Based on the heat equation of the Mathematica Manual tutorial, I wrote the complex counterpart (Schrödinger) equation, for the free particle propagation of an initial wavepacket.

  I D[u[t, x, y], t] == -D[u[t, x, y], {x, 2}] - 
    D[u[t, x, y], {y, 2}], u[0., x, y] == Exp[-(x^2. + y^2.)], 
  u[t, -5., y] == u[t, 5., y], u[t, x, -5.] == u[t, x, 5.]
  }, u, {t, 0., 1.}, {x, -5., 5.}, {y, -5., 5.}

However the solver chokes with serveral warnings, the most serious being that the Maximum number of iterations has been reached that stops the calculation at t == 0.48. But the worst is that the solutions (plotted with Table[Plot3D[ Evaluate[Abs[u[t, x, y] /. First[sol]]^2], {x, -5, 5}, {y, -5, 5}, PlotRange -> All, PlotPoints -> 100, Mesh -> False], {t, 0.0, 0.05, 0.01}]) looks completely wrong with diverging values.

I know the NDSolve is not magic but does anybody know of an option to pass that will make this problem tractable with NDSolve?

I couldn't find a coherent explanation of the NDSolve options in the manual or anywhere else. Otherwise I would be playing with known methods of propagation, like Crank-Nicolson (for time propagation) and spectral methods (for spatial coordinates). A first nice step would to control the spatial (x and y) resolution and the time (t) resolution independently.

Note 1: One knows that the time propagation of this problem for that particular initial condition is a simple spreading Gaussian probability, in particular the solution is well defined and smooth.

Note 2: I tried with and without periodic boundary conditions in both cases the result is numerically wrong (diverging values).

Note 3: I did some progress for the simpler 1+1D equation, that problem is well undercontrol with most of the defaults/automatics of NDSolve, I can propagate for a while:

sol = NDSolve[
   I D[u[t, x], t] == -D[u[t, x], {x, 2}], u[0., x] == Exp[-(x^2.)], 
   u[t, 5.] == 0, u[t, -5.] == 0
   }, u, {t, 0., 20.}, {x, -5., 5.}, MaxStepSize -> 0.1
Animate[Plot[Evaluate[Abs[u[t, x] /. First[sol]]^2], {x, -5, 5}, 
  PlotRange -> {0, 1}], {t, 0, 17, 0.01}]
  • $\begingroup$ Do you absolutely insist on using NDSolve for this, or are other approaches OK? $\endgroup$ – acl Jun 29 '13 at 11:43
  • $\begingroup$ @acl, other methods with Mathematica is fine, but if I can use the leverage of NDSolve for part of the problem that would be fine too. The main question is how to fine control NDSolve for this particular problem, for example things like MaxStepSize control the discretizations of all dimensions equaly. $\endgroup$ – alfC Jun 29 '13 at 18:34
  • 2
    $\begingroup$ Please use MaxStepSize -> 0.01, AccuracyGoal -> 3, PrecisionGoal -> 3. This will solve your problem. If you want me to become more specific I could try to post an answer, but not today. This will be a short night... $\endgroup$ – Stefan Jun 30 '13 at 22:07
  • $\begingroup$ @Stefan +1 Are the AccuracyGoal/PrecisionGoal needed? $\endgroup$ – Eric Brown Jul 1 '13 at 4:00
  • $\begingroup$ @Stefan, I tried your suggested options (in the 2D+1 version) and after 3 minutes I get a nomem error and no points in the result (interpolation in the time range {0.,0.}). I tried MaxStepSize->0.1 and I get same as at the beginning, some quickly diverging answer. I am using Mathematica 8. $\endgroup$ – alfC Jul 1 '13 at 4:29

I think it's worth pointing out that the problem can be solved "straightforwardly" (i.e., really using only NDSolve) once you know the options that Stefan used in ProcessEquations (which I upvoted because those options are the main ingredient):

Below I show the original problem of a Gaussian wave packet with no initial momentum, and then a modified case where an initial momentum has been imparted, making the initial condition complex as well. I call the complex wave function $\psi$ and plot its absolute value:

ψ = u /. 
   First@NDSolve[{I D[u[t, x, y], t] == -D[u[t, x, y], {x, 2}] - 
        D[u[t, x, y], {y, 2}], u[0., x, y] == Exp[-(x^2. + y^2.)], 
      u[t, -5., y] == u[t, 5., y], u[t, x, -5.] == u[t, x, 5.]}, 
     u, {t, 0., 2.}, {x, -5., 5.}, {y, -5., 5.}, 
     Method -> {"MethodOfLines", 
       "SpatialDiscretization" -> {"TensorProductGrid", 
         "DifferenceOrder" -> "Pseudospectral"}}];

pl =
  Table[Plot3D[Abs[ψ[t, x, y]], {x, -5, 5}, {y, -5, 5}, 
    PlotRange -> {0, 1}], {t, 0, 2, .1}];

Export["spreading.gif", pl, AnimationRepetitions -> Infinity, 
 "DisplayDurations" -> .4]


ψ = 
 u /. First@
   NDSolve[{I D[u[t, x, y], t] == -D[u[t, x, y], {x, 2}] - 
       D[u[t, x, y], {y, 2}], 
     u[0., x, y] == Exp[-(x^2. + y^2.)] Exp[3 I x], 
     u[t, -10., y] == u[t, 10., y], u[t, x, -10.] == u[t, x, 10.]}, 
    u, {t, 0., 1.}, {x, -10., 10.}, {y, -10., 10.}, 
    Method -> {"MethodOfLines", 
      "SpatialDiscretization" -> {"TensorProductGrid", 
        "DifferenceOrder" -> "Pseudospectral"}}]

pl =
  Table[Plot3D[Abs[ψ[t, x, y]], {x, -10, 10}, {y, -10, 10}, 
    PlotRange -> {0, 1}], {t, 0, 1, .1}];

Export["moving.gif", pl, AnimationRepetitions -> Infinity, 
 "DisplayDurations" -> .4]


Edit 2: Adding a potential energy term.

The above numerical solutions are basically for a free particle, except that the spatial grid is forcing us to choose some boundary conditions on the sides of the square. Periodic boundary conditions are a common choice. But the whole effort is overkill for a free particle because the solutions can be obtained analytically. It gets more interesting if we add an arbitrary potential energy to see how the wave packet is deflected over time.

The periodic boundary conditions in this calculation allow you to add a potential energy to the Hamiltonian, as long as it doesn't conflict with the periodicity of box. Here is an example where I added the potential

$$V(x, y) = - 20 \cos(\frac{\pi x}{10}) \cos(\frac{\pi y}{10})$$

with a box of side length $10$. This potential vanishes on the box boundaries, and has an attractive center at the origin.

Also, I started the Gaussian slightly offset from the center, with a momentum tangential to the equipotential lines, so we expect it to go around with some angular momentum:

ψ = u /. 
   First@NDSolve[{I D[u[t, x, y], t] == -D[u[t, x, y], {x, 2}] - 
        D[u[t, x, y], {y, 2}] - 
        20 Cos[Pi x/10] Cos[Pi y/10] u[t, x, y], 
      u[0., x, y] == Exp[-((x - 1)^2. + y^2.)] Exp[I y], 
      u[t, -5., y] == u[t, 5., y], u[t, x, -5.] == u[t, x, 5.]}, 
     u, {t, 0., 3.}, {x, -5., 5.}, {y, -5., 5.}, 
     Method -> {"MethodOfLines", 
       "SpatialDiscretization" -> {"TensorProductGrid", 
         "DifferenceOrder" -> "Pseudospectral"}}];

pl = Table[
   Plot3D[Abs[ψ[t, x, y]], {x, -5, 5}, {y, -5, 5}, 
    PlotRange -> {0, 1}], {t, 0, 3, .1}];

Export["revolve.gif", pl, AnimationRepetitions -> Infinity, 
 "DisplayDurations" -> .1]


The packet still disperses but is clearly trapped in the potential minimum, as expected.

  • $\begingroup$ Great! (+1)...i was still fighting with my Iter approach ... close to despair. now that there is your answer i know that there must be a solution with my approach as well, but the OP's question is perfectly answered by you, which gives me some time to find out, what i'm doing wrong actually. $\endgroup$ – Stefan Jul 3 '13 at 6:16
  • 1
    $\begingroup$ Thanks, those are the pretty much options I came up with. I don't understand what stopped Stefan for continuing, at some point I will have to use the iteration approach to save memory (I'll wait for his answer). BTW, this is just the first step the next step will be to a add a potential (which even in the simplest cases seem to make NDSolve choke) and change the metric. $\endgroup$ – alfC Jul 3 '13 at 6:52
  • 1
    $\begingroup$ If you have problems with added potentials, that could happen if the modifications aren't consistent with the periodic boundary conditions. I'll add an example that works for me where the potential matches the periodicity of the box. $\endgroup$ – Jens Jul 3 '13 at 17:44
  • $\begingroup$ @Jens, exactly I was using a parabolic potential that was not compatible with the PBC (probably creating effective cusps in the potential). Next question will be about absorbing boundary conditions, necessary for the problem I have: esnt.cea.fr/Images/Page/27/besse.pdf $\endgroup$ – alfC Jul 4 '13 at 20:23
  • 1
    $\begingroup$ @user21 The answer is yes, but it's a vast subject. Some good and reasonable simple applications could be found in physical chemistry and mesoscopic physics (picking some fields at random). Or maybe I don't know what you mean by "application". The better question is perhaps: where are ideal 2D rectangular geometries ever found in applications? Most applications don't have the symmetries of the rectangle. $\endgroup$ – Jens May 2 '16 at 5:39

Solving 1D and 2D complex Schroedinger wave equations with NDSolve

I do not agree with you when you write:

I know the NDSolve is not magic...

My opinion is that NDSolve is one of the most complex functionality I've met so far in the Mathematica environment, with its millions of options and special function this is a real complex thing and it is hard indeed to get a proper result. NDSolve is big, you could even sell this as a standalone program.

This post is somewhat long and at one side it is an answer (or an approximation of answer) and on the other side it is opening a huge amount of new questions...

It's a pity that guys like Rob Knapp are not available to ask. So it's only me who tries to give a look under the hood of NDSolve while scratching only at the surface.)

NDSolve includes a general solver for partial differential equations based on the methods of lines. There are several ways to control the selection of the spatial grid (I'll list here only those who are relevant for the 1D case):

AccuracyGoal $\rightarrow$ the number of digits of absolute tolerance

PrecisionGoal $\rightarrow$ the number of digits of relative tolerance

MinStepSize $\rightarrow$ the minimum grid spacing to use

MaxStepSize $\rightarrow$ the maximum grid spacing to use

The discretization is done (not for pseudospectral methods) with uniform grids with the following direct correspondences (with interval length L):

MaxPoints $\rightarrow$ n $\Leftrightarrow$ MaxStepSize $\rightarrow$ L/n

MinPoints $\rightarrow$ n $\Leftrightarrow$ MinStepSize $\rightarrow$ L/n

1) The 1D case:

sol = NDSolve[{I D[u[t, x], t] == -D[u[t, x], {x, 2}], 
    u[0., x] == Exp[-(x^2.)], u[t, 5.] == 0, u[t, -5.] == 0}, 
    u, {t, 0., 20.}, {x, -5., 5.}, MaxStepSize -> 0.01, 
    AccuracyGoal -> 3, PrecisionGoal -> 3]

Animate[Plot[Evaluate[Abs[u[t, x] /. First[sol]]^2], {x, -5, 5}, 
    PlotRange -> {0, 1}], {t, 0, 17, 0.01}]

enter image description here

In order to increase the grid size you have to decrease MaxStepSize. Now your plot will work nicely.

Warming up:

Now let's turn to a 2D heat equation. To get warm we'll solve a SIAM 100 Challenge using NDSolve for that, although an analytical solution would be the best way.

(I only want to show that the solver is quite capable to get an answer)

The SIAM 100 Challenge #8 states the following problem:

A square plate [-1,1] is at temperature u = 0. At time t = 0 the temperature is increased to u = 5 along one of the four sides while being held at u = 0 along the other three sides, and heat then flows into the plate according $u_i = \Delta{u}$. When does the temperture reach u = 1 at the center of the plate? (Folkmar Bornemann)

The initial condition in that problem is discontinuous, so we'll provide a specific grid spacing.

Quiet[Block[{n = 25, he = 5 UnitStep[-(x + 1)]}, hsol = NDSolve[
 D[u[t, x, y], t] == D[u[t, x, y], x, x] + D[u[t, x, y], y, y],
 he == u[0, x, y],
 u[t, -1, y] == 5, u[t, 1, y] == 0, u[t, x, -1] == he, 
 u[t, x, 1] == he
 }, u, {t, 0, 1}, {x, -1, 1}, {y, -1, 1},
 Method -> {"MethodOfLines",
  Method -> {"EventLocator",
    "Event" -> u[t, 0, 0] - 1,
    "EventAction" :> Throw[end = t, "StopIntegration"]},
    "SpatialDiscretization" -> {"TensorProductGrid", 
        "MinPoints" -> {n, n}, "MaxPoints" -> {n, n}}}]]]

which yields an answer quickly:


We've specified the EventLocator controller method. Every time the Event option is zero EventAction is evaluated, which in this case will stop the integration.

The most important suboptions for method option = MethodOfLines are:

"SpatialDiscretization $\rightarrow$ "TensorProductGrid"

"MinPoints" $\rightarrow$ list Of Discreatization points for each spatial variable

"MaxPoints" $\rightarrow$ dito

"DifferenceOrder" $\rightarrow$ positive integer or "PseudoSpectral"

Now let's Plot3D the solution of the heat equation at t = 0.424014, along with a DensityPlot:

    Plot3D[Evaluate[u[0.424014, x, y] /. hsol], {x, -1, 1}, {y, -1, 1}],
    DensityPlot[Evaluate[Abs[u[0.424014, x, y]] /. hsol], {x, -1, 1}, {y, -1, 1}, 
        PlotPoints -> 200, Mesh -> False]

enter image description here

The 1D complex Schroedinger wave equation

We will model the quantum-mechanical scattering of a Gaussian wave packet by using a 1D time-dependent Schroedinger equation.

Let's define the the Schroedinger equation:

The time-dependent Schroedinger equation is defined as:

$i\hbar \frac{d}{dt} \Psi(x,t)=[\frac{-\hbar^2}{2m} \Delta^2+V(x,t)] \psi(x,t)$

As you already wrote in your Note 1, the inital condition is a simple spreading Gaussian probability.

And its potential/kinetic energy is defined as:

$e^{-(x+3)^{2}+3 i x}$

If we integrate this with $\Psi$ we get the maximum value of the potential, which is $\approx 6$

Setting $\hbar$ and $m$ = 1, and the potential to be 6, we get:

schroedingerEq = I D[u[x, t], {t, 1}] ==  -1/2 D[u[x, t], {x, 2}] + 6
    Exp[-x^2] u[x, t]

In order to make sure that at $\pm xMax$ our wave function behaves identically we define a Dirichlet boundary condition which adds a cosine to u:

DirichletBC[u_, x_, xM_] := u - (((u /. x -> -xM) - (u /. x -> xM))/
  2*(Cos[(x + xM)/(2 xM) Pi] + 1) + (u /. x -> xM))

Now let's solve this wave equation numerically:

With[{xMax = 15},(nsol = 
        u[x, 0] == DirichletBC[Exp[-(x + 3)^2], x, xMax] Exp[3 I x], 
        u[xMax, t] == 0, u[-xMax, t] == 0},
        u[x, t], {x, -xMax, xMax}, {t, 0, 5}, 
        AccuracyGoal -> 3, PrecisionGoal -> 3]) // Timing]

Plotting nsol yields:

DensityPlot[Evaluate[Abs[u[x, t]] /. nsol], {x, -15, 15}, {t, 0, 5},
    PlotPoints -> 200, Mesh -> False]

Here we can see that at t $\approx$ 3.2 the wave packet reaches the right boundary and gets reflected there.

enter image description here

The 2D complex Schroedinger wave equation

When we look at the byte size of sol (see the 1D part of this answer) we realize, that the size is about 72 MB large.

In order to avoid such huge InterpolatingFunction objects and since I found out, in our prior discussion, that a low memory consumption is important to you I want to show another way to tackle the problem.

Under the hood NDSolve is broken up actually in several parts:

  1. Equation processing and method selection
  2. Method initialization
  3. Numerical solution
  4. Solution processing

Normally, if you use NDSolve you won't even notice these steps, but there exist low-level functions which you can use on your own, to break up the steps.

  • ProcessEquations
  • Iterate
  • ProcessSolutions

ProcessEquations is used to set up the problem and to create the StateData data structure.

Iterate advances the numerical solution.

ProcessSolutions converts the numerical data into a InterpolationFunction

Please see the documentation for NDSolve steps and components

Let's first set up the problem using a homogenous Dirichlet boundary condition on all four edges, with length L, the initial condition ics0 and the DifferenceOrder option do:

makeNDSolveStateDataObject[ics0_, do_, L_: 5, opts___] := 
    I D[u[t, x, y], t] == -D[u[t, x, y], {x, 2}] - D[u[t, x, y], {y, 2}], 
    u[0., x, y] == ics0,
    u[t, -L, y] == u[t, L, y], u[t, x, -L] == u[t, x, L]},
    u, {t, 0., 2.}, {x, -L, L}, {y, -L, L}, opts, 
    Method -> {"MethodOfLines", 
         "SpatialDiscretization" -> {"TensorProductGrid", 
         "DifferenceOrder" -> do}}]] 

and build the StateData data structure:

stateD = makeNDSolveStateDataObject[Exp[-(x^2 + y^2)], "Pseudospectral"];

stateD ==> NDSolve`StateData[<0.>]

The nice thing about the StateData data structure is that we have more control of the integration. Sometimes it is appropriate to check the solution and change maybe some parameters and to start it over again.

Iterate on its own does not return a value but it modifies the StateData data structure.

We can integrate using Iterate and if we want to integrate further we have to call Iterate again but with a larger value of time.

For the sake of brevity I gonna create now a GraphicsGrid iterating through the StateData structure:

GraphicsGrid[Partition[sols = Table[
NDSolve`Iterate[stateD, t];
  Abs@u[t, x, y] /.
   NDSolve`ProcessSolutions[stateD, "Forward"]],
 {x, -5, 5}, {y, -5, 5}, PlotRange -> {0, 1}],
{t, 0, 2, .1}], 2]]

enter image description here

And here the canonical animation:

enter image description here

This solution may not be the "straightforward" one, but if you are short on ressources (memory) this is the way to go.

According to Weierstrass, the ultimative goal is always the representation of a function.

I guess I fulfilled his dictum now...

I hope this helps.

  • $\begingroup$ This excellent answer is very valuable because the iterative approach is eventually necessary for 2D+1 problems, because memory ends up being a constrain. I accepted the other answer because it answers the posed question more directly. $\endgroup$ – alfC Jul 4 '13 at 20:19
  • $\begingroup$ Can you please give me more information on how/why you constructed the Dirichlet boundary conditions? Was it to ensure that the initial condition was consistent with the boundary conditions? $\endgroup$ – Buddhapus Nov 27 '16 at 1:15

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.