0% found this document useful (0 votes)
12 views38 pages

Boolean Functions, Invariance Groups, and Parallel Complexity

This paper investigates the invariance groups of boolean functions and their implications for parallel complexity in computational languages. It establishes conditions for a permutation group to represent the invariance group of a boolean function and demonstrates that most boolean functions have trivial invariance groups. Additionally, it provides algorithms for determining group representability and explores the relationship between group theory and the complexity of languages with polynomial index.

Uploaded by

Anjali Maurya
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
12 views38 pages

Boolean Functions, Invariance Groups, and Parallel Complexity

This paper investigates the invariance groups of boolean functions and their implications for parallel complexity in computational languages. It establishes conditions for a permutation group to represent the invariance group of a boolean function and demonstrates that most boolean functions have trivial invariance groups. Additionally, it provides algorithms for determining group representability and explores the relationship between group theory and the complexity of languages with polynomial index.

Uploaded by

Anjali Maurya
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 38

SIAM J. COMPUT.

© 1991 Society for Industrial and Applied Mathematics


Yo!. ::?O, No. 3, pp. 553-590, June 1991 Oli

BOOLEAN FUNCTIONS, INVARIANCE GROUPS, AND


PARALLEL COMPLEXITY*
PETER CLOTEt AND EVANGELOS KRANAKIS:j:

Abstract. This paper studies the invariance groups S(J) of boolean functions f E Bn (i.e., f: {O, 1}"..,.
{ 0, 1}) on n variables, i.e., the set of all permutations on n elements which leave f invariant. After building
intuition by presenting several examples that suggest relations between algebraic properties of groups and
computational complexity of languages, necessary and sufficient conditions are given via P6lya's cycle index
for an arbitrary finite permutation group to be of the form S(f), for some/ E B". It is shown that asymptotically
'"almost all" boolean functions have trivial invariance groups. For cyclic groups G;::;; S,, a logs pace algorithm
for determining whether the given group is of the form S(J), for some fE B,, is given. The applicability of
group theoretic techniques in the study of the parallel complexity of languages is demonstrated. For any
language L let L,, be the characteristic function of the set of all strings in L which have length exactly n
and let S,,(L) be the invariance group of L,,. The index IS" :S,,(LJI are considered as a function of n and
the class of languages whose index is polynomial in n is studied. Bochert's lower bound on the index of
primitive permutation groups is used together with the O'Nan-Scott theorem, a deep result in the classification
of finite simple groups, in order to show that any language with polynomial index is in (nonuniform) TC 0
and hence in (nonuniform) NC 1• As a corollary, an extension is given of a result of Fagin-Klawe-Pippenger-
Stockmeyer, giving necessary and sufficient conditions for a language with polynomial index to be computable
by a constant depth polynomial size circuit family. As another corollary, it is shown that the problem of
""weight-swapping" for a sequence of groups of polynomial index is in (nonuniform) NC 1•

Key words. abelian group, boolean function, circuit, classification theory, cyclic-, dihedral-, hyperocta-
hedral-groups, index of a group, invariance group of boolean function, NC, parallel complexity, permutation
group, P61ya cycle index, pumping lemma, representable group, regular language, symmetric boolean
function, wreath product

AMS(MOS) subject classifications. 68Ql5, 68Q25, 68Q45

1. Introduction. The aim of this paper is to study the invariance groups of boolean
functions, provide efficient algorithms for determining the representability of a given
group as the invariance group of a boolean function, and use group theoretic techniques
in order to deduce results about the parallel complexity of formal languages.
Given n input values, each of which can assume one of two possible states 0, 1,
a "module" M outputs a value which assumes one of the states 0, l. The output of
the module when the input values are x 1 , • • • , Xn depends in general on the order of
the inputs. There are certain permutations of the input states which leave the output
state invariant or unchanged. For example, it may be that the output is independent
of any permutation of the input states, in which case the given module is called
symmetric. In general, for a given module, the set of permutations which, when applied
to any set of input states, leave the output invariant is easily seen to form a permutation
group.

*Received by the editors November 7, 1988; accepted for publication (in revised form) June 28, 1990.
This article first appeared as an extended abstract in the Proceedings of the Fourth Annual IEEE Conference
on Structure in Complexity Theory, 1989, pp. 55-65.
t Department of Computer Science, Boston College, Chestnut Hill, Massachusetts 02167. This author's
research was supported in part by National Science Foundation grant DCR-8606165. Some of this research
was performed while the author was visiting the Universite de Paris VII, Equipe de Logique Mathematique,
Centre Nationale de la Recherche Scientifique-UA 753, 2 Place Jussieu, Paris, France.
t Centrum voor Wiskunde en lnformatica, P.O. Box 4079, 1009 AB Amsterdam, the Netherlands.
553
554 P. CLOTE AND E. KRANAKIS

Formally, the operation performed by such an n-ary module M is usually repre-


sented by an n-ary boolean function 1 f:2n ~ 2. For fixed n, let the set of all such n-ary
boolean functions be denoted by Bn. If the input states of the module are assigned
the boolean values x 1 , • • • , Xn then by definition f(x 1 , • • • , xn) is the value of the
output state of the module M on input x 1 , • • • , xn- Given such an n-ary boolean
function f let S(f) be the set of all permutations on the n elements 1, 2, · · · , n such
that for all input values (xl> · · ·, Xn) E 2n, f(x1> · · ·, Xn) = f(xcr(lh · · ·, Xa(n))- Clearly,
the group S(f) equals the full symmetric group S,, exactly in the special case when
the boolean function f is symmetric.
By a counting argument Lupanov, Shannon, and Strassen have shown that almost
all boolean functions have exponential size circuit complexity. Despite this result, very
little is known concerning specific languages or families of boolean functions. Our
interest in the present study arose from attempting to use group theoretic techniques
in order to generalize the simple observation that any family Un: f,, E B", n EN} of
symmetric boolean functions is computable by a logarithmic depth, polynomial size
circuit family. Probabilistic techniques have been successfully used by several authors
(Furst, Saxe, and Sipser [FSS84], Yao [Yao85], etc.) in order to obtain lower bounds
on the size and/ or depth of circuit families which compute certain symmetric languages
(families of symmetric boolean functions). However, there are few results giving tight
upper bounds, apart from the above cited fact that any family of symmetric boolean
functions is computable by a nonuniform circuit family of logarithmic depth and
polynomial size (formula size bounds have been obtained by various authors in this
case). In this paper we indicate the applicability of group theory in obtaining upper
bounds for the parallel complexity of families of boolean functions. Our work is
different from, but somewhat related to, studies on the automorphism groups of
error-correcting codes (e.g., kth order Reed-Muller codes, which are specific k-
dimensional subspaces of 2" [MS78]), as well as to work in [Har64] where group
theoretic methods are used to calculate the number ofnonequivalent boolean functions,
where the equivalence relation is defined by f = g if and only if there exists u E Sn
such that for all X1, • • ·, Xn E {0, 1} (f(Xi. · · ', Xn) = g(Xo-(1)> • • ·, Xo-(n))).
In [FKL88] it was indicated how the classification theorem for finite simple groups
could be applied to VLSI technology by giving an algorithm to minimize pin-count in
a sequence of circuits. Here we consider the problem of placement of modules on a
chip where permutation of input wires is allowed. It is expected that study of the
invariance groups of boolean functions may lead to algorithms for optimizing space
in VLSI design, e.g., knowledge that certain modules leading into a block can be
permuted without changing the function computed.
It is interesting to point out that invariance groups are also relevant to the
computability problem for boolean functions in anonymous networks as used in
distributed computing. For example, we are interested in computing n-ary boolean
functions in an n-node anonymous network Jf. To compute the value of a given
function fat the input (bi.···, b,.) the processors p 1 , • • • ,pn are initialized with
the inputs b 1 , • • • , b", respectively. By exchanging messages through the links all the
processors must eventually compute the same bit b =f(bi. · · ·, b,.). It has been the
focus of several papers to determine and study networks for which
f is computable in Jf ~ S(f) 2 Aut(Jf),

1 Throughout the paper we identify a positive integer n with the set {O, l, · · ·, n -1}, e.g., 2 = {O, l}; in

general, however, we will prefer the set-notation when we want to emphasize the elements of the language
under consideration.
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 555

where Aut(.N') denotes the group of automorphisms of .N'. In fact, this is the case for
several types of networks, like directed and unlabeled rings [ASW85], labeled tori
[BB89], and labeled hypercubes [KK89].
1.1. Results of the paper. Following is an outline of the main results and contents
of the paper. We begin in § 2 by providing some preliminary results regarding the size
of the index of a permutation group. We remind the reader of the essential parts of
P6lya's beautiful enumeration theory that will be used in the present study.
In§§ 3 and 4, to build intuition for the reader, we present a number of examples
concerning the invariance groups of certain types of languages, such as palindromes,
parentheses, and regular languages, and study the reverse problem of constructing
languages realizing specific types of groups. We compute the invariance groups of
Dyck palindrome languages and give an efficient algorithm for determining membership
in the invariance group of regular languages. We show that each of the cyclic (for
n >6 3, 4, 5), dihedral, and hyperoctahedral sequences of groups are representable by
regular languages and construct groups which cannot be represented by regular
languages.
In § 5 we study the representation problem for general permutation groups. We
define a subgroup G ~Sn to be strongly representable if G is the invariance group of
an n-ary boolean function-Le., there exists f E Bn for which G = S(f). We distinguish
between groups which are "strongly representable" and groups which are "isomorphic
to strongly representable." In the latter case, we show that every permutation subgroup
of Sn is isomorphic to a strongly representable group S(f), for some f:2"< 108 n+l)-? 2;
but as stated, this isomorphism is at the expense of increasing the number of variables
in the boolean function from n to n(log n + 1). The problem is more interesting in the
former case, where we give a necessary and sufficient condition in terms of the P6lya
index, for an arbitrary subgroup of Sn to be of the form S(f), for some n-ary boolean
function f: 2n-? 2. Using the classification theorem for maximal permutation groups
we show that "with few exceptions" (essentially, only the alternating group An, for
n ~ 10) all maximal permutation groups on n letters are strongly representable. This
contrasts with the fact that there are numerous nonrepresentable permutation groups.
We also give a logspace algorithm which, on input of a cyclic group G ~Sn, decides
whether G is strongly representable, in which case it outputs a boolean function
f: 2n-? 2 such that G = S(f). Our last result in this section concerns asymptotics. For
any sequence of nonidentity permutation groups ( Gn ~Sn: n ~ 1) we prove that

It then immediately follows that asymptotically "almost all" boolean functions have
a trivial invariance group; i.e., they are equal to the identity permutation group.
Given a language L~ {O, 1}*, let Ln be the characteristic function of the set of
words of L oflength exactly n. Section 6 is concerned with the complexity of languages
of polynomial index, i.e., languages L for which there exists a polynomial p(n) such
that !Sn:Sn(L)l~p(n), where Sn(L) denotes the invariance group of the boolean
function Ln. We study the closure properties of the class of these languages and apply
the NC algorithm for permutation group membership of [BLS87] in order to show
that languages of polynomial index are in (nonuniform) NC. By using the O'Nan-Scott
theorem, a deep result in classification theory of finite simple groups, we improve the
last result to show that any language of polynomial index is in (nonuniform) TC 0 and
hence NC 1•
556 P. CLOTE AND E. KRANAKIS

In [FKPS85], Fagin, Klawe, Pippenger, and Stockmeyer used group theoretic


techniques together with the exponential size lower bound for constant depth circuits
accepting parity [Yao85] to give a necessary and sufficient condition for a symmetric
language Ls; {O, 1}* to belong to AC 0 ; i.e., for L to be computable by a nonuniform
circuit family of constant depth and polynomial size. Our characterization of languages
of polynomial index allows an immediate extension of this result. Namely, for Ls;
{O, l}* of polynomial index, L is in AC 0 if and only if the least number of input bits
which must be set to a constant in order for the resulting language L" = Ln {O, lt to
be constant is polylogarithmic in n.
As mentioned in the introduction, we believe that group theoretic considerations
may possibly play a role in VLSI design. In particular, knowledge of the invariance
group of "modules" might allow minimization of the surface area for automated circuit
layout. Toward a mathematical formalization of this idea, we introduce some notation.
For any sequence G = {Gn: Gn ~Sn, n EN} of permutation groups the problem
SWAP ( G) is given by the following.
Input. n EN, al, ... 'an positive rationals.
Output. A permutation <I E G,. such that for all 1 ~ i < n, a,ru) + a,,.(i+i);;:; 2, if such
a permutation exists, and the response "NO" otherwise.
The intuition behind the problem SWAP ( G) is that the output wires of modules
M 1 , • · · , M,. are the inputs to module M, and that the invariance group of M is 0 11 •
The "width" of module M; is the rational number a;. Modules Mi and Mj can be
placed next to each other if they do not "overlap"; i.e., exactly when a;+ aj;;:; 2, where
we imagine an average size of 1 per module. Thus, the output for SWAP ( G) indicates
whether there exists a permutation of the input modules Mi which does not change
the output of Mand which allows a layout of Mall)•···, M,,(n) without overlap. A
simple application of our work yields an NC 1 algorithm for the problem SWAP ( G),
where G = {Gn: Gn ~ S", n EN} is of polynomial index.
Recall that the stipulation of the layout problem is to find an optimal layout given
a number of modules together with their connections. A popular algorithm that attempts
to solve the layout problem is due to Kernighan and Lin [KL82] and partitions the
chip into an upper and a lower half, swapping modules on either side, trying to
minimize a certain parameter, then recursively partitioning simultaneously the top and
bottom into left and right parts, swapping modules between left and right parts to
minimize a parameter, etc. Our problem stipulation in SWAP is quite different: instead
of being given a list of modules and their connections (including which input port of
a target module), we allow the input ports of the target module to be swapped, provided
that the resultant function is not changed.
Finally, in § 7, we discuss some open problems and give directions for further
research.
An acquaintance with the standard results on group theory and finite permutation
groups, as presented for example in [Hal57] and [Wie64], will be essential for an
adequate understanding of the results of the present paper.
2. Preliminaries. Here we give some introductory definitions and results regarding
permutation groups and complexity of circuits that will be used in our subsequent
investigations. The three topics we will discuss are:
• the size of the group index,
• the size of the cycle index and its computation via P6lya's formula, and
• complexity of boolean functions with respect to the size and/ or depth of boolean
circuits computing them.
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 557

2.1. Index of a permutation group. In the sequel it will be convenient to think of


permutations on the set {1, 2, · · ·, n} as bijective mappings on the set of all positive
integers such that a( k) = k for all k > n. Part of this paper is primarily concerned with
"large" permutation subgroups of the full symmetric group. Let Sn denote the group
of all permutations of n elements, and An be the subgroup of even permutations (also
known as the alternating group on n letters). In general, for any nonempty set n let
Sn denote the set of all permutations of n. For any group G the symbol H ~ G means
that H is a subgroup of G. Regarding the sizes of permutation groups the following
theorem summarizes some known results on the sizes permutation groups.
THEOREM 1. Let H ~Sn be a permutation group which does not contain An.
(1) !Sn: Hj ~ n.
(2) If the order of H is maximal then ISn: Hj = n. In fact, for n ;&: 6 the subgroups
Hof Sn with !Sn: Hj = n are exactly the one point stabilizers of Sn.
(3) If H is primitive then
(Bochert) !Sn:Hl~[(n+l)/2]!.
(Praeger and Sax!) IHI <4n.
(Cameron) either His a "known" group or IHI< n iotoglogn.
Proof. For all three parts and further information, consult [Wie64], [Tzu82], as
well as the references in [KL88] (in particular, the proof of (3) is very hard). Part (1)
follows from the following claim.
CLAIM. If His a subgroup of G and IG: HI= n then there exists a normal subgroup
N of G such that N ~ H and IG: NI divides n ! .
Indeed, consider the set n ={Hg: g E G} of cosets of the quotient group G/ H.
By assumption, this set has size n. Let Sn be the group of permutations on n. For
each x E G consider the permutation ef>(x): 0 ' 0, where <P(x)(Hg) = Hgx. Clearly,
<P: G' Sn is a group homomorphism. Moreover, it is easy to see that
N:= Ker(ef>) = n Hg
gEG

is a normal subgroup of G, where Hg= g- 1 Hg. By the homomorphism theorem, the


order of the quotient group G / N divides the order of the permutation group Sn. This
proves the claim.
Now let us prove ( 1) by the above claim there exists a normal subgroup N of Sn
such that N ~ H and !Sn: NI divides (n -1) !. It follows that N ~ 1. Since the only
normal subgroups of Sn are An, Sn, and 1, the result is clear. D
2.2. Cycle index of a permutation group. Let G be a permutation group on n
elements. Define an equivalence relation i = j if and only if for some a E G, cr(i) = j.
The equivalence classes under this equivalence relation are called orbits. Let Gi =
{a E G: u(i) = i} be the stabilizer of i, and let i 0 be the orbit of i. An elementary
theorem asserts that jG: G;j = ji 0 j. Using this, we can obtain the well-known theorem
of Burnside and Frobenius, which states that for any permutation group G on n
elements, the number of orbits of G is equal to the average number of fixed points of
a permutation er E G,

(1)

where wn( G) is the number of orbits of G [Com70]. Any permutation u ES,, can be
identified with a permutation on 2n defined as follows:
X = (X1,' ' ', Xn) ' X " = (Xcr(I), . ' ' , Xcr(n)).
558 P. CLOTE AND E. KRANAKIS

Hence, any permutation group G on n elements can also be thought of as a permutation


group on the set 2n. It follows from (1) that

j{x 0 : x e2n}I =fGI1 "~a l{x e2": xa- =x}I,


where x 0 = {x"": u e G} is the orbit of x. We would like to find a more explicit formula
for the right-hand side of the above equation. To do this, note that x"" = x if and only
if x is invariant on the orbits of u. It follows that l{x E 2n: xa = x}I = 2°<a>, where o( u)
is the number of orbits of (the group generated by) u. Using the fact that o(<7) =
c 1 (<7)+· · ·+cn(<7), where c;(C7) is the number of i-cycles in CJ" {i.e., in the cycle
decomposition of u), we obtain P6lya's formula:

(2) l{xa:xe2"}1=-1- L 20(0-J=_l_ L 2"1<0->+·+<·.,<(T>.


IGI o-EG IGI aeG
The number l{x 0 : x e 2"}1 is called the cycle index of the permutation group G and
will be denoted by 0( G). If we want to stress that G is a permutation group on n
letters, then we write 0.(G), instead of E>(G). For more information on P61ya's
enumeration theory the reader should consult [Ber71] and [PR87].
Since the invariance group S(f) of a function f E Bn contains G if and only if it
is invariant on each of the different orbits x 0 , x e 2", we obtain that
j{f E Bn: S(f) ~ G}I = 2El(G)•
It is also not difficult to compare the size of 0( G) and ISn: GI. Indeed, let H ~ G;;;; Sn.
If
Hg 1 , Hg 2 , • • ·, Hgk
are the distinct right cosets of G modulo H then for any x E 2" we have that
x0 = XHg, u XHg u ... u XHgk.
2

It follows that 0n(H);;;;:E>n(G) · IG:HI. Using the fact that E>n(S,,)= n+l we obtain
as a special case that en (G);;;;: (n + l)IS 11 : Gj. In addition, using a simple argument
concerning the size of the orbits of a permutation group we obtain that if A1 , • • • , Aw
are different orbits of the group G;;;; Sn acting on {1, 2, · · · , n} then

We summarize these results in the following useful theorem.


THEOREM 2. For any permutation groups H;;;;: G;;;; S,, we have
(1) E>n(G);;;;0,,(H);;;;0,,(G) · IG:HI.
(2) E>,,{G);;;; (n+l) ·IS,,: GI.
(3) 11+1;;;;:0 11 (G);;;;2".
(4) If '11, · · -, llw are different orbits of G then (jA 1I + 1) · · · (lllwl + 1);;;; 0 11 ( G).
It is easy to see that in general IS.: GI and E>n ( G) can diverge widely. For example,
letf(n) = n -log n and let G be the group {<7 ES,,: Vi> f(n)(u(i) = i)}. It is then clear
that 0 11 ( G) = (f(n) + 1) · 2' 0 gn is of order n 2, while IS.: GI is of order 11 10& 11 • Another
simpler example is obtained when G is the identity subgroup of S,,.
2.3. Circuits. An n-circuit a 11 is a labeled, directed acyclic graph whose nodes
are labeled by x 1 , • • • , x11 (input bits), 1, 11, v. The input nodes are of in-degree 0 and
there is a unique output node whose out-degree is 0. The size c( a) of a,, is the number
of internal (i.e., noninput) nodes, while the depth d (a) of an is the maximal length
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 559

of a path from an input node to the output node. A word x E {O, 1}" is accepted by an
n-circuit an if each input node labeled by xi has as value the ith bit of x. An n-circuit
a,, recognizes or computes a language L,, s {O, 1}" (respectively, boolean function/ E B,,)
if and only if for all words x in {O, l}",
x EL,, (respectively /(x) = 1) <=> a,, accepts x.
A circuit family (an: a,, is an n-circuit, n EN) recognizes or computes a language
Lt::; {O, l}* if and only if for all n (a,, accepts Ln{O, 1r). In this paper, we usually
consider nonuniform circuit families as defined above-of course, such families can
recognize nonrecursive languages. A circuit family (a,,: n E N) is logspace uniform if
there is a logspace computable function F: 1" ~a,, for constructing the circuits. There
are stronger and weaker uniformity notions. See [Coo85] for further discussion and
for a survey of parallel complexity theory. The class SIZE-DEPTH(/, g) is the collection
of languages accepted by a family (a,,: n EN) where c( a,,) -;;2/(n) and d (a,,)~ g( n ).
The class ACk (respectively, NCk) is the collection of languages 2 belonging to SIZE-
DEPTH(n0(1>, O(logk (n))) where the in-degree of nodes labeled by /\,vis arbitrary
(respectively, 2). Of importance to this paper is the class AC 0 of languages accepted
by (nonuniform) circuit families of constant depth and polynomial size with arbitrary
fanin, and the class NC 1 of languages accepted by (nonuniform) circuit families of
logarithmic depth (and a fortiori polynomial size) withfanin 2. By unwinding a circuit
into an equivalent boolean formula (circuit with fanout 1), NC 1 is easily seen to be
the class of languages computable by (non uniform) polynomial size boolean formulas.
The class TC 0 is the collection of languages computable by (nonuniform) circuit
families with constant depth and polynomial size, whose gates are arbitrary fanin
threshold gates. NC is defined to be u nEN NCk. Trivially, NCk s Ack, and by replacing
an arbitrary fanin gate by a binary tree of fanin 2 gates, it is clear that ACk s NC k+i.
A language Ls {O, 1}* is said to have (or be computable by) polynomial size circuits,
denoted LE SIZE(n°( 11 ), if there is a circuit family (a,,: n EN) where a,, computes the
characteristic function of L,,=Ln{O, I}" and c(a,,)~p(n) for some polynomial p.
Note that SIZE(n°( 1 l) is the same class, whether one considers arbitrary fanin or fanin
2 circuits. Since the out-degree of a node is arbitrary, partial computations may be
reused; thus the circuit provides a model for parallel computation. Stockmeyer and
Vishkin [SV84] have shown that ACk is the class of languages computed in O(Iogk (n))
time with a polynomial number of processors on a parallel random access machine
(PRAM).
For a boolean function f: 2" "'2, we define
c(f) =min {c(a): a computes/}
where a has fanin 2. The following results are well known (e.g., see [Sav76] or [Yab83]).
In particular, we shall use the second fact in a later proof.
(1) For any symmetric functionfEB,,, c(f)= O(n).
(2) (Lupanov-Shannon-Strassen) J{f E Bn: c(f) < q}J = O(qq+ 1 ).
(3) For any s>O, the ratio of fEB,, such that c(f)>(l-s)2n-i;n tends to 1
as n "'co.
3. Invariance groups of certain languages. The main objects of study in this paper
are boolean functions and their invariance groups. Let Bn,k be the set of all k-valued

2 Usually these classes are defined to be classes of functions rather than languages. Since we will not

discuss function computations in this paper, we adopt the above definition.


560 P. CLOTE AND E. KRANAKIS

functions f: 2" ~ k on n boolean variables. If k = 2 then we abbreviate Bn, 2 by Bn. If


Z 2 denotes the finite two-element field then it is clear that

(x7- X;, i = 1, 2, · · · , n) ·
For x = (x1, · · ·, Xn) E 2" and <r E Sn, let x" = (xcr(l), · · ·, Xcr(n)). For any n-ary boolean
function f E B" let fa be defined by

The invariance group off is defined by

S(J) = {<rE Sn:f =f"}


={O"E Sn: 'Vx E 2"f"(x 1 , • • · , Xn) =f(xa(JJ> · · · , Xa(n))}.
If K £ {O, 1}" is a set of words of length n, then by abuse of notation we shall write
S(K) for the invariance group of the characteristic function of the set K. If Li;;;; {O, 1}*
is a set of finite words and n ~ 1 then Sn(L) denotes the invariance group of the n-ary
boolean function Ln. Clearly, S(f), being nonempty and closed under multiplication,
is a subgroup of Sn.
Here we compute the invariance groups of well-known formal languages. We
begin with the Dyck (or parenthesis) and palindrome languages and conclude with an
"efficient" algorithm for computing the invariance group of regular languages.
3.1. Dyck languages. The semiDyck language D [Harr78] is defined as the least
set of strings in the alphabet 0, 1 such that A ED and (for all x, y ED) (xy ED and
Oxl ED). The semiDyck language is not regular, as can be seen from the fact that the
elements O" give rise to infinitely many distinct equivalence classes in the right
congruence relation for D. The Dyck languages Dr, r ~ 1, are defined in the alphabet
L, = {O;, l;: i = 1, · · ·, r} in a similar fashion: D' is the least set of strings in the
alphabet L, such that A ED' and (for all x, y ED') (for all i ~ r) (xy E D' "O;xl; ED').
Clearly, D = D 1• Next we determine the invariance group of the Dyck languages.
THEOREM 3. For the Dyck language D' defined above we have that

S (D')={l ifnisoddorr~2
" ((i, i+l): i<n is even) ifn is even and r= L
Proof First, notice that D is a homomorphic image of D'. The homomorphism
h,: L,' ~is defined by setting h,(b;) = b, where b E {O, 1}. It follows that for all strings
x of length n, and all permutations <rES", h,(xa)=(h,(x))a, which in turn implies
that Sn(D') £: Sn(D). Now, if n is odd, then trivially S(D) = 1 and so S(Dr) = L
Suppose that n = 4, r = 2, and, respectively, write "(", "['', ") ", "]" in place of Oi. 0 2 ,
11' 12. Then ([])ED~, but(] [)~D~. Similar examples can be constructed to verify
that S(D') = 1 for 2 ~ r. To prove the theorem, it is enough to show that, for n even,

Sn(D)=((i, i+l): i< n is even).


For any string x = x 1 • • • xk let l(x) = k be its length and s(x) its signature, where
k
s(x)= L (-lf',
i=l

Then we can prove the following claims.


BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 561

CLAIM 1. For any string x, XE D~s(x) =O and for all i~ l(x)(s(x Ii) S;O).
Proof of Claim 1. The direction from left to right is trivial by induction on the
construction of x ED. To prove the other direction, assume the right-hand side is true.
We use induction on the length of x. If for some k < l(x), s(x I k) = O then x = (x I k)y,
for some y. Clearly, the induction hypothesis applies to x I k and y. Consequently,
both x I k, y ED and hence also x ED. Otherwise, for all k < /(x), s(x I k) > O. Clearly,
X1cxl = l (otherwise s(x) > 0). We also know that x 1 =0. Hence, x = Oyl, for some y.
Clearly, this y satisfies the induction hypothesis stated in the right-hand side of Claim
1. Hence, y E D and consequently also x ED.
As mentioned above, if n is odd the theorem is trivial. Hence, in all the proofs
below we assume that n is even.
CLAIM 2. For any b E {O, I} and any l < i < n there exists a string x E Dn such
that X; =b.
Proof of Claim 2. The proof is by induction on n. The claim is trivial if n = 2. So
assume n > 2. If i = 2 then consider the strings Oly, 001 lz E Dn. If i = n -1, then consider
the strings yOl, zOOl 1 E Dn. Hence, without loss of generality, we can assume that
2 < i < n -1. But then consider strings of the form Oyl, where y ED" _2 , and use the
induction hypothesis.
CLAIM 3. aE Sn(D)~a(l) = 1, u(n)= n.
Proof of Claim 3. Assume a(l) = i ?61. Consider an x E Dn such that X; = l (use
Claim 2). Then note that x(J" = Iy e D", for some stringy, which is a contradiction. A
similar proof shows that a( n) = n.
CLAIM4. IfuESn(D) ando-[{I,···,i-1} ]=[{1,···,i-1}] anda-(i)<i then
(a) i is even, (b) a-(i)=i+l, (c) u(i+l)=i.
Proof of Claim 4. To prove (a) assume on the contrary that i is odd. Consider an
x ED" such that x = yO · · · 1 z, where X; = 0 and x"uJ = 1 and s(y) = 0. Applying a- to
x we obtain that x" = y" 1 · · · . But then s(y" 1) = s(y") - 1 = s(y) - 1 = -1 0. Hence,
x" ~ D,,, by Claim 1, a contradiction.
To prove (b) assume on the contrary that a-( i) > i + 1. For simplicity, assume that
lT( i) = i + 2 (a similar proof will work if a( i) ~ i + 2). We distinguish several cases. If
a(i + 1) = i + l then consider the string x = yOOl l · · · E Dn, with l(y) = i - 2, X;-.i = X; = 0
and X;+i = X;+ 2 = 1. Then it is clear that x" = y"Oll · · · e Dn, a contradiction. If
a(i+l)=i+3 then consider the string x=yOOOlll · · ·EDn, with /(y)=i-2, X;_ 1 =
X; = X;+ 1 =0 and X;+ 2 = X;+ 3 = X;+ 4 = 1. Then it is clear that x" = yu011 · · · e D,,, a contra-
diction. If a( i + 1) > i + 3 then consider the string x = yOOll · · · 1 · · · E Dn, with l(y) =
i-2, X;_ 1 =X;=0 and X;+ 1 = X;+ 2 = Xau+ 11 = 1. Then it is clear that x" = y"Oll · ··ED,,,
a contradiction. Thus, we obtain a contradiction in all cases considered above. Hence,
<T(i) = i + 1. This completes the proof of (b ).
To prove (c) use an argument similar to (b). Indeed, assume on the contrary,
CT(i + 1) 7" i. It follows that a-(i + 1) ~ i + 2. If a(i + 1) = i + 2 then take x = yOOll · · · E
Dn, with X;-i = X; = 0, X;+ 1 = X;+ 2 == l. Ifwe apply a to x then we obtain x" = y"Ol 1 · · ·~
Dn, which is a contradiction. If a-(i + 1) = i + 3 then take x = y00101 · · · E Dn, with
x;_ 1 = x; = X;+ 2 = 0, xi+ 1 = X;+ 3 = 1. Ifwe apply u to x then we obtain x.,. = y"Oll · · ·ED,,,
which is a contradiction. In general, a similar proof works if u( i + 1) s i + 3. This
completes the proof of (c).
Now we are ready to complete the proof of the theorem. Let a ED,,. We know
that a(l) = 1. Let i 1 be minimal such that u(i 1) 7" i 1 and for all i < i1(a(i) < i1). By
minimality a-(i 1 ) > i 1 • It follows from Claim 4 that i 1 is even and a(i 1 ) = i1+1 and
O"( i 1 + 1) = i 1 • Let i 2 be minimal i 1 such that a( i 2 ) 7" i 2 and for all i < i2( u( i) i 2 ). By
minimality a-( i 2) i2 . Hence, Claim 4 applies again to show that i2 is even and a( i2) = i2 + l
562 P. CLOTE AND E. KRANAKIS

and a( i 2 + l) = i 2 . Proceeding in this fashion we show that Sn ( D) <;;::: ( ( i, i + 1): i n is


even). It remains to show that, in fact, equality holds. Indeed, let i < n be even. There
are four possibilities for X;X;+ 1 in the string x:
X 1 = yOO · · ·, X 2 = yOl · · ·,
where y is a string of odd length. But then it is easy to see that for all j = 1, 2, 3, 4,
x.J EDn ~ x(i,i+I)
J
ED
n'

which completes the proof of the theorem. 0


3.2. Palindrome language. The palindrome language is defined as the set of all
strings (in the alphabet l, with at least two elements) u = u 1 • • • un such that for all i
( U; = Un-i+l ).
THEOREM 4. If L is the palindrome then
uESn(L) ~ ('v'i~n)(a(n=n-a(i)+l)).
Moreover, S"(L) is isomorphic to S[n/Zl x (Z 2 )["/ 2 l.
Proof. (=>) Let aE Sn(L). Suppose that a(i) = j. Consider the string u = u 1 • • • un
such that u; = un-;+ 1 =0, and uk = 1, for all k ¥- i, n - j-1. Clearly, u E Lm. Hence, also
ua E Ln. It follows that Uaui = u1 =0 and consequently ua(n-i+I) = 0. But this is true
only if a(n - i+ 1) = n -j+ 1, as desired.
(~) This direction is obvious from the very definition of the palindrome.
To determine the group Sn(L), notice that by the previous result, a permutation
uES"(L), is determined by the values u(l), · · ·, a([n/2]). Furthermore, note that if
n is odd then a((n + 1)/2) = (n + 1)/2. Now consider the permutation a 0 such that for
all i~n, a 0 (i)=n+l-i and put Gn={aa0 aa 01 : aES[n; 2 i}. It is easy to see that G"
is isomorphic to S(n/zJ, moreover the group Hn generated by G" and the transpositions
( i, n - i + 1) is exactly the group
Gn X (1, n) x (2, n -1) x · · · X ([n/2], n -[n/2]-1).
Moreover, H" = Sn(L). This completes the proof of the theorem. 0
3.3. An algorithm for the invariance group of regular languages. Here we are
interested in studying the complexity of membership in the invariance group of a
regular language. To this end consider a term t(x, y) built up from the variables x, y
by concatenation. For example, t(x, y) = xyx, t(x, y) = x 2 yx 5 y 3 , etc. are such terms. The
number of occurrences of x and y in the term t(x, y) is called the length of t and is
denoted by Itl, e.g., Itl = 3 and Itl = 11, in the two previous examples. For any permuta-
tions u, r let the permutation t( a, T) be obtained from the term t(x, y) by substituting
each occurrence of x, y by a, r, respectively, and interpreting concatenation as the
product of permutations. We know that the symmetry group S" is generated by the
cyclic permutation en= (1, 2, · · ·, n) and the transposition T = (1, 2) (in fact any trans-
position will do) [Wie64]. A sequence a-= (an: n ~ 1> of permutations is term-generated
by the permutations en, T if there is a term t(x, y) such that for all n ~ 2, an = t( en, T ).
We have the following theorem.
THEOREM 5. (1) Let a-= (O"n: n;;;: 1) be a sequence of permutations which is term-
generated by the permutations Cn = (1, 2, · · · , n), T = (1, 2). Then for any regular language
L, La is also regular.
(2) For any term t of length Itl the problem of testing whether, for a regular language
L, L = L ", where a= {un: n;;;: 1) is a sequence of permutations generated by the term t
via the permutations c" = (1, 2, · · ·, n), T = (1, 2), is decidable; in fact it has com-
plexity 0(21' 1) .
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY
563

Proof Part (2) is an immediate consequence of the proof of part (1) and the
solvability of the equality problem for regular languages [Harr 78]. So we concentrate
only on the proof of (1). To prove the theorem we need the following claim, whose
proof is easy and left to the reader.
CLAIM.

LE REG~{x: Ox EL} E REG.


LE REG~{x: xl EL} E REG.
LEREG~{x: Oxl E L}eREG.
LEREG~{x: lxOEL}EREG.
First we show how to prove the theorem when un = (1, n). Indeed,

and this last set is the union of the following four sets:
{x E 2": Ox 2 • • • Xn- 10 EL}, {xe2": lx2 • • • Xn- 1 1 EL},
{x E 2n: Ox2 • • • Xn-11 EL}, {xE2": lx2 • • • Xn-10E L}.
This completes the proof in view of the above claim. A similar proof will yield the
result when each un = (1, 2). Next we use the above result for the transpositions (1, n)
to prove the result for the n-cycles, un =en. Indeed,
LE REG~{x 1 • • • Xn: xl E L}E REG

~{X1 · · · Xn: X1 · · · Xnl E L}EREG


~{X1 · · · x,,: 1X2 · · · XnX1 E L}E REG
~{X1 · • · Xn: X2 • • • XnXI EL} E REG.
Finally, the theorem follows by using the following product formula, which is valid
for any permutations T 1 , r 2 E Sn,

This completes the proof of the theorem. 0


The assumption on term generation of the sequence (un: n >- 1) of permutations,
made in the last theorem, is necessary as the following example shows.
Example 6. Let R be a recursively enumerable but nonrecursive set. Consider the
permutation un, which is equal to (1, n), if n ER, and is equal to id,,, if n e R, where
idn is the identity permutation on n letters. Consider the regular language defined by
L= 10*. Then it is easy to see that L~={lO": n+le R}U{O"l: n + 1 ER}. It follows
that n ER <:?0"- 11 E Lcr. Hence, La is not even a recursive language, although L is
regular.
4. Constructing languages with given invariance groups. This section is concerned
with the problem of realizing specific sequences of finite permutation groups by
languages Ls; {O, l}*. A language L is said to realize a sequence G = <Gn: n ~ 1) of
permutation groups Gn ~Sn if it is true that Sn(L) = Gn, for all n. We consider the
following types of groups and determine regular as well as nonregular languages
realizing them.
564 P. CLOTE AND E. KRANAKIS

Reflection. Rn= (p), where p(i) = n + 1- i is the reflection permutation,


Cyclic. Cn = ((1, 2, · · ·, n)).
Dihedral. Dn = Cn X Rn.
Hyperoctahedral. 0 n = (( i, i + 1): i is even ~ n).
THEOREM 7. (1) Each of the identity, reflection, cyclic (for n i' 3, 4, 5), dihedral,
and hyperoctahedral groups can be realized by regular languages.
(2) Each of the identity, cyclic, and dihedral groups can be realized by languages L
such that Le SIZE(n° 0 >).
Proof ( 1) For each of the above-mentioned types of groups we provide a regular
language realizing it.
Identity. This case is simple: take L = O*l*.
Dihedral. Let L = O* 1*O* U 1*O*1 *. It is clear that Dn s;;; S,, ( L). Let p be the reflec-
tion permutation defined by p( i) = n + 1- i and let u = (1, 2, · · · , n ). It is easy to check
that upu=p. It follows that Dn={ukp 1: k~n, l=O, 1}. Next we prove the following
claim.
CLAIM. For all TE Sn, if addition is modulo n,

or
Vi~ n( T(i) = T(i + 1) + 1).
Proof of the claim. From left to right the equivalence is easily verified for the
permutations ukp 1 (l~k~n, l=0,1). For example, u(i+l)=u(i)+l and p(i)=
p( i + 1) + 1. To prove the other direction, assume that T satisfies the right-hand side.
Say, T( 1) = k It is then easy to see that either T = uk-t or T = ukp. This completes the
proof of the claim.
It remains to show that Sn(L) s;;; Dn. If n ~3 the result is trivial. So assume that
n 6 4. Let Te Dn. There exists an i ~ n -1 such that \T( i + 1) - T( i)l 6 2. Let us suppose
that 1 ~ T( i) + 1 < T( i + 1) ~ n. Then we have that

Reflection. Let L=O*l*O*. It is clear that Rns;;;Sn(L). We want to show that


Sn(L) s;;; Rn. By the proof given in the case of dihedral groups we have that Sn(Ln) s;;; Dn.
Assume on the contrary that TE S,,(L), but TED,, - Rn- It follows that T = u;p, for
some i 61. Since p E S,,(L) we obtain that u; E Sn(L), which is a contradiction.
Cyclic. First assume that n = 2. Then consider the regular language
L=(OlU 10)0*1*
and notice that Sn(L) = (1, 2).
Next assume that n 6 6. Consider the regular language L= L 1 n L2 where L 1 is
the language
1*O*l * U O*l *O* U 101000*1U0*1101000* U 0*011010
u 0*001101u10*00110 u 010*0011
and L2 is the language
10*00101.
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 565

Clearly, C,, £ S,,(L). In view of the result on dihedral groups we have that S,,(L) £ D,,.
Let x = 101000"- 6 1 EL,,. Then xP = 10"-6 00101 ~ L,,, where p(i) = n + 1- i. Hence, C,, =
S,,(L), for n ~ 6.
It is interesting to note that for 3 ~ n ~ 5 the groups C,, are not representable. This
is obvious for n = 3, since C 3 = A3 • For n = 4, 5, one can show directly that for any
boolean function f E B,,, if C,, £ S(f) £ D,, then S(f) = D,,.
Hyperoctahedral. Consider the language L consisting of the set of all finite strings
x = (x1, · · ·, xk) such that for some i ~ k/2, x 2i-I = x2 ;. The regularity of the language
follows from the obvious equality

For any set I= { i,j} of indices, let f 1 be the n-ary boolean function defined by

= {~ if X; ,C. X;.
if x; =xj
f,(x)

Put m = [n/2]. For each i = 1, · · ·, m consider the two-element sets I;= {2i -1, 2i}
and the functions f,; defined above. Consider the boolean function
f = f,, v ... v f,..,.

It is then clear that S,,(L) = S(f). It is also easy to see that this last group consists of
all permutations a-ES,, which permute the blocks I;, i = 1, · · · , m. In fact this last
group has exactly 2[" 121 • [n/2]! elements.
To prove part (2) of the theorem we use Lupanov's theorem (see § 2.3), i.e.,
j{f E B": c(f) < q}I = O(q 4 + 1).
Identity. By Lupanov's theorem we have that
J{f E B,,: c(f) ~ nlog11}J = 2 01n"'""ClogII1' I« 22"

- l{f E B,,: S(f) = l}J.


It follows that for all but a finite number of n there exists f,, E B,, such that L(f,,) ~ n 10s"
and S(f,,) = 1. If we define a language L such that for all n, L,, =f,,, then the proof is
complete.
Cyclic. The result will follow by a proof similar to the above if we could prove that
(3)
Indeed, the left part of the above inequality is true because one may independently
assign a value of 0, 1 to each orbit, except for orbits of words having 2 or 3 occurrences
of the symbol 1. Let a-= ( 1, 2, · · · , n) be the n-cycle and let p be the reflection on n
letters. We agree to have f( v) ~ f( w ), where JvJ 1 =Iwl 1 = 2 and

This removes n. Choose 2 independent choices while adding one choice of 0 or 1. We


agree to havef(v);t.f(w), where lvl 1 =lwl1=3 and
v E {(101000"- 6 1)'": 0 ~ i ~ n -1}, w E {(10"- 6 00101)'": 0 ~ i ~ n -1}.
Again, this removes n. Choose 2 independent choices while adding one choice of O
or 1. Hence, the proof of the desired lower bound ( 1) is complete.
566 P. CLOTE AND E. KRANAKIS

Dihedral. By [Ber71, p.171], 6(D11 )~r- 1 /n. An argument similar to the one for
cyclic groups used above shows that
I{/ E Bn: S(/) =D,,}I ~ 22"-'/n-nl n-1)/2 » 2 O(nlog"/ n(logn)2).

This completes the proof of the theorem. 0


There is another interesting way for realizing the cyclic groups C 11 , for n ~ 4. For
any groups G, H,put[G, HJ ={g. 1 h- 1 gh: g E G, h E H}.Let G, H~S 11 be two permuta-
tion groups. Consider the set of words in G* defined by
Lo,H = { w E G*: w EH}.
(The reader should be warned of the different interpretation of w in the expressions
w E G* and w EH; the former is a word in G* and the latter is an element of a group.)
THEOREM 8. For any permutation groups G, H ;;£Sn, if [ G, G] is not a subset of
the normal subgroup H of G, then Sn ( L 0 , H) = C,,, for n ~ 4.
Proof First we show that Cn s S7,(Lo,H ). Indeed, consider the cyclic permutation
Cn =(1, 2, · · ·, n) and notice that for w = u 1 • · • lTn E G*,

It follows from the normality of Hin G that en E S7,(Lo,H ). This completes the proof
of en s s;(Lo,H)· Next we prove that Sn(Lc;,H) s en. Indeed, let p be a permutation
in S 11 - Dn. It follows from the proof of Theorem 7 that either (A) there exists an i
such that lp(i+l)-p(i)lmodn>l, or (B) lp(n)-p(l)lmodn>l. We show that
p E S 11 (Lc;,H ). First we consider case (A) and distinguish four sub cases.
Case 1. l~p(i)<p(i+l)n.
Let u, 7 be given such that [cr,7]=cr7cr- 1 T- 1 eH. Let j=p- 1(p(i)+l), k=
p- 1(p(i+l)+l). Consider w = cr 1 • • • cr11 E G", where lT; = u, lT;+ 1 = u-1, u1 =7, crk = T-· 1 ,
and all other u/s are equal to 1. Then we have that w = uu - 1 TT - l or uu - 1 T · 1 7 depending,
respectively, on whether or not}< k or k <j. In either case w = 1, but wP = lTTlT- 1 7- 1 e H.
Case 2. p(i)<p(i+l);;£n.
Let cr,T be given such that [cr,7]=uw- 1 T- 1 eH. Letj=p- 1 (p(i)-1) and k=
p- 1(p(i) + 1). Choose w such that w = cr 1 · · · lTn E G", where u1 = cr, lT;+ 1 = 7- 1, lT; = 7,
fh = u - 1 and all other a/s are equal to 1. Then it is clear that w = 1, while wP e H.
Case 3. 1~p(i+1) < p(i) < n. This is similar to case 1.
Case 4. l < p( i + 1) < p( i) ~ n. This is similar to case 1.
Case (B) is handled exactly as before. Hence, we have proved that S 11 (Lo,H) s Dn.
It remains to show that in fact S 11 (Lc;,H) = Cn. Since [ G, G] is not a subset of H, G / H
cannot be abelian. Therefore, there exist elements g 1 , g2 , g 3 , g 4 E G such that
gig2g3g4E H, but g4g3g2g1E H.
It follows that the reflection permutation does not belong to S11 (Lo,H ), which completes
the proof of the theorem. 0
Given a language Ls 2.* over the alphabet 2. the syntactic semigroup GL of L is
defined as follows. Define w = w' mod L if for all u, v E 2.*, uwv EL<.;;> uw' v E L. Then
let GL be the quotient of 2.* modulo the equivalence relation= mod L. Recall that the
Krohn-Rhodes theorem [Arb69] states that the syntactic semigroup GL of any given
regular language L is the homomorphic image of a wreath product of cyclic simple
groups, noncyclic simple groups, and three particular nongroup semigroups called
"units." If G is abelian and H = 1, then it is clear that S" (Lo,H) = Sw If G is a
nonabelian group and H = 1, then Theorem 8 yields that S 11 (Lo,H) = Cn. We have seen
families of these groups as invariance groups of regular languages. However, we have
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 567

examples of representable groups whose homomorphic image is not representable,


(e.g., (1, 2, 3) is the homomorphic image of (1, 2, 3 )( 4, 5, 6) ), thus indicating that it is
unlikely that the Krohn-Rhodes theorem can be used to characterize those families
of invariance groups of regular languages. Similarly, from the examples given in the
paper, there is no invariance group structure preserved when taking regular operations:
from Sn(L) and Sn(L'), we cannot say anything in general about Sn(M), where
M = L* L' and * is a boolean operation or language concatenation or where M = L *
(Kleene star). This blocks a natural attempt to inductively define the families of
invariance groups of regular languages.
It is not known whether there is a characterization of those sequences of groups
which can be realized by regular languages. However, it is interesting to note that for
regular languages L the invariance group S 2 n(L) can never be equal to the {1, 2, · · ·, n}
point-stabilizer of S 2 n.
THEOREM 9. (1) There is no regular language L such that for all but a.finite number
of n we have that

(2) There is a regular language L such that for all n we have that

S2n(L) = (S2n){2i;i,,;;n/2}·
Proof ( 1) By the pumping lemma for regular languages [Harr78] there exist words
a;, b;, i < m and iiJ, ~, j <m and languages L;, ij such that
L= U a;bf L;,
im

where 1L = {O, 1}* - L is the complement of L. Let r be the least common multiple of
the lengths of all the above words. Put i = r + 1, j = i + r, and n0 = 3 r. Consider the
transposition r = (i,j) and let n ~ n 0 • Then for any word w of length n we consider
the following two cases.
Case l. w E Ln.
Then for some i 0 < m and some s we have that w must be of the form a; 0 0 c; 0 • b:
The ith position in the word w falls within the block b; 0 • Since the length of b; 0 divides
r the jth position of the word w falls in exactly the same position with respect to the
block b; 0 • It follows that W; = w1 and hence wr = w.
Case 2. w i. Ln.
This is similar to the proof of Case 1.
It follows from the above that TE Sn(L), as desired. This completes the proof of
part (1 ).
(2) Consider the languages L'=O* and L"= l*O*. It is clear that for all n, Sn(L') =
S and S,,(L") = 1. Let L be the set of all words w of even length 2n such that
11 ,

Clearly, L is a regular language and S 2 ,, (L) ;:;;:> (S 2 n ){ 2 i: i"2n/ 2 l. It remains to show that
in fact S 211 (L)s;:; (S 2 ,,) 12 i,;""n; 2}· Indeed, let U'ES 2 n(L) and decompose U' as a product
of the disjoint cycles a 1 • • • U'k· Assume on the contrary that there exists an io such
that U';o = (a1' . .. 'ar) and
(i) either there exists a 1 ~}0 < r such that a10 is even and aj0 +1 is odd,
(ii) or ar is even and a 1 is odd.
We treat only case (ii), the other case being entirely similar. Consider a word w
defined as follows. Let w1 = w3 = · · · = w2 n-i = 0 and w2 = W4 = · · · = Wa, = 1 and the
568 P. CLOTE AND E. KRANAKIS

remaining W; 's equal to 0. Then w EL. However, ( wu)a, = 1, where a 1 is odd, and so
(w")1(w"h · · · (w.,.hn-1e L'.
It follows that Wu e L. Hence, (}" e S2n(L), a contradiction. 0
5. Representations of permutation groups. The aim of this section is to give general
results on permutation groups G ~Sn which can be represented as the invariance
groups of boolean functions, i.e., G = S(f) for some f E Bn. It will be seen in the sequel
that there is a rich class of permutation groups which are representable in this way.
The main motivation for the results of the present section is the simple observation
that the alternating group A" is not the invariance group of any boolean function
f E Bn, provided that n ~ 3. Although this will follow directly from our representation
theorem it will be instructive to give a direct proof. Suppose that the invariance group
off E Bn contains An. Given x E 2n, for 3 ~ n, there exist 1 ~ i <j ~ n such that X; = Xj.
It follows that the alternating group An and a transposition fix f on x, and hence Sn
does as well. As this holds for every x E 2n, it follows that S(f) =Sn. In fact it is clear,
using part (1) of Theorem 1, that An is not isomorphic to the invariance group S(f)
of any f E Bn. However, An is isomorphic to the invariance group S(f) for some boolean
function f E B,, 00 gn+o (see Theorem 11 below).
One can generalize the notion of invariance group for any language Ls;;
{O, 1, · · ·, k}* by setting Ln = L n {O, · · ·, kf and S(Ln) to be

We leave the details of the proof of the following fact as an exercise for the reader.
FACT. For all n, there exist groups Gn ~Sn which are strongly representable as
Gn = S(Ln) for some Ls;; {O, 1, · · ·, n -1}" but which are not so representable for any
language L' s;; {O, 1, · · · , n -2}".
Proof The alternating group A,,= S(L,, ), where L,, = {w E {O, · · ·, n -1r: u..,,. EA,,},
where u"': i ~ w( i -1) + 1. By a variant of the previous argument, A,, is not so represent-
able by any language L' s;; {O, 1, · · · , n - 2}". 0
Compared to the difficulties regarding the question of representing permutation
groups G ~ S,, in the form G = S(f ), for some f E B,., it is interesting to note that a
similar representation theorem for the groups S(x) = {u E Sn: xu = x}, where x E 2", is
relatively easy. It turns out that these last groups are exactly the permutation groups
which are isomorphic to sk x s,,_k for some k. Indeed, given x E 2n let
X={i: 1~i~n and X;=O}, Y={i: l~i~n and X;=l}.
It is then easy to see that S(x) is isomorphic to Sx x SY· In fact, u E S(x) if and only
if Xu= X and Y" = Y.
5.1. Elementary properties. Before we proceed with the general results we will
prove several simple observations that will be used frequently in the sequel. We begin
with a few useful definitions. For any f E B", let s+ (f) = { u E Sn: for all x E
2"(f(x) = 0=? f(x") = O)}. For any permutation group G ~Sn and any~ s;; {l, 2, · · ·, n}
let G:.. be the set of permutations u E G such that (for all i E ~)( u(i) = i). G.:;. is called
the pointwise stabilizer of G on~. Notice that (S,,){k+t.··-,nl= Sk. fork~ n. For any
permutation u and permutation group G let G" = u- 1 Gu, also called a conjugate of
G by u. For any f E 8 let lE8f E B,, be defined by (1E8f)(x) = lEBJ(x), for x E 2".
11

If f 1 ,···,fk.EB,, and fEBk then g=f(fi,···,fk)EB,, is defined by g(x)=


f(f;(x), · · · ,fk.(x)). The following theorem contains several useful observations that
will be used frequently in the sequel.
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 569

THEOREM 10. ( 1) If f E Bn is symmetric then S(f) =Sn.


(2) S(f) = S(l EBJ), for all f E Bn-
(3) For any permutation er, S(fu) = S(f)u.
(4) ForeachfEBn, S(f)=S+(f).
(5) If Ii. ... ,fk E Bn and f E Bk and g = f(fi, ... ,fk) E Bn then SCf1) n ... n
S(fk) c;; S(g).
(6) (For all k;:£n)(3fEBn)S(f)=Sk.
Proof. The proofs of (1)-(3), (5) are easy and are left as an exercise to the reader.
To prove (4), notice that S+(f) is a group and trivially S(J) s:; S+(f). Now let er E S+(f)
and suppose thatf(x(T) = 0 holds. Since, o-- 1 E s+(f) we have thatf(x) = f((x")lT- 1) = 0.
It follows that s+(f) <;; S(f), as desired. To prove (6) we consider two cases. If k+ 22 n,
define f by

f(x)={l ifxk+1~Xk+ 2 ;:£· · ·2xn


0 otherwise.
Let aES(f). First notice that for all i> k(cr(i)> k). Next, it is easy to show that if er
is a nontrivial permutation then there can be no k ~ i <j ~ n such that cr(j) < cr(i).
This proves the desired result. If k = n - 1, then the function! must be defined as follows.

f(x)={l ifx1,·: · ,Xn-1~Xn


0 otherwise.

A similar proof will show that S(f) = Sn-i · This completes the proof of the
theorem. D
We define a permutation group G ;::£Sn to be representable (respectively, strongly
representable) if there exists an integer k and a function f E Bn,k (respectively, with
k = 2) such that G = S(f). G ~Sn is called weakly representable if there exists an integer
k, an integer m < n, and a function f: mn-? k such that G = S(f). It will be seen in
the sequel (representability theorem) that the distinction between representable and
strongly representable is superfluous since these two notions coincide.
Notice the importance of assuming m < n in the above definition of weak rep-
resentability. If m = n were allowed, then every permutation group would be weakly
representable. Indeed, given any permutation group G ~Sn define the function f as
follows:

(here, we think of (x 1 , • • · , Xn) as the function i ~ xi in n") and notice that for all
er E Sn, er E S(J) if and only if for all TE Sn (TE G~ rcr E G). Hence G = S(f), as
desired.
Another issue concerns the number of variables allowed in a boolean function in
order to represent a permutation group G 2 Sn. We can also consider representing
functions by using additional variables, but as the following theorem shows, every
group becomes representable if enough variables are allowed.
THEOREM 11 (Isomorphism Theorem). Every finite permutation group G ;::£Sn is
isomorphic to the invariance group of a boolean function f E Bnoogn+I>.
Proof. First, let us give some notation. Let w be a word in {O, l}*. lwl 1 is the
number of occurrences of 1 in w, and wi is the ith symbol in w, where 1 2 i ;::£I wl =length
of w. The word w is monotone iffor all 12 i <J2 lwl, w; = l==>wj = 1. The complement
of w, denoted by w is the word which is obtained from w by "flipping" each bit wi,
570 P. CLOTE AND E. KRANAKIS

i.e., \w\ = \w\ and W; = 1 EB W;, for all l ;;:;ii ;;:;i \w\. Fix n and let s =log n + 1. View each
word w E {O, 1 ys (of length ns) as consisting of n-many blocks, each of length s, and
let w(i) = w(i-JJs+i · • · W;s denote the ith such block. For a given permutation group
G ;;:;i Sn let L 0 be the set of all words w E {O, l}"s such that
• either (i) Iw\ 1 = s, and if the word w is divided into n-many blocks
w(l), w(2), · · · , w(n), each of lengths, then exactly one of these blocks consists of
l 's, while the rest of the blocks consist only of O's,
• or (ii) \w\ 1 ;;:;is -1 and for each 1 ;;:;ii ;;:;in, the complement w of the ith block of w is
monotone (this implies that each w( i) consists of a sequence of l 's concatenated
with a sequence of O's),
• or (iii) lw\ 1 6 n and for each 1 ~ i~ n, w(i) 1 =0 (i.e., the first bit of w(i) is O) and
the binary representations of the words w(i), say bin( w, i), are mutually distinct
integers and <Tw E G, where aw: {l, · · ·, n}~ {l, · · ·, n} is the permutation defined by

<Tw(i) = bin(w, i).

The intuition for items (i) and (ii) above is the following. The words with exactly
s-many l 's have all these l 'sin exactly one block. This guarantees that any permutation
"respecting" the language L 0 must map blocks to blocks. By considering words with
a single 1 (which by monotonicity must be located at the first position of a block) we
guarantee that each permutation "respecting" L 0 must map the first bit of a block to
the first bit of some other block. Inductively, by considering the word with exactly
( r - 1)-many l 's, all located at the beginning of a single block, while all other bits of
the word are O's, we guarantee that each permutation "respecting" La must map the
( r- l )st bit of each block to the ( r -1 )st bit of some other block. It follows that any
permutation respecting L 0 must respect blocks as well as the order of elements in the
blocks; i.e., for every permutation TE S 11s(L0 ),

('v'O;;:;i k< n)(30~ m < n)('v'l ~ i~ n)'T(ks+ i) =ms+ i.

Call such a permutation "s-block invariant." Given a permutation TE Sn,(La) let 7 E Sn


be the induced permutation defined by

i(k) = m ~ ('v'l ;;:;ii~ n)T(ks+ i) =ms+ i.

We claim that G={7: TES7.s(L 0 )}. Indeed, to prove(<;) notice that every element 7
of G gives rise to a unique "s-block invariant" permutation T. If w E L 0 and Iwl 1 ~ s,
then bys-block invariance of T, W E La. This proves(<;). If w E L 0 and <Tw E G, then
7

a-< w • 1 = <T,,.i E G (composition is from the right). To prove ( 2) let w E L 0 be such that
<Tw is the identity on S". Then for any TES 11s(Lo), W 7 ELa, so <T<w'J=awi'=i'EG,
which proves the above claim. This completes the proof of the theorem. 0
Clearly, the idea of the proof of the previous theorem can also be used to show
that for any alphabet l, if L r;;. l", then Sn (L) (the set of permutations in Sn "respecting"
the language L) is isomorphic to Sn,(L'), for some L' !;:; {O, iys, where s = l +log Ill.
We conclude by comparing the different definitions of representability given above.
THEOREM 12. For any permutation group G~Sn the following statements are
equivalent:
(1) G is representable.
(2) G is the intersection ofa finite family of strongly representable permutation groups.
(3) For some m, G is a pointwise stabilizer of a strongly representable group over
Sn+m, i.e., G = (Sn+m(f)){n+!,···,n+m), for some f E Bn+m and m ~ n.
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 571

Proof First we prove that (1)=>(2). Indeed, let fe Bn,k such that G=S(f). For
each b < k define as follows a 2-valued function fb: 2" ~ {b, k}:

h(x)={~ if f(x) = b
if f(x) ;t. b.
It is straightforward to show that
S(f) = S(fo) n · · · n S(fk-1).
But also, conversely, we can prove that (2) => ( 1). Indeed, assume that .fi, E Bn, b < k,
is a given family of boolean valued functions such that G is the intersection of the
strongly representable groups S(fb). Define f E Bn, 2 k as follows:
f(x) = <fo(x), · · · ,fk-1(x)),
where for any integers n 0 , • • ·, nk-i. the symbol (n 0, • · ·, nk_ 1) represents a standard
coding of the k-tuple (n 0 , • • · , nk_ 1 ). It is then clear that S(f) = S(f0 ) n· · ·
n S(fk_ 1 ),
as desired.
To prove that (3) is equivalent to statements (1) and (2) it is enough to show that
(i) for any family {f;: 0 ;a; i ~ k} of boolean functions f; E B" there exists an integer
0 ~ m ;a; log k and a boolean function f E Bn+m such that
(4) (S(f)){n+l,-·-,n+m} = SU1) n ... nS(fk),
and (ii) also conversely, for any integer m;;; 0, and any boolean function f E Bn+m
there exist boolean functions {f;: 0 ;a; i ;a; k}, with k ~ zm such that equation (1) holds.
Indeed, part (i) of the above statement follows by repeated application of part
(6) of Theorem 10 and the case k = 2 of the above statement. To prove the case k = 2,
define f(x 1 , • • • , Xn, i) =f;(x 1 , • • • , Xn ). The desired equality is now easily proved. To
prove the converse part (ii), let m, f be as in the hypothesis and define the desired
family of functions fb,,-·-,b,,. as follows.
h,,. ..,bJX1, · · ·, Xn) = f(X1, · · ·, Xn, b1, · · ·, bm).
It is now easy to see that equation ( 1) is satisfied. This completes the proof of the
theorem. D
5.2. Representation theorems for general permutation groups. Here we study the
representability problem for general permutation groups, give a necessary and sufficient
condition via P6lya's cycle index for a permutation group to be representable, and
show that the notions of representable and strongly representable coincide. In order
to state the first general representation theorem we define, for any n + 1 ~ 0 ;a; 2" and
any permutation group G~S", the set G~"> = {M;;a; G: ®"(M) = O}. Also, for any
H s; S", and any g E Sn, the notation (H, g) denotes the least subgroup ofS" containing
the set HU {g}.
THEOREM 13 (Representation Theorem). The following statements are equivalent
for any permutation groups H < G~Sn.
n
( 1) H = G K, for some strongly representable permutation group K ;a; Sn.
n
(2) H = G K, for some representable permutation group K ~S,..
(3) (for all g E 0-H)(@n((H, g)) < 0,.(H)).
(4) His maximal in G~"\ where <E>n(H) = 0.
Proof. We prove the equivalence of the above statements by showing the following
sequence of implications: (1)=>(2)=>(3)=>(1) and (4)=>(3)=>(4). The proof of (1)=>
(2) is trivial. First, we prove (2)=>(3). By Theorem 12, K is the intersection of a family
572 P. CLOTE AND E. KRANAKIS

of strongly representable groups. Hence, by assumption let S(/;), where {/;} s;; Bn, be
a finite family of invariance groups such that
H =n s(_t;) n a.
Assume on the contrary that there exists an HK~ G such that 0(K) = @(H). This
last statement is equivalent to the statement
'Vx E 2n (xK = xH).
We show that in fact
K s;; n S(f;) n G,
which is a contradiction, since the right-hand side of the above inequality is equal
to H. Indeed, let u E K and x E 2n. Then we know that
XK = (x")K = (x")H.

It follows that x = (x'T, for some TE H. Consequently, f;(x) = /;((x"f) = f;(x,,.), as


desired.
Next we prove that (3)~(1). Let Pn(X) be the property of subgroups stated by
X ~Sn i\ (for all L > X)(0n(L) < 0n(X)). (When n and X ~Sn are clear from context,
we say simply that X satisfies property P.)
CLAIM. For all n and subgroups X of Sn,

Pn(X) ~ X is strongly representable.

Proof As the direction from right to left is obvious, we only consider the direction
from left to right. Suppose, in order to obtain a contradiction, that this direction fails.
Let X ~Sn be of maximal size such that Pn(X) holds, but that X is not strongly
representable. It follows that
(V L > X)(L satisfies P=::?) Lis strongly representable).
Since the full symmetric group Sn is strongly representable we can assume, without
loss of generality, that X <Sn- In particular, there is a strongly representable group
L > X of minimal size. Leth E Bn be such that L = S(h). Thus,
(*) V M(X < M < L=::?) M does not satisfy P).
Since P,,(X) holds, we have that 0,,(L)<0,,(K). It follows that there exist x,yE2"
such that
x= ymod L, x#ymodX,
where for H ~ S,, and x, y E 2" the symbol x = y mod H means that y = xo-, for some
u E H. Define a boolean function g E B,, as follows, for w E 2n,

h(w) ifw#xmodX, w#ymodX


g( w) = { 0 if w = x mod X
1 if w = y mod X.

It follows from the definition of g that X ~ S(g) < S( h) = L. Since every strongly
representable group satisfies property P, an immediate consequence of (*) is that
X = S(g ). This completes the proof of the claim. D
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 573

Now returning to the proof of (3)~(1), by assumption, for all gE G-H,


2® .. «H.z>l < 2® .. <H>. In particular, for all g E G- H, there exists a boolean functionfg E Bn
such that H ~ S" (.fg ), but (H, g) is not a subset of Sn(fg ). Consider the representable
group K defined by

K = n S(fg).
gEG-H

It is now trivial to check that H =Kn G. Moreover, as in the implication (2)~(3)


above, it follows that the permutation group K satisfies property P. By the above claim,
K is strongly representable. This concludes the proof (3)~(1).
It remains to prove the equivalence of the last statement of the theorem. First we
prove (4)~(3). Assume that H is a maximal element of G\:l, but that for some
g E G-H, we have that 0n((H, g)) = 0"(H). But then H < (H, g) ~ G, contradicting
the maximality of H. Finally, we prove (3)~(4). Assume on the contrary that (3) is
true but that His not maximal in G~"l. This means there exists H < K ~ G such that
0n(K) = 0,,(H). Take any g EK - H and notice that
0"((H, g)) ~ 0,,(K) = 8 = 0,,(H) ~ 0,,((H, g)).
Hence, 0,,(H) = 0,,((H, g)), contradicting (3). 0
A "naive" algorithm for testing the representability of a general permutation group
G ~ S,, is to test all boolean functions f E B,, to see if G = S,, (f).Clearly, this requires
time 22 ". An immediate consequence of the representation theorem is the following
algorithm whose running time is O((n!)2)=2°<niognl.
Algorithm for Deciding the Representability of Permutation Groups Input
A permutation group G ~ S 11 •

for each uE S,, -G do


if 0,,((G, u)) =0,,(G)
then output G is not representable.
od
else output G is representable.
end
The well-known graph nonisomorphism problem (NGIP) is related to the above
group representation problem. Indeed, let
G = ({v 1 , • • ·, v,,}, Ea),
be two graphs on n vertices each. Consider the permutation group ISO( G, H) ~ S,,+ 3
whose generators u satisfy:
'v'l ~ i,

and in addition the permutation n + i ~ u(n + i), i = 1, 2, 3, belongs to the group


C3 =(n+1, n + 2, n + 3). It is easy to show that if G, H are isomorphic, then there
exists a group K ~ S,, such that ISO( G, H) = K x C 3 • On the other hand, if G, H are
not isomorphic, then ISO( G, H) = (id,,+ 3 ). As a consequence of the nonrepresentability
of C3 , and the representability theorem of direct products, it follows that G, H are
not isomorphic if and only if ISO( G, H) = (id,,d.
Remark. An idea similar to that used in the proof of the representation theorem
can also be used to show that for any representable permutation groups G < H ~ S,,,
2· l{heB,,: H=S(h)}l~j{geB,,: G=S(g)}I.
574 P. CLOTE AND E. KRANAKIS

Indeed, assume that G, H are as above. Without loss of generality we may assume
that there is no representable group K such that G < K < H. As in the proof of the
representation theorem there exist x, y E 2" such that x = y mod H, x ¥ y mod G. Define
two boolean functions hh E B", b = 0, 1, as follows from w E 2",

h(w) ifw ¥ x mod G, w¥ymod G


{
hh(w)= ~ ifw=xmodG
b if w = y mod G.

Since G ~ S(hb) < S(h ), it follows from the above definition that each h E Bn with
H = S(h) gives rise to two distinct hb E Bn, b = 0, 1, such that G = S(hb). Moreover, it
is not difficult to check that the mapping h ~ {h 0 , h 1}, where H = S(h), is 1-1. It is
now easy to complete the proof of the assertion.
An immediate consequence of the representation theorem is that all cycle indices
0"( G) can in fact be realized by representable permutation groups. The previous
theorem also has a consequence concerning the representatjon of "maximal" permuta-
tion groups.
THEOREM 14 (Maximality Theorem). (1) If His a maximal proper subgroup of
G~Sn then

(2) All maximal subgroups of Sn are strongly representable, the only exceptions
being: (a) the alternating group A", for all n ~ 3; (b) the I-dimensional, linear, affine
group AGL 1 (5) over the.field offive elements, for n = 5; (c) the group of linear transforma-
tions PGL 2 (5) of the projective line over the field offive elements, for n = 6; ( d) the group
of semilinear transformations Pf L 2 (8) of the projective line of the field of eight elements,
forn=9.
Proof To prove ( 1) let H be a maximal proper subgroup of G such that 0" ( G) <
E>n(H). Put () = E>n(H). Since condition (4) of the representation theorem is satisfied,
H is of the form S(f), for some f E Bn. This completes the proof of(=:;>). To prove
the other direction, assume that E>n( G) = 0n(H). Then for all g E G- H, E>n((H, g)) =
0" ( H). Hence, again by the representation theorem, there is no f E B" such that
H = G n S(f). This completes the proof of (1 ).
To prove (2) let M be a maximal subgroup of Sn. We distinguish two cases.
Case 1. E>n(M)>n+l.
In this case, part (1) of this theorem implies that M is strongly representable,
since E>n(S") = n + 1. (Note that by Theorem 2(4), the condition of Case 1 is satisfied
by all intransitive groups M, i.e., groups with wn(M) ~ 2.)
Case 2. 0n(M) = n+ 1.
In this case we know from the main theorem of [BP55] that M is of one of the
forms in the statement of the theorem. D
As noted above, all maximal permutation groups with the exception of An are of
the form S(f ), provided that n ~ 10. Such maximal permutation groups include: the
cartesian products SkxSn-k (k~n/2), the wreath products SdS 1 (n=kl, k,l>l),
the affine groups AGLd ( p ), for n =pd, etc. The interested reader will find a complete
survey of classification results for maximal permutation groups in [KL88]. It should
also be pointed out that there are plenty of nonmaximal permutation groups which
are not representable. In fact, it can be verified that examples of such groups are the
wreath products G l An. In general we can prove the following theorem. For any
permutation groups G~Sm, H~Sn.
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 575

THEOREM 15. Let G~Sm, H~Sn. Then


(1) G and H representable==}G I His representable.
(2) G l His representable==:} H is representable.
(3) G I His representable and 2n < m==} G is weakly representable.
(4) For p prime, a p-Sylow subgroup P of Sn is representable~ p ,c 3, 4, 5.
Proof ( 1) Suppose we are given two representable groups G = S(L0 ) ~Sm, H =
S(LH) ~Sn, where Lo£ {O, l}m, LH £ {O, lf. We want to show that the wreath product
G ! H ~ Smn is representable. The wreath product G l H consists of all permutations
p = [CT; T i , · · · , Tm], where CTE G and T i . · · · , Tm EH, such that

p((k- l)n + i) = CT(k)n + Ta-(kJ(i),


for 1 ~ k ~ m, 1 ~ i ~ n. (Intuitively speaking, p acts on m x n matrices in such a way
that T; acts only on the ith row and <r permutes rows.) Without loss of generality we
can assume that om, 1m E La and on, l" E LH. Define a set L £ {O, l}mn of words w by
the disjunctoin of the following three clauses:
(a) lw/1 = n, and for Some 0~ k < m, Wkn+I = ... = Wkn+n = 1 (i.e., the (k+ l)st row
consists only of 1's).
(b) Iwl 1 > n, and w is of the form e~ e; · · · e;:,, where the word e 1 e2 • • • em e La.
(c) lwl 1>n and w is not of the form e~e; · · · e::,, but wkn+I • • • Wkn+nELH, for
all o~ k< m.
We claim that Sm (L) = G I H. Indeed, the inequality G I H £ Smn(L) is clear. To
11

prove the other direction assume that p E Smn(L). By clause (a), p respects then-blocks
of words of length mn. Hence, p is of the form p = [u; T 1 , • • • , Tm], and T; E Sn, u E G,
where i = 1, · · · , m. If <re G, then there is a word v of length m, with v E La and
v,,. e La. Then (using clause (b) above) we have that w = v~ v; · · · v;:, EL, but wP t L,
which is a contradiction. If for some i, T; e H, then there is a word v of length n such
that v E LH and v e LH. It follows (by clause ( c) above) that the word w = v · · · v e L,
7
'

but wP e L, a contradiction. This completes the proof of (1).


(2) By assumption, GIH=Smn(f), for somefEBmn· Hence,
G IH ={[ u; T1,. •. ' T mJ E Sm I Sn: (V Xi. ... , Xm )f(X~(l)' .. ., x;;m))
= f(Xi. · · ·, X"')}.
In particular, we have that
TE H ~ [ idm; T, idn, · · · , idn] E G ! H
~ 'VX1['VX2, · · ·, X,,,(fx 2 , •••• xJX~) =fx2 •••.• xJX1))]

~ TE n . S(fx2 .... ,xJ,


X2.···,XmE2

as desired.
The proof of (3) is similar and uses the simple observation that for any permu-
tation uESm,
[u; idn, . .. 'idn] E G 11 ~ (V Xi. ... 'Xrn)f(Xu(l)' ... , Xu(m)) = f(Xi. . .. 'Xm).
( 4) Let p be a prime p ~ n. By Sylow's theorem, all the p-Sylow subgroups of Sn
are conjugates of one another. Moreover, by [ Pas66, pp. 8-11 ], if C is the cyclic group
(1, 2, · · · , p ), then there exists an integer r such if we iterate the wreath product r
times on C then the group CIC · · · I C obtained is a p-Sylow subgroup of Sn.
Combining this with the previous assertions of the theorem, as well as part (3) of
Theorem 10, we obtain the desired result. 0
576 P. CLOTE AND E. KRANAKIS

The converse of part (1) of the above theorem is not necessarily true. This is easy
to see from the following example. We show that the wreath product A3 1 S2 is
representable, but that A3 is not. Indeed, consider the language
L = {001101, 010011, 110100, 001110, 100011, 111000} s 26 •
We already proved that A3 is not representable. We claim that A3 1S 2 = S 6 (L). Consider
the three-cycle 7=({1,2}, {3,4}, {5,6}). It is easy to see A3 1S 2 consists of the 24
permutations()" in s6 which permute the two-element sets {l, 2}, {3, 4}, {5, 6} as in the
three-cycles 7, 72 , 7 3 • A straightforward (but tedious) computation shows that S6(L)
also consists of exactly the above 24 permutations.
Another class of examples of nonrepresentable groups is given by the direct
products of the form A"' x G, G x A,,,, where G is any permutation group acting on a
set which is disjoint from {1, 2, · · ·, m}, m;:::;; 3 (for a proof of this, see the next
subsection).
We conclude this section by showing the representability of the normalizers of
groups G generated by a family of "disjoint" transpositions. Let G be a subgroup of
Sn and let H = (H (x): x E 2") be a family of normal subgroups of N( G) (the normalizer
of Gin S,,) such that for all uE N( G), x E 2", H(x) = H(u(x)). (This last condition
is satisfied if, for example, each H(x) = 1 or each H(x) =G.) For any x E 2" let
Gx = {u E G: x" = x} be the stabilizer of G at x. Define the function f o,H : 2" _,,. 2 as
follows:
l ifG,=H(x)
{
fo,H(x) = 0 if G, .,c. H(x).
Normalizers of certain permutation groups can be written in the form S(f). To see
this observe the following two claims.
(1) N(G)~S(fo,H).
(2) If(forall uES,,) [(for all XE2") (Gx=H(x)<=>Gv-(xl=H(x))~G"=G]
then there exists an f E B,, such that N( G) = S(f).
For convenience, let a-(x) denote xa-. To prove (1) let uE N(G). This means that
G" =G. We want to show that
VxE2"(G,=H(x) <::> Gcr(x 1 =H(x)).
To prove the implication (~)notice that
H(x) = G, = ( G")x = ( Ga(x,)" = H(x)".
Hence, H (x) = G.,(xl, as desired. The converse (~) is similar.
The proof of assertion (2) is immediate. The hypothesis is simply a restatement
of the condition SUo.H) ~ N ( G).
5.3. A logspace algorithm for the representability of cyclic groups. This section is
devoted to the proof of the existence and correctness of a logspace algorithm which,
when given as input a cyclic group G ~S,,, decides whether the group is representable,
in which case it outputs a boolean function f E B,,,k such that G = S(f). The algorithm
is as follows.
Algorithm for Representing Cyclic Groups
Input
G = (u) cyclic group.
Step 1
Decompose u = 0- 1 u 2 · · · uk. where 0- 1 , o-2 , · · · , uk are disjoint cycles of lengths
/1, !2, · · · , lk s:; 2, respectively.
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 577

Step 2
if for all 1 ;a! i ;a! k,
l; = 3=>(3j ;I= i)(3 Ili) and
l; = 4=>(3j 7'= i)(gcd (4, ~);I= 1) and
l; = 5=>(3j ;1= i)(S I~)
then output G is representable.
else output G is not representable.
end
At the present time, we do not know how to efficiently test the representability of
arbitrary abelian groups (or other natural classes of groups such as solvable, nilpotent,
etc.). If a given abelian group K can be decomposed into disjoint cyclic factors, then
we have the following NC algorithm for testing representability: (1) use an NC
algorithm [LM85], [MC85], [Mul86] to "factor" K into its cyclic factors and then (2)
apply the "cyclic-group" algorithm to each of the cyclic factors of K. In view of the
lemma below, the group K is representable exactly when each of its disjoint, cyclic
factors is.
LEMMA 16. Let G ;a! Sm, H ;a! Sn be permutation groups. Then G x H is represent-
able<=> both G, H are representable.
Proof (==>) By the representability of the groups G, H there exist boolean functions
f E Bm and g E Bn such that G x H = S(f) x S(g ). By the maximality theorem there
exists a function h : 2 m+n "'2 such that S( h) =Sm x Sn. Hence, if we put F(x, y) =
(f(x ), g(y )), then it is easy to see that
S(f) xS(g) = S(h) ns(F).
This implies that G x H is representable, and hence also strongly representable.
To prove (~) assume that G x H = S(f), for some f: 2m+n"' k. It is then easy to
see that
G={aE Sm: (a, idn)E Gx H}
={aESm: ('v'x,y)(f(xa,y)=f(x,y))}
={a E Sm: ('v'y)(f'; = J;,)}
= n S(J;,).
ye2"

A similar proof works for the group H. D


The main result of the present section is the following theorem.
THEOREM 17 (Cyclic Group Representability Theorem). There is a logspace
algorithm which, when given as input a cyclic group G ;a! Sn. decides whether the group
is representable, in which case it outputs a function f E Bn such that G = S(f).
The rest of this section is dedicated to the proof (sketch) of correctness of the
above algorithm. The proof is in a series of lemmas. For technical reasons, we intro-
duce two definitions. A boolean function f E B" is called special if for all words w of
length n,

Let u 1 , • • · , ak be a collection of cycles. We say that the group G = a 1 , • • • , uk>


generated by the permutations a 1 , • • · , uk, is specially representable if there exists a
special boolean functionf:2°"'2 (where n is the union of the supports of the a;'s)
such that G = S(f). The support of a permutation u, denoted by Supp(u), is the set
578 P. CLOTE AND E. KRANAKIS

of i such that cr(i) ¥i. The support of a permutation group G, denoted Supp( G), is
the union of the supports of the elements of G.
5.4. Main ideas of the proof. Before proceeding with the details; it will be instruc-
tive to give an outline of the main ideas needed for the corectness proof. We are given
a cyclic group G generated by a permutation er. Decompose er into disjoint cycles
er i. cr2 , • · • , erk of lengths l 1, 12, · · · , lk ~ 2, respectively.
If k = 1 then we know that G is specially representable exactly when 11 ¥ 3, 4, 5.
(The representability of the cyclic group C., for s ¥ 3, 4, 5 is proved in§ 4; for s = 3, 4, 5
observe that for any f E B., if c. s; S(f) then Ds s; S(f), where D. is the dihedral
group. We refrain from repeating the proof and refer the reader to § 4 for the details.)
If k = 2 then the result will follow by considering several possibilities for the
pairs U1 , 12):
• if gcd (/i. 12 ) = 1, then G=<u 1)x(cr2) is the direct product of cr1 and u 2 • Hence, G
is specially representable exactly when both factors are specially representable,
• if (!1 , 12 ) = (3, 3) or (4, 4) or (5, 5) then G is specially representable,
•if Ui. 12 ) = (3, m) (with 31 m) or (4, m) (with gcd (4, m) ¥ 1) or (5, m) (with 51 m),
then G is specially representable.
This will take care of deciding the representability of G for all possible pairs (/i. 12 ).
A similar argument will work for k ~ 3. This concludes the outline of the proof of
correctness.
5.4.1. Sketch of proof. The details of the above constructions are rather tedious
but a sufficient indication is given in the sequel.
LEMMA 18. Suppose that cr1 , • • • , crn+i is a collection of cycles such that both
(cr1, • • · , crn) and (crn+i) are specially representable and have disjoint supports. Then
(cr1 , • • • , O"n+ 1) is specially representable.
Proof.
Put
n
0 0 = U Supp(cr;),
i=l

and let IOol = m, I0 1I = k. Suppose that Jo: 2°0 ...,.. 2 and fi: 2° 1 ...,.. 2 are special boolean
functions representing the groups (cri. · · ·, crn) and <crn+ 1), respectively. Without loss
of generality, we may assume that 1 = fo(Om) ¥ fi(Ok) = 0. Let 0 = 0 0 U 0 1 and define
the function f: 2°...,.. 2 by

Clearly, (o-1 , • • • , O"n+ 1) s; S 0 (j). Hence, it remains to prove that


Sn(/) s; (u1, · · · , O"n+1).
Assume on the contrary that TE S 0 (f)-<cri. · · ·, D"n+ 1). We distinguish two cases.
Case 1. (3i E 0 0 )(3j E 0 1 )( -r(i) =j).
Let w E {O, 1} 0 be defined by w ~ 0 0 = om, and

(w~0 1 )(l)={~ ~~:~.


for I E 0 1 • Since f is a special boolean function and using the fact that f 0 (0m) ¥ f 1(0k)
we obtain that f( w) = 1 ¥ f( wT) = 0, which is a contradiction.
Case 2. (For all i E 0 0)(-r(i) E 0 0).
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 579

Put To= (TI flo) E S 00 and T1= (TI 0 1) E S0 ,. By hypothesis, for all w E 2'1, we have
that

which implies To E S7i0 (fo) and T1E S7iJf1). This completes the proof of the lemma.
An immediate consequence of the previous lemma is the following.
LEMMA 19. If G, H have disjoint support and are specially representable then G x H
is specially representable.
Next we will be concerned with the problem of representing cyclic groups. In
view of Theorem 7 in§ 4, we know that the cyclic group ((1, 2, · · · , n)) is representable
exactly when n #- 3, 4, 5. In particular, the groups ((1, 2, 3)), ((1, 2, 3, 4)), ((1, 2, 3, 4, 5))
are not representable. The following lemma may be somewhat surprising, since
it implies that the group ((1,2,3)(4,5,6)), though isomorphic to ((1,2,3)), is
representable.
LEMMA 20. Let the cyclic group G be generated by a permutation er which is the
product of two disjoint cycles of lengths 11 , 12 , respectively. Then G is specially representable
exactly when the following conditions are satisfied: ( 11= 3 :=;, 3 I12 ) and ( 12 = 3:=;, 3 I11),
U1=4=;>gcd (4, 12) #- 1) and U2 = 4:=;,gcd (4, 11 ) >6 1), (/1=5:=;,5I12) and (12 =5:=;,5I11).
Sketch of proof It is clear that the assertion of the lemma will follow if we can
prove that the three assertions below are true.
(1) The groups ((1, 2, · · ·, n)(n + 1, n + 2, · · ·,kn)> are specially representable
when n = 3, 4, 5.
(2) The groups ((1,2,3,4)(5,· .. ,m+4)) are specially representable when
gcd ( 4, rn) #- 1.
(3) Let m, n be given integers such that either rn = n = 2 or m = 2 and n ~ 6 or
n =2 and rn ~6 or m, n ~6. Then ((1, 2, · · ·, m)(m+l, m+2, · · ·, m+n)) is specially
representable.
Proof of ( 1). We give the proof only for the case n = 5 and k = 2. The other cases
n = 3, n = 4, and k ~ 3 are treated similarly. Details of these constructions are left to
the reader. Let er= cr0 0'1, where u 0 = (1, 2, 3, 4, 5) and cr 1 = (6, 7, 8, 9, 10). From the
proof of Theorem 7 in § 4 we know that

D 5 = S 5 (L') = Ss(L"),
where L'= O*l*O*U l*O*l* and L"= {w E L':1wl 0 ~ l}. Let L consist of all words w of
length 10 such that
-either lwl1=1
-or lwl 1 =2 and (31 ~ i~5) (w; = Ws+; and ('Vj~ i, 5+ i) (w1 =O)) .
-or lwl 1 =3 and (30~i~4) (w=(lOOOOllOOO)<T' or w=(llOOOlOOOO)<T')
-or lwl 1 3 and w1 • • · w5 E L' and w6 • • • w1oE L".
We want to show that in fact ((1, 2, 3, 4, 5)(6, 7, 8, 9, 10)) = S 10 (L). It is clear that

((1, 2, 3, 4, 5)(6, 7, 8, 9, 10)) £;; S 10 (L).

Conversely, suppose that TE S 10 (L). Assume on the contrary there exists an 1~i~5
and a 6 ~j ~ 10 such that T(i) = j. Let the word w be defined such that w1= 0, if l = j,
and = 1 otherwise. It follows from the last clause in the definition of L and the fact
that 0 5 .e L" that w .e L and wr EL, contradicting the assumption TE S 10 (L). It follows
that r is the product of two disjoint permutations To and T 1 acting on 1, 2, · · · , 5 and
6, 7, · · · , 10, respectively. It follows from the last clause in the definition of L that
r 0 E D 5 and T 1 E1T- 1D 5 1T, where 7r(i) = 5+ i, for i = 1, · · ·, 5. Let p 0 = (1, 5)(2, 4) and
580 P. CLOTE AND E. KRANAKIS

p 1 =(6,10)(7, 9) be the reflection permutations on 1, 2, · · ·, 5 and 6, 7, · · ·, 10, respec-


tively. To complete the proof of (1), it is enough to show that none of the permutations
Po, Pi. PoPi. PoaL u~pi, u~u{,
for i "'j, belong to S 10 (L). To see this, let x = 1000011000 EL. Then for the permutations
T = p 0 , p 1 , p 0 p 1 , p0 cr\ for i = 1, 2, 3, 5, and T = u~p 1 for i = 1, 2, 4, 5 it is easy to check
that xT e L. Let x = 110001000. Then for T = p 0 ui and T = a~p 1 it is easy to check that
X 7 e L. Finally, for x = 1000010000 EL and u~u{, where i ~ j, we have that X 7 e L. This
completes the proof of part ( 1) of the lemma.
Proof of (2). Put u 0 =(1,2,3,4), u 1 =(5,6,- · -,m+4), u=cr0 cr 1 • Let L be the
set of words of length m + 4 such that
either lwl1=1
or lwl1=2 and (30~ i ~ lcm (4, m)-1)( w = (lOOOlOm-t)"')
or Jwl 1 =3 and (30 ~ i ~ lcm (4, m)-1)( w = (110010m-i)"')
or Jwl 1 =3 and w1 · · · w4 E L' and w5 • • • Wm+sE L",
where L' = O*l *O* U 1*O*l * and L" are as in Theorem 7 of§ 4 satisfying Sm(L") =Cm
and moreover for all i ~ 1, O; e L". Clearly, <(1, 2, 3, 4)(5, 6, · · ·, m +4)> s;; Sm+4 (L). It
remains to prove that
Sm+iL) £;; ((1, 2, 3, 4)(5, 6, · · · , m +4)).
Let TE (( 1, 2, 3, 4)(5, 6, · · · , m + 4)). As before, T = T 0T 1 , where To E D 4 and
T 1 E 7T- 1Dm1T, where 7T(i) = 4+ i for i = 1, 2, · · ·, m. Let p = (1, 4)(2, 3) be the reflection

on 1, 2, 3, 4. It suffices to show that none of the permutations


j ; j
PCT1, CToa1,
for i 1': j mod 4 are in Sm+4 (L). Indeed, if T = u~u{, then let x = 100010m-l. So it is
clear that x EL, but X e L. Next assume that T = pu\. We distinguish the following
7

two cases.
Case 1. m = 4k, i.e., a multiple of 4.
Let x = 100010m-l. Then x EL, but X 7 e L unless xT = x"; for some j. In this case
j=3mod4 and j=imod4k. So it follows that i=3,7,11, .. -,4k-l. Now let
y = 110010m-i. Then y EL, but y 7 e L for the above values of i, unless y 7 = y" 1 for
some I. In that case we have that l = 2 mod 4 and l = i mod 4k. So it follows that
i = 2, 6, 10, ... '4k- 2. Consequently, Te Sm+iL).
Case 2. gcd (4, m) = 2.
Let x = 100010m-l. Then x EL, but X 7 e L unless X 7 = xcr' for some j. In this case
j = 3 mod 4 and j = i mod 4k. So it follows that for even values of i, Te Sm+ 4 (L). Let
y = 110010m-i. Then y EL, but y' ~ L unless yT = y" 1 for some 1. In that case we have
that l = 2 mod 4 and l = i mod m. So it follows that for odd values of i, Te Sm+ 4(L).
This completes the proof of (2).
Proof of (3). A similar technique can be used to generalize the representability
result to more general types of cycles. Details are left as an exercise to the reader.
A straightforward generalization of Lemma 20 is given in the next lemma.
LEMMA 21. Let G be a permutation group generated by a permutation a which can
be decomposed into k-many disjoint cycles of lengths 11 , / 2 , • • • , lk> respectively. The group
G is specially representable exactly when the following conditions are satisfied for all
l;;ii~k,

l; = 3~(3j 1': i)(3 \ l) and


l; = 4~(3j 1': i)(gcd (4, 1) ,<: 1) and
l; = 5~(3j 1': i)(S \ lj).
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 581

Now the correctness of the algorithm is an immediate consequence of Lemmas


1-5. This completes the proof of Theorem 17. 0

5.5. Asymptotic behavior. Finally, for any sequence (On 2 Sn: n s 1) of permuta-
tion groups we consider the value of the limit

We have the following theorem.


THEOREM 22. (Almost all boolean functions have trivial invariance groups.) For
any family (On: n s 1) of permutation groups such that each On 2 S,,, we have that

Moreover, if lim inf IOn I> 1 then

lim l{f E Bn: S~,(H:;: On}I = lim j{f E Bn: S1\ ( ) = On}I = O.
n-oc 2- n-w 2-

Proof During the course of this proof we use the abbreviation 0( m) :=


0m(((l, 2, · · ·, m))). First we prove the second part of the theorem. By assumption
there exists an n0 such that for all n s n 0 , IOn I> 1. Hence, for each n s n 0 , On contains
a permutation of order k( n) s 2, say <Tn. Without loss of generality we can assume that
each k(n) is a prime number. Since k(n) is prime, <Tn is a product of k(n)-cycles. If
(i,, · · · , ik(n 1) is the first k(n )-cycle in this product then it is easy to see that

It follows that

Recall from [Ber71] that the formula

0(m)=__!__· I <jJ(k) · 2m/k


m klm

gives the P6lya cycle index of the group ((1, 2, · · ·, m)) acting on {1, 2, · · ·, m}, where
<f;(k) is Euler's totient function. However, it is easy to see that fork prime

0(k) 1 2 2
-2k- -- - +---
k 2k k2k.

In fact the function in the right-hand side of the above equation is decreasing in k.
Hence, for k prime,
0(k) 0(2) 3
--:$--=-
2k - 22 4·
582 P. CLOTE AND E. KRANAKIS

It follows that

Since the right-hand side of the above inequality converges to 0 the proof of the second
part of the theorem is complete. To prove the first part notice that
{fEBn:S(f)~idn}s;:; U {feBn:ueS(f)},
u7"idn

where a ranges over cyclic permutations of order a prime number ~ n. Since there are
at most n t permutations on n letters we obtain from the last inequality that
l{fEBn:S(f)~{idn}}I< .r2"-2= 2 o(nlogn), 2 -2•- 2 O
22" = n.1 ~ ,

as desired. 0
As a consequence of the above theorem we obtain that asymptotically almost all
boolean functions have trivial invariance group.
6. Invariance groups of languages and circuits. In this section we classify languages
according to the size of their invariance groups. Furthermore, we consider questions
concerning their structural properties and complexity. Recall that for each Ls;:; {O, 1}*
and n, Ln is the set of strings in L of length exactly n. By abuse of notation we also
denote the characteristic function of Ln with the same symbol. Let Sn(L) denote the
invariance group of the n-ary boolean function Ln. For any language L and any
sequence a= (an: n ~ 1) of permutations such that each <Fn E Sn we define the language
L: = {x E 2n: Xu" E Ln}.
For each n let Gn ~Sn and put G = (Gn: n ~ 1). Define
LG= u L'::·.

For each 1 ~ k ~ oo, let F k be the class of functions n clog< kl n, c > 0, where log< I) n =
log n, log<k+lJ n =log log<kl n, and log«'°l n = 1. Clearly, F00 is the class P of polynomial
functions. We also define F 0 as the class of functions 2cn, c > 0. Let L(F d be the set
languages Ls;:; {O, 1}* such that there exists a function f E F k satisfying
'v'n(ISn: Sn(L)I ~f(n)).
We will also use the notation L(EXP) and L(P) for the classes L(F0 ) and L(F00 ) ,
respectively. Occasionally, a language LE L(P) will also be called a language which
has polynomial index or is even almost symmetric.
6.1. Structural properties. The following theorem gives some of the structural
properties of the classes of languages L(F k).
THEOREM 23. For any 0 ~ k ~ oo and any language LE L(F k),
( 1) L(F k) is closed under boolean operations and homomorphisms,
(2) (L · ~) E L(Fd,
(3) La E L(F k), where a= (an: n ~ 1), with each Un E Sn,
(4) if ISn: Ns.( Gn)I ~f(n) andf EFk then LG E L(Fk), where G = (Gn: n ~ 1).
Proof We use extensively (even without explicit mention) the results of Theorem
10. To prove (1) notice first that Sn(-iL) =Sn(L). To prove that L(Fk) is closed under
union and intersection use the following inequality from group theory: for K, K' ~ G,
IG:KnK'l~IG:KI · IG:K'I·
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 583

For example, for closure under intersection we have that Sn ( L) n S,, ( L') £Sn ( L n L'),
which implies that

To prove closure under a homomorphism h: L--'> L' note that Sn(L) £ S,,(h(L)). Hence,

IS,,: S,, (L')I =IS,,: S,,( h(L) )I 2i !Sn : S., (L )I.


To prove (2) let L' = L · L = {xa: xE L, a EL} and note that

IS,, :S,,(L')I ~ n · IS,,_1 :S,,-1(L)I.


To prove (3) note that S,,(L)'rn = S,,(L°.). To prove (4), note that we have Ns,, ( Gn) n
Sn(L)£S,,(Lc). Indeed, for rENs,,(G,,)nS,,(L) we have that G,,r=rG,,, which in
turn implies that

L~n 7 = L~G" = LJ L~O"" = LJ L~11 = L';! 11 •

0' 11 ECiu UnEGn

Hence,

as desired. 0
The classes L(P) and L(EXP) enjoy the closure properties mentioned below.
THEOREM 24.

LE L(P) and p E P:;>!Sp(n): sp(niCL)I = n°(1>.


Proof. The proof is obvious, since the class of polynomials is closed under
composition. 0
THEOREM 25.

L1, L 2 E L(EXP):;> L = {xy: x EL 1, y E L2, l(x) = l(y)} E L(EXP).


Proof. It is clear that S,,(L 1) xS,,(L2)s; S2 ,,(L). It follows from Stirling's formula
that

(2n) !
IS2,, :S2,,(L)I;:;;; /S,,(L)I · IS,,(L)I

= ( 2 n)! ·IS,,:S,,(L)i2
n! · n !

2i (2n)! ·20(nl=20(n).
0
n! · n!

Let REG denote the class of regular languages.


THEOREM 26. The following properties hold for any 1 ~ k < cx:i,

( 1) L(F<X,) = L(P) c · · · c L(Fk+i) c L(Fd c · · · c L(EXP) = L(F 0 ),


(2) REG n L(P) ~ 0, REG-L(EXP) ~0, L(P)-REGr= 0.
Proof. To prove L(Fk+i) c L(Fk),forl ~ k<ro, putf(n) = n -log<kl n and consid
the language
584 P. CLOTE AND E. KRANAKIS

Then we have that

IS n.. S n (L)I =~=


J(n) !
no(1og('ln)
.

It follows that L(F k+i) c L(F k). (Note that by the pumping lemma for regular languages
L cannot be regular.) The proof of L(F k) c L(F 0 ) is more delicate. The group S,, x S,,
is maximal in S 2 n. It follows from our representation theorem for maximal groups that
there exists a language L such that for all n,
S2n(L) =Sn x Sn.
It follows from Stirling's formula that IS 2": S 2 " (L)I = 2 Otn \ as desired. The proof of
L(F00 ) c L(F d, k ~ l, follows from the above remarks. This completes the proof of ( 1 ).
To prove REG n L(P)"" 0, consider the trivial language L = {O, I}*. To prove REG -
L(EXP)"" 0, consider the language L = O*l *.To prove L(P)- REG"" 0. For any set
S of positive integers let L 5 ={On: n ES}. Clearly, L~(x) = 1 if n ES and x = O", and
=0 otherwise. It is easy to see that for all S, Ls E L(P), and hence L(P) is uncoutable.
(In fact, Sn(L 5 ) = S,,, for all n and S.) In particular, the nonregular language L ={OP: p
is a prime number} E L(P). D
A few useful and illuminating examples are now in order.
Examples. (1) Let Lk={xE{O, 1}*: l(x)~k, x 1 ~· • ·~xd. Then S,,(Lk)=Sn-k
and therefore IS": Sn (L)I = n !/ (n - k) ! = O(nk). Hence, for all k, L k E L(P).
(2) For each wordx =x 1 • • • Xn letxT = x,, · · · x 1 and LT ={xT: XE L}. Puter,,(i) =
n - i + 1. Then Lcr =LT, where er= (er11 : n ~ 1).
(3) There exist languages L 0 , L 1 E L(P) such that L 0 • L 1 E L(EXP). Indeed, put
L = {O}*, L 1 = {l}*. Then L = L 0 • L 1 ={O"1 m: n, m ~ O}. It is easy to see that
0

IS:S,,(L)l=n!.
(4) There exists a language LEL(P) such that L*EL(P). Indeed, put L={Ol}.
Then for n even, er ES if and only if for all i ~ n ( i is even if and only if er( i) is even).
It follows that ISn: S 11 (L)I = n !/ (n/2) !(n/2) !. Hence, L * E L(EXP) - L(P).
(5) L(P) is not closed under inverse homomorphism. Indeed, let D be the Dyck
language on one parenthesis and h: D-'> L be the homomorphism h(O) = h(l) = 0. In
view of the results of§ 3, DE L(P).
(6) For each function f: N--" N such that for all n ~ 1, f( n) ~ n, we define the
language

Using the pumping lemma for regular languages we can show that Lf E REG=?
sup 11 f(n) «:X:l.
Similar classes of languages corresponding to the cycle index can be defined as
follows. Let L0 (F k) be the set of languages L such that there exists a function f E F k
satisfying
\in( ®(S( 1 11 ) ) ~f( n )).
Since, 0(Sn(L))~(n+l) · ISn:Sn(L)I, it is clear that L(Fk)~ L 0 (Fk)· In fact we can
show that L(Fk) c L 0 (Fk)· To see this takef(n) = n -logCkl n. Define xE L 11 if and only
if x 1 ~x1 ~· • ·~xf(n)· Then it is easy to see that Sn(L)=Sf(n)· Hence, ISn:S (L)I= 11

O(n 10 g"), while E>(S (L)) = (f(n)+ l)iogl'l" = O(n 2).


11

6.2. Circuit complexity of formal languages. In this section, we study the com-
plexity of languages LE L(P). The following result is proved by applying the intricate
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 585

NC algorithm of [BLS87] for permutation group membership. By delving into a deep


result in classification theory of finite simple groups, we improve the conclusion to
that of Theorem 29. For clarity however, we present the following.
THEOREM 27. For any language L£ {O, 1}*, if LE L(P) then Lis nonuniform NC.
Proof. As a first step in the proof we will need the following claim.
CLAIM. There is an NC 1 algorithm which, when given xE{O, ir, outputs crES"
such that x<T = 1mon-m' for some m.
Proof of the claim. Before giving the proof of the claim, we illustrate the idea by
citing an example. Suppose that x = 101100111. By simultaneously going from left to
right and from right to left, we swap an "out-of-place" 0 with an "out-of-place" 1,
keeping track of the respective positions. 3 This gives rise to the desired permutation
cr. In the case at hand we find a-= (2, 9)(5, 8)( 6, 7) and xcr = 16 03 .
Now we proceed with the proof of the main claim. Define the predicates Ek b ( u ),
to hold when there are exactly k occurrences of b in the word u ( b = 0, 1) are in, N C 1•
The predicates Ek.b are obviously computable in constant depth, polynomial size
threshold circuits, i.e., in TC 0 . By the work of Ajtai, Komlos, and Szemeredi [AKS83]
TC 0 £ NC 1 • Fork= 1, · · ·, [n/2] and 1 ~ i <j;:;; n, let a:;,1,k be a log depth circuit which
outputs l exactly when the kth "out-of-place" 0 is in position i and the kth "out-of-
place" 1 is in position j. It follows that o:;,J,k(x) = 1 if and only if "there exist k-1
zeros to the left of position i, the ith bit of x is zero, and there exist k ones to the
right of position i" and "there exist k - 1 ones to the right of position j, the jth bit of
x is one, and there exist k zeros to the left of position j." This in tum is equivalent to
Ek-1,o(x 1,···,X;- 1) andx;=O andE1<, 1(X;+ 1 ,···,x") and
Ek-1,l(XJ+i. · · ·, x") and x1 = l and Ek,o(X 1 • • • xJ_ 1 ).
This implies that the required permutation can be defined by
[n/2] }
a-=fl { (i,j): i<j and V O\;.k .
k~I

Converting the fanin, [n/2]-v-gate into a log ([n/2]) depth tree of fanin, 2-v-gates,
we have an NC 1 procedure for computing er. This completes the proof of the claim.
Next we continue with the proof of the main theorem. Put G" = S,,(L) and let
Rn= {h 1 , • • • , hq} be a complete set of representatives for the left cosets of G"' where
q~p(n) and p(n) is a polynomial such that ISn:Gnl~p(n). Fix xE{O, By the 1r.
previous claim there is a permutation er which is the product of disjoint transpositions
and an integer O;;;k~n such that x"=Ikon-k. So x=(lkon-k)u. In parallel for
i= 1, · · ·, q test whether hj 1 erE Gn by using the principal result of [BLS87], thus
determining i such that a= h;g, for some g E Gn. Then we obtain that
Ln(x) = Ln((l kon-k)o-) = Ln((l kon·-k)h,g) = Ln((lkon-k)h•).
By hardwiring the polynomially many values L 11 ((lko"-k)h 1 ) for O~ k~ n and 1 ~ i~ q,
we produce a polynomial size polylogarithmic depth circuit family for L. 0
Theorem 27 involves a straightforward application of the beautiful NC algorithm
of Babai, Luks, and Seress [BLS87] for testing membership in a finite permutation
group. By using the deep structure consequences of the O'Nan-Scott theorem below,
together with Bochert's result on the size of the index of primitive permutation groups

3 This is a well-known trick for improving the efficiency of the "'partition" or "split" algorithm used in

quick-sort.
586 P. CLOTE AND E. KRANAKIS

(see Theorem 1(3) in § 2), we can improve the NC algorithm of Theorem 27 to an


optimal TC0 algorithm (and hence Nc1). First, we take the following discussion and
statement of the O'Nan-Scott theorem from [KL88, p. 376).
Let I= {l, 2, · · ·, n} and let Sn act naturally on I. Consider all subgroups of the
following five classes of subgroups of Sn.
a 1 : Sk xSn-k, where 1 ~ k~ n/2,
a 2 : Sal Sb, where either (n = ab and a, b 1) or (n = ab and a~ 5, b ~ 2),
a 3 : the affine groups AGLa ( p ), where n =pa,
a 4 : Tk · (Out( T) x Sk), where Tisa nonabelian simple group, k s;;; 2 and n =I TI k-i,
as well as all groups in the class,
a 5 : almost simple groups acting primitively on I.
THEOREM 28 (O'Nan-Scott). Every subgroup of Sn not containing An is a member
ofa1U · · ·U a5.
Now we can improve the result of Theorem 27 in the following way.
THEOREM 29 (Parallel complexity of Languages of Polynomial Index). For any
language L £; {O, 1}*, ifLE L(P) then Lis in 9-nonuniform TC0 and hence in (nonuniform)
NC 1•
Proof. The proof requires the following consequence of the O'Nan-Scott theorem.
CLAIM. Suppose that ( Gn ~Sn: n s;;; 1) is a family of permutation groups such that
for all n, !Sn: Gnl ~ n\ for some k. Then for sufficiently large N, there exists an in~ kfor
which Gn = U" x Vn with the supports of Un, Vn disjoint and Un~ Sin• Vn =Sn-in.
Before proving the claim we complete the details of the proof of Theorem 29.
Apply the claim to Gn = Sn(L) and notice that given x E 2", the question of whether x
belongs to Lis decided completely by the number of 1's in the support of Kn= Sn-i.,,
together with information about the action of a finite group Hn ~Sin• for i,, ~ k. Using
the counting predicates as in the proof of Theorem 27, it is clear that this is a TC 0
and hence NC 1 algorithm. Thus, the proof of the theorem is complete, assuming the
claim.
Proof of the claim. We have already observed at the beginning of§ 5 that G" -;t An.
By the O'Nan-Scotttheorem, Gn is a member of £¥ 1 U · · · U a 5 • Using Bochert's theorem
on the size of the index of primitive permutation groups (§ 2, Theorem 1(3)), the
observations of [LPS88] concerning the primitivity of the maximal groups in a 3 U a 4 U
a 5 and the fact that G" has polynomial index with respect to Sn, we conclude that the
subgroup Gn cannot be a member of the class a 3 U a 4 U a 5 • It follows that G" E a 1 U a 2 •
We show that in fact Gn e a 2 • Assume on the contrary that G" ~ Hn =Sal Sb. It follows
that IHnl =a !(b !t. We distinguish the following two cases.
Case 1. n = ab, for a, b > 1.
In this case it is easy to verify using Stirling's interpolation formula
(n/ e)nvn < n! < (n/ e)n3vn
that
n! an-a
ISn: Hnl =a !(b !t 3ba/2(3/ at.Ja.'
Moreover, it is clear that the right-hand side of this last inequality cannot be asymptoti-
cally polynomial in n, since a~ n is a proper divisor of n, which is a contradiction.
Case 2. n =ab, for a6;;;5, b6;;;2.
A similar calculation shows that asymptotically
n! n!
JS": Hnl =a !(b !)a= a !(b'!)a,
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 587

where b' = ab-i. It follows from the argument of Case 1 that this last quantity cannot
be asymptotically polynomial in n, which is a contradiction. It follows that G" E a 1 •
Let Gn ~ S; x Sn-i• for some I~ in n/2. We claim that, in fact, i" ~ k for all but a finite
number of n's. Indeed, put in= i and notice that
n! ·
ISn:S;XSn-i I=. (
k
") .G(n')~ISn:Gnl~n,
1! n-1 !
which proves that i~ k. It follows that Gn =Un x Vn, where Un ~Si and Vn ~Sn-i .
Since in~ k and IS,.: G,.I ~ nk it follows that for n large enough" Vn =Sn-in. This
completes the proof of the claim. Now let L£ {O, 1}* have polynomial index. Given
a word x E {O, lt, in TC 0 one can test whether the number of 1's occurring in the n - in
positions (where V,. = Sn-;J is equal to a fixed value, hardwired into the nth circuit.
This, together with a finite look-up table corresponding to the U,. part, furnishes a
TC 0 algorithm for testing membership in L. D
6.3. Applications. An immediate consequence of our analysis is that if ( G,. ~
S,.: n;?; 1) is a family of transitive permutation groups such that IS,.: Gnl = n° 0 l then
Gn =Sn, for all but a finite number of n's (this answers a conjecture of Perrin). It is
also possible to give a more algebraic formulation of the main consequence of Theorem
29. For Pn a polynomial in the variables x 1 , • • • , x" and with coefficients from the two
element field Z 2 , let
S(pn) ={a E Sn: 'v'xi, ... 'Xn(p,.(xi, . .. 'Xn) =Pn(Xcr(I)> ... 'Xu(n)) mod 2)}.
A family (pn: n;?; 1) of multivariate polynomials in Z 2 [x 1 , • • · , Xn] is of polynomial
index if ISn :S(pn)I = n° 0 l.
THEOREM 30. If (pn: n ~ 1) is family of multivariate polynomials (in
Z 2 [Xi. · · ·, xn]) of polynomial index then there is a family (q,,: n;?; 1} of multivariate
polynomials (in Z 2 [x 1 , • • • , x,.]) of polynomial length such that p,, = q,,.
Because of the limitations of families of groups of polynomial index proved in
the claim above, we obtain a generalization of the principal results of [FKPS85].
Namely, for Le;; {O, 1}* let µdn) be the least number of input bits which must be set
to a constant in order for the resulting language Ln = L n {O, 1}" to be constant (see
[FKPS85] for more details). Then we can prove the following theorem.
THEOREM 31. If LE L(P) (i.e., L is a language of polynomial index) then

µ.L(n)~(logn) 00 l ~ LEAC 0 •

Our characterization of permutation groups of polynomial index given during the


proof of Theorem 29 can also be used to determine the parallel complexity of the
following problem concerning "weight-swapping." Let G = (Gn: n EN) denote a
sequence of permutation groups such that G,, ~Sn, for all n. By SWAP ( G) we
understand the following problem:
Input. n EN, a 1 , • • ·,a,. positive rationals, each of whose (binary) representations
is of length at most n.
Output. A permutation a E Gn such that for all 1 ~ i n, aa(i) + au(i+I) ~ 2, if such
a permutation exists, and the response "NO" otherwise.
THEOREM 32. For any sequence G of permutation groups of polynomial index, the
problem SWAP (G) is in nonuniform NC 1 •
Proof By the characterization of sequences of groups of polynomial index, there
exist integers k, N such that for all n;?; N, G" = H,, x Kn, where Hn ~Si" and Kn= Sn-in,
588 P. CLOTE AND E. KRANAKIS

with in ;:;2 k. Given n ~ N, and n positive rational weights a 1 , • • • , a,, test whether there
exist permutations u EH,, and TE K,, such that for 1 ;:;2 i ;:;2 n, aco-x-r)(il + a(<T+-r)(i+ll ;:;2 2,
as follows. For T, sort the set of weights {ai: i E Supp( K,,)} in decreasing order. Assume
wlog that Supp(K,,) = {l, · · ·, n - i,,}. Let p EK,, be a "sorting" permutation such that
aP 0 i ~ ap( 2 ) ~ • • • ~ ap(n-i.,l. Test in parallel whether

aµ(l) + ap(n-i.,) ;:;2 2, apc 2l + ap(n-i.,-I) ;:;2 2, · · ·,etc.


If so, then let T be the appropriate permutation such that

i - p(l), 2- p(n - i,,), · · ·, n- i,, -1- p ( ~-1 ,


n -i )
n - in 1--7
i)
n-
p(~ ,

if n - i,, is even, and a variant of this, if n - i,, is odd. Since sorting n many n-bit
numbers is in NC 1, computing T is in NC 1 • Since H,, ;:;2S;,,, where i,, ;:;2 k, there are only
a finite number of possibilities to test for u. These are hard wired (by non uniformity)
into the circuit. D
The following conjecture would relate the cycle index of a sequence G = ( G,,: n ~ 1>
of groups with the circuit complexity of the language L.
CONJECTURE 33. For any language Le:; {O, 1}*, if LE L 0 (P) then Lis nonuniform
NC.
This conjecture appears somewhat plausible, since it follows from the next theorem
that if G=(G,, ;:;2S,,: nsl) is a sequence of groups whose cycle index 0,,(G,,), as a
function of n, majorizes all polynomials, then there is a language L with S,,(L) 2 G,,
and Le SIZE(n° 0 l).
THEOREM 34. For any sequence G = (G,,: n s 1) of permutation groups G,, ;:;2 S,, it
is possible to find a language L such that
Lit SIZE(.JE>( G,,)), and Yn(S(L,,) 2 G,,).
Proof. By Lupanov's theorem i{f E B,,: c(f) ;:;2 q }j = O(qq+i) = 2 O(qlogq>_ Hence, if
q,, -HX) then i{f E B,,: c(f) ;:;2 q,,}i < 2q;" In particular, setting q,, = .JE>( G,,) we obtain
IU E B,,: c(f) ;:;2 .J0( G,,)}I < 2 6 (G.,l = i{f E B,,: S(f) 2 G,,}i.
It follows that for n big enough there exists an f,, E B,, such that S(f,,) 2 G,, and
c(f,,) > .J@( G,,). This completes the proof of the theorem. D
7. Discussion and open problems. Three of the main questions we have tried to
answer in the present paper are ( l) which permutation groups arise as (or are isomorphic
to) the invariance groups of boolean functions, (2) determining the complexity of
deciding the representability of a permutation group, (3) determining the relation
between the family of invariance groups of a formal language L and the parallel
complexity of L.
Concerning question (1), we saw that most (i.e., with a few exceptions) maximal
permutation subgroups ofS,, are representable. We have shown that every permutation
group GS,, is isomorphic to the invariance group of a boolean function f E Bn(Jogn+u.
However, we do not know if this last "upper bound" can be improved to f E B""' for
some constant c independent of n. In the case of question (2), we gave a logspace
algorithm for deciding the representability of cyclic groups. In general however, we
do not know of any efficient algorithm for deciding the representability of any other
natural classes of permutation groups (e.g., abelian, nilpotent, solvable, etc.). The
existence of a polynomial time algorithm for testing representability of an arbitrary
permutation group is related to the question of whether graph nonisomorphism is in
polynomial time.
BOOLEAN FUNCTIONS AND PARALLEL COMPLEXITY 589

Concerning question (3), we have shown a relation between the size of the index
of the invariance group of a formal language and its complexity. We showed that any
language of "polynomial size index" is in (nonuniform) TC 0 • It is possible that a finer
analysis of the structure results for maximal permutation groups will yield a similar
result for other classes of languages, like the ones with subexponential or even
exponential size index. We conjecture that a similar result is true for any language of
"polynomial size P6lya index." We believe as well that there should be a relation
between the algebraic structure of the syntactic monoid of a regular language L £ {O, 1}*
(Krohn-Rhodes theorem) and the family of invariance groups of Ln. As indicated by
our preliminary work, straightforward approaches to such an investigation are not
likely-the property of a group being representable is not preserved under homomorph-
ism. Our parallel complexity results concern nonuniform families of boolean circuits.
A natural sequel to our work might investigate uniform versions of some of our results.
For instance, if L£ {O, 1}* is a regular (or context free, or logspace computable, etc.)
language \.\'.ith polynomial index (or polynomial size P6lya index) then is Lin logspace
uniform TC 0 ?
Another interesting question concerns the problem of giving an efficient algorithm
A which on input a formal language L, a permutation o- ES", and an integer n,
determines whether or not o- E Sn(L), i.e.,

A( L, n, o- ) ={1 ifo-ESn(L)
0 otherwise.
We investigated this question in the present paper for regular languages. The obvious
algorithm has complexity 0(2") (to check membership of a permutation u in Sn(L)
test whether for all x E 2", x E Ln ~ xa E Ln ). A similar question applies to right-quotient
representatives of S"(L). It would also be interesting to investigate these questions for
other types of languages, such as CFL, etc.
Acknowledgments. Discussions with Peter van Emde Boas, Danny Krizanc,
Dominique Perrin, Paul Schupp, and Paul Vitanyi are gratefully acknowledged. A. M.
Cohen was extremely helpful with the literature on maximal permutation groups.
Lambert Meertens made comments that significantly improved the presentation and
pointed out that essentially our original n 2 -upper-bound proof of the isomorphism
theorem could yield the improved, n(log n + 1)-upper-bound.

REFERENCES

[AKS83] M. AlTAI, J. KOMLOS, AND E. SZEMEREDI, An O(n log (n)) sorting network, in Proc. 15th
Annual ACM Symposium on Theory of Computing, 1983, pp. 1-9.
[Arb69] M. A. ARBIB, Theories of Abstract Automata, Prentice-Hall Series in Automatic Computation,
Prentice-Hall, Englewood Cliffs, NJ, 1969.
[ASW85] c. ATTIYA, M. SNIR, AND M. WARMUTH, Computing on an anonymous ring, in Proc. 4th
Annual ACM Symposium on Principles of Distributed Computation, 1985.
[BLS87] L. BABA!, E. LUKS, AND A. SERESS, Permutation Groups in NC, in Proc. 19th ACM Symposium
on Theory of Computing, New York, 1987.
[BB89] P. BEAME AND H. BODLAENDER, Distributed computing on transitive networks: The torus, in
Proc. 6th Annual Symposium on Theoretical Aspects of Computer Science (STACS), 1989.
[BPSS] R. A. BEAUMONT AND R. P. PETERSON, Set-transitive permutation groups, Canad. J. Math., 7
{1955), pp. 35-42.
[Ber71] C. BERGE, Principles of Combinatorics, Academic Press, New York, 1971.
[Com70] L. COMTET, Analyse Combinatoire, Deuxieme Tome; Collection SUP, Presses Universitaires de
France, 1970.
590 P. CLOTE ANO E. KRANAKIS

[Coo85] S. A. COOK, A taxonomy of problems with fast parallel algorithms, Inform. and Control, 64
(1985), pp. 2-22.
(FKPS85] R. FAGIN, M. KLAWE, N. PIPPENGER, AND L. STOCKMEYER, Bounded-Depth, Polynomial-Size
Circuits for Symmetric Functions, Theoret. Comput. Sci., 36 (1985), pp. 239-250.
[FKL88] L. FINKELSTEIN, D. KLEITMAN, AND T. LEIGHTON, Applying the classification theorem for
finite simple groups to minimize pin count in uniform permutation architectures, VLSI
Algorithms and Architectures, 3rd Aegean Workshop on Computing (A WOC)88, Corfu,
Greece, June/July 1988, J. H. Reif, ed., Lecture Notes in Computer Science 319, Springer-
Verlag, Berlin, New York, pp. 247-256.
[FHL80] M. FURST, J. HoPCROFT, AND E. LUKS, Polynomial-time algorithms for permutation groups, in
Proc. 2lst IEEE Symposium on Foundations of Computer Science, Syracuse, NY, 1980.
[FSS84] M. FURST, J. SAXE, AND M. SJPSER, Parity circuits and the polynomial time hierarchy, Math.
Systems Theory, 17 (1984), pp. 13-27.
[Hal57] M. HALL, The Theory of Groups, Macmillan, New York, 1957.
[Har64] M. HARRISON, On the classification of Boolean functions by the general linear and affine groups,
J. Soc. Indust. Appl. Math., 12 (1964), pp. 285-299.
[Harr78] M. HARRISON, Introduction to Formal Language Theory, Addison Wesley, New York, 1978.
[KL82] B. W. KERNIGHAN AND S. LIN, An efficient heuristic procedure for partitioning graphs, Bell
Systems Tech. J., 49 (1982).
[KL88] P. B. KLEID MAN AND M. W. LIEBECK, A survey of the maximal subgroups of the finite simple
groups, in Geometries and Groups, M. Aschbacher, A. M. Cohen, and W. M. Kantor, eds.,
reprinted from Geometriae Dedicata 25(1-3), pp. 375-389, D. Reidel, Dordrecht, the
Netherlands, 1988.
[KK89] E. KRANAKIS AND D. KRIZANC, Computing Boolean functions on anonymous networks, Centrum
voor Wiskunde en lnformatica, Tech. Report CS-8935, Amsterdam, the Netherlands, Sep-
tember 1989.
[LPS88] M. W. LIEBECK, C. E. PRAEGER, AND J. SAXL, On the 0' Nan-Scott theorem for finite primitive
permutation groups, J. Austral. Math. Soc. Ser. A, 44 (1988), pp.389-396.
[LM85] E. M. LUKS AND P. McKENZIE, Fast parallel computation with permutation groups, in Proc.
26th IEEE Symposium on Foundations of Computer Science, 1985.
[MS78] F. J. MACWILLIAMS AND N. J. A. SLOANE, The Theory of Error Correcting Codes, North-
Holland, Amsterdam 1978.
[McC56] E. J. McCLUSKEY, JR., Detection of group invariance or total symmetry of a Boolean function,
Bell Systems Tech. J., 35 (1956), pp. 1445-1453.
[McK84] P. McKENZIE, Parallel complexity and permutation groups, Ph.D. thesis, Department of
Computer Science, University of Toronto, Toronto, Ontario, Canada, 1984.
[MC85] P. McKENZIE AND S. A. COOK, The parallel complexity of abelian permutation group problems,
Tech. Report 181/85, Department of Computer Science, University of Toronto, Toronto,
Ontario, Canada, 1985.
[Mul86] K. MULMULEY, A fast parallel algorithm to compute the rank of a matrix over an arbitrary field,
in Proc. 18th ACM Symposium on Theory of Computing, 1986, pp. 338-339.
[Pas66] D.S. PASSMAN, Permutation Groups, W. A. Benjamin, New York, Amsterdam, 1966.
(Pip79] N. PIPPENGER, On simultaneous resource bounds, in Proc. 20th IEEE Symposium on Foundations
of Computer Science, 1979, pp. 307-311.
[PR87] G. P6L YA AND R. C. READ, Combinatorial Enumeration of Groups, Graphs and Chemical
Compounds, Springer-Verlag, Berlin, New York, 1987.
[Sav76] J.E. SAVAGE, The Complexity o.fComputing, John Wiley, New York, 1976.
[Sha49] C. SHANNON, The synthesis of two-terminal switching circuits, Bell Systems Tech. J. (1949),
pp. 59-98.
[SV84] L. STOCKMEYER AND U. V1SHKIN, Simulation of parallel random access machines by circuits,
SIAM J. Comput., 13 (1984), pp. 409-422.
[Tzu82] T. TSUZUKU, Finite Groups and Finite Geometries, Cambridge University Press, Cambridge,
U.K., 1982.
[Wie64] H. WIELANDT, Finite Permutation Groups, Academic Press, New York, 1964.
[Yab83] S. YABLONSKY, Introduction aux mathematiques discretes, MIR, (translated from the Russian),
1983.
[Yao85] A. YAo, Separating the polynomial time hierarchy by oracles, in Proc. 26th IEEE Symposium on
Foundations of Computer Science, 1985, pp. 1-10.

You might also like