Optimizing the Matrices V {V} V and V E {V}_{E} VE
In this subsection, the TPC matrix
V
{\bf{V}}
V and matrix
V
E
{\bf{V}}_{E}
VE are optimized by fixing
Φ
\mathbf{\Phi }
Φ. Specifically, the unit modulus constraint on the phase shifts
Φ
\mathbf{\Phi }
Φ is removed, and the updated optimization problem reduced from Problem is given by
min
V
,
V
E
−
Tr
(
W
I
V
H
H
^
I
H
U
I
)
−
Tr
(
W
I
U
I
H
H
^
I
V
)
+
Tr
(
V
H
H
V
V
)
−
Tr
(
W
E
V
E
H
H
^
E
H
U
E
)
−
Tr
(
W
E
U
E
H
H
^
E
V
E
)
+
Tr
(
V
E
H
H
V
E
V
E
)
s.t.
T
r
(
V
V
H
+
V
E
V
E
H
)
≤
P
T
.
\begin{align*} & \ \ \underset{{\bf{V}} ,{{\bf{V}}_E}}{\mathop\text{min}} \ \ -\text{Tr}({{\mathbf{W}}_{I}}{{\mathbf{V}}^{H}}{{{\mathbf{\hat{H}}}}_{I}}^{H}{{\mathbf{U}}_{I}})-\text{Tr}({{\mathbf{W}}_{I}}{{\mathbf{U}}^{H}_{I}}{{{\mathbf{\hat{H}}}}_{I}}\mathbf{V})+\text{Tr}({{\mathbf{V}}^{H}}{{\mathbf{H}}_{V}}\mathbf{V}) \nonumber \\ &\quad \quad \quad \quad -\text{Tr}({{\mathbf{W}}_{E}}{{\mathbf{V}}^{H}_{E}}{{{\mathbf{\hat{H}}}}_{E}}^{H}{{\mathbf{U}}_{E}})-\text{Tr}({{\mathbf{W}}_{E}}{{\mathbf{U}}^{H}_{E}}{{{\mathbf{\hat{H}}}}_{E}}{{\mathbf{V}}_{E}})+\text{Tr}({{\mathbf{V}}^{H}_{E}}{{\mathbf{H}}_{VE}}{{\mathbf{V}}_{E}}) \\ & \ \ \text{s.t.} \quad {\rm{Tr(}}{\bf{V}}{{\bf{V}}^H}{\rm{ + }}{{\bf{V}}_{E}}{{\bf{V}}^{H}_{E}}{\rm{)}} \le P_{T}. \end{align*}
V,VEmin −Tr(WIVHH^IHUI)−Tr(WIUIHH^IV)+Tr(VHHVV)−Tr(WEVEHH^EHUE)−Tr(WEUEHH^EVE)+Tr(VEHHVEVE) s.t.Tr(VVH+VEVEH)≤PT.
The above problem is a convex QCQP problem, and the standard optimization packages, such as CVX can be exploited to solve it. However, the calculation burden is heavy. To reduce the complexity, the near-optimal closed form expressions of the TPC matrix and AN covariance matrix are provided by applying the Lagrangian multiplier method.
Since Problem is a convex problem, the Slater’s condition is satisfied, where the duality gap between Problem and its dual problem is zero. Thus, Problem can be solved by addressing its dual problem if the dual problem is easier. For this purpose, by introducing Lagrange multiplier
λ
\lambda
λ to combine the the constraint and OF of Problem, the Lagrangian function of Problem is obtained as
L
(
V
,
V
E
,
λ
)
≜
−
T
r
(
W
I
V
H
H
^
I
H
U
I
)
−
T
r
(
W
I
U
I
H
H
^
I
V
)
+
T
r
(
V
H
H
V
V
)
−
T
r
(
W
E
V
E
H
H
^
E
H
U
E
)
−
T
r
(
W
E
U
E
H
H
^
E
V
E
)
+
T
r
(
V
E
H
H
V
E
V
E
)
+
λ
[
T
r
(
V
V
H
+
V
E
V
E
H
)
−
P
T
]
=
−
T
r
(
W
I
V
H
H
^
I
H
U
I
)
−
T
r
(
W
I
U
I
H
H
^
I
V
)
+
T
r
[
V
H
(
H
V
+
λ
I
)
V
]
−
T
r
(
W
E
V
E
H
H
^
E
H
U
E
)
−
T
r
(
W
E
U
E
H
H
^
E
V
E
)
+
T
r
[
V
E
H
(
H
V
E
+
λ
I
)
V
E
]
−
λ
P
T
.
\begin{align} \mathcal{L}\left( {{\bf{V}},{{\bf{V}}_E},\lambda } \right) &\!\triangleq \!- {\rm{Tr}}\left( {{\bf{W}}_I}{{\bf{V}}^H}{\bf{\hat H}}_I^H{{\bf{U}}_I} \right)\! - \!{\rm{Tr}}\left( {{{\bf{W}}_I}{\bf{U}}_I^H{{{\bf{\hat H}}}_I}{\bf{V}}} \right)\! +\! {\rm{Tr}}\left( {{{\bf{V}}^H}{{\bf{H}}_V}{\bf{V}}} \right) \!-\! {\rm{Tr}}\left( {{{\bf{W}}_E}{\bf{V}}_E^H{\bf{\hat H}}_E^H{{\bf{U}}_E}} \right)\nonumber \\ & \quad- {\rm{Tr}}\left( {{{\bf{W}}_E}{\bf{U}}_E^H{{{\bf{\hat H}}}_E}{{\bf{V}}_E}} \right) + {\rm{Tr}}\left( {{\bf{V}}_E^H{{\bf{H}}_{VE}}{{\bf{V}}_E}} \right)+ \lambda[ {{\rm{Tr}}\left( {{\bf{V}}{{\bf{V}}^H} + {{\bf{V}}_E}{\bf{V}}_E^H} \right)} -P_{T}] \nonumber \\ &= - {\rm{Tr}}\left( {{{\bf{W}}_I}{{\bf{V}}^H}{\bf{\hat H}}_I^H{{\bf{U}}_I}} \right) - {\rm{Tr}}\left( {{{\bf{W}}_I}{\bf{U}}_I^H{{{\bf{\hat H}}}_I}{\bf{V}}} \right) + {\rm{Tr}}\left[ {{{\bf{V}}^H}\left( {{{\bf{H}}_V} + \lambda{\bf{I}}} \right){\bf{V}}} \right] \nonumber \\ &\quad- {\rm{Tr}}\left( {{{\bf{W}}_E}{\bf{V}}_E^H{\bf{\hat H}}_E^H{{\bf{U}}_E}} \right)\! - \!{\rm{Tr}}\left( {{{\bf{W}}_E}{\bf{U}}_E^H{{{\bf{\hat H}}}_E}{{\bf{V}}_E}} \right)\! +\! {\rm{Tr}}\left[ {{\bf{V}}_E^H\left( {{{\bf{H}}_{VE}} \!+ \!\lambda {\bf{I}}} \right){{\bf{V}}_E}} \right] \!-\! \lambda {P_T}. \end{align}
L(V,VE,λ)≜−Tr(WIVHH^IHUI)−Tr(WIUIHH^IV)+Tr(VHHVV)−Tr(WEVEHH^EHUE)−Tr(WEUEHH^EVE)+Tr(VEHHVEVE)+λ[Tr(VVH+VEVEH)−PT]=−Tr(WIVHH^IHUI)−Tr(WIUIHH^IV)+Tr[VH(HV+λI)V]−Tr(WEVEHH^EHUE)−Tr(WEUEHH^EVE)+Tr[VEH(HVE+λI)VE]−λPT.
Then the dual problem of Problem is
max
λ
h
(
λ
)
s.t.
λ
≥
0
,
\begin{align} \mathop {\max }\limits_\lambda \quad \quad {\rm{ }}h\left( \lambda \right) \\ \text{s.t.}\quad \quad {\rm{ }}\lambda \ge 0, \end{align}
λmaxh(λ)s.t.λ≥0,
where
h
(
λ
)
h\left( \lambda \right)
h(λ) is the dual function given by
h
(
λ
)
≜
min
V
,
V
E
L
(
V
,
V
E
,
λ
)
.
\begin{align} h\left( \lambda \right) \triangleq \mathop {\min }\limits_{{\bf{V}},{{\bf{V}}_E}} {\rm{ }}\mathcal{L}\left( {{\bf{V}},{{\bf{V}}_E},\lambda } \right). \end{align}
h(λ)≜V,VEminL(V,VE,λ).
Note that Problem is a convex quadratic optimization problem with no constraint, which can be solved in closed form. The optimal solution
V
⋆
,
V
⋆
E
{\bf{V}^{\star}},{{\bf{V}^{{\star}}}_E}
V⋆,V⋆E for Proble is
[
V
⋆
,
V
⋆
E
]
=
arg
min
V
,
V
E
L
(
V
,
V
E
,
λ
)
.
\begin{align} [{\bf{V}^{\star}},{{\bf{V}^{{\star}}}_E}]=\text{arg}\mathop {\min }\limits_{{\bf{V}},{{\bf{V}}_E}} {\rm{ }}\mathcal{L}\left( {{\bf{V}},{{\bf{V}}_E},\lambda } \right). \end{align}
[V⋆,V⋆E]=argV,VEminL(V,VE,λ).
By setting the first-order derivative of
L
(
V
,
V
E
,
λ
)
\mathcal{L}\left( {{\bf{V}},{{\bf{V}}_E},\lambda } \right)
L(V,VE,λ) w.r.t.
V
{{{\bf{V}}}}
V to zero matrix, we can obtain the optimal solution of
V
{\bf{V}}
V as follows:
∂
L
(
V
,
V
E
,
λ
)
∂
V
=
0
,
∂
L
(
V
,
V
E
,
λ
)
∂
V
E
=
0.
\begin{align} \frac{\partial{\mathcal{L}\left( {{\bf{V}},{{\bf{V}}_E},\lambda } \right)}}{{\partial {\bf{V}}}}=\bf{0}, \\ \frac{\partial{\mathcal{L}\left( {{\bf{V}},{{\bf{V}}_E},\lambda } \right)}}{{\partial {{\bf{V}}_E}}}=\bf{0}. \end{align}
∂V∂L(V,VE,λ)=0,∂VE∂L(V,VE,λ)=0.
The left hand side of Equation can be expanded as
∂ L ( V , V E , λ ) ∂ V = ∂ T r [ V H ( H V + λ I ) V ] ∂ V − ( W I U I H H ^ I ) H − ( H ^ I H U I W I ) = 2 ( H V + λ I ) V − 2 ( H ^ I H U I W I ) . \begin{align} \frac{\partial{\mathcal{L}\left( {{\bf{V}},{{\bf{V}}_E},\lambda } \right)}}{{\partial {\bf{V}}}}&=\frac{{\partial {\rm{Tr}}\left[ {{{\bf{V}}^H}\left( {{{\bf{H}}_V} + \lambda {\bf{I}}} \right){\bf{V}}} \right]}}{{\partial {\bf{V}}}}-\left( {{{\bf{W}}_I}{\bf{U}}_I^H{{{\bf{\hat H}}}_I}} \right)^H-\left( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{{\bf{W}}_I}} \right) \nonumber \\ &=2\left( {{{\bf{H}}_V} + \lambda {\bf{I}}} \right){\bf{V}}-2\left( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{{\bf{W}}_I}} \right). \end{align} ∂V∂L(V,VE,λ)=∂V∂Tr[VH(HV+λI)V]−(WIUIHH^I)H−(H^IHUIWI)=2(HV+λI)V−2(H^IHUIWI).
这里理论上需要使用复数求导的公式,但是这里实际上将函数当作普通的实数变量来求导了,根据 matrix-cookbook 当中的这个公式 ∂ ∂ X Tr ( X T B X ) = B X + B T X \frac{\partial}{\partial \mathbf{X}} \operatorname{Tr}\left(\mathbf{X}^{T} \mathbf{B X}\right)=\mathbf{B} \mathbf{X}+\mathbf{B}^{T} \mathbf{X} ∂X∂Tr(XTBX)=BX+BTX,并且 H V + λ I \mathbf{H}_V+\lambda {\bf{I}} HV+λI 实际上是共轭对称的,因此可以写成 ∂ T r [ V H ( H V + λ I ) V ] ∂ V = 2 ( H V + λ I ) V . \frac{{\partial {\rm{Tr}}\left[ {{{\bf{V}}^H}\left( {{{\bf{H}}_V} + \lambda {\bf{I}}} \right){\bf{V}}} \right]}}{{\partial {\bf{V}}}}\nonumber =2\left( {{{\bf{H}}_V} + \lambda {\bf{I}}} \right){\bf{V}}. ∂V∂Tr[VH(HV+λI)V]=2(HV+λI)V.,我们可以将所有的共轭转置符号 H H H 当作 T T T 来理解。同理,因为 ∂ ∂ X Tr ( X A ) = A T , \frac{\partial}{\partial \mathbf{X}} \operatorname{Tr}(\mathbf{X A})=\mathbf{A}^{T}, ∂X∂Tr(XA)=AT,我们有 − ∂ ∂ X Tr ( W I U I H H ^ I V ) = − ( W I U I H H ^ I ) H . -\frac{\partial}{\partial \mathbf{X}} \operatorname{Tr}\left(\mathbf{W}_{I} \mathbf{U}_{I}^{H} \hat{\mathbf{H}}_{I} \mathbf{V}\right)=-\left( {{{\bf{W}}_I}{\bf{U}}_I^H{{{\bf{\hat H}}}_I}} \right)^H. −∂X∂Tr(WIUIHH^IV)=−(WIUIHH^I)H.目前这一系列的结论都是观察得到的,至于是不是确实这样计算的有待商榷。
The equation becomes
(
H
V
+
λ
I
)
V
=
(
H
^
I
H
U
I
W
I
)
.
\begin{align} \left( {{{\bf{H}}_V} + \lambda {\bf{I}}} \right){\bf{V}}=\left( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{{\bf{W}}_I}} \right). \end{align}
(HV+λI)V=(H^IHUIWI).
Then the optimal solution
V
⋆
{\bf{V}^{\star}}
V⋆ for Problem is
V
⋆
=
(
H
V
+
λ
I
)
†
(
H
^
I
H
U
I
W
I
)
≜
Θ
V
(
λ
)
(
H
^
I
H
U
I
W
I
)
.
\begin{align} {{\bf{V}}^{\star}}&=\left( {{{\bf{H}}_V} + \lambda {\bf{I}}} \right)^{ \dag }\left( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{{\bf{W}}_I}} \right) \nonumber \\ &\triangleq{{\bf{\Theta }}_V}\left( \lambda \right)\left( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{{\bf{W}}_I}} \right). \end{align}
V⋆=(HV+λI)†(H^IHUIWI)≜ΘV(λ)(H^IHUIWI).
Similarly, we solve Problem by setting the first-order derivative of
L
(
V
,
V
E
,
λ
)
\mathcal{L}\left( {{\bf{V}},{{\bf{V}}_E},\lambda } \right)
L(V,VE,λ) w.r.t.
V
E
{{{\bf{V}}_E}}
VE to zero matrix, which becomes
2
(
H
V
E
+
λ
I
)
V
E
−
2
H
^
E
H
U
E
W
E
H
=
0.
\begin{align} 2\left( {{{\bf{H}}_{VE}} + \lambda {\bf{I}}} \right){{\bf{V}}_E} - 2{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H = \bf{0}. \end{align}
2(HVE+λI)VE−2H^EHUEWEH=0.
Then the optimal solution
V
E
⋆
{\bf{V}}_E^{\star}
VE⋆ for Problem is
V
E
⋆
=
(
H
V
E
+
λ
I
)
†
H
^
E
H
U
E
W
E
H
≜
Θ
V
E
(
λ
)
H
^
E
H
U
E
W
E
H
.
\begin{align} {\bf{V}}_E^{\star} &= \left( {{{\bf{H}}_{VE}} + \lambda {\bf{I}}} \right)^{ \dag }{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H \nonumber \\ &\triangleq {{\bf{\Theta }}_{VE}}\left( \lambda \right){\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H. \end{align}
VE⋆=(HVE+λI)†H^EHUEWEH≜ΘVE(λ)H^EHUEWEH.
Once the optimal solution
λ
⋆
{\lambda}^{\star}
λ⋆ for Problem is found, the final optimal
V
⋆
,
V
E
⋆
{\bf{V}^{\star}},{\bf{V}}_E^{\star}
V⋆,VE⋆ can be obtained. The value of
λ
⋆
{\lambda}^{\star}
λ⋆ should be chosen in order to guarantee the complementary slackness condition as
λ
[
T
r
(
V
⋆
V
⋆
H
+
V
E
⋆
V
E
⋆
H
)
−
P
T
]
=
0.
\begin{align} \lambda[ {\rm{Tr(}}{\bf{V}^{\star}}{{\bf{V}}^{{\star}H}}{\rm{ + }}{{\bf{V}}_{E}^{{\star}}}{{\bf{V}}^{{\star}H}_{E}}{\rm{)}}-P_{T}]=0. \end{align}
λ[Tr(V⋆V⋆H+VE⋆VE⋆H)−PT]=0.
We define
P
(
λ
)
≜
T
r
(
V
⋆
V
⋆
H
+
V
E
⋆
V
E
⋆
H
)
=
T
r
(
V
⋆
V
⋆
H
)
+
T
r
(
V
E
⋆
V
E
⋆
H
)
,
\begin{align} P(\lambda)&\triangleq {\rm{Tr(}}{\bf{V}^{\star}}{{\bf{V}}^{{\star}H}}{\rm{ + }}{{\bf{V}}_{E}^{{\star}}}{{\bf{V}}^{{\star}H}_{E}}{\rm{)}} ={\rm{Tr(}}{\bf{V}^{\star}}{{\bf{V}}^{{\star}H}}{\rm{)}}+{\rm{Tr(}}{{\bf{V}}_{E}^{{\star}}}{{\bf{V}}^{{\star}H}_{E}}{\rm{)}}, \end{align}
P(λ)≜Tr(V⋆V⋆H+VE⋆VE⋆H)=Tr(V⋆V⋆H)+Tr(VE⋆VE⋆H),
where
T
r
(
V
⋆
V
⋆
H
)
=
T
r
(
Θ
V
(
λ
)
(
H
^
I
H
U
I
W
I
H
)
(
H
^
I
H
U
I
W
I
H
)
H
Θ
V
H
(
λ
)
)
=
T
r
(
Θ
V
H
(
λ
)
Θ
V
(
λ
)
(
H
^
I
H
U
I
W
I
H
)
(
H
^
I
H
U
I
W
I
H
)
H
)
,
\begin{align} {\rm{Tr}}\left( {{\bf{V}}^{\star}{{\bf{V}}^{{\star}H}}} \right)&={\rm{Tr}}\left( {{{\bf{\Theta }}_V}\left( \lambda \right)( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{\bf{W}}_I^H} )}( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{\bf{W}}_I^H} )^H{{\bf{\Theta }}^H_V}\left( \lambda \right) \right) \nonumber \\ &={\rm{Tr}}\left( {{{\bf{\Theta }}^H_V}\left( \lambda \right){{\bf{\Theta }}_V}\left( \lambda \right)( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{\bf{W}}_I^H} )} ( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{\bf{W}}_I^H} )^H \right), \end{align}
Tr(V⋆V⋆H)=Tr(ΘV(λ)(H^IHUIWIH)(H^IHUIWIH)HΘVH(λ))=Tr(ΘVH(λ)ΘV(λ)(H^IHUIWIH)(H^IHUIWIH)H),
T
r
(
V
E
⋆
H
V
E
⋆
)
=
T
r
(
Θ
V
E
(
λ
)
(
H
^
E
H
U
E
W
E
H
)
(
H
^
E
H
U
E
W
E
H
)
H
Θ
V
E
H
(
λ
)
)
=
T
r
(
Θ
V
E
H
(
λ
)
Θ
V
E
(
λ
)
(
H
^
E
H
U
E
W
E
H
)
(
H
^
E
H
U
E
W
E
H
)
H
)
.
\begin{align} {\rm{Tr}}\left( {{{\bf{V}}_E^{{\star}H}}{\bf{V}}_E^{\star}} \right)&={\rm{Tr}}\left( {{{\bf{\Theta }}_{VE}}\left( \lambda \right)( {{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H} )}( {{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H} )^H{{\bf{\Theta }}^H_{VE}}\left( \lambda \right) \right) \nonumber \\ &={\rm{Tr}}\left( {{{\bf{\Theta }}^H_{VE}}\left( \lambda \right){{\bf{\Theta }}_{VE}}\left( \lambda \right)( {{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H} )} ( {{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H} )^H \right). \end{align}
Tr(VE⋆HVE⋆)=Tr(ΘVE(λ)(H^EHUEWEH)(H^EHUEWEH)HΘVEH(λ))=Tr(ΘVEH(λ)ΘVE(λ)(H^EHUEWEH)(H^EHUEWEH)H).
Then
P
(
λ
)
P(\lambda)
P(λ) becomes
P
(
λ
)
=
T
r
(
Θ
V
n
(
H
^
I
H
U
I
W
I
H
)
(
H
^
I
H
U
I
W
I
H
)
H
)
+
T
r
(
Θ
V
E
n
(
H
^
E
H
U
E
W
E
H
)
(
H
^
E
H
U
E
W
E
H
)
H
)
,
\begin{align} P(\lambda)={\rm{Tr}}\left( {{{\bf{\Theta }}^n_V}( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{\bf{W}}_I^H} ) ( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{\bf{W}}_I^H} )^H }\right) +{\rm{Tr}}\left( {{{\bf{\Theta }}^n_{VE}}( {{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H} ) ( {{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H} )^H }\right), \end{align}
P(λ)=Tr(ΘVn(H^IHUIWIH)(H^IHUIWIH)H)+Tr(ΘVEn(H^EHUEWEH)(H^EHUEWEH)H),
where
Θ
V
n
=
Θ
V
H
(
λ
)
Θ
V
(
λ
)
=
(
H
V
+
λ
I
)
†
H
(
H
V
+
λ
I
)
†
,
\begin{align} {{\bf{\Theta }}^n_V}&={{\bf{\Theta }}^H_V}\left( \lambda \right){{\bf{\Theta }}_V}\left( \lambda \right)=\left( {{{\bf{H}}_V} + \lambda {\bf{I}}} \right)^{ \dag H }\left( {{{\bf{H}}_V} + \lambda {\bf{I}}} \right)^{ \dag}, \end{align}
ΘVn=ΘVH(λ)ΘV(λ)=(HV+λI)†H(HV+λI)†,
Θ
V
E
n
=
Θ
V
E
H
(
λ
)
Θ
V
E
(
λ
)
=
(
H
V
E
+
λ
I
)
†
H
(
H
V
E
+
λ
I
)
†
.
\begin{align} {{\bf{\Theta }}^n_{VE}}&={{\bf{\Theta }}^H_{VE}}\left( \lambda \right){{\bf{\Theta }}_{VE}}\left( \lambda \right) =\left( {{{\bf{H}}_{VE}} + \lambda {\bf{I}}} \right)^{ \dag H}\left( {{{\bf{H}}_{VE}} + \lambda {\bf{I}}} \right)^{ \dag }. \end{align}
ΘVEn=ΘVEH(λ)ΘVE(λ)=(HVE+λI)†H(HVE+λI)†.
To find the optimal
λ
⋆
≥
0
{\lambda^{\star}}\geq0
λ⋆≥0, we first check whether
λ
=
0
\lambda=0
λ=0 is the optimal solution or not. If
P
(
0
)
=
T
r
(
V
⋆
H
(
0
)
V
⋆
(
0
)
)
+
T
r
(
V
E
⋆
H
(
0
)
V
E
⋆
(
0
)
)
≤
P
T
,
\begin{align} P(0)= {\rm{Tr}}\left( {{{\bf{V}}^{{\star}H}}(0){\bf{V}}^{\star}(0)} \right) + {\rm{Tr}}\left( {{\bf{V}}_E^{{\star}H}(0){{\bf{V}}_E}^{\star}(0)} \right)\le {P_{T}}, \end{align}
P(0)=Tr(V⋆H(0)V⋆(0))+Tr(VE⋆H(0)VE⋆(0))≤PT,
then the optimal solutions are given by
V
⋆
=
V
(
0
)
{{\bf{V}}^{\star}}={{\bf{V}}}(0)
V⋆=V(0) and
V
E
⋆
=
V
E
(
0
)
{{\bf{V}}_{E}^{\star}}={{\bf{V}}_{E}}(0)
VE⋆=VE(0). Otherwise, the optimal
λ
⋆
>
0
\lambda^{\star}>0
λ⋆>0 is the solution of the equation
P
(
λ
)
=
0
P(\lambda)=0
P(λ)=0.
It is ready to verify that
H
V
{{\bf{H}}_V}
HV and
H
V
E
{{\bf{H}}_{VE}}
HVE is a positive semidefinite matrix. Let us define the rank of
H
V
{{\bf{H}}_V}
HV and
H
V
E
{{\bf{H}}_{VE}}
HVE as
r
V
=
r
a
n
k
(
H
V
)
≤
N
T
r_{V}={\rm{rank}}({\bf{H}}_{V})\le N_T
rV=rank(HV)≤NT and
r
V
E
=
r
a
n
k
(
H
V
E
)
≤
N
T
r_{VE}={\rm{rank}}({\bf{H}}_{VE}) \le N_T
rVE=rank(HVE)≤NT respectively. By decomposing
H
V
{{\bf{H}}_V}
HV and
H
V
E
{{\bf{H}}_{VE}}
HVE by using the singular value decomposition (SVD), we have
H
V
=
[
P
V
,
1
,
P
V
,
2
]
Σ
V
[
P
V
,
1
,
P
V
,
2
]
H
,
H
V
E
=
[
P
V
E
,
1
,
P
V
E
,
2
]
Σ
V
E
[
P
V
E
,
1
,
P
V
E
,
2
]
H
,
{{\bf{H}}_V} = \left[ {{{\bf{P}}_{V,1}},{{\bf{P}}_{V,2}}} \right]{{\bf{\Sigma }}_V}{\left[ {{{\bf{P}}_{V,1}},{{\bf{P}}_{V,2}}} \right]^{\rm{H}}},{{\bf{H}}_{VE}} = \left[ {{{\bf{P}}_{{VE},1}},{{\bf{P}}_{{VE},2}}} \right]{{\bf{\Sigma }}_{VE}}{\left[ {{{\bf{P}}_{{VE},1}},{{\bf{P}}_{{VE},2}}} \right]^{\rm{H}}},
HV=[PV,1,PV,2]ΣV[PV,1,PV,2]H,HVE=[PVE,1,PVE,2]ΣVE[PVE,1,PVE,2]H, where
P
V
,
1
{\bf{P}}_{V,1}
PV,1 comprises the first
r
V
r_V
rV singular vectors associated with the
r
V
r_V
rV positive eigenvalues of
H
V
{{\bf{H}}_V}
HV, and
P
V
,
2
{\bf{P}}_{V,2}
PV,2 includes the last
N
T
−
r
V
N_T-r_V
NT−rV singular vectors associated with the
N
T
−
r
V
N_T-r_V
NT−rV zero-valued eigenvalues of
H
V
{{\bf{H}}_V}
HV,
Σ
V
=
d
i
a
g
{
Σ
V
,
1
,
0
(
N
T
−
r
V
)
×
(
N
T
−
r
V
)
}
{{\bm{\Sigma}} _V} = {\rm{diag}}\left\{ {{{\bm{\Sigma}} _{V,1}},{{\bf{0}}_{\left( {{N_T} - {r_V}} \right) \times \left( {{N_T} - {r_V}} \right)}}} \right\}
ΣV=diag{ΣV,1,0(NT−rV)×(NT−rV)} with
Σ
V
,
1
{\bm{\Sigma}} _{V,1}
ΣV,1 representing the diagonal submatrix collecting the first
r
V
r_V
rV positive eigenvalues. Similarly, the first
r
V
E
r_{VE}
rVE singular vectors corresponding to the
r
V
E
r_{VE}
rVE positive eigenvalues of
H
V
E
{{\bf{H}}_{VE}}
HVE are contained in
P
V
E
,
1
{\bf{P}}_{VE,1}
PVE,1, while the last
N
T
−
r
V
E
N_T-r_{VE}
NT−rVE singular vectors corresponding to the
N
T
−
r
V
E
N_T-r_{VE}
NT−rVE zero-valued eigenvalues of
H
V
E
{{\bf{H}}_{VE}}
HVE are held in
P
V
E
,
2
{\bf{P}}_{VE,2}
PVE,2.
Σ
V
E
=
d
i
a
g
{
Σ
V
E
,
1
,
0
(
N
T
−
r
V
E
)
×
(
N
T
−
r
V
E
)
}
{{\bm{\Sigma}} _{VE}} = {\rm{diag}}\left\{ {{{\bm{\Sigma}} _{{VE},1}},{{\bf{0}}_{\left( {{N_T} - {r_{VE}}} \right) \times \left( {{N_T} - {r_{VE}}} \right)}}} \right\}
ΣVE=diag{ΣVE,1,0(NT−rVE)×(NT−rVE)} is a diagonal matrix with
Σ
V
E
,
1
{\bm{\Sigma}} _{{VE},1}
ΣVE,1 representing the diagonal submatrix gathering the first
r
V
E
r_{VE}
rVE positive eigenvalues. By defining
P
V
≜
[
P
V
,
1
,
P
V
,
2
]
{{\bf{P}}_V} \triangleq \left[ {{{\bf{P}}_{V,1}},{{\bf{P}}_{V,2}}} \right]
PV≜[PV,1,PV,2] and
P
V
E
≜
[
P
V
E
,
1
,
P
V
E
,
2
]
{{\bf{P}}_{VE}}\triangleq \left[ {{{\bf{P}}_{{VE},1}},{{\bf{P}}_{{VE},2}}} \right]
PVE≜[PVE,1,PVE,2], and substituting \eqref{svdforHVHVE} into and ,
P
(
λ
)
P(\lambda)
P(λ) becomes
P
(
λ
)
=
T
r
(
[
(
P
V
Σ
V
P
V
H
+
λ
P
V
P
V
H
)
−
1
(
P
V
Σ
V
P
V
H
+
λ
P
V
P
V
H
)
−
1
]
(
H
^
I
H
U
I
W
I
H
)
(
H
^
I
H
U
I
W
I
H
)
H
)
+
T
r
(
[
(
P
V
E
Σ
V
E
P
V
E
H
+
λ
P
V
E
P
V
E
H
)
−
1
(
P
V
E
Σ
V
E
P
V
E
H
+
λ
P
V
E
P
V
E
H
)
−
1
]
(
H
^
E
H
U
E
W
E
H
)
(
H
^
E
H
U
E
W
E
H
)
H
)
=
T
r
(
[
(
Σ
V
+
λ
I
)
−
2
]
Z
V
)
+
T
r
(
[
(
Σ
V
E
+
λ
I
)
−
2
]
Z
V
E
)
=
∑
i
=
1
r
V
[
[
Z
V
]
i
,
i
(
[
Σ
V
]
i
,
i
+
λ
)
2
]
+
∑
i
=
1
r
V
E
[
[
Z
V
E
]
i
,
i
(
[
Σ
V
E
]
i
,
i
+
λ
)
2
]
+
∑
i
=
r
V
+
1
N
T
[
[
Z
V
]
i
,
i
(
λ
)
2
]
+
∑
i
=
r
V
E
+
1
N
T
[
[
Z
V
E
]
i
,
i
(
λ
)
2
]
,
\begin{align} & P(\lambda) ={\rm {Tr}}\left({[{\left({{{\bf {P}}_{V}}{{\bf {\Sigma}}_{V}}{\bf {P}}_{V}^{H}+\lambda{{\bf {P}}_{V}}{\bf {P}}_{V}^{H}}\right)^{-1}}{\left({{{\bf {P}}_{V}}{{\bf {\Sigma}}_{V}}{\bf {P}}_{V}^{H}+\lambda{{\bf {P}}_{V}}{\bf {P}}_{V}^{H}}\right)^{-1}}]({{\bf {\hat{H}}}_{I}^{H}{{\bf {U}}_{I}}{\bf {W}}_{I}^{H}})({{\bf {\hat{H}}}_{I}^{H}{{\bf {U}}_{I}}{\bf {W}}_{I}^{H}})^{H}}\right)\nonumber \\ & \!\!+\!\!{\rm {Tr}}\!\left(\!{[\!{\left(\!{{{\bf {P}}_{VE}}{{\bf {\Sigma}}_{VE}}{\bf {P}}_{VE}^{H}\!+\!\lambda{{\bf {P}}_{VE}}{\bf {P}}_{VE}^{H}}\!\right)^{-1}}\!\!{\left({{{\bf {P}}_{VE}}{{\bf {\Sigma}}_{VE}}{\bf {P}}_{VE}^{H}\!+\!\lambda{{\bf {P}}_{VE}}{\bf {P}}_{VE}^{H}}\right)^{-1}}\!]\!(\!{{\bf {\hat{H}}}_{E}^{H}{{\bf {U}}_{E}}{\bf {W}}_{E}^{H}}\!)\!({{\bf {\hat{H}}}_{E}^{H}{{\bf {U}}_{E}}{\bf {W}}_{E}^{H}})^{H}}\!\right)\nonumber \\ & ={\rm {Tr}}\left({[{\left({{{\bf {\Sigma}}_{V}}+\lambda{\bf {I}}}\right)^{-2}}]{\bf {Z}}_{V}}\right)+{\rm {Tr}}\left({[{\left({{{\bf {\Sigma}}_{VE}}+\lambda{\bf {I}}}\right)^{-2}}]{\bf {Z}}_{VE}}\right)\nonumber \\ & {=}\sum\limits _{i=1}^{r_{V}}\left[{\frac{{{\left[{{\bf {Z}}_{V}}\right]}_{i,i}}}{{{\left({{{\left[{{\Sigma}_{V}}\right]}_{i,i}}\!+\!\lambda}\right)}^{2}}}}\right]+\sum\limits _{i=1}^{r_{VE}}\left[{\frac{{{\left[{{\bf {Z}}_{VE}}\right]}_{i,i}}}{{{\left({{{\left[{{\Sigma}_{VE}}\right]}_{i,i}}\!+\!\lambda}\right)}^{2}}}}\right]+\sum\limits _{i={r_{V}}+1}^{{N_{T}}}{\left[{\frac{{{\left[{{\bf {Z}}_{V}}\right]}_{i,i}}}{{{\left({\lambda}\right)}^{2}}}}\right]}\!+\!\sum\limits _{i={r_{VE}}+1}^{{N_{T}}}{\left[{\frac{{{\left[{{\bf {Z}}_{VE}}\right]}_{i,i}}}{{{\left({\lambda}\right)}^{2}}}}\right]}, \end{align}
P(λ)=Tr([(PVΣVPVH+λPVPVH)−1(PVΣVPVH+λPVPVH)−1](H^IHUIWIH)(H^IHUIWIH)H)+Tr([(PVEΣVEPVEH+λPVEPVEH)−1(PVEΣVEPVEH+λPVEPVEH)−1](H^EHUEWEH)(H^EHUEWEH)H)=Tr([(ΣV+λI)−2]ZV)+Tr([(ΣVE+λI)−2]ZVE)=i=1∑rV
([ΣV]i,i+λ)2[ZV]i,i
+i=1∑rVE
([ΣVE]i,i+λ)2[ZVE]i,i
+i=rV+1∑NT[(λ)2[ZV]i,i]+i=rVE+1∑NT[(λ)2[ZVE]i,i],
where
Z
V
=
P
V
H
(
H
^
I
H
U
I
W
I
H
)
(
H
^
I
H
U
I
W
I
H
)
H
P
V
{{\bf{Z}}_{V}}={\bf{P}}_V^H( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{\bf{W}}_I^H} ) ( {{\bf{\hat H}}_I^H{{\bf{U}}_I}{\bf{W}}_I^H} )^H {{\bf{P}}_V}
ZV=PVH(H^IHUIWIH)(H^IHUIWIH)HPV and
Z
V
E
=
P
V
E
H
(
H
^
E
H
U
E
W
E
H
)
(
H
^
E
H
U
E
W
E
H
)
H
P
V
E
{{\bf{Z}}_{VE}}={\bf{P}}_{VE}^H( {{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H} ) ( {{\bf{\hat H}}_E^H{{\bf{U}}_E}{\bf{W}}_E^H} )^H {{\bf{P}}_{VE}}
ZVE=PVEH(H^EHUEWEH)(H^EHUEWEH)HPVE.
[
Z
V
]
i
,
i
{\left[ {{{\bf{Z}}_V}} \right]}_{i,i}
[ZV]i,i,
[
Z
V
E
]
i
,
i
{\left[ {{{\bf{Z}}_{VE}}} \right]}_{i,i}
[ZVE]i,i,
[
Σ
V
]
i
,
i
{\left[ {{{\Sigma}}_{V}} \right]}_{i,i}
[ΣV]i,i, and
[
Σ
V
E
]
i
,
i
{\left[ {{{\Sigma}}_{VE}} \right]}_{i,i}
[ΣVE]i,i represent the
i
i
ith diagonal element of matrices
Z
V
{{{\bf{Z}}_V}}
ZV,
Z
V
E
{{\bf{Z}}_{VE}}
ZVE,
Σ
V
{{{\Sigma}}_{V}}
ΣV, and
Σ
V
E
{{{\Sigma}}_{VE}}
ΣVE, respectively. The first line of (\ref{Plamda}) is obtained by substituting (\ref{svdforHVHVE}) into the expression of
P
(
λ
)
P({\lambda})
P(λ) in (\ref{Plambdanew}). It can be verified from the last line of (\ref{Plamda}) that
P
(
λ
)
P({\lambda})
P(λ) is a monotonically decreasing function.
Then, the optimal
λ
⋆
\lambda^{\star}
λ⋆ can be obtained by solving the following equation,
∑
i
=
1
r
V
[
[
Z
V
]
i
,
i
(
[
Σ
V
]
i
,
i
+
λ
)
2
]
+
∑
i
=
1
r
V
E
[
[
Z
V
E
]
i
,
i
(
[
Σ
V
E
]
i
,
i
+
λ
)
2
]
+
∑
i
=
r
V
+
1
N
T
[
[
Z
V
]
i
,
i
(
λ
)
2
]
+
∑
i
=
r
V
E
+
1
N
T
[
[
Z
V
E
]
i
,
i
(
λ
)
2
]
=
P
T
.
\begin{align} \sum\limits_{i = 1}^{r_{V}}\!\left[{\frac{{{\left[ {{{\bf{Z}}_{V}}} \right]}_{i,i}}}{{{{\left( {{{\left[ {{{\Sigma}}_{V}} \right]}_{i,i}} + \lambda} \right)}^2}}}}\right]\!+\!\!\sum\limits_{i = 1}^{r_{VE}}\!\left[{\frac{{{\left[ {{{\bf{Z}}_{VE}}} \right]}_{i,i}}}{{{{\left( {{{\left[ {{{\Sigma}}_{VE}} \right]}_{i,i}} + \lambda} \right)}^2}}}}\right]\!+\!\! \sum\limits_{i = {r_{V}} + 1}^{{N_T}}\! {\left[{\frac{{{\left[ {{{\bf{Z}}_{V}}} \right]}_{i,i}}}{{{{\left( {\lambda} \right)}^2}}}}\right]}\!+ \!\!\sum\limits_{i = {r_{VE}} + 1}^{{N_T}}\! {\left[{\frac{{{\left[ {{{\bf{Z}}_{VE}}} \right]}_{i,i}}}{{{{\left( {\lambda} \right)}^2}}}}\right]}=P_{T}. \end{align}
i=1∑rV
([ΣV]i,i+λ)2[ZV]i,i
+i=1∑rVE
([ΣVE]i,i+λ)2[ZVE]i,i
+i=rV+1∑NT[(λ)2[ZV]i,i]+i=rVE+1∑NT[(λ)2[ZVE]i,i]=PT.
To solve it, the bisection search method is utilized. Since P ( ∞ ) = 0 P(\infty )=0 P(∞)=0, the solution to Equation (\ref{lambdaequat}) must exist. The lower bound of λ ⋆ \lambda^{\star} λ⋆ is a positive value approaching zero, while the upper bound of λ ⋆ \lambda^{\star} λ⋆ is given by
λ
⋆
<
∑
i
=
1
N
T
[
Z
V
]
i
,
i
+
∑
i
=
1
N
T
[
Z
V
E
]
i
,
i
P
T
≜
λ
u
b
.
{\lambda^{\star}} < \sqrt {\frac{{\sum\limits_{i = 1}^{{N_T}} {{{\left[ {{{\bf{Z}}_V}} \right]}_{i,i}}} }+{\sum\limits_{i = 1}^{{N_T}} {{{\left[ {{{\bf{Z}}_{VE}}} \right]}_{i,i}}} }}{{{P_{T}}}}} \triangleq \lambda^{{\rm{ub}}}.
λ⋆<PTi=1∑NT[ZV]i,i+i=1∑NT[ZVE]i,i≜λub.
which can be proved as
P
(
λ
u
b
)
=
∑
i
=
1
r
V
[
Z
V
]
i
,
i
(
[
Σ
V
]
i
,
i
+
λ
u
b
)
2
+
∑
i
=
1
r
V
E
[
Z
V
E
]
i
,
i
(
[
Σ
V
E
]
i
,
i
+
λ
u
b
)
2
+
∑
i
=
r
V
+
1
N
T
[
[
Z
V
]
i
,
i
(
λ
u
b
)
2
]
+
∑
i
=
r
V
E
+
1
N
T
[
[
Z
V
E
]
i
,
i
(
λ
u
b
)
2
]
<
∑
i
=
1
N
T
[
Z
V
]
i
,
i
(
λ
u
b
)
2
+
∑
i
=
1
N
T
[
Z
V
E
]
i
,
i
(
λ
u
b
)
2
=
P
T
.
\begin{align} {P}(\lambda^{{\rm{ub}}})&=\sum\limits_{i = 1}^{r_{V}} {\frac{{{{\left[ {{{\bf{Z}}_V}} \right]}_{i,i}}}}{{{{\left({{{\left[{{{\Sigma}}_{V}} \right]}_{i,i}} + {\lambda^{{\rm{ub}}}}} \right)}^2}}}}+\sum\limits_{i = 1}^{r_{VE}} {\frac{{{{\left[ {{{\bf{Z}}_{VE}}}\right]}_{i,i}}}}{{{{\left( {{{\left[ {{{\Sigma}}_{VE}} \right]}_{i,i}} + {\lambda^{{\rm{ub}}}}} \right)}^2}}}}+\sum\limits_{i={r_{V}}+1}^{{N_{T}}}{\left[{\frac{{{\left[{{\bf{Z}}_{V}}\right]}_{i,i}}}{{{\left({\lambda^{{\rm{ub}}}}\right)}^{2}}}}\right]}\!+\!\sum\limits _{i={r_{VE}}+1}^{{N_{T}}}{\left[{\frac{{{\left[{{\bf{Z}}_{VE}}\right]}_{i,i}}}{{{\left({\lambda^{{\rm{ub}}}}\right)}^{2}}}}\right]}\nonumber \\ & < \sum\limits_{i = 1}^{{N_T}} {\frac{{{{\left[ {{{\bf{Z}}_V}} \right]}_{i,i}}}}{{{{\left( {\lambda^{{\rm{ub}}}} \right)}^2}}}}+\sum\limits_{i = 1}^{{N_T}} {\frac{{{{\left[ {{{\bf{Z}}_{VE}}} \right]}_{i,i}}}}{{{{\left( {\lambda^{{\rm{ub}}}} \right)}^2}}}} = {P_{T}}. \end{align}
P(λub)=i=1∑rV([ΣV]i,i+λub)2[ZV]i,i+i=1∑rVE([ΣVE]i,i+λub)2[ZVE]i,i+i=rV+1∑NT[(λub)2[ZV]i,i]+i=rVE+1∑NT[(λub)2[ZVE]i,i]<i=1∑NT(λub)2[ZV]i,i+i=1∑NT(λub)2[ZVE]i,i=PT.
When the optimal λ ⋆ \lambda^{{\star}} λ⋆ is found, the optimal matrices V ⋆ {{\bf{V}}^{{\star}}} V⋆ and V E ⋆ {{\bf{V}}_{E}^{{\star}}} VE⋆ can be obtained by substituting λ ⋆ \lambda^\star λ⋆ into (10) and (12).