Reproduced from the Proceedings of the Mathematics 2000 Festival, Melbourne, 10 13 January 2000

MUSIC FROM FRACTAL NOISE

Michael Bulmer
University of Queensland
<[email protected]>
There are many interesting connections between music and mathematics, though these are
rarely used when teaching maths in schools. In this paper we look at one example which
involves using musical motivations to introduce some mathematical ideas. The aim is to
develop a random method for making music which produces pleasant results.

INTRODUCTION
Folklore has it that music and mathematics are somehow related as human skills.
However when it comes to teaching mathematics we rarely use examples from music,
preferring more practical applications from physics, engineering, or finance. This is
perhaps a shame since it gives little credit to the artistic side of mathematics.
In this paper we will look at a short teaching activity that introduces a variety of ideas,
such as autocorrelation, through the goal of creating pleasant random music. While
having an obvious conclusion, this is very much an open-ended task, reflecting the desire
to encourage an ongoing artistic appreciation of mathematics.
The materials required for this activity are some dice and an instrument for listening to the
generated music. Alternatively, a computer can be used to simulate the dice rolls or to play
the music. Some experience with statistical measures will be useful for the mathematical
aspects. The activity also works well after a discussion of the more traditional kinds of
fractals.
RANDOM MUSIC
This activity should start with some discussion and brainstorming about how music could
be created randomly and the kind of properties that such music should have. There can
also be some discussion about how maths might help in this task. Below we describe
three methods for creating music. Each is based on some kind of noise, a random process
in time. Students may well come up with other methods that go beyond these.
White Noise
One of the easiest methods is to generate notes one at a time using dice. Suppose we have
6 dice, so when we roll them together and add up the results we get a number between 6
and 36. For each of these possibilities, assign it a note of some music scale. For example,
6 could be the C below middle C, 7 could be the following D, 8 the E, and so on up to
36. This is a standard major scale, but you can also use other scales or modes, as
described by Kandell (1984). You can write the result on normal music paper, or just
make a plot of the raw numbers.
If you listen to this music it will sound pretty bad, almost like static that has been slowed
down. This noise is termed white because of this. Figure 1 shows a plot of 256 notes
generated by this method.

Reproduced from the Proceedings of the Mathematics 2000 Festival, Melbourne, 10 13 January 2000

40
35
30
25
20
15
10
0

50

100

150

200

250

Figure1. Time plot of white noise

After you have generated a sequence of pitches, you can also generate a sequence of
durations for you notes using a similar method.
Brown Noise
An obvious problem with white noise is that there is no connection between successive
notes. To overcome this failing, students may suggest methods that result in various
kinds of brown noise. This noise gets its name because it corresponds to the random
walks of physical Brownian motion. The standard example is of a drunk who staggers
randomly back and forth, sometimes moving a bit in one direction and sometimes moving
a bit in the other direction.
For example, we might start our music at middle C. To make each new note we roll a die.
If the die comes up with a 1 then we go down two notes from where we are; if it comes
up 2 we go down one note; for 3 we go up one note; for 4 we go up two notes; and for 5
and 6 we stay where we are. Figure 2 gives an example of noise generated by this
method.
50
45
40
35
30
25
20
0

50

100

150

200

250

Figure 2. Time plot of brown noise

Again, you can use a similar process to produce lengths of notes. Brown music is less
painful to the ear but is still rather boring.
Pink Noise
White noise and brown noise can be seen as two extremes for random music. In white
noise there is no association between successive notes while for brown noise there is a
very strong association. White noise is dull because it is too unpredictable but brown

Reproduced from the Proceedings of the Mathematics 2000 Festival, Melbourne, 10 13 January 2000

noise is also dull because it is too predictable. Interestingly, neither is predictable in the
long run.
Traditional music, on the other hand, seems to achieve a balance between these extremes.
A composer might sit down with a vision for a whole piece of music, devise finer
structure for smaller sections, and then write the notes for each section. This gives the
pattern over time a long-range dependence while still involving short-term randomness.
This construction is reminiscent of the construction in the plane of the Koch snowflake,
as described in Mandelbrot (1982). The first four steps in making the snowflake are
shown in Figure 3. The result is an object which is self-similar, possessing similar
structures as you look closer and closer at it. Such objects, whether they are in space, like
the snowflake, or in time, like musical notes, are called fractals.

Figure 3. Creating a Koch snowflake

Neither white nor brown noise has the long-range dependence or self-similarity required
by nice music. We will create a type of noise that lies in between the extremes of white
and brown, sometimes known as pink noise (or 1/f noise, as described by Voss and
Clarke (1978)).
To achieve true pink noise is actually very difficult, as described by Mandelbrot (1971).
However, Gardner (1978) describes a simple method, invented by Richard Voss, which
approximates pink noise and which is easy enough for students to both carry out and
understand. Using n dice, this method will generate 2n notes. We will illustrate it here
with 3 dice, labelled A, B, and C. Make a list, as in Table 1, of the numbers from 0 up to
2n-1 with their binary representations. Start by rolling all three dice and adding up the
results to give the first note (note 0). To generate each subsequent note, look at the binary
digits that change from row to row in the table. For example, when moving from note 0
to note 1, the C digit changes while the A and B digits stay the same. Follow this by
rolling the C die again while leaving the A and B dice as they were. Add up the three
results to give note 1. To get note 2, roll both B and C but leave A alone. Continue doing
this until all 8 notes have been generated.

Reproduced from the Proceedings of the Mathematics 2000 Festival, Melbourne, 10 13 January 2000

Table 1
Binary method for pink noise
Note
A
B
0
1
2
3
4
5
6
7

0
0
0
0
1
1
1
1

0
0
1
1
0
0
1
1

0
1
0
1
0
1
0
1

It is clear that this method will give a series of notes that exhibit long-range dependence.
The higher digits change less frequently and so the corresponding dice provide a longterm stability in the sequence of notes. Compare Figure 4, showing an example of pink
noise created with 8 dice, with the pictures in Figures 1 and 2.

50
45
40
35
30
25
20
15
0

50

100

150

200

250

Figure 4. Time plot of pink noise

DESCRIBING NOISE
So far we have motivated and described the three colours of noise in general terms. Of
course we can also use mathematics to help explore these ideas more concretely. In
particular, the notion of musical notes being related over time is captured by the
definition of autocorrelation. More advanced students can also use a variance calculation
to try and determine the fractal dimension of our pink noise.
Autocorrelation
This activity requires an understanding of the standard correlation coefficient between two
variables. This is an easy idea to introduce in isolation and can be motivated by looking at
real data sets. For example, you could get the class to measure the lengths of their feet and
their heights, display the data, and then look at the calculated correlation.
Autocorrelation uses the same calculation, measuring the correlation coefficient rk between
values in the noise sequence and the values k time points ahead. (Write this down as a
usual data set to convey the idea.) This number k is called the lag. For example, the lag
would have no effect for white noise.

Reproduced from the Proceedings of the Mathematics 2000 Festival, Melbourne, 10 13 January 2000

Since the two data sets in autocorrelation are actually the same, the standard correlation
formula can be simplified to the following, as described by Chatfield (1996):
N k

(x

x )( xt + k x )

t =1

rk =

(x

x )2

t =1

To visual the autocorrelation structure of noise, students can make a correlogram, a plot
of autocorrelation against lag. This realistically requires a computer; even if students
cannot make the plots themselves there is still much room for discussing the plots.
The correlograms for white and brown noises follow patterns that are easy to guess.
Students can be encouraged to sketch the pattern they would expect to see beforehand.
The white noise correlogram in Figure 5 captures the fact that there really is no association
between values in the sequence, giving autocorrelations close to 0. Figure 6 shows the
correlogram for brown noise. The random walk nature of the noise means that values
close together will be highly correlated. As time increases values wander away from each
other and correlation declines, ultimately tending to 0.
0.4
0.2
0
-0.2
-0.4
0

10

15

20

Figure 5. Correlogram for white noise

1
0.8
0.6
0.4
0.2
0

10

15

20

Figure 6. Correlogram for brown noise

The picture for pink noise in Figure 7 shows what we would like, moderate correlation
over the short term which do not disappear to 0 over the long term.

Reproduced from the Proceedings of the Mathematics 2000 Festival, Melbourne, 10 13 January 2000

0.5
0.4
0.3
0.2
0.1
5

15

10

20

Figure 7. Correlogram for pink noise

The patterns in these correlograms are more easily visible when longer runs of values are
generated. (A run of length 1024 works well, requiring 10 dice for the pink noise.)
Variance Plots
The correlograms described above give an intuitive feel for the correlation structure of the
different types of noise, requiring only a minimal background in statistical ideas when
presented in conjunction with a discussion of correlation. However, the activity can be
extended if students know the important rule for the effect of sample size on the variability
of the average :
x2

1
n

Unlike correlation, it is probably unwise to use this activity to introduce this activity since
it turns out not to always hold! To see this, think of the run of values as a series of
samples of size m, for each of which we can calculate the mean. That is, for the sequence
x1, x2, , calculate the averages (x1++xm)/m, (xm+1++x2m)/m, , and then calculate
the sample variance of these numbers, Var(m). (This is certainly something that is better
suited to a computer!) Repeat this for a range of m values, say 2 to 30, and then plot a
graph of log(Var(m)) against log(m).
If the standard rule held we would expect to see a line of points with slope equal to 1.
Figure 8 shows this plot for the white noise, where the least-squares line through the
points is 1.29- 1.07x, a slope very close to 1. This is not surprising since in white noise
there is no association between adjacent values and so the consecutive samples really are
independent.
0.8
0.6
0.4
0.2
0
-0.2
-0.4
0.4

0.6

0.8

1.2

1.4

Figure 8. Variance plot for white noise

Reproduced from the Proceedings of the Mathematics 2000 Festival, Melbourne, 10 13 January 2000

The picture for brown noise, shown in Figure 9, is quite different. Here the variability of
the sample average seems to be independent of the sample size! This happens because the
variability of the values in a sample increases as the sample size increases, since the
random walk can cover more ground, and this perfectly cancels with the decrease in
variability from having a larger sample.
3
2.5
2
1.5
1
0.5
0.4

0.6

0.8

1.2

1.4

Figure 9. Variance plot for brown noise

The Koch snowflake is made up of a line (an object of dimension 1) which seems to fill
up the plane (dimension 2). Thus it is said to have a fractional dimension, somewhere
between 1 and 2 (in fact it is 1.26). The analogous fractal dimension of a noise is given
by the Hurst parameter, H. If is the slope of the line in the variance plot, then H is equal
to 1+ /2. White noise has H = 0.5 while brown noise has H = 1.0. As is to be expected,
the parameter is somewhere in between for pink noise. Figure 10 shows the variance plot
for pink noise, giving the least-squares line 1.40 - 0.29x. Thus the pink noise has a
fractal dimension of H = 0.86.

1.3
1.2
1.1
1
0.4

0.6

0.8

1.2

1.4

Figure 10. Variance plot for pink noise

EXTENSIONS
One of the great advantages in involving an obviously creative area such as music in
mathematics teaching is that it leads immediately to many extensions. Students are
naturally keen to come up with improvements on the basic algorithm which make the
resulting music more pleasant to the ear. These improvements could be mathematically
motivated, such as trying to change the autocorrelation structure. They could also be

Reproduced from the Proceedings of the Mathematics 2000 Festival, Melbourne, 10 13 January 2000

musically oriented, such as changing the scale used to produce modal music or
introducing a second instrument.
There are also extensions away from the original musical task, looking at other time series
to develop an understanding of the kind of patterns that might emerge. Students could be
set a project of making some measurement over time and then discussing the behaviour
that they observe. Stock prices typically give brown noise while anything to do with
independent observations will give white noise. Mandelbrot and Wallis (1968) originally
found pink noise when looking at the flooding patterns of rivers. Other patterns observed
may motivate a general discussion of time series analysis, including seasonal variation
and trend.
The study and use of pink noise and 1/f phenomena is currently of broad interest. For
example, traffic on the Internet exhibits the long-range dependence of pink noise that
current models do not properly capture; Jeong et. al (1999) give an overview of the role
of such self-similar noise in teletraffic research. There is an extensive bibliography of
other 1/f phenomena, in such areas as astronomy, ecology, economics, electronics, and
DNA sequences, on the web at http://linkage.rockefeller.edu/wli/1fnoise.

Technical Notes
The fractal noise used in this paper was generated by simulating the structured dice rolls
in Mathematica. This could have been done easily in almost any programming language.
To listen to the resulting music the numbers were converted into a standard MIDI file
using the Perl package MIDI-Perl by Sean Burke. This was then imported into the
QuickTime Player on the Macintosh and played. The instrument sounds available in
QuickTime are lovely, but again there are many other simpler ways to generate notes on a
computer.
REFERENCES
Chatfield, C. (1996), The Analysis of Time Series: An Introduction (4th ed.). London: Chapman & Hall.
Gardner, M. (1978), White and brown music, fractal curves and one-over-f fluctuations. Scientific
American, April, 1631.
Jeong, H.D.J., McNickle, D., & Pawlikowski, K. (1999), Fast Self-Similar Teletraffic Generation Based
on FGN and Wavelets. Proceedings of IEEE International Conference on Networks (ICON99),
Brisbane, Australia.
Kandell, J. (1984), Computer Music Worth Listening To. inCider, February, 6872.
Mandelbrot, B. (1971), A Fast Fractional Gaussian Noise Generator. Water Resources Research, 7,
543553.
Mandelbrot, B. (1982), The Fractal Geometry of Nature. New York: W.H. Freeman and Company.
Mandelbrot, B., & Wallis, J.R. (1968), Noah, Joseph and operational hydrology. Water Resources
Research, 4, 909918.
Voss, R.F., & Clarke, J. (1978), 1/f noise in music: Music from 1/f noise. Journal of the Acoustical
Society of America, 63 (1), 258263.