PaperlayerView exact revision on arXiv

Source-linked AI summary

Beam-STeP: Generative Site-Specific Beam Prediction with Spatio-Temporal Priors

Ziyue Wang, Zongyao Zhao, Zhaolin Wang, Yuanwei Liu

arXiv:2610.20282v1eess.SP

TL;DR

Beam prediction must handle multiple plausible future beams from the same history while limiting probing overhead. Beam-STeP combines site-specific spatial and mobility priors with historical beam sequences in a diffusion model, and simulations report stronger baseline performance across three ray-tracing scenarios with robust multi-step prediction.

  • Problem

    A single beam history can correspond to multiple future beam modes, and deterministic MSE prediction can average incompatible modes and reduce achievable beamforming gain.

    Source: §I · p. 2

  • Method

    Beam-STeP learns a site-specific, history-conditioned future-beam distribution using spatial and temporal priors, then uses DDIM to generate multiple candidates for probing and selection.

    Source: §I · p. 2, §III · p. 4

  • Results

    Across three ray-tracing scenarios, Beam-STeP generally outperforms the considered baselines and maintains robust multi-step prediction performance, with larger probing budgets mitigating degradation over recursive horizons.

    Source: §V · p. 13

  • Takeaways & Limitations

    In the Philadelphia scenario, K = 10 cuts candidate-generation latency by 78.1% versus K = 50, with only a 0.13 dB gain reduction.

    Source: F. Online Inference Latency Analysis · p. 12

Abstract

from arXiv · show

A generative site-specific beam prediction (Beam-STeP) method is proposed. Spatial priors from site-specific propagation and mobility information are jointly exploited with temporal priors from historical beam observations for beam prediction. To address the beam prediction ambiguity from multimodal spatio-temporal priors, the generative model is adopted, where a history-conditioned diffusion model is proposed to learn the multimodal conditional distribution of future beamforming vectors. Multiple beam candidates are sampled from the distribution using a denoising diffusion implicit model (DDIM). The candidate with the highest received power is then selected for transmission. Simulation results across three urban scenarios demonstrate that: i) Beam-STeP generally achieves higher normalized beamforming gains than the baselines under the same probing budget, particularly as beam prediction progresses; ii) site-specific propagation and mobility priors improve recursive beam prediction performance; and iii) DDIM sampling enables a flexible gain-latency trade-off.

I. INTRODUCTION

Narrow beams are vulnerable to mobility, while historical-beam prediction can reduce the overhead of repeated beam training and CSI acquisition. Beam-STeP combines site-specific spatial and mobility priors with beam history in a generative model to predict plausible future beams.

I. INTRODUCTION

  • Beam tracking from historical observations can reduce beam-training and CSI-acquisition overhead compared with repeated exhaustive sweeping or instantaneous-CSI redesign.Prediction also permits inference during data transmission, relaxing inference-time constraints.Source: §I · pp. 1–2
  • Multimodal future beams can make deterministic MSE predictors average incompatible beam modes, reducing achievable beamforming gain.Mobility alternatives and uncertain propagation states can produce different future beamformers for the same history.Source: §I · pp. 1–2
  • Existing site-specific generative beam methods rely on instantaneous probing and incur probing needs and inference time, motivating history-conditioned prediction.Beam-STeP learns future-beam distributions from historical observations reflecting channel evolution and geometry-constrained mobility.Source: §I · p. 2
  • Beam-STeP jointly uses spatial priors from site-specific beam data and temporal priors from ordered historical beam vectors, then generates multiple candidates for RSRP-based selection.The history-conditioned diffusion model addresses multimodal future-beam evolution, and DDIM generates candidates with fewer denoising steps than conventional DDPM sampling.Source: §I · p. 2
  • Simulations across three urban scenarios show improved prediction performance over the deterministic ODE-LSTM baseline, with benefits from continuous beam generation and site-specific training.The evaluation incorporates a road-aware mobility model and different probing budgets and site-specific configurations.Source: §I · p. 3

A. Site-Specific Downlink Channel

The downlink model represents the site-specific channel through propagation paths and a normalized array response. Analog beamforming is constrained by the hardware, and predicting candidates can avoid explicit estimation of the full instantaneous channel.

A. Site-Specific Downlink Channel

  • The site-specific channel is described by propagation paths whose power, phase, and departure angle vary over time with UE motion and whose distribution is governed by the deployment site.The site is characterized by BS placement and orientation, surrounding geometry and materials, and road topology.Source: A. Site-Specific Downlink Channel · p. 3
  • Analog phase-shifter hardware imposes a constant-modulus constraint, while prediction from historical observations enables limited-probing selection without explicitly estimating the full instantaneous channel.With perfectly known instantaneous channel information, the MRT solution maximizes the beamforming gain and serves as the system upper bound.Source: A. Site-Specific Downlink Channel · p. 3

B. Beam Prediction and Problem Formulation

For each next slot, the BS generates feasible beam candidates conditioned on stored beam history and site-specific information. Candidates are probed under a common channel realization, and the objective maximizes expected best-candidate normalized gain.

B. Beam Prediction and Problem Formulation

  • At each slot, candidates generated in the preceding slot are evaluated by CSI-RS probing, and the highest-RSRP candidate is used for transmission.The channel is assumed quasi-static within a probing round but evolves across slots; candidate generation finishes during the current data-transmission period.Source: B. Beam Prediction and Problem… · p. 4
  • The temporal prior for predicting slot t+1 is the ordered history of selected beamforming vectors over an observation window of length T_h.The BS combines available history with site-specific prior information because future beams are unavailable at prediction time.Source: B. Beam Prediction and Problem… · p. 4
  • The formulation samples P feasible candidates conditioned on history and site information, then maximizes expected best-of-P normalized beamforming gain.Normalization compares the best candidate with the optimal constant-modulus gain for each channel realization.Source: B. Beam Prediction and Problem… · p. 4

III. PROPOSED BEAM-STEP METHOD

Beam-STeP trains a site-specific, history-conditioned diffusion model offline and uses DDIM online to produce feasible beam candidates for probing and selection. Its angular-domain representation makes beam direction explicit while retaining an invertible mapping to the spatial domain.

III. PROPOSED BEAM-STEP METHOD

  • Offline training uses site-specific historical and future beams for a conditional noise predictor; online DDIM generates candidates from independent Gaussian initializations, projects them to feasible analog beams, and selects the highest-RSRP candidate.Beam-STeP names the joint use of spatial priors from the deployment site and temporal priors from beam evolution for beam prediction.Source: §III · p. 4

A. Angular-Domain Beam Representation

  • A unitary DFT transforms each beam into an angular representation that retains all coefficients, making direction more explicit while remaining invertible.The model input represents the real and imaginary parts of the angular-domain beam sequence across the history window.Source: A. Angular-Domain Beam Representation · p. 5
  • Generated real-valued angular samples are reconstructed as complex spatial beams and projected onto the feasible analog-beam set.The inverse representation reconstructs the complex angular vector from its real and imaginary parts before projection.Source: A. Angular-Domain Beam Representation · p. 5

B. History-Conditioned Diffusion Training and DDIM Sampling

Beam-STeP trains a history-conditioned denoiser to predict noise in corrupted angular-domain beams, then uses DDIM to generate future-beam candidates from independent Gaussian initializations. The noise-prediction loss trains the denoiser, while best-of-P normalized beamforming gain evaluates performance after candidate probing and selection.

B. History-Conditioned Diffusion Training and DDIM Sampling

  • DDIM generates multiple plausible future-beam candidates from independent Gaussian initializations conditioned on beam history.Its shortened schedule uses K reverse updates, with K much smaller than the full diffusion-step count.Source: B. History-Conditioned Diffusion… · p. 5
  • The denoiser predicts injected noise conditioned on noisy beams, diffusion step, and historical beam observations.Training samples a diffusion step and Gaussian noise, constructs the noisy target directly, and minimizes the noise-prediction loss.Source: B. History-Conditioned Diffusion… · p. 5
  • The best-of-P normalized beamforming gain after probing and selection is the performance criterion, while the noise-prediction loss is the training surrogate.Source: B. History-Conditioned Diffusion… · p. 5

C. Network Architecture

The Beam-STeP noise predictor combines a Transformer history encoder, diffusion-step encoder, and FiLM-conditioned one-dimensional U-Net denoiser.

C. Network Architecture

  • The noise predictor comprises a Transformer-based history encoder, a diffusion-step encoder, and a FiLM-conditioned one-dimensional U-Net denoiser.Source: C. Network Architecture · p. 6
  • The history encoder supplies sequence context to the predictor, while the diffusion-step encoder represents the current diffusion step.Source: C. Network Architecture · p. 6
  • The FiLM-conditioned one-dimensional U-Net serves as the denoiser in the noise-prediction network.Source: C. Network Architecture · p. 6

1) History Encoder:

The architecture encodes recent beam state and trajectory history, conditions denoising on diffusion noise level, and uses a U-Net to recover angular-domain noise predictions.

1) History Encoder:

  • The history encoder projects angular real/imaginary vectors into positioned tokens, then pools the last token and mean-pooled trajectory into a unified context.The last token retains the most recent beam state; mean pooling summarizes evolution across the observation window.Source: 1) History Encoder · p. 6

2) Diffusion-Step Encoder:

3) FiLM Conditioning:

  • The diffusion-step embedding is fused with history context, and the resulting condition is injected into every U-Net residual convolution block through FiLM.FiLM modulates feature maps using learned channel-wise scaling and shifting projections.Source: 3) FiLM Conditioning · p. 6

4) U-Net:

  • The U-Net processes a two-channel real/imaginary angular feature map; its encoder expands receptive fields and its decoder restores resolution with skip connections.A final 1 × 1 convolution returns the two-channel predicted noise.Source: 4) U-Net · pp. 6–7

D. Offline Training and Online Deployment

Beam-STeP learns a site-specific denoiser offline from angular-domain history–target pairs, then generates and probes candidate beams online. The selected highest-RSRP beam updates the history for recursive prediction.

1) Offline Training:

  • Offline training learns site-specific propagation and mobility information implicitly from dataset history–target pairs, without requiring an explicit map during online inference.Each mini-batch samples a diffusion step and Gaussian noise, and the network learns to predict the injected noise.Source: 1) Offline Training · p. 7

2) Online Inference:

  • Online inference independently initializes P candidates, denoises them in parallel with cached history context, and probes them for transmission selection.Source: 2) Online Inference · p. 7
  • The highest-RSRP candidate is transmitted, then its angular representation replaces the oldest history entry to condition the next prediction.Candidate beams are reconstructed as feasible physical beams before probing.Source: 2) Online Inference · p. 7

IV. NUMERICAL RESULTS

The evaluation uses three DeepMIMO urban scenarios and a road-aware mobility model to assess site-specific beam prediction. Trajectories are generated from local road structure and partitioned by trajectory, while the model is explicitly not intended to reproduce exact real-user behavior.

A. Simulation Setup

  • The simulations use three DeepMIMO urban scenarios at 28 GHz, with a 64-element base-station array and 15 historical beam observations for each prediction.Adjacent trajectory points are about 1.5 m apart, with a 100 ms sampling interval corresponding to 15 m/s.Source: A. Simulation Setup · pp. 7–8

B. Road-Aware Site-Specific Trajectory Construction

  • Road-aware trajectories approximate plausible urban mobility while respecting local road layout, providing temporal information distinct from site geometry.The construction identifies road branches and intersections, then samples movement actions such as going straight or turning.Source: B. Road-Aware Site-Specific Trajectory… · p. 8
  • Each scenario contains 18,000 trajectories, split into training, validation, and test sets at 70%, 15%, and 15%, respectively.The split is performed at the trajectory level, and the training and test sets are checked for identical history-target windows.Source: B. Road-Aware Site-Specific Trajectory… · p. 9
  • The road-aware mobility model is a controllable approximation, not a reproduction of the exact statistical behavior of real users.The paper notes that other mobility models may also be used to evaluate the method.Source: B. Road-Aware Site-Specific Trajectory… · p. 9

C. Baselines

The supplied passages identify normalized beamforming-gain comparisons across three urban scenarios. The figures specify the scenarios but do not provide comparison outcomes.

C. Baselines

  • Fig. 7 presents normalized beamforming gain for different schemes in the City 17 Seattle 28 scenario.The caption identifies the scenario and metric but gives no scheme ranking or values.Source: Figure 7 · p. 9
  • Fig. 8 presents normalized beamforming gain for different schemes in the City 2 Chicago 28 scenario.The caption identifies the scenario and metric but gives no scheme ranking or values.Source: Figure 8 · p. 9

1) Last-Beam Reuse:

Beam-STeP’s multi-step gains exceed the ODE-LSTM baseline, and the experiments examine candidate diversity, codebook comparison, and the contribution of site-specific training information. At the first prediction slot, it also improves on reusing the preceding beam.

2) Continuous Multi-Candidate ODE-LSTM:

  • ODE-LSTM takes historical beamforming vectors and predicts multiple continuous-valued candidates, assigning each a confidence logit for probing.The candidates are selected for probing using probabilities obtained by softmax-normalizing the logits.Source: 2) Continuous Multi-Candidate ODE-LSTM · pp. 9–10

4) Site-Agnostic Methods:

  • The site-agnostic variants remove road-aware mobility, weaken the target-site propagation prior, or remove both sources of site-specific knowledge.The variants keep the architecture, training budget, and probing protocol unchanged.Source: 4) Site-Agnostic Methods · p. 10

1) Last-Beam Reuse:

  • At the first prediction slot, P = 4 Beam-STeP improves on last-beam reuse by 0.81 dB, 0.63 dB, and 0.48 dB in Philadelphia, Seattle, and Chicago.Table II reports last-beam reuse alongside Beam-STeP under different probing budgets.Source: D. Performance of Beamforming Gain · p. 10, Table II · p. 10

D. Performance of Beamforming Gain

  • At h = 10 with P = 64, Beam-STeP exceeds ODE-LSTM by approximately 2.34 dB, 1.61 dB, and 0.78 dB in Philadelphia, Seattle, and Chicago.Across the same steps, Beam-STeP’s gain decreases less than ODE-LSTM’s in all three scenarios.Source: D. Performance of Beamforming Gain · p. 11

3) Codebook-Based Method:

D. Performance of Beamforming Gain

  • At an intersection, independently denoised candidates converge to distinct beam states and reproduce rich sidelobe and multiple-main-lobe patterns.The example uses three independent Gaussian noise samples conditioned on the same history.Source: Figure 9 · p. 10, D. Performance of Beamforming Gain · p. 11

E. Site-Specific Analysis

The site-specific approach outperforms variants that weaken propagation or mobility priors across the three scenarios. The latency evaluation finds that 10-step DDIM retains gain close to longer sampling procedures while reducing candidate-generation latency.

E. Site-Specific Analysis

  • Across all three scenarios, the fully site-specific approach outperforms both the single-site random-walk and mixed-site road-aware models.The reported advantage persists as the prediction horizon increases; the mixed-site random-walk model generally performs worst.Source: E. Site-Specific Analysis · p. 12

F. Online Inference Latency Analysis

  • In Philadelphia at P=16 and h=10, K=10 achieves −1.85 dB at 40.45 ms; K=20 achieves −1.76 dB at 82.81 ms, while K=50 takes 184.41 ms for a further 0.04 dB improvement.K=10 reduces latency by 78.1% relative to K=50, with a gain reduction of only 0.13 dB.Source: F. Online Inference Latency Analysis · p. 12, Figure 11 · p. 12
  • At K = 10, DDIM candidate generation takes around 40 ms across the three scenarios, with gain within 0.13 dB of the 500-step DDPM sampler.Table III reports the latency and normalized gain for P = 16 and h = 10 on an NVIDIA A100-SXM4-80GB GPU.Source: F. Online Inference Latency Analysis · p. 12

V. CONCLUSION

  • Beam-STeP combines site-dependent propagation and mobility information with historical beamforming sequences to generate diverse future beam candidates.The conclusion reports robust multi-step prediction and a gain-latency trade-off for the 10-step DDIM sampler.Source: §V · p. 13
Loading 2610.20282v1…