Numerical integration methods

[parisse]

(07-24-2022 08:55 PM)Wes Loewer Wrote:  (Note: It's possible that I am mistranslating "ordre". What the paper called "ordre 30" seems to be "degree 29" (ie, 30 terms), but I could be wrong about this. Francophones, please correct me if I am mistaken.)

"order" may mean max degree of polynomial where the method is exact, or the exponent in the step of the error, up to the author choice...

Quote:Since the 15-node Gauss-Kronrod is exact up to degree 22 and the Prime's method is exact up to degree 29, it would seem that the Prime's method is a bit better with only a small added overhead. A slightly more accurate calculation per iteration could occasionally reduce the need for further recursion. It would be interesting to do some comparisons of these two methods with some well-behaved functions as well as more "temperamental" ones.

On a final note, I am curious why Hairer chose only 6 nodes for the 3rd calculation instead of 7. Any insight on that would be welcomed. [Edit: Perhaps it is important to have the 3rd calculation's 6 nodes be a subset of the 2nd calculation's 14 nodes. This would not be the case if 7 nodes were used for the 3rd calculation. It is not clear to me why this subset would be necessary or even desirable.]

Yes, it is important not to re-evaluate the function for efficiency reasons. The idea is to estimate the error of the order 30 method in h^30 by
err1*(err1/err2)^2
where err1=abs(i30-i14) is in h^14 and err2=abs(i30-i6) in h^6