Interpolation and Polynomial Approximation: Professor Henry Arguello

Interpolation and Polynomial Approximation
Professor Henry Arguello
Universidad Industrial de Santander

Colombia
March 12, 2018
High Dimensional Signal Processing Group

www.hdspgroup.com
henarfu@uis.edu.co
Professor Henry Arguello Numerical methods March 12, 2018 1 / 77

Outline: Chapter 4
1 Interpolation and Polynomial Approximation
2 Taylor Series and Calculation of Functions
3 Introduction to Interpolation
4 Lagrange Approximation
5 Newton Polynomials
6 Lagrange Approximation
7 Newton Polynomials

Interpolation and Polynomial Approximation
Interpolation is used to approximate different functions as sin(x),

cos(x), etc.. using polynomials.
y y
2.5 2.5
y = ex y = ex
2.0 2.0
y = pH x L y = qH xL
1.5 1.5
1.0 1.0
0.5 0.5
x x
- 1.0 - 0.5 0.0 0.5 1.0 - 1.0 - 0.5 0.0 0.5 1.0
Figure: Comparison between 2 approximations (a) The Taylor polynomial

p(x) = 1.000000 + 1.000000x + 0.500000x2 , which approximates f (x) = ex over [-1,1]. (b) The
Chebyshev approximation q(x) = 1.000000 + 1.129772x + 0.532042x2 for f (x) = ex over [-1,1].

Taylor Series and Calculation of Functions
Represent the elementary functions: sin(x), cos(x), ex , ln(x), etc. The

next table gives several of the common Taylor series expansions. The
partial sums can be accumulated until the accuracy specified.
Table 4.1 Taylor Series Expansions for Some Common Functions

x3 x5 x7
sin(x) = x − 3!
+ 5!
− 7!
+ ··· for all x
2 4 6
x x x
cos(x) = 1 − 2!
+ 4!
− 6!
+ ··· for all x
2 3 4
ex = 1 + x + x
2!
+ x
3!
+ x
4!
+ ··· for all x
2 3
x x x4
ln(1 + x) = x − 2
+ 3
− 4
+ ··· −1 ≤ x ≤ 1
x3 x5 x7
arctan(x) = x − 3
− 5
− 7
+ ··· −1 ≤ x ≤ 1
p p(p−1) 2 p(p−1)(p−2) 3
(1 + x) = 1 + px + 2!
x + 3!
x + ··· for |x| < 1

Table 4.2 Partial Sums Sn Used to Determine e
n 1 + 1 + ··· + 1
Sn = 1 + 1! 2! n!
0 1.0
Example: Compute the number e 1 2.0
using the Taylor series in Table 4.1 2 2.5
with x = 1, 3 2.666666666666...
4 2.708333333333...
5 2.716666666666...
2 3 4 k
1 1 1 1 1 1 6 2.718055555555...
e = 1+ + + + +···+ +···
1! 2! 3! 4! k! 7 2.718253968254...
8 2.718278769841...
The sum of an infinite series 9 2.718281525573...
requires that the partial sums SN 10 2.718281801146...
tend to a limit. The values of these 11 2.718281826199...

12 2.718281828286...
sums are given in Table 4.2.
13 2.718281828447...
14 2.718281828458...
15 2.718281828459...

Theorem 9: Taylor Polynomial Approximation

Assume that f ∈ CN+1 [a, b] and x0 ∈ [a, b] is a fixed value. If x ∈ [a, b],
then
f (x) = PN (x) + EN (x), (1)
where PN (x) is a polynomial that can be used to approximate f (x):
N
X f (k) (x0 )
f (x) ≈ PN (x) = (x − x0 )k . (2)
k=0
k!
The error term EN (x) has the form

f N+1 (c)
En (x) = (x − x0 )N+1 (3)
(N + 1)!
for some value c = c(x) that lies between x and x0 .

EN (x) is used to determine a bound for the accuracy of the
approximation.

Example: Show why 15 terms are all that are needed to obtain the
13-digit approximation e = 2.718281828459 in Table 4.2.

Example: Show why 15 terms are all that are needed to obtain the
13-digit approximation e = 2.718281828459 in Table 4.2.
Expand f (x) = ex in a Taylor polynomial of degree 15 using the fixed
value x0 = 0 and involving the powers (x − 0)k = xk . The derivatives
required are f 0 (x) = f 00 (x) = ... = f (16) = ex . The first 15 derivatives are
used to calculate the coefficients ak = e0 /k! and are used to write the
polynomial
x2 x3 x15
P15 (x) = 1 + x + + + ··· + (4)
2! 3! 15!

Setting x = 1 in (4) gives the partial sum S1 5 = P1 5.Then the accuracy

of the approximation is:
f (16) (c)x16
E15 (x) = (5)
16!


f (16) (c)x16
E15 (x) = (5)
16!
Since we chose x0 = 0 and x = 1, the value c lies between them, which
implies that ec < e1 . Notice that the partial sums in Table 4.2 are
bounded above by 3. Combining these two inequalities yields ec < 3,
which is used in the following calculation
|f (16) (c)| ec 3
|E15 (1)| = ≤ < < 1.433844x10−13
16! 16! 16!


f (16) (c)x16
E15 (x) = (5)
16!
Since we chose x0 = 0 and x = 1, the value c lies between them, which
implies that ec < e1 . Notice that the partial sums in Table 4.2 are
bounded above by 3. Combining these two inequalities yields ec < 3,
which is used in the following calculation
|f (16) (c)| ec 3
|E15 (1)| = ≤ < < 1.433844x10−13
16! 16! 16!
Therefore, all the digits in the approximation e ≈ 2.718281828459 are
correct, because the actual error must be less than 2 in the thirteenth
decimal place.

Observe that the approximation
ex ≈ 1 + x is good near the center
x0 = 0 and that the distance
between the curves grows as x
y moves away from 0. The slopes of
5
y= ex
the curves agree at (0,1). The
4
study of curvature shows that if two
3 curves y = f (x) and y = g(x) have
2 y = P1 H x L the property that
1
f (x0 ) = g(x0 ), f 0 (x0 ) = g0 (x0 ) and
f 00 (x0 ) = g00 (x0 ) then they have the
x
-2 -1 1 2 same curvature at x0 . This property
-1
would be desirable for a polynomial
function that approximates f (x).
Corollary 4.1 shows that the Taylor
polynomial has this property for
N ≥ 2.
Corollary
If PN (x) is the Taylor polynomial of degree N given in Theorem 9, then
(k)
PN (x0 ) = f (k) (x0 ) for k = 0, 1, ..., N. (6)
Proof. Set x = x0 in equations (2) and (3), and the result is
PN (x0 ) = f (x0 ). thus statement (6) is true for k = 0. Now differentiate
the right-hand side of (2) and get
N N−1
X f (k) (x0 ) X f (k+1) (x0 )
P0N (x) = (x − x0 ) k−1
= (x − x0 )k . (7)
(k − 1)! k!
k=1 k=0
Set x = x0 in (7) to obtain P0N (x0 ) = f 0 (x0 ). Thus statement (6) is true for
k = 1. Successive differentiations of (7) will establish the other
identities in (6). The details are left as an exercise.

The accuracy of a Taylor polynomial is increased when we choose N

large. The accuracy of any given polynomial will generally decrease as
the value of x moves away from the center x0 . Hence we must choose
N large enough and restrict the maximum value of |x − x0 | so that the
error does not exceed a specified bound.
MRN+1
|error| = |EN (x)| ≤ (8)
(N + 1)!

Table 4.3 Values for the Error Bound |error| < eR RN + 1/(N + 1)!
R = 2.0, R = 1.5, R = 1.0, R = 0.5,

|x| ≤ 2.0 |x| ≤ 1.5 |x| ≤ 1.0 |x| ≤ 0.5
ex ≈ P5 (x) 0.65680499 0.07090172 0.00377539 0.00003578

ex ≈ P6 (x) 0.18765857 0.01519323 0.00053934 0.00000256
ex ≈ P7 (x) 0.04691464 0.00284873 0.00006742 0.00000016
ex ≈ P8 (x) 0.01042548 0.00047479 0.00000749 0.00000001
y
10
8
y = ex
y = P4 H x L
y = P3H xL
y = P2 H x L
6
x
-2 -1 0 1 2

Example: Establish the error bounds for the approximation ex ≈ P8 (x)

on each of the intervals |x| ≤ 1.0 and |x| ≤ 0.5.

Example: Establish the error bounds for the approximation ex ≈ P8 (x)

on each of the intervals |x| ≤ 1.0 and |x| ≤ 0.5.
if |x| ≤ 1.0, then letting R = 1.0 and |f (9) (c)| = |ec | ≤ e1.0 = M in (8)
implies that
e1.0 (1.0)9
|error| = |E8 (x)| ≤ ≈ 0.00000749.
9!
If |x| ≤ 0.5, then letting R = 0.5 and |f (9) (c)| = |ec | ≤ e0.5 = M in (8)
implies that
e0.5 (0.5)9
|error| = |E8 (x)| ≤ ≈ 0.00000001.
9!

Methods for Evaluating a Polynomial
Horner’s Method (Nested Multiplication)
Consider, for example, the function

f (x) = (x − 1)8 (9)
The binomial formula can be used to expand f (x) in powers of x :

8
!
X 8 8−k
f (x) = x (−1)k = x8 − 8x7 + 28x6 − 56x5 + 70x4 − 56x3 + 28x2 − 8x + 1 (10)
k=0
k
Now, Horner’s Method can be used to evaluate the polynomial in (10).

when applied to formula (10), nested multiplication permits us to write
f (x) = (((((((x − 8)x + 28)x − 56)x + 70)x − 56)x + 28)x − 8)x + 1 (11)
to evaluate f (x) now requires seven multiplications and eight additions

or subtractions. The necessity of using an exponential function to
evaluate the polynomial has now been eliminated.
Methods for Evaluating a Polynomial
Theorem 10: Taylor Series
Assume that f (x) is analytic on an interval (a, b) containing x0 . Suppose that
the Taylor polynomials (2) tend to a limit
N
X f (k) (x0 )
S(x) = lim PN (x) = lim (x − x0 )k (12)
N→∞ N→∞
k=0
k!
then f (x) has the Taylor series expansion

∞ (k)
X f (x0 )
f (x) = (x − x0 )k (13)
k=0
k!
Proof. This follows directly form the definition of convergence of series. The
limit condition is often stated by saying that the error term must go to zero as
N goes to infinity. Therefore, a necessary and sufficient condition for (18) to
hold is that
f (N+1) (c)(x − x0 )N+1
lim EN (x) = lim = 0. (14)
N→∞ N→∞ (N + 1)!
where c depends on N and x.

Introduction to Interpolation
Suppose that the function y = f (x) is known at the N + 1 points

(x0 , y0 ), ..., (xN , yN ),where the values xk are spread out over the interval
[a, b] and satisfy
a ≤ x0 < x1 < · · · < xN ≤ b and yk = f (xk ).
A polynomial P(x) of degree N will be constructed that passes through
these N + 1 points. In the construction, only the numerical values xk
and yk are needed. Hence the higher-order derivatives are not
necessary.The polynomial P(k) can be used to approximate f (x) over
the entire interval [a, b]. However, if the error function
E(x) = f (x) − P(x) is required, then we will need to know f (N+1) (x) and
a bound for its magnitude, that is,
M = max{|f (N+1) (x)| : a ≤ x ≤ b}.

Let us briefly mention how to evaluate the polynomial P(x):
P(x) = aN xN + aN−1 xN−1 + · · · + a2 x2 + a1 x + a0 (1)
Horner’s Method of synthetic division is an efficient way to evaluate

P(x). The derivative P0 (x) is
P0 (x) = NaN xN−1 + (N − 1)aN−1 xN−2 + · · · + 2a2 x + a1 (2)

P(x) dx , which satisfies I 0 (x) = P(x),
R
and the indefinite integral I(x) =
is
aN xN+1 aN−1 xN a2 x 3 a1 x 2
I(x) = + + ··· + + + a0 x + C (3)
N+1 N 3 2
Where C is the constant of integration. Algorithm 4.1 shows how to
adapt Horner’s method to P0 (x) and I(x).

Example: The polynomial P(x) = −0.02x3 + 0.2x2 − 0.4x + 1.28 passes

through the four points (1, 1.06), (2, 2.12), (3, 1.34), and (5, 1.78). Find
R4
(a) P(4), (b) P’(4), (c) 1 P(x) dx, and (d) P(5.5). Finally, (e) show how
to find the coefficients of P(x).

Example: The polynomial P(x) = −0.02x3 + 0.2x2 − 0.4x + 1.28 passes

through the four points (1, 1.06), (2, 2.12), (3, 1.34), and (5, 1.78). Find
R4
(a) P(4), (b) P’(4), (c) 1 P(x) dx, and (d) P(5.5). Finally, (e) show how
to find the coefficients of P(x).
y
2.0
(5.5,P(5.5))
(a) (4,P(4))
1.5
b3 = a3 = −0.02
y = P(x)
b2 = a2 + b3 x = 0.2 + (−0.02)(4) = 0.12
1.0
b1 = a1 + b2 x = −0.4 + (0.12)(4) = 0.08
b0 = a0 + b1 x = 1.28 + (0.08)(4) = 1.60
0.5
The interpolated value is P(4) = 1.60

x
0 1 2 3 4 5 6

(b) y
2.0 The tangent line has

slople P'(4).
d2 = 3a3 = −0.06
d1 = 2a2 + d2 x = 0.4 + (−0.06)(4) = 0.16 1.5
(4,P(4))
d0 = a1 + d1 x = −0.4 + (0.16)(4) = 0.24 y = P(x)
1.0
The numerical derivative is
P0 (4) = 0.24 0.5
x
0 1 2 3 4 5 6
(c) y
2.0
i4 = a43 = −0.005
a2
i3 = 3 + i4 x = 0.06666667 + (−0.005)(4) = 1.5
0.04666667 y=P(x)
i2 = a21 + i3 x = −0.2 + (0.04666667)(4) = 1.0

−0.01333333
i1 = a0 + i2 x = 1.28 + (−0.01333333)(4) = 0.5
1.22666667
i0 = 0 + i1 x = 0 + (1.22666667)(4) = 0.0 x
0 1 2 3 4 5 6
4.90666667.
y
2.0
(5.5,P(5.5))
(d) Use Algorithm 4.1(i) with x = 5.5 (4,P(4))
1.5
b3 = a3 = −0.02
y = P(x)
b2 = a2 + b3 x = 0.2 + (−0.02)(5.5) = 0.09
1.0
b1 = a1 + b2 x = −0.4 + (0.09)(5.5) = 0.095
b0 = a0 + b1 x = 1.28 + (0.095)(5.5) = 1.8025
0.5
The extrapolated value is P(5.5) = 1.8025.

x
0 1 2 3 4 5 6
(e) Assume that P(x) = A + Bx + Cx2 + Dx3 ; then at each value x = 1, 2, 3, and
5 we get a linear equation involving A, B, C and D.
Atx = 1 : A + 1B + 1C + 1D = 1.06
Atx = 2 : A + 2B + 4C + 8D = 1.12 (4)
Atx = 3 : A + 3B + 9C + 27D = 1.34
Atx = 5 : A + 5B + 25C + 125D = 1.78
The solution to (4) is A = 1.28, B = −0.4, C = 0.2 and D = −0.2.

Algorithm 4.1 (Polynomial Calculus). To evaluate the polynomial P(x),
its derivative P0 (x), and its integral P(x) dx by performing synthetic divi-
R
sion.
INPUT N Degree of P(x)
INPUT A(0), A(1), ..., A(N) Coefficients of P(x)
INPUT C Constant of integration
INPUT X Independent variable
(i) Algorithm to Evaluate P(x) Space-saving version:

B(N) := A(N) Poly := A(N)
FOR K = N − 1 DOWNTO 0 DO FOR K = N − 1 DOWNTO 0 DO
B(K) := A(K) + B(K + 1) ∗ X Poly := A(K)+ Poly∗X
PRINT "The Value P(x) is", B(0) PRINT "The Value P(x) is", Poly
(ii) Algorithm to Evaluate P0 (x) Space-saving version:

D(N − 1) := N ∗ A(N) Deriv := N ∗ A(N)
FOR K = N − 1 DOWNTO 1 DO FOR K = N − 1 DOWNTO 1 DO
D(K − 1) := K ∗ A(K) + D(K) ∗ X Deriv := K ∗ A(K)+Deriv∗X
PRINT "The Value P0 (x) is", D(0) PRINT "The Value P0 (x) is", Deriv
(iii) Algorithm to Evaluate I(x) Space-saving version:

I(N + 1) := A(N)/(N + 1) Integ := A(N)/(N + 1)
FOR K = N DOWNTO 1 DO FOR K = N DOWNTO 1 DO
I(K) := A(K − 1)/K + I(K + 1) ∗ X Integ := A(K − 1)/K + Integ∗X
I(0) := C + I(1) ∗ X Integ:= C + Integ ∗ X
PRINT "The Value I(x) is", I(0) PRINT "The Value I(x) is", Integ

Lagrange Approximation
Interpolation means to estimate a missing function value by taking a

weighted average of known function values at neighboring points.
Linear interpolation uses a line segment that passes through two

points. The slope between (x0 , y0 ) and (x1 , y1 ) is m = (y1 − y0 )/(x1 − x0 ),
and the point-slope formula for the line y = m(x − x0 ) + y0 can be
rearranged as
x − x0
y = P(x) = y0 + (y1 − y0 ) (5)
x1 − x0
When formula (1) is expanded, the result is a polynomial of degree
≤ 1. Evaluation of P(x) at x0 and x1 produces y0 and y1 , respectively:
P(x0 ) = x0 + (y1 − y0 )(0) = y0 ,

P(x1 ) = y0 + (y1 − y0 )(1) = y1 . (6)

Lagrange used a slightly different method to find this polynomial. He

noticed that it could be written as
x − x1 x − x0
y = P1 (x) = y0 + y1 . (7)
x0 − x1 x1 − x0
Each term on the right side of (3) involves a linear factor; hence the
sum is a polynomial of degree ≤ 1. The quotients in (3) are denoted by
x − x1 x − x0
L1,0 (0) = and L1,1 (x) = . (8)
x0 − x1 x1 − x0
Computation reveals that L1,0 (x0 ) = 1, L1,0 (x1 ) = 0, L1,1 (x0 ) = 0, and
L1,1 (x1 ) = 1 so that the polynomial P1 (x) in (3) also passes through the
two given points:
P1 (x0 ) = y0 + y1 (0) = y0 and P1 (x1 ) = y0 (0) + y1 = y1 (9)

The terms L1,0 (x) and L1,1 (x) in (4) are called Lagrange coefficient
polynomials based on the nodes x0 and x1 . Using this notation, (3)
can be written in summation form
1
X
P1 (x) = yk L1,k (x). (10)
k=0
Suppose that the ordinates yk are computed with the formula

yk = f (xk ). If P1 (x) is used to approximate f (x) over the interval [x0 , x1 ],
we call the process interpolation. If x < x0 (or x1 < x), then using P1 (x)
is called extrapolation.

Example 4.6: Consider the graph y = f (x) = cos(x) over [0.0, 1.2].
(a) Use the nodes x0 = 0.0 and x1 = 1.2 to construct a linear
interpolation polynomial P1 (x).
(b) Use the nodes x0 = 0.2 and x1 = 1.0 to construct a linear
approximating polynomial Q1 (x).

(a) Using (3) with the abscissas x0 = 0.0 and x1 = 1.2 and the ordinates
y0 = cos(0.0) = 1.000000 and y1 = cos(1.2) = 0.362358 produces
x − 1.2 x − 0.0
P1 (x) = 1.000000 + 0.362358
0.0 − 1.2 1.2 − 0.0
= −0.833333(x − 1.2) + 0.301965(x − 0.0).

x − 1.2 x − 0.0
P1 (x) = 1.000000 + 0.362358
0.0 − 1.2 1.2 − 0.0
= −0.833333(x − 1.2) + 0.301965(x − 0.0).
(b) When the nodes x0 = 0.2 and x1 = 1.0 with y0 = cos(0.2) = 0.980067
and y1 = cos(1.0) = 0.540302 are used, the result is
x − 1.0 x − 0.2
Q1 (x) = 0.980067 + 0.540302
0.2 − 1.0 1.0 − 0.2
= −1.225083(x − 1.0) + 0.675378(x − 0.2).

Table 4.6 Comparison of f (x) = cos(x) and the Linear Approximations P1 (x) and Q1 (x)
xk f (xk ) = cos(xk ) P1 (xk ) f (xk ) − P1 (xk ) Q1 (xk ) f (xk ) − Q1 (xk )
0.0 1.000000 1.000000 0.000000 1.090008 −0.090008

0.1 0.995004 0.946863 0.048141 1.035037 −0.040033
0.2 0.980067 0.893726 0.086340 0.980067 0.000000
0.3 0.955336 0.840589 0.114747 0.925096 0.030240
0.4 0.921061 0.787453 0.133608 0.870126 0.050935
0.5 0.877583 0.734316 0.143267 0.815155 0.062428
0.6 0.825336 0.681179 0.144157 0.760184 0.065151
0.7 0.764842 0.628042 0.136800 0.705214 0.059628
0.8 0.696707 0.574905 0.121802 0.650243 0.046463
0.9 0.621610 0.521768 0.099842 0.595273 0.026337
1.0 0.540302 0.468631 0.071671 0.540302 0.000000
1.1 0.453596 0.415495 0.038102 0.485332 −0.031736
1.2 0.362358 0.362358 0.000000 0.430361 −0.068003

Numerical Computations are given in Table 4.6 and reveal that Q1 (x) has less
error at the points xk that satisfy 0.1 ≤ xk ≤ 1.1. The largest tabulated error,
f (0.6) − P1 (0.6) = 0.144157, is reduced to f (0.6) − Q1 (0.6) = 0.065151 by using
Q1 (x).
y y
1.0 1.0
y = f(x)
y = f(x)
0.8 0.8
y = P1 HxL y = Q1 HxL
0.6 0.6
0.4 0.4
0.2 0.2
x x
0.0 0.2 0.4 0.6 0.8 1.0 1.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2
The linear approximation y = P1 (x) where The linear approximation y = Q1 (x) where
the nodes x0 = 0.0 and x1 = 1.2 are the the nodes x0 = 0.2 and x1 = 1.0 lie inside
endpoints of the interval [a, b]. the interval [a, b].

The generalization of (6) is the construction of a polynomial PN (x) of
degree at most N that passes through the N + 1 points
(x0 , y0 ), (x1 , y1 ), ..., (xN , yN ) and has the form
N
X
PN (x) = yk LN,k (x), (11)
k=0
where LN,k is the Lagrange coefficient polynomial based on these

nodes:
(x − x0 ) · · · (x − xk−1 )(x − xk+1 ) · · · (x − xN )
LN,k (x) = (12)
(xk − x0 ) · · · (xk − xk−1 )(xk − xk+1 ) · · · (xk − xN )
The product notation for (8), is written as

QN
j=0 (x − xj )
j6=k
LN,k (x) = QN (13)
j=0 (xk − xj )
j6=k

For each fixed k, the Lagrange coefficient polynomial LN,k (x) has the
property
LN,k (xj ) = 1 when j = k and LN,k (xj ) = 0 when j 6= k. (14)
Then direct substitution of these values into (7) is used to show that
the polynomial curve y = PN (x) goes through (xj , yj ):
PN (xj ) = y0 LN,0 (xj ) + · · · + yj LN,j (xj ) + · · · + yN LN,N (xj )

= y0 (0) + · · · + yj (1) + · · · + yN (0) = yj (15)
To show that PN (x) is unique, we invoke the fundamental theorem of

algebra, which states that a nonzero polynomial T(x) of degree ≤ N
has at most N roots. In other words, if T(x) is zero at N + 1 distinct
abscissas, it is identically zero.

Suppose that PN (x) is not unique and that there exist another
polynomial QN (x) of degree ≤ N that also passes through the N + 1
points.
Form the difference polynomial T(x) = PN (x) − QN (x).
Observe that the polynomial T(x) has degree ≤ N and that
T(xj ) = PN (xj ) − QN (xj ) = yj − yj = 0, for j = 0, 1, ..., N. Therefore,
T(x) ≡ 0 and it follows that QN (x) = PN (x).

When (7) is expanded, the result is similar to (3). The Lagrange

quadratic interpolating polynomial through the three points
(x0 , y0 ), (x1 , y1 ), and (x2 , y2 ) is
(x − x1 )(x − x2 ) (x − x0 )(x − x2 ) (x − x0 )(x − x1 )
P2 (x) = y0 + y1 + y2 (16)
(x0 − x1 )(x0 − x2 ) (x1 − x0 )(x1 − x2 ) (x2 − x0 )(x2 − x1 )
The Lagrange cubic interpolating polynomial through the four points

(x0 , y0 ), (x1 , y1 ), (x2 , y2 ) and (x3 , y3 ) is
(x − x1 )(x − x2 )(x − x3 ) (x − x0 )(x − x2 )(x − x3 )
P3 (x) = y0 + y1 (17)
(x0 − x1 )(x0 − x2 )(x0 − x3 ) (x1 − x0 )(x1 − x2 )(x1 − x3 )
(x − x0 )(x − x1 )(x − x3 ) (x − x0 )(x − x1 )(x − x2 )
+y2 + y3
(x2 − x0 )(x2 − x1 )(x2 − x3 ) (x3 − x0 )(x3 − x1 )(x3 − x2 )

Example 4.7: Consider y = f (x) = cos(x) over [0.0, 1.2].
(a) Use the three nodes x0 = 0.0, x1 = 0.6 and x2 = 1.2 to construct a
quadratic interpolation polynomial P2 (x).
(b) Use the four nodes x0 = 0.0, x1 = 0.4, x2 = 0.8 and x3 = 1.2 to
construct a cubic interpolation polynomial P3 (x).

(a) Using x0 = 0.0, x1 = 0.6, x2 = 1.2 and

y0 = cos(0.0) = 1.0, y1 = cos(0.6) = 0.825336, and
y2 = cos(1.2) = 0.362358 in equation (12) produces
(x − 0.6)(x − 1.2) (x − 0.0)(x − 1.2)
P2 (x) = 1.0 + 0.825336
(0.0 − 0.6)(0.0 − 1.2) (0.6 − 0.0)(0.6 − 1.2)
(x − 0.0)(x − 0.6)
+0.362358
(1.2 − 0.0)(1.2 − 0.6)
= 1.388889(x − 0.6)(x − 1.2) − 2.292599(x − 0.0)(x − 1.2)
+0.503275(x − 0.0)(x − 0.6).

(b) Using x0 = 0.0, x1 = 0.4, x2 = 0.8, x3 = 1.2 and
y0 = cos(0.0) = 1.0, y1 = cos(0.4) = 0.921061, y2 = cos(0.8) = 0.696707,
and y3 = cos(1.2) = 0.362358 in equation (13) produces
(x − 0.4)(x − 0.8)(x − 1.2)

P3 (x) = 1.000000
(0.0 − 0.4)(0.0 − 0.8)(0.0 − 1.2)
(x − 0.0)(x − 0.8)(x − 1.2)
+0.921061
(0.4 − 0.0)(0.4 − 0.8)(0.4 − 1.2)
(x − 0.0)(x − 0.4)(x − 1.2)
+0.696707
(0.8 − 0.0)(0.8 − 0.4)(0.8 − 1.2)
(x − 0.0)(x − 0.4)(x − 0.8)
+0.362358
(1.2 − 0.0)(1.2 − 0.4)(1.2 − 0.8)
= −2.604167(x − 0.4)(x − 0.8)(x − 1.2)

+7.195789(x − 0.0)(x − 0.8)(x − 1.2)
−5.443021(x − 0.0)(x − 0.4)(x − 1.2)
+0.943641(x − 0.0)(x − 0.4)(x − 0.8).
y y
1.0 1.0
0.5 0.5
y = f (x)
y = f(x)
x x
0.6 1.2 1.8 0.4 0.8 1.2 1.6 2
y = P2 H x L
- 0.5 - 0.5 y = P3H xL
The quadratic approximation polynomial The cubic approximation polynomial

y = P2 (x) based on the nodes y = P3 (x) based on the nodes
x0 = 0.0, x1 = 0.6 and x2 = 1.2. x0 = 0.0, x1 = 0.4, x2 = 0.8 and x3 = 1.2.

Theorem 4.3 (Lagrange Polynomial Approximation)

Assume that f ∈ CN+1 [a, b] and that x0 , x1 , ..., xN ∈ [a, b] are N + 1
nodes. If x ∈ [a, b], then
f (x) = PN (x) + EN (x), (18)
where PN is a polynomial that can be used to approximate f (x):

N
X
f (x) ≈ PN (x) = f (xk )LN,k (x). (19)
k=0
(x − x0 )(x − x1 ) · · · (x − xN )f (N+1) (c)

EN (x) = (20)
(N + 1)!
for some value c = c(x) that lies in the interval [a, b].

Theorem 4.4 (Error Bounds for Lagrange Interpolation, Equally
Spaced Nodes)
Assume that f (x) is defined on [a, b], which contains equally spaced nodes
xk = x0 + hk. Additionally, assume that f (x) and the derivates of f (x), up to the
order N + 1, are continuous and bounded on the special subintervals
[x0 , x1 ], [x0 , x2 ], and [x0 , x3 ], respectively; that is,
|f (N+1) (x)| ≤ MN+1 for x0 ≤ x ≤ xN , (21)
for N = 1, 2, 3. The error terms (16) corresponding to the cases N = 1, 2, and

3 have the following useful bounds on their magnitude:
h2 M2
|E1 (x)| ≤ valid for x ∈ [x0 , x1 ], (22)
8
h3 M3
|E2 (x)| ≤ √ valid for x ∈ [x0 , x2 ], (23)
9 3
h4 M4
|E3 (x)| ≤ valid for x ∈ [x0 , x3 ]. (24)
24
Comparison of Accuracy and O(hN+1 )
The significance of Theorem 4.4 is to understand a simple relationship
between the size of the error terms for linear, quadratic, and cubic
interpolation. In each case the error bound |EN (x)| depends on h in two ways.
First hN+1 is explicitly present to that |EN (x)| is proportional hN+1 . Second, the
values MN+1 generally depend on h and tend to |f (N+1) (x0 )| as h goes to zero.
Therefore, as h goes to zero, |EN (x)| converges to zero with the same rapidity
that hN+1 converges to zero. The notation O(hN+1 ) is used when discussing
this behavior. For example, the error bound (18) can expressed as
|E1 (x)| = O(h2 ) valid for x ∈ [x0 , x1 ]
The notation O(h2 ) stands in place of h2 M2 /8 in relation (18) and is meant to
convey the idea that the bound for the error term is approximately a multiple
of h2 ; that is,
|E1 (x)| ≤ Ch2 ≈ O(h2 ).
As a consequence, if the derivatives of f (x) are uniformly bounded on the
interval [a, b] and |h| < 1, the choosing N large will make hN+1 small, and the
higher-degree approximating polynomial will have less error.
Example 4.8: Consider y = f (x) = cos(x) over [0.0, 1.2]. Use formula
(18) through (20) and determine the error bounds for the Lagrange
polynomials P1 (x), P2 (x), and P3 (x) that were constructed in Examples
4.6 and 4.7.

4.6 and 4.7. First, determine the bounds M2 , M3 , and M4 for the
derivatives |f (2) (x)|, |f (3) (x)| and |f (4) (x)|, respectively, taken over the
interval [0.0, 1.2]:
|f (2) (x)| = | − cos(x)| ≤ | − cos(0.0)| = 1.000000 = M2 ,

(3)
|f (x)| = |sin(x)| ≤ |sin(1.2)| = 0.932039 = M3 ,
(4)
|f (x)| = |cos(x)| ≤ |cos(0.0)| = 1.000000 = M4 .
For P1 (x) the spacing of the nodes is h = 1.2, and its error bound is
h2 M2 (1.2)2 (1.000000)
|E1 (x)| ≤ ≤ = 0.180000 (25)
8 8

h 3 M3 (0.6)3 (0.932039)
|E2 (x)| ≤ √ ≤ √ = 0.012915 (26)
9 3 9 3
h 4 M4 (0.4)4 (1.000000)
|E3 (x)| ≤ ≤ = 0.001067 (27)
24 24
From Example 4.6 we saw that
|E1 (0.6)| = |cos(0.6) − P1 (0.6)| = 0.144157, so the bound 0.180000 in
(21) is reasonable.

Table 4.7 Comparison of f (x) = cos(x) and the Quadratic and Cubic
Polynomial Approximations P2 (x) and P3 (x)
xk f (xk ) = cos(xk ) P2 (xk ) E2 (xk ) P3 (xk ) E3 (xk )
0.0 1.000000 1.000000 0.000000 1.000000 0.000000

0.1 0.995004 0.990911 0.004093 0.995835 −0.000831
0.2 0.980067 0.973813 0.006253 0.980921 −0.000855
0.3 0.955336 0.948707 0.006629 0.955812 −0.000476
0.4 0.921061 0.915592 0.005469 0.921061 0.000000
0.5 0.877583 0.874468 0.003114 0.877221 0.000361
0.6 0.825336 0.825336 0.000000 0.824847 0.000890
0.7 0.764842 0.768194 −0.003352 0.764491 0.000351
0.8 0.696707 0.703044 −0.006338 0.696707 0.000000
0.9 0.621610 0.629886 −0.008276 0.622048 −0.000438
1.0 0.540302 0.548719 −0.008416 0.541068 −0.000765
1.1 0.453596 0.459542 −0.005946 0.454320 −0.000724
1.2 0.362358 0.362358 0.000000 0.362358 0.000000
y y
y = E3H xL
0.008 0.0008
y = E2 H x L
0.004 0.0004
x x
0.2 0.4 0.6 0.8 1.0 1.2 0.2 0.4 0.6 0.8 1.0 1.2
- 0.004 - 0.0004
- 0.008 - 0.0008
The error function E2 (x) = cos(x) − P2 (x). The error function E3 (x) = cos(x) − P3 (x).

Interpolation means to estimate a missing function value by taking a

weighted average of known function values at neighboring points.
Linear interpolation uses a line segment that passes through two
points. The slope between (x0 , y0 ) and (x1 , y1 ) is m = (y1 − y0 )/(x1 − x0 ),
and the point-slope formula for the line y = m(x − x0 ) + y0 can be
rearranged as
x − x0
y = P(x) = y0 + (y1 − y0 ) (28)
x1 − x0
When formula (1) is expanded, the result is a polynomial of degree
≤ 1. Evaluation of P(x) at x0 and x1 produces y0 and y1 , respectively:
P(x0 ) = x0 + (y1 − y0 )(0) = y0 ,

P(x1 ) = y0 + (y1 − y0 )(1) = y1 . (29)

The French mathematician Joseph Louis Lagrange used a slightly

different method to find this polynomial. He noticed that it could be
written as
x − x1 x − x0
y = P1 (x) = y0 + y1 . (30)
x0 − x1 x1 − x0
Each term on the right side of (3) involves a linear factor; hence the
sum is a polynomial of degree ≤ 1. The quotients in (3) are denoted by
x − x1 x − x0
L1,0 (0) = and L1,1 (x) = . (31)
x0 − x1 x1 − x0
Computation reveals that L1,0 (x0 ) = 1, L1,0 (x1 ) = 0, L1,1 (x0 ) = 0, and
L1,1 (x1 ) = 1 so that the polynomial P1 (x) in (3) also passes through the
two given points:
P1 (x0 ) = y0 + y1 (0) = y0 and P1 (x1 ) = y0 (0) + y1 = y1 (32)

The terms L1,0 (x) and L1,1 (x) in (4) are called Lagrange coefficient
polynomials based on the nodes x0 and x1 . Using this notation, (3)
can be written in summation form
1
X
P1 (x) = yk L1,k (x). (33)
k=0
Suppose that the ordinates yk are computed with the formula

yk = f (xk ). If P1 (x) is used to approximate f (x) over the interval [x0 , x1 ],
we call the process interpolation. If x < x0 (or x1 < x), then using P1 (x)
is called extrapolation.


x − 1.2 x − 0.0
P1 (x) = 1.000000 + 0.362358
0.0 − 1.2 1.2 − 0.0
= −0.833333(x − 1.2) + 0.301965(x − 0.0).

x − 1.2 x − 0.0
P1 (x) = 1.000000 + 0.362358
0.0 − 1.2 1.2 − 0.0
= −0.833333(x − 1.2) + 0.301965(x − 0.0).
(b) When the nodes x0 = 0.2 and x1 = 1.0 with y0 = cos(0.2) = 0.980067
and y1 = cos(1.0) = 0.540302 are used, the result is
x − 1.0 x − 0.2
Q1 (x) = 0.980067 + 0.540302
0.2 − 1.0 1.0 − 0.2
= −1.225083(x − 1.0) + 0.675378(x − 0.2).

Table 4.6 Comparison of f (x) = cos(x) and the Linear Approximations P1 (x) and Q1 (x)
xk f (xk ) = cos(xk ) P1 (xk ) f (xk ) − P1 (xk ) Q1 (xk ) f (xk ) − Q1 (xk )
0.0 1.000000 1.000000 0.000000 1.090008 −0.090008

0.1 0.995004 0.946863 0.048141 1.035037 −0.040033
0.2 0.980067 0.893726 0.086340 0.980067 0.000000
0.3 0.955336 0.840589 0.114747 0.925096 0.030240
0.4 0.921061 0.787453 0.133608 0.870126 0.050935
0.5 0.877583 0.734316 0.143267 0.815155 0.062428
0.6 0.825336 0.681179 0.144157 0.760184 0.065151
0.7 0.764842 0.628042 0.136800 0.705214 0.059628
0.8 0.696707 0.574905 0.121802 0.650243 0.046463
0.9 0.621610 0.521768 0.099842 0.595273 0.026337
1.0 0.540302 0.468631 0.071671 0.540302 0.000000
1.1 0.453596 0.415495 0.038102 0.485332 −0.031736
1.2 0.362358 0.362358 0.000000 0.430361 −0.068003

Numerical Computations are given in Table 4.6 and reveal that Q1 (x) has less
error at the points xk that satisfy 0.1 ≤ xk ≤ 1.1. The largest tabulated error,
f (0.6) − P1 (0.6) = 0.144157, is reduced to f (0.6) − Q1 (0.6) = 0.065151 by using
Q1 (x).
y y
1.0 1.0
y = f(x)
y = f(x)
0.8 0.8
y = P1 HxL y = Q1 HxL
0.6 0.6
0.4 0.4
0.2 0.2
x x
0.0 0.2 0.4 0.6 0.8 1.0 1.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2
The linear approximation y = P1 (x) where The linear approximation y = Q1 (x) where
the nodes x0 = 0.0 and x1 = 1.2 are the the nodes x0 = 0.2 and x1 = 1.0 lie inside
endpoints of the interval [a, b]. the interval [a, b].

The generalization of (6) is the construction of a polynomial PN (x) of
degree at most N that passes through the N + 1 points
(x0 , y0 ), (x1 , y1 ), ..., (xN , yN ) and has the form
N
X
PN (x) = yk LN,k (x), (34)
k=0
where LN,k is the Lagrange coefficient polynomial based on these

nodes:
(x − x0 ) · · · (x − xk−1 )(x − xk+1 ) · · · (x − xN )
LN,k (x) = (35)
(xk − x0 ) · · · (xk − xk−1 )(xk − xk+1 ) · · · (xk − xN )
It is understood that the terms (x − xk ) and xk − xk do not appear on the
right side of equation (8). It is appropriate to introduce the product
notation for (8), and we write
QN
j=0 (x − xj )
j6=k
LN,k (x) = QN (36)
j=0 (xk − xj )
j6=k

A Straightforward calculation shows that for each fixed k, the Lagrange
coefficient polynomial LN,k (x) has the property
LN,k (xj ) = 1 when j = k and LN,k (xj ) = 0 when j 6= k. (37)
Then direct substitution of these values into (7) is used to show that
the polynomial curve y = PN (x) goes through (xj , yj ):
PN (xj ) = y0 LN,0 (xj ) + · · · + yj LN,j (xj ) + · · · + yN LN,N (xj )
= y0 (0) + · · · + yj (1) + · · · + yN (0) = yj (38)
To show that PN (x) is unique, we invoke the fundamental theorem of
algebra, which states that a nonzero polynomial T(x) of degree ≤ N
has at most N roots. In other words, if T(x) is zero at N + 1 distinct
abscissas, it is identically zero. Suppose that PN (x) is not unique and
that there exist another polynomial QN (x) of degree ≤ N that also
passes through the N + 1 points. Form the difference polynomial
T(x) = PN (x) − QN (x). Observe that the polynomial T(x) has degree
≤ N and that T(xj ) = PN (xj ) − QN (xj ) = yj − yj = 0, for j = 0, 1, ..., N.
Therefore, T(x) ≡ 0 and it follows that QN (x) = PN (x).
When (7) is expanded, the result is similar to (3). The Lagrange

quadratic interpolating polynomial through the three points
(x0 , y0 ), (x1 , y1 ), and (x2 , y2 ) is
(x − x1 )(x − x2 ) (x − x0 )(x − x2 ) (x − x0 )(x − x1 )
P2 (x) = y0 + y1 + y2 (39)
(x0 − x1 )(x0 − x2 ) (x1 − x0 )(x1 − x2 ) (x2 − x0 )(x2 − x1 )
The Lagrange cubic interpolating polynomial through the four points

(x0 , y0 ), (x1 , y1 ), (x2 , y2 ) and (x3 , y3 ) is
(x − x1 )(x − x2 )(x − x3 ) (x − x0 )(x − x2 )(x − x3 )
P3 (x) = y0 + y1 (40)
(x0 − x1 )(x0 − x2 )(x0 − x3 ) (x1 − x0 )(x1 − x2 )(x1 − x3 )
(x − x0 )(x − x1 )(x − x3 ) (x − x0 )(x − x1 )(x − x2 )
+y2 + y3
(x2 − x0 )(x2 − x1 )(x2 − x3 ) (x3 − x0 )(x3 − x1 )(x3 − x2 )


(a) Using x0 = 0.0, x1 = 0.6, x2 = 1.2 and

y0 = cos(0.0) = 1.0, y1 = cos(0.6) = 0.825336, and
y2 = cos(1.2) = 0.362358 in equation (12) produces
(x − 0.6)(x − 1.2) (x − 0.0)(x − 1.2)
P2 (x) = 1.0 + 0.825336
(0.0 − 0.6)(0.0 − 1.2) (0.6 − 0.0)(0.6 − 1.2)
(x − 0.0)(x − 0.6)
+0.362358
(1.2 − 0.0)(1.2 − 0.6)
= 1.388889(x − 0.6)(x − 1.2) − 2.292599(x − 0.0)(x − 1.2)
+0.503275(x − 0.0)(x − 0.6).

(b) Using x0 = 0.0, x1 = 0.4, x2 = 0.8, x3 = 1.2 and
y0 = cos(0.0) = 1.0, y1 = cos(0.4) = 0.921061, y2 = cos(0.8) = 0.696707,
and y3 = cos(1.2) = 0.362358 in equation (13) produces
(x − 0.4)(x − 0.8)(x − 1.2)

P3 (x) = 1.000000
(0.0 − 0.4)(0.0 − 0.8)(0.0 − 1.2)
(x − 0.0)(x − 0.8)(x − 1.2)
+0.921061
(0.4 − 0.0)(0.4 − 0.8)(0.4 − 1.2)
(x − 0.0)(x − 0.4)(x − 1.2)
+0.696707
(0.8 − 0.0)(0.8 − 0.4)(0.8 − 1.2)
(x − 0.0)(x − 0.4)(x − 0.8)
+0.362358
(1.2 − 0.0)(1.2 − 0.4)(1.2 − 0.8)
= −2.604167(x − 0.4)(x − 0.8)(x − 1.2)

+7.195789(x − 0.0)(x − 0.8)(x − 1.2)
−5.443021(x − 0.0)(x − 0.4)(x − 1.2)
+0.943641(x − 0.0)(x − 0.4)(x − 0.8).
y y
1.0 1.0
0.5 0.5
y = f (x)
y = f(x)
x x
0.6 1.2 1.8 0.4 0.8 1.2 1.6 2
y = P2 H x L
- 0.5 - 0.5 y = P3H xL
The quadratic approximation polynomial The cubic approximation polynomial

y = P2 (x) based on the nodes y = P3 (x) based on the nodes
x0 = 0.0, x1 = 0.6 and x2 = 1.2. x0 = 0.0, x1 = 0.4, x2 = 0.8 and x3 = 1.2.

Theorem 4.3 (Lagrange Polynomial Approximation)

Assume that f ∈ CN+1 [a, b] and that x0 , x1 , ..., xN ∈ [a, b] are N + 1
nodes. If x ∈ [a, b], then
f (x) = PN (x) + EN (x), (41)
where PN is a polynomial that can be used to approximate f (x):

N
X
f (x) ≈ PN (x) = f (xk )LN,k (x). (42)
k=0
(x − x0 )(x − x1 ) · · · (x − xN )f (N+1) (c)

EN (x) = (43)
(N + 1)!
for some value c = c(x) that lies in the interval [a, b].

Theorem 4.4 (Error Bounds for Lagrange Interpolation, Equally
Spaced Nodes)
Assume that f (x) is defined on [a, b], which contains equally spaced nodes
xk = x0 + hk. Additionally, assume that f (x) and the derivates of f (x), up to the
order N + 1, are continuous and bounded on the special subintervals
[x0 , x1 ], [x0 , x2 ], and [x0 , x3 ], respectively; that is,
|f (N+1) (x)| ≤ MN+1 for x0 ≤ x ≤ xN , (44)
for N = 1, 2, 3. The error terms (16) corresponding to the cases N = 1, 2, and

3 have the following useful bounds on their magnitude:
h2 M2
|E1 (x)| ≤ valid for x ∈ [x0 , x1 ], (45)
8
h3 M3
|E2 (x)| ≤ √ valid for x ∈ [x0 , x2 ], (46)
9 3
h4 M4
|E3 (x)| ≤ valid for x ∈ [x0 , x3 ]. (47)
24
Comparison of Accuracy and O(hN+1 )
The significance of Theorem 4.4 is to understand a simple relationship
between the size of the error terms for linear, quadratic, and cubic
interpolation. In each case the error bound |EN (x)| depends on h in two ways.
First hN+1 is explicitly present to that |EN (x)| is proportional hN+1 . Second, the
values MN+1 generally depend on h and tend to |f (N+1) (x0 )| as h goes to zero.
Therefore, as h goes to zero, |EN (x)| converges to zero with the same rapidity
that hN+1 converges to zero. The notation O(hN+1 ) is used when discussing
this behavior. For example, the error bound (18) can expressed as
|E1 (x)| = O(h2 ) valid for x ∈ [x0 , x1 ]
The notation O(h2 ) stands in place of h2 M2 /8 in relation (18) and is meant to
convey the idea that the bound for the error term is approximately a multiple
of h2 ; that is,
|E1 (x)| ≤ Ch2 ≈ O(h2 ).
As a consequence, if the derivatives of f (x) are uniformly bounded on the
interval [a, b] and |h| < 1, the choosing N large will make hN+1 small, and the
higher-degree approximating polynomial will have less error.
4.6 and 4.7.

4.6 and 4.7. First, determine the bounds M2 , M3 , and M4 for the
derivatives |f (2) (x)|, |f (3) (x)| and |f (4) (x)|, respectively, taken over the
interval [0.0, 1.2]:
|f (2) (x)| = | − cos(x)| ≤ | − cos(0.0)| = 1.000000 = M2 ,

(3)
|f (x)| = |sin(x)| ≤ |sin(1.2)| = 0.932039 = M3 ,
(4)
|f (x)| = |cos(x)| ≤ |cos(0.0)| = 1.000000 = M4 .
h2 M2 (1.2)2 (1.000000)
|E1 (x)| ≤ ≤ = 0.180000 (48)
8 8

h 3 M3 (0.6)3 (0.932039)
|E2 (x)| ≤ √ ≤ √ = 0.012915 (49)
9 3 9 3
h 4 M4 (0.4)4 (1.000000)
|E3 (x)| ≤ ≤ = 0.001067 (50)
24 24
From Example 4.6 we saw that
|E1 (0.6)| = |cos(0.6) − P1 (0.6)| = 0.144157, so the bound 0.180000 in
(21) is reasonable.

Table 4.7 Comparison of f (x) = cos(x) and the Quadratic and Cubic
Polynomial Approximations P2 (x) and P3 (x)
xk f (xk ) = cos(xk ) P2 (xk ) E2 (xk ) P3 (xk ) E3 (xk )
0.0 1.000000 1.000000 0.000000 1.000000 0.000000

0.1 0.995004 0.990911 0.004093 0.995835 −0.000831
0.2 0.980067 0.973813 0.006253 0.980921 −0.000855
0.3 0.955336 0.948707 0.006629 0.955812 −0.000476
0.4 0.921061 0.915592 0.005469 0.921061 0.000000
0.5 0.877583 0.874468 0.003114 0.877221 0.000361
0.6 0.825336 0.825336 0.000000 0.824847 0.000890
0.7 0.764842 0.768194 −0.003352 0.764491 0.000351
0.8 0.696707 0.703044 −0.006338 0.696707 0.000000
0.9 0.621610 0.629886 −0.008276 0.622048 −0.000438
1.0 0.540302 0.548719 −0.008416 0.541068 −0.000765
1.1 0.453596 0.459542 −0.005946 0.454320 −0.000724
1.2 0.362358 0.362358 0.000000 0.362358 0.000000
y y
y = E3H xL
0.008 0.0008
y = E2 H x L
0.004 0.0004
x x
0.2 0.4 0.6 0.8 1.0 1.2 0.2 0.4 0.6 0.8 1.0 1.2
- 0.004 - 0.0004
- 0.008 - 0.0008
The error function E2 (x) = cos(x) − P2 (x). The error function E3 (x) = cos(x) − P3 (x).

Newton Polynomials
We take a new approach and construct Newton polynomials that have
the recursive pattern
P1 (x) = a0 + a1 (x − x0 ), (1)
P2 (x) = a0 + a1 (x − x0 ) + a2 (x − x0 )(x − x1 ), (2)
P3 (x) = a0 + a1 (x − x0 ) + a2 (x − x0 )(x − x1 ) (3)
+a3 (x − x0 )(x − x1 )(x − x2 ),
..
.
PN (x) = a0 + a1 (x − x0 ) + a2 (x − x0 )(x − x1 ) (4)
+a3 (x − x0 )(x − x1 )(x − x2 )
+a4 (x − x0 )(x − x1 )(x − x2 )(x − x3 ) + · · ·
+aN (x − x0 ) · · · (x − xN−1 ).
Here the polynomial PN (x) is obtained from PN−1 (x) using the recursive
relationship
PN (x) = PN−1 (x) + aN (x − x0 )(x − x1 )(x − x2 ) · · · (x − xN−1 ). (5)

Newton Polynomials
The polynomial (4) is said to be a Newton polynomial with N centers

x0 , x1 , ..., xN−1 . It involves sums of products of linear factors up to
aN (x − x0 )(x − x1 )(x − x2 ) · · · (x − xN−1 ),
so PN will simply be an ordinary polynomial of degree ≤ N.

Newton Polynomials
Example 4.10: Given the centers x0 = 1, x1 = 3, x2 = 4, and x3 = 4.5
and the coefficients a0 = 5, a1 = −2, a2 = 0.5, a3 = −0.1 and a4 = 0.003,
find P1 (x), P2 (x), P3 (x), P4 (x) and evaluate Pk (2.5) for k = 1, 2, 3, 4.

Newton Polynomials
Using formulas (1) through (4), we have
P1 (x) = 5 − 2(x − 1),

P2 (x) = 5 − 2(x − 1) + 0.5(x − 1)(x − 3),
P3 (x) = P2 (x) − 0.1(x − 1)(x − 3)(x − 4),
P4 (x) = P3 (x) + 0.003(x − 1)(x − 3)(x − 4)(x − 4.5).

Newton Polynomials
P1 (x) = 5 − 2(x − 1),

P2 (x) = 5 − 2(x − 1) + 0.5(x − 1)(x − 3),
P3 (x) = P2 (x) − 0.1(x − 1)(x − 3)(x − 4),
P4 (x) = P3 (x) + 0.003(x − 1)(x − 3)(x − 4)(x − 4.5).
Evaluating the polynomials at x = 2.5 results in

Newton Polynomials
P1 (x) = 5 − 2(x − 1),

P2 (x) = 5 − 2(x − 1) + 0.5(x − 1)(x − 3),
P3 (x) = P2 (x) − 0.1(x − 1)(x − 3)(x − 4),
P4 (x) = P3 (x) + 0.003(x − 1)(x − 3)(x − 4)(x − 4.5).
Evaluating the polynomials at x = 2.5 results in
P1 (2.5) = 5 − 2(1.5) = 2,
P2 (2.5) = P1 (2.5) + 0.5(1.5)(−0.5) = 1.625,
P3 (2.5) = P2 (2.5) − 0.1(1.5)(−0.5)(−1.5) = 1.5125,
P4 (2.5) = P3 (2.5) + 0.003(1.5)(−0.5)(−1.5)(−2.0) = 1.50575.

Newton Polynomials
Nested Multiplication
If N is fixed and the polynomial PN (x) is evaluated many times, then
nested multiplication should be used. The process is similar to nested
multiplication for ordinary polynomials, except that the centers xk must
be subtracted from the independent variable x. The nested
multiplication form for P3 (x) is
P3 (x) = ((a3 (x − x2 ) + a2 )(x − x1 ) + a1 )(x − x0 ) + a0 . (6)
To evaluate P3 (x) for a given value of x, start with innermost grouping
and form successively the quantities
S3 = a3 ,
S2 = S3 (x − x2 ) + a2 ,
S1 = S2 (x − x1 ) + a1 , (7)
S0 = S1 (x − x0 ) + a0 .
The quantity S0 is now P3 (x).
Newton Polynomials
Example 4.11: Compute P3 (2.5) in Example 4.10 using nested

multiplication.

Newton Polynomials
Example 4.11: Compute P3 (2.5) in Example 4.10 using nested

multiplication.
Using (6), we write
P3 (x) = ((−0.1(x − 4) + 0.5)(x − 3) − 2)(x − 1) + 5.
The values in (7) are
S3 = −0.1,
S2 = −0.1(2.5 − 4) + 0.5 = 0.65,
S1 = 0.65(2.5 − 3) − 2 = −2.325,
S0 = −2.325(2.5 − 1) + 5 = 1.5125.
Therefore, P3 (2.5) = 1.5125.

Newton Polynomials
Polynomial Approximation, Nodes, and Centers
Suppose that we want to find the coefficients ak for all the polynomials
P1 (x), ..., PN (x) that approximate a given function f (x). Then Pk (x) will
be based on the centers x0 , x1 , ..., xk and have the nodes x0 , x1 , ..., xk+1 .
For the polynomial P1 (x) the coefficients a0 and a1 have a familiar
meaning. In this case
P1 (x0 ) = f (x0 ) and P1 (x1 ) = f (x1 ). (8)
Using (1) and (8) to solve for a0 , we find that
f (x0 ) = P1 (x0 ) = a0 + a1 (x0 − x0 ) = a0 . (9)
Hence a0 = f (x0 ). Next, using (1), (8), and (9), we have
f (x1 ) = P1 (x1 ) = a0 + a1 (x1 − x0 ) = f (x0 ) + a1 (x1 − x0 ),
which can be solved for a1 , and we get
f (x1 ) − f (x0 )
a1 = (10)
x1 − x0
Newton Polynomials
Hence a1 is the slope of the secant line passing through the two points
(x0 , f (x0 )) and (x1 , f (x1 )).
The coefficients a0 and a1 are the same for both P1 (x) and P2 (x).
Evaluating (2) at the node x2 , we find that
f (x2 ) = P2 (x2 ) = a0 + a1 (x2 − x0 ) + a2 (x2 − x0 )(x2 − x1 ). (11)
The values for a0 and a1 in (9) and (10) can be used in (11) to obtain
f (x2 ) − a0 − a1 (x2 − x0 )
a2 =
(x2 − x0 )(x2 − x1 )

f (x2 ) − f (x0 ) f (x1 ) − f (x0 )
= − /(x2 − x1 ).
x2 − x0 x1 − x0
For computational purposes we prefer to write this last quantity as

f (x2 ) − f (x1 ) f (x1 ) − f (x0 )
a2 = − /(x2 − x0 ). (12)
x2 − x1 x1 − x0

Newton Polynomials
Definition 4.1 (Divided differences)

The divided differences for a function f (x) are defined as follows:
f [xk ] = f (xk ),
f [xk ] − f [xk−1 ]
f [xk−1 , xk ] = ,
xk − xk−1
f [xk−1 , xk ] − f [xk−2 , xk−1 ]
f [xk−2 , xk−1 , xk ] = , (13)
xk − xk−2
f [xk−2 , xk−1 , xk ] − f [xk−3 , xk−2 , xk−1 ]
f [xk−3 , xk−2 , xk−1 , xk ] = .
xk − xk−3
(14)
The recursive rule for constructing higher-order divided differences is

f [xk−j+1 , ..., xk ] − f [xk−j , ..., xk−1 ]
f [xk−j , xk−j+1 , ..., xk ] = (15)
xk − xk−j

Newton Polynomials
Table 4.8 Divided-Difference Table for y = f (x)
xk f [xk ] f[ , ] f[ , , ] f[ , , , ] f[ , , , , ]
x0 f [x0 ]
x1 f [x1 ] f [x0 , x1 ]
x2 f [x2 ] f [x1 , x2 ] f [x0 , x1 , x2 ]
x3 f [x3 ] f [x2 , x3 ] f [x1 , x2 , x3 ] f [x0 , x1 , x2 , x3 ]
x4 f [x4 ] f [x3 , x4 ] f [x2 , x3 , x4 ] f [x1 , x2 , x3 , x4 ] f [x0 , x1 , x2 , x3 , x4 ]
The coefficients ak of Pn (x) depend on the values f (xj ), for j = 0, 1, ..., k.

Newton Polynomials
Theorem 4.5 (Newton Polynomial)

Suppose that x0 , x1 , ..., xN are N + 1 distinct numbers in [a, b]. There
exists a unique polynomial PN (x) of degree at most N with the property
that
f (xj ) = PN (xj ) for j = 0, 1, ..., N
The Newton form of this polynomial is
PN (x) = a0 + a1 (x − x0 ) + · · · + aN (x − x0 )(x − x1 ) · · · (x − xN−1 ), (16)
where ak = f [x0 , x1 , ..., xk ], for k = 0, 1, ..., N.

Remark. If {(xj , yj )}Nj=0 is a set of points whose abscissas are distinct, the values
f (xj ) = yj can be used to construct the unique polynomial of degree ≤ N that passes
through the N + 1 points.

Newton Polynomials
Corollary 4.2 (Newton Approximation)

Assume that PN (x) is the Newton polynomial given in Theorem 4.5 and
is used to approximate the function f (x), that is,
f (x) = PN (x) + EN (x). (17)
If f ∈ CN+1 [a, b], then for each x ∈ [a, b] there corresponds a number
c = c(x) in (a, b), so that the error term has the form
(x − x0 )(x − x1 ) · · · (x − xN )f (N+1) (c)

EN (x) = . (18)
(N + 1)!
Remark. The error term EN (x) is the same as the one for Lagrange interpolation,
which was introduced in equation (16) of Section 4.3.

Newton Polynomials
Example 4.12. Let f (x) = x3 − 4x. Construct the divided-difference

table based on the nodes x0 = 1, x1 = 2, ..., x5 = 6, and find the Newton
polynomial P3 (x) based on x0 , x1 , x2 , and x3 .

Newton Polynomials
Example 4.12. Let f (x) = x3 − 4x. Construct the divided-difference

table based on the nodes x0 = 1, x1 = 2, ..., x5 = 6, and find the Newton
polynomial P3 (x) based on x0 , x1 , x2 , and x3 .
Table 4.9 Divided-Difference Table Used for Constructing the Newton Polynomial P3 (x)
First Second Third Fourth Fifth
divided divided divided divided divided
xk f [xk ] difference difference difference difference difference
x0 = 1 −3
x1 = 2 0 3
x2 = 3 15 15 6
x3 = 4 48 33 9 1
x4 = 5 105 57 12 1 0
x5 = 6 192 87 15 1 0 0

Newton Polynomials
The coefficients a0 = −3, a1 = 3, a2 = 6, and a3 = 1 of P3 (x) appear on

the diagonal of the divided-difference table. The centers x0 = 1, x1 = 2,
and x2 = 3 are the values in the first column. Using formula (3),we write
P3 (x) = −3 + 3(x − 1) + 6(x − 1)(x − 2) + (x − 1)(x − 2)(x − 3).

Newton Polynomials
Example 4.13. Construct a divided-difference table for f (x) = cos(x)
based on the five points (k, cos(k)), for k = 0, 1, 2, 3, 4. Use it to find the
coefficients ak and the four Newton interpolating polynomials Pk (x), for
k = 1, 2, 3, 4.

Newton Polynomials
Example 4.13. Construct a divided-difference table for f (x) = cos(x)
based on the five points (k, cos(k)), for k = 0, 1, 2, 3, 4. Use it to find the
coefficients ak and the four Newton interpolating polynomials Pk (x), for
k = 1, 2, 3, 4.
For simplicity we round off the values to seven decimal places, which
are displayed in Table 4.10. The nodes x0 , x1 , x2 , x3 and the diagonal
elements a0 , a1 , a2 , a3 , a4 in Table 4.10 are used in formula (16), and we
write down the first four Newton polynomials
P1 (x) = 1.0000000 − 0.4596977(x − 0.0),
P2 (x) = 1.0000000 − 0.4596977(x − 0.0) − 0.2483757(x − 0.0)(x − 1.0),
P3 (x) = 1.0000000 − 0.4596977(x − 0.0) − 0.2483757(x − 0.0)(x − 1.0)
+0.1465592(x − 0.0)(x − 1.0)(x − 2.0),
P4 (x) = 1.0000000 − 0.4596977(x − 0.0) − 0.2483757(x − 0.0)(x − 1.0),
+0.1465592(x − 0.0)(x − 1.0)(x − 2.0)
−0.0146568(x − 0.0)(x − 1.0)(x − 2.0)(x − 3.0)

Newton Polynomials
Table 4.10 Divided-Difference Table Used for Constructing the Newton Polynomials Pk (x)
xk f [xk ] f[ , ] f[ , , ] f[ , , , ] f[ , , , , ]
x0 = 0.0 1.0000000
x1 = 1.0 0.5403023 −0.4596977
x2 = 2.0 −0.4161468 −0.9564491 −0.2483757
x3 = 3.0 −0.9899925 −0.5738457 0.1913017 0.1465592
x4 = 4.0 −0.6536436 0.3363499 0.4550973 0.0879318 −0.0146568

Newton Polynomials
The following sample calculation shows how to find the coefficient a2 .

f [x1 ] − f [x0 ] 0.5403023 − 1.0000000
f [x0 , x1 ] = = = −0.4596977,
x1 − x0 1.0 − 0.0
f [x2 ] − f [x1 ] −0.4161468 − 0.5403023
f [x1 , x2 ] = = = −0.9564491,
x2 − x1 2.0 − 1.0
f [x1 , x2 ] − f [x0 , x1 ] −0.9564491 + 0.4596977
a2 = f [x0 , x1 , x2 ] = = = −0.2483757.
x2 − x0 2.0 − 0.0

Newton Polynomials
y y
1.0 1.0
y = P1 H x L
0.5 0.5
x x
1 2 3 1 2 3
- 0.5 - 0.5
y = cos(x)
y = P2 H x L
y = cos(x)
-1.0 -1.0
y
1.0
0.5
x
1 2 3
y = P3H xL
- 0.5
y = cos(x)
- 1.0
Figure: Newton polynomials Pk (x) for k = 1, 2 and 3

Interpolation and Polynomial Approximation: Professor Henry Arguello

Uploaded by

Copyright:

Available Formats

Interpolation and Polynomial Approximation: Professor Henry Arguello

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

Interpolation and Polynomial Approximation: Professor Henry Arguello

Uploaded by

Copyright:

Available Formats

Interpolation and Polynomial Approximation

Professor Henry Arguello

Universidad Industrial de Santander

March 12, 2018

High Dimensional Signal Processing Group

Professor Henry Arguello Numerical methods March 12, 2018 1 / 77

1 Interpolation and Polynomial Approximation

2 Taylor Series and Calculation of Functions

Professor Henry Arguello Numerical methods March 12, 2018 2 / 77

Interpolation is used to approximate different functions as sin(x),

Figure: Comparison between 2 approximations (a) The Taylor polynomial

Professor Henry Arguello Numerical methods March 12, 2018 3 / 77

Represent the elementary functions: sin(x), cos(x), ex , ln(x), etc. The

Table 4.1 Taylor Series Expansions for Some Common Functions

Professor Henry Arguello Numerical methods March 12, 2018 4 / 77

Table 4.2 Partial Sums Sn Used to Determine e

tend to a limit. The values of these 11 2.718281826199...

Professor Henry Arguello Numerical methods March 12, 2018 5 / 77

Theorem 9: Taylor Polynomial Approximation

The error term EN (x) has the form

for some value c = c(x) that lies between x and x0 .

Professor Henry Arguello Numerical methods March 12, 2018 6 / 77

Professor Henry Arguello Numerical methods March 12, 2018 7 / 77

Professor Henry Arguello Numerical methods March 12, 2018 7 / 77

Setting x = 1 in (4) gives the partial sum S1 5 = P1 5.Then the accuracy

Professor Henry Arguello Numerical methods March 12, 2018 8 / 77

Setting x = 1 in (4) gives the partial sum S1 5 = P1 5.Then the accuracy

Professor Henry Arguello Numerical methods March 12, 2018 8 / 77

Setting x = 1 in (4) gives the partial sum S1 5 = P1 5.Then the accuracy

Professor Henry Arguello Numerical methods March 12, 2018 8 / 77

Professor Henry Arguello Numerical methods March 12, 2018 10 / 77

The accuracy of a Taylor polynomial is increased when we choose N

Professor Henry Arguello Numerical methods March 12, 2018 11 / 77

R = 2.0, R = 1.5, R = 1.0, R = 0.5,

ex ≈ P5 (x) 0.65680499 0.07090172 0.00377539 0.00003578

Professor Henry Arguello Numerical methods March 12, 2018 12 / 77

Example: Establish the error bounds for the approximation ex ≈ P8 (x)

Professor Henry Arguello Numerical methods March 12, 2018 13 / 77

Example: Establish the error bounds for the approximation ex ≈ P8 (x)

Professor Henry Arguello Numerical methods March 12, 2018 13 / 77

Horner’s Method (Nested Multiplication)

Consider, for example, the function

The binomial formula can be used to expand f (x) in powers of x :

Now, Horner’s Method can be used to evaluate the polynomial in (10).

to evaluate f (x) now requires seven multiplications and eight additions

then f (x) has the Taylor series expansion

where c depends on N and x.

Suppose that the function y = f (x) is known at the N + 1 points

M = max{|f (N+1) (x)| : a ≤ x ≤ b}.

Professor Henry Arguello Numerical methods March 12, 2018 16 / 77

Let us briefly mention how to evaluate the polynomial P(x):

P(x) = aN xN + aN−1 xN−1 + · · · + a2 x2 + a1 x + a0 (1)

Horner’s Method of synthetic division is an efficient way to evaluate

P0 (x) = NaN xN−1 + (N − 1)aN−1 xN−2 + · · · + 2a2 x + a1 (2)

Professor Henry Arguello Numerical methods March 12, 2018 17 / 77

Example: The polynomial P(x) = −0.02x3 + 0.2x2 − 0.4x + 1.28 passes

Professor Henry Arguello Numerical methods March 12, 2018 18 / 77

Example: The polynomial P(x) = −0.02x3 + 0.2x2 − 0.4x + 1.28 passes