Source-linked AI summary
Fast Beam Training for IRS-Assisted Multiuser Communications
Changsheng You, Beixiong Zheng, Rui Zhang
TL;DR
IRS beam training is costly because large reflecting surfaces require exhaustive single-beam searches. The paper partitions the IRS into sub-arrays and coordinates simultaneous beam steering over time, enabling users to identify directions through received-power comparisons. Simulations show reduced training time with comparable passive beamforming performance, although larger sub-array counts create a training-time–accuracy trade-off.
Problem
Large IRS codebooks make conventional single-beam reflect training costly for multiuser communication.
Method
The method divides IRS elements into sub-arrays, steers multiple directions over time, and identifies each user’s best direction from received-power or SNR comparisons.
Results
With M = 4, training time falls 50% versus single-beam training, from 160 to 80 symbols, while Psuc ≈92% at SNR = 46.4 dB and rate remains close.
Takeaways & Limitations
The proposed multi-beam design significantly reduces training overhead while achieving comparable passive beamforming performance for data transmission.
Takeaways & Limitations
Increasing M reduces training time but lowers sub-array gain and increases inter-beam interference, degrading beam-identification accuracy and passive beamforming performance.
Abstract
from arXiv · showhide
In this letter, we consider an intelligent reflecting surface (IRS)-assisted multiuser communication system, where an IRS is deployed to provide virtual line-of-sight (LoS) links between an access point (AP) and multiple users. We consider the practical codebook-based IRS passive beamforming and study efficient design for IRS reflect beam training, which is challenging due to the large number of IRS reflecting elements. In contrast to the conventional single-beam training, we propose a new multi-beam training method by dividing the IRS reflecting elements into multiple sub-arrays and designing their simultaneous multi-beam steering over time. By simply comparing the received signal power over time, each user can detect its optimal IRS beam direction with a high probability, even without searching over all possible beam directions as the single-beam training. Simulation results show that our proposed multi-beam training significantly reduces the training time of conventional single-beam training and yet achieves comparable IRS passive beamforming performance for data transmission.
I. INTRODUCTION
IRS can create virtual LoS links for blocked mmWave AP–user channels, but beam training becomes costly with many reflecting elements. The paper therefore studies a faster IRS reflect-beam training design for multiuser communication.
- IRS provides virtual LoS AP–IRS–user links when direct mmWave AP–user channels suffer blockage and propagation loss.
- IRS reflect beam training must coordinate with AP transmit beam training to establish high-SNR links.
- The system considers an IRS serving a multi-antenna AP and nearby single-antenna users, with fixed AP and IRS locations.
- The proposed method divides IRS elements into sub-arrays and steers different beams simultaneously over training symbols.
- Users identify their optimal IRS beam directions through received-power or SNR comparisons without exhaustively searching all directions.
- Simulations report substantially lower training time while largely preserving passive beamforming performance for data transmission.
II. SYSTEM MODEL
The system models an IRS-assisted downlink with a multi-antenna AP, single-antenna users, and LoS AP–IRS and IRS–user channels. IRS reflection phases form the passive beamforming vector, whose optimal direction aligns with each user’s cascaded spatial direction.
- The IRS has NI = Nx × Nz reflecting elements, while the AP has NA antennas and serves K single-antenna users.
- The model assumes limited scattering, blocked direct AP–user links, and deterministic LoS paths through the deployed IRS.
- Because the AP–IRS link is quasi-static, AP transmit beamforming is aligned with its LoS channel and treated as an equivalent single antenna.
- The IRS reflection matrix is diagonal with unit reflection amplitudes, and each element is controlled through a reflection phase shift.
- For user k, the optimal IRS beamforming vector aligns azimuth and elevation with the cascaded IRS spatial directions.
- Vertical and horizontal beam training are decoupled, and the analysis focuses on horizontal training because the vertical direction is common across users.
III. SINGLE-BEAM TRAINING
Conventional single-beam training samples Nx horizontal directions sequentially with the full IRS array. Although it identifies each user’s strongest beam by received power, its exhaustive search creates prohibitive training overhead for large IRSs.
- The horizontal spatial domain is divided into J = Nx equal sectors, each represented by a central beam direction.
- The single-beam codebook contains full-array codewords whose main-lobe width is 2/Nx and maximum beam gain is Nx.
- Each user selects the sampled direction nearest its effective cascaded spatial direction as the optimal beam index.
- Training sequentially sweeps all codewords, and each user chooses the direction producing maximum received signal power or SNR.
- Nx training symbols are required, creating large overhead and delay that is unsuitable for delay-sensitive or short-packet transmissions.
IV. MULTI-BEAM TRAINING
The proposed method partitions the IRS into sub-arrays that steer multiple directions simultaneously, then identifies each user’s best direction from received-power comparisons. Its two-phase design reduces training time while retaining directional beam-search capability.
- The IRS is divided into M adjacent sub-arrays, each containing L = Nx/M reflecting elements and its own direction-covering codebook.
- When all sub-arrays steer the same direction, their composite codeword equals the corresponding full-array single-beam codeword.
- Each sub-array has beam width 2M/Nx and beam gain Nx/M, trading narrower coverage and higher gain for wider coverage and lower gain.
- Different sub-arrays steer different directions simultaneously, with their directions changing across training symbols according to the multi-beam codebook.
- The method uses IRS beam sweeping followed by IRS beam identification, with users selecting directions through received-power or SNR comparisons.
1) IRS Beam Sweeping:
The proposed beam-sweeping phase divides IRS elements into sub-arrays and steers multiple beam-direction sets over several rounds. This design separates simultaneous beams while reducing training symbols as the number of sub-arrays increases.
- Beam-sweeping design: The method divides IRS reflecting elements into M sub-arrays and sweeps multiple sub-array beam-direction sets across 1+log2 M rounds.Each round uses training symbols reflected by different direction bins.
- Beam-sweeping design: For the first round, each bin contains M directions spaced by L, maximizing the minimum intra-bin distance to L.The spacing reduces inter-beam interference among simultaneously trained directions.
- Beam-sweeping design: For Nx = 32 and M = 4, the directions {1, 9, 17, 25} produce four well-separated beams with strong directionality after inter-beam interference.The example illustrates the spatial separation achieved by the bin construction.
- Beam-sweeping design: Subsequent rounds combine beam directions to narrow candidate sets while simultaneously searching subsets from different initial bins.Only one subset per relevant bin is tested, and multiple sub-arrays support simultaneous searches.
- Training cost: The total training-symbol count decreases monotonically as M increases and is smaller than single-beam training for M > 1.The reduction comes at the cost of using more sub-arrays.
2) IRS Beam Identification:
Beam identification uses received-power comparisons to progressively eliminate candidate directions until each user selects a single IRS beam. Increasing the number of sub-arrays accelerates training but can reduce identification accuracy through weaker sub-array gains and stronger interference.
- Candidate reduction: Each user first selects the bin with the largest received power, then examines only bins sharing directions with that initial candidate set.This avoids searching every beam direction individually.
- Candidate reduction: A received-power threshold determines whether the inspected subset contains the best beam direction, updating the candidate set after each round.The threshold is set to half the expected power associated with the best direction.
- Candidate reduction: The candidate-set size decreases logarithmically according to |b_I^k(r)| = M/(2^(r−1)) across the beam-identification rounds.The procedure progressively narrows the possible directions.
- Accuracy trade-off: Increasing M reduces training time but decreases sub-array beam gains and increases inter-beam interference, creating a trade-off with passive beamforming performance.Receiver noise and NLoS components can also cause the identified direction to differ from the optimal direction.
V. SIMULATION RESULTS
Simulations evaluate the proposed multi-beam training under varying sub-array counts and SNR, measuring training overhead, beam identification success, and average achievable rate. The method substantially reduces overhead while preserving performance for moderate sub-array counts, but excessive subdivision causes interference and accuracy loss.
- Simulation setup: The simulations use a 30 GHz system with 160 IRS elements, five users at 2 m, and 64 AP antennas.Results are averaged over 1500 Rician fading channel realizations.
- Evaluation metrics: The evaluation compares training overhead, success beam identification rate, and average achievable rate across sub-array counts and SNR.The proposed method is compared with conventional single-beam training and RH based multi-beam training.
- Results by sub-array count: 25% lower training overhead, 120 versus 160, is achieved with M = 2 while maintaining very close success and average-rate performance to single-beam training.This setting demonstrates a favorable reduction-performance trade-off.
- Results by sub-array count: 80 versus 160 training symbols are used with M = 4, while Psuc ≈92% at SNR = 46.4 dB and rate performance remains close to single-beam training.Increasing M from 2 to 4 further reduces the overhead while retaining high-SNR identification accuracy.
- Results by sub-array count: M = 8 substantially degrades beam identification accuracy and passive beamforming gain because of more severe inter-beam interference.The results show that increasing the number of sub-arrays creates a training-time versus performance trade-off.
- Benchmark comparison: The proposed method significantly outperforms RH based multi-beam training in beam identification accuracy and passive beamforming gain for M ∈ {2, 4, 8}.The RH benchmark with M = 2 is limited because its random coverage misses some beam directions.
VI. CONCLUSIONS
The paper proposes fast IRS reflect beam training by partitioning IRS elements into sub-arrays and coordinating their beam directions over training symbols. Users identify beams from received-power or SNR comparisons, reducing conventional training overhead while retaining comparable passive beamforming performance.
- Conclusion: The proposed method divides IRS elements into sub-arrays and designs sub-array beam directions over different training symbols.Users independently identify beams using received power or SNR comparisons.
- Conclusion: The method significantly reduces conventional single-beam training overhead while achieving comparable passive beamforming performance for data transmission.The approach can also be applied to vertical IRS beam training and AP transmit beam training in related settings.