# Composite Simpson's Rule vs. Adaptive Simpson's Method

I was playing in Matlab when the following occured: I had an integral $I$ which I computed with the (a) Composite Simpson's Rule and (b) Adaptive Simpson's Method (my teacher told me that the built-in function quad in Matlab uses this method). These computations resulted in two approximations $I_1,I_2$ of $I$ for (a) and (b) respectively.

Prior to these computations I computed a reference solution to the problem, that is: an approximation of $I$ with a very low tolerance level (for example $10^{-15}$, compared to $10^{-6}$ which was used in later computations). Call this reference solution (approximation) $I_*$.

After finding $I_*$ I was able to compare $|I_1-I_*|$ with $|I_2-I_*|$. To my surprise $|I_1-I_*|<|I_2-I_*|$, and so: method (a) was more effective than method (b) in this particular example.

Now follows a general question: is there some general condition(s) for which it is true that $|I_1-I_*|<|I_2-I_*|$?

EDIT: Put

$y(x)=\frac{1-\arctan(p(x-1))/\pi}{2-\cos(\pi x)}$.

I am integrating $\pi \cdot y^2$ over $0 \leq x \leq 2.6$:

$$\pi \int_0^{2.6} y(x)^2 dx.$$

For $p=1$ I found that $|I_1-I_*|>|I_2-I_*|$, for $p=1000$ I found that $|I_1-I_*|<|I_2-I_*|$.

Below is the graph of $\pi \cdot y(x)^2$ for $p=1, 1000$

• Could you give the integral ? – Claude Leibovici Jan 8 '14 at 6:04
• @ClaudeLeibovici Done. – bobbo Jan 8 '14 at 6:10

• This (see relation (33)) tells me (?) that given a tolerance level $\epsilon$, we should check that $|I_{[a,b]}-(I_{[a,c]}+I_{[c,b]})|<15\epsilon$ (where $I_{[a,b]}$ is the approximation of the integral over $[a,b]$) before proceeding. I do not understand how this answers my question, or hints at an answer to my question, or any other relation provided in the pdf-file. – bobbo Jan 8 '14 at 6:37
• Yes, I also think it depends on the specific integrand. That is why there is such a change when we change $p$ from 1 to 1000. I can compute the integral with quad in Matlab, and using my own implementation of Composite Simpson's rule. What options do you mean? Here is another example (p. 5 of 6) for which $|I_1-I_*|<|I_2-I_*|$: math.usm.edu/lambers/mat460/fall09/lecture30.pdf ("Although the composite Simpson method contains half the error [!] of the adaptive quadrature method, 176 more function evaluations are required."). – bobbo Jan 8 '14 at 7:29