III.

5 Interpolation and quadrature

Polynomial interpolation is the process of finding a polynomial that equals data at a
precise set of points. Quadrature is the act of approximating an integral by a weighted
sum꞉
b n

∫ f (x)w(x)dx ≈ ∑ wj f (xj ).
a
j=1

In these notes we see that the two concepts are intrinsically linked꞉ interpolation leads
naturally to quadrature rules.
1. Polynomial Interpolation꞉ we describe how to interpolate a function by a polynomial
and a set of points.
2. Interpolatory quadrature rule꞉ polynomial interpolation leads naturally to ways to
integrate
functions, but onely realisable in the simplest cases.
1. Polynomial Interpolation
We already saw a special case of polynomial interpolation, where we saw that the
polynomial
n−1
n k
f (z) ≈ ∑ f̂ z
k
k=0

equaled at evenly spaced points on the unit circle꞉

f . But here we consider the
e
i2πj/n

following꞉
Definition 1 (interpolatory polynomial) Given distinct points
n and x1 , … , xn ∈ R n

samples f1 , … , fn ∈ R, a degree interpolatory polynomial satisfies

n − 1 p(x)

p(xj ) = fj

The easiest way to solve this problem is to invert the Vandermonde system꞉
Definition 2 (Vandermonde) The Vandermonde matrix associated with distinct points n

x1 , … , xn ∈ R is the matrix
n−1
1 x1 ⋯ x
⎡ 1 ⎤

⎢ ⎥
V ꞉= ⎢ ⎥
⎢ ⋮ ⋮ ⋱ ⋮ ⎥

⎣ n−1 ⎦
1 xn ⋯ xn
 Proposition 1 (interpolatory polynomial uniqueness) The interpolatory polynomial is
unique, and the Vandermonde matrix is invertible.
Proof Suppose and are both interpolatory polynomials. Then
p p̃ vanishes
p(x) − p̃(x)

at distinct points . By the fundamental theorem of algebra it must be zero, i.e.,

n xj

p = p̃.
For the second part, if for
Vc = 0 then for
c ∈ R we have
q(x) = c1 + ⋯ + cn x
n−1

⊤
q(xj ) = e Vc = 0
j

hence vanishes at distinct points and is therefore 0, i.e.,

q n c = 0 .
∎

Thus a quick‑and‑dirty way to to do interpolation is to invert the Vandermonde matrix

(which we saw in the least squares setting with more samples then coefficients)꞉
In [1]: using Plots, LinearAlgebra
f = x -> cos(10x)
n = 5

x = range(0, 1; length=n)# evenly spaced points (BAD for interpolation)

V = x .^ (0:n-1)' # Vandermonde matrix
c = V \ f.(x) # coefficients of interpolatory polynomial
p = x -> dot(c, x .^ (0:n-1))

g = range(0,1; length=1000) # plotting grid

plot(g, f.(g); label="function")
plot!(g, p.(g); label="interpolation")
scatter!(x, f.(x); label="samples")
Out[1]:

But it turns out we can also construct the interpolatory polynomial directly. We will use
the following which equal at one grid point and zero at the others꞉
1

Definition 3 (Lagrange basis polynomial) The Lagrange basis polynomial is defined as

x − xj (x − x1 ) ⋯ (x − xk−1 )(x − xk+1 ) ⋯ (x − xn )
ℓk (x) ꞉= ∏ =
j≠k xk − xj (xk − x1 ) ⋯ (xk − xk−1 )(xk − xk+1 ) ⋯ (xk − xn )

Plugging in the grid points verifies the following꞉

Proposition 2 (delta interpolation)
ℓk (xj ) = δkj

We can use these to construct the interpolatory polynomial꞉

Theorem 1 (Lagrange interpolation) The unique polynomial of degree at most n − 1

that interpolates at distinct points is꞉

f n xj

p(x) = f (x1 )ℓ1 (x) + ⋯ + f (xn )ℓn (x)

Proof Note that

n

p(xj ) = ∑ f (xj )ℓk (xj ) = f (xj )

j=1

so we just need to show it is unique. Suppose is a polynomial of degree at most

p̃(x)

n − 1 that also interpolates . Then

f vanishes at distinct points. Thus by the
p̃ − p n

fundamental theorem of algebra it must be zero.

∎

Example 1 We can interpolate exp(x) at the points 0, 1, 2 ꞉

(x − 1)(x − 2) x(x − 2) x(x − 1)
2 2
p(x) = ℓ1 (x) + eℓ2 (x) + e ℓ3 (x) = + e + e
(−1)(−2) (−1) 2

2 2 2
= (1/2 − e + e /2)x + (−3/2 + 2e − e /2)x + 1

Remark Interpolating at evenly spaced points is a really bad idea꞉ interpolation is

inheritely ill‑conditioned. The labs have explored this issue experimentally.
2. Interpolatory quadrature rules
By integrating an interpolant exactly we get a simple approach to approximating
integrals. Using the Lagrange basis we can rewrite this procedure as a simple weighted
sum꞉
Definition 4 (interpolatory quadrature rule) Given a set of points the x = [x1 , … , xn ]

interpolatory quadrature rule is꞉

n
w,x
Σn [f ] ꞉= ∑ wj f (xj )
j=1

where
b

wj ꞉= ∫ ℓj (x)w(x)dx
a

Proposition 3 (interpolatory quadrature is exact for polynomials) Interpolatory

quadrature is exact for all degree polynomials ꞉
n − 1 p

w,x
∫ p(x)w(x)dx = Σn [p]
a

Proof The result follows since, by uniqueness of interpolatory polynomial꞉

n

p(x) = ∑ p(xj )ℓj (x)

j=1

Example 2 (arbitrary points) Find the interpolatory quadrature rule for w(x) = 1 on
with points
[0, 1] ? We have꞉
[x1 , x2 , x3 ] = [0, 1/4, 1]
 1 1
(x − 1/4)(x − 1)
w1 = ∫ w(x)ℓ1 (x)dx = ∫ dx = −1/6
0 0
(−1/4)(−1)

1 1
x(x − 1)
w2 = ∫ w(x)ℓ2 (x)dx = ∫ dx = 8/9
0 0
(1/4)(−3/4)

1 1
x(x − 1/4)
w3 = ∫ w(x)ℓ3 (x)dx = ∫ dx = 5/18
0 0
3/4

That is we have
f (0) 8f (1/4) 5f (1)
w,x
Σn [f ] = − + +
6 9 18

This is indeed exact for polynomials up to degree (and no more)꞉ 2

w,x w,x w,x 2 w,x 3

Σn [1] = 1, Σn [x] = 1/2, Σn [x ] = 1/3, Σn [x ] = 7/24 ≠ 1/4.

Example 3 (Chebyshev roots) Find the interpolatory quadrature rule for

w(x) = 1/√1 − x on 2
with points equal to the roots of
[−1, 1] . This is a special T3 (x)

case of Gaussian quadrature which we will approach in another way below. We use꞉
1 1 1

2
∫ w(x)dx = π, ∫ xw(x)dx = 0, ∫ x w(x)dx = π/2
−1 −1 −1

Recall from before that x1 , x2 , x3 = √3/2, 0, −√3/2 . Thus we have꞉

1 1
x(x + √3/2) π
w1 = ∫ w(x)ℓ1 (x)dx = ∫ dx =
(√3/2)√3√1 − x2 3
−1 −1

1 1
(x − √3/2)(x + √3/2) π
w2 = ∫ w(x)ℓ2 (x)dx = ∫ dx =
(−3/4)√1 − x2 3
−1 −1

1 1
(x − √3/2)x π
w3 = ∫ w(x)ℓ3 (x)dx = ∫ dx =
(−√3)(−√3/2)√1 − x2 3
−1 −1

(It's not a coincidence that they are all the same but this will differ for roots of other
OPs.) That is we have
π
w,x
Σn [f ] = (f (√3/2) + f (0) + f (−√3/2)
3

This is indeed exact for polynomials up to degree n − 1 = 2 , but it goes all the way up
to2n − 1 = 5 ꞉
π
w,x w,x w,x 2
Σn [1] = π, Σn [x] = 0, Σn [x ] = ,
2
3π
w,x 3 w,x 4 w,x 5
Σn [x ] = 0, Σn [x ] = , Σn [x ] = 0
8

9π 5π
w,x 6
Σn [x ] = ≠
32 16
We shall explain this miracle in the next chapter.