Fast Fourier Transform (FFT)

Elena Punskaya
www-sigproc.eng.cam.ac.uk/~op205
Some material adapted from courses by Prof. Simon Godsill, Dr. Arnaud Doucet, Dr. Malcolm Macleod and Prof. Peter Rayner

55

Computations for Evaluating the DFT

Discrete Fourier Transform and its Inverse can be computed on a digital computer

What are computational requirements?

56

Operation Count
How do we evaluate computational complexity? count the number of real multiplications and additions by real we mean either fixed-point or floating point operations depending on the specific hardware subtraction is considered equivalent to additions divisions are counted separately Other operations such as loading from memory, storing in memory, loop counting, indexing, etc are not counted (depends on implementation/architecture) present overhead 57

Operation Count Modern DSP

Modern DSP: a real multiplication and addition is a single machine cycle y=ax+b called MAC (multiply/accumulate) maximum number of multiplications and additions must be counted, not their sum traditional claim multiplication is much more timeconsuming than addition becomes obsolete
58

Discrete Fourier Transform as a Vector Operation

Again, we have a sequence of sampled values x n and transformed values Xp , given by

Let us denote:

Then:
x

59

Matrix Interpretation of the DFT

This Equation can be rewritten in matrix form:
x

as follows:
X0 X1
0 0 0 1

complex numbers

XN-1

2
0

0
N-1

x0 x1

x N-1

can be computed off-line and stored

0
N-1

N x N matrix

(N-1) 2

2(N-1)

60

DFT Evaluation Operation Count

Multiplication of a complex N x N matrix by a complex N-dimensional vector. No attempt to save operations N2 complex multiplications (xs) N(N-1) complex additions (+s) First row and first column are 1 however 2 save 2N-1 xs, (N-1) complex xs & N(N-1) +s Complex x: 4 real xs and 2 real +s; complex + : 2 real +s Total: 4(N-1) real xs & 4(N-0.5)(N-1) real +s
2 61

DFT Computational Complexity

Thus, for the input sequence of length N the number of arithmetic operations in direct computation of DFT is proportional to N 2 . For N=1000, about a million operations are needed! In 1960s such a number was considered prohibitive in most applications.

62

Discovery of the Fast Fourier Transform (FFT)

When in 1965 Cooley and Tukey first announced discovery of Fast Fourier Transform (FFT) in 1965 it revolutionised Digital Signal Processing. They were actually 150 years late the principle of the FFT was later discovered in obscure section of one of Gauss (as in Gaussian) own notebooks in 1806. Their paper is the most cited mathematical paper ever written.
63

The Fast Fourier Transform (FFT)

The FFT is a highly elegant and efficient algorithm, which is still one of the most used algorithms in speech processing, communications, frequency estimation, etc one of the most highly developed area of DSP. There are many different types and variations. They are just computational schemes for computing the DFT not new transforms! Here we consider the most basic radix-2 algorithm which requires N to be a power of 2.

64

FFT Derivation 1
Lets take the basic DFT equation: Now, split the summation into two parts: one for even n and one for odd n:
Take outside

65

FFT Derivation 2
Thus, we have:

66

FFT Derivation 3
Now, take the same equation again where we split the summation into two parts: one for even n and one for odd n:

And evaluate at frequencies p+N/2

67

FFT Derivation 4
simplified

68

FFT Derivation 5

69

FFT Derivation - Butterfly

Or, in more compact form:

(Butterfly)

70

FFT Derivation - Redundancy

71

FFT Derivation - Computational Load

for particular p

72

Flow Diagram for a N=8 DFT

Input: Output:

73

Further decomposition
We obtained Now A0, A1, A2 , A3 - N/2-point DFT Ap =p +WN/2p p Ap+N/2 =p -WN/2p p

- using redundancy just discovered

B0, B1, B2 , B3 - N/2-point DFT Bp = /p +WN/2p /p Finally Bp+N/2 = /p -WN/2p /p

- using redundancy just discovered

WN/2p =e-j2p/N/2=e-j22p/N=WN2p

74

Flow Diagram for a N=8 DFT - Decomposition

0 1 0 1 /0 /1 /0 /1
75

Further decomposition
We obtained Now A0, A1, A2 , A3 - N/2-point DFT Ap =p +WN/2p p Ap+N/2 =p -WN/2p p

- using redundancy just discovered

B0, B1, B2 , B3 - N/2-point DFT Bp = /p +WN/2p /p Finally Bp+N/2 = /p -WN/2p /p

- using redundancy just discovered

WN/2p =e-j2p/N/2=e-j22p/N=WN2p

76

Flow Diagram for a N=8 DFT Decomposition 2

0 1 0 1 /0 /1 /0 /1
77

Bit-reversal

78

FFT Derivation Summary

The FFT derivation relies on redundancy in the calculation of the basic DFT A recursive algorithm is derived that repeatedly rearranges the problem into two simpler problems of half the size Hence the basic algorithm operates on signals of length a power of 2, i.e. (for some integer M) At the bottom of the tree we have the classic FFT `butterfly structure
79

Computational Load of full FFT algorithm

The type of FFT we have considered, where N = 2M, is called a radix-2 FFT. It has M = log2 N stages, each using N / 2 butterflies Complex multiplication requires 4 real multiplications and 2 real additions Complex addition/subtraction requires 2 real additions Thus, butterfly requires 10 real operations. Hence the radix-2 N-point FFT requires 10( N / 2 )log2 N real operations compared to about 8N2 real operations 80 for the DFT.

Computational Load of full FFT algorithm

The radix-2 N-point FFT requires 10( N / 2 )log2 N real operations compared to about 8N2 real operations for the DFT. This is a huge speed-up in typical applications, where N is 128 4096:
Direct DFT

FFT
81

Further advantage of the FFT algorithm

Input

Output

82

The Inverse FFT (IFFT)

The IDFT is different from the DFT:

it uses positive power of instead of negative ones There is an additional division of each output value by N Any FFT algorithm can be modified to perform the IDFT by using positive powers instead of negatives Multiplying each component of the output by 1 / N Hence the algorithm is the same but computational load increases due to N extra multiplications

83

Many types of FFT

The form of FFT we have described is called decimation in time; there is a form called decimation in frequency (but it has no advantages). The "radix 2" FFT must have length N a power of 2. Slightly more efficient is the "radix 4" FFT, in which 2-input 2-output butterflies are replaced by 4input 4-output units. The transform length must then be a power of 4 (more restrictive). A completely different type of algorithm, the Winograd Fourier Transform Algorithm (WFTA), can be used for FFT lengths equal to the product of a number of mutually prime factors (e.g. 9*7*5 = 315 or 5*16 = 80). The WFTA uses fewer multipliers, but more adders, than a similar-length FFT. Efficient algorithms exist for FFTing real (not complex) data at about 60% the effort of the same-sized complex-data FFT. The Discrete Cosine and Sine Transforms (DCT and DST) are similar real-signal algorithms used in image coding.
84

Applications of the FFT

There FFT is surely the most widely used signal processing algorithm of all. It is the basic building block for a large percentage of algorithms in current usage Specific examples include: Spectrum analysis used for analysing and detecting signals Coding audio and speech signals are often coded in the frequency domain using FFT variants (MP3, ) Another recent application is in a modulation scheme called OFDM, which is used for digital TV broadcasting (DVB) and digital radio (audio) broadcasting (DAB). Background noise reduction for mobile telephony, speech and audio signals is often implemented in the frequency 85 domain using FFTs

Practical Spectral Analysis

Say, we have a symphony recording over 40 minutes long about 2500 seconds Compact-disk recordings are sampled at 44.1 kHz and are in stereo, but say we combine both channels into a single one by summing them at each time point Are we to compute the DFT of a sequence one hundred million samples long? a) unrealistic speed and storage requirements are too high b) useless we will get a very much noiselike wide-band spectrum covering the range from 20 Hz to 20 kHz at high resolution including all notes of all instruments with their harmonics 86

Short-Time Spectral Analysis

We actually want a sequence of short DFTs, each showing the spectrum of a signal of relatively short interval If violin plays note E during this interval spectrum would exhibit energies at not the frequency of note E (329 Hz) and characteristic harmonics of the violin If the interval is short enough we may be able to track note-to-note changes If the frequency resolution is high enough we will be able to identify musical instruments playing Remember: our ear excellent spectrum analyser?
87

Short-Time Spectral Analysis Computational Load

If we take a time interval of about 93 milliseconds (signal of length 4096) the frequency resolution will be 11 Hz (adequate) The number of distinct intervals in the symphony is 40 x 60/0.093 - about 26,000 We may choose to be conservative and overlap the intervals (say 50%) In total, need to compute 52,000 FFTs of length 4096 the entire computation will be done in considerably less time than the music
(more on
88

itself

Short-time spectral analysis also use windowing this later)

Case Study: Spectral analysis of a Musical Signal

Sample rate is 10.025 kHz (T=1/10,025 s)

Extract a short segment Load this into Matlab as a vector x Take an FFT, N=512 X=fft(x(1:512));
Looks almost periodic over short time interval
89

Case Study: Spectral analysis of a Musical Signal, FFT

Symmetric

Note Conjugate symmetry as data are real:

Symmetric

Anti-Symmetric

90

Case Study: Spectral analysis of a Musical Signal - Frequencies

91

The Effect of data length N

N=32

FFT

Low resolution

N=128

FFT

N=1024

FFT

High resolution
92

The DFT approximation to the DTFT

DTFT at frequency : DFT:

Ideally the DFT should be a `good approximation to the DTFT Intuitively the approximation gets better as the number of data points N increases This is illustrated in the previous slide resolution gets better as N increases (more, narrower, peaks in spectrum). How to evaluate this analytically? View the truncation in the summation as a multiplication by a rectangle window function
93

Analysis

94

Pre-multiplying by rectangular window

95

Spectrum of the windowed signal

96

Fourier Transform of the Rectangular Window

97

Rectangular Window Spectrum

N=32

Central `Lobe Sidelobes

98

Rectangular Window Spectrum Lobe Width

N=32 N=16 N=8 N=4

Lobe width inversely Proportional to N

99

Now, imagine what happens when the sum of two frequency components is DFT-ed:

The DTFT is given by a train of delta functions:

Hence the windowed spectrum is just the convolution of the window spectrum with the delta functions:

100

DFT for the data

Both components separately

Both components Together

101

Brief summary

The rectangular window introduces broadening of any frequency components (`smearing) and sidelobes that may overlap with other frequency components (`leakage). The effect improves as N increases However, the rectangle window has poor properties and better choices of wn can lead to better spectral properties (less leakage, in particular) i.e. instead of just truncating the summation, we can pre-multiply by a suitable window function wn that has better frequency domain properties. More on window design in the filter design section of the course see later

102

Summary so far
The Fourier Transform (FT) is a way of transforming a continuous signal into the frequency domain The Discrete Time Fourier Transform (DTFT) is a Fourier Transform of a sampled signal The Discrete Fourier Transform (DFT) is a discrete numerical equivalent using sums instead of integrals that can be computed on a digital computer The FFT is a highly elegant and efficient algorithm, which is still one of the most used algorithms in speech processing, communications, frequency estimation, etc one of the most highly developed area of DSP
103

Thank you!

104