Many numerical integration formulas over a finite interval have the form
That is, the integral on the left can be approximated by evaluating the integrand f at particular nodes and taking the weighted sum, and the error is some multiple of a derivative of f evaluated at a point in the interval [a, b].
This post will give several examples, showing how they all fit into to framework above, then discuss the impossibility of extending this framework to infinite integrals.
Simpson’s 3/8 rule
Simpson’s rule says
In this case the x‘s are evenly spaced a distance of h apart. The weights are 3h/8, 9h/8, 9h/8, and 3h/8.
We have n = 4, k = 4, and C = −3h5/80.
We don’t know a priori what the value of ξ will be, only that there will be some value of ξ between x0 and x3 that makes the equation hold. If we have an explicit bound on the 4th derivative of f then we have an explicit bound on the error in approximating the integral by the weighted sum.
Bode’s rules
A sequence of quadrature rules go by the name “Bode’s rule.” Here is one example, also known as Weddle’s rule.
As with Simpson’s 3/8 rule, you could map the formula for Bode’s rule(s) to the template at the top of the post.
Gauss quadrature
Gauss’ formula says
for some ξ in [−1, 1]. Here the limits of integration are fixed, though you could use a change of variables to integrals over other finite integrals into this form.
Unlike Bode’s rule and Simpson’s rule, the x‘s are not evenly spaced but are the zeros of Pn, the Legendre polynomial of degree n. The weights are related to the x‘s and the derivative P′n evaluated at the x‘s. The constant C is a complicated function of n but is independent of f.
Note that the error term involves the (2n)th derivative of f. This explains why Gaussian integration can be more accurate than other methods using the same number of function evaluations. The non-uniform spacing of the integration nodes enables higher-order error terms.
Non-existence theorem
Although many integration rules over a finite interval have the form
Davis and Rabinowitz [1] proved that there cannot be an integration rule of the form
The proof, given in [1], takes only about one page. The entire article is a little more than a page, and about half the article is preamble to the proof.
Related posts
- Three surprises with the trapezoid rule
- Romberg integration
- Lobatto quadrature
- Orthogonal polynomials
[1] P. J. Davis and P. Rabinowitz. On the Nonexistence of Simplex Integration Rules for Infinite Integrals. Mathematics of Computation, Vol. 26, No. 119 (July, 1972), pp. 687–688
It would be worth mentioning integration limits from -Inf to Inf, as covered by Gauss-Hermite quadrature. It seems the impossibility theorem does not extend there?
The reason Gauss-Hermite is not an exception is that the function “f” is not the same on both sides of the equation. The left side is exp(-x^2) f(x) and the right side is f(x). Or if you say the left side is g(x) = exp(-x^2) f(x), then the right side is not g(x) but exp(x^2) g(x).
Great post! I had forgotten these rules.
What happens for poorly-behaved functions? Is there a test to check when above rules apply?
The prohibition of poorly-behaved functions is implicit in the formulas. When the error term depends on the kth derivative of f, the implicit assumption is that the kth derivative exists and is continuous.
For a field experiment in which we had to record the average outdoor air temperature for each day, I suggested that we measure the temperature at 5:04:18 AM and 6:55:41 PM and average the two readings. Nobody got the joke.