2 title: "Syntactic Unification Is Easy!"
3 date: 2020-02-19 Europe/Paris
4 updated: 2020-11-26 15:36:12 America/Detroit
9 summary: "Last year I tried implementing syntactic unification and got completely stuck on the algorithms. Last week I read the MicroKanren paper and discovered that unification could be very simple. Let me show you how!"
10 next: posts/2020/logic-programming
13 Syntactic unification is the process of trying to find an assignment of variables that makes two terms the same.
14 For example, attempting to unify the terms in the equation `(X, Y) = (1, 2)` would give the substitution `{X: 1, Y: 2}`, but if we tried to unify `(X, X) = (1, 2)` it would fail since one variable can't be equal to two different values.
15 Unification is used to implement two things I'm very interested in: type systems, and logic programming languages.
17 Earlier last year I tried implementing unification in order to make a toy Prolog, and I got completely stuck on the algorithms.
18 I eventually puzzled throught it by reading lots of papers and articles, but I don't recall understanding it fully.
19 You can take a look at [the equations that the Wikipedia article uses][wiki] to see what I was dealing with---it's certainly intimidating!
21 Cut to last week, I happened to read [the MicroKanren paper] (which is very readable) and discovered that unification could be very simple!
24 [wiki]: https://en.wikipedia.org/wiki/Unification_(computer_science)#A_unification_algorithm
25 [the microkanren paper]: http://webyrd.net/scheme-2013/papers/HemannMuKanren2013.pdf
27 ### The Terms To Unify
29 We have to define what kinds of things we're going to unify.
30 There are a few basic things we *need* to have.
32 - Variables, because otherwise unification is just equality comparison.
33 - Terms with multiple subterms, because otherwise unification is just.
34 - Other atomic "base" terms are good to have, because they let you write more interesting terms that are easily understandable.
36 I'm going to write my examples in Python, so the types I'm going to work with are:
38 - Variables: A `Var` class defined for this purpose.
39 - Composite terms: Tuples of terms.
40 - Atomic terms: Any Python type that can be compared for equality.
42 The scheme example they demonstrate in the MicroKanren article uses variables, pairs, and arbitrary objects.
43 The Rust version I wrote just has variables, pairs, and integers.
44 Exactly what you want to work with is completely up to you: hopefully at the end of this, you'll be able to figure out how to do it for some custom type.
46 My `Var` type is very simple, with only one custom method to make the printing more compact:
49 from dataclasses import dataclass
51 @dataclass(frozen=True)
59 And everything else is base Python types, so we're done with this part.
63 Let's work through unifying things.
64 We want to unify two terms, and return a substitution that makes them equal, which I'll call the "environment."
65 If they can't possibly be equal, we'll return `None`.
66 Let's start out simple: two things can be unified if they're already equal.
76 Here we return an empty environment because they're already equal, and nothing needs to be substituted.
78 The variable cases can also be solved very easily.
79 Trying to unify a variable with another term, we can just substitute that for the variable.
80 When unifying `x = 1`, we get back `{ x: 1 }`.
83 if isinstance(a, Var):
85 if isinstance(b, Var):
89 Finally, let's handle tuples.
90 For this, we want to unify each individual component of the tuple.
91 We want the union of each of these unifications, so we'll update the dictionary each time.
92 We have to check if the sub-unification returned None and pass that along.
95 if isinstance(a, tuple) and isinstance(b, tuple):
99 for (ax, bx) in zip(a, b):
107 Now you might see where this could go wrong, and we'll get to that in a moment with an example.
108 Repeating that function all in one piece:
114 if isinstance(a, Var):
116 if isinstance(b, Var):
118 if isinstance(a, tuple) and isinstance(b, tuple):
122 for (ax, bx) in zip(a, b):
131 Now we can unify stuff:
134 >>> x, y = Var("x"), Var("y")
139 >>> unify((x, 2), (1, 2))
141 >>> unify((x, y), (1, 2))
143 >>> unify((x, x), (1, 2))
148 That last one is wrong!
149 So what can we do to fix that?
151 Well, we need to take into account context.
152 Let's change our `unify` function to accept an environment to work inside.
153 I'm gonna make this a pure function, so we don't update the environment, we just make a new version of it.
154 Some things are changed a bit when doing this.
155 We include `env` in all of our outputs, instead of starting from nothing.
159 def unify(a, b, env={}):
162 if isinstance(a, Var):
164 if isinstance(b, Var):
166 if isinstance(a, tuple) and isinstance(b, tuple):
169 for (ax, bx) in zip(a, b):
170 env = unify(a, b, env)
177 This still doesn't do the right thing though, because we're not actually looking in the environment.
178 We need to look stuff up, so let's write a function to do that.
179 We'll just always call it, and if the term isn't a key for the environment, we'll just return it unchanged.
188 Now we can add these two lines at the beginning.
195 With these modifications, we correctly handle the case we were stuck on earlier.
198 >>> unify((x, x), (1, 2))
199 >>> # look ma no solution!
202 It's still not perfect though.
203 It doesn't properly reject this case where the conflict is two steps away:
206 >>> unify((x, y, x), (y, 8, 9))
210 Here the problem is that we have two facts around, that `x = y`, and that `y = 8`.
211 Even so, it doesn't see that that means `x = 8` when it gets to the assertion of `x = 9`.
212 Luckily this is another simple fix.
213 We just keep walking through the environment as long as we have a binding.
214 Just change the `if` to a `while`!
223 And now this unifier should be able to handle any [(finite)](#postscript-unifying-infinity) hard cases you throw at it!
225 ### The Whole Unifier
227 Here's the whole program we wrote in one piece, in just 30 lines.
230 from dataclasses import dataclass
232 @dataclass(frozen=True)
243 def unify(a, b, env={}):
248 if isinstance(a, Var):
249 return { **env, a: b }
250 if isinstance(b, Var):
251 return { **env, b: a }
252 if isinstance(a, tuple) and isinstance(b, tuple):
255 for (a, b) in zip(a, b):
256 env = unify(a, b, env)
263 Consider writing your own in your language of choice, or adding support for new datatypes.
264 It would be nice to be able to unify namedtuples and dictionaries and things, or do unification in the web browser.
265 Try it out for yourself, because next time I'll be talking about how you can go from here and use unification to build logic programs.
267 ### Postscript: Unifying Infinity
269 What was that earlier?
270 About only the finite cases?
271 Well, it turns out this unifier is good enough for implementing well-behaved logic programs, but it leaves out one complexity: the "Occurs Check."
272 This checks that a variable does not "occur" inside a term you unify it with, to avoid infinite solutions and non-termination.
273 Many old Prolog implementations left out the occurs check for efficiency reasons, and we can leave it out for simplicity reasons if we want to trust our programmers to only write finite terms.
275 But let's go look at the problem specifically.
276 We can write a unification like:
283 If you think through this, you can see that `a = (1, (1, (1, (1, ...))))`, going on infinitely deep.
284 There are languages that work with logic on infinite terms like this, but applied naively it can end up giving us incorrect results or non-termination.
285 For instance, with proper support for infinite terms, the following should unify:
288 >>> env = unify(a, (1, a))
289 >>> env = unify(b, (1, b), env)
290 >>> env = unify(a, b)
293 But instead, it gets stuck on the last step, descending forever.
294 There are two ways to solve this.
295 One is to embrace infinite terms and include fixpoints in your logic, giving you terms like `a = fix x. (1, x)` and `b = fix x. (1, x)`.
296 I don't know how to do this, though.
298 The other way is to detect unifications that would lead to infinite terms, and reject them.
299 We take our code that adds variables to the substitution, and add the occurs check there:
302 if isinstance(a, Var):
303 if occurs(a, b, env): return None
304 return { **env, a: b }
305 if isinstance(b, Var):
306 if occurs(b, a, env): return None
307 return { **env, b: a }
310 If we're using the occurs-check here, then we know a few things:
312 - `a` is a variable that's already been `walk()`-ed.
313 - `a` and `b` aren't equal to each other, because earlier in the unify checked that.
315 This means that we can check if the literal variable `a` occurs literally in walked terms and subterms of `b`.
316 If there's a variable in `b` that's a synonym of `a`, then it will walk to `a` because it was the result of walking.
317 While it's safe to assume to that `a` has already been walked, note that we can't assume that `b` has already been walked since we're recursing over it.
318 Then, the implementation looks like this:
321 def occurs(a, b, env={}):
325 if isinstance(b, tuple):
326 return any(occurs(a, x) for x in b)
337 We correctly don't unify this.
338 We've got sound unification over finite terms.
339 This is useful to clean the situtation up, but I'm still a bit sad to see infinite terms go...
343 *Edit November 26, 2020.*
344 Thanks to Onel Harrison for giving some clarity feedback and pointing out some typos.