See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224103413

A walk along the branches of the extended Farey Tree

Article in Ibm Journal of Research and Development · June 1995

DOI: 10.1147/rd.393.0283 · Source: IEEE Xplore

CITATIONS READS

35 593

2 authors:

Jeffrey C. Lagarias Charles Tresser

University of Michigan Aperio, Inc.
371 PUBLICATIONS 19,149 CITATIONS 156 PUBLICATIONS 4,727 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Commutators of dilated floor functions View project

Nonlinear Circuits and Systems View project

All content following this page was uploaded by Charles Tresser on 05 March 2015.

The user has requested enhancement of the downloaded file.

 by J. C. Lagarias
A walk along C. P. Tresser

the branches
of the extended
tarey tree

The rational numbers can be presented as the of formula (1) can be generalized to allow the construction
set of vertices of a degree-three tree.If p/q of a free tree of degree three, which we call the extended
and p'/q'are two rational numbers written in Farey tree or Q-tree, having the set Q of rational numbers
lowest terms, the differencepq' - p'q depends as its set of vertices, as we show in Section 2.
only on the shape of the path joining p/q to The Q-tree embeds in the plane in such a way that its
p'/q'on this tree. set of ends, naturally labeled by real numbers (in general,
not rational), forms a nondecreasing sequence when read
1. Introduction from left to right. Using the embedded tree, we give a
Back in elementary school, many of us thought that life topological interpretation of the numerator pq' - p ' q
would be much simpler if adding fractions required simply of the difference of two fractions p / q and p ' l q ' , before
adding the numerators to get the numerator of the sum, reduction to lowest terms, and an equivalent interpretation
and adding the denominators to get the denominator of the of the same quantity in terms of symbols associated
sum. In fact, the pairing with the abstract tree. This extends the following

t?$) w.
P fP'
reinterpretation of n - m , the difference of two integers:
Consider the real line marked by the lattice Z of integer
points. When going from n to m along the line, initialize
is useful in number theory, in topology, and in dynamical a counter to zero; add 1 for each lattice pointreached
systems theory under the name Farey sum of p / q and if walking toward --m (to the left); subtract 1 if walking
p ' / q ' . More precisely, this is so when p / q and p'lq' are toward +-m(to the right). The final number on the counter
both in the unit interval, are in lowest terms, and are is n - m.
Farey neighbors, i.e., are not separated, on the real line, Keeping in mind that n and m can be rewritten as n / l
by any fraction with denominator smaller than max(q, 9 ' ) . and mll (as rational numbers written in lowest terms), we
[For example, we would associate 10113 with the pair (3/4, wish to rewrite n - m as n * 1 - m * 1 . The aimof this
7/9) because no other rational number with denominator paper is to use the Q-tree to interpret topologically, as we
smaller than 9 can be found in the interval (314, 7/9); did for n - m , the quantity p q ' - p ' q associated with the
however, we would not associate 2/7 with the pair pair ( p / q , p ' l q 'of
) rational numbers reduced to lowest
(U5, 1/2) because both 1/4 and 1/3 are found in the interval terms. In other words, we give a pictorial perspective to
(U5, 1/2).] The Farey sum operation allows one to present the solutions of the Diophantine equations pq' - p ' q = k ,
the set of all rational numbers in(0, 1) as the set of vertices where the case k = 1 is solved by consecutive terms of
of an infinite rooted tree known as the Farey tree. The use Farey series.
"Copyright 1995 by International Business Machines Corporation. Copying in printed form for private use is permitted without payment of royalty provided that (1) each
reproduction is done without alteration and (2) the Journal reference and IBM copyright notice are included on the first page. The title and abstract, but no other portions, of
this paper may be copied or distributed royalty free without further permission by computer-based and other information-service systems. Permission to republish any other
portion of this paper must be obtained from the Editor.
001&8646/95/$3.00 8 1995 IBM

IBM DEVELOP.
J. RES. VOL. 39 NO. 3
MAY 1995 J. C. LAGARIAS
C. AND P. TRESSER
 Since the formal language makes comprehension of our
elementary result artificially difficult, we have tried to
capture this result in Figure 1, which displays the
embedded Q-tree. Vertices are labeled with rational
numbers. Five paths are marked. The three red paths
[from (-5)/3 to 1 4 , from (-3)/4 to (- 1)/4, and from
(-10)/7 to (-6)/5] have the same shape and the same value
o f p q ' - p ' q : ( - 5 ) ' 1 - 1 3 = (-3) * 4 - (-1) * 4 =
(- 10) 5 - (- 6 ) 7 = -8. Similarly, the two blue paths
have the same value o f p q ' - p ' q : 2 1 - 2 * 3 =
7 . 3 - 5 . 5 = -4.
In the sequel, all rational numbers are written in lowest
terms (with nonnegative denominators), except when
otherwise stated, and with the proviso that both n and
d l are considered as being written in lowest terms.
We write [x] for the integer part of x, and {x} for its
fractional part, so {x} = x - [x]. For rational number
x = p l q , we define
N(x) =p , D(x) 3 q,
and if x' = p ' / q ' is rational, we define
(x, x') = pq' - p'q.
For any integer n , and for x E [ n , n + 11, we write
Illustration of the main results: The difference pq' - p ' q depends
only on the shape of the oriented path from p / q to p'lq' on the ({x) ifn c : x < n + 1,
Q-tree.

Then, for x and x' both in [n, n + 11, it is easy to verify

that
In Section 2, we construct the Q-tree and formulate (x,x') = ({XI,, {Xf>">. (2)
what we call the Q-tree theorem. Proofs and
complementary results are provided in Section 3; the As a consequence, all classical results concerning Farey
proofs given there also furnish algorithms to compute sequences extend readily to fractions in Q.
pq' - p ' q along the Q-tree paths emanating from the
vertex p / q . Sections 4 and 5 give less algorithmic but more Remark It is convenient to extend the set of rational
structural and natural proofs of the result in Section 2. numbers to contain the ideal number 1/0. Since this
More precisely, Section 4 is an elementary proof, which convention allows us to shorten many discussions, we
depends on a nice homogeneity property of the Q-tree. set 6 = Q U {UO}. The usefulness of adjoining 1/0 to Q
Section 5 relates the Q-tree theorem to the action of in the context of Farey theory was recognized, e.g., by
PSL(2, Z) on trees, which is related to a well-developed P. Bachmann [ l l l , E. Lucas [121, and A. Denjoy [131'.
theory (e.g., [ l , 21). In Section 5, the Q-tree is labeled
by the tiles of a tessellation of the hyperbolic plane, Extended Farey sequences The extendedFarey
introduced by H. J. S. Smith [3] and studied at length by sequence of order i is the ordered set X iof fractions x in
G.Humbert [4-61. For another discussion of the relations 6,written in lowest terms, whose denominators do not
between the Farey tree and PSL(2, h),see [7]. See also exceed i . Hence,
[81 for related material. ... = X-3 =-2X= X- I =a,
2. Definitions and the Q-tree theorem 1
We first describe an extension of the elementary concepts
= 6,
in the theory of Farey sequences from Q n [O, 11 to Q.
1 As a matter of fact, we could have chosen to adjoin (- l)/O, instead of 110. to 0
For the classical theory, see, e.g., [9], pp. 23-26, or for much of the discussion: A geometrical meaning of the ambiguity appears in
[lo], pp. 7-11. Then, the Q-tree theorem is stated in two Section 5. However, a reason to choose 110 as we did, and not (-])/a, that is
meaningful to our main purpose of exploring the 0-tree, appears in Section 4.
284 different forms, once the required language is in place. See also [131.

J. C. LAGARIAS AND C. P. TRESSER IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995
 1 Theorem I
X,=ZU-, Any noninteger rational number admits a unique Farey
0
sum decomposition. SpeciJcally, given p l q in ( n , n + l ) ,
there exists a unique pair of Farey neighbors
6*
( p o / q o , p l / qinl ) such that

and PI
p- = -Po
$ -.
4 so 41
-5 -3
-4 -1 -2 -1 -1
2,= . . . - - - - - - - Furthermore, the pair (po/qo, pl/q,) belongs to [n, n + 11,.
’ 3 ’ 2 ’ 3 ’ 1 ’ 3 ’ 2 ’ 3 ’
The numbers p o / q oand p l / q l are called the Farey
0 1 1 2 1 4 3 5 1 parents of p l q . In the sequel, each time we use the symbol
”””“... - p l q , we associate it with the ordered pair ( p o / q o , p , / q , )
1’3’2’3’1’3’2’3’ ’0’
of its Farey parents, so that
This extension of the definition of Farey sequence to all of
6 is preferred here to the more usual one (see Section 6), p Po PI
”
” -,
because our main object is the extended Farey tree, 414 40
constructed below. With our definition, X,and X, (but with p o / q o< p , / q , . To represent d l , we use the
not X,)appear at the first stages of the construction of convention
this tree.
n n-1 1
_-
=- @ -.
Fareyneighbors We say that two fractions p / q and 1 1 0
p’lq‘ are Farey neighbors if they are consecutive in some
extended Farey sequence, or equivalently, as is easy to
Young and old parents For any rational number, one of
verify, if
its two Farey parents, called the young Farey parent, has
a bigger denominator than the other, called the old Farey
Ipq’ - p’q1 =
It I):
-, 7 = 1. parent. For instance, the young parent of (n + 1)/1 is d l ,
and its old parent is 1/0. This terminology will make more
sense after we use it in the description of the Q-tree. To
Fareysum avoid inflating our list of definitions even more, we freely
It is clear that if p / q and p ’ / q ’ are Farey neighbors, there use genealogical relations that are coherent with the ones
exists some integer n such that both numbers belong to precisely defined so far.
[n,n +
11. Given Farey neighbors p / q and p ‘ / q ‘ in
[ n , n + 11, we define their Farey sum as @tree
The extended Farey tree or Q-tree, T, is defined as the
P P’ P +P’ free tree of degree three with vertices labeled by rational
- $ -=-
4 q’ 4 + q’. numbers in such a way that p l q and p ’ / q ’ are consecutive
(or bound an edge) on the tree if and only if one of them is
Equivalently, we have
the young Farey parent of the other.
Once the ordered set of integers has been used to
consecutively label all vertices along an infinite path of
the tree T, the remaining vertices of T can be labeled
where, for Farey neighbors p o / q oand p , l q , in [O, 11, we inductively. In the induction process, for any p / q , the
define young parent of p / q is used to label a vertex one or more
steps after the old parent has been used to label a vertex.,
Po PI + PI The restriction of 9 to the set Q rl(0, 1) is often called
- @ -=- Po the Farey tree (see Section 6).
40 O 41 qo + 41
Formula (3) is the key to an immediate generalization of Walking on the Q-tree We say that a path on the Q-tree
Farey theory from [O, 11 to 8. goes down an edge if it goes from the parent to the child;
The well-known uniqueness of the Farey sum otherwise it goes up. The path goes left along an edge if
decomposition for numbers in Q fl [O, 11 (the combination the label of the end point of this edge is smaller than the
of Theorems 28-30 in [9]; see also [14]) combined with (2)
and (3) gives Theorem 1. * The label 110 appears morally first, but at infinity.

IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995 J . C. LAGARIAS AND C . P. TRESSER
 We say that D or U is the principal part of an elementary
symbol, and 1 or r its index.
) Q2, the path from p / q to
For any pair ( p l q , p ’ / q ’ E
p’lq’ is completely described by its symbol,

2 E Ua, - * -
UamDb,* Dbn,
($.$)
where both ai and bj are in (1, r } , and m 2 0 and n 2 0.

Q-tree theorem (symbolic version)

For ( x , x’) E Q 2 , the quantity ( x , x f ) depends only on
the symbol 2.(x,x,).3

Fifteenexamples
According to the Q-tree theorem, we can unambiguously
assign the number (x, x’) to the symbol 2(r,x,). We call the
number (x, x ’ ) the value of the symbol 2(X,X,). The list of
values u ( 2 ) corresponding to the 15 shortest symbols is as
follows:
40) = 0,
u(UJ= u(DJ = 1,
u(U) = u(D) = -1,

u(UrU,)= u(D,D) = 1,
U( UIV) = u(D,D,) = - 1,
u(U,V,) = u(D,D,) = 2,
v(UrV) = u(D,D) = -2,

4UiD,) = 3,
u(U,D) = -3,
where (4b) and (4c) are simply consequences of the
definition of generalized Farey neighbors.

Embedding the Q-tree in Rz

To limit the amount of formalism, we use figures to

h indicate how to inductively construct a topological

embedding of the Q-tree in the real plane R2: Figure 2
shows the starting configuration (integer vertices) together
with the next generation4, while Figure 3 explains by
examples how to grow the tree. The upper path in the
figure is used to label the vertex (already labeled)
Initialization of the construction of the Q-tree: (a) Q is augmented ( a + e ) / ( b + f ) . The young Farey parent of that vertex is
8 by 110; (b) Q isaugmented by (- l)/O. the one labeled elf. The path is used to find the old Farey
parent. Because the path from elf to the vertex to be
labeled goes to the right, the path used to find the old
Farey parent starts to the right of vertex elf and moves
label at its starting point; otherwise, we say it goes right. upward to the right of edges until it is forced to turn down.
Thus, an edge can be labeled unequivocally by one of the The old Farey parent, a l b , is the label of the vertex at
following elementary symbols:
3 That is, (x, x‘) does not depend on the values x and only on the symbol Z ~ x , x ~ ~ .
Figure 2(b) shows how the diagram must be changed if Q is augmented by (-l)/O
286 instead of 110: see footnote I

1. C. LAGARIAS AND C. P. TRESSER IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995
 a
b

0
d+ h
X ..( .( X

__""."I~."...~."_~I"__.- . ...

1 Examples of howtogrowtheQ-tree.

which the downturn occurs. The Farey sum of alb and

elf is used as the new label. Similarly, the lower path in
Figure 3 is used to label the vertex (already labeled)
+ +
(c g)l(d h ) . Because the path from glh to the vertex
...
to be labeled goes to the left, the path used to find the old
parent starts to the left of vertex glh and moves upward to
I All five possible initializations for Algorithm 2 of Section 3.

the left of edges until it is forced to turn down. The old

Farey parent is cld in this case.
There are, of course, infinitely many geometrical
realizations of this topological embedding. Some examples,
based on hyperbolic geometry, are discussed in Section 5 ,
but in these cases the edges do not go left and right or up
and down as desired. As we have already mentioned, our
Q-tree theorem is illustrated in Figure 1, which suggests
a different geometrical realization, using Euclidean
geometry: All edges at generation n have length 1/2".
The contents of Figure 1 can be formulated as follows,
with the word "shape" referring to the way in which a
sequence of directed edges, say oriented line segments,
goes up and down, and left and right, with no distances
involved.

Q-tree theorem (graphical version)

For ( x , x ' ) E QZ, (x, x ' ) depends only on the shape of
the path going from x to x' on the embedded Q-tree.

3. Evaluation of the paths

We prove the Q-tree theorem by giving two equivalent
algorithms for computing the value associated with a
1 Implementation of Algorithm 2 of Section 3.

symbol.

IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995 I. C. LAGARIAS AND C. P. TRESSER
 In the first algorithm, knowing the value of a symbol C
of length n, and the value of its truncation of length n - 1,
we compute the value of all symbols of length n + 1
beginning with Z. In the second (equivalent) algorithm, we For example, if Sn and Sn+, have the same principal part
construct a directed, weighted tree with marked vertex x, and the same index, then
so that in order to determine the value of the symbol
C(x,x,),we sum the weights along the path from x to x'.
To avoid formal writing, Algorithm 2 is illustrated in
Figures 4 and 5. Checking that the steps of these
algorithms are correct amounts to elementaryalgebra, Thus, it is a straightforward process to calculate C n + , ,
and is essentially left to the reader, except for one case beginning with the values of 2, and C, from Equations (4).
(chosen at random)of the first algorithm. We now justify the formula F, the proofs for the other
two formulas being similar. From Equation ( 9 , we have
Algorithm 1
Given two symbols,

2"= s,s2 * S,-,S, Thus, it is only necessary to verify that

v(Z,,,) = 2v(Cn) - u(Xn-,). We assume that
and
both X, andterminate with U , ; all othercases
would be treated similarly. Then, choose any u,/v, so
that the symbols Cn and Cn+lcan be followed when starting
from u,/u,. Let

Si E {D,,
DI, U,, U ) , u , h , be the vertex reached by following the symbol
X,,-,, starting from u,Iv,;
we say that Xn+,is obtained from X, by either u,/v, be the vertex reached by following the symbol Xn,
starting from u,/v,;
following (Xn+,= FCn) if Sn and S,+,have the same u 3 h 3be the vertex reached by following the symbol
principal part and the same index; X n t l , starting from u,Iu,.
changing (Xnt1 = CC,) if Sn and Sntlhave the same
principal part and different indices; or Thus, we have
turning (Xntl = TCn) if S,, and S n + ,have different
principal parts.

We denote the truncation of Zn by Xn-, = S,S,* Sfl-,.

We can then calculate the column vector and

from the column vector On the other hand, from the 15 examples, we have

7J1 7J3

by multiplying the latter by the one of the following and

matrices that corresponds to the preceding list of
directions:

which allows us to solve for u 3 and v3 in terms of u , , v , ,

u,, and v 2 . It is then easy to verify that

200

I. C. LAGARIAS AND C. P. TRESSER IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995
 Algorithm 2
All of the possibilities used in the initialization for the
construction of a directed weighted tree with marked
vertex x are given in Figure 4, and the way to compute the
remaining weights inductively is indicated in Figure 5;
weights are integers associated with directed edges. To
compute (x, x’), we sum the weights along the path from x
to x‘. Like the first one, this second algorithm has a matrix
representation, the relevant transformations, as seenin
Figure 5, being

Figure 6 displays an example.

4. Copies of the Farey tree in the Q-tree

For any rational number, Theorem 1 asserts the existence
of a unique decomposition

p_ = Po
- @
PI
-
4 40 91
of p / q , as a Farey sum of its two parents p o / q o< p , / q , .
Let us denote polqo by (0, 1) and p l / q l by (1, 0). Then
p o / q o , p I / q l , pand
/ q , all of the descendants of p l q can
be written uniquely as ( m p , + n p , ) / ( m q , + n q , ) and
represented by the pair of integers (n,m).Figure 7 shows
how the pairs can be organized on a tree. If the symbol
( n , m ) is replaced by d m , one obtains the piece of Y
generated from the pair 0/1 < 1/0, which thus reappears
everywhere in 5,justifying the choice we made of
1/0 to extend Q (see footnote 1). On the basis of this
observation, the proof of the Q-tree theorem is now
straightforward: Just notice that if p / q is the youngest
common ancestor of x and x ’ , setting

x=-
aP, + PI
aq, + bql’
CPO + dP,
x’ = ~

cqo + dq, ’
we have

(x, x‘) = (ap, + bp,)(cqo+ dq,) - (aq, + bq,)(cpo+ dp,)

= (bc - ad)(p,q0 - P04J
= bc - ad
Illustration of the construction in Section 4.
independently of x, x ) , but depending on the path that
joins them on 9.

J. C. LAGARIASAND C. P. TRESSER
IBM J. RES. DEVELOP.VOL. 39 NO. 3 MAY 1995
 -2 -1 0 1 2

Tesselations M and T.

5. PSL(2, Z) and the Q-tree Figure 8 represents two tilings of H together:

The tiling of H . J . S. Smith The tiling M, corresponding to the coloring, is a tiling

As usual, let SL(2, W) stand for the set of 2 X 2 matrices by the modular group PSL(2, H) = SL(2, Z ) l k l , with
of real numbers with determinant equal to 1. It is well fundamental region
known that the group PSL(2, W) = SL(2, W ) l k I , which
acts on C by the Mobius transforms as

[::I- az
GT-2
+b

is in fact the set of the orientation-preserving isometries of

The tiling T,corresponding to the solid lines, we call the
Q-tiling. Each of its tiles is made of three tiles of M.
[Figure 9 is an embedding of T,invariant by PSL(2, Z).
The dotted lines are the same as in Figure 8, but more
Poincark’s model of the hyperbolic plane, i.e., the upper easily seen.]
halfof the complex plane H = { x + iy E C: y > 0 } ,
equipped with the metric The Q-tiling, obtained by joining all pairs of Farey
a k 2 + dy2 neighbors by a geodesic in H, is not a tiling by a subgroup
ds2 = of PSL(2, R); in fact, it corresponds to an order-two
Y2 *
extension (by z b -Z) of the congruence subgroupr(2)of
For this metric, geodesics are pieces of circles orthogonal PSL(2, Z) [see, e.g., [15], p. 82, for a fundamental domain of
to the real axis, or pieces of vertical lines. For a discrete U2)1*
subgroup r of PSL(2, W), a fundamental region R is
defined by the following properties: Remark The relevance of T in the study of continued
fractions was recognized by H. J. S. Smith [3] in 1877 and
Its interior does not contain any pair (x, g(x)), g E r. studied in detail by A. Hurwitz [16] and G. Humbert
Its closure contains at least a point of each orbit. [4-61.

The set of images of any fundamental region R by the Marked tiles Each tile of T is an ideal triangle, i.e.,
elements of r provides a tessellation of H, called the tiling a triangle with all vertices at infinity (the real axis is at
290 by I’, with fundamental region R . infinity for H). The middle vertex of the ideal triangle can

J. C. CAGARIAS AND C. P. TRESSER IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995
be chosen to label the tile, but there is an ambiguity for
the upper tiles with vertical sides, regarding how to label
the upper ideal vertex. This is the geometrical ambiguity in
the choice among 110 and (- 1)/0 we mentioned in footnote
1. We choose 110 as m. With each tile, let us associate the c*:-4..
,*eu *-4.4..,*" *.*..A.,*.- *.*:-?*.* e-*.A...*"a
collection of its vertices, ordered so that one goes around -2 -1 0 1 2 3
the tile counterclockwise. A threefold ambiguity then
remains, corresponding to the original vertex of the triplet.
Hence, for any p l q , we have three marked tiles,

1 Embedding of T, invariantby PSL(2, 25). The arc PQ is a funda-

1 mentaldomainfortheaction of PSL(2,Z) onthistree.

with r , I , m standing respectively for right, left, and

middle. -2 -1 0 1 2 3
In the following computations from [6],we set
a.O+ b b
"

c-O+d-d'
and NOT 1 Anotherembedding of 3 in €I.

a-O+b a€ +b
"
- lim -
c - 0+d CE + d '

contains a vertex of Tg(see the black dots in Figures 8

which matters (only) when d = 0. A simple computation
and 9). Hence, from the previous discussion about the tiles
yields
of T, the vertices of Tgare in one-to-one correspondence
with the rational numbers, and we can easily verify that
the labeling so defined is such that Tgis a geometrical
realization of T.

Remark The boundary of U also contains an embedding

of T, shown in Figure 10.

PSL(2, Z) actionand the Q-tree Theorem 2 yields an

action of PSL(2, Z) on T X 2/32, defined by matrix
from which it is then easy to prove the following result multiplication, once the three matrices Miiq, Miiq, MiSlq
of G . Humbert (cf. [ 6 ] ,pp. 105-110). are used to label the vertex p / q of 9,
In general,
M:;::(Mi) = M:,':, with no simple general rule for s"'
Theorem 2 and x"', because the order relation of numbers with
For any s, S I in r, I , m, and any x, x' in Q, there is a their young Farey parents can be found in either way.
singlematrix M E PSL(2, Z) suchthat M T: = 4 : . However, it is easy to verify the following result.
More precisely, M = Mi::: = M:: (MI)". -
Lemma I
PSL(2, Z) and degree-three trees The pieces of the Consider any s in r, I, m, and any x, x' in 6.
Assume that
boundary of M that do not belong to the boundary of T x is the young Farey parent of y and z, with y <: x < z,
(Le., the dotted lines in Figure 8 and Figure 9) form a and that x' is the young Farey parent of y' and z', with
(geometrical) degree-three tree Tgon which PSL(2, 2) y ' < x' < z' . Then
acts with fundamental domain P Q (see [ I ] , p. 35 and
M::x,(M:) = MZ ,
[2], pp. 21-24). Notice that each vertex of Tgbelongs to
the interior of a single tile of T,and that each tile of T M::x,(M:) = M;: ,

".
IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995 J. C. LAGARIAS PND C . P. TRESSER
 and noticed in [19] that if all “vulgar fractions” with
denominator smaller than n are written in order, with
M : ; ~ , ( M ~=) sf. alb < a‘lb’ < al’lb”, then a’ = a +
a” and b’ = b + b”
Now let 3; stand for the rooted sub-tree of 3 labeled by (Theorem 29 in [9]). A wrong claim of proof was made by
x and all its descendants, each vertex p l q being labeled by an anonymous person with the signature “S.A.” in the
MPSiq.Iterating Lemma 1, we obtain the following equality. following volume of the same journal [20]. Farey also
presented his remark anonymously in [21]. His remark
was shortly afterward proved by A. Cauchy [22], who
mentioned only the French version [21] but was possibly
familiar with [19], since he associated the name of Farey
with the subject, followed in that regard by most
PSL(2, Z)and the Q-tree theorem mathematicians since. Cauchy in fact proved that with
The Q-tree theorem can be deduced from Theorem 3, the previous notations, a ’ b - ab’ = 1, from which he
specialized to the case s = r . More precisely, let deduced the property noticed by Farey.’ Cauchy did
not limit himself to [0, 11 but considered, without giving
them a name, what we call extended Farey sequences; it
occurred to us that the name “Cauchy sequences” would
Then, for u , v among the descendants of x, and u ‘ , u’ the not necessarily be well accepted in this context. In 1879,
corresponding points among the descendants of x ) , we J. W. L. Glaisher [23] (see, in particular, the historical
have comment, pp. 329-336) examined the history of the subject
and concluded that at least part of the credit should go to
au +b H. Goodwyn. He seems to have ignored the following fact:
“
- u’
cu + d Most of the Farey-Cauchy theory was presented in 1802
by le Citoyen Haros6 in [24] (see in particular the bottom
and
of p. 367 and the top of p. 368 of [24]). We refer the reader
an +b to L. E. Dickson’s book ([25], pp. 155-158, 162) for the
”
- u’. early history of Farey sequences up to 1919. Important
cu +d
applications were made by A. Hurwitz to Diophantine
Now, for any rational numbers p l q and p ’ l q ’ , we have approximation [26] and the reduction of binary quadratic
forms [16], and a relation to the Riemann hypothesis
was found by J. Franel [27] (see also [28, 291). The
Hardy-Littlewood circle method (see, e.g., [30]) makes
essential use of a Farey series dissection of [0, 11; it led, in
particular, to the celebrated result of I. M. Vinogradov [31]
that every sufficiently large odd number is expressible as
= (ap + bq)(cp’ + dq’) - (cp + dq)(ap’ + dq’) the sum of three prime numbers. For more concrete
= (ad - bc)(pq’ - p ’ q ) applications, see, e.g., [32].

= (Pq’ - p ’ q ) , Fareytree
so that The first occurrences we could find of this tree are in
[33-351, where the Farey tree is presented as a way to
(u’, u’) = (u, u). analyze applications of Diophantine approximation to
dynamics, and in [36], where the tree is implicit but
6. Historical notes parenthood is mentioned explicitly to describe a dynamical
result. For recent constructions, uses of the Farey tree in
Fareysequences dynamics, and lists of applications of Farey theory to
The Farey sequence of order i is usually defined as the dynamical systems theory, see for instance [7, 8, 14, 371.
ordered set 8, of fractions in x in Q n [O, 11 whose The Farey tree is strongly reminiscent of what were
denominators D ( x ) do not exceed i. This is usually called Brocot sequences in treatises on number theory in
extended to as the ordered set 8:of fractions x in 6,
6
with max(l(N(x)/, D ( x ) ) 5 i (see, e.g., [ l l , 17, 181). Theorem 28 in 191. See p. 24 of 191 for the proof of the equivalence of the two
In 1816, J. Farey (a geologist), studying the privately properties, and pp. 24-26 for proofs of both results.
6 Citoyen is the French word for citizen: The paper was published in Messidor,
circulated early version of the table of “complete decimal year X of the Republican calendar; Dickson refers to “C. Haros,” as well as most
authors after him, but we could not find any evidence that C stands for the initial of
292 quotients” by Henry Goodwyn, Esq., of Blackheath, Haros’s first name.

I . C. LAGARIAS AND C. P. TRESSER IBM I. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995
France around the turn of the century, after the clock References and notes
maker who described them in 1862 [38] (see also [12] I . J. P. Serre, Trees, Springer-Verlag, Berlin, 1980.
and [39]). The Brocot sequence of order i is the ordered 2. W. Dicks and M. J. Dunwoody, Groups Acting on
Graphs, Cambridge University Press, Cambridge,
set aiof fractions x in Q f l [0, 11 (or in bt)such that England, 1989.
= { O h , 1/1} (or a0= { O h , 14, 1/0}), and Bj+,is 3. H. J. S . Smith, “MCmoire sur les Equations Modulaires,
obtained from aiby including the Farey sum of all pairs Ac. de Lincei 1877,” Collected papers of H . J . S. Smith,
Chelsea, New York, 1965, Vol. 2, pp. 224-241.
of successive elements in !?hi. Thus, in bt , 4. G. Humbert, “Sur les Fractions Continues et les Formes
Quadratiques Binaires IndCfinies,” C . R . Acad. Sci. Paris
a =[! 1 ! 2 !) 162, 23-26 (1916).
5 . G. Humbert, “Sur les RCduites d’Herrnites,” C . R . A c a d .
’ 1’2’1’ 1’0 Sci. Paris 162, 67-73 (1916).
6. G . Humbert, “Sur les Fractions Continues Ordinaires et
and les Formes Quadratiques Binaires IndCfinies,” J . Math.
Pures Apll. 2, 104-154 (1916).
7. J. C. Lagarias, “Number Theory and Dynamical
Systems,” Proceedings of Symposia in Applied
Mathematics 46, 35-72 (1992).
8. R. Siegel, C. Tresser, and G. Zettler, “A Decoding
In [13], Denjoy generalizes this definition, and in Problem in Dynamics and in Number Theory,” Chaos 2,
473-493 (1992).
particular extends it to 6’. Brocot sequences and their 9. G. H. Hardy and E. M. Wright, An Introduction to the
generalizations are a primary object of study in the theory Theory of Numbers, Fifth edition, Clarendon, Oxford,
of the [?] function of Minkowski [40] (see also [7, 13, 41]), England, 1979.
10. H. Rademacher, Lectures in Elementary Number Theory,
which is defined as follows: If the path on the Farey tree Blaisdell, New York, 1964.
f r o m 1/2 to x E [0, 11 is labeled a , , a 2 , a 3 , - , then 1 1 . P. Bachmann, Niedere Zahlentheorie, Vol. 1 , 1902,
[?I( x) = O . a , a , a , in base 2. We alsonoticethatthe pp. 121-150; Vol. 2, 1910, pp. 55-96; reprinted in one
volume: Chelsea, New York, 1968.
ordered set of denominators of the fractions appearing 12. E. Lucas, The‘orie des Nombres, Gauthiers-Villars, Pans,
at depth n of the Farey tree form the Stein diatomic 1891, Vol. 1 , pp. 467-475, 508-510.
sequences [42] (see also [43-451). 13. A. Denjoy, “Sur une Fonction RCelle de Minkowski,”
J . Math. Pures Apll. 17, 105-151 (1938).
14. L. Goldberg and C. Tresser, “Rotation Orbits and the
Farey tiling Farey Tree,” Research Report RC-19376 (84294),
The tiling T is often called the Farey tiling in the recent IBM Thomas J. Watson Research Center, Yorktown
Heights, NY, 1994. To appear in Ergod. Syst. & Dynam.
literature, e.g., [17, 181, although G. Humbert (cited in [17] Syst.
and [18]) did his best to associate U with the name Smith. 15. L. R. Ford, Automorphic Functions, Second edition,
It is well known that some of the relations established Chelsea, New York, 1951.
16. A. Hurwitz, “Ueber die Reduction der binaren
between the Q-tiling and the theory of continued fractions quadratichen Formen,” Math. Ann. 45, 85-117 (1894).
can be reformulated in terms of the Q-tree (see, e.g., 17. R. Moeckel, “Geodesics on Modular Surfaces and
[14, 33, 341, as well as [46-481 for related results). Continued Fractions,” Ergod. Theor. & Dynam. Syst. 2,
There are, of course, many aspects of Farey theory and 69-83 (1982).
18. C. Series, “The Modular Surface and Continued
its applications not discussed here (see, e.g., [49-561). Fractions,” J . London Math. SOC.31, 69-80 (1985).
19. J. Farey, “On a Curious Property of Vulgar Fractions,”
Conclusion Philos. Mag. & Journal, London 47, 385-386 (1816).
20. Anonymous author, “On Vulgar Fractions,” Philos. Mag.
As J. Farey said at the end of his first paper on the & Journal, London 48, 204 (1816).
subject: “ I am not acquainted, whether this curious 21. J. Farey, “PropriCtC Curieuse des Fractions Ordinaires,”
property of vulgar fractions has been before pointed out?” Bull. Sc. SOC.Philomatique 3, No. 3, 112 (1816). (Page
numbers 105-1 12 occur twice in the volume: This is the
first page 112.)
Acknowledgments 22. A. L. Cauchy, “Demonstration d’un Theoreme Curieux
Recent discussions of one of the authors (C. T.) with Lisa sur les Nombres, Bull. SOC.Philomatique 1816,” Oeuvres,
Gauthiers-Villars, Paris, 1882-1958, Ser. 2, Vol. 6, pp.
Goldberg, Matt Grayson, and George Zettler about the 146-148.
Farey tree played a crucial role in discovering (perhaps 23. J. W. L. Glaisher, “On a Property of Vulgar Fractions,”
rediscovering) the elementary result reported in this paper. Lond. Ed. Dub. Phil. Mag., Ser. 5 , Vol. 7, 1879, pp.
321-336.
Section 4 is a modification of a simplified proof suggested 24. le C.“ Haros, “TABLES pour Evaluer une Fraction
by Don Coppersmith, who is gratefully acknowledged. Ordinaire avec Autan de DCcimales qu’on Voudra; et pour
Thanks also to Linda Keen for a helpful discussion. Trouver la Fraction Ordinaire la Plus Simples, et qui
Approche Sensiblement d’une Fraction DBcimale,”
“Annonces et Notices d’ouvrages,” J . de I’Ecole
Polytechnique Cah. 11, t. 4, 364-368 (1802).
’ We could have started with such a definition right away, but extended Farey
sequences, which do not involve Farey sums, are more elementary than the 25. L. E. Dickson, History of the Theory of Numbers, Vol. i,
sequences defined by the Brocot construction. Chelsea, New York, 1952 (or Carnegie Institution, 1919). $93

IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995 J. C. LAGARIAS AND C. P. TRESSER
 26. A. Hurwitz, “Ueber die angenaherte Darstellung der 53. S. Kanemitsu, “On Some Sums Involving Farey
Zahlen durch rationale Briiche,” Math. Ann. 44, 417-436 Fractions,” Mar. J . Okayama Univ. 20, 101-113(1978).
(1 894). 54. S. Kanemitsu, R. Sita Rama Chandra Rao, and A. Siva
27. J. Franel, “Les Suites de Farey et le Probltme des Rama Sarma, “Some Sums Involving Farey Fractions,”
Nombres Premiers,” Giittinger Nachrichten, pp. 198-201 J . Mat. SOC.Jpn. 34, 125-142 (1982).
(1924). 55. J. Lehner and M. Newman, “Sums Involving Farey
28. E. Landau, “Bemerkung zu der vorstehenden Arbeit von Fractions,” Acta Arith. 15, 181-187 (1968).
Herrn Franel,” Giittinger Nachrichten, pp. 202-206 (1924). 56. S. H. Kim and S. Ostlund, “Simultaneous Rational
29. M. N. Huxley, The Distribution of Prime Numbers. Large Approximations in the Study of Dynamical Systems,”
Sieves and Zero Density Theorems, Clarendon, Oxford, Phys. Rev. A 34, 3426-3434 (1986).
England, 1972.
30. R. C. Vaughan, The Hardy-Littlewood Method, University
Press, Cambridge, England, 1981. Received June 4 , 1993; accepted for publication March
31. I. M. Vinogradov, “Representation of an Odd Number as 15. 1994
a Sum of Three Primes,” Dokl. Akad. nauk. USSR 15,
169-172 (1937).
32. M. D. McIlroy, “Number Theory in Computer Graphics,”
Proceedings of Symposia in Applied Mathematics 46,
105-121 (1992). Jeffrey c. Lagarias AT&T Bell Laboratories, Murray Hill,
33. P. CvitanoviC, “Farey Organization of the Fractal Hall New Jersey 07974 (jcl@research.att.com). Dr. Lagarias
Effect,” Phys. Scripta T9, 202 (1985). received his Ph.D. in mathematics at the Massachusetts
34. P. CvitanoviC, B. Shraiman, and B. Soderberg, “Scaling Institute of Technology in 1974, with a thesis in algebraic
Laws for Mode Locking in Circle Maps,” Phys. Scripta number theory. Since then he has worked at AT&T Bell
32, 263-270 (1985). Laboratories, where he is currently a Distinguished Member
35. S. Ostlund and S. H. Kim, “Renormalization of of the Technical Staff in the Mathematical Research Center.
Quasiperiodic Mappings,” Phys. Scripta T9, 193-198 He held visiting positions at the University of Maryland
(1985). (mathematics) in 1978-1979 and at Rutgers University
36. J. M. Gambaudo, P. Glendinning, and C. Tresser, (computer science) in 1982. Dr. Lagarias has worked in a
“Collage de Cycles et Suites de Farey,” C. R. Acad. Sci. variety of areas of pure and applied mathematics, theoretical
Paris t. 299, Serie I, 711-714 (1984). computer science, and operations research. Most recently his
37. K. M. Brucks and C. Tresser, “A Farey Tree research interests have centered on problems in number theory
Organization of Locking Regions for Simple Circle Maps,” and dynamical systems, Fourier analysis and self-similarity,
Research Report RC-19136 (83340),IBM Thomas J. and mathematical programming.
Watson Research Center, Yorktown Heights, NY, 1993.
To appear in Proc. Amer. Math. SOC.
38. A. Brocot, Calcul des Rouages par Approximations,
Nouvelles Me‘thodes, par Brocot, horloger (Paris), 1862. Charles P. Tresser IBM Thomas J . Watson Research
39. E. Cahen, Ele‘ments de la The‘orie des Nombres, Center, P.O. Box 218, Yorktown Heights, New York 10598
Gauthiers-Villars, Paris, 1900, pp. 331-335. (TRESSER at YORKTOWN, tresser@watson.ibm.com). Dr.
40. H. Minkowski, “Zur Geometrie der Zahlen,” Proceedings Tresser received his “Thtse de troisitme cycle” in 1976 and
of the International Congress, Heidelberg, 1904. Also in his “Thtse d’Etat” in 1981, both from the University of Nice,
Gesammelte Abhandlungen, Vol. 2, pp. 33-52, reprinted in France. He worked at the Centre National de la Recherche
one volume: Chelsea, New York, 1967. Scientifique in Nice from 1978 to 1989 and held visiting
41. B. B. Mandelbrot, “The Minkowski Measure and positions at Columbia University (Department of Astronomy),
Multifractal Anomalies in Invariant Measures of Parabolic Courant Institute, Stevens Institute of Technology
Dynamical Systems,” Chaos in Australia, G. Brown and (Department of Physics), the Institute for Mathematics
A. Opie, Eds., World Publishing, Singapore, 1993. and its Applications in Minneapolis, 1’Institut des Hautes
42. M. A. Stern, “Ueber eine Zalentheoretische Funktion,” Etudes Scientifiques, the Hebrew University (Departments
J . reine angew. Math. 55, 193-220 (1858). of Mathematics and Physics), the Weizmann Institute
43. C. Giuli and R. Giuli, “A Primer on Stern’s Diatomic (Departments of Mathematics and Chemical Physics), City
University of New York (Department of Mathematics), and
Sequence, I and 11,” Fibonacci Quart. 17, 103-108,
246-248 (1979). the University of Arizona (Department of Mathematics). Since
44. D. H. Lehmer, “On Stern’s Diatomic Series,” Amer. 1989, he has been a research staff member in the Mathematical
Math. Monthly 36, 59-67 (1929). Sciences Department of the IBM Thomas J. Watson Research
45. D. A. Lind, “An Extension of Stem’s Diatomic Series,” Center. He is a former member of the editorial board of
Duke Math. J . 36, 55-60 (1969). Nonlinearity and is currently serving on the editorial board of
46. H. Appelgate and H. Onishi, “Continued Fractions and Chaos. His first thesis was on molecular dynamics simulations
the Conjugacy Problem in SL,(Z),” Commun. Algebra 9, of diatomic fluids, and his second thesis was on simple models
1121-1130 (1981). for the transition to turbulence. While still involved in several
47. G. N. Raney, “On Continued Fractions and Finite research projects related to various natural sciences, Dr.
Tresser has turned his focus to pure mathematics and spends
Automata,” Math. Ann. 206, 206-283 (1973).
most of his research activity on dynamical systems theory.
48. C. Series, “The Geometry of Markoff Numbers,” Math.
Intell. 7, 20-29 (1985).
49. T. Allen, “On the Arithmetic of Phase Locking: Coupled
Neurons as a Lattice on R’,” Phys. D 6, 305-320 (1983).
50. R. R. Hall, “A Note on Farey Series,” J . Lond. Math.
SOC.2, 139-148 (1970).
51. R. R. Hall and G . Tenenbaum, “On Consecutive Farey
Arcs,” Acta Arith. 44, 397-405 (1984).
52. M. N. Huxley, “The Distribution of Farey Points. I,”
Acta Arith. 18, 281-287 (1971).

J. C. LAGARIASAND C. P. TRESSER IBM J. RES. DEVELOP. VOL. 39 NO. 3 MAY 1995

View publication stats