Fast Fourier Transform

最新推荐文章于 2024-07-05 14:02:35 发布

一苇何如

最新推荐文章于 2024-07-05 14:02:35 发布

阅读量212

点赞数

分类专栏：碎岁睢穂文章标签： FFT algebro matrix 数学快速傅里叶变换

本文链接：https://blog.csdn.net/yiweiheru/article/details/77863309

版权

碎岁睢穂专栏收录该内容

2 篇文章 0 订阅

订阅专栏

快速傅里叶变换的数学本质

Defination of Fourier matrix

F n = ⎡ ⎣ ⎢ ⎢ ⎢ ⎢ ⎢ 11 ⋮ 1 1 ω 1 \cdot 1 ⋮ ω (n - 1) \cdot 1 \dots \dots ⋱ \dots 1 ω 1 \cdot (n - 1) ⋮ ω (n - 1) \cdot (n - 1) ⎤ ⎦ ⎥ ⎥ ⎥ ⎥ ⎥ w i t h ω n = 1

$F_n=\left[ \begin{matrix} 1 & 1 & \cdots &1\\ 1 & \omega^{1\cdot1} & \cdots &\omega^{1\cdot(n-1)}\\ \vdots & \vdots & \ddots &\vdots \\ 1 & \omega^{(n-1)\cdot1} & \cdots & \omega^{(n-1)\cdot(n-1)} \end{matrix} \right]with\quad\omega^n=1$

n次单位根周期规律:

$\omega_{32}=e^{i\cdot2\pi/32}=\omega_{64}^2=e^{(i\cdot2\pi/64)\cdot2}$
$\omega_{64}^{32}=-\omega_{64}^{0},\omega_{64}^{33}=-\omega_{64}^{1},\cdots,\omega_{64}^{63}=-\omega_{64}^{32}$

Fourier矩阵 $F_n$ 递归原理:

F 64 = [I I D - D] [F 32 0 0 F 32] ⎡ ⎣ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ 10 ⋮ 00 ⋮ 0 00 ⋮ 01 ⋮ 0 01 ⋮ 00 ⋮ 0 \dots \dots ⋱ \dots \dots ⋱ \dots 00 ⋮ 00 ⋮ 1 00 ⋮ 10 ⋮ 0 ⎤ ⎦ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ w h e r e D = ⎡ ⎣ ⎢ ⎢ ⎢ ⎢ ⎢ 1 ω 64 ⋱ ω n - 1 64 ⎤ ⎦ ⎥ ⎥ ⎥ ⎥ ⎥

$F_{64}=\left[\begin{matrix}I & D\\I & -D\end{matrix}\right] \begin{bmatrix}F_{32} &0\\0 & F_{32}\end{bmatrix} \begin{bmatrix}1&0&0&\cdots&0&0\\0&0&1&\cdots&0&0\\\vdots&\vdots&\vdots&\ddots&\vdots&\vdots\\0&0&0&\cdots&0&1\\0&1&0&\cdots&0&0\\\vdots&\vdots&\vdots&\ddots&\vdots&\vdots\\0&0&0&\cdots&1&0\end{bmatrix}\\ where\quad D=\begin{bmatrix} 1&\ &\ &\ &\ \\ \ &\omega_{64}&\ &\ &\ \\ \ &\ &\ddots&\ &\ \\ \ &\ &\ &\ &\omega_{64}^{n-1} \end{bmatrix}$

The proof of the decomposition of $F_{n}$

$\left[\begin{matrix}I & D\\I & -D\end{matrix}\right] \begin{bmatrix}F_{32} &0\\0 & F_{32}\end{bmatrix} \begin{bmatrix}1&0&0&\cdots&0&0\\0&0&1&\cdots&0&0\\\vdots&\vdots&\vdots&\ddots&\vdots&\vdots\\0&0&0&\cdots&0&1\\0&1&0&\cdots&0&0\\\vdots&\vdots&\vdots&\ddots&\vdots&\vdots\\0&0&0&\cdots&1&0\end{bmatrix}\cdot\begin{bmatrix}c_0\\c_1\\\vdots\\c_{63}\end{bmatrix}\\ =\left[\begin{matrix}I & D\\I & -D\end{matrix}\right] \begin{bmatrix}F_{32} &0\\0 & F_{32}\end{bmatrix}\cdot \begin{bmatrix}{c_0\\c_2\\\vdots\\c_{62}\\c_1\\c_3\\\vdots\\c_{63}} \end{bmatrix}\\ =\begin{bmatrix}I & D\\I & -D\end{bmatrix}\cdot \begin{bmatrix}{\sum_{j=0}^{31}c_{j\cdot2}\\ \sum_{j=0}^{31}c_{j\cdot2}\omega_{32}^{j}\\\vdots\\ \sum_{j=0}^{31}c_{j\cdot2}\omega_{32}^{j\cdot31}\\ \sum_{j=0}^{31}c_{j\cdot2+1}\\ \sum_{j=0}^{31}c_{j\cdot2+1}\omega_{32}^{j}\\\vdots\\ \sum_{j=0}^{31}c_{j\cdot2+1}\omega_{32}^{j\cdot31}}\end{bmatrix}\\ =\begin{bmatrix}{ \sum_{j=0}^{31}c_{2j}+\sum_{j=0}^{31}c_{2j+1}\\ \sum_{j=0}^{31}c_{2j}\omega_{32}^{j\cdot1}+\sum_{j=0}^{31}c_{2j+1}\omega_{32}^{j\cdot1}\cdot\omega_{64}\\\vdots\\ \sum_{j=0}^{31}c_{2j}\omega_{32}^{j\cdot31}+\sum_{j=0}^{31}c_{2j+1}\omega_{32}^{j\cdot31}\cdot\omega_{64}^{31}\\ \sum_{j=0}^{31}c_{2j}-\sum_{j=0}^{31}c_{2j+1}\\ \sum_{j=0}^{31}c_{2j}\omega_{32}^{j\cdot1}-\sum_{j=0}^{31}c_{2j+1}\omega_{32}^{j\cdot1}\cdot\omega_{64}\\\vdots\\ \sum_{j=0}^{31}c_{2j}\omega_{32}^{j\cdot31}-\sum_{j=0}^{31}c_{2j+1}\omega_{32}^{j\cdot31}\cdot\omega_{64}^{31} }\end{bmatrix}\\ =\begin{bmatrix}{ \sum_{j=0}^{31}c_{2j}+\sum_{j=0}^{31}c_{2j+1}\\ \sum_{j=0}^{31}c_{2j}\omega_{64}^{2j\cdot1}+\sum_{j=0}^{31}c_{2j+1}\omega_{64}^{(2j+1)\cdot1}\\\vdots\\ \sum_{j=0}^{31}c_{2j}\omega_{64}^{2j\cdot31}+\sum_{j=0}^{31}c_{2j+1}\omega_{64}^{(2j+1)\cdot31}\\ \sum_{j=0}^{31}c_{2j}\omega_{64}^{32\cdot2j}-\sum_{j=0}^{31}c_{2j+1}(-\omega_{64}^{32\cdot(2j+1)})\\ \sum_{j=0}^{31}c_{2j}\omega_{64}^{33\cdot2j}-\sum_{j=0}^{31}c_{2j+1}(-\omega_{64}^{33\cdot(2j+1)})\\\vdots\\ \sum_{j=0}^{31}c_{2j}\omega_{64}^{63\cdot2j}-\sum_{j=0}^{31}c_{2j+1}(-\omega_{64}^{63\cdot(2j+1)}) }\end{bmatrix}\\= \begin{bmatrix}{\sum_{k=0}^{63}c_k\cdot1\\ \sum_{k=0}^{63}c_k\cdot\omega_{64}^k\\\vdots\\\sum_{k=0}^{63}c_k\cdot\omega_{64}^{k\cdot63}}\end{bmatrix}= F_{64}\cdot\begin{bmatrix}c_0\\c_1\\\vdots\\c_{63}\end{bmatrix}$

大幅度减少计算量

$F_{64} \cdot C$ costs $64^2$ multiplication
Decompose it once, taking $[2*(32^2)+32]$
Decompose it twice, taking $[2*(2*16^2+16)+32]=[4*16^2+32*2]$
…
Decompose it $\log_264=6$ , taking $[64*1^2+32*6]$ multiplication