Least Square Criteria
模型
-
输入导频序列, x ( n ) , n = 0 , 1 , ⋯ , N − 1 x(n),n=0,1,\cdots,N-1 x(n),n=0,1,⋯,N−1
-
加性噪声序列: ω ( n ) , n = 0 , 1 , ⋯ , N − 1 \omega(n), n= 0, 1, \cdots, N-1 ω(n),n=0,1,⋯,N−1
-
输出序列,即接收序列: y ( n ) , n = 0 , 1 , ⋯ , N − 1 y(n),n=0,1,\cdots,N-1 y(n),n=0,1,⋯,N−1
-
信道系数: h ( n ) , n = 0 , 1 , ⋯ , N − 1 h(n), n=0,1,\cdots,N-1 h(n),n=0,1,⋯,N−1
-
于是模型为:
y ( n ) = x ( n ) ⋅ h ( n ) + ω ( n ) y(n) = x(n) \cdot h(n) + \omega(n) y(n)=x(n)⋅h(n)+ω(n)
推导
-
y = [ y ( 0 ) , y ( 1 ) , ⋯ , y ( N − 1 ) ] T \mathbf{y} = [y(0), y(1), \cdots,y(N-1)]^T y=[y(0),y(1),⋯,y(N−1)]T
-
x = [ x ( 0 ) , x ( 1 ) , ⋯ , x ( N − 1 ) ] T \mathbf{x} = [x(0), x(1), \cdots,x(N-1)]^T x=[x(0),x(1),⋯,x(N−1)]T
-
X = diag ( x ) \mathbf{X} = \text{diag}(\mathbf{x}) X=diag(x)
-
h = [ h ( 0 ) , h ( 1 ) , ⋯ , h ( N − 1 ) ] T \mathbf{h} = [h(0), h(1), \cdots,h(N-1)]^T h=[h(0),h(1),⋯,h(N−1)]T
-
w = [ ω ( 0 ) , ω ( 1 ) , ⋯ , ω ( N − 1 ) ] T \mathbf{w} = [\omega(0), \omega(1), \cdots, \omega(N-1)]^T w=[ω(0),ω(1),⋯,ω(N−1)]T
-
y = X h + w \mathbf{y} = \mathbf{X}\mathbf{h} + \mathbf{w} y=Xh+w
-
代价函数: f = ∣ ∣ y − X h ^ ∣ ∣ 2 f = ||\mathbf{y} - \mathbf{X}\hat{\mathbf{h}}||^2 f=∣∣y−Xh^∣∣2
f = ∣ ∣ y − X h ^ ∣ ∣ 2 = ( y − X h ^ ) H ( y − X h ^ ) = y H y − y H X h ^ − h ^ H X H y + h ^ H X H X h ^ \begin{aligned} f &= ||\mathbf{y} - \mathbf{X}\hat{\mathbf{h}}||^2 \\ &= (\mathbf{y} - \mathbf{X}\hat{\mathbf{h}})^H(\mathbf{y} - \mathbf{X}\hat{\mathbf{h}})\\ &= \mathbf{y}^H \mathbf{y} - \mathbf{y}^H\mathbf{X}\hat{\mathbf{h}}- \hat{\mathbf{h}}^H\mathbf{X}^H\mathbf{y} + \hat{\mathbf{h}}^H\mathbf{X}^H\mathbf{X}\hat{\mathbf{h}} \end{aligned} f=∣∣y−Xh^∣∣2=(y−Xh^)H(y−Xh^)=yHy−yHXh^−h^HXHy+h^HXHXh^
∂ f ∂ h ^ = − y H X − ( X H y ) H + ( X H X h ^ ) H + h ^ H X H X = − 2 y H X + 2 h ^ H X H X \begin{aligned} \frac{\partial f}{\partial \hat{\mathbf{h}}} &= -\mathbf{y}^H\mathbf{X} - (\mathbf{X}^H\mathbf{y})^H + (\mathbf{X}^H\mathbf{X}\hat{\mathbf{h}})^H + \hat{\mathbf{h}}^H\mathbf{X}^H\mathbf{X}\\ &= -2\mathbf{y}^H\mathbf{X} + 2\hat{\mathbf{h}}^H\mathbf{X}^H\mathbf{X} \end{aligned} ∂h^∂f=−yHX−(XHy)H+(XHXh^)H+h^HXHX=−2yHX+2h^HXHX
令
∂
f
∂
h
^
=
0
\frac{\partial f}{\partial \hat{\mathbf{h}}} = 0
∂h^∂f=0,
−
2
y
H
X
+
2
h
^
H
X
H
X
=
0
h
^
H
X
H
X
=
y
H
X
X
H
X
h
^
=
X
H
y
h
^
=
(
X
H
X
)
−
1
X
H
y
\begin{aligned} -2\mathbf{y}^H\mathbf{X} + 2\hat{\mathbf{h}}^H\mathbf{X}^H\mathbf{X} &= 0 \\ \hat{\mathbf{h}}^H\mathbf{X}^H\mathbf{X} &= \mathbf{y}^H\mathbf{X} \\ \mathbf{X}^H\mathbf{X}\hat{\mathbf{h}} &= \mathbf{X}^H \mathbf{y} \\ \hat{\mathbf{h}} &= (\mathbf{X}^H\mathbf{X})^{-1}\mathbf{X}^H \mathbf{y} \end{aligned}
−2yHX+2h^HXHXh^HXHXXHXh^h^=0=yHX=XHy=(XHX)−1XHy
如果
X
\mathbf{X}
X满秩(
x
\mathbf{x}
x是导频序列,应该会满足),那么
h
^
=
X
−
1
y
\hat{\mathbf{h}} = \mathbf{X}^{-1} \mathbf{y}
h^=X−1y
Mean Square Error:
E
(
∣
∣
h
−
h
^
∣
∣
2
)
=
E
(
∣
∣
X
−
1
w
∣
∣
2
)
=
σ
ω
2
σ
x
2
=
1
SNR
\begin{aligned} \text{E}(||\mathbf{h} - \hat{\mathbf{h}}||^2) &= \text{E}(||\mathbf{X}^{-1}\mathbf{w}||^2)\\ &= \frac{\sigma^2_{\omega}}{\sigma^2_{x}} \\ &= \frac{1}{\text{SNR}}\end{aligned}
E(∣∣h−h^∣∣2)=E(∣∣X−1w∣∣2)=σx2σω2=SNR1
这个与SNR的反比的关系,导致在SNR很低的时候,MSE以反比的方式放大的很大。