Source-linked AI summary
Designing Multi-User MIMO for Energy Efficiency: When is Massive MIMO the Answer?
Emil Björnson, Luca Sanguinetti, Jakob Hoydis, Mérouane Debbah
TL;DR
The paper asks how to choose antennas, active users, and transmit power for maximum energy efficiency in a multi-user MIMO system. It derives closed-form optimization results with a realistic power model and finds that the optimum is a massive-MIMO configuration using increasing transmit power and interference-suppressing precoding.
Problem
The central problem is selecting M, K, and transmit power to maximize energy efficiency while accounting for circuit power that scales with system size.
Method
The paper derives closed-form optimal values and parameter interactions using a realistic power-consumption model with high-order terms, primarily under ZF precoding.
Results
M = 165, K = 85, and ρ = 4.6097 produce the global optimum in the reported ZF numerical illustration.
Takeaways & Limitations
Energy-efficient macro-cell systems use hundreds of antennas and relatively many users, while transmit power should increase with M and interference-suppressing precoding is required.
Abstract
from arXiv · showhide
Assume that a multi-user multiple-input multiple-output (MIMO) communication system must be designed to cover a given area with maximal energy efficiency (bit/Joule). What are the optimal values for the number of antennas, active users, and transmit power? By using a new model that describes how these three parameters affect the total energy efficiency of the system, this work provides closed-form expressions for their optimal values and interactions. In sharp contrast to common belief, the transmit power is found to increase (not decrease) with the number of antennas. This implies that energy efficient systems can operate at high signal-to-noise ratio (SNR) regimes in which the use of interference-suppressing precoding schemes is essential. Numerical results show that the maximal energy efficiency is achieved by a massive MIMO setup wherein hundreds of antennas are deployed to serve relatively many users using interference-suppressing regularized zero-forcing precoding.
I. INTRODUCTION
The paper studies how to choose antennas, active users, and transmit power to maximize energy efficiency in downlink multi-user MIMO. It addresses limitations of simulation-based guidance with a realistic power model that captures scaling with system size and propagation conditions.
- Motivation: Current wireless networks prioritize spectral efficiency, but this can produce poor energy efficiency and large disparities between peak and average rates.Energy efficiency is measured in bits transferred per Joule and depends on architecture, spectral efficiency, transmit power, and circuit consumption.
- Problem: Circuit power cannot be treated as constant in massive MIMO because digital processing and RF/baseband circuitry scale with M and K.A transmit-power-only model could otherwise imply unbounded energy efficiency as M increases.
- Research gap: Prior downlink studies found concavity in M or K mainly through simulations, leaving the joint parameter interactions incompletely characterized.The paper seeks analytical guidance rather than only numerical evidence.
- Contribution: Closed-form expressions are derived for the EE-optimal M, K, and transmit power ρ using a power model with high-order processing terms.The analysis is developed for zero-forcing precoding, with simulations indicating similar results for other common schemes.
- System model: The system consists of an M-antenna base station serving K single-antenna users selected from a distribution of channel variances.The downlink uses TDD pilots and channel reciprocity to provide instantaneous CSI at the base station.
- Objective: The design objective is to guarantee a uniform rate for selected users while maximizing energy efficiency through the choices of M, K, and R.The users are selected in a round-robin fashion from a larger population within the coverage area.
A. General Energy Efficiency Metric
The paper defines energy efficiency as average achievable sum rate divided by average total power, using a model that includes transmit, circuit, and processing consumption. The resulting formulation explicitly captures how power scales with antennas and users.
- Metric: Energy efficiency is measured in bit/Joule as the ratio of average achievable sum information rate to total average power consumption.A transmit-power-only model can misleadingly predict infinite energy efficiency as M tends to infinity.
- Power model: The proposed power model includes RF power-amplifier consumption and circuit power from digital processing and RF/baseband analog filters.Its purpose is to describe realistic dependence on M and K.
- Power model: The total power contains a static term C0,0 and terms of the form Ci,j K^i M^j, with amplifier efficiency η and expectations over channels and user locations.This structure allows high-order processing costs to be represented explicitly.
- Metric: The metric uses the average achievable sum rate and total average power specified by the system model.The power formulation is motivated by the separate contributions of transmission hardware, processing, and fixed infrastructure.
- Hardware terms: Transceiver consumption includes MPtx for base-station antenna components, KPrx for user receivers, and a shared oscillator term Psyn.This separates per-antenna, per-user, and shared hardware costs.
2) Coding and Decoding:
The model accounts for user coding, channel estimation, precoding, data transmission, and fixed infrastructure costs. Under ZF with perfect CSI and M ≥ K, the resulting tractable EE expression supports optimization of M, K, and ρ.
- Coding and Decoding:: Coding and decoding consume K(Pcod+Pdec) Joule/channel use because the base station handles K information streams and each user decodes its own stream.
- Channel estimation: Channel estimation consumes power proportional to MK/(LT), reflecting M received signals per user and estimation once per coherence period.
- Fixed consumption: The architecture also incurs fixed power P0 for control signaling, backhaul, and load-independent baseband processing.
- ZF model: Under ZF, the power coefficients combine static, user-dependent, antenna-dependent, and processing terms determined by hardware and computational efficiency.The listed coefficients include C0,0, C1,0, C3,0, C0,1, C1,1, and C2,1.
- Optimization: The optimization assumes M ≥ K, ZF precoding, and perfect CSI from pilot signaling, and uses the resulting tractable expression to compute optimal M, K, and ρ.ZF is highly suboptimal at low SNR, but the paper argues that low SNR is not the optimal operating regime.
A. Preliminaries
The preliminaries establish the Lambert W function as the analytical tool for solving the EE optimization. The relevant objective is strictly quasi-concave, so its unique stationary solution is the global optimum.
- A. Preliminaries: The Lambert W function is defined by x = W(x)e^W(x) and appears repeatedly in the paper’s closed-form solutions.
- A. Preliminaries: The optimization problem has a strictly quasi-concave objective and a unique solution under its stated parameter conditions.
- A. Preliminaries: The objective increases below zopt and decreases above zopt, establishing zopt as the global maximizer.The natural number e appears in the closed-form solution.
- A. Preliminaries: The paper uses the monotonic behavior of e^W(x)+1 to analyze how the optimal solution changes with model parameters.
- A. Preliminaries: The function W(x) is increasing for x ≥ 0, while e^W(x)+1 is approximately e for small x and grows almost linearly for large x.
B. Optimal System Parameters
The analysis optimizes one of M, K, or ρ for energy efficiency while fixing the other two parameters.
- Each optimization isolates one design parameter while holding the other two fixed.
1) Optimal Number of BS Antennas:
The optimal antenna count is characterized explicitly and lies at a finite feasible value. Its behavior depends on transmit power, circuit-power coefficients, and the propagation environment.
- M_opt lies in the feasible set K ≤ M < ∞ because the objective is quasiconcave and zero at M = K and M →∞.
- M increases sublinearly with normalized transmit power ρ, becoming almost linear when ρ is large.
- M increases with circuit coefficients independent of M and decreases with coefficients multiplied by M in the EE metric.
- M increases almost linearly with Aλ, which is proportional to d_max^κ in circular cells.
- The integer-optimal M is one of the two integers nearest the generally non-integer M_opt.
2) Optimal Transmit Power:
The optimal transmit power is positive and, unlike the commonly reported massive-MIMO scaling, generally increases with the number of antennas when circuit power grows with system size.
- ρ_opt is always positive because the objective is quasiconcave and zero at ρ = 0 and ρ →∞.
- The optimal transmit power and SINR ρ(M − K) increase with the circuit-power coefficients.
- Massive-MIMO power reduction proportional to 1/M is generally not the most energy-efficient strategy in practice; EE instead favors increasing ρ with M.
- For large M, Corollary 1 gives an asymptotic expression for the optimal transmit power.
- The optimal transmit power should increase almost linearly with M when circuit power grows with M.
- When circuit power is independent of antenna count, power decreases proportional to log_e(M), more slowly than the linear reduction reported previously.
3) Optimal Number of UEs:
The optimal number of active users is obtained from a quasiconcave optimization problem whose solution is tied to a quartic polynomial and power-consumption assumptions.
- The EE optimization over K is quasiconcave and solved by a root of a quartic polynomial.
- The quartic formulation uses c_1 = C_1,0 + βC_0,1, c_2 = C_2,0 + βC_1,1, and c_3 = C_3,0 + βC_2,1.
- The EE-maximizing root is generally non-integer, so the optimal K is one of the two closest integers.
- With fixed β = M/K and highest-order power terms omitted, K_opt decreases with coefficients of circuit-power terms scaling with M and/or K.
- K_opt increases with static hardware power C_0,0 and propagation parameter Aλ, which scales with coverage area.
C. Joint and Sequential Optimization of M, K, ρ
The paper combines closed-form coordinate updates for M, K, and ρ with exhaustive search or alternating optimization to approach joint energy-efficiency maximization.
- Joint optimization: The joint optimum can be guaranteed by exhaustively searching reasonable integer combinations of M and K and optimizing ρ for each combination.This approach is feasible for offline cell planning.
- Sequential optimization: Alternating optimization updates K, M, and ρ sequentially using Theorems 3, 1, and 2.The sequence repeats until convergence.
- Simulation setup: The numerical study uses the simulation parameters summarized in Table I.The supplied passage identifies the table but does not provide its parameter values.
- Sequential optimization: The alternating algorithm is guaranteed to converge because EE is nondecreasing at each step and has a finite upper bound.Convergence does not necessarily imply a global optimum.
IV. NUMERICAL ILLUSTRATIONS
Numerical illustrations compare energy efficiency, transmit power, and spectral efficiency across antenna counts, user counts, and precoding schemes. They identify massive-MIMO operation with interference suppression as the strongest configuration, while MRT favors few antennas and one user.
- ZF optimization: M = 165, K = 85, and ρ = 4.6097 give the global ZF energy-efficiency optimum in the modeled cell.The optimum is interpreted as a massive-MIMO setup.
- MRT comparison: MRT achieves its highest energy efficiency with few BS antennas and one active UE because strong inter-user interference dominates.Increasing M reduces interference through channel decorrelation, but added circuit and computational power costs dominate.
- Transmit-power behavior: The EE-maximizing transmit power increases with M for ZF, RZF, and MRT, including ZF with imperfect CSI.The similarity between RZF and ZF indicates operation in a high-SNR regime.
- Precoding comparison: RZF and ZF have similar maximal energy efficiency, while MRT is substantially worse under perfect CSI.The comparison includes corresponding spectral efficiencies across BS antenna counts.
- Precoding comparison: There is a 3-fold optimal-EE difference and a 100-fold spectral-efficiency difference between RZF/ZF and MRT under perfect CSI.Most of the spectral-efficiency gain also appears with imperfect CSI.
V. CONCLUSIONS
The paper derives and validates closed-form energy-efficiency design laws for antennas, users, and transmit power under a realistic power model. Its numerical conclusions favor hundreds of antennas with interference-suppressing precoding, while results remain sensitive to circuit-power assumptions and hardware feasibility.
- Contributions: The study derives closed-form scaling laws for EE-maximizing M, K, and ρ under ZF with perfect CSI, then verifies them for other precoders and imperfect CSI.The analysis uses a power model whose total consumption depends explicitly on M, K, and ρ.
- Main conclusions: Hundreds of antennas serving relatively many UEs maximize macro-cell energy efficiency with current circuit technology.The EE is quasiconcave in M and K, yielding finite global optima.
- Main conclusions: Transmit power should increase with M to compensate for increasing circuit power rather than decrease as commonly believed.Energy-efficient operation is therefore not confined to low SNR.
- Main conclusions: Massive MIMO improves energy efficiency only when paired with interference-suppressing precoding such as ZF or RZF.The conclusion explicitly rules against MRT for this energy-efficiency objective.
- Scope and limitations: Numerical results are stable to small circuit-coefficient changes but can change drastically otherwise.Large antenna arrays and fast, energy-efficient ZF/RZF processors remain practical bottlenecks.