1: \section{Our class of groups in context}
2: In this paper we study groups, together with a finite ordered set of
3: monoid generators, with the property that their set of
4: universally minimal \shl rules is a regular language.
5: In this section, we explain what this rather daunting sentence means,
6: and we set this class of groups in the context of various other related
7: classes, investigating which of these classes is included in which.
8: In the next section, we will prove that groups in this class are finitely
9: presented.
10:
11: Throughout we will work with a group $G$ generated by a fixed finite set $A$,
12: and a fixed finite set of defining relations.
13: Formally, we are given a map $A\to G$, but our language will
14: sometimes (falsely) pretend that $A$ is a subset of $G$. The reader is
15: urged to remain aware of the distinction, remembering that, as a result of
16: the insolubility of the word problem, it is not in general possible to tell
17: whether the given map $A\to G$ is injective. We assume we are
18: given an involution
19: $\inverse:A\to A$ such that, for each $x\in A$, $\inverse(x)$ represents
20: $x^{-1}\in G$. By $\Astar$ we mean the set of words (strings) over
21: $A$. (Formally a word is a function $\{1,\ldots,n\}\to A$, where $n\ge 0$.)
22: We also write
23: $\inverse:\Astar\to\Astar$ for the formal inverse map defined by
24: $\inverse(x_1\ldots x_p) =\inverse(x_p)\ldots\inverse(x_1)$.
25:
26: We assume we are given a fixed total order on $A$.
27: This allows us to define the \textit{\shl} order on $\Astar$ as
28: follows.
29: We denote by $|u|$ the length of $u \in \Astar$.
30: If $u,v\in \Astar$, we
31: say that $u<v$ if either
32: $|u| < |v|$ or $u$ and $v$ have the same
33: length and $u$ comes before $v$ in lexicographical order.
34: The \textit{\shl representative}
35: of $g\in G$ is the smallest $u\in\Astar$ such
36: that $u$ represents $g$. This is also called the \textit{\shl normal
37: form} of $g$. If $u\in\Astar$, we write $\overline u \in G$ for the element
38: of $G$ which it represents. If $u$ is the \shl representative of
39: $\overline u$, we say that $u$ is \textit{in \shl normal form}.
40:
41: Suppose we have $(G,A)$ as above. Then there may or may not be an algorithm
42: that has a word $u\in\Astar$ as input and as output the \shl
43: representative of $\overline u\in G$. The existence of such an algorithm is
44: equivalent to the solubility of the word problem for $G$,
45: since there are only a finite number of words $v$ such that $v<u$.
46:
47: A natural attempt to construct such an algorithm is to find a set
48: $R$ of \textit{replacement rules}, also known as \textit{Knuth--Bendix rules}.
49: In this paper, a replacement rule
50: will be called simply a \textit{rule}, and we will restrict our
51: attention to rules of a rather special kind.
52: A \textit{rule} is a pair $(u,v)$ with $u>v$
53: Given a rule $(u,v)$, $u$ is called the \textit{left-hand
54: side} and $v$ the \textit{right-hand side}.
55: The idea of the
56: algorithm is to start with an arbitrary word $w$ over $A$ and to \textit{reduce}
57: it as follows: we change it
58: to a smaller word by looking in $w$ for some left-hand side $u$ of some rule
59: $(u,v)$ in $R$. We then replace $u$ by $v$ in $w$ (this is called an
60: \textit{elementary reduction}) and repeat the operation until no
61: further elementary reductions are possible (the repeated process is
62: called a \textit{reduction}).
63: Eventually the process must stop with an
64: \textit{$R$-irreducible} word, that is a word which contains no subword
65: which is a left-hand side of $R$.
66:
67: \subsection{Thue equivalence.}\label{Thue}
68: Given a set of rules $R$, we write $u\rightarrow_R v$ if there is an elementary
69: reduction from $u$ to $v$, that is, if there are words $\alpha$ and
70: $\beta$ over $A$ and a rule
71: $(\lambda,\rho)\in R$ such that $u = \alpha \lambda \beta$ and
72: $v=\alpha \rho\beta$. \textit{Thue equivalence} is the equivalence
73: relation on $\Astar$ generated by elementary reductions.
74:
75: There is a multiplication in $\Astar$ given by concatenation. This
76: induces a multiplication on the set of Thue equivalence
77: classes. We will work with rules where the set of equivalence classes
78: is isomorphic to the group $G$.
79:
80: By no means every set of rules can be used to find the
81: \shl normal form of a word constructively. We now discuss the
82: various properties that a set of rules should have in order that reduction
83: to an irreducible always
84: gives the \shl normal form of a word.
85: First we give the assumptions that we will always make about every set
86: of rules we consider. When constructing a new set of rules, we will
87: always ensure that these assumptions are correct for the new set.
88:
89: \subsection{Standard assumptions about rules.}\label{Standard
90: assumptions about rules.}
91: \begin{enumerate}
92: \item\label{reduction to epsilon}[Condition]
93: For each $x\in A$, $x.\inverse
94: (x)$ is Thue equivalent to the trivial word $\epsilon$.
95: The preceding condition
96: is enough to ensure that the set of Thue equivalence classes is a
97: group.
98: If $r=s$ is a defining relation for
99: $G$, then $r$ is Thue equivalent to $s$.
100: This ensures that the group of Thue equivalence classes is a quotient
101: of $G$.
102: \item\label{same element}[Condition]
103: If $(u,v)$ is a rule of $R$, then $u>v$ and $\overline u =
104: \overline v \in G$.
105: This ensures that the group of Thue equivalence classes is isomorphic
106: to $G$.
107: \end{enumerate}
108:
109: \subsection{Confluence.}\label{Confluence.}[Condition]
110: This property is one which we certainly desire, but which is hard
111: to achieve. Given $w$,
112: there may be different ways to reduce $w$. For example we could look in
113: $w$ for the first subword that is a left-hand side,
114: or for the last subword, or just
115: look for a left-hand side which is some random subword of $w$.
116: We say that $R$ is \textit{confluent} if the result of fully reducing
117: $w$ gives an irreducible that is independent of which elementary
118: reductions were used.
119:
120: \begin{lemma}\label{properties imply group}[Lemma]
121: If a set $R$ of rules satisfies the conditions of
122: \ref{Standard assumptions about rules.} and \ref{Confluence.} then
123: the set of $R$-irreducibles is mapped bijectively to $G$ and
124: multiplication corresponds to concatenation followed by reduction.
125: Under these assumptions,
126: an $R$-irreducible is in \shl normal form, and conversely;
127: moreover, each
128: Thue equivalence class contains a unique irreducible.
129: \end{lemma}
130: \begin{proof} The homomorphism $\Astar\to G$ is surjective and, by
131: \thref{same element}, elementary
132: reduction does not change the image in $G$. It follows that the
133: induced map from the set of irreducibles to $G$ is surjective.
134: Suppose $u$
135: and $v$ are irreducibles such that $\overline u = \overline v \in G$.
136: Then $\overline{u .\inverse(v)} = 1_G$. Therefore $u.\inverse(v)$ is
137: equal in the free group generated by $A$ (with $\inverse(x)$ equated
138: to the formal inverse of $x$, for each $x\in A$)
139: to a word $s$ which is
140: a product of formal conjugates of the defining
141: relators. Now $u.\inverse(v)$ and $s$ reduce to the same
142: word, using only reductions that replace $x.\inverse(x)$, where $x\in A$,
143: by the trivial word $\epsilon$.
144: By Condition~\ref{reduction to epsilon}, $s$ can be reduced to
145: $\epsilon$. It follows from Condition~\ref{Confluence.} that
146: $u.\inverse(v)v$ can be reduced to $v$. It can also be reduced to $u$,
147: using Condition~\ref{reduction to epsilon} again, and the fact that
148: $\inverse:A\to A$ is an involution.
149: It follows from Condition~\ref{Confluence.} that $u=v$, as required.
150:
151: The description of the multiplication of irreducibles follows from the
152: fact that multiplication in $\Astar$ is given by concatenation
153: and the fact that the map $\Astar\to G$ is a homomorphism of monoids.
154:
155: Since reduction reduces the \shl order of a word, a word in
156: \shl least normal form must be $R$-irreducible. Conversely, if $u$
157: is $R$-irreducible, let $v$ be the \shl normal form of $\overline
158: u$. Then $v$ is also $R$-irreducible, as we have just pointed out,
159: and $u$ and $v$ represent the same element of $G$. Since the map from
160: irreducibles to $G$ is injective, we deduce that $u=v$. Therefore $u$
161: is in \shl normal form.
162:
163: To show that each Thue equivalence class contains a unique
164: irreducible, we note that if there is an elementary reduction of $u$
165: to $v$, then, in case of confluence, any reduction of $u$ gives the
166: same answer as any reduction of $v$.
167: \end{proof}
168:
169: \subsection{Recursive sets of rules.}\label{Recursive sets of
170: rules.}[Condition]
171: Another important property (lacked by some of the sets of rules we
172: discuss)
173: is the condition that the set of rules be a recursive
174: set. As opposed to the usual setup when discussing rewrite systems,
175: we do not require $R$ to be a finite set of
176: rules---in fact, in this paper $R$ will normally be infinite.
177: To say that $R$ is \textit{recursive} means that
178: there exists a Turing machine which can decide whether or
179: not a given pair $(u,v)$ belongs to $R$.
180:
181: \begin{definition}\label{U}[Definition]
182: We denote by $U$ the
183: set of all rules of the form $(u,v)$, where $u > v$ and $\overline u
184: = \overline v\in G$. $U$ is called the
185: \textit{universal set of rules}.
186: Note that a word is $U$-irreducible if and only if it is in
187: \shl normal form.
188: \end{definition}
189:
190: \begin{lemma}\label{rules and word problem}
191: The existence of a set of rules $R$ satisfying the conditions of
192: \ref{Standard assumptions about rules.},
193: \ref{Confluence.} and \ref{Recursive sets of rules.}
194: is equivalent to the solubility of the word problem in $G$ and in
195: this case $U$ defined in \ref{U} is such a set of rules.
196: \end{lemma}
197: \begin{proof}
198: On the one hand,
199: if we have such a set $R$, then we can solve the word problem by reduction---%
200: according to Lemma~\ref{properties imply group} a word $w$ reduces to the
201: trivial word if and only if $\overline w = 1_G$.
202:
203: On the other hand, if the word problem is solvable, then
204: the set $U$ of Definition~\ref{U}
205: is recursive. The various conditions on a set of
206: rules follow for $U$.
207: \end{proof}
208:
209: $U$ can be difficult to manipulate,
210: even for a very well-behaved
211: group $G$ and a finite ordered set $A$ of generators, and we therefore
212: restrict our attention to a much smaller subset, namely the set of
213: $U$-minimal rules, which we now define.
214:
215: \begin{definition}\label{minimal}[Definition]
216: Let $R$ be a set of rules for a group $G$ with generators
217: $A$. We say that a rule $(u,v)\in R$ is \textit{$R$-minimal} if $v$ is
218: $R$-irreducible and if every proper subword of $u$ is $R$-irreducible.
219: \end{definition}
220:
221: \begin{proposition}\label{conditions and minimals}[Proposition]
222: \par
223: \begin{enumerate}
224: \setlength\itemsep{0pt}
225: \item\label{conditions minimals satisfy}
226: The set of $U$-minimal rules satisfies the conditions of
227: \ref{Standard assumptions about rules.} and \ref{Confluence.}.
228: In particular they are confluent.
229: \item\label{length conditions}
230: Let $(u,v)$ be a $U$-minimal rule and let $u=u_{1}\ldots
231: u_{n+r}$ and $v=v_{1}\ldots v_{n}$. Then the following must hold:
232: $0\le r \le 2$; if $n>0$,
233: $u_{1} \neq v_{1}$; if $n>0$, then $u_{n+r}\neq v_{n}$;
234: if $r=0$ and $n>0$, then $u_{1} > v_{1}$;
235: if $r=2$ and $n>0$, then $u_{1}<v_{1}$ and $u_{2} < \inverse(u_{1})$;
236: if $r=2$ and $n=0$, then $u_1 \le \inverse(u_2)$ and $u_2\le
237: \inverse(u_1)$.
238: \item\label{minimals recursive}
239: The set of $U$-minimal rules
240: is recursive if and only if $G$ has a solvable word problem.
241: \end{enumerate}
242: \end{proposition}
243: \begin{proof}
244: If $w$ is $U$-reducible, let $u$ be the shortest prefix of $w$ which
245: is $U$-reducible. Then every subword of $u$ which does not contain the
246: last letter is $U$-irreducible.
247: Let $v$ be the shortest suffix of $u$ which is
248: $U$-reducible. Then every proper subword of $v$ is $U$-irreducible.
249: Let $s$ be the \shl normal form for $v$.
250: Then $(v,s)$ is a $U$-minimal rule. Replacing $v$ in $w$ by $s$ gives
251: an elementary reduction by a $U$-minimal rule. It follows that reduction
252: of $w$ using only $U$-minimal rules eventually gives us a $U$-irreducible
253: word, and this must be the \shl normal form of $w$.
254: Therefore the conditions of \ref{Standard assumptions about rules.} and
255: \ref{Confluence.} are satisfied by the set of $U$-minimal rules.
256:
257: We now prove \ref{length conditions}. Since $u>v$ in the \shl
258: order, $|u|\ge|v|$. So $r\ge0$. If $r>2$, then $\overline u =\overline
259: v$ gives rise to $\overline{u_{2}\ldots u_{n+r}} =
260: \overline{\inverse(u_{1})v_{1}\ldots v_{n}}$. Therefore $u_{2}\ldots
261: u_{n+r}$ is not in \shl normal form. It follows that
262: $u_{2}\ldots u_{n+r}$ is $U$-reducible. Therefore $(u,v)$ is
263: not $U$-minimal. Similar arguments work for the other cases.
264: This completes the proof of \ref{length conditions}.
265:
266: Clearly $U$-minimality of a rule can be detected by a Turing machine if the
267: word problem is solvable. Conversely, if the set of $U$-minimal rules
268: is recursive, then the word problem can be solved by reduction using
269: only $U$-minimal rules.
270: \end{proof}
271:
272: Now we have a uniqueness result for the set of minimal rules.
273: \begin{lemma}\label{uniqueness}
274: Let $R$ satisfy the conditions of \ref{Standard assumptions about rules.}
275: and \ref{Confluence.}. Suppose every rule of $R$ is
276: $R$-minimal. Then $R$ is equal to the set of $U$-minimal rules.
277: \end{lemma}
278: \begin{proof}
279: By Lemma~\ref{properties imply group}, the $R$-irreducibles are the same
280: as the words in \shl normal form.
281: Let $(u,v)$ be a rule in $R$.
282: Then $v$ is $R$-irreducible and therefore in \shl normal form.
283: Also every proper subword of $u$ is in \shl normal form.
284: Therefore $(u,v)$ is in $U$ and is $U$-minimal.
285:
286: Conversely, suppose $(u,v)$ is $U$-minimal.
287: Then $v$ is the \shl normal form of $\overline u$.
288: By Lemma~\ref{properties imply group} for $R$, $u$ must be $R$-reducible.
289: Every proper subword of $u$ is already in \shl normal form.
290: It follows that there is a rule $(u,w)$ in $R$.
291: Since this rule is $R$-minimal, $w$ is $R$-irreducible.
292: Therefore $w$ is the \shl normal form of $\overline u$.
293: It follows that $v=w$. Therefore every $U$-minimal rule is in
294: $R$.
295: \end{proof}
296:
297: We are interested in those pairs $(G,A)$, where $G$ is a group and $A$ is
298: an ordered set of generators, such that the set of $U$-minimal rules is
299: not only recursive, but is in fact regular. We now explain what we
300: mean by \textit{regular} in this context.
301:
302: We recall that a subset of $\Astar$
303: is called \textit{regular} if it is equal to $L(M)$, the language accepted by
304: some finite state automaton over $A$.
305: (See Definition~\ref{FSA}, where finite state automata are discussed.)
306: We need to formalize what it means for
307: an automaton to accept pairs of words over an alphabet $A$.
308: If the pair of
309: words is $(abb,ccdc)$, then we have to \textit{pad} the shorter of the two
310: words to make them the same length, regarding this pair as
311: the word of length four $(a,c)(b,c)(b,d)(\$,c)$.
312: In general, given an arbitrary pair of words $(u,v) \in \Astar\times\Astar$,
313: we regard this instead as a word of pairs by adjoining a {\it padding symbol}
314: $\$$ to $A$ and then ``padding'' the shorter of $u$ and $v$ so that
315: both words have the same length.
316: We obtain a word over $A\cup \{\$\} \times A\cup\{\$\}$.
317: The alphabet $A\cup\{\$\}$ is denoted $A^+$
318: and is called the {\it padded extension} of $A$.
319: The result of padding an arbitrary pair $(u,v)$
320: is denoted $(u,v)^+$. A word $w \in (A^+)\uast\times(A^+)\uast$
321: is called {\it padded} if there exists $u,v \in A\uast$
322: with $w = (u,v)^+$ (that is, at most one of the two components
323: of $w$ ends with a padding symbol and there are no padding symbols in
324: the middle of a word).
325:
326: A set $R$
327: of pairs of words over $A$ is called {\it regular} if the corresponding
328: set of padded words is a regular language over the product
329: alphabet $A^+\times A^+$.
330: We say that $R$ is accepted by a two-variable finite state automaton over $A$.
331:
332: \begin{theorem}\label{automatic implies regular}
333: Let $G$ be a group and let $A$ be a finite set of
334: generators, closed under taking inverses.
335: If $(G,A)$ is \shl automatic, then the set of $U$-minimal rules is regular.
336: \end{theorem}
337:
338: Having a finite confluent
339: set of rules does not imply \shl automatic. A counter-example is
340: given in \cite[page 118]{WPiG}. So the converse of this theorem is not
341: true.
342:
343: \begin{proof}
344: Since we have a \shl automatic structure, the set $L$ of
345: \shl normal forms is a regular language.
346: If $x\in A$, the automatic structure includes the multiplier $M_x$,
347: which is a two-variable automaton over $A$. The language $L(M_x)$ is
348: the set of pairs $(u,v)$, such that $u,v\in L$ and $\overline{ux}=
349: \overline v$.
350: It is not hard to construct from the union of the
351: $M_x$ an automaton whose language
352: $P$ is the set of $(u,v)$ such that
353: $\overline u = \overline v \in G$, $u\in L.A$ and $v\in L$.
354:
355: We know that
356: $ (L.A \cap A.L)\cap (\Astar\setminus L)$
357: is a regular language. Clearly,
358: this is the set of left-hand sides of $U$-minimal
359: rules, since it is the set of $U$-reducible
360: words such that each proper subword is $U$-irreducible.
361: The set of pairs $(u,v)\in P$, such that $u$ is a left-hand side of a
362: $U$-minimal rule is easily seen to be the set of all $U$-minimal rules.
363: \end{proof}
364:
365: \subsection{Question.}
366: Suppose $(G,A)$ has a finite confluent set $R$ of \shl reducing
367: rules which define $G$.
368: Then it is easy to construct from this a finite confluent set $R'$ of
369: $R'$-minimal rules defining $G$.
370: The method is to use minimization, as described in \ref{minimization}.
371: This set of rules is equal to the set of $U$-minimal rules by
372: \thref{uniqueness}.
373:
374: Suppose now that $(G,A)$ has an infinite confluent set $R$ of \shl-reducing
375: rules defining $G$, and this set is regular. Is the set of
376: $U$-minimal rules also regular? We know that it is confluent and
377: recursive by \thref{conditions and minimals}, since $R$ provides a
378: solution to the word problem.
379:
380: If $R$ contains all $U$-minimal rules,
381: then the answer is easily seen to be \textit{yes}. The answer is not
382: clear to us if $R$ does not contain all minimal rules. There is no loss
383: of generality in making $R$ smaller so that each proper subword of
384: each left-hand side is irreducible. But we see no way of
385: changing $R$ so as to ensure that each right-hand side is irreducible,
386: while maintaining $R$'s property of being regular.
387:
388: \subsection{Objective.}\label{Objective.}
389: In this paper we present a procedure which,
390: given a set of rules satisfying
391: the conditions of
392: \ref{Standard assumptions about rules.}, changes the set of rules so that it
393: becomes ``more confluent''. More precisely, the set of words for
394: which all reductions give the same irreducible, and this irreducible
395: is in \shl normal form, increases with time. If we fix attention on
396: a single word this will eventually be included in the set. However, in
397: general, because of the insolubility of the word problem, it is not
398: in general possible to know when that time has been arrived at.
399:
400: For a group where the set of all $U$-minimal rules
401: (see Definition~\ref{U}) is the set of all pairs accepted by a
402: two-variable minimal PDFA $M$ (these concepts are defined in
403: \ref{FSA}), our procedure gives
404: rise to $M$ after a finite number of steps.
405:
406: For many undecidable problems, there is a ``one-sided'' solution. The
407: technical language is that a certain set is recursively enumerable,
408: but not recursive. For
409: example, consider a fixed group for which the word problem is undecidable.
410: Given a word $w$ in the generators, if you are correctly
411: informed that $\overline w = 1_{G}$, then this can be verified by a
412: Turing machine. All that you have to do is to enumerate products of
413: conjugates of the defining relators, reduce them in the free group
414: on the generators, and see if you get $w$, also reduced in the free
415: group. If $w$ represents the identity then you will prove this
416: sooner or later. If it's not the identity, the process continues for
417: ever.
418:
419: We know that there is no algorithm which has as input a finite
420: presentation of a group and outputs whether the group is trivial or
421: not (see \cite{Rabin:unsolvability}).
422: It follows easily that there is no algorithm which has as input a
423: finite presentation and outputs either an FSA accepting the set of
424: $U$-minimal rules or correctly answers \textit{There is no such FSA}.
425: For, in the case of the trivial group,
426: the set of $U$-minimal rules is finite---for each element $x\in A$, we
427: have the rule $(x,\epsilon)$---and so it is certainly regular.
428:
429: But the situation is even worse than this. We do not even know of a
430: one-sided solution to the problem of whether the set of $U$-minimal
431: rules is regular. If the set of $U$-minimal rules is regular,
432: our procedure will eventually produce a candidate
433: with some indication that it is correct, but we will not know \textit{for sure}
434: whether the answer is correct or incorrect.
435:
436: What is at issue is whether there is an algorithm which has as its
437: input a regular set of \shl rules for a group
438: and outputs whether or not the set of rules is confluent.
439: For finite sets of rules the question of confluence is
440: decidable by classical critical pair analysis
441: which we describe in \thref{Standard Knuth--Bendix}.
442: However, for
443: {\it infinite} rewriting systems the confluence question is, in
444: general,
445: undecidable. Examples exhibiting undecidability are given in
446: \cite{InfiniteRegularThue}.
447: They are length-reducing rewriting
448: systems $R$ which are regular in a very strong sense:
449: $R$ contains only a finite
450: number of right-hand sides and for each right-hand side $r$, the set
451: $\{l : (l,r) \in R\}$ is a regular language.
452: These examples are in the context of rewriting for monoids.
453: As far as we know,
454: there is no known example of undecidability if we add to the hypothesis
455: that the monoid defined by $R$ is in fact a group.
456:
457: In the special case where $(G,A)$ is \shl automatic, there
458: is a test for confluence of a set of rules satisfying the conditions
459: of \ref{Standard assumptions about rules.},
460: namely the axiom-checking procedure described in theory in
461: \cite{WPiG} and carried out in practice in Derek Holt's
462: \textit{kbmag} programs \cite{Holt:KBMAG}.
463: %$Log: Sec2.tex,v $
464: %Revision 1.12 1999/12/16 19:55:02 dbae
465: %shortening for IJAC
466: %
467: %Revision 1.12 1999/06/01 16:17:48 dbae
468: %*** empty log message ***
469: %
470: %Revision 1.11 1999/06/01 10:00:12 dbae
471: %version from goliath
472: %
473: %Revision 1.11 1999/05/30 19:43:06 dbae
474: %finished this version
475: %
476: %Revision 1.10 1998/11/24 12:47:05 dbae
477: %done for now
478: %
479: %Revision 1.9 1998/11/24 10:07:20 dbae
480: %*** empty log message ***
481: %
482: %Revision 1.8 1998/11/18 16:32:37 dbae
483: %*** empty log message ***
484: %
485: %Revision 1.7 1998/11/16 17:17:08 dbae
486: %*** empty log message ***
487: %
488: %Revision 1.6 1998/11/15 21:20:05 dbae
489: %*** empty log message ***
490: %
491: %Revision 1.5 1998/11/13 11:48:51 dbae
492: %compiles correctly
493: %
494: %Revision 1.4 1998/11/12 11:05:15 dbae
495: %*** empty log message ***
496: %
497: %Revision 1.3 1998/11/11 21:34:27 dbae
498: %fixed up bib trouble
499: %
500: