WIP - Introducing Functional Reactive Programming



FRP has been around for a long time. Originally discovered and coined by Conal Elliot and Paul Hudak in the late 90's it represented a new approach to writing dynamic systems, wherein the programmer describes how the system changes over time using continuous semantics. What this meant practically is that FRP code tended to look more like interelated equations and less like a list of steps to achieve a goal.

As time went on more implementations were made and the term got thrown around freely, eventually even being applied to other technologies like Javascript's React framework. These days when folks talk about React, bacon.js, etc. they tend to drop the "functional" part of FRP in favor of "reactive", which suits the nature of those frameworks better. Check out the wikipedia article on FRP for more info on what it is, specifically

Prep Work

An FRP system deals in terms of continuous values and discrete events. Continuous values are often referred to as Behaviors while discrete events are simply ... you guessed it ... Events! What this means is that your values always come wrapped in these contexts. For this reason I've been known to call writing FRP code as "advanced burrito making". The newly popular (as of the first writing of this article) reflex package adds another context - one that combines Behaviors and Events, called Dynamics. Reflex's brand of FRP is "very advanced burrito making" ;)

In any implementation there are lots of functions for turning a value in one context into a value in another context ... so if you are going to write in FRP I would recommend being really, really familiar with the basic typeclasses:

and if you've got the time it would really help to get these bad boys in as well -

The haskell wiki has more info on how FRP libs use these typeclasses.

Then once you have a grasp on the supporting machinery we can talk a bit about Behaviors and Events.


A behavior is a value that changes over some domain. In most reactive settings like games and simulations a behavior is likely to be a value that changes over time, but it could just as easily be a value that changes over user input, database notifications or any other input.

I find it beneficial to think of a Behavior as a function - a mathematical function.

Let's use Newton's law of gravitation, f = g * (m1 * m2) / r ** 2, where f is the force between two bodies of mass m1 and m2 respectively, g is a gravitational constant, which is just a number, and r is the distance between the centers of the two bodies. In a Behavioral context we can express this equation quite simply and quite literally by saying that if each term (g, m1, m2 and r) are Behaviors themselves, then f is also a Behavior. Even better, if the FRP lib you choose defines Num and Fractional instances for its Behavior type then we can define f as the equation itself:

f g m1 m2 r = g * (m1 * m2) / r ** 2

In this case g is a constant, and due to Fractional's fromRational we should be able to construct a Behavior for g simply using a float literal:

f m1 m2 r = 6.674e−11 * (m1 * m2) / r ** 2

So let's say the bodies we're finding the force between are planets. Then let's exclaim "we have a type for that!":

data Planet = Planet { planetMass     :: Float
                     , planetPosition :: V2 Float -- this is just a vector like (Float, Float), see the linear package

planetA = Planet 10 $ V2 0 0
planetB = Planet 100 $ V2 250 250

POOF! Planetary formation. Now we need to know how these planets change over time. Oooor we can assume they don't. Either way we need a planet in a Behavioral context. Behaviors are often (always?) Applicatives. They have a first order kind (or more) like * -> *. In plain terms the Behavior type takes another type and wraps it in its context. This is Functor and Applicative stuff. So a planet in a Behavioral context would be something like Behavior Planet. Up until now I've left off the type variable in Behavior, but it should really be Behavior a, where a is any type.

Now, since many (all?) FRP implementations provide an Applicative instance for their Behavior a type, we can construct some Behavior Planets:

planetBehaviorA = pure planetA
planetBehaviorB = pure planetB

These planet's don't change over time. These rocks just sit. Which reminds me of a poem:

Nobody sits like this rock sits. You rock, rock. The rock just sits - and is. You show us how to just sit here And that's what we need. -- Albert Markovski

Now that we have what we need (a couple of sitty rocks) we can write the other Behaviors we need, using Functor to great lengths ... and masses.

posBehaviorA  = fmap planetPos planetBehaviorA
posBehaviorB  = fmap planetPos planetBehaviorB
massBehaviorA = fmap planetMass planetBehaviorA
massBehaviorB = fmap planetMass planetBehaviorB

Great! Now, we have Behavior (V2 Float)s for the planet's positions over time, but what we need is the vector between them. We're going to assume that V2 has a Num instance (and it does, if we're talking about this V2). Assuming this means we can treat it just like any other number because remember Behavior a also has a Num instance, at least for as that have a Num instance (in this case V2) ... so:

vectorAB = posBehaviorA - posBehaviorB

And that's enough to get us the rest of the way:

gForceAB = f massBehaviorA massBehaviorB vectorAB


Events are values at a specific domain input. Again, in most cases the domain is time, but it doesn't have to be.

For me it's easiest to think about operating over time because it's quite close to my human experience.

Events in FRP are just like your intuition about the word "event". Each event is a thing that happens at an exact moment. Unlike behaviors, which describe how things change - events describe how things are right now. Or maybe in three seconds. Or possibly in a couple months.

Events & Behaviors

Events and behaviors are bound to each other by a couple of novel ideas. The first is that given an input like a time (or whatever the domain is) we can "sample" a behavior and get a result value. In math this is just evaluating a function with an input. Combining the input value with the result value gives us an event!

makeEvent :: input -> output -> (input, output)
makeEvent domain range = (domain, range)

Using both events and behaviors we can define complicated systems and relationships succinctly.

A tour of FRP flavors

Evan Czaplicki of Elm fame did a great talk on the different flavors of FRP:


Hopefully that gives a little background to the situation. FRP has come a long way since 1997 but it's really still an area of active research.

advanced burrito making