APM2613/LN5/0/2023

Tutorial Letter LN5/0/2023

Numerical Analysis I
APM2613

Year Module

Mathematical Sciences Department

This tutorial letter contains important information about …

Polynomial Interpolation and Approximation (Lesson 5)
It contains additional notes to supplement prescribed content in Chapter 3 and 8.1 of the
textbook. The notes should be read alongside relevant chapters in the textbook.

BARCODE

Open Rubric
1 Lesson 5: Polynomial Interpolation and Approximation
1.1 Objectives
The objectives of this Lesson are:
• To highlight further useful theory and properties of functions;

• To highlight the study of various interpolating polynomials as a technique to approximate the

value of a function at a given point within an interval;

• To highlight the study of using least-squares functions as a technique to approximate the

value of a function at a given point within an interval.

1.2 Learning Outcomes

At the end of this topic you should be able
1. To use the following interpolating polynomials to approximate the value of a function whose
values are given at certain points, the nodes/data points:

• Lagrange interpolating polynomial of any degree

• Neville’s tabular form of interpolating polynomials of various degrees
• Newton’s difference formulas of various degrees
• Hermite osculating polynomials of various degrees
• Parametric form of interpolating polynomials of various degrees: Cubic splines and
Bezier curves.

2. To compute the actual error incurred in using the interpolating polynomial approximations at
a given point;

3. To approximate the error in the approximation by the Lagrange interpolating polynomial;

4. To use the following functions to approximate a function whose values are tabulated at certain
points in the least-squares sense:

• Discrete least-squares polynomial of different degrees

• The exponential functions of the form y = beax and y = bxa ;

5. To compute the actual error in the least-squares approximations

1.3 Introduction
It is common that data is generated from experiments, observations or any such activity that gen-
erates data or tabulated points, (x0 , f (x0 )), (x1 , f (x1 )), . . . , (xn , f (xn )). Since functions are used
to summarise relationships associated with points, it would be desirable to obtain estimations of
values that are not in the collected data. The function f (x) would normally not be known. Two
contexts of using other functions to approximates a function represented by data are discussed:
interpolation and least-squares approximation. Attributed to their nice functional properties of con-
tinuity, differentiability, and integrability polynomials once again are popularly first candidates in the

2
 APM2613/LN05/0/2022

search for the approximating functions.

Interpolation is discussed first (Chapter 3 in the textbook). This technique is about finding a func-
tional expression that approximates the true function whose data are given and require that the
approximating function coincides with the data points or tabulated points. Hence the main focus
of interpolation is to find polynomials of suitable degree that interpolate the given data. Different
forms of polynomials given different names are presented in the textbook, but the bottom line is
that they are just different ways of expressing the same polynomial. The fundamental theorem on
which all the formulation is based is the Weierstrass Approximation Theorem (Theorem 3.1).

Least squares approximation (Chapter 8.1) on the other hand relaxes the requirement that the
approximating polynomial equals the tabulated functional values but only requires that the approx-
imating function minimises the total error in its deviations from the tabulated points in the square
sense. Here the approximating function is extended to other classes than polynomials.

The chapter begins with the Lagrange form of interpolating polynomials, then tackles the Neville’s
form and divided difference forms. All these are standard forms of the interpolating polynomial.
Section 3.4 onwards presents the so called osculating polynomials. These are characterised by
the fact that they do not only coincide with the function f (x) at the nodes x0 , x1 , . . . , xn but also have
the derivatives up to order less than or equal to n that coincide with the same order derivatives of
the function.

1.4 Polynomial Interpolation

The discussion in this section forms the foundation of the thinking behind different forms of the
interpolating polynomials, before specific methods as presented in the textbook are presented.

Although normally more than one data points would be used to construct interpolating functions
(polynomials), the basic building block of polynomials is the linear polynomial. We know that we
need only two points to construct/define a line, so only two of the data points would be needed.
Without going into the implications of using only two points out of the many, let us start with the
linear polynomial to illustrate the notion of polynomial interpolation.

First we consider passing a first degree polynomial (or straight line) through any two successive
tabulated points (xk , f (xk )) and (xk+1 , f (xk+1 )) to approximate f (x̂), x ∈ [xk , xk+1 ].
The equation of a straight line that passes through (xi , f (xi )) and (xk+1 , f (xk+1 )) is

y − f (xk ) f (xk+1 ) − f (xk )

=
x − xk xk+1 − xk
or
f (xk+1 ) − f (xk )
f (x) = y = f (xk ) + (x − xk ) = P1 (x) (1)
xk+1 − xk
where by P1 (x) we denote a first degree interpolating polynomial in x. So that at x = x̂, the
interpolated value is
f (xk+1 ) − f (xk )
f (x̂) = f (xk ) + (x̂ − xk )
xk+1 − xk

3
We note that if in the interpolating polynomial P1 (x) we let x = xk , a tabulated node, then
P1 (x) = P1 (xk ) = f (xk ) + 0 = f (xk )
and in the same way x = xk+1 yields
f (xk+1 ) − f (xk )
P1 (x) = P (xk+1 ) = f (xk ) + (xk+1 − xk )
xk+1 − xk
= f (xk+1 )
and thus P1 (x) is said to interpolate exactly at xk and xk+1 .

Rearranging 1 above as
   
f (xk+1 ) − f (xk ) f (xk+1 ) − f (xk )
P1 (x) = x + f (xk ) − xk .
xk+1 − xk xk+1 − xk
which is in turn written as
P1 (x) = a1 x + a0 ,
a linear expression in x, and hence the term linear interpolation. Once more the requirement that
the straight line passes through (xk , f (xk )) and (xk+1 , f (xk+1 )) can easily be confirmed for this lin-
ear equation form.

Suppose now that a second degree polynomial a2 x2 + a1 x + a0 was used to interpolate the data.
Then f (x) would be required to coincide with Pn (x) = P2 (x) at three points, say xi , xk+1 adn xk+2 .
The system of equations resulting from this requirement is
P2 (xk ) = a2 x2k + a1 xk + a0 = f (xk )
P2 (xk+1 ) = a2 x2k+1 + a1 xk+1 + a0 = f (xk+1 )
P2 (xk+2 ) = a2 x2k+2 + a1 xk+2 + a0 = f (xk+2 )
which when solved yields the values of the coefficients a0 , a1 , and a2 .

In general, if m out of n + 1 points are chosen, a polynomial of degree m − 1 could be passed

through these points, to obtain an (m − 1)-th degree interpolating polynomial
Pm−1 (x) = am−1 xm−1 + am−1 xm−2 + · · · + a2 x2 + a1 x + a0
The main task is to evaluate the coefficients a0 , a1 , . . . , am−1 . Using methods such as undetermined
coefficients or any suitable method for solving linear systems to solve for these coefficients can be
cumbersome for higher degree polynomials.

The focus of this topic is in presenting other, hopefully more efficient ways to construct the interpo-
lating polynomial. Beginning with the Lagrange interpolating polynomial, other forms are presented
and discussed in the sequel.

1.5 Various Forms of the Interpolating Polynomial

The textbook the interpolating polynomial in various forms, starting with the Lagrange form and
discussing other forms as well. The gist of the discussions is the notion of interpolation, meaning
the polynomials coincide with the function being approximated at selected points called nodes.
For the same degree polynomial, it can be observed that the approximations obtained using the
different forms are the same.

4
 APM2613/LN05/0/2022

1.5.1 Lagrange Interpolating Polynomial

Here is how the Lagrange form of the linear interpolating polynomial is derived. We start with the
linear polynomial
P1 (x) = a1 x + a0
and impose the interpolation requirement f (xi ) = P1 (xi ) at two given nodes x0 and x1 to obtain the
equations

P1 (x) − a1 x − a0 = 0
f (x0 ) − a1 x0 − a0 = 0
f (x1 ) − a1 x1 − a0 = 0

From the results of linear algebra we know that this system of equations has a non-trivial solution
only if the determinant of the system is zero. That is, if

P1 (x) −x −1
f (x0 ) −x0 −1 =0 (2)
f (x1 ) −x1 −1

Expanding on the first column yields

P1 (x)(x0 − x1 ) − f (x0 )(x − x1 ) + f (x1 )(x − x0 ) = 0.

Solving for P1 (x) yields

f (x0 )(x − x1 ) − f (x1 )(x − x0 )

P1 (x) = =
x0 − x1
x − x1 x − x1
= f (x0 ) + f (x1 )
x0 − x1 x1 − x0
which can be written in a more general form as

P1 (x) = f (x0 )L0 (x) + f (x1 )L1 (x) (3)

where
x − x1 x − x0
L0 (x) = , L1 (x) =
x0 − x1 x x − x0
are the coefficients associated with f (x0 ) and f (x1 ), respectively. This is the first degree Lagrange
interpolating polynomial.

NB: The textbook notation, Ln,k (x) is used to denote the coefficient associated with the node xk .
A similar evaluation of a higher order determinant would yield similar expressions for the corre-
sponding polynomial. Using the pattern observed for the coefficients of the linear polynomial, the
general n-th degree Lagrange interpolating polynomial through n + 1 data points

{(x0 , f (x0 )), (x1 , f (x1 )), (x2 , f (x2 )), . . . , (xn , f (xn ))}

is written as

Pn (x) = f (x0 )L0 (x) + f (x1 )L1 (x) + f (x2 )L2 (x) + · · · + f (xn )Ln (x) (4)

5
In order to ensure that at each nodes point xk , the requirement
Pn (xk ) = f (xk )
is met, the coefficients L are chosen so that at xk only Lk (x) is 1 and the other L’s are zero. Hence
the general expression for Lk (x) is
(x − x0 )(x − x1 )(x − x2 ) . . . (x − xk−1 )(x − xk+1 ) . . . (x − xn )
Lk (x) = (5)
(xk − x0 )(xk − x1 )(xk − x2 ) . . . (xk − xk−1 )(xk − xk+1 ) . . . (xk − xn )
We note the following about the above expression:
• The factor (x − xk ) is missing in the expression of Lk (x);
• At any xj such that xj ̸= xk , one of the factors in the numerator is zero, and Lk (x) is zero;
• At x = xk . on the other hand, the numerator and denominator are exactly equal, and Lk (x) =
1;
• Since the factor (x − xk ) is missing for Lk (x), its numerator is an n-th degree polynomial in x;
thus
• since x does not appear in the denominator, each Lk (x) itself is an n-th degree polynomial in
x, and so is Pn (x).
It is also worth noting that the two forms of the interpolating polynomial represent the same poly-
nomial. There are other forms as well but the interpolating polynomial of degree n is unique.

1.5.2 Neville’s Form of the Interpolating Polynomial

For a given set of nodes x0 , x1 , . . . , xn the method starts with two sets of consecutive pairs of nodes,
x0 , x1 and x1 , x2 and finding linear approximating polynomials through these pairs of nodes. Higher
degree approximating are generated recursively by adding previously computed polynomial values
multiplied by factors involving the other higher indexed nodes.
The first set of polynomial values of zero degree are the tabulated functional values. Again since
the approximations are based on the tabulated values, accuracy of the approximations are deter-
mined by comparing the change in the approximated values.

The method is discussed with examples on pages 116 - 120.

1.5.3 Newton’s Difference Forms of the Interpolating Polynomial

On pages 122-130 the textbook discusses several difference-formulas for the interpolating poly-
nomial. Four divided difference formulae are presented, with three attributed to Newton and one
attributed to Stirling: Divided Difference formula (3.10), Newton Forward Difference Formula (3.12),
Newton Backward Difference Formula (3.13) and Stirling’s Centered-difference Formula. For the
Newton’s formulas, starting with the divided-difference formula the discussion goes on to present
adaptations of this formula to forward-difference and backward-difference formulas, as well as the
central difference formula. These are different ways of writing the Newton difference interpolating
polynomial.All these formulae can be based on the same values computed using divided differ-
ences defined in (3.9). The difference is in the way that the values from the table are read (see
Table 3.12 and 3.13):

6
 APM2613/LN05/0/2022

• Forward-difference: From top left diagonally to bottom right;

• Backward-difference: From bottom left diagonally to top right;

• Centered-difference: From centre to the right using 1 or 2 differences across the table.

Formula (3.9) suggests that you can start with any node xi , depending on the degree of interest of
the interpolating polynomial, and move diagonally for the respective differences. The denominator
in f [xi , xi+1 , . . . , xi+k ] in (3.9) is xi+k − xi and the general interpolating polynomial is (3.5):

Pn (x) = a0 + a1 (x − x0 ) + a2 (x − x0 )(x − x1 ) + · · · + an (x − x0 )(x − x1 ) . . . (x − xn−1 ),

with ak = f [x0 , x1 , . . . , xk ]. In these formulae for the case of equally spaced nodes with xi+1 −xi = h,
the coefficients in the expressions make use of the value s calculated from the the target value
x = x0 + sh, which solves to give s = (x − x0 )/h.

NB: The value of the difference f (x0 , x1 , . . . , xn ) is independent of the order of the numbers x0 , x1 , . . . , xk
for the data points; i.e. the nodes do not have to be in increasing order to be used in the formulae.

Note also the formulae for computing the respective binomial coefficients:

Forward-difference formula:  
s s(s − 1) . . . (s − k + 1)
=
k k!
Backward-difference formula:
 
s s(s + 1) . . . (s + k − 1)
= (−1)k
k k!

As to which points to use for a desired degree polynomial, the aim is to use a formula that makes the
earliest use possible use of the data points closest to x and make use of the n-th divided difference;
e.g. if x is closer to the lower/upper end of the nodes, use the forward/backward difference formula;
if x is near the centre, the centered-difference formula may be more appropriate.

Understanding the s-k form of the Newton’s difference formulas

The basic formula for the difference formulas with respect to x0 , x1 , . . . , xn is given in (3.5) as

Pn (x) = a0 + a1 (x − x0 ) + a2 (x − x0 )(x − x1 ) + · · · + an (x − x0 ) . . . (x − xn−1 ).

or n
X
Pn (x) = ak (x − x0 )(x − x1 ) . . . (x − xk−1 ).
k=0

The rest of the discussion focuses on the expressions of the coefficients ak , which are first given in
terms of divided differences,
f [x1 , x2 , . . . , xk ] − f [x0 , x1 , . . . , xk−1 ]
ak = f [x0 , x1 , . . . , xk ] = ,
xk − x0

7
for k = 0, 1, 2, . . . , n. This leads to the formula (3.10),
n
X
Pn (x) = f [x0 , x1 , . . . , xk ](x − x0 )(x − x1 ) . . . (x − xk−1 ).
k=0

Hence the divided-difference form of the polynomial consists in calculating the differences, f [x0 , x1 , . . . , xk ]
and putting them together in the summation expression (3.10). In this form of the divided-difference
interpolating polynomial, it doesn’t matter how the data points xi are ordered and the spacing be-
tween them.

However, the discussion in p126 continues with a simplification of the above formula by required
that the data points be ordered consecutively and assuming that the spacing between the xi is
equal, say h = xi+1 − xi .

If you follow the textbook’s derivation, the formula (3.12), comes from this ordering and equi-
spacing of the xi . By letting x = x0 +sh, the value of s is derived from the difference x−xi = (s−i)h.
This determination of the value s leads to the expression of the interpolating polynomial in (3.10)
as:
Pn (x) = f (x0 + sh) = f [x0 ] + shf [x0 , x1 ] + . . . + s(s − 1)(s − 2). . . (s − n + 1)hn f [x0 , x1 , x2 , . . . , xn ],
which in turn is written in summation notation as in (3.11) as
n  
X s
Pn (x) = Pn (x0 + sh) = f [x0 ] + k!hk f [x0 , x1 , . . . , xk ],
k=1
k

noting that here the product s(s − 1)(s − 2). . . (s − n + 1) is replaced using the binomial coefficient
and written as  
s
s(s − 1)(s − 2). . . (s − n + 1) = k!.
k
Thus (3.10) and (3.11) are equivalent expressions of the divided-difference formula for the interpo-
lating polynomials.

To obtain the Newton forward-difference and backward-difference formulas, the Aitken’s notation ∆
is reintroduced and used to write the differences, f [x0 , x1 , x2 , . . . , xn ] as
1
f [x0 , x1 , x2 , . . . , xn ] = ∆k f (x0 )
k!hk
By so doing, the k!hk factors cancel out, leading to (3.12),
n  
X s
Pn (x) = f (x0 ) + ∆k f (x0 ),
k=1
k

now referred to as the Newton Forward-Difference formula.

The Backward-difference formula is derived from the same principles but now the differences in
(3.10) are taken backward, starting with f [xn ], and
1
f [xn , xn−1 , . . . , xn−k ] = ∇k f (xn ).
k!hk

8
 APM2613/LN05/0/2022

Equation (3.10) is now written backward to give (3.13) as

s(s + 1) s(s + 1) . . . (s + n − 1) n
Pn (x) = f [xn ] + s∇f (xn ) + + ··· + ∇ f (xn )
2 n!
n  
X −s
= f [xn ] + (−1)k ∇k f (xn )
k=1
k

The difference between (3.12) and (3.13) is two-fold:

• the binomial coefficient in the former involves s while in the latter it involves −s;

• the direction in which the values in the difference table are read - from top left in a downward
diagonal line to the right in the forward-difference and from bottom left in an upward diagonal
in the backward-difference formula.

1.6 Error Estimation

Error estimation for the Lagrange polynomial and Neville’s formula are based on the (n + 1)th
derivative of f (x) and given by Theorem 3.3, by finding an upper bound of the (n + 1)th derivative
in the interval containing the nodes.

NB: The error estimation formula can only be applied if the function f (x) whose value is being
approximated is known - this is the difficulty with the Lagrange polynomial error term.

1.7 Osculating Polynomials

An osculating polynomial is a polynomial of least degree whose values not only coincide with those
of f (x) at the nodes x0 , x1 , . . . , xn but also has the same values of its derivatives of order less than
or equal to some integer values mi ∈ (x0 , xn ) as f (x) at each xi ; i.e.

dk P (xi ) dk f (xi ))
= for eachi = 0, 1, . . . , n and k = 0, 1, . . . , mi
dxk dxk
Thus the criteria for constructing these polynomials involve conditions on the derivative values at
the respective nodes.

Three types of osculating polynomials are discussed: Hermite polynomials, Cubic Splines and
Bezier curves. Hermite polynomials are the basis for the latter. While Hermite interpolation seeks
to find a single polynomial to approximate the desired function over the whole interval containing
the nodes, the cubic splines and Bezier curves generate pieces of cubic polynomials that satisfy
certain conditions pertaining to f (x) the polynomials Sj (x). What is common for these polynomials
is that they not only agree with f (x) at the nodes, but their derivatives at the nodes also agree with
those of f (x). The sets of nodes used to construct these cubic polynomials should span the entire
interval containing the nodes.

Because of the involvement of the derivative(s), the derivative values used in the computation
would be included in the tabulated values.

9
1.7.1 Hermite Polynomials
Two forms of the Hermite polynomials are given in the text:
n
X n
X
H2n+1 = f (xj )Hn,j (x) + f ′ (xj )Ĥn,j (x),
j=0 j=0

based on the coefficients of the Lagrange polynomial (p134); and the form
2n+1
X
H2n+1 (x) = f [z0 ] + f [z0 , . . . , zk ](x − z + 0)(x − z1 ) . . . (x − zk−1 ),
k=1

based on Newton’s interpolatory divided difference formula. These are discussed in detail on
pages 134-139. What is worth noting for these is that the data points include an additional column
of derivatives f ′ (xk ).

1.7.2 Cubic Splines

The literal (dictionary) meaning of ’spline’ is: a flexible strip of material, that may be bent into a
curve and used in a similar manner to a ruler to draw smooth curves between points. This explains
the underlying concept of cubic splines in interpolation.

The main form of the cubic splines Sj (x), is (the first equation on p145):
Sj (x) = aj + bj (x − xj ) + cj (x − xj )2 + dj (x − xj )3
for each j-th piece (arc), j = 0, 1, . . . , n − 1 between nodes (xj , yj ) and (xj+1 , yj+1 ), and spacing
hj = xj+1 − xj . These spacings may differ between any two successive nodes.

Constructing cubic splines consists is calculating the coefficients aj , bj , cj , dj for each of the arcs
for the selected nodes x0 , x1 , . . . , xn−1 .

Understanding the Design

The spline is a curve S similar to a string with endpoints of the interval [a, b] = [x0 , xn ]. The interval
[x0 , xn ] is divided into n + 1 nodes {x0 , x1 , x2 , . . . , xn }. The spline comprises smaller/shorter pieces
{Sj , j = 0, 1, 2, . . . , n − 1}, each defined for a pair of adjacent nodes xj , xj+1 : j = 0, 1, 2, . . . , n − 1.
It is important to note that the length of the subintervals [xj , xj+1 ] in NOT assumed to be the same
- this is the key to the flexibility of the spline. So there are n such pieces, {S0 , S1 , S2 , . . . , Sn−1 } that
are joined together to form the cubic spline S.

To design the spline we take two adjacent pieces, Sj (x) and Sj+1 (x), j = 0, 1, 2, . . . , n − 1 at a time,
that interpolate three points xj , xj+1 , xj+2 , j = 0, 1, 2, . . . , n − 2 to define a ’cubic piece’ over two ad-
jacent subintervals [xj , xj+1 ] and [xj+1 , xj+2 ]. Note that each pair of subintervals share a common
node xj+1 , but do not overlap with other subintervals.

Each piece Sj (x), j = 0, 1, 2, . . . , n − 1 of the spline S is defined as a cubic polynomial

Sj = aj + bj (x − xj ) + cj (x − xj )2 + dj (x − xj )3

10
 APM2613/LN05/0/2022

The definition specifies conditions to be met by each pair of adjacent ’spline pieces’ and used to
determine the coefficient sets {aj , bj , cj , dj }, j = 0, 1, 2 . . . , n − 1. The meaning of these conditions
is what we use to unpack the equations that should be solved in order to determine the set of
coefficients for each ’spline piece’.

Unpacking the conditions

Let us put a voice over each of the conditions listed in Definition 3.10, that interpretes each of them.

(a) This condition reiterates the description of the spline in terms of the ’spline pices’.

(b) Sj (xj ) = f (xj ) and Sj (xj+1 ) = f (xj+1 )

This condition means that for each piece Sj (x) suspended on the subinterval [xj , xj+1 ] on
which it is defined, the ’spline piece’ coincides with (or interpolates) the function at the end
nodes of the subinterval.

(c) Sj (xj+1 ) = Sj+1 (xj+1 )

This condition means the two adjacent pieces have the same value at the right end node xj+1
of the subinterval [xj , xj+1 ].
′
(d) Sj′ (xj+1 ) = Sj+1 (xj+1 )
This condition means that the slopes of the any two adjacent pieces are the same at the node
where they are joined. this is the ensure smoothness of the spline at each node.
′′
(e) Sj′′ (xj+1 ) = Sj+1 (xj+1 )
This condition means that the curvature of the any adjacent pieces is the same at the node
where they are joined. This takes care of the ’flexibility’ of the pieces.

(f) The final condition applies to the whole spline S(x) at the endpoints x0 and xn , the boundary
of the domain of the function f (x). The interpretation of the two parts of the condition are as
follows:

(i) The natural boundary condition S ′′ (x0 ) = S ′′ (xn ) = 0 can be interpreted in terms of
calculus to mean that the endpoints are inflection points of the spline.
(ii) The clamped boundary condition S ′ (x0 ) = f ′ (x0 ) and S ′ (xn ) = f ′ (xn ) can be interpreted
to mean the slopes of the spline at the endpoints x0 and xn coincides with the slopes of
the function in question at those points. This can also be seen to mean the direction of
the spline coincides with that of the function at the endpoints.

Applying the conditions given in Definition 3.10 yield a set of equations in the various coefficients.
These are simplified first by observing that the coefficients aj are determined from the given f (xj ).
With some ingenuity, this system of equations reduce to a system Ax = b where the solution vector
x = (c0 , c1 , . . . , cn ) is obtained with A consisting of expressions in terms of hj , and b involves only
the coefficients aj . Note the different formulations of A and b for the natural spline (p146-147) and
the clamped spline (p152).
Once the a’s and the c’s are determined, the c’s are related to the d’s in
cj+1 − cj
cj+1 = cj + 3dj hj ⇒ dj =
3hj

11
and the b’s are computable from the relation
1 hj
bj = (aj+1 − aj ) − (cj+1 + 2cj ).
hj 3
This now fully determines all the coefficients aj , bj , cj and dj defining the j-th piece Sj , j = 0, 1, . . . , n−
1 of the polynomial, to obtain the cubic spline

 S0 (x) on [x0 , x1 ]



 S1 (x) on [x1 , x2 ]

 ..
.

S(x) =
 S j (x) on [xj , xj+1 ]
.

..






Sn−1 (x) on [xn , xn−1 ]

Example 2 (p148) is a detailed procedure for constructing a spline (natural spline in this case).
Note that the examples used in the textbook use points with nodes that are equally spaced. this is
for ease of computation in demonstrating the method, but this is not a requirement.

1.7.3 Bezier Curves

A general parametric form of Hermite cubics is given in equations (3.23) and (3.24) of the textbook.
Bezier curves are a specific modification of the parametric curves which incorporate a scaling fac-
tor of 3 when computing the derivatives at the endpoints (see paragraph on p166). This scaling
affects the shape of the parametric curve in relation to the tangent at the nodes.

Two choices need to be made in the construction of the Bezier curves: the guidepoints; and the
step size of the parameter t. The step size determines the smoothness of the curve, while the
guidepoints control the approximation of the derivative, and hence the tangent line (computation of
the coefficients).

Note the scale factor of 3 incorporated in equations (3.25) and (3.26) for the Bezier curve, the
general parametric equations. Algorithm 3.6 is simply for generating the coefficients ai and bi for
the i-th curve using two nodes (xi , yi ) and (xi+1 , yi+1) and two guidepoints (xi + α0 , yi + β0 ) and
− −
(xi+1 − α1 , yi+1 − β1 ). The guidepoints in the text algorithm are denoted by (x+ +
i , yi ) and (xi+1 , yi+1 ),
respectively. The coefficients, ai , bi ; i = 0, . . . , 3 generated in the algorithm correspond to the coef-
ficients in (3.25) and (3.26) as summaries of the coefficients there, used in the first equation of the
algorithm:
(i) (i) (i) (i) (i) (i) (i) (i)
(xi (t), yi (t)) = (a0 + a1 t + a2 t2 + a3 t3 , b0 + b1 t + b2 t2 + b3 t3 ),
for the Ci , the pieces of the curve between (xi , yi ) and (xi+1 , yi+1 ), i = 0, 1, . . . , n − 1. Note that the
coordinates (x(t), y(t)) yield a parametric curve, Si between (xi , yi ) and (xi+1 , yi+1 ) as t varies in
the interval 0 ≤ t ≤ 1 at the predetermined steps.

1.8 Least Squares Approximation

Still on the problem of estimating the values of a function at nontabulated points, given a set of
m data points (xi , yi ) : i = 1, 2, . . . , m, the least-squares approximation approach does not require

12
 APM2613/LN05/0/2022

that the approximating function agrees with the sought function at these points.

Let g(x) be such an approximating function. This function is required to minimise the least squares
error (square of the deviations from the tabulated points)
m
X
E = (yi − g(xi ))2
i=1
m
X m
X m
X
= yi2 −2 g(xi )yi + (g(xi ))2
i=1 i=1 i=1

The general requirement for minimising this error is that the set of partial derivatives with respect
to the unknown coefficients in g(x) be equal to zero. This leads to a set of normal equations which
are solved and used in the expression of g(x).

Two cases of the function g(x) are considered: a polynomial Pn (x) and an exponential function
y = beax or y = bxa . The n−th degree least squares polynomial approximation uses

Pn (x) = an xn + an−1 xn−1 + · · · + a1 x + a0

The task here is to determine the coefficients ai , i = 0, 1, . . . , n that satisfy the system of normal
equations
n j+k n
X X j+k
X
ak xi = yi xji , for each j = 0, 1, . . . , n.
k=0 xi=1 i=1

The system of normal equations is

 Pm Pm 1 Pm 2
··· P m
   Pm
 
0 n 0
P
i=1 x i i=1 x i i=1 xi i=1 xi a0 i=1 yi xi
 Pm x1 Pm x2 Pm 3 m
· · · Pi=1 xi n+1   Pm yi x1
i=1 xi a1
  
 Pmi=1 i i=1 i     Pi=1 i 
2
P m 3
P m 4 m n+2 m 2

 i=1 xi i=1 xi i=1 xi ··· i=1 xi



 a2 =
 
i=1 yi x i


 .. .. ..   ..  .. 
.
Pm. n Pm n+1 Pm n+2 Pm . 2n . Pm . n
     
i=1 xi i=1 xi i=1 xi ··· i=1 xi an i=1 yi xi

The values in the coefficient matrix and the right hand side can all be calculated from the tabulated
points. Once calculated, the system can be solved by suitable methods.

Finding the approximating exponential functions y = beax or y = bxa requires is more manageable
if they are converted to linear form by taking the logarithm on both sides. This reduces the expo-
nential forms to linear forms which can then be treated as for linear least squares polynomials by
letting
z = ln y, c = ln b and or z = ax + c
for y = beax , and for y = bxa
ln y = ln b + a ln x
can be replaced by
z = c + a ln x.
It is critical to remember to substitute back to obtain the original values:

c = ln b ⇒ b = ec .

13
An example of this kind of calculation is given on p512.

The error in the approximation of f (x̂) is estimated by substituting the approximating function in the
error term m
X
E= (yi − g(xi ))2 ,
i=1

and taking the square root. i.e.

√ p
e = E = (y1 − g(x1 ))2 + ()y2 − g(x2 ))2 + · · · + (yn − g(xn ))2

using the appropriate expression for g(x).

1.9 In Summary
Different forms of the interpolating polynomial have been presented. Although the expressions are
different, the interpolating polynomial is unique. Hence the different forms are different expressions
of the same polynomial. Generally, there is no requirement that the spacing between the nodes be
equal. If the nodes are equally spaced, it may simplify the formulae by replacing xi+1 − xi by h in
the formulae.
For n + 1 nodes given on [x0 , xn ], the various degree polynomials make use of some of the
points/nodes - k + 1 for k-th degree polynomial, k ≤ n. Thus the challenge remains, as to which
set of k nodes would give the closest approximation of a particular value f (x), x ∈ (x0 , xn ). The
highest degree polynomial for n + 1 nodes is n. It may be necessary to compare different polyno-
mials of the same degree using different sets of nodes.

Intuitively, the selection of the nodes to be used in the interpolating polynomial is based on en-
suring that the value where the function is approximated is included in the interval containing the
nodes.

For least-squares approximation using Pn (x) = a0 + a1 x + · · · + an xn , the unknown coefficients

a0 , a1 , . . . , an are obtained from solving a linear systems of normal equations. Exponential functions
are converted to linear polynomials by taking the logarithm of the exponential form.

2 Useful Formulas
1. Lagrange interpolating polynomial on the nodes (x0 , f (x0 ), (x1 , f (x1 )), (xn , f (xn ))

Pn (x) = Ln,0 (x)f (x0 ) + Ln,1 (x)f (x1 ) + · · · + Ln,n f (xn )

where
(x − x0 )(x − x1 ) . . . (x − xk−1 )(x − xk+1 ) . . . (x − xn )
Lk (x) =
(xk − x0 )(xk − x1 ) . . . (xk − xk−1 )(xk − xk+1 ) . . . (xk − xn )

14