Efficient, Adaptive Near-Field Beam Training based on Linear Bandit

This paper proposes a linear bandit-based beam training framework for near-field multi-path channels that utilizes Thompson Sampling with a correlated Gaussian prior and three distinct search strategies to significantly reduce pilot overhead while maximizing beamforming gain and approaching full-CSI performance.

The Gist

Imagine you are trying to tune an old-fashioned radio to find a specific station in a crowded city. In the past (the "far-field" era), the radio waves were like straight laser beams. You just had to point your antenna in the right direction (left, right, up, down), and if you got the angle right, the music would play clearly.

But now, we are entering the world of 6G and Near-Field Communication. Think of this like moving from a flat, open field into a bustling, cluttered indoor shopping mall.

Here is the problem:

1. The Waves are Curved: In the mall, the sound doesn't travel in straight lines; it curves around corners and bounces off walls. To hear the music clearly, you can't just point your antenna; you have to know exactly where the person is (the angle) AND how far away they are (the distance).
2. The "Search" is Huge: To find the perfect spot, you used to have to check every single possible angle and distance combination. If you had 1,000 spots to check, you'd have to try them all one by one. This takes forever and wastes a lot of battery (pilot overhead).
3. The Echoes (Multipath): In the mall, the sound bounces off mannequins and pillars. You hear the direct voice, but also echoes. Your radio needs to combine all these sounds to make the voice clear, not just pick the loudest one.

The Paper's Solution: The "Smart Detective" (Linear Bandit)

The authors of this paper propose a new way to find the signal, which they call a Linear Bandit Framework using Thompson Sampling.

Let's break this down with a simple analogy: The "Guess-and-Learn" Detective.

Instead of checking every single spot in the mall (which is slow), imagine you have a Smart Detective who is trying to find the source of a whisper.

1. The "Gut Feeling" (The Prior)

The detective starts with a "Gut Feeling" (a mathematical model called a Gaussian Prior).

• Old way: The detective thinks every spot in the mall is equally likely.
• New way: The detective knows that if the whisper is coming from near the "Coffee Shop," it's also likely to be coming from the "Bookstore" right next to it because sound leaks and spreads. The detective uses a Gaussian Kernel to understand that nearby spots are connected. This helps the detective learn faster.

2. The "Balancing Act" (Exploration vs. Exploitation)

The detective has two modes:

• Exploration: "I'm not sure where the voice is. Let me check a few random spots to see if I hear anything."
• Exploitation: "I think the voice is near the Coffee Shop. Let me focus my attention there to get the clearest sound."

The paper uses Thompson Sampling to perfectly balance these two. It's like a smart coin flip: if the detective is very unsure, the coin lands on "Explore." As the detective gets more clues, the coin starts landing on "Exploit" (focusing on the best spot).

3. The Three Strategies (The Detective's Tools)

The paper suggests three ways the detective can do the job, depending on how fast they need to be:

• Strategy A: The "Checklist" (Codebook-Constrained)
  The detective has a printed checklist of 1,000 specific spots to check. They stick to the list.

  • Pros: Very fast to get started. You won't get lost.
  • Cons: You might miss the perfect spot because it's not on the list (like missing a spot between two checklist items).

• Strategy B: The "Free Roam" (Continuous Space)
  The detective throws the checklist away and can point to any spot in the mall, even the tiny gaps between the checklist items.

  • Pros: Can find the absolute perfect spot.
  • Cons: In a noisy mall (low signal), the detective gets confused and wastes time wandering aimlessly. It takes too long to find the voice.

• Strategy C: The "Hybrid" (The Best of Both Worlds)
  This is the paper's star strategy.

  • Phase 1: The detective starts with the Checklist (Strategy A) to quickly narrow down the search to the right general area.
  • Phase 2: Once they are close, they throw away the checklist and use Free Roam (Strategy B) to fine-tune the exact spot.
  • Result: You get the speed of the checklist and the precision of free roaming.

The Results: Why Does This Matter?

The authors ran simulations (virtual tests) to see how well this works.

• Speed: Their "Hybrid Detective" found the signal using 90% fewer checks than the old method of checking every single spot. It's like finding a needle in a haystack in 10 seconds instead of 100.
• Quality: Even though they checked fewer spots, the signal quality (SNR) was actually better (by about 2 dB) than the old methods.
• Robustness: It works great even when there are lots of echoes (multipath), which is the reality of real-world 6G networks.

The Bottom Line

This paper solves a major headache for future 6G networks. As we use bigger antennas and higher frequencies, the "search" for the signal becomes incredibly complex and slow.

By treating the search like a smart learning game (using Thompson Sampling) and combining a rough map with fine-tuning, the authors created a system that finds the signal much faster and more accurately than before. This means your future 6G phone will connect instantly, even in crowded, echoey places, without draining your battery.

Technical Summary

Here is a detailed technical summary of the paper "Efficient, Adaptive Near-Field Beam Training based on Linear Bandit".

1. Problem Statement

The paper addresses the critical challenge of beam training in Near-Field (NF) communications for Extremely Large-Scale MIMO (XL-MIMO) systems, particularly under multipath propagation conditions.

• Context: As 6G moves to higher frequencies (mmWave/THz) and utilizes XL-MIMO, the radiative near-field region expands. In this region, channel steering vectors are spherical rather than planar, requiring parameterization by both angle and distance.
• Challenge: Traditional near-field beam training requires an exhaustive sweep over a 2D polar-domain codebook (angle + distance). This leads to a codebook size that scales drastically, resulting in prohibitive pilot overhead and training latency.
• Gap: Existing efficient strategies often assume a dominant Line-of-Sight (LoS) path or rely on deterministic searches. Methods handling multipath (e.g., multi-beam linear combination) often still require exhaustive sweeps to ensure reconstruction accuracy, failing to reduce overhead significantly. There is a lack of robust, low-overhead solutions for multipath near-field beam alignment.

2. Methodology

The authors propose a Linear Bandit-based beam training framework utilizing Thompson Sampling (TS) to adaptively balance exploration and exploitation.

A. System & Channel Model

• Setup: A Base Station (BS) with a Uniform Linear Array (ULA) serves a single-antenna user in a multipath environment (1 LoS + $L-1$ Non-Line-of-Sight paths).
• Transformation: To exploit spatial correlations, the system transforms the spatial-domain channel vector ($\mathbf{h}$) into the DFT domain ($\mathbf{g}$) using a unitary DFT matrix. The received signal is modeled as a linear Gaussian observation: $y_t = \mathbf{g}^H \mathbf{v}_t + n_t$.

B. Key Innovations

1. Correlated Gaussian Prior:
   • Unlike far-field scenarios where energy focuses on a single DFT bin, near-field paths exhibit energy leakage across adjacent angular bins.
   • The authors incorporate a Gaussian (RBF) kernel into the prior covariance matrix ($\mathbf{D}_0$) to model these angular correlations. This allows the algorithm to learn about adjacent beams even without explicit probing, accelerating posterior convergence.
2. Thompson Sampling (TS) Strategy:
   • The beam training is modeled as a sequential decision-making problem.
   • At each time slot, the BS samples a channel vector from the posterior distribution (updated via Bayesian recursion) and selects a beamformer to maximize the expected reward (beamforming gain).
   • TS naturally balances exploration (when uncertainty/variance is high) and exploitation (when the posterior concentrates).

C. Three Proposed Schemes

The paper develops three specific TS strategies to address different trade-offs:

1. Scheme I (Codebook-constrained TS): Restricts the search to a predefined near-field polar codebook. This provides structural regularization, preventing deviation due to initial uncertainty and ensuring rapid convergence, especially in low SNR.
2. Scheme II (Continuous-space TS): Searches over the full unit sphere (continuous space) without codebook constraints. This avoids grid quantization errors and achieves near-optimal performance but suffers from slower convergence and noise sensitivity in low SNR due to unconstrained exploration.
3. Scheme III (Hybrid Refinement TS): A two-stage approach.
   • Stage 1: Uses Scheme I (Codebook) for rapid initial stabilization ("warm-start").
   • Stage 2: Switches to Scheme II (Continuous space) for high-precision refinement once the estimate is stable.
   • This balances speed and accuracy.

3. Key Contributions

• Framework: First application of a Linear Bandit framework with Thompson Sampling specifically tailored for multipath near-field beam training.
• Prior Modeling: Introduction of a Gaussian kernel-based prior in the DFT domain to explicitly capture near-field energy leakage and angular correlations, improving data efficiency.
• Algorithmic Diversity: Development of three distinct strategies (Codebook, Continuous, Hybrid) offering flexibility for different latency and accuracy requirements.
• Theoretical Insight: Demonstration that the continuous-space search is asymptotically optimal, approaching the Full-CSI bound when pilot overhead is unconstrained.

4. Simulation Results

Simulations were conducted with $N=256$ antennas, 30 GHz carrier, and 4 multipath components.

• Pilot Overhead Reduction: The proposed framework reduces pilot overhead by up to 90% compared to exhaustive near-field codebook search.
• Performance Gain:
  • At 15 dB SNR, the Hybrid Scheme (III) achieves 12.8 bps/Hz, outperforming the multi-beam baseline (12.4 bps/Hz) and exhaustive search (12.1 bps/Hz).
  • To reach a target rate of 13.9 bps/Hz, Scheme III offers >2 dB SNR gain over the exhaustive search baseline.
• Robustness: Scheme III maintains the narrowest gap to the Full-CSI upper bound across all SNR regimes.
• Overhead Efficiency: Scheme III achieves its performance with an average of ~101 pilots, compared to 256 (multi-beam) and 1280 (exhaustive) required by baselines.
• Asymptotic Behavior: When pilot constraints are removed, the continuous-space search (Scheme II) converges to the Full-CSI bound, confirming that performance limits in constrained scenarios are due to budget, not estimation precision.

5. Significance

This work provides a critical solution for the deployment of 6G XL-MIMO systems in realistic, multipath-rich environments. By moving away from exhaustive 2D sweeps and leveraging adaptive learning (Bandits) with spatial correlation modeling, the proposed framework:

• Drastically reduces training latency, which is vital for dynamic 6G scenarios.
• Maintains high spectral efficiency even without perfect Channel State Information (CSI).
• Offers a scalable architecture that can adapt to varying SNR conditions and hardware constraints through its hybrid strategy.

In summary, the paper successfully bridges the gap between theoretical near-field beamforming and practical, low-overhead implementation in complex multipath channels.