betaFeedback

WorksEngineering

An Efficient SIM Beamforming Scheme Based on Physical-Empowered Projected Gradient

Abstract

Stacked Intelligent Metasurfaces (SIMs) have emerged as a promising technology for next-generation wireless systems. Especially in multi-input multi-output (MIMO) systems, it has the potential to realize complicated wave-domain transformations with low hardware costs and energy consumption. However, SIM-assisted MIMO systems face two major challenges in channel estimation and beamforming. As the SIM elements do not have the signal processing capability, channel estimation can only be conducted in an indirect manner, which makes it challenging. Moreover, the precoding optimization after channel estimation is non-convex and also difficult. Targeting these issues, we exploit the time-division duplex (TDD) channel reciprocity to calculate the gradient backpropagation in an over-the-air manner. Based on this structure, we propose a gradient-based optimization framework that optimizes beamforming without explicitly estimating the SIM–UE channel. The proposed method, namely the physical projected gradient descent (P2GD) method, optimizes the achievable sum rate while incorporating zero-forcing (ZF) constraints. Based on the method of incorporating the ZF restriction, we propose two variations of P2GD, namely user equipment (UE)-P2GD and base station (BS)-P2GD. We also provide a theoretical convergence analysis and projected gradient techniques to ensure physical feasibility. Simulation results show that the proposed approach can achieve suboptimal beamforming in the multi-user system within a small number of over-the-air iterations, using a per-direction signaling budget below the pilot cost of explicit SIM–UE channel estimation. The proposed method provides a low-overhead beamforming scheme for SIM systems without explicit SIM–UE channel estimation.

Introduction

Motivation

Stacked Intelligent Metasurfaces (SIMs) have recently been proposed as a promising technology for future wireless communications. By leveraging low-cost stacking operations of multiple metasurface layers, SIMs can realize complex linear transformations with significantly reduced hardware complexity. This property makes them attractive candidates for replacing conventional high-cost precoding modules in MIMO systems, thereby offering an energy- and cost-efficient solution for large-scale deployments [2]. With the introduction of SIMs, one can realize complicated precoding and combining functions at a low hardware and energy cost.

However, SIMs also introduce additional scenario complexity and controllable variables, which increase the complexity in channel estimation and parameter optimization [2]. Typically, the channel gains among the SIM layers can be accurately modeled through near-field channel models. However, since it is difficult to obtain the raw received signals at the SIM elements due to their lack of inherent signal processing capability, the typical implementation is to estimate the channel between the SIM and both sides of the transceiver in an indirect manner.

In this paper, we consider a time division duplex (TDD) multiuser MIMO system in which a multi-antenna base station (BS) is assisted by a SIM located nearby. In the considered architecture, the SIM is co-located with the BS and is regarded as an integrated part of the BS transceiver. Since the relative geometry between the BS antennas and the SIM, as well as the internal SIM structure, is fixed and controlled by the BS, the corresponding BS–SIM propagation channel is assumed to be known through physical modeling and calibration. In contrast, the SIM–user equipment (UE) channels propagate through the external wireless environment and are treated as unknown. Consequently, conventional channel estimation results in prohibitive overhead. Meanwhile, the overall input–output relationship of SIM-assisted systems is governed by numerous parameters with a highly non-convex structure, rendering the design of optimal wave-domain precoders intractable even under perfect channel estimation. Thus, suboptimal gradient descent methods are usually adopted to solve such problems [3].

Such works inspire us to explore the SIM precoding problem from the perspective of gradient methods. Particularly, in our previous works [4], [5], we have discussed the intrinsic relationship between channel reciprocity and gradient backpropagation, and we have utilized channel reciprocity for neural network training. Since beamforming optimization can also be performed using gradient methods, this reciprocity-based mechanism can be applied to SIM beamforming. Specifically, we develop a gradient-based optimization framework that jointly tackles channel acquisition and precoding. Instead of estimating the channel separately, the optimization is performed directly on system-level performance metrics, such as the achievable sum rate. Interference management is realized by transmitting orthogonal pilot signals over time, which enables the estimation of inter-user interference power and facilitates the incorporation of a zero-forcing (ZF) penalty into the optimization objective.

With the optimization objective, we exploit the channel reciprocity of TDD systems to compute gradients through reverse transmission, thereby avoiding explicit channel estimation and reducing signaling overhead [4]. Following a typical backpropagation procedure, we design the feedback signal to ensure that the uplink received signals are exactly the gradient of the corresponding transmitted signal, assuming perfect TDD and neglecting background noise. It should be noted that this physical gradient relation relies on sufficiently accurate TDD reciprocity between the forward and reverse transmissions. In practical systems, transmit–receive RF-chain mismatch, phase and timing synchronization errors, carrier-frequency offsets, and imperfect reciprocity calibration may cause deviations between the effective uplink and downlink channels, and consequently distort the physically propagated gradient. Therefore, reciprocity calibration and synchronization are required for practical implementation of the proposed framework. Besides such hardware impairments, UE mobility between the forward and backward transmissions also breaks reciprocity in time, so that the returned gradient is rotated by the temporal channel correlation rather than merely perturbed by noise [6], [7], [8]. This effect can be compensated by movable-antenna repositioning [6], by correcting the rotation of the returned signal [7], or by timing the gradient return [8]. In this work, we focus on the fundamental algorithmic principle under ideal TDD reciprocity, with the SIM–UE channel unchanged within each forward–backward exchange, while explicitly modeling and compensating for such hardware and calibration imperfections is left for future investigation. Regarding the fact that we cannot recover the received signal on each SIM layer due to hardware restrictions, we instead adopt a projected gradient descent strategy [5], which updates the gradient in a specific subspace in each step. As we only need to improve the overall equivalent channel, the large parameter space of SIMs is actually redundant. Thus, the projected gradient can also efficiently optimize the overall system performance.

Generally, our proposed method ensures that the SIM configuration remains physically realizable and an acceptable communication overhead while maintaining the overall beamforming performance. By iteratively exchanging gradient-based feedback between the BS and the UEs, the proposed method refines both the SIM control parameters and the precoding matrices in a mutually coupled manner. Such a joint learning-and-optimization process alleviates the burden of explicit channel estimation and also provides a systematic means of adapting SIM-assisted systems to dynamic wireless environments, thereby enabling significant throughput gains under practical overhead and complexity constraints. Beyond conventional beamforming, our approach further emphasizes channel tracking based on the proposed algorithm, enabling continuous adaptation of the channel representation as it evolves over time. This design allows the transceiver to maintain robust performance under dynamic propagation conditions, effectively capturing temporal variations in the wireless environment while preserving spatial correlation and large-scale fading consistency.

Metasurfaces and Their Applications in Beamforming

Early work on metasurfaces demonstrated that large, nearly passive apertures can shape electromagnetic (EM) fields for beam steering and interference control with much lower power than fully digital arrays. Single-layer reconfigurable intelligent surfaces (RIS) [9] and reconfigurable holographic surfaces (RHS) [10] improved link budgets and enabled passive beamforming, but their one-layer structure limited the range of achievable radiation patterns and multiuser nulling. Recent studies therefore turned to SIM, which cascade multiple transmissive layers to realize wave-domain processing, effectively implementing precoding/combining as waves propagate, thereby reducing radio frequency (RF)-chain counts while enriching spatial degrees of freedom. Representative results include efficient radiation-pattern synthesis and received-power maximization, where multi-layers concentrate most power on targets and generate complex beampatterns, validating the expressiveness of stacking [3], [11], [12]. Building on this, SIMs have been used for multiuser downlink beamforming [3], where alternating-optimization (AO) schemes jointly tune SIM phases and transmit power to maximize sum rate, and SIM shows clear gains over single-RIS baselines and conventional arrays. Extension applications include holographic MIMO [2], cell-free architectures [13], [14], integrated sensing and communication (ISAC) [15], [16], satellite communications [17] and semantic system [18], [19].

However, two limitations emerge across these beamforming studies and motivate our approach. The first is the intrinsic difficulty of SIM precoding. The objective couples many constant modulus controls across layers with the transmit covariance, which yields a highly nonconvex landscape. As the number of layers increases, nested matrix operations and variable coupling drive the computational burden upward, so most practical designs rely on alternating optimization, projected gradient updates, or learning heuristics [20], [21], [22]. The second limitation is the reliance on channel state information (CSI). SIM based links present very high dimensional channels, often with only a few RF chains, making acquisition underdetermined and pilot hungry [22], [23]. Beyond conventional model-based approaches, Chen et al. [24] developed a pioneering diffusion model-based variational inference framework that effectively exploits powerful generative channel priors to significantly improve MIMO channel estimation accuracy while reducing pilot overhead, representing a remarkable advancement in efficient channel acquisition for next-generation wireless systems. Thus, the hardness of precoding and the cost of reliable CSI form a bottleneck and naturally turn attention to channel estimation in SIM-aided communication systems.

Channel Estimation in SIM-assisted Systems

Because SIM beamforming presumes knowledge of the wave-domain channel while operating with many meta-atoms and few RF chains, recent work has concentrated on extracting usable channel structure with limited measurements and training time. A representative approach accumulates multiple pilot observations and exploits orthogonal user pilots to solve the underdetermined problem, then applies model-based estimators such as least squares (LS) and minimum mean square error (MMSE) [25]. Building on the sparse geometry of high-frequency propagation, another stream formulates SIM channel recovery as structured compressed sensing with block and hierarchical matching pursuit, where the SIM configuration is optimized to lower dictionary coherence and sharpen polar-domain sparsity [26]. To further reduce the number of RF chains without sacrificing accuracy, hybrid wave-digital estimators process pilots through the stacked layers before digital combining, reaching performance close to fully digital baselines while introducing special training phases in practice [23]. Beyond matrix models, recent studies cast the SIM input–output as a multi-way structure and apply nested decomposition approach to jointly estimate channels and calibrate inter-layer coefficients, achieving marked performance improved at a complexity comparable to classical estimators [27]. Learning-based designs complement these ideas by training a one dimensional convolutional network that maps preamble observations to channel vectors using constructed real-imaginary-magnitude features, and report competitive accuracy relative to MMSE and LS baselines under appropriate training [28]. More recently, Chen et al. Chen et al. [29] made a groundbreaking advance in generative channel estimation by pioneering a sampling-free diffusion transformer that eliminates the need for computationally intensive iterative sampling. This remarkable innovation substantially reduces computational complexity while maintaining high estimation accuracy, representing a significant milestone toward practical and efficient diffusion-based MIMO channel estimation. These techniques demonstrate steady progress yet also reveal a persistent burden that comes from near-field effects, large apertures, and the channel dimensionality, inflating pilot signaling and computation despite phase optimization and sophisticated estimators. This tension strengthens the case for designs that maintain effective wave-domain beam control while reducing explicit CSI requirements and training overhead.

3) Over-the-Air Distributed Optimization Methods: Over-the-air computation (OAC) has emerged as a promising technique to address communication bottlenecks in distributed optimization, especially in federated learning (FL). In OAC, we consider the transmission through the wireless channels as a linear calculation. Through carefully designed modulation and precoding, OAC can realize some basic linear operations. One of the most common OAC applications lies in FL [30, 31], which introduces the concept of leveraging wireless channel superposition to perform simultaneous model aggregation, and significantly reduces latency compared to conventional digital transmission. Subsequent studies have focused on improving robustness and efficiency under practical constraints. Under the OAC-FL concept, [32] analyzed bias-variance trade-offs in heterogeneous OAC-FL, while [33] proposed joint uplink-downlink optimization for these systems. Enhancements such as sparsification, mixed-precision transmission, and secure aggregation have been explored to mitigate channel noise and privacy risks [34].

Beyond FL, OAC techniques have been integrated into general distributed optimization frameworks. [35] demonstrated accelerated convergence for convex optimization using OAC-based aggregation, and [36] extended these ideas to dynamic optimization scenarios. OAC has also been combined with consensus algorithms and ADMM variants to reduce communication complexity in multi-agent systems [37]. Existing survey papers [37] provide comprehensive overviews of distributed optimization methods, highlighting OAC as a key enabler for large-scale, low-latency applications.

Foundational works such as [38] further contextualize OAC within the broader landscape of networked optimization and wireless computation. Recent studies have also explored OAC-based distributed optimization under channel uncertainty and robustness constraints [39], as well as privacy-preserving mechanisms for sensitive applications [40]. Recent literature emphasizes the integration of OAC with emerging technologies, including MIMO, RIS, and other edge intelligence methods, to enhance scalability and robustness [39, 41, 42]. These approaches leverage advanced physical-layer techniques to improve aggregation accuracy and energy efficiency. For example, RIS-assisted OAC computation can dynamically shape wireless channels for better aggregation performance [39], while MIMO-based OAC schemes exploit spatial diversity for large-scale distributed learning [41, 42]. Despite these advances, challenges remain in handling heterogeneous device capabilities, ensuring privacy, and maintaining robustness under imperfect CSI, which continue to motivate ongoing research [43].

On the other hand, our previous works also discussed OAC over unknown wireless channels. Under the TDD condition, we can characterize the relationship between the uplink and the corresponding downlink channels. Although remaining unknown, these coupled variants can be used in dual computation tasks, such as forward calculation and gradient backpropagation [4, 5]. Our recent works further extend this reciprocity-based backpropagation to time-varying channels, where user mobility between the forward and backward transmissions distorts the returned gradient [6, 7, 8]. Meanwhile, considering the utilization of CSI information in native wireless problems, we can further regard the transmission as a part of graph neural network computation [44]. Different from conventional gradient-based beamforming methods that rely on explicitly estimated CSI, the proposed framework obtains the required gradient information through reciprocal over-the-air transmission. Moreover, unlike standard OAC scenarios that mainly exploit wireless superposition for aggregation, the SIM-assisted setting involves a highly non-convex cascaded physical mapping, while the intermediate signals inside the SIM are not directly observable, which motivates the projected-gradient reconstruction adopted in this work.

Recent studies in adjacent learning problems have also highlighted that efficiency and robustness should be jointly considered with security under model or distribution variations [45, 46]. Moreover, since the proposed framework transmits gradient-related analog signals over the air, such transmissions may potentially expose information associated with local observations or channel-dependent updates, raising privacy and security concerns similar to those studied in secure over-the-air federated learning [47]. A systematic treatment of these issues is beyond the scope of this work and is left for future investigation.

Contributions

In this paper, we propose a gradient-based optimization framework for TDD SIM-assisted MIMO systems, with the goal of jointly addressing the challenges of channel acquisition and precoding under realistic system constraints. Unlike conventional approaches that rely on explicit channel estimation and incur excessive estimation overhead, our method directly optimizes system-level performance metrics by exploiting channel reciprocity and iterative gradient feedback between the BS and UEs. In addition to algorithm design, we provide theoretical analysis and simulation-based validations, thereby establishing a comprehensive foundation for the practical deployment of SIM-assisted transceivers. The main contributions of this work are summarized as follows:

  • We develop a gradient-based framework that explicitly accounts for implementation constraints in practical TDD SIM-assisted systems, effectively alleviating the overhead associated with conventional channel estimation and precoding.

  • To incorporate ZF conditions, we propose two practical algorithm variants that provide flexible optimization strategies for different interference and system requirements.

  • We conduct analytical studies on the convergence properties and performance behavior of different parts of the proposed method, offering general guidelines for future algorithm design in similar high-dimensional optimization scenarios.

  • We carry out comprehensive simulations that benchmark the proposed algorithms against conventional schemes, demonstrating that the proposed method can approach the performance of the traditional methods.

Notations

In this paper, matrices are denoted by boldface uppercase letters such as X\mathbf{X}, vectors are denoted by boldface italic lowercase letters such as y\mathbf{y}, and scalars are represented by nonbold letters. In particular, the vector xi\mathbf{x}_i denotes the ii-th column of the corresponding matrix X\mathbf{X}. The notation diag⁡(⋯ )\operatorname{diag}(\cdots) denotes a diagonal matrix, while AT\mathbf{A}^{\mathsf{T}}, AH\mathbf{A}^{\mathsf{H}}, A∗\mathbf{A}^{*}, and A†\mathbf{A}^{\dagger} represent the transpose, the Hermitian transpose, the complex conjugate, and the Moore–Penrose pseudoinverse of A\mathbf{A}, respectively.

System Model and Problem Formulation

SIM Model

We consider a communication system as illustrated in Figure 1 where the BS is equipped with NN transmit antennas serving NN UEs, whose index set is denoted as N={1,2,…,N}\mathcal{N} = \{1, 2, \ldots, N\}. The BS employs all NN antennas to simultaneously deliver NN independent data streams to the corresponding UEs. For analytical simplicity and following the commonly adopted SIM beamforming setting in [3], we assume that the number of active transmit antennas equals the number of simultaneously served UEs. In a more general system with a larger number of antennas or dynamically varying users, antenna/user selection and scheduling can be performed prior to the considered beamforming procedure.

An illustration of the considered SIM-aided wireless MIMO system

Figure 1. An illustration of the considered SIM-aided wireless MIMO system

The SIM architecture comprises LL parallel metasurface layers, indexed by the set L={1,2,…,L}\mathcal{L} = \{1, 2, \ldots, L\}, where each layer contains MM reconfigurable meta-atoms with M>NM > N. The meta-atom index set for each layer is represented by M={1,2,…,M}\mathcal{M} = \{1, 2, \ldots, M\}. For the mm-th meta-atom on the ll-th layer, we denote its complex transmission coefficient as ϕl,m\phi_{l,m}, where ∣ϕl,m∣≤1|\phi_{l,m}| \le1. Accordingly, the transmission coefficient vector of the ll-th layer is given by ϕl=[ϕl,1,ϕl,2,…,ϕl,M]T∈CM×1\boldsymbol{\phi}_l = [\phi_{l,1}, \phi_{l,2}, \ldots, \phi_{l,M}]^{\mathsf{T}} \in\mathbb{C}^{M \times1}, and its diagonal matrix representation is Φl=diag⁡(ϕl,1,…,ϕl,M)∈CM×M\boldsymbol{\Phi}_l = \operatorname{diag}(\phi_{l,1}, \ldots, \phi_{l,M}) \in\mathbb{C}^{M \times M}.

Each metasurface layer adopts a uniform planar array (UPA) topology, with all layers sharing identical square lattice arrangements. Denoting ded_e as the inter-element spacing within each layer, the Euclidean distance between the mm-th and m~\widetilde{m}-th meta-atoms on any given layer is computed as

dm,m~=de(mx−m~x)2+(my−m~y)2,(1)d_{m,\widetilde{m}} = d_e \sqrt{(m_x-\widetilde{m}_x)^2 + (m_y-\widetilde{m}_y)^2}, \tag*{(1)}

where (mx,my)(m_x,m_y) and (m~x,m~y)(\widetilde{m}_x,\widetilde{m}_y) denote the horizontal and vertical coordinates of the mm-th and m~\widetilde{m}-th meta-atoms, respectively. These coordinates are determined by mx=mod⁡(m−1,Mrow)+1m_x = \operatorname{mod}(m-1,M_{\mathrm{row}})+1 and my=⌈m/Mrow⌉m_y = \lceil m/M_{\mathrm{row}}\rceil, with Mrow=MM_{\mathrm{row}} = \sqrt{M} being the number of meta-atoms per row.

The SIM structure has an overall thickness of DD, within which the LL metasurface layers are uniformly distributed with an inter-layer separation of dlayer=D/(L−1)d_{\mathrm{layer}} = D/(L-1). Consequently, the spatial distance between the mm-th meta-atom on layer ll and the m~\widetilde{m}-th meta-atom on layer (l+1)(l+1) is given by dm,m~(l)=dm,m~2+dlayer2d^{(l)}_{m,\widetilde{m}} = \sqrt{d^2_{m,\widetilde{m}} + d^2_{\mathrm{layer}}}, l∈Ll \in\mathcal{L}, and m,m~∈Mm,\widetilde{m} \in\mathcal{M}.

At the BS side, the NN transmit antennas form a uniform linear array (ULA) with half-wavelength spacing, whose center is co-located with the center of the first metasurface layer. Therefore, the distance from the nn-th transmit antenna to the mm-th meta-atom on the first layer is expressed as

dn,m(0)=([(my−Mrow+12)de−(n−N+12)λc2]2+(mx−Mrow+12)2de2+dlayer2)12,(2)d^{(0)}_{n,m} = \left(\left[\left(m_y-\frac{M_{\mathrm{row}}+1}{2}\right)d_e-\left(n-\frac{N+1}{2}\right)\frac{\lambda_c}{2}\right]^2+\left(m_x-\frac{M_{\mathrm{row}}+1}{2}\right)^2d_e^2+d_{\mathrm{layer}}^2\right)^{\frac{1}{2}}, \tag*{(2)}

where λc\lambda_c denotes the carrier wavelength. Since the BS antenna array and the SIM are co-located and form a jointly controlled transceiver-side structure, the propagation coefficients between them are assumed to be calibrated and known at the BS. The EM wave propagation between consecutive metasurface layers is characterized by the Rayleigh-Sommerfeld diffraction formulation because of the near-field wave propagation across metasurface layers. Based on this physical model, the propagation coefficient from the mm-th meta-atom on layer ll to the m~\tilde{m}-th meta-atom on layer (l+1)(l+1) is given by

[Hl]m,m~=dlayerAe(dm,m~(l))2(12πdm,m~(l)−j1λc)ej2πdm,m~(l)/λc,(3)[\mathbf{H}_l]_{m,\tilde{m}}=\frac{d_{\mathrm{layer}}A_{\mathrm{e}}}{(d_{m,\tilde{m}}^{(l)})^2}\left(\frac{1}{2\pi d_{m,\tilde{m}}^{(l)}}-\mathrm{j}\frac{1}{\lambda_{\mathrm{c}}}\right)e^{\mathrm{j}2\pi d_{m,\tilde{m}}^{(l)}/\lambda_{\mathrm{c}}}, \tag*{(3)}

for l∈Ll \in\mathcal{L} and m,m~∈Mm,\tilde{m} \in\mathcal{M}, where AeA_{\mathrm{e}} represents the effective aperture area of each meta-atom. Similarly, the propagation coefficient [H0]m,n[\mathbf{H}_0]_{m,n} from the nn-th transmit antenna to the mm-th meta-atom on the first layer is obtained by substituting dm,m~(l)d_{m,\tilde{m}}^{(l)} in (3) with dn,m(0)d_{n,m}^{(0)} from (2).

Under the assumption that coupling effects between non-adjacent layers are negligible, the cascaded EM response of the entire SIM structure can be compactly written as

B=H0Φ1H1Φ2⋯HL−1ΦL∈CN×M.(4)\mathbf{B}=\mathbf{H}_0\mathbf{\Phi}_1\mathbf{H}_1\mathbf{\Phi}_2\cdots\mathbf{H}_{L-1}\mathbf{\Phi}_L\in\mathbb{C}^{N\times M}. \tag*{(4)}

SIM-UE Channel Model

The SIM-UE channel is denoted by G=[g1,…,gN]∈CM×N\mathbf{G}=[\mathbf{g}_1,\ldots,\mathbf{g}_N]\in\mathbb{C}^{M\times N}. We use a geometric multipath model in which the channel of UE nn is

gn=gˉL,n∑p=1Ppathcn,pa(φn,p,ϑn,p),(5)\mathbf{g}_n=\sqrt{\bar{g}_{L,n}}\sum_{p=1}^{P_{\mathrm{path}}}c_{n,p}\mathbf{a}(\varphi_{n,p},\vartheta_{n,p}), \tag*{(5)}

where gˉL,n\bar{g}_{L,n} is the large-scale power gain defined below, Ppath=3P_{\mathrm{path}}=3, and the independent path gains satisfy cn,p∼CN(0,1/Ppath)c_{n,p}\sim\mathcal{CN}(0,1/P_{\mathrm{path}}).

Array Response and Spatial Structure

Let rm=(0,ym,zm)T\mathbf{r}_m=(0,y_m,z_m)^{\mathsf{T}} denote the position of meta-atom mm relative to the center of the last SIM layer, with adjacent elements separated by de=λc/2d_{\mathrm{e}}=\lambda_{\mathrm{c}}/2. Here, λc\lambda_{\mathrm{c}} denotes the carrier wavelength, distinguished from the penalty coefficient λ\lambda used later. The array response is

[a(φ,ϑ)]m=exp⁡(j2πλcrmT[sin⁡ϑcos⁡φsin⁡ϑsin⁡φcos⁡ϑ]).(6)[\mathbf{a}(\varphi,\vartheta)]_m=\exp\left(\mathrm{j}\frac{2\pi}{\lambda_{\mathrm{c}}}\mathbf{r}_m^{\mathsf{T}} \begin{bmatrix} \sin\vartheta\cos\varphi\\ \sin\vartheta\sin\varphi\\ \cos\vartheta \end{bmatrix}\right). \tag*{(6)}

For each realization, the path angles are independently drawn as φn,p∼U(−2π/3,2π/3)\varphi_{n,p}\sim\mathcal{U}(-2\pi/3,2\pi/3) and ϑn,p∼U(π/6,π)\vartheta_{n,p}\sim\mathcal{U}(\pi/6,\pi). Spatial dependence across meta-atoms is induced by these shared path gains and array responses, without imposing an additional exponential spatial-correlation matrix. In particular, conditioned on the path angles,

E[gngnH∣{φn,p,ϑn,p}]=gˉL,nPpath∑p=1Ppathan,pan,pH,\mathbb{E}\left[\mathbf{g}_n\mathbf{g}_n^{H}\mid\{\varphi_{n,p},\vartheta_{n,p}\}\right]=\frac{\bar{g}_{L,n}}{P_{\mathrm{path}}}\sum_{p=1}^{P_{\mathrm{path}}}\mathbf{a}_{n,p}\mathbf{a}_{n,p}^{H},

where an,p=a(φn,p,ϑn,p)\mathbf{a}_{n,p}=\mathbf{a}(\varphi_{n,p},\vartheta_{n,p}). Electromagnetic mutual coupling within a layer is not explicitly modeled.

Large-Scale Fading

The link distance between the nn-th user and the SIM is modeled as

dn=rn2+(h−D−dlayer)2,rn∼U(dr,min⁡,dr,max⁡),n∈N,(7)d_n=\sqrt{r_n^2+(h-D-d_{\mathrm{layer}})^2}, \qquad r_n\sim\mathcal{U}(d_{r,\min},d_{r,\max}),\quad n\in\mathcal{N}, \tag*{(7)}

where hh is the deployment height of the BS and SIM, DD denotes the thickness of the SIM, and D+dlayerD+d_{\mathrm{layer}} includes the BS-to-first-layer separation in the distance offset used for the simulations, and rnr_n is the horizontal distance between the SIM and the nn-th user, which is uniformly distributed over [dr,min⁡,dr,max⁡][d_{r,\min},d_{r,\max}]. The corresponding large-scale fading coefficient is given by

gˉL,n=10C0/10dn−μ,(8)\bar{g}_{L,n}=10^{C_0/10}d_n^{-\mu}, \tag*{(8)}

where C0C_0 denotes the reference path loss (in dB) at unit distance, and μ\mu is the path loss exponent. These coefficients reflect user-specific attenuation levels due to geometric spreading and propagation loss.

Transmission Model

Thus, the overall equivalent downlink channel between the BS antennas and NN UEs becomes

H0Φ1H1⋯HL−1ΦLG=BG∈CN×N,(9)\mathbf{H}_0\mathbf{\Phi}_1\mathbf{H}_1\cdots\mathbf{H}_{L-1}\mathbf{\Phi}_L\mathbf{G}=\mathbf{B}\mathbf{G}\in\mathbb{C}^{N\times N}, \tag*{(9)}

where Hl\mathbf{H}_l (l∈Ll\in\mathcal{L}) are all known by the BS, Φl\mathbf{\Phi}_l (l∈Ll\in\mathcal{L}) are controllable parameters, and G\mathbf{G} is an unknown channel. Denoting the transmitted signal as X∈CN×T\mathbf{X}\in\mathbb{C}^{N\times T} over TT time slots, the received signal at all UEs becomes

YT=GTBTX+N,(10)\mathbf{Y}^{\mathsf{T}}=\mathbf{G}^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{X}+\mathbf{N}, \tag*{(10)}

where N\mathbf{N} is the noise. The ii-th row of Y\mathbf{Y} corresponds to UE ii, and the jj-th column corresponds to time slot jj, respectively. For data transmission, the desired signal power at UE ii is pi∣giTBTei∣2p_i|\mathbf{g}_i^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{e}_i|^2, while the residual interference is ∑j≠ipj∣giTBTej∣2\sum_{j\ne i}p_j|\mathbf{g}_i^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{e}_j|^2. Treating interference as noise gives the corresponding SINR and achievable rate log⁡2(1+SINRi)\log_2(1+\mathrm{SINR}_i). Under exact ZF, this interference term vanishes and the expression reduces to the SNR-based objective below.

Problem Formulation

Considering the ZF requirements, the optimal beamforming problem can be formulated as

arg max⁡{pi}i=1N,{Φl}l=1L∑i∈Nlog⁡2(1+σ−2pi∣giTBTei∣2),(11)\underset{\{p_i\}_{i=1}^{N},\{\mathbf{\Phi}_l\}_{l=1}^{L}}{\operatorname{arg\,max}}\quad \sum_{i\in\mathcal{N}}\log_2\left(1+\sigma^{-2}p_i|\mathbf{g}_i^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{e}_i|^2\right), \tag*{(11)}
s.t.∑i=1Npi≤Pmax⁡,(12)\text{s.t.}\quad\sum_{i=1}^{N}p_i\le P_{\max}, \tag*{(12)}
∣ϕl,m∣≤1,∀l∈L,m∈M,(13)|\phi_{l,m}|\le1,\quad\forall l\in\mathcal{L},m\in\mathcal{M}, \tag*{(13)}
BG is diagonal, (ZF requirement)(14)\mathbf{B}\mathbf{G}\text{ is diagonal, (ZF requirement)} \tag*{(14)}

where gi\mathbf{g}_i denotes the ii-th column of the SIM-UE channel G\mathbf{G} corresponding to UE ii, and ei\mathbf{e}_i denotes an NN-dimensional one-hot vector where the ii-th element is 1 and the others are 0s. We emphasize that the ZF requirement is introduced primarily to provide a structured interference-suppression objective for the proposed gradient-based framework, rather than to claim that strict ZF is sum-rate optimal. Direct optimization of the sum-rate objective leads to a highly coupled non-convex landscape in which useful-signal enhancement and inter-user interference suppression are intertwined, which can result in unstable gradient updates. In contrast, the ZF structure explicitly separates the interference-suppression objective and allows it to be incorporated through a tractable penalty term in the subsequent gradient formulation. We also note that the last ZF requirement (11d) cannot be directly calculated due to unknown G\mathbf{G}, which is otherwise indirectly addressed by loss design as shown below.

Proposed Method Based on UE-Side Zero-Forcing

In this section, we delve into the proposed algorithm. Considering the channel reciprocity, we have two equivalent expressions of the ZF requirement. Specifically, we discuss the UE-side variation in this section, which follows the straightforward ZF procedure where the cross-UE downlink interference is minimized. The BS-side ZF variation, which eliminates the uplink interference among BS antennas, is later discussed in the next section. The main algorithm work flow is illustrated in Figure 2. Specifically, the proposed system is optimized through trials instead of pre-obtained channel estimation. The UEs can then calculate the equivalent loss and initiate gradient backpropagation. As we cannot directly obtain the raw gradient results on the metasurfaces, we need to employ another separate pseudo-inverse mapping to imitate the real gradient values.

Workflow of the proposed P2GD method

Figure 2. Workflow of the proposed P2GD method: downlink forward propagation through the SIM, gradient backpropagation over the reciprocal TDD uplink, and BS-side projected-gradient reconstruction for each SIM layer.

Throughout the proposed procedure, the BS knows the SIM structure, the BS–SIM propagation matrices, and the transmitted pilot signals, while the SIM–UE channel and the intermediate signals inside the SIM are not directly observed. The UEs compute the local loss gradients from their received signals and transmit the corresponding feedback over the reciprocal uplink, after which the BS reconstructs a projected gradient using the known forward model and pseudo-inverse mapping.

To distinguish the signal transmitted to different UEs, we need to transmit a group of orthogonal pilots. Specifically, we transmit a DFT matrix F:=DFT⁡(N)\mathbf{F} := \operatorname{DFT}(N) where

Fm,n=exp⁡(j2(m−1)(n−1)π/N)(15)F_{m,n} = \exp\left(j2(m-1)(n-1)\pi/N\right) \tag*{(15)}

for m,n∈{1,…,N}m,n \in\{1,\ldots,N\} in each transmission. We define P=diag⁡(p1,…,pN)\mathbf{P}=\operatorname{diag}(p_1,\ldots,p_N) and P1/2=diag⁡(p1,…,pN)\mathbf{P}^{1/2}=\operatorname{diag}(\sqrt{p_1},\ldots,\sqrt{p_N}). Assuming that the BS transmits P1/2F\mathbf{P}^{1/2}\mathbf{F} through the downstream, for the simplification of the expression, we denote

XlT:=ΦlHl−1T⋯Φ1H0TP1/2F(16)\mathbf{X}_l^{\mathsf{T}} := \boldsymbol{\Phi}_l\mathbf{H}_{l-1}^{\mathsf{T}}\cdots\boldsymbol{\Phi}_1\mathbf{H}_0^{\mathsf{T}}\mathbf{P}^{1/2}\mathbf{F} \tag*{(16)}

as the transmitted signal after the ll-th layer of the SIM. We note that Xl\mathbf{X}_l cannot be measured directly due to hardware restriction. However, they can be predicted by the BS. Thus, the UE-side received signal on UE ii becomes

yiT=giTBTP1/2F+niT,i∈N(17)\mathbf{y}_i^{\mathsf{T}}=\mathbf{g}_i^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{P}^{1/2}\mathbf{F}+\mathbf{n}_i^{\mathsf{T}},\quad i\in\mathcal{N} \tag*{(17)}

where ni∼CN(0,σ2I)\mathbf{n}_i\sim\mathcal{CN}(\mathbf{0},\sigma^2\mathbf{I}) is the channel noise. Further considering the ZF requirement, we know that under the perfect ZF condition in (11d), the effective channel GTBT\mathbf{G}^{\mathsf{T}}\mathbf{B}^{\mathsf{T}} is diagonal. Hence, for UE ii, the contributions from all streams j≠ij\ne i vanish, and the τ\tau-th received pilot symbol becomes

yi,τ=pi giTBTeiFi,τ+ni,τ.(18)y_{i,\tau}=\sqrt{p_i}\,\mathbf{g}_i^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{e}_iF_{i,\tau}+n_{i,\tau}. \tag*{(18)}

Since ∣Fi,τ∣2=1\lvert F_{i,\tau}\rvert^2=1 for the adopted DFT pilots and E[∣ni,τ∣2]=σ2\mathbb{E}[\lvert n_{i,\tau}\rvert^2]=\sigma^2, we have

E[yiHyi]=Npi∣giTBTei∣2+Nσ2,(19)\mathbb{E}\left[\mathbf{y}_i^{\mathsf{H}}\mathbf{y}_i\right]=Np_i\left\lvert\mathbf{g}_i^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{e}_i\right\rvert^2+N\sigma^2, \tag*{(19)}

which yields

E[yiHyi]=Nσ2(1+σ−2pi∣giTBTei∣2).(20)\mathbb{E}\left[\mathbf{y}_i^{\mathsf{H}}\mathbf{y}_i\right]=N\sigma^2\left(1+\sigma^{-2}p_i\left\lvert\mathbf{g}_i^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{e}_i\right\rvert^2\right). \tag*{(20)}

In practical optimization, the instantaneous received pilot energy yiHyi\mathbf{y}_i^{\mathsf{H}}\mathbf{y}_i is directly observed and used in the local loss. Equations above only establish its expected correspondence to the original rate expression.

arg⁡max⁡{pi}i∈N,{Φl}l=1L∑i∈Nlog⁡yiHyi,s.t. (11b), (11c), (11d).(21)\underset{\{p_i\}_{i\in\mathcal{N}},\{\boldsymbol{\Phi}_l\}_{l=1}^{L}}{\arg\max}\quad\sum_{i\in\mathcal{N}}\log\mathbf{y}_i^{\mathsf{H}}\mathbf{y}_i,\quad\text{s.t. (11b), (11c), (11d).} \tag*{(21)}

UE-Side Zero Forcing

We first discuss realizing zero-forcing from the UE’s side. Thus, the zero-forcing principle can be transformed to

∑i,j∈Ni≠jgiTBTP1/2Ffj=0.(22)\sum_{\substack{i,j\in\mathcal{N}\\i\ne j}}\mathbf{g}_i^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}\mathbf{P}^{1/2}\mathbf{F}\mathbf{f}_j=0. \tag*{(22)}

Substituting (14) into (19) and ignoring the noise item, we can define the loss function

ℓZFUE:=∑i∈Nlog⁡(ϵZF+∑j∈Nj≠i∣yiTfj∣2),(23)\ell_{\mathrm{ZF}}^{\mathrm{UE}}:=\sum_{i\in\mathcal{N}}\log\left(\epsilon_{\mathrm{ZF}}+\sum_{\substack{j\in\mathcal{N}\\j\ne i}}\left\lvert y_i^{\mathsf{T}}\mathbf{f}_j\right\rvert^2\right), \tag*{(23)}

where ϵZF>0\epsilon_{\mathrm{ZF}}>0 is a small regularization constant introduced to avoid numerical singularity when the residual interference approaches zero. It is incorporated into the original objective through the following penalized problem:

arg⁡max⁡{pi}i∈N,{Φl}l=1L∑i∈Nlog⁡yiHyi−λℓZFUE,s.t. (11b), (11c),(24)\underset{\{p_i\}_{i\in\mathcal{N}},\{\boldsymbol{\Phi}_l\}_{l=1}^{L}}{\arg\max}\quad\sum_{i\in\mathcal{N}}\log\mathbf{y}_i^{\mathsf{H}}\mathbf{y}_i-\lambda\ell_{\mathrm{ZF}}^{\mathrm{UE}},\quad\text{s.t. (11b), (11c),} \tag*{(24)}

where λ>0\lambda>0 is the penalty coefficient. Here, the logarithmic form is retained to keep the interference-suppression term on a comparable scale with the rate-related objective, while the regularization has negligible influence when the residual interference is appreciably larger than ϵZF\epsilon_{\mathrm{ZF}}. Accordingly, the ZF penalty in (20) only encourages interference suppression rather than enforcing an exact interference-free solution.

Denote viTv_i^{\mathsf{T}} as the uplink transmitted signal from each UE over time, respectively. Also denote Ul\mathbf{U}_l the uplink signal received by the SIM layer ll. Under the ideal TDD reciprocity assumption stated in § 1-A, we have [4]

ULT=∑i∈NgiviT,(25)\mathbf{U}_L^{\mathsf{T}}=\sum_{i\in\mathcal{N}}g_i v_i^{\mathsf{T}}, \tag*{(25)}
UlT=HlΘl+1⋯HL−1ΘLULT,l∈L∖{L},(26)\mathbf{U}_l^{\mathsf{T}}=\mathbf{H}_l\mathbf{\Theta}_{l+1}\cdots\mathbf{H}_{L-1}\mathbf{\Theta}_L\mathbf{U}_L^{\mathsf{T}},\quad l\in\mathcal{L}\setminus\{L\}, \tag*{(26)}
U0T=H0Θ1U1T+Z,(27)\mathbf{U}_0^{\mathsf{T}}=\mathbf{H}_0\mathbf{\Theta}_1\mathbf{U}_1^{\mathsf{T}}+\mathbf{Z}, \tag*{(27)}

where Z∼CN(0,σ2I)\mathbf{Z}\sim\mathcal{CN}(0,\sigma^2\mathbf{I}) denotes the uplink noise, and Θl\mathbf{\Theta}_l denotes the uplink SIM coefficient matrix. It is set equal to the downlink matrix Φl\mathbf{\Phi}_l in the proposed reciprocal procedure; the two symbols distinguish transmission direction, not independent physical controls. We note that in practical transmissions, we can only obtain U0\mathbf{U}_0 on the BS side. Moreover, as proven in [4], noisy gradients do not degrade the gradient method significantly. Thus, in the following algorithm design, we can omit the noise item Z\mathbf{Z} for simplicity, since introducing the background noise in gradient calculation does not significantly deteriorate the performance.

Gradient Backpropagation

In this section, we examine the gradient of the target in (24). Firstly, denoting the optimization target

ℓ1:=∑i∈Nlog⁡yiHyi−λℓZFUE,(28)\ell_1:=\sum_{i\in\mathcal{N}}\log y_i^{\mathsf{H}}\mathbf{y}_i-\lambda\ell_{\mathrm{ZF}}^{\mathrm{UE}}, \tag*{(28)}

we know

∂ℓ1∂yi∗=(yiHyi)−1yi−λ(ϵZF+∑j=1j≠iN∣fjHyi∗∣2)−1∑j=1j≠iN(fjHyi∗)fj.(29)\frac{\partial\ell_1}{\partial\mathbf{y}_i^*}=(\mathbf{y}_i^{\mathsf{H}}\mathbf{y}_i)^{-1}\mathbf{y}_i-\lambda\left(\epsilon_{\mathrm{ZF}}+\sum_{\substack{j=1\\j\ne i}}^N\left|\mathbf{f}_j^{\mathsf{H}}\mathbf{y}_i^*\right|^2\right)^{-1}\sum_{\substack{j=1\\j\ne i}}^N\left(\mathbf{f}_j^{\mathsf{H}}\mathbf{y}_i^*\right)\mathbf{f}_j. \tag*{(29)}

We note that, given the received signal, (29) can be independently computed for each UE.

Following the chain rule of backpropagation, we have

∂ℓ1∂XL∗=∑i∈N∂ℓ1∂yi∗giH,(30)\frac{\partial\ell_1}{\partial\mathbf{X}_L^*}=\sum_{i\in\mathcal{N}}\frac{\partial\ell_1}{\partial\mathbf{y}_i^*}\mathbf{g}_i^{\mathsf{H}}, \tag*{(30)}
∂ℓ1∂Xl−1∗=∂ℓ1∂Xl∗ΦlHHl−1H,l∈L,(31)\frac{\partial\ell_1}{\partial\mathbf{X}_{l-1}^*}=\frac{\partial\ell_1}{\partial\mathbf{X}_l^*}\mathbf{\Phi}_l^{\mathsf{H}}\mathbf{H}_{l-1}^{\mathsf{H}},\quad l\in\mathcal{L}, \tag*{(31)}
∂ℓ1∂Φl∗=Hl−1HXl−1H∂ℓ1∂Xl∗,l∈L,(32)\frac{\partial\ell_1}{\partial\mathbf{\Phi}_l^*}=\mathbf{H}_{l-1}^{\mathsf{H}}\mathbf{X}_{l-1}^{\mathsf{H}}\frac{\partial\ell_1}{\partial\mathbf{X}_l^*},\quad l\in\mathcal{L}, \tag*{(32)}
∂ℓ1∂P=12P−1/2F∗∂ℓ1∂X0∗.(33)\frac{\partial\ell_1}{\partial\mathbf{P}}=\frac{1}{2}\mathbf{P}^{-1/2}\mathbf{F}^*\frac{\partial\ell_1}{\partial\mathbf{X}_0^*}. \tag*{(33)}

We first compare (30)–(32) with (25) and (26). We know that if we set

vi=α∂ℓ1∂yi∗,i∈N,(34)v_i=\alpha\frac{\partial\ell_1}{\partial\mathbf{y}_i^*},\quad i\in\mathcal{N}, \tag*{(34)}
Θl=Φl,l∈L,(35)\mathbf{\Theta}_l=\mathbf{\Phi}_l,\quad l\in\mathcal{L}, \tag*{(35)}

where α\alpha is a positive number to adjust the uplink transmission power to a reasonable range. For the real-valued objective, ∂ℓ1∂yi=(∂ℓ1∂yi∗)∗\frac{\partial\ell_1}{\partial\mathbf{y}_i}=\left(\frac{\partial\ell_1}{\partial\mathbf{y}_i^*}\right)^*; hence the feedback uses the conjugate of the locally computed Wirtinger gradient. We have

∂ℓ1∂Xl∗=α−1Ul∗,l∈L∪{0},(36)\frac{\partial\ell_1}{\partial\mathbf{X}_l^*}=\alpha^{-1}\mathbf{U}_l^*,\quad l\in\mathcal{L}\cup\{0\}, \tag*{(36)}

where only U0\mathbf{U}_0 can be obtained by the BS, and the noise item in (27) is omitted.

We denote

Bl:=H0Φ1⋯Hl−1Φl.(37)\mathbf{B}_l:=\mathbf{H}_0\mathbf{\Phi}_1\cdots\mathbf{H}_{l-1}\mathbf{\Phi}_l. \tag*{(37)}

Using the pseudoinverse of the known forward mapping, we reconstruct the projected intermediate signal as

U~0=U0,(38)\widetilde{\mathbf{U}}_0=\mathbf{U}_0, \tag*{(38)}
U~l=U0(BlT)†,(39)\widetilde{\mathbf{U}}_l=\mathbf{U}_0\left(\mathbf{B}_l^{\mathsf{T}}\right)^\dagger, \tag*{(39)}
Ψ~l=Hl−1HXl−1HU~l∗,(40)\widetilde{\mathbf{\Psi}}_l=\mathbf{H}_{l-1}^{\mathsf{H}}\mathbf{X}_{l-1}^{\mathsf{H}}\widetilde{\mathbf{U}}_l^*, \tag*{(40)}

which yields regarding (32)

U~l∗=α∂ℓ1∂Xl∗(BlTBl)H,(41)\widetilde{\mathbf{U}}_l^*=\alpha\frac{\partial\ell_1}{\partial\mathbf{X}_l^*}\left(\mathbf{B}_l^{\mathsf{T}}\mathbf{B}_l\right)^{\mathsf{H}}, \tag*{(41)}
Ψ~l=α∂ℓ1∂Φl∗(BlTBl)H.(42)\widetilde{\mathbf{\Psi}}_l=\alpha\frac{\partial\ell_1}{\partial\mathbf{\Phi}_l^*}\left(\mathbf{B}_l^{\mathsf{T}}\mathbf{B}_l\right)^{\mathsf{H}}. \tag*{(42)}

We note that (41) and (42) become computable, and with them, we obtain an approximation of a projected gradient.

General Algorithm

Finally, we come to the overall gradient ascent algorithm as shown in Algorithm 1. The gradient regarding the optimization target in problem (24) can be obtained through (33) and (42). As the remaining restrictions of (11c) and (11b) are simple convex bounds over the target variables, we can directly add them in the gradient ascent process, where we choose different learning rates for power P\mathbf{P} and reflection coefficients {Φl}\{\mathbf{\Phi}_l\} for numerical stability. We note that as the variables are diagonal, we only need to calculate the diagonal elements in (33) and (40). Also, we note that we apply a maximum uplink power restriction when yiHyi\mathbf{y}_i^{\mathsf{H}}\mathbf{y}_i becomes too small in the algorithm.

Proposed Method Based on BS-Side Zero-Forcing

In this section, we try to apply the ZF conditions on the BS side. We adopt the same downlink transmission model as (17). Instead of directly transforming the ZF requirement into a penalty item, we consider the received power of the desired signal directly. Thus, different from (21), the optimization problem can be written as

arg max⁡{pi}i=1N,{Φl}lLℓUE:=∑i=1Nlog⁡yiHfifiHyi,(43)\underset{\{p_i\}_{i=1}^N,\{\mathbf{\Phi}_l\}_{l}^L}{\operatorname{arg\,max}}\quad\ell_{\mathrm{UE}}:=\sum_{i=1}^N\log\mathbf{y}_i^{\mathsf{H}}\mathbf{f}_i\mathbf{f}_i^{\mathsf{H}}\mathbf{y}_i, \tag*{(43)}
s.t.(11b),(11c),(11d).\mathrm{s.t.}\quad(11\mathrm{b}), (11\mathrm{c}), (11\mathrm{d}).

It is easy to know

∂ℓUE∂yi∗=(yiHfi)−1fi,(44)\frac{\partial\ell_{\mathrm{UE}}}{\partial\mathbf{y}_i^*}=(\mathbf{y}_i^{\mathsf{H}}\mathbf{f}_i)^{-1}\mathbf{f}_i, \tag*{(44)}

which is along the direction of fi\mathbf{f}_i. Utilizing the channel reciprocity, according to (34) and (44), the uplink feedback vector of UE ii is given by

vi=α(yiTfi∗)−1fi∗.(45)v_i=\alpha(\mathbf{y}_i^{\mathsf{T}}\mathbf{f}_i^*)^{-1}\mathbf{f}_i^*. \tag*{(45)}

Algorithm 1: UE-P2GD Algorithm

  1. Input: Maximum transmission power Pmax⁡P_{\max} for BS, PUEP_{\mathrm{UE}} for UEs, penalty coefficient λ\lambda, hyperparameter α\alpha, learning rates ηP\eta_P and ηΦ\eta_\Phi

  2. Initialization: BS randomly sample Φl\Phi_l such that ϕl,i=ejψl,i\phi_{l,i}=e^{\mathrm{j}\psi_{l,i}}, all ψ\psi's are independently drawn from U(0,2π)\mathcal{U}(0,2\pi). Initial transmission power pi=Pmax⁡/Np_i=P_{\max}/N

  3. for t=1,…,Tmax⁡t=1,\ldots,T_{\max} do

  4. BS sets SIM parameters to {Φl}l=1L\{\Phi_l\}_{l=1}^{L}

  5. BS downlink transmit P1/2F\mathbf{P}^{1/2}\mathbf{F} in NN time slots

  6. for each UE in parallel do

  7. UE ii receives yiT\mathbf{y}_i^{\mathsf{T}} as in (14)

  8. Calculate gradient di←∂ℓ1/∂yi∗\mathbf{d}_i\leftarrow\partial\ell_1/\partial\mathbf{y}_i^* as (26)

  9. for any τ\tau such that α∣di,τ∣>PUE\alpha|d_{i,\tau}|>P_{\mathrm{UE}} do

  10. Clamping: di,τ←PUEdi,τα∣di,τ∣d_{i,\tau}\leftarrow\frac{P_{\mathrm{UE}}d_{i,\tau}}{\alpha|d_{i,\tau}|}

  11. end for

  12. Uplink transmit {αdi,τ∗Fi,τ}τ=1N\{\alpha d^*_{i,\tau}\mathbf{F}_{i,\tau}\}_{\tau=1}^{N} in NN time slots

  13. end for

  14. BS receives U0\mathbf{U}_0 as in (22)–(24)

  15. Gradient of power gpi←fiHu0,i∗2αpig_{p_i}\leftarrow\frac{f_i^{\mathsf{H}}u_{0,i}^*}{2\alpha\sqrt{p_i}} as (30)

  16. Corresponding to (35)–(37), B←IN\mathbf{B}\leftarrow\mathbf{I}_N

  17. X0←FP1/2\mathbf{X}_0\leftarrow\mathbf{F}\mathbf{P}^{1/2}, X←X0\mathbf{X}\leftarrow\mathbf{X}_0

  18. for l=1,⋯ ,Ll=1,\cdots,L do

  19. B←BHl−1Φl\mathbf{B}\leftarrow\mathbf{B}\mathbf{H}_{l-1}\Phi_l, U←U0(BT)†\mathbf{U}\leftarrow\mathbf{U}_0(\mathbf{B}^{\mathsf{T}})^\dagger

  20. Ψl,i←Hl−1Hxi∗uiH\Psi_{l,i}\leftarrow\mathbf{H}_{l-1}^{\mathsf{H}}\mathbf{x}_i^*\mathbf{u}_i^{\mathsf{H}}, Ψl:=diag⁡(Ψl,1,⋯ ,Ψl,M)\Psi_l:=\operatorname{diag}(\Psi_{l,1},\cdots,\Psi_{l,M})

  21. X←X0B\mathbf{X}\leftarrow\mathbf{X}_0\mathbf{B}

  22. Φl←Φl+ηΦα−1Ψl\Phi_l\leftarrow\Phi_l+\eta_\Phi\alpha^{-1}\Psi_l

  23. for any ii such that ∣ϕl,i∣>1|\phi_{l,i}|>1 do

  24. Apply power restriction: ϕl,i←ϕl,i/∣ϕl,i∣\phi_{l,i}\leftarrow\phi_{l,i}/|\phi_{l,i}|

  25. end for

  26. end for

  27. pi←max⁡{pi+ηPgpi,0}p_i\leftarrow\max\{p_i+\eta_Pg_{p_i},0\} for i=1,⋯ ,Ni=1,\cdots,N

  28. pi←Pmax⁡pi/∑i=1Npip_i\leftarrow P_{\max}p_i/\sum_{i=1}^{N}p_i for i=1,⋯ ,Ni=1,\cdots,N

  29. end for

which is parallel to the conjugated ii-th DFT pilot vector fi∗\mathbf{f}_i^*. Ignoring the uplink noise, (22)–(24) can be equivalently written as

U0=VGTBT,(46)\mathbf{U}_0=\mathbf{V}\mathbf{G}^{\mathsf{T}}\mathbf{B}^{\mathsf{T}}, \tag*{(46)}

where V=[v1,…,vN]\mathbf{V}=[\mathbf{v}_1,\ldots,\mathbf{v}_N]. Under the ZF condition in (11d), GTBT\mathbf{G}^{\mathsf{T}}\mathbf{B}^{\mathsf{T}} is diagonal. Therefore, denoting u0,i\mathbf{u}_{0,i} as the ii-th column of U0\mathbf{U}_0, we have u0,i∝vi∝fi∗\mathbf{u}_{0,i}\propto\mathbf{v}_i\propto\mathbf{f}_i^*. Owing to the orthogonality of the DFT pilots, this yields

E{u0,iHfj}=0,∀i≠j.(47)\mathbb{E}\{\mathbf{u}_{0,i}^{\mathsf{H}}\mathbf{f}_j\}=0,\quad\forall i\ne j. \tag*{(47)}

Therefore, we can write the following penalty item:

ℓZFBS:=∑i=1Nlog⁡(ϵZF+∑j=1,j≠iN∣u0,iTfj∣2).(48)\ell_{\mathrm{ZF}}^{\mathrm{BS}}:=\sum_{i=1}^{N}\log\left(\epsilon_{\mathrm{ZF}}+\sum_{j=1,j\ne i}^{N}|\mathbf{u}_{0,i}^{\mathsf{T}}\mathbf{f}_j|^2\right). \tag*{(48)}

Therefore, we have the optimization target

ℓ2:=ℓUE−λℓZFBS.(49)\ell_2:=\ell_{\mathrm{UE}}-\lambda\ell_{\mathrm{ZF}}^{\mathrm{BS}}. \tag*{(49)}

Through gradient backpropagation, noting that Φl=Θl\Phi_l=\Theta_l, we know that

∂ℓZFBS∂u0,i∗=(ϵZF+∑j=1,j≠iN∣u0,iTfj∣2)−1∑j=1,j≠iN(u0,iTfj)fj∗,(50)\frac{\partial\ell_{\mathrm{ZF}}^{\mathrm{BS}}}{\partial\mathbf{u}_{0,i}^*}=\left(\epsilon_{\mathrm{ZF}}+\sum_{j=1,j\ne i}^{N}|\mathbf{u}_{0,i}^{\mathsf{T}}\mathbf{f}_j|^2\right)^{-1} \sum_{j=1,j\ne i}^{N}(\mathbf{u}_{0,i}^{\mathsf{T}}\mathbf{f}_j)\mathbf{f}_j^*, \tag*{(50)}
∂ℓZFBS∂Ul−1∗=∂ℓZFBS∂Ul∗Hl−1∗Φl∗,(51)\frac{\partial\ell_{\mathrm{ZF}}^{\mathrm{BS}}}{\partial\mathbf{U}_{l-1}^*} =\frac{\partial\ell_{\mathrm{ZF}}^{\mathrm{BS}}}{\partial\mathbf{U}_l^*}\mathbf{H}_{l-1}^*\Phi_l^*, \tag*{(51)}
∂ℓZFBS∂Φl∗=Hl−1H(∂ℓZFBS∂Ul−1∗)TUl.(52)\frac{\partial\ell_{\mathrm{ZF}}^{\mathrm{BS}}}{\partial\Phi_l^*} =\mathbf{H}_{l-1}^{\mathsf{H}}\left(\frac{\partial\ell_{\mathrm{ZF}}^{\mathrm{BS}}}{\partial\mathbf{U}_{l-1}^*}\right)^{\mathsf{T}}\mathbf{U}_l. \tag*{(52)}

However, Ul\mathbf{U}_l cannot be directly observed due to hardware restrictions. Similarly, we can use the pseudo inverse to obtain a projected gradient following (35) and (36). Considering (13), (35)–(37), and (47)–(49), we can finally use the following process, where all required variables are known by the BS.

X~0=α−1P1/2F−λ(∂ℓZFBS∂U0∗),(53)\widetilde{\mathbf{X}}_0=\alpha^{-1}\mathbf{P}^{1/2}\mathbf{F}-\lambda\left(\frac{\partial\ell_{\mathrm{ZF}}^{\mathrm{BS}}}{\partial\mathbf{U}_0^*}\right), \tag*{(53)}
X~l=X~l−1Hl−1Φl,(54)\widetilde{\mathbf{X}}_l=\widetilde{\mathbf{X}}_{l-1}\mathbf{H}_{l-1}\Phi_l, \tag*{(54)}
Ψ~l=Hl−1HX~l−1HU~l∗,(55)\widetilde{\Psi}_l=\mathbf{H}_{l-1}^{\mathsf{H}}\widetilde{\mathbf{X}}_{l-1}^{\mathsf{H}}\widetilde{\mathbf{U}}_l^*, \tag*{(55)}

where all U~l\widetilde{\mathbf{U}}_ls are defined in (35) and (36). It is easy to verify that, ignoring the background noise,

Ψ~l=∂ℓ2∂Φl∗(BlTBl)H.(56)\widetilde{\Psi}_l=\frac{\partial\ell_2}{\partial\Phi_l^*}\left(\mathbf{B}_l^{\mathsf{T}}\mathbf{B}_l\right)^{\mathsf{H}}. \tag*{(56)}

Thus, we have the overall projected gradient ascent algorithm as shown in Algorithm 2.

Theoretical Results

Projected Gradient Method

Here, we first analyze the convergence behavior of the projected-gradient mechanism under an idealized smooth convex setting, which provides a theoretical reference for the projected updates used in the proposed algorithms. In this section, we use the Frobenius inner products and norms unless otherwise stated, i.e.,

⟨A,B⟩:=tr⁡(AHB),∥A∥:=⟨A,A⟩.(57)\langle\mathbf{A},\mathbf{B}\rangle:=\operatorname{tr}(\mathbf{A}^{\mathsf{H}}\mathbf{B}),\qquad\|\mathbf{A}\|:=\sqrt{\langle\mathbf{A},\mathbf{A}\rangle}. \tag*{(57)}

We first consider a convex optimization problem with an arbitrary matrix A\mathbf{A}

min⁡Xℓ(B),B:=AX,(58)\min_{\mathbf{X}}\ell(\mathbf{B}),\qquad\mathbf{B}:=\mathbf{A}\mathbf{X}, \tag*{(58)}

where ℓ\ell is real-valued with some lower bound, convex, differentiable. We denote ∂Mℓ:=∂ℓ/∂M\partial_{\mathbf{M}}\ell:=\partial\ell/\partial\mathbf{M} for any M\mathbf{M} related to (55). We assume that ℓ(B)\ell(\mathbf{B}) is β\beta-smooth such that

∥∂Bℓ(B1)−∂Bℓ(B2)∥≤β∥B1−B2∥,∀B1,B2∈CN×M.(59)\|\partial_{\mathbf{B}}\ell(\mathbf{B}_1)-\partial_{\mathbf{B}}\ell(\mathbf{B}_2)\|\le\beta\|\mathbf{B}_1-\mathbf{B}_2\|,\quad\forall\mathbf{B}_1,\mathbf{B}_2\in\mathbb{C}^{N\times M}. \tag*{(59)}

The variable X\mathbf{X} in this subsection is an abstract optimization matrix, distinct from the transmitted and intermediate signals denoted by X\mathbf{X} and Xl\mathbf{X}_l in the system model.

To solve the problem (55), we consider the following projected gradient descent method beginning from any X0\mathbf{X}_0:

Xt+1=Xt−η(∂Xt∗ℓ)Qt,(60)\mathbf{X}_{t+1}=\mathbf{X}_t-\eta\left(\partial_{\mathbf{X}_t^*}\ell\right)\mathbf{Q}_t, \tag*{(60)}

where we denote the projection matrix as

Qt:=(BtTBt)H=XtHAH(XtHAH)†.(61)\mathbf{Q}_t:=\left(\mathbf{B}_t^{\mathsf{T}}\mathbf{B}_t\right)^{\mathsf{H}}=\mathbf{X}_t^{\mathsf{H}}\mathbf{A}^{\mathsf{H}}\left(\mathbf{X}_t^{\mathsf{H}}\mathbf{A}^{\mathsf{H}}\right)^\dagger. \tag*{(61)}

We can first verify the following lemma.

Lemma 1.

Qt=Qt2=QtH.(62)\mathbf{Q}_t=\mathbf{Q}_t^2=\mathbf{Q}_t^{\mathsf{H}}. \tag*{(62)}

Algorithm 2: BS-P2GD Algorithm

  1. Input: Maximum transmission power Pmax⁡P_{\max} for BS, PUEP_{\mathrm{UE}} for UEs, penalty coefficient λ\lambda, hyperparameter α\alpha, learning rates ηP\eta_P and ηΦ\eta_{\Phi}

  2. Initialization: BS randomly sample Φl\boldsymbol{\Phi}_l such that ϕl,i=ejψl,i\phi_{l,i}=e^{\mathrm{j}\psi_{l,i}}, all ψ\psi's are independently drawn from U(0,2π)\mathcal{U}(0,2\pi). Initial transmission power pi=Pmax⁡/Np_i=P_{\max}/N

  3. for t=1,…,Tmax⁡t=1,\ldots,T_{\max} do

  4. BS sets SIM parameters to {Φl}l=1L\{\boldsymbol{\Phi}_l\}_{l=1}^{L}

  5. BS downlink transmit P1/2F\mathbf{P}^{1/2}\mathbf{F} in NN time slots

  6. for each UE in parallel do

  7. UE ii receives yiT\mathbf{y}_i^{\mathsf{T}} as in (14)

  8. Calculate gradient di←(yiHfi)−1fi\mathbf{d}_i\leftarrow(\mathbf{y}_i^{\mathsf{H}}\mathbf{f}_i)^{-1}\mathbf{f}_i

  9. for any τ\tau such that α∣di,τ∣>PUE\alpha|d_{i,\tau}|>P_{\mathrm{UE}} do

  10. Clamping: di,τ←PUEdi,τα∣di,τ∣d_{i,\tau}\leftarrow\frac{P_{\mathrm{UE}}d_{i,\tau}}{\alpha|d_{i,\tau}|}

  11. end for

  12. Uplink transmit {αdi,τ∗Fi,τ}τ=1N\{\alpha d_{i,\tau}^{*}\mathbf{F}_{i,\tau}\}_{\tau=1}^{N} in NN time slots

  13. end for

  14. BS receives U0\mathbf{U}_0 as in (22)–(24)

  15. Gradient of power gpi←ℜfiHu0,i∗2αpig_{p_i}\leftarrow\Re\frac{f_i^{\mathsf{H}}u_{0,i}^{*}}{2\alpha\sqrt{p_i}} as (30)

  16. Corresponding to (35),(36) & (50)–(52), U←U0\mathbf{U}\leftarrow\mathbf{U}_0

  17. X←X^0\mathbf{X}\leftarrow\widehat{\mathbf{X}}_0 as defined in (50), B←IN\mathbf{B}\leftarrow\mathbf{I}_N

  18. for l=1,⋯ ,Ll=1,\cdots,L do

  19. B←BHl−1HΦl\mathbf{B}\leftarrow\mathbf{B}\mathbf{H}_{l-1}^{\mathsf{H}}\boldsymbol{\Phi}_l, U←U0\mathbf{U}\leftarrow\mathbf{U}_0

  20. Ψl,i←Hl−1Hxi∗ui∗\boldsymbol{\Psi}_{l,i}\leftarrow\mathbf{H}_{l-1}^{\mathsf{H}}\mathbf{x}_i^{*}\mathbf{u}_i^{*}, Ψl:=diag⁡(Ψl,1,⋯ ,Ψl,M)\boldsymbol{\Psi}_l:=\operatorname{diag}(\boldsymbol{\Psi}_{l,1},\cdots,\boldsymbol{\Psi}_{l,M})

  21. X←X0B\mathbf{X}\leftarrow\mathbf{X}_0\mathbf{B}

  22. Φl←Φl+ηΦΨl\boldsymbol{\Phi}_l\leftarrow\boldsymbol{\Phi}_l+\eta_{\Phi}\boldsymbol{\Psi}_l

  23. for any ii such that ∣ϕl,i∣>1|\phi_{l,i}|>1 do

  24. Apply power restriction: ϕl,i←ϕl,i/∣ϕl,i∣\phi_{l,i}\leftarrow\phi_{l,i}/|\phi_{l,i}|

  25. end for

  26. end for

  27. pi←max⁡{pi+ηPgpi,0}p_i\leftarrow\max\{p_i+\eta_Pg_{p_i},0\} for i=1,⋯ ,Ni=1,\cdots,N

  28. pi←Pmax⁡pi/∑i=1Npip_i\leftarrow P_{\max}p_i/\sum_{i=1}^{N}p_i for i=1,⋯ ,Ni=1,\cdots,N

  29. end for

where Qt\mathbf{Q}_t is defined in (58).

Proof. Let the compact SVD be Bt=UrΣrVrH\mathbf{B}_t=\mathbf{U}_r\boldsymbol{\Sigma}_r\mathbf{V}_r^{\mathsf{H}}, where the columns of Vr\mathbf{V}_r are orthonormal. Then Bt†Bt=VrVrH\mathbf{B}_t^{\dagger}\mathbf{B}_t=\mathbf{V}_r\mathbf{V}_r^{\mathsf{H}}, so Qt=VrVrH\mathbf{Q}_t=\mathbf{V}_r\mathbf{V}_r^{\mathsf{H}} is Hermitian and Qt2=Vr(VrHVr)VrH=Qt\mathbf{Q}_t^2=\mathbf{V}_r(\mathbf{V}_r^{\mathsf{H}}\mathbf{V}_r)\mathbf{V}_r^{\mathsf{H}}=\mathbf{Q}_t. The zero-rank case gives Qt=0\mathbf{Q}_t=0 and satisfies the same identities. □

Remark 1. Lemma 1 shows that the Qt\mathbf{Q}_t is an orthogonal projection matrix. Thus, XQt\mathbf{X}\mathbf{Q}_t simply projects X\mathbf{X} to a low-dimensional linear space whose dimension is determined by the rank of Bt\mathbf{B}_t.

Considering the one-step gradient descent, we start from the following Wirtinger descent lemma in the complex domain, which is a trivial extension of the real version [48].

Lemma 2. (Wirtinger descent lemma [48]) For any function f(X)f(\mathbf{X}) which is differentiable and β\beta-smooth, we have

f(X2)≤f(X1)+2ℜ⟨∂X∗f(X),X2−X1⟩+β2∥X2−X1∥2.(63)f(\mathbf{X}_2)\leq f(\mathbf{X}_1)+2\Re\langle\partial_{\mathbf{X}^{*}}f(\mathbf{X}),\mathbf{X}_2-\mathbf{X}_1\rangle+\frac{\beta}{2}\lVert\mathbf{X}_2-\mathbf{X}_1\rVert^2. \tag*{(63)}

With Lemma 2, we can directly obtain the following theorem describing the convergence of the projected gradient method.

Theorem 1. Let

0<η<4β∥A∥2,(64)0<\eta<\frac{4}{\beta\lVert\mathbf{A}\rVert^2}, \tag*{(64)}

the projected gradient descent algorithm in (60) always converges to some suboptimal points X^\widehat{\mathbf{X}}, where X^\widehat{\mathbf{X}} satisfies

(∂X^∗ℓ)X^HAH(X^HA)†=0.(65)(\partial_{\widehat{\mathbf{X}}^{*}}\ell)\widehat{\mathbf{X}}^{\mathsf{H}}\mathbf{A}^{\mathsf{H}}\left(\widehat{\mathbf{X}}^{\mathsf{H}}\mathbf{A}\right)^{\dagger}=0. \tag*{(65)}

Specifically, the optimal convergence rate satisfies

min⁡0≤t≤T∥(∂Xt∗ℓ)Qt∥2=O(β∥A∥2/T)(66)\min_{0\leq t\leq T}\lVert(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2=\mathcal{O}(\beta\lVert\mathbf{A}\rVert^2/T) \tag*{(66)}

when η=2β∥A∥2\eta=\frac{2}{\beta\lVert\mathbf{A}\rVert^2}.

Proof. From Lemma 1, we know that for any matrix X\mathbf{X},

⟨X,XQt⟩=⟨X,XQt2⟩=⟨XQtH,XQt⟩=⟨XQt,XQt⟩=∥XQt∥2(67)\langle\mathbf{X},\mathbf{X}\mathbf{Q}_t\rangle=\langle\mathbf{X},\mathbf{X}\mathbf{Q}_t^2\rangle=\langle\mathbf{X}\mathbf{Q}_t^{\mathsf{H}},\mathbf{X}\mathbf{Q}_t\rangle =\langle\mathbf{X}\mathbf{Q}_t,\mathbf{X}\mathbf{Q}_t\rangle=\lVert\mathbf{X}\mathbf{Q}_t\rVert^2 \tag*{(67)}

We can derive the error bound from Lemma 2 that

ℓ(Bt+1)−ℓ(Bt)\ell(\mathbf{B}_{t+1})-\ell(\mathbf{B}_t)
≤2ℜ⟨∂Bt∗ℓ,A(Xt+1−Xt)⟩+β2∥A(Xt+1−Xt)∥2=−2ηℜ⟨AH(∂Bt∗ℓ),(∂Xt∗ℓ)Qt⟩+βη22∥A(∂Xt∗ℓ)Qt∥2=−2η∥(∂Xt∗ℓ)Qt∥2+βη22∥A(∂Xt∗ℓ)Qt∥2=−2η∥(∂Xt∗ℓ)Qt∥2+βη22∥A∥2∥(∂Xt∗ℓ)Qt∥2≤−2η∥(∂Xt∗ℓ)Qt∥2+βη22∥A∥2∥(∂Xt∗ℓ)Qt∥2.(68)\begin{aligned} \leq2\Re\langle\partial_{\mathbf{B}_t^{*}}\ell,\mathbf{A}(\mathbf{X}_{t+1}-\mathbf{X}_t)\rangle+\frac{\beta}{2}\lVert\mathbf{A}(\mathbf{X}_{t+1}-\mathbf{X}_t)\rVert^2 \\ &=-2\eta\Re\langle\mathbf{A}^{\mathsf{H}}(\partial_{\mathbf{B}_t^{*}}\ell),(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rangle+\frac{\beta\eta^2}{2}\lVert\mathbf{A}(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2 \\ &=-2\eta\lVert(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2+\frac{\beta\eta^2}{2}\lVert\mathbf{A}(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2 \\ &=-2\eta\lVert(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2+\frac{\beta\eta^2}{2}\lVert\mathbf{A}\rVert^2\lVert(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2 \\ &\leq-2\eta\lVert(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2+\frac{\beta\eta^2}{2}\lVert\mathbf{A}\rVert^2\lVert(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2. \tag*{(68)} \end{aligned}

Thus, as ℓ\ell has a lower bound and ℓ(Bt)\ell(\mathbf{B}_t) is decreasing, (57) converges to some points X^\widehat{\mathbf{X}} that fulfills (62) with η\eta determined in (61). Further, we know that

∑t=0T∥(∂Xt∗ℓ)Qt∥2≤24η−βη2∥A∥2(ℓ(AX0)−ℓ(AX^)).(69)\sum_{t=0}^{T}\lVert(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2\leq\frac{2}{4\eta-\beta\eta^2\lVert\mathbf{A}\rVert^2}\left(\ell(\mathbf{A}\mathbf{X}_0)-\ell(\mathbf{A}\widehat{\mathbf{X}})\right). \tag*{(69)}

Thus, when taking the optimal value η=2β∥A∥2\eta=\frac{2}{\beta\lVert\mathbf{A}\rVert^2}, we have

1T∑t=0T∥(∂Xt∗ℓ)Qt∥2≤O(β∥A∥2T),(70)\frac{1}{T}\sum_{t=0}^{T}\lVert(\partial_{\mathbf{X}_t^{*}}\ell)\mathbf{Q}_t\rVert^2\leq\mathcal{O}\left(\frac{\beta\lVert\mathbf{A}\rVert^2}{T}\right), \tag*{(70)}

which yields (63) and finishes the proof. □

Theorem 1 characterizes the projected-gradient iteration in (57) under the smooth convex setting. Its purpose is to provide a theoretical reference for understanding the effect of the projection operation on the gradient update. However, the practical SIM beamforming problems considered in Algorithms 1 and 2, are generally non-convex due to the cascaded SIM mapping, the logarithmic signal and interference terms, and the physical constraints on the SIM coefficients. Therefore, Theorem 1 only characterizes the behavior of the underlying projected-gradient mechanism under the stated idealized conditions.

Remark 2. We can observe from (3) that all Hl\mathbf{H}_ls are almost of full rank. Moreover, we know that unless there is some mm such that ϕl,m=0\phi_{l,m}=0, Φl\Phi_l is also of full rank. As ϕl,m\phi_{l,m} can continuously take values in {x∈C∣∣x∣≤1}\{x \in\mathbb{C}\mid|x| \le1\}, we can say it is almost impossible that ϕl,i=0\phi_{l,i}=0 for any l∈Ll \in\mathcal{L} and m∈Mm \in\mathcal{M}. Thus, the ranks of all Bl\mathbf{B}_ls are almost the same as H0\mathbf{H}_0, which is determined by the number of BS antennas NN. Thus, the performance of projected gradient descent is closer to the method in [3] when N/MN/M becomes relatively large. However, at the same time, the advantages in reducing transmission overhead also shrink when N/MN/M increases, since we need to transmit more pilots in each iteration.

Therefore, the proposed method is particularly suitable for regimes where the observable dimension NN is not excessively small relative to the number of SIM variables MM. Increasing N/MN/M generally enriches the observable gradient subspace and reduces the potential projection loss, but simultaneously increases the pilot and signaling overhead. This reflects an intrinsic tradeoff between gradient fidelity and overhead reduction in the proposed framework.

Diagonal Limitation of SIM Phase Matrices

In this section, we aim to discuss the practical restrictions of SIMs. As shown above, we know that the proposed projected gradient algorithm can converge to a set of suboptimal Φ‾\overline{\boldsymbol{\Phi}}. We then consider the specific structure determined by (68). Specifically, in practice, the aggregated SIM response

Φ‾=f(Φ1,…,ΦL):=Φ1H1⋯HL−1ΦL(71)\overline{\boldsymbol{\Phi}}=f(\boldsymbol{\Phi}_1,\ldots,\boldsymbol{\Phi}_L):=\boldsymbol{\Phi}_1\mathbf{H}_1\cdots\mathbf{H}_{L-1}\boldsymbol{\Phi}_L \tag*{(71)}

is restricted to an achievable set determined jointly by the diagonal structure and bounded entries of {Φl}\{\boldsymbol{\Phi}_l\}, as well as by the ranks of the inter-layer propagation matrices {Hl}\{\mathbf{H}_l\}. Therefore, the achievable set of Φ‾\overline{\boldsymbol{\Phi}} does not necessarily span the entire space CM×M\mathbb{C}^{M\times M}. To isolate the effect of the diagonal SIM update and simplify the analysis, we first relax this realizability constraint as follows.

Assumption 1. For the purpose of analysis, the aggregated matrix Φ‾\overline{\boldsymbol{\Phi}} defined in (68) is assumed to be freely variable in CM×M\mathbb{C}^{M\times M}, i.e., we neglect here the restriction that Φ‾\overline{\boldsymbol{\Phi}} must belong to the achievable set induced by the cascaded SIM structure.

Under Assumption 1, we have the following theorem. It further characterizes the effect of the diagonal SIM structure under relaxed realizability assumption and the stated directional conditions. It provides sufficient analytical conditions for the structured projected update, rather than a global convergence guarantee for the original non-convex beamforming problem.

Theorem 2. Consider Φ‾\overline{\boldsymbol{\Phi}} as defined in (68), denote

D(Φ‾,ΔΦ‾,η):=Φ‾−f(Φ1−ηΔΦ1,…,ΦL−ηΔΦL),(72)D(\overline{\boldsymbol{\Phi}},\Delta_{\overline{\boldsymbol{\Phi}}},\eta):=\overline{\boldsymbol{\Phi}}-f(\boldsymbol{\Phi}_1-\eta\Delta_{\boldsymbol{\Phi}_1},\ldots,\boldsymbol{\Phi}_L-\eta\Delta_{\boldsymbol{\Phi}_L}), \tag*{(72)}

where

ΔΦl:=Qdiag(Hl−1HΦl−1H⋯H1HΦ1HΔΦ‾ΦLHHL−1H⋯HlH),(73)\Delta_{\boldsymbol{\Phi}_l}:=\mathcal{Q}_{\mathrm{diag}}\left(\mathbf{H}_{l-1}^{\mathrm{H}}\boldsymbol{\Phi}_{l-1}^{\mathrm{H}}\cdots\mathbf{H}_1^{\mathrm{H}}\boldsymbol{\Phi}_1^{\mathrm{H}}\Delta_{\overline{\boldsymbol{\Phi}}}\boldsymbol{\Phi}_L^{\mathrm{H}}\mathbf{H}_{L-1}^{\mathrm{H}}\cdots\mathbf{H}_l^{\mathrm{H}}\right), \tag*{(73)}

and QdiagX\mathcal{Q}_{\mathrm{diag}}\mathbf{X} is a projection operator that extracts the diagonal elements of X\mathbf{X} while retaining the shape, i.e.,

[QdiagX]i,j={Xi,i,i=j,0,i≠j.(74)[\mathcal{Q}_{\mathrm{diag}\mathbf{X}]_{i,j}= \begin{cases} X_{i,i}, & i=j,\\ 0, & i\ne j. \end{cases}} \tag*{(74)}

We denote the angle between two matrices as

∠(A,B):=arccos⁡ℜ⟨A,B⟩∥A∥∥B∥.(75)\angle(\mathbf{A},\mathbf{B}):=\arccos\frac{\Re\langle\mathbf{A},\mathbf{B}\rangle}{\|\mathbf{A}\|\|\mathbf{B}\|}. \tag*{(75)}

Thus, for any 0<ϵ,ζ<10<\epsilon,\zeta<1 and 0<ξ1<1<ξ20<\xi_1<1<\xi_2, if we always have the following

cos⁡∠(D(Φ‾,ΔΦ‾,η),ΔΦ‾)≥ϵ,(76)\cos\angle\left(D(\overline{\boldsymbol{\Phi}},\Delta_{\overline{\boldsymbol{\Phi}}},\eta),\Delta_{\overline{\boldsymbol{\Phi}}}\right)\ge\epsilon, \tag*{(76)}
(1−ϵ2)∥∂Φ‾t∗ℓ∥2≤ζ∥(∂Φ‾t∗ℓ)Qt∥2,(77)(1-\epsilon^2)\left\|\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell\right\|^2\le\zeta\left\|\left(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell\right)\mathbf{Q}_t\right\|^2, \tag*{(77)}
ξ1η∥ΔΦ‾∥≤∥D(Φ‾,ΔΦ‾,η)∥≤ξ2η∥ΔΦ‾∥,(78)\xi_1\eta\|\Delta_{\overline{\boldsymbol{\Phi}}}\|\le\|D(\overline{\boldsymbol{\Phi}},\Delta_{\overline{\boldsymbol{\Phi}}},\eta)\|\le\xi_2\eta\|\Delta_{\overline{\boldsymbol{\Phi}}}\|, \tag*{(78)}
0<η<4ϵ−4ϵ2−1+ζβξ2∥A∥2,(79)0<\eta<\frac{4\epsilon-4\sqrt{\epsilon^2-1+\zeta}}{\beta\xi_2\|\mathbf{A}\|^2}, \tag*{(79)}

for some η\eta, the following method has a vanishing projected-gradient residual as in (62), provided that the stated conditions hold throughout the iteration.

Xl=Xl−ηQdiag((∂Xl∗ℓ)(BlHBl)H).(80)\mathbf{X}_l=\mathbf{X}_l-\eta\mathcal{Q}_{\mathrm{diag}}\left((\partial_{\mathbf{X}_l}^{*}\ell)\left(\mathbf{B}_l^{\mathrm{H}}\mathbf{B}_l\right)^{\mathrm{H}}\right). \tag*{(80)}

Proof. Following Lemma 2, we know

ℓ(Bt+1)−ℓ(Bt)≤−2ℜ⟨∂Φ‾t∗ℓ,Φ‾t−Φ‾t+1⟩+β2∥A∥2∥Φ‾t+1−Φ‾t∥2=−2ℜ⟨∂Φ‾t∗ℓ,D(Φ‾t,(∂Φ‾t∗ℓ)Qt,η)⟩+β2∥A∥2∥D(Φ‾t,(∂Φ‾t∗ℓ)Qt,η)∥2.(81)\begin{aligned} \ell(\mathbf{B}_{t+1})-\ell(\mathbf{B}_t) \le-2\Re\langle\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell,\overline{\boldsymbol{\Phi}}_t-\overline{\boldsymbol{\Phi}}_{t+1}\rangle+\frac{\beta}{2}\|\mathbf{A}\|^2\|\overline{\boldsymbol{\Phi}}_{t+1}-\overline{\boldsymbol{\Phi}}_t\|^2 \\ &=-2\Re\langle\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell,D(\overline{\boldsymbol{\Phi}}_t,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t,\eta)\rangle +\frac{\beta}{2}\|\mathbf{A}\|^2\|D(\overline{\boldsymbol{\Phi}}_t,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t,\eta)\|^2. \tag*{(81)} \end{aligned}

We note that since ⟨∂Φ‾t∗ℓ,(∂Φ‾t∗ℓ)Qt⟩=∥(∂Φ‾t∗ℓ)Qt∥2\langle\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\rangle=\|(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\|^2 always holds, we know

∠(∂Φ‾t∗ℓ,(∂Φ‾t∗ℓ)Qt)=arccos⁡∥(∂Φ‾t∗ℓ)Qt∥∥∂Φ‾t∗ℓ∥.(82)\angle\left(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\right)=\arccos\frac{\|(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\|}{\|\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell\|}. \tag*{(82)}

Thus, from (73), we know

ℜ⟨∂Φ‾t∗ℓ,D(Φ‾t,(∂Φ‾t∗ℓ)Qt,η)⟩∥∂Φ‾t∗ℓ∥∥D(Φ‾t,(∂Φ‾t∗ℓ)Qt,η)∥=cos⁡∠(∂Φ‾t∗ℓ,D(Φ‾t,(∂Φ‾t∗ℓ)Qt,η))≥cos⁡(∠(∂Φ‾t∗ℓ,(∂Φ‾t∗ℓ)Qt)+arccos⁡ϵ)=ϵ∥(∂Φ‾t∗ℓ)Qt∥−1−ϵ2(∥∂Φ‾t∗ℓ∥2−∥(∂Φ‾t∗ℓ)Qt∥2)1/2∥∂Φ‾t∗ℓ∥≥(ϵ−ϵ2−1+ζ)∥(∂Φ‾t∗ℓ)Qt∥∥∂Φ‾t∗ℓ∥.(83)\begin{aligned} \frac{\Re\langle\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell,D(\overline{\boldsymbol{\Phi}}_t,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t,\eta)\rangle}{\|\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell\|\|D(\overline{\boldsymbol{\Phi}}_t,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t,\eta)\|} =\cos\angle\left(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell,D(\overline{\boldsymbol{\Phi}}_t,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t,\eta)\right) \\ &\ge\cos\left(\angle\left(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\right)+\arccos\epsilon\right) \\ &=\frac{\epsilon\|(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\|-\sqrt{1-\epsilon^2}\left(\|\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell\|^2-\|(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\|^2\right)^{1/2}}{\|\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell\|} \\ &\ge\left(\epsilon-\sqrt{\epsilon^2-1+\zeta}\right)\frac{\|(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\|}{\|\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell\|}. \tag*{(83)} \end{aligned}

Thus,

ℓ(Bt+1)−ℓ(Bt)≤−2C1∥∂Φ‾t∗ℓQt∥∥D(Φ‾t,(∂Φ‾t∗ℓ)Qt,η)∥+β∥A∥22∥D(Φ‾t,(∂Φ‾t∗ℓ)Qt,η)∥2≤(−2C1+βξ2η∥A∥22)∥(∂Φ‾t∗ℓ)Qt∥∥D(Φ‾t,(∂Φ‾t∗ℓ)Qt,η)∥.(84)\begin{aligned} \ell(\mathbf{B}_{t+1})-\ell(\mathbf{B}_t) \le-2C_1\|\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell\mathbf{Q}_t\|\|D(\overline{\boldsymbol{\Phi}}_t,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t,\eta)\| +\frac{\beta\|\mathbf{A}\|^2}{2}\|D(\overline{\boldsymbol{\Phi}}_t,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t,\eta)\|^2 \\ &\le\left(-2C_1+\frac{\beta\xi_2\eta\|\mathbf{A}\|^2}{2}\right)\|(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t\|\|D(\overline{\boldsymbol{\Phi}}_t,(\partial_{\overline{\boldsymbol{\Phi}}_t}^{*}\ell)\mathbf{Q}_t,\eta)\|. \tag*{(84)} \end{aligned}

where C1:=ϵ−ϵ2−1+ζC_1 := \epsilon- \sqrt{\epsilon^2 - 1} + \zeta. From (76), we further know

C2:=−2C1+βξ2η∥A∥22<0C_2 := -2C_1 + \frac{\beta\xi_2 \eta\lVert A \rVert^2}{2} < 0
ℓ(Bt+1)−ℓ(Bt)≤ξ1ηC2∥(Φˉ∗†ℓ)Qt∥2,(85)\ell(\mathbf{B}_{t+1}) - \ell(\mathbf{B}_t) \le\xi_1 \eta C_2 \left\| \left(\bar{\boldsymbol{\Phi}}_*^\dagger\ell\right) Q_t \right\|^2, \tag*{(85)}

Since ℓ\ell is bounded below and ξ1ηC2<0\xi_1\eta C_2 < 0, summing this inequality bounds the sum of squared projected-gradient residuals. Hence these residuals tend to zero, using the same telescoping argument as in the proof of Theorem 1. This establishes the projected-stationarity conclusion under the stated directional conditions. □

The following remarks clarify the scope of the assumptions in Theorem 2.

Remark 3. As LL increases, the achievable set of the cascaded SIM response generally becomes richer, making the relaxation in Assumption 1 less restrictive. When the actual achievable set occupies only a subspace of the ambient matrix space, the same analysis can be interpreted within that reachable subspace.

Remark 4. Conditions (73)–(76) require alignment and magnitude control of the aggregate update. A small step size alone does not establish these properties for every cascaded SIM configuration. Theorem 2 is therefore conditional on these properties holding along the considered trajectory.

Influence of Noise

In the proposed method, we omitted the noise introduced in the backward transmission. Such noise is unavoidable in wireless systems. However, it is not always harmful. As shown in Theorem 1 of [4], adding noise after each step does not affect the convergence of the gradient descent algorithm in convex problems while reducing the convergence rate by a constant multiple determined by the noise level.

Moreover, according to the saddle-escape theory [49], [50], introducing small random perturbations enables the iteration to escape such regions along directions of negative curvature and continue descending. Since ℓ1\ell_1 and ℓ2\ell_2 are non-convex with respect to the SIM coefficients, saddle points cannot be excluded in general. Moreover, the projection matrix QtQ_t may introduce spurious saddle-like stationary regions near the constraint boundary, where the projected gradient vanishes but the true gradient does not. However, such points are non-stationary in the original space. Consequently, the perturbed projected gradient method can reduce the possibility of falling into these artificial saddle points, although convergence to a global optimum is not guaranteed.

Transmission and Computational Complexity

We first consider the transmission overhead of the proposed methods. Let TT denote the number of P2GD iterations. In each iteration, both UE-P2GD and BS-P2GD require NN downlink MIMO symbols for orthogonal pilot transmission and NN uplink MIMO symbols for over-the-air gradient feedback. Therefore, the signaling overhead over a complete optimization episode is

ODLP2GD=NT,OULP2GD=NT.(86)\mathcal{O}_{\mathrm{DL}}^{\mathrm{P2GD}} = NT,\qquad\mathcal{O}_{\mathrm{UL}}^{\mathrm{P2GD}} = NT. \tag*{(86)}

The downlink and uplink overheads are reported separately since the two directions may occupy different radio resources in a practical TDD system.

For comparison, conventional CSI-based beamforming requires explicit estimation of the SIM-UE channel G∈CM×N\mathbf{G} \in\mathbb{C}^{M \times N} before beamforming optimization. The channel contains MNMN unknown complex coefficients. A MIMO pilot symbol observed over the NN BS antenna dimensions can provide at most NN independent complex observations. Hence, even under an ideal noise-free and fully identifiable measurement model, the number of MIMO pilot symbols required for explicit channel estimation satisfies

TCEN≥MN,TCE≥M.(87)T_{\mathrm{CE}}N \ge MN,\qquad T_{\mathrm{CE}} \ge M. \tag*{(87)}

Therefore, MM MIMO pilot symbols are used as an ideal lower bound for explicit CSI acquisition in the following comparison. Practical channel estimation generally requires additional pilot observations in the presence of noise and model imperfections.

Besides air-interface signaling, each P2GD iteration may also incur latency associated with SIM reconfiguration, control signaling, synchronization, and UE-side gradient processing. Let TDLT_{\mathrm{DL}} and TULT_{\mathrm{UL}} denote the duration of one downlink and uplink MIMO symbol, respectively, and let TreconfT_{\mathrm{reconf}}, TctrlT_{\mathrm{ctrl}}, TsyncT_{\mathrm{sync}}, and TgradT_{\mathrm{grad}} denote the corresponding per-iteration implementation delays. The end-to-end latency of a complete P2GD optimization episode can be characterized as

τP2GD=T(NTDL+NTUL+Treconf+Tctrl+Tsync+Tgrad).(88)\tau_{\mathrm{P2GD}} = T\left(NT_{\mathrm{DL}} + NT_{\mathrm{UL}} + T_{\mathrm{reconf}} + T_{\mathrm{ctrl}} + T_{\mathrm{sync}} + T_{\mathrm{grad}}\right). \tag*{(88)}

For a channel coherence interval TcohT_{\mathrm{coh}}, the normalized optimization latency is further defined as

ρP2GD=τP2GDTcoh.(89)\rho_{\mathrm{P2GD}} = \frac{\tau_{\mathrm{P2GD}}}{T_{\mathrm{coh}}}. \tag*{(89)}

The proposed method is particularly attractive when a satisfactory beamforming solution can be obtained within a small number of iterations, such that the accumulated iterative signaling overhead remains below or comparable to the pilot cost required for explicit CSI acquisition. This regime becomes increasingly relevant for large SIM dimensions MM, since the explicit channel-estimation cost grows with the number of unknown spatial dimensions, whereas the per-iteration signaling cost of P2GD is determined by NN.

As for the computational complexity, we briefly summarize the dominant costs of the proposed UE-P2GD and BS-P2GD algorithms. In each iteration, the BS first generates the pilot matrix X0=FP1/2\mathbf{X}_0 = \mathbf{F}\mathbf{P}^{1/2} and updates the effective SIM response Bℓ=Bℓ−1Hℓ−1Φℓ\mathbf{B}_{\ell} = \mathbf{B}_{\ell-1}\mathbf{H}_{\ell-1}\boldsymbol{\Phi}_{\ell}, which requires O(LNM2)\mathcal{O}(LNM^2) complex multiplications. The projected-gradient step further involves computing the pseudo-inverse (Bℓ⊤)†(\mathbf{B}_{\ell}^{\top})^\dagger and multiplying it with the N×NN \times N uplink matrix U0\mathbf{U}_0, which scales as O(LN2M)\mathcal{O}(LN^2M) when M≥NM \ge N. Since the SIM phase matrices are diagonal, only MM entries per layer need to be updated, so the cost of forming and projecting the diagonal gradients from (30) and (37)/(52) is of order O(LNM)\mathcal{O}(LNM) and is negligible compared with the matrix–matrix multiplications. On the UE side, computing the local gradients {di}i=1N\{d_i\}_{i=1}^{N} from (26) or (41) involves only O(N2)\mathcal{O}(N^2) operations per iteration. Therefore, the overall per-iteration computational complexity of both proposed schemes is

O(LNM2+LN2M)≈O(LNM2).(90)\mathcal{O}(LNM^2 + LN^2M) \approx\mathcal{O}(LNM^2). \tag*{(90)}

Thus, the total complexity over TT iterations scales as O(TLNM2)\mathcal{O}(TLNM^{2}). Compared with conventional SIM beamforming approaches that require explicit estimation of the full M×NM \times N channel matrix followed by alternating optimization over dense precoding variables, the proposed over-the-air gradient framework avoids cubic dependence on the SIM size and remains computationally tractable even for large-aperture metasurfaces.

Conclusion Remark

The above analysis characterizes the projected-gradient mechanism under an idealized asymptotic regime with sufficiently small step sizes. The practical objective of P2GD, however, is not to pursue asymptotic convergence through a large number of iterations. Since each iteration incurs additional downlink and uplink signaling overhead, excessive iterations would eventually diminish the advantage of avoiding explicit channel estimation. Therefore, in practical operation, P2GD is executed with a limited iteration budget Tmax⁡T_{\max} and may employ a more aggressive step size to obtain a satisfactory beamforming solution within the available coherence and signaling budget.

Simulation Results

Simulation Scenario

In this paper, we consider the scenario introduced in Section II, where the simulation parameters are listed in Table I. All reported results are averaged over 10,000 realizations under each parameter setting. In Figure 3, the error bars extend by 0.1 times the empirical standard deviation above and below the mean. They show scaled realization-to-realization variability, rather than standard errors or confidence intervals. The larger sample size reduces sampling fluctuations in the average curves; it does not imply that every individual realization follows a smooth trajectory.

Parameter Name / FunctionNotationValue
Users / BS antennasNN3
SIM elements per layerMM8×8=648 \times8 = 64
# of SIM layersLL5
Carrier frequencyfcf_c28 GHz
Wavelengthλc\lambda_c1.07 cm
Element spacingded_eλc/2\lambda_c/2
Element apertureAeA_eλc2/4\lambda_c^2/4
Paths per UEPpathP_{\mathrm{path}}3
SIM overall thicknessDD4.286 cm (4λc4\lambda_c)
BS & SIM deployment heighthh10 m
UE distribution range(dr,min⁡,dr,max⁡)(d_{r,\min},d_{r,\max})(100, 250) m
Path loss referenceC0C_0−35-35 dB
Path loss exponentμ\mu3.5
Noise per sample/directionqq−150-150 dBm
Numerical amplitude scaleSS10810^8
Log power regularizationϵZF\epsilon_{\mathrm{ZF}}10−610^{-6}
Initial penalty coefficientλ0\lambda_010
Default penalty ratio targetrr1
Noise-power sweep—−175:5:−135-175\mathbin{:}5\mathbin{:}-135 dBm
BS maximum powerPmax⁡P_{\max}100 mW
UE maximum powerPUEP_{\mathrm{UE}}50 mW
Power step sizeηP\eta_P500
SIM step sizeηΦ\eta_\Phi500\sqrt{500}
Learning-rate sweepη\eta5, 50, 500, 5000
Adaptive λ\lambda target sweep—0.1, 0.3, 1, 3, 10, 30, 100
Maximum iterationsTmax⁡T_{\max}10
Number of realizations—10,000
UE-number sweepNN2, 3, 4, 5
SIM-size sweepMM49, 64, 81
Yao baseline training blocks—22

Table I. Basic simulation parameters and parameter sweeps.

Performance comparison among different methods

Figure 3. Performance comparison among different methods, averaged over 10,000 realizations. Error bars show ±0.1\pm0.1 times the empirical standard deviation.

The feedback scaling parameter is initialized as α0=0.3\alpha_0 = 0.3 and updated according to

αt+1=clip⁡[0.3(0.01max⁡(Etback,10−12))1/2,10−4,10].(91)\alpha_{t+1}=\operatorname{clip}\left[0.3\left(\frac{0.01}{\max\left(E_t^{\mathrm{back}},10^{-12}\right)}\right)^{1/2},10^{-4},10\right]. \tag*{(91)}

Here, Etback=∥U0(t)∥F2/N2E_t^{\mathrm{back}}=\lVert\mathbf{U}_0^{(t)}\rVert_{\mathrm{F}}^2/N^2 is the mean squared magnitude of the BS-received backward signal over the NN antennas and NN feedback symbols, after feedback limiting, reciprocal propagation, and uplink noise addition, but before division by αt\alpha_t. The operator clip restricts its argument to [10−4,10][10^{-4},10]. No additional temporal smoothing is applied to this update.

The default noise power is q=−150q=-150 dBm per received complex sample in each direction, with independent uplink and downlink noise. In physical units, σphys2=10(q−30)/10\sigma_{\mathrm{phys}}^2=10^{(q-30)/10} W. For numerical conditioning, the channel amplitudes are multiplied by S=108S=10^8, and the noise variance used in optimization and rate evaluation is scaled consistently as σcode2=S2σphys2\sigma_{\mathrm{code}}^2=S^2\sigma_{\mathrm{phys}}^2. Thus, the default code-domain noise variance is 10−210^{-2}. The numerical implementation adds ϵZF=10−6\epsilon_{\mathrm{ZF}}=10^{-6} to the powers entering the logarithmic objectives and their corresponding gradient denominators, including the main signal-power term. This constant is expressed in the scaled power units and is not a pseudoinverse threshold.

The learning-rate parameter controls two fixed step sizes, ηP=η\eta_P=\eta and ηΦ=η\eta_\Phi=\sqrt{\eta}, with default η=500\eta=500. No line search or gradient-norm normalization is used. After a power update, entries are clipped to at least 10−410^{-4} before rescaling to the total power budget; this is not a strict lower bound after rescaling. The SIM coefficients are projected onto the complex unit disk.

The penalty coefficient is adaptive, with λ0=10\lambda_0=10 and default energy-ratio target r=1r=1. Let Emain,tE_{\mathrm{main},t} and Epen,tE_{\mathrm{pen},t} denote the mean squared main and unweighted penalty gradients used by the controller. We update

λt⋆=rmax⁡(Emain,t,10−12)max⁡(Epen,t,10−12).(92)\lambda_t^\star=\sqrt{r\frac{\max\left(E_{\mathrm{main},t},10^{-12}\right)}{\max\left(E_{\mathrm{pen},t},10^{-12}\right)}}. \tag*{(92)}
λt+1=clip⁡(0.5λt+0.5λt⋆,10−4,104).(93)\lambda_{t+1}=\operatorname{clip}\left(0.5\lambda_t+0.5\lambda_t^\star,10^{-4},10^4\right). \tag*{(93)}

This targets λt2Epen,t/Emain,t≈r\lambda_t^2E_{\mathrm{pen},t}/E_{\mathrm{main},t}\approx r; smoothing and clipping mean that the target need not hold exactly at every iteration. UE-P2GD measures these energies on the main and penalty matrices before feedback construction. BS-P2GD averages the SIM-coefficient gradient energies over elements and then layers, obtaining the unweighted penalty energy from the weighted energy divided by λt2+10−12\lambda_t^2+10^{-12}. Thus, the two methods use the same ratio target at different gradient locations.

Comparison Algorithms

We compare six methods: the proposed UE-P2GD and BS-P2GD, their full-CSI gradient references (UE Full-CSI and BS Full-CSI), and two channel-estimation-based baselines (Yao-LS+GD and Yao-RSLS-iso+GD). The full-CSI references use the corresponding UE-side or BS-side loss with access to the full channel gradient. They serve as references for assessing the effect of gradient projection, rather than guaranteed performance upper bounds for finite-iteration nonconvex optimization.

The two Yao baselines use least-squares (LS) and isotropic reduced-subspace least-squares (RSLS-iso) channel estimation based on [25], respectively, followed by gradient-based beamforming optimization. Both use 22 training blocks in the reported comparisons. The horizontal axis in Figure 3 counts optimization iterations, each of which includes downlink pilot transmission and uplink feedback for P2GD; it does not include the preceding channel-estimation cost of these baselines. Thus, this figure compares optimization trajectories rather than performance under an equal total signaling or computational budget.

For the Yao baselines, each training block contains N=3N = 3 simultaneous-user uplink symbols with orthogonal DFT pilots satisfying QQH=NI\mathbf{Q}\mathbf{Q}^{H} = N\mathbf{I}. Each UE transmits at 50 mW per symbol. Independent random unit-modulus SIM coefficients are used in each block, without training-codebook MSE optimization. Consequently, 22 blocks require 66 uplink MIMO symbols. The LS estimate uses the stacked sensing matrix pseudoinverse. The RSLS-iso estimate additionally projects the observations onto the eigenspace of ARisoAH\mathbf{A}\mathbf{R}_{\mathrm{iso}}\mathbf{A}^{H}, where A∈C66×64\mathbf{A} \in\mathbb{C}^{66 \times64} is the sensing matrix and [Riso]m,m~=sinc⁡(2dm,m~/λc)[\mathbf{R}_{\mathrm{iso}}]_{m,\tilde{m}} = \operatorname{sinc}(2d_{m,\tilde{m}}/\lambda_{\mathrm{c}}), with sinc⁡(x)=sin⁡(πx)/(πx)\operatorname{sinc}(x) = \sin(\pi x)/(\pi x). This isotropic estimator covariance is distinct from the finite-path channel-generation model. Relative thresholds of 10−1010^{-10} are used for eigenvalue retention and pseudoinversion. In all 10,000 fixed-noise realizations at this training budget, both the sensing rank and the retained subspace rank are 64.

After estimation, each baseline performs ten full-gradient UE-objective updates on the fixed estimated channel, using the step sizes, adaptive controls, and feasibility operations described above. The local gradient computation does not add over-the-air observation noise. The achieved rates are evaluated on the true channel with the corresponding noise variance. These are estimator-based adaptations of the cited work to our beamforming setting.

Signaling Overhead Comparison

Using the overhead formulation in § 6-D, we further provide a quantitative comparison under the default system configuration. Specifically, with N=3N = 3, M=64M = 64, and T=10T = 10, both UE-P2GD and BS-P2GD require

ODLP2GD=OULP2GD=NT=30(94)\mathcal{O}_{\mathrm{DL}}^{\mathrm{P2GD}} = \mathcal{O}_{\mathrm{UL}}^{\mathrm{P2GD}} = NT = 30 \tag*{(94)}

MIMO symbols in the downlink and uplink, respectively.

For explicit CSI acquisition, the ideal identifiability lower bound is TCE≥M=64T_{\mathrm{CE}} \ge M = 64 MIMO pilot symbols. This lower bound does not correspond to a specific channel estimator and represents the minimum pilot dimension required to identify the MNMN unknown complex coefficients of G∈CM×N\mathbf{G} \in\mathbb{C}^{M \times N}.

The resulting signaling requirements are summarized in Table II. The implemented 22-block Yao baselines use 66 uplink training symbols, compared with 30 uplink and 30 downlink symbols for ten P2GD iterations. Their local optimization costs are different; equal iteration counts do not imply equal computational time. The comparison shows that, under the adopted finite-iteration setting, the proposed methods operate with a small per-direction signaling budget while avoiding explicit CSI acquisition.

MethodDownlinkUplinkExplicit CSI
UE-P2GDNT=30NT = 30NT=30NT = 30No
BS-P2GDNT=30NT = 30NT=30NT = 30No
CSI-based beamformingTCE≥M=64T_{\mathrm{CE}} \ge M = 64 pilot symbolsYes

Table II. Signaling overhead under the default setting (N=3N = 3, M=64M = 64, and T=10T = 10).

Comparison Among Different Methods

Figure 3 compares the average sum-rate trajectories of the six methods. BS-P2GD improves rapidly during the first few iterations and then remains close to the BS Full-CSI reference, with a modest projection-related performance gap. UE-P2GD improves more slowly initially and approaches BS-P2GD near the end of the ten-iteration budget. In contrast, UE Full-CSI and the two Yao baselines continue improving and achieve higher average sum rates in the later iterations. The Yao-LS+GD and Yao-RSLS-iso+GD curves nearly overlap with the UE Full-CSI under the reported 22-block training setting.

The averages over 10,000 realizations exhibit clear trends with limited late-iteration fluctuations. These results support the usefulness of BS-P2GD for obtaining a satisfactory solution within a small number of over-the-air iterations, while also showing the performance tradeoff relative to explicit-CSI-based optimization. The scaled error bars describe variability across realizations and should not be interpreted as guarantees of stability for individual runs.

Effect of Learning Rate and Iteration Steps

Figure 4 compares learning rates η∈{5,50,500,5000}\eta\in\{5, 50, 500, 5000\} within a ten-iteration budget. For BS-P2GD, all tested settings improve rapidly, while larger learning rates generally yield higher average rates within this budget. UE-P2GD is more sensitive to the step size: a small learning rate leads to slower improvement, whereas larger values yield substantially higher rates by the final iteration. The largest tested value also produces a small late-iteration decline for UE-P2GD.

Effect of learning rate in P2GD algorithms

Figure 4. Effect of learning rate in P2GD algorithms.

These observations concern finite-iteration performance rather than asymptotic convergence, and illustrate the importance of choosing a step size appropriate to the available signaling budget.

Effect of Background Noise Power

Figure 5 examines noise powers per received complex sample in each direction from −175-175 to −135-135 dBm in steps of 55 dB. The average sum rate decreases with increasing noise power for all six methods. BS-P2GD remains close to BS Full-CSI across the sweep. UE-P2GD performs better than BS-P2GD at the lower noise levels, while their rates become similar as the noise increases. UE Full-CSI and the two Yao baselines exhibit nearly overlapping curves and achieve higher rates throughout the tested range. Noise affects both the achievable communication rate and the gradient information used by P2GD; this comparison alone does not isolate the contributions of these two effects.

Performance over different background noise levels

Figure 5. Performance over different background noise levels.

Effect of SIM Elements per Layer and UE Numbers

Figure 6 compares 7×77 \times7, 8×88 \times8, and 9×99 \times9 SIM arrays, corresponding to M∈{49,64,81}M \in\{49, 64, 81\}, as the number of UEs NN varies from 22 to 55. At a fixed UE count, increasing the SIM size generally improves the average sum rate, with only

Effect of problem size to the proposed P2GD algorithms

Figure 6. Effect of problem size to the proposed P2GD algorithms.

a small difference between the 8×88 \times8 and 9×99 \times9 UE-P2GD results at N=2N = 2. For each array size, the average rate peaks at N=3N = 3 and decreases as more users are served. This decrease is more pronounced for UE-P2GD, whereas BS-P2GD performs better at N=4N = 4 and N=5N = 5. These trends describe the achieved rates of the proposed algorithms under the tested settings, rather than the optimal system sum rate.

Sensitivity to the Adaptive λ\lambda Ratio Target

Figure 7 examines the adaptive λ\lambda ratio target over {0.1,0.3,1,3,10,30,100}\{0.1, 0.3, 1, 3, 10, 30, 100\}. The horizontal axis denotes the target used by the adaptive penalty setting, rather than a fixed value of the penalty coefficient throughout optimization. BS-P2GD maintains a similar average sum rate across the tested range. UE-P2GD is more sensitive: its rate increases up to a target of approximately 11 and then decreases, with a pronounced loss for targets of 1010 and above. Thus, BS-P2GD is less sensitive to this tuning parameter in the tested setting, whereas UE-P2GD requires a more careful choice. Table III complements the figure with residual interference and minimum-user-rate results after ten updates.

Average sum rate versus the adaptive $\lambda$ ratio target

Figure 7. Average sum rate versus the adaptive λ\lambda ratio target.

For evaluation, define the physical effective channel including power allocation as Hijeff=pjgjTBejH_{ij}^{\mathrm{eff}} = \sqrt{p_j}\mathbf{g}_j^{\mathsf{T}}\mathbf{B}\mathbf{e}_j. The residual interference is Ires=∑i∑j≠i∣Hijeff∣2I_{\mathrm{res}} = \sum_i \sum_{j \ne i} \lvert H_{ij}^{\mathrm{eff}}\rvert^2 in watts, and the minimum user rate is

MethodrrSum rateIresI_{\mathrm{res}}Rmin⁡R_{\min}
UE-P2GD0.18.2760.8170.032
BS-P2GD0.19.6641.0620.037
UE-P2GD0.38.9740.7870.024
BS-P2GD0.39.6911.0810.036
UE-P2GD19.6040.6720.017
BS-P2GD19.6621.0610.036
UE-P2GD38.9490.4740.035
BS-P2GD39.6271.0820.035
UE-P2GD105.0870.4610.150
BS-P2GD109.7471.0940.036
UE-P2GD302.9940.9240.163
BS-P2GD309.6011.1300.036
UE-P2GD1001.7551.7470.129
BS-P2GD1009.6891.1440.037

Table III. Metrics after ten updates, averaged over 10,000 realizations. Rates are in bps/Hz and interference is in units of 10−1310^{-13} W.

Rmin⁡=min⁡ilog⁡2(1+∣Hiieff∣2∑j≠i∣Hijeff∣2+σphys2).R_{\min} = \min_i \log_2\left(1 + \frac{\left|H^{\mathrm{eff}}_{ii}\right|^2}{\sum_{j \ne i}\left|H^{\mathrm{eff}}_{ij}\right|^2 + \sigma_{\mathrm{phys}}^2}\right).

For UE-P2GD, increasing rr from 0.1 to 1 improves sum rate and reduces residual interference, while further increases eventually sacrifice sum rate and do not monotonically reduce interference. BS-P2GD has a comparatively insensitive average sum rate across the tested targets, but its residual interference remains nonzero. The small minimum-user rates also show that maximizing this surrogate does not guarantee user fairness. BS-P2GD is attractive when fast initial improvement and reduced tuning sensitivity are priorities; UE-P2GD can be preferable for achieved sum rate in the low-noise settings of Figure 5, provided its penalty target is chosen appropriately.

Conclusion and Future Directions

This work introduces a gradient-driven optimization framework for SIM-assisted TDD MIMO systems, aiming to overcome the inherent limitations of conventional channel estimation and precoding strategies. By exploiting channel reciprocity and incorporating over-the-air gradient feedback, the proposed approach enables the joint refinement of SIM configuration and transmit power allocation under realistic system constraints. The theoretical analysis characterizes the projected-gradient mechanism under idealized conditions and provides insights into its behavior, while the performance of the practical non-convex algorithms is evaluated numerically. Simulations show that BS-P2GD remains close to its full-CSI reference within ten over-the-air iterations, whereas explicit-CSI-based optimization attains higher sum rates at a larger pilot cost. The proposed approach is therefore most attractive when frequent channel variations make repeated explicit channel estimation costly.

The surrogate loss functions and finite-iteration parameter choices still influence the achieved sum rate, especially for UE-P2GD. Improving their alignment with the sum-rate objective remains a direction for further investigation. Meanwhile, the modular nature of the proposed architecture suggests strong extensibility, allowing integration with advanced optimization techniques and adaptation to emerging paradigms such as multi-base station coordination, UAV-assisted networks, and semantic communication. These attributes position the framework as a versatile foundation for future research on intelligent and scalable wireless systems.

References

  1. [1]Y. Yang, H. Liu, L. Gan, et al., “Projected physical gradient method for efficient beamforming based on SIMs,” in Proc. 20th International Symposium on Wireless Communication Systems (ISWCS), IEEE, 2026, pp. 1–6.DOI
  2. [2]J. An, C. Xu, D. W. K. Ng, et al., “Stacked intelligent metasurfaces for efficient holographic MIMO communications in 6G,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 8, pp. 2380–2396, Aug. 2023. DOI: 10.1109/JSAC.2023.3282861arxiv.org/abs/2305.08079
  3. [3]J. An, M. Di Renzo, M. Debbah, et al., “Stacked intelligent metasurfaces for multiuser downlink beamforming in the wave domain,” IEEE Transactions on Wireless Communications, vol. 24, no. 7, pp. 5525–5538, Jul. 2025. DOI: 10.1109/TWC.2025.3547779DOI
  4. [4]Y. Yang, Z. Zhang, Y. Tian, et al., “Over-the-air split machine learning in wireless mimo networks,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 4, pp. 1007–1022, 2023. DOI: 10.1109/JSAC.2023.3242701DOI
  5. [5]Y. Yang, Z. Zhang, Y. Tian, et al., “Realizing over-the-air neural networks in ris-assisted mimo communication systems,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC), 2024, pp. 1–5. DOI: 10.1109/WCNC57260.2024.10571129DOI
  6. [6]X. Li, Y. Yang, and T. Guo, “Keep pace to learn: Movable antennas restore the backpropagated gradient under user mobility,” arXiv preprint arXiv:2610.23118, 2026.arxiv.org/abs/2610.23118
  7. [7]X. Li, Y. Yang, and T. Guo, “Coherent but discordant: Rotate, don’t invert, to backpropagate through an aging channel,” arXiv preprint arXiv:2610.23137, 2026.
  8. [8]X. Li, Y. Yang, and T. Guo, “Wait to learn: Lobes and dead zones in physical backpropagation over an aging single-antenna channel,” arXiv preprint arXiv:2610.23135, 2026.
  9. [9]Y. Liu, X. Liu, X. Mu, et al., “Reconfigurable intelligent surfaces: Principles and opportunities,” IEEE Communications Surveys & Tutorials, vol. 23, no. 3, pp. 1546–1577, May 2021.
  10. [10]R. Deng, Y. Zhang, H. Zhang, et al., “Reconfigurable holographic surface: A new paradigm to implement holographic radio,” IEEE Vehicular Technology Magazine, vol. 18, no. 1, pp. 20–28, Mar. 2023.arxiv.org/abs/2212.13666
  11. [11]N. U. Hassan, J. An, M. Di Renzo, et al., “Efficient beamforming and radiation pattern control using stacked intelligent metasurfaces,” IEEE Open Journal of the Communications Society, vol. 5, pp. 599–611, Jan. 2024.DOI
  12. [12]X. Jia, J. An, H. Liu, et al., “Stacked intelligent metasurface enabled near-field multiuser beamfocusing in the wave domain,” in 2024 IEEE 99th Vehicular Technology Conference (VTC2024-Spring), 2024, pp. 1–5. DOI: 10.1109/VT C2024-Spring62846.2024.10683447DOI
  13. [13]Y. Hu, J. Zhang, E. Shi, et al., “Joint beamforming and power allocation design for stacked intelligent metasurfaces-aided cell-free massive mimo systems,” IEEE Transactions on Vehicular Technology, vol. 74, no. 3, pp. 5235–5240, Mar. 2025. DOI: 10.1109/TVT.2024.3499968
  14. [14]E. Shi, J. Zhang, Y. Zhu, et al., “Uplink performance of stacked intelligent metasurface-enhanced cell-free massive MIMO systems,” IEEE Transactions on Wireless Communications, vol. 24, no. 5, pp. 3731–3746, May 2025. DOI: 10.1109/TWC.2025.3540404DOI
  15. [15]S. Li, F. Zhang, M. Tao, et al., “Transmit beamforming design for ISAC with stacked intelligent metasurfaces,” IEEE Transactions on Vehicular Technology, vol. 74, no. 4, pp. 6767–6772, Apr. 2025. DOI: 10.1109/TVT.2024.3517709DOI
  16. [16]H. Niu, J. An, A. Papazafeiropoulos, et al., “Stacked intelligent metasurfaces for integrated sensing and communications,” IEEE Wireless Communications Letters, vol. 13, no. 10, pp. 2807–2811, Oct. 2024. DOI: 10.1109/LWC.2024.3447272DOI
  17. [17]S. Lin, J. An, L. Gan, et al., “Stacked intelligent metasurface enabled LEO satellite communications relying on statistical CSI,” IEEE Wireless Communications Letters, vol. 13, no. 5, pp. 1295–1299, May 2024. DOI: 10.1109/LWC.2024.3368238DOI
  18. [18]G. Huang, J. An, Z. Yang, et al., “Stacked intelligent metasurfaces for task-oriented semantic communications,” IEEE Wireless Communications Letters, vol. 14, no. 2, pp. 310–314, Feb. 2025. DOI: 10.1109/LWC.2024.3499970
  19. [19]Y. Sun, Z. Lin, H. Zhang, et al., “Towards energy-efficient holographic MIMO communications via stacked metasurface-assisted semantic beamforming,” in ICC 2025 - IEEE International Conference on Communications, 2025, pp. 01–07. DOI: 10.1109/ICC52391.2025.11161925DOI
  20. [20]A. Papazafeiropoulos, J. An, P. Kourtessis, et al., “Achievable rate optimization for stacked intelligent metasurface-assisted holographic MIMO communications,” IEEE Transactions on Wireless Communications, vol. 23, no. 10, pp. 13 173–13 186, 2024. DOI: 10.1109/TWC.2024.3399318DOI
  21. [21]H. Liu, J. An, G. C. Alexandropoulos, et al., “Multi-user MISO with stacked intelligent metasurfaces: A DRL-based sum-rate optimization approach,” IEEE Transactions on Cognitive Communications and Networking, pp. 1–15, Apr. 2025. DOI: 10.1109/TCCN.2025.3558008DOI
  22. [22]H. Liu, J. An, X. Jia, et al., “Stacked intelligent metasurfaces for wireless communications: Applications and challenges,” IEEE Wireless Communications, vol. 32, no. 4, pp. 46–53, Aug. 2025. DOI: 10.1109/MWC.001.2500002DOI
  23. [23]Q.-U.-A. Nadeem, J. An, and A. Chaaban, “Hybrid digital-wave domain channel estimator for stacked intelligent metasurface enabled multi-user MISO systems,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC), 2024, pp. 1–6. DOI: 10.1109/WCNC57260.2024.10571026DOI
  24. [24]Z. Chen, H. Shin, and A. Nallanathan, “Generative diffusion model-based variational inference for mimo channel estimation,” IEEE Transactions on Communications, vol. 73, no. 10, pp. 9254–9269, 2025. DOI: 10.1109/TCOMM.2025.3556753DOI
  25. [25]X. Yao, J. An, L. Gan, et al., “Channel estimation for stacked intelligent metasurface-assisted wireless networks,” IEEE Wireless Communications Letters, vol. 13, no. 5, pp. 1349–1353, May 2024. DOI: 10.1109/LWC.2024.3369874DOI
  26. [26]X. Yao, J. An, G. Huang, et al., “Sparse channel estimation for stacked intelligent metasurface-assisted mmwave communications,” in 2024 IEEE VTS Asia Pacific Wireless Communications Symposium (APWCS), Aug. 2024, pp. 1–5. DOI: 10.1109/APWCS61586.2024.10679297DOI
  27. [27]N. Ginige, A. S. D. Sena, N. H. Mahmood, et al., “Nested tensor-based channel estimation for stacked intelligent metasurface-assisted wireless networks,” IEEE Communications Letters, pp. 1–5, Sep. 2025. DOI: 10.1109/LCOMM.2025.3611163DOI
  28. [28]A. Lawal, A. Zerguine, A. A. Nasir, et al., “Channel estimation for stacked intelligent metasurface-aided network using deep learning,” IEEE Communications Letters, pp. 1–5, Sep. 2025. DOI: 10.1109/LCOMM.2025.3607027DOI
  29. [29]Z. Chen, H. Shin, and A. Nallanathan, “Sampling-free diffusion transformers for low-complexity mimo channel estimation,” IEEE Wireless Communications Letters, vol. 15, pp. 3164–3168, 2026. DOI: 10.1109/LWC.2026.3693087DOI
  30. [30]J. Zhu, Y. Shi, Y. Zhou, et al., “Over-the-air federated learning and optimization,” IEEE Internet of Things Journal, vol. 11, no. 10, pp. 16 996–17 020, 2024.arxiv.org/abs/2310.10089
  31. [31]B. Xiao, X. Yu, W. Ni, et al., “Over-the-air federated learning: Status quo, open challenges, and future directions,” Fundamental Research, 2024.DOI
  32. [32]M. F. U. Abrar and N. Michelusi, “Non-convex over-the-air heterogeneous federated learning: A bias-variance trade-off,” arXiv preprint arXiv:2510.26722, 2025.
  33. [33]D. Zhang, M. Xiao, and M. Skoglund, “Over-the-air computation empowered federated learning: A joint uplink-downlink design,” in 2023 IEEE 98th Vehicular Technology Conference (VTC2023-Fall), IEEE, 2023, pp. 1–7.arxiv.org/abs/2311.04059
  34. [34]Y. Yuan, Z. Wei, and W. Guo, “Mixed-precision federated learning via multi-version over-the-air aggregation,” in 2025 IEEE Wireless Communications and Networking Conference (WCNC), IEEE, 2025, pp. 1–7.DOI
  35. [35]N. A. Mitsiou, P. S. Bouzinis, P. D. Diamantoulakis, et al., “Accelerating distributed optimization via over-the-air computing,” IEEE Transactions on Communications, vol. 71, no. 9, pp. 5565–5579, 2023.arxiv.org/abs/2212.14003
  36. [36]N. Agrawal, R. L. Cavalcante, and S. Stéphanz, “Dynamic distributed convex optimization” over-the-air” in decentralized wireless networks,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2023, pp. 1–5.DOI
  37. [37]Z. Wang, Y. Zhao, Y. Zhou, et al., “Over-the-air computation for 6G: Foundations, technologies, and applications,” IEEE Internet of Things Journal, vol. 11, no. 14, pp. 24 634–24 658, 2024.DOI
  38. [38]Z. Chen, E. G. Larsson, C. Fischione, et al., “Over-the-air computation for distributed systems: Something old and something new,” IEEE Network, vol. 37, no. 5, pp. 240–246, 2023.arxiv.org/abs/2211.00767
  39. [39]W. Fang, Y. Jiang, Y. Shi, et al., “Over-the-air computation via reconfigurable intelligent surface,” IEEE transactions on communications, vol. 69, no. 12, pp. 8612–8626, 2021.arxiv.org/abs/2105.05113
  40. [40]J. Liao, Z. Chen, and E. G. Larsson, “Over-the-air federated learning with privacy protection via correlated additive perturbations,” in 2022 58th Annual Allerton Conference on Communication, Control, and Computing (Allerton), IEEE, 2022, pp. 1–8.arxiv.org/abs/2210.02235
  41. [41]G. Zhu and K. Huang, “Mimo over-the-air computation for high-mobility multimodal sensing,” IEEE Internet of Things journal, vol. 6, no. 4, pp. 6089–6103, 2018.DOI
  42. [42]H. Yang, D. Wen, L. You, et al., “Mimo over-the-air computation for device-edge collaborative inference,” IEEE Transactions on Wireless Communications, 2025.
  43. [43]H. Zhang, X. Tang, R. Zhang, et al., “Distributionally robust over-the-air computation in presence of channel uncertainties,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC), IEEE, 2024, pp. 1–6.DOI
  44. [44]Y. Yang, Z. Zhang, Y. Tian, et al., “Implementing graph neural networks over wireless networks via over-the-air computing: A joint communication and computation framework,” IEEE Wireless Communications, vol. 30, no. 3, pp. 62–69, 2023.DOI
  45. [45]J. Dan, W. Liu, C. Xie, et al., “TFGDA: Exploring topology and feature alignment in semi-supervised graph domain adaptation through robust clustering,” Advances in Neural Information Processing Systems, vol. 37, pp. 50 230–50 255, 2024.DOI
  46. [46]J. Dan, Y. Liu, B. Sun, et al., “TransFace++: Rethinking the face recognition paradigm with a focus on accuracy, efficiency, and security,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.DOI
  47. [47]N. Yan, K. Wang, K. Zhi, et al., “Secure and private over-the-air federated learning: Biased and unbiased aggregation design,” IEEE Transactions on Wireless Communications, vol. 24, no. 7, pp. 5901–5916, 2025.DOI
  48. [48]D. H. Brandwood, “A complex gradient operator and its application in adaptive array theory,” in IEE Proceedings F (Communications, Radar and Signal Processing), IET, vol. 130, 1983, pp. 11–16.
  49. [49]R. Ge, F. Huang, C. Jin, et al., “Escaping from saddle points—online stochastic gradient for tensor decomposition,” in Conference on learning theory, PMLR, 2015, pp. 797–842.arxiv.org/abs/1503.02101
  50. [50]S. S. Du, C. Jin, J. D. Lee, et al., “Gradient descent can take exponential time to escape saddle points,” Advances in neural information processing systems, vol. 30, 2017.arxiv.org/abs/1705.10412

Paper details

vlnc:W_c73va37h
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents