Segmented Pinching Antenna (SWAN)

• SWAN is a reconfigurable, segmented antenna system that partitions a dielectric waveguide into independently fed segments for precise beamforming.
• It employs mechanically repositionable pinching antennas and diverse segment protocols (SS, SA, SM) to balance hardware complexity and array gain.
• The architecture integrates joint beamforming, placement optimization, and learning-based control for enhanced connectivity and reduced propagation loss.

A segmented pinching-antenna system (SWAN) is a reconfigurable, spatially flexible wireless aperture that partitions a dielectric waveguide into discrete, independently fed segments, each hosting mechanically repositionable “pinching antennas.” This architecture enables dynamic user-centric connectivity, spatially adaptive beamforming, reduced in-waveguide propagation loss, enhanced maintainability, and tractable modeling free from inter-antenna re-radiation (IAR). SWAN is a foundational element in next-generation wireless communication, sensing, and computation systems. The framework combines advances in channel modeling, placement optimization, hybrid beamforming, and learning-based control.

1. Physical and Channel Model of SWAN

The canonical SWAN comprises a rigid dielectric waveguide of length $L$ parallel to the $x$-axis at height $d_v$, subdivided into $S$ ("segments") of length $L/S$ (Xu et al., 30 Jun 2025). Each segment possesses a feed port and supports a sliding or statically mounted pinching antenna, which can be positioned at any coordinate $x_m$ within its segment $[(m-1)L/S, mL/S]$. Activation of antennas is managed independently, providing spatial degrees of freedom for beam steering and user-centric optimization.

The end-to-end channel coefficient from segment $m$ to a user at $(x_u, d_u)$ combines in-waveguide attenuation and free-space loss: $$h(x_m) = C_0 \cdot e^{-\alpha x_m} \Big((x_m - x_u)^2 + d_u^2 \Big)^{-\beta/2} e^{-j \Phi(x_m)}$$ where $x$0, $x$1 is the waveguide attenuation coefficient (e.g., $x$2 dB/m $x$3 m$x$4), $x$5 is the free-space path-loss exponent, and the phase term $x$6 includes guided and radiative propagation (Xu et al., 30 Jun 2025).

When segments are electrically isolated, IAR is fully eliminated: PAs on one segment cannot couple signal into or reradiate via other segments (Ouyang et al., 12 Sep 2025). Thus, SWAN supports physically consistent uplink models for large, distributed, or user-centric array deployments.

2. Segment Protocols and Hardware Complexity

Three key segment operating protocols are established for both uplink and downlink, each presenting a unique trade-off between hardware complexity and array gain (Ouyang et al., 12 Sep 2025, Gan et al., 20 Nov 2025, Jiang et al., 8 Dec 2025, Gu et al., 4 May 2026):

Protocol Abbreviation	Segment Control	RF Complexity	Achievable Gain
SS	Segment Selection	1 RF chain	No inter-segment array gain
SA	Segment Aggregation	1 RF chain + combiner	Array gain, phase alignment needed
SM	Segment Multiplexing	$x$7 RF chains	Full digital (MIMO), max gain

• SS: Only one segment is active per resource slot, minimizing RF hardware but precluding inter-segment array gain.
• SA: All $x$8 segments are simultaneously connected via combiner, enabling coherent (or aggregate) array gain. Optional per-segment phase shifter networks further enhance phase alignment (Type-II SA).
• SM: Each segment is connected to a dedicated RF chain, supporting full digital beamforming.

Optimal segment size $x$9 and the number of segments $d_v$0 balance in-guide loss against complexity. In practice, $d_v$1–$d_v$2 cm is typical ($d_v$3–$d_v$4) (Ouyang et al., 12 Sep 2025, Gan et al., 20 Nov 2025).

3. Placement, Beamforming, and Algorithmic Optimization

A core advantage of SWAN is flexibility in the spatial deployment of pinching antennas and the joint optimization of beamforming and placement. Key algorithmic developments include:

Single-user closed-form optimization:

Given attenuation, the optimal PA position $d_v$5 maximizes $d_v$6:

• If $d_v$7: $d_v$8 (feed).
• Otherwise: $d_v$9 (Xu et al., 30 Jun 2025).

Multi-user joint optimization:

Two families dominate:

• WMMSE-based joint beamforming and placement: Alternating updates over receive combiners, auxiliary weights, beamformers (QCQP), and PA locations (1-D grid search). Sum-rate optimization: $S$0 (Xu et al., 30 Jun 2025).
• Two-stage approach (e.g., MRC): Fix MRC beamformers, alternately optimize over antenna positions and power allocation using block coordinate descent and projected gradients for fast convergence (Xu et al., 30 Jun 2025, Jiang et al., 4 Mar 2026).

Tri-hybrid architectures:

SWAN supports digital, analog (RF phase-shifters), and spatial ("pinching") degrees of freedom. For fully connected (FC) analog combining, Riemannian manifold optimization is adopted; for partially connected (PC), element-wise phase calibration suffices (Jiang et al., 4 Mar 2026). Pinching beamforming is solved by Gauss-Seidel coordinate search over feasible segment locations.

Placement scaling laws:

• In FC case for large $S$1: maximum SNR decays $S$2 (non-monotonic).
• In PC (MRC): SNR saturates to a constant as $S$3 (Jiang et al., 4 Mar 2026).

4. System Performance: Attenuation, Array Gain, and Maintainability

Attenuation impact and optimality: Inclusion of the in-waveguide coefficient reveals nontrivial dependencies of optimal PA location and achievable rate on $S$4 and $S$5. Average rate-loss under attenuation-ignorant placement is $S$6, with explicit design bounds on “safe neglect” of attenuation (Xu et al., 30 Jun 2025). For typical values (e.g., $S$7 m$S$8, $S$9 m), attenuation-ignorant placement is viable for user regions $L/S$0 m.

Maintainability: SWAN greatly improves operational reliability under random segment failures, described by failure (hazard) rate $L/S$1 and repair rate $L/S$2, both scaled per unit length (Ouyang et al., 12 Feb 2026). Probability of nonzero rate (PNR) for SWAN in SS is $L/S$3 ($L/S$4), offering up to $L/S$5 gain over monolithic PASS and even $L/S$6 gain in SA.

Propagation loss and scalability: Segmentation strictly increases average in-waveguide gain: $L/S$7 is monotonically increasing in $L/S$8, with diminishing returns for large $L/S$9 (Gan et al., 20 Nov 2025).

5. Applications in Communications, Computation, Sensing, and Security

Massive MIMO and Uplink/Downlink: SWAN outperforms conventional fixed and PASS-based ULAs in uplink sum-rate. Gains of 20–35% in sum-rate and per-user rates are reported for realistic $x_m$0 and deployment geometries (Xu et al., 30 Jun 2025, Gu et al., 23 Dec 2025, Ouyang et al., 12 Sep 2025). SA and SM protocols deliver additional array gain and SNR, especially in large apertures, with SA optimal for moderate complexity and SM for maximal spectral efficiency.

Over-the-Air Computation (AirComp): SWAN reduces mean-squared error for analog function computation. Type-II SA (with segment phase-shifters) yields up to 30–40% further MSE reduction versus phase-free aggregation and far outperforms conventional PASS (Gu et al., 4 May 2026).

Integrated Sensing and Communications (ISAC): The flexible reallocation of segments to transmit (Tx) or receive (Rx) and precision PA control enable joint optimization for downlink beamforming and monostatic (echo) reception (Jiang et al., 8 Dec 2025). CRLB minimization for target position estimation delivers superior localization versus PASS and multi-waveguide benchmarks (Geng et al., 1 Apr 2026). Hybrid and learning-based reinforcement schemes (HSSM + SHRL) yield robust performance under hardware constraints (Gao et al., 28 Jan 2026).

Physical-layer security: Game-theoretic amplitude/phase PA coordination, optimized via Shapley value-based algorithms, significantly boosts secrecy rate by destructive combining at eavesdroppers—a capability directly arising from segment-level spatial resolution (Wang et al., 14 Jul 2025).

6. Learning-based and Data-driven Optimization

SWAN’s nonconvex, mixed-integer design space motivates deep learning solutions, particularly for ISAC regimes (Gao et al., 11 Apr 2026, Gao et al., 11 Apr 2026, Guo et al., 3 Dec 2025). Notable advances include:

• CSI-induced self-graph encoders for permutation-invariant extraction of user–target relationships.
• Transformer/LLM backbones with LoRA adaptation and split deployment/beamforming heads, enabling joint optimization of antenna positions, segment partitioning (Tx/Rx mode), and multi-beamforming weights.
• User-count transfer mechanisms decouple spatial deployment from specific user/target configurations, allowing near-zero retraining for dynamic deployments.
• GNN architectures (e.g., SWISAC-GNN) achieve near-optimal communication/sensing trade-off at real-time inference speeds, with natural handling of system-size variation and constraint enforcement.

Empirically, advanced learning frameworks achieve higher communication rates, better CRLBs, and vastly lower complexity than conventional AO/CVX, MLP, or vanilla transformer baselines.

7. Implementation, Limitations, and Research Directions

Implementation guidelines:

• Segment lengths and pinning sites are optimized to avoid grating lobes and mutual coupling; $x_m$1 is typical.
• Segment aggregation (SA) protocols require phase-shifter calibration and careful combinational design to ensure coherent sum.
• Hardware trade-offs between increased segmentation (better reliability, lower attenuation) and hardware cost/complexity (RF chains, switches) are core to practical instantiations of SWAN (Gao et al., 28 Jan 2026).

Open challenges:

• Precision actuator and MEMS for pinching antenna repositioning, particularly for rapid user tracking and 2D/3D waveguide layouts.
• Per-segment S-parameter measurement for real-time calibration.
• Rapid machine learning-driven adaptation to time-varying user/target scenarios, including federated or transfer learning across sites.
• Extensions to multi-mode, multiport, or T-junction-fed segmented waveguides for even higher spatial degrees of freedom (Liu et al., 26 Jan 2026).

SWAN thus constitutes a foundational architecture for flexible, reliable, and high-performance wireless arrays, underpinning advances in next-generation MIMO, ISAC, secure communications, and over-the-air computation (Xu et al., 30 Jun 2025, Jiang et al., 4 Mar 2026, Gao et al., 11 Apr 2026, Jiang et al., 8 Dec 2025, Ouyang et al., 12 Feb 2026).