Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SCAN-BEST: Sub-6GHz-Aided Near-field Beam Selection with Formal Reliability Guarantees

Weicao Deng, Binpu Shi, Min Li, Osvaldo Simeone

TL;DR

SCAN-BEST couples deep learning-based near-field beam prediction with conformal risk control to guarantee a target suboptimality level with high probability when selecting mmWave beams from sub-6G data. It constructs a candidate beam subset using CRC (and optionally weighted CRC for covariate shifts), ensuring a specified coverage rate before performing limited mmWave training to finalize the beam. The framework is predictor-agnostic, supports offline calibration, and demonstrates reliability across calibration sizes, sub-6G antenna counts, and transition scenarios between calibration and deployment. Results show superior efficiency and reliability over baselines, with meaningful reductions in beam training overhead and robust performance under distribution shifts. The work provides a practical, theoretically grounded path for scalable sub-6G aided near-field beam selection in 6G-like systems.

Abstract

As millimeter-wave (mmWave) MIMO systems adopt larger antenna arrays, near-field propagation becomes increasingly prominent, especially for users close to the transmitter. Traditional far-field beam training methods become inadequate, while near-field training faces the challenge of large codebooks due to the need to resolve both angular and distance domains. To reduce in-band training overhead, prior work has proposed to leverage the spatial-temporal congruence between sub-6 GHz (sub-6G) and mmWave channels to predict the best mmWave beam within a near-field codebook from sub-6G channel estimates. To cope with the uncertainty caused by sub-6G/mmWave differences, we introduce a novel Sub-6G Channel Aided Near-field BEam SelecTion (SCAN-BEST) framework that wraps around any beam predictor to produce candidate beam subset with formal suboptimality guarantees. The proposed SCAN-BEST builds on conformal risk control (CRC), and is calibrated offline using limited calibration data. Its performance guarantees apply even in the presence of statistical shifts between calibration and deployment. Numerical results validate the theoretical properties and efficiency of SCAN-BEST.

SCAN-BEST: Sub-6GHz-Aided Near-field Beam Selection with Formal Reliability Guarantees

TL;DR

SCAN-BEST couples deep learning-based near-field beam prediction with conformal risk control to guarantee a target suboptimality level with high probability when selecting mmWave beams from sub-6G data. It constructs a candidate beam subset using CRC (and optionally weighted CRC for covariate shifts), ensuring a specified coverage rate before performing limited mmWave training to finalize the beam. The framework is predictor-agnostic, supports offline calibration, and demonstrates reliability across calibration sizes, sub-6G antenna counts, and transition scenarios between calibration and deployment. Results show superior efficiency and reliability over baselines, with meaningful reductions in beam training overhead and robust performance under distribution shifts. The work provides a practical, theoretically grounded path for scalable sub-6G aided near-field beam selection in 6G-like systems.

Abstract

As millimeter-wave (mmWave) MIMO systems adopt larger antenna arrays, near-field propagation becomes increasingly prominent, especially for users close to the transmitter. Traditional far-field beam training methods become inadequate, while near-field training faces the challenge of large codebooks due to the need to resolve both angular and distance domains. To reduce in-band training overhead, prior work has proposed to leverage the spatial-temporal congruence between sub-6 GHz (sub-6G) and mmWave channels to predict the best mmWave beam within a near-field codebook from sub-6G channel estimates. To cope with the uncertainty caused by sub-6G/mmWave differences, we introduce a novel Sub-6G Channel Aided Near-field BEam SelecTion (SCAN-BEST) framework that wraps around any beam predictor to produce candidate beam subset with formal suboptimality guarantees. The proposed SCAN-BEST builds on conformal risk control (CRC), and is calibrated offline using limited calibration data. Its performance guarantees apply even in the presence of statistical shifts between calibration and deployment. Numerical results validate the theoretical properties and efficiency of SCAN-BEST.

Paper Structure

This paper contains 32 sections, 5 theorems, 58 equations, 16 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works and Motivations
  3. Main Contributions
  4. System Model and Problem Formulation
  5. MmWave Downlink Communication and Channel Model
  6. Sub-6G Channel Model and Uplink Channel Estimation
  7. Problem Formulation
  8. Sub-6G Channel Aided Near-field Beam Selection Framework
  9. Overview of SCAN-BEST
  10. Deep Learning-Based Near-Field Beam Prediction
  11. Sub-6G-Based Candidate Beam Subset Selection
  12. Addressing Covariate Shifts between Calibration and Test Data
  13. MmWave Beam Training
  14. Numerical Results
  15. Simulation Scenario and System Parameter Setup
  16. ...and 17 more sections

Key Result

Proposition 1

The subset $\mathcal{C}_{\hat{\lambda}}(\widehat{\underline{\mathbf{H}}})$ in def candidate beam subset with the threshold attains the target coverage rate $1-\alpha$ in target coverage rate for any input $\widehat{\underline{\mathbf{H}}}$ and "any" predictive model $\mathscr{P}_{\boldsymbol{\theta}^{\star}}(\cdot)$.

Figures (16)

  • Figure 1: An illustration of spatial and temporal correlation between sub-6G and mmWave channels. Each column represents a pair of sub-6G (upper row) and mmWave (lower row) augmented angle and delay profiles. The first column describes a line-of-sight (LoS) condition, while the second column represents a non-LoS (NLoS) condition Remcom.
  • Figure 2: An illustration of system model and beamforming architecture.
  • Figure 3: The overall design of the SCAN-BEST framework.
  • Figure 4: A simple illustration of $\hat{R}(\lambda)$ (blue line), the orange line denotes the desired risk level.
  • Figure 5: An illustration of simulation setup.
  • ...and 11 more figures

Theorems & Definitions (6)

  • Proposition 1
  • Proposition 2
  • Proposition 3
  • proof
  • Lemma 1: From tibshirani2019conformal
  • Lemma 2: Lemma 3 of tibshirani2019conformal