THE GEOMETRY OF MUSICAL CHORDS

Dmitri Tymoczko, Princeton University

Musical chords have a non-Euclidean geometry that has been exploited by
Western composers in many different styles.

A musical chord can be represented as a point in a geometrical space called an
orbifold. Line segments represent mappings from the notes of one chord to those
of another. Composers in a wide range of styles have exploited the non-Euclidean
geometry of these spaces, typically by utilizing short line segments between
structurally similar chords. Such line segments exist only when chords are nearly
symmetrical under translation, reflection, or permutation. Paradigmatically
consonant and dissonant chords possess different near-symmetries, and suggest
different musical uses.

Western music lies at the intersection of two seemingly independent disciplines:
harmony and counterpoint. Harmony delimits the acceptable chords (simultaneously
occurring notes) and chord sequences. Counterpoint (or voice leading) is the technique of
connecting the individual notes in a series of chords so as to form simultaneous melodies.
Chords are usually connected so that these lines (or voices) move independently (not all in
the same direction by the same amount), efficiently (by short distances), and without voice
crossings (along non-intersecting paths) (Fig. 1A-C). These features facilitate musical
performance, engage explicit aesthetic norms (1-2), and enable listeners to distinguish
multiple simultaneous melodies (3).
How is it that Western music can satisfy harmonic and contrapuntal constraints at
once? What determines whether two chords can be connected by efficient voice leading?
Musicians have been investigating these questions for almost three centuries. The circle of
fifths (Fig. S1), first published in 1728 (4), depicts efficient voice leadings among the
twelve major scales. The Tonnetz (Fig. S2), originating with Euler in 1739, represents
efficient voice leadings among the twenty-four major and minor triads (2, 5). Recent work
(513) investigates efficient voice leading in a variety of special cases. Despite tantalizing
hints (610), however, no theory has articulated general principles explaining when and
why efficient voice leading is possible. This report provides such a theory, describing
geometrical spaces in which points represent chords and line segments represent voice
2

leadings between their endpoints. These spaces show us precisely how harmony and
counterpoint are related.
Human pitch perception is both logarithmic and periodic: frequencies f and 2f are
heard to be separated by a single distance (the octave) and to possess the same quality or
chroma. To model the logarithmic aspect of pitch perception, I associate a pitchs
fundamental frequency f with a real number p according to the equation

p = 69 + 12log
2
(f/440). (1)

The result is a linear space (pitch space) in which octaves have size 12, semitones (the
distance between adjacent keys on a piano) have size 1, and middle C is assigned the
number 60. Distance in this space reflects physical distance on keyboard instruments,
orthographical distance in Western musical notation, and musical distance as measured in
psychological experiments (1415).
Musically, the chroma of a note is often more important than its octave. It is
therefore useful to identify all pitches p and p + 12. The result is a circular quotient space
(pitch-class space) that mathematicians call R/12Z (Fig. S3). (For a glossary of terms and
symbols, see Table S1.) Points in this space (pitch classes) provide numerical alternatives
to the familiar letter-names of Western music theory: C = 0, Cs/Df = 1, D = 2, D quarter
tone sharp = 2.5, etc. Western music typically uses only a discrete lattice of points in this
space. Here I consider the more general, continuous case. This is because the symmetrical
chords that influence voice-leading behavior need not lie on the discrete lattice.
The content of a collection of notes is often more important than their order.
Chords can therefore be modeled as multisets of either pitches or pitch classes. (Chord
will henceforth refer to a multiset of pitch classes unless otherwise noted.) The musical
term transposition is synonymous with the mathematical term translation and is
represented by addition in pitch or pitch-class space. Transpositionally related chords are
the same up to translation: thus the C major chord, {C, E, G} or {0, 4, 7}, is
transpositionally related to the F major chord, {F, A, C} or {5, 9, 0}, since {5, 9, 0}
{0 + 5, 4 + 5, 7 + 5} modulo 12Z. The musical term inversion is synonymous with the
3

mathematical term reflection, and corresponds to subtraction from a constant value.
Inversionally related chords are the same up to reflection: thus, the C major chord is
inversionally related to the C minor chord {C, Ef, G}, or {0, 3, 7}, since {0, 3, 7}
{7 7, 7 4, 7 0} modulo 12Z. Musically, transposition and inversion are significant
because they preserve the character of a chord: transpositionally related chords sound
extremely similar, inversionally related chords fairly so (Movie S1).
A voice leading between two multisets {x
1
, x
2
, ..., x
m
} and {y
1
, y
2
, ..., y
n
} is a
multiset of ordered pairs (x
i
, y
j
), such that every member of each multiset is in some pair. A
trivial voice leading contains only pairs of the form (x, x). The notation (x
1
, x
2
, ..., x
n
)
(y
1
, y
2
, ..., y
n
) identifies the voice leading that associates the corresponding items in each
list.

Thus, the voice leading (C, C, E, G)(B, D, F, G) associates C with B, C with D, E
with F, and G with G. Music theorists have proposed numerous ways of measuring voice-
leading size. Rather than adopting one, I will require only that a measure satisfy a few
constraints reflecting widely acknowledged features of Western music (16). These
constraints make it possible to identify, in polynomial time, a minimal voice leading (not
necessarily bijective) between arbitrary chords (16). Every music-theoretical measure of
voice-leading size satisfies these constraints.
I now describe the geometry of musical chords. An ordered sequence of n pitches
can be represented as a point in R
n
(Fig. S4). Directed line segments in this space represent
voice leadings. A measure of voice-leading size assigns lengths to these line segments. I
will use quotient spaces to model the way listeners abstract from octave and order
information. To model an ordered sequence of n pitch classes, form the quotient space
(R/12Z)
n
, also known as the n-torus T
n
. To model unordered n-note chords of pitch
classes, identify all points (x
1
, x
2
, ... x
n
) and (x
(1)
, x
(2)
, ... x
(n)
), where is any
permutation. The result is the global-quotient orbifold T
n
/S
n
(1718), the n-torus T
n

modulo the symmetric group S
n
. It contains singularities at which the local topology is not
that of R
n
.
Figure 2 shows the orbifold T
2
/S
2
, the space of unordered pairs of pitch classes. It
is a Mbius strip, a square whose left edge is given a half twist and identified with its right.
The orbifold is singular at its top and bottom edges, which act like mirrors (18). Any
4

bijective voice leading between pairs of pitches or pairs of pitch classes can be associated
with a path on Fig. 2 (Movie S2). Measures of voice-leading size determine these paths
lengths. They are the images of line segments in the parent spaces T
n
and R
n
, and are
either line segments in the orbifold, or reflected line segments that bounce off its singular
edges. For example, the voice leading (C, Df)(Df, C) reflects off the orbifolds upper
mirror boundary (Fig. 2).
Generalizing to higher dimensions is straightforward. To construct the orbifold
T
n
/S
n
, take an n-dimensional prism whose base is an (n1)-simplex, twist the base so as to
cyclically permute its vertices, and identify it with the opposite face (Figs. S5, S6) (16).
The boundaries of the orbifold are singular, acting as mirrors and containing chords with
duplicate pitch classes. Chords that divide the octave evenly lie at the center of the
orbifold, and are surrounded by the familiar sonorities of Western tonality. Voice leadings
parallel to the height coordinate of the prism act as transpositions. A free computer
program written by the author allows readers to explore these spaces (19).
In many Western styles, it is desirable to find efficient, independent voice leadings
between transpositionally or inversionally related chords. The progressions in Fig. 1 are all
of this type (Movie S3). A chord can participate in such progressions only if it is nearly
symmetrical under transposition, permutation, or inversion (16). I conclude by describing
these symmetries, explaining how they are embodied in the orbifolds geometry, and
showing how they have been exploited by Western composers.
A chord is transpositionally symmetrical (T-symmetrical) if it either divides the
octave into equal parts or is the union of equally sized subsets that do so (20). Nearly T-
symmetrical chords are close to these T-symmetrical chords. Both types of chord can be
linked to at least some of their transpositions by efficient bijective voice leadings. As one
moves toward the center of the orbifold, chords become increasingly T-symmetrical, and
can be linked to their transpositions by increasingly efficient bijective voice leading. The
perfectly even chord at the center of the orbifold can be linked to all of its transpositions by
the smallest possible bijective voice leading; a related result covers discrete pitch-class
spaces (16). Efficient voice leadings between perfectly T-symmetrical chords are typically
5

not independent. Thus composers have reason to prefer near T-symmetry to exact T-
symmetry.
It follows that the acoustically consonant chords of traditional Western music can
be connected by efficient voice leading. Acoustic consonance is incompletely understood;
however, theorists have long agreed that chords approximating a few consecutive elements
of the harmonic series are particularly consonant, at least when played with harmonic tones
(21). Since elements n to 2n of the harmonic series evenly divide an octave in frequency
space, they divide the octave nearly evenly in log-frequency space. These chords are
therefore clustered near the center of the orbifolds (Table 1), and can typically be linked by
efficient, independent voice leadings. Traditional tonal music exploits this possibility (Fig.
1A-C, Movie S4). This central feature of Western counterpoint is made possible by
composers interest in the harmonic property of acoustic consonance.
A chord with duplicate pitch classes is permutationally symmetrical (P-
symmetrical), since there is some nontrivial permutation of its notes that is a trivial voice
leading. These chords lie on the singular boundaries of the orbifolds. Nearly P-
symmetrical chords, such as {E, F, Gf}, are near these chords, and contain several notes
that are clustered close together. Efficient voice leadings permuting the clustered notes
bounce off the nearby boundaries (Fig. 2. Movies S2 and S4). Such voice leadings can be
efficient, independent, and non-trivial. Since trivial voice leadings are musically inert,
composers have reason to prefer near P-symmetry to exact P-symmetry.
Nearly P-symmetrical chords such as {B, C, Df} are considered to be extremely
dissonant. They are well-suited to static music in which voices move by small distances
within an unchanging harmony (Fig. 1D). Such practices are characteristic of recent atonal
composition, particularly the music of Ligeti and Lutoslawski. From the present
perspective, these avant-garde techniques are closely related to those of traditional tonality:
they exploit one of three fundamental symmetries permitting efficient, independent voice
leading between transpositionally or inversionally related chords.
A chord is inversionally symmetrical (I-symmetrical) if it is invariant under
reflection in pitch class space. Nearly I-symmetrical chords are near these chords, and can
be found throughout the orbifolds (16). For example, the Fs half-diminished seventh chord
6

{6, 9, 0, 4} and the F dominant seventh chord {5, 9, 0, 3} are related by inversion, and are
very close to the I-symmetrical chord {5.5, 9, 0, 3.5}. Consequently, we can find an
efficient voice leading between them (6, 9, 0, 4)(5, 9, 0, 3) (Fig. 1C) (16). Nearly T-
symmetrical chords, such as the C major triad, and nearly P-symmetrical chords, such as
{C, Df, Ef}, can also be nearly I-symmetrical. Consequently, I-symmetry is exploited in
both tonal and atonal music. It plays a salient role in the nineteenth-century, particularly in
the music of Schubert (22), Wagner (23), and Debussy (Fig. 1C).
The preceding ideas can be extended in several directions. First, one might
examine in detail how composers have exploited the geometry of musical chords (Movie
S4). Second, one could generalize the geometrical approach by considering quotient
spaces that identify transpositionally and inversionally related chords (24). Third, since
cyclical rhythmic patterns can also be modeled as points on T
n
/S
n
, one could use these
spaces to study African and other non-Western rhythms. Fourth, one could investigate how
distances in the orbifolds relate to perceptual judgments of chord similarity. Finally,
understanding the relation between harmony and counterpoint may suggest new techniques
to contemporary composers.
7

REFERENCES

1. C. Masson, Nouveau Trait des Rgles pour la Composition de la Musique (Da Capo,
New York, 1967).

2. O. Hostinsky, Die Lehre von den musikalischen Klngen (H. Dominicus, Prague, 1879).

3. D. Huron, Mus. Percept. 19, 1 (2001).

4. J. D. Heinichen, Der General-Bass in der Composition (G. Olms, New York, 1969).

5. R. Cohn, J. Mus. Theory 41, 1 (1997).

6. J. Roeder, thesis, Yale University (1984).

7. E. Agmon, Musikometrica 3, 15 (1991).

8. R. Cohn, Mus. Anal. 15, 9 (1996).

9. C. Callender, Mus. Theory Online 10 (2004).

10. G. Mazzola, The Topos of Music (Birkhuser, Boston, 2002).

11. R. Morris, Mus. Theory Spectrum 20, 175 (1998).

12. J. Douthett, P. Steinbach, J. Mus. Theory 42, 241(1998).

13. J. Straus, Mus. Theory Spectrum 25, 305 (2003).

14. F. Attneave, R. Olson, Am. Psychol. 84, 147 (1971).

15. R. Shepard, Psychol. Rev. 89, 305 (1982).

16. Materials and methods are available as supporting material on Science Online.

17. I. Satake, Proc. Natl. Acad. Sci. U.S.A. 42, 359 (1956).

18. W. Thurston, The Geometry and Topology of Three Manifolds, available at
http://www.msri.org/publications/books/gt3m/.

8

19. http://music.princeton.edu/~dmitri/ChordGeometries.html.

20. R. Cohn, J. Mus. Theory 35, 1 (1991).

21. W. Sethares, Tuning, Timbre, Spectrum, Scale (Springer, New York, 2005).

22. R. Cohn, 19
th
-Cent. Mus. 22, 213 (1999).

23. B. Boretz, Perspect. New Mus. 11, 146 (1972).

24. C. Callender, I. Quinn, D. Tymoczko, paper presented to the John Clough Memorial
Conference, sponsored by the University of Chicago, in Chicago, IL, July 9, 2005.

25. Thanks to D. Biss, C. Callender, E. Camp, N. Elkies, P. Godfrey Smith, R. Hall, A.
Healy, I. Quinn, N. Weiner, and M. Weisberg.

Supporting Online Material
Materials and Methods
Figs. S1 to S12
Table S1
Movies S1 to S4
Soundfile S1
References

&
?

# n
b
&
b


# n
#
b
#
b

#

b


b b

C F C G C D
7
G
7
C
7
F
7
Fs
7
F
7
As
7
Bf
7
{B, C, Df}
A B C D

Fig. 1. Efficient voice leading between transpositionally and inversionally related chords.
These progressions exploit three near-symmetries: transposition (A-B), inversion (C), and
permutation (D). Sources: classical music (A), jazz (B), Wagners Parsifal (C), Debussys
Faun (C), and contemporary atonality (D) (Soundfile S1).
9

01
00
02
03
04
05
06
70
80
90
t0
e0
[00]
11
12
13
14
15
16
17
81
91
t1
[e1]
e1
e2
e3
e4
e5
6e
7e
8e
9e
te
ee
22
23
24
25
26
27
28
92
[t2]
t2
t3
t4
5t
6t
7t
8t
9t
tt
33
34
35
36
37
38
[39] 93
49
59
69
79
89
99
44
45
46
47
[48]
48
58
68
78
88
55
56
[57]
57
67
77
[66]
66
unison
transposition
tritone
unison
minor second
minor second
major second
major second
minor third
minor third
major third
major third
perfect fourth
perfect fourth


Fig. 2. The orbifold T
2
/S
2
. C = 0, Cs = 1, etc., with Bf = t, and B = e. The left edge is
identified with the right. The voice leadings (C, Df)(Df, C) and (C, G)(Cs, Fs) are
shown; the first reflects off the singular boundary.

Table 1. Common sonorities in Western tonal music. The middle column lists the best
equal-tempered approximation to the first n pitch classes of the harmonic series. The right
column lists other good approximations. All divide the octave nearly evenly.

2 notes CG CFs
3 notes CEG CEfGf
CEfG
CEGs
4 notes CEGBf CEfGfA
CEfGfBf
CEfGBf
CEGB
5 notes CDEGBf CDEGA
CDEGB
6 notes CDEFsGBf CDEfFGBf
CDEFsGsBf
7 notes CDEFsGABf CDEFGABf
CDEfFsGABf

SUPPORTING ONLINE MATERIAL
TABLE OF CONTENTS

1. Overview of the argument 1
2. Pitch and pitch class 1
3. Comparing voice leadings 2
4. Minimal voice leadings and voice crossings 4
5. Derivation of the voice-leading orbifolds 7
6. Efficient voice leading and symmetry 8
7. Maximal evenness and minimal voice leading 11
8. A polynomial-time algorithm for finding a minimal 13
voice leading between arbitrary chords
9. Supplementary Figures and Tables 15
10. References 28

1. Overview of the argument. Our goal is to model composers judgments about the
relative sizes of voice leadings. One cannot assume that these judgments will be
consistent with any mathematical norm or metric (3). However, it is reasonable to
stipulate that a theoretical model should reflect widely recognized features of Western
music. Since musicians treat transposition and inversion as preserving musical distance, I
will require that voice-leading comparisons be invariant under transposition and inversion
of any of their individual musical voices (2). Since Western pedagogues instruct
composers to minimize voice-leading size while eschewing voice crossings (13), I
will require that crossed voice leadings be no smaller than their natural, uncrossed
alternatives (34). These minimal requirements suffice to establish a number of
interesting results about voice leading (68).
Sections 2 and 3 articulate mathematical constraints on methods of comparing
voice leadings. Section 4 shows that these constraints formalize my two musical
requirements. Section 5 derives the voice-leading orbifolds. Sections 6 and 7 relate the
structure of a chord to its voice-leading possibilities. Section 8 introduces a further
constraint on methods of comparing voice leadings, and uses it to derive a polynomial-
time algorithm for finding a minimal voice leading (not necessarily bijective) between
arbitrary chords.

2. Pitch and pitch class. Pitches are modeled as real numbers. The distance between two
pitches p, q is the absolute value of their difference, |q p|. Pitch classes are modeled as
points in the quotient space R/12Z. They are sets of real numbers {p + 12k | k Z}, with
p representing some pitch in the pitch class. We can label these sets using real numbers
in the range 0 x < 12. The elements of R/12Z form a group under addition of their
labels modulo 12Z. The distance between pitch classes a and b, written ||b a||
12Z
, is the
2

smallest nonnegative real number x such that, if p is a pitch belonging to pitch class a,
then either p + x or p x belongs to pitch class b.
Let (p
1
, p
2
, ..., p
n
)(q
1
, q
2
, ..., q
n
) be a voice leading between multisets of pitches,
and let (a
1
, a
2
, ..., a
n
)(b
1
, b
2
, ..., b
n
) be a voice leading between multisets of pitch
classes. I will say that the two voice leadings are associated if the pitches p
i
and q
i
all
belong to the pitch classes a
i
and b
i
, respectively. The displacement multiset of the voice
leading (p
1
, p
2
, ..., p
n
)(q
1
, q
2
, ..., q
n
) is the multiset of distances {|q
i
p
i
|}. Similarly, the
displacement multiset of the voice leading (a
1
, a
2
, ..., a
n
)(b
1
, b
2
, ..., b
n
) is the multiset of
distances {||b
i
a
i
||
12Z
}. I will require that a method of comparing voice leadings be
invariant under transposition and inversion of its individual voices: in any voice leading,
one can replace the voice (p, q) with (q + x, p + x) without changing the voice leadings
size. It follows that the size of a voice leading depends only on its displacement multiset.
I will also require that the size of a displacement multiset be nondecreasing in each of its
members: we can replace any member of a displacement multiset with a smaller number
without increasing the size of the resulting voice leading (3). This ensures that a voice
leading between multisets of pitch classes will be the same size as the smallest associated
voice leading between multisets of pitches.

3. Comparing voice leadings. A method of comparing voice leadings is a relation
over multisets of nonnegative real numbers that is reflexive, transitive, and total (a total
preorder; see Table S1 for more information). The relation must satisfy what I call the
distribution constraint:

{x
1
+ c, x
2
, ..., x
n
} {x
1
, x
2
+ c, ..., x
n
} {x
1
, x
2
, ..., x
n
}, for x
1
> x
2
, c > 0

(NB: since multisets are unordered, the numerical subscripts do not have ordinal
significance: x
1
is no more first than x
2
or x
n
.) The first inequality requires that the total
preorder not consider an uneven distribution of values to be smaller than a more even
distribution with the same total sum: if X is an n-member displacement multiset whose
members sum to x, then {x, 0, ..., 0} X {x/n, x/n, ..., x/n}. Thus, x semitones of
motion in a single voice yields at least as large a voice leading as x semitones of motion
distributed over multiple voices. This requirement is a weakened relative of the triangle
inequality (4). The distribution constraints second inequality requires that the size of a
multiset be nondecreasing in each of its members: increasing the size of any number in a
multiset never decreases that multisets size. If a total preorder satisfies both inequalities
strictly, I will say that it strictly satisfies the distribution constraint.
Every music-theoretical method of comparing voice leadings satisfies the
distribution constraint.

3

A. Smoothness. The size of a voice leading is the sum of the objects in the
displacement multiset (57). Thus {2, 2} > {3.999} > {1, 1, 1}.
Smoothness is sometimes called the taxicab norm. It reflects aggregate
physical distance on keyboard instruments. Smoothness satisfies the
distribution constraint non-strictly.

B. L
p
vector norms. Smoothness is analogous to the L
1
vector norm,
though the components of vectors are ordered whereas the members of
displacement multisets are not. The analogues to the L
p
vector norms
strictly satisfy the distribution constraint for finite p > 1. The Euclidean
vector norm L
2
has been used by Callender (8).

C. Semitonal and stepwise voice leadings. According to the L

vector
norm, the size of a displacement multiset is its largest member. The
musical terms semitonal voice leading and stepwise voice leading
refer to this measure of voice-leading size. Semitonal voice leadings have
an L

norm of 1; stepwise voice leadings have an L

norm less than or

equal to 2. The L

norm measures the largest physical distance moved by

any single voice on a keyboard instrument. It satisfies the distribution
constraint non-strictly.

D. Parsimony. Parsimony is related to the lexicographic ordering. It
generalizes a notion introduced by Richard Cohn and developed by Jack
Douthett and Peter Steinbach (910). Given two voice leadings, and ,
is smaller (or more parsimonious) than iff there exists some real
number d such that
1) for all real numbers c > d, c appears the same number of times
in the displacement multisets of and ; and
2) d appears fewer times in the displacement multiset of than .
Thus, according to Parsimony, {3 + } > {3, 2, 1} (in this case, d = 3 + ),
and {4, 3, 3} > {4, 3} (here, d = 3). Parsimony strictly satisfies the
distribution constraint.

The first three methods of comparing voice leadings resemble familiar mathematical
metrics. Parsimony, however, does not. This is because there is no function from
multisets to real numbers, such that f(X) f(Y) if and only if X Y according to
Parsimony. Nevertheless, Parsimony represents a musically viable way of thinking about
voice-leading size.

4

4. Minimal voice leadings and voice crossings. This section connects the formalism of
23 to the avoidance of voice crossings. I begin by demonstrating that if a method of
comparing voice leadings obeys the distribution constraint, then there is a minimal
bijective voice leading between any two chords that is crossing free. I then argue in the
opposite direction, showing that if a method of comparing voice leadings violates the
distribution constraint, some crossed voice leading will be smaller than its natural
uncrossed alternative. I conclude with a few remarks about the relation between the
avoidance of voice crossings and judgments about voice-leading size.
Let (p
1
, p
2
, ..., p
n
)(q
1
, q
2
, ..., q
n
) be a voice leading between multisets of pitches.
The voice leading has no voice crossings (is crossing free) if p
i
> p
j
implies q
i
q
j
for all
i, j. I will say that it is strongly crossing free if (p
1
+ 12k
1
, p
2
+ 12k
2
, ..., p
n
+ 12k
n
)
(q
1
+ 12k
1
, q
2
+ 12k
2
, ..., q
n
+ 12k
n
) is crossing free, for all integers k
i
. Thus one cannot
introduce crossings into a strongly crossing-free voice leading simply by shifting the
octaves in which its voices appear. A voice leading between multisets of pitch classes is
crossing free if it is associated with a strongly crossing-free voice leading between
multisets of pitches in which no voice moves by more than six semitones. Intuitively, a
voice leading is crossing free if the notes of the source chord can be connected to those of
the target along minimal-length line-segments that intersect only at their endpoints.
The following theorem shows that one can always find a minimal voice leading
that is crossing free. Though I state the argument using chords of pitches, it generalizes
straightforwardly to chords of pitch classes.

THEOREM 1. Let P and Q be any two n-note multisets of pitches, and
let our method of comparing voice leadings be a total preorder satisfying
the distribution constraint. Then there will exist a minimal bijective voice
leading from P to Q that is crossing free. If the total preorder strictly
satisfies the distribution constraint, then every minimal bijective voice
leading from P to Q will be crossing free.

Suppose (p
1
, p
2
, ..., p
n
)(q
1
, q
2
, ..., q
n
) contains a crossing, with p
1
< p
2
and q
1
> q
2

[Fig. S7(a)]. Let (r, r) be the intersection, in R
2
, of the line segment (p
1
, p
2
)(q
1
, q
2
)
with the line x = y. Since a lines slope is constant, | r p
2
| / | r p
1
| = | q
2
r | / | q
1
r |.
There are three cases: either both numerators are greater than their denominators, both are
equal to their denominators, or both are less than their denominators. In all three cases,
the distribution constraint implies that, for any numbers d
i
,

{|q
1
r| + |r p
1
|, |q
2
r| + |r p
2
|, d
1
, ..., d
m
} {|q
1
r| + |r p
2
|, |q
2
r| + |r p
1
|, d
1
, ..., d
m
},

since the multiset on the left is more uneven than that on the right. By construction
|q
1
r| + |r p
1
| = | q
1
p
1
| and | q
2
r| + | r p
2
| = | q
2
p
2
|. By the triangle inequality,
5

|q
1
p
2
| |q
1
r| + |r p
2
| and | q
2
p
1
| |q
2
r| + |r p
1
|. Thus, using the distribution
constraint again, {|q
1
p
1
|, |q
2
p
2
|, d
1
, ..., d
m
} {|q
1
p
2
|, |q
2
p
1
|, d
1
, ..., d
m
} and the
voice leading (p
1
, p
2
, ..., p
n
)(q
1
, q
2
, ..., q
n
) is no smaller than (p
1
, p
2
, ..., p
n
)
(q
2
, q
1
, ..., q
n
). We do not increase the size of a voice leading when we replace the
crossed voices (p
1
, q
1
) and (p
2
, q
2
) with the uncrossed (p
1
, q
2
) and (p
2
, q
1
).
Suppose now that our total preorder strictly satisfies the distribution constraint. In
this case, I claim that the crossed voice leading (p
1
, p
2
, ..., p
n
)(q
1
, q
2
, ..., q
n
) is strictly
larger than the uncrossed (p
1
, p
2
, ..., p
n
)(q
2
, q
1
, ..., q
n
). This can be proved by
contradiction. Suppose the two voice leadings were the same size. Since the total
preorder strictly satisfies the distribution constraint, | r p
2
| / | r p
1
| = | q
2
r| / | q
1
r| = 1.
Furthermore, |q
1
r| + |r p
2
| = |q
1
p
2
| and |q
2
r| + |r p
1
| = |q
2
p
1
|. Therefore,
|q
1
p
1
| = |q
1
p
2
| and |q
2
p
1
| = |q
2
p
2
|. In other words, q
1
and q
2
are each equidistant
from p
1
and p
2
. This contradicts the hypothesis that the voices (p
1
, q
1
) and (p
2
, q
2
) cross.
We conclude that if a method of comparing voice leadings strictly satisfies the
distribution constraint, then removing any voice crossing will make the resulting voice
leading strictly smaller.
It remains to be shown that replacing (p
1
, q
1
) and (p
2
, q
2
)

with (p
1
, q
2
) and (p
2
, q
1
)
does not introduce additional crossings into the voice leading. Recall that p
1
< p
2
and
q
1
> q
2
. Now suppose some voice (p
3
, q
3
) crosses (p
1
, q
2
) [Fig. S7(b)]. Then either p
3
< p
1

and q
3
> q
2
, in which case (p
3
, q
3
) crosses (p
2
, q
2
); or p
3
> p
1
and q
3
< q
2
, in which case
(p
3
, q
3
) crosses (p
1
, q
1
). Similarly, if (p
3
, q
3
) crosses (p
2
, q
1
) it crosses either (p
1
, q
1
) or
(p
2
, q
2
). Finally, suppose (p
3
, q
3
) crosses both (p
1
, q
2
) and (p
2
, q
1
). Then, either p
3
< p
1
and
q
3
> q
1
, or p
3
> p
2
and q
3
< q
2
; in either case (p
3
, q
3
) crosses both (p
1
, q
1
) and (p
2
, q
2
). The
voice (p
3
, q
3
) therefore crosses at least as many voices in the old, crossed voice leading as
it does in the new, uncrossed one. Replacing (p
1
, q
1
) and (p
2
, q
2
) with (p
1
, q
2
) and (p
2
, q
1
)
therefore decreases the total number of crossings in the voice leading.
This establishes Theorem 1. The result immediately generalizes to chords of
pitch classes.
I now argue in the opposite direction, showing that violations of the distribution
constraint imply that some crossed voice leading is smaller than its natural uncrossed
alternative.
Suppose a method of comparing voice leadings depends only on the displacement
multiset and violates the distribution constraints first inequality. That is, some voice
leading with displacement multiset {x
1
+ c, x
2
} is smaller than some voice leading with
displacement multiset {x
1
, x
2
+ c}, with x
1
> x
2
and c > 0. It follows that the crossed voice
leading (p, p + x
1
x
2
)(p + x
1
+ c, p + x
1
) is smaller than the uncrossed voice leading
(p, p + x
1
x
2
)(p + x
1
, p + x
1
+ c) [Figure S8(a)]. Removing the crossing therefore
increases the size of the voice leading. Additional voices do not affect the logic of the
argument.
6

Now suppose that a method of comparing voice leadings violates the distribution
constraints second inequality: some voice leading with displacement multiset
{x
1
+ c, x
2
, x
3
} is smaller than some voice leading with displacement multiset {x
1
, x
2
, x
3
},
with x
3
x
2
> 0 and c > 0. Consider the uncrossed voice leading

(p + x
1
+ x
2
, p + x
1
+ x
3
+ c)(p + x
1
, p + x
1
+ c).

Suppose we decide to add pitch p to the first chord, and to map it to some note in the
second chord. As shown in Figure S8(b), the crossed alternative

(p, p + x
1
+ x
2
, p + x
1
+ x
3
+ c)(p + x
1
+ c, p + x
1
, p + x
1
+ c)

is smaller than the uncrossed (p, p + x
1
+ x
2
, p + x
1
+ x
3
+ c)(p + x
1
, p + x
1
, p + x
1
+ c).
Thus we can conclude that, for multisets with three or more nonzero members, total
preorders satisfying the distribution constraint are the only methods of comparing voice
leadings that are transpositionally and inversionally invariant, and that embody the
principle that voice crossings do not make a voice leading smaller.
It remains possible that composers considered crossed voice leadings to be
smaller than uncrossed voice leadings, yet avoided crossings for reasons unrelated to
voice-leading size. There are two reasons to reject this hypothesis, however. First, if
there were a conflict between the goal of avoiding voice crossings and the goal of
minimizing voice-leading size, there would presumably be some evidence of this in the
musical, theoretical, or pedagogical literature. In particular, one would not expect voice-
leading preferences to be even approximately invariant under transposition of their
individual musical voices: when voices were close together, the goal of avoiding voice
crossings would trump the goal of minimizing voice-leading size, while when voices
were farther apart minimal voice leadings could be used freely. Western music would
therefore manifest two distinct sets of voice-leading preferences, depending on whether
voices were close together or far apart. There is no evidence that this is so. No
pedagogue or theorist has ever proposed a method of comparing voice leadings that
favors voice crossings. Second, methods of comparing voice leadings that favor
crossings tend to be undesirable for independent reasons: they cannot, for example,
represent aggregate or maximal keyboard distance (3)physical quantities that
presumably help shape composers intuitive judgments of voice-leading size. Violations
of the distribution constraint are also violations of the triangle inequality (4) and hence
conflict with basic geometric intuitions about distance. Finally, it is difficult to develop
simple, robust, and musically plausible heuristics that help composers to reduce voice-
leading size when using total preorders that violate the distribution constraint. By
contrast, when a method of measuring voice-leading size obeys the distribution
constraint, then a composer can always reduce a voice leadings size simply by removing
7

crossings. We therefore have good reason to require that the goal of avoiding voice
crossings not conflict with the goal of minimizing voice-leading size.

5. Derivation of the voice-leading orbifolds. I now turn from methods of comparing
voice leadings to the geometry of musical chords. I begin by deriving Figure 2 in the
main text. Figure S9(a) shows the 2-torus T
2
, representing the space of ordered pairs of
pitch classes. To form the space of unordered pairs we need to identify all points (x, y)
and (y, x). As can be seen from Figure S9(a), this involves folding the 2-torus along the
diagonal line-segment AB. The result is a triangle with two sides identified, shown in
Figure S9(b). This figure is a Mbius strip. To see why, cut Figure S9(b) along the line-
segment CD, creating two detached triangles. Then attach AC on one triangle to CB on
the other, turning one piece of paper over so that chords on the shared edge match. The
result is the main texts Figure 2. Figure S4 shows how Figure 2 tiles the infinite plane
representing ordered pairs of pitches.
I now proceed more abstractly, deriving the orbifolds T
n
/S
n
for arbitrary n. Since
R
n
/12Z
n
is the n-torus T
n
, we can describe T
n
/S
n
as the quotient of R
n
by the semidirect
product 12Z
n
S
n
. We can construct our orbifolds by describing fundamental domains
of 12Z
n
S
n
in R
n
, and showing how their boundary points are to be identified.
To construct a fundamental domain of S
n
in R
n
, consider only those points with
coordinates in nondecreasing order: {(x
1
, x
2
, ..., x
n
) | x
1
x
2
... x
n
}. To incorporate the
12Z
n
action, require that x
n
x
1
+ 12, and 0 x
i
12. The resulting fundamental
domain is an n-dimensional prism whose base is an (n1)-dimensional simplex. To see
why, observe that the inequalities x
1
x
2
... x
n
x
1
+ 12 define an (n1)-simplex in
every plane x
i


= c. Every point in the fundamental domain lies in exactly one of these
simplexes and every one of these simplexes is either entirely within, or entirely outside
of, the fundamental domain. Since addition by (d, d, ..., d) sends the simplex in the plane
x
i


= c to the simplex in the plane x
i


= c + nd, the vector (1, 1, ..., 1) points in the
direction of the height coordinate of the prism. Figure S6 illustrates three lower-
dimensional cases.
It remains to be determined how the two simplicial faces of the prism are to be
identified. Define the function

O(x
1
, x
2
, ..., x
n
) = (x
2
12/n, x
3
12/n, ..., x
n
12/n, x
1
+ 12 12/n).

O is an automorphism of the prism that cyclically permutes the vertices of the simplex in
each plane x
i


= c. By repeatedly applying O to a chord X, we can obtain chords X,
O(X), O
2
(X), ... O
n-1
(X), all related by transposition, and with pitch classes all summing to
the same value. In Euclidean space, O is a rotation when the prism has odd dimension,
and a rotation-plus-reflection otherwise. On Figure S6(a), O reflects the square around
the line at its center. On Figure S6(b), O rotates the prism around the central line. On
8

Figure S6(c), O is a rotation-plus-reflection that cyclically permutes the vertices of each
tetrahedron.
Now suppose that (x
1
, x
2
, ..., x
n
) lies in the x
i


= 0 face of the prism (the base).
Then (x
1
+ 12/n, x
2
+ 12/n, ..., x
n
+ 12/n) is the corresponding point on the opposite face.
O(x
1
+ 12/n, x
2
+ 12/n, ..., x
n
+ 12/n) = (x
2
, x
3
, ..., x
n
, x
1
+ 12) also lies on that same face.
When we disregard order and octave, (x
1
, x
2
, ..., x
n
) and (x
2
, x
3
, ..., x
n
, x
1
+ 12) represent
the same chord, and should be identified. Thus for any point on the base of the prism x
and corresponding point on the opposite face y, we identify x with O(y). Figures S5 and
S6 illustrate, as does a free computer program written by the author (11).

6. Efficient voice leading, chord structure, and symmetry. This section relates the
internal structure of a chord to its voice-leading capabilities. Let F be any function over
pitch classes. I begin by showing how the internal structure of a chord A determines
whether the chords A and F(A) can be linked by efficient bijective voice leading. I then
show how a chords voice-leading capabilities relate to its distance from nearby
symmetrical chords. Suppose G is an isometry of pitch class space, and S
G
is invariant
under G. I first show that the size of any voice leading from A to S
G
sets an upper bound
on the size of the minimal voice leading(s) from A to G(A). I then argue in the opposite
direction, showing that the size of a bijective crossing-free voice leading from A to G(A)
sets an upper bound on the size of the smallest of the voice leadings from A to any chord
S
G
that is invariant under G. Since there is always a minimal voice leading that is
crossing free (4), these results show that there is a close relationship between the size of
the minimal bijective voice leading(s) from A to G(A) and the size of the smallest of the
bijective voice leadings from A to the chords S
G
that are invariant under G.
In 67, I write T
x
(a) = a + x to refer to transposition by x semitones, and
I
x
(a) = x a to refer to the inversion that sends 0 to x. If F
1
and F
2
are two functions over
pitch classes, and A = {a
1
, a
2
, ..., a
n
} is any chord, then I will say that F
1
resembles F
2

(with respect to A) when F
1
(a
i
) F
2
(a
i
) for all a
i
. Finally, I will say that a chord A is
invariant under if there is a trivial voice leading (a
1
, a
2
, ..., a
n
)(a
(1)
, a
(2)
, ..., a
(n)
).
Observation A. Let F be any function over pitch classes. Chords A and F(A) can
be linked by efficient, bijective voice leading if and only if some permutation of As notes
resembles F (with respect to A).
Any bijective voice leading from A to F(A) can be written (a
1
, a
2
, ..., a
n
)
(F(a
(1)
), F(a
(2)
), ..., F(a
(n)
)) where is some permutation. This voice leading will have
displacement multiset {||F(a
(i)
) a
i
||
12Z
}, or, equivalently, {||F(a
i
) a

1
(i)
||
12Z
}. The size
of this voice leading depends on the difference between the effects of the function F and
the permutation
-1
. The voice leading will be small when F(a
i
) a

1
(i)
, for all i, and
will be trivial with strict equality.
Thus, the internal structure of any chord A, as represented by the size of the voice
leadings (a
1
, a
2
, ..., a
n
)(a
(1)
, a
(2)
, ..., a
(n)
), for all permutations , determines the
9

functions F such that A and F(A) can be linked by efficient voice leading. For example,
when the notes {a
1
, a
2
, ..., a
n
} are all close together, the voice leadings (a
1
, a
2
, ..., a
n
)
(a
(1)
, a
(2)
, ..., a
(n)
) will resemble the trivial voice leading. In this case, A and F(A) can be
linked by efficient voice leading only when F resembles the identity. By contrast, when
the notes {a
1
, a
2
, ..., a
n
} divide the octave nearly evenly, then there will be a voice leading
(a
1
, a
2
, ..., a
n
)(a
(1)
, a
(2)
, ..., a
(n)
) that resembles transposition by 12/n semitones. In
this case, A and T
12/n
(A) can be linked by efficient voice leading.
Observation B. Let G be an isometry of pitch class space, and let S
G
be invariant
under G. The size of any voice leading from A to S
G
sets an upper bound on the size of
the minimal voice leading(s) from A to G(A).
Suppose that G is an isometry. It follows that for any voice leading from A to S
G
,
with S
G
invariant under G, we can find an equally large voice leading from S
G
to G(A). If
a voice leading from A to S
G
has displacement multiset {d
1
, d
2
, ..., d
n
}, then there is some
permutation such that the minimal voice leading from A to G(A) has displacement
multiset less than or equal to {d
1
+ d
(1)
, d
2
+ d
(2)
, ..., d
n
+ d
(n)
}. The distribution
constraint tells us that no multiset {d
1
+ d
(1)
, d
2
+ d
(2)
, ..., d
n
+ d
(n)
} can be larger than
{2d
1
, 2d
2
, ..., 2d
n
}. The size of any voice leading from A to S
G
thus sets an upper bound
on the size of the minimal voice leading(s) from A to G(A). Note that the same reasoning
can be used when we require that a bijective voice leading from A to A involve the
permutation : the size of a voice leading from A to S

, with S

invariant under , sets

an upper bound on the size of the smallest voice leading of the form (a
1
, a
2
, ..., a
n
)
(a
(1)
, a
(2)
, ..., a
(n)
).
Observation C. Let G be an isometry of pitch class space. The size of a bijective
crossing-free voice leading from A to G(A) sets an upper bound on the size of the smallest
of the voice leadings from A to any chord S
G
that is invariant under G.
Transposition and inversion are the isometries of pitch class space. When G is an
inversion, the reasoning is straightforward. A bijective, crossing-free voice leading from
A to I
x
(A) maps pitch class a
i
to x a
(i)
, where is a permutation involution. Therefore,
it also maps a
(i)
to x a
i
. Let pitch class b
i
be d/2 semitones away from both a
i
and
x a
(i)
. Pitch class x b
i
will therefore be d/2 semitones away from both a
(i)
and x a
i
,
and we can label it b
(i)
. It follows that if a bijective, crossing-free voice leading from A
to I
x
(A) has displacement multiset {d
1
, d
2
, ..., d
n
}, then we can find a voice leading from A
to S
I
x
= (b
1
, b
2
, ..., b
n
) with displacement multiset {d
1
/2, d
2
/2, ..., d
n
/2} (12).
When G is a transposition, things are more complicated. In this case, we cannot
use the size of a bijective crossing-free voice leading from A to T
x
(A) to determine the
size of a voice leading from A to S
T
x
, with S
T
x
invariant under T
x
. Figure S10 illustrates.
The chords {0, 3, 6, 8}, or {C, Ef, Gf, Af}, and {0, 3, 5, 8}, or {C, Ef, F, Af}, can be
arranged as cycles of pitch classes whose elements are 2, 3, 3, and 4 semitones apart.
Consequently, there is a minimal bijective voice leading between each chord and its T
3
-
form having displacement multiset {||2 3||
12Z
, || 3 3||
12Z
, || 3 3||
12Z
, || 4 3||
12Z
}, or
10

{1, 0, 0, 1}. However, one can transform {0, 3, 6, 8} into a T
3
-invariant chord by moving
only one note, whereas one needs to move two notes to transform {0, 3, 5, 8} into a T
3
-
invariant chord. Many methods of comparing voice leadings therefore consider
{0, 3, 6, 8} to have a smaller minimal bijective voice leading to the nearest T
3
-invariant
chord than does {0, 3, 5, 8}. But this difference is not reflected in the size of the minimal
bijective voice leadings between the two chords and their T
3
-forms. Thus, unlike the
inversional case considered above, the size of a bijective crossing-free voice leading from
A to T
x
(A) does not directly determine the size of a voice leading from A to S
T
x
, where
S
T
x
is invariant under T
x
.
However, the size of a bijective crossing-free voice leading from A to T
x
(A) does
set an upper bound on the size of the smallest of the bijective voice leadings from A to
any S
T
x
. Let A be a chord with n notes, and let be a bijective voice leading of the form
(a
0
, a
1
, ..., a
n-1
)(a
n-1
+ x, a
0
+ x, a
1
+ x, ..., a
n-2
+ x), which combines transposition by x
semitones with a permutation consisting of a single cycle. Furthermore, let nx 0
modulo 12Z, so that there is an S
T
x
that is invariant under T
x
. The voice leading has
displacement multiset {||(a
i
+ x) a
i+1 (mod n)
||
12Z
}. Write d
i
= (a
i
+ x a
i+1 (mod n)
), so that
the displacement multiset is {||d
i
||
12Z
}, and a
i+1 (mod n)
a
i
= d
i
+ x. Figure S10(c) shows that
there is a voice leading from A to some S
T
x
having displacement multiset

{0, ||d
0
||
12Z
, ||d
n-1
||
12Z
, ||d
0
+ d
1
||
12Z
, ||d
n-1
+ d
n-2
||
12Z
, ||d
0
+ d
1
+ d
2
||
12Z
, ...}.

This multiset has one zero member, two members of the form ||d
i
||
12Z
, two members of the
form ||d
i
+ d
j
||
12Z
, two members of the form ||d
i
+ d
j
+ d
k
||
12Z
, and so on. By the
distribution constraint, this multiset will be smaller than

{0, ||d
0
||
12Z
, ||d
n-1
||
12Z
, ||d
0
||
12Z
+ ||d
1
||
12Z
, ||d
n-1
||
12Z
+ ||d
n-2
||
12Z
, ||d
0
||
12Z
+ ||d
1
||
12Z
+ ||d
2
||
12Z
, ...},

where we have used the fact that ||a + b||
12Z
||a||
12Z
+ ||b||
12Z
. Now arrange the quantities
||d
i
||
12Z
in a sequence from largest to smallest, labeling the elements of this sequence
c
1
, c
2
, ..., c
n
such that c
i
c
i+1
. By the distribution constraint,

{0, ||d
0
||
12Z
, ||d
n-1
||
12Z
, ||d
0
||
12Z
+ ||d
1
||
12Z
, ||d
n-1
||
12Z
+ ||d
n-2
||
12Z
, ||d
0
||
12Z
+ ||d
1
||
12Z
+ ||d
2
||
12Z
, ...}
{0, c
1
, c
2
, c
1
+ c
2
, c
3
+ c
4
, c
3
+ c
4
+ c
5
, ...}.

Return to the voice leading : AT
x
(A), which has the form (a
0
, a
1
, ..., a
n-1
)
(a
n-1
+ x, a
0
+ x, a
1
+ x, ..., a
n-2
+ x). Label the members of its displacement multiset
c
1
, c
2
, ..., c
n
such that c
i
c
i+1
. The preceding paragraph shows that we can find a voice
leading : AS
T
x
with a displacement multiset that is smaller than the first n elements of
the sequence (0, c
2
, c
3
, c
2
+ c
3
, c
4
+ c
5
, c
2
+ c
3
+ c
4
, c
5
+ c
6
+ c
7
, ...). (The circular nature of
pitch class space allows us to choose a
0
so as to avoid any terms involving c
1
.) We
11

conclude that the size of our voice leading from A to T
x
(A) sets a limit on the size of the
smallest of the voice leadings from A to any S
T
x
.
This argument has considered only simple bijective voice leadings whose
permutational component consists of a single cycle. However, the reasoning generalizes
to bijective voice leadings whose permutational component consists of multiple such
cycles, and hence to all bijective voice leadings. (Indeed, the limit we have derived also
applies to these voice leadings.) We can also use the argument to show that the size of a
bijective voice leading (a
1
, a
2
, ..., a
n
)(a
(1)
, a
(2)
, ..., a
(n)
) limits the size of the smallest
of the voice leadings from A to any chord S

that is invariant under .

7. Maximal evenness and minimal voice leading. I now prove that no bijective voice
leading from A to T
x
(A) can be smaller than the minimal bijective voice leading from E to
T
x
(E), where E has the same cardinality as A and divides the octave evenly. Thus the
perfectly even chord lying at the center of the voice-leading orbifold has the smallest
possible minimal bijective voice leading to all of its transpositions. As one moves away
from this chord, and toward the singular boundary of the orbifold, the size of the voice
leadings between chords and their transpositions will in general tend to increase.
However, for a given transposition T
x
, the rate of increase of these voice leadings
depends on the direction in which one is moving, since the size of the minimal voice
leading from A to T
x
(A) depends not just on the distance from chord A to the nearest T
x
-
invariant chord, but also on the internal structure of the chord itself (6).

THEOREM 2. Let A be any multiset of cardinality n, and let our method
of comparing voice leadings be a total preorder satisfying the distribution
constraint. For all x, the minimal bijective voice leading between A and
T
x
(A) can be no smaller than the minimal bijective voice leading between
E and T
x
(E), where E divides pitch-class space into n equal parts.

Since E is invariant under transposition by 12/n semitones, the crossing-free voice
leadings between E and T
x
(E) will have the form (e
1
, e
2
, ..., e
n
)(e
1
+ c, e
2
+ c, ... e
n
+ c),
where c x modulo (12/n)Z. (NB: c is congruent to x modulo (12/n)Z, not modulo 12Z.)
Choose c so that ||c||
12Z
is as small as possible. The displacement multiset will have n
members each equal to ||c||
12Z
. By the distribution constraint, this multiset is as small as
any n-member multiset with the same or greater sum.
Now consider any bijective voice leading between transpositionally related n-note
chords (a
1
, a
2
, ..., a
n
)(a
(1)
+ x, a
(2)
+ x, ..., a
(n)
+ x). The sum of the members in the
displacement multiset is

||a
(i)
+ x a
i
||
12Z
. Since ||a||
12Z
+ ||b||
12Z
||a + b||
12Z
, this sum is
greater than or equal to ||nx + (a
(i)
a
i
)||
12Z
= ||nx||
12Z
. By construction, ||nx||
12Z
= ||nc||
12Z
,
since x c modulo (12/n)Z, and ||nc||
12Z
= n||c||
12Z
, since we can always choose c such that
12

||c||
12Z
6/n. Thus the voice leading can be no smaller than the minimal voice leading
between E and T
x
(E).
I now prove a corollary that applies to collections of k equally spaced pitch
classes. Musically, these pitch classes represent an equal-tempered chromatic scale.
Such discrete musical universes will not always contain chords that divide the octave into
n perfectly even pieces. For example, no five-note chord in twelve-tone equal
temperament divides the octave perfectly evenly. But as Douthett and Clough have
shown (13), equal-tempered scales will always contain a unique collection of maximally
even chords dividing the octave as evenly as possible (14). These maximally even
chords are the discrete analogues to the perfectly even chords we have been considering.
In what follows, it will be convenient to relabel the pitch classes so that the
equally-spaced points in our chromatic scale have integer co-ordinates. We will therefore
temporarily abandon R/12Z in favor of R/kZ, where k is an integer.

COROLLARY. Let A and M be integer-valued n-note submultisets of
R/kZ, and let M be maximally even. Then, for any integer x, the minimal
bijective voice leading between A and T
x
(A) can be no smaller than the
minimal bijective voice leading between M and T
x
(M).

The proof is very similar to the proof of Theorem 2. Since M is maximally even
(14), we can find a (not necessarily integer-valued!) chord {e
1
, e
2
, ..., e
n
} that divides the
octave into n precisely even parts, such that the crossing-free voice-leadings between M
and T
x
(M) can be written

(

e
1
,

e
2
, ...,

e
n
)(

e
1
+ c

e
2
+ c

, ...

e
n
+ c

), where c x modulo (k/n)Z

Choose c so that ||c||
kZ
is as small as possible. The displacement multiset will have n
members chosen from the set {||

||
kZ
, ||

||
kZ
}, and summing to n||c||
kZ
(15). By the
distribution constraint, this multiset is as small as any integer-valued n-note multiset with
the same or greater sum. As we saw in Theorem 2, the minimal bijective voice leading
between A and T
x
(A) will have a displacement multiset summing to at least n||c||
kZ
. Hence
it can be no smaller than the minimal bijective voice leading between M and T
x
(M).
The preceding corollary can also be applied to scales that do not divide pitch class
space evenly. To see why, note that we can always devise a metric on pitch class space
such that a given scale evenly divides the octave. For example, given the C major scale
(C, D, E, F, G, A, B) we can define a metric according to which the distances CD, DE,
EF, FG, GA, AB, and BC all equal one [Fig. S11(a)]. Musicians use the term
scale step to refer to this scale-dependent unit of distance. Two subsets of the scale are
related by scalar transposition if they are related by rotation relative to this new metric
13

[Fig. S11(b)]. Since our scale evenly divides the octave according to the new metric, the
preceding corollary applies.
Relative to this metric, the familiar diatonic tertian triads {C, E, G}, {D, F, A},
{E, G, B}, {F, A, C}, {G, B, D}, {A, C, E}, and {B, D, F} are maximally even and
related by transposition [Fig. S11(b)]. It follows from our corollary that the minimal
bijective voice leadings between these tertian triads will be as small as possible: a
bijective voice leading between any pair of transpositionally related three-note diatonic
subsets can be no smaller than the minimal bijective voice leading between diatonic
tertian triads related by that transposition. Analogous facts hold for the tertian seventh
chord {C, E, G, B} and its transpositions, which are also maximally even. Our corollary
therefore generalizes a result of Agmon (16), who first noted the special voice-leading
properties of diatonic tertian triads and seventh chords.

8. A polynomial-time algorithm for finding a minimal voice leading between two
chords. Given two chords A and B, how do we find a minimal voice leading between
them? The question is nontrivial, since a minimal voice leading need not be bijective.
For example, using any of the standard methods of comparing voice leadings, the
voiceleading (0, 0, 4, 6)(10, 0, 6, 6) is smaller than any of the bijective voice leadings
between {0, 4, 6} and {6, 10, 0} (17). Adding additional voices therefore allows us to
decrease the size of the voice leading. The large number of non-bijective voice leadings
between any two chordsroughly 2
mn
, where m and n are their cardinalitiesmeans that
an exhaustive search may be impractical, particularly in time-critical applications such as
interactive computer music.
Suppose, however, that our method of comparing voice leadings is a total
preorder satisfying both the distribution constraint and what I will call the recursion
constraint:

{x
1
, x
2
, ..., x
m
} {y
1
, y
2
, ..., y
n
} implies {x
1
, x
2
, ..., x
m
, c} {y
1
, y
2
, ..., y
n
, c}

The recursion constraint mandates a straightforward relationship between the size of a
multiset and the size of its sub-multisets. Every music-theoretical method of comparing
voice leadings satisfies this constraint. When a total preorder satisfies both the
distribution and recursion constraints, we can use the technique of dynamic
programming, common in computer science, to determine a minimal voice leading
between arbitrary chords in polynomial time (order n
2
m).
Define the ascending distance from pitch class a to pitch class b as the smallest
nonnegative real number x such that, if p is a pitch belonging to pitch class a, then p + x
belongs to pitch class b. Let(a
1
, a
2
, ..., a
m
, a
m+1
= a
1
) order the notes of chord A based on
increasing ascending distance from arbitrarily chosen a
1
. (Note that I repeat the first
element a
1
as the last element of the list.) Similarly for (b
1
, b
2
, ..., b
n
, b
n+1
= b
1
). The
14

notation [a
1
, ..., a
i
][b
1
, ... b
j
] will refer to all voice leadings from {a
1
, a
2
, ..., a
i
} to
{b
1
, b
2
, ... b
j
} that can be written with both chords subscripts in nondecreasing order.
Thus [a
1
, a
2
][b
1
, b
2
, b
3
] refers to (a
1
, a
1
, a
2
)(b
1
, b
2
, b
3
), (a
1
, a
1
, a
2
, a
2
)(b
1
, b
2
, b
3
, b
3
),
and so on.
If a crossing-free voice leading contains the pair (a
i
, b
j
), with i, j > 1, then it must
contain at least one of the following: (a
i-1
, b
j
), (a
i
, b
j-1
), or (a
i-1
, b
j-1
). By the recursion
constraint, the smallest voice leading of the form [a
1
, ..., a
i
][b
1
, ..., b
j
] will be the voice
leading that adds the pair (a
i
, b
j
) to the smallest voice leading of the form [a
1
, ..., a
i-1
]
[b
1
, ..., b
j
], [a
1
, ..., a
i
][b
1
, ..., b
j-1
], or [a
1
, ..., a
i-1
][b
1
, ..., b
j-1
]. Thus, once we have fixed
the pair (a
1
, b
1
) we can recursively compute the smallest voice leading between A and B
containing that pair. We do this by creating a matrix whose entries e
i, j
record the size of
the minimal voice leading of the form [a
1
, ..., a
i
][b
1
, ..., b
j
]. It is trivial to fill in the first
row and column of the matrix. For any other entry e
i, j
, we simply add the voice (a
i
, b
j
) to
the smallest of the voice leadings in the entrys upper, left, and upper-left neighbors.
Figure S12 illustrates the technique, identifying the smallest voice leading
between the C and E major-seventh chords, {4, 7, 11, 0} and {4, 8, 11, 3}, such that the
voice leading contains the pair (4, 4). In constructing this matrix I have used the L
1

taxicab norm to measure voice-leading size. The voice leading in the bottom-right
entry, (4, 4, 7, 11, 0)(3, 4, 8, 11, 11), is one of the minimal voice leadings between the
two chords that contains (4, 4). To remove this last restriction, we would need to repeat
the calculation three more times, each time cyclically permuting the order of one of the
chords so as to fix a different initial pair. As it happens, however, the voice leading
shown in Figure S12 is minimal. This follows from the fact that the mapping in the top-
left position, (4, 4), contributes nothing to the overall size of the voice leading; we can
therefore add it to any voice leading without increasing its L
1
size.
Figure S12 includes in each entry both the numerical size of the voice leading and
the voice leading itself. With the L
1
norm this is unnecessary: we need to keep track of
the size, but not the voice leading. To determine the value of entry e
i, j
we can simply add
the distance between the pair (a
i
, b
j
) to the minimum value in the entries e
i-1, j
,
e
i, j-1
, and e
i-1, j-1
. (For the other L
p
norms we can calculate the p
th
power of the voice-
leading size in this way, taking the p
th
root before output.) Having filled in the matrix, we
can recover a minimal voice leading between the two chords by tracing back a path that
moves from the bottom-right entry to the top left, moving only up, left, and diagonally
up-and-left, such that the size of the voice leading decreases as much as possible with
each step. The entries in boldface indicate the path such a traceback algorithm would
take. Due to the circular structure of pitch-class space, the voice leading in the lower
right-hand corner of the matrix counts the pair (a
1
, b
1
) = (a
m+1
, b
n+1
) twice; this can easily
be corrected prior to output. The resulting algorithm is easy to implement and suitable
for time-critical applications such as interactive computer music.
15





G















D
















A
















E














B
/
C
f
F s / G f C
s
/
D
f










A
f















E
f
















B
f














F




.
.
.
.
.
.
.
.
.
.
.
.
C
{5

6
}
{
t
e
}
{
3

4
}
{
8

9
}
{
1

2
}
{
6

7
}
{
e

0
}{
4

5
}
{
2

3
}
{
7

8
}
{
0

1
}

{
9

t
}
{024579e}
{
0
2
4
6
7
9
e
}
{
1
2
4
6
7
9
e
}
{
1
2
4
6
8
9
e
}
{
1
3
4
6
8
9
e
}
{
1
3
4
6
8
t
e
}
{ 1 3 5 6 8 t e }
{
1
3
5
6
8
t
0
}
{
1
3
5
7
8
t
0
}
{
2
3
5
7
8
t
0
}
{
2
3
5
7
9
t
0
}
{
2
4
5
7
9
t
0
}

Figure S1. The circle of fifths depicts minimal voice leadings between diatonic collections
(major scales). Each diatonic collection can be transformed into its neighbors by moving
one pitch class by one semitone. For example, the C major scale can be transformed into the
G major scale by moving the pitch class 5 (F) to 6 (Fs). Here and elsewhere, the letters t
and e refer to 10 (Bf) and 11 (B), respectively.

16

C c
f F
a A
[e] [E]
e E
[g] [G]
d D
fs
gs
Fs
[as]
[cs]
cs
[Cs]
Af
Ef
bf
g
G
B
ds
b
Bf
[D]
[Fs]
[Bf]
[Bf]
[Bf]
Df

Figure S2. The Tonnetz. Nineteenth-century theorists such as Hostinsky, Oettingen, and
Riemann explored a graph that is the geometric dual of the one shown here. The graph
displays efficient voice leadings among the 24 familiar major and minor triads. Uppercase
and lowercase letters indicate major and minor triads, respectively. Triads connected by
horizontal lines share both root and fifth, and can be linked by voice leading in which one
note moves by one semitone. (For example, the C major triad can be transformed into a C
minor triad by changing E to Ef.) Triads connected by a NE/SW diagonal also share two
notes and can be linked by single-semitone voice leading. (For example, the C major triad
can be transformed into an E minor triad by changing C to B.) Triads connected by a
NW/SE diagonal share two notes and can be linked by voice leading in which one note
moves by two semitones. (For example, the C major triad can be transformed into an A
minor triad by changing G to A.)

1
















2

















3

















4

















5
6
7
















8

















9

















1
0
















1
1
0
0.17
2.5
C C
s
/
D
f














D

















D
s

/
E
f















E





















F
F s / G f G















G
s

/
A
f














A















A
s

/
B
f
















B


Figure S3. The quotient space R/12Z is a circle whose circumference is twelve units long.
The twelve familiar pitch classes of Western equal-temperament evenly divide this circle.
Since the circle is continuous, it contains a point for every conceivable pitch class. The
figure shows the locations of the pitch class 0.17, which is seventeen cents (hundredths of a
semitone) above pitch class C, and pitch class 2.5 (D quarter tone sharp), which is halfway
between D and Ef.
17

(C4, C4) (D4, D4) (E4, E4) (F4, F4) (Fs4, Fs4) (Cs4, Cs4) (Ef4, Ef4)
(D5, D4) (E5, E4) (F5, F4) (Fs5 Fs4) (Cs5, Cs4) (Ef5, Ef4)
(D4, D5) (E4, E5) (F4, F5) (Fs4, Fs5) (Cs4, Cs5) (Ef4, Ef5) (Fs3, Fs4) (Gs3 Gs4) (Bf3, Bf4) (B3, B4) (C4, C5) (G3, G4) (A3, A4)
(Gs4 Gs4) (Bf4, Bf4) (B4, B4) (C5, C5) (G4, G4) (A4, A4)
(Fs4, Fs3) (Gs4 Gs3) (Bf4, Bf3) (B4, B3) (C5, C4) (G4, G3) (A4, A3)
(C4, Cs4)
(B3, Cs4)
(D4, Ef4) (E4, F4) (F4, Fs4) (Cs4, D4) (Ef4, E4)
(C5, Cs4) (D5, Ef4) (E5, F4) (F5, Fs4) (Cs5, D4) (Ef5, E4)
(Cs4, C5) (Ef4, D5) (F4, E5) (Fs4, F5) (D4, Cs5) (E4, Ef5)
(Cs4, C4) (Ef4, D4) (F4, E4) (Fs4, F4) (D4, Cs4) (E4, Ef4)
(C4, D4) (D4, E4) (E4, Fs4) (F4, G4) (Cs4, Ef4) (Ef4, F4)
(B4, Cs4) (C5, D4) (D5, E4) (E5, Fs4) (Cs5, Ef4) (Ef5, F4)
(D4, C5) (E4, D5) (Fs4, E5) (G4, F5) (Ef4, Cs5) (F4, Ef5)
(Fs4, Gs4) (G4, A4) (A4, B4) (B4, Cs5) (Gs4, As4) (Bf4, C5)
(F4, G3) (Fs4, Gs3) (G4, A3) (A4, B3) (B4, Cs4) (Gs4, As3) (Bf4, C4)
(B3, D4) (C4, Ef4) (D4, F4) (E4, G4) (Cs4, E4) (Ef4, Gf4)
(B4, D4) (C5, Ef4) (D5, F4) (E5, G4) (Cs5, E4) (Ef5, Gf4)
(D4, B4) (Ef4, C5) (F4, D5) (G4, E5) (E4, Cs5) (Gf4, Ef5)
(D4, B3) (Ef4, C4) (F4, D4) (G4, E4) (E4, Cs4) (Gf4, Ef4)
(F4, Af4) (Fs4, A4) (Gs4, B4) (As4, Cs5) (G4, Bf4) (A4, C5)
(F4, Af3) (Fs4, A3) (Gs4, B3) (As4, Cs4) (G4, Bf3) (A4, C4)
(Bf3, D4) (B3, Ds4) (Df4, F4) (Ef4, G4) (E4, Gs4) (C4, E4) (D4, Fs4)
(B4, Ds4) (Df5, F4) (Ef5, G4) (E5, Gs4) (C5, E4) (D5, Fs4)
(Ds4, B4) (F4, Df5) (G4, Ef5) (Gs4, E5) (E4, C5) (Fs4, D5)
(Fs4, As4) (Af4, C5) (Bf4, D5) (A4, Cs5) (F4, A4) (G4, B4)
(E4, Gs3) (Fs4, As3) (Af4, C4) (Bf4, D4) (A4, Cs4) (F4, A3) (G4, B3)
(Bf3, E4) (A3, Ds4) (B3, F4) (Df4, G4) (Ef4, A4) (C4, Fs4) (D4, Gs4)
(Bf4, E4) (B4, F4) (Df5, G4) (Ef5 A4) (C5, Fs4) (D5, Gs4)
(F4, B4) (E4, Bf4) (Fs4, C5) (Gs4, D5) (G4, Cs5) (A4, Ds5)
(Ef4, A3) (F4, B3) (E4, Bf3) (Fs4, C4) (Gs4, D4) (G4, Cs4) (A4, Ds4)
(Bf3, Gf4) (A3, F4) (B3, G4) (Cs4, A4) (Gs3, E4) (C4, Af4)
(Bf3, G4) (A3, Fs4) (B3, Gs4) (Cs4, As4) (Af3, F4) (C4, A4)
(Bf4, G4) (A4, Fs4) (B4, Gs4) (Cs5, As4) (Af4, F4) (C5, A4)
(D4, Bf4)
(Bf4, Gf4) (A4, F4) (B4, G4) (Cs5, A4) (C5, Af4) (D5, Bf4) (Ds4, B3) (F4, Df4) (G4, Ef4) (Gs4, E4) (E4, C4) (Fs4, D4) (D4, Bf3)
(Bf3, Af4) (A3, G4) (B3, A4) (Cs4, B4)
(D4, C4) (E4, D4) (Fs4, E4) (G4, F4) (Ef4, Cs4) (F4, Ef4) (Cs4, B3)
(Af3, Gf4) (C4, Bf4)
(Bf4, Af4) (A4, G4) (B4, A4) (Cs5, B4) (Af4, Gf4) (C5, Bf4)
(G3, F4)
(Bf3, A4) (A3, Gs4) (B3, As4) (Af3, G4) (C4, B4) (G3, Fs4)
(Bf4, A4) (A4, Gs4) (B4, As4) (Af4, G4) (C5, B4) (G4, Fs4)
(Bf4, B4) (A4, Bf4) (B4, C5) (Af4, A4) (G4, Af4) (Fs4, G4)
(Bf4, B3) (A4, Bf3) (B4, C4) (Af4, A3) (G4, Af3) (Fs4, G3)
(Bf3, F4) (A3, E4) (B3, Fs4) (Df4, Af4) (C4, G4) (D4, A4)
(Bf4, F4) (A4, E4) (B4, Fs4) (Df5, Af4) (C5, G4) (D5, A4)
(Bf3, Ef4) (B3, E4) (Df4, Gf4) (Ef4, Af4) (C4, F4) (D4, G4)
(Bf4, Ef4) (B4, E4) (Df5, Gf4) (Ef5, Af4) (C5, F4) (D5, G4)
(Ef4, Bf4) (E4, B4) (Gf4, Df5) (Af4, Ef5) (F4, C5) (G4, D5)
(Ef4, Bf3) (E4, B3) (Gf4, Df4) (Af4, Ef4) (F4, C4) (G4, D4)
(Fs4, B4) (G4, C5) (A4, D5) (Gs4, Cs5) (E4, A4) (F4, Bf4)
(Fs4, B3) (G4, C4) (A4, D4) (Gs4, Cs4) (E4, A3) (F4, Bf3)
P
P
P
P
(a)
(b)

Figure S4. (a) A portion of the infinite plane representing ordered two-note chords of
pitches. The four quadrants are equivalent to within octave displacement and permutation.
Each is equivalent to Figure 2 in the main paper. (b) An abstract representation of the
symmetries relating the four quadrants. The lower-left quadrant is related to the upper-left
by a reflection that preserves their common border. This action permutes each dyad.
Translation of any quadrant diagonally up and right transposes the first element of each
dyad by an ascending octave. Translation of any quadrant diagonally down and right
transposes the second element of each dyad by an ascending octave. These operations
suffice to generate the infinite, periodic figure. The space can be described, metaphorically,
as wallpaper; Figure 2 in the main paper provides the pattern, Figure S4(b) shows how the
pattern is to be assembled, and Figure S4(a) shows a portion of the result. The diamond in
the center of Fig. S4(a) corresponds to the 2-torus shown in Figure S9.
18



Figure S5. The orbifold T
3
/S
3
is a prism whose two triangular faces are identified by way
of a 120 rotation. Several familiar equal-tempered chords are depicted on the figure.
Augmented triads, which divide the octave into three equal parts, are shown as dark cubes.
Minor chords are light spheres and major chords are dark spheres. Lines connecting
augmented, minor, and major chords indicate that they can be linked by voice leading in
which a single voice moves by a single semitone. Since minor and major chords divide the
octave nearly evenly, they are clustered near the center of the orbifold. Triple unisons,
which contain only one pitch class, are found on the edge of the figure.

19

(a) (b) (c)

CCC
CCC
EEE
EEE
GsGsGs
CEGs
CEGs
GsGsGs
transposition
CC
CFs CFs
CC
FsFs
FsFs
CCCC
AAAA
EfEfEfEf
FsFsFsFs
EfEfEfEf
FsFsFsFs
AAAA
CCCC
t
r
a
n
s
p
o
s
i
t
i
o
n
t
r
a
n
s
p
o
s
i
t
i
o
n

Figure S6. (a) The orbifold T
2
/S
2
is a two-dimensional prism (parallelogram) whose base (a
line segment) is glued to the opposite face. Before gluing, the base must be twisted so that
chords on the left edge match those on the right. This twist is a reflection that can be
represented as a rotation in three Euclidean dimensions. The line at the center of the figure
contains chords that divide the octave evenly. (b) The orbifold T
3
/S
3
is a three-dimensional
prism whose two triangular faces are glued together. Before gluing, rotate one face by 120,
so that the chords match. The result is the bounded interior of a twisted triangular 2-torus.
Augmented triads, which divide the octave into three equal parts, lie on the line at the center
of the figure. Major and minor chords are close to this line, as shown in Figure S5.
Rotating the prism around the central line by 120 transposes every chord by major third.
(c) The orbifold T
4
/S
4
is a four-dimensional prism whose two tetrahedral faces are glued
together. The dashed lines extend into the fourth dimension. Before identifying the two
faces, twist one so that the chords match. The twist is a reflection, as in the two-
dimensional case. Diminished seventh chords, which divide the octave into four equal
pieces, lie at the center of the orbifold. Familiar four-note tonal chords lie close to this
chord.



20

(a) (b)

p
1
p
2
r r
p
1
< p
2
, q
1
> q
2
q
1
q
2
p
1
p
2
q
1
q
2


Figure S7. (a) | r p
2
| / | r p
1
| = | q
2
r | / | q
1
r |. (b) Any line segment that crosses (p
1
, q
2
)
crosses either (p
1
, q
1
) or (p
2
, q
2
).

(a) (b)
p
p

+ x
1
x
2
{x
1
+ c, x
2
} {x
1
+ c, x
2
, x
3
}
{x
1
, x
2
, x
3
} {x
1
, x
2
+ c}
p

+ x
1
+ c
p

+ x
1
p
p

+ x
1
+ x
2
p

+ x
1
+ x
3
+ c
p

+ x
1
+ c
p

+ x
1
p
p

+ x
1
+ x
2
p

+ x
1
+ x
3
+ c
p

+ x
1
+ c
p

+ x
1
p
p

+ x
1
x
2
x
1
> x
2
, c > 0 x
3
x
2
> 0, c > 0
p

+ x
1
+ c
p

+ x
1


Figure S8. For any violation of the distribution constraint, we can find a crossed voice
leading that is smaller than its natural uncrossed alternative. The crossed voice leading at
the bottom of each column is smaller than the uncrossed voice leading above it.
21

(a) (b)
10 00 20 30 40 50 60 70 80 90 t0 e0 [00]
11 01 21 31 41 51 61 71 81 91 t1 e1 [01]
22 12 02
23 13 03
32 42 52 62 72 82 92 t2 e2 [02]
33
24 14 04 34
43 53 63 73 83 93 t3 e3 [03]
44
25 15 05 35 45
54 64 74 84 94 t4 e4 [04]
55
26 16 06 36 46 56
65 75 85 95 t5 e5 [05]
66
27 17 07 37 47 57 67
76 86 96 t6 e6 [06]
77
28 18 08 38 48 58 68 78
87 97 t7 e7 [07]
88
29 19 09 39 49 59 69 79 89
98 t8 e8 [08]
99
2t 1t 0t 3t 4t 5t 6t 7t 8t 9t
t9 e9 [09]
tt
2e 1e 0e 3e 4e 5e 6e 7e 8e 9e te
et [0t]
ee
[20] [10] [00] [30] [40] [50] [60] [70] [80] [90] [t0] [e0]
[0e]
[00]
B
A
01 00 02 03 04 05 06 07 08 09 0t 0e [00]
11 12 13 14 15 16 17 18 19 1t 1e [10]
22 23 24 25 26 27 28 29 2t 2e [20]
33 34 35 36 37 38 39 3t 3e [30]
44 45 46 47 48 49 4t 4e [40]
55 56 57 58 59 5t 5e [50]
66 67 68 69 6t 6e [60]
77 78 79 7t 7e [70]
88 89 8t 8e [80]
99 9t 9e [90]
tt te [t0]
ee [e0]
[00]
A
D
B
C


Figure S9. (a) The space of ordered two-note chords of pitch-classes is a 2-torus. To
identify points (x, y) and (y, x), we need to fold the torus along the AB diagonal. The
resulting figure, shown in (b), is a triangle with two of its sides identified. This is a Mbius
strip. To see why, cut figure (b) along the line CD and glue AC to CB. (To make this
identification in Euclidean 3-space, you will need to turn over one of the pieces of paper.)
The result is a square with opposite sides identified, as in Figure 2 of the main paper.
22


.... .
0
3
(a) the dominant-seventh chord {0, 3, 6, 8}
(b) the minor-seventh chord {0, 3, 5, 8}
(c)
3 2 4
[0] 3 6 8
.... .
0
3 2 3 4
[0] 3 5 8
.... .
a
n-2
a
n-1
x + d
n-2
x + d
n-1
x + d
0
x + d
1
a
2
a
0 a
1
.... .
a
0
2x
x
a
0
+ 2x a
0
x a
0
a
0
+ x
x x x
||

d
0
||
12Z
||

d
n-1
||
12Z
|| d
0
+ d
1
||
12Z
|| d
n-2
+ d
n-1
||
12Z
0


Figure S10. (ab) The dominant-seventh chord {0, 3, 6, 8} and minor-seventh chord
{0, 3, 5, 8} can both be arranged as cycles whose notes are 2, 3, 3, and 4 semitones apart.
To transform the dominant-seventh chord into a chord that divides the octave evenly, one
need only move pitch class 8 to pitch class 9. However, to transform the minor-seventh
chord into a chord that divides the octave evenly, one must move at least two pitch classes.
(For example, one can move 5 and 8 to 6 and 9, respectively.) Consequently, although both
chords have equally small minimal voice leadings to their T
3
forms, the first is closer to the
nearest T
3
-invariant chord according to many metrics. (c) To find a bijective voice leading
from any chord to a chord whose intervals are all equal to x, fix a
0
, move note a
1
by d
0

semitones, note a
n-1
by d
n-1
semitones, note a
2
by d
0
+ d
1
semitones, and so on. Since pitch-
class space is circular, the term d

n/2

will not be involved in this voice leading.

23

(a) (b)


1
















2

















3

















4

















5

6


7
















8

















9

















1
0
















1
1
0
C C
s
/
D
f














D

















D
s

/
E
f















E





















F
F s / G f G















G
s

/
A
f














A















A
s

/
B
f
















B
1





















1

















1








1













1























1












1
C












D


























E


























F
G

























A


























B


Figure S11. For any scale, we can define a metric such that the scales notes divide the
octave evenly. (a) shows the C major scale, as it appears in circular pitch class space. In
(b), we apply a new metric, so that the scales notes are equally spaced. The new unit of
distance is called a scale step. Relative to this new metric, the C major triad {C, E, G}
and the D minor triad {D, F, A} are related by rotation, which musicians call scalar
transposition.

4 8 11 3 4
4 (4)(4)

Size: 0
(4, 4)(4, 8)

Size: 4
(4, 4, 4)
(4, 8, 11)
Size: 9
(4, 4, 4, 4)
(4, 8, 11, 3)
Size: 10
(4, 4, 4, 4, 4)
(4, 8, 11, 3, 4)
Size: 10
7 (4, 7)(4, 4)

Size: 3
(4, 7)(4, 8)

Size: 1
(4, 7, 7)
(4, 8, 11)
Size: 5
(4, 7, 7, 7)
(4, 8, 11, 3)
Size: 9
(4, 7, 7, 7, 7)
(4, 8, 11, 3, 4)
Size: 12
11 (4, 7, 11)
(4, 4, 4)
Size: 8
(4, 7, 11)
(4, 8, 8)
Size: 4
(4, 7, 11)
(4, 8, 11)
Size: 1
(4, 7, 11, 11)
(4, 8, 11, 3)
Size: 5
(4, 7, 11, 11, 11)
(4, 8, 11, 3, 4)
Size: 10
0 (4, 7, 11, 0)
(4, 4, 4, 4)
Size: 12
(4, 7, 11, 0)
(4, 8, 8, 8)
Size: 8
(4, 7, 11, 0)
(4, 8, 11, 11)
Size: 2
(4, 7, 11, 0)
(4, 8, 11, 3)
Size: 4
(4, 7, 11, 0, 0)
(4, 8, 11, 3, 4)
Size: 8
4 (4, 7, 11, 0, 4)
(4, 4, 4, 4, 4)
Size: 12
(4, 7, 11, 0, 4)
(4, 8, 8, 8, 8)
Size: 12
(4, 7, 11, 0, 4)
(4, 8, 11, 11, 11)
Size: 7
(4, 7, 11, 0, 4)
(4, 8, 11, 11, 3)
Size: 3
(4, 7, 11, 0, 4, 4)
(4, 8, 11, 11, 3, 4)
Size: 3

Figure S12. Using dynamic programming to find a minimal voice leading between
{4, 7, 0, 11} and {4, 8, 11, 3}.

24

SYMBOL OR TERM DEFINITION
multiset,
object,
member,
element
A multiset is an unordered collection in which duplications are
permitted. {0, 1, 1} is the same multiset as {1, 0, 1} but is different
from {0, 1}. The multiset {0, 1, 1} contains three objects and has
three members. However, it has only two elements, 0 and 1.
{a, b, c} The multiset with members a, b, c, some of which may be identical.
(a, b, c) An ordered list. (a, b, c) and (b, c, a) are distinct.
i I Object i belongs to set or multiset I.
group
A group is a set whose elements can be combined so as to satisfy
certain axioms. See a group theory textbook for details.
R The real numbers.
R
n

The set of ordered n-tuples (x
1
, x
2
, ... x
n
) such that each x
i
R.
Z The integers.
nZ, where n R The set {nk | k Z}. Thus 12Z is the set {..., 24, 12, 0, 12, 24, ...}.
The elements of this set form a group under addition.
mZ
n
, where m R and n Z
The set of ordered n-tuples (x
1
, x
2
, ... x
n
) such that each x
i
mZ.
These n-tuples form a group under vector addition.
S
n
The symmetric group of degree n, consisting of the group of
permutations of n objects.
quotient space A quotient space is formed by identifying (or gluing together) points
in another space.
A/G, where G is some group of
transformations acting on A
The quotient space that identifies all points a and ga, where a A
and g G.
R/12Z The circular quotient space in which all real numbers x and x + 12
are identified. Points in this space are infinite sets of the form
{..., x 24, x 12, x, x + 12, x + 24, ...}, where x R. These sets can
be labeled using real numbers in the range 0 x < 12. The elements
of R/12Z form a group under addition of their labels modulo 12Z.
T
n
The n-torus, or the Cartesian product of n circles. Since R/12Z is a
circle, (R/12Z)
n
is an n-torus.
topological equivalence Two spaces are topologically equivalent if one can be continuously
deformed into the other.
manifold A space that is locally topologically equivalent to R
n
.
global-quotient orbifold A global-quotient orbifold (or orbit manifold) is a quotient space M/G,
where M is a manifold and G is a group acting discontinuously on M.
Global-quotient orbifolds inherit many geometrical properties from
their parent spaces.

Table S1, part 1. A glossary of mathematical terms and symbols.
25

singularity, singular locus A point in an orbifold at which the local geometry or topology is not
that of the corresponding point in the parent space. If M/G is a global-
quotient orbifold, then its singularities are fixed points of the group G.
The singularities of the orbifolds T
n
/S
n
act like mirrors.
boundary
(of the orbifolds T
n
/S
n
)
The singularities of the orbifolds T
n
/S
n
, which enclose the nonsingular
points (see Figs. 2 and S6).
simplex, simplicial A simplex is a generalized triangle. The 1-simplex is a line segment,
the 2-simplex is a triangle, the 3-simplex is a tetrahedron, and so forth.
prism,
height coordinate of the prism,
base
In this paper, the term prism refers to the n-dimensional region
formed when an (n1)-dimensional polyhedron (the base) is dragged
along a line segment extending into the remaining dimension. The
direction of the dragging is the height coordinate of the prism.
||b a||
12Z
The distance between two points of R/12Z, equal to the smallest
nonnegative real number x such that, if p belongs to a, then either
p + x or p x belongs to the set b.
a b modulo nZ The real numbers a and b are congruent modulo nZ: that is, there
exists an integer k such that a = b + kn.
a - b a is approximately equal to b.
iff If and only if.
total preorder
reflexive relation
transitive relation
total relation
A total preorder is a relation > over set A that is reflexive (a > a, for
all a e A), transitive (a > b and b > c imply a > c, for all a, b, c e A)
and total (either a > b or b > a, or both, for all a, b e A).
a < b means a s b but not b s a. a b means a s b and b s a.
Note that a b does not imply that a and b are identical.
fundamental domain A fundamental domain of the group G acting on the space S is a region
R that tiles S under the action of the group: S is the union of the
regions gR, for all g e G, and any two regions gR and hR, for g = h,
intersect only at their boundaries.

semidirect product
Let G be a group with subgroup F and normal subgroup N. G is the
semidirect product N F iff every element in G can be written in one
and only one way as the product fn, with f e F and n e N. See a
group theory textbook for details.
involution A function F is an involution iff F(F(a)) = a.
isometry A function F is an isometry iff the distance between F(a) and F(b) is
equal to the distance between a and b, for all a and b.

x
]
and

x
1
The greatest integer s x and the smallest integer > x, respectively.

Table S1, part 2. A glossary of mathematical terms and symbols.
26


pitch Pitch is the perceptual correlate of fundamental frequency. Pitches
can be modeled as real numbers such that middle C is 60, the
octave has size 12, and semitones have size 1.
pitch class A set consisting of all pitches separated by an integral number of
octaves. A220 and A440 both belong to the pitch class A. Pitch
classes can be modeled as elements of the quotient space R/12Z.
chord A multiset of either pitches or pitch classes.
transposition,
translation
In both pitch and pitch-class space, transposition (or translation)
corresponds to addition by a constant value. If a is a pitch or pitch
class then a + x is the transposition of a by x semitones.
T
x
(A) The transposition of the chord A by x semitones.
transpositionally related Two chords are transpositionally related if one is the transposition
of the other.
inversion,
reflection
In both pitch and pitch-class space, inversion (reflection)
corresponds to subtraction from a constant value. If a is a pitch or
pitch class then x a is the inversion of a with index number x.
I
x
(A) The inversion that maps a to x a. I
x
maps 0 to x and vice versa.
inversionally related Two chords are inversionally related if one is the inversion of the
other.
voice leading A voice leading between two multisets {a
1
, a
2
, ..., a
m
} and
{b
1
, b
2
, ..., b
n
} is a multiset of ordered pairs (a
i
, b
j
), such that every
member of each multiset is in some pair.
(a
1
, a
2
, ..., a
n
)(b
1
, b
2
, ..., b
n
) The voice leading that contains all and only the pairs (a
i
, b
i
). Voice
leadings do not uniquely determine the chords they connect. For
example, the voice leading (0, 0, 4, 7)(11, 2, 5, 7) is both a
bijective voice leading from {0, 0, 4, 7} to {11, 2, 5, 7} and a non-
bijective voice leading from {0, 4, 7} to {11, 2, 5, 7}.
distance The distance between two pitches p and q is the absolute value of
their difference, |q p|. The distance between two pitch classes a
and b, written ||b a||
12Z
, is the smallest nonnegative real number x
such that, if p is a pitch belonging to pitch class a, then either p + x
or p x belongs to pitch class b.
ascending distance The ascending distance from pitch class a to pitch class b is the
smallest nonnegative real number x such that, if p is a pitch
belonging to pitch class a, then p + x belongs to pitch class b.

Table S1, part 3. A glossary of musical terms and symbols.
27


associated voice leading (p
1
, p
2
, ..., p
n
)(q
1
, q
2
, ..., q
n
) is associated with (a
1
, a
2
, ..., a
n
)
(b
1
, b
2
, ..., b
n
) if pitches p
i
and q
i
all belong to the pitch classes a
i
and b
i
,
respectively.
displacement multiset The multiset of distances between notes in the source chord and their
images in the target chord.
distribution constraint {x
1
+ c, x
2
, ..., x
n
} {x
1
, x
2
+ c, ..., x
n
} {x
1
, x
2
, ..., x
n
} for x
1
> x
2
and
c > 0
recursion constraint {x
1
, x
2
, ..., x
m
} {y
1
, y
2
, ..., y
n
} implies {x
1
, x
2
, ..., x
m
, c}
{y
1
, y
2
, ..., y
n
, c}
trivial voice leading A voice leading containing only pairs of the form (x, x).
bijective voice leading A voice leading from A to B such that every member of A is mapped to
exactly one member of B, and vice versa.
independent voice leading A voice leading that cannot be written in the form (a
1
, a
2
, ..., a
n
)
(a
1
+ x, a
2
+ x, ..., a
n
+ x).
parallel voice leading A voice leading that is not independent, and hence acts as a
transposition.
crossing free,
strongly crossing free
Intuitively, a voice leading is crossing free if the notes of the source
chord can be connected to those of the target along minimal-length
line-segments intersecting only at their endpoints. A voice leading
between multisets of pitches (p
1
, p
2
, ..., p
n
)(q
1
, q
2
, ..., q
n
) has no voice
crossings if p
i
> p
j
implies q
i
q
j
, for all i, j n. It is strongly crossing
free if (p
1
+ 12k
1
, p
2
+ 12k
2
, ..., p
n
+ 12k
n
)(q
1
+ 12k
1
, q
2
+ 12k
2
, ..., q
n
+ 12k
n
)
is crossing free, for all integers k
i
. Thus one cannot introduce crossings
into a strongly crossing-free voice leading simply by shifting the octave
in which its voices appear. A voice leading between multisets of pitch
classes is crossing free if it is associated with a strongly crossing-free
voice leading between multisets of pitches in which no voice moves by
more than six semitones.
voice crossing Two voices (p
1
, q
1
) and (p
2
, q
2
) cross if the voice leading (p
1
, p
2
)
(q
1
, q
2
) is not crossing free.
maximally even Let {e
0
, e
1
, ..., e
n-1
} be a subset of R/kZ that divides the octave into n
precisely even parts. {m
0
, m
1
, ..., m
n-1
} is maximally even if m
i
=

e
i
.
For any maximally even chord {m
0
, m
1
, ..., m
n-1
}, and any integer d, the
set {m
i + d (mod n)
m
i
} will contain either a single integer-valued point in
R/kZ or a pair of adjacent integer-valued points in R/kZ. The n-note
maximally even subsets of a chromatic scale are related by
transposition.

Table S1, part 4. A glossary of musical terms and symbols.
28

REFERENCES

1. R. Gauldin, Harmonic Practice in Tonal Music (Norton, New York, 2004).
2. S. Laitz, The Complete Musician (Oxford, New York, 2003).
3. D. Huron, Music Perception 19, 1 (2001).
4. It is possible to show that the triangle inequality implies the distribution constraint, given
a few minimal assumptions. Let P, Q, and R be points in R
n
, representing ordered series of
pitches, and let PQ be the voice leading that maps the i
th
component of P to the i
th

component of Q. Suppose our method of comparing voice leadings is a metric that assigns a
(real-valued) size to every displacement multiset. Let D(PR) refer to the size of the
displacement multiset associated with the voice leading PR. Suppose that D(PR)
D(PQ) + D(QR), for all P, Q, and R (the triangle inequality), and that D(PR) =
D(PQ) + D(QR) whenever Q lies on the line segment PR. It follows that the
distribution constraints first inequality will be satisfied.
5. D. Lewin, Journal of Music Theory 42, 15 (1998).
6. R. Cohn, Journal of Music Theory 42, 283 (1998).
7. J. Straus, Music Theory Spectrum 25, 305 (2003).
8. C. Callender, Music Theory Online 10 (2004).
9. R. Cohn, Journal of Music Theory 41, 1 (1997).
10. J. Douthett, P. Steinbach, Journal of Music Theory 42, 241 (1998).
11. http://music.princeton.edu/~dmitri/ChordGeometries.html.
12. Consequently, for all of the standard voice-leading metrics, we can always find minimal
bijective voice leadings from A to I
x
(A) and A to S
I
x
with displacement multisets equal to
{d
1
, d
2
, ..., d
n
} and {d
1
/2, d
2
/2, ..., d
n
/2}.
13. J. Clough, J. Douthett, Journal of Music Theory 35, 93 (1991).
14. Let {e
0
, e
1
, ..., e
n-1
} be a subset of R/kZ that divides the octave into n precisely even
parts. {m
0
, m
1
, ..., m
n-1
} is maximally even if m
i
=

e
i
. For any maximally even chord
{m
0
, m
1
, ..., m
n-1
}, and any integer d, the set {m
i + d (mod n)
m
i
} will contain either a single
integer-valued point in R/kZ or a pair of adjacent integer-valued points in R/kZ. The n-note
maximally even subsets of a chromatic scale are related by transposition. In twelve-tone
equal temperament, the maximally even chords are {0, 6}, {0, 4, 8}, {0, 3, 6, 9}, {0, 2, 4, 5, 7},
{0, 2, 4, 6, 8, 10}, their transpositions, and the complements of these chords.
15. There is an integer d such that e
i+d (mod n)
+ x = e
i
+ c. So n||c||
kZ
= ||e
i+d (mod n)
+ x e
i
||
kZ
=
||

e
i+d (mod n)
+ x

e
i
||
kZ
= ||

e
i
+ c

e
i
||
kZ
, where we use the fact that ||c||
kZ
k/2n.
16. E. Agmon, Musikometrica 3, 15 (1991).
17. Note that voice leadings do not uniquely determine the chords they connect: the voice
leading (0, 0, 4, 6)(10, 0, 6, 6) is both a non-bijective voice leading between {0, 4, 6} and
{6, 10, 0} and a bijective voice leading between {0, 0, 4, 6} and {6, 6, 10, 0}.