Forecasting the Fading Field for Movable Antennas

Movable antennas are radio elements that a motor, pump, or sliding contact can reposition along a track. They improve a link by placing the antenna at a spot where the reflected copies of the signal add up instead of cancelling. The difficulty is timing. A motor needs milliseconds to seconds to move, while the signal pattern along the track changes much faster, so the best spot measured now is usually the wrong spot by the time the antenna arrives. Recent work handles this by training a neural network to predict where the antenna should go next. Those networks learn to imitate coordinates produced offline by a heuristic optimizer, and they do not see the signal pattern along the track that made those coordinates good.

We propose to change what is predicted. The network forecasts the future signal-strength map along the whole track, from the sparse samples the antennas have actually measured. A small exact planner then reads that map and chooses where to move, accounting for motor energy and minimum antenna spacing. The forecasting network is built around one physical fact. Along the track, fast fading has a correlation that depends only on the distance between two points, which is exactly the assumption under which convolution is the right operator. The slow components (user motion, path loss, blockage transitions) are not local in that sense and suit attention. A Conformer-style network pairs the two. Conformer is the speech architecture that interleaves short convolutions with global self-attention. Here it is used as a principled match to this physics, not as a drop-in replacement for an existing Transformer.

Core Claim

The quantity the network predicts is the future gain map $G_{t+\tau}(x)$. This is the channel gain at track position $x$, sampled on a sub-wavelength grid, at the time $t+\tau$ when a move started at slot $t$ would finish. Under the classical isotropic-scattering model, small-scale fading along a line satisfies

$$\mathbb{E}[h(x)\,h^*(x+\Delta)] \approx J_0(2\pi\Delta/\lambda),$$

where $h(x)$ is the complex channel at position $x$, $\Delta$ is the displacement, $\lambda$ is the carrier wavelength, and $J_0$ is the zeroth-order Bessel function. Spatially correlated fading models used for fluid antennas, such as Performance Analysis of Fluid Antenna System Under Spatially-Correlated Rician Fading Channels, rest on correlation functions of this kind.

Because the correlation depends only on $\Delta$, a convolution kernel whose width is about one correlation length is the natural local operator. Line-of-sight and blockage transitions, cluster births and deaths, and user trajectories break this stationarity at meter and multi-slot scales, and that is where attention earns its place.

The specific claim follows from this. On realistic 3GPP channels, a local-global map forecaster combined with an exact planner will reach a higher achieved rate after actuation delay, at equal motor energy, than two alternatives. The first is direct coordinate regression trained on optimizer labels. The second is the classical linear minimum-mean-square-error (LMMSE) predictor, meaning spatial interpolation with a Bessel kernel plus temporal Kalman prediction. We also predict that the margin over the linear predictor will appear only where the channel departs from the stationary Gaussian model. If that pattern fails, the claim fails.

Relation to Prior Work

Three papers motivate the problem by documenting the latency gap and the value of movement.

Both assume the channel is known at the moment of positioning. Our work exists to drop that assumption.

The closest learned systems predict coordinates directly.

  • Predictive Position Control for Movable Antenna Arrays in UAV Communications chains per-antenna LSTMs, cross-antenna attention, and a bidirectional LSTM head to forecast 3D coordinates up to 60 slots ahead. It reports at least 49% lower error than LSTM, NARX, and Transformer baselines.
  • RoleAware-MAPP decodes whole future coordinate sequences non-autoregressively with role-aware cross-attention between receiver and eavesdropper. It reports an average secrecy rate of 0.3569 bps/Hz and 81.52% strictly positive secrecy capacity on 28 GHz QuaDRiGa urban macro channels.

Both learn from coordinates found by particle swarm optimization, so their ceiling is that optimizer's quality. Both also enforce spacing through loss penalties or post-hoc correction. Our proposal differs in three ways. The supervision target is the simulated channel itself rather than optimizer output. The forecast is a map rather than a point. Constraints and energy are satisfied exactly by a planner rather than encouraged by a penalty.

The research file suggests that depthwise convolutions might "inherently encode" minimum-spacing constraints. We do not adopt this. A convolution limits which inputs interact; it does not bound the distance between outputs. We therefore keep constraint satisfaction out of the network.

Two other papers are adjacent rather than competing.

- Two-Timescale Joint Transmit and Pinching Beamforming separates slow pinch placement, solved by stochastic successive convex approximation, from fast precoding, solved by a Transformer predicting dual variables. It reports a 36.4% sum-rate gain over iterative WMMSE (weighted minimum mean-square error precoding). Its learned component never touches positioning, which is the part we learn.

Three speech papers supply the architecture.

  • Pushing the Limits of Non-Autoregressive Speech Recognition shows that Conformer encoders work well for single-pass, non-autoregressive prediction. It also warns that the largest models overfit small corpora, which constrains us to modest model sizes on simulated data.
  • Attention is All You Need in Speech Separation contributes the dual-path pattern: attention within short chunks plus attention across chunks. We use it for the coarse-scale branch along long tracks.
  • OWSM v3.1 shows that parallel convolution and attention branches (E-Branchformer) lowered multilingual error from 18.8% to 15.2% relative to a plain Transformer.

That last result is speech evidence only. We treat it as a hypothesis for radio channels, not as support.

Research Program

The setting is a base station with $K$ movable antennas on a one-dimensional track several meters long, discretized at $\lambda/8$, serving one or a few mobile users. The forecaster's input is a time-by-position tensor holding the channel samples observed at the positions the antennas actually occupied over the last $T$ slots. A mask marks which entries were observed.

Each block of the network contains three operators:

  • a depthwise convolution along position, with kernel width set to roughly one to two wavelengths;
  • a short depthwise convolution along time;
  • self-attention over time, plus inter-chunk attention across pooled track segments for the coarse scale.

The network is trained on full simulated maps and outputs $\hat G_{t+\tau}(x)$ for the actuation horizon $\tau$. An optional head outputs per-position uncertainty.

The planner then solves

$$\max_{x_1<\dots<x_K}\ \sum_{k=1}^{K}\hat G_{t+\tau}(x_k)\;-\;\mu\sum_{k=1}^{K}E(|x_k-x_k^{\text{now}}|)\quad \text{s.t.}\ x_{k+1}-x_k\ge \lambda/2.$$

Here $x_k$ is the target position of antenna $k$, $x_k^{\text{now}}$ its current position, $E(\cdot)$ the motor-energy cost from the flexible-position power model, and $\mu$ a price on energy. Because positions on a line are ordered, dynamic programming solves this exactly in time linear in $K$ and in the number of grid points. For a single user the sum of gains is the received-SNR objective of maximum-ratio combining. Multiuser and secrecy variants replace it with a per-position score map that the network predicts per user, and the planner then scores sets of positions.

We will study three things:

  1. how forecast error translates into achieved rate after the actual actuation delay;
  2. how the gain depends on convolution kernel width relative to $\lambda$;
  3. whether a light decision-aware fine-tuning stage helps beyond plain map regression. In that stage, the loss is the rate lost by the planner's choice.

Baselines

The naive reference keeps the antennas at the best position measured at slot $t$, ignoring the actuation delay. It shows how much the delay actually hurts. Next to it, we run a fixed dense array with the same number of radio chains doing electronic antenna selection. This tells us whether movement is worth it at all under the energy model.

The strongest classical competitor is the linear predictor. Spatial interpolation uses the Bessel or Rician correlation kernel fitted to the training data, and a Kalman filter predicts in time with the Doppler matched to the user's estimated speed. It feeds the same planner. It is the optimal linear predictor if the channel really were stationary and Gaussian, so it is the natural competitor to a learned local-global model. We find no reported result on Lacuna for this predictor on movable-antenna position control. We will implement it and tune its kernel and noise parameters on the validation split.

The strongest published learned competitors are RoleAware-MAPP and the Transformer-LSTM controller. We will re-implement both on our data, trained on particle-swarm labels generated as their papers describe. On the secrecy variant, we will reproduce RoleAware-MAPP's setup: QuaDRiGa urban macro, non-line-of-sight, 28 GHz, 43,200 samples, 10–100 km/h. Its reported 0.3569 bps/Hz average secrecy rate and 81.52% strictly positive secrecy capacity are the published reference points. Because reproductions can drift, the comparison of record is against our own re-run under identical channels. For the Transformer-LSTM controller, the file reports only relative improvements, so no absolute number found on Lacuna can serve as a target.

Fairness rules apply to all learned models. They see the same channel realizations and the same actuation delay. Their parameter counts are matched within 10%, and each gets the same hyperparameter search budget in trials and GPU-hours.

Three ablations isolate the key ingredients:

  • Representation. The same Conformer backbone regresses coordinates directly instead of forecasting the map. This is the "obvious" combination and tests whether the representation change matters.
  • Local operator. The convolution is replaced by attention (pure Transformer), and separately the attention is removed (pure convolutional U-Net style network).
  • Constraint handling. The planner is replaced by penalty-based constraint handling.

Beating the field means a statistically significant gain in rate after delay, measured over held-out drives with at least ten seeds, at equal or lower motor energy. The gain must hold against both the linear predictor and the best re-run coordinate regressor.

Why not just…?

Why not just use the linear interpolation-plus-Kalman predictor? This is the most serious threat, and we treat it as one. Under pure Rayleigh fading with the Bessel correlation, nothing should beat it meaningfully. Our first decisive check runs on exactly such synthetic channels. We expect a tie there, and a large win would point to a bug or leakage, not to a better model. The claim lives entirely on QuaDRiGa channels with cluster dynamics, line-of-sight/non-line-of-sight switching, and changing speed, where stationarity fails. If the learned model's advantage does not grow with the rate of such transitions, the honest conclusion is that classical prediction suffices, and we will report it.

Why not just swap RoleAware-MAPP's Transformer for a Conformer? Doing so keeps the particle-swarm label ceiling and the point-prediction bottleneck. The coordinate-regression ablation measures this directly with the identical backbone. If it matches the map-plus-planner system, the map representation is not pulling its weight.

Why not just use a plain Transformer or a plain convolutional network on the map? The speech evidence from OWSM v3.1 favors mixed operators, but that is a different signal. A mechanism test settles it here. If convolution is useful because fading is stationary, accuracy should peak when the kernel width is near the correlation length, and the gap to a pure Transformer should shrink as training data grows. A pure convolutional network should degrade mainly at long horizons and blockage transitions.

Why not just skip movement and use a dense fixed array with selection? Flexible-Position MIMO argues that movement wins on hardware count. It also shows that motor energy can erase the gain. Our energy-priced planner and the fixed-array baseline answer this per scenario rather than by assumption.

Proposed Experiments

Data. Data come from QuaDRiGa urban macro channels at 28 GHz with line-of-sight and non-line-of-sight mixes, plus a 3.5 GHz configuration in which the wavelength and thus the correlation length are longer. They also include synthetic correlated Rayleigh and Rician channels as the control. Train, validation, and test splits are by drive route, not by time slot, to avoid temporal leakage. Actuation delays range from 10 ms to 500 ms.

Metrics. We report:

  • map forecast error, as normalized mean square error;
  • rate or SNR achieved at the moment the move completes;
  • net energy efficiency including motor energy;
  • the fraction of slots in a deep fade;
  • for the secrecy variant, average secrecy rate and strictly positive secrecy capacity;
  • inference latency, which must stay well under the roughly 7–9 ms the coordinate regressors report so that it is not a bottleneck.

Decisive checks. Four results decide the claim:

  1. A tie with the linear predictor on stationary synthetic channels, and a gain that grows with non-stationarity on QuaDRiGa.
  2. Map-plus-planner beating coordinate regression with the same backbone, especially at long actuation delays, where coordinate imitation should suffer most.
  3. A kernel-width curve peaking near the correlation length, with the peak shifting between the 28 GHz and 3.5 GHz settings as the wavelength scaling predicts.
  4. A positive net energy-efficiency gain over the fixed array once motor energy is charged.

If the third check fails while the others pass, the system may still be useful. In that case the "physics-matched convolution" explanation would be wrong, and we would say so.