\(\renewcommand{\vec}[1]{{\bf #1}}\)

Literate Programming

Write prose and code intermixed. Not just some choice snippets: all code is included! This document is a rendering of a completely self-contained Markdown file.

About Entangled

Build with Pandoc

Pandoc is the universal document converter. This demo is converted from Markdown to a responsive Bootstrap website using Pandoc.

About Pandoc

Strongly typed Physics

This demonstrates some strongly typed physics, combining Hamiltonian dynamics with the elegant type system of PureScript.

About PureScript

This demo shows an interactive model of the chaotic double pendulum, written in PureScript, using Pandoc and Entangled to create the static web page. The entire demo runs in your web browser, so unlike with Python or R no special server is needed, just host from Github pages! Since PureScript is not the most well known programming language, I will explain some details about the use of PureScript in a bit more detail. You can recognize these by their blue shade.

PureScript is a strongly typed functional language similar to Haskell that compiles to JavaScript. It has a highly formal approach to data types that matches really well with the topic of modelling dynamic systems.

Other than that, I have been able to build this demo without great knowledge of JavaScript.

I understand the danger of explaining two difficult things at once. Non-linear dynamics is a difficult topic and PureScript is a language with high levels of abstraction. I’m hoping you can cherry pick from this demo whatever you’re trying to learn.

Why we care

I’ve heard some rumours that people really like Jupyter notebooks and MyBinder. Rightly so! You can show your work to colleages around the world and they don’t have to setup anything. People can learn from what you have done because the code is right there! MyBinder has some problems though: it only works with languages that have a Jupyter kernel available. While there are kernels available for languages other than Python, Julia and R, outside those two your options are really limited especially in terms of visualisation. Also MyBinder has been getting slower due to large demands on a free service. The computations are being run on a remote server and one way or another, this costs money.

If you’re willing to learn JavaScript, there is the option of writing Observable notebooks. Observable is great, but it requires you to work in JavaScript. Also, to my best knowledge Observable is not (yet) open source, so you’ll still be relying on an external service to work with your code.

Thirdly, you may not always want an interactive web environment to do your coding in. Coding is hard, and we got used to working with code, runtimes and debuggers from editors, consoles and IDEs. Why change all that? There is a solution where you can mix and match languages, write your story in one or more Markdown files, have your cake and eat it too! This setup is still under development, which is in part why I made this demo.

Get to see

  • Experimental science blogging with live interactive demos
  • How to use Literate Programming to learn and teach
  • See a chaotic double pendulum in action

Get to learn

The physics used in this demo assumes some understanding of Lagrangian/Hamiltonian dynamics. We will first model the time evolution of a simple pendulum, and then extend our model to that of a double pendulum.


The pendulum

It was found by Christiaan Huygens that a pendulum can make an excelent clock. We assume a solid massless rod suspending a weight of mass \(m\) at a length \(l\) from the origin. The Lagrangian for such a system (being \(\mathcal{L} = T - U\), where \(T\) is the kinetic energy and \(U\) is the potential energy) is written as

\[\mathcal{L}(\theta, \dot{\theta}) = \frac{1}{2}ml^2\dot{\theta}^2 + mgl \cos\theta.\]

Using the Euler-Lagrange equations, in very terse notation \(D_t \partial_{\dot{\theta}} \mathcal{L} = \partial_{\theta} \mathcal{L}\), leads to

\[\ddot{\theta} = -\frac{g}{l} \sin\theta.\]

This is a second order non-linear differential equation. To solve this numerically we write down the equation in phase-space using position \(\vec{q}(t) = \theta(t)\) and momentum \(\vec{p}(t) = D_{\dot{\theta}} \mathcal{L} = ml^2\dot{\theta}(t)\). This is best done using the Hamiltonian formalism. The Hamiltonian becomes

\[\mathcal{H} = \frac{\vec{p}^2}{2ml^2} - mgl \cos \vec{q}.\]

Having \(\dot{\vec{p}} = - \partial_q \mathcal{H}\) and \(\dot{\vec{q}} = + \partial_p \mathcal{H}\).

\[\begin{pmatrix}\dot{\vec{q}} \\ \dot{\vec{p}}\end{pmatrix} = \begin{pmatrix}\vec{p} / {ml^2} \\ -mgl \sin \vec{q}\end{pmatrix}.\]

In PureScript this system can be expressed as follows:

The first line of this definition is the type declaration of the simplePendulum function: it takes a number (the gravitational accelleration) another number (the length of the pendulum) and returns a HamiltonianSystem Scalar, which is saying a Hamiltonian system of equations in one dimension. These types will be defined when we program the leap-frog integrator.

We are now going to compute the time evolution of the single pendulum using the leap-frog method.

Leap-frog method

We can describe this methods in terms of types in code. Later on, we will deal with systems of more than one dimension. To deal with this I assume the position and momentum vectors are stored in some Apply data structure. By doing this, I can write down the leap-frog integrator in generic terms. For the one-dimensional single pendulum, I have to wrap the Number into an Apply which I call Scalar.

A Functor is a type that can be mapped over. This mapping should hold to the Functor laws. It should preserve the identity function:

and be linear over function composition

An Applicative is a type allows for combining Functors over functions with Functors over arguments. Combined with curried function application this lets us map over functions of multiple arguments. This is explained better than I ever could in the “Learn you a Haskell” book chapter on Functors, Applicatives and Monoids. Haskell and PureScript are similar enough for this reference to apply.

We can describe the State of a system by giving time, position and momentum. Here a is the type-parameter that indicates the container type. For instance, the position member of a State Scalar will have type Scalar Number.

Then, a Hamiltonian system is given by two sets of first-order differential equations.

And in general, a Solver for the Hamiltonian system takes a state object and returns a new state.

The leap-frog method is a solver that uses a kick and drift phase. The kick phase integrates only the momentum part of the system,

while the drift phase only integrates the position part.

Note that, for instance in the drift function, the operators <$> and <*> are our way of applying the lambda-function (\q dq -> q + dt * dq) to the position member of the state and the change in position given by the governing system of equations of motion. The <$> operator takes a function (of two arguments) and an applicative, call it a vector, and creates a vector of functions of one argument. Then the <*> operator applies the vector of functions to another vector of numbers. This way we can apply a function of \(n\) arguments, point-wise to \(n\) vectors, as long as these vectors are an instance of Apply.

We add a wait phase to control the clock, in case we encounter time-dependent systems, and because we move the clock by dt/2 after each kick or drift, which is specific to the leap-frog method.

The leap-frog method follows a kick by a drift, with a separation of \(dt/2\). We choose to start with a kick, then wait, drift, and wait again.

Halting condition


Integrating the single pendulum

We had the following equations of motion,

Integrating this gives the time evolution of pendulum. We can compare it with the small-angle approximation. Here is a plot of the behaviour, you can change the initial angle with the slider.

The source code for this interactive plot can be expanded below.

module Pendulum where

import Prelude

import Effect (Effect)
import Math (pow, sin, sqrt, cos)
import Data.Array ((..))
import Data.Int (toNumber, round)
import Hamilton (HamiltonianSystem, Scalar (..), State, leapFrog, integrateSystem, haltAtTime)
import Plotting as Plotting

import Flare (runFlareWith, numberSlider)


type Model =
        { initAngle :: Number
        , dt :: Number
        , tEnd :: Number
        , length :: Number }

updatePlot :: Number -> Effect Unit
updatePlot iq = do
    let m = initModel { initAngle = iq }
        result = integrateSystem (leapFrog m.dt) (haltAtTime m.tEnd)
                                 (simplePendulum 9.81 m.length)
                                 { time: 0.0, position: Scalar m.initAngle, momentum: Scalar 0.0 }
        small = smallAngle m
    Plotting.restyle "pendulum-plot" "y" [map (\{position: Scalar q} -> q) result] [0]
    Plotting.restyle "pendulum-plot" "y" [small.y] [1]

initModel :: Model
initModel = { initAngle: 3.0, dt: 0.1, tEnd: 10.0, length: 1.0 }

convertResult :: Array (State Scalar) -> Plotting.PlotData
convertResult rs =
    { x: map (\{time} -> time + initModel.dt/2.0) rs
    , y: map (\{position: Scalar q} -> q) rs
    , mode: "lines+markers" }

smallAngle :: Model -> Plotting.PlotData
smallAngle m =
    let period = sqrt $ m.length / 9.81
        ts = (\i -> m.dt * (toNumber i)) <$> 0..(round $ m.tEnd / m.dt)
        ys = (\t -> m.initAngle * (cos $ t / period)) <$> ts
    in { x: ts, y: ys, mode: "lines" }

main :: Effect Unit
main = do
    let result = integrateSystem (leapFrog initModel.dt) (haltAtTime 10.0)
                                 (simplePendulum 9.81 1.0)
                                 { time: 0.0, position: Scalar 3.0, momentum: Scalar 0.0 }
        slider = numberSlider "initial angle" 0.01 3.14 0.01 1.57

    Plotting.lineChart "pendulum-plot" [convertResult result, smallAngle initModel] { title: "pendulum" }
    runFlareWith "pendulum-control" updatePlot slider

The double pendulum

The double pendulum is more complicated than the single one. We now have two angles \(\theta\) and \(\varphi\), two lengths \(k\) and \(l\), and two masses \(m\) and \(n\). The potential energy is

\[U(\theta, \varphi) = - m g k \cos{\theta} - n g (k \cos{\theta} + l \cos{\varphi}),\]

and the kinetic energy (takes a bit of calculating),

\[T(\theta, \varphi, \dot{\theta}, \dot{\varphi}) = \frac{1}{2} m k^2 \dot{\theta}^2 + \frac{1}{2} n \left(\dot{\theta}^2 k^2 + \dot{\varphi}^2 l^2 + 2\dot{\theta}\dot{\varphi}kl \cos(\theta - \varphi)\right).\]

We describe the pendulum in terms of the coordinates \(\theta\) and \(\varphi\). Just as we did for the Scalar type, we need to define how to map over the Coordinates type, meaning we have to create an Applicative instance.

We choose the position vector \(\vec{q} = (\theta, \varphi)\), then the canonical momentum is \(\vec{p} = \partial_{\dot{\vec{q}}} \mathcal{L}\). This gives the terrible momentum vector,

\[\vec{p} = \begin{pmatrix} mk^2\dot{\theta} + nk^2\dot{\theta} + nkl \dot{\varphi} \cos(\theta - \varphi)\\ nl^2\dot{\varphi} + nkl \dot{\theta} \cos(\theta - \varphi) \end{pmatrix}.\]

We need to write the Hamiltonian \(\mathcal{H}\) in terms of \(\vec{q}\) and \(\vec{p}\). This is possible. I stole and rewrote this derivation from Diego Assencio’s blog.

We can rewrite our expression for \(\vec{p}\) in a matrix-vector form,

\[\vec{p} = B \dot{\vec{q}},\]


\[B = \begin{pmatrix} (m + n)k^2 && nkl \cos(\theta - \varphi)\\ nkl \cos(\theta - \varphi) && nl^2 \end{pmatrix}.\]

Then we can write the velocities \(\dot{\vec{q}} = B^{-1} \vec{p}\). The inverse of a \(2\times2\) symmetric matrix is given by

\[B^{-1} = \begin{pmatrix} a && c \\ c && b\end{pmatrix}^{-1} = \frac{1}{ab - c^2} \begin{pmatrix} b && -c \\ -c && a \end{pmatrix}.\]

We will write \[D = \det B = nk^2l^2 \left(m + n \sin^2(\theta - \varphi)\right).\] From this expression for the determinant it is clear that \(D \ge nmk^2l^2\), and that the inverse \(B^{-1}\) exists.

Now it is helpful to write \(T(\theta, \varphi, \dot{\theta}, \dot{\varphi})\) in terms of \(a, b, c\) and \(d\),

\[T = \frac{1}{2}\left(a\dot{\theta}^2 + b\dot{\varphi} + 2c\dot{\theta}\dot{\varphi}\right).\]

Inserting \(\dot{\theta} = (p_{\theta}b - p_{\varphi}c)/D\) and \(\dot{\varphi} = (p_{\varphi}a - p_{\varphi}c)/D\), things magically work out (making me feel I did too much work) and we get,

\[T = \frac{bp_{\theta}^2 + ap_{\varphi}^2 - 2cp_{\theta}p_{\varphi}}{2D},\]

and the Hamiltonian becomes

\[\mathcal{H} = \frac {nl^2 p_{\theta}^2 + (m + n)k^2 p_{\varphi}^2 - 2nkl p_{\theta} p_{\varphi} \cos(\theta - \varphi)} {2nk^2l^2 \left(m + n \sin^2(\theta - \varphi)\right)} - (m + n)gk \cos{\theta} - ngl \cos{\varphi}.\]

Now we get the Hamilton equations of motion, \(\dot{\vec{p}} = - \partial_q \mathcal{H}\) and \(\dot{\vec{q}} = + \partial_p \mathcal{H}\),

\[\begin{align} \dot{\theta} = \partial_{p_{\theta}} \mathcal{H} =& \frac{nl}{D} \left(l p_{\theta} - k\cos(\theta - \varphi) p_{\varphi}\right),\\ \dot{\varphi} = \partial_{p_{\varphi}} \mathcal{H} =& \frac{k}{D} \left((m+n)k p_{\varphi} - nl\cos(\theta - \varphi) p_{\theta}\right),\\ \dot{p}_{\theta} = -\partial_{\theta} \mathcal{H} =&\xi - (m+n)gk \sin \theta,\\ \dot{p}_{\varphi} = -\partial_{\varphi} \mathcal{H} =& -\xi - ngl \sin \varphi, \end{align}\]


\[\xi = \frac{nkl}{2D}\left(T\ nkl\sin(2(\theta - \varphi)) - 2\sin(\theta - \varphi) p_{\theta} p_{\varphi}\right).\]

Terrible indeed. It is amazing how quickly the addition of another pendulum to the system just explodes in your face. Since we have a HamiltonianSystem defined for the double pendulum, integration of this system is the same as integrating the single pendulum. The next section describes how I made the animation at the top of this page.

Animation using Flare

We have an animation canvas and a button that controls if the animation plays or not. The Model is as follows:

The model is updated by passing messages to an update function.

Updating the model:

And we use the following initial state:


Using Flare we can create a set of UI controls that generate a stream of messages. The events function takes a Signal Number argument that generates a signal for every animation frame.

In the main function we can obtain the animation signal, and pass it to events. Using Flare.foldp the series of Msg messages is converted into a series of Model objects. These are then rendered to a canvas in the draw function.