Segmented Waveguide Pinching Antenna (SWAN)

• SWAN is a reconfigurable antenna architecture that segments a long dielectric waveguide into independently controlled short segments, reducing propagation loss and inter-antenna interference.
• It employs segment-level protocols like switching, aggregation, and multiplexing to optimize communication, integrated sensing, and over-the-air computation.
• SWAN enhances maintainability and reliability by modularizing the waveguide, improving repair rates while enabling precise, metric-dependent control strategies.

A segmented waveguide-enabled pinching-antenna system (SWAN) is a reconfigurable antenna architecture derived from the pinching-antenna system (PASS) in which a single long dielectric waveguide is replaced by multiple short, independently fed segments, each supporting reconfigurable pinching antennas (PAs) and segment-level control. In the literature, SWAN is used to mitigate inter-antenna radiation in uplink reception, reduce in-waveguide propagation loss, simplify reception and sensing models, and create new control dimensions such as segment switching, aggregation, selection, multiplexing, and Tx/Rx partitioning for communications, integrated sensing and communication (ISAC), over-the-air computation, and maintainability analysis (Gu et al., 23 Dec 2025, Ouyang et al., 12 Feb 2026, Jiang et al., 8 Dec 2025).

1. Architectural foundations and relation to PASS

PASS is built on a dielectric waveguide that carries RF energy as a guided wave and a set of PAs realized by locally disturbing the dielectric structure with small dielectric particles. The disturbance creates a local mismatch in effective permittivity and wave impedance, so part of the guided wave leaks into free space and behaves as a radiating element. In the foundational formulation, the guided wavelength is written as $\lambda_g=\lambda_0/\sqrt{\varepsilon_r}$, and the effective antenna position is determined by the coordinate of the pinch along the waveguide. This makes antenna position control equivalent to waveguide-coordinate control and enables LoS-oriented placement close to users or targets (Yang et al., 18 Jan 2025).

The principal architectural difference between PASS and SWAN is segmentation. Instead of one monolithic waveguide of length $D_x$, SWAN divides the structure into $M$ short segments of length $L$, typically arranged end-to-end so that $D_x=LM$. Each segment has its own feed point, and the segments are not electromagnetically continuous. In the uplink SWAN formulations, at most one PA is active per segment, with $\psi_0^m\le \psi_m\le \psi_0^m+L$ and $|\psi_m-\psi_{m'}|\ge \Delta$ for mutual-coupling control. This isolates segment-wise guided propagation and removes the inter-antenna radiation that makes multi-PA uplink analysis on a single long PASS waveguide intractable (Gu et al., 23 Dec 2025).

The motivations for segmentation recur across the literature. One long waveguide provides very long coverage, but in-waveguide attenuation grows with length, a fault anywhere may force replacement of the whole waveguide, and uplink modeling with many PAs becomes difficult because signal captured by one PA can be re-radiated by others. Segmentation converts the waveguide into multiple short branches, each with bounded guided distance, modular maintenance, and controllable segment-level participation in signal transmission, reception, or sensing (Ouyang et al., 12 Feb 2026).

2. Electromagnetic and signal models

A common SWAN channel model uses a cascaded free-space and in-waveguide representation. For a user at $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$ and the PA on segment $m$ at $\boldsymbol{\psi}_m=[\psi_m,0,d]^{\mathsf T}$, the free-space LoS channel is written as

$D_x$0

with $D_x$1 and $D_x$2. The in-waveguide channel from the segment feed point $D_x$3 to the PA is commonly modeled as

$D_x$4

where $D_x$5 is the in-waveguide attenuation factor and $D_x$6 is the effective refractive index (Gu et al., 13 May 2026).

Under this model, the cascaded user-to-segment channel is

$D_x$7

and, when $D_x$8, it reduces to a pure phase-weighted spherical-wave term whose amplitude is governed by the distance $D_x$9 with $M$0. This same structure is reused in uplink sum-rate analysis, AirComp, and maintainability modeling, with different segment-control policies applied on top of the same segment-wise channel law (Gu et al., 23 Dec 2025).

In downlink ISAC formulations, SWAN introduces an additional layer of segment-level transmission control. One representative model uses a segmented dielectric waveguide of length $M$1 divided into $M$2 independently controllable intervals $M$3, with segment state $M$4 indicating Tx or Rx mode. PAs have coordinates $M$5, collected in a deployment vector $M$6, subject to range and minimum-spacing constraints. The baseband ISAC transmit vector is

$M$7

so SWAN geometry, segment partition, and beamforming are jointly coupled through the near-field channel $M$8 induced by PA positions (Gao et al., 11 Apr 2026).

For sensing-oriented monostatic ISAC with separate transmit and receive segmented waveguides, the transmit side uses $M$9 transmit segmented waveguides (TSWs), each with $L$0 transmit PAs, while the receive side uses $L$1 receive segmented waveguides (RSWs), each with one receive PA. The in-waveguide channels on TSWs and RSWs are

$L$2

which makes segmentation explicit in both the transmit and echo paths and gives a direct route to sensing-SNR and CRLB analysis (Jiang et al., 8 Dec 2025, Geng et al., 1 Apr 2026).

3. Operating protocols and performance regimes

Three segment-control paradigms appear repeatedly. Segment switching/selection (SS) activates only one segment through a switch network, so the effective channel is that of a single short segment chosen for the current user or objective. Segment aggregation (SA) combines the outputs of all segments into one RF chain, typically with a factor $L$3 to reflect noise aggregation or power splitting. Segment multiplexing (SM) assigns one RF chain per segment, enabling fully digital multi-segment beamforming. In ISAC, the same architectural idea also appears as segment-wise Tx/Rx partitioning, where each segment is assigned to transmission or reception through a binary variable $L$4 (Jiang et al., 8 Dec 2025).

For maintainability, SWAN admits closed-form performance laws that differ sharply from those of monolithic PASS. With failure-repair rate ratio $L$5 and service width $L$6, conventional PASS has

$L$7

where $L$8 is the probability of non-zero rate and $L$9 is outage probability. For SS-based SWAN with $D_x=LM$0,

$D_x=LM$1

while for SA-based SWAN,

$D_x=LM$2

The maintainability result is unambiguous: both SS-based and SA-based SWAN achieve higher PNR and lower OP than conventional PASS, and SA is stronger than SS (Ouyang et al., 12 Feb 2026).

For uplink sum-rate, the conclusions are subtler. In multiuser SA with a single RF chain, the achievable rate depends on the balance between coherent signal growth and noise accumulation. An upper bound analysis shows that, as the number of activated segments grows, the useful amplitude increases only logarithmically while combined noise grows linearly, so there exists at least one finite optimal activation level and full segment aggregation is generally suboptimal. This led to hybrid segment selection and aggregation (HSS/A), which jointly optimizes the active set and PA locations; the resulting greedy algorithms outperform conventional full-segment aggregation (Gu et al., 13 May 2026).

In SWAN-assisted ISAC, the sensing-side scaling differs again by protocol. For segment selection, the sensing gain over PASS increases with the number of segments and converges to a limit determined by $D_x=LM$3 and the attenuation coefficient. For segment aggregation, the gain can first decrease and then increase with segment count because splitting and aggregated noise compete with geometric gain; the paper derives critical segment counts $D_x=LM$4 and $D_x=LM$5 beyond which sensing SNR grows. For segment multiplexing, sensing SNR increases monotonically with segment number and achieves the best sensing-communication Pareto region, at the cost of one RF chain per segment (Jiang et al., 8 Dec 2025).

These results jointly establish a recurring theme: segmentation is beneficial, but the relevant optimality criterion is metric-dependent. Maintainability improves monotonically with segmentation under SS and SA, whereas uplink sum-rate under full SA and sensing-SNR under some SA regimes exhibit non-monotonic behavior. The literature therefore treats segment activation, aggregation, and multiplexing not as interchangeable options but as distinct operating regimes with different asymptotics (Ouyang et al., 12 Feb 2026, Gu et al., 13 May 2026, Jiang et al., 8 Dec 2025).

4. Reliability, maintainability, and failure–repair analysis

A distinctive contribution of the SWAN literature is the explicit treatment of maintainability. Each dielectric waveguide or segment is modeled as a repairable component with lifetime $D_x=LM$6 and repair time $D_x=LM$7, yielding a two-state continuous-time Markov chain with working state $D_x=LM$8 and failed state $D_x=LM$9. The transition probabilities are available in closed form, and the steady-state probabilities are

$\psi_0^m\le \psi_m\le \psi_0^m+L$0

Length dependence enters through unit-length parameters $\psi_0^m\le \psi_m\le \psi_0^m+L$1 and $\psi_0^m\le \psi_m\le \psi_0^m+L$2, so a monolithic waveguide of length $\psi_0^m\le \psi_m\le \psi_0^m+L$3 has $\psi_0^m\le \psi_m\le \psi_0^m+L$4 and $\psi_0^m\le \psi_m\le \psi_0^m+L$5, while a segment of length $\psi_0^m\le \psi_m\le \psi_0^m+L$6 has $\psi_0^m\le \psi_m\le \psi_0^m+L$7 and $\psi_0^m\le \psi_m\le \psi_0^m+L$8 (Ouyang et al., 12 Feb 2026).

This length scaling implies that shorter segments are simultaneously more reliable and easier to repair. In the same model, each segment has

$\psi_0^m\le \psi_m\le \psi_0^m+L$9

so segmentation improves both mean lifetime and mean repair time. The maintainability gain is then propagated into communication performance through the binary-rate model, where a nonfailed segment supports deterministic rate and a failed segment yields rate zero. Closed-form PNR and OP expressions show that the maintainability gain of SWAN over conventional PASS increases with the number of segments and that SA reaches the asymptotic gain much faster than SS (Ouyang et al., 12 Feb 2026).

The maintainability analysis also clarifies a common misconception: the advantage of SWAN is not confined to communication-theoretic path-loss considerations. In the failure-repair model, segmentation remains advantageous even when physical SNR is held constant, because the architecture changes the stochastic availability process itself. This suggests that SWAN should be interpreted not only as a beamforming structure but also as a modular reliability architecture (Ouyang et al., 12 Feb 2026).

5. Optimization, learning, and algorithmic control

SWAN design problems are generally mixed discrete-continuous and highly non-convex because segment states, PA positions, and beamforming vectors are jointly coupled through spherical-wave channels, in-waveguide phase, minimum-spacing constraints, and, in ISAC, sensing metrics such as CRLB or illumination. One representative near-field ISAC formulation jointly optimizes antenna deployment $|\psi_m-\psi_{m'}|\ge \Delta$0, segment-wise Tx/Rx partition $|\psi_m-\psi_{m'}|\ge \Delta$1, and beamforming matrices $|\psi_m-\psi_{m'}|\ge \Delta$2 under communication-rate and sensing-accuracy constraints. The paper addresses scenario variability through a CSI-induced self-graph, an SGNN encoder, and a GPT-2-style LLM backbone with LoRA, followed by a deployment/partition head and a beamforming head. The resulting framework achieves higher communication rates while maintaining reliable sensing accuracy, and its user-count transfer mechanism preserves a deployment MSE around $|\psi_m-\psi_{m'}|\ge \Delta$3 while reducing trainable parameters from approximately $|\psi_m-\psi_{m'}|\ge \Delta$4 in full retraining to $|\psi_m-\psi_{m'}|\ge \Delta$5–$|\psi_m-\psi_{m'}|\ge \Delta$6 when only the beamforming head is adapted; best-epoch convergence accelerates from 27–30 epochs to 8 epochs (Gao et al., 11 Apr 2026).

A complementary line of work formulates SWAN-ISAC as reinforcement learning. In the hybrid segment selection and multiplexing (HSSM) protocol, each segment has a selection variable $|\psi_m-\psi_{m'}|\ge \Delta$7, and the optimization variables are PA positions, segment selection, and beamforming. The segment hysteresis based reinforcement learning (SHRL) method embeds these variables into an A2C framework and applies a probabilistic hysteresis gate so that segment selections are not updated at every step. The reward is the sum communication rate minus a penalty for violating target-illumination thresholds. Simulations reported for a $|\psi_m-\psi_{m'}|\ge \Delta$8 area, $|\psi_m-\psi_{m'}|\ge \Delta$9, $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$0, $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$1 PAs per segment, and $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$2 GHz show that SWAN-HSSM outperforms SWAN-SM and PASS in both sparse and dense scenarios, while SHRL converges more smoothly and to higher reward than A2C, PPO, SPRL, and random control (Gao et al., 28 Jan 2026).

For sensing-driven ISAC, CRLB minimization leads to geometric optimization on non-Euclidean constraint sets. In a SWAN-assisted downlink ISAC system with $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$3 transmit segmented waveguides, $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$4 receive segmented waveguides, $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$5 transmit PAs per transmit segment, and one receive PA per receive segment, the target-location CRLB is minimized under communication-rate constraints by jointly optimizing beamforming and PA positions. The proposed method constructs a Riemannian product manifold, applies a penalty method, and uses a Riemannian Broyden-Fletcher-Goldfarb-Shanno algorithm to obtain feasible solutions. The reported result is superior CRLB performance for target localization compared with existing schemes including multi-waveguide-enabled pinching-antenna-assisted ISAC systems (Geng et al., 1 Apr 2026).

SWAN has also been adapted to over-the-air computation. In the AirComp setting, the effective uplink channels $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$6 determine the computation MSE

$\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$7

after optimal receive scaling. Three architectures are considered: segment selection, phase-shifter-free segment aggregation, and phase-shifter-enabled segment aggregation. Low-complexity algorithms jointly optimize PA positions and, when present, per-segment phase shifts. The key empirical conclusion is that both SS and SA achieve lower computation mean-squared error than conventional PASS, while segment-wise phase control further improves SA (Gu et al., 4 May 2026).

6. Applications, design tradeoffs, and limitations

SWAN appears in several application classes. In multiuser uplink, it supports TDMA and NOMA under both SS and SA, with low-complexity PA-placement algorithms. The reported numerical results show that SWAN achieves higher sum-rate performance than conventional PASS and that SA provides additional performance gains over SS. In the same setting, NOMA generally attains the highest sum-rate, while SA with optimized PA positions can deliver substantial gains over SS (Gu et al., 23 Dec 2025).

In sensing and ISAC, segmentation is repeatedly used to suppress uplink-model complexity, improve sensing accuracy, and enlarge the feasible sensing-communication tradeoff region. The sensing-focused SWAN-ISAC study reports that SM has the best Pareto front, followed by SA and SS, and that larger segment counts improve SA and SM much more than SS. In low-altitude wireless networks, the segmented waveguide-enabled formulation is explicitly adopted to mitigate in-waveguide attenuation; increasing the number of segments from $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$8 to $\mathbf{u}_k=[u_k^x,u_k^y,0]^{\mathsf T}$9 raises sensing rate from approximately $m$0 to $m$1 for SWISAC-GNN and from $m$2 to $m$3 for SWISAC-AO under the reported setup, while the GNN achieves sensing performance comparable to or better than AO with much lower implementation complexity (Jiang et al., 8 Dec 2025, Guo et al., 3 Dec 2025).

The broader PASS literature clarifies why segmentation became necessary. Closed-form outage and rate analysis for single-waveguide PASS shows that waveguide attenuation becomes a major bottleneck as guided distance grows, and that optimal pinching location balances free-space loss against guided loss. PASS multicast work further shows that spatial flexibility from movable radiators can outperform simply adding more fixed antennas. These results do not by themselves define SWAN, but they establish the propagation and optimization pressures that segmentation addresses: long-waveguide attenuation, wide-area deployment, and the need for tractable multi-radiator uplink models (Tyrovolas et al., 10 Feb 2025, Mu et al., 23 Feb 2025).

The limitations of current SWAN models are consistent across papers. Many formulations assume dominant LoS propagation, narrowband signaling, perfect CSI or exact position information, ideal synchronization, and simplified coupling captured only through minimum-spacing constraints. Several works either set $m$4 in the main optimization or treat in-waveguide attenuation as modest, even when numerical sections later examine nonzero loss. Multi-cell interference, severe blockage, quantized or slow PA repositioning, and hardware impairments are generally outside the main analytical models. This suggests that future SWAN research will likely focus on robust design under imperfect CSI, correlated or non-exponential failure models, multi-band or multi-waveguide generalizations, richer sensing metrics beyond illumination or single-target CRLB, and online control architectures that preserve the structural advantages of segmentation without assuming ideal actuation or channel knowledge (Ouyang et al., 12 Feb 2026, Gao et al., 28 Jan 2026, Geng et al., 1 Apr 2026).

Taken together, the literature portrays SWAN as a modular large-aperture architecture in which segmentation is the central design variable. It reduces in-waveguide loss, suppresses inter-antenna radiation, and creates segment-level control knobs that can be exploited differently for reliability, uplink sum-rate, sensing, ISAC, AirComp, and learning-based adaptation. The strongest recurring conclusion is not that one protocol dominates universally, but that segmentation makes protocol choice itself an optimization variable: SS emphasizes simplicity and bounded loss, SA trades extra combining for higher diversity and stronger maintainability, and SM converts the segmented structure into a fully digital near-field array with the largest performance envelope and the highest hardware cost (Ouyang et al., 12 Feb 2026, Gu et al., 13 May 2026, Jiang et al., 8 Dec 2025).