Source-linked AI summary
Beamforming Tradeoffs for Initial UE Discovery in Millimeter-Wave MIMO Systems
Vasanthan Raghavan, Juergen Cezanne, Sundar Subramanian, Ashwin Sampath, Ozge Koymen
TL;DR
The paper addresses low-complexity initial UE discovery in millimeter-wave MIMO, where beamformer structure and physical-scattering connections are not well understood. It interprets beamformers through physical angles, develops directional beam-broadening codebooks, and shows substantial discovery-latency tradeoffs with slow performance roll-off.
Problem
The work addresses limited understanding of how beamformer structure connects to the underlying physical scattering environment, alongside the beam-broadening tradeoff between UE discovery latency and peak gain.
Method
The paper interprets optimal and phase-only beamformer structures through physical departure and arrival angles, then develops directional codebooks using a virtual subarray architecture.
Results
The proposed constructions remain within a couple of dB of the best beam-broadening tradeoff at all useful factors and substantially reduce UE discovery latency for many users with slow performance roll-off.
Takeaways & Limitations
Directional beamforming provides a practical low-complexity approach for trading received gain against UE discovery latency in millimeter-wave MIMO systems.
Abstract
from arXiv · showhide
Millimeter-wave MIMO systems have gained increasing traction towards the goal of meeting the high data-rate requirements in next-generation wireless systems. The focus of this work is on low-complexity beamforming approaches for initial UE discovery in such systems. Towards this goal, we first note the structure of the optimal beamformer with per-antenna gain and phase control and the structure of good beamformers with per-antenna phase-only control. Learning these beamforming structures in mmW systems is fraught with considerable complexities such as the need for a non-broadcast system design, the sensitivity of the beamformer approximants to small path length changes, etc. To overcome these issues, we establish a physical interpretation between these beamformer structures and the angles of departure/arrival of the dominant path(s). This physical interpretation provides a theoretical underpinning to the emerging interest on directional beamforming approaches that are less sensitive to small path length changes. While classical approaches for direction learning such as MUSIC have been well-understood, they suffer from many practical difficulties in a mmW context such as a non-broadcast system design and high computational complexity. A simpler broadcast solution for mmW systems is the adaptation of directional codebooks for beamforming at the two ends. We establish fundamental limits for the best beam broadening codebooks and propose a construction motivated by a virtual subarray architecture that is within a couple of dB of the best tradeoff curve at all useful beam broadening factors. We finally provide the received SNR loss-UE discovery latency tradeoff with the proposed constructions. Our results show that users with a reasonable link margin can be quickly discovered by the proposed design with a smooth roll-off in performance as the link margin deteriorates.
I. INTRODUCTION
The paper studies low-complexity beamforming for initial UE discovery in mmW MIMO, linking optimal beamformer structures to physical propagation directions. It compares directional and codebook-based approaches against practical constraints, showing near-optimal beam broadening and useful latency-performance tradeoffs.
- Motivation: MmW systems combine large bandwidths and dense antenna arrays with substantial propagation and shadowing losses that complicate coverage.Small wavelengths permit many antennas in a fixed aperture, while higher frequencies make losses more significant.
- Beamformer structures: The optimal RSV beamformer corresponds to beam steering across paths with appropriate power allocation and phase compensation.In general, realizing it requires per-antenna gain and phase control, whereas good phase-only structures are also characterized.
- Directional beamforming: Directional beamforming along the dominant path offers a low-complexity proxy whose performance incurs only minimal loss relative to optimal structures in sparse mmW channels.The approach is motivated by the physical interpretation of RSV structures and is not expected to be optimal for nonsparse cellular channels.
- Practical constraints: RSV learning, MUSIC, and related direction-learning methods face sensitivity to path-length changes and practical issues including non-broadcast operation, low-link-margin performance, and computational demands.At mmW frequencies, small path changes can produce full-cycle phase changes across paths.
- Codebook design: The paper studies globally known directional codebooks that trade UE discovery latency against peak beam gain through beam broadening.It establishes fundamental limits and realizable constructions motivated by a virtual subarray architecture.
- Results: The proposed constructions remain within a couple of dB of the best tradeoff curve at all useful beam broadening factors and provide substantial latency reductions with slow performance roll-off as link margin deteriorates.Their simplicity, adaptability, and scalability to beam refinement support initial UE discovery in practical mmW implementations.
II. SYSTEM SETUP
The system models downlink mmW MIMO beamforming over a geometric multipath channel, with transmit and receive beamformers applied to the channel matrix. The formulation specifies antenna, steering-vector, noise, normalization, and ideal-channel assumptions.
- System model: The downlink model uses an Nt-antenna transmitter, an Nr-antenna receiver, transmit beamformer f, and receive combiner g.The channel is represented by an Nr × Nt matrix H.
- Evaluation objective: The analysis focuses on rank-1 signaling over H and compares beamforming schemes through received SNR.The paper explicitly sets aside normalization technicalities when comparing schemes.
- Signal model: The received signal uses pre-beamforming SNR ρf, zero-mean unit-energy symbol s, and proper complex Gaussian noise n at the UE.The noise covariance is the identity matrix.
- Channel model: The channel follows an extended Saleh–Valenzuela model with L scattering clusters or paths and no near-field impairments.Each path has a complex gain and transmit and receive array steering vectors.
- Normalization: The path normalization preserves standard MIMO channel-power normalization, while the paper notes that massive-MIMO scaling can require a physical-law correction.That correction is not expected to alter the comparison between beamforming schemes.
- Array geometry: For uniform linear arrays, steering vectors correspond to azimuthal AoA and AoD, with receive and transmit inter-antenna spacings dR and dT.The formulation assumes 90° elevation angles and uses CPO notation to capture constant phase offsets.
III. OPTIMAL BEAMFORMING AND RSV LEARNING
This section characterizes optimal beamforming through singular vectors and path steering vectors, then motivates directional beamforming as a lower-complexity alternative for sparse mmW channels. It also highlights low-link-margin conditions and the sensitivity of RSV learning to practical channel changes.
- Objective: The study compares optimal channel-RSV beamforming with low-complexity directional beamforming using received SNR as the performance metric.Received SNR captures achievable rate and symbol-estimation error probability.
- Link margin: Low pre-beamforming SNRs are expected in mmW systems, with an illustrative range of −30 to −15 dB for Nt = 64.The example assumes 40–55 dBm EIRP and 64-level time-repetition gain.
- Optimal beamforming: With full amplitude and phase control, the optimal transmit beamformer is the dominant right singular vector, while the receive combiner is a matched filter.The beamformers are unit-norm when energy is fully used.
- Optimal beamforming: Theorem 1 states that all eigenvectors of H^H H, including the optimal transmit beamformer, lie in the span of the transmit steering vectors.The optimal receive combiner similarly lies in the span of the receive steering vectors.
- Directional interpretation: For sparse mmW channels, the steering-vector structure supports directional beamforming because optimal beamformers coherently combine dominant paths through phase compensation.This contrasts with dense cellular channels, where many paths wash away the steering-vector Fourier structure.
- Technical interpretation: Theorem 1 provides a non-unitary basis for the positive-eigenvalue eigenspace when L ≤ Nt; for L > Nt, the steering vectors span the eigenspace without forming a basis.This technical distinction does not change the physical steering-vector interpretation.
- Two-path structure: In a two-path example, optimal power allocation approaches a single dominant path when steering vectors are orthogonal and proportional allocation when they are nearly parallel.The same limiting behavior is described for the transmit-side coefficient βopt as v1 and v2 vary from orthogonal to parallel.
B. Optimal Beamforming with Phase-Only Control
With phase-only control, the optimal transmit beamformer has equal-magnitude entries, while two practical candidates approximate the received-SNR optimum using phase-only beamforming at both ends.
- Practical constraint: The phase-only problem is motivated by practical power amplifiers that disallow per-antenna power control.
- Optimal structure: Theorem 2 identifies the optimal phase-only transmit beamformer as an equal-gain transmission scheme.
- Candidate solutions: Two good phase-only solutions are proposed for received-SNR maximization, including the equal-gain RSV with its matched-filter combiner.
- Candidate solutions: The second solution is characterized through the channel’s column vectors and provides a phase-only beamforming solution that can be designed with simple uplink training.
C. Issues with RSV Learning
RSV learning is difficult for millimeter-wave initial discovery because it is sensitive to channel perturbations and requires iterative, non-broadcast procedures. Numerical results show that noise averaging helps more than iteration at low pre-beamforming SNR, but noisy power iteration can remain poor.
- Noisy learning: At low pre-beamforming SNR, noise averaging is more important than beamformer iteration, yet noisy power iteration performs poorly for many users because noise is amplified through iteration.
- System-level issues: RSV learning is not directly amenable to common broadcast downlink discovery because each user requires bidirectional iteration.
- System-level issues: Sampling antennas individually can slow iteration under RF-chain constraints, while calibration, phase coherence, and TDD reciprocity are also required.
- Power-efficiency tradeoff: Combining multiple directional beams incurs PA inefficiency: median PAR loss exceeds 2 dB for Nt ≥8, outweighing the RSV gain relative to directional beamforming of less than 1 dB.
- Sensitivity to perturbations: RSV reconstruction is highly sensitive to relative path phases, whereas the directional scheme remains approximately stable when the dominant-path phase changes.
IV. DIRECTIONAL BEAMFORMING AND DIRECTION LEARNING
Directional beamforming offers a low-complexity alternative for initial discovery by learning dominant AoAs/AoDs, while MUSIC provides a more computationally demanding subspace-based direction-learning approach. Directional beamforming stays close to optimal over broad channel conditions, whereas MUSIC is weak at low link margins.
- Directional beamforming: The results motivate learning the dominant AoAs/AoDs along which the UE and millimeter-wave base station should beamform for initial discovery.
- Directional beamforming: Directional beamforming incurs less than 1 dB loss for over 50% of users and no more than 2.5 dB loss for up to 90% of users, even with L = 5 clusters.
- MUSIC: MUSIC estimates signal directions from the signal and noise subspaces of a received covariance matrix, locating peaks of a pseudospectrum.
- MUSIC: The evaluated MUSIC scheme uses bidirectional covariance-based training, allocating samples between uplink AoD and downlink AoA learning.
- MUSIC limitations: MUSIC performs poorly at low link margins but improves as the link margin increases, partly because reliable covariance estimation is difficult with few samples and large arrays.
- MUSIC limitations: MUSIC also requires a non-broadcast design and high computational complexity, making its utility for UE discovery questionable despite possible usefulness for later beam refinement.
B. Beam Broadening for Initial UE Discovery
The paper frames beam broadening as a gain–coverage tradeoff for initial UE discovery and proposes virtual-subarray beams that approach the best achievable tradeoff. Different codebook widths then exchange beamforming gain against discovery latency according to UE link margin.
- Beam-broadening tradeoff: Broad beams shorten discovery latency, whereas narrow beams preserve link margin for cell-edge or heavily blocked UEs.Intermediate beams trade between these two properties.
- Virtual-subarray construction: The virtual-subarray construction combines subarray patterns to enlarge coverage while limiting peak-gain loss from reduced effective aperture.The design partitions the transmitter array into virtual subarrays aimed at selected directions.
- Virtual-subarray construction: For small beamspace coverage areas, f = 0 is optimal; as Ω0 increases, the middle-subarray length decreases and beam orientation moves away from f = 0.These trends are reported for the optimized M = 2, 3, and 4 constructions.
- Achievable tradeoff: The M = 4 subarray scheme is within a couple of dB of the upper bound across useful beam-count regimes for Nt = 64.The comparison covers a 120° field of view.
C. Learning Dominant Directions via Beam Sweep with Broadened Beam Codebooks
Beam sweep with broadened directional codebooks provides a broadcast-compatible way to learn dominant directions while exposing an explicit received-SNR versus discovery-latency tradeoff. Codebook size can be matched to UE link margin, but weaker users require longer sweeps and remain more vulnerable to loss.
- Broadcast beam sweep: A codebook of beamforming vectors at the MWB and UE enables broadcast SNR estimation for initial UE discovery.The selected vectors can then support subsequent beamforming or beam refinement.
- Broadcast beam sweep: The beam sweep approach performs worse than noisy power iteration and MUSIC, but its simplicity yields a better system design.The comparison is made at the same ρf value.
- Codebook-size tradeoff: Nmwb = 20 can outperform Nmwb = 40 at small coverage levels because codebook choice changes the steepness of the achievability curve.The tradeoff is evaluated across beam-broadening factors and shifted versions.
- Codebook-size tradeoff: Higher codebook size improves SNRrx, while users with better link margin can be discovered with lower latency using smaller codebooks.Users with worse link margin require larger codebooks and experience a smooth latency roll-off.
- Limitations and mitigations: The beam sweep can suffer significant performance loss for cell-edge or blocked users, motivating coding, MWB densification, or low-frequency overlays.These alternatives are presented as possible enhancements for such users.
A. Finite-Bit Phase Shifters
The study evaluates finite-bit phase shifters against an infinite-precision ideal and finds that practical quantization can preserve performance. It also examines how receiver-array size and channel dynamics affect directional and RSV-type schemes.
- Phase quantization: The practical analysis assumes infinite-precision phase shifters, although actual MWB and UE beamformers use finite-bit phase shifters.The reported quantization study indicates that B = 3 or B = 4 bits can be sufficient.
- Phase quantization: B = 4-bit phase shifters incur less than 0.25 dB SNRrx loss relative to infinite-precision optimal beamforming.The result is reported for L = 2, Nr = 4, and Nt = 64.
- Array architecture: The study’s Nr = 4 UE-side assumption is presented as realistic because multiple UE subarrays are needed to cover sectors despite their aperture constraints.The paper contrasts this with the MWB’s softer aperture constraints and larger antenna arrays.
- Array size: Directional schemes improve relative to the optimal scheme as Nr increases, with absolute SNRrx also benefiting from increased array gain.The comparison uses Nt = 64 with L = 2 and L = 5 clusters and Nr ∈ {4, 8, 12, 16}.
- Channel dynamics: Short mmW coherence times favor directional approaches over RSV-type schemes because RSV implementation requires bidirectional feedback.Directional performance is described as essentially stable under small path-length changes, whereas RSV learning is not robust to them.
D. Planar Antenna Arrays
The paper develops its beam-broadening results using ULAs for tractability but states that the construction can extend to planar arrays. It positions directional codebooks as a broadcast-compatible alternative to more complex direction-learning and RSV approaches, while identifying unresolved scan-latency questions.
- Planar-array extension: The ULA geometry is used primarily to illustrate beam-broadening tradeoffs, with more general steering vectors available for other array geometries.The proposed development is stated to extend easily to planar arrays.
- Directional information: Omnidirectional scanning does not reveal AoD after discovery, whereas beam sweep provides AoD information that can support subsequent data delivery.Broadened codebooks still leave AoD uncertainty requiring later refinement.
- Open design question: A detailed latency study for broad-beam or omnidirectional discovery followed by AoD refinement remains of interest.The paper also notes that omnidirectional scanning could prevent discovery of multiple MWBs in small-cell settings.
- System-level comparison: RSV learning is sensitive to small path-length changes, while MUSIC requires a non-broadcast system design that may be unattractive at the system level.These constraints motivate learning dominant directions instead.
- System-level comparison: Directional codebooks provide a broadcast solution for initial UE discovery, but their performance is slightly poorer than RSV learning and MUSIC.The paper emphasizes simplicity as outweighing this suboptimality for the target system design.
APPENDIX
The appendix characterizes optimal beamformers through eigenvector combinations and proves equal-gain transmission optimality under the stated constraints. It also derives phase-only and beam-broadening bounds, while noting limits on closed-form analysis for larger beam counts.
- fopt is a linear combination of v1, · · · , vL, while gopt is a linear combination of u1, · · · , uL.
- The eigenvector representation differs between L ≤ Nt and L > Nt because the number of distinct eigenvectors of X is bounded by L and Nt, respectively.
- The receive-side optimum is attained by a matched-filter structure, and the resulting choice satisfies the constraints and the upper SNR bound.
- Equal-gain transmit beamforming is optimal under the considered per-antenna constraints.
- For recursive phase-only beamformers, each objective term is maximized, although the effect on their sum remains unclear.
- Closed-form worst-case beamforming-gain expressions become difficult for J = 3, 4 and impossible for J > 4.