PaperlayerView exact revision on arXiv

Source-linked AI summary

Scaling up MIMO: Opportunities and Challenges with Very Large Arrays

Fredrik Rusek, Daniel Persson, Buon Kiong Lau, Erik G. Larsson, Thomas L. Marzetta, Ove Edfors, Fredrik Tufvesson

arXiv:1201.3210v1cs.IT

TL;DR

Very large MIMO promises practical gains from arrays with hundreds of antennas, but channel acquisition, propagation, interference, and detection remain important challenges. This survey synthesizes advances across theory, antennas, propagation, CSI, precoding, and detection, finding that large-array systems can support high throughput and robust operation while leaving realistic-channel issues open.

  • Problem

    Practical operation of very large MIMO remains insufficiently understood beyond ideal asymptotic capacity laws, particularly regarding CSI acquisition, interference, propagation, and detection.

  • Method

    The paper surveys theoretical, propagation, antenna, CSI, precoding, and detection aspects, including measured channels and receiver comparisons for large arrays.

  • Results

    More than 40 single-antenna users can achieve 17 Mbits per second average throughput and 3.6 Mbits per second with 95% probability in a 20 MHz TDD system.

  • Takeaways & Limitations

    Very large MIMO offers high-throughput, stable, and robust operation, while MMSE-SIC and TS detectors can approach optimal maximum-likelihood detection in challenging large-user scenarios.

  • Takeaways & Limitations

    Realistic channels remain incompletely understood, and measured performance loses relative to ideal IID-channel models.

Abstract

from arXiv · show

This paper surveys recent advances in the area of very large MIMO systems. With very large MIMO, we think of systems that use antenna arrays with an order of magnitude more elements than in systems being built today, say a hundred antennas or more. Very large MIMO entails an unprecedented number of antennas simultaneously serving a much smaller number of terminals. The disparity in number emerges as a desirable operating condition and a practical one as well. The number of terminals that can be simultaneously served is limited, not by the number of antennas, but rather by our inability to acquire channel-state information for an unlimited number of terminals. Larger numbers of terminals can always be accommodated by combining very large MIMO technology with conventional time- and frequency-division multiplexing via OFDM. Very large MIMO arrays is a new research field both in communication theory, propagation, and electronics and represents a paradigm shift in the way of thinking both with regards to theory, systems and implementation. The ultimate vision of very large MIMO systems is that the antenna array would consist of small active antenna units, plugged into an (optical) fieldbus.

I. INTRODUCTION … 1) Channel model:

Very large MIMO scales antenna arrays far beyond current systems, creating deterministic channel behavior, improved conditioning and noise averaging while introducing power, complexity, propagation, and hardware challenges. The paper develops information-theoretic foundations, propagation analysis, and practical low-complexity transmission and reception schemes, beginning with the point-to-point channel model.

  • I. INTRODUCTION: Very large MIMO uses arrays with roughly an order of magnitude more elements than current systems, while serving fewer terminals because channel-state information limits simultaneous users.Additional terminals can be accommodated through conventional time- and frequency-division multiplexing via OFDM.
  • I. INTRODUCTION: ∝1/nt per-antenna power maintains constant total transmit power as antenna count grows, and single-terminal operation can support total power inversely proportional to nt.Practical savings are limited by multi-user multiplexing, CSI errors, and interference, but an order-of-magnitude transmit-power reduction is presented as important.
  • I. INTRODUCTION: Larger arrays make channel behavior more deterministic, improve matrix conditioning, accelerate some inversions, and increase spatial resolution for distinguishing scattering centers.Communication performance depends increasingly on aggregated propagation properties, including asymptotic orthogonality between distinct terminals’ channel vectors.
  • I. INTRODUCTION: Array scaling can average out thermal noise, leaving interference as the dominant limitation, while physical-model breakdown and engineering difficulties constrain antenna counts.The uplink effect follows from coherent averaging of quantities uncorrelated across receive antennas; the downlink behavior requires additional conditions.
  • A. Outline and key results: The paper surveys information-theoretic limits, channel and propagation effects, antenna arrangements, and practical algorithms for very large MIMO.Its organization proceeds from information theory to antennas and propagation, then to transmit and receive schemes, emphasizing approximate low-complexity processing because optimal complexity scales unfavorably.
  • II. INFORMATION THEORY FOR VERY LARGE MIMO ARRAYS: Large-array Shannon theory simplifies and suggests capacity-approaching strategies, with multi-user MIMO identified as the setting where very large MIMO may provide greatest utility.The discussion begins with point-to-point MIMO to expose limitations of compactly clustered antennas before moving to multi-user systems.
  • A. Point-to-point MIMO: A point-to-point MIMO link connects nt transmit antennas and nr receive antennas through a channel where every receive antenna experiences the combined action of all transmit antennas.The narrowband memoryless model uses transmitted vector s, received vector x, propagation matrix G of size nr × nt, and receiver-noise vector w.
  • 1) Channel model:: Wideband frequency-dependent channels can be decomposed into parallel independent narrowband channels, an operation rigorously performed by OFDM.The narrowband model assumes IID zero-mean unit-variance circularly symmetric complex-Gaussian noise, while ρ measures link SNR.

2) Achievable rate:

The section characterizes achievable rate through mutual information and singular values, identifying capacity conditions and best- versus worst-case singular-value distributions. It also relates the rate to parallel links and describes propagation conditions associated with rank-1 and equal-singular-value cases.

  • Achievable rate: Capacity is achieved with water-filling inputs, while C equals capacity when GGH is a scaled identity matrix.The transmitter need not know the channel, but must know the achievable-rate value; with known channel statistics, it can select a rate meeting an acceptable outage probability.
  • Achievable rate: The achievable rate equals the combined rate of parallel links whose ℓ-th link has SNR ρν2ℓ/nt.The singular values are the diagonal elements of Dν in the propagation-matrix decomposition.
  • Achievable rate: The worst case has all but one singular value zero, whereas the best case has all min(nt, nr) singular values equal, bounding the achievable rate.This ordering follows from the concavity of the logarithm under the trace constraint.
  • Achievable rate: Rank-1 behavior arises with compact-array LOS propagation or extreme keyholes, while IID propagation approaches the equal-singular-value case.Under favorable propagation and high SNR, the achievable rate is proportional to the smaller of the number of t…

3) Limiting cases:

The limiting cases show that asymptotically orthogonal propagation vectors make excess transmit or receive antennas highly desirable, with achievable rates reaching the upper bound. Without such gains, multiple transmit antennas may provide no rate benefit.

  • No multiplexing gain: Even under the most favorable propagation conditions, multiplexing gains are lost when the expression is independent of n_t, making multiple transmit antennas worthless for achievable rate.This limiting behavior is stated from the perspective of achievable rate.
  • Many transmit antennas: With a fixed number of receive antennas, asymptotically orthogonal propagation-matrix rows make the achievable rate n_r · log2(1 + ρ), matching upper bound (9).The limiting case assumes the number of transmit antennas grows large while the number of receive antennas remains constant.
  • Many receive antennas: With a fixed number of transmit antennas, asymptotically orthogonal propagation-matrix columns again make the achievable rate match upper bound (9).This case uses the identity det(I + AA^H) = det(I + A^HA) together with the achievable-rate expressions.
  • Implications: Excess transmit or receive antennas with asymptotically orthogonal propagation vectors constitute a highly desirable scenario.Extra receive antennas boost effective SNR and could compensate for low SNR, restoring multiplexing gains otherwise lost.

B. Multi-user MIMO · 1) Propagation: · 2) Reverse link:

Multi-user MIMO uses a large antenna array to serve many autonomous single-antenna terminals, with TDD enabling practical channel-state acquisition. Under favorable propagation, large arrays separate users asymptotically and provide reverse-link throughput at least as high as terminal collaboration, with each user’s SNR scaling with antenna count.

  • B. Multi-user MIMO: MU-MIMO splits a point-to-point array into autonomous antennas, using M antennas to simultaneously serve K single-antenna terminals.The array may be located at a base station, while each terminal has one antenna.
  • B. Multi-user MIMO: TDD makes the reverse-link propagation matrix the transpose of the forward-link matrix and avoids forward-pilot overhead proportional to antenna count.Reverse-link pilot duration is independent of the number of antennas, whereas forward-link pilot duration grows with that number.
  • 1) Propagation:: Large-scale fading captures path loss and shadow fading, while each propagation-matrix column describes a terminal’s small-scale fading across the M antennas.The antenna array is assumed compact enough that a terminal’s propagation paths share the same large-scale fading, with small-scale coefficients typically normalized to magnitude one.
  • 1) Propagation:: With large arrays, the number of antennas greatly exceeds the number of terminals, and favorable propagation makes propagation-matrix columns asymptotically orthogonal.This orthogonality is the key favorable-propagation condition for separating users.
  • 2) Reverse link:: Reverse-link MU-MIMO achieves total throughput no lower than a system in which the terminals collaborate.Collaboration could simplify channel coding and decoding but would not change the ultimate sum-rate, which need not be equally shared among terminals.
  • 2) Reverse link:: Under favorable propagation and many more antennas than terminals, the reverse-link sum-rate has an asymptotic expression determined by the terminals’ fading coefficients.The supplied passage introduces this asymptotic sum-rate but does not include the formula itself.
  • 2) Reverse link:: Matched-filter processing separates terminal signals when propagation-matrix columns are nearly orthogonal, giving terminal k an SNR of Mρrβk.The resulting individual rate corresponds to terminal k’s term in the reverse-link sum-rate.

3) Forward link:

Under favorable propagation and many antennas, the forward-link sum-capacity simplifies asymptotically because propagation-matrix columns become nearly orthogonal. A matched-filter precoder with suitable power allocation achieves the same sum-capacity expression.

  • 3) Forward link:: Under favorable propagation conditions and a large excess of antennas, the forward-link sum-capacity has a simple asymptotic form.The result relies on the propagation matrix becoming nearly orthogonal as the number of antennas grows.
  • 3) Forward link:: Nearly orthogonal propagation-matrix columns enable a simple MF linear precoder for transmission.The transmitter can exploit the asymptotic orthogonality of the channel columns.
  • 3) Forward link:: Σ_{k=1}^K log2 (1 + Mρ_f p_k β_k) is identical to the sum-capacity when p = γ.The MF scheme yields an achievable sum-rate matching the sum-capacity expression under the stated identification.

III. ANTENNA AND PROPAGATION ASPECTS OF VERY LARGE MIMO · A. Spatial focus with more antennas · B. Antenna aspects

Very large MIMO performance depends on how antenna arrays exploit propagation complexity, but nonideal elements, limited propagation richness, fixed apertures, correlation, and mutual coupling constrain the available gains. More antennas can sharpen spatial focusing and reduce interference, while dense 2D/3D implementations introduce additional resolution and practical limitations.

  • III. ANTENNA AND PROPAGATION ASPECTS OF VERY LARGE MIMO: With ideal, well-separated elements in sufficiently complex propagation, each additional antenna adds a usable degree of freedom; real arrays may not.Nonideal elements, insufficient separation, and limited propagation complexity can prevent large arrays from exploiting many degrees of freedom.
  • A. Spatial focus with more antennas: A 100-element ULA focuses the field far better than a 10-element ULA in the same scattering environment, reducing spatial interference.The example uses 400 uniformly distributed scatterers over an 800λ × 800λ square and MF precoding with d = λ/2.
  • B. Antenna aspects: Non-isotropic antenna patterns alter spatial correlation, while directive antennas pointing in distinct directions tend to experience lower correlation than nondirective antennas.The lower correlation arises because directive antennas see signals arriving from distinct angular sectors.
  • B. Antenna aspects: Fixed array apertures force smaller antenna separations as element counts increase, making correlation and mutual coupling central constraints in very large arrays.The aperture may be constrained by the supporting structure or aesthetic considerations, while the no-coupling assumption is valid only for well-separated antennas.
  • B. Antenna aspects: Optimal multiport impedance-matching RF circuits can perfectly mitigate coupling among co-polarized antennas, supporting capacity improvement in compact arrays.Coupling compensation is motivated by implementing MIMO arrays in compact volumes such as mobile terminals.
  • B. Antenna aspects: Dense 2D/3D arrays avoid linear-array front-back ambiguity and resolve azimuth and elevation, but increase implementation complexity and may offer little vertical benefit outdoors.Outdoor cellular signals often arrive within a narrow elevation-angle range, limiting the usefulness of elevation resolution for vertical signaling.
  • B. Antenna aspects: In fixed-aperture MU-MIMO, correlation and coupling prevent asymptotic channel orthogonality, whereas ignoring coupling leaves correlation with only a minor penalty relative to IID channels.The comparison considers K = 15 single-antenna users and evaluates the normalized effective channel and reverse-link maximum rate.

C. Real propagation - measured channels

Measured channels from a 128-antenna indoor base station serving six users show stable, closely grouped eigenvalues, with far less spread than a conventional 6 × 6 MIMO system. The observed 7 dB spread corresponds to that of a 6 × 24 IID conventional MIMO system.

  • Propagation factors: Propagation correlation and terminal/base-station antenna counts determine channel-matrix orthogonality and the ability to separate users or data streams.Conventional MU-MIMO typically has a base-station-to-terminal antenna ratio near 1 and rarely above 2.
  • Measurement setup: A measured 128-antenna base station served 6 single-antenna users at 2.6 GHz over 50 MHz, using 100 channel snapshots.The array comprised four stacked, double-polarized 16-element circular patch arrays; three users were indoors and three outdoors near the base station.
  • Measured-channel eigenvalues: The large array produced eigenvalues with low variance and relatively small spread across the measured 6 × 128 channel.The comparison included measured 6 × 6 MIMO and simulated IID 6 × 6 and 6 × 128 systems.
  • Measured-channel eigenvalues: 7 dB was the measured difference between the smallest and largest eigenvalue, versus around 26 dB for conventional 6 × 6 MIMO.The eigenvalues were not normalized by the number of base-station antennas in the plotted comparison.
  • Measured-channel eigenvalues: The measured 6 × 128 eigenvalue spread corresponded to that of a 6 × 24 conventional MIMO system with IID complex Gaussian channel entries.This comparison indicates that the tall channel-matrix structure relaxes performance requirements for very large MIMO.

IV. TRANSCEIVERS · A. Acquiring CSI at the base station · B. Precoding in the forward link: Collection of results for single cell systems

The section describes reciprocity-based CSI acquisition for large arrays and compares forward-link precoding methods in large-system single-cell settings. As antenna arrays scale, linear precoding approaches the interference-free benchmark, while nonlinear methods matter most when antennas are not much more numerous than terminals.

  • A. Acquiring CSI at the base station: Large arrays must rely on channel reciprocity: TDD pilots let NCoh terminals transmit simultaneously over one OFDM symbol, requiring K/NCoh time slots.The channel response is assumed constant over NCoh consecutive subcarriers, and K terminals are served.
  • B. Precoding in the forward link: Collection of results for single cell systems: The analysis considers IID CN (0, 1) channels with M, K →∞ at fixed α = M/K and derives asymptotic SNR/SINR expressions for standard precoders.Each precoded data component has average power ρf/M.
  • B. Precoding in the forward link: Collection of results for single cell systems: The interference-free benchmark yields an SNR per receiving unit that converges to ρfα as M →∞.This benchmark assumes all channel energy reaches each terminal without inter-user interference.
  • B. Precoding in the forward link: Collection of results for single cell systems: ZF precoding approaches the optimal interference-free SNR for an array with M −K transmit antennas, but gives SNR = 0 when M = K.ZF requires inversion of a K × K matrix, motivating simpler approximations as arrays grow.
  • B. Precoding in the forward link: Collection of results for single cell systems: As M grows, channel Gram matrices approach the identity, so the ZF precoder tends to the matched-filter precoder and matrix inversion may be unnecessary.The matched-filter approximation is enabled by increasing antenna-array size.
  • B. Precoding in the forward link: Collection of results for single cell systems: MF SINR can grow arbitrarily through antenna scaling but has an error floor, approaching α as ρf →∞.With imperfect MMSE-based CSI, SINR can likewise be made as high as desired for any estimate reliability 0 ≤ξ ≤1.
  • B. Precoding in the forward link: Collection of results for single cell systems: Nonlinear precoding is most valuable when M ≈K because ZF is then far from the interference-free benchmark; at M = 2K, the gap is only 3 dB.The cited nonlinear methods include DPC, vector perturbation, and lattice-aided techniques.
  • B. Precoding in the forward link: Collection of results for single cell systems: For K = 15 users, DPC is about 3 dB from the interference-free benchmark at M = 15, but its gain over ZF falls to about 1 dB at M = 40.The comparison reports ergodic sum-rate capacities for MF, ZF, DPC, and interference-free systems at M = 15, 40, and 100.

C. Precoding in the forward link: The ultimate limit of non-cooperative multi cell MIMO with large arrays

As the number of base-station antennas M grows without limit, thermal noise and small-scale Rayleigh fading vanish, but pilot-contamination interference persists in multi-cell systems. MF and ZF precoding therefore approach limits governed by pilot contamination, with ZF additionally depending on pilot SNR and distributed processing offering further improvement.

  • Asymptotic behavior: As M grows without limit, thermal noise and small-scale Rayleigh fading vanish in single-cell and multi-cell MIMO.The large-array limit removes these effects from the received signal.
  • Pilot contamination: Pilot contamination persists because channel estimates include synchronized pilot transmissions from other cells, causing base stations to beamform partially toward unintended terminals.Using different pilots across cells does not substantially improve the situation because pilots remain confined to a finite-dimensional signal space.
  • Pilot contamination: Pilot-contamination interference remains even with more pilot-transmission power when pilots overlap across cells, although staggered pilots can make additional training power beneficial.Some pilot transmissions necessarily collide with data or pilots in other cells in a multi-cell system.
  • Precoding limits: With ZF precoding, the asymptotic limit is independent of forward-link SNR ρf but depends on pilot SNR ρp, converging to MF performance as ρp →0.Regularization connects the ZF and MF extremes: δ = 0 gives ZF, while δ →∞ gives MF.
  • Potential improvements: Power allocation can further improve the limit, while centralized distributed MIMO could apply ZF across base stations to reduce pilot-contamination effects.This requires estimating the slowly changing large-scale fading factors {βkjℓ}.

1) Numerical results:

The numerical results show that MF precoding has higher asymptotic mean capacity, while ZF provides more concentrated SIR and decisively outperforms MF at finite antenna counts. With ρp = ρf = 10 dB, MF suffers a roughly 5 dB SIR loss that vanishes as M grows without limit.

  • Numerical results:: 13.3 bits/channel use versus 9.6 bits/channel use: MF achieves the higher mean capacity despite ZF producing a more concentrated SIR distribution.The comparison is E{log2(1 + SIR)} for MF and ZF precoding as M grows without limit.
  • Numerical results:: 11 dB: MF’s asymptotic mean SIR exceeds ZF’s as M →∞ under pilot-contamination-limited SNRs.The stated infinite-SNR condition means ρp and ρf are large enough that pilot contamination limits performance.
  • Numerical results:: At finite M, ZF decisively outperforms MF in mean SIR when ρp and ρf are effectively infinite.The finite-M curves are compared against asymptotic limits, with performance limited by pilot contamination.
  • Numerical results:: 5 dB: with ρp = ρf = 10 dB, MF’s SIR is worse than with infinite ρp and ρf across the plotted M range.As M →∞, this SIR loss vanishes.

D. Detection in the reverse link: Survey of algorithms for single cell systems

In very large MIMO detection, simple linear detectors are near-optimal when M ≫ K under favorable propagation, while practical systems with M ≈ K motivate advanced iterative and random-step methods. Iterative filtering detects the signaling vector through propagated hard or soft information across iterations.

  • D. Detection in the reverse link: Survey of algorithms for single cell systems: When M ≫ K under favorable propagation, simple linear detectors are close to optimal.The passage contrasts this regime with operating points where M ≈ K.
  • D. Detection in the reverse link: Survey of algorithms for single cell systems: Operating points with M ≈ K remain important in practical systems with many users, motivating iterative filtering and random step detection methods.These methods are compared with linear and tree-search methods.
  • D. Detection in the reverse link: Survey of algorithms for single cell systems: Iterative filtering detects the signaling vector q by repeatedly using newly propagated information from previous estimates.Propagated information may be hard decisions on signal vectors or soft probabilistic measures of transmitted symbols.

1) Iterative linear filtering schemes: · 2) Random step methods:

The paper surveys iterative linear filtering and matrix-inversion-free search methods for very large MIMO detection. These methods trade filtering updates, neighborhood searches, tabu exploration, or restricted tree searches against computational complexity and local-minimum or search-radius limitations.

  • 1) Iterative linear filtering schemes:: MMSE-SIC initializes with a linear MMSE estimate and iteratively constructs interference-canceled signals for each user, while residual interference remains because symbol estimates are imperfect.The cancellation is performed per user and iteration.
  • 1) Iterative linear filtering schemes:: KNIter matrix inversions are required per decoded vector, but the matrix inversion lemma can reduce this to 1 per iteration through rank-one updates.The original inversions arise across every realization, user symbol, and iteration.
  • 1) Iterative linear filtering schemes:: BI-GDFE filters depend on iteration-varying input-decision correlation rather than the received vector, allowing precomputation when channel G remains fixed across many signaling vectors.Filters still vary across users and iterations.
  • 1) Iterative linear filtering schemes:: Matrix-inversion-free search methods, except possibly at initialization, evaluate MSE over NNeigh neighboring vectors and repeatedly choose the lowest-MSE neighbor for NIter iterations.The usual initialization uses the MMSE solution.
  • 2) Random step methods:: Tabu Search improves on LAS by permitting transitions to larger-MSE states, helping it avoid local minima during search.Its exploration is designed to move away from locally optimal regions.
  • 2) Random step methods:: TS maintains a recently traversed-vector list with at most NTabu entries, temporarily forbidding those moves to drive exploration toward new search-space regions.This tabu-list strategy gives the algorithm its name.
  • 2) Random step methods:: Sphere Decoding is an ML decoder restricted to points inside a specified-radius sphere, increasing the radius when no signaling point is found.Tree-based low-complexity methods further reduce search by expanding only apparently promising nodes; stack decoding is one example.

3) Tree-based algorithms: · 4) Numerical comparisons of the algorithms: · 5) Soft-input soft-output detection:

The paper presents FCSD as a constant-complexity, parallelizable tree-based detector best suited to M ≈ K, then compares detectors experimentally and extends hard detection to soft-input soft-output methods. TS and MMSE-SIC offer the strongest reported complexity-performance results, while soft detection relies on incorporating a-priori information or constructing candidate lists for approximate LLRs.

  • 3) Tree-based algorithms:: FCSD fully searches the first r symbols and detects the remaining K − r symbols with ZF-DF, enabling |S|^r parallel hardware chains and constant complexity.Its hybrid enumeration and ZF-DF structure makes the method low-complexity and suboptimal relative to sphere decoding.
  • 3) Tree-based algorithms:: FCSD’s column elimination improves conditioning, but provides little benefit when M ≫ K because the channel matrix is already well conditioned.The method should therefore mainly be used when M ≈ K.
  • 4) Numerical comparisons of the algorithms:: The detector comparisons use QPSK over independent-component CN (0, 1) Rayleigh channels and continue simulations until 500 symbol errors are counted.Transmit power is denoted ρ, and an interference-free genie solution provides a benchmark with equal receive signaling power but no multi-user interference.
  • 4) Numerical comparisons of the algorithms:: For K = 15 with ρ ∼ 1/M, Figure 12 evaluates uncoded BER versus the antenna-to-user ratio α = M/K and identifies when excess antennas make linear detection effective.The comparison uses NIter = 6 for MMSE-SIC and NIter = 4 for BI-GDFE.
  • 4) Numerical comparisons of the algorithms:: At BER 0.002 for M = K = 40 and ρ = 12 dB without preprocessing, TS is 1000 times less complex than FCSD, while TS and MMSE-SIC perform best.Complexities are taken from Table II and summed for each scheme.
  • 4) Numerical comparisons of the algorithms:: For 40 × 40 MIMO, TS and MMSE-SIC remain best across the presented SNR range and operate no more than 0.9 dB from ML.The ML detector cannot outperform the interference-free benchmark despite its search space of size 2^80.
  • 5) Soft-input soft-output detection:: MMSE-SIC incorporates a-priori information by conditioning the MMSE filter, requiring filters for each user, symbol interval, and decoder iteration.Unconditional incorporation instead produces filters varying by user and iteration, similarly to BI-GDFE.
  • 5) Soft-input soft-output detection:: TS and FCSD approximate max-log LLRs from candidate lists containing hard-detection results and searched vectors, adding nearby vectors when bit values are missing or accuracy is insufficient.Candidate lists should include both 0 and 1 for every bit.

V. SUMMARY

Very large MIMO can reduce transmit power by an order of magnitude or more while averaging out small-scale fading, but pilot contamination and antenna coupling remain fundamental design challenges. Measurements and detector evaluations support robust performance and feasible operation, while approximate matrix inversion offers a lower-complexity implementation path.

  • Summary: An order of magnitude or more in transmit power can be saved, while small-scale fading averages out and only large-scale fading remains.The paper identifies this as a potential radical change to wireless communication.
  • Summary: Pilot contamination increasingly limits systems as base-station antennas grow, making pilot reuse in neighboring cells a fundamental design challenge.The paper identifies this issue as warranting future research.
  • Summary: Antenna interaction can significantly reduce channel orthogonality and link capacity, with severity depending on reduced spacing and array geometry.The problem is especially relevant for large arrays with fixed overall aperture.
  • Summary: Uplink detection remains feasible when about 40 users and base-station antennas are involved, despite a search space exponential in the number of users.Receiver tests compared several state-of-the-art detectors and demonstrated that this challenging scenario can be handled.
  • Summary: An indoor measurement campaign with a 128 antenna element base station and 6 single antenna users showed stable and robust performance despite losses from non-IID channels.The measurements also leave open issues concerning realistic-channel behavior.
  • Approximate matrix inversion: When M/K is large, say 5-10 or so, the Neumann-series approximation converges rapidly and requires only a few terms.Approximate inversion is motivated by the complexity of directly inverting the K × K matrix Z.
Loading 1201.3210v1…