Active STAR-RIS: Amplifying Intelligent Surfaces

• Active STAR-RIS is a reconfigurable intelligent surface that simultaneously transmits, reflects, and amplifies incident signals to overcome path loss and double-fading issues.
• It is applied in various domains such as downlink communications, integrated sensing and communication, SWIPT, massive MIMO, UAV-assisted IoT, and edge computing to improve signal quality and coverage.
• Active STAR-RIS systems balance amplification gains with practical challenges like thermal noise, circuit power constraints, and hardware nonlinearities, requiring careful tradeoff optimization.

Active STAR-RIS denotes a simultaneously transmitting-and-reflecting reconfigurable intelligent surface whose elements not only split an incident wave into transmitted and reflected components but also amplify it. In the hardware-oriented literature, this is realized by integrating active circuitry such as reflection-type amplifiers, which removes the strictly passive conservation regime, introduces power consumption and thermal noise, and enables either coupled or independent transmission/reflection phase control depending on the element architecture. The concept has been developed for downlink communication, integrated sensing and communication (ISAC), simultaneous wireless information and power transfer (SWIPT), massive MIMO, UAV-assisted IoT, and edge computing, while a parallel body of work uses the phrase “active” only for base-station beamforming and studies a passive STAR-RIS instead (Xu et al., 2023, Noh et al., 24 Jul 2025, Zhu et al., 2024).

1. Definitions and terminological scope

A precise distinction is necessary because the literature uses closely related but non-equivalent meanings. In hardware-centric work, an active STAR-RIS is an amplifying surface: each element can transmit and reflect simultaneously, and the surface injects additional power into the scattered field. By contrast, several papers on “active and passive beamforming” optimize active beamforming coefficients at the base station and passive transmission/reflection coefficients at the STAR-RIS; these papers explicitly state that they do not study an active STAR-RIS hardware architecture in the sense of amplifying or relaying elements at the surface (Zhang et al., 2022, Le et al., 2024, Huang et al., 2024, Li et al., 2023, Zuo et al., 2021, Kavianinia et al., 2022).

Usage	Meaning	Representative papers
Active STAR-RIS hardware	Surface elements amplify and split waves	(Xu et al., 2023, Noh et al., 24 Jul 2025, Zhu et al., 2024)
Joint active/passive beamforming	BS beamforming is active; STAR-RIS remains passive	(Zhang et al., 2022, Le et al., 2024, Huang et al., 2024, Li et al., 2023, Zuo et al., 2021)
Hybrid architectures	Active communication node plus passive or partially active STAR-RIS operation	(Huang et al., 2024, Li et al., 7 Jan 2025)

Within the hardware-oriented definition, the active surface is usually motivated by the double-fading or multiplicative fading limitation of passive RIS and passive STAR-RIS links. The active design compensates path loss through amplification, but it also creates a new design space involving circuit bias power, amplification budgets, active-noise terms, and hardware feasibility constraints. This tradeoff, rather than amplification alone, is central to the subject (Zhu et al., 2024, Huang et al., 8 Sep 2025).

2. Element models, coefficients, and physical realizations

A common system-level representation assigns separate transmission and reflection coefficient matrices to each STAR-RIS. In the multi-STAR-RIS ISAC model, the $l$-th surface uses

$$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$

with

$$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$

where $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$ and $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$. In the passive case, energy conservation imposes $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$; in the active case, the unit-modulus or passive conservation restriction is removed and amplification is allowed, while thermal noise from the active circuitry is explicitly modeled (Noh et al., 24 Jul 2025).

At the element level, the hardware model in (Xu et al., 2023) uses a quadrature hybrid coupler, PIN diode phase shifters, and reflection-type amplifiers. For dual-sided incidence,

$$\begin{pmatrix} y^A_m\ y^B_m \end{pmatrix} = \begin{pmatrix} \widetilde{R}^A_m & \widetilde{T}^{AB}_m\ \widetilde{T}^{BA}_m & \widetilde{R}^B_m \end{pmatrix} \begin{pmatrix} s^A_m\ s^B_m \end{pmatrix},$$

or $\mathbf y_m=\mathbf\Xi_m\mathbf s_m$, and an active element satisfies $\mathbf\Xi_m^H\mathbf\Xi_m \succ \mathbf I_2$. The same paper distinguishes two architectures. In the coupled phase-shift case, one delay line and one reflection-type amplifier produce

$$R^A_m = -\widetilde{G}/2,\quad R^B_m = \widetilde{G}/2,\quad T^{AB}_m=T^{BA}_m=j\widetilde{G}/2,$$

with phase coupling

$$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$0

In the independent phase-shift case, two amplifiers with gains $$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$1 and $$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$2 yield

$$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$3

so transmission and reflection amplitudes and phases can be independently configured; transmission-only and reflection-only modes appear as $$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$4 and $$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$5, respectively (Xu et al., 2023).

A system-level active model often introduces an amplification matrix

$$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$6

so that active coefficients are written as $$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$7 and $$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$8. In SWIPT formulations, this active model removes the passive coupled-phase restriction and leaves amplitude sharing $$\boldsymbol{\Theta}_{\mathrm{T},l},\ \boldsymbol{\Theta}_{\mathrm{R},l}\in\mathbb{C}^{M\times M},$$9, while adding amplified thermal noise $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$0 and active-surface power constraints (Zhu et al., 2024). Related models for SWIPT and ISAC also include binary element selection matrices, per-element gain bounds, and explicit noise terms such as $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$1, $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$2, or $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$3 (Faramarzi et al., 2024, Zhang et al., 2023).

3. Network architectures and functional roles

The most elaborate communication-sensing deployment in the provided literature is a secure ISAC network empowered by multiple active STAR-RISs arranged in cascade: $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$4 There, reflected paths serve users on the reflection side, transmission paths continue toward downstream STAR-RISs and finally the sensing target, and the target echo propagates back through the STAR-RIS chain to the base station. The surfaces form a cooperative multi-hop relay layer in an NLoS ISAC network, creating full-space coverage, establishing LoS-like cascaded links for communication and sensing, compensating multi-hop path loss using active gain, and supporting secure ISAC by steering sensing energy toward the target while suppressing leakage of user data to that same target when it acts as a potential eavesdropper (Noh et al., 24 Jul 2025).

Active STAR-RIS has also been embedded in single-surface ISAC architectures. In the full-duplex dual-function base-station model, the base station simultaneously transmits communication and radar signals, receives the radar echo with a receive beamformer, and jointly optimizes transmit beamforming with active STAR-RIS reflection and transmission beamforming under radar SINR, power, and hardware constraints. In that formulation, sensing is based on the direct DFBS-target path because the target is assumed far from the STAR-RIS, so the STAR-RIS primarily assists communications while remaining part of the joint ISAC resource allocation (Zhang et al., 2023). An OFDM-based ASTARS-aided ISAC variant places users on the transmission side and the sensing target on the reflection side, introduces frequency-selective fading, and evaluates sensing through range and velocity mean-squared error in addition to communication BER (Ge et al., 2024).

In SWIPT, the active STAR-RIS is used to create virtual LoS links to reflection-side and transmission-side users while each user applies power splitting for simultaneous information decoding and energy harvesting. One line of work maximizes energy efficiency by jointly optimizing BS beamforming, active gain, element selection, phase shifts, and user power-splitting ratios under SINR, harvested-energy, BS-power, output-power, and active-surface constraints (Faramarzi et al., 2024). Another line compares active and passive STAR-RIS for SWIPT under near-field and far-field propagation, minimizing access-point transmit power subject to information-user SINR and energy-device harvested-power constraints (Zhu et al., 2024).

Beyond communication and power transfer, active STAR-RIS can act as a sensing-capable node. In the mmWave dynamic-scatterer tracking architecture, the surface has $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$5 RF chains and can receive and process impinging signals from both sides during the preamble. Specifically, $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$6 RF chains are used to alternately activate partial STAR-RIS elements for indoor communication-channel acquisition, and one RF chain is dedicated to a fixed element for outdoor sensing. After preamble-based parameter acquisition, dynamic paths are identified from all scattering paths, dynamic targets are classified with respect to radar cross sections, and a Gaussian mixture-probability hypothesis density filter tracks the scatterers; the data phase then uses full passive communication mode with predicted BS precoding and STAR-RIS refraction phase shifts (Li et al., 7 Jan 2025).

Other architectures extend the active STAR-RIS idea to mobility and computation. A UAV-mounted active STAR-RIS uses reflection amplifiers, a power divider, and two phase shifters per element to provide full-space coverage and power compensation for IoT NOMA links, while the UAV trajectory is jointly optimized with power allocation and active STAR-RIS beamforming (Zhao et al., 5 Jan 2025). In multi-access edge computing, the active STAR-RIS is co-designed with partial task offloading, admitted tasks, and queue stability so that amplification, amplitude splitting, and phase shifts become part of the edge-resource control loop rather than merely a propagation aid (Aung et al., 2024). In massive MIMO, ASTARS is treated under a two-timescale statistical-CSI framework with correlated fading, explicit channel estimation, and simultaneous optimization of amplitudes, phase shifts, and active amplifying coefficients (Papazafeiropoulos et al., 2024).

4. Optimization frameworks and signal processing methods

The governing optimization problems are typically highly non-convex because rates, sensing metrics, and active-surface power terms couple beamforming variables multiplicatively. In the multiple-active-STAR-RIS secure ISAC problem, the objective is weighted sum-rate maximization,

$$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$7

subject to a BS transmit-power constraint, per-STAR-RIS power constraints, secrecy constraints $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$8, a sensing constraint $$\boldsymbol{\Theta}_{m,l}=\mathbf{B}_{m,l}\boldsymbol{\Phi}_{m,l},\quad m\in\{\mathrm{T},\mathrm{R}\},$$9, and phase bounds. The solution uses alternating optimization: WMMSE variables transform the sum-rate, the sensing beamformer is obtained in closed form from KKT conditions in an MVDR-like expression, communication beamformers are derived from Lagrangian/KKT conditions with dual variables for BS power, secrecy, sensing, and active STAR-RIS power, and each STAR-RIS coefficient block is updated via SCA and SDR after channel re-expression in terms of an affine coefficient vector (Noh et al., 24 Jul 2025).

ISAC formulations with one active STAR-RIS use closely related but not identical decompositions. In (Zhang et al., 2023), fractional programming converts the sum-of-log-SINR objective to a tractable form via auxiliary variables, the radar receive beamformer is the principal generalized eigenvector of a generalized Rayleigh quotient, the DFBS transmit beamforming subproblem becomes a QCQP with a first-order Taylor approximation for the radar constraint, and the active STAR-RIS subproblem depends on the work mode: UED is solved by convex optimization, EED by QCQP plus MM or complex circle manifold methods for unit-modulus phases, and SD by the same alternating framework under partition constraints. In the OFDM-based ASTARS-aided ISAC system, the main problem maximizes radar SNR subject to communication SINR, BS power, elementwise energy splitting, phase bounds, and active-gain constraints; the beamforming subproblem is a convex QSDP, while the ASTARS phase design uses a fractional-programming reformulation and MM-style majorization after vectorization (Ge et al., 2024).

Energy- and queue-oriented formulations introduce different objective functions but similar decompositions. The SWIPT energy-efficiency maximization in (Faramarzi et al., 2024) treats the problem as a non-convex MINLP and splits it into a beamforming/power-splitting block solved by SDR, DC programming, first-order Taylor approximation, and Dinkelbach fractional programming, and an active-STAR-RIS block solved by a hybrid reinforcement-learning framework: modified DDPG for binary element selection, SAC for continuous gain and phase control, and MAML-style meta-learning to produce MMDS. The MEC formulation minimizes long-term energy under queue stability by combining sequential fractional programming for transmit power, convex optimization for partial task offloading, and Lyapunov optimization with DDQN for amplitude coefficients, phase shifts, amplification coefficients, and task admission (Aung et al., 2024).

Large-scale and data-driven methods broaden this algorithmic space. In ASTARS-aided massive MIMO, the two-timescale design updates the BS precoder on the fast timescale from end-to-end channel estimates while optimizing the STAR-RIS amplitudes, phases, and active coefficients on the slow timescale from statistical CSI. The paper derives a closed-form achievable sum-rate expression and applies projected gradient ascent with explicit projections for phase, amplitude, and amplification variables; the pilot length is $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$0, so the channel-estimation overhead does not scale with $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$1 or $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$2 (Papazafeiropoulos et al., 2024). By contrast, the distributed STAR-RIS HGNN framework in (Le et al., 2024) learns joint active beamforming at the BS and passive STAR-RIS coefficient control through a heterogeneous graph, but the surface there remains passive; the result is relevant to algorithm design for STAR-RIS systems, not to active STAR-RIS hardware.

5. Scaling laws, comparative regimes, and tradeoffs

The foundational analytical comparison between active and passive STAR-RIS appears in the two-user downlink analysis of (Xu et al., 2023). There, active STAR-RIS with independent phase shifts gives full diversity $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$3 for both users, whereas the coupled active design gives diversity $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$4 for user A and only $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$5 for user B. The same paper proves that average received SNR scales with $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$6 for active STAR-RIS and with $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$7 for passive STAR-RIS, so passive STAR-RIS has the better asymptotic scaling law, while active STAR-RIS can still yield lower outage probability when the number of elements is small and the practical amplifier gain dominates the passive asymptotics (Xu et al., 2023).

Later comparative studies sharpened this regime dependence. For SWIPT, (Zhu et al., 2024) shows that, given the same aperture sizes, active STAR-RIS exhibits superior performance over passive STAR-RIS when the aperture size is small, but the performance gap decreases with increasing aperture size; given identical power budgets, passive STAR-RIS is generally preferred because the active architecture must balance hardware power $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$8 and amplification power $$\mathbf{B}_{m,l}=\mathrm{diag}(\boldsymbol{\beta}_{m,l})$$9. In hardware-impairment-aware resource-efficiency optimization, (Huang et al., 8 Sep 2025) reports that the achievable RE of active STAR-RIS first increases with the number of elements $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$0 and then decreases when $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$1 becomes too large, because array gain is eventually outweighed by active-surface power consumption, amplified noise, and hardware impairments.

Within specific applications, active STAR-RIS often outperforms passive or single-surface baselines under the stated constraints. The multiple-active-STAR-RIS secure ISAC study reports substantial sum-rate gains over passive-RIS and single STAR-RIS baselines while rigorously meeting sensing and security constraints; adding more cooperative STAR-RISs increases path diversity and robustness, and three active STAR-RISs can substantially outperform one active STAR-RIS (Noh et al., 24 Jul 2025). In massive MIMO, the ASTARS paper reports the superiority of 16 ASTARS compared to passive STAR-RIS for a practical number of surface elements, while also noting that too many active elements can hurt because active amplification consumes more power and amplifies thermal noise (Papazafeiropoulos et al., 2024).

Energy-centric evaluations are similarly conditional. The SWIPT design in (Faramarzi et al., 2024) reports about $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$2 average energy-efficiency improvement over the passive STAR-RIS counterpart, and the meta-learning-enhanced MMDS improves EE by about $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$3 over the original MDS. The MEC study reports that the proposed active STAR-RIS-assisted system outperforms the conventional STAR-RIS-assisted system by $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$4 and the conventional RIS-assisted system by $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$5, respectively (Aung et al., 2024). A plausible implication is that active STAR-RIS is most attractive in NLoS, multi-hop, or small-aperture regimes where passive array gain alone is insufficient, whereas large-aperture or strict power-budget regimes may favor passive STAR-RIS.

6. Prototyping, non-idealities, and implementation constraints

The subject moved from abstract coefficient models to hardware prototyping with the filtering active STAR-RIS of (Song et al., 19 Jan 2025). That design implements independent transmission and reflection phase control, adjustable power splitting, active amplification, and out-of-band harmonic suppression in a multilayer unit cell operating around $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$6 GHz. The prototype uses an $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$7 array assembled from two $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$8 subarrays, 1-bit phase selection via MXD8625 SPDT switches, a tunable power divider based on $$\boldsymbol{\Phi}_{m,l}=\mathrm{diag}(\boldsymbol{\phi}_{m,l})$$9 coupled lines and BB857 varactors, and a Mini-Circuits GALI-S66+ power amplifier. The reported results include equal power division at $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$0 pF in simulation and $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$1 V bias in measurement, about $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$2 dB measured tuning range in port amplitudes, phase error within $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$3, transmission zeros at $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$4 GHz and $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$5 GHz, gain improvement of more than $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$6 dB over the passive response in simulation, and up to $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$7 dB active gain in the conclusion. The same work extends an RCS-based path-loss model to the STAR-RIS and validates it experimentally with agreement within $$\beta_{\mathrm{T},l,i}^2+\beta_{\mathrm{R},l,i}^2=1$$8 dB in transmission mode (Song et al., 19 Jan 2025).

These hardware benefits are inseparable from non-idealities. Active STAR-RIS models consistently include thermal noise generated at the surface, explicit amplification-power constraints, and circuit-power terms; some formulations add per-element linear-range limits, amplifier bias power, and transceiver hardware impairments at the BS and user side (Noh et al., 24 Jul 2025, Zhu et al., 2024, Huang et al., 8 Sep 2025). The sensing-capable active STAR-RIS of (Li et al., 7 Jan 2025) introduces another implementation direction: rather than making every element fully active, it uses a small number of RF chains to enable hybrid active preamble reception for channel acquisition and dynamic-target tracking, then switches to full passive communication mode in the data phase. This suggests a spectrum of realizations between fully amplifying surfaces and partially active, measurement-capable STAR-RIS architectures.

Practical limitations remain explicit in the available prototypes and models. The prototype in (Song et al., 19 Jan 2025) uses only 1-bit phase control, requires power amplifiers, switches, varactors, drivers, and multiple control lines, and shows beamforming degradation at large steering angles. Algorithmically, many designs still rely on AO, SDR, SCA, or penalty methods, which provide local optima under perfect or near-perfect CSI assumptions. The literature therefore presents active STAR-RIS not as a single settled architecture but as a family of amplifying, full-space, power-constrained metasurfaces whose value depends on how amplification gain, noise injection, aperture size, and hardware overhead are balanced in a given deployment.