Fast Fourier Transform

Key Papers in Computer Science

Seminar 2005

Dima Batenkov
Weizmann Institute of Science

dima.batenkov@gmail.com

p.1/33
 Fast Fourier Transform - Overview

» Fast Fourier Transform -

Overview
J. W. Cooley and J. W. Tukey. An algorithm for the machine
calculation of complex Fourier series. Mathematics of
Computation, 19:297–301, 1965

p.2/33
 Fast Fourier Transform - Overview

» Fast Fourier Transform -

Overview
J. W. Cooley and J. W. Tukey. An algorithm for the machine
calculation of complex Fourier series. Mathematics of
Computation, 19:297–301, 1965
 A fast algorithm for computing the Discrete Fourier
Transform

p.2/33
 Fast Fourier Transform - Overview

» Fast Fourier Transform -

Overview
J. W. Cooley and J. W. Tukey. An algorithm for the machine
calculation of complex Fourier series. Mathematics of
Computation, 19:297–301, 1965
 A fast algorithm for computing the Discrete Fourier
Transform
 (Re)discovered by Cooley & Tukey in 1965 1 and widely
adopted thereafter

p.2/33
 Fast Fourier Transform - Overview

» Fast Fourier Transform -

Overview
J. W. Cooley and J. W. Tukey. An algorithm for the machine
calculation of complex Fourier series. Mathematics of
Computation, 19:297–301, 1965
 A fast algorithm for computing the Discrete Fourier
Transform
 (Re)discovered by Cooley & Tukey in 1965 1 and widely
adopted thereafter
 Has a long and fascinating history

p.2/33
» Fast Fourier Transform -
Overview

Fourier Analysis
» Fourier Series
» Continuous Fourier Transform
» Discrete Fourier Transform
» Useful properties 6
» Applications Fourier Analysis

p.3/33
 Fourier Series

» Fast Fourier Transform -

Overview
 Expresses a (real) periodic function x (t) as a sum of
Fourier Analysis
trigonometric series (− L < t < L)
» Fourier Series
∞
» Continuous Fourier Transform
1 πn πn
x (t) = a0 + ∑ ( an cos t + bn sin t)
» Discrete Fourier Transform

L L
» Useful properties 6
» Applications 2 n =1

 Coefficients can be computed by

Z
1 L πn
an = x (t) cos t dt
L −L L
Z
1 L πn
bn = x (t) sin t dt
L −L L

p.4/33
 Fourier Series

» Fast Fourier Transform -

Overview
 Generalized to complex-valued functions as
Fourier Analysis ∞
c n ei
πn t
x (t) = ∑
» Fourier Series
» Continuous Fourier Transform
L
» Discrete Fourier Transform
» Useful properties 6
n =− ∞
Z L
» Applications
1
x ( t ) e −i L t
πn
cn = dt
2L −L

p.4/33
 Fourier Series

» Fast Fourier Transform -

Overview
 Generalized to complex-valued functions as
Fourier Analysis ∞
c n ei
πn t
x (t) = ∑
» Fourier Series
» Continuous Fourier Transform
L
» Discrete Fourier Transform
» Useful properties 6
n =− ∞
Z L
» Applications
1
x ( t ) e −i L t
πn
cn = dt
2L −L

 Studied by D.Bernoulli and L.Euler

 Used by Fourier to solve the heat equation

p.4/33
 Fourier Series

» Fast Fourier Transform -

Overview
 Generalized to complex-valued functions as
Fourier Analysis ∞
c n ei
πn t
x (t) = ∑
» Fourier Series
» Continuous Fourier Transform
L
» Discrete Fourier Transform
» Useful properties 6
n =− ∞
Z L
» Applications
1
x ( t ) e −i L t
πn
cn = dt
2L −L

 Studied by D.Bernoulli and L.Euler

 Used by Fourier to solve the heat equation
 Converges for almost all “nice” functions (piecewise smooth,
L2 etc.)

p.4/33
 Continuous Fourier Transform

» Fast Fourier Transform -

Overview
 Generalization of Fourier Series for infinite domains
Z ∞
F ( f )e−2πi f t d f
Fourier Analysis
» Fourier Series x (t) =
» Continuous Fourier Transform
−∞
» Discrete Fourier Transform
» Useful properties 6
Z ∞
» Applications
F(f) = x (t)e2πi f t dt
−∞

 Can represent continuous, aperiodic signals

 Continuous frequency spectrum

p.5/33
 Discrete Fourier Transform

» Fast Fourier Transform -

Overview
 If the signal X (k) is periodic, band-limited and sampled at
Fourier Analysis
Nyquist frequency or higher, the DFT represents the CFT
» Fourier Series
» Continuous Fourier Transform
exactly 14
» Discrete Fourier Transform
» Useful properties 6
N −1
∑ X (k)Wrk
» Applications
A (r ) = N
k =0

2πi
where W N = e− N and r = 0, 1, . . . , N − 1
 The inverse transform:
N −1
1 − jk
X ( j) =
N ∑ A(k)W N
k =0

p.6/33
 Useful properties 6

» Fast Fourier Transform -

Overview
 Orthogonality of basis functions
Fourier Analysis
N −1
WrNk Wrk
0
∑
» Fourier Series
0
» Continuous Fourier Transform
N = Nδ N ( r − r )
» Discrete Fourier Transform
» Useful properties 6 k =0
» Applications

 Linearity
 (Cyclic) Convolution theorem
N −1
( x ∗ y)n = ∑ x (k)y(n − k)
k =0
⇒ F {( x ∗ y)} = F { X }F {Y }
 Symmetries: real ↔ hermitian symmetric, imaginary ↔
hermitian antisymmetric
 Shifting theorems

p.7/33
 Applications

» Fast Fourier Transform -

Overview
 Approximation by trigonometric polynomials
 Data compression (MP3...)
Fourier Analysis
» Fourier Series
» Continuous Fourier Transform
 Spectral analysis of signals
Frequency response of systems
» Discrete Fourier Transform
» Useful properties 6 
» Applications
 Partial Differential Equations
 Convolution via frequency domain
 Cross-correlation
 Multiplication of large integers
 Symbolic multiplication of polynomials

p.8/33
» Fast Fourier Transform -
Overview

Methods known by 1965

» Available methods
» Goertzel’s algorithm 7

Methods known by 1965

p.9/33
 Available methods

» Fast Fourier Transform -

Overview
 In many applications, large digitized datasets are becoming
Methods known by 1965
available, but cannot be processed due to prohibitive
» Available methods
» Goertzel’s algorithm 7
running time of DFT
 All (publicly known) methods for efficient computation utilize
the symmetry of trigonometric functions, but still are O( N 2 )
 Goertzel’s method 7 is fastest known

p.10/33
 Goertzel’s algorithm 7

» Fast Fourier Transform -

Overview
 Given a particular 0 < j < N − 1, we want to evaluate
Methods known by 1965
N −1   N −1  
» Available methods 2πj 2πj
» Goertzel’s algorithm 7 A( j) = ∑ x (k) cos
N
k B( j) = ∑ x (k) sin
N
k
k =0 k =0

p.11/33
 Goertzel’s algorithm 7

» Fast Fourier Transform -

Overview
 Given a particular 0 < j < N − 1, we want to evaluate
Methods known by 1965
N −1   N −1  
» Available methods 2πj 2πj
» Goertzel’s algorithm 7 A( j) = ∑ x (k) cos
N
k B( j) = ∑ x (k) sin
N
k
k =0 k =0

 Recursively compute the sequence {u(s)} sN=+01

 
2πj
u(s) = x (s) + 2 cos u ( s + 1) − u ( s + 2)
N
u ( N ) = u ( N + 1) = 0

p.11/33
 Goertzel’s algorithm 7

» Fast Fourier Transform -

Overview
 Given a particular 0 < j < N − 1, we want to evaluate
Methods known by 1965
N −1   N −1  
» Available methods 2πj 2πj
» Goertzel’s algorithm 7 A( j) = ∑ x (k) cos
N
k B( j) = ∑ x (k) sin
N
k
k =0 k =0

 Recursively compute the sequence {u(s)} sN=+01

 
2πj
u(s) = x (s) + 2 cos u ( s + 1) − u ( s + 2)
N
u ( N ) = u ( N + 1) = 0

 Then

2πj
B( j) = u(1) sin
N
 
2πj
A( j) = x (0) + cos u (1) − u (2)
N
p.11/33
 Goertzel’s algorithm 7

» Fast Fourier Transform -

Overview
 Requires N multiplications and only one sine and cosine
Methods known by 1965  Roundoff errors grow rapidly 5
» Available methods
» Goertzel’s algorithm 7  Excellent for computing a very small number of coefficients
(< log N)
 Can be implemented using only 2 intermediate storage
locations and processing input signal as it arrives
 Used today for Dual-Tone Multi-Frequency detection (tone
dialing)

p.11/33
» Fast Fourier Transform -
Overview

Inspiration for the FFT

» Factorial experiments
» Good’s paper 9
» Good’s results - summary

Inspiration for the FFT

p.12/33
 Factorial experiments

» Fast Fourier Transform -

Overview
Purpose: investigate the effect of n factors on some measured
Inspiration for the FFT
quantity
» Factorial experiments  Each factor can have t distinct levels. For our discussion let
» Good’s paper 9
» Good’s results - summary t = 2, so there are 2 levels: “+” and “−”.
 So we conduct t n experiments (for every possible
combination of the n factors) and calculate:
 Main effect of a factor: what is the average change in the
output as the factor goes from “−” to “+”?
 Interaction effects between two or more factors: what is
the difference in output between the case when both
factors are present compared to when only one of them
is present?

p.13/33
 Factorial experiments

» Fast Fourier Transform -

Overview
Example of t = 2, n = 3 experiment
 Say we have factors A, B, C which can be “high” or “low”. For
Inspiration for the FFT
» Factorial experiments
» Good’s paper 9
example, they can represent levels of 3 different drugs given
» Good’s results - summary to patients
 Perform 23 measurements of some quantity, say blood
pressure
 Investigate how each one of the drugs and their combination
affect the blood pressure

p.13/33
 Factorial experiments

» Fast Fourier Transform -

Overview

Inspiration for the FFT

Exp. Combination A B C Blood pressure
» Factorial experiments
» Good’s paper 9
» Good’s results - summary
1 (1) − − − x1
2 a + − − x2
3 b − + − x3
4 ab + + − x4
5 c − − + x5
6 ac + − + x6
7 bc − + + x7
8 abc + + + x8

p.13/33
 Factorial experiments

» Fast Fourier Transform -

Overview
 The main effect of A is
Inspiration for the FFT
( ab − b) + ( a − (1)) + ( abc − bc) + ( ac − c)
» Factorial experiments
» Good’s paper 9
 The interaction of A, B is
» Good’s results - summary
( ab − b) − ( a − (1)) − (( abc − bc) − ( ac − c))
 etc.

p.13/33
 Factorial experiments

» Fast Fourier Transform -

Overview
 The main effect of A is
Inspiration for the FFT
( ab − b) + ( a − (1)) + ( abc − bc) + ( ac − c)
» Factorial experiments
» Good’s paper 9
 The interaction of A, B is
» Good’s results - summary
( ab − b) − ( a − (1)) − (( abc − bc) − ( ac − c))
 etc.
If the main effects and interactions are arranged in vector y,
then we can write y = Ax where
 
1 1 1 1 1 1 1 1
 
1 −1 −1 1 1 −1 −1 1 
 
1 1 −1 −1 1 1 − 1 − 1 
 
1 −1 −1 1 −1 1 −1 1 
 
A= 
1 1 1 1 −1 −1 −1 −1 
 
1 −1 1 −1 −1 1 −1 1 
 
 
 1 1 − 1 − 1 − 1 − 1 1 1 
1 −1 −1 1 −1 1 1 −1 p.13/33
 Factorial experiments

» Fast Fourier Transform -

Overview
 F.Yates devised 15 an iterative n-step algorithm which
Inspiration for the FFT simplifies the calculation. Equivalent to decomposing
» Factorial experiments
» Good’s paper 9
A = Bn , where (for our case of n = 3):
» Good’s results - summary
 
1 1 0 0 0 0 0 0
 
 0 0 1 1 0 0 0 0 
 
 0 0 0 0 1 1 0 0 
 
 0 0 0 0 0 0 1 1 
 
B= 
 −1 −1 0 0 0 0 0 0 
 
 0 0 −1 −1 0 0 0 0 
 
 
 0 0 0 0 −1 −1 0 0 
0 0 0 0 0 0 −1 −1

so that much less operations are required

p.13/33
 Good’s paper 9

» Fast Fourier Transform -

Overview
I. J. Good. The interaction algorithm and practical Fourier
Inspiration for the FFT
analysis. Journal Roy. Stat. Soc., 20:361–372, 1958
» Factorial experiments  Essentially develops a “prime-factor” FFT
» Good’s paper 9
» Good’s results - summary
 The idea is inspired by Yates’ efficient algorithm, but Good
proves general theorems which can be applied to the task of
efficiently multiplying a N-vector by certain N × N matrices
 The work remains unnoticed until Cooley-Tukey publication
 Good meets Tukey in 1956 and tells him briefly about his
method
 It is revealed later 4 that the FFT had not been discovered
much earlier by a coincedence

p.14/33
 Good’s paper 9

» Fast Fourier Transform -

Overview Definition 1 Let M (i) be t × t matrices, i = 1 . . . n. Then the
Inspiration for the FFT direct product
M (1) × M (2) × · · · × M ( n )
» Factorial experiments
» Good’s paper 9
» Good’s results - summary

is a tn × tn matrix C whose elements are

n
(i )
Cr,s = ∏ Mri ,si
i =1

where r = (r1 , . . . , rn ) and s = (s1 , . . . sn ) are the base-t

representation of the index

p.14/33
 Good’s paper 9

» Fast Fourier Transform -

Overview Definition 1 Let M (i) be t × t matrices, i = 1 . . . n. Then the
Inspiration for the FFT direct product
M (1) × M (2) × · · · × M ( n )
» Factorial experiments
» Good’s paper 9
» Good’s results - summary

is a tn × tn matrix C whose elements are

n
(i )
Cr,s = ∏ Mri ,si
i =1

where r = (r1 , . . . , rn ) and s = (s1 , . . . sn ) are the base-t

representation of the index
Theorem 1 For every M, M [n] = An where
s s
Ar,s = { Mr1 ,sn δr21 δrs32 . . . δrnn−1 }

Interpretation: A has at most t n+1 nonzero elements. So if, for

a given B we can find M such that M [n] = B, then we can
compute Bx = An x in O(ntn ) instead of O(t2n ). p.14/33
 Good’s paper 9

» Fast Fourier Transform -

Overview
Good’s observations
 For 2n factorial experiment:
Inspiration for the FFT
» Factorial experiments
» Good’s paper 9
!
» Good’s results - summary 1 1
M=
1 −1

p.14/33
 Good’s paper 9

» Fast Fourier Transform -

Overview
Good’s observations
 n-dimensional DFT, size N in each dimension
Inspiration for the FFT
» Factorial experiments
» Good’s paper 9 ( N −1,...,N −1)
r s
∑ x (r)W N1 1 × · · · × WrNn sn
» Good’s results - summary
A ( s1 , . . . , s n ) =
(r1 ,...,rn )

can be represented as

A = Ω[n] x, Ω = {Wrs
N} r, s = 0, . . . , N − 1

By the theorem, A = Ψ n x where Ψ is sparse

p.14/33
 Good’s paper 9

» Fast Fourier Transform -

Overview
Good’s observations
 n-dimensional DFT, size N in each dimension
Inspiration for the FFT
» Factorial experiments
» Good’s paper 9 ( N −1,...,N −1)
r s
∑ x (r)W N1 1 × · · · × WrNn sn
» Good’s results - summary
A ( s1 , . . . , s n ) =
(r1 ,...,rn )

can be represented as

A = Ω[n] x, Ω = {Wrs
N} r, s = 0, . . . , N − 1

By the theorem, A = Ψ n x where Ψ is sparse

 n-dimensional DFT, Ni in each dimension

A = ( Ω (1) × · · · × Ω ( n ) ) x = Φ (1) Φ (2) . . . Φ ( n ) x

Ω(i) = {Wrs
Ni }

Φ(i) = IN1 × · · · × Ω(i) × · · · × I Nn

p.14/33
 Good’s paper 9

» Fast Fourier Transform -

Overview
If N = N1 N2 and ( N1 , N2 ) = 1 then N-point 1-D DFT can be
Inspiration for the FFT
represented as 2-D DFT
» Factorial experiments
» Good’s paper 9
» Good’s results - summary
r = N2 r1 + N1 r2
−1
(Chinese Remainder Thm.) s = N2 s 1 ( N2 ) mod N1 + N1 s2 ( N1−1 ) mod N2
⇒ s ≡ s1 mod N1 , s ≡ s2 mod N2

These are index mappings s ↔ (s1 , s2 ) and r ↔ (r1 , r2 )

N N1 −1 N2 −1
A ( s ) = A ( s1 , s2 ) = ∑ x(r)WrsN = ∑ ∑ x (r1 , r2 )Wrs
N1 N2
r =0 r1 =0 r2 =0
N1 −1 N2 −1
r s
(substitute r, s from above) = ∑ ∑ x (r1 , r2 )W N11 1 WrN22s2
r1 =0 r2 =0

This is exactly 2-D DFT seen above!

p.14/33
 Good’s paper 9

» Fast Fourier Transform -

Overview
So, as above, we can write
Inspiration for the FFT
» Factorial experiments A = Ωx
» Good’s paper 9
» Good’s results - summary
= P ( Ω (1) × Ω (2) ) Q −1 x
= PΦ(1) Φ(2) Q−1 x

where P and Q are permutation matrices. This results in

N ( N1 + N2 ) multiplications instead of N 2 .

p.14/33
 Good’s results - summary

» Fast Fourier Transform -

Overview
 n-dimensionanal, size Ni in each dimension
Inspiration for the FFT ! ! !2
» Factorial experiments
» Good’s paper 9
» Good’s results - summary ∑ Ni ∏ Ni instead of ∏ Ni
n n n

 1-D, size N = ∏n Ni where Ni ’s are mutually prime

!
∑ Ni N instead of N 2
n

 Still thinks that other “tricks” such as using the symmetries

and sines and cosines of known angles might be more
useful

p.15/33
» Fast Fourier Transform -
Overview

The article
» James W.Cooley (1926-)
» John Wilder Tukey
(1915-2000)
» The story - meeting
» The story - publication
» What if...
The article
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.16/33
 James W.Cooley (1926-)

» Fast Fourier Transform -

Overview
 1949 - B.A. in Mathematics
The article
(Manhattan College, NY)
» James W.Cooley (1926-)
» John Wilder Tukey
 1951 - M.A. in Mathematics
(1915-2000)
» The story - meeting
(Columbia University)
» The story - publication
» What if...  1961 - Ph.D in Applied Math-
» Cooley-Tukey FFT
» FFT properties ematics (Columbia Univer-
» DIT signal flow
» DIF signal flow
sity)
 1962 - Joins IBM Watson Research Center
 The FFT is considered his most significant contribution to
mathematics
 Member of IEEE Digital Signal Processing Committee
 Awarded IEEE fellowship on his works on FFT
 Contributed much to establishing the terminology of DSP

p.17/33
 John Wilder Tukey (1915-2000)

» Fast Fourier Transform -

Overview
 1936 - Bachelor in Chem-
The article
istry (Brown University)
» James W.Cooley (1926-)
» John Wilder Tukey
 1937 - Master in Chemistry
(1915-2000)
» The story - meeting (Brown University)
» The story - publication
» What if...  1939 - PhD in Mathemat-
ics (Princeton) for “Denu-
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow
merability in Topology”
 During WWII joins Fire Control Research Office and
contributes to war effort (mainly invlolving statistics)
 From 1945 also works with Shannon and Hamming in AT&T
Bell Labs
 Besides co-authoring the FFT in 1965, has several major
contributions to statistics

p.18/33
 The story - meeting

» Fast Fourier Transform -

Overview
 Richard Garwin (Columbia University IBM Watson Research
The article
center) meets John Tukey (professor of mathematics in
» James W.Cooley (1926-)
» John Wilder Tukey
Princeton) at Kennedy’s Scientific Advisory Committee in
(1915-2000)
» The story - meeting
1963
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.19/33
 The story - meeting

» Fast Fourier Transform -

Overview
 Richard Garwin (Columbia University IBM Watson Research
The article
center) meets John Tukey (professor of mathematics in
» James W.Cooley (1926-)
» John Wilder Tukey
Princeton) at Kennedy’s Scientific Advisory Committee in
(1915-2000)
» The story - meeting
1963
» The story - publication  Tukey shows how a Fourier series of length N when
N = ab is composite can be expressed as a-point series
» What if...
» Cooley-Tukey FFT

of b-point subseries, thereby reducing the number of

» FFT properties
» DIT signal flow

computations from N 2 to N ( a + b)
» DIF signal flow

p.19/33
 The story - meeting

» Fast Fourier Transform -

Overview
 Richard Garwin (Columbia University IBM Watson Research
The article
center) meets John Tukey (professor of mathematics in
» James W.Cooley (1926-)
» John Wilder Tukey
Princeton) at Kennedy’s Scientific Advisory Committee in
(1915-2000)
» The story - meeting
1963
» The story - publication  Tukey shows how a Fourier series of length N when
N = ab is composite can be expressed as a-point series
» What if...
» Cooley-Tukey FFT

of b-point subseries, thereby reducing the number of

» FFT properties
» DIT signal flow

computations from N 2 to N ( a + b)
» DIF signal flow

 Notes that if N is a power of two, the algorithm can be

repeated, giving N log2 N operations

p.19/33
 The story - meeting

» Fast Fourier Transform -

Overview
 Richard Garwin (Columbia University IBM Watson Research
The article
center) meets John Tukey (professor of mathematics in
» James W.Cooley (1926-)
» John Wilder Tukey
Princeton) at Kennedy’s Scientific Advisory Committee in
(1915-2000)
» The story - meeting
1963
» The story - publication  Tukey shows how a Fourier series of length N when
N = ab is composite can be expressed as a-point series
» What if...
» Cooley-Tukey FFT

of b-point subseries, thereby reducing the number of

» FFT properties
» DIT signal flow

computations from N 2 to N ( a + b)
» DIF signal flow

 Notes that if N is a power of two, the algorithm can be

repeated, giving N log2 N operations
 Garwin, being aware of applications in various fields that
would greatly benefit from an efficient algorithm for DFT,
realizes the importance of this and from now on,
“pushes” the FFT at every opportunity

p.19/33
 The story - publication

» Fast Fourier Transform -

Overview
 Garwin goes to Jim Cooley (IBM Watson Research Center)
The article
with the notes from conversation with Tukey and asks to
» James W.Cooley (1926-)
» John Wilder Tukey
implement the algorithm
(1915-2000)
» The story - meeting
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.20/33
 The story - publication

» Fast Fourier Transform -

Overview
 Garwin goes to Jim Cooley (IBM Watson Research Center)
The article
with the notes from conversation with Tukey and asks to
» James W.Cooley (1926-)
» John Wilder Tukey
implement the algorithm
(1915-2000)
» The story - meeting
 Says that he needs the FFT for computing the 3-D Fourier
» The story - publication
» What if...
transform of spin orientations of He3 . It is revealed later that
» Cooley-Tukey FFT
» FFT properties
what Garwin really wanted was to be able to detect nuclear
» DIT signal flow
» DIF signal flow
exposions from seismic data

p.20/33
 The story - publication

» Fast Fourier Transform -

Overview
 Garwin goes to Jim Cooley (IBM Watson Research Center)
The article
with the notes from conversation with Tukey and asks to
» James W.Cooley (1926-)
» John Wilder Tukey
implement the algorithm
(1915-2000)
» The story - meeting
 Says that he needs the FFT for computing the 3-D Fourier
» The story - publication
» What if...
transform of spin orientations of He3 . It is revealed later that
» Cooley-Tukey FFT
» FFT properties
what Garwin really wanted was to be able to detect nuclear
» DIT signal flow
» DIF signal flow
exposions from seismic data
 Cooley finally writes a 3-D in-place FFT and sends it to
Garwin

p.20/33
 The story - publication

» Fast Fourier Transform -

Overview
 Garwin goes to Jim Cooley (IBM Watson Research Center)
The article
with the notes from conversation with Tukey and asks to
» James W.Cooley (1926-)
» John Wilder Tukey
implement the algorithm
(1915-2000)
» The story - meeting
 Says that he needs the FFT for computing the 3-D Fourier
» The story - publication
» What if...
transform of spin orientations of He3 . It is revealed later that
» Cooley-Tukey FFT
» FFT properties
what Garwin really wanted was to be able to detect nuclear
» DIT signal flow
» DIF signal flow
exposions from seismic data
 Cooley finally writes a 3-D in-place FFT and sends it to
Garwin
 The FFT is exposed to a number of mathematicians

p.20/33
 The story - publication

» Fast Fourier Transform -

Overview
 Garwin goes to Jim Cooley (IBM Watson Research Center)
The article
with the notes from conversation with Tukey and asks to
» James W.Cooley (1926-)
» John Wilder Tukey
implement the algorithm
(1915-2000)
» The story - meeting
 Says that he needs the FFT for computing the 3-D Fourier
» The story - publication
» What if...
transform of spin orientations of He3 . It is revealed later that
» Cooley-Tukey FFT
» FFT properties
what Garwin really wanted was to be able to detect nuclear
» DIT signal flow
» DIF signal flow
exposions from seismic data
 Cooley finally writes a 3-D in-place FFT and sends it to
Garwin
 The FFT is exposed to a number of mathematicians
 It is decided that the algorithm should be put in the public
domain to prevent patenting

p.20/33
 The story - publication

» Fast Fourier Transform -

Overview
 Garwin goes to Jim Cooley (IBM Watson Research Center)
The article
with the notes from conversation with Tukey and asks to
» James W.Cooley (1926-)
» John Wilder Tukey
implement the algorithm
(1915-2000)
» The story - meeting
 Says that he needs the FFT for computing the 3-D Fourier
» The story - publication
» What if...
transform of spin orientations of He3 . It is revealed later that
» Cooley-Tukey FFT
» FFT properties
what Garwin really wanted was to be able to detect nuclear
» DIT signal flow
» DIF signal flow
exposions from seismic data
 Cooley finally writes a 3-D in-place FFT and sends it to
Garwin
 The FFT is exposed to a number of mathematicians
 It is decided that the algorithm should be put in the public
domain to prevent patenting
 Cooley writes the draft, Tukey adds some references to
Good’s article and the paper is submitted to Mathematics of
Computation in August 1964

p.20/33
 The story - publication

» Fast Fourier Transform -

Overview
 Garwin goes to Jim Cooley (IBM Watson Research Center)
The article
with the notes from conversation with Tukey and asks to
» James W.Cooley (1926-)
» John Wilder Tukey
implement the algorithm
(1915-2000)
» The story - meeting
 Says that he needs the FFT for computing the 3-D Fourier
» The story - publication
» What if...
transform of spin orientations of He3 . It is revealed later that
» Cooley-Tukey FFT
» FFT properties
what Garwin really wanted was to be able to detect nuclear
» DIT signal flow
» DIF signal flow
exposions from seismic data
 Cooley finally writes a 3-D in-place FFT and sends it to
Garwin
 The FFT is exposed to a number of mathematicians
 It is decided that the algorithm should be put in the public
domain to prevent patenting
 Cooley writes the draft, Tukey adds some references to
Good’s article and the paper is submitted to Mathematics of
Computation in August 1964
 Published in April 1965
p.20/33
 What if...

» Fast Fourier Transform -

Overview
 Good reveals in 1993 4 that Garwin visited him in 1957, but:
The article  “Garwin told me with great enthusiasm about his
» James W.Cooley (1926-)
» John Wilder Tukey experimental work... Unfortunately, not wanting to change
(1915-2000)
» The story - meeting the subject..., I didn’t mention my FFT work at that dinner,
although I had intended to do so. He wrote me on February
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
9, 1976 and said: “Had we talked about it in 1957, I could
» DIT signal flow
» DIF signal flow
have stolen [publicized] it from you instead of from John
Tukey...”... That would have completely changed the history
of the FFT”

p.21/33
Cooley-Tukey FFT

jk
A( j) = ∑kN=−01 X (k)W N N = N1 N2

p.22/33
Cooley-Tukey FFT

jk
A( j) = ∑kN=−01 X (k)W N N = N1 N2
 Convert the 1-D indices into 2-D

j = j0 + j1 N1 j0 = 0, . . . , N1 − 1 j1 = 0, . . . , N2 − 1
k = k 0 + k1 N2 k0 = 0, . . . , N2 − 1 k 1 = 0, . . . , N1 − 1

p.22/33
Cooley-Tukey FFT

jk
A( j) = ∑kN=−01 X (k)W N N = N1 N2
 Convert the 1-D indices into 2-D

j = j0 + j1 N1 j0 = 0, . . . , N1 − 1 j1 = 0, . . . , N2 − 1
k = k 0 + k1 N2 k0 = 0, . . . , N2 − 1 k 1 = 0, . . . , N1 − 1

 “Decimate” the original sequence into N2 N1 -subsequences:

p.22/33
Cooley-Tukey FFT

jk
A( j) = ∑kN=−01 X (k)W N N = N1 N2
 Convert the 1-D indices into 2-D

j = j0 + j1 N1 j0 = 0, . . . , N1 − 1 j1 = 0, . . . , N2 − 1
k = k 0 + k1 N2 k0 = 0, . . . , N2 − 1 k 1 = 0, . . . , N1 − 1

 “Decimate” the original sequence into N2 N1 -subsequences:

N −1
jk
A( j1 , j0 ) = ∑ X (k)W N
k =0

p.22/33
Cooley-Tukey FFT

jk
A( j) = ∑kN=−01 X (k)W N N = N1 N2
 Convert the 1-D indices into 2-D

j = j0 + j1 N1 j0 = 0, . . . , N1 − 1 j1 = 0, . . . , N2 − 1
k = k 0 + k1 N2 k0 = 0, . . . , N2 − 1 k 1 = 0, . . . , N1 − 1

 “Decimate” the original sequence into N2 N1 -subsequences:

N −1
jk
A( j1 , j0 ) = ∑ X (k)W N
k =0
N2 −1 N1 −1
( j + j N1 )(k 0 +k1 N2 )
= ∑ ∑ X (k1 , k 0 )W N01 N21
k 0 =0 k 1 =0

p.22/33
Cooley-Tukey FFT

jk
A( j) = ∑kN=−01 X (k)W N N = N1 N2
 Convert the 1-D indices into 2-D

j = j0 + j1 N1 j0 = 0, . . . , N1 − 1 j1 = 0, . . . , N2 − 1
k = k 0 + k1 N2 k0 = 0, . . . , N2 − 1 k 1 = 0, . . . , N1 − 1

 “Decimate” the original sequence into N2 N1 -subsequences:

N −1
jk
A( j1 , j0 ) = ∑ X (k)W N
k =0
N2 −1 N1 −1
( j + j N1 )(k 0 +k1 N2 )
= ∑ ∑ X (k1 , k 0 )W N01 N21
k 0 =0 k 1 =0
N2 −1 N1 −1
j k j k j k
= ∑ W N0 0 W N1 2 0 ∑ X (k1 , k 0 )W N0 1 1
k 0 =0 k 1 =0

p.22/33
Cooley-Tukey FFT

jk
A( j) = ∑kN=−01 X (k)W N N = N1 N2
 Convert the 1-D indices into 2-D

j = j0 + j1 N1 j0 = 0, . . . , N1 − 1 j1 = 0, . . . , N2 − 1
k = k 0 + k1 N2 k0 = 0, . . . , N2 − 1 k 1 = 0, . . . , N1 − 1

 “Decimate” the original sequence into N2 N1 -subsequences:

N −1
jk
A( j1 , j0 ) = ∑ X (k)W N
k =0
N2 −1 N1 −1
( j + j N1 )(k 0 +k1 N2 )
= ∑ ∑ X (k1 , k 0 )W N01 N21
k 0 =0 k 1 =0
N2 −1 N1 −1 N2 −1
j k j k j k j k j k
= ∑ W N0 0 W N1 2 0 ∑ X (k1 , k 0 )W N0 1 1 = ∑ W N0 0 X j0 (k0 )W N1 2 0
k 0 =0 k 1 =0 k 0 =0

p.22/33
Cooley-Tukey FFT

 The sequences X j0 have N elements and require NN1 operations to

compute
 Given these, the array A requires NN2 operations to compute
 Overall N ( N1 + N2 )
 In general, if N = ∏in=0 Ni then N (∑in=0 Ni ) operations overall.

p.22/33
Cooley-Tukey FFT

If N1 = N2 = · · · = Nn = 2 (i.e. N = 2n ):

j = jn−1 2n−1 + · · · + j0 , k = k n −1 2n −1 + · · · + k 0
1 1
jk
A( jn−1 , . . . , j0 ) = ∑ ··· ∑ X ( k n −1 , . . . , k 0 ) W N
k0 =0 k n −1 = 0
jk jk ×2 jkn−1 ×2n−1
=∑ W N0 ∑ W N1 · · · ∑ WN X ( k n −1 , . . . , k 0 )
k0 k1 k n −1
j k n −1
⇒ X j0 (k n−2 , . . . , k0 ) = ∑ X (k n−1 , . . . , k0 )W20
k n −1
j k n −2 j k n −2
X j1 ,j0 (k n−3 , . . . , k0 ) = ∑ W40 X j0 (k n−2 , . . . , k0 )W21
k n −2
( j0 +···+ jm−2 ×2m−2 ) kn−m
X jm−1 ,...,j0 (k n−m−1, . . . , k0 ) = ∑ W 2m X jm−2 ,...,j0 (k n−m , . . . , k0 )×
k n−m
j k n−m
× W2m−1

p.22/33
Cooley-Tukey FFT

Finally, A( jn−1 , . . . , j0 ) = X jn−1 ,...,j0

 Two important properties:
 In-place: only two terms are involved in the calculation of every
intermediate pair X jm−1 ,...,j0 (. . . ) jm−1 = 0, 1:
• Xj (0, . . . )
m−2 ,...,j0
• Xj (1, . . . )
m−2 ,...,j0
So the values can be overwritten and no more additional storage be
used
 Bit-reversal
• It is convenient to store X j (k n−m−1 , . . . , k 0 ) at index
m−1 ,...,j0
j0 2n−1 + · · · + jm−1 2n−m + k n−m−1 2m−n−1 + · · · + k 0
• The result must be reversed
 Cooley & Tukey didn’t realize that the algorithm in fact builds large
DFT’s from smaller ones with some multiplications along the way.
( j0 +···+ jm−2 ×2m−2 )k n−m
 The terms W 2m are later called 6 “twiddle factors”.
p.22/33
 FFT properties

» Fast Fourier Transform -

Overview
 Roundoff error significantly reduced compared to defining
The article
formula 6
A lower bound of 12 n log2 n operations over C for linear
» James W.Cooley (1926-)
» John Wilder Tukey 
(1915-2000)
» The story - meeting
» The story - publication
algorithms is proved 12 , so FFT is in a sense optimal
» What if...
» Cooley-Tukey FFT
 Two “canonical” FFTs
» FFT properties  Decimation-in-Time: the Cooley & Tukey version.
» DIT signal flow
» DIF signal flow Equivalent to taking N1 = N/2, N2 = 2. Then X j0 (0)
are the DFT coefficients of even-numbered samples and
X j0 (1) - those of odd-numbered. The final coefficients are
simply linear combination of the two.
 Decimation-in-Frequency (Tukey-Sande 6 ):
N1 = 2, N2 = N/2. Then Y (∗) = X0 (∗) is the sum of
even-numbered and odd-numbered samples, and
Z (∗) = X1 (∗) is the difference. The even Fourier
coefficients are the DFT of Y, and the odd-numbered -
“weighted” DFT of Z.
p.23/33
 DIT signal flow

» Fast Fourier Transform -

Overview

The article
» James W.Cooley (1926-)
» John Wilder Tukey
(1915-2000)
» The story - meeting
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.24/33
 DIT signal flow

» Fast Fourier Transform -

Overview

The article
» James W.Cooley (1926-)
» John Wilder Tukey
(1915-2000)
» The story - meeting
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.24/33
 DIT signal flow

» Fast Fourier Transform -

Overview

The article
» James W.Cooley (1926-)
» John Wilder Tukey
(1915-2000)
» The story - meeting
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.24/33
 DIF signal flow

» Fast Fourier Transform -

Overview

The article
» James W.Cooley (1926-)
» John Wilder Tukey
(1915-2000)
» The story - meeting
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.25/33
 DIF signal flow

» Fast Fourier Transform -

Overview

The article
» James W.Cooley (1926-)
» John Wilder Tukey
(1915-2000)
» The story - meeting
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.25/33
 DIF signal flow

» Fast Fourier Transform -

Overview

The article
» James W.Cooley (1926-)
» John Wilder Tukey
(1915-2000)
» The story - meeting
» The story - publication
» What if...
» Cooley-Tukey FFT
» FFT properties
» DIT signal flow
» DIF signal flow

p.25/33
» Fast Fourier Transform -
Overview

Newly discovered history of FFT

» What happened after the
publication
» Danielson-Lanczos
» Gauss
Newly discovered history of FFT

p.26/33
 What happened after the publication

» Fast Fourier Transform -

Overview
 Rudnick’s note 13 mentions Danielson-Lanczos paper from
Newly discovered history of FFT 1942 3 . “Although ... less elegant and is phrased wholly in
» What happened after the
publication terms of real quantities, it yields the same results as the
» Danielson-Lanczos
» Gauss
binary form of the Cooley-Tukey algorithm with a comparable
number of arithmetical operations”.
 “Historical Notes on FFT” 2
 Also mentions Danielson-Lanczos
 Thomas and Prime Factor algorithm (Good) - quite
distinct from Cooley-Tukey FFT
 A later investigation 10 revealed that Gauss essentially
discovered the Cooley-Tukey FFT in 1805(!)

p.27/33
Danielson-Lanczos

 G. C. Danielson and C. Lanczos. Some improvements in practical

Fourier analysis and their application to X-ray scattering from liquids. J.
Franklin Institute, 233:365–380 and 435–452, 1942

p.28/33
Danielson-Lanczos

 G. C. Danielson and C. Lanczos. Some improvements in practical

Fourier analysis and their application to X-ray scattering from liquids. J.
Franklin Institute, 233:365–380 and 435–452, 1942
 “... the available standard forms become impractical for a large number
of coefficients. We shall show that, by a certain transformation process,
it is possible to double the number of ordinates with only slightly more
than double the labor”

p.28/33
Danielson-Lanczos

 G. C. Danielson and C. Lanczos. Some improvements in practical

Fourier analysis and their application to X-ray scattering from liquids. J.
Franklin Institute, 233:365–380 and 435–452, 1942
 “... the available standard forms become impractical for a large number
of coefficients. We shall show that, by a certain transformation process,
it is possible to double the number of ordinates with only slightly more
than double the labor”
 Concerned with improving efficiency of hand calculation

p.28/33
Danielson-Lanczos

 G. C. Danielson and C. Lanczos. Some improvements in practical

Fourier analysis and their application to X-ray scattering from liquids. J.
Franklin Institute, 233:365–380 and 435–452, 1942
 “... the available standard forms become impractical for a large number
of coefficients. We shall show that, by a certain transformation process,
it is possible to double the number of ordinates with only slightly more
than double the labor”
 Concerned with improving efficiency of hand calculation
 “We shall now describe a method which eliminates the necessity of
complicated schemes by reducing ... analysis for 4n coefficients to two
analyses for 2n coefficients”

p.28/33
Danielson-Lanczos

 G. C. Danielson and C. Lanczos. Some improvements in practical

Fourier analysis and their application to X-ray scattering from liquids. J.
Franklin Institute, 233:365–380 and 435–452, 1942
 “... the available standard forms become impractical for a large number
of coefficients. We shall show that, by a certain transformation process,
it is possible to double the number of ordinates with only slightly more
than double the labor”
 Concerned with improving efficiency of hand calculation
 “We shall now describe a method which eliminates the necessity of
complicated schemes by reducing ... analysis for 4n coefficients to two
analyses for 2n coefficients”
 “If desired, this reduction process can be applied twice or three times”
(???!!)
 “Adoping these improvements the approximate times... are: 10 minutes
for 8 coefficients, 25 min. for 16, 60 min. for 32 and 140 min. for 64”
(≈ 0.37N log2 N)
p.28/33
Danielson-Lanczos

 Take the DIT Cooley-Tukey with N1 = N/2, N2 = 2. Recall

N/2−1 N/2−1
N j0 k 1 j0 + N j j k
A( j0 + j1 ) =
2 ∑ X (2k 1 )W N/2 + WN 2 1 ∑ X (2k 1 + 1)W N/2
0 1

k 1 =0 k 1 =0

p.28/33
Danielson-Lanczos

 Take the DIT Cooley-Tukey with N1 = N/2, N2 = 2. Recall

N/2−1 N/2−1
N j0 k 1 j0 + N j j k
A( j0 + j1 ) =
2 ∑ X (2k 1 )W N/2 + WN 2 1 ∑ X (2k 1 + 1)W N/2
0 1

k 1 =0 k 1 =0

 Danielson-Lanczos showed completely analogous result, under the

framework of real trigonometric series:
 “...the contribution of the even ordinates to a Fourier analysis of 4n
coefficients is exactly the same as the contribution of all the
ordinates to a Fourier analysis of 2n coefficients
 “..contribution of the odd ordinates is also reducible to the same
scheme... by means of a transformation process introducing a
phase difference”
 “...we see that, apart from the weight factors 2 cos( π/4n ) k and
2 sin(π/4n)k, the calculation is identical to the cosine analysis of
half the number of ordinates”

p.28/33
Danielson-Lanczos

 Take the DIT Cooley-Tukey with N1 = N/2, N2 = 2. Recall

N/2−1 N/2−1
N j0 k 1 j0 + N j j k
A( j0 + j1 ) =
2 ∑ X (2k 1 )W N/2 + WN 2 1 ∑ X (2k 1 + 1)W N/2
0 1

k 1 =0 k 1 =0

 Danielson-Lanczos showed completely analogous result, under the

framework of real trigonometric series:
 “...the contribution of the even ordinates to a Fourier analysis of 4n
coefficients is exactly the same as the contribution of all the
ordinates to a Fourier analysis of 2n coefficients
 “..contribution of the odd ordinates is also reducible to the same
scheme... by means of a transformation process introducing a
phase difference”
 “...we see that, apart from the weight factors 2 cos( π/4n ) k and
2 sin(π/4n)k, the calculation is identical to the cosine analysis of
half the number of ordinates”
j
 These “weight factors” are exactly the “twiddle factors” W above
N
p.28/33
 Gauss

» Fast Fourier Transform -

Overview
 Herman H. Goldstine writes in a footnote of “A History of
Newly discovered history of FFT
Numerical Analysis from the 16th through the 19th Century”
» What happened after the
publication
(1977) 8 :
» Danielson-Lanczos
» Gauss
“This fascinating work of Gauss was neglected and
rediscovered by Cooley and Tukey in an important
paper in 1965”
 This goes largely unnoticed until the research by Heideman,
Johnson and Burrus 10 in 1985
 Essentially develops an DIF FFT for two factors and real
sequences. Gives examples for N = 12, 36.
 Declares that it can be generalized to more than 2
factors, although no examples are given
 Uses the algorithm for solving the problem of determining
orbit parameters of asteroids
 The treatise was published posthumously (1866). By
then other numerical methods were preferred to DFT, so
nobody found this interesting enough...
p.29/33
» Fast Fourier Transform -
Overview

Impact
» Impact
» Further developments
» Concluding thoughts

Impact

p.30/33
 Impact

» Fast Fourier Transform -

Overview
 Certainly a “classic”
Impact
 Two special issues of IEEE trans. on Audio and
Electroacoustics devoted entirely to FFT 11
» Impact
» Further developments
» Concluding thoughts
 Arden House Workshop on FFT 11
 Brought together people of very diverse specialities
 “someday radio tuners will operate with digital processing
units. I have heard this suggested with tongue in cheek,
but one can speculate.” - J.W.Cooley
 Many people “discovered” the DFT via the FFT

p.31/33
 Further developments

» Fast Fourier Transform -

Overview
 Real-FFT (DCT...)
Impact
 Parallel FFT
» Impact
» Further developments  FFT for prime N
» Concluding thoughts
 ...

p.32/33
 Concluding thoughts

» Fast Fourier Transform -

Overview
 It is obvious that prompt recognition and publication of
Impact
significant achievments is an important goal
» Impact
» Further developments
» Concluding thoughts

p.33/33
 Concluding thoughts

» Fast Fourier Transform -

Overview
 It is obvious that prompt recognition and publication of
Impact
significant achievments is an important goal
» Impact
» Further developments
 However, the publication in itself may not be enough
» Concluding thoughts

p.33/33
 Concluding thoughts

» Fast Fourier Transform -

Overview
 It is obvious that prompt recognition and publication of
Impact
significant achievments is an important goal
» Impact
» Further developments
 However, the publication in itself may not be enough
» Concluding thoughts
 Communication between mathematicians, numerical
analysts and workers in a very wide range of applications
can be very fruitful

p.33/33
 Concluding thoughts

» Fast Fourier Transform -

Overview
 It is obvious that prompt recognition and publication of
Impact
significant achievments is an important goal
» Impact
» Further developments
 However, the publication in itself may not be enough
» Concluding thoughts
 Communication between mathematicians, numerical
analysts and workers in a very wide range of applications
can be very fruitful
 Always seek new analytical methods and not rely solely on
increase in processing speed

p.33/33
 Concluding thoughts

» Fast Fourier Transform -

Overview
 It is obvious that prompt recognition and publication of
Impact
significant achievments is an important goal
» Impact
» Further developments
 However, the publication in itself may not be enough
» Concluding thoughts
 Communication between mathematicians, numerical
analysts and workers in a very wide range of applications
can be very fruitful
 Always seek new analytical methods and not rely solely on
increase in processing speed
 Careful attention to a review of old literature may offer some
rewards

p.33/33
 Concluding thoughts

» Fast Fourier Transform -

Overview
 It is obvious that prompt recognition and publication of
Impact
significant achievments is an important goal
» Impact
» Further developments
 However, the publication in itself may not be enough
» Concluding thoughts
 Communication between mathematicians, numerical
analysts and workers in a very wide range of applications
can be very fruitful
 Always seek new analytical methods and not rely solely on
increase in processing speed
 Careful attention to a review of old literature may offer some
rewards
 Do not publish papers in Journal of Franklin Institute

p.33/33
 Concluding thoughts

» Fast Fourier Transform -

Overview
 It is obvious that prompt recognition and publication of
Impact
significant achievments is an important goal
» Impact
» Further developments
 However, the publication in itself may not be enough
» Concluding thoughts
 Communication between mathematicians, numerical
analysts and workers in a very wide range of applications
can be very fruitful
 Always seek new analytical methods and not rely solely on
increase in processing speed
 Careful attention to a review of old literature may offer some
rewards
 Do not publish papers in Journal of Franklin Institute
 Do not publish papers in neo-classic Latin

p.33/33
References
1. J. W. Cooley and J. W. Tukey. An algorithm for the machine calculation of complex Fourier series.
Mathematics of Computation, 19:297–301, 1965.

2. J. W. Cooley, P. A. Lewis, and P. D. Welch. History of the fast Fourier transform. In Proc. IEEE,
volume 55, pages 1675–1677, October 1967.

3. G. C. Danielson and C. Lanczos. Some improvements in practical Fourier analysis and their appli-
cation to X-ray scattering from liquids. J. Franklin Institute, 233:365–380 and 435–452, 1942.

4. D.L.Banks. A conversation with I. J. Good. Statistical Science, 11(1):1–19, 1996.

5. W. M. Gentleman. An error analysis of Goertzel’s (Watt’s) method for computing Fourier coeffi-
cients. Comput. J., 12:160–165, 1969.

6. W. M. Gentleman and G. Sande. Fast Fourier transforms—for fun and profit. In Fall Joint Computer
Conference, volume 29 of AFIPS Conference Proceedings, pages 563–578. Spartan Books, Washington,
D.C., 1966.

7. G. Goertzel. An algorithm for the evaluation of finite trigonometric series. The American Mathe-
matical Monthly, 65(1):34–35, January 1958.

8. Herman H. Goldstine. A History of Numerical Analysis from the 16th through the 19th Century.
Springer-Verlag, New York, 1977. ISBN 0-387-90277-5.

9. I. J. Good. The interaction algorithm and practical Fourier analysis. Journal Roy. Stat. Soc., 20:
361–372, 1958.

10. M. T. Heideman, D. H. Johnson, and C. S. Burrus. Gauss and the history of the Fast Fourier
Transform. Archive for History of Exact Sciences, 34:265–267, 1985.

11. IEEE. Special issue on fast Fourier transform. IEEE Trans. on Audio and Electroacoustics, AU-17:
65–186, 1969.

12. Morgenstern. Note on a lower bound of the linear complexity of the fast Fourier transform. JACM:
Journal of the ACM, 20, 1973.

13. Philip Rudnick. Note on the calculation of Fourier series (in Technical Notes and Short Papers).
j-MATH-COMPUT, 20(95):429–430, July 1966. ISSN 0025-5718.

14. C. E. Shannon. Communication in the presence of noise. Proceedings of the IRE, 37:10–21, January
1949.

15. F. Yates. The design and analysis of factorial experiments, volume 35 of Impr. Bur. Soil Sci. Tech. Comm.
1937.

33-1