Introduction
The evolution of multiple-input multiple-output (MIMO) technologies for sixth-generation (6G) and beyond wireless network has continuously expanded the spatial degree-of-freedom (DoF) available for wireless transmission [1 , 2 ] . Nevertheless, conventional fixed-position antenna (FPA) arrays employ predetermined element locations that cannot adapt to the propagation environment, thereby limiting their ability to exploit spatial channel variations. To overcome this limitation, reconfigurable antenna technologies, such as movable antenna (MA) [3 , 4 ] and fluid antenna systems (FAS) [5 , 6 ] , have emerged as promising approaches for exploiting spatial channel variations. In representative position-reconfigurable FAS models, selecting favorable antenna locations, termed ports, provides spatial diversity [5 , 7 ] . Such spatial selection can also mitigate multi-user interference by exploiting locations where interfering signals undergo deep fading [8 ] . In MA systems, spatial reconfiguration is implemented by adaptively adjusting antenna positions within prescribed regions according to the propagation environment [3 , 9 ] . This enables MA to reshape the effective channel, circumvent unfavorable fading regions, and consequently enhance MIMO performance [10 ] .
Recent studies have demonstrated substantial performance gains of MA over FPA. In [9 ] , a field-response channel model characterized spatial multipath variations and revealed the fundamental spatial-diversity gains of MA. Building on this foundation, [10 ] characterized MA-MIMO capacity by jointly optimizing transmit/receive antenna positions and transmit covariance. For multi-user systems, [11 ] substantially reduced transmit power through joint antenna-position, power-control, and receive-combining optimization. In array processing, [12 ] jointly optimized antenna positions and weights to enhance desired multi-beam gains and suppress interference toward undesired directions. For efficient position optimization, [13 ] developed a discrete-sampling framework with sequential updates and Gibbs sampling to achieve near-optimal performance at low complexity, while [14 ] used differential evolution and adaptive constraint penalties to reduce the complexity of conventional population-based methods.
The synergy between MA and other technologies has also attracted growing attention. In [15 ] , MA positioning was jointly designed with intelligent reflecting surface (IRS) beamforming, revealing that the additional gain offered by antenna mobility depends on the channel-reconfiguration capability of the IRS. Furthermore, [16 ] developed a large-timescale MA positioning based on statistical channel-state information (CSI), demonstrating a communication-sensing tradeoff through reduced inter-user channel correlation and sensing ambiguity. In the context of integrated-sensing-and-communications (ISAC), [17 ] jointly optimized antenna positions and beamforming to enhance the ISAC performance. Overall, these studies have demonstrated the substantial benefits of MA, establishing it as a promising reconfigurable architecture for 6G and beyond.
Despite these advantages, conventional MA primarily exploit spatial DoF through antenna repositioning. With fixed antenna radiation characteristics, however, repositioning changes propagation distances and path-dependent spatial phases [10 ] . Position optimization alone therefore cannot directly reshape the radiation pattern to favor useful propagation paths or suppress interfering directions. This limitation is particularly relevant in multi-user and multipath environments, where paths arrive or depart along distinct angular directions [18 , 19 ] .
To handle this, pattern-reconfigurable antenna (PRA) offers complementary electromagnetic (EM)-domain DoF by dynamically adjusting their radiation patterns [20 ] , redistributing directional gains to enhance desired paths and suppress less useful or interfering directions [21 , 22 ] . Recent studies have demonstrated that PRA can substantially improve the performance of MIMO systems; In [23 ] , joint analog-phase and radiation-state optimization on FAS improved spectral-efficiency (SE), received power, and bit-error-rate (BER), as validated by full-wave simulations and hardware experiments. In [24 ] , spherical-harmonic-based radiation-pattern synthesis was jointly configured with digital precoding, achieving significant SE gains with low channel-estimation overhead. These results highlight PRA as a promising means of adapting the angular response to the propagation environment.
Realizing the benefits of PRA requires jointly optimizing their radiation states with multiple precoding stages. In hardware-efficient multi-antenna systems, precoding is conventionally implemented through hybrid beamforming across the digital baseband and analog radio-frequency (RF) domains [25 ] . Building upon this two-layer architecture, additional performance gains can be achieved by explicitly incorporating the antenna-domain DoFs provided by PRA, which has motivated the emergence of tri-hybrid beamforming [26 ] , which augments the hybrid layers with an EM-domain precoding layer [27 ] . By configuring the antenna characteristics, it directly shapes the radiated field and further reduces reliance on costly RF-domain hardware [19 , 28 , 29 ] , thereby enabling flexible and energy-efficient transmission.
Furthermore, PRA naturally complements MA: MA selects favorable spatial locations, while PRA adapts the angular response at each location. Since both reshape the effective multi-user channel, their gains are best realized through joint design with precoders. Tri-hybrid beamforming therefore supports this integration by combining digital/analog RF processing with EM-domain antenna-state control, alongside spatial position adaptation. However, this couples antenna positions, radiation states, and hybrid precoders through the multipath channel, while movement and pattern switching incur time and energy costs. Existing studies largely address these DoFs separately, leaving their joint optimization under hardware-inspired reconfiguration costs unexplored. Although [30 ] electronically emulates movement through discrete radiation-position selection, it neither treats physical displacement and pattern shaping as independent DoFs nor accounts for multilayer precoding and realistic reconfiguration overhead.
Motivated by these considerations, we introduce a movable pattern-reconfigurable antenna (MPRA) architecture and develop an energy-efficient tri-hybrid framework for multi-user system that explicitly accounts for hardware-inspired reconfiguration delays and energy costs. We formulate and solve an energy-efficiency (EE) maximization problem that balances the channel gains offered by spatial and pattern reconfiguration against their hardware-inspired costs. The main contributions are summarized as follows:
We propose an MPRA-enabled tri-hybrid transmission architecture that jointly leverages antenna positioning, radiation-pattern reconfiguration, and digital-analog precoding. A unified position- and pattern-dependent field-response model is established, in which antenna displacement controls the path-dependent spatial phases, whereas pattern selection reshapes the path-dependent directional gains. The resulting cascade explicitly incorporates pattern selection as an additional EM-domain precoding layer and establishes the tri-hybrid architecture.
We formulate a hardware-inspired EE model that captures both the energy expenditure and the loss of data
Figure 1. Illustration of a multi-user downlink system employing a tri-hybrid MPRA architecture at BS.
transmission time caused by antenna reconfiguration. We further demonstrate that operating each antenna at its maximum feasible synchronized speed is EE-optimal, thereby eliminating the movement-speed variables. Based on this model, the joint tri-hybrid MPRA design is formulated as a nonsmooth mixed discrete-continuous fractional program.
We develop an alternating optimization (AO) framework tailored to the problem. The hybrid-precoder block combines Dinkelbach’s method with weighted minimum mean-squared error (WMMSE) optimization, using closed-form digital updates with bisection and Riemannian conjugate-gradient (RCG) analog updates. Thereafter, a switching-aware cyclic search optimizes discrete patterns while accounting for transition costs and avoiding an exponential number of state enumeration. The antenna-position optimization block uses a quadratic transform and WMMSE-based successive convex approximation (SCA) to yield convex subproblems.
We establish that every block update preserves feasibility and does not decrease the original EE; consequently, the resulting EE sequence is monotonically nondecreasing and convergent. Numerical results further show that the proposed framework closely approaches the brute-force search (BFS) benchmark and consistently outperforms baselines in various scenarios.
System Model and Problem Formulation
Network Architecture
As illustrated in Fig. 1, we consider a multi-user downlink system in which a base station (BS) with N R F N_{\mathrm{RF}} N RF RF chains and N N N MPRAs simultaneously serves K K K single-antenna users. The user, MPRA, RF chains, and radiation-pattern index sets are denoted by K ≜ { 1 , ⋯ , K } \mathcal{K} \triangleq\{1,\cdots,K\} K ≜ { 1 , ⋯ , K } , N ≜ { 1 , ⋯ , N } \mathcal{N} \triangleq\{1,\cdots,N\} N ≜ { 1 , ⋯ , N } , N R F ≜ { 1 , ⋯ , N R F } \mathcal{N}_{\mathrm{RF}} \triangleq\{1,\cdots,N_{\mathrm{RF}}\} N RF ≜ { 1 , ⋯ , N RF } , and S ≜ { 1 , ⋯ , S } \mathcal{S} \triangleq\{1,\cdots,S\} S ≜ { 1 , ⋯ , S } , respectively, where K ≤ N R F ≤ N K \leq N_{\mathrm{RF}} \leq N K ≤ N RF ≤ N . Let s = [ s 1 ⋯ s K ] T ∼ C N ( 0 , I K ) \mathbf{s}=[s_1\cdots s_K]^{\mathrm{T}}\sim\mathcal{CN}(0,\mathbf{I}_K) s = [ s 1 ⋯ s K ] T ∼ C N ( 0 , I K ) denote the data-symbol vector. The digital/analog precoders are denoted by F B B = [ f 1 ⋯ f K ] ∈ C N R F × K \mathbf{F}_{\mathrm{BB}}=[\mathbf{f}_1\cdots\mathbf{f}_K]\in\mathbb{C}^{N_{\mathrm{RF}}\times K} F BB = [ f 1 ⋯ f K ] ∈ C N RF × K and F R F ∈ C N × N R F \mathbf{F}_{\mathrm{RF}}\in\mathbb{C}^{N\times N_{\mathrm{RF}}} F RF ∈ C N × N RF , respectively. The transmit vector delivered to the N N N MPRA ports is therefore x = F R F F B B s \mathbf{x}=\mathbf{F}_{\mathrm{RF}}\mathbf{F}_{\mathrm{BB}}\mathbf{s} x = F RF F BB s , with:
E [ ∥ x ∥ 2 ] = ∥ F R F F B B ∥ F 2 ≤ P max , (1) \mathbb{E}[\|\mathbf{x}\|^2]=\|\mathbf{F}_{\mathrm{RF}}\mathbf{F}_{\mathrm{BB}}\|_{\mathrm{F}}^2\leq P_{\max}, \tag*{(1)} E [ ∥ x ∥ 2 ] = ∥ F RF F BB ∥ F 2 ≤ P m a x , (1)
with power constraint P max P_{\max} P m a x . We assume a partially-connected phase-shifter network, where each MPRA is connected to exactly one RF chain [25 ] . Let M ≜ N / N R F ∈ Z M\triangleq N/N_{\mathrm{RF}}\in\mathbb{Z} M ≜ N / N RF ∈ Z denote the number of MPRAs connected to each RF chain and define N r ≜ { ( r − 1 ) M + 1 , ⋯ , r M } N_r \triangleq\{(r-1)M+1,\cdots,rM\} N r ≜ {( r − 1 ) M + 1 , ⋯ , r M } (r ∈ N R F r\in\mathcal{N}_{\mathrm{RF}} r ∈ N RF ).2 The corresponding F R F \mathcal{F}_{\mathrm{RF}} F RF -feasible set is [25 ] :
F R F ≜ { F R F = blkdiag ( f 1 , ⋯ , f N R F ) ∈ C N × N R F , f r ∈ C M , ∣ [ f r ] m ∣ = 1 M ( m = 1 , ⋯ , M ) } . (2) \mathcal{F}_{\mathrm{RF}}\triangleq\left\{\mathbf{F}_{\mathrm{RF}}=\operatorname{blkdiag}(\mathbf{f}_1,\cdots,\mathbf{f}_{N_{\mathrm{RF}}})\in\mathbb{C}^{N\times N_{\mathrm{RF}}},\mathbf{f}_r\in\mathbb{C}^{M},\left|[\mathbf{f}_r]_m\right|=\frac{1}{\sqrt{M}}\ (m=1,\cdots,M)\right\}. \tag*{(2)} F RF ≜ { F RF = blkdiag ( f 1 , ⋯ , f N RF ) ∈ C N × N RF , f r ∈ C M , ∣ [ f r ] m ∣ = M 1 ( m = 1 , ⋯ , M ) } . (2)
Since F R F ∗ F R F = I N R F \mathbf{F}_{\mathrm{RF}}^{*}\mathbf{F}_{\mathrm{RF}}=\mathbf{I}_{N_{\mathrm{RF}}} F RF ∗ F RF = I N RF by (2), (1) becomes
∥ F R F F B B ∥ F 2 = ∥ F B B ∥ F 2 ≤ P max . (3) \left\|\mathbf{F}_{\mathrm{RF}}\mathbf{F}_{\mathrm{BB}}\right\|_{F}^{2}=\left\|\mathbf{F}_{\mathrm{BB}}\right\|_{F}^{2}\leq P_{\max}. \tag*{(3)} ∥ F RF F BB ∥ F 2 = ∥ F BB ∥ F 2 ≤ P m a x . (3)
Movable Pattern-Reconfigurable Antenna (MPRA) Model
The n n n th MPRA moves inside { C n ⊂ R 2 } \{\mathcal{C}_n\subset\mathbb{R}^{2}\} { C n ⊂ R 2 } , a non-overlapping A × A A\times A A × A squares to prohibit physical collisions [10 , 32 ] . Its given original and target positions are denoted by r n 0 = [ x n 0 y n 0 ] T \mathbf{r}_n^{0}=[x_n^{0}\ y_n^{0}]^{\mathrm{T}} r n 0 = [ x n 0 y n 0 ] T and r n = [ x n y n ] T \mathbf{r}_n=[x_n\ y_n]^{\mathrm{T}} r n = [ x n y n ] T , respectively, and let R 0 ≜ [ r 1 0 ⋯ r N 0 ] ∈ R 2 × N \mathbf{R}^{0}\triangleq[\mathbf{r}_1^{0}\cdots\mathbf{r}_N^{0}]\in\mathbb{R}^{2\times N} R 0 ≜ [ r 1 0 ⋯ r N 0 ] ∈ R 2 × N and R ≜ [ r 1 ⋯ r N ] ∈ R 2 × N \mathbf{R}\triangleq[\mathbf{r}_1\cdots\mathbf{r}_N]\in\mathbb{R}^{2\times N} R ≜ [ r 1 ⋯ r N ] ∈ R 2 × N . To prevent excessive EM coupling [10 , 33 ] , { r n } \{\mathbf{r}_n\} { r n } satisfy
r n ∈ C n ( ∀ n ∈ N ) , ∥ r n − r m ∥ 2 ≥ d min ( ∀ n < m ) , (4) \mathbf{r}_n\in\mathcal{C}_n\ (\forall n\in\mathcal{N}),\quad\left\|\mathbf{r}_n-\mathbf{r}_m\right\|_2\geq d_{\min}\ (\forall n<m), \tag*{(4)} r n ∈ C n ( ∀ n ∈ N ) , ∥ r n − r m ∥ 2 ≥ d m i n ( ∀ n < m ) , (4)
with minimum separation d min > 0 d_{\min}>0 d m i n > 0 and so for { r n 0 } \{\mathbf{r}_n^{0}\} { r n 0 } but > d min >d_{\min} > d m i n instead of ≥ \geq ≥ ; this ensures the feasible initialization developed in Section III-E.
For pattern reconfiguration, each MPRA selects a pattern from a calibrated codebook of size S S S [19 , 21 , 23 , 29 ] :3
ψ ‾ ( Ω ) ≜ [ ψ ‾ 1 ( Ω ) ⋯ ψ ‾ S ( Ω ) ] T , (5) \overline{\boldsymbol{\psi}}(\Omega)\triangleq[\overline{\psi}_1(\Omega)\cdots\overline{\psi}_S(\Omega)]^{\mathrm{T}}, \tag*{(5)} ψ ( Ω ) ≜ [ ψ 1 ( Ω ) ⋯ ψ S ( Ω ) ] T , (5)
where Ω = ( θ , ϕ ) \Omega=(\theta,\phi) Ω = ( θ , ϕ ) denotes an elevation and azimuth departure direction, respectively, and ψ ‾ s ( Ω ) ∈ R + \overline{\psi}_s(\Omega)\in\mathbb{R}_{+} ψ s ( Ω ) ∈ R + denotes the s s s th normalized field-amplitude response, shared by all MPRAs.4 With state selection z n ∈ { 0 , 1 } S \mathbf{z}_n\in\{0,1\}^{S} z n ∈ { 0 , 1 } S , the configured pattern is
ψ n ( Ω ; z n ) = ψ ‾ T ( Ω ) z n , 1 S T z n = 1 , (6) \psi_n(\Omega;\mathbf{z}_n)=\overline{\boldsymbol{\psi}}^{\mathrm{T}}(\Omega)\mathbf{z}_n,\quad\mathbf{1}_{S}^{\mathrm{T}}\mathbf{z}_n=1, \tag*{(6)} ψ n ( Ω ; z n ) = ψ T ( Ω ) z n , 1 S T z n = 1 , (6)
and we collect { z n } \{\mathbf{z}_n\} { z n } in Z ≜ [ z 1 ⋯ z N ] ∈ { 0 , 1 } S × N \mathbf{Z}\triangleq[\mathbf{z}_1\cdots\mathbf{z}_N]\in\{0,1\}^{S\times N} Z ≜ [ z 1 ⋯ z N ] ∈ { 0 , 1 } S × N .
Tri-Hybrid Architecture with R- and Z-Dependency
We adopt a narrowband geometric block-fading model [18 , 19 , 34 ] . The channel between the BS and user k k k , varies with ( R , Z ) (\mathbf{R},\mathbf{Z}) ( R , Z ) [10 , 32 , 33 ] , consists of L k L_k L k far-field propagation paths. The ℓ \ell ℓ th path is characterized by Ω k , ℓ = ( θ k , ℓ , ϕ k , ℓ ) \Omega_{k,\ell}=(\theta_{k,\ell},\phi_{k,\ell}) Ω k , ℓ = ( θ k , ℓ , ϕ k , ℓ ) , i.i.d. channel gain α k , ℓ ∈ C \alpha_{k,\ell}\in\mathbb{C} α k , ℓ ∈ C with E [ ∣ α k , ℓ ∣ 2 ] = 1 \mathbb{E}[|\alpha_{k,\ell}|^2]=1 E [ ∣ α k , ℓ ∣ 2 ] = 1 , and path-loss β k \beta_k β k [19 ] . For Ω = ( θ , ϕ ) \Omega=(\theta,\phi) Ω = ( θ , ϕ ) , define u ( Ω ) = [ sin θ cos ϕ sin θ sin ϕ cos θ ] T \mathbf{u}(\Omega)=[\sin\theta\cos\phi\ \sin\theta\sin\phi\ \cos\theta]^{\mathrm{T}} u ( Ω ) = [ sin θ cos ϕ sin θ sin ϕ cos θ ] T . We use k 0 ≜ 2 π / λ k_0\triangleq2\pi/\lambda k 0 ≜ 2 π / λ for the wavenumber, and embed r n \mathbf{r}_n r n into r ~ n = [ r n T 0 ] T \widetilde{\mathbf{r}}_n=[\mathbf{r}_n^{\mathrm{T}}\ 0]^{\mathrm{T}} r n = [ r n T 0 ] T . The antenna-wise pattern selections are collected into:
F s e l ( Z ) ≜ blkdiag { z 1 , ⋯ , z N } ∈ { 0 , 1 } N S × N . (7) \mathbf{F}_{\mathrm{sel}}(\mathbf{Z})\triangleq\operatorname{blkdiag}\{\mathbf{z}_1,\cdots,\mathbf{z}_N\}\in\{0,1\}^{NS\times N}. \tag*{(7)} F sel ( Z ) ≜ blkdiag { z 1 , ⋯ , z N } ∈ { 0 , 1 } N S × N . (7)
For ( k , ℓ ) (k,\ell) ( k , ℓ ) , define the R \mathbf{R} R -dependent spatial-response as
a k , ℓ ( R ) ≜ [ e − j k 0 u T ( Ω k , ℓ ) r ~ 1 ⋮ e − j k 0 u T ( Ω k , ℓ ) r ~ N ] ∈ C N . \mathbf{a}_{k,\ell}(\mathbf{R})\triangleq\begin{bmatrix}e^{-\mathrm{j}k_0\mathbf{u}^{\mathrm{T}}(\Omega_{k,\ell})\widetilde{\mathbf{r}}_1}\\ \vdots\\ e^{-\mathrm{j}k_0\mathbf{u}^{\mathrm{T}}(\Omega_{k,\ell})\widetilde{\mathbf{r}}_N}\end{bmatrix}\in\mathbb{C}^{N}. a k , ℓ ( R ) ≜ e − j k 0 u T ( Ω k , ℓ ) r 1 ⋮ e − j k 0 u T ( Ω k , ℓ ) r N ∈ C N .
To represent all S S S states of every MPRA before pattern selection, introduce g ‾ k , ℓ ≜ 1 N ⊗ ψ ‾ ( Ω k , ℓ ) ∈ R N S \overline{\mathbf{g}}_{k,\ell}\triangleq\mathbf{1}_N\otimes\overline{\boldsymbol{\psi}}(\Omega_{k,\ell})\in\mathbb{R}^{NS} g k , ℓ ≜ 1 N ⊗ ψ ( Ω k , ℓ ) ∈ R N S and a ‾ k , ℓ ( R ) ≜ a k , ℓ ( R ) ⊗ 1 S ∈ C N S \overline{\mathbf{a}}_{k,\ell}(\mathbf{R})\triangleq\mathbf{a}_{k,\ell}(\mathbf{R})\otimes\mathbf{1}_S\in\mathbb{C}^{NS} a k , ℓ ( R ) ≜ a k , ℓ ( R ) ⊗ 1 S ∈ C N S . Their Hadamard product v k , ℓ ( R ) ≜ g ‾ k , ℓ ∘ a ‾ k , ℓ ( R ) ∈ C N S \mathbf{v}_{k,\ell}(\mathbf{R})\triangleq\overline{\mathbf{g}}_{k,\ell}\circ\overline{\mathbf{a}}_{k,\ell}(\mathbf{R})\in\mathbb{C}^{NS} v k , ℓ ( R ) ≜ g k , ℓ ∘ a k , ℓ ( R ) ∈ C N S gives the pattern-expanded row channel:
h k , s e l ∗ ( R ) ≜ β k L k ∑ ℓ = 1 L k α k , ℓ v k , ℓ T ( R ) ∈ C 1 × N S . (8) \mathbf{h}_{k,\mathrm{sel}}^{*}(\mathbf{R})\triangleq\sqrt{\frac{\beta_k}{L_k}}\sum_{\ell=1}^{L_k}\alpha_{k,\ell}\mathbf{v}_{k,\ell}^{\mathrm{T}}(\mathbf{R})\in\mathbb{C}^{1\times NS}. \tag*{(8)} h k , sel ∗ ( R ) ≜ L k β k ℓ = 1 ∑ L k α k , ℓ v k , ℓ T ( R ) ∈ C 1 × N S . (8)
The physical N N N -dimensional channel after pattern selection can then be factorized as h k ∗ ( R , Z ) = h k , s e l ∗ ( R ) F s e l ( Z ) ∈ C 1 × N \mathbf{h}_{k}^{*}(\mathbf{R},\mathbf{Z})=\mathbf{h}_{k,\mathrm{sel}}^{*}(\mathbf{R})\mathbf{F}_{\mathrm{sel}}(\mathbf{Z})\in\mathbb{C}^{1\times N} h k ∗ ( R , Z ) = h k , sel ∗ ( R ) F sel ( Z ) ∈ C 1 × N . Equivalently, its n n n th entry is
[ h k ∗ ( R , Z ) ] n = β k L k ∑ ℓ = 1 L k α k , ℓ ψ ‾ T ( Ω k , ℓ ) z n e − j k 0 u T ( Ω k , ℓ ) r ~ n . (9) [\mathbf{h}_{k}^{*}(\mathbf{R},\mathbf{Z})]_n=\sqrt{\frac{\beta_k}{L_k}}\sum_{\ell=1}^{L_k}\alpha_{k,\ell}\overline{\boldsymbol{\psi}}^{\mathrm{T}}(\Omega_{k,\ell})\mathbf{z}_n e^{-\mathrm{j}k_0\mathbf{u}^{\mathrm{T}}(\Omega_{k,\ell})\widetilde{\mathbf{r}}_n}. \tag*{(9)} [ h k ∗ ( R , Z ) ] n = L k β k ℓ = 1 ∑ L k α k , ℓ ψ T ( Ω k , ℓ ) z n e − j k 0 u T ( Ω k , ℓ ) r n . (9)
Finally, stacking the user channels gives H s e l ( R ) ≜ [ h 1 , s e l ∗ ( R ) ⋮ h K , s e l ∗ ( R ) ] ∈ C K × N S \mathbf{H}_{\mathrm{sel}}(\mathbf{R})\triangleq\begin{bmatrix}\mathbf{h}_{1,\mathrm{sel}}^{*}(\mathbf{R})\\ \vdots\\ \mathbf{h}_{K,\mathrm{sel}}^{*}(\mathbf{R})\end{bmatrix}\in\mathbb{C}^{K\times NS} H sel ( R ) ≜ h 1 , sel ∗ ( R ) ⋮ h K , sel ∗ ( R ) ∈ C K × N S .5 Consequently, the overall downlink channel exhibits the following tri-hybrid cascade:
y = H s e l ( R ) F s e l ( Z ) F R F F B B s ⏟ ≜ H e f f ( R , Z ) + n . (10) \mathbf{y}=\underbrace{\mathbf{H}_{\mathrm{sel}}(\mathbf{R})\mathbf{F}_{\mathrm{sel}}(\mathbf{Z})\mathbf{F}_{\mathrm{RF}}\mathbf{F}_{\mathrm{BB}}\mathbf{s}}_{\triangleq\mathbf{H}_{\mathrm{eff}}(\mathbf{R},\mathbf{Z})}+\mathbf{n}. \tag*{(10)} y = ≜ H eff ( R , Z ) H sel ( R ) F sel ( Z ) F RF F BB s + n . (10)
Herein, with ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) , F s e l ( Z ) \mathbf{F}_{\mathrm{sel}}(\mathbf{Z}) F sel ( Z ) offers the third precoding layer: switches the path amplitudes through the selected directional magnitude-gain patterns in (5) [26 , 27 ] .6
Received Signal and Achievable SE
From (10), the received signal at user k k k is
y k = h k , s e l ∗ ( R ) F s e l ( Z ) F R F f k s k ⏟ desired signal + ∑ j ≠ k h k , s e l ∗ ( R ) F s e l ( Z ) F R F f j s j ⏟ multi-user interference + n k . (11) y_k=\underbrace{\mathbf{h}_{k,\mathrm{sel}}^{*}(\mathbf{R})\mathbf{F}_{\mathrm{sel}}(\mathbf{Z})\mathbf{F}_{\mathrm{RF}}\mathbf{f}_ks_k}_{\text{desired signal}}+\underbrace{\sum_{j\ne k}\mathbf{h}_{k,\mathrm{sel}}^{*}(\mathbf{R})\mathbf{F}_{\mathrm{sel}}(\mathbf{Z})\mathbf{F}_{\mathrm{RF}}\mathbf{f}_js_j}_{\text{multi-user interference}}+n_k. \tag*{(11)} y k = desired signal h k , sel ∗ ( R ) F sel ( Z ) F RF f k s k + multi-user interference j = k ∑ h k , sel ∗ ( R ) F sel ( Z ) F RF f j s j + n k . (11)
where n k ∼ C N ( 0 , σ k 2 ) n_k\sim\mathcal{CN}(0,\sigma_k^2) n k ∼ C N ( 0 , σ k 2 ) . Hence, the SINR of user k k k is
γ k = ∣ h k , s e l ∗ ( R ) F s e l ( Z ) F R F f k ∣ 2 ∑ j ≠ k ∣ h k , s e l ∗ ( R ) F s e l ( Z ) F R F f j ∣ 2 + σ k 2 , (12) \gamma_k=\frac{\left|\mathbf{h}_{k,\mathrm{sel}}^{*}(\mathbf{R})\mathbf{F}_{\mathrm{sel}}(\mathbf{Z})\mathbf{F}_{\mathrm{RF}}\mathbf{f}_k\right|^2}{\sum_{j\ne k}\left|\mathbf{h}_{k,\mathrm{sel}}^{*}(\mathbf{R})\mathbf{F}_{\mathrm{sel}}(\mathbf{Z})\mathbf{F}_{\mathrm{RF}}\mathbf{f}_j\right|^2+\sigma_k^2}, \tag*{(12)} γ k = ∑ j = k h k , sel ∗ ( R ) F sel ( Z ) F RF f j 2 + σ k 2 h k , sel ∗ ( R ) F sel ( Z ) F RF f k 2 , (12)
and the sum-SE is therefore R w ≜ ∑ k = 1 K log 2 ( 1 + γ k ) R_{\mathrm{w}}\triangleq\sum_{k=1}^{K}\log_2(1+\gamma_k) R w ≜ ∑ k = 1 K log 2 ( 1 + γ k ) .
Reconfiguration Overhead and Block Energy
Reconfiguration Overhead and Data-Transmission Time
The physical movement of MPRA consumes a portion of each coherence block, giving rise to a tradeoff between the channel improvement obtained through reconfiguration and the remaining time available for data transmission [33 , 38 ] . At the beginning, the BS occupies a known original configuration R 0 \mathbf{R}^{0} R 0 and Z 0 ≜ [ z 1 0 ⋯ z N 0 ] ∈ { 0 , 1 } S × N \mathbf{Z}^{0} \triangleq[\mathbf{z}^{0}_{1}\ \cdots\ \mathbf{z}^{0}_{N}] \in\{0,1\}^{S \times N} Z 0 ≜ [ z 1 0 ⋯ z N 0 ] ∈ { 0 , 1 } S × N . After channel acquisition and optimization, the MPRAs reconfigure from R 0 , Z 0 \mathbf{R}^{0},\mathbf{Z}^{0} R 0 , Z 0 to R , Z \mathbf{R},\mathbf{Z} R , Z before data transmission begins.
To explicitly account for the biaxial positioning, let r n , a r_{n,a} r n , a denote the a ( ∈ A ) a(\in\mathcal{A}) a ( ∈ A ) th coordinate of r n \mathbf{r}_{n} r n . Extending [33 ] , each MPRA is mounted on a Cartesian positioning stage driven by two orthogonal linear actuators. Herein, the two axes are synchronized to generate a constant-speed straight-line trajectory from r n 0 \mathbf{r}^{0}_{n} r n 0 to r n \mathbf{r}_{n} r n [39 , 40 ] .7 Define the axis-wise displacement and the travel distance respectively as d n , a ( R ) ≜ ∣ r n , a − r n , a 0 ∣ d_{n,a}(\mathbf{R}) \triangleq\lvert r_{n,a}-r^{0}_{n,a}\rvert d n , a ( R ) ≜ ∣ r n , a − r n , a 0 ∣ ( a ∈ A ) (a \in\mathcal{A}) ( a ∈ A ) and D n ( R ) ≜ ∥ r n − r n 0 ∥ 2 ≜ ∑ a ∈ A d n , a 2 ( R ) D_{n}(\mathbf{R}) \triangleq\lVert\mathbf{r}_{n}-\mathbf{r}^{0}_{n}\rVert_{2} \triangleq\sqrt{\sum_{a\in\mathcal{A}}d_{n,a}^{2}(\mathbf{R})} D n ( R ) ≜ ∥ r n − r n 0 ∥ 2 ≜ ∑ a ∈ A d n , a 2 ( R ) . For D n ( R ) > 0 D_{n}(\mathbf{R})>0 D n ( R ) > 0 , let v n > 0 v_{n}>0 v n > 0 denote the resultant linear speed of MPRA n n n along its trajectory, where v n = 0 ( ↔ ∀ v n , a = 0 ) v_{n}=0(\leftrightarrow\forall v_{n,a}=0) v n = 0 ( ↔ ∀ v n , a = 0 ) is adopted by convention when D n ( R ) = 0 D_{n}(\mathbf{R})=0 D n ( R ) = 0 . The corresponding speed of actuator a a a is
v n , a ( R , v n ) ≜ v n d n , a ( R ) D n ( R ) ( a ∈ A , D n ( R ) > 0 ) , (13) v_{n,a}(\mathbf{R},v_{n}) \triangleq v_{n}\frac{d_{n,a}(\mathbf{R})}{D_{n}(\mathbf{R})}\quad(a\in\mathcal{A},D_{n}(\mathbf{R})>0), \tag*{(13)} v n , a ( R , v n ) ≜ v n D n ( R ) d n , a ( R ) ( a ∈ A , D n ( R ) > 0 ) , (13)
which ensures that both axes reach their target coordinates simultaneously [40 ] . Let v n , a max > 0 v^{\max}_{n,a}>0 v n , a m a x > 0 denote the maximum admissible speed of actuator a a a [33 ] . From (13) , the feasible v n v_{n} v n is bounded by 0 < v n ≤ v ˉ n ( R ) ≜ min a ∈ A : d n , a ( R ) > 0 D n ( R ) d n , a ( R ) v n , a max 0<v_{n}\le\bar{v}_{n}(\mathbf{R})\triangleq\min_{a\in\mathcal{A}:d_{n,a}(\mathbf{R})>0}\frac{D_{n}(\mathbf{R})}{d_{n,a}(\mathbf{R})}v^{\max}_{n,a} 0 < v n ≤ v ˉ n ( R ) ≜ min a ∈ A : d n , a ( R ) > 0 d n , a ( R ) D n ( R ) v n , a m a x . The movement time of MPRA n n n is therefore T n ( R , v n ) ≜ { D n ( R ) v n ( D n ( R ) > 0 ) 0 ( D n ( R ) = 0 ) T_{n}(\mathbf{R},v_{n})\triangleq\begin{cases}\frac{D_{n}(\mathbf{R})}{v_{n}} & (D_{n}(\mathbf{R})>0)\\ 0 & (D_{n}(\mathbf{R})=0)\end{cases} T n ( R , v n ) ≜ { v n D n ( R ) 0 ( D n ( R ) > 0 ) ( D n ( R ) = 0 ) .
Assuming that all MPRA positioning stages operate in parallel, the overall mechanical reconfiguration time is
T m o v ( R , v ) = max n ∈ N T n ( R , v n ) , v ≜ [ v 1 ⋯ v N ] T . (14) T_{\mathrm{mov}}(\mathbf{R},\mathbf{v})=\max_{n\in\mathcal{N}}T_{n}(\mathbf{R},v_{n}),\quad\mathbf{v}\triangleq[v_{1}\ \cdots\ v_{N}]^{\mathrm{T}}. \tag*{(14)} T mov ( R , v ) = n ∈ N max T n ( R , v n ) , v ≜ [ v 1 ⋯ v N ] T . (14)
Thereafter, the number of MPRAs whose pattern state changes is N s w ( Z ) ≜ 1 2 ∑ n = 1 N ∥ z n − z n 0 ∥ 1 N_{\mathrm{sw}}(\mathbf{Z})\triangleq\frac{1}{2}\sum_{n=1}^{N}\lVert\mathbf{z}_{n}-\mathbf{z}^{0}_{n}\rVert_{1} N sw ( Z ) ≜ 2 1 ∑ n = 1 N ∥ z n − z n 0 ∥ 1 . Assuming simultaneous electronic switching, the pattern-switching delay is T p a t ( Z ) = τ s w 1 N s w ( Z ) > 0 T_{\mathrm{pat}}(\mathbf{Z})=\tau_{\mathrm{sw}}\mathbf{1}_{N_{\mathrm{sw}}(\mathbf{Z})>0} T pat ( Z ) = τ sw 1 N sw ( Z ) > 0 with hardware switching time τ s w \tau_{\mathrm{sw}} τ sw . The total overhead and data-transmission time are thus respectively [33 ] : T o h ( R , Z , v ) = T m o v ( R , v ) + T p a t ( Z ) T_{\mathrm{oh}}(\mathbf{R},\mathbf{Z},\mathbf{v})=T_{\mathrm{mov}}(\mathbf{R},\mathbf{v})+T_{\mathrm{pat}}(\mathbf{Z}) T oh ( R , Z , v ) = T mov ( R , v ) + T pat ( Z ) and T d ( R , Z , v ) = T c − T o h ( R , Z , v ) T_{\mathrm{d}}(\mathbf{R},\mathbf{Z},\mathbf{v})=T_{\mathrm{c}}-T_{\mathrm{oh}}(\mathbf{R},\mathbf{Z},\mathbf{v}) T d ( R , Z , v ) = T c − T oh ( R , Z , v ) .
Total Block Energy
For each actuator, we adopt the speed-dependent stepper-motor model and allow its parameters to depend on both the antenna and the motion axis [33 , 40 ] . In particular, the power consumption is given by
P M , n , a ( v ) = v ℓ 0 , n , a M n , a ( v ℓ 0 , n , a ) , P_{\mathrm{M},n,a}(v)=\frac{v}{\ell_{0,n,a}}\mathcal{M}_{n,a}\left(\frac{v}{\ell_{0,n,a}}\right), P M , n , a ( v ) = ℓ 0 , n , a v M n , a ( ℓ 0 , n , a v ) ,
where ℓ 0 , n , a \ell_{0,n,a} ℓ 0 , n , a is the linear-displacement factor of the lead-screw transmission, such that v = ℓ 0 , n , a ω v=\ell_{0,n,a}\omega v = ℓ 0 , n , a ω with angular speed ω \omega ω , and M n , a ( ω ) \mathcal{M}_{n,a}(\omega) M n , a ( ω ) is the available motor torque at ω \omega ω :
M n , a ( ω ) = p n , a ω ψ M , n , a V n , a R m , n , a 2 + ω 2 L m , n , a 2 − p n , a ω ψ M , n , a 2 R m , n , a R m , n , a 2 + ω 2 L m , n , a 2 , \mathcal{M}_{n,a}(\omega)=\frac{p_{n,a}\omega\psi_{\mathrm{M},n,a}V_{n,a}}{\sqrt{R_{m,n,a}^{2}+\omega^{2}L_{m,n,a}^{2}}}-\frac{p_{n,a}\omega\psi_{\mathrm{M},n,a}^{2}R_{m,n,a}}{R_{m,n,a}^{2}+\omega^{2}L_{m,n,a}^{2}}, M n , a ( ω ) = R m , n , a 2 + ω 2 L m , n , a 2 p n , a ω ψ M , n , a V n , a − R m , n , a 2 + ω 2 L m , n , a 2 p n , a ω ψ M , n , a 2 R m , n , a ,
where p n , a p_{n,a} p n , a is the number of rotor teeth, ψ M , n , a \psi_{\mathrm{M},n,a} ψ M , n , a is the peak magnet flux, and V n , a V_{n,a} V n , a , R m , n , a R_{m,n,a} R m , n , a , and L m , n , a L_{m,n,a} L m , n , a are the motor voltage, phase resistance, and phase inductance, respectively. Applying P M , n , a ( v ) P_{\mathrm{M},n,a}(v) P M , n , a ( v ) independently to the two orthogonal actuators, the total movement energy is8
E m o v ( R , v ) = ∑ n = 1 N T n ( R , v n ) ∑ a ∈ A P M , n , a ( v n , a ( R , v n ) ) . (15) E_{\mathrm{mov}}(\mathbf{R},\mathbf{v})=\sum_{n=1}^{N}T_{n}(\mathbf{R},v_{n})\sum_{a\in\mathcal{A}}P_{\mathrm{M},n,a}(v_{n,a}(\mathbf{R},v_{n})). \tag*{(15)} E mov ( R , v ) = n = 1 ∑ N T n ( R , v n ) a ∈ A ∑ P M , n , a ( v n , a ( R , v n )) . (15)
The pattern-switching energy is E p a t ( Z ) = ε s w N s w ( Z ) E_{\mathrm{pat}}(\mathbf{Z})=\varepsilon_{\mathrm{sw}}N_{\mathrm{sw}}(\mathbf{Z}) E pat ( Z ) = ε sw N sw ( Z ) , where ε s w \varepsilon_{\mathrm{sw}} ε sw is the energy required to switch one MPRA state.
Remark 1. A stepper-motor-driven positioning admits only discrete locations determined by its axis-wise step sizes d s , a d_{s,a} d s , a [33 , 40 ] . Thus, the implementable position set of MPRA n n n is C n s t e p = C n ∩ { r n 0 + [ d s , x q x d s , y q y ] : [ q x q y ] T ∈ Z 2 } \mathcal{C}^{\mathrm{step}}_{n}=\mathcal{C}_{n}\cap\left\{\mathbf{r}^{0}_{n}+\left[\frac{d_{s,x}q_{x}}{d_{s,y}q_{y}}\right]:[q_{x}\ q_{y}]^{\mathrm{T}}\in\mathbb{Z}^{2}\right\} C n step = C n ∩ { r n 0 + [ d s , y q y d s , x q x ] : [ q x q y ] T ∈ Z 2 } . For tractability, we optimize R \mathbf{R} R over the continuous { C n } \{\mathcal{C}_{n}\} { C n } , providing a lower-bound on the implementable discrete-position performance. The resulting R ⋆ \mathbf{R}^{\star} R ⋆ can subsequently be mapped onto C n s t e p \mathcal{C}^{\mathrm{step}}_{n} C n step through the maximum-likelihood (ML) criterion.
At f c = 3.55 f_{c}=3.55 f c = 3.55 GHz, for instance, λ ≃ 84.5 \lambda\simeq84.5 λ ≃ 84.5 mm, and the step size d s = 1.2 d_{s}=1.2 d s = 1.2 mm from [33 ] is only 0.014 λ 0.014\lambda 0.014 λ , yielding a maximum two-dimensional error of d s / 2 ≃ 0.010 λ d_{s}/\sqrt{2}\simeq0.010\lambda d s / 2 ≃ 0.010 λ . Although such fine resolution suggests a minor quantization loss, independent rounding may violate the spacing constraint, necessitating the discrete-position selection [13 , 15 ] . We therefore leave a comprehensive fully-discrete MPRA design for future work; nevertheless, the close agreement with BFS in Section IV-B indicates that the fully-discrete framework is expected to incur only marginal performance degradation.
The data-transmit power is modeled as [27 , 28 , 41 ] :
P d a t a = ∥ F B B ∥ F 2 η P A ⏟ ≜ P P A + P L O + N R F ( 2 P D A C + P R F ) + N P P S , (16) P_{\mathrm{data}}=\underbrace{\frac{\lVert\mathbf{F}_{\mathrm{BB}}\rVert_{\mathrm{F}}^{2}}{\eta_{\mathrm{PA}}}}_{\triangleq P_{\mathrm{PA}}}+P_{\mathrm{LO}}+N_{\mathrm{RF}}(2P_{\mathrm{DAC}}+P_{\mathrm{RF}})+NP_{\mathrm{PS}}, \tag*{(16)} P data = ≜ P PA η PA ∥ F BB ∥ F 2 + P LO + N RF ( 2 P DAC + P RF ) + N P PS , (16)
where η P A ∈ ( 0 , 1 ] \eta_{\mathrm{PA}}\in(0,1] η PA ∈ ( 0 , 1 ] denotes the power-amplifier (PA) efficiency, and P L O P_{\mathrm{LO}} P LO , P P A P_{\mathrm{PA}} P PA , P D A C P_{\mathrm{DAC}} P DAC , P R F P_{\mathrm{RF}} P RF , and P P S P_{\mathrm{PS}} P PS denote the local-oscillator (LO), PA, per digital-to-analog converter (DAC), per-RF-chain, and per-phase-shifter circuit powers, respectively [27 Table II] . Hence, the total block energy is [33 ] :
E b l k = E m o v ( R , v ) + E p a t ( Z ) + T d ( R , Z , v ) P d a t a . (17) E_{\mathrm{blk}}=E_{\mathrm{mov}}(\mathbf{R},\mathbf{v})+E_{\mathrm{pat}}(\mathbf{Z})+T_{\mathrm{d}}(\mathbf{R},\mathbf{Z},\mathbf{v})P_{\mathrm{data}}. \tag*{(17)} E blk = E mov ( R , v ) + E pat ( Z ) + T d ( R , Z , v ) P data . (17)
For a system bandwidth W W W , the EE is
η E E = W T d ( R , Z , v ) R w E b l k [ b i t / J ] . (18) \eta_{\mathrm{EE}}=\frac{WT_{\mathrm{d}}(\mathbf{R},\mathbf{Z},\mathbf{v})R_{\mathrm{w}}}{E_{\mathrm{blk}}}\ [\mathrm{bit/J}]. \tag*{(18)} η EE = E blk W T d ( R , Z , v ) R w [ bit/J ] . (18)
We aim to maximize η E E \eta_{\mathrm{EE}} η EE within each coherence block by jointly designing X ≜ { F B B , F R F , R , Z , v } \mathcal{X} \triangleq\{\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}},\mathbf{R},\mathbf{Z},\mathbf{v}\} X ≜ { F BB , F RF , R , Z , v } :
max X ≜ { F B B , F R F , R , Z , v } η E E ( X ) s.t. ( 3 ) , F R F ∈ F R F , z n ∈ { 0 , 1 } S , 1 S T z n = 1 , ( 4 ) v n ∈ V n ( R ) ( ∀ n ∈ N ) , T o h ( R , Z , v ) ≤ T c − ϵ T , (19) \begin{aligned}
\max_{\mathcal{X}\triangleq\{\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}},\mathbf{R},\mathbf{Z},\mathbf{v}\}}\quad&\eta_{\mathrm{EE}}(\mathcal{X})\\
\text{s.t.}\quad&(3),\ \mathbf{F}_{\mathrm{RF}}\in\mathcal{F}_{\mathrm{RF}},\ z_n\in\{0,1\}^{S},\ \mathbf{1}_S^{\mathsf{T}}\mathbf{z}_n=1,\ (4)\\
&v_n\in\mathcal{V}_n(\mathbf{R})\ (\forall n\in\mathcal{N}),\ T_{\mathrm{oh}}(\mathbf{R},\mathbf{Z},\mathbf{v})\leq T_c-\epsilon_T, \tag*{(19)}
\end{aligned} X ≜ { F BB , F RF , R , Z , v } max s.t. η EE ( X ) ( 3 ) , F RF ∈ F RF , z n ∈ { 0 , 1 } S , 1 S T z n = 1 , ( 4 ) v n ∈ V n ( R ) ( ∀ n ∈ N ) , T oh ( R , Z , v ) ≤ T c − ϵ T , (19)
where ϵ T > 0 \epsilon_T>0 ϵ T > 0 guarantees a strictly positive T d T_d T d and
V n ( R ) ≜ { { 0 } , ( D n ( R ) = 0 ) ( 0 , v ‾ n ( R ) ] , ( D n ( R ) > 0 ) . \mathcal{V}_n(\mathbf{R})\triangleq
\begin{cases}
\{0\}, & (D_n(\mathbf{R})=0)\\
(0,\overline{v}_n(\mathbf{R})], & (D_n(\mathbf{R})>0).
\end{cases} V n ( R ) ≜ { { 0 } , ( 0 , v n ( R )] , ( D n ( R ) = 0 ) ( D n ( R ) > 0 ) .
Problem (19) is a mixed-integer fractional programming, which is nonconvex due to the fractional objective involving both continuous and discrete variables, as well as their nonsmooth coupling through the reconfiguration-time and mechanical-energy models. Therefore, we develop an AO framework for solving (19).
Proposed EE Optimization Framework
Optimal { v n , a } \{v_{n,a}\} { v n , a } of MPRA
We first examine { v n , a } \{v_{n,a}\} { v n , a } for fixed ( F B B , F R F , R , Z ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}},\mathbf{R},\mathbf{Z}) ( F BB , F RF , R , Z ) . It is known that from [33 ] , M n , a ( ω ) \mathcal{M}_{n,a}(\omega) M n , a ( ω ) is nonnegative and non-increasing over 0 ≤ ω ≤ ω n , a max 0\leq\omega\leq\omega_{n,a}^{\max} 0 ≤ ω ≤ ω n , a m a x , where ω n , a max = v n , a max ℓ 0 , n , a \omega_{n,a}^{\max}=\frac{v_{n,a}^{\max}}{\ell_{0,n,a}} ω n , a m a x = ℓ 0 , n , a v n , a m a x . Thereafter, we can consider the following theorem:
Theorem 1. For every fixed ( F B B , F R F , R , Z ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}},\mathbf{R},\mathbf{Z}) ( F BB , F RF , R , Z ) , η E E \eta_{\mathrm{EE}} η EE is non-decreasing with respect to each v n v_n v n over V n ( R ) \mathcal{V}_n(\mathbf{R}) V n ( R ) :
v n ⋆ ( R ) = { v ‾ n ( R ) ( D n ( R ) > 0 ) 0 ( D n ( R ) = 0 ) ( n ∈ N ) . (20) v_n^\star(\mathbf{R})=
\begin{cases}
\overline{v}_n(\mathbf{R}) & (D_n(\mathbf{R})>0)\\
0 & (D_n(\mathbf{R})=0)
\end{cases}
\quad(n\in\mathcal{N}). \tag*{(20)} v n ⋆ ( R ) = { v n ( R ) 0 ( D n ( R ) > 0 ) ( D n ( R ) = 0 ) ( n ∈ N ) . (20)
Proof. Consider an MPRA n n n satisfying D n ( R ) > 0 D_n(\mathbf{R})>0 D n ( R ) > 0 . The contribution E m o v , n ( R , v n ) E_{\mathrm{mov},n}(\mathbf{R},v_n) E mov , n ( R , v n ) of MPRA n n n in (15) is
E m o v , n ( R , v n ) ≜ T n ( R , v n ) ∑ a ∈ A P M , n , a ( v n , a ( R , v n ) ) = ∑ a ∈ A : d n , a ( R ) > 0 d n , a ( R ) ℓ 0 , n , a M n , a ( v n d n , a ( R ) D n ( R ) ℓ 0 , n , a ) . (21) \begin{aligned}
E_{\mathrm{mov},n}(\mathbf{R},v_n)&\triangleq T_n(\mathbf{R},v_n)\sum_{a\in\mathcal{A}}P_{\mathrm{M},n,a}(v_{n,a}(\mathbf{R},v_n))\\
&=\sum_{\substack{a\in\mathcal{A}:\\d_{n,a}(\mathbf{R})>0}}\frac{d_{n,a}(\mathbf{R})}{\ell_{0,n,a}}\mathcal{M}_{n,a}\left(\frac{v_nd_{n,a}(\mathbf{R})}{D_n(\mathbf{R})\ell_{0,n,a}}\right). \tag*{(21)}
\end{aligned} E mov , n ( R , v n ) ≜ T n ( R , v n ) a ∈ A ∑ P M , n , a ( v n , a ( R , v n )) = a ∈ A : d n , a ( R ) > 0 ∑ ℓ 0 , n , a d n , a ( R ) M n , a ( D n ( R ) ℓ 0 , n , a v n d n , a ( R ) ) . (21)
Its derivative with respect to v n v_n v n is
∂ E m o v , n ∂ v n = ∑ a ∈ A : d n , a ( R ) > 0 d n , a 2 ( R ) D n ( R ) ℓ 0 , n , a 2 M n , a ′ ( v n d n , a ( R ) D n ( R ) ℓ 0 , n , a ) ≤ 0 , (22) \frac{\partial E_{\mathrm{mov},n}}{\partial v_n}
=\sum_{\substack{a\in\mathcal{A}:\\d_{n,a}(\mathbf{R})>0}}
\frac{d_{n,a}^{2}(\mathbf{R})}{D_n(\mathbf{R})\ell_{0,n,a}^{2}}
\mathcal{M}_{n,a}'\left(\frac{v_nd_{n,a}(\mathbf{R})}{D_n(\mathbf{R})\ell_{0,n,a}}\right)\leq0, \tag*{(22)} ∂ v n ∂ E mov , n = a ∈ A : d n , a ( R ) > 0 ∑ D n ( R ) ℓ 0 , n , a 2 d n , a 2 ( R ) M n , a ′ ( D n ( R ) ℓ 0 , n , a v n d n , a ( R ) ) ≤ 0 , (22)
since M n , a \mathcal{M}_{n,a} M n , a is non-increasing.
Moreover, T n ( R , v n ) = D n ( R ) v n T_n(\mathbf{R},v_n)=\frac{D_n(\mathbf{R})}{v_n} T n ( R , v n ) = v n D n ( R ) is decreasing in v n v_n v n . Thus, increasing v n v_n v n cannot grow T m o v ( R , v ) = max m ∈ N T m ( R , v m ) T_{\mathrm{mov}}(\mathbf{R},\mathbf{v})=\max_{m\in\mathcal{N}}T_m(\mathbf{R},v_m) T mov ( R , v ) = max m ∈ N T m ( R , v m ) , and thus cannot decrease T d T_d T d . For fixed variables, R w R_w R w , P d a t a P_{\mathrm{data}} P data , T p a t T_{\mathrm{pat}} T pat , and E p a t E_{\mathrm{pat}} E pat are independent of v \mathbf{v} v . Dividing the numerator and denominator of (18) by T d > 0 T_d>0 T d > 0 :
η E E = W R w P d a t a + E m o v + E p a t T d . (23) \eta_{\mathrm{EE}}=\frac{W R_w}{P_{\mathrm{data}}+\frac{E_{\mathrm{mov}}+E_{\mathrm{pat}}}{T_d}}. \tag*{(23)} η EE = P data + T d E mov + E pat W R w . (23)
Since increasing v n v_n v n cannot increase E m o v + E p a t E_{\mathrm{mov}}+E_{\mathrm{pat}} E mov + E pat and cannot decrease T d T_d T d , it cannot increase the denominator. Hence, η E E \eta_{\mathrm{EE}} η EE is non-decreasing in v n v_n v n , and its maximum over V n ( R ) \mathcal{V}_n(\mathbf{R}) V n ( R ) is attained at (20). For D n ( R ) = 0 D_n(\mathbf{R})=0 D n ( R ) = 0 , it follows from V n ( R ) = { 0 } \mathcal{V}_n(\mathbf{R})=\{0\} V n ( R ) = { 0 } . ■
Theorem 1 extends the result in [33 ] to independently controlled MPRAs undergoing synchronized biaxial motion. By defining v ^ ( R ) ≜ [ v 1 ⋆ ( R ) ⋯ v N ⋆ ( R ) ] T \widehat{\mathbf{v}}(\mathbf{R})\triangleq[v_1^\star(\mathbf{R})\cdots v_N^\star(\mathbf{R})]^{\mathsf{T}} v ( R ) ≜ [ v 1 ⋆ ( R ) ⋯ v N ⋆ ( R ) ] T hereafter, v \mathbf{v} v is set to v = v ^ ( R ) \mathbf{v}=\widehat{\mathbf{v}}(\mathbf{R}) v = v ( R ) , and we use T ^ d ( R , Z ) ≜ T d ( R , Z , v ^ ( R ) ) \widehat{T}_d(\mathbf{R},\mathbf{Z})\triangleq T_d(\mathbf{R},\mathbf{Z},\widehat{\mathbf{v}}(\mathbf{R})) T d ( R , Z ) ≜ T d ( R , Z , v ( R )) and E ^ r e c ( R , Z ) ≜ E m o v ( R , v ^ ( R ) ) + E p a t ( Z ) \widehat{E}_{\mathrm{rec}}(\mathbf{R},\mathbf{Z})\triangleq E_{\mathrm{mov}}(\mathbf{R},\widehat{\mathbf{v}}(\mathbf{R}))+E_{\mathrm{pat}}(\mathbf{Z}) E rec ( R , Z ) ≜ E mov ( R , v ( R )) + E pat ( Z ) to denote the resulting T d T_d T d and total reconfiguration energy, respectively. Accordingly, η E E \eta_{\mathrm{EE}} η EE becomes η E E = W T ^ d ( R , Z ) R w E ^ r e c ( R , Z ) + T ^ d ( R , Z ) P d a t a \eta_{\mathrm{EE}}=\frac{W\widehat{T}_d(\mathbf{R},\mathbf{Z})R_w}{\widehat{E}_{\mathrm{rec}}(\mathbf{R},\mathbf{Z})+\widehat{T}_d(\mathbf{R},\mathbf{Z})P_{\mathrm{data}}} η EE = E rec ( R , Z ) + T d ( R , Z ) P data W T d ( R , Z ) R w .
Dinkelbach-WMMSE Optimization of ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF )
We next optimize ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) for fixed ( R , Z ) (\mathbf{R},\mathbf{Z}) ( R , Z ) . For brevity, define N ( F B B , F R F ) ≜ W T ^ d R w N(\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}})\triangleq W\widehat{T}_dR_w N ( F BB , F RF ) ≜ W T d R w and D ( F B B , F R F ) ≜ E ^ r e c + T ^ d P d a t a D(\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}})\triangleq\widehat{E}_{\mathrm{rec}}+\widehat{T}_dP_{\mathrm{data}} D ( F BB , F RF ) ≜ E rec + T d P data , where the dependence of T ^ d \widehat{T}_d T d and E ^ r e c \widehat{E}_{\mathrm{rec}} E rec on ( R , Z ) (\mathbf{R},\mathbf{Z}) ( R , Z ) is temporarily suppressed since ( R , Z ) (\mathbf{R},\mathbf{Z}) ( R , Z ) are fixed. For a Dinkelbach parameter λ D ≥ 0 \lambda_D\geq0 λ D ≥ 0 , consider J λ D ≜ N − λ D D \mathcal{J}_{\lambda_D}\triangleq N-\lambda_DD J λ D ≜ N − λ D D . The classical Dinkelbach characterization states that the globally optimal EE η E E , D ⋆ \eta^\star_{\mathrm{EE},D} η EE , D ⋆ satisfies max F B B , F R F { N − η E E , D ⋆ D } = 0 \max_{\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}}\{N-\eta^\star_{\mathrm{EE},D}D\}=0 max F BB , F RF { N − η EE , D ⋆ D } = 0 over feasible ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) [42 ] . Since T ^ d > 0 \widehat{T}_d>0 T d > 0 is constant herein, maximizing J λ D \mathcal{J}_{\lambda_D} J λ D is equivalent to maximizing
W R w − λ D P d a t a , (24) W R_w-\lambda_D P_{\mathrm{data}}, \tag*{(24)} W R w − λ D P data , (24)
since − λ D E ^ r e c / T ^ d -\lambda_D\widehat{E}_{\mathrm{rec}}/\widehat{T}_d − λ D E rec / T d is independent of ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) .
To handle the interference-coupled R w R_w R w , define g k ⋆ ≜ h k , s e l ⋆ ( R ) F s e l ( Z ) F R F ∈ C 1 × N R F \mathbf{g}_k^\star\triangleq\mathbf{h}_{k,\mathrm{sel}}^\star(\mathbf{R})\mathbf{F}_{\mathrm{sel}}(\mathbf{Z})\mathbf{F}_{\mathrm{RF}}\in\mathbb{C}^{1\times N_{\mathrm{RF}}} g k ⋆ ≜ h k , sel ⋆ ( R ) F sel ( Z ) F RF ∈ C 1 × N RF . For equalizer u k ∈ C u_k\in\mathbb{C} u k ∈ C , the estimate of s k s_k s k is s ^ k = u k ⋆ y k \widehat{s}_k=u_k^\star y_k s k = u k ⋆ y k , and the corresponding MSE is
e k ≜ E [ ∣ u k ⋆ y k − s k ∣ 2 ] = ∣ u k ∣ 2 ( ∑ j = 1 K ∣ g k ⋆ f j ∣ 2 + σ k 2 ) − 2 ℜ { u k ⋆ g k ⋆ f k } + 1. (25) \begin{aligned}
e_k\triangleq\mathbb{E}\left[\left|u_k^\star y_k-s_k\right|^2\right] \\
&=|u_k|^2\left(\sum_{j=1}^{K}|\mathbf{g}_k^\star\mathbf{f}_j|^2+\sigma_k^2\right)-2\Re\{u_k^\star\mathbf{g}_k^\star\mathbf{f}_k\}+1. \tag*{(25)}
\end{aligned} e k ≜ E [ ∣ u k ⋆ y k − s k ∣ 2 ] = ∣ u k ∣ 2 ( j = 1 ∑ K ∣ g k ⋆ f j ∣ 2 + σ k 2 ) − 2ℜ { u k ⋆ g k ⋆ f k } + 1. (25)
Introducing an auxiliary MSE weight ϖ k > 0 \varpi_k>0 ϖ k > 0 , the WMMSE identity for SE gives [19 ] :
R w = 1 ln 2 max { u k } , { ϖ k > 0 } ∑ k = 1 K ( ln ϖ k − ϖ k e k + 1 ) . (26) R_w=\frac{1}{\ln2}\max_{\{u_k\},\{\varpi_k>0\}}\sum_{k=1}^{K}(\ln\varpi_k-\varpi_ke_k+1). \tag*{(26)} R w = ln 2 1 { u k } , { ϖ k > 0 } max k = 1 ∑ K ( ln ϖ k − ϖ k e k + 1 ) . (26)
Hence, substituting (26) into (24):
max F B B , F R F , { u k ∈ C , ϖ k > 0 } k = 1 K W ln 2 ∑ k = 1 K ( ln ϖ k − ϖ k e k + 1 ) − λ D η P A ∥ F B B ∥ F 2 s.t. ∥ F B B ∥ F 2 ≤ P max , F R F ∈ F R F . (27) \begin{aligned}
\max_{\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}},\{u_k\in\mathbb{C},\varpi_k>0\}_{k=1}^{K}}\quad&
\frac{W}{\ln2}\sum_{k=1}^{K}(\ln\varpi_k-\varpi_ke_k+1)-\frac{\lambda_D}{\eta_{\mathrm{PA}}}\|\mathbf{F}_{\mathrm{BB}}\|_F^2\\
\text{s.t.}\quad&\|\mathbf{F}_{\mathrm{BB}}\|_F^2\leq P_{\max},\quad\mathbf{F}_{\mathrm{RF}}\in\mathcal{F}_{\mathrm{RF}}. \tag*{(27)}
\end{aligned} F BB , F RF , { u k ∈ C , ϖ k > 0 } k = 1 K max s.t. ln 2 W k = 1 ∑ K ( ln ϖ k − ϖ k e k + 1 ) − η PA λ D ∥ F BB ∥ F 2 ∥ F BB ∥ F 2 ≤ P m a x , F RF ∈ F RF . (27)
Update of { u k , ϖ k } \{u_k,\varpi_k\} { u k , ϖ k } : For fixed ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) , the optimal auxiliary variables of (27) are given by
u k ⋆ = g k ⋆ f k ∑ j = 1 K ∣ g k ⋆ f j ∣ 2 + σ k 2 , ϖ k ⋆ = ( e k M M S E ) − 1 = 1 + γ k . (28) u_k^\star=\frac{\mathbf{g}_k^\star\mathbf{f}_k}{\sum_{j=1}^{K}|\mathbf{g}_k^\star\mathbf{f}_j|^2+\sigma_k^2},\quad
\varpi_k^\star=(e_k^{\mathrm{MMSE}})^{-1}=1+\gamma_k. \tag*{(28)} u k ⋆ = ∑ j = 1 K ∣ g k ⋆ f j ∣ 2 + σ k 2 g k ⋆ f k , ϖ k ⋆ = ( e k MMSE ) − 1 = 1 + γ k . (28)
Update of F B B \mathbf{F}_{\mathrm{BB}} F BB : For fixed ( F R F , { u k , ϖ k } ) (\mathbf{F}_{\mathrm{RF}},\{u_k,\varpi_k\}) ( F RF , { u k , ϖ k }) , define A ≜ ∑ k = 1 K ϖ k ∣ u k ∣ 2 g k g k ⋆ ⪰ 0 \mathbf{A}\triangleq\sum_{k=1}^{K}\varpi_k|u_k|^2\mathbf{g}_k\mathbf{g}_k^\star\succeq0 A ≜ ∑ k = 1 K ϖ k ∣ u k ∣ 2 g k g k ⋆ ⪰ 0 , B ≜ [ ϖ 1 u 1 g 1 ⋯ ϖ K u K g K ] \mathbf{B}\triangleq[\varpi_1u_1\mathbf{g}_1\cdots\varpi_Ku_K\mathbf{g}_K] B ≜ [ ϖ 1 u 1 g 1 ⋯ ϖ K u K g K ] , and ρ D ≜ λ D ln 2 W η P A \rho_D\triangleq\frac{\lambda_D\ln2}{W\eta_{\mathrm{PA}}} ρ D ≜ W η PA λ D l n 2 . The F B B \mathbf{F}_{\mathrm{BB}} F BB -subproblem then becomes
min F B B tr [ F B B ⋆ ( A + ρ D I ) F B B ] − 2 ℜ { tr ( B ⋆ F B B ) } s.t. ∥ F B B ∥ F 2 ≤ P max . (29) \begin{aligned}
\min_{\mathbf{F}_{\mathrm{BB}}}\quad&
\operatorname{tr}[\mathbf{F}_{\mathrm{BB}}^\star(\mathbf{A}+\rho_D\mathbf{I})\mathbf{F}_{\mathrm{BB}}]-2\Re\{\operatorname{tr}(\mathbf{B}^\star\mathbf{F}_{\mathrm{BB}})\}\\
\text{s.t.}\quad&\|\mathbf{F}_{\mathrm{BB}}\|_F^2\leq P_{\max}. \tag*{(29)}
\end{aligned} F BB min s.t. tr [ F BB ⋆ ( A + ρ D I ) F BB ] − 2ℜ { tr ( B ⋆ F BB )} ∥ F BB ∥ F 2 ≤ P m a x . (29)
Problem (29) is a convex quadratic program with its solution:
F B B ⋆ ( μ ) = [ A + ( ρ D + μ ) I ] − 1 B , (30) \mathbf{F}_{\mathrm{BB}}^\star(\mu)=[\mathbf{A}+(\rho_D+\mu)\mathbf{I}]^{-1}\mathbf{B}, \tag*{(30)} F BB ⋆ ( μ ) = [ A + ( ρ D + μ ) I ] − 1 B , (30)
where μ ≥ 0 \mu\ge0 μ ≥ 0 is chosen to satisfy μ ( ∥ F B B ⋆ ( μ ) ∥ F 2 − P max ) = 0 \mu\left(\left\|\mathbf{F}_{\mathrm{BB}}^{\star}(\mu)\right\|_{F}^{2}-P_{\max}\right)=0 μ ( ∥ F BB ⋆ ( μ ) ∥ F 2 − P m a x ) = 0 . If the unconstrained solution at μ = 0 \mu=0 μ = 0 violates the power budget, the unique μ > 0 \mu>0 μ > 0 satisfying ∥ F B B ⋆ ( μ ) ∥ F 2 = P max \left\|\mathbf{F}_{\mathrm{BB}}^{\star}(\mu)\right\|_{F}^{2}=P_{\max} ∥ F BB ⋆ ( μ ) ∥ F 2 = P m a x can be found by scalar bisection.
3) Update of F R F \mathbf{F}_{\mathrm{RF}} F RF : We next optimize F R F \mathbf{F}_{\mathrm{RF}} F RF for fixed ( F B B , { u k , ϖ k } ) (\mathbf{F}_{\mathrm{BB}},\{u_k,\varpi_k\}) ( F BB , { u k , ϖ k }) . Let r ( n ) r(n) r ( n ) denote the unique RF chain connected to MPRA n n n , i.e., n ∈ N r ( n ) n\in\mathcal{N}_{r(n)} n ∈ N r ( n ) , and collect the N N N active analog coefficients into q ≜ M [ [ F R F ] 1 , r ( 1 ) ⋯ [ F R F ] N , r ( N ) ] T \mathbf{q}\triangleq\sqrt{M}[[\mathbf{F}_{\mathrm{RF}}]_{1,r(1)}\ \cdots\ [\mathbf{F}_{\mathrm{RF}}]_{N,r(N)}]^{\mathrm{T}} q ≜ M [[ F RF ] 1 , r ( 1 ) ⋯ [ F RF ] N , r ( N ) ] T . The fixed zero entries of F R F \mathbf{F}_{\mathrm{RF}} F RF are excluded from q \mathbf{q} q . By (2), the feasible set of q \mathbf{q} q is the product complex-circle manifold M c c N ≜ { q ∈ C N : ∣ q n ∣ = 1 ( ∀ n ∈ N ) } \mathcal{M}_{\mathrm{cc}}^{N}\triangleq\{\mathbf{q}\in\mathbb{C}^{N}:|q_n|=1\;(\forall n\in\mathcal{N})\} M cc N ≜ { q ∈ C N : ∣ q n ∣ = 1 ( ∀ n ∈ N )} . The mapping from q \mathbf{q} q to F R F \mathbf{F}_{\mathrm{RF}} F RF is denoted by F R F ( q ) \mathbf{F}_{\mathrm{RF}}(\mathbf{q}) F RF ( q ) , whose entries satisfy
[ F R F ( q ) ] n , r = { q n M ( r = r ( n ) ) 0 ( r ≠ r ( n ) ) . [\mathbf{F}_{\mathrm{RF}}(\mathbf{q})]_{n,r}=
\begin{cases}
\dfrac{q_n}{\sqrt{M}} & (r=r(n))\\
0 & (r\ne r(n)).
\end{cases} [ F RF ( q ) ] n , r = ⎩ ⎨ ⎧ M q n 0 ( r = r ( n )) ( r = r ( n )) .
For notational convenience, define [ c k , j ] n ≜ [ h k ( R , Z ) ] n [ f j ] r ( n ) M [\mathbf{c}_{k,j}]_n\triangleq\dfrac{[\mathbf{h}_k(\mathbf{R},\mathbf{Z})]_n[\mathbf{f}_j]_{r(n)}}{\sqrt{M}} [ c k , j ] n ≜ M [ h k ( R , Z ) ] n [ f j ] r ( n ) ( n ∈ N ) (n\in\mathcal{N}) ( n ∈ N ) . It then follows that h k ∗ ( R , Z ) F R F ( q ) f j = c k , j ∗ q \mathbf{h}_k^{\ast}(\mathbf{R},\mathbf{Z})\mathbf{F}_{\mathrm{RF}}(\mathbf{q})\mathbf{f}_j=\mathbf{c}_{k,j}^{\ast}\mathbf{q} h k ∗ ( R , Z ) F RF ( q ) f j = c k , j ∗ q . Thus, the terms of the Dinkelbach-WMMSE problem that depend on q \mathbf{q} q can then be:
min q ∈ M c c N f R F ( q ) ≜ q ∗ C q − 2 ℜ { d ∗ q } , (31) \min_{\mathbf{q}\in\mathcal{M}_{\mathrm{cc}}^{N}} f_{\mathrm{RF}}(\mathbf{q})\triangleq\mathbf{q}^{\ast}\mathbf{C}\mathbf{q}-2\Re\{\mathbf{d}^{\ast}\mathbf{q}\}, \tag*{(31)} q ∈ M cc N min f RF ( q ) ≜ q ∗ Cq − 2ℜ { d ∗ q } , (31)
where C ≜ ∑ k = 1 K ϖ k ∣ u k ∣ 2 ∑ j = 1 K c k , j c k , j ∗ ⪰ 0 \mathbf{C}\triangleq\sum_{k=1}^{K}\varpi_k|u_k|^{2}\sum_{j=1}^{K}\mathbf{c}_{k,j}\mathbf{c}_{k,j}^{\ast}\succeq0 C ≜ ∑ k = 1 K ϖ k ∣ u k ∣ 2 ∑ j = 1 K c k , j c k , j ∗ ⪰ 0 and d ≜ ∑ k = 1 K ϖ k u k c k , k \mathbf{d}\triangleq\sum_{k=1}^{K}\varpi_ku_k\mathbf{c}_{k,k} d ≜ ∑ k = 1 K ϖ k u k c k , k . Problem (31) can therefore be addressed by an RCG method with Armijo backtracking [25 , 43 ] . It generates a non-increasing and converging sequence of { q } \{\mathbf{q}\} { q } and terminates when ∥ grad f R F ( q ) ∥ 2 ≤ ϵ R F \|\operatorname{grad}f_{\mathrm{RF}}(\mathbf{q})\|_{2}\le\epsilon_{\mathrm{RF}} ∥ grad f RF ( q ) ∥ 2 ≤ ϵ RF with Riemannian gradient grad \operatorname{grad} grad [25 ] . The resulting F R F \mathbf{F}_{\mathrm{RF}} F RF is then F R F = F R F ( q ) \mathbf{F}_{\mathrm{RF}}=\mathbf{F}_{\mathrm{RF}}(\mathbf{q}) F RF = F RF ( q ) .
4) Dinkelbach-Parameter Update and Termination: Let i ≥ 0 i\ge0 i ≥ 0 denote the Dinkelbach iteration index and initialize λ D ( 0 ) = N ( F B B ( 0 ) , F R F ( 0 ) ) D ( F B B ( 0 ) , F R F ( 0 ) ) \lambda_{\mathrm{D}}^{(0)}=\dfrac{N(\mathbf{F}_{\mathrm{BB}}^{(0)},\mathbf{F}_{\mathrm{RF}}^{(0)})}{D(\mathbf{F}_{\mathrm{BB}}^{(0)},\mathbf{F}_{\mathrm{RF}}^{(0)})} λ D ( 0 ) = D ( F BB ( 0 ) , F RF ( 0 ) ) N ( F BB ( 0 ) , F RF ( 0 ) ) . For a fixed λ D ( i ) \lambda_{\mathrm{D}}^{(i)} λ D ( i ) , we alternately update { u k ( i ) , ϖ k ( i ) } k = 1 K \{u_k^{(i)},\varpi_k^{(i)}\}_{k=1}^{K} { u k ( i ) , ϖ k ( i ) } k = 1 K , F B B ( i ) \mathbf{F}_{\mathrm{BB}}^{(i)} F BB ( i ) , and F R F ( i ) \mathbf{F}_{\mathrm{RF}}^{(i)} F RF ( i ) according to (28), (30), and the RCG procedure in Section III-B3, respectively. Let ( F B B ( i + 1 ) , F R F ( i + 1 ) ) (\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)}) ( F BB ( i + 1 ) , F RF ( i + 1 ) ) denote the resulting feasible hybrid precoders. The associated Dinkelbach residual is J λ D ( i ) ( F B B ( i + 1 ) , F R F ( i + 1 ) ) = N ( F B B ( i + 1 ) , F R F ( i + 1 ) ) − λ D ( i ) D ( F B B ( i + 1 ) , F R F ( i + 1 ) ) \mathcal{J}_{\lambda_{\mathrm{D}}^{(i)}}(\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)})=N(\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)})-\lambda_{\mathrm{D}}^{(i)}D(\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)}) J λ D ( i ) ( F BB ( i + 1 ) , F RF ( i + 1 ) ) = N ( F BB ( i + 1 ) , F RF ( i + 1 ) ) − λ D ( i ) D ( F BB ( i + 1 ) , F RF ( i + 1 ) ) . Thereafter, λ D ( i + 1 ) \lambda_{\mathrm{D}}^{(i+1)} λ D ( i + 1 ) is updated:
λ D ( i + 1 ) = N ( F B B ( i + 1 ) , F R F ( i + 1 ) ) D ( F B B ( i + 1 ) , F R F ( i + 1 ) ) . \lambda_{\mathrm{D}}^{(i+1)}=\frac{N(\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)})}{D(\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)})}. λ D ( i + 1 ) = D ( F BB ( i + 1 ) , F RF ( i + 1 ) ) N ( F BB ( i + 1 ) , F RF ( i + 1 ) ) .
The overall Dinkelbach routine terminates when ∣ λ D ( i + 1 ) − λ D ( i ) ∣ ≤ ϵ D |\lambda_{\mathrm{D}}^{(i+1)}-\lambda_{\mathrm{D}}^{(i)}|\le\epsilon_{\mathrm{D}} ∣ λ D ( i + 1 ) − λ D ( i ) ∣ ≤ ϵ D for given tolerance ϵ D > 0 \epsilon_{\mathrm{D}}>0 ϵ D > 0 . Since { u k ( i ) , ϖ k ( i ) } k = 1 K \{u_k^{(i)},\varpi_k^{(i)}\}_{k=1}^{K} { u k ( i ) , ϖ k ( i ) } k = 1 K and F B B \mathbf{F}_{\mathrm{BB}} F BB -updates optimize their respective blocks and the Armijo RCG update does not decrease the objective [25 ] , the resulting ( F B B ( i + 1 ) , F R F ( i + 1 ) ) (\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)}) ( F BB ( i + 1 ) , F RF ( i + 1 ) ) satisfies
J λ D ( i ) ( F B B ( i + 1 ) , F R F ( i + 1 ) ) ≥ J λ D ( i ) ( F B B ( i ) , F R F ( i ) ) = 0 , (32) \mathcal{J}_{\lambda_{\mathrm{D}}^{(i)}}(\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)})\ge\mathcal{J}_{\lambda_{\mathrm{D}}^{(i)}}(\mathbf{F}_{\mathrm{BB}}^{(i)},\mathbf{F}_{\mathrm{RF}}^{(i)})=0, \tag*{(32)} J λ D ( i ) ( F BB ( i + 1 ) , F RF ( i + 1 ) ) ≥ J λ D ( i ) ( F BB ( i ) , F RF ( i ) ) = 0 , (32)
where the right-side equality follows from λ D ( i ) = N ( F B B ( i ) , F R F ( i ) ) D ( F B B ( i ) , F R F ( i ) ) \lambda_{\mathrm{D}}^{(i)}=\dfrac{N(\mathbf{F}_{\mathrm{BB}}^{(i)},\mathbf{F}_{\mathrm{RF}}^{(i)})}{D(\mathbf{F}_{\mathrm{BB}}^{(i)},\mathbf{F}_{\mathrm{RF}}^{(i)})} λ D ( i ) = D ( F BB ( i ) , F RF ( i ) ) N ( F BB ( i ) , F RF ( i ) ) .
λ D ( i + 1 ) − λ D ( i ) = J λ D ( i ) ( F B B ( i + 1 ) , F R F ( i + 1 ) ) D ( F B B ( i + 1 ) , F R F ( i + 1 ) ) ≥ 0 , (33) \lambda_{\mathrm{D}}^{(i+1)}-\lambda_{\mathrm{D}}^{(i)}
=\frac{\mathcal{J}_{\lambda_{\mathrm{D}}^{(i)}}(\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)})}{D(\mathbf{F}_{\mathrm{BB}}^{(i+1)},\mathbf{F}_{\mathrm{RF}}^{(i+1)})}\ge0, \tag*{(33)} λ D ( i + 1 ) − λ D ( i ) = D ( F BB ( i + 1 ) , F RF ( i + 1 ) ) J λ D ( i ) ( F BB ( i + 1 ) , F RF ( i + 1 ) ) ≥ 0 , (33)
and EE generated by ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) -block is nondecreasing.
Switching-Aware Iterative Optimization of Z \mathbf{Z} Z
We then optimize Z \mathbf{Z} Z for fixed R \mathbf{R} R and ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) . Let R ‾ \overline{\mathbf{R}} R denote the position configuration held fixed during this Z \mathbf{Z} Z -update block. Unlike exhaustive search over all S N S^{N} S N configurations, we successively update one MPRA pattern at a time using a cyclic candidate-testing procedure inspired by [28 ] .
1) Candidate Pattern Configuration and Channel Update: Let Z [ p ] ≜ [ z 1 [ p ] ⋯ z N [ p ] ] \mathbf{Z}^{[p]}\triangleq[\mathbf{z}_{1}^{[p]}\cdots\mathbf{z}_{N}^{[p]}] Z [ p ] ≜ [ z 1 [ p ] ⋯ z N [ p ] ] denote the pattern configuration after p p p th pattern-coordinate iteration, and let s n [ p ] s_{n}^{[p]} s n [ p ] be the unique index satisfying z n [ p ] = e s n [ p ] \mathbf{z}_{n}^{[p]}=\mathbf{e}_{s_{n}^{[p]}} z n [ p ] = e s n [ p ] . For fixed r n r_n r n , define the channel column associated with MPRA n n n operating in s s s :
h n , s ( r ‾ n ) ≜ [ χ 1 , n , s ( r n ) ⋮ χ K , n , s ( r n ) ] ∈ C K , \mathbf{h}_{n,s}(\overline{\mathbf{r}}_n)\triangleq
\begin{bmatrix}
\chi_{1,n,s}(r_n)\\
\vdots\\
\chi_{K,n,s}(r_n)
\end{bmatrix}\in\mathbb{C}^{K}, h n , s ( r n ) ≜ χ 1 , n , s ( r n ) ⋮ χ K , n , s ( r n ) ∈ C K ,
where
χ k , n , s ( r n ) = β k L k ∑ ℓ = 1 L k α k , ℓ ψ ‾ s ( Ω k , ℓ ) e − j κ k , ℓ T r n , \chi_{k,n,s}(r_n)=\sqrt{\frac{\beta_k}{L_k}}\sum_{\ell=1}^{L_k}\alpha_{k,\ell}\overline{\psi}_{s}(\boldsymbol{\Omega}_{k,\ell})e^{-\mathrm{j}\boldsymbol{\kappa}_{k,\ell}^{\mathrm{T}}\mathbf{r}_n}, χ k , n , s ( r n ) = L k β k ℓ = 1 ∑ L k α k , ℓ ψ s ( Ω k , ℓ ) e − j κ k , ℓ T r n ,
with \boldsymbol{\kappa}_{k,\ell}\triangleq k_0\begin{bmatrix}[\mathbf{u}(\boldsymbol{\Omega}_{k,\ell})]_x\\[\mathbf{u}(\boldsymbol{\Omega}_{k,\ell})]_y\end{bmatrix}\in\mathbb{R}^{2} is the channel contribution of MPRA n n n toward user k k k . For Z = [ e s 1 ⋯ e s N ] \mathbf{Z}=[\mathbf{e}_{s_1}\cdots\mathbf{e}_{s_N}] Z = [ e s 1 ⋯ e s N ] , H e f f \mathbf{H}_{\mathrm{eff}} H eff can be expressed:
H e f f ( R ‾ , Z ) = [ h 1 , s 1 ( r ‾ 1 ) ⋯ h N , s N ( r ‾ N ) ] ∈ C K × N . (34) \mathbf{H}_{\mathrm{eff}}(\overline{\mathbf{R}},\mathbf{Z})=[\mathbf{h}_{1,s_1}(\overline{\mathbf{r}}_1)\cdots\mathbf{h}_{N,s_N}(\overline{\mathbf{r}}_N)]\in\mathbb{C}^{K\times N}. \tag*{(34)} H eff ( R , Z ) = [ h 1 , s 1 ( r 1 ) ⋯ h N , s N ( r N )] ∈ C K × N . (34)
At p ≥ 1 p\ge1 p ≥ 1 , the MPRA to be updated is selected cyclically as n p ≜ 1 + mod ( p − 1 , N ) n_p\triangleq1+\operatorname{mod}(p-1,N) n p ≜ 1 + mod ( p − 1 , N ) . For s ∈ S ≜ { 1 , ⋯ , S } s\in\mathcal{S}\triangleq\{1,\cdots,S\} s ∈ S ≜ { 1 , ⋯ , S } , define
Z n p ← s [ p − 1 ] ≜ [ z 1 [ p − 1 ] ⋯ z n p − 1 [ p − 1 ] e s z n p + 1 [ p − 1 ] ⋯ z N [ p − 1 ] ] . (35) \mathbf{Z}_{n_p\leftarrow s}^{[p-1]}\triangleq[\mathbf{z}_{1}^{[p-1]}\cdots\mathbf{z}_{n_p-1}^{[p-1]}\ \mathbf{e}_s\ \mathbf{z}_{n_p+1}^{[p-1]}\cdots\mathbf{z}_{N}^{[p-1]}]. \tag*{(35)} Z n p ← s [ p − 1 ] ≜ [ z 1 [ p − 1 ] ⋯ z n p − 1 [ p − 1 ] e s z n p + 1 [ p − 1 ] ⋯ z N [ p − 1 ] ] . (35)
Since only the n p n_p n p th pattern state is replaced, substituting (35) into (34) gives
H e f f ( R ‾ , Z n p ← s [ p − 1 ] ) = H e f f ( R ‾ , Z [ p − 1 ] ) + [ h n p , s ( r ‾ n p ) − h n p , s n p [ p − 1 ] ( r ‾ n p ) ] e n p T , (36) \mathbf{H}_{\mathrm{eff}}(\overline{\mathbf{R}},\mathbf{Z}_{n_p\leftarrow s}^{[p-1]})
=\mathbf{H}_{\mathrm{eff}}(\overline{\mathbf{R}},\mathbf{Z}^{[p-1]})
+\left[\mathbf{h}_{n_p,s}(\overline{\mathbf{r}}_{n_p})-\mathbf{h}_{n_p,s_{n_p}^{[p-1]}}(\overline{\mathbf{r}}_{n_p})\right]\mathbf{e}_{n_p}^{\mathrm{T}}, \tag*{(36)} H eff ( R , Z n p ← s [ p − 1 ] ) = H eff ( R , Z [ p − 1 ] ) + [ h n p , s ( r n p ) − h n p , s n p [ p − 1 ] ( r n p ) ] e n p T , (36)
which shows that a single pattern replacement changes only one column of H e f f \mathbf{H}_{\mathrm{eff}} H eff .
2) Cyclic Candidate-Testing: Let ( R ‾ , Z [ 0 ] , F R F [ 0 ] , F B B [ 0 ] ) (\overline{\mathbf{R}},\mathbf{Z}^{[0]},\mathbf{F}_{\mathrm{RF}}^{[0]},\mathbf{F}_{\mathrm{BB}}^{[0]}) ( R , Z [ 0 ] , F RF [ 0 ] , F BB [ 0 ] ) denote the feasible initial input configuration obtained from the preceding outer AO iteration. At iteration p p p , construct the feasible testing set: T [ p ] ≜ { Z n p ← s [ p − 1 ] : s ∈ S , T d ( Z n p ← s [ p − 1 ] ) ≥ ϵ T } \mathcal{T}^{[p]}\triangleq\{\mathbf{Z}_{n_p\leftarrow s}^{[p-1]}:s\in\mathcal{S},\ T_d(\mathbf{Z}_{n_p\leftarrow s}^{[p-1]})\ge\epsilon_{\mathrm{T}}\} T [ p ] ≜ { Z n p ← s [ p − 1 ] : s ∈ S , T d ( Z n p ← s [ p − 1 ] ) ≥ ϵ T } . Since s = s n p [ p − 1 ] s=s_{n_p}^{[p-1]} s = s n p [ p − 1 ] reproduces the current pattern configuration, the current point satisfies Z [ p − 1 ] ∈ T [ p ] \mathbf{Z}^{[p-1]}\in\mathcal{T}^{[p]} Z [ p − 1 ] ∈ T [ p ] . Thus, T [ p ] ≠ ∅ \mathcal{T}^{[p]}\ne\varnothing T [ p ] = ∅ .
For every Z ~ ∈ T [ p ] \widetilde{\mathbf{Z}}\in\mathcal{T}^{[p]} Z ∈ T [ p ] , H e f f \mathbf{H}_{\mathrm{eff}} H eff is first generated using (36). Thereafter, ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) are then reoptimized as ( F ^ B B ( Z ~ ) , F ^ R F ( Z ~ ) ) = A F ( R ‾ , Z ~ ) (\widehat{\mathbf{F}}_{\mathrm{BB}}(\widetilde{\mathbf{Z}}),\widehat{\mathbf{F}}_{\mathrm{RF}}(\widetilde{\mathbf{Z}}))=\mathcal{A}_{\mathrm{F}}(\overline{\mathbf{R}},\widetilde{\mathbf{Z}}) ( F BB ( Z ) , F RF ( Z )) = A F ( R , Z ) , where A F \mathcal{A}_{\mathrm{F}} A F denotes the framework in Section III-B. The corresponding locally optimized value is V ^ [ R ‾ , Z ~ ] ≜ η E E ( F ^ R F ( Z ~ ) , F ^ B B ( Z ~ ) , R ‾ , Z ~ ) \widehat{\mathcal{V}}[\overline{\mathbf{R}},\widetilde{\mathbf{Z}}]\triangleq\eta_{\mathrm{EE}}(\widehat{\mathbf{F}}_{\mathrm{RF}}(\widetilde{\mathbf{Z}}),\widehat{\mathbf{F}}_{\mathrm{BB}}(\widetilde{\mathbf{Z}}),\overline{\mathbf{R}},\widetilde{\mathbf{Z}}) V [ R , Z ] ≜ η EE ( F RF ( Z ) , F BB ( Z ) , R , Z ) , and the best configuration in T [ p ] \mathcal{T}^{[p]} T [ p ] is selected as
Z ^ [ p ] ( R ‾ ) ∈ arg max Z ~ ∈ T [ p ] V ^ [ R ‾ , Z ~ ] . (37) \widehat{\mathbf{Z}}^{[p]}(\overline{\mathbf{R}})\in\underset{\widetilde{\mathbf{Z}}\in\mathcal{T}^{[p]}}{\arg\max}\ \widehat{\mathcal{V}}[\overline{\mathbf{R}},\widetilde{\mathbf{Z}}]. \tag*{(37)} Z [ p ] ( R ) ∈ Z ∈ T [ p ] arg max V [ R , Z ] . (37)
Let the EE at the beginning of iteration p p p be η E E [ p − 1 ] ≜ η E E ( F ‾ R F [ p − 1 ] , F ‾ B B [ p − 1 ] , R ‾ , Z [ p − 1 ] ) \eta_{\mathrm{EE}}^{[p-1]} \triangleq\eta_{\mathrm{EE}}\left(\overline{\mathbf{F}}_{\mathrm{RF}}^{[p-1]},\overline{\mathbf{F}}_{\mathrm{BB}}^{[p-1]},\overline{\mathbf{R}},\mathbf{Z}^{[p-1]}\right) η EE [ p − 1 ] ≜ η EE ( F RF [ p − 1 ] , F BB [ p − 1 ] , R , Z [ p − 1 ] ) . From (32)–(33) and Z [ p − 1 ] ∈ T [ p ] \mathbf{Z}^{[p-1]}\in\mathcal{T}^{[p]} Z [ p − 1 ] ∈ T [ p ] : V ^ [ p ] ( R ‾ , Z [ p − 1 ] ) ≥ η E E [ p − 1 ] \widehat{\mathcal{V}}^{[p]}\left(\overline{\mathbf{R}},\mathbf{Z}^{[p-1]}\right)\geq\eta_{\mathrm{EE}}^{[p-1]} V [ p ] ( R , Z [ p − 1 ] ) ≥ η EE [ p − 1 ] . Furthermore, by (37) :
V ^ [ p ] ( R , Z ^ [ p ] ‾ ) = max Z ^ ∈ T [ p ] V ^ [ p ] ( R , Z ^ ‾ ) ≥ V ^ [ p ] ( R , Z [ p − 1 ] ‾ ) ≥ η E E [ p − 1 ] . (38) \widehat{\mathcal{V}}^{[p]}\left(\overline{\mathbf{R},\widehat{\mathbf{Z}}^{[p]}}\right)=\max_{\widehat{\mathbf{Z}}\in\mathcal{T}^{[p]}}\widehat{\mathcal{V}}^{[p]}\left(\overline{\mathbf{R},\widehat{\mathbf{Z}}}\right)\geq\widehat{\mathcal{V}}^{[p]}\left(\overline{\mathbf{R},\mathbf{Z}^{[p-1]}}\right)\geq\eta_{\mathrm{EE}}^{[p-1]}. \tag*{(38)} V [ p ] ( R , Z [ p ] ) = Z ∈ T [ p ] max V [ p ] ( R , Z ) ≥ V [ p ] ( R , Z [ p − 1 ] ) ≥ η EE [ p − 1 ] . (38)
Therefore, the selected Z ^ [ p ] \widehat{\mathbf{Z}}^{[p]} Z [ p ] together with its reoptimized ( F ^ R F [ p ] ( Z ^ ) , F ^ B B [ p ] ( Z ^ ) ) (\widehat{\mathbf{F}}_{\mathrm{RF}}^{[p]}(\widehat{\mathbf{Z}}),\widehat{\mathbf{F}}_{\mathrm{BB}}^{[p]}(\widehat{\mathbf{Z}})) ( F RF [ p ] ( Z ) , F BB [ p ] ( Z )) cannot decrease the EE since by (37) : η E E [ p ] ← V ^ [ p ] ( R ‾ , Z ^ [ p ] ) \eta_{\mathrm{EE}}^{[p]}\leftarrow\widehat{\mathcal{V}}^{[p]}\left(\overline{\mathbf{R}},\widehat{\mathbf{Z}}^{[p]}\right) η EE [ p ] ← V [ p ] ( R , Z [ p ] ) . Hence for fixed R ‾ \overline{\mathbf{R}} R :
( Z [ p ] , F R F [ p ] , F B B [ p ] ) = ( Z ^ [ p ] , F ^ R F [ p ] ( Z ^ [ p ] ) , F ^ B B [ p ] ( Z ^ [ p ] ) ) . (39) \left(\mathbf{Z}^{[p]},\mathbf{F}_{\mathrm{RF}}^{[p]},\mathbf{F}_{\mathrm{BB}}^{[p]}\right)=\left(\widehat{\mathbf{Z}}^{[p]},\widehat{\mathbf{F}}_{\mathrm{RF}}^{[p]}\left(\widehat{\mathbf{Z}}^{[p]}\right),\widehat{\mathbf{F}}_{\mathrm{BB}}^{[p]}\left(\widehat{\mathbf{Z}}^{[p]}\right)\right). \tag*{(39)} ( Z [ p ] , F RF [ p ] , F BB [ p ] ) = ( Z [ p ] , F RF [ p ] ( Z [ p ] ) , F BB [ p ] ( Z [ p ] ) ) . (39)
Define the pattern-change indicator b p a t [ p ] ≜ 1 { Z [ p ] ≠ Z [ p − 1 ] } b_{\mathrm{pat}}^{[p]}\triangleq\mathbb{1}\{\mathbf{Z}^{[p]}\ne\mathbf{Z}^{[p-1]}\} b pat [ p ] ≜ 1 { Z [ p ] = Z [ p − 1 ] } . The procedure terminates when no pattern state has changed during the latest complete sweep, ∑ i = p − N + 1 p b p a t [ i ] = 0 \sum_{i=p-N+1}^{p}b_{\mathrm{pat}}^{[i]}=0 ∑ i = p − N + 1 p b pat [ i ] = 0 (p ≥ N p\geq N p ≥ N ), and ∣ η E E [ p ] − η E E [ p − N ] ∣ ≤ ϵ C O D \left|\eta_{\mathrm{EE}}^{[p]}-\eta_{\mathrm{EE}}^{[p-N]}\right|\leq\epsilon_{\mathrm{COD}} η EE [ p ] − η EE [ p − N ] ≤ ϵ COD , for given tolerance ϵ C O D \epsilon_{\mathrm{COD}} ϵ COD .
WMMSE-SCA-Based Optimization of R \mathbf{R} R
We then optimize R \mathbf{R} R while holding others. Let ( F ‾ R F , F ‾ B B ) (\overline{\mathbf{F}}_{\mathrm{RF}},\overline{\mathbf{F}}_{\mathrm{BB}}) ( F RF , F BB ) and Z ‾ ≜ [ e ‾ s 1 ⋯ e ‾ s N ] \overline{\mathbf{Z}}\triangleq[\overline{\mathbf{e}}_{s_1}\ \cdots\ \overline{\mathbf{e}}_{s_N}] Z ≜ [ e s 1 ⋯ e s N ] denote the fixed hybrid precoders and pattern, and define
P ‾ ≜ F ‾ R F F ‾ B B = [ p ‾ 1 ⋯ p ‾ K ] , P ‾ d a t a ≜ ∥ F ‾ B B ∥ 2 η P A + P c i r c , \overline{\mathbf{P}}\triangleq\overline{\mathbf{F}}_{\mathrm{RF}}\overline{\mathbf{F}}_{\mathrm{BB}}=[\overline{\mathbf{p}}_1\cdots\overline{\mathbf{p}}_K],\quad\overline{P}_{\mathrm{data}}\triangleq\frac{\|\overline{\mathbf{F}}_{\mathrm{BB}}\|^2}{\eta_{\mathrm{PA}}}+P_{\mathrm{circ}}, P ≜ F RF F BB = [ p 1 ⋯ p K ] , P data ≜ η PA ∥ F BB ∥ 2 + P circ ,
T ‾ p a t ≜ T p a t ( Z ‾ ) , E ‾ p a t ≜ E p a t ( Z ‾ ) . (40) \overline{T}_{\mathrm{pat}}\triangleq T_{\mathrm{pat}}(\overline{\mathbf{Z}}),\quad\overline{E}_{\mathrm{pat}}\triangleq E_{\mathrm{pat}}(\overline{\mathbf{Z}}). \tag*{(40)} T pat ≜ T pat ( Z ) , E pat ≜ E pat ( Z ) . (40)
Let r ≜ vec ( R ) \mathbf{r}\triangleq\operatorname{vec}(\mathbf{R}) r ≜ vec ( R ) . Under Theorem 1, define the r \mathbf{r} r -dependent:
T ^ m o v ( r ) ≜ T m o v ( R , v ^ ( R ) ) , E ^ m o v ( r ) ≜ E m o v ( R , v ^ ( R ) ) , \widehat{T}_{\mathrm{mov}}(\mathbf{r})\triangleq T_{\mathrm{mov}}\left(\mathbf{R},\widehat{\mathbf{v}}(\mathbf{R})\right),\quad\widehat{E}_{\mathrm{mov}}(\mathbf{r})\triangleq E_{\mathrm{mov}}\left(\mathbf{R},\widehat{\mathbf{v}}(\mathbf{R})\right), T mov ( r ) ≜ T mov ( R , v ( R ) ) , E mov ( r ) ≜ E mov ( R , v ( R ) ) ,
E ^ b l k ( r ) ≜ E ^ m o v ( r ) + E ‾ p a t + T d ( r ) P ‾ d a t a , \widehat{E}_{\mathrm{blk}}(\mathbf{r})\triangleq\widehat{E}_{\mathrm{mov}}(\mathbf{r})+\overline{E}_{\mathrm{pat}}+T_{\mathrm{d}}(\mathbf{r})\overline{P}_{\mathrm{data}}, E blk ( r ) ≜ E mov ( r ) + E pat + T d ( r ) P data ,
T d ( r ) ≜ T c − T ^ m o v ( r ) − T ‾ p a t , η R ( r ) ≜ W T d ( r ) R w ( r ) E ^ b l k ( r ) . (41) T_{\mathrm{d}}(\mathbf{r})\triangleq T_{\mathrm{c}}-\widehat{T}_{\mathrm{mov}}(\mathbf{r})-\overline{T}_{\mathrm{pat}},\quad\eta_{\mathbf{R}}(\mathbf{r})\triangleq\frac{W T_{\mathrm{d}}(\mathbf{r})R_{\mathrm{w}}(\mathbf{r})}{\widehat{E}_{\mathrm{blk}}(\mathbf{r})}. \tag*{(41)} T d ( r ) ≜ T c − T mov ( r ) − T pat , η R ( r ) ≜ E blk ( r ) W T d ( r ) R w ( r ) . (41)
Here, R w ( r ) R_{\mathrm{w}}(\mathbf{r}) R w ( r ) denotes R w R_{\mathrm{w}} R w evaluated at R = unvec ( r ) \mathbf{R}=\operatorname{unvec}(\mathbf{r}) R = unvec ( r ) . To start, the quadratic transform gives [44 ]
η E E = A ( r ) B ( r ) = max ξ ≥ 0 [ 2 ξ A ( r ) − ξ 2 B ( r ) ] , (42) \eta_{\mathrm{EE}}=\frac{A(\mathbf{r})}{B(\mathbf{r})}=\max_{\xi\geq0}\left[2\xi\sqrt{A(\mathbf{r})}-\xi^2B(\mathbf{r})\right], \tag*{(42)} η EE = B ( r ) A ( r ) = ξ ≥ 0 max [ 2 ξ A ( r ) − ξ 2 B ( r ) ] , (42)
wherein A ( r ) = W T d ( r ) R w ( r ) A(\mathbf{r})=WT_{\mathrm{d}}(\mathbf{r})R_{\mathrm{w}}(\mathbf{r}) A ( r ) = W T d ( r ) R w ( r ) and B ( r ) = E ^ b l k ( r ) B(\mathbf{r})=\widehat{E}_{\mathrm{blk}}(\mathbf{r}) B ( r ) = E blk ( r ) , and the optimal ξ ⋆ ( r ) \xi^\star(\mathbf{r}) ξ ⋆ ( r ) is ξ ⋆ ( r ) = W T d ( r ) R w ( r ) E ^ b l k ( r ) \xi^\star(\mathbf{r})=\frac{\sqrt{WT_{\mathrm{d}}(\mathbf{r})R_{\mathrm{w}}(\mathbf{r})}}{\widehat{E}_{\mathrm{blk}}(\mathbf{r})} ξ ⋆ ( r ) = E blk ( r ) W T d ( r ) R w ( r ) . Thus, at iteration q q q with corresponding r ( q ) \mathbf{r}^{(q)} r ( q ) , we set ξ ( q ) = W T d ( r ( q ) ) R w ( r ( q ) ) E ^ b l k ( r ( q ) ) \xi^{(q)}=\frac{\sqrt{WT_{\mathrm{d}}(\mathbf{r}^{(q)})R_{\mathrm{w}}(\mathbf{r}^{(q)})}}{\widehat{E}_{\mathrm{blk}}(\mathbf{r}^{(q)})} ξ ( q ) = E blk ( r ( q ) ) W T d ( r ( q ) ) R w ( r ( q ) ) .
WMMSE Representation and r \mathbf{r} r -Dependent MSE
For fixed s ‾ n \overline{s}_n s n , the coefficient of stream j j j observed by user k k k is a k , j ( r ) ≜ h k ∗ ( R , Z ‾ ) p ‾ j = ∑ n = 1 N p ‾ n , j χ k , n ( r n ) a_{k,j}(\mathbf{r})\triangleq\mathbf{h}_k^{*}(\mathbf{R},\overline{\mathbf{Z}})\overline{\mathbf{p}}_j=\sum_{n=1}^{N}\overline{p}_{n,j}\chi_{k,n}(r_n) a k , j ( r ) ≜ h k ∗ ( R , Z ) p j = ∑ n = 1 N p n , j χ k , n ( r n ) with p ‾ n , j ≜ [ p ‾ j ] n \overline{p}_{n,j}\triangleq[\overline{\mathbf{p}}_j]_n p n , j ≜ [ p j ] n and χ k , n ≜ χ k , n , s ‾ n \chi_{k,n}\triangleq\chi_{k,n,\overline{s}_n} χ k , n ≜ χ k , n , s n . Same with Section III-B, for equalizer u ~ k ∈ C \widetilde{u}_k\in\mathbb{C} u k ∈ C , the MSE of user k k k is e k ( r , u ~ k ) = ∣ u ~ k ∣ 2 ( ∑ j = 1 K ∣ a k , j ( r ) ∣ 2 + σ k 2 ) − 2 ℜ { u ~ k ∗ a k , k ( r ) } + 1 e_k(\mathbf{r},\widetilde{u}_k)=|\widetilde{u}_k|^2\left(\sum_{j=1}^{K}|a_{k,j}(\mathbf{r})|^2+\sigma_k^2\right)-2\Re\{\widetilde{u}_k^{*}a_{k,k}(\mathbf{r})\}+1 e k ( r , u k ) = ∣ u k ∣ 2 ( ∑ j = 1 K ∣ a k , j ( r ) ∣ 2 + σ k 2 ) − 2ℜ { u k ∗ a k , k ( r )} + 1 , and the sum-SE satisfies [19 ] :
R w ( r ) = 1 ln 2 max { u ~ k , ϖ ~ k > 0 } ∑ k = 1 K ( ln ϖ ~ k − ϖ ~ k e k ( r , u ~ k ) + 1 ) , (43) R_{\mathrm{w}}(\mathbf{r})=\frac{1}{\ln2}\max_{\{\widetilde{u}_k,\widetilde{\varpi}_k>0\}}\sum_{k=1}^{K}\left(\ln\widetilde{\varpi}_k-\widetilde{\varpi}_ke_k(\mathbf{r},\widetilde{u}_k)+1\right), \tag*{(43)} R w ( r ) = ln 2 1 { u k , ϖ k > 0 } max k = 1 ∑ K ( ln ϖ k − ϖ k e k ( r , u k ) + 1 ) , (43)
where ϖ ~ k > 0 \widetilde{\varpi}_k>0 ϖ k > 0 is the auxiliary MSE weight, and at r ( q ) \mathbf{r}^{(q)} r ( q ) , the optimal u ~ k \widetilde{u}_k u k and ϖ ~ k \widetilde{\varpi}_k ϖ k are respectively u ~ k ( q ) = a k , k ( r ( q ) ) ∑ j = 1 K ∣ a k , j ( r ( q ) ) ∣ 2 + σ k 2 \widetilde{u}_k^{(q)}=\frac{a_{k,k}(\mathbf{r}^{(q)})}{\sum_{j=1}^{K}|a_{k,j}(\mathbf{r}^{(q)})|^2+\sigma_k^2} u k ( q ) = ∑ j = 1 K ∣ a k , j ( r ( q ) ) ∣ 2 + σ k 2 a k , k ( r ( q ) ) and ϖ ~ k ( q ) = 1 e k ( r ( q ) , u ~ k ( q ) ) \widetilde{\varpi}_k^{(q)}=\frac{1}{e_k(\mathbf{r}^{(q)},\widetilde{u}_k^{(q)})} ϖ k ( q ) = e k ( r ( q ) , u k ( q ) ) 1 . However, since e k ( r , u ~ k ) e_k(\mathbf{r},\widetilde{u}_k) e k ( r , u k ) remains nonconvex in r \mathbf{r} r , we construct its convex upper-approximation.
Convex MSE Majorization
Using g k , n ( r n ) ≜ ∇ r n χ k , n ( r n ) = − j β k L k ∑ ℓ = 1 L k α k , ℓ ψ s ‾ n ( Ω k , ℓ ) κ k , ℓ e − j κ k , ℓ r n g_{k,n}(r_n)\triangleq\nabla_{r_n}\chi_{k,n}(r_n)=-\mathrm{j}\sqrt{\frac{\beta_k}{L_k}}\sum_{\ell=1}^{L_k}\alpha_{k,\ell}\psi_{\overline{s}_n}(\Omega_{k,\ell})\kappa_{k,\ell}e^{-\mathrm{j}\kappa_{k,\ell}r_n} g k , n ( r n ) ≜ ∇ r n χ k , n ( r n ) = − j L k β k ∑ ℓ = 1 L k α k , ℓ ψ s n ( Ω k , ℓ ) κ k , ℓ e − j κ k , ℓ r n :
∇ r e k ( r , u ~ k ) = 2 ∣ u ~ k ∣ 2 { ∑ j = 1 K ℜ { a k , j ∗ ( r ) p ‾ n , j g k , n ( r n ) } } n = 1 N − 2 ℜ { u ~ k ∗ p ‾ n , k g k , n ( r n ) } , (44) \nabla_{\mathbf{r}}e_k(\mathbf{r},\widetilde{u}_k)=2|\widetilde{u}_k|^2\left\{\sum_{j=1}^{K}\Re\left\{a_{k,j}^{*}(\mathbf{r})\overline{p}_{n,j}g_{k,n}(r_n)\right\}\right\}_{n=1}^{N}-2\Re\left\{\widetilde{u}_k^{*}\overline{p}_{n,k}g_{k,n}(r_n)\right\}, \tag*{(44)} ∇ r e k ( r , u k ) = 2∣ u k ∣ 2 { j = 1 ∑ K ℜ { a k , j ∗ ( r ) p n , j g k , n ( r n ) } } n = 1 N − 2ℜ { u k ∗ p n , k g k , n ( r n ) } , (44)
and stacking (44) over n n n gives ∇ r e k \nabla_{\mathbf{r}}e_k ∇ r e k . Since { C n } \{\mathcal{C}_n\} { C n } are compact and the field-response functions are twice continuously differentiable, the gradient of e k e_k e k is Lipschitz continuous. Hence, for a sufficiently large curvature L e , k ( q ) > 0 L_{e,k}^{(q)}>0 L e , k ( q ) > 0 [45 ] .11
e k ( r , u ~ k ( q ) ) ≤ e k ( r ( q ) , u ~ k ( q ) ) + ∇ r e k ( r ( q ) , u ~ k ( q ) ) T ( r − r ( q ) ) + L e , k ( q ) 2 ∥ r − r ( q ) ∥ 2 2 ≜ e ‾ k ( r ; r ( q ) ) . (45) e_k(\mathbf{r},\widetilde{u}_k^{(q)})\leq e_k(\mathbf{r}^{(q)},\widetilde{u}_k^{(q)})+\nabla_{\mathbf{r}}e_k(\mathbf{r}^{(q)},\widetilde{u}_k^{(q)})^{\mathsf{T}}(\mathbf{r}-\mathbf{r}^{(q)})+\frac{L_{e,k}^{(q)}}{2}\|\mathbf{r}-\mathbf{r}^{(q)}\|_2^2\triangleq\overline{e}_k(\mathbf{r};\mathbf{r}^{(q)}). \tag*{(45)} e k ( r , u k ( q ) ) ≤ e k ( r ( q ) , u k ( q ) ) + ∇ r e k ( r ( q ) , u k ( q ) ) T ( r − r ( q ) ) + 2 L e , k ( q ) ∥ r − r ( q ) ∥ 2 2 ≜ e k ( r ; r ( q ) ) . (45)
Herein, e ‾ k ( r ; r ( q ) ) \overline{e}_k(\mathbf{r};\mathbf{r}^{(q)}) e k ( r ; r ( q ) ) is convex in r \mathbf{r} r , upper-bounds e k e_k e k , and satisfies e ‾ k ( r ( q ) ; r ( q ) ) = e k ( r ( q ) , u ~ k ( q ) ) \overline{e}_k(\mathbf{r}^{(q)};\mathbf{r}^{(q)})=e_k(\mathbf{r}^{(q)},\widetilde{u}_k^{(q)}) e k ( r ( q ) ; r ( q ) ) = e k ( r ( q ) , u k ( q ) ) and ∇ r e ‾ k ( r ( q ) ; r ( q ) ) = ∇ r e k ( r ( q ) , u ~ k ( q ) ) \nabla_{\mathbf{r}}\overline{e}_k(\mathbf{r}^{(q)};\mathbf{r}^{(q)})=\nabla_{\mathbf{r}}e_k(\mathbf{r}^{(q)},\widetilde{u}_k^{(q)}) ∇ r e k ( r ( q ) ; r ( q ) ) = ∇ r e k ( r ( q ) , u k ( q ) ) . Substituting (45) into (43) yields
R ‾ w ( r ; r ( q ) ) ≜ 1 ln 2 ∑ k = 1 K [ ln ϖ ~ k ( q ) − ϖ ~ k ( q ) e ‾ k ( r ; r ( q ) ) + 1 ] . (46) \overline{R}_{\mathrm{w}}(\mathbf{r};\mathbf{r}^{(q)})\triangleq\frac{1}{\ln2}\sum_{k=1}^{K}\left[\ln\widetilde{\varpi}_k^{(q)}-\widetilde{\varpi}_k^{(q)}\overline{e}_k(\mathbf{r};\mathbf{r}^{(q)})+1\right]. \tag*{(46)} R w ( r ; r ( q ) ) ≜ ln 2 1 k = 1 ∑ K [ ln ϖ k ( q ) − ϖ k ( q ) e k ( r ; r ( q ) ) + 1 ] . (46)
Since − e ‾ k -\overline{e}_k − e k is concave, R ‾ w \overline{R}_{\mathrm{w}} R w is concave in r \mathbf{r} r . Moreover, R ‾ w ( r ; r ( q ) ) ≤ R w ( r ) \overline{R}_{\mathrm{w}}(\mathbf{r};\mathbf{r}^{(q)})\leq R_{\mathrm{w}}(\mathbf{r}) R w ( r ; r ( q ) ) ≤ R w ( r ) ( ∀ r ) (\forall\mathbf{r}) ( ∀ r ) , while equality and first-order consistency hold at r = r ( q ) \mathbf{r}=\mathbf{r}^{(q)} r = r ( q ) .
Convex Majorization of E b l k E_{\mathrm{blk}} E blk
Although T d ( r ) T_{\mathrm{d}}(\mathbf{r}) T d ( r ) is concave, the denominator term T d ( r ) P ‾ d a t a T_{\mathrm{d}}(\mathbf{r})\overline{P}_{\mathrm{data}} T d ( r ) P data must be upper-bounded to construct a valid minorizer of the transformed EE objective. Define I ( q ) ≜ arg max ( n , a ) ∈ N { r n , a ( q ) − r n , a 0 v n , a max } \mathcal{I}^{(q)}\triangleq\arg\max_{(n,a)\in\mathcal{N}}\left\{\frac{r_{n,a}^{(q)}-r_{n,a}^{0}}{v_{n,a}^{\max}}\right\} I ( q ) ≜ arg max ( n , a ) ∈ N { v n , a m a x r n , a ( q ) − r n , a 0 } the active set of ( n , a ) (n,a) ( n , a ) that determine the overall mechanical reconfiguration time at r ( q ) \mathbf{r}^{(q)} r ( q ) . A subgradient of T ^ m o v \widehat{T}_{\mathrm{mov}} T mov at r ( q ) \mathbf{r}^{(q)} r ( q ) [46 ] :
g m o v ( q ) = ∑ ( n , a ) ∈ I ( q ) ϑ n , a ( q ) sgn ( r n , a ( q ) − r n , a 0 ) v n , a max e n , a , (47) \mathbf{g}_{\mathrm{mov}}^{(q)}=\sum_{(n,a)\in\mathcal{I}^{(q)}}\vartheta_{n,a}^{(q)}\frac{\operatorname{sgn}(r_{n,a}^{(q)}-r_{n,a}^{0})}{v_{n,a}^{\max}}\mathbf{e}_{n,a}, \tag*{(47)} g mov ( q ) = ( n , a ) ∈ I ( q ) ∑ ϑ n , a ( q ) v n , a m a x sgn ( r n , a ( q ) − r n , a 0 ) e n , a , (47)
where ϑ n , a ( q ) ≥ 0 \vartheta_{n,a}^{(q)}\geq0 ϑ n , a ( q ) ≥ 0 , ∑ ( n , a ) ∈ I ( q ) ϑ n , a ( q ) = 1 \sum_{(n,a)\in\mathcal{I}^{(q)}}\vartheta_{n,a}^{(q)}=1 ∑ ( n , a ) ∈ I ( q ) ϑ n , a ( q ) = 1 , and e n , a ∈ R 2 N \mathbf{e}_{n,a}\in\mathbb{R}^{2N} e n , a ∈ R 2 N is the canonical basis associated with ( n , a ) (n,a) ( n , a ) . Accordingly, g d ( q ) ≜ − g m o v ( q ) \mathbf{g}_{\mathrm{d}}^{(q)}\triangleq-\mathbf{g}_{\mathrm{mov}}^{(q)} g d ( q ) ≜ − g mov ( q ) is a supergradient of the concave function T d T_{\mathrm{d}} T d , and its first-order affine upper-bound is therefore
T d ( r ) ≤ T d ( r ( q ) ) + ( g d ( q ) ) T ( r − r ( q ) ) ≜ T ‾ d ( r ; r ( q ) ) . (48) T_{\mathrm{d}}(\mathbf{r})\leq T_{\mathrm{d}}(\mathbf{r}^{(q)})+\left(\mathbf{g}_{\mathrm{d}}^{(q)}\right)^{\mathsf{T}}\left(\mathbf{r}-\mathbf{r}^{(q)}\right)\triangleq\overline{T}_{\mathrm{d}}(\mathbf{r};\mathbf{r}^{(q)}). \tag*{(48)} T d ( r ) ≤ T d ( r ( q ) ) + ( g d ( q ) ) T ( r − r ( q ) ) ≜ T d ( r ; r ( q ) ) . (48)
The inequality direction follows from the concavity of T d T_{\mathrm{d}} T d . Within every differentiable operating region, the gradient of E m o v E_{\mathrm{mov}} E mov is Lipschitz continuous. Thus, with g E ( q ) ≜ ∇ r E m o v ( r ( q ) ) \mathbf{g}_{E}^{(q)}\triangleq\nabla_{\mathbf{r}}E_{\mathrm{mov}}(\mathbf{r}^{(q)}) g E ( q ) ≜ ∇ r E mov ( r ( q ) ) and for a sufficiently large L E ( q ) > 0 L_{E}^{(q)}>0 L E ( q ) > 0 ,
E m o v ( r ) ≤ E m o v ( r ( q ) ) + ( g E ( q ) ) T ( r − r ( q ) ) + L E ( q ) 2 ∥ r − r ( q ) ∥ 2 2 ≜ E ‾ m o v ( r ; r ( q ) ) . (49) E_{\mathrm{mov}}(\mathbf{r})\leq E_{\mathrm{mov}}(\mathbf{r}^{(q)})+\left(\mathbf{g}_{E}^{(q)}\right)^{\mathsf{T}}\left(\mathbf{r}-\mathbf{r}^{(q)}\right)+\frac{L_{E}^{(q)}}{2}\|\mathbf{r}-\mathbf{r}^{(q)}\|_2^2\triangleq\overline{E}_{\mathrm{mov}}(\mathbf{r};\mathbf{r}^{(q)}). \tag*{(49)} E mov ( r ) ≤ E mov ( r ( q ) ) + ( g E ( q ) ) T ( r − r ( q ) ) + 2 L E ( q ) ∥ r − r ( q ) ∥ 2 2 ≜ E mov ( r ; r ( q ) ) . (49)
Combining (48) and (49) gives the upper-bound of E b l k E_{\mathrm{blk}} E blk :
E ‾ b l k ( r ; r ( q ) ) ≜ E ‾ m o v ( r ; r ( q ) ) + E ‾ p a t + P ‾ d a t a T ‾ d ( r ; r ( q ) ) . (50) \overline{E}_{\mathrm{blk}}(\mathbf{r};\mathbf{r}^{(q)})\triangleq\overline{E}_{\mathrm{mov}}(\mathbf{r};\mathbf{r}^{(q)})+\overline{E}_{\mathrm{pat}}+\overline{P}_{\mathrm{data}}\overline{T}_{\mathrm{d}}(\mathbf{r};\mathbf{r}^{(q)}). \tag*{(50)} E blk ( r ; r ( q ) ) ≜ E mov ( r ; r ( q ) ) + E pat + P data T d ( r ; r ( q ) ) . (50)
which is convex, satisfies E ‾ b l k ( r ; r ( q ) ) ≥ E b l k ( r ) \overline{E}_{\mathrm{blk}}(\mathbf{r};\mathbf{r}^{(q)}) \ge E_{\mathrm{blk}}(\mathbf{r}) E blk ( r ; r ( q ) ) ≥ E blk ( r ) and is tight and first-order consistent at r ( q ) \mathbf{r}^{(q)} r ( q ) .
Successive Inner Approximation of ∥ r n − r m ∥ 2 \lVert\mathbf{r}_n-\mathbf{r}_m\rVert^2 ∥ r n − r m ∥ 2 : For each n < m n<m n < m , define q n , m ( q ) ≜ r n ( q ) − r m ( q ) \mathbf{q}_{n,m}^{(q)} \triangleq\mathbf{r}_n^{(q)}-\mathbf{r}_m^{(q)} q n , m ( q ) ≜ r n ( q ) − r m ( q ) . Since ∥ r n − r m ∥ 2 \lVert\mathbf{r}_n-\mathbf{r}_m\rVert^2 ∥ r n − r m ∥ 2 is convex, its first-order lower-bound at current position is ∥ r n − r m ∥ 2 ≥ ∥ q n , m ( q ) ∥ 2 + 2 ( q n , m ( q ) ) T [ ( r n − r m ) − q n , m ( q ) ] \lVert\mathbf{r}_n-\mathbf{r}_m\rVert^2\geq\lVert\mathbf{q}_{n,m}^{(q)}\rVert^2+2(\mathbf{q}_{n,m}^{(q)})^{\mathsf{T}}\left[(\mathbf{r}_n-\mathbf{r}_m)-\mathbf{q}_{n,m}^{(q)}\right] ∥ r n − r m ∥ 2 ≥ ∥ q n , m ( q ) ∥ 2 + 2 ( q n , m ( q ) ) T [ ( r n − r m ) − q n , m ( q ) ] . Consequently, the affine constraint
∥ q n , m ( q ) ∥ 2 + 2 ( q n , m ( q ) ) T [ ( r n − r m ) − q n , m ( q ) ] ≥ d min 2 (51) \lVert\mathbf{q}_{n,m}^{(q)}\rVert^2+2(\mathbf{q}_{n,m}^{(q)})^{\mathsf{T}}\left[(\mathbf{r}_n-\mathbf{r}_m)-\mathbf{q}_{n,m}^{(q)}\right]\geq d_{\min}^2 \tag*{(51)} ∥ q n , m ( q ) ∥ 2 + 2 ( q n , m ( q ) ) T [ ( r n − r m ) − q n , m ( q ) ] ≥ d m i n 2 (51)
is sufficient for satisfying the original spacing constraint, which is also tight at ( r n ( q ) , r m ( q ) ) (\mathbf{r}_n^{(q)},\mathbf{r}_m^{(q)}) ( r n ( q ) , r m ( q ) ) .
Define the convex inner position set D r ( q ) ≜ { r : r n ∈ C n ( ∀ n ) , ( 51 ) ( ∀ n < m ) , T d ( r ) ≥ ϵ T , ∥ r − r ( q ) ∥ 2 ≤ Δ ( q ) } \mathcal{D}_{\mathbf{r}}^{(q)}\triangleq\left\{\mathbf{r}:\mathbf{r}_n\in\mathcal{C}_n\ (\forall n),\ (51)\ (\forall n<m),\ T_d(\mathbf{r})\geq\epsilon_T,\ \lVert\mathbf{r}-\mathbf{r}^{(q)}\rVert^2\leq\Delta^{(q)}\right\} D r ( q ) ≜ { r : r n ∈ C n ( ∀ n ) , ( 51 ) ( ∀ n < m ) , T d ( r ) ≥ ϵ T , ∥ r − r ( q ) ∥ 2 ≤ Δ ( q ) } , where Δ ( q ) > 0 \Delta^{(q)}>0 Δ ( q ) > 0 is an optional trust-region radius used to ensure the accuracy of the local quadratic majorizers, and T d ( r ) ≥ ϵ T T_d(\mathbf{r})\geq\epsilon_T T d ( r ) ≥ ϵ T , a superlevel set of the concave T d T_d T d , is convex.
SCA Transformation: For fixed ξ ( q ) \xi^{(q)} ξ ( q ) , define the transformed surrogate: Q R ( r ; r ( q ) ) ≜ 2 ξ ( q ) W T d ( r ) R w ( r ; r ( q ) ) − ( ξ ( q ) ) 2 E ‾ b l k ( r ; r ( q ) ) − ρ R 2 ∥ r − r ( q ) ∥ 2 \mathcal{Q}_{R}(\mathbf{r};\mathbf{r}^{(q)})\triangleq2\xi^{(q)}\sqrt{W T_d(\mathbf{r})R_w(\mathbf{r};\mathbf{r}^{(q)})}-(\xi^{(q)})^2\overline{E}_{\mathrm{blk}}(\mathbf{r};\mathbf{r}^{(q)})-\frac{\rho_R}{2}\lVert\mathbf{r}-\mathbf{r}^{(q)}\rVert^2 Q R ( r ; r ( q ) ) ≜ 2 ξ ( q ) W T d ( r ) R w ( r ; r ( q ) ) − ( ξ ( q ) ) 2 E blk ( r ; r ( q ) ) − 2 ρ R ∥ r − r ( q ) ∥ 2 , where ρ R > 0 \rho_R>0 ρ R > 0 is a proximal regularization coefficient. Herein, since T d T_d T d and R w R_w R w are nonnegative concave functions, their geometric mean T d ( r ) R w ( r ; r ( q ) ) \sqrt{T_d(\mathbf{r})R_w(\mathbf{r};\mathbf{r}^{(q)})} T d ( r ) R w ( r ; r ( q ) ) is concave. Moreover, E ‾ b l k \overline{E}_{\mathrm{blk}} E blk is convex. Hence, Q R \mathcal{Q}_R Q R is concave in r \mathbf{r} r .
To obtain a solver-ready formulation, introduce auxiliary variables x ≥ 0 x\geq0 x ≥ 0 , y ≥ 0 y\geq0 y ≥ 0 , and ζ ≥ 0 \zeta\geq0 ζ ≥ 0 , representing a lower-bound on data time and rate, and geometric-mean, respectively. The q q q th SCA subproblem is
max r , x , y , ζ ≥ 0 2 ξ ( q ) ζ − ( ξ ( q ) ) 2 E ‾ b l k ( r ; r ( q ) ) − ρ R 2 ∥ r − r ( q ) ∥ 2 s . t . r ∈ D r ( q ) , ϵ T ≤ x ≤ T d ( r ) , ϵ R ≤ y ≤ R w ( r ; r ( q ) ) , ∥ [ 2 ζ / W x − y ] ∥ 2 ≤ x + y . (52) \begin{aligned}
\max_{\mathbf{r},x,y,\zeta\geq0}\quad&2\xi^{(q)}\zeta-(\xi^{(q)})^2\overline{E}_{\mathrm{blk}}(\mathbf{r};\mathbf{r}^{(q)})-\frac{\rho_R}{2}\lVert\mathbf{r}-\mathbf{r}^{(q)}\rVert^2\\
\mathrm{s.t.}\quad&\mathbf{r}\in\mathcal{D}_{\mathbf{r}}^{(q)},\quad\epsilon_T\leq x\leq T_d(\mathbf{r}),\quad\epsilon_R\leq y\leq R_w(\mathbf{r};\mathbf{r}^{(q)}),\\
&\left\lVert\begin{bmatrix}2\zeta/\sqrt{W}\\x-y\end{bmatrix}\right\rVert_2\leq x+y. \tag*{(52)}
\end{aligned} r , x , y , ζ ≥ 0 max s.t. 2 ξ ( q ) ζ − ( ξ ( q ) ) 2 E blk ( r ; r ( q ) ) − 2 ρ R ∥ r − r ( q ) ∥ 2 r ∈ D r ( q ) , ϵ T ≤ x ≤ T d ( r ) , ϵ R ≤ y ≤ R w ( r ; r ( q ) ) , [ 2 ζ / W x − y ] 2 ≤ x + y . (52)
Here, ϵ R > 0 \epsilon_R>0 ϵ R > 0 ensures that R w R_w R w remains positive. The last constraint of (52) is equivalent to ζ 2 ≤ W x y \zeta^2\leq Wxy ζ 2 ≤ W x y , and is a rotated second-order-cone constraint. Therefore, (52) is a convex optimization problem that can be solved by CVX [46 ] .
Minorization Property and Monotonicity: From R w ( r ; r ( q ) ) ≤ R w ( r ) R_w(\mathbf{r};\mathbf{r}^{(q)})\leq R_w(\mathbf{r}) R w ( r ; r ( q ) ) ≤ R w ( r ) , E ‾ b l k ( r ; r ( q ) ) ≥ E b l k ( r ) \overline{E}_{\mathrm{blk}}(\mathbf{r};\mathbf{r}^{(q)})\geq E_{\mathrm{blk}}(\mathbf{r}) E blk ( r ; r ( q ) ) ≥ E blk ( r ) :
Q R ( r ; r ( q ) ) ≤ 2 ξ ( q ) W T d ( r ) R w ( r ) − ( ξ ( q ) ) 2 E b l k ( r ) ≤ η R ( r ) , (53) \mathcal{Q}_{R}(\mathbf{r};\mathbf{r}^{(q)})\leq2\xi^{(q)}\sqrt{W T_d(\mathbf{r})R_w(\mathbf{r})}-(\xi^{(q)})^2E_{\mathrm{blk}}(\mathbf{r})\leq\eta_R(\mathbf{r}), \tag*{(53)} Q R ( r ; r ( q ) ) ≤ 2 ξ ( q ) W T d ( r ) R w ( r ) − ( ξ ( q ) ) 2 E blk ( r ) ≤ η R ( r ) , (53)
where the second inequality follows from the optimality condition of (42). Furthermore, Q R ( r ( q ) ; r ( q ) ) = η R ( r ( q ) ) \mathcal{Q}_{R}(\mathbf{r}^{(q)};\mathbf{r}^{(q)})=\eta_R(\mathbf{r}^{(q)}) Q R ( r ( q ) ; r ( q ) ) = η R ( r ( q ) ) . Let r ( q + 1 ) \mathbf{r}^{(q+1)} r ( q + 1 ) be the optimizer of (52). Since r ( q ) ∈ D r ( q ) \mathbf{r}^{(q)}\in\mathcal{D}_{\mathbf{r}}^{(q)} r ( q ) ∈ D r ( q ) ,
Q R ( r ( q + 1 ) ; r ( q ) ) ≥ Q R ( r ( q ) ; r ( q ) ) . (54) \mathcal{Q}_{R}(\mathbf{r}^{(q+1)};\mathbf{r}^{(q)})\geq\mathcal{Q}_{R}(\mathbf{r}^{(q)};\mathbf{r}^{(q)}). \tag*{(54)} Q R ( r ( q + 1 ) ; r ( q ) ) ≥ Q R ( r ( q ) ; r ( q ) ) . (54)
Combining (53)-(54) yields
η R ( r ( q + 1 ) ) ≥ η R ( r ( q ) ) . (55) \eta_R(\mathbf{r}^{(q+1)})\geq\eta_R(\mathbf{r}^{(q)}). \tag*{(55)} η R ( r ( q + 1 ) ) ≥ η R ( r ( q ) ) . (55)
Thus, every valid SCA update is guaranteed to be nondecreasing in the original EE. The SCA procedure terminates when ∣ η R ( r ( q + 1 ) ) − η R ( r ( q ) ) ∣ ≤ ϵ S C A \left|\eta_R(\mathbf{r}^{(q+1)})-\eta_R(\mathbf{r}^{(q)})\right|\leq\epsilon_{\mathrm{SCA}} η R ( r ( q + 1 ) ) − η R ( r ( q ) ) ≤ ϵ SCA , for given tolerance ϵ S C A \epsilon_{\mathrm{SCA}} ϵ SCA .
Remark 2. At an iterate corresponding to a zero displacement or an axis tie, E m o v ( r ) E_{\mathrm{mov}}(\mathbf{r}) E mov ( r ) may be nondifferentiable. In this case, let E m o v , τ ( r ) E_{\mathrm{mov},\tau}(\mathbf{r}) E mov , τ ( r ) denote a continuously differentiable approximation of E m o v ( r ) E_{\mathrm{mov}}(\mathbf{r}) E mov ( r ) , obtained by replacing its absolute-value and maximum operators with smooth approximations ϕ τ ( x ) \phi_\tau(x) ϕ τ ( x ) and m τ ( z ) m_\tau(\mathbf{z}) m τ ( z ) parameterized by τ > 0 \tau>0 τ > 0 , respectively [46 ] : ϕ τ ( x ) = x 2 + τ 2 \phi_\tau(x)=\sqrt{x^2+\tau^2} ϕ τ ( x ) = x 2 + τ 2 and m τ ( z ) = τ ln ∑ i e z i / τ m_\tau(\mathbf{z})=\tau\ln\sum_i e^{z_i/\tau} m τ ( z ) = τ ln ∑ i e z i / τ so that lim τ → 0 + E m o v , τ ( r ) = E m o v ( r ) \lim_{\tau\to0^+}E_{\mathrm{mov},\tau}(\mathbf{r})=E_{\mathrm{mov}}(\mathbf{r}) lim τ → 0 + E mov , τ ( r ) = E mov ( r ) . For a fixed τ q > 0 \tau_q>0 τ q > 0 , a candidate r ( q + 1 ) \mathbf{r}^{(q+1)} r ( q + 1 ) is generated using the quadratic majorizer
E ‾ m o v , τ q ( r ; r ( q ) ) ≜ E m o v , τ q ( r ( q ) ) + ∇ E m o v , τ q ( r ( q ) ) T ( r − r ( q ) ) + L E , τ q ( q ) 2 ∥ r − r ( q ) ∥ 2 . (56) \overline{E}_{\mathrm{mov},\tau_q}(\mathbf{r};\mathbf{r}^{(q)})\triangleq E_{\mathrm{mov},\tau_q}(\mathbf{r}^{(q)})+\nabla E_{\mathrm{mov},\tau_q}(\mathbf{r}^{(q)})^{\mathsf{T}}(\mathbf{r}-\mathbf{r}^{(q)})+\frac{L_{E,\tau_q}^{(q)}}{2}\lVert\mathbf{r}-\mathbf{r}^{(q)}\rVert^2. \tag*{(56)} E mov , τ q ( r ; r ( q ) ) ≜ E mov , τ q ( r ( q ) ) + ∇ E mov , τ q ( r ( q ) ) T ( r − r ( q ) ) + 2 L E , τ q ( q ) ∥ r − r ( q ) ∥ 2 . (56)
Since (56) does not necessarily majorize the original nonsmooth E m o v E_{\mathrm{mov}} E mov , the resulting r ( q + 1 ) \mathbf{r}^{(q+1)} r ( q + 1 ) at the end of the optimization is accepted only after the evaluation:
r ( q + 1 ) ← { r ( q + 1 ) , if r ( q + 1 ) is feasible and η R ( r ( q + 1 ) ) ≥ η R ( r ( q ) ) , r ( q ) , otherwise . (57) \mathbf{r}^{(q+1)}\leftarrow
\begin{cases}
\mathbf{r}^{(q+1)}, & \text{if }\mathbf{r}^{(q+1)}\text{ is feasible and }\eta_R(\mathbf{r}^{(q+1)})\geq\eta_R(\mathbf{r}^{(q)}),\\
\mathbf{r}^{(q)}, & \text{otherwise}.
\end{cases} \tag*{(57)} r ( q + 1 ) ← { r ( q + 1 ) , r ( q ) , if r ( q + 1 ) is feasible and η R ( r ( q + 1 ) ) ≥ η R ( r ( q ) ) , otherwise . (57)
Upon rejection, τ q \tau_q τ q may be reduced by fixed portion and the surrogate problem may be re-solved. Consequently, η R ( r ( q + 1 ) ) ≥ η R ( r ( q ) ) \eta_R(\mathbf{r}^{(q+1)})\geq\eta_R(\mathbf{r}^{(q)}) η R ( r ( q + 1 ) ) ≥ η R ( r ( q ) ) holds. The reliability of the R-convergence will be numerically verified in Section IV-B.
Overall AO Framework
By combining Section III-A-III-D, we propose the AO framework for EE maximization with details in Algorithm 1, where the single-F notation corresponds to ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) .{sup}11` For deterministic initialization X ( 0 ) \mathcal{X}^{(0)} X ( 0 ) , we assume that S ≥ 2 S\geq2 S ≥ 2 and T c − ϵ T > τ s w T_c-\epsilon_T>\tau_{\mathrm{sw}} T c − ϵ T > τ sw with sufficiently small ϵ T \epsilon_T ϵ T .
We first select a prescribed anchor r ^ n ( ≠ r n 0 ) ∈ C n \hat{\mathbf{r}}_n(\ne\mathbf{r}_n^0)\in\mathcal{C}_n r ^ n ( = r n 0 ) ∈ C n and define Δ r n ≜ r ^ n − r n 0 \Delta\mathbf{r}_n\triangleq\hat{\mathbf{r}}_n-\mathbf{r}_n^0 Δ r n ≜ r ^ n − r n 0 and Δ r n , m ≜ Δ r n − Δ r m \Delta\mathbf{r}_{n,m}\triangleq\Delta\mathbf{r}_n-\Delta\mathbf{r}_m Δ r n , m ≜ Δ r n − Δ r m . To ensure that Z ( 0 ) ≠ Z 0 \mathbf{Z}^{(0)}\ne\mathbf{Z}^0 Z ( 0 ) = Z 0 , select a prescribed MPR n ˉ \bar{\mathbf{n}} n ˉ and cyclically change its state: s n ˉ ( 0 ) = 1 + mod ( s n ˉ ( 0 ) , S ) \mathbf{s}_{\bar n}^{(0)}=1+\operatorname{mod}(s_{\bar n}^{(0)},S) s n ˉ ( 0 ) = 1 + mod ( s n ˉ ( 0 ) , S ) , s n ( 0 ) = s n ( 0 ) \mathbf{s}_n^{(0)}=s_n^{(0)} s n ( 0 ) = s n ( 0 ) ( n ≠ n ˉ ) (n\ne\bar n) ( n = n ˉ ) , where z n 0 = e s n 0 \mathbf{z}_n^0=\mathbf{e}_{s_n^0} z n 0 = e s n 0 , and set Z ( 0 ) = [ e s 1 ( 0 ) ⋯ e s N ( 0 ) ] \mathbf{Z}^{(0)}=[\mathbf{e}_{s_1^{(0)}}\cdots\mathbf{e}_{s_N^{(0)}}] Z ( 0 ) = [ e s 1 ( 0 ) ⋯ e s N ( 0 ) ] . Thus, provided that S ≥ 2 S\geq2 S ≥ 2 , Z ( 0 ) ≠ Z 0 \mathbf{Z}^{(0)}\ne\mathbf{Z}^0 Z ( 0 ) = Z 0 , T p a t ( Z ( 0 ) ) = τ s w T_{\mathrm{pat}}(\mathbf{Z}^{(0)})=\tau_{\mathrm{sw}} T pat ( Z ( 0 ) ) = τ sw .
Now define T m o v ≜ max n ∈ N ∥ Δ r n ∥ a v n , a max T_{\mathrm{mov}}\triangleq\max_{n\in\mathcal{N}}\frac{\lVert\Delta\mathbf{r}_n\rVert_a}{v_{n,a}^{\max}} T mov ≜ max n ∈ N v n , a m a x ∥ Δ r n ∥ a and the time-feasible scaling bound:
α T ≜ { T c − ϵ T − τ s w T ‾ m o v , T ‾ m o v > 0 , + ∞ , T ‾ m o v = 0 ( > 0 ) . \alpha_T\triangleq
\begin{cases}
\dfrac{T_c-\epsilon_T-\tau_{\mathrm{sw}}}{\overline{T}_{\mathrm{mov}}}, & \overline{T}_{\mathrm{mov}}>0,\\
+\infty, & \overline{T}_{\mathrm{mov}}=0
\end{cases}
\ (>0). α T ≜ ⎩ ⎨ ⎧ T mov T c − ϵ T − τ sw , + ∞ , T mov > 0 , T mov = 0 ( > 0 ) .
A conservative spacing-feasible scaling bound is also α s p ≜ min n < m : ∥ Δ r n , m ∥ 2 > 0 ∥ r n 0 − r m 0 ∥ 2 − d min ∥ Δ r n , m ∥ 2 ( > 0 ) \alpha_{\mathrm{sp}}\triangleq\min_{\substack{n<m:\\\lVert\Delta\mathbf{r}_{n,m}\rVert_2>0}}\frac{\lVert\mathbf{r}_n^0-\mathbf{r}_m^0\rVert_2-d_{\min}}{\lVert\Delta\mathbf{r}_{n,m}\rVert_2}(>0) α sp ≜ min n < m : ∥ Δ r n , m ∥ 2 > 0 ∥ Δ r n , m ∥ 2 ∥ r n 0 − r m 0 ∥ 2 − d m i n ( > 0 ) , where the minimum
Figure 1 (PDF p. 8).
over an empty set is interpreted as + ∞ +\infty + ∞ . For a prescribed χ ∈ ( 0 , 1 ) \chi\in(0,1) χ ∈ ( 0 , 1 ) , set α 0 = χ min { 1 , α T , α s p } ( > 0 ) \alpha_0=\chi\min\{1,\alpha_T,\alpha_{\mathrm{sp}}\}(>0) α 0 = χ min { 1 , α T , α sp } ( > 0 ) and r n ( 0 ) = r n 0 + α 0 Δ r n \mathbf{r}_n^{(0)}=\mathbf{r}_n^0+\alpha_0\Delta\mathbf{r}_n r n ( 0 ) = r n 0 + α 0 Δ r n . Since every C n \mathcal{C}_n C n is convex, r n ( 0 ) ∈ C n \mathbf{r}_n^{(0)}\in\mathcal{C}_n r n ( 0 ) ∈ C n . Moreover,
∥ r n ( 0 ) − r m ( 0 ) ∥ 2 = ∥ r n 0 − r m 0 + α 0 Δ r n , m ∥ 2 ≥ ∥ r n 0 − r m 0 ∥ 2 − α 0 ∥ Δ r n , m ∥ 2 ≥ d min , (58) \left\|\mathbf{r}_n^{(0)}-\mathbf{r}_m^{(0)}\right\|_2=\left\|\mathbf{r}_n^0-\mathbf{r}_m^0+\alpha_0\Delta\mathbf{r}_{n,m}\right\|_2
\geq\left\|\mathbf{r}_n^0-\mathbf{r}_m^0\right\|_2-\alpha_0\left\|\Delta\mathbf{r}_{n,m}\right\|_2\geq d_{\min}, \tag*{(58)} r n ( 0 ) − r m ( 0 ) 2 = r n 0 − r m 0 + α 0 Δ r n , m 2 ≥ r n 0 − r m 0 2 − α 0 ∥ Δ r n , m ∥ 2 ≥ d m i n , (58)
obtained by the reverse triangle inequality. Furthermore,
T ^ m o v ( R ( 0 ) ) + T ^ p a t ( Z ( 0 ) ) = α 0 T ^ m o v + τ s w ≤ T c − ϵ T . (59) \widehat{T}_{\mathrm{mov}}(\mathbf{R}^{(0)})+\widehat{T}_{\mathrm{pat}}(\mathbf{Z}^{(0)})=\alpha_0\widehat{T}_{\mathrm{mov}}+\tau_{\mathrm{sw}}\leq T_c-\epsilon_T. \tag*{(59)} T mov ( R ( 0 ) ) + T pat ( Z ( 0 ) ) = α 0 T mov + τ sw ≤ T c − ϵ T . (59)
Hence R ( 0 ) ≠ R 0 \mathbf{R}^{(0)}\ne\mathbf{R}^0 R ( 0 ) = R 0 and Z ( 0 ) ≠ Z 0 \mathbf{Z}^{(0)}\ne\mathbf{Z}^0 Z ( 0 ) = Z 0 , and the initialized position and EM-domain variables are feasible.
Furthermore, F R F ( 0 ) \mathbf{F}_{\mathrm{RF}}^{(0)} F RF ( 0 ) is deterministically initialized as F R F ( 0 ) = blkdiag { ϕ 1 ( 0 ) , … , ϕ N R F ( 0 ) } \mathbf{F}_{\mathrm{RF}}^{(0)}=\operatorname{blkdiag}\{\phi_1^{(0)},\ldots,\phi_{N_{\mathrm{RF}}}^{(0)}\} F RF ( 0 ) = blkdiag { ϕ 1 ( 0 ) , … , ϕ N RF ( 0 ) } with [ ϕ r ( 0 ) ] m = e j θ r , m ( 0 ) M [\phi_r^{(0)}]_m=\frac{e^{\mathrm{j}\theta_{r,m}^{(0)}}}{\sqrt{M}} [ ϕ r ( 0 ) ] m = M e j θ r , m ( 0 ) , where we set θ r , m ( 0 ) = 2 π ( m − 1 ) M \theta_{r,m}^{(0)}=\frac{2\pi(m-1)}{M} θ r , m ( 0 ) = M 2 π ( m − 1 ) . Finally, since N R F ≥ K N_{\mathrm{RF}}\geq K N RF ≥ K :
F B B ( 0 ) = P max K [ I K 0 ( N R F − K ) × K ] , \mathbf{F}_{\mathrm{BB}}^{(0)}=\sqrt{\frac{P_{\max}}{K}}\begin{bmatrix}\mathbf{I}_K\\\mathbf{0}_{(N_{\mathrm{RF}}-K)\times K}\end{bmatrix}, F BB ( 0 ) = K P m a x [ I K 0 ( N RF − K ) × K ] ,
which satisfies ∥ F B B ( 0 ) ∥ F 2 ≤ P max \left\|\mathbf{F}_{\mathrm{BB}}^{(0)}\right\|_F^2\leq P_{\max} F BB ( 0 ) F 2 ≤ P m a x . Therefore, { F B B ( 0 ) , F R F ( 0 ) , R ( 0 ) , Z ( 0 ) } \{\mathbf{F}_{\mathrm{BB}}^{(0)},\mathbf{F}_{\mathrm{RF}}^{(0)},\mathbf{R}^{(0)},\mathbf{Z}^{(0)}\} { F BB ( 0 ) , F RF ( 0 ) , R ( 0 ) , Z ( 0 ) } constitutes a deterministic feasible initialization.
The monotonicity and convergence of EE is given by following theorem:
Theorem 2. { η E E ( t ) } \{\eta_{\mathrm{EE}}^{(t)}\} { η EE ( t ) } is nondecreasing and converges.
Proof. Let η F ( t ) \eta_F^{(t)} η F ( t ) and η Z ( t ) \eta_Z^{(t)} η Z ( t ) denote the original EEs after the update of ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) and Z \mathbf{Z} Z , respectively, during outer iteration t t t . Since (33) implies that the update of ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) does not decrease the original EE: η F ( t ) ≥ η E E ( t ) \eta_F^{(t)}\geq\eta_{\mathrm{EE}}^{(t)} η F ( t ) ≥ η EE ( t ) . In Z \mathbf{Z} Z -block, by (38)–(39): η Z ( t ) ≥ η F ( t ) \eta_Z^{(t)}\geq\eta_F^{(t)} η Z ( t ) ≥ η F ( t ) . For R \mathbf{R} R -block, (55) ensure that the update is nondecreasing in the original EE. Even at a nondifferentiable iterate, (57) accepts the candidate only when it is feasible and does not decrease the original nonsmooth EE. Consequently, η E E ( t + 1 ) ≥ η Z ( t ) \eta_{\mathrm{EE}}^{(t+1)}\geq\eta_Z^{(t)} η EE ( t + 1 ) ≥ η Z ( t ) . Combining the results yields η E E ( t ) ≤ η F ( t ) ≤ η Z ( t ) ≤ η E E ( t + 1 ) \eta_{\mathrm{EE}}^{(t)}\leq\eta_F^{(t)}\leq\eta_Z^{(t)}\leq\eta_{\mathrm{EE}}^{(t+1)} η EE ( t ) ≤ η F ( t ) ≤ η Z ( t ) ≤ η EE ( t + 1 ) . Thus, { η E E ( t ) } \{\eta_{\mathrm{EE}}^{(t)}\} { η EE ( t ) } is nondecreasing.
It remains to show that { η E E ( t ) } \{\eta_{\mathrm{EE}}^{(t)}\} { η EE ( t ) } is bounded above. From the data-time constraint and P c i r c > 0 P_{\mathrm{circ}}>0 P circ > 0 , E b l k = E ^ r e c + T ^ d P d a t a ≥ T ^ d P c i r c ≥ ϵ T P c i r c ( > 0 ) E_{\mathrm{blk}}=\widehat{E}_{\mathrm{rec}}+\widehat{T}_{d}P_{\mathrm{data}}\geq\widehat{T}_{d}P_{\mathrm{circ}}\geq\epsilon_TP_{\mathrm{circ}}(>0) E blk = E rec + T d P data ≥ T d P circ ≥ ϵ T P circ ( > 0 ) . Moreover, ∥ F B B ∥ F 2 ≤ P max \left\|\mathbf{F}_{\mathrm{BB}}\right\|_F^2\leq P_{\max} ∥ F BB ∥ F 2 ≤ P m a x , F R F \mathbf{F}_{\mathrm{RF}} F RF and { C n } \{\mathcal{C}_n\} { C n } are compact, and the feasible set of Z \mathbf{Z} Z is finite. Hence, there exists a finite constant R ‾ w \overline{R}_w R w such that R w ≤ R ‾ w R_w\leq\overline{R}_w R w ≤ R w over the feasible set. Furthermore, since 0 < T ^ d ≤ T c 0<\widehat{T}_d\leq T_c 0 < T d ≤ T c , it follows that η E E ( t ) ≤ W T c R ‾ w ϵ T P c i r c < ∞ \eta_{\mathrm{EE}}^{(t)}\leq\frac{W T_c\overline{R}_w}{\epsilon_TP_{\mathrm{circ}}}<\infty η EE ( t ) ≤ ϵ T P circ W T c R w < ∞ . Therefore, { η E E ( t ) } \{\eta_{\mathrm{EE}}^{(t)}\} { η EE ( t ) } is nondecreasing and bounded above, and the theorem follows. ■
Computational Complexity
Let L Σ ≜ ∑ k = 1 K L k L_\Sigma\triangleq\sum_{k=1}^{K}L_k L Σ ≜ ∑ k = 1 K L k and B R B_R B R be the maximum Armijo trials per RCG iteration. Let I A O I_{\mathrm{AO}} I AO , I D I_D I D , I μ I_\mu I μ , I R I_R I R , I P I_P I P , and I S C A I_{\mathrm{SCA}} I SCA denote, respectively, the numbers of outer AO iterations, Dinkelbach-WMMSE iterations for A F \mathcal{A}_F A F , bisection iterations per F B B \mathbf{F}_{\mathrm{BB}} F BB -update, RCG iterations per F R F \mathbf{F}_{\mathrm{RF}} F RF -update, individual pattern-coordinate updates per Z \mathbf{Z} Z -update, and SCA iterations per R \mathbf{R} R -update.
( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) -Update: For each i i i , computing the effective channels, equalizers, and MSE weights requires O ( K N + K 2 N R F ) \mathcal{O}(KN+K^2N_{\mathrm{RF}}) O ( K N + K 2 N RF ) operations. Constructing A \mathbf{A} A and B \mathbf{B} B and evaluating (30) require O ( N R F 3 + K N R F 2 + I μ K N R F ) \mathcal{O}(N_{\mathrm{RF}}^3+KN_{\mathrm{RF}}^2+I_\mu KN_{\mathrm{RF}}) O ( N RF 3 + K N RF 2 + I μ K N RF ) operations, where the eigendecomposition of A \mathbf{A} A is reused during the scalar bisection. For F R F \mathbf{F}_{\mathrm{RF}} F RF -update, explicitly constructing C \mathbf{C} C requires O ( K 2 N 2 ) \mathcal{O}(K^2N^2) O ( K 2 N 2 ) operations. Given C \mathbf{C} C , each RCG gradient evaluation requires O ( N 2 ) \mathcal{O}(N^2) O ( N 2 ) operations [25 ] , and the corresponding Armijo search requires at most O ( B R N 2 ) \mathcal{O}(B_RN^2) O ( B R N 2 ) operations [43 ] . The tangent-space projection, vector transport, and retraction require only O ( N ) \mathcal{O}(N) O ( N ) operations. Consequently,
C F = O ( I D [ K 2 N R F + K N R F 2 + N R F 3 + I μ K N R F + K 2 N 2 + I R ( 1 + B R ) N 2 ] ) . (60) C_F=\mathcal{O}\left(I_D\left[K^2N_{\mathrm{RF}}+KN_{\mathrm{RF}}^2+N_{\mathrm{RF}}^3+I_\mu KN_{\mathrm{RF}}+K^2N^2+I_R(1+B_R)N^2\right]\right). \tag*{(60)} C F = O ( I D [ K 2 N RF + K N RF 2 + N RF 3 + I μ K N RF + K 2 N 2 + I R ( 1 + B R ) N 2 ] ) . (60)
Z \mathbf{Z} Z -Update: For fixed R \mathbf{R} R , constructing all { h n , s } n , s \{\mathbf{h}_{n,s}\}_{n,s} { h n , s } n , s requires O ( N S L Σ ) \mathcal{O}(NSL_\Sigma) O ( N S L Σ ) operations. At each p p p , at most S S S candidates are tested with invoking A F \mathcal{A}_F A F . Therefore,
C Z = O ( N S L Σ + I P S [ C F + K + N ] ) , (61) C_Z=\mathcal{O}\left(NSL_\Sigma+I_P S[C_F+K+N]\right), \tag*{(61)} C Z = O ( N S L Σ + I P S [ C F + K + N ] ) , (61)
where the O ( K + N ) \mathcal{O}(K+N) O ( K + N ) term accounts for the rank-one channel update, feasibility test, and switching-overhead evaluation.
R \mathbf{R} R -Update: For each q q q , evaluating the r \mathbf{r} r -dependent channels and constructing the WMMSE surrogates require O ( N L Σ + K 2 N ) \mathcal{O}(NL_\Sigma+K^2N) O ( N L Σ + K 2 N ) operations. Problem (52) has 2 N + 3 2N+3 2 N + 3 primary variables, and after standard epigraph and second-order cone reformulations, the number remains n R = O ( N ) n_R=\mathcal{O}(N) n R = O ( N ) , whereas the number of scalar and conic constraints is m R = N ( N − 1 ) 2 + O ( N ) = O ( N 2 ) m_R=\frac{N(N-1)}{2}+\mathcal{O}(N)=\mathcal{O}(N^2) m R = 2 N ( N − 1 ) + O ( N ) = O ( N 2 ) . A generic primal-dual interior-point method has the conservative complexity bound [46 Chapter 11] : C I P = O ( N 5 log 1 ϵ I P ) C_{\mathrm{IP}}=\mathcal{O}\left(N^5\log\frac{1}{\epsilon_{\mathrm{IP}}}\right) C IP = O ( N 5 log ϵ IP 1 ) with accuracy ϵ I P > 0 \epsilon_{\mathrm{IP}}>0 ϵ IP > 0 . If at most B S B_S B S convex subproblems are solved at each SCA iteration due to Remark 2,13 the R \mathbf{R} R -update requires
C R = O ( I S C A [ N L Σ + K 2 N + B S C I P ] ) . (62) C_R=\mathcal{O}\left(I_{\mathrm{SCA}}\left[NL_\Sigma+K^2N+B_SC_{\mathrm{IP}}\right]\right). \tag*{(62)} C R = O ( I SCA [ N L Σ + K 2 N + B S C IP ] ) . (62)
Hence, the overall computational complexity is
C A O = O ( I A O [ ( 1 + I P S ) C F + N S L Σ + I P S ( K + N ) + I S C A ( N L Σ + K 2 N + B S C I P ) ] ) , (63) C_{\mathrm{AO}}=\mathcal{O}\left(I_{\mathrm{AO}}\left[(1+I_PS)C_F+NSL_\Sigma+I_PS(K+N)+I_{\mathrm{SCA}}(NL_\Sigma+K^2N+B_SC_{\mathrm{IP}})\right]\right), \tag*{(63)} C AO = O ( I AO [ ( 1 + I P S ) C F + N S L Σ + I P S ( K + N ) + I SCA ( N L Σ + K 2 N + B S C IP ) ] ) , (63)
where the deterministic initialization in Section III-E requires O ( N 2 + N R F K ) \mathcal{O}(N^2+N_{\mathrm{RF}}K) O ( N 2 + N RF K ) operations and is thus negligible.
Simulation Results
Simulation Setup
We consider a multi-user downlink system with f c = 3.55 f_c=3.55 f c = 3.55 GHz and σ k 2 = − 80 \sqrt{\sigma_k^2}=-80 σ k 2 = − 80 dBm, in which a BS equipped with N = 16 N=16 N = 16 MPRAs with { C n } \{\mathcal{C}_n\} { C n } arranged as a 4 × 4 4\times4 4 × 4 uniform planar array (UPA) form, serves K = 4 K=4 K = 4 single-antenna users using N R F = 4 N_{\mathrm{RF}}=4 N RF = 4 or 8 8 8 RF chains, and P max = 30 P_{\max}=30 P m a x = 30 dBm. Let r ~ i , j \tilde{\mathbf{r}}_{i,j} r ~ i , j denote the center of C i , j \mathcal{C}_{i,j} C i , j , where i , j ∈ { 1 , 2 , 3 , 4 } i,j\in\{1,2,3,4\} i , j ∈ { 1 , 2 , 3 , 4 } . Hence, by appropriately mapping C i , j ← C n r ~ i , j = [ ( i − 5 2 ) A , ( j − 5 2 ) A ] T \mathcal{C}_{i,j}\leftarrow\mathcal{C}_n\tilde{\mathbf{r}}_{i,j}=\left[\left(i-\frac{5}{2}\right)A,\left(j-\frac{5}{2}\right)A\right]^{\mathsf{T}} C i , j ← C n r ~ i , j = [ ( i − 2 5 ) A , ( j − 2 5 ) A ] T , C i , j = r ~ i , j + [ − A 2 , A 2 ] 2 \mathcal{C}_{i,j}=\tilde{\mathbf{r}}_{i,j}+\left[-\frac{A}{2},\frac{A}{2}\right]^2 C i , j = r ~ i , j + [ − 2 A , 2 A ] 2 , where A = λ 1.5 A=\frac{\lambda}{1.5} A = 1.5 λ . Hence, { C i , j } \{\mathcal{C}_{i,j}\} { C i , j } are contiguously tiled without gaps, and their union forms the overall movement aperture C = ⋃ i = 1 4 ⋃ j = 1 4 C i , j \mathcal{C}=\bigcup_{i=1}^{4}\bigcup_{j=1}^{4}\mathcal{C}_{i,j} C = ⋃ i = 1 4 ⋃ j = 1 4 C i , j . We set d min = λ 2 d_{\min}=\frac{\lambda}{2} d m i n = 2 λ , { r i , j 0 } \{\mathbf{r}_{i,j}^0\} { r i , j 0 } are set to the cell centers, i.e., r i , j 0 = r ~ i , j \mathbf{r}_{i,j}^0=\tilde{\mathbf{r}}_{i,j} r i , j 0 = r ~ i , j , and ∀ z n [ 0 ] = e 1 \forall\mathbf{z}_n^{[0]}=\mathbf{e}_1 ∀ z n [ 0 ] = e 1 . For channel, the BS-user distance follows i.i.d. ∀ d k ∼ U [ 20 , 100 ] \forall d_k\sim\mathcal{U}[20,100] ∀ d k ∼ U [ 20 , 100 ] m, while ∀ ϕ k , l ∼ U [ − π 2 , π 2 ] \forall\phi_{k,l}\sim\mathcal{U}\left[-\frac{\pi}{2},\frac{\pi}{2}\right] ∀ ϕ k , l ∼ U [ − 2 π , 2 π ] , ∀ θ k , l ∼ U [ 0 , π ] \forall\theta_{k,l}\sim\mathcal{U}[0,\pi] ∀ θ k , l ∼ U [ 0 , π ] , L = 10 L=10 L = 10 , and13 T c = 0.25 T_{\mathrm{c}} = 0.25 T c = 0.25 s. The path coefficients follow α k , ℓ ∼ C N ( 0 , 1 ) \alpha_{k,\ell} \sim\mathcal{CN}(0,1) α k , ℓ ∼ C N ( 0 , 1 ) and β k = β 0 d k − α P L \beta_k = \beta_0 d_k^{-\alpha_{\mathrm{PL}}} β k = β 0 d k − α PL with β 0 = − 43.45 \beta_0 = -43.45 β 0 = − 43.45 dB and α P L = 2.8 \alpha_{\mathrm{PL}} = 2.8 α PL = 2.8 . For pattern reconfiguration, we employ the S = 8 S = 8 S = 8 distinct radiation patterns from PRA in [21 ] as in [29 ] .14 For position reconfiguration, each MPRA is mounted on a biaxial positioning stage driven by two identical AM2224 stepper motors, with parameters in [33 Section II.B] . Thereby, the maximum axis speed v n , a max v_{n,a}^{\max} v n , a m a x is given by 2.76 m/s. Motivated by the compact piezoelectric micropump discussed in [23 Section VI-A] , we adopt a pump power of P p u m p = 50 P_{\mathrm{pump}} = 50 P pump = 50 mW and a maximum flow rate of Q p u m p = 14 Q_{\mathrm{pump}} = 14 Q pump = 14 mL/min. Assuming an effective fluid volume of V s w = 0.2 V_{\mathrm{sw}} = 0.2 V sw = 0.2 μL per pattern transition, τ s w \tau_{\mathrm{sw}} τ sw and ε s w \varepsilon_{\mathrm{sw}} ε sw are respectively set as τ s w = V s w Q p u m p ≃ 0.86 \tau_{\mathrm{sw}} = \frac{V_{\mathrm{sw}}}{Q_{\mathrm{pump}}} \simeq0.86 τ sw = Q pump V sw ≃ 0.86 ms and ε s w = P p u m p τ s w ≃ 42.9 \varepsilon_{\mathrm{sw}} = P_{\mathrm{pump}}\tau_{\mathrm{sw}} \simeq42.9 ε sw = P pump τ sw ≃ 42.9 μJ. The power-related parameters follow [41 Table I and Equation (13)] , and all results are averaged over 1,000 Monte Carlo realizations, with every convergence tolerance of 10 − 4 10^{-4} 1 0 − 4 , which satisfies T c − ϵ T > τ s w T_{\mathrm{c}}-\epsilon_T > \tau_{\mathrm{sw}} T c − ϵ T > τ sw .
For baselines, we compare the proposed tri-hybrid MPRA (TH-MPRA) framework with the following schemes:
Partially-connected hybrid beamforming MA scheme (HBF-MA) [33 ] : R \mathbf{R} R and ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) are optimized, assuming that no pattern-reconfigurable ability on FPA, inspired by [33 ] . This incurs mechanical movement time and energy but no pattern-switching overhead.
Tri-hybrid pattern-reconfigurable-antenna scheme (TH-PRA) [19 ] : Z \mathbf{Z} Z (hence F s e l ( Z ) \mathbf{F}_{\mathrm{sel}}(\mathbf{Z}) F sel ( Z ) ) and ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) are optimized, while R \mathbf{R} R is fixed as R 0 \mathbf{R}^{0} R 0 . This incurs pattern-switching overhead but no movement cost.
Tri-hybrid beamforming with FPA (TH-FPA) [27 ] : R \mathbf{R} R is fixed as R 0 \mathbf{R}^{0} R 0 , ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) and corresponding EM-domain precoder F ~ E M \widetilde{\mathbf{F}}_{\mathrm{EM}} F EM are jointly optimized by [27 ] , with no pattern-reconfigurable ability on FPAs on BS.
Reliability and Optimality of the Proposed Framework
Figure 2 provides a physical interpretation of optimized configurations. In Figure 2 a, every optimized R \mathbf{R} R remains inside its assigned { C n } \{\mathcal{C}_n\} { C n } , while preserving d min d_{\min} d m i n -separation. The MPRA therefore exploits the spatial DoF by proposed framework without producing collision violations. Figure 2 b further demonstrates how optimizing Z \mathbf{Z} Z works synergistically with channel environment. The left-panel depicts given { θ k , l , ϕ k , l } \{\theta_{k,l},\phi_{k,l}\} { θ k , l , ϕ k , l } , and the right-panel shows a corresponding ψ ‾ n ( Ω ) \overline{\psi}_n(\Omega) ψ n ( Ω ) selected by the proposed framework [21 , 29 ] . The selected ψ ‾ n ( Ω ) \overline{\psi}_n(\Omega) ψ n ( Ω ) redistributes the radiated gain toward favorable angular regions while attenuating less useful directions. Overall, these validate the physical feasibility and effectiveness of the proposed framework, which returns reasonable results thereafter.
Figure 2. Representative configurations optimized by the proposed framework.
Figure 3 evaluates the optimality of the proposed TH-MPRA framework against BFS, which exhaustively searches all configurations with the continuous variables discretized over a fine grid, for two tractable instances. As shown, the cumulative
Figure 3. Empirical CDFs of η E E \eta_{\mathrm{EE}} η EE achieved by the BFS reference, the proposed TH-MPRA framework, and baselines. (a) N = 3 N = 3 N = 3 , N R F = 3 N_{\mathrm{RF}} = 3 N RF = 3 , K = 2 K = 2 K = 2 , and L = 5 L = 5 L = 5 ; and (b) N = 4 N = 4 N = 4 , N R F = 2 N_{\mathrm{RF}} = 2 N RF = 2 , K = 2 K = 2 K = 2 , and L = 6 L = 6 L = 6 .
distribution function (CDF) of the proposed TH-MPRA framework almost overlaps with that of the BFS, demonstrating that it attains near-optimal EE with substantially lower complexity compared to BFS.15 In contrast, the HBF-MA curve is shifted markedly to the left since it cannot exploit pattern-domain reconfiguration. TH-PRA closes a considerable portion of this gap, while the remaining separation from the proposed framework confirms the additional benefit of optimizing R \mathbf{R} R .
Figure 4 verifies the convergence reliability of the proposed framework. Figure 4 a plots the successive change of ∣ λ D ( i + 1 ) − λ D ( i ) ∣ / W \lvert\lambda_{\mathrm{D}}^{(i+1)} - \lambda_{\mathrm{D}}^{(i)} \rvert/W ∣ λ D ( i + 1 ) − λ D ( i ) ∣ / W , which rapidly vanishes as ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) -subproblem approaches its stationary solution. Figure 4 b shows a similar decay in ∣ η R ( q + 1 ) − η R ( q ) ∣ / W \lvert\eta_{\mathrm{R}}^{(q+1)} - \eta_{\mathrm{R}}^{(q)} \rvert/W ∣ η R ( q + 1 ) − η R ( q ) ∣ / W , confirming stable convergence of the R \mathbf{R} R -subproblem. For Z \mathbf{Z} Z -subproblem, the upper-panel of Figure 4 c reports the number of antenna-state changes during each complete sweep over N N N MPRAs ((p − 1 ) N ∼ p N p-1)N \sim pN p − 1 ) N ∼ pN ). The lower panel shows the corresponding sweep-wise variation in the normalized EE. Specifically, both ∑ i = ( q − 1 ) N + 1 q N b p a t [ i ] \sum_{i=(q-1)N+1}^{qN} b_{\mathrm{pat}}^{[i]} ∑ i = ( q − 1 ) N + 1 q N b pat [ i ] and ∣ η E E [ q N ] − η E E [ ( q − 1 ) N ] ∣ / W \lvert\eta_{\mathrm{EE}}^{[qN]}-\eta_{\mathrm{EE}}^{[(q-1)N]}\rvert/W ∣ η EE [ q N ] − η EE [( q − 1 ) N ] ∣ / W progressively approach zero, indicating that the cyclic pattern search reaches a coordinate-wise stationary point. Finally, Figure 4 d also shows that ∣ η E E ( t ) − η E E ( t − 1 ) ∣ / W \lvert\eta_{\mathrm{EE}}^{(t)}-\eta_{\mathrm{EE}}^{(t-1)}\rvert/W ∣ η EE ( t ) − η EE ( t − 1 ) ∣ / W decreases to zero within only a few iterations. Overall, the stable convergence behavior clearly demonstrates the reliability.
Figure 4. (a) ( F B B , F R F ) (\mathbf{F}_{\mathrm{BB}},\mathbf{F}_{\mathrm{RF}}) ( F BB , F RF ) -update. (b) R \mathbf{R} R -update. (c) Z \mathbf{Z} Z -update. (d) Outer AO framework.
Figure 5 evaluates initialization sensitivity using ρ E E ≜ η E E ( R 0 , Z 0 ) / η E E , 0 \rho_{\mathrm{EE}}\triangleq\eta_{\mathrm{EE}(\mathbf{R}^{0},\mathbf{Z}^{0})}/\eta_{\mathrm{EE},0} ρ EE ≜ η EE ( R 0 , Z 0 ) / η EE , 0 , where η E E , 0 \eta_{\mathrm{EE},0} η EE , 0 corresponds to the reference initialization in Section IV-A. Herein, the CDFs concentrate around ρ E E = 1 \rho_{\mathrm{EE}}=1 ρ EE = 1 ; provides representative, although not uniformly optimal, performance. Randomizing only Z 0 \mathbf{Z}^{0} Z 0 causes the smallest variation since pattern switching incurs relatively lower overhead and the subsequent Z \mathbf{Z} Z -update effectively corrects unfavorable states. Conversely, randomizing only R 0 \mathbf{R}^{0} R 0 produces a wider distribution by altering both channel and movement cost. Its close overlap with the jointly randomized case confirms that sensitivity is dominated by R 0 \mathbf{R}^{0} R 0 . The narrower distributions for N R F = 8 N_{\mathrm{RF}}=8 N RF = 8 further indicate that additional RF-domain flexibility mitigates unfavorable initialization, though its higher P c i r c P_{\mathrm{circ}} P circ does not improve EE.
Figure 5. Empirical CDFs of the EE sensitivity under random ( R 0 , Z 0 ) (\mathbf{R}^{0},\mathbf{Z}^{0}) ( R 0 , Z 0 ) .
Figure 6 a compares the bandwidth-normalized EE, η E E / W \eta_{\mathrm{EE}}/W η EE / W versus P max P_{\max} P m a x . As shown, η E E / W \eta_{\mathrm{EE}}/W η EE / W initially increases since the SE gain outweighs the additional PA consumption, but saturates since the EE-optimal transmit power is reached and ∥ F B B ∥ F 2 ≤ P max \lVert\mathbf{F}_{\mathrm{BB}}\rVert_{F}^{2}\le P_{\max} ∥ F BB ∥ F 2 ≤ P m a x becomes inactive. The proposed TH-MPRA framework consistently achieves the highest η E E / W \eta_{\mathrm{EE}}/W η EE / W ; its gain over TH-PRA demonstrates the net benefit of optimizing R \mathbf{R} R , even after accounting for E m o v E_{\mathrm{mov}} E mov . The observed improvement therefore indicates that the additional spatial DoF can be effectively exploited to improve η E E / W \eta_{\mathrm{EE}}/W η EE / W despite the associated movement energy overhead. Meanwhile, the larger separation from HBF-MA and TH-FPA further highlights the effectiveness of adapting Z \mathbf{Z} Z . In addition, N R F = 4 N_{\mathrm{RF}}=4 N RF = 4 outperforms N R F = 8 N_{\mathrm{RF}}=8 N RF = 8 since the additional DAC and RF-chain power consumption exceeds the corresponding SE gain, demonstrating that greater RF-domain flexibility does not necessarily improve η E E / W \eta_{\mathrm{EE}}/W η EE / W .
Figure 6. η E E / W \eta_{\mathrm{EE}}/W η EE / W versus (a) P max P_{\max} P m a x (b) the actual transmit power ∥ F B B ∥ 2 \lVert\mathbf{F}_{\mathrm{BB}}\rVert_2 ∥ F BB ∥ 2 (c) S S S and (d) T c T_c T c .
Figure 6 b further reveals the EE-power tradeoff by plotting η E E / W \eta_{\mathrm{EE}}/W η EE / W against the actual transmit power ∥ F B B ∥ F 2 \lVert\mathbf{F}_{\mathrm{BB}}\rVert_{F}^{2} ∥ F BB ∥ F 2 . Unlike Figure 6 a, each curve is unimodal; η E E / W \eta_{\mathrm{EE}}/W η EE / W first increases since additional ∥ F B B ∥ F 2 \lVert\mathbf{F}_{\mathrm{BB}}\rVert_{F}^{2} ∥ F BB ∥ F 2 substantially improves the achievable SEs, reaches its maximum at an intermediate power level, and then decreases as the PA energy grows approximately linearly whereas the SE increases only logarithmically. The peak for N R F = 8 N_{\mathrm{RF}}=8 N RF = 8 occurs at a slightly higher ∥ F B B ∥ F 2 \lVert\mathbf{F}_{\mathrm{BB}}\rVert_{F}^{2} ∥ F BB ∥ F 2 than N R F = 4 N_{\mathrm{RF}}=4 N RF = 4 , since a larger radiated power is required to amortize the higher fixed circuit-power expenditure of the additional RF chains; nevertheless, its maximum η E E / W \eta_{\mathrm{EE}}/W η EE / W remains lower. The proposed TH-MPRA framework retains a clear advantage over the baselines, since Z \mathbf{Z} Z -update concentrates the radiated energy along favorable multipath directions, while R \mathbf{R} R -update provides a further channel-conditioning gain.
Figure 6 c illustrates the effect of enlarging Ψ ‾ ( Ω ) \overline{\Psi}(\Omega) Ψ ( Ω ) . As shown, η E E / W \eta_{\mathrm{EE}}/W η EE / W of the proposed TH-MPRA framework increases noticeably when S S S grows from 8 to 16, since a richer codebook provides finer angular selectivity and increases the probability of finding ψ ‾ s ( Ω ) \overline{\psi}_{s}(\Omega) ψ s ( Ω ) that reinforces the favorable multipath components while suppressing interference. The improvement becomes marginal beyond S = 32 S=32 S = 32 , however, since the dominant angular regions have already been covered [19 , 21 , 23 , 29 ] , leads to a clear diminishing return. By construction, HBF-MA and TH-FPA do not optimize Z \mathbf{Z} Z and therefore remain essentially invariant with S S S . The persistent gap between the proposed TH-MPRA framework and TH-PRA represents the complementary gain obtained from R \mathbf{R} R -update. These observations suggest that even a moderate S = 16 S=16 S = 16 or 32 offers an attractive performance-search-complexity tradeoff.
Figure 6 d illustrates how T c T_{c} T c affects the cost effectiveness. Herein, η E E / W \eta_{\mathrm{EE}}/W η EE / W of the proposed TH-MPRA and HBF-MA increases with T c T_{c} T c since T m o v T_{\mathrm{mov}} T mov and E m o v E_{\mathrm{mov}} E mov are amortized over a longer data interval. In contrast, TH-PRA varies only slightly since T p a t T_{\mathrm{pat}} T pat is much smaller than T m o v T_{\mathrm{mov}} T mov herein (T p a t = 0.86 m s T_{\mathrm{pat}}=0.86\ \mathrm{ms} T pat = 0.86 ms vs. T m o v ≈ 10.2 m s T_{\mathrm{mov}}\approx10.2\ \mathrm{ms} T mov ≈ 10.2 ms ), whereas TH-FPA remains constant since it incurs no reconfiguration overhead. TH-MPRA nevertheless achieves the highest η E E / W \eta_{\mathrm{EE}}/W η EE / W even for short T c T_c T c , indicating that its joint optimization gain compensates for the associated overhead. Moreover, the persistent advantage of N R F = 4 N_{\mathrm{RF}}=4 N RF = 4 shows that increasing T c T_c T c does not offset P c i r c P_{\mathrm{circ}} P circ .
Figure 7 a shows how N N N affects η E E / W \eta_{\mathrm{EE}}/W η EE / W . When N N N increases from 9 to 16, the array gain improves the effective channel strength and interference-suppression capability, yielding a clear increase in η E E / W \eta_{\mathrm{EE}}/W η EE / W . Beyond this, however, the achievable-SE gain exhibits diminishing returns, whereas P c i r c P_{\mathrm{circ}} P circ continues to grow with N N N . Consequently, the proposed TH-MPRA framework, though it maintains superior performance over baselines, attains their largest η E E / W \eta_{\mathrm{EE}}/W η EE / W around N = 16 N=16 N = 16 and then gradually decreases, and so is TH-PRA. This degradation is more pronounced for HBF-MA and TH-FPA since they cannot use pattern reconfiguration to extract sufficient channel gain from the enlarged N N N . The result therefore reveals an η E E \eta_{\mathrm{EE}} η EE -optimal N N N rather than an “as-large-as-possible” rule.
Figure 7. η E E / W \eta_{\mathrm{EE}}/W η EE / W versus (a) N N N and (b) K K K (c) L L L and (d) A / λ A/\lambda A / λ .
Figure 7 b depicts the effect of K K K .16 As shown, η E E / W \eta_{\mathrm{EE}}/W η EE / W increases with K K K since the spatial multiplexing delivers more aggregate information during the same coherence block and amortizes P c i r c P_{\mathrm{circ}} P circ and reconfiguration costs over a larger SE. The increase becomes progressively smaller at large K K K owing to stronger multi-user interference, transmit-power sharing, and the finite DoF. The proposed TH-MPRA framework consistently outperforms the benchmarks, with the gap over TH-PRA reflecting the additional channel orthogonalization provided by position control and the larger gaps over HBF-MA and TH-FPA demonstrating the importance of pattern-domain interference management. Over K = 2 K=2 K = 2 –4 4 4 , N R F = 4 N_{\mathrm{RF}}=4 N RF = 4 achieves
higher η E E / W \eta_{\mathrm{EE}}/W η EE / W because of its lower circuit-power expenditure.
Figure 7 c demonstrates that richer multipath is beneficial but produces only a moderate improvement on η E E / W \eta_{\mathrm{EE}}/W η EE / W . As L L L increases, the additional directions provide greater spatial diversity [18 ] , [19 ] , allowing TH-MPRA to jointly select favorable R \mathbf{R} R and Z \mathbf{Z} Z . The gain gradually saturates since the channel model normalizes the path powers with respect to L L L , hence additional paths increase diversity rather than the received power; moreover, newly added paths can be angularly correlated, or interference generating. Nevertheless, the proposed TH-MPRA framework maintains the largest η E E / W \eta_{\mathrm{EE}}/W η EE / W throughout, with its persistent gains over TH-PRA, HBF-MA, and TH-FPA highlighting the benefits of jointly exploiting reconfiguration of R \mathbf{R} R and Z \mathbf{Z} Z , which provides a complementary phase-domain DoF even in rich scattering scenarios.
Figure 7 d illustrates the effect of A / λ A/\lambda A / λ . Herein, the result reflects the combined effects of array geometry and, for frameworks with position adaptation, antenna movement. For TH-MPRA, increasing A / λ A/\lambda A / λ from 0.6 to 0.8–0.9 improves η E E / W \eta_{\mathrm{EE}}/W η EE / W , as the expanded array geometry and positioning flexibility can enhance spatial phase diversity and user separation. As A / λ A/\lambda A / λ approaches 1, however, it yields a slightly lower EE. Larger positioning regions can also entail longer movement distances, increasing T m o v T_{\mathrm{mov}} T mov and E m o v E_{\mathrm{mov}} E mov , although their contribution cannot be isolated from the accompanying geometry changes in this experiment. Accordingly, A / λ = 0.8 A/\lambda=0.8 A / λ = 0.8 –0.9 0.9 0.9 provides the most favorable balance under the considered setup. Throughout the evaluated range, TH-MPRA consistently outperforms the baselines by jointly exploiting position and pattern adaptation.
Conclusion
In this paper, we proposed an energy-efficient multi-user framework with tri-hybrid MPRA. A hardware-inspired power and latency model captured circuit consumption, mechanical and pattern actuation, and transmission-time overhead. A unified AO framework addressed the coupled EE maximization through block-specific optimization techniques, guaranteeing a nondecreasing and convergent EE sequence. Simulations demonstrated performance close to exhaustive search and consistent EE gains over the baselines. These highlight the potential of joint spatial and pattern reconfiguration on MA for energy-efficient wireless network in 6G and beyond.