本文继续进行MMSE均衡器的理论推导。
本文需要根据上一节推导的一般MMSE均衡器表达式
x
^
n
=
E
(
x
n
)
+
Cov
(
x
n
,
z
n
)
Cov
(
z
n
,
z
n
)
−
1
(
z
n
−
E
(
z
n
)
)
\hat{x}_n = \mathbb{E}(x_n) + \text{Cov}(x_n, \mathbf{z}_n) \text{Cov}(\mathbf{z}_n, \mathbf{z}_n)^{-1} (\mathbf{z}_n - \mathbb{E}(\mathbf{z}_n))
x^n=E(xn)+Cov(xn,zn)Cov(zn,zn)−1(zn−E(zn))
证明在ISI信道模型下，该式可转化为：
x
^
n
=
x
‾
n
+
v
n
s
H
(
σ
w
2
I
N
+
H
V
n
H
H
)
−
1
(
z
n
−
H
x
‾
n
)
\hat{x}_n=\overline{x}_n+v_n\mathbf{s}^H(\sigma_w^2\mathbf{I}_N+\mathbf{H}\mathbf{V}_n\mathbf{H}^H)^{-1}(\mathbf{z}_n-\mathbf{H}\overline{\mathbf{x}}_n)
x^n=xn+vnsH(σw2IN+HVnHH)−1(zn−Hxn)
论文原文如下，可跳过
- Exact Implementation:
The MMSE equalizer for this novel approach is an LE consisting of a length N N N filter with time-varying coefficients c n , k c_{n,k} cn,k, k = − N 1 , 1 − N 1 , … , N 2 , k=-N_1,1-N_1,…,N_2, k=−N1,1−N1,…,N2,where N = N 1 + N 2 + 1 N=N_1+N_2+1 N=N1+N2+1, which are defined by the linear MMSE estimate x ^ n \hat{x}_n x^n of x n x_n xn given the observation z n ≜ [ z n − N 2 z n − N 2 + 1 ⋯ z n + N 1 ] T {{\bf{z}}_n} \triangleq {[{z_{n - {N_2}}}{z_{n - {N_2} + 1}} \cdots {z_{n + {N_1}}}]^T} zn≜[zn−N2zn−N2+1⋯zn+N1]T:
x ^ n = E ( x n ) + Cov ( x n , z n ) ⋅ Cov ( z n , z n ) − 1 ⋅ ( z n − E ( z n ) ) \hat{x}_n = \mathbb{E}(x_n) + \text{Cov}(x_n, \mathbf{z}_n) \cdot \text{Cov}(\mathbf{z}_n, \mathbf{z}_n)^{-1} \cdot \left( \mathbf{z}_n - \mathbb{E}(\mathbf{z}_n) \right) x^n=E(xn)+Cov(xn,zn)⋅Cov(zn,zn)−1⋅(zn−E(zn))
For data transmission over an ISI channel, this becomes
x ^ n = x ‾ n + v n s H ( σ w 2 I N + H V n H H ) − 1 ( z n − H x ‾ n ) \hat{x}_n=\overline{x}_n+v_n\mathbf{s}^H(\sigma_w^2\mathbf{I}_N+\mathbf{H}\mathbf{V}_n\mathbf{H}^H)^{-1}(\mathbf{z}_n-\mathbf{H}\overline{\mathbf{x}}_n) x^n=xn+vnsH(σw2IN+HVnHH)−1(zn−Hxn)
where H \mathbf{H} H is the N × ( N + M − 1 ) N\times(N+M-1) N×(N+M−1) channel convolution matrix
H ≜ [ h M − 1 h M − 2 ⋯ h 0 0 ⋯ 0 0 h M − 1 h M − 2 ⋯ h 0 0 ⋯ 0 ⋱ 0 ⋯ 0 h M − 1 h M − 2 ⋯ h 0 ] \mathbf{H}\triangleq\begin{bmatrix}h_{M-1} & h_{M-2} & \cdots & h_0 & 0 & \cdots & & 0\\ 0 & h_{M-1} & h_{M-2} & \cdots & h_0 & 0 & \cdots & 0\\ & & & \ddots & & & & \\ 0 & & \cdots & 0 & h_{M-1} & h_{M-2} & \cdots & h_0\end{bmatrix} H≜ hM−100hM−2hM−1⋯hM−2⋯h0⋯⋱00h0hM−1⋯0hM−2⋯⋯00h0
and
x ‾ n ≜ [ x ‾ n − M − N 2 + 1 x ‾ n − M − N 2 + 2 ⋯ x ‾ n + N 1 ] T \overline{\mathbf{x}}_n\triangleq [\overline{x}_{n-M-N_2+1}\overline{x}_{n-M-N_2+2}\cdots\overline{x}_{n+N_1}]^T xn≜[xn−M−N2+1xn−M−N2+2⋯xn+N1]T V n ≜ Cov ( x n , x n ) = D i a g ( v n − M − N 2 + 1 v n − M − N 2 + 2 ⋯ v n + N 1 ) \mathbf{V}_n\triangleq\text{Cov}(\mathbf{x}_n, \mathbf{x}_n)=\mathrm{Diag}(v_{n-M-N_2+1}v_{n-M-N_2+2}\cdots v_{n+N_1}) Vn≜Cov(xn,xn)=Diag(vn−M−N2+1vn−M−N2+2⋯vn+N1) s ≜ H [ 0 1 × ( N 2 + M − 1 ) 1 0 1 × N 1 ] T \mathbf{s}\triangleq\mathbf{H}\left[\mathbf{0}_{1\times(N_2+M-1)}\quad1\quad\mathbf{0}_{1\times N_1}\right]^T s≜H[01×(N2+M−1)101×N1]T
证明过程如下：
为了推导ISI信道模型下线性MMSE均衡器表达式，关键步骤在于推导 Cov ( z n , z n ) − 1 \text{Cov}(\mathbf{z}_n, \mathbf{z}_n)^{-1} Cov(zn,zn)−1和 Cov ( x n , z n ) \text{Cov}(x_n, \mathbf{z}_n) Cov(xn,zn)：
1. 定义ISI信道模型
假设接收信号
z
n
\mathbf{z}_n
zn 由以下模型生成：
z
n
=
H
x
n
+
w
n
\mathbf{z}_n = \mathbf{H} \mathbf{x}_n + \mathbf{w}_n
zn=Hxn+wn
其中：
-
H
\mathbf{H}
H 是
N
×
(
N
+
M
−
1
)
N \times (N + M - 1)
N×(N+M−1) 的信道卷积矩阵。
H ≜ [ h M − 1 h M − 2 ⋯ h 0 0 ⋯ 0 0 h M − 1 h M − 2 ⋯ h 0 0 ⋯ 0 ⋱ 0 ⋯ 0 h M − 1 h M − 2 ⋯ h 0 ] \mathbf{H}\triangleq\begin{bmatrix}h_{M-1} & h_{M-2} & \cdots & h_0 & 0 & \cdots & & 0\\ 0 & h_{M-1} & h_{M-2} & \cdots & h_0 & 0 & \cdots & 0\\ & & & \ddots & & & & \\ 0 & & \cdots & 0 & h_{M-1} & h_{M-2} & \cdots & h_0\end{bmatrix} H≜ hM−100hM−2hM−1⋯hM−2⋯h0⋯⋱00h0hM−1⋯0hM−2⋯⋯00h0 - x n = [ x n − M − N 2 + 1 , x n − M − N 2 + 2 , … , x n + N 1 ] T \mathbf{x}_n = [x_{n-M-N_2+1}, x_{n-M-N_2+2}, \dots, x_{n+N_1}]^T xn=[xn−M−N2+1,xn−M−N2+2,…,xn+N1]T 是发送符号向量。
- w n \mathbf{w}_n wn 是 0 0 0均值噪声向量， E [ w n ] = 0 \mathbb{E}[\mathbf{w}_n]=0 E[wn]=0
- w n \mathbf{w}_n wn的协方差矩阵为 E [ w n w n H ] = σ w 2 I N \mathbb{E}[\mathbf{w}_n \mathbf{w}_n^H] = \sigma_w^2 \mathbf{I}_N E[wnwnH]=σw2IN，其中 I N \mathbf{I}_N IN是 N N N阶单位阵。
- x ‾ n = E [ x n ] \overline{\mathbf{x}}_n = \mathbb{E}[\mathbf{x}_n] xn=E[xn] 是发送符号的均值向量。
- V n ≜ Cov ( x n , x n ) = Diag ( v n − M − N 2 + 1 , … , v n + N 1 ) \mathbf{V}_n \triangleq \text{Cov}(\mathbf{x}_n, \mathbf{x}_n) = \text{Diag}(v_{n-M-N_2+1}, \dots, v_{n+N_1}) Vn≜Cov(xn,xn)=Diag(vn−M−N2+1,…,vn+N1) 是对角协方差矩阵（符号独立）。
2. 计算 Cov ( z n , z n ) \text{Cov}(\mathbf{z}_n, \mathbf{z}_n) Cov(zn,zn)
Cov ( z n , z n ) = H Cov ( x n , x n ) H H + Cov ( w n , w n ) \text{Cov}(\mathbf{z}_n, \mathbf{z}_n) = \mathbf{H} \text{Cov}(\mathbf{x}_n, \mathbf{x}_n) \mathbf{H}^H + \text{Cov}(\mathbf{w}_n, \mathbf{w}_n) Cov(zn,zn)=HCov(xn,xn)HH+Cov(wn,wn)
计算噪声协方差矩阵：
Cov
(
w
n
,
w
n
)
=
E
[
w
n
w
n
H
]
=
σ
w
2
I
N
\text{Cov}(\mathbf{w}_n, \mathbf{w}_n)=\mathbb{E}[\mathbf{w}_n \mathbf{w}_n^H] = \sigma_w^2 \mathbf{I}_N
Cov(wn,wn)=E[wnwnH]=σw2IN
因此
Cov
(
z
n
,
z
n
)
=
H
V
n
H
H
+
σ
w
2
I
N
\text{Cov}(\mathbf{z}_n, \mathbf{z}_n)= \mathbf{H} \mathbf{V}_n \mathbf{H}^H + \sigma_w^2 \mathbf{I}_N
Cov(zn,zn)=HVnHH+σw2IN
该式中前一项代表符号间干扰，后一项代表噪声。
3.计算 Cov ( x n , z n ) \text{Cov}(x_n, \mathbf{z}_n) Cov(xn,zn)
-
展开协方差：
Cov ( x n , z n ) = E [ ( x n − x ‾ n ) ( z n − E [ z n ] ) H ] \text{Cov}(x_n, \mathbf{z}_n) = \mathbb{E}\left[ (x_n - \overline{x}_n)(\mathbf{z}_n - \mathbb{E}[\mathbf{z}_n])^H \right] Cov(xn,zn)=E[(xn−xn)(zn−E[zn])H] 代入 z n = H x n + w n \mathbf{z}_n = \mathbf{H} \mathbf{x}_n + \mathbf{w}_n zn=Hxn+wn：
= E [ ( x n − x ‾ n ) ( H x n + w n − E [ z n ] ) H ] = \mathbb{E}\left[ (x_n - \overline{x}_n) \left( \mathbf{H} \mathbf{x}_n+ \mathbf{w}_n-\mathbb{E}[\mathbf{z}_n]\right)^H \right] =E[(xn−xn)(Hxn+wn−E[zn])H]
由于 E ( z n ) = H E ( x n ) + E [ w n ] = H x ‾ n \mathbb{E}(\mathbf{z}_n) = \mathbf{H} \mathbb{E}(\mathbf{x}_n)+\mathbb{E}[\mathbf{w}_n] = \mathbf{H} \overline{\mathbf{x}}_n E(zn)=HE(xn)+E[wn]=Hxn，代入得
= E [ ( x n − x ‾ n ) ( H ( x n − x ‾ n ) + w n ) H ] = \mathbb{E}\left[ (x_n - \overline{x}_n) \left( \mathbf{H} (\mathbf{x}_n - \overline{\mathbf{x}}_n) + \mathbf{w}_n \right)^H \right] =E[(xn−xn)(H(xn−xn)+wn)H] = E [ ( x n − x ‾ n ) ( ( x n − x ‾ n ) H H H + w n H ) ] = \mathbb{E}\left[ (x_n - \overline{x}_n) \left((\mathbf{x}_n - \overline{\mathbf{x}}_n)^H\mathbf{H}^H + \mathbf{w}_n^H \right) \right] =E[(xn−xn)((xn−xn)HHH+wnH)] = E [ ( x n − x ‾ n ) ( x n − x ‾ n ) H H H + ( x n − x ‾ n ) w n H ] = \mathbb{E}\left[ (x_n - \overline{x}_n) (\mathbf{x}_n - \overline{\mathbf{x}}_n)^H\mathbf{H}^H + (x_n - \overline{x}_n)\mathbf{w}_n^H \right] =E[(xn−xn)(xn−xn)HHH+(xn−xn)wnH] = E [ ( x n − x ‾ n ) ( x n − x ‾ n ) H H H ] + E [ ( x n − x ‾ n ) w n H ] = \mathbb{E}\left[ (x_n - \overline{x}_n) (\mathbf{x}_n - \overline{\mathbf{x}}_n)^H\mathbf{H}^H\right] + \mathbb{E}\left[ (x_n - \overline{x}_n)\mathbf{w}_n^H \right] =E[(xn−xn)(xn−xn)HHH]+E[(xn−xn)wnH]
由于 x n x_n xn 与噪声 w n \mathbf{w}_n wn 独立，交叉项 E [ ( x n − x ‾ n ) w n H ] = 0 \mathbb{E}\left[ (x_n - \overline{x}_n) \mathbf{w}_n^H \right] = 0 E[(xn−xn)wnH]=0，因此：
Cov ( x n , z n ) = E [ ( x n − x ‾ n ) ( x n − x ‾ n ) H ] H H \text{Cov}(x_n, \mathbf{z}_n)= \mathbb{E}\left[ (x_n - \overline{x}_n) (\mathbf{x}_n - \overline{\mathbf{x}}_n)^H\right]\mathbf{H}^H Cov(xn,zn)=E[(xn−xn)(xn−xn)H]HH -
利用符号独立性：
x n \mathbf{x}_n xn 的协方差为对角矩阵 V n \mathbf{V}_n Vn，则 x n x_n xn 的方差为 v n v_n vn。
设 x n x_n xn 在 x n \mathbf{x}_n xn 中的位置对应单位向量 e \mathbf{e} e，则：
e = [ 0 , … , 1 , … , 0 ] T = [ 0 1 × ( N 2 + M − 1 ) 1 0 1 × N 1 ] T \mathbf{e} = [0, \dots, 1, \dots, 0]^T=\left[\mathbf{0}_{1\times(N_2+M-1)}\quad1\quad\mathbf{0}_{1\times N_1}\right]^T e=[0,…,1,…,0]T=[01×(N2+M−1)101×N1]T
定义零均值变量 x ~ n = x n − x ‾ n \tilde{x}_n=x_n - \overline{x}_n x~n=xn−xn
类似地，再定义零均值符号向量：
x ~ n = x n − x ‾ n = [ x ~ n − M − N 2 + 1 , … , x ~ n , … , x ~ n + N 1 ] T \tilde{\mathbf{x}}_n = \mathbf{x}_n - \overline{\mathbf{x}}_n = [\tilde{x}_{n-M-N_2+1}, \dots,\tilde{x}_n,\dots ,\tilde{x}_{n+N_1}]^T x~n=xn−xn=[x~n−M−N2+1,…,x~n,…,x~n+N1]T
其中 x ~ i = x i − x ‾ i \tilde{x}_i = x_i - \overline{x}_i x~i=xi−xi，满足 E [ x ~ i ] = 0 \mathbb{E}[\tilde{x}_i] = 0 E[x~i]=0 且 E [ x ~ i x ~ j ∗ ] = { v i , i = j 0 , i ≠ j \mathbb{E}[\tilde{x}_i \tilde{x}_j^*] = \begin{cases} v_i, & i=j \\ 0, & i \neq j \end{cases} E[x~ix~j∗]={vi,0,i=ji=j
则代求协方差变为 Cov ( x n , z n ) = E [ x ~ n x ~ n H ] H H \text{Cov}(x_n, \mathbf{z}_n)= \mathbb{E}\left[ \tilde{x}_n\tilde{\mathbf{x}}_n^H \right] \mathbf{H}^H Cov(xn,zn)=E[x~nx~nH]HH
E [ x ~ n x ~ n H ] = E [ x ~ n ⋅ [ x ~ n − M − N 2 + 1 ∗ , … , x ~ n ∗ , … , x ~ n + N 1 ∗ ] ] \mathbb{E}\left[ \tilde{x}_n\tilde{\mathbf{x}}_n^H \right] = \mathbb{E}\left[ \tilde{x}_n\cdot\begin{bmatrix} \tilde{x}_{n-M-N_2+1}^* , \dots,\tilde{x}_n^*,\dots ,\tilde{x}_{n+N_1}^* \end{bmatrix} \right] E[x~nx~nH]=E[x~n⋅[x~n−M−N2+1∗,…,x~n∗,…,x~n+N1∗]]
由于符号之间是相互独立的， E [ x ~ i x ~ j ∗ ] = { v i , i = j 0 , i ≠ j \mathbb{E}[\tilde{x}_i \tilde{x}_j^*] = \begin{cases} v_i, & i=j \\ 0, & i \neq j \end{cases} E[x~ix~j∗]={vi,0,i=ji=j
则
E [ x ~ n x ~ n H ] = E [ [ 0 , … , v n , … , 0 ] ] \mathbb{E}\left[ \tilde{x}_n\tilde{\mathbf{x}}_n^H \right] = \mathbb{E}\left[ \begin{bmatrix} 0 , \dots,v_n,\dots ,0 \end{bmatrix} \right] E[x~nx~nH]=E[[0,…,vn,…,0]] = e T ⋅ v n = \mathbf{e}^T \cdot v_n =eT⋅vn
代入协方差化简得 Cov ( x n , z n ) = v n ⋅ e T H H = v n ⋅ s H \text{Cov}(x_n, \mathbf{z}_n)= v_n \cdot \mathbf{e}^T \mathbf{H}^H= v_n \cdot \mathbf{s}^H Cov(xn,zn)=vn⋅eTHH=vn⋅sH
其中 s ≜ H e \mathbf{s}\triangleq\mathbf{H} \mathbf{e} s≜He 是信道对 x n x_n xn 的响应向量。
s ≜ H [ 0 1 × ( N 2 + M − 1 ) 1 0 1 × N 1 ] T \mathbf{s} \triangleq\mathbf{H}\left[\mathbf{0}_{1\times(N_2+M-1)}\quad1\quad\mathbf{0}_{1\times N_1}\right]^T s≜H[01×(N2+M−1)101×N1]T
- s = H e \mathbf{s} = \mathbf{H} \mathbf{e} s=He 表示信道对符号 x n x_n xn 的冲激响应。在卷积矩阵 H \mathbf{H} H 中， e \mathbf{e} e 对应 x n x_n xn 在发送向量 x n \mathbf{x}_n xn 中的位置，因此 s \mathbf{s} s 是 x n x_n xn 对接收信号 z n \mathbf{z}_n zn 的贡献。
- Cov ( x n , z n ) \text{Cov}(x_n, \mathbf{z}_n) Cov(xn,zn)的表达式说明，由于符号独立性，只有 x n x_n xn 的方差 v n v_n vn 对协方差有贡献，且通过信道响应 s \mathbf{s} s 映射到接收端。
4. 代入一般MMSE公式
将上述结果代入原式：
x
^
n
=
x
‾
n
+
Cov
(
x
n
,
z
n
)
⋅
Cov
(
z
n
,
z
n
)
−
1
⋅
(
z
n
−
H
x
‾
n
)
=
x
‾
n
+
v
n
s
H
(
σ
w
2
I
N
+
H
V
n
H
H
)
−
1
(
z
n
−
H
x
‾
n
)
\begin{aligned} \hat{x}_n &= \overline{x}_n + \text{Cov}(x_n, \mathbf{z}_n) \cdot \text{Cov}(\mathbf{z}_n, \mathbf{z}_n)^{-1} \cdot \left( \mathbf{z}_n - \mathbf{H} \overline{\mathbf{x}}_n \right) \\ &= \overline{x}_n + v_n \mathbf{s}^H \left( \sigma_w^2 \mathbf{I}_N + \mathbf{H} \mathbf{V}_n \mathbf{H}^H \right)^{-1} \left( \mathbf{z}_n - \mathbf{H} \overline{\mathbf{x}}_n \right) \end{aligned}
x^n=xn+Cov(xn,zn)⋅Cov(zn,zn)−1⋅(zn−Hxn)=xn+vnsH(σw2IN+HVnHH)−1(zn−Hxn)
证毕.
扫一扫
4万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
