  0,1, 2,..., n  2
5. S j 1 ( x j 1 )  S j ( x j 1 ), for each j  0,1, 2,..., n  2
6. One of the following sets of boundary conditions is satisfied:
(i) S ( x0 )  S ( xn )  0 (Natural (or free) boundary)

(ii) S ( x0 )  f ( x0 ) and S ( xn )  f ( xn ) (Clamped boundary)

When natural boundary conditions occur, the spline is called Natural Spline and when
clamped boundary conditions occur, we have Clamped Spline.

15
Using Definition 4.3.1, we will construct cubic splines of the form

S j ( x)  a j  b j ( x  x j )  c j ( x  x j ) 2  d j ( x  x j )3

for each j  0,1, 2,..., n  1.

Example 4.3.3

1. Construct a natural cubic spline that passes through the points (1, 2), (2,3) and (3,5).
2. Construct a clamped cubic spline that passes through the points (1, 2), (2,3) (3, 5) and
has S (1)  2 and S (3)  1.

Solutions:

From the given data, we have that x0  1, x1  2, x2  3 , S0 ( x)  [1, 2] and S1 ( x)  [2,3] where

S0 ( x)  a0  b0 ( x 1)  c0 ( x 1)2  d0 ( x 1)3 and S1 ( x)  a1  b1 ( x  2)  c1 ( x  2)2  d1 ( x  2)3

so that

 S0 ( x), x  [1, 2]
S ( x)  
 S1 ( x), x  [2,3]

1. S (1)  f (1)  2  S0 (1)  a0  2

S (2)  f (2)  3  S0 (2)  a0  b0  c0  d 0  3  b0  c0  d 0  1. (i)

Also, S1 (2)  3  a1  3

S (3)  f (3)  5  S1 (3)  S1 ( x)  a1  b1  c1  d1  5  b1  c1  d1  2 (ii)

S0 (2)  S1 (2)  b0  2c0 (2  1)  3d 0 (2  1)  b1  2c1 (2  2)  3d1 (2  2)

 b0  b1  2c0  3d0  0 (iii)

S0 (2)  S1 (2)  2c0  6d0 (2  1)  2c1  6d1 (2  2)

 c0  c1  3d0  0 (iv)

Using natural conditions, we get

S ( x0 )  S ( xn )  0  S (1)  S (3)  0
 2c0  6d 0 (1  1)  0  c0  0
And 2c1  6d1 (3  2)  0  c1  3d1  0 (v)

16
 Solving the system formed by equations (i) - (v), we get
b0  d 0  1 1 0 0 1 0   b0   1 
b1  c1  d1  2     
0 1 1 0 1   b1   2 
b0  b1  3d 0  0  1 1 0 3 0   c1    0 
    
c1  3d 0  0 0 0 1 3 0   d 0   0 
c1  3d1  0 0 0 1 0 3   d1   0 


1 0 0 1 0 1 1 0 0 1 0 1  1 0 0 1 0 1
     
0 1 1 0 1 2 0 1 1 0 1 2 0 1 1 0 1 2
r3  r3  r1 r3  r2  r3
1 1 0 3 0 0  0 1 0 2 0 1  0 0 1 2 1 1
  r  r4  r5   r  13 r5  
0 0 1 3 0 0  5 0 0 1 3 0 0  5 0 0 1 3 0 0 
0 0 1 0 3 0  0 0 0 3 3 0  0 0 0 1 1 0 
  

1 0 0 1 0 1 1 0 0 1 0 1 
   
0 1 1 0 1 2 0 1 1 0 1 2 
r4  r3  r4  0 0 1 2 1 1  r5  r5  15 r4 0 0 1 2 1 1 
   
0 0 0 5 1 1 0 0 0 5 1 1 
0 0 0 1 1 0  0 0 0 0 1
5 5
4
 
4 1 1
 d1    d1  
5 5 4
1
5d 0  d1  1  d 0 
4
3
c1  2d 0  d1  1  c1 
4
3
b1  c1  d1  2  b1 
2
3
b0  d 0  1  b0 
4

2  4 ( x  1)  4 ( x  1) ,
3 1 3
x  [1, 2]
 S ( x)  
3  2 ( x  2)  4 ( x  2)  4 ( x  2) ,
 x  [2,3]
3 3 2 1 3

2. The first five conditions are as in part 1, i.e.

S0 (1)  a0  2
S0 (2)  S1 (2)  3  a0  b0  c0  d 0  3  b0  c0  d 0  1 and S1 (2)  3  a1  3
S1 (3)  5  a1  b1  c1  d1  5  b1  c1  d1  2

S0 (2)  S1 (2)  b0  b1  2c0  3d0  0 and S0 (2)  S1 (2)  c0  c1  3d0  0

17
Using clamped condition, we get
S ( x0 )  f ( x0 )  S (1)  2  b0  2c0 (1  1)  3d0 (1  1)  2  b0  2

And S ( xn )  f ( xn )  S (3)  1  b1  2c1 (3  2)  3d1 (3  2)  1  b1  2c1  3d1  1.

Thus,
 01 0 1 0   b1   1   01 0 1 0 1   10 1 01 2 
  c   2     
 10 1 0 1  0     10 1 0 1 2  01 0 1 0 1 
1 2 0 3 0   c1    2     1 2 0 3 0 2  r1  r2 1 2 0 3 0 2 
        
 01 1 3 0   d 0   0   01 1 3 0 0   01 1 3 0 0 
 10  
2 0 3   d1   1   10 2 0 3 1   10 2 0 3 1 
  
1 0 1 0 1 2 1 0 1 0 1 2
   
r3  r3  r1  0 1 0 1 0 1  0 1 0 1 0 1
r3  r3  2r2 r4  r3  r4
r5  r5  r1  0 2 1 3 1 0 0 0 1 1 1 2
  r  r5  r4   r  12 r5
r4  r2  r4  0 0 1 2 0 1  5 0 0 1 2 0 1 5
0 0 1 1 0 0 
 0 2  0 0 2 2

1 0 101 2 1 0 1 0 1 2
   
0 1 01 0 1  0 1 0 1 0 1 
0 0 11 1 2  r5  r5  13 r4 0 0 11 1 2
   
0 0 03 1 3 0 0 0 3 1 3
0 0  0 0 0 0 23 1
 0 01 1 

2 3
 d1  1  d1  
3 2
3
3d 0  d1  3  d 0 
2
c1  d0  d1  2  c1  2

5
c0  d 0  1  c0  
2
3
b1  c1  d1  2  b1 
2


2  2( x  1)  2 ( x  1)  2 ( x  1) ,
5 2 3 3
x  [1, 2]
 S ( x)  
3  2 ( x  2)  2( x  2)  2 ( x  2) ,
 x  [2,3]
3 2 3 3

18
4.4. LEAST-SQUARES APPROXIMATION

Suppose that a mathematical equation is to be fitted to experimental data by plotting the data
on a graph paper and then passing a line through the data points. The method of least squares
endeavours to determine the best approximating line by minimising the sum of the squares of
error, i.e. if ( xi , yi ), i  1, 2,3,..., n are data points and Y  f ( x) is the curve to be fitted to
this data, then the error is

ei  yi  f ( xi )

and the sum of the squares of the errors is

S  [ y1  f ( x1 )]2  [ y2  f ( x2 )]2  ...  [ yn  f ( xn )]2 . (4.4)

Then the method of least-squares involves minimising S .

Straight line: Let Y  a0  a1 x be the straight line to be fitted to the given data. Then,

S  [ y1   a0  a1 x1 ]2  [ y2   a0  a1 x2 ]2  ...  [ yn   a0  a1 xn ]2

and minimising S implies that

S
 0  2[ y1   a0  a1 x1 ]  2[ y2   a0  a1 x2 ]  ...  2[ yn   a0  a1 xn ]  0
a0

n
 2 [ yi   a0  a1 xi ]  0 (4.5)
i 1

and

S
 0  2 x1[ y1   a0  a1 x1 ]  2 x2 [ y2   a0  a1 x2 ]  ...  2 xn [ yn   a0  a1 xn ]  0
a1

n
 2 xi [ yi   a0  a1 xi ]  0. (4.6)
i 1

Simplifying (4.5) and (4.6) leads to what are known as normal equations:
n n n n
2 [ yi   a0  a1 xi ]  0   yi   a0  a1  xi  0
i 1 i 1 i 1 i 1

n n
  yi  na0  a1  xi  0 (4.7)
i 1 i 1

19
and
n n n n
2 xi [ yi   a0  a1 xi ]  0   xi yi  a0  xi  a1  xi 2  0
i 1 i 1 i 1 i 1

n n n
  xi yi  a0  xi  a1  xi 2 (4.8)
i 1 i 1 i 1

Solving (4.7) and (4.8) simultaneously, we get

n
 n  n 
n xi yi    xi   yi 
a1  i 1  i 1  i 1 
2
 n 2  n 
n   xi     xi 
 i 1   i 1 
 n 2  n   n  n 
  i   i    i i   xi 
x y  x y
a0   i 1  i 1   i 1  i 1  .
2
 n 2  n 
n   xi     xi 
 i 1   i 1 

n n

 xi y i
Letting x  i 1
and y  i 1
, we can rewrite the formulas for a0 and a1 as
n n
n

 x y  n.x . y
i i
a1  i 1
n

x
i 1
i
2
 n.x 2

 n   n 
y   xi 2   x   xi yi 
a0   i 1 n   i 1 .
 xi 2  n.x 2 i 1

2S 2S
Clearly,  0 and  0 showing that these values of a0 and a1 provide a minimum
a0 2 a12
of S .
Example 4.4.1

The table below gives the temperature T (in o C ) and length l (in mm) of a heated rod. If
l  a  bT , find the best values of a and b.
T 20 30 40 50 60 70
l 800.3 800.4 800.6 800.7 800.9 801.0

20
Solution:

T l T2 Tl
20 800.3 400 16006
30 800.4 900 24012
40 800.6 1600 32024
50 800.7 2500 40035
60 800.9 3600 48054
70 801.0 4900 56070
Sum 270 4803.9 13900 216201

 n  6,  T  270,  l  4803.9,  T 2
 13900,  Tl  216201
  
n Tl    T   l 
b     6(216201)  (270)(4803.9)  0.014571428  0.0146
 2  
2
6(13900)  (270) 2
n  T    T 
   
 2     
  T   l     Tl   T  (13900)(4803.9)  (216201)(270)
a       799.9942857  800
 2  
2
6(13900)  (270) 2
n  T    T 
   

 l  800  0.0146T

Polynomial of the k th degree: Let Y  a0  a1 x  a2 x 2  ...  ak x k be the polynomial of the

k th  order to be fitted to the data ( xi , yi ), i  1, 2,..., n. Then, we minimise

S  [ y1   a0  a1 x1  a2 x12  ...  ak x1k ]2  [ y2   a0  a1 x2  a2 x2 2  ...  ak x2 k ]2 

...  [ yn   a0  a1 xk  a2 xk 2  ...  ak xk k ]2

S
by setting  0, for i  0,1, 2,..., k. This leads to the following k  1 normal equations in
ai
k  1 unknowns:

21
 n n n n
na0  a1  xi  a2  xi 2  ...  ak  xi k   yi
i 1 i 1 i 1 i 1
n n n n n
a0  xi  a1  xi 2  a2  xi 3  ...  ak  xi k 1   xi yi
i 1 i 1 i 1 i 1 i 1

n n n n n
a0  xi k  a1  xi k 1  a2  xi k  2  ...  ak  xi 2 k   xi k yi
i 1 i 1 i 1 i 1 i 1

Example 4.4.2
Fit a polynomial of the second degree to the data points given below:
x 0 1.0 2.0
y 1.0 6.0 17.0
Solution:

Let Y  a0  a1 x  a2 x 2 , where n  3 and k  2 . The normal equations are

n n n n n n
na0  a1  xi  a2  xi 2   yi 3a0  a1  xi  a2  xi 2   yi
i 1 i 1 i 1 i 1 i 1 i 1
n n n n n n n n
a0  xi  a1  xi 2  a2  xi 3   xi yi  a0  xi  a1  xi 2  a2  xi 3   xi yi
i 1 i 1 i 1 i 1 i 1 i 1 i 1 i 1
n n n n n n n n
a0  xi k  a1  xi k 1  a2  xi 2 k   xi k yi a0  xi 2  a1  xi 3  a2  xi 4   xi 2 yi
i 1 i 1 i 1 i 1 i 1 i 1 i 1 i 1

x y x2 x3 x4 xy x2 y
0 1 0 0 0 0 0
1 6 1 1 1 6 6
2 17 4 8 16 34 68

: 3 24 5 9 17 40 74

Using normal equations, we get a system of linear equations

3a0  3a1  5a2  24

3a0  5a1  9a2  40
5a0  9a1  17a2  74

whose solution is a0  1, a1  2 and a2  3.

 Y  1  2 x  3x 2

22
Exponential function: Let the curve y  a0ea1x be fitted to the data ( xi , yi ), i  1, 2,..., n.
Then, taking natural logarithms on both sides, we get

 
ln y  ln a0e a1x  ln a0  a1 x,

which can be written as

Y  A0  A1 x,

where Y  ln y, A0  ln a0 , A1  a1 and this can be treated as a straight line problem.

Example 4.4.3

Fit an exponential function to the data given in the table below:

x 1.00 1.25 1.50 1.75 2.00

y 5.10 5.79 6.53 7.45 8.46

Solution:

We transform y  a0e 1 into Y  A0  A1 x, where Y  ln y, A0  ln a0 , A1  a1 and n  5.

ax

5 A0  A1  xi  Yi  ln yi

A0  xi  A1  xi 2   xiYi   xi ln yi

x y ln y x2 x ln y
1.00 5.10 1.629 1.00 1.629
1.25 5.79 1.756 1.5625 2.195
1.50 6.53 1.876 2.2500 2.814
1.75 7.45 2.008 3.0625 3.514
2.00 8.46 2.135 4.0000 4.270

: 7.5 9.404 11.875 14.422

5 A0  A1  xi  ln yi
5 A0  7.5 A1  9.404

7.5 A0  11.875 A1  14.422
A0  xi  A1  xi 2   xi ln yi

23
 9.404  5 A0  9.404  5 A0 
A1   7.5 A0  11.875    14.422  A0  1.1224
7.5  7.5 
Since A0  ln a0 , we have that a0  e1.1224  3.072

9.404  5(1.1224)
 A1  a1   0.5056
7.5
 y  3.072e0.5056 x

4.4.1. Weighted Least-Squares Approximation

If the given data is not of equal quality, the fit by minimising the sum of squares of the errors
may not be very accurate. To improve the fit, a more general approach is to minimise the
weighted sum of squares of the errors taken over all data points

S  w1[ y1  f ( x1 )]2  w2 [ y2  f ( x2 )]2  ...  wn [ yn  f ( xn )]2

 w1e12  w2e2 2  ...  wnen 2 .

The wi 's are prescribed positive numbers and are called weights. A weight is prescribed
according to the relative accuracy of a data point. If all the data points are accurate, we set
wi  1 for all i.

Example 4.4.4

In Example 4.4.1 we got the linear fit l  800  0.0146T . Suppose that the point (60,800.9)

is known to be more reliable than the others. Then we prescribe a weight (say 10)
corresponding to this point only and set wi  1, for all other points so that

n n n
a0  wi  a1  wiTi   wili
i 1 i 1 i 1
n n n
a0  wiTi  a1  wiTi 2   wiTi li
i 1 i 1 i 1

24
 T l w T2 wT wl Tl wT 2 wTl
20 800.3 1 400 20 800.3 16006 400 16006
30 800.4 1 900 30 800.4 24012 900 24012
40 800.6 1 1600 40 800.6 32024 1600 32024
50 800.7 1 2500 50 800.7 40035 2500 40035
60 800.9 10 3600 600 8009.0 48054 36000 480540
70 801.0 1 4900 70 801.0 56070 4900 56070

: 270 4803.9 15 13900 810 12012 216201 46300 648687

n n n
a0  wi  a1  wiTi   wili
i 1 i 1 i 1 15a0  810a1  12012

n n n
810a0  46300a1  648687
a0  wiTi  a1  wiTi 2   wiTi li
i 1 i 1 i 1

12012  15a0
a1 
810
 12012  15a0 
 810a0  46300    648687  a0  800
 810 
12012  15(800)
 a1   0.0148
810
 l  800  0.0148T

Note that the first fit gives

l (60)  800  0.0146(60)  800.876

and the weighted fit gives

l (60)  800  0.0148(60)  800.888

Note that the approximation becomes better when the weight is increased.

4.4.2. Least-Squares for Continuous Functions

We discuss the least squares approximation of a continuous function on an interval [a, b].
Let Y ( x)  a0  a1 x  ...  an x n be chosen to minimise

S   w( x)  y ( x)   a0  a1 x  ...  an x n   dx
b 2

25
The necessary conditions for a minimum yield

2  w( x)  y ( x)   a0  a1 x  ...  an x n  dx  0
b

2  x.w( x)  y ( x)   a0  a1 x  ...  an x n  dx  0

b

2  x n w( x)  y ( x)   a0  a1 x  ...  an x n  dx  0
b

Rearrangement of terms gives the following normal equations:

b b b b b
a0  w( x)dx  a1  x.w( x)dx  a2  x 2 .w( x)dx  ...  an  x n .w( x)dx   w( x) y ( x)dx
a a a a a
b b b b b
a0  x.w( x)dx  a1  x 2 .w( x)dx  a2  x 3 .w( x)dx  ...  an  x n1.w( x )dx   x.w( x ) y ( x )dx
a a a a a

b b b b b
a0  x n .w( x)dx  a1  x n1.w( x)dx  a2  x n 2 .w( x)dx  ...  an  x 2 n .w( x)dx   x n .w( x) y ( x)dx
a a a a a

Example 4.4.5
Find the least-squares approximating polynomial of degree two for the function f ( x)  sin  x
on the interval [0,1] with respect to the weight function w( x)  1.
Solution:

The normal equations for Y ( x)  a0  a1 x  a2 x 2 are

1 1 2
1 1 1
a0  dx  a1  x dx  a2  x 2 dx   sin  xdx
1
a0  a1  a2 
0 0 0 0 2 3 
 1 1 1 2
1 1 1 1
a0  x dx  a1  x 2 dx  a2  x 3 dx   x sin  xdx a0  a1  a2 
0 0 0 0
2 3 4 
1 1 1 1
a0  x 2 dx  a1  x 3 dx  a2  x 4 dx   x 2 sin  xdx 1 1 1 2 4
0 0 0 0 a0  a1  a2 x 
3 4 5 3
Solving this system gives

12 2  120
a0   0.050465
3
720  60 2
a1   4.12251
3
60 2  720
a2   4.12251
3
 Y ( x)  0.050465  4.12251x  4.12251x 2

26
4.5. CHEBYSHEV POLYNOMIALS AND ECONOMISATION OF POWER SERIES

Let f1 , f 2 ,..., f n be values of the given function and 1 ,2 ,...,n be the corresponding values

of the approximating function. Then the error vector e, where the components of e are given
by ei  fi  i . The approximation may be chosen using least-squares method or may be

chosen in such a way that the maximum component of e is minimised. The later method leads
to Chebyshev polynomials.

4.5.1. Chebyshev Polynomials

The Chebyshev polynomial of degree n over the interval [ 1,1] is defined by

Tn ( x)  cos(n cos1 x)

from which we get the relation

Tn ( x)  T n ( x).

Letting cos1 x   implies that cos  x so that

Tn ( x)  cos n .

Hence, T0 ( x)  1 and T1 ( x)  x. Using the trigonometric identity

cos  (n  1)   cos (n  1)   2cos  n  cos 

we have that

Tn1 ( x)  Tn1 ( x)  2 xTn ( x)

 Tn1 ( x)  2 xTn ( x)  Tn 1 ( x)

which is the recurrence relation that can be used to complete successively all Tn ( x) since we
know T0 ( x) and T1 ( x).

T0 ( x)  1
T1 ( x)  x
T2 ( x)  2 xT1 ( x)  T0 ( x)  2 x 2  1
T3 ( x)  2 xT2 ( x)  T1 ( x)  2 x  2 x 2  1  x  4 x 3  3 x
T4 ( x)  8 x 4  8 x 2  1
T5 ( x)  16 x5  20 x 3  5 x
T6 ( x)  32 x 6  48 x 4  18 x 2  1

27
The graphs of the first four Chebyshev polynomials are:

Note that the coefficient of x n in Tn ( x) is always 2 n1 and Tn ( x)  1, for n  1,  1  x  1.

If Pn ( x) is a monic polynomial such that Pn ( x)  21n Tn ( x), then Pn ( x) has the least upper

bound 21 n since Tn ( x)  1 . Thus, in Chebyshev approximation, the maximum error is kept

down to a minimum.

It is possible to express powers of x in terms of Chebyshev polynomials. Then

1  T0 ( x)
x  T1 ( x)
1
x 2  [T0 ( x)  T2 ( x)]
2
1
x3  3T1 ( x)  T3 ( x) 
4
1
x 4  3T0 ( x)  4T2 ( x)  T4 ( x) 
8
1
x5  10T1 ( x)  5T3 ( x)  T5 ( x) 
16
1
x 6  10T0 ( x)  15T1 ( x)  6T4 ( x)  T6 ( x) 
32

Chebyshev polynomials can be used to reduce the degree of an approximating polynomial

with a minimal loss of accuracy. This is known as economisation of power series.

28
Example 4.5.1

Economise the power series

x3 x5 x7
a) sin x  x   
6 120 5040
while keeping the error less than 0.005.
x 2 x3 x 4
b) e x  1  x   
2! 3! 4!
while keeping the error less than 0.05.

Solution:

1
a) Since  0.000198... is the first value which is numerically less than 0.005,
5040
we have that
x3 x5
sin x  x   .
6 120
We now express P5 ( x) in terms of Chebyshev polynomials.

1 1 1 1
Pn ( x)  T1 ( x)  . 3T1 ( x)  T3 ( x)   . 10T1 ( x)  5T3 ( x)  T5 ( x) 
6 4 120 16
169 5 1
 T1 ( x)  T3 ( x)  T5 ( x).
192 128 1920
1
Since  0.00052083  0.005, the economised power series is
1920

P3 ( x) 
169
192
T1 ( x) 
5
128
T3 ( x) 
169
192
x
5
128
 4 x 3  3x 

383 5
 x  x3 .
384 32

b) Remember that the upper bound of the error is

f (5) ( ( x)) x 5 e
R4 ( x)    0.023 for 1  x  1.
5! 120

We now express P4 ( x) in terms of Chebyshev polynomials.

29
 1 1 1 1 1 1 1
P4 ( x)  T0 ( x)  T1 ( x)  T0 ( x)  T2 ( x)  T1 ( x)  T3 ( x)  T0 ( x)  T2 ( x)  T4 ( x)
4 4 8 24 64 48 192
81 9 13 1 1
 T0 ( x)  T1 ( x)  T2 ( x)  T3 ( x)  T4 ( x)
64 8 48 24 192
1 1
But T4 ( x)   0.0053. Thus,
192 192

1
R4 ( x)  T4 ( x)  0.023  0.0053  0.0283  0.05.
192
Also,

1
R4 ( x)  T3 ( x)  0.023  0.04173  0.0647  0.05.
24

Therefore,

81 9 13 1
P3 ( x)  T0 ( x)  T1 ( x)  T2 ( x)  T3 ( x)
64 8 48 24

is the lowest-degree polynomial possible, i.e.

191 13 2 1 3
P3 ( x )  x x  x
192 24 6

THE END!

30