Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
PhilosophicalSociety
http://journals.cambridge.org/PSP
AdditionalservicesforMathematicalProceedingsofthe
CambridgePhilosophicalSociety:
Emailalerts:Clickhere
Subscriptions:Clickhere
Commercialreprints:Clickhere
Termsofuse:Clickhere
Orderedorbitsoftheshift,squareroots,andthedevil's
staircase
ShaunBullettandPierretteSentenac
MathematicalProceedingsoftheCambridgePhilosophicalSociety/Volume115/Issue03/May1994,pp
451481
DOI:10.1017/S0305004100072236,Publishedonline:24October2008
Linktothisarticle:http://journals.cambridge.org/abstract_S0305004100072236
Howtocitethisarticle:
ShaunBullettandPierretteSentenac(1994).Orderedorbitsoftheshift,squareroots,andthe
devil'sstaircase.MathematicalProceedingsoftheCambridgePhilosophicalSociety,115,pp
451481doi:10.1017/S0305004100072236
RequestPermissions:Clickhere
Downloadedfromhttp://journals.cambridge.org/PSP,IPaddress:132.248.41.83on22May2013
451
Ordered orbits of the shift, square roots, and the devil's staircase
B Y SHAUN BULLETT
School of Mathematical Sciences, Queen Mary and Westfield College, Mile End Road,
London Ei 4NS
AND PIERRETTE SENTENAC
Mathe'matique, Bdtiment 425, Universite de Paris-Sud, 91405 Orsay, France
(Received 25 January 1993; revised 3 June 1993)
Abstract
An orbit of the shift <r: t*-*2t on the circle T = R/Z is ordered if and only if it is
contained in a semi-circle C =\ji,/i + |]. We investigate the 'devil's staircase'
associating to each / t e T the rotation number v of the unique minimal closed o~invariant set contained in C ; we present algorithms for fi in terms of v, and we prove
(after Douady) that if v is irrational then ji is transcendental. We apply some of this
analysis to questions concerning the square root map, and mode-locking for families
of circle maps, we generalize our algorithms to orbits of the shift having 'sequences
of rotation numbers', and we conclude with a characterization of all orders of points
around T realizable by orbits of cr.
Resume
Une orbite du shift a: t^-2t sur le cercle T = [R/Z est ordonnee si et seulement si
elle est contenue dans un demi-cercle C = \ji,/i + \\. Pur chaque fisT, il existe un
unique sou-ensemble minimal ferme cr-invariant contenu dans Gp ayant un nombre
de rotation v; la correspondance (/*, v) est un ' escalier du diable'; nous donnons un
algorithme pour calculer /i en fonction de v et nous prouvons (apres Douady) que si
v est irrationnel, alors /i est transcendant. Nous donnons une application de cette
analyse a l'etude de fonctions 'racine carree et racine A-ieme' sur le cercle.
Nous generalisons l'algorithme a l'etude d'orbites du shift ayant une chaine de
nombres de rotation, et caracterisons les ensembles de points ordonnes sur T,
realisables par une orbite de o~.
Introduction
We consider orbits of the 'doubling map' o~: t^-2t (mod 1) on the circle U./Z (the
'squaring map' when we think of (R/Z as the unit circle in the complex plane).
A closed subset A of U/Z is described as ordered under a it A is invariant (that is,
o-(A) = A) and if a preserves the cyclic order of the points of A. Such a subset has a
rotation number, defined as the rotation number of any degree 1 order-preserving selfmap of [R/Z extending cr on A, or, equivalently, as the frequency with which the digit
' 1' occurs in the binary expansion of any chosen point teA (see Part 1: Section 1).
This paper is concerned with the complete classification of such subsets A, explicit
452
Fig. 1.
We prove that the attractors Ag for the restrictions of cr'1 to the intervals Ig are
the unique closed subsets of IR/Z which are invariant under the shift a and which
have C3rclic order preserved by a. Furthermore we show that if Ag has rational
rotation number p/q then Ag is one of three sets A^'9, AZlq, Aplq which depend only
on p/q, and that if Ae has irrational rotation number v then Ag = A", a Cantor set
which depends only on v. The three possibilities with rotation number 1/2 are (as
453
+-+-
Fig. 2.
indicated by the example above) A1!2 = A1/3, A]!2 = A2/3 and A1/2 = Ag for any
1/3 < 6 < 2/3. The first two are maximal and the third (their intersection) minimal.
We give explicit algorithms for the construction of the sets AVJ9 and A" for both
rational and irrational v, including algorithms based on continued fraction
expansions.
The graph which assigns to each Ig the rotation number (under a) of the set Ao is
a devil's staircase (Fig. 1), that is to say a continuous weakly monotonic map of the
interval to itself which is locally constant on a set of full measure without being
globally constant. The horizontal steps on this graph (which together occupy full
measure) correspond to rational rotation numbers p/q. At the left-hand end of the
step for p/q the attractor Ae for cr"1 is AZl9, at the right-hand end it is Afla, and in
between it is Avlq,
An identical graph appears in the combinatorial theory of the Mandelbrot set [6,
8, 9]. It arises in the following way. The Mandelbrot set is defined to be
M = {ceC: the orbit of 0 underfc: zi^z2 + c is bounded}.
See Fig. 2. Douady and Hubbard[8, 9] showed that the complement of M in the
Riemann sphere C is conformally equivalent to the disc C D (where D denotes the
closed unit disc) via:
(where care is taken to choose the appropriate 2 n th root in this expression). The
inverse image under <t> of the line {z: arg (2) = 8} is known as the external ray of
454
S H A U N B U L L E T T AND P I E R B E T T E S E N T E N A C
argument 6 e IR/Z in C M. The cardioidi>f0, the main component of the interior ofilf,
is also conformally equivalent to a disc [8, 9]. Mo consists of the values of c where /
has an attractive fixed point other than oo, and it is parametrized by the derivative
of fc at that fixed point. This parametrization defines the internal rays in Mo. Each
internal ray of rational argument has end point on dM0 which is also the end point
of two external rays. The devil's staircase (Fig. 1) appears as the assignment of
internal to external arguments along dM0. The underlying reason for this is as
follows. When ceM the nlled-in Julia set Kc of/c has complement parametrized by
(as shown in [8, 9]). This parametrization defines the dynamic rays in CKc. Note
that fc acts on the arguments of these rays as the binary shift. When fc has a
rationally indifferent fixed point, say of derivative e2niplQ, Kc is a 'flower' with q
'petals', separated by q dynamic rays. The map fc permutes these dynamic rays
cyclically, with rotation number p/q, and the analysis [6, 8, 9] of this action and that
of the DouadyHubbard map O yields Fig. 1 as the graph of the assignment of
internal to external arguments.
Yet another appearance of the same combinatorics is in the study of kneading
sequences of Lorenz maps of the interval [12, 15]. In these articles, Glendinning,
Sparrow and Hubbard consider pairs (a,/?) of binary numbers, 0 < a < /? < 1,
satisfying the condition
(*)
a ^ an(oc) <P, a < <rn{P) ^ /?, Vn ^ 0.
These they call allowed pairs. It is shown in [15] that (*) is both necessary and
sufficient for (a, /?) to be realizable as the kneading invariant of an expanding Lorenz
map. But the allowed pairs of form (a,a+1/2) are easily seen to be the pairs of
minima and maxima of our closed ordered sets A", with v irrational. It follows, for
example, that the points on the 'diagonal' (/? = a + 1 / 2 ) of figure 3 of [15],
correspond to precisely the points on the devil's staircase (our Fig. 1) where the
rotation number is irrational. (See also Remark 2 at the end of Part 1, Section 3,
where we discuss the relationship between 'order' and 'trivial renormalizability'.)
The ordered sets considered in this paper have been studied in one guise or another
since the earliest days of symbolic dynamics. Hedlund and Morse [17] investigated
Sturmian trajectories, namely sequences (infinite or bi-infinite) in two symbols a and
b with the property that the numbers of bs in any two segments of the same length
differ by at most one. Their basic examples of such sequences were given by the
following construction:
Let a be a positive real number, let (S = I/a, and let c be an arbitrary real number. On
the real axis oo < x < oo consider the set of points
Let T(c,a) (respectively T'(c,a)), denote the sequence ... abj<abj<+i... where j i denotes the
number of points of (*) in the interval i ^ x < i +1 (respectively i < x ^ i + 1). For a
rational and c = m integer let S(m, a) (respectively S'(m, a)) be as above except that the
single value j m is now to be the number of points of (*) in the interval m ^ x ^ m+ 1
(respectively m < x < m+ 1).
456
S H A U N B T J L L E T T AND P I E R R E T T E S E N T E N A C
are obtained by using as building blocks periodic sequences for different d^"19". The
method is closely related to that of Theorem 4 (and to renormalization, that is,
Douady tuning, in the terminology of the Mandelbrot set). Our final result, Theorem
8, is a complete characterization of all the cyclic orders of points that can be realized
by orbits of a.
I t is in the nature of this subject that many of the results have been discovered and
rediscovered at different times by different authors from different points of view. We
give references just before the statements of our theorems, wherever we know of
relevant earlier work. Where we give no references we believe our results to be new.
We apologize for any attributions we have omitted through ignorance.
Part 1. Orbits of the shift a
1. Introduction
Let A be a subset of the circle T = R/Z and / be a map A -> T (not necessarily
continuous). We say t h a t / i s order-preserving if for each triple (a, b,c) in A the triple
(f(a),f(b),f(c)) lies in the same cyclic order around T, or else is degenerate (i.e. two,
or all three, of the points coincide).
1. Every order-preserving map f:A^T,
to an order-preserving map F: T -> T.
LEMMA
Proof. Let n denote the projection R-s-IR/Z. Choose a 'base point' aeA and
representatives a,beR for a and b=f(a) (that is, n(a) = a and n(b) = b). Now
define F(d) = b, F(d+l) = b+l, and extend I to a (weakly) monotonic map
[a, d + 1] -> [b, b + 1]. There may be many ways to make such an extension: we give
an explicit one in detail as we shall use a similar construction much later, in the
proof of Theorem 8.
(i) if x = n(x)eA and/(x) =t= b then define F(x) to be the unique representative of
f(x) in ]b, 6 + l [ ; if xeA (x #= a) has/(a;) = b define F{x) to be b + l if there is a yeA
with x > y &ndf(y) #= b, else set F(x) = b;
(ii) if xeJA is a limit of points xteA with xi<x (in [a,d+l[) then set
F(x) = \imF(xt) (defined by (i));
(iii) if xeJ A is not a limit of the form in (ii), but is a limit of points xteA with
xt > x, then again set F(x) = lim-F(iy;
(iv) sinceF(x) has now been denned for all xeA, by (i), (ii) and (iii), we may extend
F to the whole of [a, a + 1] by linear extrapolation.
Finally, we extend F to a map F: M.-+M by periodicity, that is by sending each
[d + n,d + n+l] to [b + n,b + n+l] in the same way. It follows from the orderpreservation of/ on A that F is (weakly) monotonic increasing and hence induces an
order-preserving map F: T-s*T. I
Any order-preserving map of the circle F: T -> T has a well-defined rotation
number, the fractional part of
pn
lim
n-*co
for any lift/: U-+M satisfying F(y+l) =F(y)+l for all ye R, and for any xeU.
457
9/31
10/31
5/31
Rigid rotation
Fig. 3.
Proof. Any triple contained in a closed semicircle has its order preserved by <r,
since a doubles length. Conversely, if (a, b, c) are not contained in a closed semicircle,
then (a,6+1/2,c) are, and hence (a(a),a(b),cr(c)) = (cr(a), a(b+ 1/2), a(c)) has the
same cyclic order as (a, 6+ 1/2, c), namely the opposite to that of (a,b,c). I
It is easy to describe all the possibilities for the subsets A of a semicircle which
satisfy a(A) = A and contain 0:
A = {0}, A = {0} U {l/2 n } n > 0 ,
A = {0} U {1 -
l/2n}n>0.
In each case the rotation number vA is equal to 0. These are the invariant sets A0, A\
and At in the notation of Theorem 1 below.
Remark. A way to compute vA: Let A be contained in a semi-circle, and
cr(A) c A. For xoeA. let xn = an(x0), n ^ 0, and let xQeU be the unique lift of
x0 with 0 ^ x 0 < l . Inductively, let xneU be the unique lift of xn such that
0^xn-xn_1 < 1. Then
=
Moreover, if we represent a;0 as a binary number, then xn is obtained from xn_1 by:
(i) Shift the fractional part one binary place to the left;
(ii) Add the overflow (0 to 1) to the integer part.
It follows that the integer part of xn is the number of Is among the first n digits
458
S H A U N B U L L E T T AND P I E R R E T T E S E N T E N A C
of the binary expression of x0 and that vA is the frequency with which the digit ' 1'
occurs in this expression.
Example.
4
/_5_ io 20 i . 18\
4 t rJL J L j - I l
-" ~ 131 ' 31 > 31 ' 31 J 31/' **
L31 ' 31 ' 2-1"
* 0 = 00101,
^=f.
3
4
5
6f/2= 0.010101...
459
3
4
5
6>,+/2 = 0.1010...
Fig. 4.
THEOREM 3 (Algorithm for A"). (a) For each v 4= OeQ/Z, there exists a unique pair
of rationals 0 < 6~ < 0+ < 1 [constructed algorithmically), such that:
(i) 0~ and 0+ are adjacent points on the a-orbit A";
(ii) 0+/2 and 0J/2 +f lie on A", and A" c [0+/2,0;/2 +f]. / / v = p/q in reduced form,
the interval [0+/2,0;/2+i] Aas ZercgrtA (2 9 " 1 - l ) / ( 2 - l ) .
(6) .For eacA ve(IR Q)/Z <Aere erases a unique real number 0 < # < 1 swcA <Aa:
(i) 0,, lies on A" and its a-orbit is dense in A"; (ii) 0,,/2 and 6J2 + \ lie on A", and
and v0>v2e{U-Q)/Z,
460
We thank Adrien Douady for Theorem 4, the statement and proof of which he
outlined to us while this article was being prepared.
_ iF tT\
( T\
p iT\
\_ v
l~.\
00
/\
j( > y W *
with e}(x)e{0,1}, Pj(x) = ex(x) + ... +Sj(x) (if x has two binary expressions, choose the
one with an infinity of zeros).
We note that the pre-images of x under a are
x
x 1
-={0,el(x),...,ei{x),...)
a n d - + - = (i,et(x),
...,e}{x),...).
Consider the points {j'Pj(.x))j-i ' n "^2 a n d define the staircase of x by:
(i) ttPj+Ax) = PM) (ei+i(x) = ) > t h e Points {j,Pj(x)) and (j+i,pj+1(x)) are joined
by a horizontal line;
(ii) if pi+1{x) = p}(x) + l, the points are joined by a horizontal line followed by a
vertical line.
Conversely, every sequence of integers (jO;)3>1, such that 0 ^ Pj+\~Pj ^ 1 f r a u i>
provides a number x e T (Fig. 5).
2. Let xeT and let 0x denote the forward orbit of x.
(i) The orbit 0x is contained in the semi-circle Cx/2 = [x/2,x/2 + \] if and only if the
following condition is satisfied:
PROPOSITION
(ii) / / (C) is satisfied, then a\o has rotation number vx = lim^ODp;(a;/2)/j and the
staircase {pj(x/2)} satisfies
( | ) ^ ( | ) 1 .
Proof,
(C)
(i) S u p p o s e 0x cz Cx/2. L e t
w h e r e x = {elte2,
...,ep
...), so x/2 = ( 0 : e 1 : e 2 , . . . ) .
We claim that
The proof is a consequence of an obvious remark. For every ke{0,1,2,...} a word
of length k, 'et ...et ' cannot occur in the binary expression of x/2 (or x/2 +1) with
461
1 2
3
4
5
6
Fig. 5. The staircase of \ = (JT.
two zeros at its extremities on one occasion and two ones at its extremities on
another occasion, since then the points
could not lie in the same semicircle, as the difference between the middle two would
be ^ \. For instance, for k = 0, the rule is that the orbit cannot contain both '00' and
'11'.
We shall prove by induction on I that, for every JceN, 0 ^ sk t ^ 1. In fact our
inductive hypothesis is that 0 < sfcj j < 1 and
(P)
or
if sk , = 1, then
=1
{ or
but
k 1=
so
x
2'
462
S H A U N B U L L E T T AND P I E R R E T T E S E N T E N A C
So, using the fact that x/2 ^ crk(x/2) ^ x/2 + \, we are done.
Now suppose that (P) is true for 1 ^ j < I, so that 0 < sk<l_1 ^ 1 and
= e,
or
_
*,* = 1.
or
= 0,
el =
Now if e
= e,7+1 we are done, and if elc+l+1 4= e m we also get the required result,
using the remark above concerning sub-words surrounded by a pair of 0s or Is.
Conversely, suppose that Ox satisfies (C). We must show that Ox c [x/2,x/2+\].
Let ak(x/2) be the first point of the orbit lying outside this range, and consider
x
2
x 1
Let I be minimal such that efc+i =t= et. Then, since ak {x/2) lies outside [x/2, x/2 + | ] ,
if efc = 0, we have e( = 1 and ek+l = 0, and if efc = 1, we have et = 0 and ek+l = 1.
These give sk l+1 = 1 and 2 respectively, contradicting (C).
(ii) Immediate from the definition of rotation number. I
3. Let p/q be a (reduced) rational number, 0 <p/q < 1.
(i) There exist exactly two real numbers, 0 < d~/g < 6p/g < 1 (both rational) such that
dp/g/2 satisfy condition (C) and the orbits of 6^/g both have rotation number p/q.
(ii) The a-orbits of 6~/g and d*/g coincide, are q-periodic and are contained in the
PROPOSITION
interval
(iii)
[61J2,(6-lq+\)/2}.
?p/9 has an even numerator and 6p/g has an odd one.
Proof. Given 0 < p/q < 1, condition (C) of Proposition 2 defines two 'extreme'
staircases (as in Fig. 4). The first is where p}(x/2) always takes the larger value
whenever there is a choice, and the second is where it always takes the smaller value.
Denote these two staircases by (j,pf) and (j,pj) respectively. Explicitly we define
the staircase algorithm by
>+ = pj = integer part ofjp/q if j is not a multiple of q;
p+ = kp and pj = kp1 if kq =j.
The following elementary properties are self-evident from a consideration of the two
extreme staircases.
(I)
(a)
VJfc,
463
We define 0*/9 and #~/9 to be the binary numbers associated to the staircases
'Pf) a r ) d (j,Pj) respectively, that is to say (by (I) above):
ft+
=
VlQ
Vvia
F
F
1
\ 1 ? 2 ' " ' ' ' ff2-' '
IF
(iii) This follows at once from the explicit expressions for 6~/g and 0/9 above.
0J/B = ( o , l . o . o , i ) =
10 20 9 181 f- r 5 20T
,3i, 31, 51,31/ ^ L3T'3lJPROPOSITION 4. Let v = M. Q, 0 < v < 1. (i) TAere is a unique real number
0 < dv < 1 5MCA <M< 0,,/2 satisfies condition (C) cmd <Ae or6i/ o/ 5W has rotation number v.
(ii) The cr-orbit of 6V is contained in the semicircle [6J2,6J2 + \].
(iii) dv is recurrent under a.
Proof. Condition (C) this times defines a unique staircase, since 0 < jvpj < 1 => p}
is the integer part oijv. This staircase satisfies condition (C) of Proposition 2, so we
at once have (i) and (ii).
To show that 6V is recurrent, given keN set
mk =
with
0 < mk < 1.
464
15
ei,/5 = 01010
Fig. 6.
Proof. This follows directly from the staircase algorithm. Note that 8~,/b, (resp.
da/b) have periodic binary expression of period b' (resp. b) and that
Given a sequence (finite or infinite) of positive integers (alta2,...),
be associated sequence of rationals
1
Pi
3
1'
a2
define
to
1
s
1)
% +
a2 + ...I/a,
If ai+1 * 0,
? m = 13
> q<>
= 1,
= 0,
P i = 1.
Pm = P i-i+a-j+iP'P Po
It is well-known that two consecutive fractions in such a sequence are adjacent,
and that the sequence splits into two monotonic subsequences
465
0^01=^
so
a, digits
<%,/,,= M i - d i O
so 0 + ^ = ^ ^
a2copies
and
V./-
2_
__
1_
_
2+
1 + 1/1
^= 0
(2,2)-. ( ^ , = 01
16*2/5 = 01001.
Proof of Proposition 5.
(i) By Lemma 1, Q\la =d1 = (00... 01); then apply Lemma 2 and induction.
a,
(ii) F o r j = l
*;,/, - ^ , = d,-d2,
7
1
1
~2>-1'
d2 = a^d,
j _ 2^ ^ 1
2 9 -2
2
~ 2-l^_ 1 2^~ (2>
Hence d1 d2= l/((2 9 ' 1) (292 1)) and the general result follows by induction.
COROLLARY 1 (how to compute 6f for v = p/q). For rational v the continued fraction
has finite sequence (a1,a2, ...,ar = 1).
(a)
.
\&l = &t . and apply Proposition 5,
Pr/Qr
x
If ris even <
TC
466
(b)
S H A U N B U L L E T T AND P I E R R E T T E S E N T E N A C
If r
to = 2*-<w
COROLLARY 2 (how to compute 6,, for ve(U Q)/Z).IfveM
fraction has an infinite sequence of positive integers, then
(i)
^ < ^ < ^
+ 1
/ W
^>
fraction
Remark 1. Our construction (above) of ordered orbits of the shift translates into
algorithmic form the set of equivalent conditions proposed in [11] in the case of
rational rotation numbers. Moreover our algorithms enable us to construct the
maximum and minimum points of all closed ordered sets, whether the rotation
number be rational or irrational.
Remark 2. A sequence of Os and Is is said to be trivially renormalizable [12, 15]
if either
0 = w_ w1 w* w3...,
Proof of Theorem 1. (a) (i) If veQ/Z, then by Proposition 3 the cr-orbit of &* is
finite (closed), periodic and therefore minimal, and has rotation number v. Hence it
provides an A".
467
(ii) If ce((R <Q)/Z, then let A" be the closure of the cr-orbit of the recurrent point
8V of Proposition 4. It is clearly a closed, <r-invariant minimal subset with rotation
number v (notice that a(A") = A"; each point has one pre-image except $). The orbit
of dv is dense in A", and A" contains no interior point, so A" is a Cantor set.
It remains to show uniqueness, in both cases. Since A" has a rotation number, it
is contained in some semicircle, and since A" is closed we can speak of its least point
x/2. Then xeA" (since A is cr-invariant) and 0x [x/2,x/2 + \\. Uniqueness now
follows from Propositions 2, 3 and 4.
(6) For veQ/Z, the set A" is finite and there is nothing to prove. For pe(U Q)/Z,
a is injective on A" except at the points 6J2,6J2+\. Remove the o-~1-orbit of 8J2
and consider Av = A" {\Jn^Qa~n(6l,/2)}. Then a is a continuous bijection of Av to
itself. But er doubles measures. It follows that m(Av) = 0 and hence m(A") = 0 (since
Av is obtained by adjoining countably many new points to A"). The same argument
proves that the rf-dimensional measure md{Av) = 0 for all d > 0, and hence that Av has
Hausdorff dimension zero.
(c) For fie]8^/2,8^/2[, the only invariant subset of C^= [/t,/t + |] is A", but for
/i = d~/2 we can adjoin to A" the point d~/2 and all its inverse images lying in Cr
This gives a maximal set we denote A"_. Similarly for /* = 6^/2 we can adjoin to
A" the point 6+/2 + \ and all its inverse images in C^, to obtain A\. In the case of
irrational v, since 8J2 and 8J2 + \ both lie in A", this set is maximal as well as
minimal. I
Example. For v = \,Av_ consists of all points of the form 0-00100100 ... 0010001001
and images under o~ (i.e. we insert at most one extra 0 into an expression for a point
of A"), and A\ consists of all points of the form 0-001001... OOIOIOOT.
Proof of'Theorem 3. (a) is immediate from Proposition 3.
(b) is immediate from Proposition 4; note that uniqueness of 8V ensures that both
8J2 and 8J2 + \ lie in A", else the ' least' and ' greatest' elements of A" would give new
(c) is immediate from Proposition 3.
Proof of Theorem 2. Since every semicircle C^ = \ji,/i + \\, /i~\0,\[ contains some
A", we have
]0,|[= U /,u
{-),
where / = [8~/2,8+/2]. To prove that the / occupy full measure we observe that by
Proposition 3
"l'\Lvl
^-J
Ct
Sl
O9
1 '
where the sum on the right-hand side is taken over all p/q in reduced form. The
calculation of this sum is standard. We give an elementary proof. Consider the subset
^2 = {(p,q)\p < q,p,qeN} cz U2 and compute the sum S(j>,9)e# 1/2* m t w o different
ways.
First, sum horizontally and vertically (Fig. 7)
y
(p,Q)e3Zl
oo
cc
k-lj-k+l*
<
1
1
468
Fig. 7.
Next, sum along each line of rational slope and deduce
1
0<p/q<l
j>\
(p/gr'e'duced)
ilq reduced) "
)q
ojq
99 _ L1
0<p/g<l
"
(p/greduced)
(p/q reduce
This set is closed (since its complement is open), and has Lebesgue measure zero.
To show that it has Hausdorff dimension zero, we consider {a^Q}j>1.
For each j , o~i is injective on 0 except in a finite set (the end points of certain /),
and Vi,j, (rl(@) 0 <T}(0) is finite. Hence the d-dimensional measure of aj(Q) satisfies
ma(<r'(0)) =
and
...)md(Q).
But U (Tj(Q) cz T, so md(Q) = 0, ~id > 0, and hence 0 has zero Hausdorff dimension. I
Proof of Theorem 4. Let ye(IR Q)/Z. From Corollary 2 of Proposition 5, we can
write
where
d-
where
a-
"~2j+l
a+
469
(a) If {cij}j^i is not bounded, then given any c > 0 it follows from the inequality
above that there exists a sequence of rationals {Nk/Dk}k>1 such that
and
h m ^ = y = -^- ~ 0-618.
k}
6--
Hence
i.e.
" D<
for given c > 0 and e > 0, and j sufficiently large, and so 6V is not (2 + y e)Diophantine. I
Remark. For the golden mean
1
v=
1 + 1 + 1/(1 + ...)
we have
Pi
p2 _ 1
p3 _ 2
p4 _ 3
~~
0r l, ^
10, 0r = 101, 03+ = ToTTo.
This sequence can be generated recursively, by the substitution rules 1 -> 10, 0 - ^ 1 .
More generally, it is not to hard to prove, using Proposition 5, that d^.a . .a > comes
from 6t
> by the same substitution rules, and so does &},.. . <. from
a
(a1;...;a2n+1)-
(7
Finally we remark that for v the golden mean, the continued fraction expansion of
6 p =(l;2;2;2 2 ;2 3 ;2 5 ;2 8 ;2 1 3 ;...;2 a '.;...),
where {an} is the Fibonacci sequence.
Part 2. Applications and generalizations
1. The square root map: an archetypal example of mode-locking
For each 6 with 0 < 0 < 1, let Sq r6 denote the map of the complex plane C to itself
defined by sending each 2 to a value of Vz having 0/2 ^ Arg y/z =% 0/2 + . Note that
Sq rg is single-valued and continuous except along the half-line Argz = 0, where it is
two-valued.
470
S H A U N BTJLLETT AND P I E R R E T T E S E N T E N A C
For instance, with 6 = \ (so the rule is 'choose a new digit to give a value ^ \ and
^ f ) , and with initial point x0 = |, we obtain an orbit
xx = 0011,
x2 = 01011,
471
(2) If f < 6 < , then all orbits of Sqr 9 tend to the period 3 orbit
h = olio
( = )
This is the orbit A" for v = , but the action of Sqrfl has rotation number 1 = .
If 8 = f, we have an alternative choice for the image of tx, namely
but then ^ has subsequent orbit tending to {t0, tx, t2} = A213. In this case, the attractor
is A2!3.
The map T2 g is equal to Rgf2S2tgRg/2, where Rg/2 denotes the rigid rotation
th>t+ d/2(mod 1) on IR/Z, so T26 has the same dynamics as Sqg. Moreover, when A
tends to 1, the map Txg tends to Re/2 rather than the identity.
Each Tx e is a continuous monotone degree 1 circle map and therefore has a welldefined rotation number /ix g. The following theorem was first proved by Veerman
([19], [20]).
6. For each real A > 1, the map 6>->nx g has graph a devil's staircase. In
particular, for each such A, the rotation number /iK g is rational for a set of 6 of full
measure.
THEOREM
472
e/2
l/A
Fig. 8.
"
the orbit of 0 has period q and the rotation number is p/q. Denote by 6p/q the value
of 6 such that the left-hand inequality of (*) becomes equality and 6^lq the value of
6 where the right-hand inequality becomes equality. Note that by the monotonicity
of {Txe)q as a function of 6 for each q, the values d/g exist and are well-defined for
each p/qeQ/Z. For 6~/g < 6 < 0+/9, increasing 8 by 8 moves (^ >e ) 9 (0) a distance
It follows that
ffr
ft-
1\_2(A-1)2
A; A(A-I)"
We compute the sum of the lengths 8p/g 6p/g over all ^/g in reduced form, with
0 < p/q ^ 1, in the same way that we computed the corresponding sum in the case
A = 2 (Theorem 2). As there, let 0t= {{p,q)\p <q,p,qeN}. Then
(A-l)2
but
(A-l)2
A
(A-l) 2
V
V
ZJ
ZJ
s
0<p/g<l
(p/qreduced)
V(A-
A- 1
V
fc-i A*
I) 2
A - l
A-l
A"
(P/<?
reduced)
(A- I) 2
"I
473
v0 = 0.0001
= 0.1000
Fig. 10.
474
Hence
2(A-1)2 2
y^ = y
(p/qreduced)
and so
X,
o<p/sa
1)
(p/g reduced)
Thus we know that for each fixed A, the rotation number /ix e is rational except on
a set of values of 6 of measure 0. It remains only to show that all irrational values
are achieved on this set. For a given ve(M <Q)/Z, consider {8+/g: p/q < v) and
{0~/g: p/q > v). The least upper bound of the first set and the greatest lower bound
of the second coincide, since rational rotation numbers occupy full measure. Denote
the common limit by 0v. By monotonicity of /tA_fl as a function of 6, we have
We remark that as A -> 1, the map TA e tends to the rigid rotation Ee/2 in the Ctopology. For each A > 1 the theorem gives us mode-locking on a set of values of 6 of
full measure, unlike the C00 case [1] where mode-locking occurs on a set of decreasing
measure as one approaches rigid rotations. Note also that as q increases the value of
A for which 8p/g 6~/g has maximum length becomes closer and closer to 1 (Fig. 9).
2. Orbits with sequences of rotation numbers: 'Douady tuning'
The smallest orbit of a: t>-> 2(mod 1) on IR/Z which is not ordered is the period 4
orbit
to = o-ooii = ,
t, = o-oTIo = f,
t2 = 01100 = |,
t3 = o-iool = f.
However, under a2: <H*4(mod 1) this splits into two ordered orbits {to,t2} and
{tlt t3}. Moreover one of these, {tx, t3}, lies on a quarter circle.
Any orbit of the ft-shift <rn on IR/Z which remains within an interval of length 2~"
has its cyclic order preserved by an, since crn multiplies lengths by 2". But the
converse is false: in the example above {t0, t2} has its order preserved by cr2 but it does
not lie within a quarter circle. Thus the situation is not as straightforward as for the
simple shift a.
We may think of {to,t1,t2,t3} above as a 'satellite' of the period 2 orbit {0-01,0-10}
of cr. The following is an example of a period 12 orbit which is a satellite of a period
4 orbit
u0 = 0-000100010010 = ^ .
This has orbit under a illustrated in Fig. 10. The orbit {v0, vlt v2, v3} in the Figure has
period 4.
Note that the points {vo,vltv2,v3} partition the a orbit of ti0 into 4 sets:
Ao = {uo,u4,us}, Ax = {u^u^Ug}, A2 = {u2>u6,u10}, A3 = {u3,u7,un}. Each of these
sets is invariant under cr4 and cr maps each A{ bijectively to Ai+1 (where i +1 is taken
mod 4).
More generally, for v = p/q, the orbit of 6~ partitions T into q intervals
^o = [&;, Of], h = <rVo, -JQ-i
^{h)
475
Fig. 11.
on!
I.
This allows us to speak of the rotation number of an orbit of cr9 on Io and to make
the following inductive definition:
A closed s u b s e t 4 c R h a s a sequence of rotation numbers
f r "> w h e r e
Proof. Consider the following explicit algorithm for the construction of A, when vn
is rational, as the orbit of a point 6jv ^ obtained by successive modifications of 6~ ,
or, when vn is irrational, as the closure of the orbit of a point djv v , obtained from
vn in the same way. Let d~ _ denote the binary block of length qn_x which gives &l
when repeated, and let d* _ denote the corresponding block for 0+ . The first step
is to replace all occurrences of 0 in d~ by copies of d~ __ and all occurrences of 1 by
copies of d* _. We now repeat the process, replacing all 0s in the new word by d~ _
and all Is by d* _ . After nl steps, we obtain a binary expression for a point 6 whose
orbit is periodic of period q1q2---qn if vneQ/I., or non-periodic if vne(M Q)/Z. In
either case, A is the closure of this orbit.
PSP115
476
0010
ooolo
00001
0100
Fig. 13.
478
x4'
Fig. 14.
479
(where ' < ' is defined by cutting open the two circles at a and 0 respectively).
If x( 6 [a, /?[ and xt e [/?, a +1 [, then by comparing the first digit of r{xi) with that of
T(X}) we have that T(xt) < T(XJ). But if both xl and x} lie in [a, /?[ (or both in [fi, a+ 1[)
we know that the order of (a, xt, x}) (or (xu xp a +1)) is preserved by h, since h maps
each of [a,/?[ and [/?, a + l [ monotonically onto the whole of T. It follows that to
complete the proof it suffices to show that if xi 4= x} their associated binary sequences
T(xi),T(xj) differ at some digit. In the case that X is infinite, this is implied by the
hypothesis that T(X0) is not preperiodic. In the case that X is finite it follows without
further hypothesis by the following argument. If two points of X have the same
480
/
'7/
xo + 2
x5 + 1
x4 + 1
r
71
X2+l
x4
7
A
x0
i
x2
|
x
x4
Fig. 15.
itinerary, then so do all points of X between them, and there is therefore an adjacent
pair of points of X with this property. Such a pair is mapped to itself by hn (where
n is the number of points in X) and indeed the line segment joining the pair is mapped
to itself by hn. Since h is transitive on X this implies that h has degree 1, contradicting
the fact that we constructed it to have degree 2. I
Example. Is X pictured in Fig. 15 realizable ? The answer is yes. We construct h
and calculate T(X0) = 0011000 = 8/21.
Remarks. 1. We need the condition of non-preperiodicity of the itinerary, in the
case of infinite X, to avoid examples such as {xn = 2~n}n>0, which is not realizable as
an orbit of cr.
2. The realization of a finite X satisfying condition (S), as an orbit of a, can also
be achieved by 'renormalization'. For n = \X\ the points hn(x0),^"(^j),...,hn(x0+
1),
in (R, are spaced along an interval of length 2" in the same order as
xo,l,...,xo+i,
and they are mapped to one another in the same way by h. However, rescaling this
481
long interval by dividing by 2 replaces h by a map very close to a linear map (of
slope 2), and repeating the process yields (in the limit) a set of points in the same
cyclic order as X and mapped to one another by a.
We are deeply indebted to Adrien Douady for many helpful discussions. The first
author would also like to thank Franco Vivaldi and Ian Percival for comments and
suggestions on a very early version of some of this work.
REFERENCES
[1] V. I. ARNOL'D. Small denominators. I. Mapping the circle onto itself. Akad. Nauk. SSSR, Ser.
Math. 25 (1961), 21-86, and Geometrical methods in the theory of ordinary differential
equations (Springer-Verlag, 1983).
[2] P. ATELA. Bifurcations of dynamic rays in complex polynomials of degree two. Erg. Th. Dyn.
Syst. 12 (1992), 401-423.
[3] C. BANDT and K. KELLER. Symbolic dynamics for angle-doubling on the circle. I. The
topology of locally connected Julia sets. In Lecture Notes in Mathematics 1514 (Springer
Verlag, 1992), 1-23.
[4] C. BANDT and K. KELLER. Symbolic dynamics for angle-doubling on the circle. II. Symbolic
[19] J. J. P. VEERMAN. Symbolic dynamics and rotation numbers. Physica 134A (1986), 543-576.
[20] J. J. P. VEERMAN. Symbolic dynamics of order-preserving orbits. Physica 29D (1987),
191-201.