Source-linked AI summary
Massive MIMO: How many antennas do we need?
Jakob Hoydis, Stephan ten Brink, Merouane Debbah
TL;DR
The paper asks how finite, large antenna arrays affect asymptotic massive-MIMO conclusions and how much sophisticated detection helps. It uses random-matrix deterministic equivalents for achievable rates, finding that MMSE can require fewer channel dimensions per user than MF for a target performance.
Problem
The paper examines how well infinite-antenna conclusions about MF optimality, vanishing transmit power, and pilot-contamination limits hold for large but finite N.
Method
The authors derive asymptotically tight deterministic equivalents for achievable-rate SINRs with MF and MMSE detection under multicell pilot contamination.
Results
At ρN = 20 dB and α = 0.3, MF requires about 90 DoF per UT for 90% of R∞, whereas MMSE requires about 60 DoF per UT.
Takeaways & Limitations
Finite-N performance depends mainly on channel DoF per UT and effective SNR, while MF and MMSE coincide with infinitely many BS antennas.
Abstract
from arXiv · showhide
We consider a multicell MIMO uplink channel where each base station (BS) is equipped with a large number of antennas N. The BSs are assumed to estimate their channels based on pilot sequences sent by the user terminals (UTs). Recent work has shown that, as N grows infinitely large, (i) the simplest form of user detection, i.e., the matched filter (MF), becomes optimal, (ii) the transmit power per UT can be made arbitrarily small, (iii) the system performance is limited by pilot contamination. The aim of this paper is to assess to which extent the above conclusions hold true for large, but finite N. In particular, we derive how many antennas per UT are needed to achieve η% of the ultimate performance. We then study how much can be gained through more sophisticated minimum-mean-square-error (MMSE) detection and how many more antennas are needed with the MF to achieve the same performance. Our analysis relies on novel results from random matrix theory which allow us to derive tight approximations of achievable rates with a class of linear receivers.
I. INTRODUCTION
Massive MIMO is proposed as an alternative densification strategy, but its benefits depend on acquiring accurate CSI. The paper defines massive MIMO operationally and studies finite-antenna performance, detection choices, and general channel models.
- Motivation: Very large BS antenna arrays offer a less-explored alternative to cell-size shrinking for increasing wireless network density.Cell shrinking requires additional equipment and can increase interference, whereas large arrays are presented as another densification option.
- Motivation: Accurate CSI is necessary to exploit additional antennas, making massive MIMO feasible primarily in TDD systems that can use channel reciprocity.The CSI challenge becomes more pronounced as the number of antennas grows, particularly in fast-fading channels.
- Asymptotic background: With infinitely many antennas, matched filtering and eigenbeamforming become optimal, transmit power can approach zero, and pilot contamination ultimately limits performance.These asymptotic conclusions motivate studying how strongly they hold for large but finite arrays.
- Paper scope: The paper defines massive MIMO as a cellular operating condition where multiuser interference and noise are small relative to pilot contamination.Whether this condition holds depends on BS antennas N, channel DoF, SNR, and path loss.
- Contributions: The analysis derives tight achievable-rate approximations for a general channel model with individual path loss and receive correlation per UT-BS link.The framework also supports future studies of antenna correlation, spacing, aperture, and downlink performance.
II. SYSTEM MODEL
The paper models a noncooperative multicell uplink in which each BS receives transmissions from single-antenna users through correlated fading channels. The channel model permits link-specific path loss and potentially rank-deficient antenna correlation.
- System configuration: The system contains L > 1 cells, one BS per cell, K single-antenna UTs per cell, and N antennas at each BS.The analysis focuses on the uplink without BS cooperation.
- Signal model: The received signal at each BS combines channel-matrix transmissions from all cells with message vectors, noise, and transmit SNR ρ.The channel matrix from cell l to BS j is H_jl, while the noise is modeled as circularly symmetric complex Gaussian.
- Channel model: Each UT-BS channel vector uses a deterministic correlation matrix multiplied by a fast-fading Gaussian vector.This formulation allows different correlation, including path loss, for every channel vector.
- Channel model: The correlation matrices need not have full rank, capturing large-array correlation caused by insufficient antenna spacing or limited scattering.Rank deficiency is important when physical channels occupy fewer effective dimensions than the antenna count.
- Physical interpretation: A physical channel with P ≤ N dimensions or angular bins can be represented as a special case using a unitary basis and inverse path loss.The paper uses a particular form of this model later in its analysis.
A. Channel estimation
During training, users transmit orthogonal pilots reused across cells, enabling local channel estimation but introducing pilot contamination from neighboring cells. MMSE estimation separates each channel into an estimate and independent error.
- Pilot training: Each cell’s UTs transmit orthogonal pilot sequences, but reuse across cells contaminates local channel estimates with neighboring-cell pilots.The BS forms estimates of its local channel matrix from these training observations.
- MMSE estimation: Under MMSE estimation, the channel equals the estimated channel plus an independent estimation error.The estimate is modeled as Gaussian with covariance Φ_jjk, while the error covariance is R_jjk − Φ_jjk.
B. Achievable rates with linear detection
The paper evaluates linear single-user detectors by filtering the received vector for each in-cell UT. It focuses on matched-filter and MMSE detection and expresses achievable rates through an SINR bound.
- Linear detection: Linear single-user detection estimates each UT’s symbol by taking an inner product between the received vector and a linear filter.The filter for UT m at BS j is denoted r_jm.
- Detector choices: The two detectors of primary practical interest are the matched filter and the MMSE detector.Both are defined as specific choices of the linear filter used for symbol detection.
- Modeling choice: The formulation permits theoretical estimation of all channel matrices, but high neighboring-cell path loss is expected to make those estimates unreliable and gains marginal.The paper therefore focuses on the stated local-estimation formulation.
- MMSE design: The MMSE filter formulation allows λ and Z_j to be treated as design parameters, with λ = 1/(ρN) identified as a natural choice.This parameterization exposes a tunable regularization choice in the detector.
- Achievable rate: A standard worst-case uncorrelated-noise bound yields the ergodic achievable rate R_jm for UT m.The associated SINR γ_jm determines the rate expression.
III. ASYMPTOTIC ANALYSIS
The paper derives deterministic equivalents for matched-filter and MMSE SINRs in a large-system multicell MIMO model, explicitly accounting for channel training and pilot contamination. These results characterize when finite systems approach the infinite-antenna performance limit and show that matched-filter and MMSE performance coincide asymptotically.
- The analysis derives deterministic SINR approximations as N and K grow infinitely large while maintaining a finite ratio K/N.The approximations target the matched-filter and MMSE detectors under general channel conditions.
- The large-system analysis assumes channel coherence time scales linearly with K to support sufficiently many orthogonal pilot sequences.The authors state that this assumption does not pose a problem when applying the results to realistic system dimensions.
- The derivations rely on technical conditions that bound correlation-matrix norms and keep normalized traces bounded away from zero.These conditions apply to the channel correlation matrices R_jlk.
- Theorem 1 provides a deterministic equivalent for the matched-filter output SINR, while Theorem 2 provides one for MMSE detection.The MMSE result uses auxiliary deterministic quantities specified through additional theorems.
- The matched-filter and MMSE performances coincide with infinitely many antennas, and the limiting quantity corresponds to an asymptotic signal-to-interference ratio.The limit is associated with the pilot-contamination-limited regime.
IV. ON THE MASSIVE MIMO EFFECT
The massive MIMO effect depends on effective SNR and the channel's degrees of freedom per UT, while pilot contamination remains the limiting factor as noise and multiuser interference vanish. Finite-antenna results quantify the DoF requirements and show that MMSE can reduce them substantially compared with MF.
- Doubling N increases effective SNR ρN linearly, allowing transmit power to be halved while maintaining the same SNR.
- Multiuser interference depends mainly on P/K, whereas noise and multiuser interference vanish as N and P grow, leaving pilot contamination as the limiting factor.
- The finite-N rate approximations are almost indistinguishable from simulations for K = 10, L = 4, ρ = 0 dB, α = 0.1, and P equal to N or N/3.The P = N/3 case performs worse because of stronger multiuser interference.
- At ρN = 20 dB and α = 0.3, MF requires about P/K = 90 DoF per UT for 90% of R∞, while MMSE requires about 60.For R∞ ≈ 2.2 b/s/Hz, this corresponds to approximately 2 b/s/Hz.
- Adding antennas has diminishing returns because the required P/K distances for higher η grow exponentially, and η = 1 requires P/K = ∞.The absolute MF–MMSE difference is marginal for small η but becomes pronounced as η approaches 1.
- At ρN = 20 dB and α = 0.1, MF achieves 80% of the ultimate performance with P/K = 90, while MMSE needs only 35 DoF per UT for the same performance.The resulting spectral efficiency is approximately 4.6 b/s/Hz, and MMSE could support about 2.5 times more UTs.
V. CONCLUSIONS
The paper derives deterministic equivalents for MF and MMSE achievable rates in a general multicell channel model and uses them to characterize the large-but-finite antenna regime. The results identify DoF per UT and effective SNR as the main dependencies and quantify antenna requirements and MMSE gains.
- The analysis derives deterministic equivalents of achievable rates with matched filtering and MMSE detection under a general channel model with individual UT correlation matrices.
- Both detectors' performance depends mainly on the channel's DoF per UT and the effective SNR in the large-but-finite N regime.
- The results determine how many antennas achieve η% of the ultimate performance and how many additional antennas MF needs to match MMSE performance.
- The results can also support downlink analysis and future studies of antenna correlation, spacing, and aperture in more realistic channel models.
APPENDIX
The appendix states random-matrix results used to obtain deterministic equivalents for the receivers' performance. These results rely on bounded covariance spectra, large-system scaling, and a uniquely solvable set of implicit equations.
- Theorem 3 considers Hermitian nonnegative definite matrices D and S and a random matrix H with independent standardized columns shaped by deterministic covariance matrices R_k.
- The theorem assumes uniformly bounded spectral norms for D, S, and R_k as N grows, with N and K tending to infinity at a controlled ratio.
- The deterministic quantities δ(ρ) are defined through K implicit equations possessing a unique nonnegative solution.
- Theorem 4 extends the framework using a Hermitian nonnegative definite matrix Θ under the same conditions as Theorem 3.
- Theorem 4 uses T(ρ), δ_k(ρ), and their derivatives together with a matrix J(ρ) and vector v(ρ).