Three Conceptions of Musical Distance

Dmitri Tymoczko
310 Woolworth Center, Princeton University, Princeton, NJ 08544.

Abstract: This paper considers three conceptions of musical distance (or

inverse similarity) that produce three different musico-geometrical spaces:
the first, based on voice leading, yields a collection of continuous quotient
spaces or orbifolds; the second, based on acoustics, gives rise to the Tonnetz
and related tuning lattices; while the third, based on the total interval content
of a group of notes, generates a six-dimensional quality space first described
by Ian Quinn. I will show that although these three measures are in principle
quite distinct, they are in practice surprisingly interrelated. This poses the
challenge of determining which model is appropriate to a given musictheoretical circumstance. Since the different models can produce comparable
results, unwary theorists could potentially find themselves using one type of
structure (such as a tuning lattice) to investigate properties more perspicuously
represented by another (for instance, voice-leading relationships).
Keywords: Voice leading, orbifold, tuning lattice, Tonnetz, Fourier transform.

1 Introduction
We begin with voice-leading spaces that make use of the log-frequency metric.1
Pitches here are represented by the logarithms of their fundamental frequencies, with
distance measured according to the usual metric on R; pitches are therefore close if
they are near each other on the piano keyboard. A point in Rn represents an ordered
series of pitch classes. Distance in this higher-dimensional space can be interpreted
as the aggregate distance moved by a collection of musical voices in passing from
one chord to another. (We can think of this, roughly, as the aggregate physical
distance traveled by the fingers on the piano keyboard.)
By disregarding
informationsuch as the octave or order of a group of noteswe fold Rn into an
non-Euclidean quotient space or orbifold.
(For example, imposing octave
equivalence transforms Rn into the n-torus Tn, while transpositional equivalence
transforms Rn into Rn1, orthogonally projecting points onto the hyperplane whose
coordinates sum to zero.) Points in the resulting orbifolds represent equivalence
classes of musical objectssuch as chords or set classeswhile generalized line

For more on these spaces, see Callender 2004, Tymoczko 2006, and Callender, Quinn, and
Tymoczko 2008.

segments represent equivalence classes of voice leadings.2 For example, Figure 1,

from Tymoczko 2006, represents the space of two-note chords, while Figure 2, from
Callender, Quinn, and Tymoczko 2008, represents the space of three-note
transpositional set classes. In both spaces, the distance between two points represents
the size of the smallest voice leading between the objects they represent.
CsCs

CC

CDf
BCs

CsD

DEf

CsDs

CD

CEf

BD
BfD

BDs

BfEf

EfEf

DD

DsE

BE

DfF
CF

EF

DF

FGf
EFs

EfGf
DFs

DfGf

[FsFs]

FF

EfF

DE

CsE

CE

EE

EG

EfG

DG

[FG]

[EGs]

EfAf

unison
minor second
major second
minor third
major third
perfect fourth

DAf [EfA] tritone

perfect fourth
DfAf
AE
BfF
BFs
DA
CG
major third
[DBf]
GsE
CsA
CAf
BfGf BG
AF
minor third
DfBf
BfG
AfF
CA
BGs
AFs
major second
AfGf
GF
BfAf
BA
AG
[CsB]
CBf
minor second
CB
BAs
AGs BfA
GFs AfG
unison
AA
GG
[CC]
BB
BfBf
AfAf
FsFs
AEf

BfE

BF

CFs

CsG

Fig. 1. The Mbius strip representing voice-leading relations among two-note chords.

Lets now turn to a very different sort of model, the Tonnetz and related structures,
which I will describe generically as tuning lattices. These models are typically
discrete, with adjacent points on a particular axis being separated by the same
interval. The leftmost lattice in Figure 3 shows the most familiar of these, where the
two axes represent acoustically pure perfect fifths and major thirds. (One can imagine
a third axis, representing either the octave or the acoustical seventh, projecting
outward from the paper.) The model asserts that the pitch G4 has an acoustic affinity
to both C4 (its underfifth) and D5 (its overfifth), as well as to Ef4 and B4 (its
underthird and overthird, respectively). The lattice thus encodes a fundamentally
different notion of musical distance than the earlier voice leading models: whereas A3
and Af3 are very close in log-frequency space, they are four steps apart our tuning
lattice. Furthermore, where chords (or more generally musical objects) are
represented by points in the voice leadings spaces, they are represented by polytopes
in the lattices.3
Finally, there are measures of musical distance that rely on chords shared interval
content. From this point of view, the chords {C, Cs, E, Fs} and {C, Df, Ef, G}
resemble one another, since they are nontrivially homometric or Z-related: that is,
they share the same collection of pairwise distances between their notes. (For
instance, both contain exactly one pair that is one semitone apart, exactly one pair that
is two semitones apart, and so on.) However, these chords are not particularly close
2

The adjective generalized indicates that these line segments may pass through one of the
spaces singular points, giving rise to mathematical complications.
3
For a modern introduction to the Tonnetz, see Cohn 1997, 1998, and 1999.

in either of the two models considered previously. It is not intuitively obvious that
this notion of similarity produces any particular geometrical space. But Ian Quinn
has shown that one can use the discrete Fourier transform to generate (in the familiar
equal-tempered case) a six-dimensional quality space in which chords that share the
same interval content are represented by the same point.4 We will explore the details
shortly.

Fig. 2. The cone representing voice-leading relations among three-note transpositional set
classes.
A3

E4

B4

Fs5

Cs6

A3

E4

B4

F5

C6

F3

C4

G4

D5

A5

F3

C4

G4

D5

A5

Df3

Af3

Ef4

Bf4

F5

D3

A3

E4

B4

F5

Fig. 3. Two discrete tuning lattices. On the left, the chromatic Tonnetz, where horizontally
adjacent notes are linked by acoustically pure fifths, while vertically adjacent notes are linked
by acoustically pure major thirds. On the right, a version of the structure that uses diatonic
intervals.

Clearly, these three musical models are very different, and it would be somewhat
surprising if there were to be close connections between them. But we will soon see
that this is in fact this case.

See Lewin 1959, 2001, Quinn 2006, 2007, Callender 2007.

EE

DD

G
{F

FF

CF

{BC}

CE

DG

BF

DA

CG
AF

F}
E
[CC]

{AB}

BB

AA

EB

CB
GG

BA

AG

{D

BG
CA

GF

DA

BF
AE

AE

BE

EG

A}
{G

[FG]
DF

BD

GD

DE

{C
D}

CG

FC

EF

CD

CC

Fig. 4. (left) most efficient voice-leadings between diatonic fifths form a chain that runs
through the center of the Mbius strip from Figure 1. (right) These voice leadings form an
abstract circle, in which adjacent dyads are related by three-step diatonic transposition, and are
linked by single-step voice leading.
C
{B

{G
A}

a
F}

{DE}

{E

G}

{C

{F

{AB}

Fig. 5. (left) most efficient voice-leadings between diatonic triads form a chain that runs
through the center of the orbifold representing three-note chords. (right) These voice leadings
form an abstract circle, in which adjacent triads are linked by single-step voice leading. Note
that here, adjacent triads are related by transposition by two diatonic steps.

2 Voice-leading lattices and acoustic affinity

Voice-leading and acoustics seem to privilege fundamentally different conceptions of
pitch distance: from a voice leading perspective, the semitone is smaller than the
perfect fifth, whereas from the acoustical perspective the perfect fifth is smaller than
the semitone. Intuitively, this would seem to be a fundamental gap that cannot be
bridged.

Fig. 6. Major, minor, and augmented triads as they appear in the orbifold representing threenote chords. Here, triads are particularly close to their major-third transpositions.

Things become somewhat more complicated, however, when we consider the

discrete lattices that represent voice-leading relationships among familiar diatonic or
chromatic chords. For example, Figure 4 records the most efficient voice leadings
among diatonic fifthswhich can be represented using an irregular, one-dimensional
zig-zag near the center of the Mbius strip T2/S2. (The zig-zag seems to be irregular
because the figure is drawn using the chromatic semitone as a unit; were we to use the
diatonic step, it would be regular.) Abstractly, these voice leadings form the circle
shown on the right of Figure 4. The figure demonstrates that there are purely
contrapuntal reasons to associate fifth-related diatonic fifths: from this perspective
{C, G} is close to {G, D}, not because of acoustics, but because the first dyad can be
transformed into the second by moving the note C up by one diatonic step. One
fascinating possibilitywhich we unfortunately cannot pursue hereis that acoustic
affinities actually derive from voice-leading facts: it is possible that the ear associates
the third harmonic of a complex tone with the second harmonic of another tone a fifth
above it, and the fourth harmonic of the lower note with the third of the upper, in
effect tracking voice-leading relationships among the partials.
Figures 5-7 present three analogous structures: Figure 5 connects triads in the C
diatonic scale by efficient voice leading, and depicts third-related triads as being
particularly close; Figure 6 shows the position of major, minor, and augmented triads
in three-note chromatic chord space, where major-third-related triads are close5;
Figure 7 shows (symbolically) that fifth-related diatonic scales are close in chromatic
space. Once again, we see that there are purely contrapuntal reasons to associate
fifth-related diatonic scales and third-related triads.

This graph was first discovered by Douthett and Steinbach (1998).

. . .
t0}

{024579e}

67
G 9e}
{0

4}

{3

1}

7}

{6

Af

t0}

68
E 9e}

{13

t}

. . .

78
135

3}

{9

{2

2}
{1

Ef

8}

{124689e}

{8
9}

Bf

{23
579
t0}

6}

{7

{23578t0}

e}
{t

{5

9e}
467
{12 D

{02

79

5
{24

{13
}
Fs/Gf
8te
5
46
Cs 68t0}
Cf
/D
{e
B/ {45}
f
0}

{13

{13568te}

Fig. 7. Fifth-related diatonic scales form a chain that runs through the center of the sevendimensional orbifold representing seven-note chords. It is structurally analogous to the circles
in Figures 4 and 5.
Correlation
Bach
MAJOR

MINOR

.96

Haydn

.93

Mozart

.91

Beethoven

.96

Bach

.95

Haydn

.91

Mozart

.91

Beethoven

.96

Fig. 8. Correlations between modulation frequency and voice-leading distances among scales,
in Bachs Well-Tempered Clavier, and the piano sonatas of Haydn, Mozart, and Beethoven.
The very high correlations suggest that composers typically modulate between keys whose
associated scales can be linked by efficient voice leading.

This observation, in turn, raises a number of theoretical questions. For instance:

should we attribute the prevalence of modulations between fifth-related keys to the
acoustic affinity between fifth-related pitches, or to the voice-leading relationships
between fifth-related diatonic scales? One way to study this question would be to
compare the frequency of modulations in classical pieces to the voice-leading
distances among their associated scales. Preliminary investigations, summarized in
Figure 8, suggest that voice-leading distances are in fact very closely correlated to
modulation frequencies. Surprising as it may seem, the acoustic affinity of perfect

fifth-related notes may be superfluous when it comes to explaining classical

modulatory practice.6
Ef

Bf
Gf
E

Af

Fig. 9. On this three-dimensional Tonnetz, the C7 chord is represented by the tetrahedron

whose vertices are C, E, G, and Bf. The C7 chord is represented by the nearby tetrahedron C,
Ef, Gf, Bf, which shares the C-Bf edge.

3 Tuning lattices as approximate models of voice leading

We will now investigate the way tuning lattices like the Tonnetz represent voiceleading relationships among familiar sonorities. Here my argumentative strategy will
by somewhat different, since it is widely recognized that the Tonnetz has something to
do with voice leading. (This is largely due to the important work of Richard Cohn,
who has used the Tonnetz to study what he calls parsimonious voice leading.7) My
goal will therefore be to explain why tuning lattices are only an approximate model of
contrapuntal relationships, and only for certain chords.
The first point to note is that inversionally related chords on a tuning lattice are
near each other when they share common tones.8 For example, the Tonnetz represents
perfect fifths by line segments; fifth-related perfect fifths, such as {C, G} and {G, D}
are related by inversion around their common note, and are adjacent on the lattice
(Figure 3). Similarly, major and minor triads on the Tonnetz are represented by
triangles; inversionally related triads that share an interval, such as {C, E, G} and {C,
E, A}, are joined by a common edge. (On the standard Tonnetz, the more common
tones, the closer the chords will be: C major and A minor, which share two notes, are
closer than C major and F minor, which share only one.) In the three-dimensional
Tonnetz shown in Figure 9, where the z axis represents the seventh, C7 is near its
inversion C7. The point is reasonably general, and does not depend on the particular
6

Similar points could potentially be made about the prevalence, in functionally tonal music, of
root-progressions by perfect fifth. It may be that the diatonic circle of thirds shown in Figure
5 provides a more perspicuous model of functional harmony than do more traditional fifthbased representations.
7
See Cohn 1997.
8
This is not true of the voice leading spaces considered earlier: for example, in three-note
chord space {C, D, F} is not particularly close to {F, Af, Bf}.

structure of the Tonnetz or on the chords involved: on tuning lattices, inversionally

related chords are close when they share common tones.9
The second point is that acoustically consonant chords often divide the octave
relatively evenly; such chords can be linked by efficient voice leading to those
inversions with which they share common notes.10 It follows that proximity on a
tuning lattice will indicate the potential for efficient voice leading when the chords in
question are nearly even and are related by inversion. Thus {C, G} and {G, D} can
be linked by the stepwise voice leading (C, G)
(D, G), in which C moves up by two
semitones. Similarly, the C major and A minor triads can be linked by the single-step
voice leading (C, E, G)
(C, E, A), and C7 can be linked to C7 by the two semitone
voice-leading (C, E, G, Bf)
(C, Ef, Gf, Bf). In each case the chords are also close
on the relevant tuning lattice. (Note that triadic distances on the diatonic Tonnetz in
Figure 3 exactly reproduce the circle-of-thirds distances from Figure 5.) This will not
be true for uneven chords: {C, E} and {E, Gs} are close on the Tonnetz, but cannot be
linked by particularly efficient voice leading; the same holds for {C, G, Af} and {G,
Af, Df}. Tuning lattices are approximate models of voice-leading only when one is
concerned with the nearly-even sonorities that are fundamental to Western tonality.
Bff

Ff

Df

Cf

Af

Gf

Ef

Df

Bf

& b

4
F

1
2

Fs

Cs

Fig. 10. On the Tonnetz, F major (triangle 3) is closer to C major (triangle 1) than F minor
(triangle 4) is. In actual music, however, F minor frequently appears as a passing chord
between F major and C major. Note that, unlike in Figure 3, I have here used a Tonnetz in
which the axes are not orthogonal; this difference is merely orthographical, however.

Furthermore, on closer inspection Tonnetz-distances diverge from voice-leading

distances even for these chords. Some counterexamples are obvious: for instance, {C,
G} and {Cs, Fs} can be linked by semitonal voice leading, but are fairly far apart on
the Tonnetz. Slightly more subtle, but more musically pertinent, is the following
example: on the Tonnetz, C major is two units away from F major but three units from
F minor (Figure 10). (Here I measure distance in accordance with neo-Riemannian
9

In general, the notion of closeness needs to be spelled out carefully, since chords can
contain notes that are very far apart on the lattice. In the cases we are concerned with, chords
occupy a small region of the tuning lattice, and the notion of closeness is fairly
straightforward.
10
See Tymoczko 2006 and 2008. The point is relatively obvious when one thinks
geometrically: the two chords divide the pitch-class circle nearly evenly into the same
number of pieces; hence, if any two of their notes are close, then each note of one chord is
near some note of the other.

theory, which considers triangles sharing an edge to be one unit apart and which
decomposes larger distances into sequences of one-unit moves.) Yet it takes only two
semitones of total motion to move from C major to F minor, and three to move from
C major to F major. (This is precisely why F minor often appears as a passing chord
between F major and C major.) The Tonnetz thus depicts F major as being closer to C
major than F minor is, even though contrapuntally the opposite is true. This means
we cannot use the figure to explain the ubiquitous nineteenth-century IV-iv-I
progression, in which the two-semitone motion ^6
^5 is broken into two singlesemitone motions ^6
f^6
^5 .
One way to put the point is that while adjacencies on the Tonnetz reflect voiceleading facts, other relationships do not. As Cohn has emphasized, two major or
minor triads share an edge if they can be linked by parsimonious voice-leading in
which a single voice moves by one or two semitones. Thus, if we are interested in
this particular kind of voice leading then the Tonnetz provides an accurate and useful
model. However, there is no analogous characterization of larger distances in the
space. In other words, we do not get a recognizable notion of voice-leading distance
by decomposing voice leadings into sequences of parsimonious moves: as we have
seen, (F, A C)
(E, G, C) can be decomposed into two parsimonious moves, while it
takes three to represent (F, Af, C)
(E, G, C); yet intuitively the first voice leading
should be larger than the second. The deep issue here is that it is problematic to assert
that parsimonious voice leadings are always smaller than non-parsimonious voiceleadings: for by asserting that (C, E, A)
(C, E, G) is smaller than (C, F, Af)
(C, E,
G), the theorist runs afoul what Tymoczko calls the distribution constraint, known
to mathematicians as the submajorization partial order.11 Tymoczko argues that
violations of the distribution constraint invariably produce distance measures that
violate our intuitions about voice leading; the problem with larger distances on the
Tonnetz would seem to illustrate this more general claim.
Nevertheless, the fact remains that the two kinds of distance are roughly consistent:
for major and minor triads, the correlation between Tonnetz distance and voiceleading distance is a reasonably high .79.12 Furthermore, since Tymoczkos
distribution constraint is not intuitively obvious, unwary theorists might well think
that they could consistently declare the parsimonious voice leading (C, E, G)
(C,
E, A) to be smaller than the non-parsimonious (C, E, G)
(Cs, E, Gs). (Indeed, the
very meaning of the term parsimonious suggests that some theorists have in fact
done so.) Consequently, Tonnetz-distances might well appear, at first or even second
blush, to reflect some reasonable notion of voice-leading distance; and this in turn
could lead the theorist to conclude that the Tonnetz provides a generally applicable
tool for investigating triadic voice-leading. I have argued that we should resist this
conclusion: if we use the Tonnetz to model chromatic music, than Schuberts major11

See Tymoczko 2006, and Hall and Tymoczko 2007. Metrics that violate the distribution
constraint have counterintuitive consequences, such as preferring crossed voice leadings to
their uncrossed alternatives. Here, the claim that A minor is closer to C major than F minor
leads to the F minor/F major problem discussed in Figure 10.
12
Here I use the L1 or taxicab metric. The correlation between Tonnetz distances and the
number of shared common tones is an even-higher .9.

third juxtapositions will seem very different from his habit of interposing F minor
between F major and C major, since the first can be readily explained using the
Tonnetz whereas the second cannot.13 The danger, therefore, is that we might find
ourselves drawing unnecessary distinctions between these two casesparticularly if
we mistakenly assume the Tonnetz is a fully faithful model of voice-leading
relationships.

4 Voice leading, quality space, and the Fourier transform

We conclude by investigating the relation between voice leading and the Fourierbased perspective.14 The mechanics of the Fourier transform are relatively simple: for
any number n from 1 to 6, and every pitch-class p in a chord, the transform assigns a
two-dimensional vector whose components are:
Vp, n = (cos (2
pn/12), sin (2
pn/12))
Adding these vectors together, for one particular n and all the pitch-classes p in the
chord, produces a composite vector representing the chord as a wholeits nth
Fourier component. The length (or magnitude) of this vector, Quinn observes,
reveals something about the chords harmonic character: in particular, chords
saturated with (12/n)-semitone intervals, or intervals approximately equal to 12/n,
tend to score highly on this index of chord quality.15 The Fourier transform thus
seems to quantify the intuitive sense that chords can be more-or-less diminishedseventh-like, perfect-fifthy, or whole-toneish. Interestingly, Z-related chordsor
chords with the same interval contentalways score identically on this measure of
chord-quality. In this sense, Fourier space (the six-dimensional hypercube whose
coordinates are the Fourier magnitudes) seems to model a conception of similarity
that emphasizes interval content, rather than voice leading or acoustic consonance.
However, there is again a subtle connection to voice leading: it turns out that the
magnitude of a chords nth Fourier component is approximately linearly related to the
(Euclidean) size of the minimal voice leading to the nearest subset of any perfectly
even n-note chord.16 For instance, a chords first Fourier component (FC1) is
approximately related to the size of the minimal voice leading to any transposition of
{0}; the second Fourier component is approximately related to the size of the minimal
voice leading to any transposition of either {0} or {0, 6}; the third component is
approximately related to the size of the minimal voice leading to any transposition of
13

See Cohn 1999.

The ideas in the following section are influenced by Robinson (2006), Hoffman (2007), and
Callender (2007).
15
Here I use continuous pitch-class notation where the octave always has size 12, no matter
how it is divided. Thus the equal-tempered five-note scale is labeled {0, 2.4, 4.8, 7.2, 9.6}.
16
Here I measure voice-leading using the Euclidean metric, following Callender 2004. See
Tymoczko 2006 and 2008 for more on measures of voice-leading size.
14

either {0}, {0, 4} or {0, 4, 8}, and so on. Figure 11 shows the location of the subsets
of the n-note perfectly even chord, as they appear in the orbifold representing threenote set-classes, for values of n ranging from 1 to 6.17 Associated to each graph is one
of the six Fourier components. For any three-note set class, the magnitude of its nth
Fourier component is a decreasing function of the distance to the nearest of these
marked points: for instance, the magnitude of the third Fourier component (FC3)
decreases, the farther one is from the nearest of {0}, {0, 4} and {0, 4, 8}. Thus,
chords in the shaded region of Figure 12 will tend to have a relatively large FC3,
while those in the unshaded region will have a smaller FC3. Figure 13 shows that this
relationship is very-nearly linear for twelve-tone equal-tempered trichords.

FC1, subsets of {0}

FC3, subsets of {0, 4, 8}

FC5, subsets of {0, 2.4, 4.8, 7.2, 9.6}

FC2, subsets of {0, 6}

FC4, subsets of {0, 3, 6, 9}

FC6, subsets of {0, 2, 4, 6, 8, 10}

Fig. 11. The magnitude of a set classs nth Fourier component is approximately linearly related
to the size of the minimal voice leading to the nearest subset of the perfectly even n-note chord,
shown here as dark spheres.

17

See Callender 2004, Tymoczko 2006, Callender, Quinn, and Tymoczko, 2008.

Fig. 12. Chords in the shaded region will have a large FC3 component, since they are near
subsets of {0, 4, 8}. Those in the unshaded region will have a smaller FC3 component.

magnitude of the 3rd

Fourier component

014 001
015 003
037 005

048
004
000

y = 1.38x + 3.16

024 002
026 006

027 013
016 036
012 025

0.5

1.5

minimal voice leading

Figure 13. For trichords, the equation FC3 = 1.38VL + 3.16 relates the third Fourier
component to the Euclidean size of the minimal voice leading to the nearest subset of {0, 4, 8}.
Table 1. Correlations between voice-leading distances and Fourier magnitudes.

Dyads
Trichords
Tetrachords
Pentachords
Hexachords
Septachords
Octachords
Nonachords
Decachords

FC1
-.97
-.98
-.96
-.96
-.96
-.96
-.96
-.96
-.96

FC2
-.96
-.97
-.96
-.96
-.96
-.96
-.96
-.96
-.96

FC3
-.97
-.97
-.95
-.95
-.95
-.96
-.95
-.96
-.96

FC4
-1
-.98
-.98
-.98
-.96
-.97
-.98
-.98
-.98

FC5
-.97
-.98
-.96
-.96
-.96
-.96
-.96
-.96
-.96

FC6
-1*
-1*
-1*
-1*
-1*
-1*
-1*
-1*
-1*

* Voice leading calculated using L1 (taxicab) distance rather than L2 (Euclidean).

Table 1 uses the Pearson correlation coefficient to estimate the relationship

between the voice-leading distances and Fourier components, for twelve-tone equaltempered multisets of various cardinalities. The strong anti-correlations indicate that
one variable predicts the other with a very high degree of accuracy. Table 2
calculates the correlation coefficients for three-to-six-note chords in 48-tone equal
temperament. These strong anticorrelations, very similar to those in Table 1, show
that there continues to be a very close relation between Fourier magnitudes and voice-

leading size in very finely quantized pitch-class space. Since 48-tone equal
temperament is so finely quantized, these numbers are approximately valid for
continuous, unquantized pitch-class space.18
Table 2. Correlations between voice-leading distances and Fourier magnitudes in 48-tone equal
temperament.
Trichords
Tetrachords
Pentachords
Hexachords

FC1
-.99
-.97
-.97
-.96

Explaining these correlations, though not very difficult, is beyond the scope of this
paper. From our perspective, the important question is whether we should measure
chord quality using the Fourier transform or voice leading.19 In particular, the issue is
whether the Fourier components model the musical intuitions we want to model: as
we have seen, the Fourier transform requires us to measure a chords harmonic
quality in terms of its distance from all the subsets of the perfectly even n-note
chord. But we might sometimes wish to employ a different set of harmonic
prototypes. For instance, Figure 14 uses a chords distance from the augmented triad
to measure the trichordal set classes augmentedness. Unlike Fourier analysis, this
purely voice-leading-based method does not consider the triple unison or doubled
major third to be particularly augmented-like; hence, set classes like {0, 1, 4} do
not score particularly highly on this index of augmentedness. This example
dramatizes the fact that, when using voice leading, we are free to choose any set of
harmonic prototypes, rather than accepting those the Fourier transform imposes on us.

Fig. 14. The mathematics of the Fourier transform requires that we conceive of chord quality
in terms of the distance to all subsets of the perfectly even n-note chord (left). Purely voiceleading-based conceptions instead allow us to choose our harmonic prototypes freely (right).
Thus we can voice leading to model a chords augmentedness in terms of its distance from
the augmented triad, but not the tripled unison {0, 0, 0} or the doubled major third {0, 0, 4}.
18

It would be possible, though beyond the scope of this paper, to calculate this correlation
analytically. It is also possible to use statistical methods for higher-cardinality chords. A
large collection of randomly generated 24- and 100-tone chords in continuous space
produced correlations of .95 and .94, respectively.
19
See Robinson 2006 and Straus 2007 for related discussion.

5 Conclusion
The approximate consistency between our three models is in one sense good news:
since they are closely related, it may not matter muchat least in practical terms
which we choose. We can perhaps use a tuning lattice such as the Tonnetz to
represent voice-leading relationships, as long as we are interested in gross contrasts
(near vs. far) rather than fine quantitative differences (3 steps away vs. 2 steps
away). Similarly, we can perhaps use voice-leading spaces to approximate the
results of the Fourier analysis, as long as we are interested in modeling generic
harmonic intuitions (very fifthy vs. not very fifthy) rather than exploring very
fine differences among Fourier magnitudes.
However, if we want to be more principled, then we need to be more careful. The
resemblances among our models mean that it might be possible to inadvertently use
one sort of structure to discuss properties that are more directly modeled by another.
And indeed, the recent history of music theory displays some fascinating (and very
fruitful) imprecision about this issue. It is striking that Douthett and Steinbach, who
first described several of the lattices found in the center of the voice-leading
orbifoldsincluding Figure 6explicitly presented their work as generalizing the
familiar Tonnetz.20 Their lattices, rather than depicting parsimonious voice leading
among major and minor triads, displayed single-semitone voice leadings among a
wide range of sonorities; and as a result of this seemingly small difference, they
created models in which all distances can be interpreted as representing voice-leading
size. However, this difference only became apparent after it was understood how to
embed their discrete structures in the continuous geometrical figures described at the
beginning of this paper. Thus the continuous voice-leading spaces evolved out of the
Tonnetz, by way of Douthett and Steinbachs discrete lattices, even though the
structures now appear to be fundamentally different. Related points could be made
about Quinns quality space, whose connection to the voice-leading spaces took
several yearsand the work of several authorsto clarify.
There is, of course, nothing wrong with this: knowledge progresses slowly and
fitfully. But the preceding investigation suggests that it we may need to be precise
about which model is appropriate for which music-theoretical purpose. I have tried to
show that the issues here are complicated and subtle: the mere fact that tonal pieces
modulate by fifth does not, for example, require us to use a tuning lattice in which
fifths are smaller than semitones. Likewise, there may be close connections between
voice-leading spaces and the Fourier transform, even though the latter associates Zrelated chords while the former does not. The present paper can thus be considered a
down-payment toward a more extended inquiry, one that attempts to determine the
relative strengths and weaknesses of our three similar-yet-different conceptions of
musical distance.

20

The same is true of Tymoczko 2004, which uses the term generalized Tonnetz to describe
another set of lattices appearing in the voice-leading spaces.

References
Callender, Clifton. 2004. Continuous Transformations. Music Theory Online 10.3.
. 2007. Continuous Harmonic Spaces. Unpublished.
Callender, Clifton, Quinn, Ian, and Tymoczko, Dmitri. 2008. Generalized VoiceLeading Spaces. Science 320: 346-348.
Cohn, Richard. 1991. Properties and Generability of Transpositionally Invariant
Sets. Journal of Music Theory 35: 1-32.
.1996. Maximally Smooth Cycles, Hexatonic Systems, and the Analysis of
Late-Romantic Triadic Progressions. Music Analysis 15.1: 9-40.
. 1997. Neo-Riemannian Operations, Parsimonious Trichords, and their
Tonnetz Representations, Journal of Music Theory 41.1: 1-66.
. 1998. Introduction to Neo-Riemannian Theory: A Survey and a Historical
Perspective, Journal of Music Theory 42.2: 167-180.
. 1999. As Wonderful as Star Clusters: Instruments for Gazing at Tonality in
Schubert, Nineteenth-Century Music 22.3: 213-232.
Douthett, Jack and Steinbach, Peter. 1998. Parsimonious Graphs: a Study in
Parsimony, Contextual Transformations, and Modes of Limited Transposition.
Journal of Music Theory 42.2: 241-263.
Hall, Rachel and Tymoczko, Dmitri. 2007. Poverty and polyphony: a connection
between music and economics. In Bridges: Mathematical Connections in Art,
Music, and Science. R. Sarhanghi, ed., Donostia, Spain.
Hoffman, Justin. 2007. On Pitch-class set cartography. Unpublished.
Lewin, David. 1959. Re: Intervallic Relations between Two Collections of Notes.
Journal of Music Theory 3: 298-301.
. 2001. Special Cases of the Interval Function between Pitch-Class Sets X and
Y. Journal of Music Theory 45: 1-29.
Quinn, Ian. 2006. General Equal Tempered Harmony (Introduction and Part I).
Perspectives of New Music 44.2: 114-158.
. 2007. General Equal-Tempered Harmony (Parts II and III). Perspectives of
New Music 45.1: 4-63.
Robinson, Thomas. 2006. The End of Similarity? Semitonal Offset as Similarity
Measure. Paper presented at the annual meeting of the Music Theory
Society of New York State, Saratoga Springs, NY.
Straus, Joseph. 2003. Uniformity, Balance, and Smoothness in Atonal Voice
Leading. Music Theory Spectrum 25.2: 305-352.
. 2007. Voice leading in set-class space. Journal of Music Theory 49.1: 45108.
Tymoczko, Dmitri. 2004. Scale Networks in Debussy. Journal of Music Theory
48.2: 215-292.
. 2006. The Geometry of Musical Chords. Science 313: 72-74.
. 2008. Scale Theory, Serial Theory, and Voice Leading. Music Analysis
27.1: 1-49.