• Background and motivation:

Suppose I want to approximate the integral $\int_{a}^{b}f(x)dx$ using evenly-spaced sampled values of $f$: $f(a + \frac{i}{n}(b-a)), i=0,\cdots,n$. By the linearity of the integral, it is natural to try a linear approximation: $$ \int_{a}^{b}f(x)dx \sim \sum_{i=0}^{n} w_i f(a + \frac{i}{n}(b-a))$$ This is the idea of the Newton-Cotes formulas. The weights $w_i$ are chosen in such a way that the formula is accurate for all polynomials of degree up to $n$ (and then the error depends on the $(n+1)$'th derivative of $f$).

In other words, just solve the following $n+1$ equations: $$ \int_{a}^{b} x^{k} dx = \frac{b^{k+1}-a^{k+1}}{k+1}= \sum_{i=0}^{n} w_i (a + \frac{i}{n}(b-a))^k, k=0,\cdots,n$$ The relevant matrix is a Vandermonde matrix, so there is indeed a solution.

  • The even miracle and my question:

According to this wikipedia page, it seems that for even $n$, this formula is also accurate for $k=n+1$, and in general the error term depends on the $(n+2)$'th derivative of the $f$. This is a non-trivial linear algebra statement.

So my question is: why is true?

  • 1
    $\begingroup$ If you're pleased to get just one degree of accuracy for free, look up Gauss quadrature. $\endgroup$ – Chris Godsil Jun 1 '13 at 11:31

It's not a miracle, just a nice property, and there are much more impressive methods out there, such as Gauss, Clenshaw-Curtis, double exponential rule. See also mathworld.

What happens is easier to see on an interval $[-h,h]$ centered around $0$. If you have an analytic function $f$, then $$\int_{-h}^h f(x)\,dx = \sum_{m\geq0} \frac{f^{(m)}(0)}{(m+1)!} 2h^{m+1}[m\text{ even}], $$ where the sum involves only terms with even $m$, because $\int_{-h}^h x^m$ vanishes when $m$ is odd.

With an equally spaced choice of points $x_j$, and $w_j = w_{n-j}$ (which happens for the Newton-Cotes rule), the sum $$ \sum_{j=0}^{n} w_j x_j^m $$ will vanish when $m$ is odd, because we will have $x_j^m = -x_{n-j}^m$. To understand why $w_j = w_{n-j}$, note that $$ w_j = \int_{-h}^h L_j(x)\,dx, $$ where $\{L_j(x)\}$ are the basis for the Lagrange interpolating polynomial of $f(x)$ on the points $x_j$.

So the rule correctly integrates all odd terms in the power series to zero, leaving only even terms in the error.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.