Fluid Antennas: Dynamic Reconfigurable Arrays

• Fluid antennas are reconfigurable systems employing fluidic, conductive, or dielectric elements to dynamically alter shape and optimize radiation characteristics.
• They leverage diverse physical realizations—from liquid metal fibers to electronically adjustable arrays—with proven benefits in diversity gain and interference suppression.
• Research focuses on overcoming hardware and channel uncertainties, with experimental results demonstrating improved SNR, outage performance, and capacity in 5G/6G applications.

Searching arXiv for recent and foundational papers on fluid antennas to ground the article in cited sources. Fluid antennas (FAs), or fluid antenna systems (FASs), denote software-controlled fluidic, conductive, or dielectric structures that dynamically alter an antenna’s shape and position to modify gain, radiation pattern, operating frequency, and related radiation characteristics. In the canonical communications abstraction, a single RF chain accesses many candidate ports inside a compact aperture and selects or combines them to exploit spatial diversity, thereby converting local spatial mobility into a controllable degree of freedom (Wu et al., 2024, Wong et al., 2020). Later surveys extend this abstraction to realistic deployments affected by channel uncertainty, hardware nonidealities, and mechanical constraints (Pakravan et al., 30 Jan 2026).

1. Physical concept and realizations

FA is used in the literature both as a physical antenna concept and, at system level, as a reconfigurable transceiver front-end. A fluid antenna is sometimes called a movable antenna and is characterized by software-controllable, flexible positioning, and shape-changing capabilities. This broader definition is intentionally not restricted to literal liquids: the category includes fluidic, conductive, dielectric, and electronically reconfigurable structures (Wu et al., 2024).

Physical realizations reported in the literature include liquid metal fibers, conductive fluid in tubes with nano-pump control, metallophobic surface patterning, stacked three-dimensional liquid metal formed by direct-write printing, water antennas with adjustable column height, stretchable clothes with liquid-metal radiators, and pixel-based reconfigurable antennas. The classification in the 6G-oriented overview distinguishes materials such as liquid metal, non-metallic liquids, and metallic pixels; shapes such as filament, planar, and three-dimensional structures; and control mechanisms such as liquid flow, pattern-controlled liquid, amount-controlled liquid, and electronic switching (Wu et al., 2024). This suggests that “fluidity” is best understood as reconfigurable electromagnetic embodiment rather than a single fabrication route.

The most concrete hardware demonstration in the supplied corpus is the mmWave surface-wave-enabled FAS prototype. That work designs a single-channel fluid antenna (SCFA) and a double-channel fluid antenna (DCFA) using Galinstan radiators inside one or two straight cylindrical channels of diameter $1.2$ mm and length $17.5$ mm, with a fluid element length of $6.5$ mm and an effective travel of about $11$ mm, roughly one free-space wavelength at $26$ GHz. The overall size is $33 \times 10 \times 2.8\ \text{mm}^3$, and the measured operating bandwidth is approximately $23.2$ GHz to $38.5$ GHz, covering the $24$–$30$ GHz 5G mmWave bands (Shen et al., 2024). In that work, the measured radiation patterns are imported into channel and network simulations, where the prototypes can vary their gain up to an averaged value of $17.5$0 dBi; in 4-user FAMA, the double-channel FAS reduces outage probability by $17.5$1 and increases the multiplexing gain to $17.5$2 compared to a static omnidirectional antenna (Shen et al., 2024).

2. Canonical mathematical models

The foundational discrete-port model places $17.5$3 ports along a line segment of length $17.5$4, where $17.5$5 is the wavelength. At any instant, only one port is active, and the effective channel is selected by maximum instantaneous channel magnitude, exactly as in selection combining over spatial positions rather than over fixed antenna branches (Wong et al., 2020). In the UAV-to-ground formulation, the receiver-side FAS has $17.5$6 closely spaced ports along a line of length $17.5$7, a single RF chain, and port-$17.5$8 SNR

$17.5$9

with output SNR

$6.5$0

Because the ports are packed into a small aperture, the channels are correlated, with Jakes-type correlation matrix

$6.5$1

and eigenvalue decomposition $6.5$2 (Zhu et al., 21 Nov 2025).

A central refinement is the distinction between raw port count and effective spatial rank. In the double-shadowing UAV model, $6.5$3 is the number of non-zero eigenvalues, and the maximum of $6.5$4 correlated ports is approximated by the maximum of $6.5$5 independent but non-identically distributed eigen-modes,

$6.5$6

where each $6.5$7 follows the single-link double-shadowing law (Zhu et al., 21 Nov 2025). This model is particularly useful because it turns a correlated-port problem into an analytically tractable eigen-mode problem.

At system level, the 6G principles paper separates two major channel abstractions. A field response-based channel is used for finite-scattering settings, especially at higher frequencies and in near-field scenarios, while a correlation-based channel is used for rich scattering with Gaussian vectors and a spatial covariance matrix (Wu et al., 2024). The near-field NF-ISCSC framework exemplifies the first class: the transmit steering vector depends explicitly on FA positions through both linear and quadratic spherical-wave phase terms, and the path-loss vector is also FA-position dependent through $6.5$8 (Yang et al., 21 Jul 2025).

A second major branch replaces discrete ports by a continuous position variable. In the continuous fluid antenna system (CFAS), the liquid position $6.5$9 is continuous over $11$0, and the received signal and corresponding SIR become a continuous random process $11$1. With one desired Rayleigh signal and $11$2 Rayleigh interferers in an interference-limited regime,

$11$3

and the FA seeks the supremum

$11$4

Spatial correlation follows either Jakes’ model $11$5 or, more generally, $11$6 (Psomas et al., 2023).

3. Fundamental performance laws

The first generation of FAS theory established exact and approximate outage laws for arbitrarily correlated Rayleigh fading. The outage probability of a single-antenna FAS can be written in a single-integral form involving the Marcum $11$7-function, and the accompanying upper bound shows that a single-antenna FAS given any arbitrarily small space can outperform an $11$8-antenna maximum ratio combining system if $11$9 is large enough (Wong et al., 2020). That result is notable because it reframes diversity as a problem of spatial sampling density and correlation, rather than exclusively of multiple simultaneous RF branches.

Capacity and second-order fading statistics were derived soon afterward. For a single-RF-chain FAS of length $26$0 with $26$1 ports, the ergodic capacity is expressed by an exact double integral involving Marcum-$26$2 terms, and a lower bound reveals the scaling with $26$3, SNR, and correlation. The same letter derives the level crossing rate

$26$4

and the average fade duration through the outage-to-LCR ratio, showing that the FAS shifts the effective envelope distribution upward while preserving the Rayleigh-form LCR at a fixed level (Wong et al., 2020).

Under Nakagami-$26$5 fading with selection combining, an ideal FA achieves full diversity $26$6. However, when the selected port is used after a post-scheduling delay, outdated CSI reduces the diversity order to $26$7. The same work proposes a linear prediction scheme based on temporal correlation $26$8, and simulation results show that prediction can restore the lost diversity, while space-time rotations over the sequential FA operation achieve maximum coded diversity with low-complexity iterative detection (Psomas et al., 2022).

More recent high-SNR analysis sharpens the role of correlation geometry. In the geometric SER framework, the asymptotic law is

$26$9

with diversity order

$33 \times 10 \times 2.8\ \text{mm}^3$0

where $33 \times 10 \times 2.8\ \text{mm}^3$1 is the effective rank inferred from the eigenvalue spectrum of the spatial correlation matrix rather than the nominal port count. In the same model, the effective rank converges to a fundamental aperture-limited value

$33 \times 10 \times 2.8\ \text{mm}^3$2

so increasing port density within a fixed aperture yields diminishing returns once the aperture-supported spatial modes are resolved (Zhu et al., 10 Sep 2025). This directly rebuts the common misconception that diversity scales linearly with $33 \times 10 \times 2.8\ \text{mm}^3$3 irrespective of aperture.

The UAV-to-ground double-shadowing analysis extends the diversity result beyond Rayleigh fading. For a single UAV transmitter and an $33 \times 10 \times 2.8\ \text{mm}^3$4-port FAS receiver over a double-shadowing channel $33 \times 10 \times 2.8\ \text{mm}^3$5, the asymptotic outage slope is

$33 \times 10 \times 2.8\ \text{mm}^3$6

where $33 \times 10 \times 2.8\ \text{mm}^3$7 is the FAS spatial rank and $33 \times 10 \times 2.8\ \text{mm}^3$8 is the intrinsic diversity order of the single link (Zhu et al., 21 Nov 2025). This is consistent with the effective-rank interpretation: in correlated fading, the FA contribution is the spatial-rank factor, not the raw number of virtual ports.

4. Continuous apertures, wideband signaling, and modulation design

The continuous-port formulation changes the mathematics from order statistics over discrete ports to extrema of a correlated random process. In CFAS, closed-form expressions for the level crossing rate and average fade duration of the continuous SIR process lead to a lower bound on the CDF of the supremum $33 \times 10 \times 2.8\ \text{mm}^3$9. The lower bound

$23.2$0

is tight in the upper tail, and numerical comparisons show that CFAS outperforms its discrete counterpart, thereby providing a performance limit or benchmark for discrete-port FA systems (Psomas et al., 2023). A plausible implication is that discrete FAS hardware can be interpreted as a spatial quantization of a continuous optimum.

Wideband modeling becomes essential once the per-port channel varies across subcarriers. In the 5G NR OFDM framework, the FAS receiver is a two-dimensional grid of $23.2$1 ports over $23.2$2, and a single port is selected for an entire subframe. The paper introduces a wideband FAS OFDM framework with a novel port selection matrix and maps the time-frequency SNR grid to a scalar selection metric via effective-SNR-style averaging; with $23.2$3, the average-SNR metric is

$23.2$4

It further derives the BICM capacity and an AMC scheme using effective SNR mapping to 3GPP CQI tables (Hong et al., 7 Mar 2025). Link-level simulations show that, for $23.2$5 and $23.2$6, the FAS can gain about $23.2$7 dB over an FPA at BLER $23.2$8; at SNR $23.2$9 dB and BLER target $38.5$0, the reported throughputs are $38.5$1 Mbit/s at $38.5$2 MHz and $38.5$3 Mbit/s at $38.5$4 MHz, compared with $38.5$5 and $38.5$6 Mbit/s for the FPA baseline (Hong et al., 7 Mar 2025).

Fluid antennas have also been merged with spatial modulation. In FA-empowered receive spatial modulation, the transmitter-side FA occupies a two-dimensional region $38.5$7 and simultaneously activates $38.5$8 ports selected from $38.5$9 candidates. Port selection is posed as a capacity maximization problem, then approximated by a trace-minimization decremental algorithm (TMD) and an MCE-TMD variant that first removes highly correlated ports. For detection, the paper proposes a two-stage maximum-energy detector and a ratio-threshold test detector, both near-optimal at much lower complexity than full MLD (Guo et al., 9 Jun 2025). That work also reports the same saturation phenomenon: activating additional ports improves performance, but the gain gradually saturates due to inherent spatial correlation.

5. System integrations and application domains

FAs have been embedded in a broad range of cross-layer and cross-domain systems. In wireless powered communications, the transmitter harvests RF energy from a power beacon through one FA port and then transmits data to the receiver through the same selected port. The end-to-end SNR is

$24$0

and the proposed joint port selection rule maximizes $24$1 across ports (Lai et al., 2024). The asymptotic outage law behaves as $24$2, which indicates diversity approximately equal to the number of ports $24$3, whereas single-hop-based selection yields diversity order only one (Lai et al., 2024).

In secure and covert communications, the transmitter adjusts multiple fluid antennas’ positions to maximize secrecy rate while satisfying a covertness constraint against a warden. The resulting optimization over beamforming covariance $24$4 and continuous antenna positions $24$5 is non-convex, so the paper uses alternating optimization, a penalty-based treatment of the rank-one constraint, and majorization-minimization for the antenna positions (Yao et al., 2024). This suggests that FA positioning can act as a physical-layer secrecy variable analogous to beamforming, but with a different geometric control surface because it reshapes the array manifold itself.

The 6G principles paper places FAs in SWIPT, ISAC, NOMA, RIS integration, physical-layer security, and mobile edge computing. In its SWIPT case study, joint beamforming and FA location optimization achieves up to $24$6 higher rate than the fixed-position baseline. In the FAS-RIS case study, outage probability decreases sharply as the number of user-side FA ports increases, and the paper argues that RIS can enrich multipath while the FAS “surfs” the local spatial field (Wu et al., 2024).

More recent ISAC-oriented work pushes the concept into massive access and semantic systems. In unsourced integrated sensing and communication, a transmitter-side FAS is used at the user equipment to improve compressed-sensing-based access, mitigate pilot collision floors under finite blocklength, and enhance angle-of-arrival sensing; the reported numerical gain is a $24$7 dB capacity gain over traditional TDMA at $24$8 active users (Xu et al., 29 Jun 2026). In near-field integrated sensing, computing, and semantic communication, the BS employs a planar FA transmit array whose positions are optimized jointly with beamforming, computing load, and semantic extraction ratio; the overall problem is solved by alternating optimization, with successive convex approximation for beamforming, projected BFGS for FA positioning, and bisection for the semantic extraction ratio, and the reported effect is higher data rates together with better privacy preservation (Yang et al., 21 Jul 2025).

A distinct application is UAV-to-ground communications over double-shadowing fading. There the receiver-side FAS enables analytical outage, ABER, and capacity expressions, and the eigenvalue-based approximation remains highly accurate except in highly rank-deficient cases, where a mild optimistic bias appears (Zhu et al., 21 Nov 2025). The larger significance is that FAs are now being studied not only in classical terrestrial Rayleigh links but also in shadowed aerial channels where both multipath fading and large-scale attenuation are severe.

6. Experimental status, practical nonidealities, and open questions

A recurring feature of the analytical literature is idealized switching and CSI assumptions. Foundational discrete-port analyses assume that the receiver can observe channel gains at all ports and instantly switch to the strongest one (Wong et al., 2020). The continuous CFAS analysis assumes instantaneous displacement and full channel knowledge, and the UAV double-shadowing study assumes perfect instantaneous SNR-based port selection with ideal port switching, no switching delay or overhead, in a single-link, noise-limited setting (Psomas et al., 2023, Zhu et al., 21 Nov 2025). These assumptions are mathematically productive, but they delimit the interpretation of published gains.

The practical-survey perspective is therefore essential. The survey on FAS under channel uncertainty and hardware impairments identifies three coupled nonidealities: spatio-temporal channel uncertainty, hardware and mechanical impairments such as RF nonlinearity, port coupling, and fluid response delay, and the need for robust design and learning-based control strategies (Pakravan et al., 30 Jan 2026). It advocates stochastic, bounded, and hybrid uncertainty models; hardware-aware selection and beamforming; predictive control under outdated CSI; and online learning or DRL for port selection and joint control. A plausible implication is that the next stage of FA research will be less about proving ideal diversity gains and more about preserving them under nonideal actuation, estimation, and RF front-end behavior.

Several broad research directions recur across surveys. One is channel estimation: high-spatial-fidelity FASs have many ports or states, so naive per-port training is costly, and exploiting spatial correlation or AI-assisted inference becomes necessary (Wu et al., 2024). Another is versatile channel modeling that unifies finite-scattering field-response formulations with rich-scattering correlation-based formulations, while also accounting for shape, polarization, and hardware-specific effects (Wu et al., 2024). A third is robust beamforming and FA location optimization under imperfect CSI, since near-field and geometric models are especially sensitive to angle and distance errors (Wu et al., 2024).

Three misconceptions are explicitly contradicted by the current literature. First, FAs are not necessarily liquid-metal tubes: the system concept includes pixel-based and other software-controlled conductive or dielectric structures (Wu et al., 2024). Second, performance does not scale with raw port count alone: both the geometric SER analysis and the UAV rank analysis show that effective rank or spatial rank governs the asymptotic gain, and fixed-aperture densification eventually saturates (Zhu et al., 10 Sep 2025, Zhu et al., 21 Nov 2025). Third, continuous movement is not equivalent to cost-free adaptation: the strongest available analyses often rely on ideal motion and ideal CSI, while practical surveys place fluid response delay and channel uncertainty at the center of robust FAS design (Psomas et al., 2023, Pakravan et al., 30 Jan 2026).

Taken together, the literature describes FAs as a reconfigurable spatial-sampling architecture that can emulate some of the diversity and interference-management benefits of larger arrays while using fewer RF chains. The most stable conclusions are that aperture-supported effective rank is the key physical quantity, that continuous or densely discretized spatial mobility can materially improve outage, BER, rate, and secrecy, and that translating those gains into deployable systems now depends on hardware-aware modeling, channel acquisition, and cross-domain control.