Equational languages

INFORMATIONANDCONTROL21, 134--147 (19.72) Equational Languages AI~DRZEJ BLIKLE Computation Center, Polish Academy of Sciences, Warsaw, Poland This paper deals with equations whose solutions are vectors of languages. Formally, solutions of equations are fix points of vectorial functions on languages. On the other hand equations (and sets of equations) can be considered as grammars. Three main groups of problems are dealt with: (1) solvability of equations in a lattice of languages, (2) relationship between type of functions used in equations and properties of languages defined by them, (3) applications to the theory of Context-free and regular languages. 1. INTRODUCTION Two principa! language-describing tools known in the theory of formal languages are generative (Chomsky's) grammars and accepting automata. On the other hand, in describing programming languages one uses mostly so-called Backus-Naur equations. These, although called equations, are not treated in a proper algebraic sense, but rather as formal expressions "to be referred to as equations." In fact, : :-- stands for -~ and [ stands for k3 (the union), but this is never improve d in practice. Moreover, to describe the meaning of such equations one associates with them context-free grammars. In effect the "equational" character of equations is lost. Sets of equations describing languages have previously been investigated by other authors. Chomsky and Schutzenberger (1963) discussed such equations understood as formal expressions. This idea was later developed in Mezei and Wright (1967) and in Shamir (1967). Equations have been treated as formal expressions which offers considerable difficulty in formulating and proving theorems. In the meantime Ginsburg and Rice (1962) and Ginsburg (1966) described Backus-Naur equations by associating with them sets of certain many-argumental functions on languages and used this tool in proving the equivalence of ALOoL-like and context-free concepts. The ideas of Ginsburg and Rice are generalized and developed in this paper. An earlier exposition of the present results was given in a technical report (Blikle, 1971). Copyright © 1972by AcademicPress, Inc. All rights of reproductionin any formreserved. 134 EQUATIONAL LANGUAGES 135 2. BASIC NOTIONS T h e reader is assumed t o be familiar with basic notions of the theory of formal languages like word, language, concatenation, substitution, homomorphism, etc. [cf. Ginsburg (1966)]. Let V be an arbitrary (finite) alphabet to be fixed for the sequel. For any positive integer n the n-dimensional lattice of languages is the set ~e~ = (2~*),, of all vectors (A 1 ,..., A~) with Ai C_ V* for i = 1,..., n. Clearly £a~ is a lattice (complete) with respect to the ordering C~ defined by the formula (A1 ..... Am) _C~ (B, .... , B~) ~ef (Vi ~< n)(A, C_B d. I n the sequel we shall omit the superscripts and write _C instead of _C~. T h e lattice operations LJ, (3, U and ~ in ~ n and the n-dimensional concatenation will be also written without superscripts, e,g., ( A l ' " " An)° (BI'"" Bn) dee~f(Al° B1 ..... An oBn), etc. Elements of ~gon will usually be denoted by boldface latin initial capitals A, B, C,.... Boldface latin terminals X, Y, Z .... will denote variables ranging over ~on. Moreover, ~ ~ = ( ~ ..... ~), where ~ is the empty set. I n the paper we shall be concerned with total functions of the type F : ~ - + ~ . T h e cases of particular interest are n ~- m--vectorial functions (abr. v functions), and m = 1--scalar functions (abr. s functions). A function F : 5¢ = - + ~oq~ is called monotonic i r a _CB implies F(A) C F ( B ) for any A and B in ~ n . F is said to be continuous, if F([)~°=I A~) = 0[=IF(A~) for any increasing sequence A I C A2 C .-- of vectors in ~?% It can be interesting to note that this notion of continuity coincides with the one of D. Scott [cf. Scott (1970)] for the lattice of languages o9.°~. However it seems to be different for arbitrary complete lattices. As it is easy tosee, each continuous function is monotonic. Examples of continuous functions are "polynomials" constructed with the operations of union, concatenation and • closure, e.g., F(X, Y) = ( A X u Y*, A X B Y * ) is a 2-dimensional continuous function and G(X, Y) = ( X - Y, Y -- X) is neither continuous nor monotonic. Consider an arbitrary vectorial function F : ~c:,~__> ~ . I f there exists a 136 BLIKLE vector A in L~°n with F(A) ----A, then A is said to be afixpoint (abr. FP) ofF. I f A is an FP o f F and for any other FP B ofF, A C B, then A is said to be the leastfix point (abr. LFP) o f F . We shall denote it by [IF[]. The following approximation theorem is well known [cf. Mezei and Wright (1967), Shamir (1967), Scott (1970) and others]. THEOREM 2.1. exists and For every continuous function F : ~ n - - ~ ~f~ the LFP HF H .% IIF[I = U Fi(~n) ' i=1 where Fi(X) = F "" F ( X ) i times. Let F(X, Y) be an arbitrary n + m-argumental function, where X ranges over ~ , Y ranges over ~ and F ranges over 5~~ again. In other words F : 0,% °'+'~ -+ ~q~n. I f there exists a function G : ~ m __+ ~q~n such that for every vector A in ~qom, G(A) is the L F P ofF(X, A), then G is said to be the resolvent o f F with respect to X and is written [ef. De Bakker (1971)] G(Y) = (t~X)F(X, V). As it is easy to see, G(Y) is a solution of the equation X = F(X, Y), therefrom the name. On the other hand if m = 0, then (/~X)F(X) = / I F ] I provided []F [I exists. The following theorem is proved in Leszczytowski (1971): THEOREM 2.2. For every continuous function F(X, Y), where X and F ranges over 5~n, the resolvent G(Y) = (/~X)F(X, Y) exists and is a continuous function. With the help of Theorem 2.1 one can easily prove the following formulas for arbitrary languages A, B and C: (~x)x = z, (~X)A = A, (I~X)(AX t3 B) = A ' B , ( ~ x ) ( x A w B) = BA*, ( ~ X ) ( X * A u A) = A+, ( ~ X ) ( A X B u C) = 0 A~CB~, i=O etc, (1) EQUATIONAL LANGUAGES 137 3. SETS OF EQUATIONS Let A e cpn. By [A]i for i ~< n we shall denote the i-th coordinate of A, i.e., if A = (A 1 ,..., An), then [A]i = A i . With every function F : ~ --+ ~ we can associate a vector (fz ,..-, f~) of n-argumental scalar functions defined as follows: for every A in ~ n and every i ~ m , f i ( A ) = [F(A)]i. For the sake of convenience we shall frequently identify the function F with the corresponding vector of functions (fl,...,.f~) and write simply F = (fl ..... f~). Clearly F is continuous if f l ,...,fr~ are all continuous. Consider now a set of n-argumental scalar functions fl ,-.., f~ and the following set of equations: X z = f l ( X l , . . . , Xn) • .. (2) X , = fn(X1 ,..., Xn). Any fix point of the function F = (fl ,..., f~) is said to be a solution of this set. If I[F [I exists, then IIF II is said to be the least solution of (2). To simplify the notation (2) will be frequently written as X = F(X). Similarly, if A = (At .... , Am) and B = (B1 ,..., B~), then by (A, B) we shall mean the vector (A 1 ,..., A n , Bz .... , B~). By a set of equations with parameters we shall mean any set of equations of the form X = F(X, Y), where X and F range over the same ~ " . Now if G(Y) = (/~X)F(X, Y) exists, then it is called the solution of this set. Clearly, F(G(Y), Y) = G(Y) for any Y. T w o sets of equations X -- F(X) and Y = H(Y) are said to be equivalent if either both 11F 1I and I[ G li do not exist or both exist and are equal. THEOREM 3.1. Consider the set of equations of the form X = F(X, Y), ¥ = H ( x , v), where F : .~n+m __~ £~n, H : ~,q~+~ --~ ~-q~, X ranges over 5q n and Y ranges over L f m. 138 BLIKLE I f F and H are continuous functions and G(Y) = (/xX)F(X, Y), then the set of equations X = a(Y), Y -~ H(G(Y), Y), is equivalent with the set (3). This theorem is proved in LeszczyIowski (1971) and permits resolving equations in ~ by eliminating variables. It is to be emphasized that the theorem is not true for arbitrary F and G, i.e., where F and G are not continuous. 4. EQUATIONALLANGUAGES Let ~ be an arbitrary family of Scalar functions. A language A is said to be equational with respect to ~ if there exists a v-function F ~ (fl ,...,f~) with f l .... ,fn in F, such that Hflj exists and A = [IIFII]~ for some i ~ n. By E Q [ ~ ] we denote the set of all languages equational with respect to ~ . THEOREM 4.1. Let ~ be an arbitrary family of continuous functions. The family of languages EQ[,~-] is closed under all functions in Y and under the resolvents of all functions in ~ . Proof. Let g ( X I ,..., Xn) be in # - and let A 1 ,..., An be in EQ[o~]. By virtue of the assumption there exists a v-function F over o~ with Ai = [/IF il]i for i ~- 1,..., n. Consider now the set of equations •x = F(X), Y = g(X, Y), and let (d 1 .... , An) -- A. By Theorem 3.1, (A, (/~Y)g(A, Y)) is the least solution of this set. Therefore (l~Y)g(A, Y ) ~ EQ[o~-], which proves the second part of the assertion. The first part follows immediately therefrom since every function in o~ is a resolvent of a function in ~-. Indeed, f ( X ) = (/xY)f(X). Q.E.D. Consider now an arbitrary family ~,~ of continuous functions and let R C C [ ~ ] (to be read: resolvent-composition closure) denotes the least set of functions that contains ~ as a subset and is dosed under the/~ operation and the composition of functions. EQUATIONAL LANGUAGES 139 THEOREM 4.2. I f Y is a family of continuous functions and contains the identity function f (~¥) ~ X, then EQ[~-] = EQ[RCC[~]]. Pro@ Let ~ be an arbitrary family of continuous functions with f ( X ) -- X. Consider an increasing sequence of sets of functions Y0, o~ ,... defined as follows: (1) J 0 = o~, (2) ~+~ is the least set with the following properties: (i) ~ C o%~ ; (ii) if F, G are in 4 , then the resolvent o f f with respect to any vector of variables, and any composition of F and G are in ~ + ~ . Clearly R C C [ ~ ] = Ui=o ~ . Consider now the set of equations x = (~z) F ( x , Y, z), (4) Y = H(X, Y), and the set z = F ( x , Y, z), x = z, (s) Y = H(X, Y). By Theorem 3.1 the last set is equivalent with Z = (~Z)F(X, Y, Z) X = (/xZ)F(X, Y, Z) (6) ¥ = H(X, V). Clearly, if (A, B) is the least solution of (4), then (A, A, B) is the least solution of (6), hence also of (5). In a similar way we can show that the set of equations x = F ( x , G(X, V), V), V = ~ ( x , v), can be "replaced" by the set Z = a ( x , V), X = F(X, G(X, Y), Y), Y = H(X, Y). 643]2112-4 140 BLIKLE In effect we have proved that every set of equations in o~+1 can be replaced (in a sense) by a corresponding set in 4 . In other words E Q [ ~ + I ] C EQ[o~], which completes the proof. Q.E.D. This theorem together with Theorem 4.1 imply immediately what follows: COROLLARY 4.1. Let ~ be an arbitrary family of continuous functions containing the identity function f (X) = X. The equational family of languages EQ[~-] is closed under all functions in R C C [ ~ ] . It should be stressed that R C C [ f f ] need not contain all functions that do not lead out of E Q [ f f ] . For appropriate examples see Blikle (1971). By an operation on languages we shall mean any function f : ~ 1 __+ oLpl. Particular operations on languages are substitutions, homomorphisms and inverse homomorphisms. As it is easy to show [see Blikle (1971)] all these operations are continuous. Given an operation h : 5¢1 _+ 5 0 we shall denote h " ( & .... , X , ) ~r (h(&),..., h(X,)). Given two n-argumental functions F and G, F o G will denote the composition, i.e., [F o G](X) = G(F(X)). TtlEORF.M 4.3. Let ~" be an arbitrary family of continuous functions and let h be an arbitrary continuous operation with h ( ~ ) = Z . I f for every n-argumental v-function F in o~ there exists a v-function F 1 in o~ with Foh" = h"oF1, then the family of equational languages E Q [ ~ ] is closed under h, i.e., h ( E Q [ ~ ] ) _C_CEQ[o~-]. Proof. Let h be an arbitrary continuous operation with h ( ~ ) = g and for some F in o~ let there exist F 1 in o~ with F o h ~ = h n o F 1 . Consider now two sets of equations x = e(x) and X : El(X) and let A be the least solution of the first one, and B of the second one. We shall show the equality B = h"(A). EQUATIONAL LANGUAGES 141 B = 0 Fli(Z~)" i=1 (7) Indeed, By T h e o r e m 2.1, On the other hand, FI( z ~) = F~(h~( z ~)) = hn(F( z ~)). Therefore, by an inductive argument, Eli( Z n) = hn(Fi( Z n)) for i = 1, 2,.... Hence, by (7), o B = h - ( F ~ ( z ~)) = h i=1 I91 F(z = ) n) = ha(A). Q.E.D. 5. INDUCTIVE FAMILIES OF FUNCTIONS Consider an arbitrary finite set eI ,..., en of scalar functions and an arbitrary family ~0 of languages. By the inductive f a m i l y of functions over e1 ..... en and ~ 0 , in symbols I N D [ e I .... ,e~, ~0], we mean the least set of many-argumental scalar functions which contains (1 °) all projections f ( X 1 ,..., X ~ ) = X i with i ~< n, (2 °) all constant functions with values in ~°0, (3 °) the functions e1 ,..., e~, and which is closed under the operation of composition. Consider now an infinite sequence o~0, ~ .... of families of functions defined as follows: (1) o~o consists of all functions of the type (1 °) and (2°), (2) o~+1 consists of all functions of the form g ( X 1 ,..., X ~ ) = ej(gl(X1 .... , X~),..., gk~(X1 .... , Xm)), w h e r e g l .... , g k ~ e ~ . 142 BLIKLE As it is easy to see, IND[el ,..-, en, ~o1 = 0 o~. (8) i=1 THEOREM 5.1. Let IND[e 1 ,..., en, ~o] be an arbitrary inductive family of functions, where e1 ,..., e~ are continuous and let h be an arbitrary continuous operation with h( ~g) = ~ . I f ~o is closed under h and if for every ei there exists anfi in IND[e 1 ,..., e . , ~fo] with h(ei(Xl ,..., Xm)) = fi(h(Xl),..., h(Xm)), then the equational family of languages EQ[IND[e 1 .... , e~, 5eo]] is closed under the operation h. Proof. Let the assumptions of the theorem be satisfied, let F ~ IND[e~ ,..., e~, ~0] and let f o , o~,.., be the sequence of families of functions in the sense of (8), where ~0 = ~ . We can easily prove by induction the following assertion: (,) for every i ~ 1 and every g in ~ there exists f in ~" with h(g(Xa ,..., X,~)) = f(h(Xa),..., h(X~,)) for any X~ ..... X~ in 5 0 . This, by Theorem 4.3, completes the proof of our theorem. Q.E.D. THEOREM 5.2. Let e1 ,..., e~ be arbitrary continuous functions and let ~o be an arbitrary family of languages. I f ~ = EQ[IND[e 1 ,..., e , , ~o]], then 5eo C ~ and = EQ[IND[e 1 ,..., e,~, ~W~]]. Proof. Let IND[e 1 ..... e,] denote the least set of functions that contains e1 ,..., e. and all projections and that is closed under the operation of composition. The following assertion can easily be proved by (8) and the induction on i. The proof is left to the reader. (**) For every f in IND[e 1 ..... e , , ~0] there exists g in IND[e 1 ,..., e~] and A 1 ,..., A~ in ~o with f(xl ,..., x ~ ) = g(X~ ,..., x ~ , n l ,..., n~). Now we can start the proper proof of the theorem. The inclusion ~o __Cis obvious since every A in ~0 is the least solution of the equation X ~ A." EQUATIONAL LANGUAGES 143 Let now IND0 ---- I N D [ e l ,..., en, ~o], I N D 1 ~- I N D [ e 1 ..... e~, ~ ] . Consider the set of equations x = F(X), where F = (fl,".,f,~) with f l .... , f ~ e l N D 1 and let A = IIFI;. W e shall show now that A ~ ( ~ ) m . L e t gx ,...,gin be these functions in I N D [ e 1 ,..., en] which correspond to f l , . . . , f ~ in the sense of (**). W i t h o u t any loss of generality we can assume now the existence of a natural n u m b e r k and a vector B in ~ with the p r o p e r t y that gl ,..., g~ are all m -}- k-argumental functions and that f,(X) = g~(X, B) (9) for i = 1.... , m. This assumption is clearly equivalent with adding to every gi a corresponding n u m b e r of "unneccessary" variables. L e t now B = (BI ,..., B~). By our assumption B 1 ,..., B~ a~qPl, thus there exist v-functions F I , . . . , F k in I N D 0 with [I]F~ II]l = Be (10) for i ---- 1,..., k. Consider now the following set of equations: z~ = & ( z ~ ) , Z~ = G(Z~), Yl = Z n , (11) . . o x = a ( x , Y), where G = (g~ ,...,g~), Y ---- (I71 ,..., Y~), Z~ = (ZI~ .... , Z,,~). This clearly a set of equations in I N D 0 . Consider now the set of equations: Z l = F~(Zl), . . . Z~ = G(Z~), Y~ - - Zl~, Y~ = z ~ is 144 BLIKLE and let (C1 ,..., Ck, D) be the least solution of this set. By Theorem 3.1, D ~ ([[IF11111.... , [liFe [[]1), thus, by (10) D = (B 1 ,..., B~) = B. Therefore, once more by Theorem 3.1, the least solution of (11) is of the form (C1 ..... C~, B, E), where E is the least solution of the set X = G(X, B). Clearly the last set is equivalent to x = F(X) [see (9)], thus (C 1 ,...,C~, B, A) is the least solution of (11). Therefore A ~(L1) ~ since (11) is clearly a set in I N D 0 . Q.E.D. T h e above theorem permits strengthening Theorem 5.1 into the following form: Tn~ORZM 5.3. Let el ,..., e, be arbitrary continuous functions, let 5eo be an arbitrary family of languages, let ~ = E Q [ I N D [ e 1 ,..., e,~, ~eo]] and let h be an arbitrary continuous operation. I f h(~o) C ~ and for every ei there exists f, in I N D [ e 1 .... , e , , 4 ] with h ( e i ( X 1 .... , X m ) ) = f i ( h ( X l ) , . . . , h(Xm)), then the family ~ is closed under h, i.e., h(~q~) C_~ . Proof. Suppose the assumptions of the theorem are satisfied and let ~ 0 = I N D [ e l ,..-, e , , ~0] and ~'1 = IND[el .... , e , , 5¢z]. I n a way analogous to that in the proof of T h e o r e m 5.1 we can easily show the following: (***) For every f in ~ 0 there exists an f l in ~ 1 with h(f(Xx .... , Xra)) = f l ( h ( X l ) , . . . , h(X,n)) for any X 1 ,..., X ~ . Let now A ~ ( ~ ) ~ . By this assumption there existsF in ~-o with 1[F 1] = A. Hence, by (***), there exists F 1 in #-1 with F o h~ = h ~ oF 1 . Let B = 11Fx[1- By T h e o r e m 5.2 B e (~LP1)'~ and the equality h~(A) = B can be proved analogously as in the proof of Theorem 4.3. Q.E.D. EQUATIONAL LANGUAGES 145 6. CONTEXT-FREELANGUAGES Let t_J, o and &oF denote, respectively, the operation of union, the operation of concatenation and the set of all finite languages, the empty languages included, over the alphabet V fixed in Section 2. Let &oc~denote the class of all context-free languages over V. By the set of all standard polynomials we mean the inductive family IND[k3, o, &or]. Clearly, standard polynomials are functions which appear in the so-called Backus-Naur equations in the definition of ALGOLand other similar languages. Therefore, the well-known theorem [see Ginsburg and Rice (1962) and also Ginsburg (1966)] to the effect that ALGoL-likelanguages and CF languages are the same, has now the following wording: THEOREM 6.1. EQ[IND[w, o, &OF]] = &OcF. According to this theorem CF languages are definable by means of polynomial equations with finite "coefficients." By Theorem 5.2 we can claim now that CF languages can be defined also by polynomial equations with CF coefficients, and by Theorem 4.2 that polynomial equations can be replaced by equations in RCC[IND[k), o, &OcF]]. In other words we have the following: THEOREM 6.2. EQ[IND[U, o, &OCF]] = &OCF" THEOREM 6.3. EQ[RCC[IND[U, o, &OcF]]] = &OCF. Note that the star closure is in RCC[u, o, &OcF]since Y* = ( f z X ) ( Y X u {~}). Consequently polynomial equations with star operation define always CF languages. For example the set of equations X = ( X Y ) * A W Y*, Y = X Y * X u B, defines two CF languages, provided A, B e &OcF• The above theorems together with theorems in Sections 4 and 5 imply immediately some well-known results [see Ginsburg (1966)] concerning closure properties of &OcF: THEOREM 6.4. &OCFis closed under the union, the concatenation and the star closure of languages. The proof is by Theorems 6.1 and 4.1. 146 BLIKLE THEOREM 6.5. &°CFis closed under the operation of substitution, Le., under every substitution s with the property that s(a) G ~ c F for every a in V. Proof. Let s be an arbitrary substitution with s(a) G ~q°cvfor every a in V. Since ooq°CFis dosed under the union and the concatenation, s(S~°F)__C~ c F Moreover, s(A: u & ) = s(&) u 4 & ) and s(AIA2) = s(A1) s(A2) for any _d1 and -//2 in coI. Therefore, by Theorem 5.3 and by the observation that s is continuous, ~CF is closed under s. Q.E.D. Also the class ~¢~ of all regular (finite-state) languages is an equational class of languages. Let SRLP (to be read: standard right-linear polynomials) denote the class of all functions of the form f ( X : ,..., Xn) = A o u A : X a u " " U A , X , ~ , (12) where A 0 ,..., A~ are finite languages. As it is proved in Blikle (1971), THEOREM 6.6. EQ[SRLP] = ~ R . It can be also proved that finite coefficients in (12) can be replaced by arbitrary regular coefficients and Theorem 6.6 remains true. However SRLP is not an inductive family of functions, thus the theorems given in Section 5 cannot be applied in this case. A practical application of the theory of equations to the theory of regular languages and finite-state automata is an effective and simple algorithm that associates with every fs automaton the corresponding regular expression. To this effect one needs only write and resolve a simple set of right-linear equations. For details see Blikle (1971). ACKNOWLEDGMENTS I wish to thank Mr. J. Leszczytowski to whom the Theorems 2.2 and 3.1 are due. I also wish to thank Mrs. B. Konikowska, Mr. A. Skowron, Mr. A, Walat, and Dr. j. Winkowski for many valuable discussions and remarks. RECEIVED: June 15, 1971 ; REVISED:April 7, 1972 EQUATIONAL LANGUAGES 147 REFERENCES DE BAKKER, J. W. (1971), Recursive procedures, Mathematical Center Tracts 24, Mathematisch Centrum Amsterdam, 1971. BLIKLE, A. (1971), Equations in a space of languages, CC PAS Reports 43, Computation Centre, Polish Academy of Sciences, Warsaw, 1971. CHO~SKY, N., AND SCHUTZENBERGER,M. P. (1963), The algebraic theory of context free languages, in "Computer Programming and Formal Systems" (P. Braffort and D. Hirschberg, Eds.), pp. 118-161, Amsterdam. GINSBURG, S. (1966), " T h e Mathematical Theory of Context Free Languages," McGraw-Hill Book Company, New York. GINSBURG, S., AND RICE, H. G. (1962), Two families of languages related to ALGOL, J. Assoc. Comput. Mach. 9, 350-371. LESZeZYLOWSKI,J. (1971), A theorem on resolving equations in the space of languages, Bull. Acad. Polon. Sci. Sgr. ScL Math. Astronom. Phys. 19, 967-970. MEZEI, J., AND WRIGHT, J. B. (1967), Algebraic automata and context free sets, Information and Control II, 3-29. SCOTT, D. (1970), Lattice-theoretic models for the A-calculus, I F I P Bulletin No. 5, 1-50, W G 2.2. SHA•IR, E. (1967), A representation theorem for algebraic and context free power series in noncommuting variables, Information and Control 11, 239-254.