# Symbolic Evaluation and Equational Reasoning

In the last chapter, we discussed the composition mechanics of machines in our new system of symbolic computation. The claim is that symbolic computation is actually a means of computation, and today we’re going to look at how that can be so. We’ll discuss a technique called equational reasoning and use it to – you guessed it – reason about all of this mumbo jumbo.

We’ve now introduced our notation for machine implementations a few times, but we will re-examine our nandn example from last chapter.

nand : Bool -> Bool -> Bool
nandn a b = not (and (not a) b)

Let’s say we’d like to evaluate (determine) the output of nandn with inputs On and Off. Recall, that we can write this output as nandn On Off. Although we are currently unaware of the value of nandn On Off, we know that it will indeed have some value, because our function tables must be total (they must be defined for all possible inputs).

We call such a thing – something which has a value, even one that we haven’t yet evaluated – an expression. Expressions are everywhere, and they are the bread and butter of our symbolic computation. On is an expression, because it has a value (which is just itself.) not On is another expression, because we can run On as an input to the not gate, and retrieve a value as its output.

However, and On is not an expression, because it is missing an input. and gates always have two inputs, but and On is only supplying one input, and so it cannot have a value, and it is therefore not an expression.

Expressions have one primary desire in life, and that is to be evaluated – that is, to determine what value they eventually end up as. Evaluating an expression is easy – so easy that even a stupid electronic circuit can do it. Let’s continue with our nandn On Off example.

The first step to evaluating an expression is to have all of the function tables it’s made out of handy. If you don’t have all of the function tables, you’re eventually going to get stuck, since a function table is necessary to reduce a machine’s expression to its final value.

Once you have all of the machine implementations in front of you, it’s time to get to work. Replace any machines in your expression with their definition, remembering to substitute the bindings in the definition with the ones you started with. We can use our definition of nandn to partially evaluate nandn On Off as follows:

not (and (not On) Off)

Don’t forget to replace bindings! Here, because On was the first input to nandn, we replaced all of the bindings a in the definition of nandn with On, and all of the bindings of b with Off. If we had had nandn Off Off instead, we would have replaced a with Off.

Make sure you replace all of your bindings before proceeding. If you don’t get all of them, you’ll probably get confused somewhere down the line when you expand another machine that uses some of the same letters for binding. Remember, just like in our machine diagrams, the same letter means different things when used in different machines. This is the most critical part of the whole technique, and it bears repeating. Many a beginner have lost their way at this step.

Now that you’ve successfully replaced all of your bindings with the correct inputs, you’re ready to rinse and repeat! But all of a sudden we’ve got this expression which is actually longer than what we started with! So much for reducing it!

This is the way that things always are. Evaluating an expression always balloons up in complexity before it settles down into something more manageable. This seems like it might be annoying, but it turns out to be one of the most important parts. It’s great that the thing we start with is small and simple, because that allows us to perform abstraction, and to keep more things in our heads simultaneously. It’s good at the end result is simple, because it has to terminate in a single value eventually. And it’s good that it balloons up, because that means our abstractions are doing lots of work for us (whenever we’re not explicitly performing evaluations by hand).

This last part is worth mentioning again – we usually don’t need to evaluate by hand. Most of the time it’s clear how a machine will behave, and so evaluation isn’t necessary. But it’s a nice technique to have under your belt, just in case you find yourself in a pickle one day, a pickle that only symbolic evaluation can get you out of. It could happen. You never know.

The good news for our interpreting selves is that this process of expanding out a machine’s definition and replacing the bindings will always result in an expression with at least one machine whose inputs are all raw values, and that means we have a place to proceed the simplification process.

However, the other good news is that I’m not your dad – you can pick any machine you’d like to expand out next. The most natural order is probably inside-out (starting from the most parenthetical machine, and working your way backwards), but you can do whatever you’d like. We’ll proceed with the expression’s evaluation here.

not (and Off Off)
not Off
On

And there you have it! A value comes forth from whence there was none originally.

## Equational Reasoning

Ready for something completely wild? Because of a thing called equational reasoning, we can actually work backwards, and build expressions that evaluate to a particular value. As long as we replace something in our expression with an expression that has the same output as what we replaced, we are guaranteed to never go wrong.

Let’s try an example:

Off = Off
Off = xor Off Off
Off = xor (and On Off) Off
Off = xor (and On Off) (not On)

This isn’t a particularly useful thing in-and-of-itself, but the point of the exercise is to demonstrate that we don’t lose anything going one way or the other. This is why we use an equals sign when defining the outputs of our machines – it’s because these things really are equal, and there’s absolutely no test “inside the system” (which is to say, not “looking at the paper you wrote down”) to distinguish xor On Off from On. It just can’t be done, and that’s a nice property to have.

Why? Because it means we are free to abstract at will, and no-one will be any the wiser. It means that at the end of the day, a smart solution to a problem works just as well as a stupid one, and that as long as we’re following all the rules, we’re safe to do whatever we please – content in the knowledge that we’re not “losing anything” when we work in more convenient settings.

Takeaway: Equational reasoning means that it’s safe to work at whatever level of abstraction we feel most comfortable in. It’s a promise that nothing can “go wrong” when we move between levels of abstraction, so long as our individual steps are all honest.

Equational reasoning fills another role in our toolkit, however, and that’s for proving algebraic identities. An algebraic identity is two expressions which are always equal. What’s that mean?

Well, suppose I go and implement some convoluted machine that computes a particular function, but evaluating it takes 100 steps. You, being the more clever of my friends, come up with a machine you swear to me is the same – in the sense that it matches outputs with mine, input for input. If my machine takes 20 inputs, that’s not a thing I want to try to verify by hand. Instead, we could work together to try and derive an algebraic identity between our machines, a proof that they are one-and-the-same.

This is why being able to work forwards and backwards is important for our reasoning when doing symbolic evaluation. Just like doing a pen-and-paper maze, it can be helpful to work forwards and backwards at the same time – attempting to meet in the middle.

As an example, let’s work through a proof of an algebraic identity for de Morgan’s laws.

and a b = not (or (not a) (not b))

We know that there are two cases for a, either On or Off, so we can try both:

and On b  = not (or (not On) (not b))

and Off b = not (or (not Off) (not b))

Remember, that whenever we replace a binding, we need to change all instances of that binding to the same expression. We can continue on, by simplifying our nots now:

and On b  = not (or Off (not b))

and Off b = not (or On (not b))

We know that anything anded with On keeps its value, and anything anded with Off has value Off, so we can evaluate our left-hand-sides:

b = not (or Off (not b))

Off = not (or On (not b))

Likewise, we know that an or against On is always On, and that an or against Off keeps its value, so we can go on:

b = not (not b)

Off = not On

and we’re left with very simple things to prove the equivalency of. Dead simple when you get down to it.

We’ll revisit equational reasoning again when we build stronger abstractions which require “laws” – inviolable algebraic identities for defining machines. Well chosen laws allow us the luxury of building loosely-defined machines “molds”, particular machine “shapes” which we can reason about for any machine definition respecting the laws. This probably doesn’t sound too powerful to you yet, but as we’ll see, it’s truly one of the most powerful forces in the logical universe.

In the next chapter, we’ll investigate moving more of our old machine diagrams to symbolic computation, and solve some of the remaining challenges that come up.

## Exercises

1. Fully evaluate (walk through all the steps of it) the expression or (xor (and On Off) Off) (or Off (not Off)).