The universal automaton

This paper is a survey on the universal automaton, which is an automaton canonically associated with every language. In the last forty years, many objects have been defined or studied, that are indeed closely related to the universal automaton. We first show that every automaton that accepts a given language has a morphic image which is a subautomaton of the universal automaton of this language. This property justifies the name “universal” that we have coined for this automaton. The universal automaton of a regular language is finite and can be effectively computed in the syntactic monoid or, more efficiently, from the minimal automaton of the language. We describe the construction that leads to tight bounds on the size of the universal automaton. Another outcome of the effective construction of the universal automaton is the computation of a minimal NFA accepting a given language, or approximations of such a minimal NFA. From another point of view, the universal automaton of a language is based on the factorisations of this language, and is thus involved in the problems of factorisations and approximations of languages. Last, but not least, we show how the universal automaton gives an elegant solution to the star height problem for some classes of languages (pure-group or reversible languages). With every language is canonically associated an automaton, called the universal automaton of the language, which is finite whenever the language is regular. It is large, it is complex, it is complicated to compute, but it contains, hopefully, many interesting informations on the language. In the last forty years, it has been described a number of times, more or less explicitly, Erich Grädel, Jörg Flum, Thomas Wilke (eds.). Logic and Automata: History and Perspectives. Texts in Logic and Games 2, Amsterdam University Press 2007, pp. 467–514. 468 S. Lombardy, J. Sakarovitch more or less approximately, in relation with one or another property of the language. This is what we review here systematically. 1 A brief history of the universal automaton The origin of the universal automaton is not completely clear. A wellpublicized note [ADN92] credits Christian Carrez of what seems to be the first definition of the universal automaton in a report that remained unpublished [Car70]. The problem at stake was the computation of the, or of a, NFA with minimal number of states that recognizes a given regular language L. And Carrez’s report states the existence of an automaton UL, very much in the way we do in Section 2, with the property that it contains a morphic image of any automaton which recognizes L, and thus a copy of any minimal NFA which recognizes L. At about the same time, Kameda and Weiner tackled the same problem and, without stating the existence of UL, described a construction for a NFA recognizing L with minimal number of states [KW70], a construction which we recognize now as being similar to the construction of UL we propose in Section s.con-uni-aut. Soon afterwards, in another context, and with no connexion of any kind with the previous problem (cf. Section 6) Conway proposed the definition of what can be seen also as an automaton attached to L and which is again equal to UL [Con71] (cf. Section 3.1). Among other work related to UL, but without reference to the previous one, let us quote [CNP91] and [MP95]. Eventually, we got interested in the universal automaton as we discovered it contains other informations on the languages that those studied before (see Section 7.1) and we made the connexion between the different instances [LS03, LS02]. 2 Creation of the universal automaton No wonder, we first fix some notations. IfX is a set, P (X) denote the power set of X , i.e. the set of subsets of X . We denote by A the free monoid generated by a set A. Elements of A are words, the identity of A is the empty word, written 1A∗ . The product in A ∗ is denoted by concatenation and is extended by additivity to P (A): XY = {uv | u ∈ X, v ∈ Y }. An automaton A is a 5-tuple A = 〈Q,A,E, I, T 〉, where Q is a finite set of states, A is a finite set of letters, E, the set of transitions, is a subset of Q × A× Q, and I (resp. T ), the set of initial (resp. terminal) states, is a subset of Q. Such an automaton A defines an action ⊲ of A on P (Q), 1 Normally, we would have denoted the action by a simple ·; but later, in Section 5, we shall need to consider an action on the right and an action on the left, hence a lateralized symbol which makes the reading easier. Moreover, when necessary, i.e. when several automata are considered at the same time, we shall even specify as a The Universal Automaton 469 by setting first for all p ∈ Q and all a ∈ A p ⊲ a = {q ∈ Q | (p, a, q) ∈ E}, and then by additivity and the definition of an action for all X ∈ P (Q) X ⊲ 1A∗ = X , X ⊲ a = ⋃ p∈X p ⊲ a , X ⊲ wa = (X ⊲ w) ⊲ a. The behaviour |A| (or the accepted language) of an automaton A = 〈Q,A,E, I, T 〉 is the set of words that label a path from an initial state to a terminal state, i.e. |A| = {w ∈ A | ∃i ∈ I, t ∈ T i w −−→ A t} = {w ∈ A | I ⊲ w ∩ T 6= ∅}. A subset of A is called a language and a language is regular if it is the behaviour of some finite automaton. Let A = 〈Q,A,E, I, T 〉 be an automaton over A. For each state p of A, the past of p is the set of labels of computation which go from an initial state of A to p, and we write it PastA(p); i.e. PastA(p) = {w ∈ A ∗ | ∃i ∈ I i w −−→ A p} = {w ∈ A | p ∈ I ⊲ w}. Dually, the future of p is the set of labels of computations that go from p to a final state of A and we write it FutA(p), i.e.: FutA(p) = {w ∈ A ∗ | ∃t ∈ T p w −−→ A t} = {w ∈ A | p ⊲ w ∩ T 6= ∅}. Likewise, for each pair of states (p, q) of A, the transitional language of (p, q) is the set of labels of computations that go from p to q and we write it TransA(p, q), i.e.: TransA(p, q) = {w ∈ A ∗ | p w −−→ A q} = {w ∈ A | q ∈ p ⊲ w}. For each p, q in Q, we clearly have [PastA(q)] [FutA(q)] ⊆ |A|. (∗) Thus, in every automaton, each state induces a set of ‘factorisations’ — which is the name we give to equations of the type (∗) — of the language it recognizes. The starting point of the construction is to prove the converse of this observation, namely that we can construct from the set of factorisations of a language L of A an automaton which accepts L. subscript the automaton that defines the action in action: p ⊲ A a. 470 S. Lombardy, J. Sakarovitch