在通信的文章中，经常会遇见信道估计的问题，而且大多数文献通常会引用S. M. Kay的著作Fundamentals of Statistical Signal Processing: Estimation Theory，却没有给出具体的缘由，这对初学者造成了很大的困扰。下面针对信道估计的问题浅谈一下自己的理解。
1. MMSE估计
以文献Simultaneous Wireless Information and Power Transfer for Downlink Multi-User Massive Antenna-Array Systems为例，经过导频估计后，用户
k
k
k的信号为
y
~
k
=
E
k
β
k
σ
2
h
k
+
n
~
k
{\widetilde {\bf{y}}_k} = \sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{\bf{h}}_k} + {\widetilde {\bf{n}}_k}
y
k=σ2Ekβkhk+n
k
其中
h
k
∼
C
N
(
0
,
I
)
{{\mathbf{h}}_k} \sim \mathcal{C}\mathcal{N}(0,{\mathbf{I}})
hk∼CN(0,I)，
n
~
k
∼
C
N
(
0
,
I
N
)
{{\mathbf{\tilde n}}_k} \sim \mathcal{C}\mathcal{N}\left( {0,{{\mathbf{I}}_N}} \right)
n~k∼CN(0,IN)。
再经过最小均方误差估计(minimum mean square error estimation,MMSE)。具体步骤如下，根据最小均方误差准则，估计量为
h
k
m
s
e
=
∫
−
∞
∞
h
k
p
(
h
k
∣
y
~
k
)
d
h
k
{{\bf{h}}_{kmse}} = \int_{ - \infty }^\infty {{{\bf{h}}_k}p\left( {{{\bf{h}}_k}\left| {{{\widetilde {\bf{y}}}_k}} \right.} \right)} d{{\bf{h}}_k}
hkmse=∫−∞∞hkp(hk∣y
k)dhk
复高斯分布的概率密度函数为
p
(
h
k
)
=
1
π
e
−
h
k
2
p\left( {{{\bf{h}}_k}} \right) = \frac{1}{\pi }{e^{ - {\bf{h}}_k^2}}
p(hk)=π1e−hk2
p
(
y
~
k
∣
h
k
)
=
1
π
e
−
(
y
~
k
−
E
k
β
k
σ
2
h
k
)
2
p\left( {{{{\bf{\tilde y}}}_k}|{{\bf{h}}_k}} \right) = \frac{1}{\pi }{e^{ - {{\left( {{{{\bf{\tilde y}}}_k} - \sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{\bf{h}}_k}} \right)}^2}}}
p(y~k∣hk)=π1e−(y~k−σ2Ekβkhk)2
p ( h k ∣ y ~ k ) = p ( y ~ k ∣ h k ) p ( h k ) p ( y ~ k ) = 1 p ( y ~ k ) 1 π e − ( y ~ k − E k β k σ 2 h k ) 2 1 π e − h k 2 = K 1 ( y ~ k ) e − ( y ~ k 2 − 2 E k β k σ 2 y ~ k h k + E k β k σ 2 h k 2 + h k 2 ) = K 2 ( y ~ k ) e − ( E k β k + σ 2 σ 2 h k 2 − 2 E k β k σ 2 y ~ k h k ) = K 2 ( y ~ k ) e − E k β k + σ 2 σ 2 ( h k 2 − 2 σ E k β k E k β k + σ 2 y ~ k h k ) = K 3 ( y ~ k ) − E k β k + σ 2 σ 2 ( h k − σ E k β k E k β k + σ 2 y ~ k ) 2 \begin{array}{l} p\left( {{{\bf{h}}_k}|{{{\bf{\tilde y}}}_k}} \right) = \frac{{p\left( {{{{\bf{\tilde y}}}_k}|{{\bf{h}}_k}} \right)p\left( {{{\bf{h}}_k}} \right)}}{{p\left( {{{{\bf{\tilde y}}}_k}} \right)}}\\ {\rm{ = }}\frac{1}{{p\left( {{{{\bf{\tilde y}}}_k}} \right)}}\frac{1}{\pi }{e^{ - {{\left( {{{{\bf{\tilde y}}}_k} - \sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{\bf{h}}_k}} \right)}^2}}}\frac{1}{\pi }{e^{ - {\bf{h}}_k^2}}\\ {\rm{ = }}{K_1}\left( {{{{\bf{\tilde y}}}_k}} \right){e^{ - \left( {{{{\bf{\tilde y}}}_k}^2 - 2\sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{{\bf{\tilde y}}}_k}{{\bf{h}}_k} + \frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}{\bf{h}}_k^2 + {\bf{h}}_k^2} \right)}}\\ = {K_2}\left( {{{{\bf{\tilde y}}}_k}} \right){e^{ - \left( {\frac{{{E_k}{\beta _k} + {\sigma ^2}}}{{{\sigma ^2}}}{\bf{h}}_k^2 - 2\sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{{\bf{\tilde y}}}_k}{{\bf{h}}_k}} \right)}}\\ = {K_2}\left( {{{{\bf{\tilde y}}}_k}} \right){e^{ - \frac{{{E_k}{\beta _k} + {\sigma ^2}}}{{{\sigma ^2}}}\left( {{\bf{h}}_k^2 - \frac{{2\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}{{{\bf{\tilde y}}}_k}{{\bf{h}}_k}} \right)}}\\ = {K_3}{\left( {{{{\bf{\tilde y}}}_k}} \right)^{ - \frac{{{E_k}{\beta _k} + {\sigma ^2}}}{{{\sigma ^2}}}{{\left( {{\bf{h}}_k^{} - \frac{{\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}{{{\bf{\tilde y}}}_k}} \right)}^2}}} \end{array} p(hk∣y~k)=p(y~k)p(y~k∣hk)p(hk)=p(y~k)1π1e−(y~k−σ2Ekβkhk)2π1e−hk2=K1(y~k)e−(y~k2−2σ2Ekβky~khk+σ2Ekβkhk2+hk2)=K2(y~k)e−(σ2Ekβk+σ2hk2−2σ2Ekβky~khk)=K2(y~k)e−σ2Ekβk+σ2(hk2−Ekβk+σ22σEkβky~khk)=K3(y~k)−σ2Ekβk+σ2(hk−Ekβk+σ2σEkβky~k)2
最后得到 h ^ k m m s e = σ E k β k E k β k + σ 2 y ~ k = σ E k β k E k β k + σ 2 ( E k β k σ 2 h k + n ~ k ) = E k β k E k β k + σ 2 h k + σ E k β k E k β k + σ 2 n ~ k {{\bf{\hat h}}_{k{\rm{mmse}}}} = \frac{{\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}{{\bf{\tilde y}}_k} = \frac{{\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}\left( {\sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{\bf{h}}_k} + {{\widetilde {\bf{n}}}_k}} \right) = \frac{{{E_k}{\beta _k}}}{{{E_k}{\beta _k} + {\sigma ^2}}}{{\bf{h}}_k}{\rm{ + }}\frac{{\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}{\widetilde {\bf{n}}_k} h^kmmse=Ekβk+σ2σEkβky~k=Ekβk+σ2σEkβk(σ2Ekβkhk+n k)=Ekβk+σ2Ekβkhk+Ekβk+σ2σEkβkn k
方差等于 ( E k β k E k β k + σ 2 ) 2 + ( σ E k β k E k β k + σ 2 ) 2 = E k β k ( E k β k + σ 2 ) ( E k β k + σ 2 ) 2 = E k β k E k β k + σ 2 {\left( {\frac{{{E_k}{\beta _k}}}{{{E_k}{\beta _k} + {\sigma ^2}}}} \right)^2}{\rm{ + }}{\left( {\frac{{\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}} \right)^2}{\rm{ = }}\frac{{{E_k}{\beta _k}\left( {{E_k}{\beta _k} + {\sigma ^2}} \right)}}{{{{\left( {{E_k}{\beta _k} + {\sigma ^2}} \right)}^2}}}{\rm{ = }}\frac{{{E_k}{\beta _k}}}{{{E_k}{\beta _k} + {\sigma ^2}}} (Ekβk+σ2Ekβk)2+(Ekβk+σ2σEkβk)2=(Ekβk+σ2)2Ekβk(Ekβk+σ2)=Ekβk+σ2Ekβk
2. MAP估计
根据最佳估计不变性，当被估计量的后验概率密度函数是高斯型时，最大后验估计(MAP)等价于MMSE。下面给出第二种方法，计算过程较为简单。
∂
ln
p
(
h
k
)
∂
h
k
=
−
2
h
k
\frac{{\partial \ln p\left( {{{\bf{h}}_k}} \right)}}{{\partial {{\bf{h}}_k}}} = - 2{{\bf{h}}_k}
∂hk∂lnp(hk)=−2hk
∂ ln p ( y ~ k ∣ h k ) ∂ h k = 2 E k β k σ 2 ( y ~ k − E k β k σ 2 h k ) \frac{{\partial \ln p\left( {{{{\bf{\tilde y}}}_k}|{{\bf{h}}_k}} \right)}}{{\partial {{\bf{h}}_k}}} = 2\sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} \left( {{{{\bf{\tilde y}}}_k} - \sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{\bf{h}}_k}} \right) ∂hk∂lnp(y~k∣hk)=2σ2Ekβk(y~k−σ2Ekβkhk)
MAP方程有
∂
ln
p
(
y
~
k
∣
h
k
)
∂
h
k
+
∂
ln
p
(
h
k
)
∂
h
k
∣
h
^
k
m
a
p
=
0
\frac{{\partial \ln p\left( {{{{\bf{\tilde y}}}_k}|{{\bf{h}}_k}} \right)}}{{\partial {{\bf{h}}_k}}}{\rm{ + }}\frac{{\partial \ln p\left( {{{\bf{h}}_k}} \right)}}{{\partial {{\bf{h}}_k}}}\left| {_{{{{\bf{\hat h}}}_{k{\rm{map}}}}}} \right. = 0
∂hk∂lnp(y~k∣hk)+∂hk∂lnp(hk)∣∣∣h^kmap=0
2 E k β k σ 2 ( y ~ k − E k β k σ 2 h ^ k m a p ) − 2 h ^ k m a p = 0 E k β k σ 2 y ~ k = E k β k σ 2 h ^ k m a p + σ 2 σ 2 h ^ k m a p h ^ k m a p = σ E k β k E k β k + σ 2 y ~ k = σ E k β k E k β k + σ 2 ( E k β k σ 2 h k + n ~ k ) = E k β k E k β k + σ 2 h k + σ E k β k E k β k + σ 2 n ~ k \begin{array}{l} 2\sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} \left( {{{{\bf{\tilde y}}}_k} - \sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{{\bf{\hat h}}}_{k{\rm{map}}}}} \right) - 2{{{\bf{\hat h}}}_{k{\rm{map}}}}{\rm{ = }}0\\ \sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{{\bf{\tilde y}}}_k} = \frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}{{{\bf{\hat h}}}_{k{\rm{map}}}} + \frac{{{\sigma ^2}}}{{{\sigma ^2}}}{{{\bf{\hat h}}}_{k{\rm{map}}}}\\ {{{\bf{\hat h}}}_{k{\rm{map}}}} = \frac{{\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}{{{\bf{\tilde y}}}_k} = \frac{{\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}\left( {\sqrt {\frac{{{E_k}{\beta _k}}}{{{\sigma ^2}}}} {{\bf{h}}_k} + {{\widetilde {\bf{n}}}_k}} \right) = \frac{{{E_k}{\beta _k}}}{{{E_k}{\beta _k} + {\sigma ^2}}}{{\bf{h}}_k}{\rm{ + }}\frac{{\sigma \sqrt {{E_k}{\beta _k}} }}{{{E_k}{\beta _k} + {\sigma ^2}}}{\widetilde {\bf{n}}_k} \end{array} 2σ2Ekβk(y~k−σ2Ekβkh^kmap)−2h^kmap=0σ2Ekβky~k=σ2Ekβkh^kmap+σ2σ2h^kmaph^kmap=Ekβk+σ2σEkβky~k=Ekβk+σ2σEkβk(σ2Ekβkhk+n k)=Ekβk+σ2Ekβkhk+Ekβk+σ2σEkβkn k
以上过程没有考虑传输 N 次的影响。
扫一扫
5162
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
