The Distance Geometry of Music

Erik D. Demaine∗ Francisco Gomez-Martin† Henk Meijer‡ David Rappaport‡

Perouz Taslakian§ Godfried T. Toussaint§¶ Terry Winogradk David R. Wood∗∗
arXiv:0705.4085v1 [[Link]] 28 May 2007

Abstract
We demonstrate relationships between the classic Euclidean algorithm and many other fields
of study, particularly in the context of music and distance geometry. Specifically, we show how
the structure of the Euclidean algorithm defines a family of rhythms which encompass over forty
timelines (ostinatos) from traditional world music. We prove that these Euclidean rhythms have
the mathematical property that their onset patterns are distributed as evenly as possible: they
maximize the sum of the Euclidean distances between all pairs of onsets, viewing onsets as
points on a circle. Indeed, Euclidean rhythms are the unique rhythms that maximize this
notion of evenness. We also show that essentially all Euclidean rhythms are deep: each distinct
distance between onsets occurs with a unique multiplicity, and these multiplicies form an interval
1, 2, . . . , k − 1. Finally, we characterize all deep rhythms, showing that they form a subclass of
generated rhythms, which in turn proves a useful property called shelling. All of our results for
musical rhythms apply equally well to musical scales. In addition, many of the problems we
explore are interesting in their own right as distance geometry problems on the circle; some of
the same problems were explored by Erdős in the plane.

∗
Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge,
Massachusetts, USA, edemaine@[Link]
†
Departament de Matemática Aplicada, Universidad Politécnica de Madrid, Madrid, Spain, fmartin@[Link]
‡
School of Computing, Queen’s University, Kingston, Ontario, Canada, {henk,daver}@[Link]
§
School of Computer Science, McGill University, Montréal, Québec, Canada, {perouz,godfried}@[Link]
¶
Centre for Interdisciplinary Research in Music Media and Technology The Schulich School of Music McGill
University. Supported by FQRNT and NSERC.
k
Department of Computer Science, Stanford University, Stanford, California, USA, winograd@[Link]
∗∗
Departament de Matemàtica Aplicada II, Universitat Politècnica de Catalunya, Barcelona, Spain,
[Link]@[Link]. Supported by the Government of Spain grant MEC SB2003-0270, and by the projects MCYT-
FEDER BFM2003-00368 and Gen. Cat 2001SGR00224.

1
1 Introduction
Polygons on a circular lattice, African bell rhythms [Tou03], musical scales [CEK99], spallation
neutron source accelerators in nuclear physics [Bjo03b], linear sequences in mathematics [LP92],
mechanical words and stringology in computer science [Lot02], drawing digital straight lines in
computer graphics [KR04], calculating leap years in calendar design [HR04, Asc02], and an ancient
algorithm for computing the greatest common divisor of two numbers, originally described by
Euclid [Euc56, Fra56]—what do these disparate concepts all have in common? The short answer
is, “patterns distributed as evenly as possible”. For the long answer, please read on.
Mathematics and music have been intimately intertwined since over 2,500 years ago when the
famous Greek mathematician, Pythagoras of Samos (circa 500 B.C.), discovered that the pleasing
experience of musical harmony is the result of ratios of small integers [Ash03]. Most of this in-
teraction between the two fields, however, has been in the domain of pitch and scales. For some
historical snapshots of this interaction, we refer the reader to H. S. M. Coxeter’s delightful ac-
count [Cox62]. In music theory, much attention has been devoted to the study of intervals used
in pitch scales [For73], but relatively little work has been devoted to the analysis of time duration
intervals of rhythm. Some notable recent exceptions are the books by Simha Arom [Aro91], Justin
London [Lon04] and Christopher Hasty [Has97].
In this paper, we study various mathematical properties of musical rhythms and scales that
are all, at some level, connected to an algorithm of another famous ancient Greek mathematician,
Euclid of Alexandria (circa 300 B.C.). We begin (in Section 2) by showing several mathematical
connections between musical rhythms and scales, the work of Euclid, and other areas of knowledge
such as nuclear physics, calendar design, mathematical sequences, and computer science. In par-
ticular, we define the notion of Euclidean rhythms, generated by an algorithm similar to Euclid’s.
Then, in the more technical part of the paper (Sections 3–5.2), we study two important properties
of rhythms and scales, called evenness and deepness, and show how these properties relate to the
work of Euclid.
The Euclidean algorithm has been connected to music theory previously by Viggo Brun [Bru64].
Brun used Euclidean algorithms to calculate the lengths of strings in musical instruments between
two lengths l and 2l, so that all pairs of adjacent strings have the same length ratios. In contrast,
we relate the Euclidean algorithm to rhythms and scales in world music.
Musical rhythms and scales can both be seen as two-way infinite binary sequences [Tou02]. In
a rhythm, each bit represents one unit of time called a pulse (for example, the length of a sixteenth
note), a one bit represents a played note or onset (for example, a sixteenth note), and a zero bit
represents a silence (for example, a sixteenth rest). In a scale, each bit represents a pitch (spaced
uniformly in log-frequency space), and zero or one represents whether the pitch is absent or present
in the scale. Here we suppose that all time intervals between onsets in a rhythm are multiples of a
fixed time unit, and that all tone intervals between pitches in a scale are multiples of a fixed tonal
unit (in logarithm of frequency).
The time dimension of rhythms and the pitch dimension of scales have an intrinsically cyclic
nature, cycling every measure and every octave, respectively. In this paper, we consider rhythms
and scales that match this cyclic nature of the underlying space. In the case of rhythms, such
cyclic rhythms are also called timelines, rhythmic phrases or patterns that are repeated throughout
a piece; in the remainder of the paper, we use the term “rhythm” to mean “timeline”. The infinite
bit sequence representation of a cyclic rhythm or scale is just a cyclic repetition of some n-bit string,
corresponding to the timespan of a single measure or the log-frequency span of a single octave. To
properly represent the cyclic nature of this string, we imagine assigning the bits to n points equally

2
spaced around a circle of circumference n [McC98]. A rhythm or scale can therefore be represented
as a subset of these n points. We use k to denote the size of this subset; that is, k is the number of
onsets in a rhythm or pitches in a scale. For uniformity, the terminology in the remainder of this
paper speaks primarily about rhythms, but the notions and results apply equally well to scales.
In this paper, we use four representations of rhythms of timespan n. The first representation
is the commonly used box-like representation, also known as the Time Unit Box System (TUBS),
which is a sequence of n ‘×’s and ‘ · ’s where ‘×’ represents an onset and ‘ · ’ denotes a silence (a zero
bit) [Tou02]. This notation was used and taught in the West by Philip Harland at the University
of California, Los Angeles, in 1962, and it was made popular in the field of ethnomusicology by
James Koetting [Koe70]. However, such box notation has been used in Korea for hundreds of
years [HK81]. The second representation of rhythms and scales we use is the clockwise distance
sequence, which is a sequence of integers that sum up to n and represent the lengths of the intervals
between consecutive pairs of onsets, measuring clockwise arc-lengths or distances around the circle
of circumference n. The third representation of rhythms and scales writes the onsets as a subset of
the set of all pulses, numbered 0, 1, . . . , n − 1, with a subscript of n on the right-hand side of the
subset to denote the timespan. Clough and Douthett [CD91] use this notation to represent scales.
For example, the Cuban clave Son rhythm can be represented as [× · · × · · × · · · × · × · · · ]
in box-like notation, (3, 3, 4, 2, 4) in clockwise distance sequence notation, and {0, 3, 6, 10, 12}16 in
subset notation. Finally, the fourth representation is a graphical clock diagram [Tou02], such as
Figure 1, in which the zero label denotes the start of the rhythm and time flows in a clockwise
direction. In such clock diagrams we usually connect adjacent onsets by line segments, forming a
polygon.

Even Rhythms. Consider the following three 12/8-time rhythms expressed in box-like notation:
[× · × · × · × · × · × · ], [× · × · ×× · × · × · ×], and [× · · · ×× · · ××× · ]. It is intuitively clear
that the first rhythm is more even (well spaced) than the second rhythm, and that the second
rhythm is more even than the third rhythm. In fact, the second rhythm is the internationally most
well known of all African timelines. It is traditionally played on an iron bell, and is known on the
world scene mainly by its Cuban name Bembé [Tou03]. Traditional rhythms tend to exhibit such
properties of evenness to some degree.
Why do many traditional rhythms display such evenness? Many are timelines (also sometimes
called claves), that is, rhythms repeated throughout a piece that serve as a rhythmic reference
point [Uri96, Ort95]. Often these claves have a call-and-response structure, meaning that the
pattern is divided into two parts: the first poses a rhythmic question, usually by creating rhythmic
tension, and the second part answers this question by releasing that tension. A good example of this
structure is the popular clave Son [× · · × · · × · · · × · × · · · ]. This clave creates such tension
through syncopation, which can be found between the second and third onsets as well as between
the third and fourth onsets. The latter is weak syncopation because the strong beat at position 8 lies
half-way between the third and fourth onsets. (The strong beats of the underlying 4/4 meter (beat)
occur at positions 0, 4, 8, and 12.) On the other hand, the former syncopation is strong because the
strong beat at position 4 is closer to the second onset than to the third onset [GMRT05]. Claves
played with instruments that produce unsustained notes often use syncopation and accentuation
to bring about rhythmic tension. Many clave rhythms create syncopation by evenly distributing
onsets in contradiction with the pulses of the underlying meter. For example, in the clave Son,
the first three onsets are equally spaced at the distance of three sixteenth pulses, which forms a
contradiction because 3 does not divide 16. Then, the response of the clave answers with an offbeat
onset, followed by an onset on the fourth strong beat of a 4/4 meter, releasing that rhythmic

3
 0 0 0
15 1 15 1 15 1
14 2 14 2 14 2

13 3 13 3 13 3

12 4 12 4 12 4

11 5 11 5 11 5

10 6 10 6 10 6
9 7 9 7 9 7
8 8 8

(a) Shiko (b) Son (c) Soukous

0 0 0
15 1 15 1 15 1
14 2 14 2 14 2

13 3 13 3 13 3

12 4 12 4 12 4

11 5 11 5 11 5

10 6 10 6 10 6
9 7 9 7 9 7
8 8 8
(d) Rumba (e) Bossa−Nova (f) Gahu

Figure 1: The six fundamental African and Latin American rhythms which all have equal sum of pairwise
geodesic distances; yet intuitively, the Bossa-Nova rhythm is more “even” than the rest.

tension.
On the other hand, a rhythm should not be too even, such as the example [× · × · × · × · × ·
× · ]. Indeed, in the most interesting rhythms with k onsets and timespan n, k and n are relatively
prime (have no common divisor larger than 1). This property is natural because the rhythmic
contradiction is easier to obtain if the onsets do not coincide with the strong beats of the meter.
Also, we find that many claves have an onset on the last strong beat of the meter, as does the clave
Son. This is a natural way to respond in the call-and-response structure. A different case is that
of the Bossa-Nova clave [× · · × · · × · · · × · · × · · ]. This clave tries to break the feeling of the
pulse and, although it is very even, it produces a cycle that perceptually does not coincide with
the beginning of the meter.
This prevalence of evenness in world rhythms has led to the study of mathematical measures of
evenness in the new field of mathematical ethnomusicology [Che02, Tou04b, Tou05], where they may
help to identify, if not explain, cultural preferences of rhythms in traditional music. Furthermore,
evenness in musical chords plays a significant role in the efficacy of voice leading as discussed in
the work of Tymoczko [Tym06].
The notion of maximally even sets with respect to scales represented on a circle was introduced
by Clough and Douthett [CD91]. According to Block and Douthett [BD94], Douthett and Entringer
went further by constructing several mathematical measures of the amount of evenness contained in
a scale; see [BD94, page 40]. One of their evenness measures simply sums the interval arc-lengths
(geodesics along the circle) between all pairs of onsets (or more precisely, onset points). This
measure differentiates between rhythms that differ widely from each other. For example, the two
four-onset rhythms [× · · · × · · · × · · · × · · · ] and [× · × · × · · × · · · · · · · · ] yield evenness

4
values of 32 and 23, respectively, reflecting clearly that the first rhythm is more evenly spaced than
the second. However, the measure is too coarse to be useful for comparing rhythm timelines such
as those studied in [Tou02, Tou03]. For example, all six fundamental 4/4-time clave/bell patterns
discussed in [Tou02] and shown in Figure 1 have an equal pairwise sum of geodesic distances,
namely 48, yet the Bossa-Nova clave is intuitively more even than, say, the Soukous and Rumba
claves.
Another distance measure that has been considered is the sum of pairwise chordal distances
between adjacent onsets, measured by Euclidean distance between points on the circle. It can
be shown that the rhythms maximizing this measure of evenness are precisely the rhythms with
maximum possible area. Rappaport [Rap05] shows that many of the most common chords and
scales in Western harmony correspond to these maximum-area sets. This evenness measure is finer
than the sum of pairwise arc-lengths, but it still does not distinguish half the rhythms in Figure 1.
Specifically, the Son, Rumba, and Gahu claves have the same occurrences of arc-lengths between
consecutive onsets, so they also have the same occurrences (and hence total) of distances between
consecutive onsets.
The measure of evenness we consider here is the sum of all pairwise Euclidean distances be-
tween points on the circle, as described by Block and Douthett [BD94]. This measure is more
discriminating than the others, and is therefore the preferred measure of evenness. For example,
this measure distinguishes all of the six rhythms in Figure 1, ranking the Bossa-Nova rhythm as
the most even, followed by the Son, Rumba, Shiko, Gahu, and Soukous. Intuitively, the rhythms
with a larger sum of pairwise chordal distances have more “well spaced” onsets.
In Section 4, we study the mathematical and computational aspects of rhythms that maximize
evenness. We describe three algorithms that generate such rhythms, show that these algorithms
are equivalent, and show that in fact the rhythm of maximum evenness is essentially unique. These
results characterize rhythms with maximum evenness. One of the algorithms is the Euclidean-
like algorithm from Section 2, proving that the rhythms of maximum evenness are precisely the
Euclidean rhythms from that section.

Deep Rhythms. Another important property of rhythms and scales that we study in this paper
is deepness. Consider a rhythm with k onsets and timespan n, represented as a set of k points on a
circle of circumference n. Now measure the arc-length/geodesic distances along the circle between
all pairs of onsets. A musical scale or rhythm is Winograd-deep if every distance 1, 2, . . . , ⌊n/2⌋ has
a unique number of occurrences (called the multiplicity of the distance). For example, the rhythm
[××× · × · ] is Winograd-deep because distance 1 appears twice, distance 2 appears thrice, and
distance 3 appears once.
The notion of deepness in scales was introduced by Winograd in an oft-cited but unpublished
class project report from 1966 [Win66], disseminated and further developed by the class instructor
Gamer in 1967 [Gam67a, Gam67b], and considered further in numerous papers and books, e.g.,
[CEK99, Joh03]. Equivalently, a scale is Winograd-deep if the number of onsets it has in common
with each of its cyclic shifts (rotations) is unique. This equivalence is the Common Tone Theorem
[Joh03, page 42], and it is originally described by Winograd [Win66] (who in fact uses this definition
as his primary definition of “deep”). Deepness is one property of the ubiquitous Western diatonic
12-tone major scale [× · × · ×× · × · × · ×] [Joh03], and it captures some of the rich structure that
perhaps makes this scale so attractive.
Winograd-deepness translates directly from scales to rhythms. For example, the diatonic major
scale is equivalent to the famous Cuban rhythm Bembé [Pre83, Tou03]. Figure 2 shows a graph-
ical example of a Winograd-deep rhythm. However, the notion of Winograd-deepness is rather

5
 0
15 1
14 2

13 3

12 4

11 5

10 6
9 7
8

Figure 2: A rhythm with k = 7 onsets and timespan n = 16 that is Winograd-deep and thus Erdős-deep.
Distances ordered by multiplicity from 1 to 6 are 2, 7, 4, 1, 6, and 5. The dotted line shows how the rhythm
is generated by multiples of m = 5.

restrictive for rhythms, because it requires half of the pulses in a timespan (rounded to a nearest
integer) to be onsets. In contrast, for example, the popular Bossa-Nova rhythm [× · · × · · × · · ·
× · · × · · ] = {0, 3, 6, 10, 13}16 pictures in Figure 1 has only five onsets in a timespan of sixteen.
Nonetheless, if we focus on just the distances that appear at least once between two onsets, then
the multiplicities of occurrence are all unique and form an interval starting at 1: distance 4 occurs
once, distance 7 occurs twice, distance 6 occurs thrice, and distance 3 occurs four times.
We therefore define a rhythm (or scale) to be Erdős-deep if it has k onsets and, for every
multiplicity 1, 2, . . . , k − 1, there is a nonzero arc-length/geodesic distance determined by the points
on the circle with exactly that multiplicity. The same definition is made by Toussaint [Tou04a].
Every Winograd-deep rhythm is also Erdős-deep, so this definition is strictly more general.
To further clarify the difference between Winograd-deep and Erdős-deep rhythms, it is useful
to consider which distances can appear. For a rhythm to be Winograd-deep, all the distances
between 1 and k − 1 must appear a unique number of times. In contrast, to be an Erdős-deep
rhythm, it is only required that any distance that appears must have a unique multiplicity. Thus,
the Bossa-Nova rhythm is not Winograd-deep because distances 1, 2 and 5 do not appear.
The property of Erdős deepness involves only the distances between points in a set, and is
thus a feature of distance geometry—in this case, in the discrete space of n points equally spaced
around a circle. In 1989, Paul Erdős [Erd89] considered the analogous question in the plane,
asking whether there exist n points in the plane (no three on a line and no four on a circle) such
that, for every i = 1, 2, . . . , n − 1, there is a distance determined by these points that occurs exactly
i times. Solutions have been found for n between 2 and 8, but in general the problem remains open.
Palásti [Pal89] considered a variant of this problem with further restrictions—no three points form
a regular triangle, and no one is equidistant from three others—and solved it for n = 6.
In Section 5, we characterize all rhythms that are Erdős-deep. In particular, we prove that all
deep rhythms, besides one exception, are generated, meaning that the rhythm can be represented
as {0, m, 2m, . . . , (k − 1)m}n for some integer m, where all arithmetic is modulo n. In the context
of scales, the concept of “generated” was defined by Wooldridge [Woo93] and used by Clough et
al. [CEK99]. For example, the rhythm in Figure 2 is generated with m = 5. Our characterization
generalizes a similar characterization for Winograd-deep scales proved by Winograd [Win66], and
independently by Clough et al. [CEK99].
In the pitch domain, generated scales are very common. The Pythagorean tuning is a good
example: all its pitches are generated from the fifth of ratio 3 : 2 modulo the octave. √ Another
example is the equal-tempered scale, which is generated with a half-tone of ratio 12 2 [Bar04a].

6
Generated scales are also of interest in the theory of the well-formed scales [Car98].
Generated rhythms have an interesting property called shellability. If we remove the “last”
generated onset 14 from the rhythm in Figure 2, the resulting rhythm is still generated, and this
process can be repeated until we run out of onsets. In general, every generated rhythm has a
shelling in the sense that it is always possible to remove a particular onset and obtain another
generated rhythm.
Shellings of rhythms play an important role in musical improvisation. For example, most African
drumming music consists of rhythms operating on three different strata: the unvarying timeline
usually provided by one or more bells, one or more rhythmic motifs played on drums, and an
improvised solo (played by the lead drummer) riding on the other rhythmic structures. Shellings
of rhythms are relevant to the improvisation of solo drumming in the context of such a rhythmic
background. The solo improvisation must respect the style and feeling of the piece which is usually
determined by the timeline. A common technique to achieve this effect is to “borrow” notes from
the timeline, and to alternate between playing subsets of notes from the timeline and from other
rhythms that interlock with the timeline [Ank97, Aga86]. In the words of Kofi Agawu [Aga86],
“It takes a fair amount of expertise to create an effective improvisation that is at the same time
stylistically coherent”. The borrowing of notes from the timeline may be regarded as a fulfillment
of the requirements of style coherence.
Of course, some subsets of notes of a rhythm may be better choices than others. For example,
it seems reasonable that, if a rhythm is deep, one should select subsets of the rhythm that are also
deep. Furthermore, a shelling seems a natural way to decrease or increase the density of the notes
in an improvisation that respects these constraints. For example, in the Bembé bell timeline [× ·
× · ×× · × · × · ×], which is deep, one possible shelling is [× · × · ×× · × · × · · ], [× · × · × · ·
× · × · · ], [× · × · · · · × · × · · ], [× · × · · · · × · · · · ]. All five rhythms sound good and are
stylistically coherent. To our knowledge, shellings have not been studied from the musicological
point of view. However, they may be useful both for theoretical analysis as well as providing formal
rules for “improvisation” techniques.
One of the consequences of our characterization that we obtain in Section 5 is that every Erdős-
deep rhythm has a shelling. More precisely, it is always possible to remove a particular onset that
preserves the Erdős-deepness property. Finally, to tie everything together, we show that essentially
all Euclidean rhythms (or equivalently, rhythms that maximize evenness) are Erdős-deep.

2 Euclid and Evenness in Various Disciplines

In this section, we first describe Euclid’s classic algorithm for computing the greatest common
divisor of two integers. Then, through an unexpected connection to timing systems in neutron
accelerators, we see how the same type of algorithm can be used as an approach to maximizing
“evenness” in a binary string with a specified number of zeroes and ones. This algorithm defines an
important family of rhythms, called Euclidean rhythms, which we show appear throughout world
music. Finally, we see how similar ideas have been used in algorithms for drawing digital straight
lines and in combinatorial strings called Euclidean strings.

2.1 The Euclidean Algorithm for Greatest Common Divisors

The Euclidean algorithm for computing the greatest common divisor of two integers is one of the
oldest known algorithms (circa 300 B.C.). It was first described by Euclid in Proposition 2 of
Book VII of Elements [Euc56, Fra56]. Indeed, Donald Knuth [Knu98] calls this algorithm the

7
“granddaddy of all algorithms, because it is the oldest nontrivial algorithm that has survived to
the present day”.
The idea of the algorithm is simple: repeatedly replace the larger of the two numbers by their
difference until both are equal. This final number is then the greatest common divisor. For example,
consider the numbers 5 and 13. First, 13 − 5 = 8; then 8 − 5 = 3; next 5 − 3 = 2; then 3 − 2 = 1;
and finally 2 − 1 = 1. Therefore, the greatest common divisor of 5 and 13 is 1; in other words, 5
and 13 are relatively prime.
The algorithm can also be described succinctly in a recursive manner as follows [CLRS01]. Let
k and n be the input integers with k < n.

Algorithm Euclid(k, n)
1. if k = 0 then return n
2. else return Euclid(n mod k, k)

Running this algorithm with k = 5 and n = 13, we obtain Euclid(5, 13) = Euclid(3, 5) =
Euclid(2, 3) = Euclid(1, 2) = Euclid(0, 1) = 1. Note that this division version of Euclid’s
algorithm skips one of the steps (5, 8) made by the original subtraction version.

2.2 Evenness and Timing Systems in Neutron Accelerators

One of our main musical motivatations is to find rhythms with a specified timespan and number
of onsets that maximize evenness. Bjorklund [Bjo03b, Bjo03a] was faced with a similar problem of
maximizing evenness, but in a different context: the operation of components such as high-voltage
power supplies of spallation neutron source (SNS) accelerators used in nuclear physics. In this
setting, a timing system controls a collection of gates over a time window divided into n equal-
length intervals. (In the case of SNS, each interval is 10 seconds.) The timing system can send
signals to enable a gate during any desired subset of the n intervals. For a given number n of time
intervals, and another given number k < n of signals, the problem is to distribute the pulses as
evenly as possible among these n intervals. Bjorklund [Bjo03b] represents this problem as a binary
sequence of k ones and n−k zeroes, where each bit represents a time interval and the ones represent
the times at which the timing system sends a signal. The problem then reduces to the following:
construct a binary sequence of n bits with k ones such that the k ones are distributed as evenly as
possible among the (n − k) zeroes.
One simple case is when k evenly divides n (without remainder), in which case we should place
ones every n/k bits. For example, if n = 16 and k = 4, then the solution is [1000100010001000].
This case corresponds to n and k having a common divisor of k. More generally, if the greatest
common divisor between n and k is g, then we would expect the solution to decompose into g
repetitions of a sequence of n/g bits. Intuitively, a string of maximum evenness should have this
kind of symmetry, in which it decomposes into more than one repetition, whenever such symmetry is
possible. This connection to greatest common divisors suggests that a rhythm of maximum evenness
might be computed using an algorithm like Euclid’s. Indeed, Bjorklund’s algorithm closely mimics
the structure Euclid’s algorithm, although this connection has never been mentioned before.
We describe Bjorklund’s algorithm by using one of his examples. Consider a sequence with
n = 13 and k = 5. Because 13 − 5 = 8, we start by considering a sequence consisting of 5 ones
followed by 8 zeroes which should be thought of as 13 sequences of one bit each:

[1][1][1][1][1][0][0][0][0][0][0][0][0]

8
If there is more than one zero the algorithm moves zeroes in stages. We begin by taking zeroes
one at a time (from right to left), placing a zero after each one (from left to right), to produce five
sequences of two bits each, with three zeroes remaining:
[10] [10] [10] [10] [10] [0] [0] [0]
Next we distribute the three remaining zeros in a similar manner, by placing a [0] sequence after
each [10] sequence:
[100] [100] [100] [10] [10]
Now we have three sequences of three bits each, and a remainder of two sequences of two bits each.
Therefore we continue in the same manner, by placing a [10] sequence after each [100] sequence:
[10010] [10010] [100]
The process stops when the remainder consists of only one sequence (in this case the sequence
[100]), or we run out of zeroes (there is no remainder). The final sequence is thus the concatenation
of [10010], [10010], and [100]:
[1001010010100]
We could proceed further in this process by inserting [100] into [10010] [10010]. However,
Bjorklund argues that, because the sequence is cyclic, it does not matter (hence his stopping rule).
For the same reason, if the initial sequence has a group of ones followed by only one zero, the
zero is considered as a remainder consisting of one sequence of one bit, and hence nothing is done.
Bjorklund [Bjo03b] shows that the final sequence may be computed from the initial sequence using
O(n) arithmetic operations in the worst case.
A more convenient and visually appealing way to implement this algorithm by hand is to perform
the sequence of insertions in a vertical manner as follows. First take five zeroes from the right and
place them under the five ones on the left:
11111000
00000
Then move the three remaining zeroes in a similar manner:
11111
00000
000
Next place the two remainder columns on the right under the two leftmost columns:
1 11
0 00
0 00
1 1
0 0
Here the process stops because the remainder consists of only one column. The final sequence is
obtained by concatenating the three columns from left to right:
1001010010100
Bjorklund’s algorithm applied to a string of n bits consisting of k ones and n − k zeros has
the same structure as running Euclid(k, n). Indeed, Bjorklund’s algorithm uses the repeated
subtraction form of division, just as Euclid did in his Elements [Euc56]. It is also well known
that applying the algorithm Euclid(k, n) to two O(n) bit numbers (binary sequences of length n)
causes it to perform O(n) arithmetic operations in the worst case [CLRS01].

9
 0 0
7 1 7 1
2
3
6 2 6 2
3

5 3 5 3
4 4
(a) (b)

Figure 3: (a) The Euclidean rhythm E(3, 8) is the Cuban tresillo. (b) The Euclidean rhythm E(5, 8) is the
Cuban cinquillo.

2.3 Euclidean Rhythms

The binary sequences generated by Bjorklund’s algorithm, as described in the preceding, may
be considered as one family of rhythms. Furthermore, because Bjorklund’s algorithm is a way
of visualizing the repeated-subtraction version of the Euclidean algorithm, we call these rhythms
Euclidean rhythms. We denote the Euclidean rhythm by E(k, n), where k is the number of ones
(onsets) and n (the number of pulses) is the length of the sequence (zeroes plus ones). For ex-
ample, E(5, 13) = [1001010010100]. The zero-one notation is not ideal for representing binary
rhythms because it is difficult to visualize the locations of the onsets as well as the duration of
the inter-onset intervals. In the more iconic box notation, the preceding rhythm is written as
E(5, 13) = [× · · × · × · · × · × · · ].
The rhythm E(5, 13) is in fact used in Macedonian music [Aro04], but having a timespan of 13
(and defining a measure of length 13), it is rarely found in world music. For contrast, let us consider
two widely used values of k and n; in particular, what is E(3, 8)? Applying Bjorklund’s algorithm
to the corresponding sequence [11100000], the reader may easily verify that the resulting Euclidean
rhythm is E(3, 8) = [× · · × · · × · ]. Figure 3(a) shows a clock diagram of this rhythm, where the
numbers by the sides of the triangle indicate the arc-lengths between those onsets.
The Euclidean rhythm E(3, 8) is one of the most famous on the planet. In Cuba, it goes by
the name of the tresillo, and in the USA, it is often called the Habanera rhythm. It was used in
hundreds of rockabilly songs during the 1950’s. It can often be heard in early rock-and-roll hits in
the left-hand patterns of the piano, or played on the string bass or saxophone [Bre99, Flo99, Mor96].
A good example is the bass rhythm in Elvis Presley’s Hound Dog [Bre99]. The tresillo pattern is
also found widely in West African traditional music. For example, it is played on the atoke bell in
the Sohu, an Ewe dance from Ghana [Kau80]. The tresillo can also be recognized as the first bar
(first eight pulses) of the ubiquitous two-bar clave Son shown in Figure 1(b).
In the two examples E(5, 13) and E(3, 8), there are fewer ones than zeros. If instead there are
more ones than zeros, Bjorklund’s algorithm yields the following steps with, for example, k = 5
and n = 8:

[1 1 1 1 1 0 0 0]
[10] [10] [10] [1] [1]
[101] [101] [10]
[1 0 1 1 0 1 1 0]

10
 0 0
11 1 11 1
10 2 10 2

9 3 9 3

8 4 8 4
7 5 7 5
6 6
Figure 4: These two rhythms are instances of the same rhythm necklace.

The resulting Euclidean rhythm is E(5, 8) = [× · ×× · ×× · ]. Figure 3(b) shows a clock diagram
for this rhythm. It is another famous rhythm on the world scene. In Cuba, it goes by the name of
the cinquillo and it is intimately related to the tresillo [Flo99]. It has been used in jazz throughout
the 20th century [Rah96], and in rockabilly music. For example, it is the hand-clapping pattern
in Elvis Presley’s Hound Dog [Bre99]. The cinquillo pattern is also widely used in West African
traditional music [Rah87, Tou02], as well as Egyptian [Hag03] and Korean [HK81] music.
We show in this paper that Euclidean rhythms have two important properties: they maximize
evenness and they are deep. The evenness property should come as no surprise, given how we
designed the family of rhythms. To give some feeling for the deepness property, we consider the
two examples in Figure 3, which have been labeled with the distances between all pairs of onsets,
measured as arc-lengths. The tresillo in Figure 3(a) has one occurrence of distance 2 and two
occurrences of distance 3. The cinquillo in Figure 3(b) contains one occurrence of distance 4, two
occurrences of distance 1, three occurrences of distance 2, and four occurrences of distance 3. Thus,
every distance has a unique multiplicity, making these rhythms Erdős-deep.

2.4 Euclidean Rhythms in Traditional World Music

In this section, we list all the Euclidean rhythms found in world music that we have collected
so far, restricting attention to those in which k and n are relatively prime. In some cases, the
Euclidean rhythm is a rotated version of a commonly used rhythm. If a rhythm is a rotated version
of another, we say that they are instances of the same necklace. Thus a rhythm necklace is a
clockwise distance sequence that disregards the starting point in the cycle. Figure 4 illustrates an
example of two rhythms that are instances of the same necklace.
Rhythms in which k and n have a common divisor larger than 1 are common all over the
planet in traditional, classical, and popular genres of music. For example, E(4, 12) = [× · · × · ·
× · · × · · ] is the 12/8-time Fandango clapping pattern in the Flamenco music of southern Spain,
where ‘×’ denotes a loud clap and ‘ · ’ denotes a soft clap [DBFG+ 04]. However, the string itself
is periodic: E(4, 12) has period 3, even though it appears in a timespan of 12. For this reason, we
restrict ourselves to the more interesting Euclidean rhythms that do not decompose into repetitions
of shorter Euclidean rhythms. We are also not concerned with rhythms that have only one onset
([× · ], [× · · ], etc.), and similarly with any repetitions of these rhythms (for example, [× · × · ]).
There are surprisingly many Euclidean rhythms with k and n relatively prime that are found
in world music. The following list includes more than 40 such rhythms uncovered so far.
E(2, 3) = [×× · ] = (12) is a common Afro-Cuban drum pattern when started on the second

11
onset as in [× · ×]. For example, it is the conga rhythm of the (6/8)-time Swing Tumbao [Klő97].
It is common in Latin American music, as for example in the Cueca [vdL95], and the coros de
clave [Rod97]. It is common in Arabic music, as for example in the Al Táer rhythm of Nu-
bia [Hag03]. It is also a rhythmic pattern of the Drum Dance of the Slavey Indians of Northern
Canada [Asc75].
E(2, 5) = [× · × · · ] = (23) is a rhythm found in Greece, Namibia, Rwanda and Central
Africa [Aro04]. It is also a 13th century Persian rhythm called Khafif-e-ramal [Wri78], as well
as the rhythm of the Macedonian dance Makedonka [Sin74]. Tchaikovsky used it as the metric
pattern in the second movement of his Symphony No. 6 [Kei91]. Started on the second onset as in
[× · · × · ] it is a rhythm found in Central Africa, Bulgaria, Turkey, Turkestan and Norway [Aro04].
It is also the metric pattern of Dave Brubeck’s Take Five, as well as Mars from The Planets by
Gustav Holst [Kei91]. Both starting points determine metric patterns used in Korean music [HK81].
E(3, 4) = [××× · ] = (112) is a pattern used in the Baiaó rhythm of Brazil [Uri93], as well as
the polos of Bali [Mon85]. Started on the second onset, as in [×× · ×], it is the Catarete rhythm
of the indigenous people of Brazil [Uri93]. Started on the third onset, as in [× · ××], it is the
archetypal pattern of the Cumbia from Colombia [Man85], as well as a Calypso rhythm from
Trinidad [Eva66]. It is also a 13th century Persian rhythm called Khalif-e-saghil [Wri78], as well
as the trochoid choreic rhythmic pattern of ancient Greece [Mat85]. Started on the silent note
(anacrusis), as in [ · ×××], it is a popular flamenco rhythm used in the Taranto, the Tiento, the
Tango, and the Tanguillo [Gam02]. It is also the Rumba clapping pattern in flamenco, as well as a
second pattern used in the Baiaó rhythm of Brazil [Uri93].
E(3, 5) = [× · × · ×] = (221), when started on the second onset, is another 13th century Persian
rhythm by the name of Khafif-e-ramal [Wri78], as well as a Romanian folk-dance rhythm [PC69],
and the Sangsa Pyǒlgok drum pattern in Korean music [HK81].
E(3, 7) = [× · × · × · · ] = (223) is a rhythm found in Greece, Turkestan, Bulgaria, and North-
ern Sudan [Aro04]. It is the Dáwer turan rhythmic pattern of Turkey [Hag03]. It is the Ruchenitza
rhythm used in a Bulgarian folk-dance [Pre83], as well as the rhythm of the Macedonian dance
Eleno Mome [Sin74]. It is also the rhythmic pattern of Dave Brubeck’s Unsquare Dance, and
Pink Floyd’s Money [Kei91]. Started on the second onset, as in [× · × · · × · ], it is a Serbian
rhythm [Aro04]. Started on the third onset, as in [× · · × · × · ], it is a rhythmic pattern found in
Greece and Turkey [Aro04]. In Yemen it goes by the name of Daasa al zreir [Hag03]. It is also the
rhythm of the Macedonian dance Tropnalo Oro [Sin74], the rhythm for the Bulgarian Makedonsko
Horo dance [Wad04], as well as the meter and clapping pattern of the tı̄vrā tāl of North Indian
music [Cla00].
E(3, 8) = [× · · × · · × · ] = (332) is the Cuban tresillo pattern discussed in the preceding [Flo99],
the most important traditional bluegrass banjo rhythm [Kei91], as well as the Mai metal-blade pat-
tern of the Aka Pygmies [Aro91]. It is common in West Africa [Bra59] and many other parts of the
world such as Greece and Northern Sudan [Aro04]. Curt Sachs [Sac53] and Willi Apel [Ape60] con-
sider it to be one of the most important rhythms in Renaissance music. Indeed, it dates back to the
Ancient Greeks who called it the dochmiac pattern [Bra59]. In India it is one of the 35 sūlādi tālas
of Karnatak music [Lon04]. Started on the second onset, it is a drum pattern of the Samhyǒn Todǔri
Korean instrumental music [HK81]. It is also found in Bulgaria and Turkey [Aro91]. Started on the
third onset, it is the Nandon Bawaa bell pattern of the Dagarti people of northwest Ghana [HAD95],
and is also found in Namibia and Bulgaria [Aro91].
E(3, 11) = [× · · · × · · · × · · ] = (443) is the metric pattern of the savārı̄ tāl of North Indian
music [Cla00].
E(3, 14) = [× · · · · × · · · · × · · · ] = (554) is the clapping pattern of the dhamār tāl of North

12
Indian music [Cla00].
E(4, 5) = [×××× · ] = (1112) is the rhythmic pattern of the Mirena rhythm of Greece [Hag03].
Started on the fourth onset, as in [× · ×××], it is the Tik rhythm of Greece [Hag03].
E(4, 7) = [× · × · × · ×] = (2221) is another Ruchenitza Bulgarian folk-dance rhythm [Pre83].
Started on the third onset, it is the Kalamátianos Greek dance rhythm [Hag03], as well as the
Shaigie rhythmic pattern of Nubia [Hag03]. Started on the fourth (last) onset, it is the rhythmic
pattern of the Dar daasa al mutawasit of Yemen [Hag03].
E(4, 9) = [× · × · × · × · · ] = (2223) is the Aksak rhythm of Turkey [Bră51] (also found in
Greece) as well as the rhythm of the Macedonian dance Kambani Bijat Oro [Sin74] and the Bul-
garian dance Daichovo Horo [Ric04]. In Bulgarian music fast tunes with this metric pattern are
called Dajchovata whereas slow tunes with this same pattern are called Samokovskata [Ric80]. It is
the rhythmic ostinato of a lullaby discovered by Simha Arom in south-western Zaı̈re [Aro04]. It is
the metric pattern used by Dave Brubeck in his well known piece Rondo a la Turk [Kei91]. When
it is started on the second onset, as in [× · × · × · · × · ], it is found in Bulgaria and Serbia [Aro04].
Started on the third onset, as in [× · × · · × · × · ], it is found in Bulgaria and Greece [Aro04].
It is the rhythm of the Macedonian dance Devojče [Sin74]. Finally, when started on the fourth
onset, as in [× · · × · × · × · ], it is a rhythm found in Turkey [Aro04], and is the metric pattern of
Strawberry Soup by Don Ellis [Kei91].
E(4, 11) = [× · · × · · × · · × · ] = (3332) is the metric pattern of the Dhruva tāla of Southern
India [Lon04]. It is also used by Frank Zappa in Outside Now [Kei91]. When it is started on the
third onset, as in [× · · × · × · · × · · ], it is a Serbian rhythmic pattern [Aro04]. When it is started
on the fourth (last) onset it is the Daasa al kbiri rhythmic pattern of Yemen [Hag03].
E(4, 15) = [× · · · × · · · × · · · × · · ] = (4443) is the metric pattern of the pañcam savārı̄ tāl
of North Indian music [Cla00].
E(5, 6) = [××××× · ] = (11112) yields the York-Samai pattern, a popular Arabic rhythm [Sta88].
It is also a handclapping rhythm used in the Al Medēmi songs of Oman [EMF90].
E(5, 7) = [× · ×× · ××] = (21211) is the Nawakhat pattern, another popular Arabic rhythm
[Sta88]. In Nubia it is called the Al Noht rhythm [Hag03].
E(5, 8) = [× · ×× · ×× · ] = (21212) is the Cuban cinquillo pattern discussed in the preced-
ing [Flo99], the Malfuf rhythmic pattern of Egypt [Hag03], as well as the Korean Nong P’yǒn drum
pattern [HK81]. Started on the second onset, it is a popular Middle Eastern rhythm [Wad04], as
well as the Timini rhythm of Senegal, the Adzogbo dance rhythm of Benin [Che79], the Spanish
Tango [Eva66], the Maksum of Egypt [Hag03], and a 13th century Persian rhythm, the Al-saghil-al-
sani [Wri78]. When it is started on the third onset it is the Müsemmen rhythm of Turkey [Bek05].
When it is started on the fourth onset it is the Kromanti rhythm of Surinam.
E(5, 9) = [× · × · × · × · ×] = (22221) is a popular Arabic rhythm called Agsag-Samai [Sta88].
Started on the second onset, it is a drum pattern used by the Venda in South Africa [Rah87], as
well as a Rumanian folk-dance rhythm [PC69]. It is also the rhythmic pattern of the Sigaktistos
rhythm of Greece [Hag03], and the Samai aktsak rhythm of Turkey [Hag03]. Started on the third
onset, it is the rhythmic pattern of the Nawahiid rhythm of Turkey [Hag03].
E(5, 11) = [× · × · × · × · × · · ] = (22223) is the metric pattern of the Savārı̄ tāla used in the
Hindustani music of India [Lon04]. It is also a rhythmic pattern used in Bulgaria and Serbia [Aro04].
In Bulgaria is is used in the Kopanitsa [Ric04]. This metric pattern has been used by Moussorgsky
in Pictures at an Exhibition [Kei91]. Started on the third onset, it is the rhythm of the Macedonian
dance Kalajdzijsko Oro [Sin74], and it appears in Bulgarian music as well [Aro04].
E(5, 12) = [× · · × · × · · × · × · ] = (32322) is a common rhythm played in the Central African
Republic by the Aka Pygmies [Aro91, Che02, CT03]. It is also the Venda clapping pattern of a

13
South African children’s song [Pre83], and a rhythm pattern used in Macedonia [Aro04]. Started
on the second onset, it is the Columbia bell pattern popular in Cuba and West Africa [Klő97],
as well as a drumming pattern used in the Chakacha dance of Kenya [Bar04b]. and also used
in Macedonia [Aro04]. Started on the third onset, it is the Bemba bell pattern used in Northern
Zimbabwe [Pre83], and the rhythm of the Macedonian dance Ibraim Odža Oro [Sin74]. Started on
the fourth onset, it is the Fume Fume bell pattern popular in West Africa [Klő97], and is a rhythm
used in the former Yugoslavia [Aro04]. Finally, when started on the fifth onset it is the Salve bell
pattern used in the Dominican Republic in a rhythm called Canto de Vela in honor of the Virgin
Mary [Far92], as well as the drum rhythmic pattern of the Moroccan Al Kudám [Hag03].
E(5, 13) = [× · · × · × · · × · × · · ] = (32323) is a Macedonian rhythm which is also played by
starting it on the fourth onset as follows: [× · × · · × · · × · × · · ] [Aro04].
E(5, 16) = [× · · × · · × · · × · · × · · · ] = (33334) is the Bossa-Nova rhythm necklace of Brazil.
The actual Bossa-Nova rhythm usually starts on the third onset as follows: [× · · × · · × · · · × · ·
× · · ] [Tou02]. However, other starting places are also documented in world music practices, such
as [× · · × · · × · · × · · · × · · ] [Beh73].
E(6, 7) = [×××××× · ] = (111112) is the Póntakos rhythm of Greece when started on the sixth
(last) onset [Hag03].
E(6, 13) = [× · × · × · × · × · × · · ] = (222223) is the rhythm of the Macedonian dance Mama
Cone pita [Sin74]. Started on the third onset, it is the rhythm of the Macedonian dance Postupano
Oro [Sin74], as well as the Krivo Plovdivsko Horo of Bulgaria [Ric04].
E(7, 8) = [××××××× · ] = (1111112), when started on the seventh (last) onset, is a typical
rhythm played on the Bendir (frame drum), and used in the accompaniment of songs of the Tuareg
people of Libya [Sta88].
E(7, 9) = [× · ××× · ×××] = (2112111) is the Bazaragana rhythmic pattern of Greece [Hag03].
E(7, 10) = [× · ×× · ×× · ××] = (2121211) is the Lenk fahhte rhythmic pattern of Turkey [Hag03].
E(7, 12) = [× · ×× · × · ×× · × · ] = (2122122) is a common West African bell pattern. For
example, it is used in the Mpre rhythm of the Ashanti people of Ghana [Tou03]. Started on the
seventh (last) onset, it is a Yoruba bell pattern of Nigeria, a Babenzele pattern of Central Africa,
and a Mende pattern of Sierra Leone [Sto05].
E(7, 15) = [× · × · × · × · × · × · × · · ] = (2222223) is a Bulgarian rhythm when started on the
third onset [Aro04].
E(7, 16) = [× · · × · × · × · · × · × · × · ] = (3223222) is a Samba rhythm necklace from Brazil.
The actual Samba rhythm is [× · × · · × · × · × · · × · × · ] obtained by starting E(7, 16) on the
last onset, and it coincides with a Macedonian rhythm [Aro04]. When E(7, 16) is started on the
fifth onset it is a clapping pattern from Ghana [Pre83]. When it is started on the second onset it
is a rhythmic pattern found in the former Yugoslavia [Aro04].
E(7, 17) = [× · · × · × · · × · × · · × · × · ] = (3232322) is a Macedonian rhythm when started
on the second onset [Sin74].
E(7, 18) = [× · · × · × · · × · × · · × · × · · ] = (3232323) is a Bulgarian rhythmic pattern [Aro04].
E(8, 17) = [× · × · × · × · × · × · × · × · · ] = (22222223) is a Bulgarian rhythmic pattern which
is also started on the fifth onset [Aro04].
E(8, 19) = [× · · × · × · × · · × · × · × · · × · ] = (32232232) is a Bulgarian rhythmic pattern
when started on the second onset [Aro04].
E(9, 14) = [× · ×× · ×× · ×× · ×× · ] = (212121212), when started on the second onset, is the
rhythmic pattern of the Tsofyan rhythm of Algeria [Hag03].
E(9, 16) = [× · ×× · × · × · ×× · × · × · ] = (212221222) is a rhythm necklace used in the Cen-
tral African Republic [Aro91]. When it is started on the second onset it is a bell pattern of the

14
Luba people of Congo [Mus02]. When it is started on the fourth onset it is a rhythm played in West
and Central Africa [Flo99], as well as a cow-bell pattern in the Brazilian samba [Sol96]. When it is
started on the penultimate onset it is the bell pattern of the Ngbaka-Maibo rhythms of the Central
African Republic [Aro91].
E(9, 22) =[× · · × · × · · × · × · · × · × · · × · × · ] = (323232322) is a Bulgarian rhythmic pat-
tern when started on the second onset [Aro04].
E(9, 23) =[× · · × · × · · × · × · · × · × · · × · × · · ] = (323232323) is a Bulgarian rhythm [Aro04].
E(11, 12) =[××××××××××× · ] = (11111111112), when started on the second onset, is the
drum pattern of the Rahmāni (a cylindrical double-headed drum) used in the Sōt silām dance from
Mirbāt in the South of Oman [EMF90].
E(11, 24) =[× · · × · × · × · × · × · · × · × · × · × · × · ] = (32222322222) is a rhythm necklace
of the Aka Pygmies of Central Africa [Aro91]. It is usually started on the seventh onset. Started
on the second onset, it is a Bulgarian rhythm [Aro04].
E(13, 24) =[× · ×× · × · × · × · × · ×× · × · × · × · × · ] = (2122222122222) is another rhythm
necklace of the Aka Pygmies of the upper Sangha [Aro91]. Started on the penultimate onset, it is
the Bobangi metal-blade pattern used by the Aka Pygmies.
E(15, 34) = [× · · × · × · × · × · · × · × · × · × · · × · × · × · × · · × · × · ] = (322232223222322)
is a Bulgarian rhythmic pattern when started on the penultimate onset [Aro04].

2.5 Aksak Rhythms

Euclidean rhythms are closely related to a family of rhythms known as aksak rhythms, which
have been studied from the combinatorial point of view for some time [Bră51, Cle94, Aro04]. Béla
Bartók [Bar81] and Constantin Brăiloiu [Bră51], respectively, have used the terms Bulgarian rhythm
and aksak to refer to those meters that use units of durations 2 and 3, and no other durations.
Furthermore, the rhythm or meter must contain at least one duration of length 2 and at least
one duration of length 3. Arom [Aro04] refers to these durations as binary cells and ternary cells,
respectively.
Arom [Aro04] generated an inventory of all the theoretically possible aksak rhythms for values
of n ranging from 5 to 29, as well as a list of those that are actually used in traditional world
music. He also proposed a classification of these rhythms into several classes, based on structural
and numeric properties. Three of his classes are considered here:

1. An aksak rhythm is authentic if n is a prime number.

2. An aksak rhythm is quasi-aksak if n is an odd number that is not prime.

3. An aksak rhythm is pseudo-aksak if n is an even number.

A quick perusal of the Euclidean rhythms listed in the preceding reveals that aksak rhythms are
well represented. Indeed, all three of Arom’s classes (authentic, quasi-aksak, and pseudo-aksak)
make their appearance. There is a simple characterization of those Euclidean rhythms that are
aksak. ¿From the iterative subtraction algorithm of Bjorklund it follows that if n = 2k all cells are
binary (duration 2). Similarly, if n = 3k all cells are ternary (duration 3). Therefore, to ensure that
the Euclidean rhythm contains both binary and ternary cells, and no other durations, it follows
that n must be between 2k and 3k.
Of course, not all aksak rhythms are Euclidean. Consider the Bulgarian rhythm with interval
sequence (3322) [Aro04], which is also the metric pattern of Indian Lady by Don Ellis [Kei91]. Here
k = 4 and n = 10, and E(4, 10) = [× · · × · × · · × · ] or (3232), a periodic rhythm.

15
 The following Euclidean rhythms are authentic aksak :
E(2, 5) = [× · × · · ] = (23) (classical music, jazz, Greece, Macedonia, Namibia, Persia, Rwanda).
E(3, 7) = [× · × · × · · ] = (223) (Bulgaria, Greece, Sudan, Turkestan).
E(4, 11) = [× · · × · · × · · × · ] = (3332) (Southern India rhythm), (Serbian necklace).
E(5, 11) = [× · × · × · × · × · · ] = (22223) (classical music, Bulgaria, Northern India, Serbia).
E(5, 13) = [× · · × · × · · × · × · · ] = (32323) (Macedonia).
E(6, 13) = [× · × · × · × · × · × · · ] = (222223) (Macedonia).
E(7, 17) = [× · · × · × · · × · × · · × · × · ] = (3232322) (Macedonian necklace).
E(8, 17) = [× · × · × · × · × · × · × · × · · ] = (22222223) (Bulgaria).
E(8, 19) = [× · · × · × · × · · × · × · × · · × · ] = (32232232) (Bulgaria).
E(9, 23) = [× · · × · × · · × · × · · × · × · · × · × · · ] = (323232323) (Bulgaria).
The following Euclidean rhythms are quasi-aksak :
E(4, 9) = [× · × · × · × · · ] = (2223) (Greece, Macedonia, Turkey, Zaı̈re).
E(7, 15) = [× · × · × · × · × · × · × · · ] = (2222223) (Bulgarian necklace).
The following Euclidean rhythms are pseudo-aksak :
E(3, 8) = [× · · × · · × · ] = (332) (Central Africa, Greece, India, Latin America, West Africa, Su-
dan).
E(5, 12) = [× · · × · × · · × · × · ] = (32322) (Macedonia, South Africa).
E(7, 16) = [× · · × · × · × · · × · × · × · ] = (3223222) (Brazilian, Macedonian, West African neck-
laces).
E(7, 18) = [× · · × · × · · × · × · · × · × · · ] = (3232323) (Bulgaria).
E(9, 22) = [× · · × · × · · × · × · · × · × · · × · × · ] = (323232322) (Bulgarian necklace).
E(11, 24) = [× · · × · × · × · × · × · · × · × · × · × · × · ] = (32222322222) (Central African and Bul-
garian necklaces).
E(15, 34) = [× · · × · × · × · × · · × · × · × · × · · × · × · × · × · · × · × · ] = (322232223222322) (Bul-
garian necklace).

2.6 Drawing Digital Straight Lines

Euclidean rhythms and necklace patterns also appear in the computer graphics literature on drawing
digital straight lines [KR04]. The problem here consists of efficiently converting a mathematical
straight line segment defined by the x and y integer coordinates of its endpoints, to an ordered
sequence of pixels that most faithfully represents the given straight line segment. Figure 5 illustrates
an example of a digital straight line (shaded pixels) determined by the two given endpoints p and
q. All the pixels intersected by the segment (p, q) are shaded. If we follow either the lower or upper
boundary of the shaded pixels from left to right we obtain the interval sequences (43333) or (33334),
respectively. Note that the upper pattern corresponds to E(5, 16), a Bossa-Nova variant. Indeed,
Harris and Reingold [HR04] show that the well-known Bresenham algorithm [Bre65] is described
by the Euclidean algorithm.

2.7 Calculating Leap Years in Calendar Design

For thousands of years human beings have observed and measured the time it takes between two
consecutive sunrises, and between two consecutive spring seasons. These measurements inspired
different cultures to design calendars [Asc02, RD01]. Let Ty denote the duration of one revolution
of the earth around the sun, more commonly known as a year. Let Td denote the duration of one

16
 5 q
4
3
2
1
0
p 0 4 8 12 16

Figure 5: The shaded pixels form a digital straight line determined by the points p and q.

complete rotation of the earth, more commonly known as a day. The values of Ty and Td are of
course continually changing, because the universe is continually reconfiguring itself. However the
ratio Ty /Td is approximately 365.242199..... It is very convenient therefore to make a year last 365
days. The problem that arises both for history and for predictions of the future, is that after a
while the 0.242199..... starts to contribute to a large error. One simple solution is to add one extra
day every 4 years: the so-called Julian calendar. A day with one extra day is called a leap year.
But this assumes that a year is 365.25 days long, which is still slightly greater than 365.242199......
So now we have an error in the opposite direction albeit smaller. One solution to this problem is
the Gregorian calendar [Sha94]. The Gregorian calendar defines a leap year as one divisible by 4,
except not those divisible by 100, except not those divisible by 400. With this rule a year becomes
365 + 1/4 - 1/100 + 1/400 = 365.2425 days long, not a bad approximation.
Another solution is provided by the Jewish calendar which uses the idea of cycles [Asc02]. Here
a regular year has 12 months and a leap year has 13 months. The cycle has 19 years including 7
leap years. The 7 leap years must be distributed as evenly as possible in the cycle of 19. The cycle
is assumed to start with Creation as year 1. If the year modulo 19 is one of 3, 6, 8, 11, 14, 17, or
19, then it is a leap year. For example, the year 5765 = 303 · 19 + 8 and so is a leap year. The year
5766, which begins at sundown on the Gregorian date of October 3, 2005, is 5766 = 303x19 + 9,
and is therefore not a leap year. Applying Bjorklund’s algorithm to the integers 7 and 19 yields
E(7, 19) = [× · · × · × · · × · · × · × · · × · · ]. If we start this rhythm at the 7th pulse we obtain
the pattern [ · · × · · × · × · · × · · × · · × · ×], which describes precisely the leap year pattern 3,
6, 8, 11, 14, 17, and 19 of the Jewish calendar. In this sense the Jewish calendar is an instance of
a Euclidean necklace.

2.8 Euclidean Strings

In the study of the combinatorics of words and sequences, there exists a family of strings called
Euclidean strings [ERSS03]. In this section we explore the relationship between Euclidean strings
and Euclidean rhythms. We use the same terminology and notation introduced in [ERSS03].
Let P = (p0 , p1 , ..., pn−1 ) denote a string of non-negative integers. Let ρ(P ) denote the right
rotation of P by one position; that is, ρ(P ) = (pn−1 , p0 , p1 , ..., pn−2 ). Let ρd (P ) denote the right
rotation of P by d positions. If P is considered as a cyclic string, a right rotation corresponds to a
clockwise rotation. Figure 6 illustrates the ρ(P ) operator with P equal to the Bembé bell-pattern
of West Africa [Tou03]. Figure 6(a) shows the Bembé bell-pattern, Figure 6(b) shows ρ(P ), which
is a hand-clapping pattern from West Africa [Pre83], and Figure 6(c) shows ρ7 (P ), which is the
Tambú rhythm of Curaçao [Ros02].
Ellis et al. [ERSS03] define a string P = (p0 , p1 , ..., pn−1 ) to be Euclidean if incrementing p0

17
 0 0 0
11 1 11 1 11 1

10 2 10 2 10 2

9 3 9 3 9 3

8 4 8 4 8 4
7 5 7 5 7 5
6 6 6

(a) (b) (c)

Figure 6: Two right-rotations of the Bembé string: (a) the Bembé, (b) rotation by one unit, (c) rotation
by seven units.

by 1 and decrementing pn−1 by 1 yields a new string τ (P ) that is the rotation of P . In other
words, P and τ (P ) are instances of the same necklace. Therefore, if we represent rhythms as
binary sequences, Euclidean rhythms cannot be Euclidean strings because all Euclidean rhythms
begin with a ‘one’. Increasing p0 by one makes it a ‘two’, which is not a binary string. Therefore,
to explore the relationship between Euclidean strings and Euclidean rhythms, we will represent
rhythms by their clockwise distance sequences, which are also strings of nonnegative integers. As
an example, consider E(4, 9) = [× · × · × · × · · ] = (2223). Now τ (2223) = (3222), which is a
rotation of E(4, 9), and thus (2223) is a Euclidean string. Indeed, for P = E(4, 9), τ (P ) = ρ3 (P ).
As a second example, consider the West African clapping-pattern shown in Figure 6(b) given by
P = (1221222). We have that τ (P ) = (2221221) = ρ6 (P ), the pattern shown in Figure 6(c), which
also happens to be the mirror image of P about the (0, 6) axis. Therefore P is a Euclidean string.
However, note that P is not a Euclidean rhythm. Nevertheless, P is a rotation of the Euclidean
rhythm E(7, 12) = (2122122).
Ellis et al. [ERSS03] have many beautiful results about Euclidean strings. They show that
Euclidean strings exist if, and only if, n and (p0 + p1 + ... + pn−1 ) are relatively prime numbers, and
that when they exist they are unique. They also show how to construct Euclidean strings using an
algorithm that has the same structure as the Euclidean algorithm. In addition they relate Euclidean
strings to many other families of sequences studied in the combinatorics of words [AS02, Lot02].
Let R(P ) denote the reversal (or mirror image) of P ; that is, R(P ) = (pn−1 , pn−2 , ..., p1 , p0 ).
Now we may determine which of the Euclidean rhythms used in world music listed in the preceding,
are Euclidean strings or reverse Euclidean strings. The length of a Euclidean string is defined as the
number of integers it has. This translates in the rhythm domain to the number of onsets a rhythm
contains. Furthermore, strings of length one are Euclidean strings, trivially. Therefore all the trivial
Euclidean rhythms with only one onset, such as E(1, 2) = [× · ] = (2), E(1, 3) = [× · · ] = (3), and
E(1, 4) = [× · · · ] = (4), etc., are both Euclidean strings as well as reverse Euclidean strings. In
the lists that follow the Euclidean rhythms are shown in their box-notation format as well as in
the clockwise distance sequence representation. The styles of music that use these rhythms is also
included. Finally, if only a rotated version of the Euclidean rhythm is played, then it is still included
in the list but referred to as a necklace.
The following Euclidean rhythms are Euclidean strings:
E(2, 3) = [×× · ] = (12) (West Africa, Latin America, Nubia, Northern Canada).
E(2, 5) = [× · × · · ] = (23) (classical music, jazz, Greece, Macedonia, Namibia, Persia, Rwanda),
(authentic aksak ).

18
E(3, 4) = [××× · ] = (112) (Brazil, Bali rhythms), (Colombia, Greece, Spain, Persia, Trinidad
necklaces).
E(3, 7) = [× · × · × · · ] = (223) (Bulgaria, Greece, Sudan, Turkestan), (authentic aksak ).
E(4, 5) = [×××× · ] = (1112) (Greece).
E(4, 9) = [× · × · × · × · · ] = (2223) (Greece, Macedonia, Turkey, Zaı̈re), (quasi-aksak ).
E(5, 6) = [××××× · ] = (11112) (Arab).
E(5, 11) = [× · × · × · × · × · · ] = (22223) (classical music, Bulgaria, Northern India, Serbia), (au-
thentic aksak ).
E(5, 16) = [× · · × · · × · · × · · × · · · · ] = (33334) (Brazilian, West African necklaces).
E(6, 7) = [×××××× · ] = (111112) (Greek necklace)
E(6, 13) = [× · × · × · × · × · × · · ] = (222223) (Macedonia), (authentic aksak ).
E(7, 8) = [××××××× · ] = (1111112) (Libyan necklace).
E(7, 15) = [× · × · × · × · × · × · × · · ] = (2222223) (Bulgarian necklace), (quasi-aksak ).
E(8, 17) = [× · × · × · × · × · × · × · × · · ] = (22222223) (Bulgaria), (authentic aksak ).
The following Euclidean rhythms are reverse Euclidean strings:
E(3, 5) = [× · × · ×] = (221) (Korean, Rumanian, Persian necklaces).
E(3, 8) = [× · · × · · × · ] = (332) (Central Africa, Greece, India, Latin America, West Africa, Su-
dan), (pseudo-aksak ).
E(3, 11) = [× · · · × · · · × · · ] = (443) (North India).
E(3, 14) = [× · · · · × · · · · × · · · ] = (554) (North India).
E(4, 7) = [× · × · × · ×] = (2221) (Bulgaria).
E(4, 11) = [× · · × · · × · · × · ] = (3332) (Southern India rhythm), (Serbian necklace), (authentic
aksak ).
E(4, 15) = [× · · · × · · · × · · · × · · ] = (4443) (North India).
E(5, 7) = [× · ×× · ××] = (21211) (Arab).
E(5, 9) = [× · × · × · × · ×] = (22221) (Arab).
E(5, 12) = [× · · × · × · · × · × · ] = (32322) (Macedonia, South Africa), (pseudo-aksak ).
E(7, 9) = [× · ××× · ×××] = (2112111) (Greece).
E(7, 10) = [× · ×× · ×× · ××] = (2121211) (Turkey).
E(7, 16) = [× · · × · × · × · · × · × · × · ] = (3223222) (Brazilian, Macedonian, West African neck-
laces), (pseudo-aksak ).
E(7, 17) = [× · · × · × · · × · × · · × · × · ] = (3232322) (Macedonian necklace), (authentic aksak ).
E(9, 22) = [× · · × · × · · × · × · · × · × · · × · × · ] = (323232322) (Bulgarian necklace), (pseudo-
aksak ).
E(11, 12) = [× · ××××××××××] = (11111111112) (Oman necklace).
E(11, 24) = [× · · × · × · × · × · × · · × · × · × · × · × · ] = (32222322222) (Central African and Bul-
garian necklaces), (pseudo-aksak ).
The following Euclidean rhythms are neither Euclidean nor reverse Euclidean strings:
E(5, 8) = [× · ×× · ×× · ] = (21212) (Egypt, Korea, Latin America, West Africa).
E(5, 13) = [× · · × · × · · × · × · · ] = (32323) (Macedonia), (authentic aksak ).
E(7, 12) = [× · ×× · × · ×× · × · ] = (2122122) (West Africa), (Central African, Nigerian, Sierra
Leone necklaces).
E(7, 18) = [× · · × · × · · × · × · · × · × · · ] = (3232323) (Bulgaria), (pseudo-aksak ).
E(8, 19) = [× · · × · × · × · · × · × · × · · × · ] = (32232232) (Bulgaria), (authentic aksak ).
E(9, 14) = [× · ×× · ×× · ×× · ×× · ] = (212121212) (Algerian necklace).

19
E(9, 16) = [× · ×× · × · × · ×× · × · × · ] = (212221222) (West and Central African, and Brazilian
necklaces).
E(9, 23) = [× · · × · × · · × · × · · × · × · · × · × · · ] = (323232323) (Bulgaria), (authentic aksak ).
E(13, 24) = [× · ×× · × · × · × · × · ×× · × · × · × · × · ] = (2122222122222) (Central African neck-
lace).
E(15, 34) = [× · · × · × · × · × · · × · × · × · × · · × · × · × · × · · × · × · ] = (322232223222322)
(Bulgarian necklace), (pseudo-aksak ).
These three groups of Euclidean rhythms reveal a tantalizing pattern. The Euclidean rhythms
that are favored in classical music and jazz are also Euclidean strings (the first group). Furthermore,
this group is not popular in African music. The Euclidean rhythms that are neither Euclidean
strings nor reverse Euclidean strings (group three) fall into two categories: those consisting of
clockwise distances 1 and 2, and those consisting of clockwise distances 2 and 3. The latter group
is used only in Bulgaria, and the former is used in Africa. Finally, the Euclidean rhythms that
are reverse Euclidean strings (the second group) appear to have a much wider appeal. Finding
musicological explanations for the preferences apparent in these mathematical properties raises
interesting ethnomusicological questions.
The Euclidean strings defined in [ERSS03] determine another family of rhythms, many of which
are also used in world music but are not necessarily Euclidean rhythms. For example, (1221222)
is an Afro-Cuban bell pattern. Therefore it would be interesting to explore empirically the re-
lation between Euclidean strings and world music rhythms, and to determine formally the exact
mathematical relation between Euclidean rhythms and Euclidean strings.

3 Definitions and Notation

Before we begin the more technical part of the paper, we need to define some precise mathematical
notation for describing rhythms.
Let Z+ denote the set of positive integers. For k, n ∈ Z+ , let gcd(k, n) denote the greatest
common divisor of k and n. If gcd(k, n) = 1, we call k and n relatively prime. For integers a < b,
let [a, b] = {a, a + 1, a + 2, . . . , b}.
Let C be a circle in the plane, and consider any two points x, y on C. The chordal distance
between x and y, denoted by d(x, y), is the length of the line segment xy; that is, d(x, y) is the
ñ
Euclidean distance between x and y. The clockwise distance from x to y, or of the ordered pair
ò
(x, y), is the length of the clockwise arc of C from x to y, and is denoted by d(x, y). Finally, the
ò ñ ñ
geodesic distance between x and y, denoted by d(x, y), is the length of the shortest arc of C between
x and y; that is, d(x, y) = min{d(x, y), d(y, x)}.
A rhythm of timespan n is a subset of {0, 1, . . . , n − 1}, representing the set of pulses that are
onsets in each repetition. For clarity, we write the timespan n as a subscript after the subset: {. . . }n .
Geometrically, if we locate n equally spaced points clockwise around a circle Cn of circumference n,
then we can view a rhythm of timespan n as a subset of these n points. We consider an element of
Cn to simultaneously be a point on the circle and an integer in {0, 1, . . . , n − 1}.
The rotation of a rhythm R of timespan n by an integer ∆ ≥ 0 is the rhythm {(i + ∆) mod n :
i ∈ R}n of the same timespan n. The scaling of a rhythm R of timespan n by an integer α ≥ 1 is
the rhythm {αi : i ∈ R}αn of timespan αn.
Let R = {r0 , r1 , . . . , rk−1 }n be a rhythm of timespan n with k onsets sorted in clockwise order.
ñ
Throughout this paper, an onset ri will mean (ri mod k ) mod n. Observe that the clockwise distance
d(ri , rj ) = (rj − ri ) mod n. This is the number of points on Cn that are contained in the clockwise

20
arc (ri , rj ] and is also known as the chromatic length [CD91].
ò
The geodesic distance multiset of a rhythm R is the multiset of all nonzero pairwise geodesic
distances; that is, it is the multiset {d(ri , rj ) : ri , rj ∈ R, ri 6= rj }. The geodesic distance multiset


has cardinality k2 . The multiplicity of a distance d is the number of occurrences of d in the geodesic
distance multiset.
A rhythm is Erdős-deep if it has (exactly)
Pk−1 one
distance of multiplicity i, for each i ∈ [1, k − 1].
Note that these multiplicities sum to i=1 i = k2 , which is the cardinality of the geodesic distance
multiset, and hence these distances are all the distances in the rhythm. Every geodesic distance is
between 0 and ⌊n/2⌋. A rhythm is Winograd-deep if every two distances from {1, 2, . . . , ⌊ n2 ⌋} have
different multiplicity.
A shelling of an Erdős-deep rhythm R is an ordering s1 , s2 , . . . , sk of the onsets in R such that
R − {s1 , s2 , . . . , si } is an Erdős-deep rhythm for i = 0, 1, . . . , k. (Every rhythm with at most two
onsets is Erdős-deep.)
The evenness of rhythm R is the sum of all inter-onset chordal distances in R; that is,
X
d(ri , rj ).
0≤i<j≤k−1

ñ The clockwise distance sequence of R is the circular sequence

+
P (d0 , d1 , . . . , dk−1 ) where di =
d(r , r
i i+1 ) for all i ∈ [0, k − 1]. Observe that each di ∈ Z and i di = n.

Observation 1. There is a one-to-one relationship between rhythms P with k onsets and timespan n
+
and circular sequences (d0 , d1 , . . . , dk−1 ) where each di ∈ Z and i di = n.

4 Even Rhythms
In this section we first describe three algorithms that generate even rhythms. We then characterize
rhythms with maximum evenness and show that, for given numbers of pulses and onsets, the three
described algorithms generate the unique rhythm with maximum evenness. As mentioned in the
introduction, the measure of evenness considered here is the pairwise sum of chordal distances.
The even rhythms characterized in this section were studied by Clough and Meyerson [CM85,
CM86] for the case where the numbers of pulses and onsets are relatively prime. This was sub-
sequently expanded upon by Clough and Douthett [CD91]. We revisit these results and provide
an additional connection to rhythms (and scales) that are obtained from the Euclidean algorithm.
Most of these results are stated in [CD91]. However our proofs are new, and in many cases are
much more streamlined.

4.1 Characterization
We first present three algorithms for computing a rhythm with k onsets, timespan n, for any k ≤ n,
that possess large evenness.
The first algorithm is by Clough and Douthett [CD91]:

Algorithm Clough-Douthett(n, k)
 
1. return { in
k : i ∈ [0, k − 1]}

Because k ≤ n, the rhythm output by Clough-Douthett(n, k) has k onsets as desired.

The second algorithm is a geometric heuristic implicit in the work of Clough and Douthett [CD91]:

21
 Algorithm Snap(n, k)
1. Let D be a set of k evenly spaced points on Cn such that D ∩ Cn = ∅.
2. For each point x ∈ D, let x′ be the first point in Cn clockwise from x.
3. return {x′ : x ∈ D}

Because k ≤ n, the clockwise distance between consecutive points in D in the execution of

Snap(n, k) is at least that of consecutive points in Cn . Thus, x′ 6= y ′ for distinct x, y ∈ D, so
Snap returns a rhythm with k onsets as desired.
The third algorithm is a recursive algorithm in the same mold as Euclid’s algorithm for greatest
common divisors. The algorithm uses the clockwise distance sequence notation described in the
introduction. The resulting rhythm always defines the same necklace as the Euclidean rhythms
from Section 2.3; that is, the only difference is a possible rotation.

Algorithm Euclidean(n, k)
1. if k evenly divides n then return ( nk , nk , . . . , nk )
| {z }
k
2. a ← n mod k
3. (x1 , x2 , . . . , xa ) ← Euclidean(k, a)
4. return (⌊ nk ⌋, . . . , ⌊ nk ⌋, ⌈ nk ⌉; ⌊ nk ⌋, . . . , ⌊ nk ⌋, ⌈ nk ⌉; . . . ; ⌊ nk ⌋, . . . , ⌊ nk ⌋, ⌈ nk ⌉)
| {z } | {z } | {z }
x1 −1 x2 −1 xa −1

As a simple example, consider k = 5 and n = 13. The sequence of calls to Euclidean(n, k)

follows the same pattern as the Euclid algorithm for greatest common divisors from Section 2.1,
except that it now stops one step earlier: (13, 5), (5, 3), (3, 2), (2, 1). At the base of the recursion,
we have Euclidean(2, 1) = (2) = [× · ]. At the next level up, we obtain Euclidean(3, 2) =
(1, 2) = [×× · ]. Next we obtain Euclidean(5, 3) = (2; 1, 2) = [× · ×× · ]. Finally, we obtain Eu-
clidean(13, 5) = (2, 3; 3; 2, 3) = [× · × · · × · · × · × · · ]. (For comparison, the Euclidean rhythm
from Section 2.2 is E(5, 13) = (2, 3, 2, 3, 3), a rotation by 5.)
We now show that algorithm Euclidean(n, k) outputs a circular sequence of k integers that
sum to n (which is thus the clockwise distance sequence of a rhythm with k onsets and timespan n).
We proceed by induction on k. If k P evenly divides n, then the claim clearly holds. Otherwise a
(= n mod k) > 0, and by induction ai=1 xi = k. Thus the sequence that is output has k terms
and sums to
lnm a
jnk X lnm jnk
a + (xi − 1) = a + (k − a)
k k k k
i=1
 j n k jnk
=a 1+ + (k − a)
j nkk k
=a+k
k
=n .

The following theorem is one of the main contributions of this paper.

Theorem 4.1. Let n ≥ k ≥ 2 be integers. The following are equivalent for a rhythm R =
{r0 , r1 , . . . , rk−1 }n with k onsets and timespan n:
(A) R has maximum evenness (sum of pairwise inter-onset chordal distances),

22
 (B) R is a rotation of the Clough-Douthett(n, k) rhythm,

(C) R is a rotation of the Snap(n, k) rhythm,

(D) R is a rotation of the Euclidean(n, k) rhythm,

ñ
(⋆) for all ℓ ∈ [1, k] and i ∈ [0, k − 1], the ordered pair (ri , ri+ℓ ) has clockwise distance d(ri , ri+ℓ ) ∈
{⌊ ℓn ℓn
k ⌋, ⌈ k ⌉}.

Moreover, up to a rotation, there is a unique rhythm that satisfies these conditions.

Note that the evenness of a rhythm equals the evenness of the same rhythm played backwards.
Thus, if R is the unique rhythm with maximum evenness, then R is the same rhythm as R played
backwards (up to a rotation).
The proof of Theorem 4.1 proceeds as follows. In Section 4.2 we prove that each of the three
algorithms produces a rhythm that satisfies property (⋆). Then in Section 4.3 we prove that there is
a unique rhythm that satisfies property (⋆). Thus the three algorithms produce the same rhythm,
up to rotation. Finally in Section 4.4 we prove that the unique rhythm that satisfies property (⋆)
maximizes evenness.

4.2 Properties of the Algorithms

We now prove that each of the algorithms has property (⋆). Clough and Douthett [CD91] proved
the following.

Proof (B) ⇒ (⋆). Say R = {r0 , r1 , . . . , rk−1 }n is the Clough-Douthett(n, k) rhythm. Consider
ñ
an ordered pair (ri , ri+ℓ ) of onsets in R. Let pi = in mod k and let pℓ = ℓn mod k. By symmetry
we can suppose that ri ≤ r(i+ℓ) mod k . Then the clockwise distance d(ri , ri+ℓ ) is
               
(i + ℓ)n in in ℓn pi + pℓ in ℓn pi + pℓ
− = + + − = + ,
k k k k k k k k
     pi +pℓ 
which is ℓn k or ℓnk , because k ∈ {0, 1}.
 
A similar proof shows that the rhythm { in k : i ∈ [0, k − 1]} satisfies property (⋆). Observe
that (⋆) is equivalent to the following property.

(⋆⋆) if (d0 , d1 , . . . , dk−1 ) is the clockwise distance sequence of R, then for all ℓ ∈ [1, k], the sum of
any ℓ consecutive elements in (d0 , d1 , . . . , dk−1 ) equals ⌈ ℓn ℓn
k ⌉ or ⌊ k ⌋.

Proof (C) ⇒ (⋆⋆). Let (d0 , d1 , . . . , dk−1 ) be the clockwise distance sequence of the rhythm deter-
mined by Snap(n, k). For the sake of contradiction, suppose that for some ℓ ∈ [1, k], the sum of ℓ
consecutive elements in (d0 , d1 , . . . , dk−1 ) is greater than ⌈ ℓn k ⌉. The case in which the sum is less
than ⌊ ℓn
k ⌋ is analogous. We can assume that these ℓ consecutive elements are (d0 , d1 , . . . , dℓ−1 ).
ñ ñ ñ
Using the notation defined in the statement of the algorithm, let x0 , x1 , . . . , xℓ be the points in D
such that d(x′i , x′i+1 ) = di for all i ∈ [0, ℓ − 1]. Thus d(x′1 , x′ℓ+1 ) ≥ ⌈ ℓn ′
k ⌉ + 1. Now d(xℓ+1 , xℓ+1 ) < 1.
ñ
Thus d(x′1 , xℓ+1 ) > ⌈ ℓn ℓn ñ ℓn
k ⌉ ≥ k , which implies that d(x1 , xℓ+1 ) > k . This contradicts the fact that
the points in D were evenly spaced around Cn in the first step of the algorithm.

23
Proof (D) ⇒ (⋆⋆). We proceed by induction on k. Let R = Euclidean(n, k). If k evenly divides n,
then R = ( nk , nk , . . . , nk ), which satisfies (D). Otherwise, let a = n mod k and let (x1 , x2 , . . . , xa ) =
Euclidean(k, a). By induction, for all ℓ ∈ [1, a], the sum of any ℓ consecutive elements in
(x1 , x2 , . . . , xa ) equals ⌊ ℓk ℓk
a ⌋ or ⌈ a ⌉. Let S be a sequence of m consecutive elements in R. By
construction, for some 1 ≤ i ≤ j ≤ a, and for some 0 ≤ s ≤ xi − 1 and 0 ≤ t ≤ xj − 1, we have

S = (⌊ nk ⌋, . . . , ⌊ nk ⌋, ⌈ nk ⌉, ⌊ nk ⌋, . . . , ⌊ nk ⌋, ⌈ nk ⌉, . . . , ⌊ nk ⌋, . . . , ⌊ nk ⌋, ⌈ nk ⌉, ⌊ nk ⌋, . . . , ⌊ nk ⌋) .
| {z } | {z } | {z } | {z }
s xi+1 −1 xj−1 −1 t
  P  mn 
It remains to prove thatP mn
k ≤mn  S ≤ k . P
We first prove that S ≥ k . We can assume the worst case for S to be minimal, which
is when s = xi − 1 and t = xj − 1. Thus by induction,
j  
X (j − i + 1)k
m+1= xα ≤ .
a
α=i

Hence
   
am a (j − i + 1)k a a (j − i + 1)k + a − 1 a 1
≤ − ≤ − = j −i+1− .
k k a k k a k k

Thus ⌊ am
k ⌋ ≤ j − i and
X jnk j n k j am k j j n k am k j m  j n k k j mn k
S=m +j−i≥m + = m + = k +a = .
k k k k k k k k
P   P
Now we prove that S ≤ mn
k . We can assume the worst case for S to be maximal, which
is when s = 0 and t = 0. Thus by induction,
j−1  
X (j − i − 1)k
m−1= xα ≥ .
a
α=i+1

Hence
   
am a (j − i − 1)k a a (j − i − 1)k − a + 1 a 1
≥ + ≥ + = j −i−1+ .
k k a k k a k k

Thus ⌈ am
k ⌉ ≥ j − i and
X jnk j n k l am m l j n k am m l m  j n k m l mn m
S=m +j−i≤m + = m + = k +a = .
k k k k k k k k

4.3 Uniqueness
In this section we prove that there is a unique rhythm satisfying the conditions in Theorem 4.1. The
following well-known number-theoretic lemmas will be useful. Two integers x and y are inverses
modulo m if xy ≡ 1 (mod m).
Lemma 4.2 ([Sti03, page 55]). An integer x has an inverse modulo m if and only if x and m are
relatively prime. Moreover, if x has an inverse modulo m, then it has an inverse y ∈ [1, m − 1].

24
Lemma 4.3. If x and m are relatively prime, then ix 6≡ jx (mod m) for all distinct i, j ∈ [0, m−1].
Proof. Suppose that ix ≡ jx (mod m) for some i, j ∈ [0, m − 1]. By Lemma 4.2, x has an inverse
modulo m. Thus i ≡ j (mod m), and i = j because i, j ∈ [0, m − 1].

Lemma 4.4. For all relatively prime integers n and k with 2 ≤ k ≤ n, there is an integer ℓ ∈
[1, k − 1] such that:
(a) ℓn ≡ 1 (mod k),

(b) iℓ 6≡ jℓ (mod k) for all distinct i, j ∈ [0, k − 1], and

(c) i⌊ ℓn ℓn
k ⌋ 6≡ j⌊ k ⌋ (mod n) for all distinct i, j ∈ [0, k − 1].

Proof. By Lemma 4.2 with x = n and m = k, n has an inverse ℓ modulo k. This proves (a). Thus k
and ℓ are relative prime by Lemma 4.2 with x = ℓ and m = k. Hence (b) follows from Lemma 4.3.
Let t = ⌊ ℓn
k ⌋. Then ℓn = kt + 1. By Lemma 4.3 with m = n and x = t (and because k ≤ n), to
prove (c) it suffices to show that t and n are relatively prime. Let g = gcd(t, n). Thus ℓ ng = k gt + g1 .
Because ng and gt are integers, g1 is an integer and g = 1. This proves (c).

The following theorem is the main result of this section.

Theorem 4.5. For all integers n and k with 2 ≤ k ≤ n, there is a unique rhythm with k onsets
and timespan n that satisfies property (⋆), up to a rotation.
Proof. Let R = {r0 , r1 , . . . , rk−1 }n be a k-onset rhythm that satisfies (⋆). Recall that the index of
an onset is taken modulo k, and that the value of an onset is taken modulo n. That is, ri = x
means that ri mod k = x mod n.
Let g = gcd(n, k). We consider three cases for the value of g.
Case 1. g = k: Because R satisfies property (⋆) for ℓ = 1, every ordered pair (ri , ri+1 ) has
clockwise distance nk . By a rotation of R we can assume that r0 = 0. Thus ri = in k for all
i ∈ [0, k − 1]. Hence R is uniquely determined in this case.
Case 2. g = 1 (see Figure 7): By Lemma 4.4(a), there is an integer ℓ ∈ [1, k − 1] such that
ℓn ≡ 1 (mod k). Thus ℓn = (k − 1)⌊ ℓn ℓn
k ⌋ + ⌈ k ⌉. Hence, of the k ordered pairs (ri , ri+ℓ ) of onsets,
ℓn
k − 1 have clockwise distance ⌊ k ⌋ and one has clockwise distance ⌈ ℓn k ⌉. By a rotation of R we
ℓn ℓn
can assume that r0 = 0 and rk−ℓ = n − ⌈ k ⌉. Thus riℓ = i⌊ k ⌋ for all i ∈ [0, k − 1]; that is,
r(iℓ) mod k = (i⌊ ℓn
k ⌋) mod n. By Lemma 4.4(b) and (c), this defines the k distinct onsets of R.
Hence R is uniquely determined in this case.
Case 3. g ∈ [2, k − 1] (see Figure 8): Let k′ = kg and let n′ = ng . Observe that both k′ and
′ ′ ñ
n′ are integers. Because R satisfies (⋆) and ⌈ kkn ⌉ = ⌊ kkn ⌋ = n′ , we have d(ri , ri+k′ ) = n′ for all
i ∈ [0, k − 1]. Thus
rik′ +j = in′ + rj (1)
for all i ∈ [0, g − 1] and j ∈ [0, n′ − 1].
Now gcd(n′ , k′ ) = 1 by the maximality of g. By Lemma 4.4(a), there is an integer ℓ′ ∈ [1, k′ − 1]
′ ′ ′ ′ ′ ′ ′ ′
such that ℓ′ n′ ≡ 1 (mod k′ ). Thus ℓ′ n′ = (k′ −1)⌊ ℓkn′ ⌋+⌈ ℓkn′ ⌉, implying ℓ′ n = (k−g)⌊ ℓkn′ ⌋+g⌈ ℓkn′ ⌉.
′ ′
Hence, of the k ordered pairs (ri , ri+ℓ′ ) of onsets, k − g have clockwise distance ⌊ ℓkn′ ⌋ and g have
′ ′ ′ ′
clockwise distance ⌈ ℓkn′ ⌉. By a rotation of R we can assume that r0 = 0 and rℓ′ = ⌈ ℓkn′ ⌉. By
Equation (1) with j = 0 and j = ℓ′ , we have
l ′ ′m
rik′ = in′ and rik′ +ℓ′ = in′ + ℓkn′ (2)

25
 0
11 1

10 2

9 3

8 4

7 5
6

Figure 7: Here we illustrate Case 2 with n = 12 and k = 7. Thus ℓ = 3 because 3 × 12 ≡ 1 (mod 7). We
have ⌈ ℓn ℓn
k ⌉ = 6 and ⌊ k ⌋ = 5. By a rotation we can assume that r0 = 0 and rk−ℓ = r4 = 6 (the darker dots).
Then as shown by the arrows, the positions of the other onsets are implied.

′ ′
for all i ∈ [0, g − 1]. This accounts for the g ordered pairs (ri , ri+ℓ′ ) with clockwise distance ⌈ ℓkn′ ⌉.
′ ′
The other k − g ordered pairs (ri , ri+ℓ′ ) have clockwise distance ⌊ ℓkn′ ⌋. Define
l ′ ′m j ′ ′k
L0 = 0 and Lj = k′ + (j − 1) ℓkn′ for all j ∈ [1, k′ − 1] .
ℓn

Thus by Equation (2),

rik′ +jℓ′ = in′ + Lj
for all i ∈ [0, g − 1] and j ∈ [0, k′ − 1]; that is, r(ik′ +jℓ′ ) mod k = (in′ + Lj ) mod n.

0
14 1
13 2

12 3

11 4

10 5

9 6
8 7

Figure 8: Here we illustrate Case 3 with′ ′n = 15 and k′ = ′

9. Thus g = 3, n′ = 5 and k ′ = 3. We have
ℓ′ = 2 because 2 × 5 ≡ 1 (mod 3). Thus ⌈ ℓkn′ ⌉ = 4 and ⌊ ℓkn′ ⌋ = 3. We have L0 = 0, L1 = 4 and L2 = 7. A
rotation fixes the first g = 3 onsets (the darker or blue dots). As shown by the arrows, these onsets imply
the positions of the next three onsets (medium or green dots), which in turn imply the positions of the final
three onsets (the light or yellow dots).

To conclude that R is uniquely determined, we must show that over the range i ∈ [0, g − 1] and
j ∈ [0, k′ − 1], the numbers ik′ + jℓ′ are distinct modulo k, and the numbers in′ + Lj are distinct
modulo n.

26
 First we show that the numbers ik′ + jℓ′ are distinct modulo k. Suppose that

ik′ + jℓ′ ≡ pk′ + jℓ′ (mod k) (3)

for some i, p ∈ [0, g − 1] and j, q ∈ [0, k′ − 1]. Because k = k′ · g, we can write (ik′ + jℓ′ ) mod k as
a multiple of k′ plus a residue modulo k′ . In particular,
 ′

(ik′ + jℓ′ ) mod k = k′ (i + ⌊ jℓ ′ ′
k ′ ⌋) mod g + (jℓ mod k ) .

Thus Equation (3) implies that

 ′
  
qℓ′
k′ (i + ⌊ jℓ
k′ ⌋) mod g + (jℓ ′
mod k ′
) = k ′
(p + ⌊ k′ ⌋) mod g + (qℓ′ mod k′ ) . (4)

Hence jℓ′ ≡ qℓ′ (mod k′ ). Thus j = q by Lemma 4.4(c). By substituting j = q into Equation (4),
it follows that i ≡ p (mod g). Thus i = p because i, p ∈ [0, g − 1]. This proves that the numbers
ik′ + jℓ′ are distinct modulo k.
Now we show that the numbers in′ + Lj are distinct modulo n. The proof is similar to the
above proof that the numbers ik′ + jℓ′ are distinct modulo k.
Suppose that
in′ + Lj ≡ pn′ + Lq (mod n) (5)
for some i, p ∈ [0, g − 1] and j, q ∈ [0, k′ − 1]. Because n = n′ · g, we can write (in′ + Lj ) mod n as
a multiple of n′ plus a residue modulo n′ . In particular,
 j k 
L
(in′ + Lj ) mod n = n′ (i + nj′ ) mod g + (Lj mod n′ ) .

Thus Equation (5) implies that

 j k   j k 
L L
n′ (i + nj′ ) mod g + (Lj mod n′ ) = n′ (p + nq′ ) mod g + (Lq mod n′ ) . (6)

Hence Lj ≡ Lq (mod n′ ). We claim that j = q. If j = 0 then Lj = 0, implying Lq = 0 and q = 0.

′ ′ ′ ′
Now assume that j, q ≥ 1. In this case, Lj = j⌊ ℓkn′ ⌋ + 1 and Lq = q⌊ ℓkn′ ⌋ + 1. Thus
j ′ ′k j ′ ′k
j ℓkn′ ≡ q ℓkn′ (mod k′ ) .

Hence j = q by Lemma 4.4(c). By substituting j = q into Equation (6), it follows that i ≡ p

(mod g). Thus i = p because i, p ∈ [0, g − 1]. This proves that the numbers in′ + Lj are distinct
modulo n.
Therefore R is uniquely determined.

We have shown that each of the three algorithms generates a rhythm with property (⋆), and
that there is a unique rhythm with property (⋆). Thus all of the algorithms produce the same
rhythm, up to rotation. It remains to prove that this rhythm has maximum evenness.

4.4 Rhythms with Maximum Evenness

We start with a technical lemma. Let v, w be points at geodesic distance d on a circle C. Obviously
d(v, w) is a function of d, independent of v and w. Let f (C, d) = d(v, w).

27
Lemma 4.6. For all geodesic lengths x ≤ d on a circle C, we have f (C, x)+f (C, d−x) ≤ 2·f (C, d2 ),
with equality only if d = 2x.

Proof. We can assume that C is a unit circle. Consider the isosceles triangle formed by the center
of C and a geodesic of length d (≤ π). We have 12 f (C, d) = sin d2 . Thus f (C, d) = 2 sin d2 . Thus our
claim is equivalent to sin x + sin(d − x) ≤ 2 sin d2 for all x ≤ d (≤ π/2). In the range 0 ≤ x ≤ d, sin x
is increasing, and sin(d − x) is decreasing at the opposite rate. Thus sin x + sin(d − x) is maximized
when x = d − x. That is, when d = 2x. The result follows.

For a rhythm R = {r0 , r1 , . . . , rk−1 }n , for each ℓ ∈ [1, k], let S(R, ℓ) be theP sum of chordal
k−1
distances taken over all ordered pairs (ri , ri+ℓ ) in R. That is, let S(R, ℓ) = i=0 d(ri , ri+ℓ ).
Pk
Property (A) says that R maximizes ℓ=1 S(R, ℓ). Before we characterize rhythms that maximize
ñ
the sum of S(R, ℓ), we first concentrate on rhythms that maximize S(R, ℓ) for each particular
value of ℓ. Let D(R, ℓ) be the multiset of clockwise distancesP{d(ri , ri+ℓ ) : i ∈ [0, k − 1]}. Then
S(R, ℓ) is determined by D(R, ℓ). In particular, S(R, ℓ) = {f (Cn , d) : d ∈ D(R, ℓ)} (where
{f (Cn , d) : d ∈ D(R, ℓ)} is a multiset).

ñ ñ
Lemma 4.7. Let 1 ≤ ℓ ≤ k ≤ n be integers. A k-onset rhythm R = {r0 , r1 , . . . , rk−1 }n maximizes
S(R, ℓ) if and only if |d(ri , ri+ℓ ) − d(rj , rj+ℓ )| ≤ 1 for all i, j ∈ [0, k − 1].
ñ
Proof. Suppose that R = {r0 , r1 , . . . , rk−1 }n maximizes S(R, ℓ). Let di = d(ri , ri+ℓ ) for all i ∈
[0, k − 1]. Suppose on the contrary that dp ≥ dq + 2 for some p, q ∈ [0, k − 1]. We can assume
that q < p, dp = dq + 2, and di = dq + 1 for all i ∈ [q + 1, p − 1]. Define ri′ = ri + 1 for all
i ∈ [q + 1, p], and define ri′ = ri for all other i. Let R′ be the rhythm {r0′ , r1′ , . . . , rk−1
′ }n . Thus
′ ′ 1
D(R, ℓ)\D(R , ℓ) = {dp , dq } and D(R , ℓ)\D(R, ℓ) = {dp −1, dq +1}. Now dp −1 = dq +1 = 2 (dp +dq ).
By Lemma 4.6, f (Cn , dp ) + f (Cn , dq ) < 2 · f (Cn , 21 (dp + dq ). Thus S(R, ℓ) < S(R′ , ℓ), which
contradicts the maximality of S(R, ℓ).
ñ ñ
For the converse, let R be a rhythm such that |d(ri , ri+ℓ ) − d(rj , rj+ℓ )| ≤ 1 for all i, j ∈ [0, k − 1].
Suppose on the contrary that R does not maximize S(R, ℓ). Thus some P rhythm T = P(t0 , t1 , . . . , tk−1 )
ñ ñ
maximizes S(T, ℓ) and T 6= R. Hence D(T, ℓ) 6= D(R, ℓ). Because D(R, ℓ) = D(T, ℓ) (= ℓn),
we have d(ti , ti+ℓ ) − d(tj , tj+ℓ ) ≥ 2 for some i, j ∈ [0, k − 1]. As we have already proved, this implies
that T does not maximize S(T, ℓ). This contradiction proves that R maximizes S(R, ℓ).
Pk−1 ñ
Because i=0 d(ri , ri+ℓ ) = ℓn for any rhythm with k onsets and timespan n, Lemma 4.7 can
be restated as follows.

ñ
Corollary 4.8. Let 1 ≤ ℓ ≤ k ≤ n be integers. A k-onset rhythm R = {r0 , r1 , . . . , rk−1 }n maximizes
S(R, ℓ) if and only if d(ri , ri+ℓ ) ∈ {⌈ ℓn ℓn
k ⌉, ⌊ k ⌋} for all i ∈ [0, k − 1].

Proof (⋆) ⇒ (A). If (⋆) holds

P for some rhythm R, then by Corollary 4.8, R maximizes S(R, ℓ) for
every ℓ. Thus R maximizes ℓ S(R, ℓ).

Proof (A) ⇒ (⋆). By Theorem 4.5, there is a unique rhythm R that satisfies property (⋆). Let R
denote the unique rhythm that satisfies property (⋆). Suppose on the contrary that there is a rhythm
ñ
T = (t0 , t1 , . . . , tk−1 ) with property (A) but R 6= T . Thus there exists an ordered pair (ti , ti+ℓ ) in
T with clockwise distance d(ti , ti+ℓ ) 6∈ {⌊ ℓn ⌋, ⌈ ℓn
Pk Pk k k ⌉}. By Corollary 4.8, S(T, ℓ) < S(R, ℓ). Because
T has property (A), ℓ=1 S(T, ℓ) ≥ ℓ=1 S(R, ℓ). Thus for some ℓ′ we have S(T, ℓ′ ) > S(R, ℓ′ ).
But this is a contradiction, because S(R, ℓ′ ) ≥ S(T, ℓ′ ) by Corollary 4.8.

28
 This completes the proof of Theorem 4.1. Note that Theorem 4.1 holds for metrics other than
the pairwise sum of all chordal distances. The only property that is needed is Lemma 4.6. For
example, the metric “pairwise sum of the squares of all geodesic distances” satisfies Lemma 4.6,
and thus Theorem 4.1 holds in this setting.

5 Deep Rhythms
Winograd [Win66], and independently Clough et al. [CEK99], characterize all Winograd-deep scales:
up to rotation, they are the scales that can be generated by the first ⌊n/2⌋ or ⌊n/2⌋ + 1 multiples
(modulo n) of a value that is relatively prime to n, plus one exceptional scale {0, 1, 2, 4}6 . In
this section, we prove a similar (but more general) characterization of Erdős-deep rhythms: up to
rotation and scaling, they are the rhythms generable as the first k multiples (modulo n) of a value
that is relatively prime to n, plus the same exceptional rhythm {0, 1, 2, 4}6 . The key difference is
that the number of onsets k is now a free parameter, instead of being forced to be either ⌊n/2⌋ or
⌊n/2⌋ + 1. Our proof follows Winograd’s, but differs in one case (the second case of Theorem 5.3).
We later prove that every Erdős-deep rhythm has a shelling and that maximally even rhythms
with n and k relatively prime are Erdős-deep.

5.1 Characterization of Deep Rhythms

Our characterization of Erdős-deep rhythms is in terms of two families of rhythms. The main
rhythm family consists of the generated rhythms Dk,n,m = {im mod n : i = 0, 1, . . . , k − 1}n of
timespan n, for certain values of k, n, and m. The one exceptional rhythm is F = {0, 1, 2, 4}6 of
timespan 6.

Fact 5.1. F is Erdős-deep.

Lemma 5.2. If k ≤ ⌊n/2⌋ + 1 and m and n are relatively prime, then Dk,n,m is Erdős-deep.

Proof. The multiset of clockwise distances in Dk,n,m is {(jm − im) mod n : i < j} = {(j − i)m mod
n : i < j}. There are k −p choices of i and j such that j −i = p, so there are exactly p occurrences of
the clockwise distance (pm) mod n in the multiset. Each of these clockwise distances corresponds
to a geodesic distance—either (pm) mod n or (−pm) mod n, whichever is smaller (at most n/2).
We claim that these geodesic distances are all distinct. Then the multiplicity of each geodesic
distance (±pm) mod n is exactly p, establishing that the rhythm is Erdős-deep.
For two geodesic distances to be equal, we must have ±pm ≡ ±qm (mod n) for some (possibly
different) choices for the ± symbols, and for some p 6= q. By (possibly) multiplying both sides
by −1, we obtain two cases: (1) pm ≡ qm (mod n) and (2) pm ≡ −qm (mod n). Because m is
relatively prime to n, by Lemma 4.2, m has a multiplicative inverse modulo n. Multiplying both
sides of the congruence by this inverse, we obtain (1) p ≡ q (mod n) and (2) p ≡ −q (mod n).
Because 0 ≤ i < j < k ≤ ⌊n/2⌋ + 1, we have 0 ≤ p = j − i < ⌊n/2⌋ + 1, and similarly for q:
0 ≤ p, q ≤ ⌊n/2⌋. Thus, the first case of p ≡ q (mod n) can happen only when p = q, and the
second case of p + q ≡ 0 (mod n) can happen only when p = q = 0 or when p = q = n/2. Either
case contradicts that p 6= q. Therefore the geodesic distances arising from different values of p are
indeed distinct, proving the lemma.

We now state and prove our characterization of Erdős-deep rhythms, which is up to rotation
and scaling. Rotation preserves the geodesic distance multiset and therefore Erdős-deepness (and

29
Winograd-deepness). Scaling maps each geodesic distance d to αd, and thus preserves multiplicities
and therefore Erdős-deepness (but not Winograd-deepness).

Theorem 5.3. A rhythm is Erdős-deep if and only if it is a rotation of a scaling of either the
rhythm F or the rhythm Dk,n,m for some k, n, m with k ≤ ⌊n/2⌋ + 1, 1 ≤ m ≤ ⌊n/2⌋, and m and
n are relatively prime.

Proof. Because a rotation of a scaling of an Erdős-deep rhythm is Erdős-deep, the “if” direction of
the theorem follows from Fact 5.1 and Lemma 5.2.
Consider an Erdős-deep rhythm R with k onsets. By the definition of Erdős-deepness, R
has one nonzero geodesic distance with multiplicity i for each i = 1, 2, . . . , k − 1. Let m be the
geodesic distance with multiplicity k − 1. Because m is a geodesic distance, 1 ≤ m ≤ ⌊n/2⌋. Also,
k ≤ ⌊n/2⌋ + 1 (for any Erdős-deep rhythm R), because all nonzero geodesic distances are between
1 and ⌊n/2⌋ and therefore at most ⌊n/2⌋ nonzero geodesic distances occur. Thus k and m are
suitable parameter choices for Dk,n,m.
Consider the graph Gm = (R, Em ) with vertices corresponding to onsets in R and with an edge
between two onsets of geodesic distance m. By the definition of geodesic distance, every vertex i in
Gm has degree at most 2: the only onsets at geodesic distance exactly m from i are (i − m) mod n
and (i + m) mod n. Thus, the graph Gm is a disjoint union of paths and cycles. The number of
edges in Gm is the multiplicity of m, which we supposed was k − 1, which is 1 less than the number
of vertices in Gm . Thus, the graph Gm consists of exactly one path and any number of cycles.
The cycles of Gm have a special structure because they correspond to subgroups generated
by single elements in the cyclic group (Z/(n), +). Namely, the onsets corresponding to vertices
of a cycle in Gm form a regular (n/a)-gon, with a geodesic distance of a = gcd(m, n) between
consecutive onsets. (a is called the index of the subgroup generated by m.) In particular, every
cycle in Gm has the same length r = n/a. Because Gm is a simple graph, every cycle must have at
least 3 vertices, so r ≥ 3.
The proof partitions into four cases depending on the length of the path and on how many
cycles the graph Gm has. The first two cases will turn out to be impossible; the third case will lead
to a rotation of a scaling of rhythm F ; and the fourth case will lead to a rotation of a scaling of
rhythm Dk,n,m .
First suppose that the graph Gm consists of a path of length at least 1 and at least one cycle.
We show that this case is impossible because the rhythm R can have no geodesic distance with
multiplicity 1. Suppose that there is a geodesic distance with multiplicity 1, say between onsets i1
and i2 . If i is a vertex of a cycle, then both (i + m) mod n and (i − m) mod n are onsets in R. If
i is a vertex of the path, then one or two of these are onsets in R, with the case of one occurring
only at the endpoints of the path. If (i1 + m) mod n and (i2 + m) mod n were both onsets in R, or
(i1 −m) mod n and (i2 −m) mod n were both onsets in R, then we would have another occurrence of
the geodesic distance between i1 and i2 , contradicting that this geodesic distance has multiplicity 1.
Thus, i1 and i2 must be opposite endpoints of the path. If the path has length ℓ, then the clockwise
distance between i1 and i2 is (ℓm) mod n. This clockwise distance (and hence the corresponding
geodesic distance) appears in every cycle, of which there is at least one, so the geodesic distance
has multiplicity more than 1, a contradiction. Therefore this case is impossible.
Second suppose that the graph Gm consists of a path of length 0 and at least two cycles. We
show that this case is impossible because the rhythm R has two geodesic distances with the same
multiplicity. Pick any two cycles C and C ′ , and let d be the smallest positive clockwise distance
from a vertex of C to a vertex of C ′ . Thus i is a vertex of C if and only if (i + d) mod n is a vertex
of C ′ . Because the cycles are disjoint, d < a. Because r ≥ 3, d < n/3, so clockwise distances of

30
d are also geodesic distances of d. The number of occurrences of geodesic distance d between a
vertex of C and a vertex of C ′ is either r or 2r, the case of 2r arising when d = a/2 (that is, C ′ is a
“half-rotation” of C). The number of occurrences of geodesic distance d′ = min{d + m, n − (d + m)}
is the same—either r or 2r, in the same cases. (Note that d < a ≤ n − m, so d + m < n, so
the definition of d′ correctly captures a geodesic distance modulo n.) The same is true of geodesic
distance d′′ = min{d − m, n − (d − m)}. If other pairs of cycles have the same smallest positive
clockwise distance d, then the number of occurrences of d, d′ , and d′′ between those cycles are also
equal. Because the cycles are disjoint, geodesic distance d and thus d + m and d − m cannot be
(pm) mod n for any p, so these geodesic distances cannot occur between two vertices of the same
cycle. Finally, the sole vertex x of the path has geodesic distance d to onset i (which must be a
vertex of some cycle) if and only if x has geodesic distance d′ to onset (i + m) mod n (which must
be a vertex of the same cycle) if and only if x has geodesic distance d′′ to onset (i − m) mod n
(which also must be a vertex of the same cycle). Therefore the multiplicities of geodesic distances
d, d′ , and d′′ must be equal. Because R is Erdős-deep, we must have d = d′ = d′′ . To have d = d′ ,
either d = d + m or d = n − (d + m), but the first case is impossible because d > 0 by nonoverlap of
cycles, so 2d + m = n. Similarly, to have d = d′′ , we must have 2d − m = n. Subtracting these two
equations, we obtain that 2m = 0, contradicting that m > 0. Therefore this case is also impossible.
Third suppose that the graph Gm consists of a path of length 0 and exactly one cycle. We show
that this case forces R to be a rotation of a scaling of rhythm F because otherwise two geodesic
distances m and m′ have the same multiplicity. The number of occurrences of geodesic distance m
in the cycle is precisely the length r of the cycle. Similarly, the number of occurrences of geodesic
distance m′ = min{2m, n − 2m} in the cycle is r. The sole vertex x on the path cannot have
geodesic distance m or m′ to any other onset (a vertex of the cycle) because then x would then be
on the cycle. Therefore the multiplicities of geodesic distances m and m′ must be equal. Because
R is Erdős-deep, m must equal m′ , which implies that either m = 2m or m = n − 2m. The first
case is impossible because m > 0. In the second case, 3m = n, that is, m = 31 n. Therefore, the
cycle has r = 3 vertices, say at ∆, ∆ + 31 n, ∆ + 32 n. The fourth and final onset x must be midway
between two of these three onsets, because otherwise its geodesic distance to the three vertices are
all distinct and therefore unique. No matter where x is so placed, the rhythm R is a rotation by
∆ + c 31 n (for some c ∈ {0, 1, 2}) of a scaling by n/6 of the rhythm F .
Finally suppose that Gm has no cycles, and consists solely of a path. We show that this case
forces R to be a rotation of a scaling of a rhythm Dk,n′ ,m′ with 1 ≤ m′ ≤ ⌊n′ /2⌋ and with m′ and n′
relatively prime. Let b be the onset such that (b−m) mod n is not an onset (the “beginning” vertex
of the path). Consider rotating R by −i so that 0 is an onset in the resulting rhythm R − i. The
vertices of the path in R − i form a subset of the subgroup of the cyclic group (Z/(n), +) generated
by the element m. Therefore the rhythm R − i = Dk,n,m = {(im) mod n : i = 0, 1, . . . , k − 1}n is
a scaling by a of the rhythm Dk,n/a,m/a = {(im/a) mod (n/a) : i = 0, 1, . . . , k − 1}n . The rhythm
Dk,n/a,m/a has an appropriate value for the third argument: m/a and n/a are relatively prime
(a = gcd(m, n)) and 1 ≤ m/a ≤ ⌊n/2⌋/a ≤ ⌊(n/a)/2⌋. Also, k ≤ ⌊(n/a)/2⌋ + 1 because the only
occurring geodesic distances are multiples of a and therefore the number k − 1 of distinct geodesic
distances is at most ⌊(n/a)/2⌋. Therefore R is a rotation by i of a scaling by a of Dk,n/a,m/a with
appropriate values of the arguments.

Corollary 5.4. A rhythm is Erdős-deep if and only if it is a rotation of a scaling of the rhythm F or
it is a rotation of a rhythm Dk,n,m for some k, n, m satisfying k ≤ ⌊n/2g⌋ + 1 where g = gcd(m, n).

Proof. First we show that any Erdős-deep rhythm has one of the two forms in the corollary. By
Theorem 5.3, there are two flavors of Erdős-deep rhythms, and the corollary directly handles

31
rotations of scalings of F . Thus it suffices to consider a rhythm R that is a rotation by ∆ of a scaling
by α of Dk,n,m where k ≤ ⌊n/2⌋+1, 1 ≤ m ≤ ⌊n/2⌋, and m and n are relatively prime. Equivalently,
R is a rotation by ∆ of Dk,n′ ,m′ where n′ = αn and m′ = αm. Now g = gcd(n′ , m′ ) = α, so n′ /g = n.
Hence, k ≤ ⌊n′ /2g⌋ + 1 as desired. Thus we have rewritten R in the desired form.
It remains to show that every rhythm in one of the two forms in the corollary is Erdős-deep.
Again, rotations of scalings of F are handled directly by Theorem 5.3. So consider a rotation
of Dk,n,m where k ≤ ⌊n/2g⌋ + 1. The value of m matters only modulo n, so we assume that
0 ≤ m ≤ n − 1.
First we show that, if ⌊n/2⌋ + 1 ≤ m ≤ n − 1, then Dk,n,m can be rewritten as a rotation of
the rhythm Dk,n,m′ where m′ = n − m ≤ ⌊n/2⌋. By reversing the order in which we list the onsets
in Dk,n,m = {im mod n : i = 0, 1, . . . , k − 1}n , we can write Dk,n,m = {(k − 1 − i) m mod n : i =
0, 1, . . . , k − 1}n . Now consider rotating the rhythm Dk,n,n−m = {i (n − m) mod n : i = 0, 1, . . . , k −
1}n by (k − 1)m. We obtain the rhythm {[i (n − m) + (k − 1) m] mod n : i = 0, 1, . . . , k − 1}n =
{[(k−1−i) m+in] mod n : i = 0, 1, . . . , k−1}n = {(k−1−i) m mod n : i = 0, 1, . . . , k−1}n = Dk,n,m
as desired.
Thus it suffices to consider rotations of Dk,n,m where 1 ≤ m ≤ ⌊n/2⌋ and k ≤ ⌊n/2g⌋. The
rhythm Dk,n′ ,m′ , where n′ = n/g and m′ = m/g, is Erdős-deep by Theorem 5.3 because n′ and m′
are relatively prime, k ≤ ⌊n′ /2⌋ + 1, and 1 ≤ m′ ≤ ⌊n′ /2⌋. But Dk,n,m is the scaling of Dk,n′ ,m′ by
the integer g, so Dk,n,m is also Erdős-deep.

An interesting consequence of this characterization is the following:

Corollary 5.5. Every Erdős-deep rhythm has a shelling.
Proof. If the Erdős-deep rhythm is Dk,n,m , we can remove the last onset from the path, resulting
in Dk−1,n,m, and repeat until we obtain the empty rhythm D0,n,m . At all times, k remains at most
⌊n/2⌋+1 (assuming it was originally) and m remains between 1 and ⌊n/2⌋ and relatively prime to n.
On the other hand, F = {0, 1, 2, 4}6 has the shelling 4, 2, 1, 0 because {0, 1, 2}6 is Erdős-deep.

5.2 Connection Between Deep and Even Rhythms

A connection between maximally even scales and Winograd-deep scales is shown by Clough et
al. [CEK99]. They define a diatonic scale to be a maximally even scale with k = (n + 2)/2 and n
a multiple of 4. They show that diatonic scales are Winograd-deep. We now prove a similar result
for Erdős-deep rhythms.
Lemma 5.6. A rhythm R of maximum evenness satsifying k ≤ ⌊n/2⌋ + 1 is Erdős-deep if and only
if k and n are relatively prime.
Proof. Recall that by property (⋆) one of the unique characterizations of an even rhythm of max-
ñ
imum evenness can be stated as follows. For all 1 ≤ ℓ ≤ k, and for every ordered pair (ri , ri+ℓ ) of
onsets in R, the clockwise distance d(ri , ri+ℓ ) ∈ {⌊ ℓn ℓn
k ⌋, ⌈ k ⌉}.
For the case in which k and n are relatively prime, by Lemma 4.2, there exists a value ℓ < k
such that ℓn ≡ 1 (mod k). Thus we can write ℓn = k⌊ℓn/k⌋ + 1. Let m = ⌊ℓn/k⌋. Now consider
the set {im mod n : i = 0, 1, . . . , k − 1}n . By Lemma 4.4(c), we get k distinct values, so R can be
ñ ñ
realized as Dk,n,m = {im mod n : i = 0, 1, . . . , k − 1}n . Thus, by Lemma 5.2, R is Erdős-deep.
Observe that F = {0, 1, 2, 4}6 does not maximize evenness because d(0, 2) = 2 and d(2, 0) = 4
yet ℓ = 2. Hence, any rhythm that maximizes evenness and that is deep must also be generated.
Now consider the case in which n and k are not relatively prime. We show that the assumption
that R is deep leads to a contradiction. Thus, assuming that R is deep implies that there is a value

32
m such that R can be realized as Dk,n,m = {im mod n : i = 0, 1, . . . , k − 1}n . This in turn implies
that there exists an integer ℓ such that km = ℓn + 1, that is, ℓn ≡ 1 (mod k). However, for this to
happen, n and k must be relatively prime, a contradiction.
Thus we have shown that R is Erdős-deep if and only if k and n are relatively prime.

Acknowledgments
This work was initiated at the 20th Bellairs Winter Workshop on Computational Geometry held
January 28–February 4, 2005 at Bellairs Research Institute in Barbados. We thank the other
participants of that workshop—Greg Aloupis, David Bremner, Justin Colannino, Mirela Damian,
Vida Dujmović, Jeff Erickson, Ferran Hurtado, John Iacono, Danny Krizanc, Stefan Langerman,
Erin McLeish, Pat Morin, Mark Overmars, Suneeta Ramaswami, Diane Souvaine, Ileana Streinu,
Remco Veltcamp, and Sue Whitesides—for helpful discussions and contributing to a fun and creative
atmosphere.
Thanks to Simha Arom for providing us with a copy of his paper on the classification of aksak
rhythms [Aro04], which provided several examples of Euclidean rhythms of which we were not aware.
Thanks to Jeff Erickson for bringing to our attention the wonderful paper by Mitchell Harris and
Ed Reingold [HR04] on digital line drawing, leap year calculations, and Euclid’s algorithm. Finally,
thanks to Marcia Ascher for pointing out that the leap-year pattern of the Jewish calendar is a
Euclidean rhythmic necklace.

References
[Aga86] V. Kofi Agawu. Gi Dunu, Nyekpadudo, and the study of West African rhythm.
Ethnomusicology, 30(1):64–83, Winter 1986.
[Ank97] Willie Anku. Principles of rhythm integration in African music. Black Music Research
Journal, 17(2):211–238, Autumn 1997.
[Ape60] Willi Apel. Vier plus vier = drei plus drei plus zwei. Acta Musicologica, 32(Fasc.
1):29–33, January-March 1960.
[Aro91] Simha Arom. African Polyphony and Polyrhythm. Cambridge University Press, Cam-
bridge, England, 1991.
[Aro04] Simha Arom. L’aksak: Principes et typologie. Cahiers de Musiques Traditionnelles,
17:12–48, 2004.
[AS02] Jean-Paul Allouche and Jeffrey O. Shallit. Automatic Sequences. Cambridge University
Press, Cambridge, England, 2002.
[Asc75] Michael I. Asch. Social context and the musical analysis of Slavey drum dance songs.
Ethnomusicology, 19(2):245–257, May 1975.
[Asc02] Marcia Ascher. Mathematics Elsewhere: An Exploration of Ideas Across Cultures.
Princeton University Press, Princeton and Oxford, 2002.
[Ash03] Anthony Ashton. Harmonograph–A Visual Guide to the Mathematics of Music. Walker
and Company, New York, 2003.
[Bar81] Béla Bartók. Ce qu’on appelle le rythme bulgare. In Musique de la vie, pages 142–155,
Stock, Paris, 1981.
[Bar04a] James M. Barbour. Tuning and Temperament: A Historical Survey. Dover, New York,
2004.

33
 [Bar04b] Gregory Barz. Music in East Africa. Oxford University Press, Oxford, England, 2004.
[BD94] Steven Block and Jack Douthett. Vector products and intervallic weighting. Journal
of Music Theory, 38:21–41, 1994.
[Beh73] Gerard Behague. Bossa and bossas: recent changes in Brazilian urban popular music.
Ethnomusicology, 17(2):209–233, 1973.
[Bek05] Tolga Bektaş. Relationships between prosodic and musical meters in the Beste form
of classical Turkish music. Asian Music, 36(1):1–26, Winter/Spring 2005.
[Bjo03a] Eric Bjorklund. A metric for measuring the evenness of timing system rep-rate pat-
terns. SNS ASD Technical Note SNS-NOTE-CNTRL-100, Los Alamos National Lab-
oratory, Los Alamos, U.S.A., 2003.
[Bjo03b] Eric Bjorklund. The theory of rep-rate pattern generation in the SNS timing system.
SNS ASD Technical Note SNS-NOTE-CNTRL-99, Los Alamos National Laboratory,
Los Alamos, U.S.A., 2003.
[Bră51] Constantin Brăiloiu. Le rythme aksak. Revue de Musicologie, 33:71–108, 1951.
[Bra59] Rose Brandel. The African hemiola style. Ethnomusicology, 3(3):106–117, September
1959.
[Bre65] Jack E. Bresenham. Algorithm for computer control of digital plotter. IBM Systems
Journal, 4:25–30, 1965.
[Bre99] Roy Brewer. The use of Habanera rhythm in rockabilly music. American Music,
17:300–317, Autumn 1999.
[Bru64] Viggo Brun. Euclidean algorithms and musical theory. Enseignement Mathématique,
10:125–137, 1964.
[Car98] Norman Carey. Distribution modulo 1 and musical scales. PhD thesis, University of
Rochester, Rochester, New York, 1998.
[CD91] John Clough and Jack Douthett. Maximally even sets. Journal of Music Theory,
35:93–173, 1991.
[CEK99] John Clough, Nora Engebretsen, and Jonathan Kochavi. Scales, sets, and interval
cycles: a taxonomy. Music Theory Spectrum, 21(1):74–104, Spring 1999.
[Che79] John Miller Chernoff. African Rhythm and African Sensibility. The University of
Chicago Press, Chicago, 1979.
[Che02] Marc Chemillier. Ethnomusicology, ethnomathematics. The logic underlying orally
transmitted artistic practices. In Gerard Assayag, Hans Georg Feichtinger, and
Jose Francisco Rodrigues, editors, Mathematics and Music, pages 161–183. Springer-
Verlag, 2002.
[Cla00] Martin Clayton. Time in Indian Music. Oxford University Press, New York, 2000.
[Cle94] Jérôme Cler. Pour une théorie de l’aksak. Revue de Musicologie, 80:181–210, 1994.
[CLRS01] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. In-
troduction to Algorithms. The MIT Press, Cambridge, Massachussetts, 2001.
[CM85] John Clough and Gerald Myerson. Variety and multiplicity in diatonic systems. Jour-
nal of Music Theory, 29:249–270, 1985.
[CM86] John Clough and Gerald Myerson. Musical scales and the generalized circle of fifths.
American Mathematical Monthly, 93(9):695–701, 1986.

34
 [Cox62] H. S. M. Coxeter. Music and mathematics. The Canadian Music Journal, VI:13–24,
1962.
[CT03] Marc Chemillier and Charlotte Truchet. Computation of words satisfying the “rhyth-
mic oddity property” (after Simha Arom’s works). Information Processing Letters,
86:255–261, 2003.
[DBFG+ 04] Miguel Dı́az-Bañez, Giovanna Farigu, Francisco Gómez, David Rappaport, and God-
fried T. Toussaint. El compás flamenco: a phylogenetic analysis. In Proceedings of
BRIDGES: Mathematical Connections in Art, Music and Science, Southwestern Col-
lege, Winfield, Kansas, July 30 - August 1 2004.
[EMF90] Issam El-Mallah and Kai Fikentscher. Some observations on the naming of musical
instruments and on the rhythm in Oman. Yearbook for Traditional Music, 22:123–126,
1990.
[Erd89] Paul Erdős. Distances with specified multiplicities. American Mathematical Monthly,
96:447, 1989.
[ERSS03] John Ellis, Frank Ruskey, Joe Sawada, and Jamie Simpson. Euclidean strings. Theo-
retical Computer Science, 301:321–340, 2003.
[Euc56] Euclid. Elements. Dover, 1956. Translated by Sir Thomas L. Heath.
[Eva66] Bob Evans. Authentic Conga Rhythms. Belwin Mills Publishing Corporation, Miami,
1966.
[Far92] Mary Farquharson. Africa in America. Discos Corazon, Mexico, 1992. [CD].
[Flo99] Samuel A. Floyd, Jr. Black music in the circum-Caribbean. American Music, 17(1):1–
38, 1999.
[For73] Allen Forte. The Structure of Atonal Music. Yale Univ. Press, New Haven, Connecti-
cut, U.S.A., 1973.
[Fra56] Philip Franklin. The Euclidean algorithm. The American Mathematical Monthly,
63(9):663–664, November 1956.
[Gam67a] Carlton Gamer. Deep scales and difference sets in equal-tempered systems. In Proceed-
ings of the 2nd Annual Conference of the American Society of University Composers,
pages 113–122, 1967.
[Gam67b] Carlton Gamer. Some combinational resources of equal-tempered systems. Journal of
Music Theory, 11:32–59, 1967.
[Gam02] José Manuel Gamboa. Cante por Cante: Discolibro Didactico de Flamenco. New
Atlantis Music, Alia Discos, Madrid, 2002.
[GMRT05] Francisco Gómez, Andrew Melvin, David Rappaport, and Godfried T. Toussaint.
Mathematical measures of syncopation. In Proc. BRIDGES: Mathematical Connec-
tions in Art, Music and Science, pages 73–84, Banff, Alberta, Canada, July 31 -
August 3 2005.
[HAD95] Royal Hartigan, Abraham Adzenyah, and Freeman Donkor. West African Rhythms
for Drum Set. Manhattan Music, 1995.
[Hag03] Kobi Hagoel. The Art of Middle Eastern Rhythm. OR-TAV Music Publications, Kfar
Sava, Israel, 2003.
[Has97] Christopher F. Hasty. Meter as Rhythm. Oxford University Press, Oxford, England,
1997.

35
 [HK81] Lee Hye-Ku. Quintuple meter in Korean instrumental music. Asian Music, 13(1):119–
129, 1981.
[HR04] Mitchell A. Harris and Edward M. Reingold. Line drawing, leap years, and Euclid.
ACM Computing Surveys, 36(1):68–80, March 2004.
[Joh03] Timothy A. Johnson. Foundations of Diatonic Theory: A Mathematically Based Ap-
proach to Music Fundamentals. Key College Publishing, Emeryville, California, 2003.
[Kau80] Robert Kauffman. African rhythm: A reassessment. Ethnomusicology, 24(3):393–415,
September 1980.
[Kei91] Michael Keith. From Polychords to Pólya: Adventures in Musical Combinatorics.
Vinculum Press, Princeton, 1991.
[Klő97] Tőm Klőwer. The Joy of Drumming: Drums and Percussion Instruments from Around
the World. Binkey Kok Publications, Diever, Holland, 1997.
[Knu98] Donald E. Knuth. The Art of Computer Programming, volume 2, 3rd edition. Addison
Wesley, Reading, Massachussets, 1998.
[Koe70] James Koetting. Analysis and notation of West African drum ensemble music. Pub-
lications of the Institute of Ethnomusicology, 1(3), 1970.
[KR04] Reinhard Klette and Azriel Rosenfeld. Digital straightness - a review. Discrete Applied
Mathematics, 139:197–230, 2004.
[Lon04] Justin London. Hearing in Time: Psychological Aspects of Musical Meter. Oxford
University Press, New York, 2004.
[Lot02] M. Lothaire. Algebraic Combinatorics on Words. Cambridge University Press, Cam-
bridge, England, 2002.
[LP92] W. F. Lunnon and Peter A. B. Pleasants. Characterization of two-distance sequences.
Journal of the Australian Mathematical Society (Series A), 53:198–218, 1992.
[Man85] Peter Manuel. The anticipated bass in Cuban popular music. Latin American Music
Review, 6(2):249–261, Autumn-Winter 1985.
[Mat85] Thomas J. Mathiesen. Rhythm and meter in ancient Greek music. Music Theory
Spectrum, 7:159–180, Spring 1985.
[McC98] Brian J. McCartin. Prelude to musical geometry. The College Mathematics Journal,
29(5):354–370, 1998.
[Mon85] Matthew Montfort. Ancient Traditions–Future Possibilities: Rhythmic Training
Through the Traditions of Africa, Bali and India. Panoramic Press, Mill Valley, 1985.
[Mor96] Craig Morrison. Go Cat Go: Rockabilly Music and Its Makers. University of Illinois
Press, Urbana, 1996.
[Mus02] Putumayo World Music. Congo to Cuba. Music CD, 2002.
[Ort95] Fernando Ortiz. La Clave. Editorial Letras Cubanas, La Habana, Cuba, 1995.
[Pal89] Ilona Palásti. A distance problem of P. Erdős with some further restrictions. Discrete
Mathematics, 76(2):155–156, 1989.
[PC69] Vera Proca-Ciortea. On rhythm in Rumanian folk dance. Yearbook of the International
Folk Music Council, 1:176–199, 1969.
[Pre83] Jeff Pressing. Cognitive isomorphisms between pitch and rhythm in world musics:
West Africa, the Balkans and Western tonality. Studies in Music, 17:38–61, 1983.

36
 [Rah87] Jay Rahn. Asymmetrical ostinatos in sub-saharan music: time, pitch, and cycles
reconsidered. In Theory Only: Journal of the Michigan Music Theory Society, 9(7):23–
37, 1987.
[Rah96] Jay Rahn. Turning the analysis around: African-derived rhythms and Europe-derived
music theory. Black Music Research Journal, 16(1):71–89, 1996.
[Rap05] David Rappaport. Geometry and harmony. In Proceedings of BRIDGES: Mathematical
Connections in Art, Music and Science, pages 67–72, Banff, Alberta, Canada, 2005.
[RD01] Edward M. Reingold and Nachum Dershowitz. Calendrical Calculations: The Mille-
nium Edition. Cambridge University Press, Cambridge, England, 2001.
[Ric80] Timothy Rice. Aspects of Bulgarian musical thought. Yearbook of the International
Folk Music Council, 12:43–66, 1980.
[Ric04] Timothy Rice. Music in Bulgaria. Oxford University Press, Oxford, England, 2004.
[Rod97] Olavo Alén Rodrı́guez. Instrumentos de la Musica Folklórico-Popular de Cuba. Centro
de Investigación y Desarrollo de la Musica Cubana, Havana, Cuba, 1997.
[Ros02] Rene V. Rosalia. Migrated Rhythm: The Tambú of Curaçao. CaribSeek, 2002.
[Sac53] Curt Sachs. Rhythm and Tempo: A Study in Music History. W. W. Norton, New
York, 1953.
[Sha94] Jeffrey O. Shallit. Pierce expansions and rules for the determination of leap years.
Fibonacci Quarterly, 32(5):416–423, 1994.
[Sin74] Alice Singer. The metrical structure of Macedonian dance. Ethnomusicology,
18(3):379–404, September 1974.
[Sol96] Doug Sole. The Soul of Hand Drumming. Mel Bay Productions Inc., Toronto, 1996.
[Sta88] James A. Standifer. The Tuareg: their music and dances. The Black Perspective in
Music, 16(1):45–62, Spring 1988.
[Sti03] John Stillwell. Elements of Number Theory. Springer, 2003.
[Sto05] Ruth M. Stone. Music in West Africa. Oxford University Press, Oxford, England,
2005.
[Tou02] Godfried T. Toussaint. A mathematical analysis of African, Brazilian, and Cuban
clave rhythms. In Proceedings of BRIDGES: Mathematical Connections in Art, Music
and Science, pages 157–168, Towson University, Towson, Maryland, U.S.A., July 27-29
2002.
[Tou03] Godfried T. Toussaint. Classification and phylogenetic analysis of African ternary
rhythm timelines. In Proceedings of BRIDGES: Mathematical Connections in Art, Mu-
sic and Science, pages 25–36, Granada, Spain, July 23-27 2003. University of Granada.
[Tou04a] Godfried T. Toussaint. Computational geometric aspects of musical rhythm. In Ab-
stracts of the 14th Annual Fall Workshop on Computational Geometry, pages 47–48,
Cambridge, Massachusetts, November 2004.
[Tou04b] Godfried T. Toussaint. A mathematical measure of preference in African rhythm. In
Abstracts of Papers Presented to the American Mathematical Society, volume 25, page
248, Phoenix, Arizona, U.S.A., January 7-10 2004. American Mathematical Society.
[Tou05] Godfried T. Toussaint. Mathematical features for recognizing preference in Sub-
Saharan African traditional rhythm timelines. In Proceedings of the 3rd International
Conference on Advances in Pattern Recognition, pages 18–27, University of Bath,
United Kingdom, August 2005.

37
[Tym06] Dmitri Tymoczko. The geometry of musical chords. Manuscript, 2006.
[Link] .
[Uri93] Ed Uribe. The Essence of Brazilian Persussion and Drum Set. CCP/Belwin Inc.,
Miami, Florida, 1993.
[Uri96] Ed Uribe. The Essence of Afro-Cuban Persussion and Drum Set. Warner Brothers
Publications, Miami, Florida, 1996.
[vdL95] Pedro van der Lee. Zarabanda: esquemas rı́tmicos de acompañamiento en 6/8. Latin
American Music Review, 16(2):199–220, Autumn-Winter 1995.
[Wad04] Bonnie C. Wade. Thinking Musically. Oxford University Press, Oxford, England,
2004.
[Win66] Terry Winograd. An analysis of the properties of ‘deep’ scales in a t-tone system.
Unpublished, May 17 1966. Term paper for music theory course at Colorado College.
[Woo93] Marc Wooldridge. Rhythmic Implications of Diatonic Theory: A Study of Scott Joplin’s
Ragtime Piano Works. PhD thesis, State University of New York, Buffalo, 1993.
[Wri78] Owen Wright. The Modal System of Arab and Persian Music AD 1250-1300. Oxford
University Press, Oxford, England, 1978.

38