# Improve accuracy of NIntegrate with GlobalAdaptive over ImplicitRegion?

Let's say that I want to integrate some arbitrarily "nice" function (uniformly $C^{\infty}$-smooth, for example) over an ImplicitRegion in more than three dimensions. For example, let's consider the integral $\int_{D\times D} 1\,\mathrm{d}\mu$, where $D$ is the unit disk in two dimensions. We can approximate this in Mathematica as follows:

disk2 = ImplicitRegion[x1^2 + y1^2 <= 1 && x2^2 + y2^2 <= 1, {x1, y1, x2, y2}]
val[n_] := NIntegrate[1, Element[{x1, y1, x2, y2}, disk2], WorkingPrecision -> n]


The exact value of this is $\pi^2$, but the numerical value returned by Mathematica 11.1 doesn't converge to $\pi^2$ for increasing $n$. Namely, val[50]-Pi^2 and val[100]-Pi^2 both give me the same numerical error, which is roughly $1.38\times 10^{-18}$. With slightly more complicated functions (that are still $C^1$ almost everywhere, at least), the error we get can be much worse; consider the following integral:

NIntegrate[Sqrt[(x1 - x2)^2 + (y1 - y2)^2], Element[{x1, y1, x2, y2}, disk2], WorkingPrecision -> n]


We should expect this integral to have a value of $\frac{128\pi}{45}$, but the output differs from that value by a little more than $4\times 10^{-5}$, regardless of the setting for WorkingPrecision. Additionally, it doesn't seem that DiscretizeRegion can really be applied to regions that can't be expressed as the product of $n$-dimensional regions for $n\leq 3$, which in the case of a more general region rules out the strategy here. Are there any general strategies I can use to get with arbitrary precision and accuracy the values of numerical integrals of $C^1$ functions on ImplicitRegions in more than three dimensions? Thanks for your time.

• Just out of curiosity, how long does it take to evaluate val[50] and val[100] for you? It seems to last forever for me... May 6 '17 at 11:25
• I get an NIntegrate::eincr error, which stops the integration with the error you report, 1.38*^-18. Don't you? If not, what version are you using? If so, why not report such information in the question? May 6 '17 at 11:34
• I don't get an error at all. The computation completes perfectly for me. Also, I'm using Mathematica 11.1, as I mentioned. May 7 '17 at 1:27
• That is to say, Mathematica doesn't complain at all. There is a slight numerical error, which is what my issue is. May 7 '17 at 1:44
• Sorry about overlooking the version. I get this (V11.1.1): i.stack.imgur.com/dnIrW.png It suggests a way to improve the result. But if you get no error, then it won't help. May 7 '17 at 2:28