Smart Radio Environment Overview

• Smart Radio Environment is a paradigm where wireless channels are engineered into controllable, programmable mediums using reconfigurable intelligent surfaces and feedback control.
• The approach leverages advanced electromagnetic designs and AI-driven algorithms to dynamically shape radio waves for enhanced coverage, capacity, and security.
• Research in SRE integrates physics-based models with machine learning techniques to optimize beamforming, localization, and energy transfer in complex propagation environments.

Searching arXiv for the core paper and closely related Smart Radio Environment work to ground the article in recent and foundational sources. A Smart Radio Environment (SRE) is a wireless propagation medium in which the channel itself becomes a controllable optimization variable and the environment is upgraded from a passive, uncontrollable channel into an active, digitally programmable participant in wireless communication and sensing. In this paradigm, walls, ceilings, facades, cavities, and other scatterers are equipped with reconfigurable electromagnetic structures so that radio waves can be steered, focused, absorbed, refracted, diffuse-reflected, or otherwise shaped to improve coverage, rate, reliability, localization, security, sensing, or energy transfer. The concept emerged through the convergence of programmable metasurfaces, wavefront shaping in complex media, and AI-based control, and later expanded to include static electromagnetic skins, fixed intelligent surfaces, active repeaters, and frequency-domain techniques such as movable signals (Frazier et al., 2021, Gradoni et al., 2021, Renzo et al., 2019, Maleki et al., 10 Aug 2025, Nerini et al., 12 Nov 2025).

1. Definitions, scope, and architectural model

In the SRE literature, the defining shift is from “adapting to wireless channels” to “changing wireless channels.” Reconfigurable Intelligent Surfaces (RISs), also called metasurfaces or intelligent reflecting surfaces, are the canonical enabling devices: planar arrays of sub-wavelength unit cells whose local reflection or transmission properties can be tuned on demand. These unit cells are described as metallic patches, dielectric resonators, graphene ribbons, and related metamaterial elements loaded with varactors, PIN diodes, phase-change materials, or micro-electromechanical switches, under the control of hardware such as FPGAs, microcontrollers, or ASICs (Gradoni et al., 2021, Liu et al., 2021).

A commonly used conceptual architecture consists of five elements: a transmitter producing a modulated signal, a complex scattering medium such as an indoor or outdoor environment or a metallic cavity, a distributed or localized reconfigurable metasurface embedded in that medium, a sensing or feedback path based on software-defined radio, and an intelligent controller implemented on a CPU, GPU, or edge-AI processor. The control loop is described in operational terms as measure, compute, command metasurface, and repeat (Frazier et al., 2021).

Although RIS has been the dominant embodiment, SRE is not restricted to online phase-programmable reflectors. The literature also includes static electromagnetic skins whose per-element phase shifts are preset at installation, fixed intelligent surfaces used jointly with movable signals, and heterogeneous deployments containing RIS, Simultaneous Transmitting and Reflecting RIS (STAR RIS), Network-Controlled Repeaters (NCR), and tri-sectoral NCR (3SNCR). This broader scope is important because it separates the SRE idea—engineering the medium itself—from any single hardware implementation (Maleki et al., 10 Aug 2025, Nerini et al., 12 Nov 2025, Ayoubi et al., 2024).

2. Electromagnetic foundations and channel models

At the electromagnetic level, an RIS unit cell is often abstracted by a local reflection coefficient

$\Gamma(x,y)=|\Gamma(x,y)|\,e^{j\Phi(x,y)},$

or, more rigorously, by a surface impedance

$Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$

In the Generalized Sheet Transition Conditions (GSTCs) formulation, surface susceptibilities determine the jump and average of the tangential fields across the sheet, providing a homogenized model for programmable wave transformations (Gradoni et al., 2021).

At the communication-theoretic level, a standard narrowband model for an RIS-assisted MIMO link is

$$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$

where $\mathbf{H}_d$ is the direct channel, $\mathbf{F}$ is the BS–RIS channel, $\mathbf{G}$ is the RIS–UE channel, and $$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$ is the RIS phase-shift matrix. In the scalar single-antenna case, the received signal is written as

$y=(h_d+h_r\,\Theta\,g)x+n.$

The SRE perspective further recasts the channel as state-dependent: $y \sim p(y\mid x,s),$ where the RIS or environmental state $s$ actively changes the transition law itself (Renzo et al., 2020, Liu et al., 2021).

Recent work in rich-scattering environments argues that the usual cascaded model is physically incomplete because it truncates multi-bounce effects. A physics-compliant model instead starts from a global scattering matrix $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$0, or from a dipole interaction matrix $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$1, with self-consistent solution

$Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$2

and channel

$Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$3

In this view, RIS tuning modifies the diagonal entries associated with tunable polarizabilities or loads, while the full matrix inversion retains arbitrarily long multi-bounce contributions. The corresponding critique is explicit: a model of the form

$Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$4

amounts to truncating multi-bounce paths at first order (Hougne, 2023).

A parallel general framework models any linear electromagnetic environment with boundary conditions as a space-variant linear feedback filter. After modal discretization, the environment becomes a graph of electromagnetic objects linked by coupling matrices $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$5 and local boundary operators $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$6, with block equations

$Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$7

This places SRE on a system-theoretic footing that is consistent with Maxwell boundary conditions rather than only with simplified phase-only abstractions (Dardari, 2023).

3. Control, optimization, and learning

The control problem in an SRE is frequently posed as an inverse problem: given a multipath-scrambled channel, choose the metasurface state vector $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$8 to maximize a metric $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$9. The objectives explicitly mentioned in the literature include beam focusing at a target location, creation of on-demand cold spots or nulls, and channel capacity maximization under interference. The practical constraints are quantized phase states, finite numbers of unit cells, and temporal coherence of the environment. Classical approaches include brute-force trial and error and stochastic search, while later work points to gradient-based algorithms and convex relaxation (Frazier et al., 2021).

The main algorithmic motivation for AI is that stochastic tuning becomes too slow in nonstationary environments. Accordingly, the SRE literature advocates deep learning and reinforcement learning, including transfer learning for on-the-fly adaptation and model compression through quantization and pruning to reduce size, weight, and power (Frazier et al., 2021). One model-free IRS controller formulates the problem as a Markov decision process and uses Double Deep Q-Network (DDQN) for coarse phase control plus extremum seeking control (ESC) for fine phase adjustment. In that framework, DRL converges in approximately $$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$0–$$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$1 episodes and outperforms random and multi-armed bandit baselines by $$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$2–$$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$3; DRL versus random reflection provides a $$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$4–$$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$5 sum-rate gain under high Rician factor; and the joint DRL + ESC scheme yields an additional $$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$6–$$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$7 gain in slow-fading blocks when $$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$8 is large (Wang et al., 2022).

For multi-user, multi-RIS systems, the same optimization logic extends to joint selection of RIS configurations and BS beams. A surveyed formulation maximizes the sum rate

$$\mathbf{H}_{\rm tot}=\mathbf{H}_d+\mathbf{G}\,\mathbf{\Theta}\,\mathbf{F},$$9

under discrete reflection phases, precoder codebooks, total power, and per-user rate requirements. Within that setting, DQN and Neural $\mathbf{H}_d$0-greedy achieve $\mathbf{H}_d$1–$\mathbf{H}_d$2 of Optimal, UCB achieves $\mathbf{H}_d$3–$\mathbf{H}_d$4 of Optimal, and Random achieves $\mathbf{H}_d$5–$\mathbf{H}_d$6. The execution-speed contrast is equally marked: UCB is reported at approximately $\mathbf{H}_d$7–$\mathbf{H}_d$8 steps/s, DQN at approximately $\mathbf{H}_d$9–$\mathbf{F}$0 steps/s, and Neural $\mathbf{F}$1-greedy at approximately $\mathbf{F}$2–$\mathbf{F}$3 steps/s (Alexandropoulos et al., 2022).

A recurring implication is that SRE control is not solely a beamforming problem. It is also a systems problem involving sample efficiency, coherence time, controller latency, and the relationship between channel measurement overhead and environmental dynamics. The literature therefore treats the control plane itself as a core part of SRE design rather than an external afterthought.

4. Physical realizations and expanding hardware classes

The most established SRE realization is the microwave or millimeter-wave RIS built from PCB-printed copper patches on dielectric substrates with varactor diodes or PIN diodes. Control can be centralized, with the BS or an edge server computing the phase profile, or distributed and autonomous, with pilot-based sensing and local intelligence. Space-time-coding digital metasurfaces extend this idea by modulating unit cells in time to produce harmonic beams and on-surface signal processing (Gradoni et al., 2021).

A distinct hardware line addresses multi-band operation. FABRIS introduces a frequency-agnostic RIS whose unit cell is a two-patch microstrip antenna with effective input impedance $\mathbf{F}$4, reflection coefficient

$\mathbf{F}$5

and discrete phase control realized through microstrip-line length

$\mathbf{F}$6

Its optimization target is the signal-to-leakage-and-noise ratio

$\mathbf{F}$7

Full-wave results report $\mathbf{F}$8 and $\mathbf{F}$9; Monte Carlo trials show a $\mathbf{G}$0–$\mathbf{G}$1 dB median SLNR gain over a naive all-on RIS when $\mathbf{G}$2 m and $\mathbf{G}$3 dB at $\mathbf{G}$4 m; and the optimized design preserves a $\mathbf{G}$5 beamwidth at $\mathbf{G}$6 with $\mathbf{G}$7 dBi main lobe, versus $\mathbf{G}$8 and $\mathbf{G}$9 dBi for the naive case (Maresca et al., 2022).

Another branch removes online control entirely. Static electromagnetic skins implement the generalized Snell’s law

$$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$0

with element phase

$$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$1

By selecting a codebook of such phase gradients offline, the installation permanently enriches channel dissimilarity without feedback or per-user adaptation (Maleki et al., 10 Aug 2025).

SRE has also been generalized into the frequency domain. In the movable-signals formulation, the signal spectrum itself is moved along the frequency axis, eliminating tunable scatterers and movable antennas. For line-of-sight MISO with fixed broadside precoder $$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$2, full co-phasing occurs at

$$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$3

Under non-line-of-sight conditions with a fixed intelligent surface and $$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$4, the optimal frequency becomes

$$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$5

and analytical results state that a FIS-aided system using movable signals can achieve up to four times the received power of a RIS-aided system using fixed-frequency signals (Nerini et al., 12 Nov 2025).

A more radical SRE realization is the surface-wave communication paradigm, in which programmable metasurfaces guide trapped surface waves rather than merely reflect free-space waves. Its path-loss law is

$$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$6

contrasted with the free-space $$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$7 law, and its proposed hardware includes a software-controlled fluidic waveguiding architecture using microtubes and liquid metal alloy to create high-impedance walls and low-loss channels for surface-wave routing (Wong et al., 2020).

5. Applications and empirical evidence

The application envelope of SRE extends beyond conventional downlink beamforming. Vision and roadmap papers explicitly highlight wireless power transfer through beam focusing, indoor localization through tailored illumination patterns, secure transmission through null creation toward eavesdroppers, imaging and computational microwave imaging, cloaking and cold spots, and backscatter communications using ambient Wi-Fi (Frazier et al., 2021, Gradoni et al., 2021).

Quantitative evidence is now available in several of these domains. For channel-charting-based localization in an $$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$8 m$$\mathbf{\Theta}=\mathrm{diag}(e^{j\phi_1},\dots,e^{j\phi_N})$$9 urban area with two $y=(h_d+h_r\,\Theta\,g)x+n.$0-element EMS panels and $y=(h_d+h_r\,\Theta\,g)x+n.$1 test points, optimized static EMS reduces the $y=(h_d+h_r\,\Theta\,g)x+n.$2th-percentile localization error from over $y=(h_d+h_r\,\Theta\,g)x+n.$3 m to less than $y=(h_d+h_r\,\Theta\,g)x+n.$4 m, lowers the median localization error by approximately $y=(h_d+h_r\,\Theta\,g)x+n.$5, and increases $y=(h_d+h_r\,\Theta\,g)x+n.$6th-percentile trustworthiness and continuity by more than $y=(h_d+h_r\,\Theta\,g)x+n.$7–$y=(h_d+h_r\,\Theta\,g)x+n.$8 percentage points compared to the no-EMS baseline (Maleki et al., 10 Aug 2025).

System-level simulation has also clarified placement and scaling rules. At $y=(h_d+h_r\,\Theta\,g)x+n.$9 GHz in the enhanced SimRIS environment, a RIS-free link peaks at approximately $y \sim p(y\mid x,s),$0 b/s/Hz at $y \sim p(y\mid x,s),$1 dBm; a single RIS with $y \sim p(y\mid x,s),$2 reaches approximately $y \sim p(y\mid x,s),$3 b/s/Hz; and a single RIS with $y \sim p(y\mid x,s),$4 reaches approximately $y \sim p(y\mid x,s),$5 b/s/Hz. Raising the RIS from $y \sim p(y\mid x,s),$6 m to $y \sim p(y\mid x,s),$7 m increases ergodic rate by approximately $y \sim p(y\mid x,s),$8–$y \sim p(y\mid x,s),$9 b/s/Hz, while two RISs with $s$0 provide about $s$1 b/s/Hz more than the single-RIS case. The same study reports that deploying $s$2 RISs of $s$3 each matches the performance of one RIS with $s$4, highlighting a concrete trade-off between a few large surfaces and many small surfaces (Toumi et al., 2021).

Physics-compliant modeling in rich-scattering indoor and factory environments yields a different type of evidence: calibration with at most $s$5 examples attains channel-prediction accuracies of $s$6–$s$7 dB, compared with less than or equal to $s$8 dB for linear cascaded models and less than or equal to $s$9 dB for generic neural surrogates. The same line reports focusing gains of approximately $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$00–$Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$01 dB without phase measurements and open-loop SNR or capacity gains of $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$02–$Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$03 dB (Hougne, 2023).

Industrial and field-trial evidence, although heterogeneous, points in the same direction. Reported demonstrations include a $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$04 GHz trial with downlink $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$05 Mbps versus $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$06 Mbps without RIS, a $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$07-element two-bit RIS with $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$08 dBi gain at $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$09 GHz and $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$10 dBi at $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$11 GHz, and an installation with $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$12 elements showing $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$13 dB indoor and $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$14 dB short-range outdoor power gain, plus $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$15 dB gain at $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$16 m with $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$17 W consumed (Liu et al., 2021).

6. Deployment, planning, standardization, and security

The transition from laboratory SREs to network infrastructure has been framed as both a planning problem and a standardization problem. Industrial analyses identify electromagnetic-consistent channel modeling, efficient RIS channel estimation and feedback protocols, control-signaling frameworks, multi-band operation, manufacturability, weather resistance, hardware non-idealities, and OSS/BSS integration as central commercialization barriers. Standardization activity has already been recorded at CCSA, ETSI ISG-RIS, ITU-R WP 5D, and 3GPP, with RIS support proposed for 5G Advanced and discussed in Release 18 and Release 19 time frames (Liu et al., 2021).

Network-level planning has moved toward heterogeneous SREs. In a Milan case study at $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$18 GHz, placement is formulated as two Mixed-Integer Linear Programs: Full Coverage Minimum Cost (FCMC) and Maximum Budget-Constrained Coverage (MBCC). The device set includes RIS, STAR RIS, NCR, and 3SNCR, with relative costs normalized so that a $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$19 RIS costs $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$20 unit, a STAR RIS costs $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$21, and NCR cost is $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$22 with $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$23 and $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$24. The reported outcome is that heterogeneous SRE yields up to $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$25 cost savings, with an optimum RIS size around $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$26 or $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$27 elements depending on the device set and an optimum NCR gain around $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$28 dB or $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$29 dB depending on the SNR threshold (Ayoubi et al., 2024).

A persistent misconception is that SRE automatically strengthens security because it gives defenders control over propagation. Security-focused experiments show that the same controllability is symmetric: both attackers and defenders can exploit metasurfaces. In Wi-Fi case studies with two identical $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$30 binary-phase metasurfaces, a defensive surface can force an attacker to increase jamming by approximately $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$31 dB; but in a secure-communication setting, an opposing surface can reduce an eavesdropper’s Symbol Error Rate from approximately $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$32 to $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$33. In sensing obfuscation, protection can drop detection from approximately $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$34 to approximately $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$35, yet an opposing surface can recover detection to approximately $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$36, and in poorly oriented deployments even to approximately $Z_s(x,y)=\frac{E_t(x,y)}{H_t(x,y)}.$37. These results support the paper’s claim that metasurface “battles” depend on timing, placement, algorithmic strategy, and hardware scale rather than on metasurface ownership alone (Staat et al., 17 Nov 2025).

The main open problems therefore remain interdisciplinary. Vision papers consistently identify real-time adaptation under short coherence times, scalability to large metasurface arrays, robustness to mobility and new scatterers, SWaP limitations for embedded AI processors, and the data-hungry nature of deep learning as core obstacles (Frazier et al., 2021). A plausible implication is that the mature form of SRE will not be a single device class, but a layered control fabric that combines passive and active electromagnetic elements, model-based and data-driven control, and network planning with physical-layer security analysis.