Pinching Antenna Systems (PASS): Principles, Positioning and Opportunities

• PASS guides electromagnetic waves through waveguides for nearby radiation, addressing large-scale fading in enclosed environments such as underground facilities and subways.Source: §I · p. 2
• The survey centers on waveguide and geometry design and an intrinsic positioning–communication–navigation loop linking uplink positioning, downlink communication, and navigation feedback.Source: §I · p. 3
• The waveguide review compares materials and geometries, modal behavior, field confinement, and attenuation, and distinguishes PASS from LCX using a unified channel model.It highlights frequency scalability and negligible coupling loss.Source: §I · p. 3
• The positioning review covers BS-side and user-side methods, single- and multi-waveguide architectures, LoS and multipath conditions, and RSSI, phase, and delay observables.It identifies multi-user scheduling, robust LoS extraction, and lightweight algorithms as research directions.Source: §I · p. 3
• The performance analysis spans uplink and downlink, LoS blockage, and mixed LoS/NLoS channels, with probabilistic and geometry-aware blockage models differing in tractability and physical interpretability.Source: §I · p. 3
• The survey reviews PA positioning, beamforming, energy efficiency, and QoS-aware optimization, alongside AI approaches including deep learning, reinforcement learning, federated learning, meta-learning, and LLMs.It also compares AI design principles and solution paradigms and identifies future challenges.Source: §I · p. 3, C. Distinction Between PASS and Other… · p. 3

A. Historical Evolution of Waveguides Toward PASS

• Dielectric waveguides’ accessible fields enabled radiation alongside guided transmission, leading to PASS’s use of waveguides as low-loss media supporting localized radiating elements.The transition runs from dielectric structures interacting with their surroundings to waveguide-fed radiating architectures and PASS.Source: A. Historical Evolution of Waveguides… · pp. 4–5

B. Effects of Waveguide Materials and Geometries on Electromagnetic Performance

• Uniform-waveguide fields are analyzed by separation of variables into a transverse distribution and an axial factor e^(-(α+jβ)z), with α denoting attenuation and β phase.The transverse eigenvalue is a cutoff wavenumber in metallic guides and a transverse phase constant in dielectric guides.Source: B. Effects of Waveguide Materials and… · p. 5
• Material and geometry affect modal behavior, attenuation, field confinement, and external-field access; the survey compares representative waveguide structures across these characteristics.The paper also establishes a unified channel model and quantitative comparison to distinguish PASS from LCX and other flexible systems.Source: B. Effects of Waveguide Materials and… · p. 5, Table I · p. 6

1) Circular Hollow Metallic Waveguides:

• Circular hollow metallic waveguides provide strong confinement and shielding but cannot support PA power extraction, making them unsuitable as PASS waveguides.Their closed, rigid walls restrict geometrical flexibility and interaction between guided fields and external space; they may still serve as shielded feeders or high-power transmission sections.Source: 1) Circular Hollow Metallic Waveguides · p. 6

2) Rectangular Hollow Metallic Waveguides:

• Rectangular hollow metallic waveguides support stable dominant-mode operation and low radiation leakage, but their closed walls prevent efficient PA power extraction in PASS.Their geometry is set by broad and narrow wall dimensions, which determine discrete cutoff characteristics.Source: 2) Rectangular Hollow Metallic… · p. 6

3) Dielectric Rod Waveguides:

• Dielectric rod waveguides use dielectric contrast and modal continuity to guide low-loss signals while retaining an accessible evanescent field for near-field PA coupling.Sensitivity to perturbations, bending, and surface contamination may require mechanical support and alignment.Source: 3) Dielectric Rod Waveguides · pp. 6–7

4) Rectangular Dielectric Waveguides:

• Rectangular dielectric guides offer low-loss propagation and localized PA coupling, but confinement trades off against evanescent-field accessibility, and excessive openness can increase leakage and scattering loss.The Marcatili approximation simplifies analysis by neglecting weak fields in the four corner regions.Source: 4) Rectangular Dielectric Waveguides · p. 7

Optical fibers

• Optical-fiber cladding stabilizes guided modes and confines most of the field, weakening external interaction needed for PASS radiation.Practical fiber attenuation can also arise from material absorption, Rayleigh scattering, bending, and fabrication imperfections.Source: 4) Rectangular Dielectric Waveguides · p. 7

5) Optical Fiber Waveguides:

• Optical-fiber cladding confines the evanescent field, weakening external interaction and limiting movable-PA power extraction; bending and fabrication imperfections further reduce suitability as a PASS waveguide.The fiber may still serve as a stable signal-delivery link in PASS-enabled networks.Source: 5) Optical Fiber Waveguides · p. 8

6) Dielectric Ribbon Waveguides:

• Dielectric ribbon waveguides’ broad, open interfaces enable external coupling, but stronger confinement reduces leakage while limiting accessibility, creating a tunable guidance–radiation trade-off.Surface roughness, perturbations, bending leakage, and radiation leakage complicate deployment; further investigation remains necessary.Source: 6) Dielectric Ribbon Waveguides · p. 8

C. Distinction Between PASS and Other Flexible Communication Architectures

• PASS uses detachable or movable PAs for reconfigurable radiation along dielectric waveguides, whereas LCX radiates through fixed, prefabricated slots or apertures.PASS combines guided transport with free-space radiation at reconfigurable points.Source: C. Distinction Between PASS and Other… · p. 8

1) Unified Model:

• The unified channel model distinguishes PASS and LCX through radiation factor and attenuation: PASS uses near-field PA coupling, while LCX slot gain depends on user direction.LCX slot coupling loss is reported at 56–77 dB, whereas PASS coupling efficiency is engineered near unity with coupling loss below 1 dB.Source: 1) Unified Model · pp. 8–9

2) Comparison:

• PASS reduces coupling loss at lower frequencies and retains scalability at higher frequencies, while LCX becomes practically unusable at millimeter-wave and terahertz frequencies.PASS’s material flexibility supports low propagation attenuation into the hundreds of gigahertz; the comparison is summarized in Table II.Source: 2) Comparison · p. 9

III. PASS-BASED POSITIONING

• Reliable user positions guide PA selection or relocation, and estimated positions can support waveguide selection, beamforming, and communication optimization.PASS’s meter-scale reconfigurability makes the location question central to choosing where PAs should be moved or activated.Source: §III · p. 9

A. Single-Waveguide BS-side Positioning

• In single-waveguide BS-side positioning, known PA locations serve as anchors for estimating users from uplink pilots; sequential or grouped activation supports a low-RF-chain-complexity architecture.The model uses a ceiling-edge, wall-edge, or corridor-boundary waveguide and a TDMA scheme with one active PA per time slot; Fig. 10 depicts the system.Source: A. Single-Waveguide BS-side Positioning · p. 10, Figure 10 · p. 10

1) RSSI-Based localization:

• PASS-specific RSSI WLS provides reliable positioning across noise powers, but more activated PAs can help at high SNR and introduce unreliable measurements at low SNR.The review recommends adaptive PA selection; Fig. 11 presents positioning errors under different PA counts.Source: Figure 11 · p. 10, 1) RSSI-Based localization · p. 11

2) Phase-Aware Localization:

• Phase-aware maximum-likelihood localization jointly uses amplitude and phase, with coarse grid initialization followed by Levenberg–Marquardt refinement; phase wrapping, synchronization, calibration, and user motion can degrade performance.RSSI WLS can supply a coarse estimate for refinement in a smaller candidate region.Source: 2) Phase-Aware Localization · p. 11

3) OMP-Based Geometry-Consistent Localization:

• OMP-GCL uses compressed measurements from PA subarrays and fuses direction-related estimates through geometry consistency to infer user and scatterer locations.The approach represents near-field channels with a distance-parameterized angular dictionary and sparse recovery.Source: 3) OMP-Based Geometry-Consistent… · p. 11

1) LoS Environments:

• MWSP uses parallel RSSI observations from non-collinear anchors for low-complexity, interpretable LoS localization, while MWMP’s superposed-power grid search trades increased signal power for possible spatial ambiguity.MWSP accuracy depends on ranging quality, anchor geometry, and user location; Fig. 12 presents its positioning error in the stated MWSP scenario.Source: 1) LoS Environments · pp. 11–12, Figure 12 · p. 12

2) Multipath Environments:

• Multipath positioning either jointly estimates channel and location or extracts LoS delays before geometric ranging; MP-based ranging favors accuracy at higher decomposition cost, while Rank-1 ranging is lighter but can retain multipath bias.BP-VI can estimate off-grid delays more accurately than grid-based OMP, but its iterative inference increases computational complexity.Source: 2) Multipath Environments · pp. 12–13
• Multi-waveguide JCEL gathers observations from diverse directions to alleviate single-waveguide half-space ambiguity; wideband delay-based methods match extracted delays to PA anchors before solving for position.BP-VI models delays as continuous latent variables, while OMP-based delay extraction uses a discretized delay dictionary.Source: 2) Multipath Environments · pp. 12–13

C. User-side Positioning

• User-side positioning uses downlink broadcasts with unknown PA coordinates and must accommodate terminals’ limited computational and hardware capabilities.This distributed self-localization approach can provide service after deployment without site surveys, manual coordinate calibration, or continuous network-side anchor-coordinate broadcasts.Source: C. User-side Positioning · pp. 13–14

1) Single-Waveguide User-side Positioning:

• Anchor-line positioning combines RSSI and ToA through the Lambert W function to estimate PA positions and pseudoranges, then uses WLS to estimate the user position.The method addresses the coupled unknown AP–PA in-waveguide and PA–user free-space distances with heterogeneous measurements.Source: 1) Single-Waveguide User-side… · p. 14
• Within breakpoint distance d0 = λ/(π tan δ), the Lambert W solution exists and is unique; beyond it, nonlinearity sharply amplifies measurement errors.A Gauss-Markov-derived weight matrix down-weights less reliable measurements from far-end PAs.Source: 1) Single-Waveguide User-side… · p. 14
• The anchor-line framework achieves centimeter-level accuracy within its working range, whose extent depends on PA spacing.It is described as the most mature single-waveguide user-side solution, with plug-and-play deployment and closed-form solvability, but it relies on both power and timing measurements.Source: 1) Single-Waveguide User-side… · p. 14, Figure 13 · p. 14

2) Multi-Waveguide User-side Positioning:

• A PLP models waveguide locations and a one-dimensional HPPP models PA locations; analysis yields localizability and RSSI-based CRLB results for dense random deployments.The localizability expression uses a dominant-interference approximation, while the CRLB approximation links waveguide and PA density, channel parameters, and probabilistic localization accuracy.Source: 2) Multi-Waveguide User-side Positioning · p. 14
• Multi-waveguide localization has a non-monotonic design trade-off: added waveguides first improve accuracy through geometric diversity, but interference can degrade it beyond an optimum.Compared with conventional uniform linear arrays, PASS localization achieves a lower CRLB.Source: 2) Multi-Waveguide User-side Positioning · p. 14

D. PASS-based Positioning Outlook

• Positioning research still needs methods that extend reliable range and identifiability, handle multipath, support multiple users with limited resources, and adapt to PASS reconfiguration.The survey also calls for downlink models that jointly characterize electromagnetic behavior, propagation loss, coupling, PA activation, and multi-user interference.Source: D. PASS-based Positioning Outlook · pp. 14–15, A. Downlink PASS · p. 16

A. Downlink PASS

• PASS consistently outperforms CASS in ergodic rate across the considered settings, while its rate depends on area width because PAs can be repositioned to match user locations.For CASS, enlarging the service area in either length or width degrades ergodic rate; PASS performance also depends on resource budget, transmit power, user distribution, and area size.Source: A. Downlink PASS · p. 15, Figure 14 · p. 15
• In multiuser downlink PASS, OMA benefits the strong user, whereas NOMA improves the weak user and fairness; neither scheme is universally superior across multi-PA settings.The comparison’s outcome depends on parameters including user count and power allocation.Source: A. Downlink PASS · p. 16
• PASS array gain is not unlimited: excessive PAs can reduce per-PA power and add weak far-away contributions, yielding an optimal PA count and spacing.With a fixed PA count, multi-waveguide deployment achieved higher spectrum efficiency but lower energy efficiency than multiple PAs on one waveguide.Source: A. Downlink PASS · p. 16

B. Uplink PASS

• Simultaneous multi-PA uplink on one waveguide can cause inter-antenna radiation: leaked signals reduce power delivered to the BS and interfere with other UEs.The uncertainty in propagation direction within the waveguide also complicates the signal model and multi-PA uplink design.Source: B. Uplink PASS · p. 16, Figure 15 · p. 16
• Existing approaches avoid this uplink effect through single-PA activation per waveguide, or segmented waveguides that require more feed points, RF-chain complexity, and deployment cost.Asynchronous uplink/downlink across different waveguides also requires multi-waveguide deployment and mode scheduling; both approaches require UE-aware activation and scheduling.Source: B. Uplink PASS · p. 16

C. PASS Under LoS Blockage

• LoS blockage reduces achievable gain, and increasing transmit power does not fully compensate; degradation becomes more pronounced as obstacle density rises.The cited analysis describes blockage impacts in outdoor and indoor environments, including at mmWave frequencies.Source: C. PASS Under LoS Blockage · p. 17, Figure 18 · p. 18

1) Probabilistic LoS blockage model:

• Probabilistic blockage models use distance-dependent LoS probability for tractable system-level analysis, but empirical fits lack explicit obstacle geometry and are poorly suited to link-level characterization or PA activation design.Their fitting parameters do not clearly map to obstacle locations, sizes, or shapes.Source: 1) Probabilistic LoS blockage model · p. 17

2) Geometry-aware blockage model:

• Geometry-aware models represent obstacle distributions explicitly, but must balance fidelity with tractability and often depend on detailed a priori obstacle maps.Dynamic LoS/NLoS state changes can make associated positioning or beamforming problems non-convex and non-smooth for gradient-based solvers.Source: 2) Geometry-aware blockage model · pp. 17–18

D. PASS Under LoS and NLoS Channels

• NLoS propagation becomes important when LoS paths are blocked, particularly in dense-obstacle or rich-multipath settings; the channel model separates waveguide gain from LoS and NLoS components.The LoS indicator selects a Rayleigh NLoS channel when absent and a Rician LoS-dominated channel when present.Source: D. PASS Under LoS and NLoS Channels · p. 18
• With weak NLoS power, Rician ergodic rate closely follows ideal LoS; stronger NLoS power widens their rate gap, while waveguide and propagation losses yield a unique optimal PA position.The Rayleigh channel shows more distinct rate variation with PA position due to in-waveguide attenuation.Source: Figure 19 · p. 18, D. PASS Under LoS and NLoS Channels · p. 19

E. Discussion and Outlook

• Uplink leakage from signals re-radiated by other activated PAs motivates UE-aware scheduling that jointly addresses leakage, hardware complexity, and deployment flexibility.Current approaches use one PA per uplink waveguide or segment and waveguide-level scheduling, which does not adapt to user distribution or traffic demand.Source: E. Discussion and Outlook • · p. 19
• NLoS scattering can perturb phases and weaken coherent combining, with phase alignment especially fragile under obstacle-rich Rayleigh fading; the multi-PA effect remains largely unexplored.Under Rician channels, a deterministic LoS component provides a stable phase reference, so random perturbations may only partially degrade combining.Source: E. Discussion and Outlook • · p. 19
• Multiuser, multi-waveguide PASS performance depends jointly on deployment, PA activation, and power allocation, with tractable analysis complicated by phase shifts, path loss, and interference.Future work is proposed to characterize outage probability, achievable rate, spectral efficiency, and energy efficiency across SNR regimes and beamforming strategies.Source: E. Discussion and Outlook • · p. 19
• Blockage-aware design remains challenging because obstacle knowledge is often unavailable and LoS/NLoS states change with moving obstacles, users, and PA repositioning.The passage identifies blockage prediction and environment tracking as essential, with AI-driven techniques described as promising.Source: E. Discussion and Outlook • · p. 19

V. RESOURCE ALLOCATION FOR PASS

• PASS resource allocation reviews position, beamforming, and energy-efficiency/QoS optimization, moving from idealized single-user models toward constrained multiuser deployments.PA relocation adds a spatial degree of freedom, making PA position a design dimension that governs channel gains, interference, and communication performance.Source: §V · p. 19

1) Continuous Position Optimization:

• Continuous position optimization yields a closed-form single-user, single-PA solution at the user’s projection onto the waveguide, subject to aperture boundaries.For multiuser, multi-PA sum-rate optimization, penalty-based alternating optimization uses a low-complexity one-dimensional position search to achieve near-optimal performance with minimal overhead.Source: 1) Continuous Position Optimization · p. 20

2) Discrete Antenna Activation:

• Discrete PA activation selects an active subset from a candidate grid, turning placement into a combinatorial problem whose complexity grows exponentially with grid size.Reported approaches use matching or low-complexity designs, and denser grids asymptotically recover continuous-placement performance.Source: 2) Discrete Antenna Activation · p. 20

3) Position Optimization Under Practical Constraints:

• Practical placement frameworks address minimum PA spacing, waveguide attenuation, and location uncertainty using projection, attenuation-aware modeling, and robust optimization.The surveyed methods include projected gradient ascent for spacing constraints and convex-program and block-coordinate approaches for uncertain user locations.Source: 3) Position Optimization Under… · p. 20

B. Transmit Beamforming Optimization

• PA positioning shapes large-scale channel gains while beamforming controls fine-grained signal distribution and interference, so their coupled design creates a highly non-convex sum-rate problem.The review describes joint frameworks rather than treating position and beamforming as independent decisions.Source: B. Transmit Beamforming Optimization · p. 20, 1) Foundational Multi-User Precoding · p. 20

1) Foundational Multi-User Precoding:

• Foundational multiuser precoding uses alternating optimization to split joint position and beamforming design into tractable subproblems, with extensions to multiwaveguide, MIMO, and element-wise methods.The surveyed approaches offer performance-complexity tradeoffs, including MRT, ZF, and MMSE baseband beamformers.Source: 1) Foundational Multi-User Precoding · pp. 20–21

2) Beamforming for Advanced PASS Architectures:

• Advanced PASS architectures introduce tailored beamforming designs, including center-fed and segmented arrays, hybrid PASS, and polarization-aware multimode modeling.The survey notes that architecture-specific algorithms fragment the methodological landscape and impede a unified framework.Source: 2) Beamforming for Advanced PASS… · p. 21

C. EE and QoS-Aware Optimization

• Energy- and QoS-aware designs account for PA repositioning’s mechanical energy and explore SE–EE tradeoffs and QoS constraints, but a unified framework offering both EE optimality and QoS guarantees remains elusive.The mechanical cost scales with movement frequency and distance; surveyed work includes joint power, position, and beamforming optimization.Source: C. EE and QoS-Aware Optimization · pp. 21–22

D. PASS Optimization Techniques Outlook

• Current PASS optimization often handles only one or two practical constraints, while dynamic users, blockages, and traffic require low-latency reconfiguration under delayed and imperfect CSI.The review calls for modular constraint handling and algorithms that jointly schedule PA movement and communication parameters.Source: D. PASS Optimization Techniques Outlook · p. 22

VI. SYNERGISTIC INTEGRATION OF AI AND PASS

• Joint optimization of beamforming, PA positions, power, and scheduling is high-dimensional and non-convex, while AI can learn configuration mappings for offline training and faster online inference.The paper also notes structural priors such as permutation equivariance and graph topology as ways to support generalization across varying system sizes.Source: §VI · p. 22

A. DL for PASS

• PASS deep-learning systems process environmental information into AI control policies for beamforming, PA placement, and resource allocation.The survey groups DL by architecture—including GNNs, transformers, CNNs, and DNNs—and by supervised, unsupervised, or generative training paradigms.Source: A. DL for PASS · pp. 22–23

1) GNNs for PASS:

• GNNs exploit PASS graph structure to learn placement and beamforming relationships, with message passing enabling permutation-invariant aggregation across different graph sizes.Examples include two-stage GPASS and bipartite BGAT; the review notes that local-neighborhood aggregation can make long-range interactions harder to capture without stacking layers.Source: 1) GNNs for PASS · p. 23

2) Transformers for PASS:

• Transformers use self-attention to model long-range dependencies among PAs and users in one layer, addressing a limitation of shallow message-passing GNNs.The survey reports that transformer-based methods outperform conventional black-box learning approaches and that KKT-guided learning balances performance and computational speed.Source: 2) Transformers for PASS · pp. 23–24

3) Other DL Models and Learning Paradigms:

• Other learning approaches match model structure to the task: CNNs handle grid-structured channel data, DNNs address moderate-dimensional problems, and generative models support underdetermined channel estimation.The survey also describes unsupervised capacity optimization and reports that a meta-learning strategy for coordinated PASS achieves a higher achievable-rate upper bound than conventional CoMP systems.Source: 3) Other DL Models and Learning… · pp. 24–25

B. RL for PASS

• Reinforcement learning maps observed environmental states to control actions and can account for long-term rewards and the different timescales of PA movement and beamforming.Surveyed applications include trajectory tracking, end-to-end placement and beamforming, NOMA MEC task offloading, and wirelessly powered PASS resource control.Source: B. RL for PASS · p. 25

C. FL for PASS

• PASS federated-learning designs progress from joint placement and client selection to straggler mitigation, over-the-air aggregation, and latency–quality optimization.The surveyed work also analyzes how PASS can shorten worst-link distance in synchronous FL and increase on-time completion probability in asynchronous FL.Source: C. FL for PASS · p. 25

D. Meta-Learning for PASS

• Meta-learning uses a few-step inner-loop adaptation and cross-task outer-loop training to address deployment shifts, while reducing—but not eliminating—the sim-to-real performance gap.Reviewed applications include GML-JO for adapting to new user distributions and a controller that adapts antenna positions and transmit power to eavesdropper mobility under security constraints,.Source: D. Meta-Learning for PASS · p. 26

E. LLMs for PASS

• LLM-based PASS work uses environmental context for beam training and system configuration, with graph-based designs aimed at transfer across changing user configurations.A reviewed framework prioritizes beam probes while maintaining alignment accuracy; SWAN designs use CSI-induced self-graphs and lightweight adaptation rather than full retraining –.Source: E. LLMs for PASS · p. 26

F. Open Challenges and Future Directions

• Open challenges include scarce real-world data, sim-to-real gaps, domain shifts, physical-constraint integration, reliable operation under mobility and partial CSI, and uncharacterized learning guarantees.The review also calls for lightweight real-time models and a theoretical framework covering sample complexity, approximation error, and optimality gap.Source: F. Open Challenges and Future Directions · p. 26

VII. CONCLUSIONS

• The survey covers PASS foundations, positioning, uplink and downlink performance, blockage models, resource allocation, and AI, and identifies field trials as essential for validating models and algorithms.Its conclusion frames PASS as a guided-wave infrastructure unifying positioning, communication, and navigation, while noting that research remains at an early stage.Source: §VII · pp. 26–27