I have a mathematica script/code that I wrote that involves a few nested NIntegrate functions but its not converging. I let it run for about a week before calling it quits. I am a beginner with Mathematica and composed this script looking through the manual, so I am hoping there is a better way of solving this and that this problem is not just numerically impossible.

Can any help?

e = 1.60217733*10^-19;
re = 2.81794092*10^-15;
gamma = 1000/0.510999060000000;
n = (200*10^-12)/e;
sigmaZ = 10*10^-6;
lambda = n/sigmaZ;
Xt = 100*10^-6;
lq = 0.1;
k = 2;
phiM = lq*k*Xt;
rhoM = lq/phiM;
drift = 10;

EFieldKernel[Xi_?NumericQ, d_?NumericQ, s_?NumericQ, sprime_?NumericQ] := 
(Xi/Xt) (2 n*lambda)/(Xt - Xi - 
    d*((d - s + sprime - Sqrt[d^2 + (Xi - Xt)^2])/(-rhoM + (
   d (rhoM + Xi - Xt))/Sqrt[d^2 + (Xi - Xt)^2])) - 
1/2 rhoM*((
  d - s + sprime - Sqrt[d^2 + (Xi - Xt)^2])/(-rhoM + (
   d (rhoM + Xi - Xt))/Sqrt[d^2 + (Xi - Xt)^2]))^2); 

EFieldLongitudinal[Xi_?NumericQ, d_?NumericQ, s_?NumericQ] := 
 NIntegrate[EFieldKernel[Xi, d, s, sprime], {sprime, 0, sigmaZ}]

EFieldNegativeSum[d_?NumericQ, s_?NumericQ] := 
 NIntegrate[-EFieldLongitudinal[Xi, d, s], {Xi, 
   Xt - rhoM (1 - Cos[phiM]) - d*Sin[phiM], Xt}]
EfieldPositiveSum[d_?NumericQ, s_?NumericQ] := 
  EFieldLongitudinal[Xi, d, s], {Xi, -Xt, 
   Xt - rhoM (1 - Cos[phiM]) - d*Sin[phiM]}]

  s_?NumericQ] := (re/gamma*phiM) (EFieldNegativeSum[d, s]) + (re/
 gamma*-phiM) (EfieldPositiveSum[d, s])
Etot[s_?NumericQ] := NIntegrate[wake[d, s], {d, 0, drift}]

Erms = Sqrt[
  NIntegrate[lambda*Etot[s]^2, {s, 0, sigmaZ}] - 
  NIntegrate[lambda*Etot[s], {s, 0, sigmaZ}]^2]
  • $\begingroup$ You have lots of undefined variables, e.g., lambda, n, re, gamma, phiM, etc. You should provide example values for them. Next, I believe the best way to compute a nested NIntegrate function is to rewrite it as a single NIntegrate, and let NIntegrate figure out what method to use. $\endgroup$
    – Carl Woll
    Jun 21 '17 at 21:30
  • $\begingroup$ You can do this Integrate[(Xi/Xt) (2 n*lambda)/(Xt - Xi - d*((d - s + sprime - Sqrt[d^2 + (Xi - Xt)^2])/(-rhoM + ( d (rhoM + Xi - Xt))/Sqrt[d^2 + (Xi - Xt)^2])) - 1/2 rhoM*(( d - s + sprime - Sqrt[d^2 + (Xi - Xt)^2])/(-rhoM + ( d (rhoM + Xi - Xt))/Sqrt[d^2 + (Xi - Xt)^2]))^2), {sprime, 0, sigmaZ}] operation analytically. Maybe this would help. $\endgroup$ Jun 21 '17 at 21:35
  • $\begingroup$ You are correct! So sorry I forgot the variables, this should be more helpful. l tried @RolfMertig suggestion, but it still does not converge. Any one else with some thoughts? $\endgroup$ Jun 22 '17 at 7:31
  • $\begingroup$ What do you mean by "does not converge" Does the integration fail with errors? Or does it not finish? $\endgroup$
    – Carl Woll
    Jun 22 '17 at 14:59
  • $\begingroup$ @CarlWoll It does not finish. No errors. I think I let it run for about a week last time. But, even if it just takes a lot of time I was hoping there was a formalism or technique that could speed it up. $\endgroup$ Jun 22 '17 at 17:02

In general, it is much better to use a single NIntegrate instead of using multiple nested integrals. Here is a contrived example:

f[x_?NumericQ] := NIntegrate[g[x t], {t, 0, 10}];
g[x_?NumericQ] := NIntegrate[Cos[u], {u, 0, x}];

NIntegrate[f[x], {x,0,1}] //AbsoluteTiming

{4.70959, 2.92526}

The equivalent single NIntegrate version is:

NIntegrate[Cos[u], {x, 0, 1}, {t, 0, 10}, {u, 0, x t}] //AbsoluteTiming

{0.441031, 2.92526}

The timing discrepancy can easily be much worse, if the nested NIntegrate has integration convergence issues, which is what I think you're experiencing.

I took a look at your code, replacing all _?NumericQ patterns with just _, and replacing all NIntegrate heads with Inactive[NIntegrate]. Then it was a relatively simple exercise to convert the nested NIntegrate objects into multiple unnested NIntegrate objects. When I did this, the NIntegrate finished quickly, but had convergence issues. I think this solves your speed issue. You could ask about the convergence issues in a different question.


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.