Numerical Differentiation

Forward and Backward Differences

Inspired by the definition of derivative:
Chapter 4
Numerical Differentiation and Integration f (x0 + h) − f (x0 )
f 0 (x0 ) = lim ,
h→0 h
choose a small h and approximate
Per-Olof Persson
[email protected] f (x0 + h) − f (x0 )
f 0 (x0 ) ≈
h
Department of Mathematics
University of California, Berkeley The error term for the linear Lagrange polynomial gives:

Math 128A Numerical Analysis f (x0 + h) − f (x0 ) h 00

f 0 (x0 ) = − f (ξ)
h 2
Also known as the forward-difference formula if h > 0 and the
backward-difference formula if h < 0

General Derivative Approximations Commonly Used Formulas

Using equally spaced points with h = xj+1 − xj , we have the

three-point formulas
Differentiation of Lagrange Polynomials
1 h2
Differentiate f 0 (x0 ) = [−3f (x0 ) + 4f (x0 + h) − f (x0 + 2h)] + f (3) (ξ0 )
2h 3
n
X 0 1 h2 (3)
(x − x0 ) · · · (x − xn ) (n+1) f (x0 ) = [−f (x0 − h) + f (x0 + h)] − f (ξ1 )
f (x) = f (xk )Lk (x) + f (ξ(x)) 2h 6
(n + 1)! 1 h2
k=0 0
f (x0 ) = [f (x0 − 2h) − 4f (x0 − h) + 3f (x0 )] + f (3) (ξ2 )
2h 3
to get
00 1 h2 (4)
f (x0 ) = 2 [f (x0 − h) − 2f (x0 ) + f (x0 + h)] − f (ξ)
n
X h 12
f (n+1) (ξ(xj )) Y
f 0 (xj ) = f (xk )L0k (xj ) + (xj − xk ) and the five-point formula
(n + 1)!
k=0 k6=j
1
This is the (n + 1)-point formula for approximating f 0 (xj ). f 0 (x0 ) = [f (x0 − 2h) − 8f (x0 − h) + 8f (x0 + h) − f (x0 + 2h)]
12h
h4 (5)
+ f (ξ)
30

Optimal h Richardson’s Extrapolation

Suppose N (h) approximates an unknown M with error

Consider the three-point central difference formula:
M − N (h) = K1 h + K2 h2 + K3 h3 + · · ·
0 1 h2
f (x0 ) = [f (x0 + h) − f (x0 − h)] − f (3) (ξ1 )
2h 6 then an O(hj ) approximation is given for j = 2, 3, . . . by
Suppose that round-off errors ε are introduced when  
h Nj−1 (h/2) − Nj−1 (h)
computing f . Then the approximation error is Nj (h) = Nj−1 +
2 2j−1 − 1
 
 f˜(x0 + h) − f˜(x0 − h)  ε h2
 0 The results can be written in a table:
f (x0 ) −  ≤ + M = e(h)
 2h  h 6
O(h) O(h2 ) O(h3 ) O(h4 )
where f˜ is the computed function and |f (3) (x)| ≤ M 1: N1 (h) ≡ N (h)
Sum of truncation error h2 M/6 and round-off error ε/h 2: N1 ( h2 ) ≡ N ( h2 ) 3: N2 (h)
p
Minimize e(h) to find the optimal h = 3 3ε/M 4: N1 ( h4 ) ≡ N ( h4 ) 5: N2 ( h2 ) 6: N3 (h)
7: N1 ( h8 ) ≡ N ( h8 ) 8: N2 ( h4 ) 9: N3 ( h2 ) 10: N4 (h)
Richardson’s Extrapolation Numerical Quadrature

Integration of Lagrange Interpolating Polynomials

Select {xP0 , . . . , xn } in [a, b] and integrate the Lagrange polynomial
Pn (x) = ni=0 f (xi )Li (x) and its truncation error term over [a, b]
If some error terms are zero, different and more efficient
to obtain
formulas can be derived
Z b Xn
Example: If
f (x) dx = ai f (xi ) + E(f )
a i=0
M − N (h) = K2 h2 + K4 h4 + · · ·
with
then an O(h2j ) approximation is given for j = 2, 3, . . . by Z b
  ai = Li (x) dx
h Nj−1 (h/2) − Nj−1 (h)
Nj (h) = Nj−1 + a
2 4j−1 − 1
and
Z bY
n
1
E(f ) = (x − xi )f (n+1) (ξ(x)) dx
(n + 1)! a i=0

Trapezoidal and Simpson’s Rules The Newton-Cotes Formulas

The Trapezoidal Rule The Closed Newton-Cotes Formulas
Linear Lagrange polynomial with x0 = a, x1 = b, h = b − a, gives Use nodes xi = x0 + ih, x0 = a, xn = b, h = (b − a)/n:
Z b Z
h h3 b n
X
f (x) dx = [f (x0 ) + f (x1 )] − f 00 (ξ) f (x) dx ≈ ai f (xi )
a 2 12 a i=0
Z xn Z xn Y (x − xj )
Simpson’s Rule ai = Li (x) dx = dx
Second Lagrange polynomial with x0 = a, x2 = b, x1 = a + h, x0 x0 (xi − xj )
j6=i
h = (b − a)/2 gives
n = 1 gives the Trapezoidal rule, n = 2 gives Simpson’s rule.
Z x2
h h5
dx = [f (x0 ) + 4f (x1 ) + f (x2 )] − f (4) (ξ) The Open Newton-Cotes Formulas
x0 3 90
Use nodes xi = x0 + ih, x0 = a + h, xn = b − h,
Definition h = (b − a)/(n + 2). Setting n = 0 gives the Midpoint rule:
The degree of accuracy, or precision, of a quadrature formula is Z x1
the largest positive integer n such that the formula is exact for xk , h3
f (x) dx = 2hf (x0 ) + f 00 (ξ)
for each k = 0, 1, . . . , n. x−1 3

Composite Rules Romberg Integration

Theorem
Let f ∈ C 2 [a, b], h = (b − a)/n, xj = a + jh, µ ∈ (a, b). The
Composite Trapezoidal rule for n subintervals is Compute a sequence of n integrals using the Composite
  Trapezoidal rule, where m1 = 1, m2 = 2, m3 = 4, . . . and
Z b n−1
X
h b − a 2 00 mn = 2n−1 .
f (x) dx = f (a) + 2 f (xj ) + f (b) − h f (µ)
a 2 12 The step sizes are then hk = (b − a)/mk = (b − a)/2k−1
j=1
The Trapezoidal rule becomes
Theorem   
Z b 2k−1
X−1
Let f ∈ C 4 [a, b], n even, h = (b − a)/n, xj = a + jh, µ ∈ (a, b). hk 
f (x) dx = f (a) + f (b) + 2  f (a + ihk )
The Composite Simpson’s rule for n subintervals is a 2
i=1
 
Z b (n/2)−1 n/2 (b − a) 2 00
h X X − hk f (µk )
f (x) dx = f (a) + 2 f (x2j ) + 4 f (x2j−1 ) + f (b) 12
a 3
j=1 j=1
b − a 4 (4)
− h f (µ)
180
Romberg Integration Romberg Integration

Let Rk,1 denote the trapezoidal approximation, then

MATLAB Implementation
h1 (b − a)
R1,1 = [f (a) + f (b)] = [f (a) + f (b)] function R=romberg(f,a,b,n)
2 2
1
R2,1 = [R1,1 + h1 f (a + h2 )] h=b-a;
2 R=zeros(n,n);
1 R(1,1)=h/2*(f(a)+f(b));
R3,1 = {R2,1 + h2 [f (a + h3 ) + f (a + 3h3 )]}
2  for i=2:n
k−2
2X R(i,1)=1/2*(R(i-1,1)+h*sum(f(a+((1:2^(i-2))-0.5)*h)));
1
Rk,1 = Rk−1,1 + hk−1 f (a + (2i − 1)hk ) for j=2:i
2 R(i,j)=R(i,j-1)+(R(i,j-1)-R(i-1,j-1))/(4^(j-1)-1);
i=1
end
Apply Richardson extrapolation to these values: h=h/2;
end
Rk,j−1 − Rk−1,j−1
Rk,j = Rk,j−1 +
4j−1 − 1

Error Estimation Adaptive Quadrature

The error term in Simpson’s rule requires knowledge of f (4) :

Z b
Rb
h5 (4) To compute a f (x) dx within a tolerance ε > 0, first apply
f (x) dx = S(a, b) − f (µ)
a 90 Simpson’s rule with h = (b − a)/2 and with h/2
If
Instead, apply it again with step size h/2:
    
 
Z b
a+b
  
a+b
  
1 h5 S(a, b) − S a, a + b − S a + b , b  < 15ε
f (x) dx = S a, +S ,b − f (4) (µ̃)  2 2 
a 2 2 16 90
then the integral is sufficiently accurate
The assumption f (4) (µ) ≈ f (4) (µ̃) gives the error estimate
Z
If not, apply the technique to [a, (a + b)/2] and [(a + b)/2, b],
 b    
 a+b a+b  and compute the integral within a tolerance of ε/2
 f (x) dx − S a, −S ,b 
 a 2 2  Repeat until each portion is within the required tolerance
    
1  a+b a+b 
≈ S(a, b) − S a, −S , b 
15  2 2

Gaussian Quadrature Legendre Polynomials

Basic idea: Calculate both nodes x1 , . . . , xn and coefficients The Legendre polynomials Pn (x) have the properties
c1 , . . . , cn such that 1 For each n, Pn (x) is a monic polynomial of degree n (leading
Z coefficient 1)
b n
X R1
2 P (x)Pn (x) dx = 0 when P (x) is a polynomial of degree
f (x) dx ≈ ci f (xi ) −1
a
less than n
i=1
The roots of Pn (x) are distinct, in the interval (−1, 1), and
Since there are 2n parameters, we might expect a degree of symmetric with respect to the origin.
precision of 2n − 1 Examples:
Example: n = 2 gives the rule
Z √ ! √ ! P0 (x) = 1, P1 (x) = x
1
− 3 3 1 3
f (x) dx ≈ f +f 2
P2 (x) = x − P3 (x) = x3 − x
−1 3 3 3 5
6 3
with degree of precision 3 P4 (x) = x4 − x2 +
7 35
Gaussian Quadrature Computing Gaussian Quadrature Coefficients

MATLAB Implementation
function [x,c]=gaussquad(n)
Theorem %GAUSSQUAD Gaussian quadrature
Suppose x1 , . . . , xn are roots of Pn (x) and
P=zeros(n+1,n+1);
Z
n
Y 1 P([1,2],1)=1;
x − xj
ci = dx for k=1:n-1
−1 x i − xj P(k+2,1:k+2)=((2*k+1)*[P(k+1,1:k+1) 0]- ...
j6=i
k*[0 0 P(k,1:k)])/(k+1);
If P (x) is any polynomial of degree less than 2n, then end
x=sort(roots(P(n+1,1:n+1)));
Z 1 n
X
P (x) dx = ci P (xi ) A=zeros(n,n);
−1 i=1 for i=1:n
A(i,:)=polyval(P(i,1:i),x)’;
end
c=A\[2;zeros(n-1,1)];

Arbitrary Intervals Double Integrals

Consider the double integral

Rb ZZ
Transform integrals a f (x) dx into integrals over [−1, 1] by a
change of variables: f (x, y) dA, R = {(x, y) | a ≤ x ≤ b, c ≤ y ≤ d}
R
2x − a − b 1 Partition [a, b] and [c, d] into even number of subintervals n, m
t= ⇔ x = [(b − a)t + a + b]
b−a 2
Step sizes h = (b − a)/n and k = (d − c)/m
Gaussian quadrature then gives Write the integral as an iterated integral
Z b Z 1   ZZ Z b Z d 
(b − a)t + (b + a) (b − a)
f (x) dx = f dt f (x, y) dA = f (x, y) dy dx
a −1 2 2 R a c

and use any quadrature rule in an iterated manner.

Composite Simpson’s Rule Double Integration Non-Rectangular Regions

The Composite Simpson’s rule gives

Z b Z  n m
d
hk X X
f (x, y) dy dx = wi,j f (xi , yj ) + E
a c 9
i=0 j=0 The same technique can be applied to double integrals of the form
where xi = a + ih, yj = c + jk, wi,j are the products of the nested Z bZ d(x)
Composite Simpson’s rule coefficients (see below), and the error is f (x, y) dy dx
a c(x)
 
(d − c)(b − a) 4 ∂ 4 f 4
4∂ f
E=− h (η̄, µ̄) + k (η̂, µ̂) The step size for x is still h = (b − a)/n, but for y it varies with x:
180 ∂x4 ∂y 4
d(x) − c(x)
d 1 4 2 4 1 k(x) =
m

4 16 8 16 4

c 1 4 2 4 1
a b
Gaussian Double Integration Improper Integrals with a Singularity

The improper integral below, with a singularity at the left

endpoint, converges if and only if 0 < p < 1 and then
Z b
For Guassian integration, first transform the roots rn,j from
b
1 (x − a)1−p  (b − a)1−p
p
dx =  =
[−1, 1] to [a, b] and [c, d], respectively a (x − a) 1−p a 1−p
The integral is then More generally, if
Z bZ n n
d
(b − a)(d − c) X X g(x)
f (x, y) dy dx ≈ cn,i cn,j f (xi , yj ) f (x) = , 0 < p < 1, g continuous on [a, b],
a c 4 (x − a)p
i=1 j=1

Similar techniques can be used for non-rectangular regions construct the fourth Taylor polynomial P4 (x) for g about a:

g 00 (a)
P4 (x) = g(a) + g 0 (a)(x − a) + (x − a)2
2!
g 000 (a) g (4) (a)
+ (x − a)3 + (x − a)4
3! 4!

Improper Integrals with a Singularity Improper Integrals with a Singularity

and write
Z b Z b Z b For the first integral, use the Composite Simpson’s rule to
g(x) − P4 (x) P4 (x)
f (x) dx = dx + dx compute the integral of G on [a, b], where
a a (x − a)p a (x − a)p
( g(x)−P (x)
4
The second integral can be computed exactly: (x−a)p , if a < x ≤ b
g(x) =
Z 4 Z 0, if x = a
b
P4 (x) X b
g (k) (a)
dx = (x − a)k−p dx
a (x − a)p k! (k)
Note that 0 < p < 1 and P4 (a) agrees with g (k) (a) for each
k=0 a
X4 k = 0, 1, 2, 3, 4, so G ∈ C 4 [a, b] and Simpson’s rule can be applied.
g (k) (a)
= (b − a)k+1−p
k!(k + 1 − p)
k=0

Singularity at the Right Endpoint Infinite Limits of Integration

R∞
For an improper integral with a singularity at the right An integral of the form a x1p dx, with p > 1, can be converted to
endpoint b, make the substitution z = −x, dz = −dx to an integral with left endpoint singularity at 0 by the substitution
obtain
t = x−1 , dt = −x−2 dx, so dx = −x2 dt = −t−2 dt
Z b Z −a
f (x) dx = f (−z) dz which gives
a −b
Z ∞ Z 0 Z 1/a
which has its singularity at the left endpoint 1 tp 1
dx = − dt = dt
For an improper integral with a singularity at c, where a xp 1/a t2 0 t2−p
a < c < b, split into two improper integrals R∞
More generally, this variable change converts a f (x) dx into
Z b Z c Z b
Z ∞ Z 1/a  
f (x) dx = f (x) dx + f (x) dx 1
a a c f (x) dx = t−2 f dt
a 0 t