SVD的应用

信号空间复用

$y=Hx+ϵ$, $H=U\Sigma V^H$.
$x^{\prime }=V^{H}x$, $y^{\prime }=U^{H}y$, $w^{\prime }=U^{H}w$.
$y^{\prime }=\Sigma x^{\prime }+w^{\prime }$

pre-filtering matrix $V$
post-filtering matrix $U^H$
gain coefficient $\Sigma$

pre_filtering = 3×3 complex    
  -0.4206 + 0.0000i   0.2913 + 0.0000i  -0.8592 + 0.0000i
  -0.4477 - 0.1281i  -0.1901 + 0.7843i   0.1547 + 0.3286i
  -0.7614 - 0.1629i  -0.0723 - 0.5086i   0.3482 - 0.0927i

post_filtering = 3×3 complex    
  -0.5074 + 0.1743i  -0.7299 + 0.1063i  -0.2862 + 0.2935i
  -0.2672 + 0.1884i  -0.2001 + 0.1837i   0.2860 - 0.8588i
   0.7607 + 0.1632i  -0.5827 - 0.2064i   0.0732 - 0.0849i

gain = 3×1    
   13.7272
    5.8734
    2.2530

MIMO AWGN

Multi Input Multi Output Additive White Gaussian Noise

equivalent parallel SISO AWGN

Single Input Single Output Additive White Gaussian Noise

注： 由于$w_1,w_2,w_3$是独立的标准高斯噪声, 且$U^H$是酉矩阵, 所以$w_1^{\prime }=U^H{w}_1,w_2^{\prime }=U^H{w}_2,w_3^{\prime }=U^H{w}_3$也是标准高斯噪声.

H = [
    1+1i,   3+1i, 7;
    5-0.1i, 4+1i, 8-1i;
    2+3i,   6-1i, 1+3i;
];
[U,S,V] = svd(H);
pre_filtering = V
post_filtering = U'
gain = diag(S)

图像压缩

灰度图像矩阵$U\Sigma V^T=M\in [0,1{]}^{n\times m}$, σ 1 ⩾ ⋯ ⩾ σ min ⁡ { n , m } ⩾ 0 \sigma_1 \geqslant \dots \geqslant \sigma_{\min\{n,m\}} \geqslant 0 σ1​⩾⋯⩾σmin{n,m}​⩾0, $k<r=\text{rank}(M)$.
压缩图像矩阵$\hat{U}\hat{\Sigma }{\hat{V}}^T=\hat{M}$, σ 1 ⩾ ⋯ ⩾ σ k ⩾ σ ^ k + 1 = ⋯ = σ ^ min ⁡ { n , m } = 0 \sigma_1 \geqslant \dots \geqslant \sigma_k \geqslant \hat \sigma_{k+1} = \dots = \hat \sigma_{\min\{n,m\}} = 0 σ1​⩾⋯⩾σk​⩾σ^k+1​=⋯=σ^min{n,m}​=0.

压缩比$\rho =\frac{nm}{kn+km+k}$
均方误差$\text{MSE}(M,\hat{M})=\frac{1}{nm}\Vert M-\hat{M}{\Vert }_F^2$
峰值信噪比$\text{PSNR}(M,\hat{M})=10{\log }_{10}\frac{{\text{peak}}^2(M)}{\text{MSE}}$

数字水印

图像$M=U\Sigma V^T$, 水印$N$.
含水印的图像的奇异值$\Sigma +\alpha N=U^{\prime }{\Sigma }^{\prime }{V}^{\prime T}$, 含水印的图像$M^{\prime }=U{\Sigma }^{\prime }{V}^T$.
提取水印$N=\frac{1}{\alpha }\left(U^{\prime }{\Sigma }^{\prime }{V}^{\prime T}-\Sigma \right)$.