PaperlayerView exact revision on arXiv

Source-linked AI summary

Intelligent Beamforming and Handover via Physics-Informed Beam-Aware CKM Diffusion

Le Zhao, Zesong Fei, Xinyi Wang, Nan Ha, Yining Wang, Yong Zeng, Rui Zhang

arXiv:2609.34704v1eess.SP

TL;DR

The paper addresses environment-aware predictive beam management when existing CKMs are restricted to fixed beams or discrete codebooks. It introduces BeamCKMDiff, a physics-informed diffusion framework for continuous beam-aware mapping and pilot-free optimization, and reports an NMSE of −23.96 dB at five sampling steps, with no observed outage samples for the proposed dual-BS handover.

  • Problem

    The paper addresses the need for environment-aware predictive communication when existing CKMs are restricted to fixed beams or discrete codebooks.

    Source: §I · p. 2

  • Method

    BeamCKMDiff combines an analytical LoS prior and environmental topology with continuous beam conditioning, then uses differentiable reverse diffusion and SCA for pilot-free beam optimization.

    Source: §I · p. 2

  • Results

    At Tstep = 5, BeamCKMDiff achieves an NMSE of −23.96 dB, while the proposed dual-BS handover produces no observed outage samples.

    Source: B. Performance of BeamCKM Construction · p. 12, C. Performance of CKM-based… · p. 14

  • Takeaways & Limitations

    The framework supports beam optimization without online pilot scanning and pre-configures the target beam before association transfer.

    Source: §I · p. 2, C. BeamCKMDiff-Based Dual-BS Handover · p. 9

Abstract

from arXiv · show

Downlink communications from base stations (BSs) to unmanned aerial vehicles (UAVs) in sixth-generation (6G) networks require precise beam alignment to overcome severe mobile communication path loss. However, traditional exhaustive beam sweeping relies on discrete codebooks and consumes valuable air-interface resources for online measurements, rendering it inefficient for highly dynamic aerial environments. In this paper, we propose BeamCKMDiff, a physics-informed generative diffusion framework designed to construct high-fidelity continuous beam-aware channel knowledge maps (BeamCKMs). Unlike existing empirical approaches, BeamCKMDiff employs a diffusion transformer (DiT) architecture featuring a dual-path conditioning mechanism. It integrates an analytical line-of-sight (LoS) beam prior with environmental topologies, while simultaneously embedding continuous beamforming vectors through an adaptive layer normalization (adaLN) mechanism. Building upon this differentiable generative mapping, we propose a CKM-based end-to-end continuous beamforming and proactive dual-BS handover algorithm. By analytically propagating gradients through the reverse diffusion process, the proposed framework enables successive convex approximation (SCA)-based beamforming optimization in the computational domain, effectively bypassing physical pilot scanning. Simulation results demonstrate that BeamCKMDiff achieves a normalized mean square error (NMSE) of --23.96 dB, establishing a highly reliable spatial prior. Compared to discrete beam sweeping and LoS-assumed baselines, the proposed CKM-based beamforming adapts to non-LoS (NLoS) conditions, providing accurate beam alignment and higher spectral efficiency. Furthermore, the integrated dual-BS handover mechanism eliminates blockage-induced link outages, ensuring robust connectivity for high-mobility aerial platforms.

I. INTRODUCTION

The introduction motivates environment-aware beam management and presents BeamCKMDiff for continuous beam-aware channel mapping and pilot-free optimization. The system model specifies a fixed-altitude UAV setting and a multipath downlink channel model.

I. INTRODUCTION

  • BeamCKMDiff conditions diffusion-generated BeamCKMs on continuous beams and uses their differentiability for pilot-free beam optimization and proactive dual-BS handover.The framework combines an analytical LoS prior and environmental topology with continuous beamforming vectors; SCA performs beam optimization in the computational domain.Source: §I · p. 2

1) Environment and UAV model:

  • The environment is a height-truncated urban grid, and the modeled BeamCKM is a horizontal map for fixed-altitude UAV flight; altitude-varying trajectories are future work.Maps at multiple altitudes can be stacked to form a full three-dimensional CKM.Source: 1) Environment and UAV model · p. 3

2) Communication model:

  • The channel model combines LoS, reflections, and diffractions, averages rapid small-scale fading, assumes orthogonal BS resources, and limits the array study to ULAs.Interference-aware simultaneous transmission and extension to uniform planar arrays are left for future work.Source: 2) Communication model · p. 3

3) Beam-Aware CKM and RoI Mapping:

This section defines BeamCKM generation as predicting spatial channel gains from the environment, BS location, and beamforming vector, with location uncertainty handled through a region-of-interest mask.

3) Beam-Aware CKM and RoI Mapping:

  • Each BeamCKM pixel represents the expected channel gain at its grid location, and a trainable generative model constructs the predicted map.Acquiring a global ground-truth map through ray tracing or exhaustive field measurements is described as prohibitive.Source: 3) Beam-Aware CKM and RoI Mapping · p. 4
  • The predicted BeamCKM is the spatial mapping function used to meet the stated prediction requirement.The function is parameterized by trainable neural-network parameters.Source: 3) Beam-Aware CKM and RoI Mapping · p. 4
  • The Region of Interest mask represents UAV coordinate uncertainty, and the predicted gain for a location is averaged over the valid region.The RoI is designed to encompass the stated spatial uncertainty bound.Source: 3) Beam-Aware CKM and RoI Mapping · p. 4

B. Problem Formulation

The formulation maximizes average spectral efficiency along a UAV trajectory using CKM-predicted spatial channel gains to optimize continuous beamforming and BS association. It separates CKM construction from gradient-based beamforming and handover optimization, avoiding online measurements.

B. Problem Formulation

  • The objective maximizes average spectral efficiency across UAV flight slots by optimizing continuous active beamforming vectors along the trajectory.The rate uses spatial channel gains computed from the generated CKM.Source: B. Problem Formulation · p. 4
  • The non-convex mapping from continuous complex beamforming vectors to generated spatial channel distributions makes discrete pilot sweeping costly and unsuitable for continuous spatial optimization.The passage identifies unaffordable latency overhead from conventional discrete pilot sweeping.Source: B. Problem Formulation · p. 4
  • The solution separates BeamCKM construction from CKM-based beamforming and handover, optimizing continuous beams by gradient propagation and BS association without online measurements.The formulation also imposes a normalized transmit-power budget and discrete BS selection.Source: B. Problem Formulation · p. 4

III. PHYSICS-INFORMED BEAMCKMDIFF

BeamCKMDiff uses a compressed latent representation and a DiT-based generative mapping conditioned on environmental topology and an analytical beam prior. The prior models LoS directivity; the generative model learns NLoS multipath structure from data.

III. PHYSICS-INFORMED BEAMCKMDIFF

  • BeamCKMDiff combines a DeepVAE, an environment-and-prior condition encoder, and a DiT backbone with adaLN beam embedding to form the generative mapping surrogate.The framework is presented as a surrogate for mapping conditions to BeamCKMs.Source: §III · p. 5

A. Perceptual Compression via DeepVAE

  • DeepVAE compresses high-dimensional CKMs into a latent Gaussian representation, reconstructs maps with a decoder, and is frozen before latent-space generation.Its training objective combines pixel-level reconstruction loss and KL-divergence regularization.Source: A. Perceptual Compression via DeepVAE · p. 5

1) Analytical Beam Prior:

  • The analytical beam prior uses 3D geometry and array directivity to construct an LoS beam-gain map, while the generative model learns NLoS multipath structure from data.The prior is explicitly limited to LoS directivity; the cited passages describe the geometry, beamforming vector, and free-space path-loss inputs.Source: 1) Analytical Beam Prior · pp. 5–6

2) Spatial Feature Extraction:

The model encodes normalized building topology, transmitter location, and analytical beam priors as spatial conditions, then fuses them with noisy latents in a DiT. A beam-aware adaLN mechanism uses the continuous beam vector and diffusion step to steer denoising.

2) Spatial Feature Extraction:

  • Building heights are normalized relative to UAV altitude, then environmental maps and the analytical beam prior are encoded as a latent-resolution spatial condition.The encoder uses strided convolutions and residual blocks; the transmitter map marks the BS location.Source: 2) Spatial Feature Extraction · p. 6

C. Continuous Beam Conditioning via DiT

1) Latent Feature Fusion:

  • Noisy latents are concatenated with spatial conditions, partitioned into 2 × 2 patches, and embedded with diffusion-step and beam-vector features for DiT conditioning.The time and beam embeddings are summed and used to modulate each DiT block through adaLN.Source: 1) Latent Feature Fusion · p. 6, 2) Beam-aware AdaLN · p. 6

2) Beam-aware AdaLN:

  • Beam-aware adaLN treats the beam configuration and diffusion step as global controls that steer denoising through scale, shift, and gating parameters.The gate is initialized to zero, so each block initially acts as an identity mapping for training stability.Source: 2) Beam-aware AdaLN · p. 6

D. Diffusion Training and Inference

Training corrupts encoded clean latents with Gaussian noise and learns to predict that noise. Inference uses DDIM reverse updates, with stochasticity disabled for a deterministic update intended to maximize inference speed.

1) Forward / Training Process:

  • Training corrupts the encoded clean latent at a randomly sampled diffusion step with Gaussian noise, and the model minimizes MSE between injected and predicted noise.The cumulative noise schedule parameter determines the noisy latent construction.Source: 1) Forward / Training Process · p. 7

2) Reverse Process via DDIM:

  • During reverse inference, DDIM estimates a clean latent from the current noisy state and transitions between diffusion steps.The transition includes a stochasticity parameter η.Source: 2) Reverse Process via DDIM · p. 7
  • Setting η = 0 removes the random noise term, yielding a fully deterministic reverse update intended to maximize inference speed for real-time beamforming.In this setting, σt becomes zero.Source: 2) Reverse Process via DDIM · p. 7

IV. CKM-BASED BEAMFORMING AND HANDOVER

The section makes BeamCKMDiff’s generated spatial prior differentiable with respect to continuous beamforming vectors. This enables gradient-based beam optimization through the unrolled reverse-diffusion process.

A. End-to-End Gradient Propagation

  • A fixed-noise DDIM path makes reverse diffusion deterministic and differentiable with respect to the beam, enabling variance-free comparisons across optimization iterations and candidate BSs.The same initial noise realization is retained throughout beam-optimization iterations and candidate-BS evaluations.Source: A. End-to-End Gradient Propagation · p. 7
  • The beam vector conditions generation through both adaLN in the DiT and the analytical beam prior shaping the spatial condition tensor.Unrolling DDIM transitions and decoding the final latent produces a composition of differentiable mappings.Source: A. End-to-End Gradient Propagation · pp. 7–8
  • The objective gradient propagates through the unrolled computational graph, combining condition-encoder backpropagation with a Wirtinger derivative of the analytical beam prior.The prior derivative links physical electromagnetic propagation to neural optimization.Source: A. End-to-End Gradient Propagation · pp. 7–8
  • The closed-form LoS-prior derivative supplies an analytical gradient field that dynamically guides the diffusion model’s spatial attention during beam optimization.The recursive formulation also makes the generative mapping continuous and amenable to reverse-mode automatic differentiation.Source: A. End-to-End Gradient Propagation · p. 8

B. BeamCKMDiff-based Beamforming

BeamCKMDiff optimizes continuous beamforming by maximizing the generative map’s expected channel gain within the target region of interest. SCA updates the beam iteratively under a power constraint.

B. BeamCKMDiff-based Beamforming

  • Because downlink rate increases monotonically with expected channel gain, maximizing the generated spatial output is equivalent to maximizing spectral efficiency within the target region of interest.The target objective is evaluated from the generative mapping.Source: B. BeamCKMDiff-based Beamforming · p. 8
  • SCA iteratively updates the continuous beamforming vector using the objective gradient with respect to the conjugate beam vector.The gradient is derived by the chain rule.Source: B. BeamCKMDiff-based Beamforming · p. 8
  • Under standard SCA conditions, the objective sequence is nondecreasing and every limit point is stationary; each update solves the proximal subproblem over the unit-power ball.The surrogate is tight at the current iterate and lower-bounds the local objective.Source: B. BeamCKMDiff-based Beamforming · p. 8

C. BeamCKMDiff-Based Dual-BS Handover

The handover procedure uses BeamCKMDiff to evaluate candidate links virtually and prepare a target beam before association transfer. The overall algorithm combines this with serving-beam optimization and centralized inference.

C. BeamCKMDiff-Based Dual-BS Handover

  • BeamCKMDiff virtually evaluates candidate-BS link quality, replacing exhaustive physical pilot sweeping as UAV mobility makes single-BS links prone to becoming outdated.The signaling procedure places RAN central-controller evaluation at the center of the proactive handover.Source: C. BeamCKMDiff-Based Dual-BS Handover · p. 8, Figure 3 · p. 9
  • When serving-link RSS crosses a degradation trigger, the UAV reports its estimated location and the controller optimizes serving and neighboring BS beams without neighbor scanning pilots.The model generates a spatial prior for the candidate-link evaluation.Source: C. BeamCKMDiff-Based Dual-BS Handover · p. 8
  • The neighboring BS is selected only when its expected gain exceeds the serving BS’s gain by the hysteresis margin; the target beam is then acquired proactively.The margin is introduced to prevent ping-pong handovers caused by transient environmental noise.Source: C. BeamCKMDiff-Based Dual-BS Handover · p. 9
  • The centralized procedure can extend to additional BSs by enlarging the candidate set without changing the handover process.Candidate links are evaluated by the RAN central controller.Source: C. BeamCKMDiff-Based Dual-BS Handover · p. 9

D. Overall Algorithm

  • In each slot, the overall algorithm optimizes the serving beam and, after severe signal degradation, evaluates candidate BSs to prepare a target beam without physical pilot scanning.The controller performs inference and SCA optimization; the UAV reports estimated horizontal coordinates and one RSS value and performs no neural-network computation.Source: D. Overall Algorithm · pp. 9–10

A. Simulation Settings

The evaluation uses ray-traced BeamCKMs from urban topologies and tests generation on unseen environments. BeamCKMDiff achieves the strongest reported map accuracy, while the visual comparison highlights its behavior under blockage.

A. Simulation Settings

  • The dataset contains 10,000 training and 1,000 testing samples from 110 urban topologies, with 100 environments for training and 10 held out for testing.The unseen test environments provide topology and transceiver coordinates, while channel gains are inferred by the model.Source: A. Simulation Settings · p. 10

1) Evaluation and System Configurations:

B. Performance of BeamCKM Construction

C. Performance of CKM-based Beamforming and Handover

Across dynamic UAV trajectories, CKM-based beamforming adapts to NLoS conditions, while dual-BS handover maintains link stability. The evaluation also examines optimization convergence, positioning error, and phase quantization.

C. Performance of CKM-based Beamforming and Handover

  • SCA stabilizes within 15 iterations across tested configurations, with Tstep = 5 achieving the highest objective value.Increasing diffusion steps further degrades optimization performance; the subsequent dynamic evaluations use re = 5 and Tstep = 5.Source: C. Performance of CKM-based… · p. 13
  • In NLoS settings, CKM-beamforming adapts continuous beams to exploit reflections and diffractions, unlike LoS-assumed steering into obstacles.In LoS settings, both CKM-beamforming and the LoS-assumption scheme align the main lobe with the UAV; NR sweeping shows codebook quantization error.Source: C. Performance of CKM-based… · p. 13
  • Across evaluated trajectories, single-BS CKM-beamforming outperforms NR sweeping and LoS-assumption, while dual-BS handover yields no observed outage samples.The reported dynamic results use exact ground-truth channels rather than predicted BeamCKMs.Source: C. Performance of CKM-based… · pp. 13–14, Figure 9 · p. 13, Figure 10 · p. 13
  • Position-error bounds re ≤10 m degrade RSS by less than 2.5 dB; low-resolution phase quantization retains robust performance in LoS and NLoS conditions.The continuous beam yields the highest RSS in the phase-quantization evaluation.Source: C. Performance of CKM-based… · p. 14, Figure 11 · p. 14, Figure 12 · p. 14
  • Virtual link evaluation and edge offloading shift beam optimization and inference from air-interface pilots and UAV computation to edge computation.The paper presents this as a potential reduction in reliance on exhaustive sweeping, with UAV battery preserved for propulsion and mission payload.Source: C. Performance of CKM-based… · p. 14

VI. CONCLUSION

  • −23.96 dB NMSE and sub-100 ms latency accompany BeamCKMDiff’s continuous-gain mapping and SCA beam optimization.The conclusion also reports that predictive dual-BS handover eliminates deep fades in NLoS environments.Source: §VI · p. 14
Loading 2609.34704v1…