Overview
The Fast Fourier Transform (FFT) is a fast method for evaluating the discrete Fourier transform (DFT) defined to be the complex exponential sum, $$ \hat{x}_n = {1 \over \sqrt{N}} \sum_{k=0}^{N-1} x_k W^{k n} \quad n = 0 \ldots N-1 $$ on the sequence $\{ x_n \: \vert \: n = 0 \ldots N-1 \}$ where $W = e^{-2 \pi i / N}$ is an nth root of unity, and i denotes $\sqrt {-1}$
The FFT performs $O(n \: lg \: N)$ operations when N is a power of 2.
Here's C++ software which implements a power of two FFT. We create it as a derived class of the STL vector type.
Features
- Clean C++ language implementation of a standard power of 2 FFT algorithm. You can define single or double precision floating point.
Download
Source code Version 1.5 and executables are distributed under the terms of the GNU General Public License.
FFT.cpp | FFT implementation. | View Download |
FFT.h | Header file containing the class definitions. | View Download |
testFFT.cpp | Main unit test or demo program. | View Download |
makefile | Makefile | View Download |
fftIn.txt | Input test file. | View Download |
fftOut.txt | Output test file. | View Download |
FFT.FOR | FFT in FORTRAN | View Download |
FFTD.FOR | FFT driver in FORTRAN | View Download |
fftOut-macOS-ARM-M1max-clang.txt, fftOut-UbuntuLinux-AMD-x86_64-gcc.txt |
Same exact results on two quite different computers: [MacOS, M1 max ARM CPU, Clang] and [Ubuntu/Linux, AMD x86_64, gcc]. All hail the IEEE floating point standard. | View Download |
Install and Run
On Mac OS X, I use the Xcode IDE; on a Windows platforms, I use the GNU Cygwin toolset for command line compiling and debugging; and on Unix systems, including Mac OS X, I use the built-in gcc compiler and gdb debugger. For online C++ language tutorials, books and references, see links to C++ documentation.
DFT on a Discretized Sine Wave, Sampling Theorem, and Windowing
Theorem. Let's show the DFT of a digitized sine wave with frequency $f$ and sampled at frequency $f_s$ which is periodic on the interval $[0, N)$, i.e. $T = N T_s$, is zero except when $n = 1$ when its value is $\frac{ N }{ 2i }$ and $n = N-1$ when its value is $-\frac{ N }{ 2i }$
Proof. $$ \mathcal{ DFT } \left( sin \left( 2 \pi \frac{ f }{ f_s } k \right) \right)(n) = $$ $$ \sum_{k=0}^{N-1} e^{ - \frac{ 2 \pi i }{ N } nk} \left[ \frac{ e^{ 2 \pi i \frac{ f }{ f_s } k} - e^{ -2 \pi i \frac{ f }{ f_s } k} }{ 2 i } \right] = $$ $$ \frac{ 1 }{ 2 i } \sum_{k=0}^{N-1} e^{ 2 \pi i k \left( \frac{ f }{ f_s} - \frac{ n }{ N } \right) } - \frac{ 1 }{ 2 i } \sum_{k=0}^{N-1} e^{ -2 \pi i k \left( \frac{ f }{ f_s} + \frac{ n }{ N } \right) } $$ By the geometric series sum formula, $$ \sum_{k=0}^{N-1} e ^{ 2 \pi i k \theta} = $$ $$ 1 + e^{2 \pi i \theta} + {(e^{2 \pi i \theta})}^2 + \ldots + {(e^{2 \pi i \theta})}^{N-1} $$ $$ = \Bigg\{ \begin{array}{rl} \frac{ 1 - {e^{2 \pi i N \theta}} }{ 1 - e^{2 \pi i \theta}} \quad \text{ for } e^{2 \pi i \theta} \neq 1 \\ N \quad \text{ for } e^{2 \pi i \theta} = 1 \end{array} $$
and let's suppose that the sine wave period and sampling period are related as $T = N T_s$ or $\frac{ f }{ f_s } = \frac{ 1 }{ N }$ then $$ \mathcal{ DFT } \left( sin \left( 2 \pi \frac{ f }{ f_s } k \right) \right)(n) = $$ $$ \frac{ 1 }{ 2 i } \Bigg\{ \begin{array}{rl} \frac{ 1 - {e^{2 \pi i (1-n)}} }{ 1 - e^{2 \pi i (1-n)/N}} \quad \text{ for } e^{2 \pi i (1-n)/N} \neq 1 \\ N \quad \text{ for } e^{2 \pi i (1-n)/N} = 1 \end{array} \Bigg\} $$ $$ - \frac{ 1 }{ 2 i } \Bigg\{ \begin{array}{rl} \frac{ 1 - {e^{2 \pi i (1+n)}} }{ 1 - e^{2 \pi i (1+n)/N}} \quad \text{ for } e^{2 \pi i (1+n)/N} \neq 1 \\ N \quad \text{ for } e^{2 \pi i (1+n)/N} = 1 \end{array} \Bigg\} $$ but $$e^{2 \pi i k} = 1 \Leftrightarrow k \in \mathbb{Z}$$ so the numerators are zero except when setting $n = 1$ on the left and and $n = N-1$ on the right in which case the result is $\frac{ N }{ 2i }$ or $-\frac{ N }{ 2i }$ $\blacksquare$
The discrete Fourier transform (DFT) is a discrete time approximation to the Fourier integral over a finite domain. It assumes the data in the domain is periodic. If it is not, any discontinuity at the beginning and end will generate artificial high frequencies which mix in with the power spectrum, called spectral leakage. To avoid that, people apply windowing functions to the data before applying the FFT.
Here's an example of a DFT on a single frequency and how to use windowing to reduce spectral leakage,
References
- The standard text on signal processing is Discrete Time Signal Processing by Oppenheim and Schafer
- My power of 2 implementation is derived from E. Oran Brigham's book
- Fast implementations: Fastest Fourier Transform in the West, KFR library (I haven't tried these).
- A note on spectral leakage and windowing Understanding FFTs and Windowing