Discrete Applied Mathematics 5 (1983) 279-297 279

North-Holland Publishing Company

ON CUBE-FREE o-WORDS GENERATED BY

BINARY MORPHISMS

Juhani KARHUMIiKI
Department of Mathematics, University of Turku, Turku, Finland

Received 14 January 1981

Revised 6 October 1981

Let h be a morphism satisfying h(a) = ax for a letter a and a nonempty word x. Then h defines
an infinite word (an w-word) when applied iteratively starting from a. Such w-words are
considered in a binary case. It is shown that only biprefixes can generate cube-free w-words, i.e.
words which do not contain a word u’, with ufl, as a subword. The same does not hold true
for fourth power-free w-words, the counterexample being the w-word defined by the Fibonacci-
morphism: h(a) = ba, h(b) = a.
As the main result it is proved that it is decidable whether a given morphism of the above form
generates a cube-free w-word. Moreover, it is shown that no more than 10 steps of iterations are
needed to solve the problem.

1. Introduction

Repetitions in words, i.e. the existence of occurrences of ui, with of A and i 22,
as subwords was first studied by Thue in [lo] and [ll]. He proved, among other
things, that there exist infinite words over a binary alphabet such that these words
do not contain cubes at all. In other words he proved the existence of an infinite
cube-free word over a two-letter alphabet. In the case of a three-letter alphabet he
proved the existence of an infinite word without any repetitions, i.e. the existence of
a square-free infinite word. On the other hand, any word over a binary alphabet and
with the length at least four contains a square.
Later on Thue’s results have been rediscovered several times in different con-
nections. As an overview we refer to [9]. In recent years the research on this field
initiated by Thue has been active. Many of his results have been generalized and a
better understanding about repetitions in words has been achieved, see e.g. [ 1,3-61.
However, many problems are still unanswered.
The Thue’s method to construct an infinite cube-free word over a binary alphabet
was that of iterating a morphism. He considered the morphism defined by h(a) = ab
and h(b) =ba. When applied this iteratively the following sequence is obtained:
a, ab, abba, abbabaab, abbabaabbaababba, . . . Continuing ‘ad infinitum’ a cube-free
infinite word is defined.
In this paper we also consider infinite words, or o-words in our terminology,

0166-218X/83/0000-0000/$03.00 0 1983 North-Holland

280 J. Karhumiiki

obtained by iterating a morphism. Such an o-word is always defined if a morphism

is of the form h(a) = ax, with x# A, for some letter a. Such morphisms are called
prefix-preserving morphisms, or pp-morphisms in short. We restrict our considera-
tions to a binary case.
After preliminaries we establish in Section 3 some simple properties of u-words.
Especially, we show that if a pp-morphism defines a cube-free o-word, then it is
necessarily a biprefix.
In Section 4 we deal with the Fibronacci-morphism: h(a) = ba, h(b) = a. We show
that the w-words (two from left to right and one from right to left) contain cubes,
but are fourth-power free. Hence, a morphism generating a fourth power-free
o-word need not be a biprefix.
Our main result is proved in Section 5. We show that it is decidable whether a
given o-word generated by a pp-morphism is cube-free. The proof is carried out by
the sequence of lemmas. The result should be compared to that of Berstel, see [2],
which says that it is decidable whether an o-word generated by a pp-morphism over
a three-letter alphabet is square-free.
In Section 6 our main result is sharpened. We present an explicit upper bound for
the number of applications of h needed to guarantee the existence of a cube if it will
ever occur. Moreover, we prove that 10 is such an upper bound in general, and that
in some special cases it is even improved. For instance, for morphisms satisfying
min{lh@)l, Ih(b an upper bound can be chosen to be 7. On the other hand,
the Fibonacci-morphism shows that, in general, at least 6 iteration steps may be
needed.

2. Preliminaries

For the purpose of this paper let Z be a binary alphabet, say Z = {a, b}. The free
monoid generated by _Z is denoted by Z* and its identity, the so-called empty word,
by A, Z+ = Z* - {A}. Elements of Z* are called words. For the length of a word x we
use the notation 1x1, specifically )A I = 0. If a word u is a prefix of a word u we write
u< u. We say that a word u is a subword in U, or alternatively a segment in U, if
u = xuy for some x and y in Z*. The notation x-‘y (resp. _YX-‘) denotes the left (resp.
right) quotient of y by x. Finally, by an o-word we mean an infinite sequence of
elements of Z (from left to right).
A word or an o-word is called cube-free (resp. square-free, fourth power-free) if
it does not contain a segment of the form uuu (resp. uu, uuuu), with of A. By an
almost cube we mean a word of the form ucucud or ducucu, with u E Z* and c, d E EC,
cf d. So a word abaabaabb is cube-free, but contains several almost cubes, namely
baa, aab,abb and abaabaabb. On the other hand, babaabaabab is not cube-free
since it has a subword (aba)3.
Our central notion is that of a morphism of X*. We consider only A-free
morphisms, i.e. morphisms for which i $ h(2). Let h(a) = (Y and h(b) = /?. We say
 On cube-free w-words 281

that h is a prefix (resp. suffix) if neither a is a prefix (resp. suffix) of /3 nor fl is a

prefix (resp. suffix) of a. A morphism which is both a prefix and a suffix is called
a biprefix.
The following notion of a cut is important in our considerations. Let a, PEZ’
andxE{a;p}*. Ifx=yr ... yl, with y; E {(L;fi) *, we say that the positions determined
by the prefixes yr *.*yj, j= 1, . . . . t - 1, in x are @-cuts, or shortly cuts if a and p are
clear. We denote cuts as [xl, [rr yz] or [yz]. Intuitively, the cuts tell how a word is
composed from its ‘bricks’. Observe also that, in general, cuts are properties of the
whole word, and hence to see whether a given position in the middle of the word is a
cut may require to check through the whole word and not only some middle
segment.
For us a and p will be images of a morphism, say a= h(a) and /3 = h(b).
Moreover, h will be a code so that cut points in each word h(x),x~Z*, are unique.
Let w E h(Z*), u E Z* and w =yuz for some y, o and z in Z*. Then we write

u Cy,z h(u) iff u = h(u) and y, z E h(Z*)

and
[ol e h(Z*) iff y, z E h(Z*).

The explanation for the relation o=,, h(u) is as follows: It tells that both the edges
of u are cut points with respect to (h(a), h(b)} in the word yoz. Moreover, it defines
how the segment u is obtained as the image under h. For simplicity we will normally
write u = h(u) instead of o =Y,zh(u) if there is no danger of confusion.
Let h be a morphism of Z* such that

h(a) = ax with xf A, (1)

i.e. a is a proper prefix of h(a). Then h(a) is also a proper prefix of h*(a) since
h*(a) = h(ax) = h(a)h(x) and, in general,

h’(a) = h’-‘(ax) = h’-‘(a)/~-r(x) for ir 1,

i.e. h’-‘(a) is a proper prefix of h’(a). Hence continuing ‘ad infinitum’ we get an
w-word. A morphism satisfying (1) or the analogous condition for b is called a
prefix preserving morphism, or a pp-morphism in short.
Prefix-preserving morphisms provide a very convenient way of defining o-words.
Indeed, an o-word is obtained simply by iterating a morphism. We refer u-words
obtained in this way to co-words generated by morphisms. In conclusion, we want to
remark that what we really have above is a DOL system (2, h,a), where h has the
prefix condition. Moreover, the way how DOL systems generate words is exactly
that of iterating a morphism. Hence these systems provide a very nice framework to
study o-words.
282 J. Karhumiiki

3. Some simple properties of cube-free words

In this section we are looking for properties satisfied by cube-free words and
w-words over a binary alphabet. As shown by Thue [ll] there exist cube-free
w-words in a binary case. However, such words have some special properties.

Lemma 1. Any cube-free word over a binary alphabet and of the length at least 17
contains aa and bb as subwords.

Proof. The step by step generation of all cube-free words which do not contain the
word aa as a subword yields the following two (one starting from a and the other
from b) terminating trees shown in Fig. 1.
a-b I
/
,aI ,a-b-b-;-;-;I;;

a-b-b-a-b-b-a)
a-b/
/
‘b-a -b-a-b-b-a-b-b-al

\b-a-b-a-b-_b-a-b~a-b-b’
‘b -a I

a-b 1
a-b-b-a-b’
‘b -a I
a-b-b-a-b-b-al
/
a-b-b1
a-b-b-a- b/
a-b-b-a-
/
b/ \b-a )
b/ ‘b a-b-b-a-b-b-al
lb--a-b-a-b-b=iLb/
\b-a 1 a-b-b I
\ b-a-b-a-b-b-a-b’
lb -a I
\bI

Fig. 1.

Above ( denotes that the continuation is not at all possible in such a way that the
required properties would remain. Because the longest sequences are of length 16
the lemma follows.

Corollary 1. Any cube-free w-word over a binary alphabet contains both aa and bb
as subwords.

From the above we obtain

Theorem 1. Any pp-morphism generating a cube-free o-word is a biprefix.

Proof. Assume the contrary that a pp-morphism h, which is not a biprefix,

 On cube-free o-words 283

generates a cube-free o-word K. Let h(a) = c$ and h(b) = (r for some cxand p in .P,
i.e. h is not a prefix. By Corollary 1, R contains bb as a subword and hence also Ma.
But h(bba) = a3/3, which shows that K is not cube-free, a contradiction. By symmetry
in Corollary 1, the cases that h(a) is a prefix of h(b) or that h is not a suffix lead
similarly to a contradiction.

Analogously to Lemma 1 one can show that any cube-free o-word, or in fact any
word of the length at least 24, contains both aba and bab as subwords. Conse-
quently, morphisms of the form h(a) = a/3a and h(b) =/3 never generate cube-free
o-words.

4. The Fihonacci-morphism

In the previous section we proved that cube-free o-words over a binary alphabet
can be regenerated, if at all, only by biprefixes. Here we show that the same does not
hold true for fourth power-free o-words. The result is obtained by studying the
famous Fibonacci-morphism.

Example 1. Let h be a morphism defined by

h(a) = ba, h(b) = a.

It is well-known that the lengths of words obtained iterating h starting at b gives the
Fibonacci-numbers. Although h itself is not a pp-morphism h* is such. So the
iteration of h gives two w-words. Observe also that h has a so-called suffix
condition, i.e. (I is a proper suffix of h(a), and hence h generates a unique u-word
from right to left. In conclusion, h can be used to define altogether three o-words.
The iteration initiates as follows:

a + ba + aba -+ baaba -+ ababaaba + baabaababaaba

+ ababaababaabaababaaba + -.. ,

where (aba)3 appears as a subword in the last written word. Hence, the o-words
obtained are not cube-free, as it must be by Theorem 1.

In what follows we will demonstrate that the o-words generated by h do not

contain fourth powers. So we will consider the sequence

(h’W)),,o. (1)
The basic idea behind the proof is that we show that if in (1) there are long enough
fourth powers there must be shorter ones, too. Repeating the argument we conclude
that in (1) there must be short fourth powers, which, finally, can be shown to be
impossible.
284 J. Karhumtiki

Assume that u”, where n 23, is a subword in (l), say yu’k is in (1). We have the
following four possibilities

[u”], [~“a], [b&z] or [bu”] E h(Z*).

We consider each of these separately.

Case I: [u”] E&E*). Clearly, there exists a word u such that, for the first
occurrence of u in u”, h(u) = u or h(u) = ua depending on whether u ends with a or
with 6. In the first case h(P) = u” since h is a code. So (1) contains a shorter cube
since the case USA+ is impossible (cf. the end of this proof). The second case is
impossible, since now u would end with b and hence u” would not be in h(Z*).
Case II: [u”a] E l&Y*). By Case I we may assume that [u”] $ h(_Z*). Hence u = u,b
for some ul EE*. Since [u“mLulbulba] E h(Z*), u1 must start with a, i.e. u =au2 for
some u2 in C*. The relation [(auZ)n-lauZa] E h(Z*) now implies that u2a (any occur-
rence of it) is an image under h, say u2a = h(u). Hence, h(P) = (u2a)n showing that a
shorter nth power has been found.
Case III: [bu”a] E /I@*). We assume that [bun] $ I@‘*), i.e. we do not include
Case IV here. As in Case II we deduce that u ends with b and starts with a, say
u = au3 b for some u3 in Z*. Hence bu”a = b(au3 b)“a = ba(u3 ba)“. So (u3 ba)” must be
in h(Z*) and therefore there exists a word u such that h(u) = u3 ba. Consequently,
h(P) = (u3ba)” yielding a shorter nth power.
Case IV: [bun] E h(Z*). First we observe that u cannot end with 6. Further u must
start with a. Hence either u = a or u = au4a for some u4 in Z*. The first possibility is
impossible, since (1) does not contain a word aaa as a subword (cf. the end of this
proof). In the second case we write u5 = u4a and obtain

bun= b(au4a)n= [ba]([us][a])“-1[u5].

(Indeed, there must always be a cut after a). Consequently, there exists a word U’
such that

bu”= h(a(u’b)“- ‘u’). (2)

Now a(u’b)“-lu’ is a segment in (1). If u’#L it must both start and end with a.
Writing u’= au;, we obtain

a(u’b)n-lu’=a(au~b)“~lu’=a[a]([u~][ba])”-l[u~].

Therefore

(U’b)“-lu’Eh(b(u”a)“~‘u”)

for some word u”. Again b(u”a)“-l U” is a subword in (1) and hence, if u”#A, then it
must start with a and end with 6. This last observation follows since the sequence (1)
does not contain a word aaa as a subword. Now writing u”=au;b we have

b(z4”a)n-1Un= b(au;ba)“-‘aurb= [ba]([U~][ba][a])“-‘[u;][b.

Remember now that the above word is a subword in h(Z*). Hence its right
 On cube-free w-words 285

neighbour is a. Consequently,

b(u”~)“-‘u”~~h(a(u”‘b)“-‘u”‘) (3)

for some word u”‘. Comparing the right hand sides of (2) and (3) we see that we have
got a cycle, however, with a shorter u-word.
What has been proved above can be summarized to the following diagram (see
Fig. 2) which shows how a shorter word (nth power or almost nth power) is
obtained. The reduction can be carried out as far as the primed u-words are non-
empty or a nonprimed u is of the length at least 2.

\
’ un ~ a("-b)"-'u* with IuI , Iu-1 < IV 1
/
/ IV

I u
I b(u--a)"-'u-* with Iuss l< I u-1
\
IV
\
\ u
\ n-l
a(u*--b) us** with Iuss+ I< Iu~-I
Fig. 2.

Now we are in the position to obtain our result. Assume that the sequence (1)
would contain a fourth power. Then, by above diagram, it would contain either
a4,b4, ba3 or ab3 as a subword, too. But this is impossible. Indeed, the sequence
(/~“(a)),,~ satisfies the recursion formula

hn+2(a) =h”(a)h”+ ‘(a) for n r0

and hence
{h”(a) 1n 20) c {a} u {h,aba}*,

which certainly does not contain any of the words a4, b4, ba3 or ab3 as a subword.
Now we are ready to state our result which should be compared to Theorem 1.

Theorem 2. Fourth power-free co-words can be generated by a nonprefix (non-

suffix) pp-morphism over a binary alphabet.

5. The main result

Here we establish our main result.

286 J. Karhumiiki

Theorem 3. It is decidable whether a given pp-morphism over a binary alphabet

generates a cube-free o-word.

The proof is carried out by the sequence of lemmas. To fix the notation let h over
{a, b) * be a pp-morphism, with h(a) = ax for some xf A, defined by

h(a)=@ and h(b)=/?

and let
L={h”(a) InrO}.
If h is not a biprefix, then, by Theorem 1, it does not generate a cube-free o-word.
Hence, we assume that h is a biprefix.
Basically, we will show that if h generates cubes it must generate short cubes (or at
least short almost cubes), too. So it remains to be checked whether any cube from a
finite set of words is a subword of a word in L, and this is easily seen to be
decidable.
We continue with

Lemma 2. Let v3 be an occurrence of a cube in L, say uv3w E L, such that it contains

both a and b. Moreover, let for some words x1, x2, x3, y1 and y2

with
v=xlyl =Y2x2Yl =Y2x3 and YlY2E 4-O

Then there exists a word v’ such that either (i) v’v’v’ is a subword in L and 1v’ 1c 1v 1,

or(ii)v’cv’cv’d, withc,dE{a,b}, c#d, isasubwordinLand Iv’cI<JvJ.

Proof. The assumptions of the lemma can be illustrated as follows:

u , v I V I
V 1
W
L A’ A Al A I
Y x2 y
x1 X3

where y = yI y2 is either a or /I, say y = a. Then for both the occurrences of y we have
y= h(a), and so there exists a word v1 such that x2= h(vl). Recalling now that h is a
biprefix we deduce that the prefix of x3 with the length 1x21 is also the image of vi
under h and that the same holds true for the suffix of xl with the length 1x2 I.
In conclusion, we get that

h(vlavlavl)=y;‘vvvy~‘. (1)
If either yl = I or y2 = A, then we are ready since in that case the other y would be y
and hence the argument on the left side of (1) could be continued to a cube. SO let
yl#A. By (1) and by the fact that UV~WEL, either a< ylw or /3-Cylw. In the first
case the result follows when we choose v’= via. In the second case an almost cube is
obtained by the choice v’= vi.
 On cube-free w-words 287

The requirements concerning the lengths of v’ follow since for pp-morphisms

max{ 1h(a) I, 1h(b) I} 12 and since o contains both a and b.

The above lemma can also be formulated for almost cubes.

Lemma 3. Let vcvcvd, with c, d E {a, b}, c# d, be a subword in L, say uvcvcvdw E L.

Moreover, let for some words x1, x2, x3, yl, y2 and y3

uvcvcvdw= ~~~l~~YlY21~~21[Y1Y21[~3w~
with
“c=xlYl =Y2x2Yl> vd=y2x3 and yly2~ h(Z).
Then there exists a word v’ such that v’c’v’c’v’d’, with c’,d’~ {a, b}, c’fd’, is a
subword in L and 1D’c’J c I v I.

Proof. The situation can now be illustrated as follows:

/I-- Y
Xl X2 y x3

If y1 #A, then we conclude as in the proof of Lemma 2 that for some word v1 and c’
in {a, b} we obtain

Now the facts that uvcvcvdwE L and that h is a prefix guarantee that
h(d’)< y,c-‘dw, with d’#c’, implying the existence of a shorter almost cube. The
case y1 =A can be handled similarly.

Intuitively, the message of the above lemmas is as follows. If a cube (or an almost
cube) exists in L and if the border lines of that cube (or that almost cube), i.e. the
positions illustrated by ) in the formula ... v Iu (II ---, are covered by the same
h-image and without any shift in the @decomposition of the whole word, then a
shorter cube (or a shorter almost cube) can be found.
In what follows we will show that if L contains cubes (or almost cubes), then the
assumptions of Lemmas 2 and 3 are satisfied. Or more precisely, we will show that if
L contains cubes at all, then either the above assumptions are satisfied for cubes and
almost cubes long enough or otherwise L contains very short cubes, too.

Lemma 4. Assume that the shortest cube in L is of the length at least 6 ICrp1, and let
v3 be such one, say uv3w E L. Moreover, let zl, z2 and 213be words such that

U”““W = kllbl[z2l[Yl[z3l~
where
Y E h(z), z,< uv< zly and zlyz2< WV< z~yz~y.
Then
luvl- Iz~l= luvul- IZIYZZI (mod IrIb (2)
288 J. Karhumtiki

Proof. By symmetry, we may assume that y = a. Assume further that (2) does not
hold true.
Now our situation is as follows:

“aw (3)
a a
Let a= ataz= a3a4 where uu =ztot and uuu = UUCZ~Z~OI~. Clearly, at most one of the
a;‘s may be empty, otherwise there is nothing to be proved. Further, by symmetry,
we may assume that 1al I> 1a3 1. Again if Icc1I = /(~31, then we already have (2).
We have two cases to be considered.
Case I: /a11 - Ia3 Is+ Ial. We consider the two different ways how the middle
part of uu may be written. We have:

Let E be the prefix of a of the length Iat / - Ia3 I as indicated in the above figure.
Then a has a prefix c2 and thus u2 has a cube e3 as a subword, a contradiction.
CaseII: lall-la31>+1al. Nowwehavethreesubcases.
(i) The first occurrence of a in (3) is followed by a. In this case the illustration is
as follows:

As in Case I let E be the prefix of a of length I al I - I a3 I and let o = c-la. Then we

have EO = oe which implies that E and cr are powers of a common word. Hence, u2
contains a cube, a contradiction.
(ii) The first occurrence of a in (3) is followed by p, and moreover
IPl~l~~l-l~31. N ow we have two possibilities depending on whether j3 or a is
followed by the second occurrence of a in (3). These can be illustrated as follows:

or
 On cube-free w-words 289

The first possibility leads to a contradiction exactly as the main Case I, only the
word ab must be considered instead of a. The second possibility means that a’s on
the lower line must precede /?, since three consecutive a’s is impossible. This, in
turn, implies that a on the upper line must precede a, otherwise we would have the
mirror image situation of the first possibility of this subcase, and it is not possible.
Finally, remembering again that three consecutive a’s is impossible we conclude that
we must have:
B CL a B
I - V i I r J

vv:
I
c A ’ A A J

B a a B

Consequently, the argument of the main Case I becomes applicable for the word
paa/?, which completes the subcase (ii).
(iii) The first occurrence of a in (3) is followed by p and ID I< [al I- 1a3 / . NOW
our illustration is:
c( %

The a on the lower line can not be preceded by a, because of the argument in Case
II(i). Hence, it must be preceded by j3.
We now note that if lall-la31-l~[z+laj, t h en, by the argument in the main
Case I, the /3 on the upper line can not be followed by a. So if a follows that /3 we
must have the situation:

““: : (4)
%
% c(

where the dot line denotes the middle of the lower a. So the upper p goes beyond it
to the right (but not beyond the lower a). We consider now the two occurrences of j3
inside the word afl. Observing that the lower a must be followed by /3 this really can
be done and we get the illustration:

The argument of the main Case I applied to pa shows that p can not be shifted at
all, i.e. both the upper and the lower j3 above must start at the same place. Hence,
290 J. Karhumtiki

(4) shows that cr= aij?oi, for some word czi, and thus (alp)3 is a subword in u*, a
contradiction.
So it remains the case that a on the upper line is followed by two p’s, and, by
symmetry, (Yon the lower line is preceded by two /3’s. Since three consecutive p’s is
impossible, we have the situation
CL B a c1
I I
I
r r Y -I

VV:
I
c ” A 6 A’ J
a B B 0.

Again the argument of the main Case I is applicable, now it must be applied to the
word a/I/k. So we have finished the main Case II, too.
Looking through the above considerations one obtains an lower bound for the
length of u. Certainly, 2 1a/? 1 is such one, which completes the proof of Lemma 4.

Using exactly the same arguments as in the proof of Lemma 4 we can show

Lemma 5. Assume that the shortest cube in L is of the length at least 6 1a/l 1, and let
vcvcud, with c, d E (a, b}, cf d, be a subword in L, say uvcvcodw E L, such that
1vcIz2 )a/3 I. Moreover, let zI, z2 and z3 be words such that

uucocudw = k11M k21 [VILz31,

where
Y E h(-O zl( uvc< zly and zlyz2< uvcvc< zlyz2y.
Then

Lemmas 2 and 4 (resp. 3 and 5) show that for cubes or almost cubes of the length
at least 6 )a/? I in L there exist shorter ones, too, if only the two border lines of u3, i.e.
the positions indicated by I in the illustration 0.. u 1o ) u a.-, are covered by the same
h-images.
Consequently, it remains to be considered the case where the border lines are
covered by different h-images.

Lemma 6. For a cube of the length at least 6 Iap I + 3 max{ )a I, //I I > in L there exists
a shorter cube or a shorter almost cube.

Proof. We use the proofs of Lemmas 2 and 4 quite heavily. Let u3 be a cube in L
with I u / L 2 Ia/3 ) + max( IaI, I/3 ) >. By the discussion before it is enough to consider
the case where the border lines of u3 are covered by different h-images. In other
words, we have the situation:
 On cube-free o-words 291

where y # 6. Let y = a, i.e. the first border line is covered by CX.By symmetry, we
may assume that lal>l/_I.
Now we recall the proof of Lemma 4. There it was not essential that we used those
occurrences of (Ywhich cover the border lines of u3. We only needed that the two
occurrences of cz were in the shifted position to each other and that they were far
enough from the edges. This in mind we have to consider the following two
possibilities
Case I. The word a covering the second border line of u3 (but which is not
forming the alp-decomposition) contains a cut meeting a, i.e. we have, for example,
the illustration:

where only the lower a’s satisfy the property [a] E h(Z*). In that case we do not have
any problems: all the conclusions in the proof of Lemma 4 are applicable using these
two a’s, since the length of u is assumed to be at least 2 )CrpI+ maxi 1al, 1/I I },
Case II. The word a covering the second border line of u3 does not contain cuts
meeting a, i.e. we have, for example, the situation:
n

Indeed, we must have a cut inside the upper a, since a was assumed to be the longer
h-image. The first occurrence of a must be surrounded by /3 at least on the one side.
Hence, we have two occurrences of p in the shifted position and with the distance
not larger than max{ Ia/, IpI} f rom the border lines. Consequently, the proof of
Lemma 4 becomes applicable now, too.

As earlier we again have the analogous result for almost cubes.

Lemma 7. For an almost cube of the length at least 6 IOrpI + 3 max{ Ia /, I/3 I } in L
ther exists in L a shorter cube or a shorter almost cube.

We still need one lemma. For the notions of a regular language, a DOL language
and EOL language we refer to [8].

Lemma 8. Given a DOL language L and a word x. It is decidable whether x occurs

as a subword in L.

Proof. Let R be a regular language Z*x.Z’*. Then x occurs as a subword in L if and

only if RrTL is nonempty. But R nL is an EOL language and for this family the
emptiness problem is decidable, see [8].
292 J. Karhumiiki

Our proof for Lemma 8 was short and it used unnecessarily complicated language
families. For our purposes better, but longer, proof of the result is presented in the
next section.
Now we are in the position to establish our main result.

Proof of Theorem 3. By Lemmas 2-7 if L contains a cube at all, it contains also a

cube or an almost cube of length shorter than 6 Ia/? I+ 3 max{ 1al, lp I >. Let

L,=(xE{a,b}*~ x is a cube or an almost cube with

lx1<6 I4+3max{l4~ IPIH.

Clearly, L, is finite. Hence, by Lemma 8, we may test whether L contains any
subword from L,. If a cube is found in this process, we are done: L contains a cube.
If, on the other hand, no word from L, is a subword in L, then we are also done: L
does not contain a cube.
So it remains the case when L contains as a subword an almost cube from L, (and
no cubes from L,). To complete the proof we must show how to decide whether such
an almost cube is obtained or not from a cube according to the proofs of Lemmas 2
and 3.
Let y=u’c’o’c’o’d’~L,, with c’ and d’ in {a,b}, c’fd’ and lyl<6laPI+
3maxM lPl> b e a subword in L. We may also assume that Ih(y)I 2 6lopI +
3 max(loJ, j/3\} so that the results of Lemmas 2 and 3 become applicable. Now we
recall the proofs of these lemmas. If y is really obtained according to these from
a longer cube or from a longer almost cube, then this longer word is determined by
h(o’)yh(o’)yh(u’)d, where y = h(c’) and 6 = h(d’). Moreover, the middle third of this
longer cube or almost cube is of the form ohs, where ~a = y. Consequently,
there are only a finite number of possibilities to be checked through. To be able to
do this the left neighbour of y (if it exists) must be known, which, of course, can
be assumed. The illustration of the above is as folows:

h(v-) Y h(v-1 Y h(v-) 6

I I I I I
t-- -I ’ A ’ ’ A ’

Here A is used to denote the middle third of a longer cube or almost cube we are
searching for. If the above leads to a cube, then we are done: L contains a cube. If,
on the other hand, even an almost cube is not found, then this particular y is not
obtained from a cube according to Lemmas 2 and 3. Finally, if an almost cube is
found the process can be repeated. Now the basic observation is that the searching
of a longer cube or almost cube is independent of u’, it depends only on letters c’ and
&and on the left neighbour of y. This means that, if the process does not terminate,
then it leads to a cycle in one or two steps.
This completes the proof of Theorem 3.
 On cube-free o-words 293

6. An effective upper bound

In this section we are looking for an upper bound for the number of iterations
needed to guarantee the existence of a cube in the sequence (h”(a)),,e if the sequence
contains cubes at all.

Lemma 9. If h is a pp-morphism which is not a biprefix, then it generates a cube in

7 steps.

Proof. Let h(a) = ax, with xf A. So 1h(a) 1L 2. If h(a) E a*, then h generates a cube in
two, and hence also in seven steps. So assume that h(a) $a*, i.e. both a and b occur
in h(a). If h(b)E b*, then necessarily h(a)=uh(b), for some word U, and thus h
generates a cube in three steps. There remains the case h(b) E Z*aZ*. In that case

1h”(a) / 2 (n + 1)st Fibonacci number,

i.e. the nth number in the sequence 1,2,3,5,8,13,21, . . . . So it follows that h

generates in six steps a word of the length at least 17, i.e. a word which contains, by
Lemma 1, a cube or words au and bb as subwords. But h is either a prefix or a
suffix. Consequently, h’(a) contains a cube.

The above upper bound is not far from the optimal one. Indeed, for the
Fibonacci-morphism a cube is obtained only after six iteration steps.
We need also the following two simple lemmas.

Lemma 10. Anypp-biprefix h, with Ih(a)1 = 1 or /h(b)1 = 1, generates a cube in two

steps.

Proof. Let h(a) = ax, with xf A. So h(b) = b, and thus h(a) = axa for some word x. If
xrsa*UaX*UZ*a, then a cube is obtained in two steps. If XE b+, say x= b’, then
h2(a) = h(ab’a) = ab’ab’ab’a which contains a cube. If x = bbx’bb for some word x’,
then necessarily x’ = ax”a, for some x”, or otherwise a cube is obtained in one step.
However,

h2(a) = h(abbux”abba) = abbax”abbabbabba ..a

and so a cube is obtained in two steps. The remaining case is x= abax”‘ba (or sym-
metrically x = abx”‘aba) for some word x”‘. Also now a cube is obtained in two steps:

h2(a) = h(abax”‘ba) = abax”bababax”’ . . . .

This ends the proof of Lemma 10.

Lemma 11. Let h, with h(a) =ax, x+1, be a pp-biprefix which does not generate a
cube in three steps. Then it generates all the subwords of the length three which it
will ever generate in at most three steps.
294 J. Karhumiiki

We left the simple but quite lengthy proof for the reader. We only want to point
out that an easier result, namely the result where it is required that the number of
iterations is five is trivial. Indeed, any subword of the length three which is
generated by a pp-morphism is generated in one step from a word of the length at
most two, and all the subwords of the length two are generated in not more than
four steps, if at all. This last observation is based on the fact that there are no more
than four words of the length two.
Now we are ready for

Theorem 4. A pp-morphism h, with h(a) = ax, x# ,I, generates a cube-free co-word if

and only if the set {h”(a) )n 5 lo} is cube-free.

Proof. By Lemma 9, we may assume that h is a biprefix. Moreover, by Lemma 10,

let /h(a)1 r2 and /h(b)1 22. If Ih(a) I = Jh(b)( = 2, then the possibilities are: either
h(a) = aa or h(a) = ab and h(b) = aa, bb or ba. It is easily seen that the theorem is
valid for each of these possibilities. Consequently, we may assume that either h(a) or
h(b) is of the length at least three.
By the considerations in the previous section we know that if h generates cubes at
all it will generate such of the length at most 6 Iclp I + 3 max{ Ia I, I/I I } or an almost
cube of the same length. Thus our problem is to try to find an upper bound for the
number of applications of h to guarantee the existence of a given word of this length
as a subword if it will ever occur.
Let lal>lflI an d assume that yO, with IyOl 16 la/II +3 max{ Ial, IpI>, occurs as a
subword in L. Let further y1 be a subword of L such that h(yl)=zly0z2 for some
words z1 and z2 with /zi I, lz21 < /oI. We may further assume that yl does not contain
cubes of the length at most 9. Indeed, if it does, then we are not interested in the
above mentioned upper bound, but an upper bound which guarantees the existence
of such short cubes in L, and the end of this proof shows that 5 is now a suitable
upper bound.
Assume now that yI contains at least 13 a’s. Then it must contain at least 7 b’s.
Indeed, by the choice of yl the number of b’s preceding the 6th occurrence of a is at
least three and hence that of b’s preceding the 12th occurrence of a is at least 6. So
yi contains at least 7 b’s. Now we have:

which shows that yi contains at most 12 a’s. To estimate the number of b’s in y1 we
note that the number of b’s preceding the third occurrence of a is at most 5. Hence
an upper bound for the whole number of b’s in yI is 22. Consequently,

lYll534.

Now we consider the ancestors of yl, i.e. the minimal subwords y2, ~3, . . . of L
such that y1 is a subword in h( y2), h2( yj), . . . . See Fig. 3.
 On cube-free w-words 295

/I A
Yl
Fig. 3.

Remember now that h(a) and h(b) are of length at least two and that at least one
of them is of length at least three. Hence, we can estimate the lengths of words in the
sequence Yl, y2, y3, y4, y5 as follows:

lY2l517, IY31’9, lY4115, IYsl53.

Now we are ready to finish the proof. By Lemma 11, a word of length three is
obtained as a subword in no more than three steps, if at all. So our chase of
ancestors of y. is complete. We needed altogether 1 + 4 + 3 (or less) iteration steps.
To guarantee the existence of a cube, and not only an almost cube, two more steps
are enough, by the proof of our main theorem. Hence the limit 10.

We want to finish this section by sharpening the above limit in some special cases.
Let h be a pp-morphism, with h(a) = ax, x f A and Ih(a) I 2 i, Ih(b) / 2 i for some is 2.
Let further o(i) be the smallest number satisfying: for any such h, it generates a
cube-free o-word if and only if {h”(a) 1n 5 a(i)} is cube-free.
With these notation we have

Theorem 5. For different values of i the following holds true: a(2) 5 10, a(3) I 9,
o(4) I 8, o(5) I 8 and a(i) I 7 for i L 6.

Proof. The case i=2 was proved in Theorem 4. For other values of i the proof can
be carried out by applying the ideas from there. Without going into details we only
mention how the value of o(i) is formed as the sum of the number of different
stages:
i=3: 1+3+3=7,

i=4: 1+2+3=6,

i=5: 1+2+3=6,

ir6: 1+1+3=5.

In all the cases two extra steps are needed to guarantee the existence of a cube and
not only of an almost cube.

The above values of o(i) are not claimed to be the best ones. On the other hand,
they are not very large either.
296 J. Karhumtiki

7. Discussion

We have considered w-words defined by iterating a morphism in a binary case.

Especially, we were interested in when cube-free u-words would be obtained in this
way. We showed that this is possible only when a morphism is a biprefix. Using the
well known Fibonacci-morphism, i.e. the morphism defined by h(a) =ba and
h(b) = a, as a counterexample we also proved that the same does not hold true for
fourth power-free o-words: such ones can be obtained by iterating a nonbiprefix.
In particular, we were interested in whether it is decidable or not whether a given
binary morphism defines a cube-free w-word. After quite lengthy considerations we
showed the problem to be decidable. Moreover, we succeeded to establish a constant
limit for the number of iterations needed to be done. The limit we obtained is 10, in
general. For some special cases the limit can be improved. For instance, if
min{ 1h(a) 1, 1h(b) / } 16, then only 7 iterations are necessary. So it would be possible,
even in practice, to use a computer to find out morphisms generating cube-free w-
words.
We also want to point out that we have not really used the fact that our
morphisms satisfy the prefix condition. If we drop out this requirement we do not
have any more a nice mechanism to define o-words. What we have is a DOL
system. From this point of view our results are also of some interest. In this
framework Theorem 4 can be formulated as: a binary DOL system generates a cube
if and only if it does it in no more than 10 steps. Actually, to make the above
conclusion we should rule out some trivial cases. The reader may check that the
result holds for these, too.
We finish this paper by discussing possibilities of using our ideas to related
problems. As already mentioned Berstel showed in [2] that the problem whether a
given morphism over a three-letter alphabet generates a square-free w-word is
decidable. Our arguments, without no essential changes, could be applied to solve
the problem, too. Further the considerations would be shorter than we had here.
Our approach would also give a constant upper bound for the number of iterations
needed to be checked, while Berstel obtained a computable upper bound depending
on a morphism. An upper bound we would get is not larger than twenty.
As another matter we want to mention a related problem solved in [7] using the
ideas of this paper. To define the problem let us recall that an w-word is called
strongly cube-free if it does not contain as a subword a word of the form uofirst(o),
where first(u) denotes the first symbol of o. In [7] it is shown that it is decidable
whether a binary pp-morphism generates a strongly cube-free w-word. Moreover, it
is enough to check 7 iteration steps only.

Acknowledgements

The author is grateful to the Academy of Finland for the excellent working
 On cube-free w-words 297

conditions under which this research was done. The author also wants to thank the
referees for their useful comments.

References

(11 D.R. Bean, A. Ehrenfeucht and G. McNulty, Avoidable patterns in strings of symbols, Pacific J.
Math. 85 (1979) 261-294.
[2] J. Berstel, SW les mots sans carte definis par une morphisme, Springer Lecture Notes in Computer
Science 71 (1979) 16-25.
[3] J. Berstel, Mots sans carre et morphismes iteres, Discrete Math. 29 (1979) 235-244.
14) F.-J. Brandenburg, Uniformily growing k-free homomorphisms, Theoret. Comput. Sci. 23 (1983),
to appear.
[5] F. Dejean, Sur un theoreme de Thue, J. Combin. Theory 13 (1972) 90-99.
[6] A. Ehrenfeucht and G. Rozenberg, On the subword complexity of square-free DOL languages,
Theoret. Comput. Sci. 16 (1981) 25-32.
[7] J. Karhumaki, On strongly cube-free w-words generated by binary morphisms, Springer Lecture
Notes in Computer Science 117 (1981) 182-189.
[8] G. Rozenberg and A. Salomaa, The Mathematical Theory of L Systems (Academic Press, London,
1980).
[9] A. Salomaa, Jewels of Formal Language Theory (Computer Science Press, 1981).
[lo] A. Thue, ijber unendliche Zeichenreihen, Videnskapsselskapets Skrifter. I. Mat.-naturv. Klasse,
Kristiania (1906) l-22.
[l l] A. Thue, ijber die gegenseitige Lage gleicher Teile gewisser Zeichenreihen, Videnskapsselskapets
Skrifter, I. Mat.-naturv. Klasse, Kristiania (1912) l-67.