PaperlayerView exact revision on arXiv

Source-linked AI summary

An Overview of Enhanced Massive MIMO with Array Signal Processing Techniques

Mingjin Wang, Feifei Gao, Shi Jin, Hai Lin

arXiv:1907.09944v3eess.SPcs.IT

TL;DR

Massive MIMO research has largely used communications-oriented designs, while sparse mmWave arrays motivate exploiting physical channel parameters more directly. This paper surveys array-signal-processing models and transceiver techniques, highlighting parameter-based designs, frequency-insensitive reciprocity, and explanations of effects such as beam squint. It concludes that these models can shorten training and improve estimation, while array-response errors and FDD calibration remain open challenges.

  • Problem

    Massive MIMO needs designs that exploit its large half-wavelength arrays and sparse physical structure beyond conventional channel-state-information approaches.

    Source: §I · p. 1, 1) CCM Based Method · p. 5, 2) DFT Based Algorithms · p. 6

  • Method

    The paper surveys array-signal-processing techniques using sparse and low-rank channel models and physical parameters such as angle, delay, and Doppler.

    Source: §I · p. 2, 1) CCM Based Method · p. 5, §IX · p. 13

  • Results

    The overview reports that parameter-based designs can reduce training overhead, improve estimation accuracy, support frequency transfer, and explain spatial- and frequency-wideband effects.

    Source: §IX · p. 13

  • Takeaways & Limitations

    Angle, delay, and Doppler parameters can guide channel estimation, synchronization, precoding, and scheduling across frequency bands, including FDD settings.

    Source: F. Uplink Downlink Reciprocity for… · p. 4, §IX · p. 13

    Parameter-based channel models are sensitive to array-response errors, and calibrating large-scale arrays in FDD systems remains an open problem.

    Source: §IX · p. 13

Abstract

from arXiv · show

In the past ten years, there have been tremendous research progresses on massive MIMO systems, most of which stand from the communications viewpoint. A new trend of investigating massive MIMO, especially for the sparse scenario like millimeter wave (mmWave) transmission, is to re-build the transceiver design from array signal processing viewpoint that could deeply exploit the half-wavelength array and provide enhanced performances in many aspects. For example, the high dimensional channel could be decomposed into small amount of physical parameters, e.g., angle of arrival (AoA), angle of departure (AoD), multi-path delay, Doppler shift, etc. As a consequence, transceiver techniques like synchronization, channel estimation, beamforming, precoding, multi-user access, etc., can be re-shaped with these physical parameters, as opposed to those designed directly with channel state information (CSI). Interestingly, parameters like AoA/AoD and multi-path delay are frequency insensitive and thus can be used to guide the down-link transmission from uplink training even for FDD systems. Moreover, some phenomena of massive MIMO that were vaguely revealed previously can be better explained now with array signal processing, e.g., the beam squint effect. In all, the target of this paper is to present an overview of recent progress on merging array signal processing into massive MIMO communications as well as its promising future directions.

I. INTRODUCTION

This introduction motivates merging array signal processing with massive MIMO, particularly for half-wavelength and sparse mmWave arrays, and frames the paper as an overview of resulting channel models and transceiver designs.

  • Massive MIMO scales base-station arrays to hundreds or thousands of antennas to increase capacity, spectrum efficiency, and power efficiency.Source: §I · p. 1
  • Half-wavelength antenna arrays make phased-array signal processing relevant to wireless communications while maintaining an implementable array aperture.Source: §I · p. 1
  • Large arrays and sparse mmWave propagation enable precise spatial identification and narrow beams for multiplexing users.Source: §I · p. 1
  • The paper reviews array-signal-processing designs for channel estimation, synchronization, beamforming, interference control, and multiple access rather than duplicating communications-oriented treatments.Source: §I · p. 2
  • The array-processing channel model represents multipath through angle, gain, delay, and Doppler parameters and builds the array response across antennas and frequencies.Source: A. Channel Modeling · pp. 2–3
  • In OFDM, frequency-dependent steering vectors for the same AoA produce the beam squint effect across subcarriers.Source: A. Channel Modeling · p. 3

B. Spatial Narrowband Modeling

Spatial narrowband modeling applies when propagation delays across the array are negligible relative to the symbol period, allowing all antennas to observe approximately synchronized waveforms. Under this modeling, beam squint disappears, whereas general time-domain modeling loses steering-vector structure because antenna waveforms are unsynchronized.

Narrowband condition

  • When propagation delay across the array is negligible compared with the symbol period, all antennas approximately observe the same synchronized waveform.This occurs when the signal bandwidth or antenna count is small enough that ∆τm,p ≪ T for every path and antenna.Source: B. Spatial Narrowband Modeling · p. 3
  • This approximation is termed narrowband modeling in array signal processing, although the passage notes that a more accurate name would be preferable.Source: B. Spatial Narrowband Modeling · p. 3

Beam squint

General modeling

C. Frequency Narrowband Modeling

The channel model distinguishes frequency-narrowband and dual-wideband behavior, while highlighting that beam squint can persist even when multipath delay is negligible. Physical parameter reciprocity can then support downlink reconstruction in FDD systems without conventional downlink training overhead.

C. Frequency Narrowband Modeling

G. Low-Rank CCM Property of Massive MIMO

Array signal processing links the angular spread of multipath arrivals to the channel covariance matrix rank. Small angular spread produces low-rank covariance, particularly in elevated-base-station and mmWave settings.

G. Low-Rank CCM Property of Massive MIMO

III. PARAMETERS ESTIMATION

The parameter-estimation discussion contrasts full-channel estimation with estimating physical channel parameters. Massive arrays and sparse mmWave channels make precise angle estimation practical, while training-based methods complement statistically demanding blind algorithms.

III. PARAMETERS ESTIMATION

  • In rich scattering, full-rank channel covariance leaves conventional LS or LMMSE estimation necessary, creating a large downlink estimation burden.Source: §III · p. 4
  • Low-rank covariance and channel sparsity in high-altitude or mmWave systems motivate reduced-overhead methods such as VCR and compressive sensing.Source: §III · p. 4
  • VCR suffers grid mismatch and power leakage, causing a channel-estimation error floor at high SNR.Source: §III · p. 5
  • Estimating physical parameters instead of the channel becomes feasible with massive arrays and sparse channels, although half-wavelength spacing is required for unambiguous AoA estimation.Source: §III · p. 5
  • Blind AoA methods use covariance-matrix EVD; MUSIC and ESPRIT provide higher resolution than DFT-based approaches, while Root-MUSIC outperformed ESPRIT under good radio conditions.Source: A. Blind Parameters Estimation · p. 5
  • Joint AoA-delay estimation improves reconstructed-channel accuracy by exploiting both spatial and temporal multipath diversity.Source: A. Blind Parameters Estimation · p. 5
  • Blind algorithms offer high spectral efficiency but require many received signals, limiting them to slow time-varying channels and increasing complexity.Source: A. Blind Parameters Estimation · p. 5
  • Training sequences enable more efficient parameter estimation in cooperative wireless terminals, while blind methods can supplement parameter tracking.Source: B. Training Based Method · p. 5

1) CCM Based Method:

CCM-based methods exploit low-rank channel structure and asymptotic orthogonality between users with non-overlapping angular supports to reduce channel dimensions and pilot overhead. DFT-based representations offer beamspace sparsity, but grid mismatch causes power leakage and estimation costs.

CCM Based Method

  • CCM eigen-decomposition reduces each user’s channel representation from M dimensions to r_k dominant eigenvector dimensions.The reduced representation uses the nonzero eigenvalues and corresponding subspace eigen-matrix.Source: 1) CCM Based Method · p. 5
  • CCMs of users with non-overlapping angular supports become asymptotically orthogonal as M grows.This property supports interference-free estimation when users reuse the same training sequence.Source: 1) CCM Based Method · p. 5
  • Non-overlapping angular supports can remove pilot contamination even when users employ the same training sequence.JSDM and MMSE-based training exploit this spatial separation.Source: 1) CCM Based Method · pp. 5–6
  • Low-rank CCM processing can reduce downlink pilot contamination, training, and feedback overhead, but requires difficult high-dimensional CCM estimation and costly EVD.The limitation becomes more severe when multiple users’ CCMs must be estimated and fed back separately.Source: 1) CCM Based Method · p. 6

DFT Based Algorithms

  • DFT beamspace representations approximate channels with P nonzero supports associated with path angles.The DFT produces sparse beamspace coefficients for massive-MIMO channels.Source: 2) DFT Based Algorithms · p. 6
  • DFT angle resolution is 1/M, so mismatch between the DFT grid and true angles leaks power into neighboring supports.The resulting grid mismatch increases channel-estimation error or training overhead.Source: 2) DFT Based Algorithms · p. 6
  • Zero-padding and spatial rotation are proposed to improve angle estimation or concentrate leaked power into fewer grids.These approaches target real rather than strictly on-grid AoA values.Source: 2) DFT Based Algorithms · p. 6
  • Two-dimensional DFT methods estimate on-grid angles and delays for wideband channels, while angle-delay reciprocity can simplify FDD downlink estimation.The cited dual-wideband approach combines 2D DFT with 2D rotation before using reciprocity.Source: 2) DFT Based Algorithms · p. 6

3) Compressive Sensing Based Algorithms:

Compressive sensing formulates sparse massive-MIMO channel estimation through angle dictionaries and sparse recovery, with on-grid and off-grid models addressing different parameter assumptions. The surveyed algorithms include convex, greedy, atomic-norm, and Bayesian approaches, each with distinct accuracy or complexity trade-offs.

Compressive Sensing Formulation

  • Compressive sensing exploits sparse mmWave channel structure to recover physical parameters from measurements.The channel is represented using steering-vector dictionaries over angle grids or continuous off-grid parameters.Source: 3) Compressive Sensing Based Algorithms · pp. 6–7
  • On-grid CS divides the continuous angle space into L uniformly spaced points and models a P-sparse coefficient vector.The model assumes true parameters align with the selected grid, with finer grids available when L ≥ M.Source: 3) Compressive Sensing Based Algorithms · p. 6
  • Off-grid CS jointly estimates sparse path gains and continuous angle parameters using an off-grid parametric dictionary.This removes the requirement that true channel parameters coincide exactly with predefined grid points.Source: 3) Compressive Sensing Based Algorithms · pp. 6–7

Convex Relaxation Algorithms

  • Convex relaxation methods replace the l_0 sparsity measure with an l_1 approximation, including LASSO-based channel and array-diagnosis methods.Applications include sparse virtual-channel estimation and joint AoA/AoD and phase-shift estimation.Source: 3) Compressive Sensing Based Algorithms · p. 7

Greedy Algorithms

  • Greedy algorithms iteratively select dictionary columns correlated with measurements and update the residual using least-squares support estimates.OMP can reduce training and feedback overhead, but its conditions are more restrictive than the restricted isometry condition.Source: 3) Compressive Sensing Based Algorithms · p. 7

Atomic Norm Algorithms

  • Atomic norm denoising provides super-resolution channel-parameter estimation and can be solved through semidefinite programming.Decoupled atomic-norm minimization reduces problem size in two-dimensional angle scenarios.Source: 3) Compressive Sensing Based Algorithms · p. 7
  • Atomic norm denoising becomes slow for large-scale problems, motivating ADMM acceleration of its semidefinite program.The limitation is computational rather than a stated loss of estimation capability.Source: 3) Compressive Sensing Based Algorithms · p. 7

Bayesian Algorithms

  • Sparse Bayesian learning infers unknown channel parameters using sparse priors and can incorporate data-adaptive priors for off-grid estimation.The surveyed applications include narrowband AoA estimation and frequency-wideband estimation using band-occupation structure.Source: 3) Compressive Sensing Based Algorithms · pp. 7–8

C. Prospects of Parameters Estimation

Parametric channel estimation reduces massive-MIMO complexity by estimating physical parameters, but its accuracy depends strongly on path-number and parameter estimates. Synchronization remains challenging in mmWave systems because Doppler spread is much larger than in classical narrowband channels.

Prospects of Parameters Estimation

  • Parametric channel modeling estimates physical channel parameters to reduce the computational complexity of massive-MIMO processing.The approach is sensitive to the number of multipaths and the accuracy of parameter estimates.Source: C. Prospects of Parameters Estimation · p. 8
  • Incorrect path-number estimates can produce distinct angle errors, especially in discrete sparse scenes with finite incident paths.MDL and AIC are commonly used for path-number estimation but have upper precision limits.Source: C. Prospects of Parameters Estimation · p. 8
  • Low-complexity, high-resolution path-number and parameter-estimation algorithms remain an open challenge.The paper identifies this as a direction for further research.Source: C. Prospects of Parameters Estimation · p. 8

Synchronization

  • In mmWave massive MIMO, Doppler spread is orders of magnitude larger than in classical narrowband wireless channels and can deteriorate system performance.This makes synchronization a distinct challenge alongside channel estimation.Source: §IV · p. 8
  • OFDM synchronization must account for delay dispersion through cyclic-prefix length and frequency dispersion through symbol duration.Beam squint can require additional cyclic-prefix length beyond the usual delay-span setting.Source: §IV · p. 8
  • Per-beam synchronization reduces wideband MIMO delay and Doppler spreads approximately by the number of antennas.Angle-domain frequency synchronization can also estimate each user’s carrier-frequency offset through joint spatial-frequency alignment.Source: §IV · p. 8

V. BEAMFORMING AND PRECODING

Array-signal-processing-based beamforming uses spatial channel structure to design hybrid analog-digital precoders for large, sparse arrays. Hybrid architectures reduce RF-chain requirements, while anglespace and beamspace designs trade continuous angle matching against lower-complexity codebook selection.

Beamforming and Precoding

  • Array-signal-processing beamforming is principally aimed at large arrays and sparse environments rather than duplicating communications-viewpoint precoding designs.The section focuses on beamforming methods that exploit array structure.Source: §V · p. 8

Hybrid Precoding

  • Fully digital beamforming is impractical for massive MIMO because one RF chain per antenna increases hardware cost, complexity, and power consumption.Hybrid analog-digital beamforming lets M antennas share M_RF much smaller than M RF chains.Source: A. Hybrid Precoding · p. 8
  • The hybrid design combines a constant-modulus analog precoder with a digital precoder optimized for spectral efficiency.The analog matrix has dimensions M × M_RF and the digital vector has dimensions M_RF × 1.Source: A. Hybrid Precoding · p. 8
  • Anglespace beamforming matches analog-precoder columns to true steering vectors, while beamspace beamforming selects beams from a fixed codebook.Anglespace designs represent continuous directions; beamspace designs use predetermined directions.Source: A. Hybrid Precoding · p. 8
  • Beamspace precoding has lower complexity but suffers power leakage because it treats angles as grid-aligned.Anglespace and beamspace methods are illustrated as non-orthogonal and orthogonal angle-space beamforming, respectively.Source: A. Hybrid Precoding · pp. 8–9, Figure 3 · p. 9
  • Hybrid beamforming can reduce RF-chain count while approaching fully digital performance when carefully designed.This result is reported for multiuser massive MIMO systems.Source: A. Hybrid Precoding · p. 9

B. Interference Control

Array signal processing reframes interference control as shaping beam-pattern lobes and nulls, while training beams balance main-lobe response against ripple. In OTFS massive MIMO, structured angle-delay-Doppler sparsity supports parameter-efficient channel estimation but motivates off-grid methods.

  • Interference cancellation can be viewed as suppressing sidelobe leakage by forming physical nulls toward undesired users’ angles of arrival.Source: B. Interference Control · p. 9
  • Training-beam design balances maximizing main-lobe magnitude response against limiting ripple to a small positive level ε.The competing objectives cannot generally both be optimized perfectly for finite antenna arrays.Source: Figure 4 · p. 9, B. Interference Control · p. 9
  • Least-squares interference control may leave large ripple and performance loss at nonquantized beamspace points, whereas maximin optimization seeks a flatter array-gain response.Source: B. Interference Control · p. 9

OTFS channel structure

  • OTFS massive MIMO represents the downlink channel as a 3D tensor structured across angle, delay, and Doppler dimensions.Source: §VI · pp. 9–10
  • The channel is sparse along delay, block-sparse along Doppler, and burst-sparse along angle, enabling estimation from limited parameters with low overhead.Source: Figure 5 · p. 10, §VI · p. 10
  • Existing OTFS channel estimation uses on-grid methods that retain power leakage, motivating future off-grid parametric algorithms.Source: §VI · p. 10

VII. MULTIPLE ACCESS

Massive MIMO enables multiple access schemes based on distinguishable physical parameters such as angle, delay, and Doppler. These schemes can be combined flexibly, but they do not create independent new degrees of freedom when they arise from the same Fourier-domain relationships.

Angle Division Multiple Access

Delay Division Multiple Access

Path Division Multiple Access

VIII. ARTIFICIAL INTELLIGENCE METHODS FOR MASSIVE MIMO

The paper surveys learning-based methods that incorporate channel structure and physical parameters into massive MIMO estimation and beamforming. It presents array signal processing as a basis for reducing training overhead and adapting transceiver techniques across frequency bands, while identifying array-response errors and calibration as open issues.

Learning-based channel estimation

Learning-based beamforming

  • Array signal processing reduces unknown parameters, shortening training overhead and improving estimation accuracy relative to approaches that simply assume channel sparsity.Source: §IX · p. 13
  • Frequency-insensitive angle, delay, and Doppler parameters allow transceiver techniques such as channel estimation, synchronization, precoding, and scheduling to transfer across frequency bands.Source: §IX · p. 13

Open issues

  • Parameter-based channel models remain sensitive to array-response errors, and calibration of large-scale arrays in FDD systems remains an open problem.Source: §IX · p. 13

Related papers

Loading 1907.09944v3…