# Global vs local adaptive integration: lattice propagator

I was doing the first exercise in the paper Lattice QCD for Novices. This is the expected result:

With the default "GlobalAdaptive" method for NIntegrate it threw errors saying that the error had increased more than 2000 times, so I switched to "LocalAdaptive". The result had some jitters, but was recognizable.

I decided to go back to "GlobalAdaptive" despite the warnings and slower time, and it actually looks really good despite the warnings.

However, the documentation says global adaptive is generally faster for multidimensional integrals. I'm looking for insight for why it is not true in this case, and in general, any settings that would give a better and faster result. I'm just getting into the options for the Monte Carlo methods. Here's the code:

stepSize = 1/2; stepCount = 8; potential[x_] := x^2/2; mass = 1;

action[x_List] :=
Sum[mass/(2 stepSize) (x[[j + 1]] - x[[j]])^2 +
stepSize potential[x[[j]]], {j, stepCount}]

ListLinePlot@
ParallelTable[
NIntegrate @@
Join[{E^-action[Join[{b}, Table[x[i], {i, stepCount - 1}], {b}]]},
Table[{x[i], -5, 5}, {i, stepCount - 1}], {Method ->
"LocalAdaptive"}], {b, 0, 2, .2}]

ListLinePlot@
ParallelTable[
Quiet[NIntegrate @@
Join[{E^-action[Join[{b}, Table[x[i], {i, stepCount - 1}], {b}]]},
Table[{x[i], -5, 5}, {i, stepCount - 1}]]], {b, 0, 2, .2}]

-
Just wanted to note that NIntegrate typically manages to use all cores of the CPU, so ParallelTable doesn't usually help with speeding things up (in fact it slows them down). I don't know what NIntegrate does in parallel internally, this is just an observation. –  Szabolcs Feb 5 '14 at 20:26
@Szabolcs The problem asks for 11 separate NIntegrate calls (each with a separate start and end boundary for the propagator), so I'm just using ParallelTable to get a quick boost on those. –  Michael Hale Feb 5 '14 at 22:13
@Szabolcs I mean, I understand your point, but in this case I go from 31 seconds to 12 seconds with ParallelTable for the local adaptive one. If anyone has insight into why this is a case where parallel does help NIntegrate I'm interested in that too. –  Michael Hale Feb 5 '14 at 22:57
In that case it's clearly worth doing the parallelization! Maybe it's the global adaptive method that has some built-in parallelization. –  Szabolcs Feb 5 '14 at 23:22