MMSE 检测算法推导

考虑发送端具有 \(N_t\) 根发送天线,接收端具有 \(N_r\) 根接收天线,在一个时隙内信道为准静态平坦衰落情况下,接收信号可表示为

\[ \mathbf{y=Hx+n} \tag{1} \]

其中 \(\mathbf{H} \in \mathbb{C}^{N_r \times N_t}\) 是 MIMO 信道矩阵,\(\mathbf{x} \in \mathbb{C}^{N_t \times 1}\) 是发送信号向量,\(\mathbf{n} \in \mathbb{C}^{N_r \times 1}\) 是与发送信号向量不相关的加性噪声向量,假设噪声向量均值为零。

预备知识

向量 \(\mathbf{x}\)\(\mathbf{n}\) 有如下自相关矩阵

\[ \mathbf{R}_x = E[\mathbf{x} \mathbf{x}^H] \tag{2} \]

\[ \mathbf{R}_n = E[\mathbf{n} \mathbf{n}^H] \tag{3} \]

已假设向量 \(\mathbf{x}\)\(\mathbf{n}\) 不相关,则 \(\mathbf{x}\)\(\mathbf{n}\) 互相关矩阵为 \(\mathbf{0}\) 矩阵

\[ \mathbf{R}_{xn} = E[\mathbf{x} \mathbf{n}^H] = \mathbf{0}_{N_t \times N_r} \tag{4} \]

\[ \mathbf{R}_{nx} = E[\mathbf{n} \mathbf{x}^H] = \mathbf{0}_{N_r \times N_t} \tag{5} \]

接收信号向量 \(\mathbf{y}\) 的自相关矩阵

\[\begin{equation} \tag{6} \begin{aligned} \mathbf{R}_y &= E\left[\mathbf{y} \mathbf{y}^H \right] \\ &= E \left[ \left(\mathbf{Hx + n} \right) \left(\mathbf{Hx + n} \right)^H \right] \\ &= E \left[\mathbf{H} \mathbf{x} \mathbf{x}^H \mathbf{H}^H + \mathbf{H} \mathbf{x} \mathbf{n}^H + \mathbf{n} \mathbf{x}^H \mathbf{H}^H + \mathbf{n} \mathbf{n}^H \right] \\ &= \mathbf{H} E\left[\mathbf{x} \mathbf{x}^H \right] + \mathbf{H} E\left[\mathbf{x} \mathbf{n}^H \right] + E\left[\mathbf{n} \mathbf{x}^H \right] \mathbf{H}^H + E\left[\mathbf{n} \mathbf{n}^H \right] \\ &= \mathbf{H} \mathbf{R}_x \mathbf{H}^H + \mathbf{R}_n \end{aligned} \end{equation}\]

这些相关矩阵将在下面的推导中使用。

MMSE检测算法

接收端估计的信号为

\[ \mathbf{\hat{x}} = \mathbf{W} \mathbf{y} \tag{7} \]

则估计信号的误差为

\[ \mathbf{e} = \mathbf{\hat{x}} - \mathbf{x} = \mathbf{W} \mathbf{y} - \mathbf{x} \tag{8} \]

估计信号的均方误差为

\[ \mathbf{e}_{MSE} = E \left\| \mathbf{Wy - x} \right\|^2 \tag{9} \]

MMSE 检测算法以最小均方误差为准则,最小化实际发送的符号和检测器输出估计值之间的均方误差。当 \(\mathbf{e}_{MSE}\) 达到最小时,接收信号 \(\mathbf{y}\) 的加权矩阵 \(\mathbf{W}\)\(\mathbf{W}_{MMSE}\)

\[ \mathbf{W}_{MMSE} = \mathop{\arg\min}_{\mathbf{w}} \ E\left\| \mathbf{Wy - x} \right\|^2 \tag{10} \]

推导 \(\mathbf{W}_{MMSE}\) 有两种方法,一是利用正交性原理,能够很方便地推出结果;二是从式(9)出发,对 \(\mathbf{W}_{MMSE}\) 求导,并令导函数为 0,转化为求极值问题。

推导方法一:正交性原理

正交性原理:估计误差 \(\mathbf{e}\) 是一个随机变量,定义代价函数为均方误差 \(\mathbf{e}_{MSE} = E\left\| \mathbf{Wy - x} \right\|\),则使均方误差 \(\mathbf{e}_{MSE}\) 获得最小值的条件是:

\[ E\left[ \mathbf{e}_0 \mathbf{y}^H \right] = \mathbf{0} \tag{11} \]

此时的 \(\mathbf{e}_0\) 是在均方误差意义上的最小值。这就是正交性原理。

将式(8)带入式(11)得

\[ E\left[ \left( \mathbf{W}_{MMSE} \mathbf{y} - \mathbf{x} \right) \mathbf{y}^H \right] = \mathbf{0} \tag{12} \]

下面对式(12)化简:

\[ E\left[ \mathbf{W}_{MMSE} \mathbf{y} \mathbf{y}^H - \mathbf{x} \mathbf{y}^H \right] = \mathbf{0} \tag{13} \]

\[ E\left[\mathbf{W}_{MMSE} \left( \mathbf{Hx + n} \right) \left(\mathbf{Hx + n}\right)^H - \mathbf{x} \left( \mathbf{Hx + n} \right)^H \right] = \mathbf{0} \tag{14} \]

\[ E\left[\mathbf{W}_{MMSE} \left( \mathbf{Hx + n} \right) \left( \mathbf{x}^H \mathbf{H}^H + \mathbf{n}^H \right) - \mathbf{x} \left( \mathbf{x}^H \mathbf{H}^H + \mathbf{n}^H \right) \right] = \mathbf{0} \tag{15} \]

\[ E\left[ \mathbf{W}_{MMSE} \left( \mathbf{H} \mathbf{x} \mathbf{x}^H \mathbf{H}^H + \mathbf{H} \mathbf{x} \mathbf{n}^H + \mathbf{n} \mathbf{x}^H \mathbf{H}^H + \mathbf{n} \mathbf{n}^H \right) - \left(\mathbf{x} \mathbf{x}^H \mathbf{H}^H + \mathbf{x} \mathbf{n}^H \right) \right] = \mathbf{0} \tag{16} \]

\[ \mathbf{W}_{MMSE} \left( \mathbf{H} E\left[ \mathbf{x} \mathbf{x}^H \right] \mathbf{H}^H + \mathbf{H} E\left[ \mathbf{x} \mathbf{n}^H \right] + \left[ \mathbf{n} \mathbf{x}^H \right] \mathbf{H}^H + E\left[ \mathbf{n} \mathbf{n}^H \right] \right) - \left( \left[ \mathbf{x} \mathbf{x}^H \right] \mathbf{H}^H + E\left[ \mathbf{x} \mathbf{n}^H \right] \right)= \mathbf{0} \tag{17} \]

\[ \mathbf{W}_{MMSE} \left( \mathbf{H} E\left[ \mathbf{x} \mathbf{x}^H \right] \mathbf{H}^H + E\left[ \mathbf{x} \mathbf{x}^H \right] \mathbf{H}^H \right) = \mathbf{0} \tag{18} \]

最后导出 \(\mathbf{W}_{MMSE}\)

\[ \mathbf{W}_{MMSE} = \mathbf{R}_x \mathbf{H}^H \left( \mathbf{H} \mathbf{R}_x \mathbf{H}^H + \mathbf{R}_n \right)^{-1} \tag{19} \]

推导方法二:矩阵求导

如果不使用正交性原理,要想找到使 \(\mathbf{e}_{MSE}\) 最小的 \(\mathbf{W}_{MMSE}\),步骤是 step 1)对 \(\mathbf{e}_{MSE}\) 求导,step 2)令 \(\mathbf{e}_{MSE}\) 的导函数等于 0,并求得极值点。

式(9)可重写为

\[ \begin{equation} \tag{20} \begin{aligned} \mathbf{e}_{MMSE} &= E\left\| \mathbf{Wy -x} \right\|^2 \\ &= E\left\{ tr \left[ \left( \mathbf{Wy - x} \right) \left( \mathbf{Wy - x} \right)^H \right] \right\} \\ &= E\left\{ tr \left[ \mathbf{W} \mathbf{y} \mathbf{y}^H \mathbf{W}^H - \mathbf{W} \mathbf{y} \mathbf{x}^H - \mathbf{x} \mathbf{y}^H \mathbf{W}^H + \mathbf{x} \mathbf{x}^H \right] \right\} \\ &= tr\left\{ \mathbf{W} E\left[ \mathbf{y} \mathbf{y}^H \right] \mathbf{W}^H - \mathbf{W} E\left[ \mathbf{y} \mathbf{x}^H \right] - E\left[ \mathbf{x} \mathbf{y}^H \right] \mathbf{W}^H + E\left[ \mathbf{x} \mathbf{x}^H \right]\right\} \end{aligned} \end{equation} \]

式(20)第一项:

\[ \mathbf{W} E\left[ \mathbf{y} \mathbf{y}^H \right] \mathbf{W}^H = \mathbf{W} \mathbf{R}_y \mathbf{W}^H \tag{21} \]

式(20)第二项:

\[ \begin{equation} \tag{22} \begin{aligned} \mathbf{W} E\left[ \mathbf{y} \mathbf{x}^H \right] &= \mathbf{W} E\left[ \left( \mathbf{Hx + n} \right) \mathbf{x}^H \right] \\ &= \mathbf{W} E\left[ \mathbf{H} \mathbf{x} \mathbf{x}^H + \mathbf{n} \mathbf{x}^H\right] \\ &= \mathbf{W} \mathbf{H} \mathbf{R}_x \end{aligned} \end{equation} \]

式(20)第三项:

\[ \begin{equation} \tag{23} \begin{aligned} E\left[ \mathbf{x} \mathbf{y}^H \right] \mathbf{W}^H &= E\left[ \mathbf{x} \left( \mathbf{Hx + n} \right)^H \right] \mathbf{W}^H \\ &= E\left[ \mathbf{x} \mathbf{x}^H \mathbf{H}^H + \mathbf{x} \mathbf{n}^H \right] \mathbf{W}^H \\ &= \mathbf{R}_x \mathbf{H}^H \mathbf{W}^H \end{aligned} \end{equation} \]

将式(2),式(21~23)带入式(20)可得

\[ \mathbf{e}_{MSE} = tr\left\{ \mathbf{W} \mathbf{R}_y \mathbf{W}^H - \mathbf{W} \mathbf{H} \mathbf{R}_x - \mathbf{R}_x \mathbf{H}^H \mathbf{W}^H + \mathbf{R}_x \right\} \tag{24} \]

\(\mathbf{e}_{MSE}\)\(\mathbf{W}\) 的偏导:

\[ \frac{\partial \mathbf{e}_{MSE} }{ \partial \mathbf{W} } = \frac{ \partial tr \left\{\mathbf{W} \mathbf{R}_y \mathbf{W}^H - \mathbf{W} \mathbf{H} \mathbf{R}_x - \mathbf{R}_x \mathbf{H}^H \mathbf{W}^H + \mathbf{R}_x \right\} }{ \partial \mathbf{W}} \tag{25} \]

式(25)分子第一项的偏导为:

\[ \frac{\partial tr \left\{ \mathbf{W} \mathbf{R}_y \mathbf{W}^H \right\}}{ \partial \mathbf{W}} = \left( \mathbf{R}_y \mathbf{W}^H \right)^T = \mathbf{W}^* \mathbf{R}_y^T \tag{26} \]

式(25)分子第二项的偏导为:

\[ \frac{\partial tr \left\{ \mathbf{W} \mathbf{H} \mathbf{R}_x \right\}}{\partial \mathbf{W}} = \left( \mathbf{H} \mathbf{R}_x \right)^T \tag{27} \]

式(25)分子第三项的偏导为:

\[ \frac{\partial tr \left\{ \mathbf{R}_x \mathbf{H}^H \mathbf{W}^H \right\}}{\partial \mathbf{W}} = \mathbf{0} \tag{28} \]

式(25)分子第四项的偏导为:

\[ \frac{\partial tr\left\{ \mathbf{R}_x \right\}}{\partial \mathbf{W}} = \mathbf{0} \tag{29} \]

将式(26~29)带入式(25)得

\[ \frac{\partial \mathbf{e}_{MSE}}{\partial \mathbf{W}} = \mathbf{W}^* \mathbf{R}_y^T - \left( \mathbf{H} \mathbf{R}_x \right)^T \tag{30} \]

\(\mathbf{e}_{MSE}\) 导函数等于 \(\mathbf{0}\)

\[ \mathbf{W}^* \mathbf{R}_y^T - \left( \mathbf{H} \mathbf{R}_x \right)^T = \mathbf{0} \tag{31} \]

对上式等号两边同时取复共轭:

\[ \mathbf{W} \mathbf{R}_y^H = \left( \mathbf{H} \mathbf{R}_x \right)^H \tag{32} \]

由于自相关矩阵是 Hermitian 矩阵,即 \(\mathbf{R}_y^H = \mathbf{R}_y\)\(\mathbf{R}_x^H = \mathbf{R}_x\),则上式改写为

\[ \mathbf{W} \mathbf{R}_y = \mathbf{R}_x \mathbf{H}^H \tag{33} \]

则有

\[ \mathbf{W} = \mathbf{R}_x \mathbf{H}^H \mathbf{R}_y^{-1} \tag{34} \]

将式(6)带入上式得

\[ \mathbf{W}_{MMSE} = \mathbf{R}_x \mathbf{H}^H \left( \mathbf{H} \mathbf{R}_x \mathbf{H}^H + \mathbf{R}_n \right)^{-1} \]

由此看见,用矩阵求导的方法得到的式(35)和用正交性原理得到的式(19)相同。

若令

\[ \mathbf{R}_x = E\left[ \mathbf{x} \mathbf{x}^H \right] = \mathbf{I} \]

\[ \mathbf{R}_n = E\left[ \mathbf{n} \mathbf{n}^H \right] = \sigma^2 \mathbf{I} \]

带入式(35)得到

\[ \mathbf{W}_{MMSE} = \mathbf{H}^H \left( \mathbf{H} \mathbf{H}^H + \sigma^2 \mathbf{I} \right)^{-1} \tag{36} \]

完毕。