原理
此处以基2频域抽取FFT为例,讲述频域抽取FFT的原理。假设FFT的长度为
N
=
2
m
N=2^m
N=2m,我们将序列
x
x
x的FFT变换分为以下两个部分:
X
(
k
)
=
∑
n
=
0
N
/
2
−
1
x
(
n
)
W
N
n
k
+
∑
n
=
N
/
2
N
−
1
x
(
n
)
W
N
n
k
X(k)=\sum_{n=0}^{N/2-1}x(n)W_N^{nk}+\sum_{n=N/2}^{N-1}x(n)W_N^{nk}
X(k)=n=0∑N/2−1x(n)WNnk+n=N/2∑N−1x(n)WNnk
对等号右边的第二项作代换:
n
=
n
+
N
/
2
n=n+N/2
n=n+N/2,则有:
X
(
k
)
=
∑
n
=
0
N
/
2
−
1
x
(
n
)
W
N
n
k
+
∑
n
=
0
N
/
2
−
1
x
(
n
+
N
/
2
)
W
N
k
(
n
+
N
/
2
)
X(k)=\sum_{n=0}^{N/2-1}x(n)W_N^{nk}+\sum_{n=0}^{N/2-1}x(n+N/2)W_N^{k(n+N/2)}
X(k)=n=0∑N/2−1x(n)WNnk+n=0∑N/2−1x(n+N/2)WNk(n+N/2)
由于
W
N
k
(
n
+
N
/
2
)
=
W
N
k
n
⋅
W
N
k
N
/
2
=
(
−
1
)
k
W
N
n
k
W_N^{k(n+N/2)}=W_N^{kn}\cdot W_N^{kN/2}=(-1)^kW_N^{nk}
WNk(n+N/2)=WNkn⋅WNkN/2=(−1)kWNnk,故有:
X
(
k
)
=
∑
n
=
0
N
/
2
−
1
x
(
n
)
W
N
n
k
+
(
−
1
)
k
∑
n
=
0
N
/
2
−
1
x
(
n
+
N
/
2
)
W
N
k
n
X(k)=\sum_{n=0}^{N/2-1}x(n)W_N^{nk}+(-1)^k\sum_{n=0}^{N/2-1}x(n+N/2)W_N^{kn}
X(k)=n=0∑N/2−1x(n)WNnk+(−1)kn=0∑N/2−1x(n+N/2)WNkn
令
k
=
2
m
k=2m
k=2m以及
k
=
2
m
+
1
k=2m+1
k=2m+1,分别有:
X
(
2
m
)
=
∑
n
=
0
N
/
2
−
1
(
x
(
n
)
+
x
(
n
+
N
/
2
)
)
W
N
2
n
m
X(2m)=\sum_{n=0}^{N/2-1}(x(n)+x(n+N/2))W_N^{2nm}
X(2m)=n=0∑N/2−1(x(n)+x(n+N/2))WN2nm
X
(
2
m
+
1
)
=
∑
n
=
0
N
/
2
−
1
(
x
(
n
)
−
x
(
n
+
N
/
2
)
)
W
N
n
(
2
m
+
1
)
X(2m+1)=\sum_{n=0}^{N/2-1}(x(n)-x(n+N/2))W_N^{n(2m+1)}
X(2m+1)=n=0∑N/2−1(x(n)−x(n+N/2))WNn(2m+1)
根据旋转因子
W
N
W_N
WN的可约性,有:
X
(
2
m
)
=
∑
n
=
0
N
/
2
−
1
(
x
(
n
)
+
x
(
n
+
N
/
2
)
)
W
N
/
2
n
m
X(2m)=\sum_{n=0}^{N/2-1}(x(n)+x(n+N/2))W_{N/2}^{nm}
X(2m)=n=0∑N/2−1(x(n)+x(n+N/2))WN/2nm
X
(
2
m
+
1
)
=
∑
m
=
0
N
/
2
−
1
(
x
(
n
)
−
x
(
n
+
N
/
2
)
)
W
N
n
⋅
W
N
/
2
n
m
X(2m+1)=\sum_{m=0}^{N/2-1}(x(n)-x(n+N/2))W_N^n\cdot W_{N/2}^{nm}
X(2m+1)=m=0∑N/2−1(x(n)−x(n+N/2))WNn⋅WN/2nm
令
x
1
(
n
)
=
x
(
n
)
+
x
(
n
+
N
/
2
)
,
x
2
(
n
)
=
(
x
(
n
)
−
x
(
n
+
N
/
2
)
)
W
N
n
x_1(n)=x(n)+x(n+N/2),x_2(n)=(x(n)-x(n+N/2))W_N^n
x1(n)=x(n)+x(n+N/2),x2(n)=(x(n)−x(n+N/2))WNn,则上式可以写作:
X
(
2
m
)
=
∑
n
=
0
N
/
2
−
1
x
1
(
n
)
W
N
/
2
n
m
X(2m)=\sum_{n=0}^{N/2-1}x_1(n)W_{N/2}^{nm}
X(2m)=n=0∑N/2−1x1(n)WN/2nm
X
(
2
m
+
1
)
=
∑
n
=
0
N
/
2
−
1
x
2
(
n
)
W
N
/
2
n
m
X(2m+1)=\sum_{n=0}^{N/2-1}x_2(n)W_{N/2}^{nm}
X(2m+1)=n=0∑N/2−1x2(n)WN/2nm
由此,我们将N点FFT转换为了两个N/2点的FFT,这就是FFT中分而治之的思想。
下图是8点频域抽取FFT的蝶形运算示意图:
实现
#include<iostream>
#include<complex.h>
#include<math.h>
#define PI 3.14159
using namespace std;
typedef complex<double> cdata_t;
void FFT(complex<double>* Xin,complex<double> *Xout,int N){
if(N<2)
Xout[0]=Xin[0];
else
{
complex<double>* X1=new complex<double>[N/2];
complex<double>* X2=new complex<double>[N/2];
complex<double>* X1_out=new complex<double>[N/2];
complex<double>* X2_out=new complex<double>[N/2];
for(int i=0;i<N;i+=2)
{
X1[i/2]=Xin[i];
X2[i/2]=Xin[i+1];
}
FFT(X1,X1_out,N/2);
FFT(X2,X2_out,N/2);
complex<double>* W=new complex<double>[N/2];
for(int i=0;i<N/2;i++){
W[i].real(cos(2*PI*i/N));
W[i].imag(-sin(2*PI*i/N));
}
for(int i=0;i<N/2;i++){
Xout[i]=X1_out[i]+W[i]*X2_out[i];
Xout[i+N/2]=X1_out[i]-W[i]*X2_out[i];
}
delete []X1;
delete []X2;
delete []X1_out;
delete []X2_out;
}
return;
}
void DIF_FFT8(complex<double>* Xin,complex<double>* Xout){
complex<double> W[4];
complex<double> TMP1[8];
complex<double> TMP2[8];
complex<double> TMP3[8];
const int N=8;
for(int i=0;i<4;i++){
W[i]=complex<double>(cos(2*PI*i/N),-sin(2*PI*i/N));
}
//stage1
for(int i=0;i<4;i++){
TMP1[i]=Xin[i]+Xin[i+4];
TMP1[i+4]=(Xin[i]-Xin[i+4])*W[i];
}
//stage2
for(int i=0;i<2;i++)
for(int j=0;j<2;j++){
TMP2[i*4+j]=TMP1[i*4+j]+TMP1[i*4+j+2];
TMP2[i*4+j+2]=(TMP1[i*4+j]-TMP1[i*4+j+2])*W[2*j];
}
//stage3
for(int i=0;i<4;i++){
TMP3[i*2]=TMP2[i*2]+TMP2[i*2+1];
TMP3[i*2+1]=(TMP2[i*2]-TMP2[i*2+1])*W[0];
}
//
Xout[0]=TMP3[0];
Xout[1]=TMP3[4];
Xout[2]=TMP3[2];
Xout[3]=TMP3[6];
Xout[4]=TMP3[1];
Xout[5]=TMP3[5];
Xout[6]=TMP3[3];
Xout[7]=TMP3[7];
}
void DIF_FFT16(cdata_t* Xin,cdata_t* Xout){
cdata_t W[8];
cdata_t T1[16];
cdata_t T2[16];
cdata_t T3[16];
cdata_t T4[16];
//
for(int i=0;i<8;i++){
W[i]=cdata_t(cos(2*PI*i/16),-sin(2*PI*i/16));
}
//stage1
for(int i=0;i<8;i++)
{
T1[i]=Xin[i]+Xin[i+8];
T1[i+8]=(Xin[i]-Xin[i+8])*W[i];
}
//stage2
for(int i=0;i<2;i++)
for(int j=0;j<4;j++){
T2[i*8+j]=T1[i*8+j]+T1[i*8+j+4];
T2[i*8+j+4]=(T1[i*8+j]-T1[i*8+j+4])*W[2*j];
}
//stage3
for(int i=0;i<4;i++)
for(int j=0;j<2;j++){
T3[i*4+j]=T2[i*4+j]+T2[i*4+j+2];
T3[i*4+j+2]=(T2[i*4+j]-T2[i*4+j+2])*W[4*j];
}
//stage4
for(int i=0;i<8;i++){
T4[2*i]=T3[2*i]+T3[2*i+1];
T4[2*i+1]=T3[2*i]-T3[2*i+1];
}
//
Xout[0]=T4[0];
Xout[1]=T4[8];
Xout[2]=T4[4];
Xout[3]=T4[12];
Xout[4]=T4[2];
Xout[5]=T4[10];
Xout[6]=T4[6];
Xout[7]=T4[14];
Xout[8]=T4[1];
Xout[9]=T4[9];
Xout[10]=T4[5];
Xout[11]=T4[13];
Xout[12]=T4[3];
Xout[13]=T4[11];
Xout[14]=T4[7];
Xout[15]=T4[15];
}
void DIF_FFT4(cdata_t* Xin,cdata_t* Xout){
cdata_t W[2];
cdata_t T1[4];
cdata_t T2[4];
for(int i=0;i<2;i++){
W[i]=cdata_t(cos(2*PI*i/4),-sin(2*PI*i/4));
}
//stage1
for(int i=0;i<2;i++){
T1[i]=Xin[i]+Xin[i+2];
T1[i+2]=(Xin[i]-Xin[i+2])*W[i];
}
//stage2
for(int i=0;i<2;i++){
T2[i*2]=T1[i*2]+T1[i*2+1];
T2[i*2+1]=(T1[i*2]-T1[i*2+1]);
}
//
Xout[0]=T2[0];
Xout[1]=T2[2];
Xout[2]=T2[1];
Xout[3]=T2[3];
}
int main(){
//FFT_LENGTH点傅里叶变换
int n=4;
complex<double> X[n];
complex<double> Y[n];
complex<double> Y2[n];
for(int i=0;i<n;i++){
X[i].real(i);
X[i].imag(0);
}
FFT(X,Y,n);
if(n==4)
DIF_FFT4(X,Y2);
else if(n==8)
DIF_FFT8(X,Y2);
else if(n==16)
DIF_FFT16(X,Y2);
else
cout<<"未实现该长度的蝶形FFT函数"<<endl;
for(int i=0;i<n;i++)
cout<<Y[i]-Y2[i]<<endl;
return 0;
}