You are on page 1of 3

Numerical Integration

Estimate Polynomial h el Name

1 00 3
A2 = h2 (f0 + f1 ) linear b−a − 12 f h trapezoidal rule

A3 = h3 (f0 + 4f1 + f2 ) quadratic 1 (4) 5


(b − a)/2 − 90 f h Simpson rule

3h 3 (4) 5
A4 = 8
(f0 + 3f1 + 3f2 + f3 ) cubic (b − a)/3 − 80 f h

Table 1: Newton-Cotes quadrature rules.

To verify these formulas, make a Taylor series expansion of f (x) about


x0 :
1 1 1 (4)
f (x) = f0 + (x − x0 )f00 + (x − x0 )2 f000 + (x − x0 )3 f0000 + (x − x0 )4 f0 + · · ·
2 3! 4!
Then the exact integral I is
Z x
I= f (x̄)dx̄ =
x0

1 1 1 1 (4)
(x−x0 )f0 + (x−x0 )2 f00 + (x−x0 )3 f000 + (x−x0 )4 f0000 + (x−x0 )5 f0 +· · ·
2 3! 4! 5!
In Simpson’s rule,

h 4h3 00 2h4 000 5h5 (4)


A3 = (f (x0 ) + 4f (x0 + h) + f (x0 + 2h)) ≈ 2hf0 +2h2 f00 + f + f + f .
3 3 0 3 0 18 0
From the formula for the exact integral (setting x = x0 + 2h),

4h3 00 2h4 000 4h5 (4)


I ≈ 2hf0 + 2h2 f00 + f + f + f
3 0 3 0 15 0
and the local error is
h5 (4)
el = I − A3 ≈ − f .
90 0
Romberg Integration
The Romberg method generates a sequence of TR estimates, each halving
the h of the previous one, combined with repeated Richardson extrapolation
to get approximations that are h2 higher in order:

e → h2 e
A00
h A01 A10
↓ A02 A11 A20
h/2 A03 A12 A21 A30 · · ·
..
.
The stopping criterion is when |Aj0 − Aj−1,1 | ≤ , where  specifies the de-
sired number of significant digits. The entries in the Romberg table are
defined by
b−a 1 1
 
A0k = k f0 + f1 + f2 + · · · + f2k −1 + f2k (TR estimates)
2 2 2
1
Ank = (4n An−1,k+1 − An−1,k ) (Richardson extrapolation).
4n − 1

The proof of the Romberg method involves the Euler-McLaurin Theorem,


that the error for the trapezoidal rule is entirely even in h:
Z xN
1 1 B2
 
f (x)dx = h f0 + f1 + f2 + · · · + fN −1 + fN − h2 (fN0 − f00 )−· · · −
x0 2 2 2!
B2k 2k  (2k−1) (2k−1)

h fN − f0 − ···
(2k)!
where B2k are the even Bernoulli numbers: B0 = 1, B2 = 1/6, B4 = −1/30,
B6 = 1/42, . . . (B1 = −1/2, B2k+1 = 0 for k = 1, 2, . . . )

2
Proof of the Romberg Method

By induction. (i) Basis step: Richardson extrapolation. Suppose

A(h) = a0 + a2 h2 + a4 h4 + · · ·

where a0 is the exact answer (this is true for TR by the Euler-McLaurin


Theorem). Then
h2 h4
A(h/2) = a0 + a2 + a4 + · · ·
4 16
Combining, we have

4A(h/2) − A(h)
= a0 + O(h4 ).
3
Simpson’s rule can be derived from the trapezoidal rule in this way.

(ii) Induction step: repeated Richardson extrapolation. Suppose

A(h) = a0 + a2n h2n + a2n+2 h2n+2 + · · ·

Then
h2n h2n+2
A(h/2) = a0 + a2n + a 2n+2 + ···
22n 22n+2
Combining, we obtain the expansion

4n A(h/2) − A(h)
= a0 + O(h2n+2 ).
4n − 1

You might also like