PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Massive MIMO 2.0: Understanding spatial correlation, interference suppression, and pilot contamination

Luca Sanguinetti, Emil Björnson, Jakob Hoydis

arXiv:1904.03406v3eess.SPcs.IT

TL;DR

The article addresses how spatial correlation, multicell interference, and pilot contamination challenge conventional Massive MIMO analyses. It explains multicell-aware processing and spatial-statistics-based limits, showing that optimized processing can avoid pilot-contamination asymptotic limits under suitable correlation conditions.

  • Problem

    Conventional Massive MIMO analyses often assume uncorrelated channels and single-cell processing, overlooking practical spatial correlation and multicell interference.

  • Method

    The article analyzes spatially correlated channels and multicell-aware signal processing, including lower-complexity schemes based on channel statistics.

  • Results

    Optimized processing with spatially correlated channels is not asymptotically limited by pilot contamination; M-MMSE SE can grow as log2(M) under asymptotically linearly independent correlation matrices.

  • Takeaways & Limitations

    Massive MIMO performance analysis and design should account for spatial correlation, multicell interference suppression, and the availability of channel statistics.

  • Takeaways & Limitations

    The unbounded log2(M) SE result requires the UE’s correlation matrix to be asymptotically linearly independent of those of pilot-contaminating UEs.

Abstract

from arXiv · show

Since the seminal paper by Marzetta from 2010, Massive MIMO has changed from being a theoretical concept with an infinite number of antennas to a practical technology. The key concepts are adopted in 5G and base stations (BSs) with $M=64$ full-digital transceivers have been commercially deployed in sub-6\,GHz bands. The fast progress was enabled by many solid research contributions of which the vast majority assume spatially uncorrelated channels and signal processing schemes developed for single-cell operation. These assumptions make the performance analysis and optimization of Massive MIMO tractable but have three major caveats: 1) practical channels are spatially correlated; 2) large performance gains can be obtained by multicell processing, without BS cooperation; 3) the interference caused by pilot contamination creates a finite capacity limit, as $M\to\infty$. There is a thin line of papers that avoided these caveats, but the results are easily missed. Hence, this tutorial article explains the importance of considering spatial channel correlation and using signal processing schemes designed for multicell networks. We present recent results on the fundamental limits of Massive MIMO, which are not determined by pilot contamination but the ability to acquire channel statistics. These results will guide the journey towards the next level of Massive MIMO, which we call ``Massive MIMO 2.0''.

I. INTRODUCTION

Massive MIMO improves spectral efficiency through spatial multiplexing, but conventional analyses often assume uncorrelated channels and single-cell processing. This tutorial defines a practical Massive MIMO framework and motivates Massive MIMO 2.0 through spatial correlation, multicell interference suppression, and statistical channel knowledge.

  • Motivation: Spectral efficiency is limited by signal attenuation and interference, motivating higher-gain antennas, scheduling, and spatial signal processing.SE increases with the SINRs of communication links.
  • Motivation: Massive MIMO uses many active BS antennas to spatially multiplex multiple UEs on the same time-frequency resource.Uplink receive combining and downlink transmit precoding combat attenuation and interference, improving per-cell spectral efficiency.
  • Limitations of conventional analyses: Most Massive MIMO analyses assume spatially uncorrelated channels and single-cell signal processing because these assumptions make analysis and optimization tractable.The paper identifies these assumptions as the basis for three major practical caveats.
  • Limitations of conventional analyses: Practical channels are spatially correlated, multicell processing can provide large gains without BS cooperation, and pilot contamination becomes a critical interference issue under simplifying assumptions.The tutorial reviews theoretical work addressing some or all of these caveats.
  • Practical definition: The article defines Massive MIMO as a synchronous multicell TDD network with M ≥64 fully digital BS antennas, K ≥8 multiplexed UEs per cell, and M/K > 1.The definition also requires at least two cells and linear combining and precoding schemes.
  • Scope: The article focuses on TDD and fully digital transceivers, excludes analog or hybrid processing, and simplifies the UE model to single-antenna terminals.Multiple-antenna UEs can be represented as virtual UEs transmitting separate data streams.
  • Transmission protocol: TDD coherence blocks support UL pilots, UL data, and DL data, while channel reciprocity enables BS channel estimation from UL pilots.The coherence block is approximated as time-invariant and flat-fading.

B. Basic signal processing

Uplink detection combines channel estimates with receive combining to separate desired signals from interference and noise. MR maximizes desired-signal power, whereas ZF uses spatial dimensions to suppress intra-cell interference.

  • Receive combining: The received uplink signal is processed with a combining vector selected to separate the desired transmission from interference and noise.The vector is multiplied with the received signal to obtain the detected signal.
  • MR and ZF: MR and ZF are popular combining choices: MR maximizes desired-signal power, while ZF partially suppresses intra-cell interference.ZF has higher complexity because it inverts a K × K matrix.
  • Spectral-efficiency bound: The use-and-then-forget bound provides an achievable spectral-efficiency expression applicable with any combining scheme and channel estimator.Channel estimates design the combining vectors and are then effectively forgotten during signal detection.
  • Interference terms: With MR and uncorrelated Rayleigh fading, desired-signal power and non-coherent interference scale differently from coherent interference caused by pilot-sharing UEs.Desired-signal power increases linearly with M; coherent interference also increases linearly with M, while non-coherent interference is independent of M.
  • ZF scaling: ZF reduces some non-coherent interference but sacrifices K spatial dimensions, causing signal and coherent-interference terms to scale with M − K.This closed-form structure does not generally extend beyond the uncorrelated-fading setting.

C. Basic spectral efficiency analysis

The basic spectral-efficiency analysis evaluates uplink performance in a multicell running example and explains how antenna scaling, combining, and pilot reuse affect the result. Under uncorrelated Rayleigh fading, pilot contamination remains the asymptotic limiting interference term.

  • Simulation setup: The running example uses 16 asymmetric wrap-around cells, averages 100 UE distributions and 50 channel realizations, and reserves fK coherence-block samples for pilots.The remaining samples are used for uplink data transmission, with τu = τc − fK.
  • Antenna scaling: 41.94 bit/s/Hz versus 5.21 bit/s/Hz: ZF uplink sum SE rises as M increases from 10 to 250.For MR, the corresponding increase is from 7.21 to 25.89 bit/s/Hz.
  • Antenna scaling: ZF exceeds 10× SE improvement for M ≥64, while MR reaches 5× improvement, supporting order-of-magnitude gains from Massive MIMO.Both schemes provide higher per-cell SE than basic LTE for M ≥16 in the reported comparison.
  • Interference behavior: At M = 100 and f = 1, non-coherent interference dominates with MR, whereas coherent interference dominates with ZF because ZF suppresses non-coherent interference.ZF suppresses the non-coherent interference by 30 dB while sacrificing a few dB of signal power.
  • Pilot reuse: Increasing the pilot reuse factor f can reduce coherent interference and improve estimation, but the reported SE decreases because the pre-log factor (τc − fK)/τc decreases.The text states that multicell signal processing is required to benefit from f > 1.
  • Asymptotic limits: With uncorrelated Rayleigh fading and M →∞, noise and non-coherent interference vanish asymptotically, leaving coherent interference from pilot contamination as the limiting factor.The same limit is reported for ZF, sophisticated variations, and analogous downlink analyses.

A. Correlated Rayleigh fading

Correlated Rayleigh fading models practical spatially correlated channels through a covariance matrix that captures macroscopic propagation characteristics. For large ULAs, its angular structure can be approximated using Fourier methods, with limits when large-scale fading varies across the array.

  • Channel model: The correlated Rayleigh model represents a no-line-of-sight channel using a spatial correlation matrix and Gaussian small-scale fading.The normalized trace gives the average channel gain, while the correlation matrix describes macroscopic propagation characteristics.
  • Channel statistics: The correlation matrix is assumed fixed over small-scale fading coherence intervals, known at the BS, and slowly varying in practice.Its estimation is discussed separately in the paper.
  • Spatial structure: The eigenstructure of the correlation matrix determines which spatial directions are statistically likely to contain signal components.Large eigenvalue variations characterize high spatial correlation, but do not generally establish that one channel is more correlated than another.
  • Spatial correlation: Spatially uncorrelated fading is a special case with a scaled identity correlation matrix, whereas practical channels differ in diagonal and off-diagonal structure.The paper attributes practical spatial correlation to propagation directions, antenna patterns and polarization, and array geometry.
  • ULA representation: For a ULA, the channel is modeled as a superposition of physical paths arriving from different angles, with each path weighted by gain and phase.The ULA array response and path-angle distribution determine the resulting correlation matrix.
  • Fourier approximation: As M →∞, the ULA correlation matrix becomes asymptotically equivalent to a circulant matrix, enabling DFT-based eigenvector and channel-generation approximations.The inverse DFT provides an angular-domain representation, and the angular resolution is governed by 1/L rather than antenna count alone.
  • Approximation boundary: The DFT approximation requires macroscopic large-scale fading to remain constant across the antenna array and therefore cannot capture near-field scattering with array-wide variations.Such variations have been observed in measurement campaigns.

C. A look at channel measurements

Measurements show that practical Massive MIMO channels have non-uniform antenna gains, structured eigenvalue distributions, and UE-specific spatial correlation matrices. In both LoS and NLoS conditions, a small set of spatial modes captures nearly all channel gain.

  • Measured channel properties: The eigenvalue distribution demonstrates that multipath components are unevenly distributed across spatial directions.The figure compares measured NLoS behavior with a hypothetical case in which off-diagonal correlation elements are zero.
  • Measurement setup: The correlation matrices are estimated from 1000 samples and normalized so their trace equals M = 128.The measurements use a 128-antenna uniform cylindrical array in LoS and NLoS scenarios.
  • Measured channel properties: Measured correlation matrices exhibit non-identical diagonal elements, non-zero off-diagonal elements, and UE-specific spatial structures.The antenna gains vary substantially around their average in a random-like pattern unique to each UE.
  • Measured channel properties: The measured gain variations have a standard deviation between 2 and 3 dB across the different cases.Periodic behavior is attributed to the cylindrical array geometry, while the random-like pattern is unique to each UE.
  • Measured channel properties: The eight largest eigenvalues contribute more than 99% of channel gains in both LoS and NLoS measurements.Thus, measured channel vectors are approximately spanned by eight eigenvectors, which differ between cases.

E. Linear independence and orthogonality of matrices

The paper distinguishes linear independence from spatial orthogonality as measures of difference between users’ correlation matrices. Practical and modeled results indicate that asymptotic linear independence is common, whereas asymptotic spatial orthogonality is generally unlikely.

  • Definitions: Linear independence means a correlation matrix cannot be written as a linear combination of the others, while asymptotic linear independence additionally requires growing difference with M.The latter is therefore more restrictive than ordinary linear independence.
  • Definitions: Spatial orthogonality requires non-overlapping eigenspaces and therefore can only occur when the correlation matrices are rank-deficient.The condition implies R1R2 = 0_M×M.
  • Finite-dimensional examples: In the finite-dimensional example, linear independence is very likely, whereas exact spatial orthogonality is very unlikely.Continuous random propagation parameters make parallel vectors and exact zero products probability-zero events.
  • Asymptotic examples: In the asymptotic example, non-zero variance in the eigenvalues yields asymptotic linear independence but not asymptotic spatial orthogonality.The conclusion follows for correlation matrices represented in the DFT angular basis with positive random eigenvalues.
  • Measurements and design implications: Measurements show that the asymptotic orthogonality metric remains far from zero and does not decrease from 32 to 64 antennas.A likely explanation is that multipath components spread randomly over the angular domain rather than occupying confined angular intervals.
  • Measurements and design implications: Signal processing can rely on asymptotic linear independence, but spatial orthogonality should be treated as a special case rather than a general property.The paper links this design conclusion to the observed irregularity of practical propagation channels.

IV. SIGNAL PROCESSING FOR MASSIVE MIMO 2.0

Massive MIMO 2.0 signal processing incorporates spatial correlation and multicell interference into channel estimation and receive combining. The resulting M-MMSE scheme uses local estimates from all cells to jointly suppress interference and combine the desired signal.

  • Framework: The signal-processing framework revises MMSE channel estimation, derives optimal receive combining, and develops simplified single-cell schemes for correlated multicell channels.Downlink precoding is treated as an analogous extension of the uplink results.
  • Channel estimation: Pilot-sharing UEs with different spatial correlation properties cause less estimation interference, and spatially orthogonal matrices eliminate that interference theoretically.A practical rule is to assign pilots so tr(R_j R_l) is small for pilot-sharing UEs.
  • Channel estimation: The NMSE decreases with M under spatially correlated fading, unlike the uncorrelated-fading NMSE, which is independent of M.Correlation lets observations across antennas provide information about individual channel elements.
  • Interference and spectral efficiency: Spatially similar correlation matrices produce stronger coherent interference, whereas spatially different matrices reduce it and spatial orthogonality makes it zero.This dependence explains why spatial correlation can improve spectral efficiency in Massive MIMO.
  • Receive combining: M-MMSE jointly maximizes the instantaneous effective SINRs of the UEs in a cell and also minimizes the conditional mean-square error.It uses both intra-cell and inter-cell channel estimates available locally at the serving BS.
  • Receive combining: M-MMSE suppresses strong interference from any cell without requiring cooperation between cells.It can be interpreted as whitening the received signal followed by maximum-ratio combining.
  • Single-cell versus multicell processing: S-MMSE has lower computational complexity than M-MMSE but provides weaker interference suppression when a cell-edge UE switches to another BS.M-MMSE continues using the UE’s channel estimate at the old BS, while S-MMSE stops suppressing that interference.

C. Downlink spectral efficiency and transmit precoding design

Downlink spectral efficiency depends on precoding vectors across the entire network, making optimal design challenging. The paper therefore considers achievable bounds and uplink–downlink duality to relate receive-combining schemes to precoding schemes.

  • Downlink model: The downlink signal contains desired, intra-cell, and inter-cell interference components generated by precoded transmissions from all cells.Each precoding vector determines transmission spatial directivity and has unit norm.
  • Spectral-efficiency bounds: Downlink capacity characterization is harder than uplink characterization because the UE must estimate its effective precoded channel for decoding.The hardening bound is one widely used achievable spectral-efficiency bound.
  • Spectral-efficiency bounds: The hardening bound can be loose for channels with little or no channel hardening, although refined bounds can estimate the effective channel more accurately.The tutorial uses the hardening bound because it suffices to demonstrate the impact of spatial correlation.
  • Precoding design: Downlink spectral efficiency depends on precoding vectors for all UEs in the network, which makes optimal precoding design challenging in practice.Uplink–downlink duality provides a heuristic way to obtain precoding from uplink combining schemes.
  • Precoding design: Uplink–downlink duality maps M-MMSE combining to M-MMSE precoding and similarly maps the other uplink combining schemes to corresponding precoders.The precoding vectors can be computed using MK complex multiplications for normalization across the UEs.

V. PERFORMANCE BENEFITS OF OPTIMIZED SIGNAL PROCESSING IN MASSIVE MIMO 2.0

In correlated Rayleigh fading, multicell-aware combining and precoding substantially outperform schemes focused on intra-cell interference, especially as the antenna count grows. M-MMSE achieves the strongest spectral-efficiency gains by balancing coherent desired-signal combining with interference suppression.

  • UL and DL spectral efficiency: M-MMSE provides the highest DL sum SE, followed by S-MMSE and ZF, while MR provides only 45-60% of M-MMSE’s SE.The reported DL trends mirror the uplink comparison.
  • Pilot reuse and interference: With M = 100, M-MMSE gains 20% in UL SE when f > 1 by improving channel estimation and suppressing interference from surrounding cells.Other schemes lose SE as f increases because they do not suppress inter-cell interference.
  • Pilot reuse and interference: M-MMSE substantially reduces coherent interference, whereas the other compared scheme can only reduce non-coherent interference.The comparison is made for the weakest UE with M = 200 and f = 2.
  • Spatial correlation across UEs: With M = 100 and f = 2, spatial correlation shifts the UL-SE CDF rightward for all combining schemes except MR, with the largest benefit for M-MMSE.Spatial correlation can still reduce the SE of a UE at a particular location, although the CDF probability is consistently higher under correlated fading as UEs move through the network.

A. Pilot contamination is not a fundamental asymptotic limitation

Pilot contamination is not a fundamental asymptotic limit under spatially correlated channels when multicell-aware M-MMSE processing is used. The decisive condition is asymptotic linear independence among the desired UE’s correlation matrix and those of pilot-sharing UEs.

  • MR versus multicell-aware processing: For MR, coherent interference depends on the correlation matrices of pilot-sharing UEs and vanishes only under asymptotic spatial orthogonality.Absent that condition, MR typically converges to a finite SE limit as M tends to infinity; UL and DL can experience different coherent-interference patterns.
  • Asymptotic result: M-MMSE UL and DL SE grows without bound as log2(M) when the desired UE’s correlation matrix is asymptotically linearly independent of pilot-contaminating UEs’ matrices.This is the stated correlated-Rayleigh condition in Theorem 5.
  • Implication: The asymptotic conclusion overturns analyses based on uncorrelated fading, where pilot contamination appears to impose a fundamental capacity limit.Schemes that only suppress intra-cell interference are not asymptotically optimal under practical spatially correlated channels.
  • Geometric mechanism: Linearly independent correlation matrices permit a combining vector that rejects pilot-sharing interference while retaining a non-zero desired-signal component.M-MMSE preserves an array gain proportional to M while suppressing coherent interference.
  • Geometric mechanism: When correlation matrices are linearly dependent, suppressing pilot-sharing interference necessarily removes an equal fraction of the desired signal, producing the uncorrelated-fading asymptotic limit.The geometric illustration shows that no suitable combining vector exists in this case.

B. Numerical validation of asymptotic analysis with spatially correlated channels

Numerical studies validate the asymptotic advantage of multicell-aware processing and show that partial statistical knowledge can preserve the same large-antenna scaling. At finite M, however, M-MMSE is more sensitive to estimator quality than simpler schemes.

  • Numerical asymptotics: In the symmetric L = 4, K = 2 setup, MR and S-MMSE converge to asymptotic limits of around 3 bit/s/Hz per UE, unlike M-MMSE.The setup places BSs at the coverage-area corners and pilot-sharing UEs pairwise close together.
  • Numerical asymptotics: Time splitting also yields unbounded SE growth but with a smaller slope than M-MMSE, and MR outperforms time splitting over the considered antenna range.The time-splitting comparison activates the L = 4 cells in different coherence blocks.
  • Finite-M interpretation: Pilot contamination remains associated with substantial estimation-error losses and sacrificed signal power, even though M-MMSE avoids convergence to a fundamental capacity limit.The absence of an asymptotic limit does not mean that pilot contamination disappears.
  • Implications for Massive MIMO 2.0: The paper frames spatial correlation exploitation as especially consequential for future systems with hundreds or thousands of antennas.The benefit of M-MMSE is expected to grow as antennas become more widely deployed.
  • Partial statistical knowledge: EW-MMSE uses only correlation-matrix diagonals, which are easier to acquire, and is equivalent to MMSE when channel-vector elements are independent.This estimator ignores inter-element correlation while retaining element-wise variance information.
  • Partial statistical knowledge: With EW-MMSE, M-MMSE loses 23% in SE, compared with 8% for MR, while S-MMSE, ZF, and MR perform almost equally with EW-MMSE and LS estimators.MMSE estimation gives the highest SE for every combining scheme in the compared setup.
  • Partial statistical knowledge: M-MMSE-EW retains log2(M) UL-SE growth when diagonal correlation matrices of pilot-sharing UEs are asymptotically linearly independent.Measured diagonal entries are non-uniform and UE-specific, so the condition is considered likely in practice.

B. Sample correlation matrices and regularization

Correlation statistics can be estimated from repeated pilot observations, but full matrices are substantially harder to estimate than their diagonals as the antenna dimension grows. Regularization and DFT structure can improve finite-sample estimation, with an eventual error floor when the imposed structure is inaccurate.

  • Sample correlation estimation: Sample-variance standard deviation decreases as 1/√N and is independent of M, so relatively few observations can provide accurate diagonal variance estimates.The pilot signals used for channel estimation also provide the observations, requiring no extra signaling.
  • Sample correlation estimation: Full sample correlation matrices are harder to estimate because errors across M^2 elements affect their eigenvalues and eigenvectors.This motivates regularizing the sample matrix when the number of observations is limited.
  • Regularization: Regularization forms a full-rank matrix for η ∈ [0, 1) while shrinking off-diagonal elements by η relative to the sample correlation matrix.The parameter can be tuned to underestimate unreliable off-diagonal elements.
  • Numerical evaluation: A small number of samples estimates correlation diagonals accurately without scaling with M, whereas estimating the full matrix remains more challenging for large antenna arrays.The paper’s NMSE study compares M ∈ {32, 64, 100}.
  • Structured estimation: DFT structure can improve estimation accuracy below 200 samples, but its mismatch creates an error floor as N increases.The switching point toward unconstrained estimation occurs at higher N when M is larger.
  • Practical acquisition boundary: Individual UE correlation matrices cannot be estimated directly from existing pilots because pilot contamination prevents interference-free channel observations.The paper notes that methods for obtaining such observations are still at an early stage.

C. Methods to estimate individual correlation matrices

The article presents several ways to estimate individual spatial correlation matrices, trading pilot overhead, contamination, and estimation quality. Direct orthogonal pilots provide interference-free observations, while alternative schemes separate or disentangle pilot-sharing users across phases or coherence blocks.

  • R direct: The “R direct” approach uses unique orthogonal pilots repeated across coherence blocks to obtain interference-free channel observations for sample correlation matrices.The extra pilots can be distributed over the blocks during which channel statistics remain fixed.
  • Via Q: “Via Q” estimates individual correlation matrices from observations containing pilot-sharing users, using a two-stage procedure and subtraction of their correlation contributions.It provides more observations than “R direct” when fewer extra pilots than pilot sequences are available, but imperfect subtraction perturbs the estimate.
  • Trade-offs: The “Via Q” and “R direct” methods differ in observation quantity and contamination: “Via Q” offers more observations when NR < N, but requires subtraction of interfering correlation matrices.Regularization can improve robustness when the resulting estimate is inaccurate.
  • Phase randomization: “Via Q” can use a specific phase for learning a correlation matrix, so pilot-sharing users’ contributions average toward zero as the number of observations grows.Random phase shifts make the interfering terms zero mean across coherence blocks.
  • Pilot reassignment: Changing pilot assignments across coherence blocks yields a linear system whose unique solution recovers individual correlation matrices when T ≥ KL/τp and the joint allocation matrix has full row rank.The same approach can estimate only the diagonal elements when desired.

D. Spectral efficiency evaluation with imperfect statistical knowledge

The section evaluates spectral efficiency under imperfect statistical knowledge and examines lower-complexity alternatives to M-MMSE. Extra pilots can bring performance close to ideal correlation knowledge, while OBE preserves favorable asymptotic scaling at substantially lower complexity but may lose finite-array spectral efficiency.

  • Imperfect statistical knowledge: A few tens of extra pilots outperform LS, while “Via Q” needs a thousand extra pilots to reach 92% of the MMSE estimator’s SE.The comparison uses M-MMSE with full or partial correlation estimates and benchmarks against ideal MMSE and no-correlation LS estimation.
  • Imperfect statistical knowledge: A few hundreds of extra pilots let M-MMSE-EW reach 98% of the MMSE estimator’s SE without overhead that grows with M.This reduced growth comes from estimating only diagonal correlation elements, at the price of lower SE than in the full-matrix case.
  • Practical limitation: Correlation estimation may be difficult when active users change rapidly because of user behavior and bursty packet transmissions.The section therefore identifies implementation under changing user activity as a practical challenge.
  • Lower-complexity processing: OBE requires one statistics-based matrix-vector multiplication using own-cell channel estimates, reducing per-block complexity to M^2 multiplications versus M^3 growth for M-MMSE.Its statistics-dependent matrix can be precomputed and stored when channel statistics change slowly.
  • Finite-array performance: For the Fig. 2 setup, OBE has 50% lower UL sum SE than M-MMSE but 40% higher sum SE than MR at M ≥64.The comparison reflects lower complexity obtained by precomputing statistical parameters.
  • Asymptotic behavior: With asymptotically linearly independent correlation matrices, OBE SE increases logarithmically as M →∞ and can achieve the same asymptotic scaling as M-MMSE.Under strong pilot contamination, OBE’s finite-array loss is around 2–7%; D-OBE is 5–7% below D-MMSE under analogous diagonal independence.

VIII. CONCLUDING REMARKS AND RESEARCH DIRECTIONS FOR BEYOND 5G

The article concludes that Massive MIMO 2.0 can sustain growing cellular spectral efficiency by exploiting spatial correlation and optimized interference suppression. It identifies computational complexity and learning correlation statistics as practical limits, and discusses large surfaces, intelligent reflecting surfaces, cell-free networks, and sub-THz operation as research directions.

  • Concluding remarks: Massive MIMO 2.0 uses spatial correlation and optimized signal processing to suppress intra-cell and inter-cell interference, removing pilot contamination as a strict asymptotic capacity limit.The article describes capacity as theoretically unlimited but practically constrained by computation and learning spatial correlation matrices.
  • Concluding remarks: Future work should address computational complexity and acquisition of spatial correlation, including data-driven learning and suitable approximations.The proposed methods are model-based but can be complemented by data-driven solutions.
  • Large intelligent surfaces: Very large intelligent surfaces place most users in the radiative near-field, requiring revised channel models in which spatial correlation carries angular and depth information.These additional spatial dimensions can support user separation through precoding and combining.
  • Intelligent reflecting surfaces: Intelligent reflecting surfaces can reconfigure reflections into narrow beams, but their beneficial use cases relative to active intelligent surfaces remain unclear.The concept also appears under names including intelligent walls, reconfigurable reflectarrays, and metasurfaces.
  • Cell-free networks: Cell-free networks can be modeled as single-cell Massive MIMO with strong spatial correlation, pilot reuse, unequal antenna powers, and subset-based service.The article asks whether Massive MIMO 2.0 asymptotic analysis extends to this setting.
  • Sub-THz Massive MIMO: At sub-THz frequencies, hardware impairments, mutual coupling, phase noise, synchronization errors, and incomplete channel characterization may constrain practical deployments.The article also anticipates a balance between analog and digital processing and continued work on low-complexity interference cancellation.
Loading 1904.03406v3…