Previous Contents Next

Term Rewriting

Chapter 3: Term Rewriting

We are all by now hopefully used to the idea of modelling systems algebraically, in terms of (i) a signature representing in some sense the interface of some system, and (ii) an algebra, representing the interpretation of the signature. We still need some practice in actually doing it, but that can come later. In general, we are quite happy to accept the idea that algebras are defined in terms of sets of equations. In this chapter, we consider some of the issues that arise when we respresent algebras using equations, and then use software tools to apply these equations.

Formally, if we represent systems using equations, we are (probably) using Equational Logic - which is difficult and counter-intuitive for most people to apply, and problematic in software as well. Fortunately, there is an alternative that we can use called term rewriting - which is also the effectively the 'natural' way most people apply equations.

3.1. The Basic Problems: an Illustrative Example

Consider the Booleans with the following signature:


signature Boolean is
        sort Bool
        ops true, false : Bool
        op not : Bool -> Bool
        op or : Bool Bool -> Bool
end

Note that as before this is a 'Maude-like' notation but it is not Maude! Notice here also that we have made it explictly clear that this is a signature (because there are no equations) instead of an algebra. A typical set of equations for this might be:


vars X, Y, Z : Bool

not(not(X)) = X
not(true) = false
not(false) = true

X or X = X
X or Y = Y or X
X or true = true
X or false = X
(X or Y) or Z = X or (Y or Z)

Given an expression like (not(true) or false) or (true or not(false)), how would we typically go about evaluating it by hand? We would search within the expression for patterns that match one side of the equations above, and we would then use that equation to modify the expression (typically by simplifying it). For example in the above case we might do:


(not(true) or false) -> not(true), because X or false = X
not(true) -> false

Equally, we could have done:


(not(true) or false) -> false or false, because not(true) = false
false or false -> false

Now suppose we look at the other half of the expression:


(true or not(false)) -> (not(false) or true), because X or Y = Y or X
(not(false) or true) -> true, because X or true = true

The mechanisation of expression evaluation based on this natural (for humans) method is what we call term rewriting - since we find and replace, or rewrite, terms using rules. However, there are some critical issues that lead to problems if we naively treat equations as rewrite rules.

3.1.1. Termination

In the second example above, where we rewrote true or not(false) we used the equation X or Y = Y or X to get not(false) or true. What would stop a machine, without intelligence, simply applying the same equation again to return to true or not(false), resulting in a non-terminating sequence of rewriting. The answer is nothing - if we are careless enough to leave in equations like X or Y = Y or X. In the example above, the equation (X or Y) or Z = X or (Y or Z) will cause the same [potential] problem.

In general, we wish to avoid term rewriting systems that have this problem - we want our systems to be terminating. Note there is a formal definition of termination - that is not especially relevant to this module.

3.1.2. Confluence

In the first example above, another [potential] problem occurs. There are two rules we can initially apply, leading to different intermediate terms. (The same is actually true of the second example: which other rule could we have applied?) In this case there is no problem: the second rule we apply in both cases leads to the same [final] result. However, consider the following set of rules:


A = B
AB = AAB

In this example, we have a number of options when considering the sequence AAAB:


AAAB -> AABB -> ABBB -> BBBB
AAAB -> AAAAB -> AAAAAB -> AAAAAB -> ...

Not only do these two sequences of rule applications not lead to the same result, the second one does not terminate. We say that systems that have the [desirable] property of converging on the same result, regardless of the choice of rules we apply, are confluent or convergent. As with termination, there is a much more formal definition - that is not particularly interesting in this module. Sometimes the term Church-Rosser (after the mathematician Alonzo Church, and his graduate student Barkeley Rosser) is used instead of confluence/convergence - strictly, this is something slightly different but the term is widely used and you may come across it. Commonly, only one sequence of rewrites is in some sense 'correct'. Again this is only a problem when we use an unintelligent machine: a human would simply avoid the incorrect rewriting sequence. (Of course in a complicated case, it may well take them some time to find the correct path!)

3.2. Implications: Equations, Rewrite Rules and some Terminology

There is a well-developed theory underlying rewriting. For the most part, we will not need to go into this in great detail, other than to note the relationship with Equational Logic, and that a range of useful introductory papers can be found here. However, a few aspects are interesting and important.

In general, and conceptually, the process of term rewriting continues until no further rules can be applied (or for ever, in the case of a non-terminating system). When we reach a point where we can apply no more rewrite rules, we have reached a normal form. If there is only one unique normal form, then our system of rewrite rules is confluent (I am bending the truth quite a bit here). In general, determining if a rewrite system is terminating and/or confluent is undecidable: fortunately, in the specific cases we are [often] interested in in practice, there are some useful rules to help us decide.

We have implied above that when we start to use software to evaluate expressions, there is some [subtle] difference between an equation and a rewrite rule. This is the case: an equation


X = Y

is a relation and is consequentially 'non-directional': there are no restrictions (other than the usual mathematical ones) on X or Y, and the equation X=Y is either true or false.

In contrast, a rewrite rule


X -> Y

is a 'directional' action: X can be replaced by Y, but not the other way around. Furthermore, there are some restrictions on X and Y:

Intuitively, these rules are to ensure that expressions 'become simpler' during the rewriting process.

3.2.1. The Effect of the Difference Between Equations and Rewrite Rules

It may have occured to you by this stage that the difference between equations and rewrite rules may have a significant impact on the implementation of systems: it is clear that we cannot just take a set of [arbitrary] equations and turn them into rewrite rules by replacing the equals sign with an arrow, expecting them to 'work'. First and most obviously, associative and commutative equations must be removed, which will obviously have an effect. Secondly, the 'one-way' nature of rewrite rules will have an effect. Both of these combine together to mean that when a set of equations is turned into a set of rewrite rules, some sequences of rewrites that were possible for an intelligent human using equations are no longer available to a machine using rewrite rules.

3.2.2. Completion Algorithms: Knuth-Bendix

The second problem - that of sets of rewrite rules not being as capable in some sense as the original equations - can be addressed by adding in further rewrite rules. These must obviously be chosen carefully, and this is typically done by applying completion algorithms of which the best known is the Knuth-Bendix algorithm.

The Knuth-Bendix algorithm actually does two things: it converts equations into rewrite rules by, and it adds any extra rules that are needed. Converting equations into rules is not as completely straightforward as it might seem - which way should the rewriting go? In some cases (e.g. not(not(a))=a) it is obvious, but in others it is not. The usual way to resolve this is to use a termination ordering - some measure of the complexity of a term. An example would be: the number of function symbols. For example, not(not(a)) has two, and x has none. So x is 'less than' not(not(a)). Once we have converted all the existing equations to rules, the algorithm works out what extra rules to add. This relies on a property called local confluence, which is illustrated in fig 1.


Fig.1. Local and Global Confluence

We can prove (but we won't) that if a term rewriting system is locally confluent, it is also globally confluent. Therefore if we have a system in which two rules A and B can be applied at a particular point, then for the system to be globally confluent it must converge later. So if it doesn't, we can simply make it do so - by adding a new rule in. The only other complication (minus all the fairly-hideous implementation details) is that whenever we add a new rule we must check to see if it is a more general case of one we have already added. If so, we must take the old rule out.

The Knuth-Bendix algorithm has some inconvenient properties (like no guarantee of termination), and has been implemented in Maude. (It would also make a nice third year project - not necessarily in Maude - if anyone is interested.)

3.2.3. Associativity and Commutativity

The problem of associative and commutative operators is generally solved by 'cheating' - that is, building in special mechanisms for dealing with them, and allowing operators to be marked with the respective properites. This is the approach that Maude takes. The usual terminology is to refer to rewriting in the absence of any commutative or associative operators as rewriting in the free theory. The addition of both associativity and commutativity is generally called AC rewriting. Incidentally, if you can get away without any AC rewriting, you will usually find things go quicker as it has a significant impact on efficiency.

3.3. In Practice...

How does this generally affect us in practice? If we are careful, not much. Typically, the sets of equations we come up with will (provided we are careful about commutative and associative operators) form a terminating and confluent set of term rewriting rules without any modifications. This in my opinion is because although we are intellectually aware we are dealing with sets of equations, our mind set is that of the programmer.

Maude recognises this implicitly by providing two forms of module: fmod modules ('functional modules') where rewrite rules appear as equations, and systems must be terminating and confluent; and mod modules ('term rewriting modules'), where a -> symbol is used to denote rules, and systems need not be terminating and confluent. Needless to say, we are interested in fmod modules.

However, it is still possible that we may end up with non-terminating, non-confluent systems:

If you wish, you can [attempt to] check if your systems are terminating by using a Maude tool.

Previous Contents Next