E

Slow Fourier Transform

by Klaus-G. Hinzen, Sharon K. Reamer, and Frank
Scherbaum
Online Material: Interactive numerical model of the Mader
Ott Harmonic Analyzer.

INTRODUCTION
For contemporary seismological observatory practice and seismic exploration work, the fast Fourier transform (FFT) has
become such a common tool that we rarely think about single
transforms. Often, we switch from time to frequency to time
domain without even examining the spectral data. Restituting
seismometer response, calculating source spectra, checking
frequency-dependent attenuation, and doing f k analysis
are only a few applications in seismology, in which the FFT
processing steps are an essential tool. Since the advent of
modern digital computers, much effort has been made to make
the Fourier transform efficient and fast.
The basics for the modern transform algorithms were first
described by the French mathematician and physicist (Baron)
Jean-Baptiste-Joseph Fourier (born 21 March 1768 in Auxerre;
died 16 May 1830 in Paris). In 1822, he published his seminal
work on heat transport in solid-state bodies, The Analytic
Theory of Heat, while at the Ecole Polytechnic Institute in
Paris. Fourier not only presented the derivation of the heat
transport equations (later called Fouriers law), but he also
proposed a method of resolution, including what we today call
a Fourier series. This paper was awarded the prize in mathematics in 1811 by the institute (Hrivel, 1975; Bracewell,
1986).
Spectral analysis has a long history in the geophysical sciences, with work on the earths free oscillations as early as the
mid-nineteenth century (e.g., Lam, 1853). The first recorded
seismograph appeared a few decades after Lams work and has
been credited to Cecchi in 1887 (Dewey and Byerly, 1969).
Time series analysis, the analysis of a sequence of signals characterized by a row vector with (usually) real components (Kanasewich, 1981), did not gain wide theoretical treatment until
published works by Wiener (1930, 1949) and Kolmogorov
(1939) made calculations with transient signals feasible for geophysical research. Various methods for increased efficiency in
doi: 10.1785/0220120139

the computing of Fourier transforms in the first half of the

twentieth century (Cooley et al., 1967) culminated in the first
widely disseminated FFT algorithm from Cooley and Tukey
(1965). As described by Kanasewich (1981), time series analysis
became the basis for significant breakthroughs in mineral
exploration with early work, for example, by Wadsworth et al.
(1953).
Gubbins (2006) described the development of instrumental geophysical investigation as going through four major stages
since the late nineteenth century: mechanical, optical, analog
magnetic, and digital. Prior to digital computational efforts,
scientists working with periodic functions relied heavily on
mechanical methods of computation, such as the harmonic
analyzer. A nicely comprehensive summary of the evolution
of such devices is given by Otnes (2008).
Perhaps the earliest example of a harmonic analyzer is the
Kelvins tide machine, based on a conception accredited to
Lord Kelvin (William Thomson) and constructed in 1872 by
the Lg Engineering Company (Parker, 2011). This machine,
operated by a system of pulleys and wheels, could be used to
compute the first 10 tidal harmonics. As noted by Parker
(2011), this device was followed by other designs on both sides
of the Atlantic. Parker (2011) gave an excellent and riveting
account about the use of these machines, especially during
wartime.
Harmonic analyzers were also used to analyze periodic
motions of a variety of periodic functions. Two examples include graphical analysis of solar apex and velocity (Nassau and
Morse, 1927) and the controversial ether investigations of
Miller (1933) as chronicled by Fickinger (2005). Both studies
were carried out on the machine designed by Olaus Henrici
(and first built by G. Coradi in 1889), which utilized glass
spheres rotating brass dials (Henrici, 1894; Miller, 1916). The
original Henrici harmonic analyzer is still on display at the
Science Museum in London. (The periodical Philosophical
Magazine and Journal of Science, in which Henricis description of his machine appeared, was coproduced by Lord
Kelvin.)
Nearly concurrently, as described by Otnes (2008), a much
simplified and economical design, based on the implementation of different-sized cogwheels attached to a planimeter, was
published by Mader (1909), based on suggestions from Yule
(1894) and implemented with the construction from Ott
(1931). It is the operation of this device that will be explored
in the present paper.

Seismological Research Letters

Volume 84, Number 2

March/April 2013

251

Mader (1909) published the idea of a simple harmonic

analyzer with arbitrary base. The instrument was designed
with several practical advantages: (1) easy handling without
the need of reduction or auxiliary calculations; (2) applicable
to time series with arbitrary physical length of the baseline
(within practical limits), avoiding the need for a transformation of original measurements (always a potential source of
error); and (3) economy. Mader (1909) mentioned in a footnote that the instrument was manufactured by Gebrder
Staerzl, Mechanische Werksttte in Munich and was available
at a price of 120 mark. (Following an estimate of Deutsche
Bundesbank, this is equivalent to up to 1700 euro.)
The following description of the theory of MOHA is
based on the instructions for the use of the instrument given
in the German language by Ott (1931). The harmonic analysis
of an oscillation requires the decomposition into a fundamental mode and higher modes for a given periodic function y

f x with the period p. Following Fouriers 1822 treatise, this
decomposition can be achieved by an expansion of trigonometric functions:
a0
2
2
 a1 cos x  a2 cos 2 x    
p
p
2
2
2
 aN cos N x  b1 sin x
p
p
2
2
 b2 sin 2 x      bN sin N x
p
p



N
a0 X
2
2
an cos n x  bn sin n x ;


p
p
2
n
1

f x

(1)

and the absolute term a0 =2 corresponds to the average ordinate

of f x:

Figure 1. (a) Completely assembled MOHA. The main elements

of the instrument are labeled. See pencil for scale. (b) One of the
planimeters with the circular nonius scale. (c) Gearbox with two
cogwheels inserted next to the cograil. The large and small
wheels are for the measurement of the first and sixth Fourier coefficient, respectively. (d) All cogwheels from the instrument. The
wheels in the upper row are the gear wheels for coefficients
larger than 6. The other 22 wheels are for the determination of
coefficients 125 (see text for details).

HARMONIC ANALYZER
The MaderOtt Harmonic Analyzer (MOHA) is a highprecision mechanical instrument for determining the coefficients of a Fourier series (Fig. 1). Measurements with this
instrument are based on a mechanical planimeter device to
mechanically determine the surface area of a plane of arbitrary
shape. (This is also a useful tool to determine the radius of an
equivalent circle for isoseismals in a macroseismic study.) An
important aspect of the analog era was the uniqueness of plotted data. Any reproduction or change of scales had to be made
by redrawing the plot by hand.
252

Seismological Research Letters

Volume 84, Number 2

a0 1


p
2

f xdx:

(2)

Fouriers coefficients an and bn are given as

Z
2 p
2
f x cos n xdx and bn
an

p 0
p
Zp
2
2


f x sin n xdx:
p 0
p

(3)

Instead of integrating over the function from 0 to p, any

other starting point can be chosen as long as the integration is
made over a full period. Important for the applicability of
MOHA is the simplification of equations (1), (2), and (3)
for the case p
2:
f x

a0
 a1 cos x  a2 cos 2x      aN cos N x
2
 b1 sin x  b2 sin 2x      bN sin N x
N
a0 X

an cos nx  bn sin nx;
2
n
1

March/April 2013

(4)

Z 2
a0
1


f xdx;
2p 0
2
Z
an 1 2
cos


f x
nxdx:
bn 0
sin

(5)
(6)

Superposition of the fundamental harmonic and the

higher harmonics with the static component, the latter in seismology commonly called offset, results in function f x. If the
independent variable is time, as is the case for any seismogram,
x
t, then the period p equals the duration of the oscillation
T , and we get
2 2



2f
;
p
T

(7)

in which f is the frequency and is the angular frequency.

MOHA accomplishes the harmonic analysis following
equation (1) by delivering the numerical values of an and
bn for n
1; 2; ; N, with N
25, by following a trace of
y
f x, with a pin as many times as coefficients are to be
determined. The instrument is composed of an analyzer guide
and a common planimeter (polar or linear). The guide is constructed so that when the pin follows the curve f x, the attached planimeter delivers the surface area proportional to the
integrals in equation (3), in which f x is multiplied with
cos n 2p x and sin n 2p x. The absolute term a20 can be determined
by tracing f x directly with the planimeter, without the use of
the analyzer guide. The analyzer guide is based on two principal ideas. (1) A 90-angle-lever ACB (Fig. 2), which can rotate
around C within a limited-angle due to the two arrestors,
projects an arbitrary period p
OO0 to a fixed period length

Figure 2. (a) Principle function of MOHA. The heavy black lines

indicate the angle lever that projects the function of period length
p between points O and O 0 to a fixed length l. The angle lever can
rotate around pivot C. (b) A movement of the indicator pin P along
the x axis is transformed into a displacement of point B on the
angle lever (after Ott, 1931).

l
BB0 . The pivot C can move along the perpendicular bisector CD of the period axis OO0 . The moving tracer pin P can be
adjusted in its position along the instrument arm AC in such a
way that when the arm CB reaches its end positions at the
arrestors, P is at the beginning and end of the period at O
and O0 , respectively. Figure 2b shows that movement of the
pin P along the x axis provokes the proportional movement
of B along the y axis by the amount z
l=px.
(2) A cogwheel rolls along a cograil of length l exact n
times. For displacement of the cograil of distance z, the cogwheel will rotate with the proportional-angle
n2=lz.
For better clarity in Figure 3, the cograil is fixed, and the cogwheel is free to move. The angle lever, cograil, and cogwheel are
arranged on a carriage whose movement is limited to the y
direction. During operation of the instrument, the pivot of
the angle lever C must move along the perpendicular bisector
CD of the period axis OO0 . The end point B of the angle lever
moves the cograil by the amount l when the moving pin P
moves along the curve of y
f x from start point O to
end point O0 , independent of the length of p
OO0 . This
is technically achieved by a parallel sliding block. For movement of P by the amount x in the x direction, the cograil moves
z
l=px, which is equal to a cogwheel rotation of

n

2
2
z
n x:
l
T

(8)

The cogwheel of radius Rl
2nR has two small boreholes positioned at a 90 angle that are marked with the letters c
and s indicating the positions for the attachment of the

Figure 3. (a) A cogwheel rotates along a cograil (gray) of length

l. A lateral movement of the cograil, z, corresponds to a proportional rotation of the cogwheel, . (b) Rotation of the cogwheel
with the change of the coordinates of the s boring in the coordinate system (see text for details) (after Ott, 1931).

Seismological Research Letters

Volume 84, Number 2

March/April 2013

253

planimeter pin to determine the cosine and sine coefficients,

respectively.
To understand the movements of the planimeter attached
to the boreholes c and s of the cogwheel, a new coordinate
system with the axes and (Fig. 3b) is introduced. At the
beginning of a measurement, the cogwheel is positioned in
such a way that the s boring lays on the negative half axis
at
R and the c boring lays on the positive axis at

R. For a fixed carriage and guiding the moving pin P from
O to O0 , P will move along a circular arc with the ordinates
gx; p (Fig. 4). If P is steered along the function
y
f x, the movement of the carriage in the y direction is
f x  gx; p. Because of this displacement and because of
the rotation of the cogwheel of
n2=px due to the displacement x in x direction, the coordinates of the s boring are
changed (Fig. 3):

R
R cos

0
f x  gx; p  R sin :

(9)

If we move the pin P along f x from O to O0 and afterward back along the straight axis from O0 to O, we have
f x
0. The s boring with respect to the stationary
coordinate system describes a closed curve with surface area:
J
d;

(10)

in which
d
Rn

2
sin dx:
p

(11)

Separating forward and backward paths, we can write

Zp
2
J

ff x  gx; p  R sin gRn sin dx
p
0
Z0
2
 fgx; p  R sin gRn sin dx:
(12)
p
p
The entire second integral cancels out with the corresponding parts of the first integral from the movement of
the tracer pin back from O0 to O, which is independent of
the forward path, and the remaining part of equation (12)
becomes
Z
2 p
2
f x sin n xdx
Rnbn :
(13)
J
Rn
p 0
p
The term Rn is the device constant K . Consequently, the
planimeter with its moving pin in the s boring of the cogwheel
shows the surface area,
J
K bn ;

Figure 4. Schematic figure of the active elements of MOHA.

The carriage can move in the y direction only. The cograil with
the attached cogwheel is shown in gray. The heavy black curve is
the function f x for which the Fourier coefficients are to be
determined. The planimeter is not shown in this graph (after
Ott, 1931).
254

Seismological Research Letters

Volume 84, Number 2

(14)

which is just a constant factor times the Fourier coefficient bn .

When the planimeter pin is placed in the c boring, the corresponding coefficient an will result.
The MOHA used in this study has a K
100 mm; a
planimeter value of 0:1 cm2 corresponds to a Fourier coefficient of 0.01 cm. The ingenuity of the MOHA construction
is the projection of all period lengths p to a fixed lateral movement of the cograil of l, which makes the resulting Fourier
coefficients an and bn independent of p. Our model of MOHA
allows period lengths between 2.5 and 36.0 cm and is equipped
with 22 cogwheels for the first 25 Fourier coefficients.
For coefficients higher than 6, a conversion cogwheel is
used in addition to the specific cogwheel of the particular coefficient. A unique feature of the instrument is the option of
installing two parallel coefficient cogwheels. In other words,
with two planimeters attached to the two wheels (Fig. 1),
two coefficients can be determined with one run of the pin
along the curve, cutting the total measuring time in half.
Theoretically, by stepwise changing the period length p of
a record, hidden periodicity could be uncovered. Furthermore,
March/April 2013

by partitioning a record in u parts of equal length, analyzing

each part separately, summing up an and bn , and dividing by u,
the coefficients for the u nth harmonic of f x can be
determined.

PRACTICE
To gain experience with MOHA, we determined coefficients
of an analytic signal. The function shown in Figure 5 is the
superposition of seven harmonic functions, including the fundamental harmonic and those of the order 2, 3, 8, and 9, with
amplitudes between 1.0 and 3.0. Students and staff of the seismological station Bensberg were challenged by one of the authors to determine the first 10 cosine and sine coefficients of
the analytic function using the MOHA device. Figure 6 shows
the result of the seven test runs. With the exception of the
coefficient 9 s, the measurements not considered as outliers
are all better than 8% error. This is compatible with earlier
tests made by Korschunow (1955). Maders (1909) own test

Figure 5. Analytic signal to test the performance of MOHA.

with a sawtooth function resulted in errors between 0% and

9% for the first seven coefficients.
A test with a more complicated function was made using a
seismogram from a local mine tremor with M L 4.1 recorded at
154-km distance. Because of the shallow source (< 2 km), the
mine tremor caused distinct surface waves. The restituted
displacement of the vertical component is shown in Figure 7.
The original digital record was sampled at 125 Hz. For comparison, a time window containing 4096 samples (Fig. 7) was
transformed with a standard FFT. The normalized spectral
amplitudes of the first 25 harmonics and the offset at the zero
frequency are shown in the bottom of Figure 7. The time window was plotted for the analysis with a horizontal scale of
36 cm, the maximum the MOHA can handle. The first 25
Fourier spectral amplitudes are also shown in Figure 7. The
value for the zero frequency was determined with the planimeter without the use of MOHA by measuring separately the
area under the positive and negative amplitudes. Coefficients
20, 22, and 24 were determined in sections by dividing the
seismogram into two parts of equal length and using cogwheels
10, 11, and 12, respectively. The averages of the two cosine and
sine coefficients of the two sections are calculated, which is a
necessity as no cogwheels for these coefficients exist. However,
as stated earlier, this method also enabled derivation of higher
coefficients than the first 25.

Figure 6. Boxplot of the bias between the theoretical first 10

Fourier coefficients of the analytic signal shown in Figure 5
and the results from the analysis with MOHA performed independently by seven people. The x axis shows the first 10 cosine (c) and
sine (s) coefficients. The horizontal lines within the gray boxes
show the median values of all measurements. The boxes indicate
the first and third quartile, the bars indicate maximum and minimum values, and the filled circle shows outliers. All coefficients at
the right side of the vertical line include the application of gear
wheels during their determination.

Seismological Research Letters

Volume 84, Number 2

March/April 2013

255

 Figure 7. (a) Displacement seismogram of the vertical component of ground motion of a shallow (< 2 km) M L 4.1 earthquake
recorded at 154-km distance. The time window marked by the
thicker line and the filled positive amplitudes were used to determine the first 25 Fourier amplitudes with MOHA, shown by diamonds in (b). For comparison, the spectral amplitudes from a
standard FFT of the digital time series are shown as open circles.
(b) The first 25 Fourier coefficients for the signal shown above from
the analysis with MOHA versus coefficients determined by FFT.
MATHMOHA
As a demonstration tool, but also to pay tribute to the engineering ingenuity of Mader and Ott, we generated mathMOHA
(Fig. 8), a fully interactive numerical model of MOHA in an
open CDF format (http://www.wolfram.com/cdf/, last accessed
January 2013). It is attached to this paper as an electronic
supplement and can be run with the freely available CDF player
(http://www.wolfram.com/cdf-player/, last accessed January
2013). mathMOHA illustrates the determination of the Fourier
series coefficients a1 to a10 and b1 to b10 for a set of selected time
series, which includes the ones discussed in this paper. mathMOHA can operate in two modes of operation.
In the Freehand mode, the users control the trace point P
(see Fig. 8) with their mouse. Once the end of the trace is
reached, the trace point can be automatically returned to
the origin by first moving vertically to the zero line and then
back to the origin by hitting the Automatic Finish button (see
Fig. 8). This way, the tracking error is due purely to the imprecision of the user in following the actual trace and not from
closing the loop to the origin.
In the Autotracking mode, the actual tracking error is further reduced to the numerical and sampling error because the
user moves the trace point with a slider along the time series
and back to the origin. This mode is particularly instructive for
256

Seismological Research Letters

Volume 84, Number 2

Figure 8. Screen copy of mathMOHA, a fully interactive

numerical model of MOHA in an open CDF format (see electronic supplement).
the illustration of the working principle of MOHA. Simultaneous with the movement of the trace point, the simulated
planimeter (the framed yellow rectangle attached to the planimeter pivot point in Fig. 8), which is a simple implementation of equation (10), displays the momentary area within the
trace of the selected cogwheel pin. The cogwheel pins and the
corresponding pin traces are colored red and blue for the b and
a coefficients, respectively. Figure 8 displays the virtual model
corresponding to the actual setup as shown in Figure 1a.

CONCLUSION
With the reactivation of the mechanical harmonic analyzer, we
have been able to reproduce earlier results (Mader, 1909; Korschunow, 1955) that showed that at least the lowest 25 Fourier
coefficients can be determined with the analog MOHA, usually with errors smaller than 8%. Considering the fact that the
analysis of the first 10 coefficients of our test signal took on
average about 2 hr, the title of this contribution seems justified.
We suggest that students in the course of learning about signal
processing should, as an exercise, perform a slow Fourier transform as a way to deeply understand that Fourier integrals are
defined by the area under the curve to be analyzed. As original
MOHAs have become rare instruments, the electronic version
serves as a viable alternative. FFT routines can easily be added
March/April 2013

to analysis codes and are indispensable to modern signal

processing. However, in many ways similar to a finely crafted
instrument in the hands of a well-trained musician, there is
something quite special and intrinsically satisfying about the
feeling of working with a masterpiece of craftsmanship such
as MOHA.
Harmonic analysis of signals is of interest in many disciplines in addition to seismology, and as alluded to in the
Introduction, it is of special interest in gravity and tidal studies.
Parker (2011) reported that harmonic analysis and synthesis of
tides were key features for the success of the D-day invasion
during World War II. Spies behind the German lines had carefully studied the local tides along the coast of Normandy. These
empirical time series were analyzed, most probably with an instrument such as the one presented in this contribution. The
resulting Fourier coefficients were the key for the precise prediction of the height and timing of the tide at beaches Omaha,
Gold, and Juno. For the synthesis of those data, mechanical
computing devices that are the size of a closet and much more
impressive than MOHA were used.

ACKNOWLEDGMENTS
We thank S. Falter, C. Fleischer, S. Gerz, H. Kehmeier, G.
Schweppe, and J. Tzislakis for taking their time to make a slow
Fourier transform. Comments from the Editor, J. Ebel, helped
to significantly improve the manuscript. Most of all, we thank
Ali Akasya for making the instrument available that was used
in our study.

REFERENCES
Bracewell, R. N. (1986). The Fourier Transform and Applications, Second
Ed. (revised), McGraw-Hill, New York, 496 pp.
Cooley, J. W., and J. W. Tukey (1965). An algorithm for the machine
calculation of complex Fourier series, Math. Comput. 19, 297301.
Cooley, J. W., P. A. W. Lewis, and P. D. Welch (1967). Historical
notes on the fast Fourier transform, IEEE Trans. Audio Electroacoust. 15, no. 2, 7679.
Dewey, J., and P. Byerly (1969). The early history of seismometry
(to 1900), Bull. Seismol. Soc. Am. 59, 183227.
Fickinger, W. (2005). Physics at a Research University: Case Western
Reserve, 18301990, Case Western Reserve University, Cleveland,
Ohio, 360 pp.
Gubbins, D. (2006). Time Series Analysis and Inverse Theory in Geophysics, Cambridge University Press, Cambridge, 255 pp.
Henrici, O. (1894). On a new harmonic analyser, Phil. Mag. J. Sci. 38,
110129.

Hrivel, J. (1975). Joseph Fourier, the Man and the Physicist, Clarendon,
London, 362 pp.
Kanasewich, E. R. (1981). Time Sequence Analysis in Geophysics,
Univ. Alberta Pr., Edmonton, 480 pp.
Kolmogorov, A. (1939). Interpolation and extrapolation, Bull. Acad. Sci.
USSR Ser. Math 5, 514.
Korschunow, A. (1955). On the reliability of harmonic analysis of
seismograms, Geophys. Prospect. 4, 303309.
Lam, M. (1853). Memoire sur lequilibre dlasticit des envelopes
sphriques, Comptes Rendus 37, 5187.
Mader, O. (1909). Ein einfacher harmonischer Analysator mit beliebiger
Basis, Elektron. Z. 36, 15.
Miller, D. C. (1916). A 32-element harmonic analyzer, J. Franklin Inst.
181, no. 1, 5181.
Miller, D. C. (1933). The ether-drift experiment and the determination
of the absolute motion of the earth, Rev. Mod. Phys. 5,
203242.
Nassau, J. J., and P. M. Morse (1927). A study of solar motion, Astrophys.
J. 65, 73.
Otnes, R. K. (2008). Notes on mechanical Fourier analyzers, J. Oughtred
Soc. 17, 3441.
Ott, A. (1931). Der harmonische Analysator MaderOtt., Druckschrift
Ad. 341 des Mathe.-mech. Instituts, Kempten Allgu.
Parker, B. (2011). The tide predictions for D-day, Phys. Today 64, 3540.
Wadsworth, G. P., E. A. Robinson, J. G. Bryan, and P. M. Hurley (1953).
Detection of reflections on seismic records by linear operators,
Geophysics 18, 539586.
Wiener, N. (1930). Generalized harmonic analysis, Acta Math. 55,
117258.
Wiener, N. (1949). Extrapolation, Interpolation and Smoothing of
Stationary Time Series, Wiley, New York, 163 pp.
Yule, G. U. (1894). On a simple form of harmonic analyser, Proc. Phys.
Soc. Lond., 13, 403412, available at http://iopscience.iop.org/
14787814/13/1/336 (last accessed January 2013).

Klaus-G. Hinzen
Sharon K. Reamer
Earthquake Geology Group and Archaeoseismology
Institute for Geology and Mineralogy
University of Cologne
Vinzenz-Pallotti-Str. 26
D-51429 Bergisch Gladbach, Germany

Seismological Research Letters

[email protected]

Frank Scherbaum
Institute of Earth and Environmental Science
Potsdam University
Karl-Liebknecht-Str. 24-25
D-14476 Potsdam-Golm, Germany

Volume 84, Number 2

March/April 2013

257