Set-Class Similarity, Voice Leading,

and the Fourier Transform

Dmitri Tymoczko

Abstract  In this article, I consider two ways to model distance (or inverse similarity) between chord types, one
based on voice leading and the other on shared interval content. My goal is to provide a contrapuntal reinterpre-
tation of Ian Quinn’s work, which uses the Fourier transform to quantify similarity of interval content. The first
section of the article shows how to find the minimal voice leading between chord types or set-classes. The second
uses voice leading to approximate the results of Quinn’s Fourier-based method. The third section explains how
this is possible, while the fourth argues that voice leading is somewhat more flexible than the Fourier transform.
I conclude with a few thoughts about realism and relativism in music theory.

twentieth-century music often moves flexibly between contrasting

harmonic regions: in the music of Stravinsky, Messiaen, Shostakovich, Ligeti,
Crumb, and John Adams, we find diatonic passages alternating with moments
of intense chromaticism, sometimes mediated by nondiatonic scales such as
the whole-tone and octatonic. In some cases, the music moves continuously
from one world to another, making it hard to identify precise bound­aries
between them. Yet we may still have the sense that a particular passage, melody,
or scale is, for instance, fairly diatonic, more-or-less octatonic, or less diatonic
than whole-tone. A challenge for music theory is to formalize these intuitions
by proposing quantitative methods for locating musical objects along the spec-
trum of contemporary harmonic possibilities.
One approach to this problem uses voice leading: from this point of
view, to say that two set-classes are similar is to say that any set of the first type
can be transformed into one of the second without moving its notes very far.
Thus, the acoustic scale is similar to the diatonic because we can transform
one into the other by a single-semitone shift; for example, the acoustic scale
{C, D, E, F≥, G, A, B≤} can be made diatonic by the single-semitone displace-
ment F≥ → F or B≤ → B. Similarly, when we judge the minor seventh chord

Thanks to Rachel Hall, Justin Hoffman, Ian Quinn, Joe Straus, and in particular Clifton Callender, whose
investigations into continuous Fourier transforms deeply influenced my thinking. Callender pursued his
approach despite strenuous objections on my part, for which I am both appropriately grateful and duly
chastened.

Journal of Music Theory  52:2, Fall 2008

DOI 10.1215/00222909-2009-017  © 2010 by Yale University 251
252 J ou r n al o f M usic T h eo r y

to be very similar to the dominant seventh, we are saying that we can relate
them by a single-semitone shift. This conception of similarity dates back to
John Roeder’s work in the mid-1980s (1984, 1987) and has been developed
more recently by Thomas Robinson (2006), Joe Straus (2007), and Clifton
Callender, Ian Quinn, and myself (2008). The approach is consistent with the
thought that composers, sitting at a piano keyboard, would judge chords to be
similar when they can be linked by small physical motions.
Another approach uses intervallic content: from this point of view, to say
that set-classes are similar is to say that they contain similar collections of inter-
vals. (That the two methods are different is shown by “Z-related” or “nontrivi-
ally homometric” sets, which contain the same intervals but are nonidentical
according to voice leading.) In a fascinating pair of papers, Quinn has dem-
onstrated that the Fourier transform can be used to quantify this approach.1
Essentially, for any number n from 1 to 6, and every pitch class p in a chord, the
Fourier transform assigns a two-dimensional vector whose components are
Vp,n 5 (cos 2ppn/12, sin 2ppn/12). (1)
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 whole—
its “nth Fourier component.” The length (or “magnitude”) of this vector,
Quinn astutely observes, reveals something about the chord’s harmonic char-
acter: in particular, chords saturated with (12/n)-semitone intervals, or inter-
vals approximately equal to 12/n, tend to score highly on this index of chord
quality.2 The Fourier transform thus seems to capture the intuitive sense that
chords can be more or less diminished-seventh-like, perfect-fifthy, or whole-
tonish. It also seems to offer a distinctive approach to set-class similarity: from
this point of view, two set-classes can be considered “similar” when their Fou-
rier magnitudes are approximately equal—a situation that obtains when the
chords have approximately the same intervals.
The interesting question is how these two conceptions relate. In recent
years, a number of theorists have tried to reinterpret Quinn’s Fourier magni-
tudes using voice-leading distances. Robinson (2006), for example, pointed
out that there is a strong anticorrelation between the magnitude of a chord’s
first Fourier component and the size of the minimal voice leading to the
nearest chromatic cluster. (See also Straus 2007, which echoes Robinson’s
point.) However, neither Robinson nor Straus found an analogous interpre-
tation of the other Fourier components. In an interesting article in this issue
(see pages 219–49), Justin Hoffman extends this work, interpreting Fourier
components in light of unusual “voice-leading lattices” in which voices move
by distances other than one semitone. But despite this intriguing idea, the

1  See Quinn 2006 and 2007. Quinn’s use of the Fourier 2  These magnitudes are the same for transpositionally or
transform develops ideas in Lewin 1959 and 2001 and Vuza inversionally related chords, so it is reasonable to speak of a
1993. set-class’s Fourier magnitudes.
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 253

relation between Fourier analysis and more traditional conceptions of voice

leading remains obscure.
The purpose of this article is to describe a general connection between
the two approaches: it turns out that the magnitude of a chord’s nth Fourier
component is approximately inversely related to the size of the minimal voice
leading to the nearest subset of any perfectly even n-note chord.3 For instance, a
chord’s first Fourier component is approximately inversely related to the size
of the minimal voice leading to any transposition of {0}; the second Fourier
component is approximately inversely related to the size of the minimal voice
leading to any transposition of either {0} or {0, 6}; the third component is
approximately inversely related to the size of the minimal voice leading to any
transposition of either {0}, {0, 4}, or {0, 4, 8}, and so on. Interestingly, however,
we can see this connection clearly only when we model chords as multisets
in continuous pitch-class space, following the approach of Callender, Quinn,
and Tymoczko (2008). (This in fact may be one reason why previous theorists
did not notice the relationship.) When we do adopt this perspective, we see
that there is a deep relationship between two seemingly very different con-
ceptions of set-class similarity, one grounded in voice leading, the other in
interval content. Furthermore, this realization allows us to generalize some of
the features of Quinn’s approach, using related methods that transcend some
of the limitations of the Fourier transform proper.

I. Voice leading and set-class similarity

Let me begin by describing the voice-leading approach to set-class similarity

(or inverse distance), reviewing along the way some basic definitions. Much of
what follows is drawn from (or implicit in) earlier essays, including Tymoczko
2006 and 2008 and Callender, Quinn, and Tymoczko 2008; readers who want
to explore these ideas further are hereby referred to these more in-depth
discussions.
We can label pitch classes using real numbers (not just integers) in the
range [0, 12), with C as 0.4 Here the octave has size 12, and familiar twelve-
tone equal-tempered semitones have size 1. This system provides labels for
every conceivable pitch class and does not limit us to any particular scale; thus,
the number 4.5 refers to “E quarter-tone sharp,” halfway between the twelve-
tone equal-tempered pitch classes E and F.
A voice leading between pitch-class sets corresponds to a phrase like “the
C major triad moves to E major by moving C down to B, holding E fixed, and
shifting G up by semitone to G≥.” We can notate this more efficiently by writing

3  By “perfectly even n-note chord” I mean the chord that 4  The notation [x, y) indicates a range that includes the lower
exactly divides the octave into n equally sized pieces, not bound x but not the upper bound y. Similarly (x, y) includes
necessarily lying in any familiar scale. For example, the per- neither upper nor lower bounds, while [x, y] includes both.
fectly even eight-note chord is {0, 1.5, 3, 4.5, 6, 7.5, 9, 10.5}.
254 J ou r n al o f M usic T h eo r y

1, 0, 1
(C, E, G) (B, E, G≥), indicating that C moves to B by one descending
semitone, E moves to E by zero semitones, and G moves to G≥ by one ascend-
ing semitone. The order in which voices are listed is not important; thus, (C,
1, 0, 1
E, G) (B, E, G≥) is the same as (E, G, C) 0, 1, 1 (E, G≥, B). The numbers
above the arrows represent paths in pitch-class space, or directed distances
such as “up two semitones,” “down seven semitones,” “up thirteen semitones,”
and so on. When the paths all lie in the range (–6, 6] I eliminate them; thus, a
notation like (C, E, G) → (B, E, G≥) indicates that each voice moves to its des-
tination along the shortest possible route, with the arbitrary convention being
that tritones ascend. Formally, voice leadings between pitch-class sets can be
modeled as multisets of ordered pairs, in which the first element is a pitch
class and the second a real number representing a path in pitch-class space.
Voice leadings are bijective when they associate each element of one
chord with precisely one element of the other. However, it matters whether we
represent chords as sets (containing no duplications) or multisets (which may
contain multiple copies of pitch classes). For example, the voice leading (C,
C, E, G) → (A, C, F, F) is simultaneously a nonbijective voice leading between
the sets {C, E, G} and {F, A, C} and also a bijective voice leading between the
multisets {C, C, E, G} and {F, F, A, C}. For the purposes of this article, it is con-
venient to represent chords as multisets and to consider only bijective voice
leadings between them. However, in other contexts, it can be useful to con-
sider sets and nonbijective voice leadings.5 It turns out to be a nontrivial task to
devise an algorithm for measuring set-class similarity when nonbijective voice
leadings are permitted. Fortunately, this complication is irrelevant here.
We measure the size of a voice leading using some function of (or partial
order on) the nondirected distances moved by the individual voices. (These are
the absolute values of the numbers above the arrows in the voice leading.) In
principle, there are many different measures of voice-leading size but no com-
pelling reason to choose one over another (Tymoczko 2006; Hall and Tymoc-
zko 2007). In this article, however, it is convenient to use the Euclidean metric,
according to which the size of a collection of real numbers x1, x2, . . . , xn is
x 12  x 22  . . .  xn2.
The reasons for this choice are that the Euclidean metric (1) provides a rea-
sonable approximation to a range of voice-leading measures (Hall and Tymoc-
zko 2007), (2) is computationally tractable, and (3) is particularly well suited
to the task of investigating the Fourier transform. The latter two points are
clarified shortly.
We can define the distance between two set-classes as the size of the minimal
voice leading between any of their transpositions or inversions. The term any

5  For example, one might consider the distance between C is in fact smaller than the smallest four-voice voice leading
and E major seventh chords to be determined by the nonbi- between them. See Callender, Quinn, and Tymoczko 2008,
jective voice leading (C, E, E, G, B) } (B, D≥, E, G≥, B), which supplementary section 7.
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 255

here means “any of their forms in continuous pitch-class space”; thus, when mea-
suring distances between set-classes we cannot necessarily confine ourselves
within any particular scale. For example, according to the Euclidean metric,
the distance between the perfect fourth and major third is given not by the
voice leading (C, F) → (D≤, F), with size 1, but by (C, F) → (C , E ) (or C
“quarter-tone sharp,” E “quarter-tone sharp”) with size

   
1 2
2

1 2
2
 1 2  0.707.

Though this may initially seem counterintuitive, it has on reflection a certain

logic: if we are really interested in intrinsic relations between set-classes, then
there is no reason to think that we can limit our attention to those that happen
to appear in any one scale.
Now the Euclidean metric is particularly convenient for the following
reason: if we are looking for the minimal voice leading between any two trans-
positions of any two chords, we need only consider those whose pitch classes
sum to the same value modulo 12. (This in turn follows from basic facts of
Cartesian geometry.)6 For example, suppose we are trying to find the minimal
Euclidean voice leading from the C augmented triad to any diminished triad.
The pitch classes {C, E, G≥} are represented by the numbers {0, 4, 8}, which sum
to 0 1 4 1 8 5 12 [ 0 (mod 12). To find the nearest diminished triad, we need
only consider those whose pitch classes sum to 0 (mod 12): {0, 3, 9}, {1, 4, 7},
and {5, 8, 11}. Observe that there are three, all related by major-third transpo-
sition. In general, we can always transpose an n-note chord by 12/n semitones
without changing its sum (mod 12), and we can repeat this procedure n times
before the initial chord reappears; thus, there will in general be n different
transpositions of each n-note chord summing to the same number.7 Note also
that when searching for minimal voice leadings, we will frequently need to
consider fractional pitch classes; for example, to find the nearest minor triad
to {0, 4, 7}, we need to look at those summing to 11. These are
1 1 1 1 1 1 1 1 1
{0 , 3 , 7 }, {4 , 7 , 11 }, {3 , 8 , 11 }
3 3 3 3 3 3 3 3 3
(or, in other words, the familiar C minor, E minor, and A≤ minor triads, trans-
posed up by one-third of a semitone). These chords, of course, do not reside
on the ordinary piano keyboard.
Finally, suppose that we have two chords (x1, x 2, . . . , xn) and (y1, y 2, . . . , yn)
with each chord’s pitch classes listed in ascending numerical order (when

6  An ordered set can be modeled as a point in Rn. Transpo- (1, 1, . . . , 1) will be equal to zero, which in turn implies that
sition corresponds to motion along the “unit diagonal” that the sum of its components is zero. Hence, the coordinates
contains both the origin and (1, 1, . . . , 1). Transpositional of its endpoints sum to the same value.
set-classes can thus be represented by lines parallel to the
7  The qualification “in general” is needed because of sym-
unit diagonal. The shortest vector between any two of these
metrical chords: when we transpose {0, 4, 8} by four semi-
lines will (according to the Euclidean metric) be perpendicu-
tones, we get the same chord again.
lar to both. This means that the vector’s dot product with
256 J ou r n al o f M usic T h eo r y

considered as real numbers). To find the minimal voice leading between

them, we need to consider n different circular permutations
(x 1, x 2, . . . , xn) → (y1, y 2, . . . , yn),
(x 1, x 2, . . . , xn) → (y 2, y 3, . . . , yn , y1),
.
.
.
(x 1, x 2, . . . , xn) → (yn , y1, y 2, . . . , yn21).
(These voice leadings have been described as the n “interscalar transposi-
tions” between the chords; that they are the only possibilities follows from
the fact that voice crossings always increase the Euclidean size of a voice lead-
ing [Tymoczko 2006].) For example, to identify the minimal voice leading
between {0, 3, 9} and {3, 10, 11}, we need to consider the three voice leadings
3, 5, 2
(0, 3, 9) (3, 10, 11),

2, 4, 6
(0, 3, 9) (10, 11, 3),

1, 0, 1
(0, 3, 9) (11, 3, 10).
Clearly, the third is the smallest, with a total size of 2. It may again seem
strange that we have to consider all these possibilities: roughly speaking, the
reason is that there is no way to determine the destination of any particular
pitch class without calculating the size of each and every one of these voice
leadings. In particular, we have no assurance that a maximally efficient voice
leading always associates a pitch class in one chord with its nearest neighbor
in the other.
Putting it all together, then, we can use the following procedure to find
the minimal Euclidean voice leading between two n-note multiset-classes A
and B:
(1) Choose a representative of A and calculate the sum of its pitch
classes.
(2) Find the n transpositions of B that sum to this same value.
(3) For each of these, calculate the (Euclidean) size of the n “intersca-
lar transpositions” described in the previous paragraph.
(4) Repeat steps 2 and 3 for the inversion of B.
(5) The minimum of these 2n 2 numbers is the Euclidean distance
between the multiset-classes.
Though it would be somewhat laborious to follow this algorithm by hand, it
is easy to program a computer to do it. The result is a single number repre-
senting the Euclidean distance between set-classes. Equivalently, this number
can be taken to represent the voice-leading distance from any particular set
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 257

(e.g., the C augmented chord) to the nearest transposition or inversion belong-

ing to some other set-class (e.g., any diminished triad in continuous space).

II. Voice-leading distances and Fourier magnitudes

To explore the connection between voice leading and the Fourier transform,
it is useful to begin with the “set-class spaces” described by Callender, Quinn,
and myself. These are n-dimensional geometrical spaces containing all the
multiset-classes of size n, where distances are as described in the preceding
section.8 Figure 1 shows the location, in three-note set-class space, of the
multiset-classes that can be constructed using the pitches of some perfectly
even n-note chord, for n ranging from 1 to 6.9 (Terminological note: I refer to
these as the “doubled subsets of the perfectly even n-note set-class.”)10 Associ-
ated 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 magni-
tude of the third Fourier component (FC3) decreases the farther one is from
the nearest of {0, 0, 0}, {0, 0, 4}, and {0, 4, 8}. Thus, set-classes in the shaded
region of Figure 2 will tend to have a relatively large FC3, while those in the
unshaded region will have a smaller FC3.
Figure 3 presents three-dimensional graphs in which the x,y-plane rep-
resents triangular set-class space, as in Figures 1 and 2, and where the z-axis
represents the magnitude of the relevant Fourier component.11 The graphs
show a series of peaks precisely at the doubled subsets of the perfectly even
n-note set-class, with valleys at the points most distant from these peaks. It is
clear from the graphs that there is a decreasing relationship between height
(nth Fourier magnitude) and distance to the nearest peak (doubled subset of
the perfectly even n-note set-class). Furthermore, the contour lines, showing
set-classes of equal Fourier magnitude, are roughly circular. This means that
the relevant measure of voice-leading size is the Euclidean metric, as this is
the metric for which a circle’s points are equidistant from the center.12 This is
quite fortunate, since Euclidean distance is also particularly easy to work with,
for the reasons discussed above.

8  See Callender 2004; Tymoczko 2006; Callender, Quinn, 11  Thanks to Cliff Callender for programming assistance. A
and Tymoczko 2008. Mathematically, these spaces are the very similar graph appears in Callender 2007, which explores
quotients of tori both by central inversion and by cyclical the Fourier transform in continuous space.
permutations of their barycentric coordinates.
12  There are many reasonable ways to measure voice lead-
9  Cliff Callender, in a personal communication, points out ing, as emphasized in both Tymoczko 2006 and Hall and
that the marked points in Figure 1 depict portions of a regu- Tymoczko 2007. Each produces a different set of points
lar lattice and that they differ only by a multiplicative factor. equidistant from a given location: for the “taxicab” metric,
this set is a diamond; for the Euclidean metric, a circle; and
10  These are not “submultisets” since they may introduce
for the “largest distance” metric, a square. See Hall and
additional duplications: {0, 0, 4} is not a submultiset of {0, 4,
Tymoczko 2007 for further discussion.
8}, since the former contains two copies of the “0” while
the latter contains only one. However, the latter chord can
be constructed by introducing doublings into a subset of the
perfectly even chord, hence the term “doubled subsets.”
 258 J ou r n al o f M usic T h eo r y

FC1, subsetsFC
of1,{0}
subsets of {0} FC2, subsetsFC
of2,{0,
subsets
6} of {0, 6}

FC3, subsetsFC
of3,{0,
subsets
4, 8} of {0, 4, 8} FC4, subsetsFC
of4,{0,
subsets
3, 6, 9}
of {0, 3, 6, 9}

FC5, subsetsFC
of5,{0,
subsets
2.4, 4.8,
of {0,
7.2,
2.4,
9.6}
4.8, 7.2, 9.6} FC6, subsetsFC
of6,{0,
subsets
2, 4, 6,
of8,
{0,10}
2, 4, 6, 8, 10}

Figure 1. The magnitude of a set-class’s nth Fourier component is approximately inversely

proportional to the distance to the nearest doubled subset of the perfectly even n-note set-class,
shown here as dark spheres. Each graph corresponds to one Fourier component and one
collection of doubled subsets.

Figure 3 provides intuitive evidence that there is a connection between

the nth Fourier magnitude and distance to the nearest doubled subset of the
perfectly even n-note set-class. The next task is to quantify and to explain this
relationship. I begin with a brute-force calculation that relates the two quanti-
ties for every twelve-tone equal-tempered multiset-class. (Other temperaments
are considered below.) The first part of the calculation is easy, requiring only
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 259

Figure 2.  Set-classes in the shaded region will have a large third Fourier component, since they
are near doubled subsets of {0, 4, 8}. Those in the unshaded region will have a smaller third
Fourier component.

Equation 1, above (Figure 4). The second is a little more difficult: in principle,
we need to repeat the algorithm in Section I for each doubled subset of the
perfectly even n-note set-class.13 However, Section III describes a shortcut that
simplifies the calculation considerably.
Once we determine both the nth Fourier component and the minimal
voice leading to the nearest doubled subset of the n-note set-class, we can plot
these two numbers for every (twelve-tone equal-tempered) multiset-class of a
given cardinality. Figure 5 shows, for trichordal multiset-classes, both the FC3
magnitude and the size of the minimal voice leading to the nearest doubled
subset of {0, 4, 8}. It is clear that there is a very nearly linear relation between
these two quantities, illustrated by the gray line:
FC3 5 21.38VL 1 3.16. (2)
Using this equation, one can estimate a trichord’s third Fourier component
(FC3) on the basis of the minimal voice leading to the nearest doubled sub-
set of any augmented triad (VL), and vice versa. The Pearson correlation
coefficient is a standard statistical measure that quantifies the “degree of fit”
between the points and the line. Here, the value –0.97 indicates that there is
a very nearly linear relation between the values.14
Table 1 correlates voice-leading distances and Fourier components, for
twelve-tone equal-tempered multiset-classes of other cardinalities. The values
in the table are determined by carrying out the computations in Figure 4 for
every equal-tempered multiset-class of size 2–10, and every Fourier component
from 1 to 6. (Appendix 1S, which appears as supplemental material [online
only] with this article at http://dx.doi.org/10.1215/00222909-2009-019, pre­
sents the raw data necessary to reconstruct Table 1.) The strong anticorrelations
indicate that one variable predicts the other with a high degree of accuracy.

13  If we are considering a k-element set-class, we need 14  A correlation of –1 indicates a perfect decreasing linear
to construct all of those doubled subsets with k elements. relation; a correlation of 11, a perfect increasing linear rela-
Thus, for the third Fourier component and three note chords, tion; and a correlation of 0, no linear relationship at all.
we need {0, 0, 0}, {0, 0, 4}, and {0, 4, 8}.
260 J ou r n al o f M usic T h eo r y

Figure 3.  Here the x,y-plane represents triangular set-class space, while the z-axis represents
Fourier magnitudes. The peaks are located at the doubled subsets of the perfectly even n-note
set-classes.

III. Understanding the correlations

We now explore this relationship in a more rigorous way. It follows from Equa-
tion 1 that the nth Fourier component represents pitch classes as unit vectors
in a “reduced” pitch-class space whose octave is only 12/n semitones large.
(The factor 2πn/12 maps pitch classes in the range [0, 12/n) to the circum-
ference of the unit circle; larger pitch-class numbers are reduced modulo
12/n.)15 Since all pitch classes p and p 1 12/n will be represented by identi-
cal vectors, we can move any note by 12/n semitones without changing the
nth Fourier component (see Figure 5; see also Hoffman 2008).16 This is illus-
trated geometrically in Figure 6. As long as pitch-class space is quantized finely
15  The reduced octave also appears in Cohn 1991. temperament containing p and p 1 12/n. For the purposes
of conceptualizing the Fourier transform, it is often useful to
16  Some equal temperaments will not contain both p and
work in this more finely quantized space, or in continuous
p 1 12/n. (E.g., twelve-tone equal temperament does not
unquantized space.
contain p and p 1 2.4.) However, we can always embed
an equal temperament into a more finely grained equal
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 261

a) Calculating the third Fourier component (FC3) of {0, 2, 5}.

Step 1: assign the vector (cos 2πpn/12, sin 2πpn/12) to each pitch class p.
0 → (cos(2π(0 3 3)/12), sin(2π(0 3 3)/12)) 5 (1, 0)
2 → (cos(2π(2 3 3)/12), sin(2π(2 3 3)/12)) 5 (21, 0)
5 → (cos(2π(5 3 3)/12), sin(2π(5 3 3)/12)) 5 (0, 1)
Step 2: add these vectors: (1, 0) 1 (21, 0) 1 (0, 1) 5 (0, 1)
Step 3: determine the length of this sum: (0, 1)  02  12  1
b) Determining the distance to the nearest doubled subset of {0, 4, 8}.

Following the procedure

in Section I, we learn that
(0, 2, 5) → (1, 1, 5) is the mini-
mal voice leading, moving two
voices by one semitone each,
and with a Euclidean size of
12  12  1.41

Figure 4.  Calculating the size of the third Fourier component of {0, 2, 5} and the minimal voice
leading from {0, 2, 5} to any doubled subset of {0, 4, 8}
magnitude of the 3rd
Fourier component

minimal voice leading

Figure 5.  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 doubled subset of {0, 4, 8}.

enough, moving any note by chromatic step will cause only a minimal change
to the Fourier components.17 To determine the magnitude of a chord’s nth
Fourier component, we add the vectors representing the all notes in the chord
and calculate the length of the result.

17  When pitch-class space is not finely quantized, this will pitch classes 0, 2.4, 4.8, 7.2, and 9.6 are assigned the same
not always be the case. For instance, consider the fifth vectors in the reduced pitch-class space of length 2.4. Mov-
Fourier component in twelve-tone equal temperament. The ing 0.2 of a semitone on the reduced circle leads to a point
262 J ou r n al o f M usic T h eo r y

Table 1.  Correlations between voice-leading distances and Fourier magnitudes

Fourier Component
Chord Size FC1 FC2 FC3 FC4 FC5 FC6
Dyads –0.97 –0.96 –0.97 –1 –0.97 –1*
Trichords –0.98 –0.97 –0.97 –0.98 –0.98 –1*
Tetrachords –0.96 –0.96 –0.95 –0.98 –0.96 –1*
Pentachords –0.96 –0.96 –0.95 –0.98 –0.96 –1*
Hexachords –0.96 –0.96 –0.95 –0.96 –0.96 –1*
Septachords –0.96 –0.96 –0.96 –0.97 –0.96 –1*
Octachords –0.96 –0.96 –0.95 –0.98 –0.96 –1*
Nonachords –0.96 –0.96 –0.96 –0.98 –0.96 –1*
Decachords –0.96 –0.96 –0.96 –0.98 –0.96 –1*

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

(Euclidean)

By contrast, when thinking about voice leading, we represent pitch

classes not as vectors but rather points on the pitch-class circle. Typically, we take
the circumference of the circle to be one octave. Suppose, however, that we
would like to find the minimal voice leading from some chord to the near-
est doubled subset of a perfectly even n-note chord. Figure 7 shows that this
problem is closely related to the problem of finding a voice leading from the
image of this pitch-class set, in a reduced pitch-class circle of circumference

Figure 6. The nth component of the Fourier transform imposes a smaller periodicity on the
pitch-class circle, representing pitches p and p + 12i/n by the same vector, for all integers i.
To determine the chord’s nth Fourier component, we add all the vectors corresponding to the
notes in the chord.

representing pitch classes 0.2, 2.6, 5, 7.4, and 9.8. Thus the lattices in which notes move by perfect fifth.) When we
perfect fourth appears to be smaller than the semitone— quantize more finely, however, motion by 0.2 of a semitone
indeed, it is the smallest twelve-tone equal-tempered inter- is seen to be just as small as motion by five semitones, and
val. (Hoffman 2008 exploits this fact to draw voice-leading motion by 0.1 of a semitone is smaller still.
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 263

Figure 7.  In searching for the minimal voice leading from any chord to the nearest doubled subset
of any transposition of the perfectly even n-note chord, it is sufficient to represent the initial
chord on a reduced pitch-class circle of size 12/n. The figure on the left represents the minimal
voice leading from {0, 5, 7} to any subset of {0, 6}, which is (0, 5, 7) → (0, 6, 6). The figure on the
right shows that this corresponds to the voice leading (0, 1, 5) → (0, 0, 0) in the reduced
pitch-class space of size 6.

12/n, to some transposition of the unison {0, . . . , 0}: any voice leading from
a set S to a doubled subset of the perfectly even n-note chord determines a
unique voice leading from the image of S to a unison in the reduced pitch-class
space.18 Thus, we need only look for voice leadings to doubled unisons in the
reduced pitch-class circle of length 12/n. This allows us to improve our algo-
rithm for identifying minimal voice leadings to the nearest doubled subset of a
perfectly even n-note chord.19
The reduced pitch-class circle of length 12/n therefore arises both
in determining the nth Fourier component and in identifying the minimal
voice leading to the nearest doubled subset of any perfectly even n-note
chord. The next task is to understand the relationship quantitatively. Figure 8
shows that a collection of vectors will yield the largest sum when they are all
pointing in the same direction or, in other words, when the chord they rep-
resent is a doubled subset of the perfectly even n-note chord. The vectors will
yield the smallest sum when they point in directions that are evenly distrib-
uted around the reduced pitch-class circle and hence cancel each other out.

18  In more mathematical terms: any voice leading in the 19  We begin by representing the chord modulo 12/n; we
larger pitch-class space, from a set S to any subset of the then consider the unisons whose pitch classes sum to the
perfectly even n-note chord, will project to an equally sized same value (mod 12/n) as those in the original chord. Thus,
voice leading in the reduced pitch-class space, from the if {x1, x2, . . . , xm} is our chord (mod 12/n), with x1 1 x2 1 . . .
image of set S to a unison; conversely, any voice leading in 1 xm [ s (mod 12/n), we need only consider voice leadings
the reduced space, from any set to a unison, can be lifted to the unison (s/m, s/m, . . . , s/m) and its transpositions by
to a collection of equally sized voice leadings in the larger 12/nm semitones.
space. These voice leadings link the preimage of the set S to
a doubled subset of some perfectly even n-note chord.
264 J ou r n al o f M usic T h eo r y

Conversely, the size of the minimal voice leading to the nearest unison will be
zero when the vectors point in the same direction and will be maximally large
when the vectors are evenly distributed around the circle. Thus, there should
be a decreasing relation between the magnitude of the nth Fourier compo-
nent and the minimal voice leading to the nearest subset of the perfectly even
n-note chord.

Figure 8.  (Left) Doubled subsets of a perfectly even n-note chord will have a large nth Fourier
component, since they will be represented by vectors pointing in the same direction. No voice
leading is necessary to transform these chords into doubled subsets of the perfectly even n-note
chord. (Right) Chords whose vectors are evenly distributed around the reduced pitch-class circle
will have an nth Fourier component of zero, since their vectors cancel out. It takes a large voice
leading to move these chords to a unison in the reduced pitch-class circle.

Importantly, however, the relation is only approximate. This is because

the Fourier transform represents pitch classes as vectors while the voice-leading
model represents them as points on the circle. Mathematically these are very
different, and there is no reason to expect the calculations to correspond
precisely. In fact, there exist chords that have the same nth Fourier compo-
nent but are not equidistant from the nearest doubled subset. Figure 9, for
example, shows two collections of vectors summing to zero: if we assemble a
chord by choosing one of the pitch classes at the head of each vector, it will
have a zero FC3 component; however, chords constructed using the circle on
the right will be slightly closer to the nearest doubled subset of the nearest
augmented triad. (See the tables in the online supplement to this article.)
Thus, the best we can hope for is an approximate correspondence between
Fourier magnitudes and voice-leading distances.
Indeed, this can be seen directly from Figure 3: the “hills” are not per-
fect cones, which implies that the relation between voice-leading distance and
Fourier magnitude is not perfectly linear. Furthermore, though it is less appar-
ent from the graph, the contour lines are not exactly circles, which means that
there is no function that will calculate Fourier magnitudes solely on the basis
of Euclidean voice-leading distances. This is shown more clearly in Figure 10,
which plots the magnitude of the first Fourier component against voice-
leading distance from the nearest tripled unison ({0, 0, 0}), for all trichordal
multiset-classes in 192-tone equal temperament. (This graph is the analogue
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 265

Figure 9.  For each circle, one can assemble a number of different multisets by choosing one pitch
class at the head of each arrow. All of these will have a vanishing third Fourier component.
However, those produced by the rightmost circle will have a slightly smaller voice leading to the
nearest subset of the nearest augmented triad.

of Figure 5 for this very finely quantized chromatic universe.) For chords close
to the triple unison, there is basically a one-to-one correspondence between
Fourier magnitude and voice-leading distance, as can be seen from the fact
that the upper-left portion of the graph is very thin. (Note that the slight
curvature indicates that the relationship is not quite linear.) The “bulge”
on the lower right shows that the relation becomes more approximate with
increasing distance: here, multiset-classes can have a range of Fourier magni-
tudes, even if they are equidistant from the triple unison. The graph tapers
again for chords maximally distant from {0, 0, 0}, indicating that the relation
between voice leading and Fourier magnitudes becomes more precise at large
distances. Figure 10 thus clearly shows both that voice-leading distance is a
reasonable predictor of the Fourier magnitude and that the relationship is
necessarily somewhat approximate. We cannot perfect our predictions simply
by using another familiar measure of voice leading, or even a simple func-
tion thereof: since there is essentially a one-to-one relationship near the triple
unison, any equation relating Fourier magnitudes to these voice-leading dis-
tances must reduce to the Euclidean metric at short range. However, because
of the bulge in Figure 10, we know that at larger distances anything resem-
bling the Euclidean metric will provide only an approximate predictor of the
magnitude of the first Fourier component.
Figure 11 contains analogous graphs relating Fourier magnitudes to
voice-leading distances for tetrachordal, pentachordal, and hexachordal
multiset-classes in 48-tone equal temperament. The graphs are all reasonably
similar in shape. Unlike Figure 10, they do not “taper” at the point of maxi-
mal distance from the perfect unison.20 The graphs are increasingly dense for

20  This is because there is only one way (within transposi- pronounced inflection point at Fourier magnitude k 2 2. This
tion) to arrange three unit vectors so that they sum to zero, may reflect the fact that there are a large number of ways to
whereas there are several ways of doing it for four or more combine k 2 2 vectors pointing in the same direction with
vectors. Note that the graphs for a k-note chord have a two other vectors pointing opposite one another.
266 J ou r n al o f M usic T h eo r y

magnitude of the 1st

Fourier component

minimal voice leading

Figure 10.  Fourier magnitudes and voice-leading distance for trichords in 192-tone equal
temperament. The correlation between the two values is –0.99.

larger chords, reflecting the fact that the number of multiset-classes grows
very quickly with cardinality. (Indeed, there are about a quarter-million hexa-
chordal multiset-classes in 48-tone equal temperament, and even more for
higher cardinalities—which is why it is difficult to produce analogous graphs
for larger multisets.) Table 2 calculates the correlation between voice leading
and Fourier magnitudes for three- to six-note chords in 48-tone equal tem-
perament. The strong anticorrelations show that relationship continues to
hold in very finely quantized pitch-class space. (In fact, 48-tone equal tempera-
ment is dense enough that these values approximate those for unquantized,
continuous pitch-class space.)21 Furthermore, in continuous space, the graphs
of all the Fourier components will be essentially identical, since in each case
vectors can point in any direction on the relevant reduced pitch-class circle.
Thus, the graphs in Figures 10 and 11, as well as the correlations in Table 2,
can be taken to represent not just the first Fourier component but the other
components, as well.
In my view, we should not be disappointed that there is only an approxi-
mate relation between voice-leading distance and Fourier magnitude. Both
the Fourier transform and the Euclidean voice-leading metric are very pre-
cise tools for modeling inherently vague musical intuitions, and we should
not become too invested in their fine quantitative structure; indeed, there is
little reason to think that very small differences in either Fourier magnitude
or Euclidean voice-leading distance correspond to anything psychologically
real for composers or listeners. What is more interesting, to my mind, is that
both the Fourier transform and voice leading provide similar, and intuitively

21  It would be possible, though beyond the scope of this and 100-tone chords in continuous space produced correla-
article, to calculate this correlation analytically. It is also pos- tions of 0.95 and 0.94, respectively. (Thanks to Rachel Hall
sible to use statistical methods for higher-cardinality chords. for performing these calculations.)
A sequence of a large number of randomly generated 24-
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 267

(a)
magnitude of the 1st
Fourier component

minimal voice leading

(b)
magnitude of the 1st
Fourier component

minimal voice leading

(c)
magnitude of the 1st
Fourier component

minimal voice leading

Figure 11.  Fourier magnitudes and voice-leading distance for tetrachords (a), pentachords (b), and
hexachords (c) in 48-tone equal temperament
268 J ou r n al o f M usic T h eo r y

Table 2.  Correlations between voice-leading

distances and Fourier magnitudes in 48-tone equal
temperament. These numbers are approximately
valid for any Fourier component in continuous
pitch-class space.
Chord Size FC1
Trichords –0.99
Tetrachords –0.97
Pentachords –0.97
Hexachords –0.96

plausible, ways of modeling the sense that chords can be very “major thirdy”
(or “whole-tonish”) without being exactly so. Here it is important that
there is a particularly strong resemblance for chords very close to doubled
subsets of perfectly even n-note chords. Thus, the two models will agree about
which chords are especially “fifthy,” “whole-tony,” and so forth—even if they
disagree somewhat about chords that are only mildly so.
Readers will have noticed that there is one circumstance in which Fou-
rier facts precisely mirror voice-leading facts: for twelve-tone equal-tempered
chords, the FC6 magnitude records the absolute value of the difference
between the number of its notes in one whole-tone scale and the number of
its notes in the other (Figure 12). (Mathematically, this is a scalar rather than
vector quantity.) One can obviously voice lead such chords to a doubled subset
of a whole-tone scale simply by moving all of the notes in the less populous
whole-tone set by semitone. It follows that the FC6 values will be perfectly
anticorrelated with the voice-leading distances obtained using the “taxicab”
(rather than Euclidean) metric. In fact, for k-note chords, the equation
FC6 5 k 2 2VL exactly determines the sixth Fourier component on the basis of
voice leading, where VL is the taxicab distance to the nearest doubled subset
of any whole-tone scale.

{C, D, E, Fs, Gs, Bf}

+1

-1

{Df, Ef, F, G, A, B}

Figure 12. The sixth Fourier component assigns the value +1 for notes in one whole-tone scale
and –1 for those in the other. The absolute value of the result represents the difference between
the number of notes in the more and less populous whole-tone scales. This Fourier component is
perfectly anticorrelated with the size of the voice leading to the nearest doubled subset of the
nearest whole-tone scale—as long as we measure voice-leading distance using the “taxicab”
metric.
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 269

IV. Discussion

Let’s return to the thought that the Fourier transform models the way chords
can be more or less saturated with particular intervals—that is, more or less
chromatic, whole-tonish, or perfect fifthy. On one level, this seems accurate:
chords such as {0, 2, 4} and {0, 0, 2} have a high sixth Fourier component, and
they are indeed saturated with major seconds. But when we think more care-
fully, we notice that the simple statement is not quite right: {0, 4, 8} also has a
very large sixth Fourier component, even though it contains no major seconds
at all! Furthermore, the Fourier components of the tripled unison {0, 0, 0} are
all maximally large, even though the multiset contains no nonzero intervals.
(By the continuity of the Fourier transform, something similar is true of such
chords as {0, e, 2e } for very small e.) Even the interpretation of the fifth Fourier
component, as representing the “perfect fifthiness,” needs to be qualified: in
very finely quantized equal temperaments, chords such as {0, 2.4, 4.8, 7.2, 9.6},
which have no perfect fifths, have a larger fifth Fourier component than the
pentatonic scale.
These examples suggest that we might sometimes want to depart from
Fourier analysis in favor of an approach based on voice leading. The Fourier
transform requires us to measure a chord’s “harmonic quality” in terms of
its distance from all the doubled subsets of the perfectly even set-classes. But
we might sometimes wish to choose a different set of harmonic prototypes.
For instance, Figure 13 uses distance from the augmented triad to measure
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.” Similarly,
we might sometimes wish to use a justly tuned diatonic scale as a harmonic
prototype, rather than accepting the fifth Fourier component as a proxy for
“diatonicness.” (Suppose we are investigating the acoustic purity of the inter-
vals in various temperaments’ best diatonic scales; here, voice leading will
produce much better results than the Fourier transform.) An approach based
on voice leading leaves us free to choose the harmonic prototypes we want,
rather than meekly accepting those the Fourier transform imposes on us.22
One way to put the point is that the Fourier transform is something
of a black box: we put a chord in, and get some numbers out. (In fact, it
can be quite hard to provide an intuitive characterization of what the
Fourier transform actually does—particularly if one makes no reference to
voice leading.) It is interesting that Quinn developed his Fourier-based tech-
nique under the influence of an avowedly “Platonist” conception of music the-
ory, according to which “chord quality” is a fundamentally objective feature

22  This is not to suggest that we should abandon Fourier

analysis in favor of approaches based on voice leading: Fou-
rier analysis provides an indispensable tool for investigat-
ing questions involving the interval vector, as suggested in
Lewin 1959 and 2001.
270 J ou r n al o f M usic T h eo r y

Figure 13. The mathematics of the Fourier transform requires that we conceive of “chord quality”
in terms of the distance to any doubled subset of some perfectly even set-class (left). If we use
voice leading, however, we can choose our harmonic prototypes freely. Thus, we can use voice
leading to model a set-class’s “augmentedness” in terms of its distance from the augmented
triad (right), but not the tripled unison {0, 0, 0} or the doubled major third {0, 0, 4}.

that is (as it were) “out there in the world.” By contrast, the voice-leading
approach is consonant with a more relativist conception according to which
we choose the musical properties that are important to us. A Platonist (e.g., the
youthful Quinn) might well be attracted to the “black box” quality of the Fou-
rier transform precisely because of its inflexibility—which could be taken to sug-
gest an idealized world of unalterable musical relationships. And conversely,
the very flexibility of the voice-leading approach might signal a (disturbing to
some, attractive to others) role for arbitrary human preferences and choice.
Beyond measuring the intervallic saturation of single set-classes, we can
of course use the Fourier transform to measure similarity between set-classes:
from this point of view, set-classes are similar when their six Fourier magni-
tudes are all similar. At first blush, this strategy seems to contrast dramati-
cally with the voice-leading approach: certainly, Fourier analysis uses very dif-
ferent mathematics, and produces results—such as the identity of Z-related
chords—that can be difficult to interpret in contrapuntal terms.23 We have
seen, however, that there is a close relationship between the two techniques:
at the most fundamental level, each individual Fourier component measures
something like a voice-leading distance. Thus what is distinctive about the
Fourier approach to chord similarity is not the conception of distance per se,
but rather the role of “harmonic prototypes”: the Fourier transform measures
the similarity of set-classes not by their distance from one another but by their
respective distances from the nearest doubled subsets of the perfectly even
n-note chords. This is why Z-related chords are judged to be identical, even
while being far apart in the set-class spaces such as Figure 1.24
From my point of view, the most interesting result is that a single concep-
tion of musical distance—voice-leading distance—turns out to underlie both
approaches. It is, I think, quite surprising that voice leading should play any

23  See Quinn 2006 and 2007. Quinn’s approach is inspired 24  Even from a voice-leading perspective, two Z-related
by earlier writers who emphasize shared subset content and chords will be approximately equidistant from the nearest
the interval vector. See Quinn 2001 for more discussion. doubled subsets of perfectly even n-note chords.
 Dmitri Tymoczko    Voice Leading and the Fourier Transform 271

role whatsoever in the Fourier transform, with its vectors, trigonometric func-
tions, and sensitivity to chords’ interval content. That we can reinterpret its
results contrapuntally says something about the power of an approach that puts
voice leading front and center. In fact, one might even take it to suggest that
Quinn’s early Platonism was not entirely misplaced: perhaps Quinn was right
to think that there is a realm of objective musical relationships that influence
us even when we are not directly aware of them. (Certainly, not everything in
music theory can be a matter of arbitrary personal preference!) If so, then I
would argue that voice leading—rather than the Fourier transform—has the
best claim to Platonic primacy. Perhaps it is spaces like Figure 1 that offer the
best glimpse of the entities casting shadows on the walls of our musical cave.

Appendix

The raw data from which Table 1 was constructed appears as supplemental
material (online only) with this article at http://dx.doi.org/10.1215/00222909-
2009-019. Appendix 1S shows the Fourier magnitudes and corresponding
minimal voice leadings for all twelve-tone equal-tempered multiset-classes.
Appendix 2S contains the data for twelve-tone equal-tempered set-classes.
An individual table is provided for set-classes and multiset-classes of each
cardinality: the first column identifies the (multi)set-class; the second shows
the first Fourier magnitude; the third, the size of the minimal voice leading
to the nearest doubled unison; the fourth, the second Fourier magnitude; the
fifth, the size of the minimal voice leading to the nearest doubled subset of
{0, 6}; and so on. Euclidean voice-leading distance is used for Fourier compo-
nents 1–5; the “taxicab” metric is used for Fourier component 6. In all cases,
voice-leading distances are calculated in continuous, unquantized pitch-class
space, as described in Section I and footnote 19.

Works Cited

Callender, Clifton. 2004. “Continuous Transformations.” Music Theory Online 10/3.

———. 2007. “Continuous Harmonic Spaces.” Journal of Music Theory 51: 277–332.
Callender, Clifton, Quinn, Ian, and Tymoczko, Dmitri. 2008. “Generalized Voice-Leading
Spaces.” Science 320: 346–48.
Cohn, Richard. 1991. “Properties and Generability of Transpositionally Invariant Sets.” Journal
of Music Theory 35: 1–32.
Hall, Rachel and Tymoczko, Dmitri. 2007. “Poverty and Polyphony: A Connection between
Music and Economics.” In Bridges Donostia: Mathematical Connections in Art, Music, and
Science, ed. Reza Sarhanghi and Javier Barallo, 259–68. Winfield, KS: Southwestern
College.
Hoffman, Justin. 2008. “On Pitch-Class Set Cartography: Relations between Voice-Leading
Spaces and Fourier Spaces.” Journal of Music Theory 52:219–49.
Lewin, David. 1959. “Re: Intervallic Relations between Two Collections of Notes.” Journal of
Music Theory 3: 298–301.
272 J ou r n al o f M usic T h eo r y

———. 2001. “Special Cases of the Interval Function between Pitch-Class Sets X and Y.” Journal
of Music Theory 45: 1–29.
Quinn, Ian. 2001. “Listening to Similarity Relations.” Perspectives of New Music 39/2: 108–58.
———. 2006. “General Equal-Tempered Harmony (Introduction and Part I).” Perspectives of
New Music 44/2: 114–58.
———. 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.
Roeder, John. 1984. “A Theory of Voice Leading for Atonal Music.” Ph.D. diss., Yale
University.
———. 1987. “A Geometric Representation of Pitch-Class Series.” Perspectives of New Music
25/1–2: 362–409.
Straus, Joseph. 2007. “Voice Leading in Set-Class Space.” Journal of Music Theory 49: 45–108.
Tymoczko, Dmitri. 2006. “The Geometry of Musical Chords.” Science 313: 72–74.
———. 2008. “Scale Theory, Serial Theory, and Voice Leading.” Music Analysis 27/1: 1–49.
Vuza, Dan Tudor. 1993. “Supplementary Sets and Regular Complementary Unending Canons
(Part Four).” Perspectives of New Music 31/1: 270–305.

Dmitri Tymoczko is a composer and music theorist who teaches at Princeton University. His book A
Geometry of Music will be published in 2010 by Oxford University Press. He is also working on an
album of pieces that combine jazz, rock, and classical styles.