4
$\begingroup$

I'm trying to solve a second order differential equation with the shooting method but, it appears to be a stiff system. This is worst for a larger parameter $\mu$. I've tried different methods: StiffnessSwitching, LSODA, BDF also, changing MaxStepFraction and WorkingPrecision options. None of these has helped. An example of the code is:

P[r_, rv_] := -(r^2/16) (1 - Sqrt[(rv/r)^3 + 1] + 3/2 (rv/r)^3 
Hypergeometric2F1[1/3, 1/2, 4/3, -(rv/r)^3])
nonintphi :=  Block[{r0 = 0.01, g = 1, λ = 1, μ = 3, σ =
10^2, rv = (4/π σ)^(1/3), inf = 10 rv}, 
ParametricNDSolve[{1/r^2 D[r^2 Φ'[r], r] - λ Φ[r]^3 
+ (μ^2 - g (P[r, rv])^2) Φ[r] ==  0, 
Φ[r0] == 0, Φ'[r0] == B}, Φ, {r, r0, inf}, B, 
Method -> {"LSODA"}, MaxStepFraction -> 1/100000, WorkingPrecision -> 30,
MaxSteps -> Infinity]]

Block[{r0 = 0.01, g = 1, λ = 1, μ = 3, σ = 10^2, 
rv = (4/π σ)^(1/3), inf = 10 rv}, 
bcs = FindRoot[Φ[B][inf] == 
Sqrt[(μ^2 - g (P[inf, rv])^2)/λ] /. nonintphi, {B, 0}, AccuracyGoal -> 3, 
PrecisionGoal -> 4]];

The problem arises when using FindRoot and gives something like:

At r == 3.5998792500342004380773953162236682802682889219723012678651`\ 30., step size is effectively zero; singularity or stiff system \ suspected.

Same thing happens if I use NDSolve with Method->"Shooting Method". At the end, I want to use the value $\mu=10^{24}$, so this problem becomes even worse. How can I get the integration to work?

$\endgroup$
  • $\begingroup$ Welcome to Mathematica.SE! I suggest the following: 1) As you receive help, try to give it too, by answering questions in your area of expertise. 2) Take the tour! 3) When you see good questions and answers, vote them up by clicking the gray triangles, because the credibility of the system is based on the reputation gained by users sharing their knowledge. Also, please remember to accept the answer, if any, that solves your problem, by clicking the checkmark sign! $\endgroup$ – Michael E2 Oct 21 '15 at 2:27
  • 1
    $\begingroup$ May I ask where your DE came from? $\endgroup$ – J. M. is away Oct 21 '15 at 2:45
  • $\begingroup$ Please reduce your code to remove extraneous issues such as the use of FindRoot and P and condense it into a self-contained example focused on the integral you having problems with. $\endgroup$ – m_goldberg Oct 21 '15 at 4:09
  • 1
    $\begingroup$ @Michael Seifert Nonlinear coefficients are not supported in NDSolve, Mathematica 10.2 $\endgroup$ – user31001 Oct 21 '15 at 16:16
  • 1
    $\begingroup$ The DE is the equation of motion for a field. The FindRoot part is the most important one since it's where I find the errors and I need it to solve the equation by shooting method. $\endgroup$ – marRrR Oct 21 '15 at 23:40
9
$\begingroup$

The goal of this question, I believe, is selecting Φ[r0] so that the solution connects smoothly to the asymptotic solution at large r. As will become apparent, large is about 3. Begin be rewriting the code as

r0 = 0.01; g = 1; λ = 1; μ = 3; σ = 10^2; rv = (4/π σ)^(1/3); inf = 5;
p = -(r^2/16) (1 - Sqrt[(rv/r)^3 + 1] + 
            3/2 (rv/r)^3 Hypergeometric2F1[1/3, 1/2, 4/3, -(rv/r)^3]);
s1 = ParametricNDSolveValue[{1/r^2 D[r^2 Φ'[r], r] + (μ^2 - g p^2 - λ Φ[r]^2) Φ[r] == 0, 
    Φ[r0] == b, Φ'[r0] == 0}, Φ, {r, r0, inf}, {b}, WorkingPrecision -> 30];

From the ODE, it is clear that the asymptotic solution is given by

(μ^2 - g p^2 - λ Φ[r]^2) == 0

Now, by trial and error I have found that b = .055035287243 is the desired parameter value. Clearly, this value can be obtained using NDSolve with Method -> "Shooting", but it is late and I am tired.

Plot[{s1[.055035287243][r], Sqrt[(μ^2 - g p^2)]}, {r, r0, inf}, 
    PlotRange -> {-3, 3}, AxesLabel -> {r, Φ}]

enter image description here

Here, the blue curve is from ParametricNDSolve, and the orange is the asymptotic solution,

Φa = Sqrt[(μ^2 - g p^2)]

They are indistinguishable to the eye for r > 2.6. A numerical comparison of the two at r -> inf is quite good.

{s1[.055035287243][r], Sqrt[(μ^2 - g p^2)] // N} /. r -> inf
(* {2.58994, 2.59165} *)

Addendum

Willinski is correct (+1) that I interchanged the Φ[r0] and Φ'[r0] boundary conditions last night. However, my inadvertent error has one advantage: It produces a solution well behaved as r approaches 0.

r0 = 10^-10; 
s3 = ParametricNDSolveValue[{1/r^2 D[r^2 Φ'[r], r] + (μ^2 - g p^2 - λ Φ[r]^2) Φ[r] == 0, 
    Φ[r0] == b, Φ'[r0] == 0}, Φ, {r, r0, inf}, {b}, WorkingPrecision -> 30];
Plot[{s3[.054949][r], Sqrt[(μ^2 - g p^2)]}, {r, r0, inf}, 
    AxesLabel -> {r, Φ}]

which produces a plot indistinguishable to the eye from my plot above but extends down to r0 = 10^-10.

Incidentally, why the problem is ill-behaved at large r can be seen by linearizing the ODE about the asymptotic solution, Φa, defined above. One finds after a small amount of algebra that small errors in Φ - Φa grow exponentially at a rate Sqrt[2] Φa, or about 10^9 over a distance of 5.

Automating "Trial and Error"

The OP understandably would prefer a better way than trial and error to find the parameter, here named b0, that yields the desired curve. Using FindRoot on s3[b][inf] is unreliable, because the integration does not reach r = inf for b much larger than b0. Below is a plot of the value of r at which the integration stops, as a function of b, with b0 superimposed.

Plot[Quiet@s3[b]["Domain"][[1, 2]], {b, .054, .056}, AxesLabel -> {b, r}, 
   Epilog -> {PointSize[Large], Red, Point[{0.0549492, inf}]}]

enter image description here

with s3 calculated without WorkingPrecison -> 30, because doing so is much faster and does not change the curve. Interestingly, Method -> StiffnessSwitching also does not change the curve. Strictly speaking, the ODE is not stiff; it is an unstable separatrix.

So, to use FindRoot, we need to operate in the flat portion of the curve above, which in turn means finding the point at which it no longer is flat. Searching for that point by successive bifurcations seems to work well.

bl = 0; bu = .1;
Do[bt = (bl + bu)/2; If[Quiet@s3[bt]["Domain"][[1, 2]] < inf, bu = bt, bl = bt], {i, 20}]
FindRoot[s3[b][inf] == Sqrt[(μ^2 - g p^2)] /. r -> inf, {b, 0, bl}]
(* {b -> 0.0549492} *)

as desired.

inf = 10

At the request of the OP, here are the corresponding results for inf = 10. With s3 computed with WorkingPrecision -> 30,

inf = 10; bl = 0; bu = .1;
Do[bt = (bl + bu)/2; If[Quiet@s3[bt]["Domain"][[1, 2]] < inf, bu = bt, bl = bt], {i, 50}]
NumberForm[{bl, Quiet@s3[bl]["Domain"][[1, 2]], bu, Quiet@s3[bu]["Domain"][[1, 2]]}, 15]
(* {0.0549492117281679, 10.000000000000, 0.0549492117281679, 9.99998284445001} *)

b0 = b /. FindRoot[s3[b][inf] == Sqrt[(μ^2 - g p^2)] /. r -> inf, {b, 5/100, bl},
    WorkingPrecision -> 30];
NumberForm[b0, 18]
(* 0.0549492117275881041 *)

Plot[Quiet@s3[b]["Domain"][[1, 2]], {b, .054, .056}, AxesLabel -> {b, r}, 
    Epilog -> {PointSize[Large], Red, Point[{b0, inf}]}]

enter image description here

Plot[{s3[b0][r], Sqrt[(μ^2 - g p^2)]}, {r, r0, inf}, AxesLabel -> {r, Φ}]

enter image description here

Further increasing inf will require even larger WorkingPrecision.

Asymptotic Series Solution

A slightly better solution can be obtained for large r by expanding the ODE as a power series in 1/r.

n = 20; app = Sum[a[i] r^-i, {i, 0, n}];
Unevaluated[1/r^2 D[r^2 D[Φ[r], r], r] + (μ^2 - g p^2 - λ Φ[r]^2) Φ[r]] /. Φ[r] -> app;
CoefficientList[Series[%, {r, Infinity, n}] // Normal, 1/r];
exp = app /. Solve[Thread[% == 0], Array[a[# - 1] &, n + 1]][[3]];

N[exp]
(* 3. -10.5543/r^2 -19.7382/r^4 + 167.977/r^5 - 86.7254/r^6 + 777.599/r^7 ... *)

The value of b for which the numerical solution smoothly matches to this series is computed as before, with the result here called b1.

b1 = b /. FindRoot[s3[b][inf] == exp /. r -> inf, {b, 5/100, bl}, 
    WorkingPrecision -> 30];
NumberForm[b1, 18]
(* 0.0549492117275880915 *)

which differs only slightly from b0 computed above. The plot of s3[b1] is, of course, indistinguishable to the eye from the plot above. Nonetheless, the series exp agrees more closely with s3[b1] than Sqrt[(μ^2 - g p^2)] does.

err = -exp[[2]]; 
LogPlot[{Sqrt[(μ^2 - g p^2)] - s3[b1][r], -exp + s3[b1][r], err}, {r, r0, 10}, 
    AxesLabel -> {r, ΔΦ}, PlotRange -> {{3, 10}, {10^-10, 1}}]

enter image description here

(The green curve is the upper bound on the error of the series expansion, obtained by noting that the series is alternating after the first few terms.) Thus, the series solution is far more accurate at larger r, where it converges well. Nonetheless, Φa = Sqrt[(μ^2 - g p^2)] is a reasonable approximation over a broad range.

A series solution for small r also can be derived.

n = 14; app = c + Sum[a[i] r^i, {i, 2, n + 2, 1/2}];
Unevaluated[1/r^2 D[r^2 D[Φ[r], r], r] + (μ^2 - g p^2 - λ Φ[r]^2) Φ[r]] /. Φ[r] -> app;
CoefficientList[Simplify[(Assuming[r >= 0, Series[%, {r, 0, n}]] // Normal) 
    /. r -> z^2, z > 0], z];
exp0 = (app /. Solve[Thread[% == 0], Array[a[(# + 3)/2] &, 2 n + 1]])[[1]];

From it, we find that the solution is indeed free of singularities at r = 0, provided that Φ'[0] == 0. Unfortunately, the series does not converge well for r > 1. Hence, it is not practical to link this series with the asymptotic series to obtain a complete analytical solution to the ODE. (I had hoped to do so, when I began deriving these series.)

$\endgroup$
  • $\begingroup$ It's a fine work! (+1) :) $\endgroup$ – user31001 Oct 21 '15 at 16:06
  • $\begingroup$ I think that the asymptotic form is wrong since the second derivative still survives when r goes to 0. So the asymptotic is Φ''[r]+(μ^2 - g p^2 - λ Φ[r]^2) == 0, but form the plots it seems that the second derivative becomes small so your asymptotics look ok. Also, the problem in this ode arises when I try to shoot to Sqrt[(μ^2 - g (P^2)/λ]. So, when I run the code using the Findroot it gives the error: At r == 3.5998792500342004380773953162236682802682889219723012678651`30., step size is effectively zero; singularity or stiff system suspected. >> $\endgroup$ – marRrR Oct 21 '15 at 19:14
  • $\begingroup$ The issue is that I want an efficient way to find this b, not just trial and error because later I will change the parameters \mu \lambda g and \sigma. $\endgroup$ – marRrR Oct 21 '15 at 23:43
  • $\begingroup$ @marRrR I have added code to automate finding the desired parameter value. It includes a procedure for selecting the range of parameters for which FindRoot works well. $\endgroup$ – bbgodfrey Oct 22 '15 at 16:52
  • $\begingroup$ Thanks for all the help, this automated way seems really useful. The only thing I'm worried of is that the initial value b needed to shoot to the desired boundary condition at inf is encountered after the flat line, which I think it's the case for the values \mu=10^24 and \sigma=10^38. I haven't had the time to confirm these but seems like it would be the case. How could I overcome these? $\endgroup$ – marRrR Oct 23 '15 at 2:41
2
$\begingroup$

I have Phi'[r0] selected, as did the OP. The value of b was determined empirically.

r0 = 1/100; g = 1; λ = 1; μ = 3; σ = 10^2; rv = (4/π σ)^(1/3); inf = 10 rv;
(*stiff system, reduced inf -> 8*) 

p = -(r^2/16) (1 - Sqrt[(rv/r)^3 + 1] + 3/2 (rv/r)^3 Hypergeometric2F1[1/3, 1/2, 4/3, -(rv/r)^3]);

s1 = ParametricNDSolveValue[
{1/r^2 D[r^2 Φ'[r], r] + (μ^2 -g p^2 - λ Φ[r]^2) Φ[r] == 0,
     Φ[r0] == 0, Φ'[r0] == b}, Φ, {r, r0, 8}, {b}, WorkingPrecision -> 30]

Plot[{s1[#][r], Sqrt[(μ^2 - g p^2)]}, {r, r0, 8}, PlotRange -> All] & /@ {5, 5.77533449, 5.7753345}

enter image description here

Check the solution:

s2 = NDSolveValue[
{1/r^2 D[r^2 Φ'[r],r] + (μ^2 -g p^2 - λ Φ[r]^2) Φ[r] == 0, 
Φ[r0] == 0, Φ'[r0] == 5.77533449`30}, Φ, {r, r0, 8}, WorkingPrecision -> 30]

Plot[{s2[r], Sqrt[(μ^2 - g p^2)]}, {r, r0, 8}, PlotRange -> All]

enter image description here

s2[r0]
0.*10^-31

s2'[r0]
5.77533449000000000000
$\endgroup$
  • $\begingroup$ This indeed gives a solution, but you found the value of b empirically. At the end, the equation I need to solve is for \mu=10^24 and finding the value by trial and error is not efficient so I would like to do it with FindRoot. $\endgroup$ – marRrR Oct 21 '15 at 23:37
  • $\begingroup$ @marRrR Many things in the world were determined empirically! It will remain you probably nothing to solve the system. You know, your system with the given values is very stiff and singular, dependent of b. There is a lot of work to find a solution at all. Look at the above plots, the system already responds to the 7th decimal place! $\endgroup$ – user31001 Oct 22 '15 at 11:14
  • $\begingroup$ If I just try to solve it empirically for my value for \mu=10^24 and \sigma=10^38 whenever I try to plot or get the value at inf for a given b it gives the stiff system error. $\endgroup$ – marRrR Oct 22 '15 at 11:54
  • $\begingroup$ @marRrR It seems that you need to work on your DE system. With your mu and sigma try out Table[{Sqrt[(\[Mu]^2 - g p^2)]}, {r, r0, 1, 0.1}]. $\endgroup$ – user31001 Oct 22 '15 at 12:52
  • $\begingroup$ I know that ([Mu]^2 - g p^2) becomes negative in that region, that's the behaviour I'm searching for. $\endgroup$ – marRrR Oct 23 '15 at 1:35

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.