文章目录
III. JOINT ANTENNA POSITION AND BEAMFORMING DESIGN ALGORITHM
(P1) min { w k , u k } ∑ k = 1 K ∥ w k ∥ 2 , s.t. ∣ t k , k ∣ 2 ∑ q = 1 , q ≠ k K ∣ t k , q ∣ 2 + σ 2 ≥ γ k , ∀ k , t k , q = h k ( u k ) H w q , ∀ k , q , u k ∈ C k , ∀ k . \begin{aligned} \text { (P1) } & \min _{\left\{\mathbf{w}_{k}, \mathbf{u}_{k}\right\}} \sum_{k=1}^{K}\left\|\mathbf{w}_{k}\right\|^{2}, \\ \text { s.t. } & \frac{\left|t_{k, k}\right|^{2}}{\sum_{q=1, q \neq k}^{K}\left|t_{k, q}\right|^{2}+\sigma^{2}} \geq \gamma_{k}, \forall k, \\ & t_{k, q}=\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q}, \forall k, q, \\ & \mathbf{u}_{k} \in \mathcal{C}_{k}, \forall k . \end{aligned} (P1) s.t. {wk,uk}mink=1∑K∥wk∥2,∑q=1,q=kK∣tk,q∣2+σ2∣tk,k∣2≥γk,∀k,tk,q=hk(uk)Hwq,∀k,q,uk∈Ck,∀k.
在新引入的等式约束中,变量 { w k } \{\mathbf{w}_{k}\} {wk}、 { u k } \{\mathbf u_k\} {uk} 相互耦合。为了解决这个问题,我们基于罚函数法将这些约束整合到目标函数中。特别地,我们通过将这些等式约束转化为二次函数并随后将其作为惩罚项合并到目标函数中,得到以下优化问题
( P 2 ) min { w k , u k , t k , q } ∑ k = 1 K ∥ w k ∥ 2 + 1 2 ρ ( ∑ k = 1 K ∑ q = 1 K ∣ h k ( u k ) H w q − t k , q ∣ 2 ) , s.t. (7a), (7c), \begin{aligned} (P2)& \min _{\left\{\mathbf{w}_{k}, \mathbf{u}_{k}, t_{k, q}\right\}} \sum_{k=1}^{K}\left\|\mathbf{w}_{k}\right\|^{2} \\ & +\frac{1}{2 \rho}\left(\sum_{k=1}^{K} \sum_{q=1}^{K}\left|\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q}-t_{k, q}\right|^{2}\right), \\ \text { s.t. } & \text { (7a), (7c), } \end{aligned} (P2) s.t. {wk,uk,tk,q}mink=1∑K∥wk∥2+2ρ1(k=1∑Kq=1∑K hk(uk)Hwq−tk,q 2), (7a), (7c),
其中 ρ > 0 ρ> 0 ρ>0 表示用于惩罚(P1)中偏离等式约束的惩罚因子。值得注意的是,即使等式约束在(P2)中是放松的,当 ρ → 0 ρ→0 ρ→0 时,从求解(P2)中得到的解通常遵循(P1)中的约束。
A. Alternating Optimization Algorithm for Solving (P2)
- 关于 { w k } \{\mathbf{w}_{k}\} {wk}:的子问题对于任意给定变量 { u k } \left\{\mathbf{u}_{k}\right\} {uk}, { t k , q } \left\{t_{k, q}\right\} {tk,q},(P2)可以转化为(P3)。
(P3) min { w k } ∑ k = 1 K ∥ w k ∥ 2 + 1 2 ρ ( ∑ k = 1 K ∑ q = 1 K ∣ h k ( u k ) H w q − t k , q ∣ 2 ) . \text { (P3) } \min _{\left\{\mathbf{w}_{k}\right\}} \sum_{k=1}^{K}\left\|\mathbf{w}_{k}\right\|^{2}+\frac{1}{2 \rho}\left(\sum_{k=1}^{K} \sum_{q=1}^{K}\left|\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q}-t_{k, q}\right|^{2}\right) \text {. } (P3) {wk}mink=1∑K∥wk∥2+2ρ1(k=1∑Kq=1∑K hk(uk)Hwq−tk,q 2).
由于这是一个无约束的二次凸问题,我们可以通过使目标函数对 w k \mathbf{w}_{k} wk 的一阶偏导数等于零来获得最优的发射波束成形向量 w k \mathbf{w}_{k} wk,其闭式解如下
w k ∗ = 1 2 ρ A − 1 ( ∑ q = 1 K t q , k h q ( u q ) ∗ ) , ∀ k , \mathbf{w}_{k}^{*}=\frac{1}{2 \rho} \mathbf{A}^{-1}\left(\sum_{q=1}^{K} t_{q, k} \mathbf{h}_{q}\left(u_{q}\right)^{*}\right), \forall k, wk∗=2ρ1A−1(q=1∑Ktq,khq(uq)∗),∀k,
其中 A = I N + 1 2 ρ ∑ q = 1 K h q ( u q ) h q ( u q ) H \mathbf{A}=\mathbf{I}_{N}+\frac{1}{2 \rho} \sum_{q=1}^{K} \mathbf{h}_{q}\left(\mathbf{u}_{q}\right) \mathbf{h}_{q}\left(\mathbf{u}_{q}\right)^{H} A=IN+2ρ1∑q=1Khq(uq)hq(uq)H。由于的所有 w k \mathbf{w}_{k} wk 目标函数中不同的用户是相互分离的,他们可以根据上面的等式同时更新。
-
{
t
k
,
q
}
,
∀
k
,
q
\left\{t_{k, q}\right\}, \forall k, q
{tk,q},∀k,q:的子问题:对于任何给定的波束形成向量
{
w
k
}
\{\mathbf{w}_{k}\}
{wk} 和MA的位置
{
u
k
}
\left\{\mathbf{u}_{k}\right\}
{uk},我们用约束(7a)来求解(P2)以优化辅助变量
{
t
k
,
q
}
\left\{t_{k, q}\right\}
{tk,q}。不难观察到,不同用户的辅助变量是相互分离的,因此由此产生的问题可以被分割成
K
K
K 个独立的并行子问题同时求解。特别是,通过忽略常数项,用户
k
k
k 对应的子问题可以简化为
(P4) min { t k , q , ∀ q } ∑ q = 1 K ∣ t ˉ k , q − t k , q ∣ 2 , s.t. ∣ t k , k ∣ 2 ∑ q = 1. q ≠ k K ∣ t k , q ∣ 2 + σ 2 ≥ γ k , \begin{align*} \text { (P4) } \quad& \min _{\left\{t_{k, q}, \forall q\right\}} \sum_{q=1}^{K}\left|\bar{t}_{k, q}-t_{k, q}\right|^{2}, \\ \text { s.t. }\quad & \frac{\left|t_{k, k}\right|^{2}}{\sum_{q=1 . q \neq k}^{K}\left|t_{k, q}\right|^{2}+\sigma^{2}} \geq \gamma_{k},\tag{11a} \end{align*} (P4) s.t. {tk,q,∀q}minq=1∑K∣tˉk,q−tk,q∣2,∑q=1.q=kK∣tk,q∣2+σ2∣tk,k∣2≥γk,(11a)
其中
t
ˉ
k
,
q
=
h
k
(
u
k
)
H
w
q
,
k
,
q
=
1
,
2
,
⋯
,
K
\bar{t}_{k, q}=\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q}, k, q=1,2, \cdots, K
tˉk,q=hk(uk)Hwq,k,q=1,2,⋯,K。(P4)是一个只有一个约束的非凸二次约束二次规划问题(QCQP)。尽管如此,[11]已经证明了该非凸问题具有很强的对偶性,这意味着可以通过拉格朗日对偶方法确定最优解。设
λ
k
≥
0
,
∀
k
λ_k≥0,∀k
λk≥0,∀k 表示对偶变量,则对应(P4)的拉格朗日对偶函数表示如下
L
(
λ
k
,
{
t
k
,
q
}
)
=
(
1
−
λ
k
)
∣
t
k
,
k
∣
2
+
∑
q
=
1
,
q
≠
k
K
(
1
+
λ
k
γ
k
)
∣
t
k
,
q
∣
2
−
2
∑
q
=
1
K
Re
{
t
ˉ
k
,
q
t
k
,
q
H
}
\begin{aligned} \mathcal{L}\left(\lambda_{k},\left\{t_{k, q}\right\}\right) & =\left(1-\lambda_{k}\right)\left|t_{k, k}\right|^{2}+\sum_{q=1, q \neq k}^{K}\left(1+\lambda_{k} \gamma_{k}\right)\left|t_{k, q}\right|^{2} \\ & -2 \sum_{q=1}^{K} \operatorname{Re}\left\{\bar{t}_{k, q} t_{k, q}^{H}\right\} \end{aligned}
L(λk,{tk,q})=(1−λk)∣tk,k∣2+q=1,q=k∑K(1+λkγk)∣tk,q∣2−2q=1∑KRe{tˉk,qtk,qH}
According to Boyd we know that a non-convex QCQP problem with one quadratic constraint has strong duality with the relaxed SDP or Lagrange counterpart. (check “Convex Optimization” by Boyd, Appendix B)
这样,相关联的对偶函数可以写成
G
(
λ
k
)
=
inf
{
t
k
,
q
}
L
(
λ
k
,
{
t
k
,
q
}
)
\mathcal{G}\left(\lambda_{k}\right)=\inf _{\left\{t_{k, q}\right\}} \mathcal{L}\left(\lambda_{k},\left\{t_{k, q}\right\}\right)
G(λk)=inf{tk,q}L(λk,{tk,q})。通过设置
式(12)中的拉格朗日函数相对于
λ
k
λ_k
λk 一阶偏导数趋于零,我们可以推导出使
L
(
λ
k
,
{
t
k
,
q
}
)
\mathcal{L}\left(\lambda_{k},\left\{t_{k, q}\right\}\right)
L(λk,{tk,q}) 最小的最优解如下
t k , k ∗ = t ˉ k , k 1 − λ k , t k , q ∗ = t ˉ k , q 1 + λ k γ k , q ≠ k t_{k, k}^{*}=\frac{\bar{t}_{k, k}}{1-\lambda_{k}} \quad, \quad t_{k, q}^{*}=\frac{\bar{t}_{k, q}}{1+\lambda_{k} \gamma_{k}}, \quad q \neq k tk,k∗=1−λktˉk,k,tk,q∗=1+λkγktˉk,q,q=k
当(11a)中的约束满足方程时,通过代入导出的
t
k
,
k
∗
,
t
k
,
q
∗
t_{k, k}^{*}, t_{k, q}^{*}
tk,k∗,tk,q∗ back (11a),这个等式约束可以转化为另一种形式如下
F
(
λ
k
)
=
∣
t
ˉ
k
,
k
∣
2
(
1
−
λ
k
)
2
−
γ
k
∑
q
=
1
,
q
≠
k
K
∣
t
ˉ
k
,
q
∣
2
(
1
+
λ
k
γ
k
)
2
−
γ
k
σ
2
=
0
\mathcal{F}\left(\lambda_{k}\right)=\frac{\left|\bar{t}_{k, k}\right|^{2}}{\left(1-\lambda_{k}\right)^{2}}-\gamma_{k} \sum_{q=1, q \neq k}^{K} \frac{\left|\bar{t}_{k, q}\right|^{2}}{\left(1+\lambda_{k} \gamma_{k}\right)^{2}}-\gamma_{k} \sigma^{2}=0
F(λk)=(1−λk)2∣tˉk,k∣2−γkq=1,q=k∑K(1+λkγk)2∣tˉk,q∣2−γkσ2=0
很容易观察到 F ( λ k ) \mathcal{F}\left(\lambda_{k}\right) F(λk) 在区域 0 ≤ λ k < 1 0≤λ_k< 1 0≤λk<1 中是 λ k λ_k λk 的单调递增函数,因此我们可以通过二分搜索[12]获得最优对偶变量。然而,如果(11a)中的约束方程不成立,即 λ k = 0 λ_k= 0 λk=0,则等式(13)退化为 t k , q ∗ = t ˉ k , q , ∀ k , q t_{k, q}^{*}=\bar{t}_{k, q}, \forall k, q tk,q∗=tˉk,q,∀k,q。
- 带 { u k } \left\{\mathbf{u}_{k}\right\} {uk} 的子问题:对于任意给定变量 { w k } \{\mathbf{w}_{k}\} {wk}, { t k , q } , ∀ k , q : \left\{t_{k, q}\right\}, \forall k, q: {tk,q},∀k,q: ,可以通过约束(7c)求解(P2)来优化天线位置 { u k } \left\{\mathbf{u}_{k}\right\} {uk} 。考虑到每个用户 { u k } \left\{\mathbf{u}_{k}\right\} {uk} 的位置变量是相互独立的,(P5)可以分解成并行的子问题。对于用户 k k k,对应的子问题如下:
(P5) min u k ∑ q = 1 K ∣ h k ( u k ) H w q − t k , q ∣ 2 , s.t. u k ∈ C k , ∀ k . \begin{aligned} \text { (P5) } \quad & \min _{\mathbf{u}_{k}} \sum_{q=1}^{K}\left|\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q}-t_{k, q}\right|^{2}, \\ \text { s.t. } \quad & \mathbf{u}_{k} \in \mathcal{C}_{k}, \forall k . \end{aligned} (P5) s.t. ukminq=1∑K hk(uk)Hwq−tk,q 2,uk∈Ck,∀k.
由于 ∣ h k ( u k ) H w q − t k , q ∣ 2 = h k ( u k ) H w q w q H h k ( u k ) − 2 Re { h k ( u k ) H w q t k , q ∗ } + ∣ t k , q ∣ 2 \left|\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q}-t_{k, q}\right|^{2}=\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q} \mathbf{w}_{q}^{H} \mathbf{h}_{k}\left(\mathbf{u}_{k}\right)-2 \operatorname{Re}\left\{\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q} t_{k, q}^{*}\right\}+\left|t_{k, q}\right|^{2} hk(uk)Hwq−tk,q 2=hk(uk)HwqwqHhk(uk)−2Re{hk(uk)Hwqtk,q∗}+∣tk,q∣2,可将(P5)中的目标函数转化为Eq.(16),其中定义 B k ≜ Σ k G k , C k , q ≜ B k ∗ w q w q H B k T , d k , q ≜ B k ∗ w q , ξ ( i , j , k ) ≜ 2 π λ ( ρ k , i r ( u k ) − ρ k , j r ( u k ) ) , f 1 ( i , j , k , q ) ≜ 2 ∣ C k , q ( i , j ) ∣ cos ( ξ ( i , j , k ) + ∠ C k , q ( i , j ) ) , f 2 ( l , k , q ) ≜ 2 ∣ t k , q ∣ ∣ d k , q l ∣ cos ( 2 π λ ρ k , l r ( u k ) + ∠ d k , q l − ∠ t k , q ) \mathbf{B}_{k} \triangleq \mathbf{\Sigma}_{k} \mathbf{G}_{k}, \mathbf{C}_{k, q} \triangleq \mathbf{B}_{k}^{*} \mathbf{w}_{q} \mathbf{w}_{q}^{H} \mathbf{B}_{k}^{T}, \mathbf{d}_{k, q} \triangleq \mathbf{B}_{k}^{*} \mathbf{w}_{q},\xi(i, j, k) \triangleq \frac{2 \pi}{\lambda}\left(\rho_{k, i}^{r}\left(\mathbf{u}_{k}\right)-\rho_{k, j}^{r}\left(\mathbf{u}_{k}\right)\right),f_{1}(i, j, k, q) \triangleq 2\left|\mathbf{C}_{k, q}(i, j)\right| \cos \left(\xi(i, j, k)+\angle \mathbf{C}_{k, q}(i, j)\right),f_{2}(l, k, q) \triangleq 2\left|t_{k, q}\right|\left|d_{k, q}^{l}\right| \cos \left(\frac{2 \pi}{\lambda} \rho_{k, l}^{r}\left(\mathbf{u}_{k}\right)+\angle d_{k, q}^{l}-\angle t_{k, q}\right) Bk≜ΣkGk,Ck,q≜Bk∗wqwqHBkT,dk,q≜Bk∗wq,ξ(i,j,k)≜λ2π(ρk,ir(uk)−ρk,jr(uk)),f1(i,j,k,q)≜2∣Ck,q(i,j)∣cos(ξ(i,j,k)+∠Ck,q(i,j)),f2(l,k,q)≜2∣tk,q∣ dk,ql cos(λ2πρk,lr(uk)+∠dk,ql−∠tk,q)。
g ( u k ) = ∑ q = 1 K ( C k , q ( 1 , 1 ) + C k , q ( 2 , 2 ) + ⋯ + C k , q ( L r , L r ) + ∑ i = 1 L r − 1 ∑ j = i + 1 L r f 1 ( i , j , k , q ) − ∑ l = 1 L r f 2 ( l , k , q ) + ∣ t k , q ∣ 2 ) (16) \mathrm{g}\left(\mathbf{u}_{k}\right)=\sum_{q=1}^{K}\left(\mathbf{C}_{k, q}(1,1)+\mathbf{C}_{k, q}(2,2)+\cdots+\mathbf{C}_{k, q}\left(L_{r}, L_{r}\right)+\sum_{i=1}^{L_{r}-1} \sum_{j=i+1}^{L_{r}} f_{1}(i, j, k, q)-\sum_{l=1}^{L_{r}} f_{2}(l, k, q)+\left|t_{k, q}\right|^{2}\right)\tag{16} g(uk)=q=1∑K(Ck,q(1,1)+Ck,q(2,2)+⋯+Ck,q(Lr,Lr)+i=1∑Lr−1j=i+1∑Lrf1(i,j,k,q)−l=1∑Lrf2(l,k,q)+∣tk,q∣2)(16)
∑ l = 1 L r f 2 ( l , k , q ) \sum_{l=1}^{L_{r}} f_{2}(l, k, q) ∑l=1Lrf2(l,k,q) 实际上对应的是 2 Re { h k ( u k ) H w q t k , q ∗ } 2 \operatorname{Re}\left\{\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q} t_{k, q}^{*}\right\} 2Re{hk(uk)Hwqtk,q∗} 这一项,将上式展开得到的是 2 Re { f k ( u k ) T d k , q t k , q ∗ } 2 \operatorname{Re}\left\{\mathbf{f}_{k}\left(\mathbf{u}_{k}\right)^T\mathbf{d}_{k, q}t_{k, q}^{*}\right\} 2Re{fk(uk)Tdk,qtk,q∗},因此可以得到 f 2 ( i , j , k , q ) f_{2}(i, j, k, q) f2(i,j,k,q) 对应的形式。 h k ( u k ) H w q w q H h k ( u k ) \mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q} \mathbf{w}_{q}^{H} \mathbf{h}_{k}\left(\mathbf{u}_{k}\right) hk(uk)HwqwqHhk(uk) 的展开稍微复杂一点,但最终可以得到 f k ( u k ) T C k , q f k ( u k ) ∗ \mathbf{f}_{k}\left(\mathbf{u}_{k}\right)^T\mathbf{C}_{k, q}\mathbf{f}_{k}\left(\mathbf{u}_{k}\right)^* fk(uk)TCk,qfk(uk)∗。观察上式特别类似于二次型的形式,因此对于上式的展开,我们可以有 [ e j 2 π λ ρ k , 1 r ( u k ) , e j 2 π λ ρ k , 2 r ( u k ) , … , e j 2 π λ ρ k , L r r ( u k ) ] [ C k , q ( 1 , 1 ) ⋯ C k , q ( 1 , j ) ⋯ C k , q ( 1 , L r ) ⋮ ⋱ ⋮ ⋮ C k , q ( i , 1 ) ⋯ C k , q ( i , j ) ⋯ C k , q ( 2 , L r ) ⋮ ⋮ ⋱ ⋮ C k , q ( L r , 1 ) ⋯ C k , q ( L r , j ) ⋯ C k , q ( L r , L r ) ] [ e j 2 π λ ρ k , 1 r ( u k ) e j 2 π λ ρ k , 2 r ( u k ) ⋮ e j 2 π λ ρ k , L r r ( u k ) ] ∗ \left[e^{j \frac{2 \pi}{\lambda} \rho_{k, 1}^{r}\left(\mathbf{u}_{k}\right)}, e^{j \frac{2 \pi}{\lambda} \rho_{k, 2}^{r}\left(\mathbf{u}_{k}\right)}, \ldots, e^{j \frac{2 \pi}{\lambda} \rho_{k, L r}^{r}\left(\mathbf{u}_{k}\right)}\right]\begin{bmatrix} \mathbf{C}_{k, q}(1,1) & \cdots & \mathbf{C}_{k, q}(1,j) & \cdots & \mathbf{C}_{k, q}(1,L_r) \\ \vdots & \ddots & \vdots & & \vdots \\ \mathbf{C}_{k, q}(i,1) & \cdots & \color{red}\mathbf{C}_{k, q}(i,j) & \cdots & \mathbf{C}_{k, q}(2,L_r) \\ \vdots & & \vdots & \ddots & \vdots \\ \mathbf{C}_{k, q}(L_r,1)& \cdots & \mathbf{C}_{k, q}(L_r,j) & \cdots & \mathbf{C}_{k, q}(L_r, L_r) \\ \end{bmatrix} \begin{bmatrix} e^{j \frac{2 \pi}{\lambda} \rho_{k, 1}^{r}\left(\mathbf{u}_{k}\right)} \\ e^{j \frac{2 \pi}{\lambda} \rho_{k, 2}^{r}\left(\mathbf{u}_{k}\right)} \\ \vdots \\ e^{j \frac{2 \pi}{\lambda} \rho_{k, L r}^{r}\left(\mathbf{u}_{k}\right)} \end{bmatrix}^* [ejλ2πρk,1r(uk),ejλ2πρk,2r(uk),…,ejλ2πρk,Lrr(uk)] Ck,q(1,1)⋮Ck,q(i,1)⋮Ck,q(Lr,1)⋯⋱⋯⋯Ck,q(1,j)⋮Ck,q(i,j)⋮Ck,q(Lr,j)⋯⋯⋱⋯Ck,q(1,Lr)⋮Ck,q(2,Lr)⋮Ck,q(Lr,Lr) ejλ2πρk,1r(uk)ejλ2πρk,2r(uk)⋮ejλ2πρk,Lrr(uk) ∗分别将对角线上,和对角线上方的元素相加,即可得到上式中的 C k , q ( 1 , 1 ) + C k , q ( 2 , 2 ) + ⋯ + C k , q ( L r , L r ) + ∑ i = 1 L r − 1 ∑ j = i + 1 L r f 1 ( i , j , k , q ) \mathbf{C}_{k, q}(1,1)+\mathbf{C}_{k, q}(2,2)+\cdots+\mathbf{C}_{k, q}\left(L_{r}, L_{r}\right)+\sum_{i=1}^{L_{r}-1} \sum_{j=i+1}^{L_{r}} f_{1}(i, j, k, q) Ck,q(1,1)+Ck,q(2,2)+⋯+Ck,q(Lr,Lr)+∑i=1Lr−1∑j=i+1Lrf1(i,j,k,q)
∂
2
g
∂
x
k
2
=
4
π
2
λ
2
∑
q
=
1
K
∑
L
=
1
L
r
f
2
(
l
,
k
,
q
)
sin
2
θ
k
,
l
r
cos
2
ϕ
k
,
l
r
−
4
π
2
λ
2
∑
q
=
1
K
∑
i
=
1
L
r
−
1
∑
j
=
i
+
1
L
r
f
1
(
i
,
j
,
k
,
q
)
(
sin
θ
k
,
i
r
cos
ϕ
k
,
i
r
−
sin
θ
k
,
j
r
cos
ϕ
k
,
j
r
)
2
(17)
\frac{\partial^{2} \mathrm{~g}}{\partial x_{k}{ }^{2}}=\frac{4 \pi^{2}}{\lambda^{2}} \sum_{q=1}^{K} \sum_{L=1}^{L_{r}} f_{2}(l, k, q) \sin ^{2} \theta_{k, l}^{r} \cos ^{2} \phi_{k, l}^{r}-\frac{4 \pi^{2}}{\lambda^{2}} \sum_{q=1}^{K} \sum_{i=1}^{L_{r}-1} \sum_{j=i+1}^{L_{r}} f_{1}(i, j, k, q)\left(\sin \theta_{k, i}^{r} \cos \phi_{k, i}^{r}-\sin \theta_{k, j}^{r} \cos \phi_{k, j}^{r}\right)^{2}\tag{17}
∂xk2∂2 g=λ24π2q=1∑KL=1∑Lrf2(l,k,q)sin2θk,lrcos2ϕk,lr−λ24π2q=1∑Ki=1∑Lr−1j=i+1∑Lrf1(i,j,k,q)(sinθk,ircosϕk,ir−sinθk,jrcosϕk,jr)2(17)
∂
2
g
∂
y
k
2
=
4
π
2
λ
2
∑
q
=
1
K
∑
l
=
1
L
r
f
2
(
l
,
k
,
q
)
cos
2
θ
k
,
l
r
−
4
π
2
λ
2
∑
q
=
1
K
∑
i
=
1
L
r
−
1
∑
j
=
i
+
1
L
r
f
1
(
i
,
j
,
k
,
q
)
(
cos
θ
k
,
i
r
−
cos
θ
k
,
j
r
)
2
(18)
\frac{\partial^{2} \mathrm{~g}}{\partial y_{k}{ }^{2}}=\frac{4 \pi^{2}}{\lambda^{2}} \sum_{q=1}^{K} \sum_{l=1}^{L_{r}} f_{2}(l, k, q) \cos ^{2} \theta_{k, l}^{r}-\frac{4 \pi^{2}}{\lambda^{2}} \sum_{q=1}^{K} \sum_{i=1}^{L_{r}-1} \sum_{j=i+1}^{L_{r}} f_{1}(i, j, k, q)\left(\cos \theta_{k, i}^{r}-\cos \theta_{k, j}^{r}\right)^{2}\tag{18}
∂yk2∂2 g=λ24π2q=1∑Kl=1∑Lrf2(l,k,q)cos2θk,lr−λ24π2q=1∑Ki=1∑Lr−1j=i+1∑Lrf1(i,j,k,q)(cosθk,ir−cosθk,jr)2(18)
∂
2
g
∂
x
k
∂
y
k
=
4
π
2
λ
2
∑
q
=
1
K
∑
l
=
1
L
r
f
2
(
l
,
k
,
q
)
sin
θ
k
,
l
r
cos
θ
k
,
l
r
cos
ϕ
k
,
l
r
−
4
π
2
λ
2
∑
q
=
1
K
∑
i
=
1
L
r
−
1
∑
j
=
i
+
1
L
r
f
1
(
i
,
j
,
k
,
q
)
sin
θ
k
,
i
r
cos
ϕ
k
,
i
r
(
cos
θ
k
,
i
r
−
cos
θ
k
,
j
r
)
+
4
π
2
λ
2
∑
q
=
1
K
∑
i
=
1
L
r
−
1
∑
j
=
i
+
1
L
r
f
1
(
i
,
j
,
k
,
q
)
sin
θ
k
,
j
r
cos
ϕ
k
,
j
r
(
cos
θ
k
,
i
r
−
cos
θ
k
,
j
r
)
(19)
\begin{aligned} \frac{\partial^{2} \mathrm{~g}}{\partial x_{k} \partial y_{k}} & =\frac{4 \pi^{2}}{\lambda^{2}} \sum_{q=1}^{K} \sum_{l=1}^{L_{r}} f_{2}(l, k, q) \sin \theta_{k, l}^{r} \cos \theta_{k, l}^{r} \cos \phi_{k, l}^{r}-\frac{4 \pi^{2}}{\lambda^{2}} \sum_{q=1}^{K} \sum_{i=1}^{L_{r}-1} \sum_{j=i+1}^{L_{r}} f_{1}(i, j, k, q) \sin \theta_{k, i}^{r} \cos \phi_{k, i}^{r}\left(\cos \theta_{k, i}^{r}-\cos \theta_{k, j}^{r}\right) \\ & +\frac{4 \pi^{2}}{\lambda^{2}} \sum_{q=1}^{K} \sum_{i=1}^{L_{r}-1} \sum_{j=i+1}^{L_{r}} f_{1}(i, j, k, q) \sin \theta_{k, j}^{r} \cos \phi_{k, j}^{r}\left(\cos \theta_{k, i}^{r}-\cos \theta_{k, j}^{r}\right) \end{aligned}\tag{19}
∂xk∂yk∂2 g=λ24π2q=1∑Kl=1∑Lrf2(l,k,q)sinθk,lrcosθk,lrcosϕk,lr−λ24π2q=1∑Ki=1∑Lr−1j=i+1∑Lrf1(i,j,k,q)sinθk,ircosϕk,ir(cosθk,ir−cosθk,jr)+λ24π2q=1∑Ki=1∑Lr−1j=i+1∑Lrf1(i,j,k,q)sinθk,jrcosϕk,jr(cosθk,ir−cosθk,jr)(19)
为了解决由此产生的非凸问题,采用连续凸逼近(SCA)方法进行优化第
k
k
k 个用户的MA位置。利用泰勒定理,我们可以构造一个二次代理函数,作为目标函数的全局上界。给定在第
i
i
i 次迭代中提供的局部点
u
k
i
\mathbf u^i_k
uki,一个上界可以被获得为
g
(
u
k
)
≤
g
(
u
k
i
)
+
∇
g
(
u
k
i
)
T
(
u
k
−
u
k
i
)
+
δ
k
2
(
u
k
−
u
k
i
)
T
(
u
k
−
u
k
i
)
\mathrm{g}\left(\mathbf{u}_{k}\right) \leq\mathrm{g}\left(\mathbf{u}_{k}^{i}\right)+\nabla \mathrm{g}\left(\mathbf{u}_{k}^{i}\right)^{T}\left(\mathbf{u}_{k}-\mathbf{u}_{k}^{i}\right)+\frac{\delta_{k}}{2}\left(\mathbf{u}_{k}-\mathbf{u}_{k}^{i}\right)^{T}\left(\mathbf{u}_{k}-\mathbf{u}_{k}^{i}\right)
g(uk)≤g(uki)+∇g(uki)T(uk−uki)+2δk(uk−uki)T(uk−uki),这是通过引入一个正实数
δ
k
δ_k
δk 使得
δ
k
I
2
⪰
∇
2
g
(
u
k
)
\delta_{k} \mathbf{I}_{2} \succeq \nabla^{2} \mathrm{~g}\left(\mathbf{u}_{k}\right)
δkI2⪰∇2 g(uk)。
∥
∇
2
g
(
u
k
)
∥
2
2
≤
∥
∇
2
g
(
u
k
)
∥
F
2
=
(
∂
2
g
∂
x
k
2
)
2
+
(
∂
2
g
∂
x
k
∂
y
k
)
2
+
(
∂
2
g
∂
y
k
2
)
2
+
(
∂
2
g
∂
y
k
∂
x
k
)
2
\left\|\nabla^{2} g\left(\mathbf{u}_{k}\right)\right\|_{2}^{2} \leq\left\|\nabla^{2} g\left(\mathbf{u}_{k}\right)\right\|_{F}^{2}=\left(\frac{\partial^{2} \mathrm{~g}}{\partial x_{k}{ }^{2}}\right)^{2}+\left(\frac{\partial^{2} \mathrm{~g}}{\partial x_{k} \partial y_{k}}\right)^{2}+\left(\frac{\partial^{2} \mathrm{~g}}{\partial y_{k}{ }^{2}}\right)^{2}+\left(\frac{\partial^{2} \mathrm{~g}}{\partial y_{k} \partial x_{k}}\right)^{2}
∇2g(uk)
22≤
∇2g(uk)
F2=(∂xk2∂2 g)2+(∂xk∂yk∂2 g)2+(∂yk2∂2 g)2+(∂yk∂xk∂2 g)2,
δ
k
δ_k
δk 可以通过计算
g
(
u
k
)
\mathrm{g}\left(\mathbf{u}_{k}\right)
g(uk) 的Hessian矩阵的Frobenius范数来选择。根据Eq.(17)-(19),我们得到
δ
k
δ_k
δk 如下:
δ
k
=
16
π
2
λ
2
∑
q
=
1
K
(
∑
l
=
1
L
r
∣
t
k
,
q
∣
∣
d
k
,
q
l
∣
+
∑
i
=
1
L
r
−
1
∑
j
=
1
L
r
∣
C
k
,
q
(
i
,
j
)
∣
)
.
\delta_{k}=\frac{16 \pi^{2}}{\lambda^{2}} \sum_{q=1}^{K}\left(\sum_{l=1}^{L_{r}}\left|t_{k, q}\right|\left|d_{k, q}^{l}\right|+\sum_{i=1}^{L_{r}-1} \sum_{j=1}^{L_{r}}\left|\mathbf{C}_{k, q}(i, j)\right|\right) .
δk=λ216π2q=1∑K(l=1∑Lr∣tk,q∣
dk,ql
+i=1∑Lr−1j=1∑Lr∣Ck,q(i,j)∣).
这样,(P5)就简化为(P6)。
(
P
6
)
min
u
k
g
(
u
k
i
)
+
∇
g
(
u
k
i
)
T
(
u
k
−
u
k
i
)
+
δ
k
2
(
u
k
−
u
k
i
)
T
(
u
k
−
u
k
i
)
s.t.
u
k
∈
C
k
.
\begin{aligned} (P6)\min _{\mathbf{u}_{k}} & \mathrm{~g}\left(\mathbf{u}_{k}^{i}\right)+\nabla \mathrm{g}\left(\mathbf{u}_{k}^{i}\right)^{T}\left(\mathbf{u}_{k}-\mathbf{u}_{k}^{i}\right) \\ & +\frac{\delta_{k}}{2}\left(\mathbf{u}_{k}-\mathbf{u}_{k}^{i}\right)^{T}\left(\mathbf{u}_{k}-\mathbf{u}_{k}^{i}\right) \\ \text { s.t. } & \mathbf{u}_{k} \in \mathcal{C}_{k} . \end{aligned}
(P6)ukmin s.t. g(uki)+∇g(uki)T(uk−uki)+2δk(uk−uki)T(uk−uki)uk∈Ck.
由于(P6)的目标函数是关于 u k \mathbf{u}_{k} uk 的二次凸函数,忽略约束会得到全局最小解的闭式解 u k ⋆ = u k i − ∇ g ( u k i ) δ k \mathbf{u}_{k}^{\star}=\mathbf{u}_{k}^{i}-\frac{\nabla \mathrm{g}\left(\mathbf{u}_{k}^{i}\right)}{\delta_{k}} uk⋆=uki−δk∇g(uki) 。如果 u k ⋆ \mathbf{u}_{k}^{\star} uk⋆ 满足(P6)中的约束条件,那么它就是问题(P6)的最优解。否则,由于天线运动区域 C k \mathcal{C}_{k} Ck 是一个线性区域,因此该问题是一个二次规划(quadratic programming, QP)问题,使用CVX可以得到局部解。
B. Update Penalty Coefficient
考虑到(P1)中的等式约束在我们提出的算法收敛时应该保持。因此,惩罚系数逐渐减小如下 ρ : = c ρ , 0 < c < 1 ρ:= cρ, 0 < c < 1 ρ:=cρ,0<c<1,其中 c c c 是一个常数比例因子。
根据上述推导,所提出的天线位置和波束形成联合设计的算法在算法1中得到了简明扼要的总结。在内层,优化变量 { w k } \{\mathbf{w}_{k}\} {wk}, { u k } \left\{\mathbf{u}_{k}\right\} {uk}, { t k , q } \left\{t_{k, q}\right\} {tk,q} 交替更新,直到实现目标函数的收敛。在外层,一个指示器 ξ ξ ξ 被定义 ξ = max { ∣ h k ( u k ) H w q − t k , q ∣ 2 , ∀ k , q } \xi=\max \left\{\left|\mathbf{h}_{k}\left(\mathbf{u}_{k}\right)^{H} \mathbf{w}_{q}-t_{k, q}\right|^{2}, \forall k, q\right\} ξ=max{ hk(uk)Hwq−tk,q 2,∀k,q} 并且惩罚系数 ρ ρ ρ 不断更新,直到 ξ ξ ξ 小于阈值为止。