Scalable Fluid Antenna System Overview

• Scalable Fluid Antenna System (SFAS) is a reconfigurable antenna architecture that dynamically adjusts spatial sampling via hardware or software.
• It enhances system performance by increasing degrees-of-freedom for DOA estimation, reducing error bounds, and enabling robust interference management.
• Hardware realizations span fluidic motion, metasurface switching, and pixel-based beamforming, achieving microsecond-scale diversity gains in practical setups.

Scalable Fluid Antenna System (SFAS) denotes a class of reconfigurable antenna architectures in which antenna position, effective aperture, or array geometry is not fixed by a half-wavelength grid, but is adjusted in hardware or software to suit a communication, sensing, or localization objective. In the recent literature, this idea appears in several closely related forms: continuous-position sparse arrays with element locations $p_n\in[0,D]$ rather than $p_n\in\{0,d_0,2d_0,\dots\}$, where $d_0=\lambda/2$ (Wu et al., 19 May 2026); scale-adjustable arrays with $d(\alpha)=\alpha d_0$ and $p_m(\alpha)=(m-1)\alpha d_0$ (Wu et al., 14 Aug 2025); and port- or region-switching architectures that retain a single RF chain while exploiting spatial diversity or interference avoidance (Wong et al., 2020, Cheng et al., 4 Jun 2026). Across these formulations, SFAS replaces fixed spatial sampling by reconfigurable spatial sampling, thereby decoupling aperture design from classical grid constraints and making the aperture itself a system degree of freedom.

1. Definition and conceptual scope

The immediate precursor of SFAS is the fluid antenna system (FAS) formulated by Kai-Kit Wong et al., where a single RF chain feeds a mechanically flexible antenna that can switch among $N$ ports placed over a line of physical length $W\lambda$, and the selected port is the one with the largest envelope (Wong et al., 2020). In that original setting, the k-th port is located at $d_k=\frac{k-1}{N-1}W\lambda$, and the channel correlation is modeled by the Bessel kernel $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$ (Wong et al., 2020). This established the basic FAS principle: spatial reconfigurability can be traded for diversity while preserving a single RF chain.

SFAS generalizes that principle in two directions. First, it relaxes the discrete-grid restriction. In sparse array processing, classical designs such as nested arrays, coprime arrays, MRAs, and ULAs are constrained to $p_n\in\{0,d_0,2d_0,\dots\}$, whereas FAS allows $p_n\in\{0,d_0,2d_0,\dots\}$0 with arbitrary real-valued locations (Wu et al., 19 May 2026). Second, it promotes aperture reconfiguration from a port-selection mechanism to a system-level design variable. In scale-adjustable SFAS, a global scaling factor $p_n\in\{0,d_0,2d_0,\dots\}$1 controls inter-element spacing and total aperture, so that $p_n\in\{0,d_0,2d_0,\dots\}$2 and the Rayleigh distance becomes $p_n\in\{0,d_0,2d_0,\dots\}$3 (Wu et al., 14 Aug 2025).

A common simplification is to equate SFAS with a mechanically moved liquid-metal radiator. The literature is broader. Proposed realizations include intrinsically elastic conductors joining rigid array nodes, mesh-spring interconnects between rigid antenna “pixels,” and microfluidic channels carrying conductive liquid (Wu et al., 14 Aug 2025). Other works implement the same functional objective through metasurface port activation or through beamforming networks that synthesize the correlation pattern of a physically moved radiator without moving it at all (Liu et al., 7 Feb 2025, Zhang et al., 3 Dec 2025). This wider design space is central to the “scalable” aspect of SFAS.

2. Scaling laws and theoretical limits

In continuous-position sparse FAS design for direction-of-arrival estimation, the central object is the difference coarray $p_n\in\{0,d_0,2d_0,\dots\}$4. Defining the coarray DOF as the largest integer $p_n\in\{0,d_0,2d_0,\dots\}$5 such that $p_n\in\{0,d_0,2d_0,\dots\}$6, the universal dual DOF bound is

$p_n\in\{0,d_0,2d_0,\dots\}$7

This bound separates two regimes: a classical $p_n\in\{0,d_0,2d_0,\dots\}$8 saturation regime and a linear-in-aperture regime in which the DOF grows with $p_n\in\{0,d_0,2d_0,\dots\}$9 (Wu et al., 19 May 2026). In the same framework, the Cramér–Rao bound for $d_0=\lambda/2$0 uncorrelated far-field sources scales as

$d_0=\lambda/2$1

yielding a $d_0=\lambda/2$2-fold improvement over the best grid design (Wu et al., 19 May 2026). For a single source, $d_0=\lambda/2$3, where $d_0=\lambda/2$4, and optimizing over $d_0=\lambda/2$5 gives $d_0=\lambda/2$6 (Wu et al., 19 May 2026).

A separate line of analysis treats scalable fluid arrays through observation entropy. In dual-configuration S-FAS, a compressed mode has effective dimension $d_0=\lambda/2$7, an extended mode has $d_0=\lambda/2$8, and the observation entropy satisfies

$d_0=\lambda/2$9

Subspace-based identifiability requires at least one noise-subspace dimension, so $d(\alpha)=\alpha d_0$0. This produces the capacity hierarchy

$d(\alpha)=\alpha d_0$1

with the last term achieved by joint stacking of compressed and extended observations (Wu et al., 1 Apr 2026). The same paper introduces the noise entropy ratio,

$d(\alpha)=\alpha d_0$2

as a diagnostic that separates fundamental DoF exhaustion from algorithmic deficiency (Wu et al., 1 Apr 2026).

In fading-channel analysis, the fundamental scaling law is expressed through symbol error rate. For a single-antenna FAS moving over $d(\alpha)=\alpha d_0$3 with $d(\alpha)=\alpha d_0$4 ports and spatial correlation matrix $d(\alpha)=\alpha d_0$5, the high-SNR asymptotic SER obeys

$d(\alpha)=\alpha d_0$6

with the leading constant depending on $d(\alpha)=\alpha d_0$7 (Zhu et al., 5 Nov 2025). As the movement space $d(\alpha)=\alpha d_0$8 increases, $d(\alpha)=\alpha d_0$9, $p_m(\alpha)=(m-1)\alpha d_0$0, and the diversity order approaches $p_m(\alpha)=(m-1)\alpha d_0$1; by contrast, increasing port density within fixed $p_m(\alpha)=(m-1)\alpha d_0$2 reduces inter-port spacing, increases correlation, and yields diminishing returns (Zhu et al., 5 Nov 2025). This complements the earlier outage analysis showing that a single-antenna FAS over any arbitrarily small space can outperform an $p_m(\alpha)=(m-1)\alpha d_0$3-antenna MRC system if $p_m(\alpha)=(m-1)\alpha d_0$4 is large enough (Wong et al., 2020).

3. Position design and inference algorithms

For source localization and array design, SFAS introduces continuous optimization where classical sparse-array design often required combinatorial search. In the single-source D-optimal problem, maximizing $p_m(\alpha)=(m-1)\alpha d_0$5 over ordered positions $p_m(\alpha)=(m-1)\alpha d_0$6 yields a closed-form optimizer: for even $p_m(\alpha)=(m-1)\alpha d_0$7, $p_m(\alpha)=(m-1)\alpha d_0$8 for $p_m(\alpha)=(m-1)\alpha d_0$9 and $N$0 for $N$1, which gives $N$2; for odd $N$3, $N$4 elements are placed at each endpoint and one at $N$5 (Wu et al., 19 May 2026). For multiple sources, the objective $N$6 is non-convex, so the design is relaxed to a probability measure $N$7 on $N$8 and optimized by a Frank–Wolfe procedure. The resulting algorithm attains $N$9-optimality in $W\lambda$0 iterations, each with a one-dimensional line search (Wu et al., 19 May 2026).

A recurring estimation strategy is two-stage processing. In sparse FAS-MUSIC, the first stage forms the sample covariance $W\lambda$1, vectorizes it to $W\lambda$2, averages redundant entries over the difference lags to construct a virtual ULA, performs spatial smoothing with subarray size $W\lambda$3, and applies classical ULA-MUSIC to obtain coarse estimates $W\lambda$4 (Wu et al., 19 May 2026). The second stage solves a local maximum-likelihood refinement

$W\lambda$5

typically with $W\lambda$6 and a quasi-Newton method such as L-BFGS-B (Wu et al., 19 May 2026). This architecture addresses the grating-lobe ambiguity of large non-uniform apertures while recovering near-CRB precision. Reported simulations show that FAS-MUSIC achieves $W\lambda$7 lower RMSE than ULA MUSIC, that FAS with $W\lambda$8 antennas outperforms MRA with $W\lambda$9 antennas, and that source separations down to $d_k=\frac{k-1}{N-1}W\lambda$0 are resolved with RMSE $d_k=\frac{k-1}{N-1}W\lambda$1 (Wu et al., 19 May 2026).

The scale-adjustable ESG-based SFAS adopts a structurally similar but model-richer approach. In the compressed mode $d_k=\frac{k-1}{N-1}W\lambda$2, the array uses a far-field manifold with mutual-coupling mitigation by spatial smoothing on the central $d_k=\frac{k-1}{N-1}W\lambda$3 elements; in the extended mode $d_k=\frac{k-1}{N-1}W\lambda$4, coupling is negligible and exact spatial geometry is used for one-dimensional range search followed by localized two-dimensional DOA-range refinement (Wu et al., 14 Aug 2025). The exact distance model

$d_k=\frac{k-1}{N-1}W\lambda$5

eliminates the need for near-field or far-field approximation (Wu et al., 14 Aug 2025).

The main caveat to two-stage processing is that it need not be capacity-optimal. Information-theoretic analysis shows that sequential compressed-then-extended processing is bottlenecked by the first stage, whereas joint MUSIC on the stacked observation

$d_k=\frac{k-1}{N-1}W\lambda$6

can approach the joint capacity bound $d_k=\frac{k-1}{N-1}W\lambda$7 (Wu et al., 1 Apr 2026). A common misconception is therefore that a compressed coarse stage followed by an extended fine stage always extracts the full benefit of reconfigurability. The derived capacity hierarchy shows that this is not generally the case.

4. Hardware realizations and implementation modalities

The hardware literature on SFAS spans fluidic motion, metasurface switching, and electronically synthesized motion. These implementations differ in actuation mechanism but pursue the same core objective: to make the antenna’s effective spatial sampling pattern reconfigurable on demand.

Realization	Mechanism	Representative result
Surface-wave enabled SCFA/DCFA (Shen et al., 2024)	Galinstan slug in single- or double-channel fluid radiator excited by a planar surface-wave launcher	In 4-user FAMA, double-channel FAS reduces outage probability by 57% and increases multiplexing gain to 2.27
Programmable meta-fluid antenna (Liu et al., 7 Feb 2025)	SIW metasurface slab with $d_k=\frac{k-1}{N-1}W\lambda$8 selectable ports, one RF feed, FPGA-controlled PIN diodes	At $d_k=\frac{k-1}{N-1}W\lambda$9 GHz, average SINR is 8.04 dB, 7.74 dB, and 9.69 dB for three transmitters; fixed-position horn is below 5 dB average
PRBFN-FAS (Zhang et al., 3 Dec 2025)	Pixel-based reconfigurable beamforming network that switches among pre-computed excitation-current vectors	Over a 5% bandwidth around 2.6 GHz, measured $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$0 dB, $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$1 dB, and two-user FAMA attains SIR $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$2 dB

The surface-wave enabled mmWave prototypes fabricated on Rogers RT5880 use a $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$3 mm$[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$4 PCB, Galinstan with $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$5 S/m, and channels with $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$6 mm and $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$7 mm (Shen et al., 2024). In the 24–30 GHz band, the measured radiation patterns can vary up to an averaged value of 11 dBi, and by shifting the radiator position one can create deep nulls exceeding 20 dB toward an interferer direction (Shen et al., 2024).

The programmable meta-fluid antenna replaces literal motion by dense spatial sampling on a $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$8 patch at 26.5 GHz. It uses 120 selectable meta-atoms, each comprising two complementary magnetic-dipole slots, a single RF chain, and FPGA switching up to 20 MHz, cycling through all 120 positions in approximately $[R]_{k,\ell}=J_0(2\pi|d_k-d_\ell|/\lambda)$9s (Liu et al., 7 Feb 2025). Under rich scattering, the measured platform demonstrates opportunistic interference avoidance through the FAMA rule $p_n\in\{0,d_0,2d_0,\dots\}$0, and achieves a full spatial multiplexing order of $p_n\in\{0,d_0,2d_0,\dots\}$1 with only $p_n\in\{0,d_0,2d_0,\dots\}$2 RF chains, one per user, without CSI or precoding at the base station (Liu et al., 7 Feb 2025).

PRBFN-FAS makes the strongest challenge to the idea that FAS must be mechanically slow. It starts from the equivalence between physical movement and switching the excitation current vector of a multi-port antenna, then realizes the desired port set through a scalable pixel-based reconfigurable beamforming network (Zhang et al., 3 Dec 2025). Two design examples are given for equivalent physical movements of 0.5 and 1.5 wavelengths. System-level experiments in a 2×2 MIMO testbed show more than 5 dB instantaneous received-power variation as the port is switched, with microsecond-scale antenna diversity (Zhang et al., 3 Dec 2025).

5. Wireless communication, multiple access, and edge intelligence

In communication theory, the simplest FAS use case is single-link reliability improvement by selecting the strongest among correlated spatial samples. The original outage analysis derived exact and approximate closed forms and a simple upper bound, and established that $p_n\in\{0,d_0,2d_0,\dots\}$3 as $p_n\in\{0,d_0,2d_0,\dots\}$4 for any fixed $p_n\in\{0,d_0,2d_0,\dots\}$5, despite having only one RF chain (Wong et al., 2020). This provided the first rigorous statement that a spatially reconfigurable single-antenna system can substitute port diversity for multi-branch combining.

The multiuser communication literature then reframed FAS as an access mechanism rather than only a diversity mechanism. In fluid antenna multiple access (FAMA), each user locally selects the port maximizing

$p_n\in\{0,d_0,2d_0,\dots\}$6

without CSI feedback to the base station (Shen et al., 2024). Experimental data from the surface-wave enabled platform show that static fixed-port or random port selection gives outage around $p_n\in\{0,d_0,2d_0,\dots\}$7–$p_n\in\{0,d_0,2d_0,\dots\}$8 and multiplexing gain around $p_n\in\{0,d_0,2d_0,\dots\}$9–$p_n\in\{0,d_0,2d_0,\dots\}$00, whereas the double-channel fluid antenna reaches outage $p_n\in\{0,d_0,2d_0,\dots\}$01 and multiplexing gain $p_n\in\{0,d_0,2d_0,\dots\}$02 for $p_n\in\{0,d_0,2d_0,\dots\}$03 users at 26 GHz (Shen et al., 2024). In the programmable meta-fluid architecture, the same opportunistic logic is implemented as exhaustive scanning over 120 positions, and average SINR gains over a fixed-position horn antenna are observed under 3-user rich-scattering measurements (Liu et al., 7 Feb 2025).

SFAS has also entered wireless learning systems. In FAir-FL, one parameter server communicates with $p_n\in\{0,d_0,2d_0,\dots\}$04 devices, each equipped with a single fluid antenna that can switch among $p_n\in\{0,d_0,2d_0,\dots\}$05 closely spaced ports during over-the-air federated learning aggregation over $p_n\in\{0,d_0,2d_0,\dots\}$06 OFDM symbols and $p_n\in\{0,d_0,2d_0,\dots\}$07 subcarriers (Park et al., 4 Mar 2025). The framework introduces a robustness-maximizing port rule,

$p_n\in\{0,d_0,2d_0,\dots\}$08

an accuracy-maximizing rule based on the number of “good” subcarriers, and a hybrid switching rule $p_n\in\{0,d_0,2d_0,\dots\}$09 determined by the effective noise ratio $p_n\in\{0,d_0,2d_0,\dots\}$10 and a threshold $p_n\in\{0,d_0,2d_0,\dots\}$11 (Park et al., 4 Mar 2025). Under assumptions of $p_n\in\{0,d_0,2d_0,\dots\}$12-smoothness, bounded variance, unbiasedness, and gradient diversity, the convergence analysis shows that choosing $p_n\in\{0,d_0,2d_0,\dots\}$13 yields

$p_n\in\{0,d_0,2d_0,\dots\}$14

(Park et al., 4 Mar 2025). On MNIST, with $p_n\in\{0,d_0,2d_0,\dots\}$15 and $p_n\in\{0,d_0,2d_0,\dots\}$16 dBm, FAir-FL improves test accuracy by 10–20 points at low power and succeeds at $p_n\in\{0,d_0,2d_0,\dots\}$17 dBm for $p_n\in\{0,d_0,2d_0,\dots\}$18 while $p_n\in\{0,d_0,2d_0,\dots\}$19 fails; on CIFAR-10, with $p_n\in\{0,d_0,2d_0,\dots\}$20, fluid-antenna schemes achieve up to 15 points higher final accuracy and faster convergence (Park et al., 4 Mar 2025).

6. Extensions, constraints, and open directions

The most explicit geometric extension of SFAS is the 3D spherical fluid antenna system. Here, candidate radiators are distributed on a spherical shell of radius $p_n\in\{0,d_0,2d_0,\dots\}$21, partitioned into regions $p_n\in\{0,d_0,2d_0,\dots\}$22, and controlled at two levels: region switching determines which spherical patch is active, and element-level reconfiguration determines how many and which ports within that patch are energized (Cheng et al., 4 Jun 2026). This structure supports flexible beamforming, concurrent multi-region transmission, blockage-adaptive aperture switching, effective-aperture reconfiguration, and high-resolution 3D aperture control (Cheng et al., 4 Jun 2026). In the numerical setup with $p_n\in\{0,d_0,2d_0,\dots\}$23 ports and $p_n\in\{0,d_0,2d_0,\dots\}$24 regions, the 3D SFAS with joint reconfiguration achieves about 9 bps/Hz at 10 dB SNR, compared with about 7 bps/Hz for a 2D FAS and about 5 bps/Hz for a fixed planar array; in a two-user scenario, activating two disjoint surface patches yields two independent beams, each with approximately 90% of the single-beam peak gain (Cheng et al., 4 Jun 2026).

Practical constraints are already prominent in the 1D and 2D literature. In sparse array design, minimum-spacing constraints can be enforced directly or through a penalty term $p_n\in\{0,d_0,2d_0,\dots\}$25, with only a small DOF loss according to Proposition 9 (Wu et al., 19 May 2026). Mutual coupling modifies steering vectors through $p_n\in\{0,d_0,2d_0,\dots\}$26, while finite position accuracy $p_n\in\{0,d_0,2d_0,\dots\}$27 produces phase noise and an SNR loss that scales as $p_n\in\{0,d_0,2d_0,\dots\}$28; sub-$p_n\in\{0,d_0,2d_0,\dots\}$29 accuracy is sufficient for less than 1 dB loss (Wu et al., 19 May 2026). In scanning-based metasurface implementations, latency must satisfy $p_n\in\{0,d_0,2d_0,\dots\}$30, and at $p_n\in\{0,d_0,2d_0,\dots\}$31 MHz with $p_n\in\{0,d_0,2d_0,\dots\}$32, the scan time is $p_n\in\{0,d_0,2d_0,\dots\}$33s (Liu et al., 7 Feb 2025).

The open problems are correspondingly multi-layered. Information-theoretic analysis identifies sequential processing bottlenecks and motivates joint processing (Wu et al., 1 Apr 2026). Error-probability analysis shows that increasing movement space is fundamentally different from merely increasing port density (Zhu et al., 5 Nov 2025). Hardware studies point to insertion loss, control-line complexity, and switch-network complexity, with the 3D SFAS survey noting $p_n\in\{0,d_0,2d_0,\dots\}$34 scaling for the switch network and the need for compressed sensing or hierarchical training for CSI acquisition across many regions (Cheng et al., 4 Jun 2026). A plausible implication is that “scalability” in SFAS is not a single property but a conjunction of scalable aperture, scalable observation space, and scalable implementation. The present literature establishes each of these aspects separately, but their unified co-design remains an open systems problem.