Suppose we have an InterpolatingFunction from NDSolve

L = 5;

sol = First@NDSolve[{D[u[x, t], t] == D[u[x, t], {x, 2}], u[x, 0] == 0, 
      u[-5, t] == Sin[t], u[5, t] == 0}, u, {x, -L, L}, {t, 0, 10}]

Defining the functions:

int1st[x_?NumericQ, t_?NumericQ] := NIntegrate[D[Evaluate[u[s, t] /. sol], s]*Cot[(π (x - s))/(2*L)], {s, -L, x, L}, Method -> "PrincipalValue"];

intNest[x_?NumericQ, t_?NumericQ] := NIntegrate[D[Evaluate[u[xp, t] /. sol], xp]*int1st[xp, t]*Cot[(π (x - xp))/(2*L)], {xp, -L, x, L}, Method -> "PrincipalValue"];

which are, respectively, similar to the circular Hilbert transform of a periodic function and its nest (please see the section 'Hilbert transform on the circle' on that webpage). Note in intNest, int1st is called with xp as its argument, which acts as the variable of integration of intNest.

For example, plotting int1st at tn=5; I obtain the curve within 25s.

Plot[int1st[x,tn], {x, -L, L}, PlotRange -> {{-L, L}, All}, ImageSize -> 400, PlotPoints -> 60, AspectRatio -> 0.5, Frame -> True, Axes -> False, PlotStyle -> {Black, Thick}]

enter image description here

My problem is on intNest: when plotting it, there is no warning but my computer runs for hours without output. Is my definition of the intNest correct. If correct, how to make it more efficient?

Plot[intNest[x, tn], {x, -L, L}, PlotRange -> {{-L, L}, All}, ImageSize -> 400, PlotPoints -> 60, AspectRatio -> 0.5, Frame -> True, Axes -> False, PlotStyle -> {Black, Thick}]

After reading this post, I modified the function definitions:

ClearAll[int1st, intNest];

Block[{x, t}, int1st[x_?NumericQ, t_?NumericQ] = NIntegrate[D[u[s, t] /. sol, s]*Cot[(π (x - s))/(2*L)], {s, -L, x, L}, Method -> "PrincipalValue"]];

Block[{x, t}, intNest[x_?NumericQ, t_?NumericQ] = NIntegrate[D[u[xp, t] /. sol, xp]*int1st[xp, t]*Cot[(π (x - xp))/(2*L)], {xp, -L, x, L}, Method -> "PrincipalValue"]];

Findings: int1st can be plotted well in about 20s (same result as above), but it is still very slow to plot intNest.


Update: Used the built-in "InterpolationPointSubdivision" method.

Update 2: Memoization helps because the subdivision is at the same points (given by xcoords in the code below) every time. So in intNest2[], the call int1st2[xp, t] will initially be made at the same xp for each subinterval, unless the subinterval contains the singular point x. I say, "initially," because the recursive subdivision of each subinterval in the intNest2[] call may be different depending on the value of x and its effect on the complete integrand.

Some explanation in the comments. If I have more time, I might be able to add more:

L = 5;

(* OP's *)
ClearAll[int1st, intNest];
int1st[x_?NumericQ, t_?NumericQ] := 
   Derivative[1, 0][u /. sol][s, t]*
    Cot[(π (x - s))/(2*L)], {s, -L, x, L}, 
   Method -> "PrincipalValue"];

intNest[x_?NumericQ, t_?NumericQ] := 
   D[Evaluate[u[xp, t] /. sol], xp]*int1st[xp, t]*
    Cot[(π (x - xp))/(2*L)], {xp, -L, x, L}, 
   Method -> "PrincipalValue"];

(* with singularity removed from NIntegrate and 
 * manual interpolating nodes subdivision *)
ClearAll[int1st2, intNest2];

(* interpolating nodes *)
{xcoords, tcoords} = u["Coordinates"] /. sol;

mem : int1st2[x_?NumericQ, t_?NumericQ] :=
  mem = Block[{s}, (* memoization seems to help with Plot[] *)
      du = Derivative[1, 0][u /. sol],
      dux = Derivative[1, 0][u /. sol][x, t]},
       Piecewise[{{du[s, t]*Cot[(π (x - s))/(2*L)] - 
           dux (2*L)/(π (x - s)), x != s}}], (* subtract singular part *)
       {s, -L, x, L},
       Method -> {"InterpolationPointsSubdivision", (* divide interval at nodes *)
         "SymbolicProcessing" -> 0},
       PrecisionGoal -> 4 (* PDE solution is low-precision *)
       ] + dux ( (2*L) Log[(L + x)/(L - x)])/π (* add PV integral of singular part *)

intNest2[x_?NumericQ, t_?NumericQ] := 
     du = Derivative[1, 0][u /. sol], 
     dux = Derivative[1, 0][u /. sol][x, t]*int1st2[x, t]}, 
      Piecewise[{{du[xp, t]*int1st2[xp, t]*
           Cot[(π (x - xp))/(2*L)] - dux (2*L)/(π (x - xp)), 
         x != xp}}], {xp, -L, x, L}, 
      Method -> {"InterpolationPointsSubdivision", 
        "SymbolicProcessing" -> 0}, PrecisionGoal -> 4] + 
     dux*((2*L) Log[(L + x)/(L - x)])/π]];

Single-call timings:

int1st2[4, 5] // Quiet // AbsoluteTiming
int1st[4, 5] // AbsoluteTiming
  {0.017778, -1.6291}
  {0.056669, -1.6291}

intNest2[4, 5] // AbsoluteTiming
intNest[4, 5] // AbsoluteTiming
  {5.96797, 1.78776}
  {52.9943, 1.78775}
Plot[int1st2[x, 5], {x, -L, L}, PlotRange -> {{-L, L}, All}, 
  ImageSize -> 400, PlotPoints -> 60, AspectRatio -> 0.5, 
  Frame -> True, Axes -> False, 
  PlotStyle -> {Black, Thick}] // AbsoluteTiming

enter image description here

There are 441 points here, so multiply by the single call timings to get an estimate of the time to plot:

Cases[%, Line[p_] :> Length@p, Infinity]
(*  {441}  *)

The nested integral is still going to be slow. You can use options like PlotPoints and MaxRecursion to reduce the number of points plotted (that is, reduce the number of function calls).

  • $\begingroup$ Nice! Could you memoize int1st2 to take advantage of the recycled points? $\endgroup$ – Chris K Feb 15 '20 at 0:50
  • $\begingroup$ @ChrisK Memoization doesn't help with intNest[4, 5]. Rather doubt it would help with Plot, either. There aren't many integration rules that recycle points. But if the same call to intNest needs to be done multiple times, then obviously memoization would speed things up here. $\endgroup$ – Michael E2 Feb 15 '20 at 3:21
  • $\begingroup$ I should have used “InterpolationPointsSubdicision”. Will update when I can log into the site (tomorrow). The rest doesn’t matter that much. Some of it is typos. (As I implied, I didn’t have much time. ) $\endgroup$ – Michael E2 Feb 15 '20 at 4:10
  • $\begingroup$ @MichaelE2 I haven't looking into why, but memoizing with int1st2[x_?NumericQ, t_?NumericQ] := int1st2[x, t] = allows Plot[intNest2[x, 5], ..., PlotPoints -> 5] to run in four minutes, whereas without memoizing it doesn't complete in at least fifteen minutes. $\endgroup$ – Chris K Feb 15 '20 at 14:55
  • $\begingroup$ @Chris K, did you mean f[x_?NumericQ, t_?NumericQ] := f[x, t] =... can memorize all the data of f[x,t] even though it is a function of x and t? $\endgroup$ – jsxs Feb 15 '20 at 15:18

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.