相关高斯分布的MIMO信道矩阵的容量统计特征推导
本文是IEEE TIT(Trans on Information Theory)一篇文章的阅读笔记,文章链接如下:
A. L. Moustakas, S. H. Simon, and A. M. Sengupta, “MIMO capacity
through correlated channels in the presence of correlated interferers and
noise: A (not so) large N analysis,” IEEE Trans. Inf. Theory, vol. 49,
no. 10, pp. 2545–2561, Oct. 2003.
文章梗概
本文的主要工作是给出了相关MIMO信道模型下容量的估计方法。
文章结构:
- I-A开始,介绍文章所用的部分数学表述及系统模型;
- II介绍本文所用的数学架构;并在II-A中讨论无interferer且AWGN的情况,在II-B中讨论有interferer且有相关噪声的情况;
- IV介绍在其他场景中的应用,此处不表。
文章部分数学表述及系统模型
数学表述:定义在矩阵元上的三种积分
1、
D
c
X
=
∏
a
=
1
m
rows
∏
α
=
1
m
cols
d
Re
X
a
α
d
Im
X
a
α
2
π
D_{c} \boldsymbol{X}=\prod_{a=1}^{m_{\text {rows }}} \prod_{\alpha=1}^{m_{\text {cols }}} \frac{d \operatorname{Re} X_{a \alpha} d \operatorname{Im} X_{a \alpha}}{2 \pi}
DcX=∏a=1mrows ∏α=1mcols 2πdReXaαdImXaα
2、
d
μ
(
T
,
R
)
=
∏
a
=
1
m
r
o
w
s
∏
α
=
1
m
c
o
l
s
d
T
a
α
d
R
α
a
2
π
i
d \mu(\boldsymbol{T}, \boldsymbol{R})=\prod_{a=1}^{m_{\mathrm{rows}}} \prod_{\alpha=1}^{m_{\mathrm{cols}}} \frac{d T_{a \alpha} d R_{\alpha a}}{2 \pi i}
dμ(T,R)=∏a=1mrows∏α=1mcols2πidTaαdRαa
3、
D
g
A
=
∏
a
=
1
m
rows
∏
α
=
1
m
cols
d
A
a
α
d
A
ˉ
a
α
D_{g} \boldsymbol{A}=\prod_{a=1}^{m_{\text {rows }}} \prod_{\alpha=1}^{m_{\text {cols }}} d A_{a \alpha} d \bar{A}_{a \alpha}
DgA=∏a=1mrows ∏α=1mcols dAaαdAˉaα
系统模型
传输过程表述如下:
y
=
ρ
s
n
t
G
s
x
s
+
ρ
i
n
i
G
i
x
i
+
z
(1)
\boldsymbol{y}=\sqrt{\frac{\rho_{s}}{n_{t}}} \boldsymbol{G}^{\boldsymbol{s}} \boldsymbol{x}^{s}+\sqrt{\frac{\rho_{i}}{n_{i}}} \boldsymbol{G}^{\boldsymbol{i}} \boldsymbol{x}^{i}+\boldsymbol{z} \tag{1}
y=ntρsGsxs+niρiGixi+z(1)其中
G
s
\boldsymbol{G}^{\boldsymbol{s}}
Gs是一个
n
r
×
n
t
n_r \times n_t
nr×nt的复矩阵,表示信号信道矩阵,
G
i
\boldsymbol{G}^{\boldsymbol{i}}
Gi是
n
r
×
n
i
n_r \times n_i
nr×ni的复矩阵,表示干扰信道矩阵;
x
s
\boldsymbol{x}^{\boldsymbol{s}}
xs与
x
i
\boldsymbol{x}^{\boldsymbol{i}}
xi表示
n
t
n_t
nt维及
n
r
n_r
nr维的发射信号与干扰信号向量,都是零均值的高斯变量。发射信号功率限制为
Q
=
E
[
x
s
x
s
†
]
,
Tr
{
Q
}
=
n
t
\boldsymbol{Q}=E\left[\boldsymbol{x}^{\boldsymbol{s}} \boldsymbol{x}^{\boldsymbol{s}^{\dagger}}\right] ,\operatorname{Tr}\{\boldsymbol{Q}\}=n_{t}
Q=E[xsxs†],Tr{Q}=nt。
数学分析架构
一般架构
互信息生成函数有: g ( ν ) = ⟨ [ det ( N + ρ i n i G i G i † + ρ s n t G s Q G s † ) ] − ν [ det ( N + ρ i n i G i G i † ) ] − ν ⟩ = ⟨ e − ν I ⟩ = 1 − ν ⟨ I ⟩ + ν 2 2 ⟨ I 2 ⟩ + ⋯ (2) \begin{aligned} g(\nu) &=\left\langle\frac{\left[\operatorname{det}\left(N+\frac{\rho_{i}}{n_{i}} G^{i} G^{i^{\dagger}}+\frac{\rho_{s}}{n_{t}} G^{s} Q G^{s \dagger}\right)\right]^{-\nu}}{\left[\operatorname{det}\left(N+\frac{\rho_{i}}{n_{i}} G^{i} G^{i^{\dagger}}\right)\right]^{-\nu}}\right\rangle \\ &=\left\langle e^{-\nu I}\right\rangle \\ &=1-\nu\langle I\rangle+\frac{\nu^{2}}{2}\left\langle I^{2}\right\rangle+\cdots \end{aligned} \tag{2} g(ν)=⟨[det(N+niρiGiGi†)]−ν[det(N+niρiGiGi†+ntρsGsQGs†)]−ν⟩=⟨e−νI⟩=1−ν⟨I⟩+2ν2⟨I2⟩+⋯(2)假设 g ( ν ) g(\nu) g(ν)至少在 ν = 0 \nu=0 ν=0时是解析的,则 l o g g ( ν ) \mathrm{log}g(\nu) logg(ν)可以表述为: log g ( ν ) = − ν ⟨ I ⟩ + ∑ p = 2 ∞ ( − ν ) p p ! C p (3) \log g(\nu)=-\nu\langle I\rangle+\sum_{p=2}^{\infty} \frac{(-\nu)^{p}}{p !} \mathcal{C}_{p} \tag{3} logg(ν)=−ν⟨I⟩+p=2∑∞p!(−ν)pCp(3)其中 C p \mathcal{C}_p Cp为 I I I的 p p p阶累计矩,例如 C 2 = Var ( I ) = ⟨ ( I − ⟨ I ⟩ ) 2 ⟩ \mathcal{C}_{2}=\operatorname{Var}(I)=\left\langle(I-\langle I\rangle)^{2}\right\rangle C2=Var(I)=⟨(I−⟨I⟩)2⟩为方差, C 3 = S k ( I ) = ⟨ ( I − ⟨ I ⟩ ) 3 ⟩ \mathcal{C}_{3}=S k(I)=\left\langle(I-\langle I\rangle)^{3}\right\rangle C3=Sk(I)=⟨(I−⟨I⟩)3⟩为分布偏态。
AWGN噪声且无干扰
此时噪声矩阵为
N
=
I
n
r
\boldsymbol{N}=\boldsymbol{I}_{n_{r}}
N=Inr,则容量表达式为:
g
(
ν
)
=
⟨
[
det
(
I
n
r
+
ρ
s
n
t
G
s
Q
G
s
†
)
]
−
ν
⟩
(4)
g(\nu)=\left\langle\left[\operatorname{det}\left(\boldsymbol{I}_{n_{r}}+\frac{\rho_{s}}{n_{t}} \boldsymbol{G}^{\boldsymbol{s}} \boldsymbol{Q} \boldsymbol{G}^{\boldsymbol{s} \dagger}\right)\right]^{-\nu}\right\rangle \tag{4}
g(ν)=⟨[det(Inr+ntρsGsQGs†)]−ν⟩(4)
设
ν
\nu
ν为任意正整数,由Appendix I中Identity 1,上式可写为:
g
(
ν
)
=
∫
D
c
X
e
−
1
2
Tr
{
X
†
X
}
⟨
e
−
ρ
s
2
n
t
Tr
{
X
†
G
s
Q
G
s
†
X
}
⟩
(5)
g(\nu)=\int D_{c} \boldsymbol{X} e^{-\frac{1}{2} \operatorname{Tr}\left\{\boldsymbol{X}^{\dagger} \boldsymbol{X}\right\}}\left\langle e^{-\frac{\rho_{s}}{2 n_{t}} \operatorname{Tr}\left\{\boldsymbol{X}^{\dagger} \boldsymbol{G}^{s} \boldsymbol{Q} \boldsymbol{G}^{\boldsymbol{s}\dagger} \boldsymbol{X}\right\}}\right\rangle \tag{5}
g(ν)=∫DcXe−21Tr{X†X}⟨e−2ntρsTr{X†GsQGs†X}⟩(5)其中
X
\boldsymbol{X}
X为复的
n
r
×
ν
n_r \times \nu
nr×ν维矩阵。
上式中括号中项利用Identity 2可以被改写为:
⟨
e
−
ρ
s
2
n
t
Tr
{
X
†
G
s
Q
G
s
†
X
}
⟩
=
∫
D
c
Y
e
−
1
2
Tr
Y
†
Y
×
⟨
e
−
ρ
S
4
n
t
Tr
{
X
†
G
s
Q
1
/
2
Y
−
Y
†
Q
1
/
2
G
s
†
X
}
⟩
(6)
\begin{array}{r} \left\langle e^{-\frac{\rho_{s}}{2 n_{t}} \operatorname{Tr}\left\{\boldsymbol{X}^{\dagger} \boldsymbol{G}^{s} \boldsymbol{Q G}^{\boldsymbol{s} \dagger} \boldsymbol{X}\right\}}\right\rangle=\int D_{c} \boldsymbol{Y} e^{-\frac{1}{2} \operatorname{Tr} \boldsymbol{Y}^{\dagger} \boldsymbol{Y}} \\ \times\left\langle e^{-\sqrt{\frac{\rho_{S}}{4 n_{t}}} \operatorname{Tr}\left\{\boldsymbol{X}^{\dagger} \boldsymbol{G}^{s} \boldsymbol{Q}^{1 / 2} \boldsymbol{Y}-\boldsymbol{Y}^{\dagger} \boldsymbol{Q}^{1 / 2} \boldsymbol{G}^{\boldsymbol{s} \dagger} \boldsymbol{X}\right\}}\right\rangle \end{array} \tag{6}
⟨e−2ntρsTr{X†GsQGs†X}⟩=∫DcYe−21TrY†Y×⟨e−4ntρSTr{X†GsQ1/2Y−Y†Q1/2Gs†X}⟩(6)则
g
(
ν
)
g(\nu)
g(ν)可以写为:
g
(
ν
)
=
∫
D
c
X
D
c
Y
⋅
e
−
1
2
Tr
{
X
†
X
+
Y
†
Y
+
ρ
s
2
n
t
Y
†
Q
1
/
2
T
s
Q
1
/
2
Y
X
†
R
s
X
}
(7)
\begin{aligned} g(\nu)=& \int D_{c} \boldsymbol{X} D_{c} \boldsymbol{Y} \\ & \cdot e^{-\frac{1}{2} \operatorname{Tr}\left\{\boldsymbol{X}^{\dagger} \boldsymbol{X}+\boldsymbol{Y}^{\dagger} \boldsymbol{Y}+\frac{\rho_{s}}{2 n_{t}} \boldsymbol{Y}^{\dagger} \boldsymbol{Q}^{1 / 2} \boldsymbol{T}^{s} \boldsymbol{Q}^{1 / 2} \boldsymbol{Y} \boldsymbol{X}^{\dagger} \boldsymbol{R}^{s} \boldsymbol{X}\right\}} \end{aligned} \tag{7}
g(ν)=∫DcXDcY⋅e−21Tr{X†X+Y†Y+2ntρsY†Q1/2TsQ1/2YX†RsX}(7)将指数的最后一项用
X
,
Y
\boldsymbol{X,Y}
X,Y表示为二次型(利用Identity 3并引入
ν
×
ν
\nu \times \nu
ν×ν维矩阵
R
1
,
T
1
\boldsymbol{R}_1,\boldsymbol{T}_1
R1,T1)则指数的最后一项可写为:
exp
[
ρ
s
2
n
t
Tr
{
Y
†
Q
1
/
2
T
s
Q
1
/
2
Y
X
†
R
s
X
}
]
=
∫
d
μ
(
T
1
,
R
1
)
exp
[
Tr
{
T
1
R
1
}
−
ρ
s
4
n
t
×
Tr
{
T
1
Y
†
Q
1
/
2
T
s
Q
1
/
2
Y
+
X
†
R
s
X
R
1
}
]
(8)
\begin{array}{l} \exp \left[\frac{\rho_{s}}{2 n_{t}} \operatorname{Tr}\left\{\boldsymbol{Y}^{\dagger} \boldsymbol{Q}^{1 / 2} \boldsymbol{T}^{s} \boldsymbol{Q}^{1 / 2} \boldsymbol{Y} \boldsymbol{X}^{\dagger} \boldsymbol{R}^{s} \boldsymbol{X}\right\}\right] \\ =\int d \mu\left(\boldsymbol{T}_{1}, \boldsymbol{R}_{1}\right) \exp \left[\operatorname{Tr}\left\{\boldsymbol{T}_{1} \boldsymbol{R}_{1}\right\}-\sqrt{\frac{\rho_{s}}{4 n_{t}}}\right. \\ \left.\times \operatorname{Tr}\left\{\boldsymbol{T}_{1} \boldsymbol{Y}^{\dagger} \boldsymbol{Q}^{1 / 2} \boldsymbol{T}^{s} \boldsymbol{Q}^{1 / 2} \boldsymbol{Y}+\boldsymbol{X}^{\dagger} \boldsymbol{R}^{s} \boldsymbol{X} \boldsymbol{R}_{1}\right\}\right] \end{array} \tag{8}
exp[2ntρsTr{Y†Q1/2TsQ1/2YX†RsX}]=∫dμ(T1,R1)exp[Tr{T1R1}−4ntρs×Tr{T1Y†Q1/2TsQ1/2Y+X†RsXR1}](8)这种方法叫Hubbard–Stratonovich变换,在“Distributions of singular values for some random matrices,”中也有应用,该方法处理上面这种指数矩阵解耦非常有用。
上述过程处理完成后,
g
(
ν
)
g(\nu)
g(ν)表述如下:
g
(
ν
)
=
∫
d
μ
(
T
1
,
R
1
)
e
−
S
(9)
g(\nu)=\int d \mu\left(\boldsymbol{T}_{1}, \boldsymbol{R}_{1}\right) e^{-\mathcal{S}} \tag{9}
g(ν)=∫dμ(T1,R1)e−S(9)其中
S
=
log
det
(
I
n
r
⊗
I
ν
+
ρ
s
n
t
R
s
⊗
R
1
)
+
log
det
(
I
n
t
⊗
I
ν
+
ρ
s
n
t
Q
1
/
2
T
s
Q
1
/
2
⊗
T
1
)
−
Tr
{
T
1
R
1
}
.
(10)
\begin{aligned} \mathcal{S}=& \log \operatorname{det}\left(\boldsymbol{I}_{n_{r}} \otimes \boldsymbol{I}_{\nu}+\sqrt{\frac{\rho_{s}}{n_{t}}} \boldsymbol{R}^{s} \otimes \boldsymbol{R}_{1}\right) \\ &+\log \operatorname{det}\left(\boldsymbol{I}_{n_{t}} \otimes \boldsymbol{I}_{\nu}+\sqrt{\frac{\rho_{s}}{n_{t}}} \boldsymbol{Q}^{1 / 2} \boldsymbol{T}^{s} \boldsymbol{Q}^{1 / 2} \otimes \boldsymbol{T}_{1}\right) \\ &-\operatorname{Tr}\left\{\boldsymbol{T}_{1} \boldsymbol{R}_{1}\right\} . \end{aligned} \tag{10}
S=logdet(Inr⊗Iν+ntρsRs⊗R1)+logdet(Int⊗Iν+ntρsQ1/2TsQ1/2⊗T1)−Tr{T1R1}.(10)到此处,后面可以用鞍点逼近来分析。
鞍点逼近法是一种获得复空间中快变函数积分系统展开式的方法。这种方法包括变形积分轮廓,使其通过一个(或多个)点,其中被积函数的指数是平稳的,即指数对所有积分变量的导数消失的点[24]。因此,沿最陡下降路径的点周围的积分渐进地携带积分的最高权重。通常,当被积函数的指数较大时,这种近似是有效的,在这种情况下,这种方法也被称为拉普拉斯方法,或Varadhan定理。
由于鞍点近似解决的是复平面积分 I ( λ ) = ∫ C d z e λ g ( z ) I(\lambda)=\int_{C} d z e^{\lambda g(z)} I(λ)=∫Cdzeλg(z)当 λ \lambda λ取大值时的近似问题。其中 g ( z ) g(z) g(z)应该是 z z z的解析函数。所以到此上述部分都是在将互信息凑成可以用鞍点分析的形式。
鞍点分析
(
7
)
(7)
(7)式中
T
1
\boldsymbol{T}_1
T1及
R
1
\boldsymbol{R}_1
R1在鞍点处的形式为
t
1
n
t
I
ν
t_{1} \sqrt{n_{t}} \boldsymbol{I}_{\nu}
t1ntIν及
r
1
n
t
I
ν
r_{1} \sqrt{n_{t}} \boldsymbol{I}_{\nu}
r1ntIν,若将除
n
t
\sqrt{n_{t}}
nt以外的项写入方差,则逼近的每一步
T
1
\boldsymbol{T}_1
T1及
R
1
\boldsymbol{R}_1
R1可写为
T
1
=
t
1
n
t
I
ν
+
δ
T
1
R
1
=
r
1
n
t
I
ν
+
δ
R
1
(11)
\begin{array}{l} \boldsymbol{T}_{1}=t_{1} \sqrt{n_{t}} \boldsymbol{I}_{\nu}+\boldsymbol{\delta} \boldsymbol{T}_{1} \\ \boldsymbol{R}_{1}=r_{1} \sqrt{n_{t}} \boldsymbol{I}_{\nu}+\boldsymbol{\delta} \boldsymbol{R}_{1} \end{array} \tag{11}
T1=t1ntIν+δT1R1=r1ntIν+δR1(11)则
(
8
)
(8)
(8)式中
S
\mathcal{S}
S可以用方差矩阵
δ
T
1
,
δ
R
1
\boldsymbol{\delta} \boldsymbol{T}_{1},\boldsymbol{\delta} \boldsymbol{R}_{1}
δT1,δR1的泰勒展开式展开:
S
=
S
0
+
S
1
+
S
2
+
S
3
+
S
4
⋯
(12)
\mathcal{S}=\mathcal{S}_{0}+\mathcal{S}_{1}+\mathcal{S}_{2}+\mathcal{S}_{3}+\mathcal{S}_{4} \cdots \tag{12}
S=S0+S1+S2+S3+S4⋯(12)其中
S
p
\mathcal{S}_p
Sp为
δ
T
1
,
δ
R
1
\boldsymbol{\delta} \boldsymbol{T}_{1},\boldsymbol{\delta} \boldsymbol{R}_{1}
δT1,δR1泰勒展开第
p
p
p项的函数:
S
0
=
ν
[
log
det
(
I
n
t
+
ρ
s
Q
T
s
t
1
)
+
log
det
(
I
n
r
+
ρ
s
R
s
r
1
)
−
n
t
r
1
t
1
]
=
ν
Γ
s
(13)
\begin{aligned} \mathcal{S}_{0}=& \nu\left[\log \operatorname{det}\left(\boldsymbol{I}_{n_{t}}+\sqrt{\rho_{s}} \boldsymbol{Q} \boldsymbol{T}^{s} t_{1}\right)\right.\\ &\left.+\log \operatorname{det}\left(\boldsymbol{I}_{n_{r}}+\sqrt{\rho_{s}} \boldsymbol{R}^{s} r_{1}\right)-n_{t} r_{1} t_{1}\right] \\ =& \nu \Gamma_{s} \end{aligned} \tag{13}
S0==ν[logdet(Int+ρsQTst1)+logdet(Inr+ρsRsr1)−ntr1t1]νΓs(13)
S
1
=
(
M
r
,
1
−
n
t
t
1
)
Tr
{
δ
R
1
}
+
(
M
t
,
1
−
n
t
r
1
)
Tr
{
δ
T
1
}
(14)
\mathcal{S}_{1}=\left(M_{r, 1}-\sqrt{n_{t}} t_{1}\right) \operatorname{Tr}\left\{\boldsymbol{\delta} \boldsymbol{R}_{1}\right\}+\left(M_{t, 1}-\sqrt{n_{t}} r_{1}\right) \operatorname{Tr}\left\{\boldsymbol{\delta} \boldsymbol{T}_{1}\right\} \tag{14}
S1=(Mr,1−ntt1)Tr{δR1}+(Mt,1−ntr1)Tr{δT1}(14)
S
2
=
−
1
2
Tr
{
M
r
,
2
(
δ
R
1
)
2
+
M
t
,
2
(
δ
T
1
)
2
+
2
δ
R
1
δ
T
1
}
=
1
2
Tr
{
[
δ
R
1
δ
T
1
]
T
V
[
δ
R
1
δ
T
1
]
}
(15)
\begin{aligned} \mathcal{S}_{2} &=-\frac{1}{2} \operatorname{Tr}\left\{M_{r, 2}\left(\boldsymbol{\delta} \boldsymbol{R}_{1}\right)^{2}+M_{t, 2}\left(\boldsymbol{\delta} \boldsymbol{T}_{1}\right)^{2}+2 \delta \boldsymbol{R}_{1} \boldsymbol{\delta} \boldsymbol{T}_{1}\right\} \\ &=\frac{1}{2} \operatorname{Tr}\left\{\left[\begin{array}{l} \boldsymbol{\delta} \boldsymbol{R}_{1} \\ \boldsymbol{\delta} \boldsymbol{T}_{1} \end{array}\right]^{T} \boldsymbol{V}\left[\begin{array}{l} \boldsymbol{\delta} \boldsymbol{R}_{1} \\ \boldsymbol{\delta} \boldsymbol{T}_{1} \end{array}\right]\right\} \end{aligned} \tag{15}
S2=−21Tr{Mr,2(δR1)2+Mt,2(δT1)2+2δR1δT1}=21Tr{[δR1δT1]TV[δR1δT1]}(15)其中Hessian矩阵
V
\boldsymbol{V}
V为:
V
=
[
−
M
r
,
2
−
1
−
1
−
M
t
,
2
]
(16)
\boldsymbol{V}=\left[\begin{array}{cc} -M_{r, 2} & -1 \\ -1 & -M_{t, 2} \end{array}\right] \tag{16}
V=[−Mr,2−1−1−Mt,2](16)当
p
>
2
p >2
p>2,则
S
p
\mathcal{S}_p
Sp可简化为:
S
p
=
(
−
1
)
p
p
Tr
{
M
r
,
p
(
δ
R
1
)
p
+
M
t
,
p
(
δ
T
1
)
p
}
(17)
\mathcal{S}_{p}=\frac{(-1)^{p}}{p} \operatorname{Tr}\left\{M_{r, p}\left(\boldsymbol{\delta} \boldsymbol{R}_{1}\right)^{p}+M_{t, p}\left(\boldsymbol{\delta} \boldsymbol{T}_{1}\right)^{p}\right\} \tag{17}
Sp=p(−1)pTr{Mr,p(δR1)p+Mt,p(δT1)p}(17)其中
M
t
,
p
=
(
ρ
s
n
t
)
p
/
2
Tr
{
[
Q
T
s
(
I
n
t
+
ρ
s
t
1
Q
T
s
)
−
1
]
p
}
M
r
,
p
=
(
ρ
s
n
t
)
p
/
2
Tr
{
[
R
s
(
I
n
r
+
ρ
s
r
1
R
s
)
−
1
]
p
}
(18)
\begin{array}{l} M_{t, p}=\left(\frac{\rho_{s}}{n_{t}}\right)^{p / 2} \operatorname{Tr}\left\{\left[\boldsymbol{Q} \boldsymbol{T}^{s}\left(\boldsymbol{I}_{n_{t}}+\sqrt{\rho_{s}} t_{1} \boldsymbol{Q} \boldsymbol{T}^{s}\right)^{-1}\right]^{p}\right\} \\ M_{r, p}=\left(\frac{\rho_{s}}{n_{t}}\right)^{p / 2} \operatorname{Tr}\left\{\left[\boldsymbol{R}^{s}\left(\boldsymbol{I}_{n_{r}}+\sqrt{\rho_{s}} r_{1} \boldsymbol{R}^{s}\right)^{-1}\right]^{p}\right\} \end{array} \tag{18}
Mt,p=(ntρs)p/2Tr{[QTs(Int+ρst1QTs)−1]p}Mr,p=(ntρs)p/2Tr{[Rs(Inr+ρsr1Rs)−1]p}(18)可以看到
(
8
)
(8)
(8)式中两个
Q
1
/
2
\boldsymbol{Q}^{1/2}
Q1/2可以合为
Q
\boldsymbol{Q}
Q。则
(
7
)
(7)
(7)式的鞍点解,即
t
1
,
r
1
t_1,r_1
t1,r1可由“针对变换的
T
1
,
R
1
\boldsymbol{T}_1,\boldsymbol{R}_1
T1,R1,
S
\mathcal{S}
S不变”得到,则
S
1
=
0
\mathcal{S}_1=0
S1=0,这类似于为了求最大值或最小值将一阶导设为0,则有:
r
1
=
1
n
t
M
t
,
1
=
ρ
s
n
t
Tr
{
Q
T
s
[
I
n
t
+
ρ
s
t
1
Q
T
s
]
−
1
}
(19)
r_{1} =\frac{1}{n_{t}} M_{t, 1}=\frac{\sqrt{\rho_{s}}}{n_{t}} \operatorname{Tr}\left\{\boldsymbol{Q} \boldsymbol{T}^{s}\left[\boldsymbol{I}_{n_{t}}+\sqrt{\rho_{s}} t_{1} \boldsymbol{Q} \boldsymbol{T}^{s}\right]^{-1}\right\} \tag{19}
r1=nt1Mt,1=ntρsTr{QTs[Int+ρst1QTs]−1}(19)
t
1
=
1
n
t
M
r
,
1
=
ρ
s
n
t
Tr
{
R
s
[
I
n
r
+
ρ
s
R
s
r
1
]
−
1
}
(20)
t_{1} =\frac{1}{n_{t}} M_{r, 1}=\frac{\sqrt{\rho_{s}}}{n_{t}} \operatorname{Tr}\left\{\boldsymbol{R}^{s}\left[\boldsymbol{I}_{n_{r}}+\sqrt{\rho_{s}} \boldsymbol{R}^{s} r_{1}\right]^{-1}\right\} \tag{20}
t1=nt1Mr,1=ntρsTr{Rs[Inr+ρsRsr1]−1}(20)
(
17
)
(
18
)
(17)(18)
(17)(18)的解在正实数
t
1
,
r
1
t_1,r_1
t1,r1的基础上最大化
Γ
s
\Gamma_{s}
Γs。
控制该逼近过程的是小参数
n
1
/
2
n^{1/2}
n1/2,所以鞍点逼近在大规模MIMO中更加精确。
平均互信息
由 g ( ν ) g(\nu) g(ν)的前几项可得: g ( ν ) ≈ exp ( − S 0 ) = exp ( − ν Γ s ) g(\nu) \approx \exp \left(-\mathcal{S}_{0}\right)=\exp \left(-\nu \Gamma_{s}\right) g(ν)≈exp(−S0)=exp(−νΓs)我们可以看到 ⟨ I ⟩ \langle I\rangle ⟨I⟩的主成分项为 Γ s \Gamma_{s} Γs,且有 ⟨ I ⟩ = O ( n ) \langle I\rangle=\mathcal{O}(n) ⟨I⟩=O(n)。
互信息的方差
为了得到
(
10
)
(10)
(10)式中
l
o
g
g
(
ν
)
\mathrm{log}g(\nu)
logg(ν)的
O
(
v
2
)
\mathcal{O}(v^2)
O(v2)项,我们需将下一个非零项包括进来,即
S
2
\mathcal{S}_2
S2。因此我们仅忽略更高项
S
p
,
p
>
2
\mathcal{S}_p,p>2
Sp,p>2。
基于
(
11
)
(11)
(11)式的均值方差表述,
g
(
ν
)
g(\nu)
g(ν)可以被表述为:
g
(
ν
)
=
e
−
S
0
∫
d
μ
(
δ
T
1
,
δ
R
1
)
e
−
S
2
=
e
−
S
0
∫
d
μ
(
δ
T
1
,
δ
R
1
)
×
exp
(
−
1
2
∑
a
,
b
=
1
ν
[
δ
T
1
,
a
b
δ
R
1
,
a
b
]
V
[
δ
T
1
,
b
a
δ
R
1
,
b
a
]
T
)
(21)
\begin{aligned} g(\nu)=& e^{-\mathcal{S}_{0}} \int d \mu\left(\boldsymbol{\delta} \boldsymbol{T}_{1}, \boldsymbol{\delta} \boldsymbol{R}_{1}\right) e^{-\mathcal{S}_{2}} \\ =& e^{-\mathcal{S}_{0}} \int d \mu\left(\boldsymbol{\delta} \boldsymbol{T}_{1}, \boldsymbol{\delta} \boldsymbol{R}_{1}\right) \\ & \times \exp \left(-\frac{1}{2} \sum_{a, b=1}^{\nu}\left[\delta T_{1, a b} \delta R_{1, a b}\right] \boldsymbol{V}\left[\delta T_{1, b a} \delta R_{1, b a}\right]^{T}\right) \end{aligned} \tag{21}
g(ν)==e−S0∫dμ(δT1,δR1)e−S2e−S0∫dμ(δT1,δR1)×exp⎝⎛−21a,b=1∑ν[δT1,abδR1,ab]V[δT1,baδR1,ba]T⎠⎞(21)为了将
(
21
)
(21)
(21)中方程的指数项对角化,利用酉变换将
δ
T
1
,
δ
R
1
\boldsymbol{\delta} \boldsymbol{T}_{1},\boldsymbol{\delta} \boldsymbol{R}_{1}
δT1,δR1改为
ν
×
ν
\nu \times \nu
ν×ν维矩阵
W
1
,
W
2
\boldsymbol{W}_1,\boldsymbol{W}_2
W1,W2,如下:
[
W
1
,
a
b
W
2
,
a
b
]
T
=
U
[
δ
T
1
,
a
b
δ
R
1
,
a
b
]
T
\left[W_{1, a b} W_{2, a b}\right]^{T}=\boldsymbol{U}\left[\delta T_{1, a b} \delta R_{1, a b}\right]^{T}
[W1,abW2,ab]T=U[δT1,abδR1,ab]T
U
\boldsymbol{U}
U是一个满足
U
V
U
T
=
diag
(
v
)
\boldsymbol{U} \boldsymbol{V} \boldsymbol{U}^{\boldsymbol{T}}=\operatorname{diag}(\boldsymbol{v})
UVUT=diag(v)的正交阵,其中
v
=
[
v
1
v
2
]
T
\boldsymbol{v}=[v_1 \space v_2]^T
v=[v1 v2]T是
V
\boldsymbol{V}
V的特征值,则
(
21
)
(21)
(21)式可以写为
g
(
ν
)
=
e
−
S
0
∫
d
μ
(
W
1
,
W
2
)
×
exp
(
−
1
2
∑
a
,
b
=
1
ν
[
v
1
W
1
,
a
b
W
1
,
b
a
+
v
2
W
2
,
a
b
W
2
,
b
a
]
)
(22)
\begin{array}{l} g(\nu)=e^{-\mathcal{S}_{0}} \int d \mu\left(\boldsymbol{W}_{1}, \boldsymbol{W}_{2}\right) \\ \times \exp \left(-\frac{1}{2} \sum_{a, b=1}^{\nu}\left[v_{1} W_{1, a b} W_{1, b a}+v_{2} W_{2, a b} W_{2, b a}\right]\right) \end{array} \tag{22}
g(ν)=e−S0∫dμ(W1,W2)×exp(−21∑a,b=1ν[v1W1,abW1,ba+v2W2,abW2,ba])(22)由于
M
t
,
2
,
M
r
,
2
<
1
M_{t, 2}, M_{r, 2}<1
Mt,2,Mr,2<1,
v
1
,
v
2
v_1,v_2
v1,v2反号,因此为了保证
M
1
,
a
b
M_{1,ab}
M1,ab按照最速下降路径逼近,我们令
W
1
=
W
1
†
\boldsymbol{W}_{1}=\boldsymbol{W}_{1}^{\dagger}
W1=W1†,同时
W
2
=
i
W
2
†
\boldsymbol{W}_{2}=i \boldsymbol{W}_{2}^{\dagger}
W2=iW2†,则
(
22
)
(22)
(22)式进一步转化为
g
(
ν
)
=
e
−
S
0
∏
k
=
1
2
{
∏
a
=
1
ν
∫
d
W
k
,
a
a
2
π
∏
b
>
a
ν
∬
d
Re
W
k
,
a
b
d
Im
W
k
,
a
b
2
π
×
exp
[
−
1
2
∣
v
k
∣
Tr
{
W
k
†
W
k
}
]
}
=
e
−
S
0
∣
v
1
v
2
∣
−
ν
2
2
=
e
−
S
0
∣
det
V
∣
−
ν
2
2
(23)
\begin{aligned} g(\nu)=& e^{-\mathcal{S}_{0}} \prod_{k=1}^{2}\left\{\prod_{a=1}^{\nu} \int \frac{d W_{k, a a}}{\sqrt{2 \pi}} \prod_{b>a}^{\nu} \iint \frac{d \operatorname{Re} W_{k, a b} d \operatorname{Im} W_{k, a b}}{2 \pi}\right.\\ &\left.\times \exp \left[-\frac{1}{2}\left|v_{k}\right| \operatorname{Tr}\left\{\boldsymbol{W}_{k}^{\dagger} \boldsymbol{W}_{k}\right\}\right]\right\} \\ =& e^{-\mathcal{S}_{0}}\left|v_{1} v_{2}\right|^{-\frac{\nu^{2}}{2}}=e^{-\mathcal{S}_{0}}|\operatorname{det} \boldsymbol{V}|^{-\frac{\nu^{2}}{2}} \end{aligned} \tag{23}
g(ν)==e−S0k=1∏2{a=1∏ν∫2πdWk,aab>a∏ν∬2πdReWk,abdImWk,ab×exp[−21∣vk∣Tr{Wk†Wk}]}e−S0∣v1v2∣−2ν2=e−S0∣detV∣−2ν2(23)将
(
3
)
(3)
(3)式与
(
23
)
(23)
(23)式逐项比较,对应上
g
(
ν
)
g(\nu)
g(ν)的
ν
\nu
ν阶泰勒展开式,则方差中的主成分为:
⟨
I
2
⟩
−
⟨
I
⟩
2
=
C
2
=
−
log
∣
det
V
∣
+
⋯
=
−
log
(
1
−
M
r
,
2
M
t
,
2
)
+
⋯
(24)
\begin{aligned} \left\langle I^{2}\right\rangle-\langle I\rangle^{2} &=\mathcal{C}_{2}=-\log |\operatorname{det} \boldsymbol{V}|+\cdots \\ &=-\log \left(1-M_{r, 2} M_{t, 2}\right)+\cdots \end{aligned} \tag{24}
⟨I2⟩−⟨I⟩2=C2=−log∣detV∣+⋯=−log(1−Mr,2Mt,2)+⋯(24)因为
M
r
,
2
,
M
t
,
2
=
O
(
1
)
M_{r, 2}, M_{t, 2}=\mathcal{O}(1)
Mr,2,Mt,2=O(1),则方差在
n
−
1
/
2
n^{-1/2}
n−1/2的扩展上也是
O
(
1
)
\mathcal{O}(1)
O(1)。同时我们发现在
S
2
\mathcal{S}_2
S2中没有哪一项是
ν
\nu
ν的函数。因此没有
O
(
1
)
\mathcal{O}(1)
O(1)中的哪一项在
⟨
I
⟩
\left\langle I \right\rangle
⟨I⟩中出现,因此有
⟨
I
⟩
=
Γ
s
+
o
(
1
)
\langle I\rangle=\Gamma_{s}+o(1)
⟨I⟩=Γs+o(1)。