[15 December 2009]
[This is another one of those ignorance-is-bliss postings. If I had studied automata theory properly, this would (I guess) have been covered in class; that would have deprived me of the fun of thinking about it without knowing the right answer. If you did study automata theory, and you know how infinite strings are handled, and it irritates you to see someone waffling on and on about it instead of just doing some homework and reading the damn literature, you might want to stop reading soon.]
Some time ago, Michael Kay suggested that it was pointless for the XSD datatypes spec to specify that the lexical representations of integers, or strings, or various other simple types, were finite sequences of characters with certain properties. No implementation, he pointed out, can reliably distinguish finite from infinite strings, so it’s a non-testable assertion.
[“True if you’re using conventional I/O and conventional representations of strings, maybe,” said Enrique. “But if you represent the sequence of characters using a description, rather than an array of characters, it’s not clear that that’s true. Instead of the sequence
"3.141592...", store an algorithm for calculating, on demand, the nth digit of the decimal expansion of π. Ditto for the square root of 2. And so on!” “You may be right,” I said. “But that wasn’t what I wanted to talk about, so be quiet.”]
The working group declined the proposal to drop the word “finite” on the grounds that if the strings in question are required to be finite, then we know that all the lexical representations of integers (for example) can in principle be recognized by a finite state automaton. Without the restriction to finite sequences, most of what people know about finite state automata isn’t guaranteed to apply.
I found myself wondering this morning about the possible application of automata to infinite and semi-infinite strings. I know that in principle automata theorists have historically not restricted their interest to finite automata; it seems plausible to assume they have also considered infinite strings. But I don’t know what they have said, without spending time looking it up; instead, I am going to enjoy myself for a little while seeing how much I can figure out for myself.
One obvious question to answer is: if you want to use an automaton to identify infinite sequences, how do you define acceptance of the sequence? For a finite sequence, you ask what state you’re in at the end of the sequence, and whether that state is an “accept state” or not. That won’t work for an infinite sequence: there is no final state.
Perhaps we can consider the sequence of states the automaton enters and define acceptance in terms of that sequence. Possible answers:
- Accept if (a) the automaton eventually ends up in a single state which it never again leaves, and (b) that state is an accept state.
- Accept if there is some point in the sequence of states such that every state following that point is an accept state.
These would work (in the sense of providing a yes/no answer).
Do these rules for acceptance of strings define sets of automata with different discriminating power?
It seems obvious that they do, but what exactly are the differences?
Consider, for example, automata for recognizing various classes of numbers written as an infinite sequence of decimal digits. Numbers seem to be on my mind, perhaps because of the tie-in to XSD datatypes.
For such infinite strings of digits (including a decimal point), integers have the property that every digit to the right of (i.e. following) the decimal point is a 0. If you build the obvious automaton, for an integer it will spend all its time in the zero-after-decimal-point state, and for a non-integer it will, eventually, end up caught in an error state.
[Enrique tells me I should pause to draw pictures of these automata, but I’m not going to take the time just yet. Apologies to those who find it hard to visualize what I’m talking about.]
So the first acceptance rule suffices for recognizing integers. It may be relevant that the same automaton can be used to recognize finite strings as representing integers: any prefix of the infinite string representing an integer will also be accepted as representing an integer.
The first rule would also suffice to allow us to build a recognizer for certain fractions, e.g. 1/3: the infinite string denoting 1/3 ends up perpetually in the “we’ve just read a 3” state.
On the other hand, it doesn’t suffice for all rationals: in decimal notation,1/7 has an infinitely repeating sequence of digits (
142857, if you were wondering). To distinguish 1/7 in decimal notation we’d need a cycle of six states in the automaton.
All rational numbers have a decimal expansion that eventually settles into an infinite series of repeated strings of digits (if only an infinitely repeating sequence of zeroes). So if we adopt the second rule for defining acceptance of the string, we can say: for every rational number, there is a finite state automaton that recognizes that rational number. And irrationals, which have no repeating sequences, aren’t recognizable by an automaton with finite states. (An automaton with infinitely many states might be able to recognize the decimal expansion of a particular irrational number, say π, but it’s hard to know what to do with that information — maybe it’s a reason to say that languages recognizable with an infinite automaton are not necessarily regular.)
That sounds like a nice pattern. It would be even nicer if we could devise an automaton to recognize the set of decimal expansions of rational numbers, but I suspect that’s not feasible, since the complement of that set is the irrationals, and being able to recognize the one set by regular means would entail being able to recognize the other, too.
Does it make sense to require that the automaton eventually end up spending all its time in accept states? (Or equivalently, that the sequence of states have a suffix in which every element in the suffix is an accept state.)
What if that is too restrictive a rule? What if we said instead
- Accept if at every point in the sequence of states there are an infinite number of accept states among the states following that point.
That is, allow the string to put the automaton into a non-accepting state, as long as it’s only temporary, and it eventually gets back into an accepting state.
Consider an automaton which has two states, A and B. Every time a B is found in the input, we go to state B; for any other symbol we go to state A. B is an accept state.
If we adopt the second story about termination, a string ending in an unending series of Bs will be accepted and is thus recognizable by an automaton. A string with an infinite number of Bs, interspersed with other symbols, will not be accepted by this automaton (nor by any other, as far as I can tell).
OK, that seems to establish (if we accept the conjecture about strings with infinitely many Bs) that the second and third rules define distinct sets of languages. I suppose that one chooses to use the second rule, or the third, or some other I haven’t thought of yet, in part based on whether it feels right to count as regular the languages one can recognize using that rule.
Hmm. OK, time to look at the bookshelves.
I’ve just checked and found that John E. Hopcroft and Jeffrey D. Ullman, in Introduction to automata theory, languages, and computation (Reading: Addison-Wesley, 1979), restrict their attention to finite strings.
Dick Grune and Ceriel J. H. Jacobs, Parsing techniques: a practical guide, second edition (New York: Springer, 2008), don’t explicitly impose this restriction. But a quick scan of their opening pages also doesn’t show any explicit consideration of infinite sequences of symbols, either. I’m guessing they do treat infinite input somewhere, if only because if you can contemplate van Wijngaarden grammars, which have infinite numbers of context-free rules (and remember, Grune didn’t just contemplate van Wijngaarden grammars, he wrote a parser generator for them), infinite strings are just not going to frighten you.
I suppose the idea of thinking seriously about infinitely long sentences in a language is one I first encountered in D. Terence Langendoen and Paul Postal, The vastness of natural languages (Oxford: Blackwell, 1984). To whom (for this, as for many other things) thanks!
I’m pretty sure that there was some treatment of infinite automata and/or infinite input strings in S. C. Kleene, “Representation of events in nerve nets and finite automata”, in Automata studies, ed. C. E. Shannon and J. McCarthy (Princeton: PUP, 1956), and V. M. Glushkov, “The abstract theory of automata”, Russian mathematical surveys: a translation of the survey articles and of selected biographical articles in Uspekhi matematicheskikh nauk 16 (1961). They are both kind of tough sledding, but I suppose I really ought to go back and read them carefully with an eye to this topic.