Fluid Antenna: Dynamic Reconfigurability

• Fluid antenna (FA) is a reconfigurable radiating structure that dynamically alters its shape or position to optimize channel performance and electromagnetic characteristics.
• FA research employs both discrete-port and continuous-position selection methods with rigorous stochastic modeling to enhance diversity, capacity, and link quality.
• Practical applications of FA include outage reduction, integrated sensing and communications, DOA estimation, and over-the-air federated learning in advanced wireless systems.

Fluid antenna (FA) denotes an antenna paradigm in which the radiating structure is position-reconfigurable, shape-reconfigurable, or both, so that the antenna can be dynamically moved or morphed within a confined region to alter the instantaneous channel, radiation pattern, gain, operating frequency, and related electromagnetic characteristics. In the canonical communications model, a single physical antenna element is dynamically positioned among a set of candidate spatial locations (“ports”) and activates the location with the most favorable channel realization; in broader formulations, FA is realized as a software-controlled fluidic, conductive, or dielectric structure whose geometry and location can be reconfigured in real time (Zhu et al., 10 Sep 2025, Wu et al., 2024). The associated fluid antenna system (FAS) has been studied as a compact alternative to conventional fixed-position antenna (FPA) arrays, with applications spanning outage reduction, multiple access, integrated sensing and communication (ISAC), wideband OFDM, direction-of-arrival (DOA) estimation, semantic communication, and over-the-air federated learning (Wong et al., 2020, Hong et al., 16 Jun 2025).

1. Definition, operating principle, and physical realizations

The basic operating principle of FA is opportunistic spatial selection. In the discrete-port model, a single antenna can switch among $N$ preset locations along a fixed-length line space of length $W\lambda$, and the effective channel magnitude is selected as

$$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$

which makes the device behave as a selection-combining receiver without requiring multiple simultaneously active RF chains (Wong et al., 2020). In a continuous formulation, the selection rule is written as

$x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$

emphasizing that the movable radiating element may be placed at any feasible position in $\mathcal{X}$ rather than at a finite set of ports (Pakravan et al., 30 Jan 2026).

The literature uses both narrow and broad definitions. The narrower definition emphasizes a dielectric holder in which a radiating liquid moves between predefined ports that serve as the transceiver’s antennas (Psomas et al., 2022). The broader definition treats FAS as any software-controlled fluidic, conductive or dielectric structure that can dynamically alter antenna shape and position to change the gain, the radiation pattern, the operating frequency, and other critical radiation characteristics (Wu et al., 2024). This broader view encompasses liquid-based antennas, pixel-reconfigurable antennas, mechanically actuated antennas, metamaterials and flexible structures, as well as electrohydrodynamic/electrocapillary, microfluidic, electronically switched, and hybrid mechanical/actuator-based architectures (Hong et al., 16 Jun 2025, Pakravan et al., 30 Jan 2026).

Experimental and architectural work has extended the concept beyond literal liquid motion. A “meta-fluid antenna” architecture uses electronically reconfigurable meta-atoms controlled by PIN diodes, with pseudo-fluid dynamics realized by microsecond-scale electromagnetic field redistribution and a substrate-integrated waveguide feed that preserves single-RF-chain operation (Liu et al., 15 Sep 2025). In that implementation, a prototype with 120 meta-atoms arranged as $8\times 15$, controlled via FPGA at $20$ MHz, supports reconfiguration times $<15~\mu s$ at $26.5$ GHz; the reported experiment achieved real-time switching between 300 distinct patterns and yielded $>10$ dB SINR for all users (Liu et al., 15 Sep 2025). This suggests that “fluidity” in current FA research functions both as a physical mechanism and as a system-level abstraction for dense, software-defined spatial reconfigurability.

2. Channel representation and stochastic modeling

Early analytical work mostly considered rich-scattering Rayleigh fading with spatial correlation determined by Jakes’ model. For a 1D FAS with $W\lambda$0 ports, a common covariance model is

$W\lambda$1

where $W\lambda$2 is the zeroth-order Bessel function and $W\lambda$3 is the normalized length in wavelengths (Hong et al., 16 Jun 2025). The same structure appears in continuous formulations through

$W\lambda$4

which models the spatial correlation of the continuous signal-to-interference ratio (SIR) process over the holder length (Psomas et al., 2023).

Two modeling lines coexist. One uses simplified closed-form parameterizations to derive outage, ergodic capacity, level crossing rate (LCR), and average fade duration (AFD) (Wong et al., 2020, Wong et al., 2020). The other uses exact or low-rank eigen-structure of the full spatial covariance matrix to more faithfully reproduce Jakes-type correlation. In the survey formulation, the exact rich-scattering model is

$W\lambda$5

while the approximation framework of (Khammassi et al., 2022) retains only the dominant eigenvalues to reduce the multi-fold outage integral to lower-dimensional or single-integral forms (Hong et al., 16 Jun 2025, Khammassi et al., 2022). This modeling distinction is central, because several later conclusions about saturation and effective rank depend precisely on how correlation is represented.

Continuous fluid antenna systems (CFAS) replace discrete-port selection by optimization over a continuous position variable $W\lambda$6. In that setting,

$W\lambda$7

and the principal analytical object becomes the supremum SIR rather than the maximum over finitely many ports (Psomas et al., 2023). The CFAS framework derives closed-form analytical expressions for the LCR and AFD of the continuous SIR process and a bound on the cumulative distribution function of $W\lambda$8, establishing CFAS as a performance limit for discrete realizations (Psomas et al., 2023).

The channel model also changes substantially in wideband and near-field settings. In 5G NR OFDM,

$W\lambda$9

so port selection must aggregate channel quality over frequency, time, and position rather than relying on a single narrowband SNR (Hong et al., 7 Mar 2025). In near-field ISAC and semantic communication, the steering vector becomes an explicit function of the FA position vector, the distance, and the angular geometry, because the waves are spherical rather than planar (Zhou et al., 2024, Yang et al., 21 Jul 2025). The literature therefore treats FA not merely as an antenna-selection problem, but as a geometry-dependent channel-design variable.

3. Diversity, outage, capacity, and fundamental limits

A major initial result was that a single-antenna FAS, by switching among many correlated ports, can outperform a conventional $$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$0-antenna maximum ratio combining (MRC) system if $$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$1 is large enough, even when the available physical space is very small (Wong et al., 2020). In the same line, ergodic-capacity analysis showed that capacity increases monotonically with $$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$2, and that with $$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$3 and $$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$4, FAS can achieve capacity on par with a 3-antenna MRC system under Rayleigh fading (Wong et al., 2020). These results established the original FA thesis: extremely fine-grained spatial sampling can create large selection diversity without multiple RF chains.

Subsequent work refined that thesis by tightening the correlation model. The approximation framework of (Khammassi et al., 2022) argued that earlier channel models may not accurately capture the correlation between ports given by Jakes’ model, and numerical results under the less-idealized correlation model showed that outage probability does not decrease indefinitely with $$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$5 for small $$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$6, but instead saturates after a certain $$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$7 (Khammassi et al., 2022). A later geometric framework made the point more explicit: the achievable diversity gain is governed not by the number of antenna ports, but by the channel’s effective rank, with

$$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$8

and diversity gain

$$|g_{\mathrm{FAS}}|=\max\{|g_1|,|g_2|,\ldots,|g_N|\},$$9

In that formulation, enlarging the explorable aperture increases effective rank, whereas increasing port density within a fixed aperture yields diminishing returns (Zhu et al., 10 Sep 2025). This suggests that “many ports in a tiny space” and “large effective aperture” should be distinguished as different asymptotic regimes rather than treated as interchangeable design principles.

Continuous-position analysis supplies a related benchmark. CFAS was shown to strictly outperform its discrete counterpart for a given finite number of ports and to provide the performance limits of FA-based systems (Psomas et al., 2023). In high-SNR error analysis under spatial correlation, the same aperture-centric logic reappears: the asymptotic symbol error rate depends on the determinant and eigenvalue spectrum of the correlation matrix, and the principal effective-rank threshold identified by a geometric knee algorithm coincides with the theoretical limit $x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$0 (Zhu et al., 10 Sep 2025).

Diversity conclusions also depend on temporal control. Under Nakagami-$x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$1 fading and perfect scheduling, the diversity order of FA selection combining is $x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$2, independent of whether the geometry is linear, circular, or wheel-shaped (Psomas et al., 2022). However, post-scheduling delay reduces the diversity order from $x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$3 to $x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$4, so diversity collapses to that of a single effective port unless temporal prediction is used; a linear prediction scheme was proposed to restore nearly all the original diversity (Psomas et al., 2022). In doubly shadowed UAV-to-ground links, the asymptotic outage analysis yields a different but related multiplicative law,

$x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$5

where $x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$6 is the FAS spatial rank and $x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$7 is the intrinsic channel diversity order (Zhu et al., 21 Nov 2025).

4. Optimization and signal-processing frameworks

A defining feature of modern FA research is that antenna position is optimized jointly with conventional physical-layer variables. In downlink FA-aided ISAC, the base station is equipped with $x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$8 FAs serving $x^\star=\arg\max_{x\in\mathcal{X}} |h^{(t)}[x]|,$9 users, while a sensing target may also act as an eavesdropper. The antenna position vector (APV)

$\mathcal{X}$0

is optimized jointly with user beamforming vectors $\mathcal{X}$1 to maximize the multiuser sum secrecy rate under sensing-power, position, minimum-spacing, and transmit-power constraints (Li et al., 26 Feb 2026). The resulting non-convex problem is handled by a block successive upper-bound minimization (BSUM) algorithm, with the proximal distance algorithm (PDA) yielding closed-form beamformer updates and extrapolated projected gradient (EPG) used for APV optimization; the reported FA-ISAC scheme achieved over $\mathcal{X}$2 sum secrecy rate gain compared to FPA systems (Li et al., 26 Feb 2026).

Near-field formulations make FA position a sensing variable as well as a communications variable. In an XL-STARS-enabled near-field ISAC system, the target employs a fluid antenna whose active position is encoded by a binary APV $\mathcal{X}$3, and the objective is to minimize the Cramér–Rao bound (CRB) for target localization while satisfying communication SINR, power, and XL-STARS constraints (Zhou et al., 2024). The problem is decomposed by penalty dual decomposition (PDD) and block coordinate descent (BCD) into subproblems over beamforming/sensing covariance, XL-STARS parameters, and FA position; simulation results reported substantial sensing gains and up to a $\mathcal{X}$4 reduction in RCRB in some cases relative to fixed-position baselines (Zhou et al., 2024).

FA optimization also appears in communication-centric designs. In FA-empowered receive spatial modulation (FA-RSM), the transmitter uses an FA with $\mathcal{X}$5 densely placed ports and activates $\mathcal{X}$6 ports, while port selection is performed either by exhaustive capacity maximization or by lower-complexity TMD and MCE-TMD procedures that explicitly exploit spatial correlation (Guo et al., 9 Jun 2025). For wideband 5G NR OFDM, port selection is generalized through a port-selection matrix that aggregates per-subcarrier and per-symbol SNR into a wideband criterion, and adaptive modulation and coding is driven by a BICM-capacity-based effective SNR mapping (Hong et al., 7 Mar 2025). In over-the-air federated learning, receiver beamforming, user selection, and FA positioning are optimized jointly through PDD to minimize an upper bound on the training loss and thus accelerate convergence (Zhao et al., 17 Feb 2025).

A separate line of work uses FA mobility to synthesize virtual apertures for array processing. Under time-constrained mobility, two specialized FA structures support aligned received signals and non-aligned received signals, respectively, and DOA estimation is performed by TMRLS-MUSIC or TMR-MUSIC with Nyström approximation (Xu et al., 14 Aug 2025). A related LoS-centric design develops an eigenvalue-ratio test for LoS path-number detection followed by a polynomial root-finding estimator, emphasizing that FA mobility can replace part of the algorithmic burden with hardware-generated spatial degrees of freedom (Xu et al., 14 Aug 2025).

5. Applications across wireless systems

The earliest multiuser application was fluid antenna multiple access (FAMA), in which each user independently selects the port where interference is in a deep fade and the desired signal is strong, thereby improving SIR without sophisticated signal processing (Wong et al., 2020). The theory derived a double-integral outage expression, a closed-form outage upper bound, and an average outage-capacity lower bound with arbitrary numbers of interferers, and concluded that it is possible for FAMA to support hundreds of users using only one fluid antenna at each user in a few wavelengths of space (Wong et al., 2020). Later hardware-oriented work on meta-fluid antennas extended this line to multi-activation with a single RF chain, higher SIR under various Rayleigh-fading environments, and CSI-free multi-user communication with optimization within a $\mathcal{X}$7 timeframe (Liu et al., 15 Sep 2025).

Sensing and localization have become a second major FA domain. In secure ISAC, the secrecy rate for user $\mathcal{X}$8 is modeled as

$\mathcal{X}$9

so FA position becomes a means of jointly improving user SINR and suppressing leakage to the sensing target acting as eavesdropper (Li et al., 26 Feb 2026). In near-field ISCSC, the FA-enabled framework jointly optimizes beamforming, FA positioning, and semantic extraction ratio to maximize worst-case semantic secrecy rate under CRB, power, computational, and latency constraints, and the reported simulations showed higher data rates and better privacy preservation (Yang et al., 21 Jul 2025). In sparse-array DOA estimation, the mobility-generated virtual aperture enabled underdetermined estimation: the reported examples state that three FAs estimate eleven directions and four hybrid antennas estimate nine directions, outcomes that are impossible for conventional fixed-position arrays with the same number of physical elements (Xu et al., 14 Aug 2025).

FA has also been integrated with system models that depart from classical narrowband terrestrial links. In UAV-to-ground communication under double-shadowing fading, an $8\times 15$0-port FAS receiver with one RF chain was analyzed through an eigenvalue-based approximation, yielding analytical expressions for outage probability, average bit error rate, and average channel capacity, together with the multiplicative diversity law $8\times 15$1 (Zhu et al., 21 Nov 2025). In 5G NR, a wideband FAS-OFDM receiver using a $8\times 15$2 two-dimensional port arrangement was shown by extensive link-level simulations to achieve striking BLER and throughput improvements in 3GPP-compliant wideband channels (Hong et al., 7 Mar 2025). In over-the-air federated learning, the FA-enabled server array reconfigures its geometry each round to reduce aggregation error and accelerate convergence, and experiments on MNIST and CIFAR-10 showed markedly faster loss decay and higher test accuracy than fixed-antenna baselines (Zhao et al., 17 Feb 2025).

6. Impairments, misconceptions, and research directions

Practical FA performance is constrained by channel uncertainty, hardware impairments, and mechanical control. A recent survey identifies estimation errors, temporal variability and outdated CSI, spatial-correlation mismatch, and feedback or quantization errors as sources of wrong port selection and SNR loss (Pakravan et al., 30 Jan 2026). The same survey also emphasizes RF nonlinearities, phase noise, insertion loss and impedance mismatch, mutual port coupling, mechanical-thermal effects, response delays, actuation energy, and position inaccuracies as system-level bottlenecks (Pakravan et al., 30 Jan 2026). These issues are not peripheral: the diversity analysis of (Psomas et al., 2022) shows concretely that post-scheduling delay can reduce the diversity order from $8\times 15$3 to $8\times 15$4, and that prediction is required to recover the lost diversity (Psomas et al., 2022).

A common misconception is that increasing the number of ports always yields commensurate diversity or error-rate gains. Early outage and capacity analyses indeed showed monotonic improvement with $8\times 15$5 and even the possibility of surpassing MRC in arbitrarily small space if $8\times 15$6 is large enough (Wong et al., 2020, Wong et al., 2020). Later works, however, stressed that realistic correlation modeling produces saturation for fixed apertures and that the effective rank is fundamentally aperture-limited (Khammassi et al., 2022, Zhu et al., 10 Sep 2025). This is not a contradiction so much as a refinement of assumptions: under simplified selection models, dense spatial sampling can dominate; under exact Jakes-type correlation and asymptotic error analysis, aperture and eigenvalue structure become the governing quantities. A plausible implication is that FA design should be interpreted through joint aperture–port-density–mobility tradeoffs rather than through port count alone.

The research agenda is correspondingly cross-disciplinary. Open directions include high-fidelity spatio-temporal channel modeling that couples electromagnetics and fluidics, advanced real-time control of fluid motion, robust beamforming and port selection under imperfect CSI, AI-driven channel estimation and configuration, integration with RIS, massive MIMO, mmWave, THz, IoT, wearables, and federated learning, cross-domain co-design across electromagnetic, mechanical, and algorithmic layers, and standardized testing and prototyping (Pakravan et al., 30 Jan 2026, Wu et al., 2024). The current literature therefore positions FA not as a single technique, but as a reconfigurable spatial-interface paradigm whose ultimate utility depends on how faithfully channel structure, actuation physics, and system optimization are jointly modeled.