A fluid antenna system can choose a small set of radiating ports from a much denser spatial grid. The apparent advantage creates a circular problem: selecting the best ports seems to require channel state information for ports that were never measured, while measuring every candidate can erase the intended pilot savings.
A new public preprint treats port selection and pilot allocation as one inference problem. Instead of assuming a complete channel map, the system maintains a probabilistic belief over all candidate ports and decides which uncertainty is worth resolving.
Spatial and temporal structure fill the unseen grid
The proposed generative model combines sub-wavelength spatial correlation with an autoregressive temporal prior. Together they maintain a Gaussian belief over the channels at all candidate ports, including those not sounded in the current slot.
This separates uncertainty from absence. An unmeasured port is not assigned a fabricated exact channel; it receives a distribution whose confidence depends on nearby measurements and recent history. The next pilot can then be directed toward a port where information has decision value.
One objective chooses service and measurement
The transmitter uses active inference to choose both the ports it serves and the subset that receives pilots. Its expected-free-energy objective trades achievable rate against information gain and port-switching cost. The public abstract describes a greedy, submodular selection procedure with per-slot complexity proportional to the product of the candidate and selected-port dimensions.
In the reported simulation on a 441-port grid with a hybrid front end, the method reaches 91% of a full-CSI reference sum rate while measuring 2.3% of candidate ports. It also reports that pilots can be withdrawn from one fifth of still-served ports for a 2% sum-rate reduction, or from two fifths for an 8% reduction. The method does not require an offline training stage.
The belief model is also the main deployment risk
The result shows how sensing effort can become part of antenna control rather than a fixed prerequisite. Its practical value depends on whether the assumed spatial correlation and temporal dynamics remain calibrated under motion, blockage, coupling, switching delay, and hardware drift.
The public evidence is numerical. It does not establish over-the-air behavior, robustness to belief-model mismatch, or the control latency of a physical large-port array.
Research notes
Active Inference for Joint Port Selection and Pilot Allocation in Fluid Antenna Systems Under Partial CSI
Authors: Kian Fotovat, Kamran Fotovat, and Zijun Wang.
Status: Public arXiv record dated 2 September 2026.
What the public evidence establishes: The work maintains a spatiotemporal Gaussian belief over fluid-antenna port channels and jointly selects served and piloted ports by balancing rate, information gain, and switching cost.
Limits: The reported results are simulation-based and do not establish measured-hardware performance, model-mismatch robustness, or physical switching latency.