integrate

    0.0.1 • Public • Published

    Integrate

    Some numerical integrators for ordinary differential equations:

    Want one that's not there? Open an issue, or better yet, a pull request adding it.

    Fast

    All of the methods are hand-written in asm.js, so they should be fast. Not that your integration method would ever be your bottleneck anyway, but hey... why not.

    Example

    There are two steps: constructing your integrator, and then calling it.

    var Integrate = require('integrate');
     
    function myODE(t, y) {
      // ordinary differential function to integrate.
      return y;  // for an initial value y=0, this function is also known as 'e^x'
    }
     
    // step 1: get the integrators for your function
    var integrate = Integrate.Integrator(myODE);
     
    var y = 1,
        t = 0,
        step = 0.25;
     
    while (true) {
      // step 2: integrate!
      console.log('t = ' + t + '   \t', y);
      if ((+= step) && t > 2) break;
      y = integrate.euler(y, t, step);
    }
     

    Should log this:

    ```
    t = 0        1
    t = 0.25     1.25
    t = 0.5      1.5625
    t = 0.75     1.953125
    t = 1        2.44140625
    t = 1.25     3.0517578125
    t = 1.5      3.814697265625
    t = 1.7      4.76837158203125
    t = 2        5.9604644775390625
    ```
    

    How bad is euler?

    Well for starters, that 5.960... at the end of the example logs should actually be 7.389....

    How far do we have to shrink our step size to get within 0.0001 of the exact solution? How much better is Runge-Kutta? The fourth-order Runge-Kutta integrator will evaulate your ODE four times for each step, so it has to converge at least 4x faster to be worthwhile...

    Here's a quick and dirty test, computing how far we have to shrink the step size for the Euler method and for fourth-order Runge-Kutta, to get within 0.0001 of the exact solution for t = 2.

    Our ODE, y' = y at y(0) = 1 is actually just e^x, so we should be converging to e^2.

    var Integrate = require('./index');
     
    var integrate = Integrate.Integrator(function(_, y) { return y; });
     
    var stopAt = 2,
        stopError = 0.0001;  // we are done when our error is <= this
     
     
    function acceptableStep(f, stopAt, stopError) {
      var target = Math.pow(Math.E, stopAt),
          step = stopAt;
     
      while (true) {
        for (var t=0, y=1; t < stopAt; t+= step)
          y = f(y, t, step);
        if (Math.abs(target - y) <= stopError) break;
        step /= 2;
      }
     
      return step;
    }
     
    var acceptableEuler = acceptableStep(integrate.euler, stopAt, stopError);
    var acceptableRk4 = acceptableStep(integrate.rk4, stopAt, stopError);
    console.log('euler step with error < '+stopError+' at '+stopAt+'', acceptableEuler);
    console.log('rk4 step with error < '+stopError+' at t='+stopAt+'', acceptableRk4);

    Should log

    ```
    euler step with error < 0.0001 at t=2:  0.00000762939453125
    rk4 step with error < 0.0001 at t=2:  0.125
    ```
    

    So, for a cost of 4x more evaluations per step, we get to run with a step size about 16,000x bigger with Runge-Kutta than with the Euler Method for similar accuracy. After our 4x evaluations per step penalty, we are still winning by about 4,000x the number of evaluations required in this example.

    So, use rk4.

    API

    Create an integrator from an ODE function

    The nice way:

    var integrator = Integrate.Integrate(myODEFunction);

    Shave off one wrapping function call:

    var integrator = Integrate.ASMIntegrators(null, {f: myODEFunction});

    Both forms will return an identical object.

    myODEFunction should accept two parameters: t, and y.

    Integrate with one of the numerical integration methods

    var yNext = integrator.euler(yLast, tNow, tStepSize);
    
    var yNext = integrator.rk4(yLast, tNow, tStepSize);
    
    var yNext = integrator.rk4general(yLast, tNow, 3);
    

    Install

    npm i integrate

    DownloadsWeekly Downloads

    9

    Version

    0.0.1

    License

    Public Domain

    Last publish

    Collaborators

    • uniphil