Fast Fluid Antenna Multiple Access

Technical Summary: Copula-Aided Fast Fluid Antenna Multiple Access

Problem Statement
Fast Fluid Antenna Multiple Access (FAMA) is a promising concept for massive connectivity that mitigates severe interference by reconfiguring the receiver's antenna position on a symbol-by-symbol basis. The core mechanism involves selecting the optimal port $k^\star$ where the aggregate interference $I_k$ naturally undergoes a deep fade, maximizing the instantaneous signal-to-interference ratio (SIR).

However, the practical implementation of fast FAMA faces a critical "genie-aided" bottleneck. The theoretical framework assumes that a User Terminal (UT) possesses:

1. Full Port Observability: Access to received signals $\{r_k\}$ and channel coefficients $\{h_k\}$ at all $K$ ports simultaneously.
2. Interference Knowledge: Instantaneous knowledge of the aggregate interference $\{I_k\}$ at all ports to compute the optimal selection metric $\gamma_k = |h_k|^2 / |I_k|^2$.

These assumptions are unrealistic due to hardware constraints (requiring $K$ RF chains or switching speeds exceeding the symbol rate), prohibitive pilot overhead for channel estimation, and the fundamental impossibility of directly observing data-dependent interference $I_k$ without knowing the interfering symbols. Previous attempts to relax these assumptions (e.g., slow FAMA or turbo FAMA) have either sacrificed multiplexing gains or retained high complexity and unrealistic observability requirements.

Methodology
This paper proposes a Copula-Aided FAMA framework that treats port selection as a probabilistic inference problem. The central hypothesis is that a UT can act as if it possesses full per-port interference knowledge by observing only a small fraction ($M \ll K$) of ports, leveraging the learned joint dependence structure of the channel and interference fields.

The proposed methodology involves a two-stage attentional copula time-series model:

1. Problem Formulation:
   The system models the joint statistics of the complex triplets $(r_k, h_k, I_k)$ across all $K$ ports. The goal is to learn the conditional distribution $p(z_{\text{miss}} | z_{\text{obs}})$, where $z_{\text{miss}}$ represents the unobserved ports and $z_{\text{obs}}$ represents the partial observations (typically $M$ ports where both $r_k$ and $h_k$ are known).

2. Model Architecture:

   • Port-Major Representation: The snapshot $(r, h, I)$ is encoded as a real-valued time series of length $T=3K$, separating real and imaginary parts.
   • Stage 1 (Marginal Modeling): Parametric univariate marginals for each scalar coordinate are fitted using Deep Sigmoidal Flows (DSF). This transforms the raw data into standard uniform variables, decoupling marginal distributions from dependence structures.
   • Stage 2 (Copula Modeling): A Transformer-based architecture parametrizes the conditional copula density. It utilizes self-attention over observed tokens (historical data) and cross-attention to predict unobserved tokens.
   • Training Strategy: The model is trained under random partial-observation masks. During training, the model is exposed to various subsets of observed ports ($M$) and must infer the missing values. This ensures the model learns to generalize to specific hardware constraints (e.g., a fixed number of RF chains) during inference.

3. Inference and Selection:
   At runtime, the UT observes $M$ ports. The model reconstructs the posterior distribution of the interference field $\{I_k\}$ and channels $\{h_k\}$ for all $K$ ports. The UT then selects the port $k^\star$ that maximizes the inferred instantaneous SIR, effectively mimicking an oracle with perfect knowledge.

Key Contributions

• Reformulation of Fast FAMA: The paper formalizes the fast FAMA port selection problem as a joint imputation task over complex-valued triplets $(r_k, h_k, I_k)$, moving away from the need for direct interference observation.
• Novel Learning Framework: It introduces an attention-copula time-series model specifically tailored to FAS interference fields. This framework explicitly models the joint distribution of channel, interference, and received signals, allowing for the inference of unobserved variables from partial data.
• Robustness Across Scenarios: The framework is evaluated under diverse conditions, including:
  • Geometries: Both 1D linear and 2D planar FAS configurations.
  • Channel Models: Rich scattering (Gaussian/Rayleigh) and finite scattering (Rician with limited paths).
  • Observation Constraints: Performance is analyzed with $M$ significantly smaller than the total ports $K$.

Simulation Results
Extensive simulations demonstrate the efficacy of the proposed approach:

• Reconstruction Accuracy: The Normalized Mean-Square Error (NMSE) for reconstructing $h$, $r$, and $I$ drops to the order of $10^{-4}$ once the number of observed ports $M$ exceeds the spatial degrees of freedom (DoF) of the aperture (approximately $2W$ for a 1D aperture of length $W$).
• System Performance: The Binary Symmetric Channel (BSC) sum-rate achieved by the proposed "Predicted" scheme closely matches the "Oracle" benchmark (which assumes perfect knowledge of all ports).
  • For 1D and 2D geometries, the performance gap vanishes once $M \approx 2W$.
  • The system maintains near-ideal performance even in finite-scattering environments and with a high number of interferers ($U$ up to 200).
• Scalability: The inference complexity is dominated by the number of observed ports $M$ rather than the total number of ports $K$, making the approach scalable to large FAS arrays.

Significance and Claims
The paper claims to bridge the gap between the theoretical potential of fast FAMA and practical hardware limitations. By demonstrating that a UT can achieve near-ideal fast FAMA performance with only a modest number of per-symbol sensing ports (e.g., $M \ll K$), the work substantially relaxes the need for "genie-aided" interference observability.

The authors argue that this approach restores the conceptual simplicity of fast FAMA—specifically, the ability to mitigate interference without precoding, power control, or Successive Interference Cancellation (SIC)—while remaining feasible under realistic switching and pilot overhead constraints. The results suggest that the rich spatial structure of FAS channels can be effectively exploited through probabilistic modeling to infer the interference field, enabling high-multiplexing massive connectivity in future 6G networks.