AI-Limited Fluid Antenna-Aided Integrated Sensing and Communication Systems

This paper characterizes the fundamental limits of integrated sensing and communication (ISAC) when the transmitter is subject to an artificial intelligence (AI) representation bottleneck and the receiver employs a fluid antenna system (FAS). Specifically, the message is first encoded into an ideal Gaussian waveform and mapped by an AI encoder into a finite-capacity latent representation that constitutes the physical channel input, while the FAS receiver selects the port experiencing the most favorable channel conditions. We reveal that the AI bottleneck is equivalent to an additive representation noise, which reduces both the communication and sensing signal-to-noise ratios (SNRs) at the selected port. We then derive the resulting ISAC capacitydistortion region and establish tight converse and achievability bounds under general fading models, including Jakes-correlated channels. Leveraging the spatial degrees of freedom (DoF) characterization of the Jakes’ model, we furthermore prove that the port-selection gain is fundamentally constrained by the physical length of the FAS region: the effective diversity order equals the numerical rank of the Jakes’ correlation matrix and increases only with the FAS length. Consequently, enlarging the FAS length allows the selected-port SNR to approach the AI-imposed ceiling, driving the achievable communication rate and sensing mean square error (MSE) toward their AI-limited fundamental bounds. Numerical results corroborate the analysis and scaling laws.

💡 Research Summary

This paper investigates the fundamental limits of integrated sensing and communication (ISAC) systems when the transmitter is constrained by an artificial‑intelligence (AI) representation bottleneck and the receiver employs a fluid‑antenna system (FAS) with port selection. The authors model the transmitter as first generating an ideal Gaussian symbol X∼𝒞𝒩(0,P) and then passing it through an AI encoder that must satisfy an information‑budget constraint I(X;Z)≤C_AI. Under a Gaussian representation model, the encoder output is Z = X + W_z, where W_z∼𝒞𝒩(0,N_z) is independent of X. Enforcing the bottleneck with equality yields the minimum representation‑noise variance N*_z = P/(2^{C_AI}−1), showing that the AI bottleneck is equivalent to an additive distortion term that degrades both communication and sensing SNRs.

The receiver consists of L closely spaced ports of a fluid antenna. Each port ℓ experiences a complex communication gain h_{c,ℓ} and a sensing gain h_{s,ℓ}. The receiver selects the “best” port according to a utility function (typically the largest instantaneous communication gain). The selected port’s effective communication SNR is Γ*c = P|h{c,ℓ*}|^2/(N_0+N_z) and the sensing SNR is Γ*s = P|h{s,ℓ*}|^2/(N_0+N_z), where N_0 is thermal noise. Consequently, the AI‑induced noise N_z appears in both SNR expressions, reducing the achievable rate and increasing the mean‑square error (MSE) of the sensed parameter.

Using this equivalence, the authors derive the exact capacity–distortion region for the AI‑constrained FAS‑assisted ISAC system under Rayleigh fading. The achievable communication rate is C = log₂(1+Γ*_c) and the sensing distortion (MSE) is D_s = σ_θ^2/(1+Γ*_s). Matching converse and achievability proofs are provided, establishing that these expressions are tight.

A major contribution is the analysis of spatial degrees of freedom (DoF) provided by the fluid antenna under the Jakes correlation model. The correlation matrix R of the L ports has a numerical rank r_eff ≈ ⌊2πW/λ⌋, where W is the physical length of the fluid antenna and λ is the carrier wavelength. This rank, not the number of ports, determines the effective diversity order of the port‑selection process. Hence, the port‑selection gain is fundamentally limited by the antenna length; increasing L without increasing W yields diminishing returns.

Importantly, the paper shows that enlarging the fluid‑antenna length allows the selected‑port SNR to approach the AI‑imposed ceiling Γ_AI = P/(N_0+N*_z). In this regime, the communication rate converges to the AI capacity C_AI = log₂(1+P/N*_z) and the sensing MSE converges to the AI‑limited lower bound σ_θ^2/(1+Γ_AI). Thus, spatial diversity from a sufficiently long fluid antenna can fully compensate for the information loss caused by the AI bottleneck.

To bridge theory and practice, the authors propose a variational information bottleneck (VIB) neural network encoder that approximates the Gaussian representation model while respecting the information budget C_AI. Simulation results confirm that the VIB encoder attains performance close to the derived theoretical limits.

Numerical experiments explore various configurations of AI capacity, antenna length, and number of ports. The results validate the scaling laws: (i) the port‑selection gain scales with W, (ii) the AI bottleneck sets an ultimate SNR ceiling, and (iii) the capacity–distortion trade‑off is jointly governed by AI information constraints and spatial DoF.

In summary, the paper provides a rigorous information‑theoretic framework that (1) models AI‑induced representation constraints as additive noise, (2) quantifies the limited diversity gain of fluid antennas via the rank of the Jakes correlation matrix, and (3) demonstrates that extending the fluid‑antenna length enables the system to reach the AI‑limited performance ceiling. These insights are valuable for the design of future 6G ISAC systems that integrate AI‑driven transceivers with reconfigurable antenna hardware.