Cell-Free Massive MIMO under a Non-Linear Power Amplifier Consumption Model

Hazem Sallouha; Robbert Beerten; Sofie Pollin; Vida Ranjbar

REVIEW 4 major objections 5 minor 16 references

Cell-Free Massive MIMO under a Non-Linear Power Amplifier Consumption Model

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Optimizing the non-linear power-amplifier consumption model directly saves up to 24.9% of amplifier power in low-demand cell-free massive MIMO, while a penalty method matches generic solvers within 0.21% error.

desk verdict Worth a serious referee: the optimization is sound and the problem is new, but the 24% saving claim rests on a low-power PA model that the paper never validates. read the letter →

arxiv 2506.06799 v1 pith:WVPYXWQ4 submitted 2025-06-07 eess.SP

Robbert Beerten , Vida Ranjbar , Hazem Sallouha , Sofie Pollin This is my paper · ORCID

classification eess.SP

keywords cell-freemassiveMIMOpoweramplifierconsumptionmodelnon-linearpenaltymethodacceleratedprojectedgradientenergyefficiencydownlinkallocationsparsity

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that cell-free massive MIMO (a dense architecture in which many distributed access points jointly serve users) has been optimized with the wrong objective: minimizing transmitted power under an idealized linear amplifier model. When the power amplifier's real behavior (efficiency increases with output power) is put into the cost function, directly minimizing consumed power changes the allocation and saves up to 24.9% of amplifier power in low-demand 50-AP networks. To make the problem tractable at scale, the authors design a penalty method that alternates between a smoothed projected gradient step and increasing penalty weights, with complexity $O(L^2 K^2 I_{\mathrm{APG}} I_{\mathrm{penalty}})$ instead of the SOCP solver's $O(\sqrt{K+L+1}\,K^4 L^3)$. The resulting solutions are within 0.21% mean relative error in total consumed power of generic convex solvers, and the non-linear model naturally concentrates transmit power on a subset of access points.

What carries the argument

The load-bearing object is the non-linear PA consumption model $P_l^{\mathrm{non\text{-}linear}}(P_l^{\mathrm{tx}}) = \eta_{\max}^{-1}\sqrt{P_l^{\mathrm{tx}} P_{\max}}$, a square-root relation in which amplifier efficiency rises with output power. This model is what makes the cost function reward concentrating power, what penalizes small per-AP powers enough to induce sparsity, and what makes the gradient $\nabla P_l = \frac{\sqrt{P_{\max}}}{\eta_{\max}\|x_l\|_2} x_l$ singular at zero transmit power. The method handles that singularity with a standard smoothing function $\psi_\mu(\|x_l\|_2)$, then solves the penalized problem via accelerated projected gradient with backtracking line search and a closed-form per-AP projection, nested in a penalty loop that raises the weight $\lambda$ until the QoS constraints are satisfied.

What would settle it

Measure the DC input power of a representative class-B PA across the transmit-power range used in the simulations, down to a few percent of $P_{\max}$, and compare the measured curve with $P = (1/\eta_{\max})\sqrt{P_{\mathrm{tx}} P_{\max}}$; if the curve deviates substantially at low power, the 24.9% savings would not reproduce, and a simulation using the measured curve in place of (9) would settle the matter.

Watch

Extended reading notes

Core claim

The central discovery is that replacing the ideal linear amplifier model $P_l^{\mathrm{ideal}} = P_l^{\mathrm{tx}}/\eta$ with the non-linear model $P_l^{\mathrm{non\text{-}linear}} = (1/\eta_{\max})\sqrt{P_l^{\mathrm{tx}} P_{\max}}$ in the constrained power-allocation problem changes the optimal allocation itself. Because the square-root model penalizes small transmit powers more heavily, the optimizer concentrates power on fewer access points and effectively switches the rest off, a sparsity pattern earlier works pursued through explicit AP-switching algorithms. In simulations, this yields 24.9% PA-power savings at 10% of the max-min rate for 50 APs, with savings declining as demand rises. The proposed penalty algorithm solves the smoothed penalized problem via accelerated projected gradient with backtracking line search and a closed-form per-AP projection, matching a generic SOCP solver to 0.21% mean relative error on total consumed power while scaling better with network size.

Load-bearing premise

The load-bearing premise is that a real class-B power amplifier consumes input power roughly as the square root of transmitted power, $P_{\mathrm{consumed}} = (1/\eta_{\max})\sqrt{P_{\mathrm{tx}} P_{\max}}$, even at the low output levels where the reported savings are largest.

Editorial extensions

If this is right

In low-demand periods, optimizing with the non-linear PA model directly rather than minimizing transmitted power gives roughly 25% lower amplifier power, with the exact saving depending on network size and target rate.
The penalty method's complexity grows as $O(L^2 K^2 I_{\mathrm{APG}} I_{\mathrm{penalty}})$ instead of the SOCP's $O(\sqrt{K+L+1}\,K^4 L^3)$, so it can re-solve the allocation often enough to track mobility and changing quality-of-service requirements.
The non-linear model induces per-AP transmit-power sparsity without any switching constraint, meaning the optimized solution itself identifies which access points can be turned off for further energy savings.
Savings shrink as the target rate approaches the max-min rate, so the benefit of PA-aware allocation is concentrated in the low-demand regime that the paper argues is under-studied.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper leaves implicit that the same penalty-plus-smoothing template could be applied to other per-AP cost functions, such as measured PA lookup tables, without changing the projection or penalty loop.
If the square-root model holds down to very low output power, the induced sparsity could be combined with an AP sleep-mode policy that uses the optimized power vector to decide which access points to switch off, potentially reducing fixed-power overhead as well.
Because the runtime comparison uses fixed stopping thresholds, a practical extension would be to loosen the inner tolerances when the network is lightly loaded, trading a small accuracy loss for faster re-optimization under user mobility.
A direct experimental test would compare the optimized power allocations produced with model (9) against those produced with a measured PA efficiency curve at low output power; the paper does not provide such hardware measurements.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here's my read on arXiv:2506.06799. The paper does something genuinely new: it formulates downlink power allocation in cell-free massive MIMO to minimize a non-linear PA consumption model rather than transmit power. The convexity argument is correct, the penalty-method approach is standard but adapted cleanly, and the reported 0.21% mean error against an SOCP solver is strong evidence that the implementation is correct. The sparsity induced at low demand is a nice observation with practical potential for AP switching. Credit where due: this is a solid engineering contribution to an understudied regime.

The soft spot is the load-bearing kind. All the headline savings—24.9% at low demand—are computed under Eq. (9), which has no idle/bias power floor and predicts zero consumption at zero transmit power. The model is taken from earlier work and described as 'quite accurate' far from saturation, but the savings are largest precisely at deep backoff where that accuracy is least documented. If real PAs have a per-AP quiescent term, the denominator in the saving ratio increases and the optimizer's preference for sparsity may not map to real consumption. The paper also lacks any sensitivity analysis to an idle-power term or to the PA model itself. That makes the headline number a property of the model rather than of real hardware, and the authors should be pushed to show the result is robust.

Other issues are minor: the runtime comparison is only against CVX and the proposed method is slower at 200 APs; the smoothness proof is omitted; no code or error bars. None of these undercut the optimization method itself.

Who this is for: researchers working on energy efficiency in cell-free massive MIMO, particularly low-traffic operation. They will find the formulation and penalty method useful, with the caveat above. I would send this to peer review: the contribution is well-scoped and the central method is sound, but the authors need to confront the PA-model uncertainty before the 24% claim is publishable as a quantitative result.

Referee Report

4 major / 5 minor

Summary. The paper considers downlink power allocation in cell-free massive MIMO under a non-linear power-amplifier consumption model. The authors formulate a convex problem that minimizes the sum of per-AP PA consumptions, where each PA consumption is proportional to the Euclidean norm of the transmit-power vector of that AP. They propose a penalty method combined with accelerated projected gradient descent and Nesterov-type smoothing of the non-smooth norm at zero. Simulation results show that the method scales better than a generic SOCP solver (CVX/SDPT3) and that, when the non-linear model is the true model, optimizing it directly rather than the ideal linear model yields up to 24.9% PA-power savings at low user demand, while inducing AP sparsity. The authors conclude that directly minimizing non-linear PA consumption is beneficial in low-demand cell-free networks and that their first-order method is suitable for large networks.

Significance. If the PA model is accepted, the paper makes a useful and timely contribution. The convex formulation is correct: the non-linear PA consumption is a sum of Euclidean norms, so the problem remains convex, and the comparison against CVX/SDPT3 with a reported 0.21% mean error supports the correctness of the implementation. The induced sparsity is a novel and falsifiable prediction that could inform AP on/off switching. The scalability claim is also relevant, since generic SOCP solvers are known to struggle with large cell-free networks. The authors are appropriately careful in noting the smoothing error bound and in referencing prior work for the penalty-method convergence. However, the headline saving of up to 24.9% is computed under a PA model that is least validated in exactly the deep-backoff operating regime where the savings are largest, and several algorithmic details needed to reproduce or verify the results are missing. The central mathematical machinery is sound, but the paper's main quantitative claim about real PA power consumption is not yet supported to the same standard.

major comments (4)

[Section II-A, Eq. (9), and Section IV-B, Fig. 2] The non-linear PA model P_nonlinear = (1/eta_max) sqrt(P_tx Pmax) has no idle/bias power floor and predicts P_nonlinear -> 0 as P_tx -> 0. The abstract's headline saving of up to 24.9% is reported at 10% of max-min rate, i.e., in deep backoff, where this model is least supported. The text says the model is 'quite accurate when working sufficiently far from the saturation point' but gives no citation or measurement for this low-power regime, and references [7]-[9] do not validate the model at the operating points selected here. Please add a sensitivity analysis with a per-AP PA idle-power term, e.g., P = P_idle + (1/eta_max) sqrt(P_tx Pmax), and re-evaluate the savings and sparsity under that variant. If no such analysis is added, the 'up to 24%' claim should be explicitly rephrased as a property of the assumed model rather than of real PA power consumption.
[Section III-B, Algorithms 1-2] The convergence of Algorithm 2 rests on the claim that the penalty functions are smooth, but the proof is 'omitted here due to space limitations' and only a similar proof in [2] is cited. Since the APG method requires a Lipschitz-continuous gradient, the omitted proof is load-bearing for the algorithmic contribution. Please provide the argument or state the Lipschitz constant, and clarify the role of the strong-convexity assumption: the sum-of-Euclidean-norms term in f^{lambda(i)}(x) is not strongly convex on the feasible set, so the conditions under which Algorithm 2 is guaranteed to converge should be stated precisely.
[Section IV-B] The reported savings and the 0.21% error versus CVX are not accompanied by a description of the simulation setup (number of users, noise power, pilot power, Pmax, number of channel realizations, confidence intervals). Figure 3 is described as a 'single realization,' which raises the question of whether the 24.9% figure in Fig. 2 is an average over many realizations or a single draw. Please specify the parameter table and report mean and spread (or a box plot) over channel realizations for the key saving numbers.
[Section II-B, Eq. (10b)] The inequality direction in the SINR constraint as printed is incorrect: it reads gamma_k <= [achievable SINR expression], whereas the intended requirement is that the achievable SINR be at least the target gamma_k. The later SOCP reformulation in Eq. (13) is correct, so this is likely a typographical error, but it should be fixed because P1 is the central problem statement.

minor comments (5)

[Section III-D, Eq. (24)] The projection operator PC is written with a slight abuse of notation: it maps x_l to the projection of [x_l]_+ onto the Euclidean ball of radius sqrt(Pmax), but the expression as written 'max(sqrt(Pmax), ||[x_l]_+||_2)' should use the norm in the denominator consistently. Please clarify the normalization.
[Section III-B, Algorithm 2] Algorithm 2 uses mu_t but mu_0 is never defined, and the update for y_t at t=1 therefore refers to an undefined quantity. Also, alpha_t appears in the projection step but its backtracking rule is only described in the text; please make the pseudocode self-contained.
[Section III-A, Algorithm 1] The termination condition 'Psi_k(x_i) approx 0' is used without a numerical tolerance. Please state the tolerance and, if possible, report the worst-case constraint violation of the final solution.
[Section IV-A] The complexity expression for the SOCP solver, O(sqrt(K+L+1) K^4 L^3), is stated without a derivation or reference. Since the comparison with the proposed method is a selling point, please justify or cite this complexity estimate.
[Section IV-B] The notion of 'fraction of max-min rate' is used but the max-min rate computation is not described. Please state how the max-min rate is obtained and how the target rate is set as a fraction of it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported power saving is an optimality-gap comparison under an externally sourced PA model, with no fitted parameter defining the result.

full rationale

The paper's central claim, up to 24.9% PA-power savings, is computed by Eq. (26), which compares the nonlinear-model objective at the solution of the ideal-power problem, x*_ideal, with the objective at the solution of the nonlinear-power problem, x*_nonlinear. Because x*_nonlinear is by definition the minimizer of the nonlinear objective, the numerator is nonnegative and the saving is structural, but the reported 24.9% figure is obtained from simulated channel realizations at a stated demand level (10% of the max-min rate, 50 APs) and is cross-checked against CVX/SDPT3 with a 0.21% mean relative error. No parameter is fitted to force the saving. The nonlinear PA model in Eq. (9) is attributed to external amplifier-aware power-allocation references [7]-[9], not to the authors' own prior work. The only self-citation is [6], used to rewrite the SINR constraint (10b) into the SOC form (13); that is a standard algebraic reformulation that can be verified from the paper's own SINR expression and does not carry the claimed result. Potential weaknesses, such as the unvalidated accuracy of sqrt(P_tx Pmax) at deep backoff, the omitted smoothness proof 'inferred from [2]', and the empirically chosen initialization and mu, are correctness or robustness concerns rather than circular steps.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new entities are introduced. The paper reuses an existing PA model and standard optimization techniques. The only model-specific object is the smoothed norm gradient, which is a numerical approximation, not a physical entity. Free parameters are algorithmic tuning constants, not physical parameters fitted to data.

free parameters (6)

Initial point x0 range = uniform random in [0, 10^-10]
The paper states the initialization is chosen empirically and 'future work might want to investigate this starting point' (Section III-A). The algorithm's convergence to the global optimum should not depend strongly on this for a convex problem, but it is an ad hoc choice.
Smoothing parameter mu = 10^-7
Chosen by hand in Section III-C; controls the smoothing approximation of the norm gradient near zero and introduces an error bounded by mu/2. The paper notes it could be optimized in future work.
Penalty scaling factor zeta = 3
Chosen in Section III-D to increase the penalty weight each outer iteration. A standard heuristic; convergence proof is referenced, not specific to this value.
Armijo parameter tau = 10^-4
Stepsize backtracking parameter in Algorithm 2, Section III-D. Standard choice, not fitted to data.
Termination tolerance epsilon = 10^-3
Relative objective decrease threshold in Algorithm 2. Affects runtime and accuracy; the paper notes lower precision improves runtime.
Initial penalty weight lambda(0) = 0.1
Starting penalty weight in Algorithm 1. Empirically chosen, not derived.

assumptions (5)

domain assumption The non-linear PA model (Eq. 9) accurately describes class B amplifier consumption when operating far from saturation.
Section II-A. The model is taken from [7]-[9]. If this physical model is inaccurate at the low-power operating points where the savings are largest, the central claim is weakened.
domain assumption The achievable downlink SINR expression (Eq. 4) and MMSE precoding (Eq. 3) correctly characterize the cell-free massive MIMO system.
Section II. These are standard results from [1], but the paper does not re-derive them and assumes the CPU has the needed statistical knowledge (bk, Cki).
standard math The penalty method converges to a solution of P1 when the penalty weight grows.
Section III-A, reference to [13]. The paper relies on a known convergence theorem, but does not prove the specific problem satisfies the required conditions (e.g., exact penalty, boundedness).
standard math The objective P2 is smooth and strongly convex, enabling APG convergence.
Section III-B. The paper states 'If P2 is smooth and strongly convex' and cites a similar proof in [2] that is omitted here. The smoothing of the non-linear term is used, but the strong convexity of the full penalized function is not established in the text.
domain assumption The channel is Rayleigh fading with a one-ring scattering covariance model.
Section II. The simulation relies on this statistical channel model from [11]; results may differ with other channel models.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cell-Free Massive MIMO under a Non-Linear Power Amplifier Consumption Model." pith.science (2026). https://pith.science/paper/WVPYXWQ4

@misc{pith2026250606799,
  author       = {Pith},
  title        = {Pith review of: Cell-Free Massive MIMO under a Non-Linear Power Amplifier Consumption Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WVPYXWQ4}},
  note         = {Machine review of arXiv:2506.06799}
}

read the original abstract

Existing works on Cell-Free Massive MIMO primarily focus on optimising system throughput and energy efficiency under high-traffic scenarios with only a limited focus on variable user demand as required by higher network layers. Additionally, existing works only minimise the transmitted power instead of the consumed power at the power amplifier. This work introduces a penalty-method-based approach to minimise the amplifier's power consumption while scaling much better with network size than current solutions and promoting sparsity in the power allocated to each access point. Furthermore, we demonstrate substantial reductions in power consumption (up to 24%) by considering the non-linear power consumption.

Figures

Figures reproduced from arXiv: 2506.06799 by the authors.

**Figure 2.** The relative power consumption saving ((26)) when accurately ac [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** The total transmit power allocated to each AP for both power [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[7]

Amplifier-Aware Multiple- Input Multiple-Output Power Allocation,

D. Persson, T. Eriksson, and E. G. Larsson, “Amplifier-Aware Multiple- Input Multiple-Output Power Allocation,” IEEE Communications Let- ters, vol. 17, p. 1112–1115, June 2013

work page 2013
[9]

Energy-Saving Precoder Design for Narrowband and Wideband Massive MIMO,

E. Peschiera and F. Rottenberg, “Energy-Saving Precoder Design for Narrowband and Wideband Massive MIMO,” IEEE Transactions on Green Communications and Networking , vol. 7, Dec. 2023

work page 2023
[2]

Energy Efficiency Maximiza- tion in Large-Scale Cell-Free Massive MIMO: A Projected Gradient Approach,

T. C. Mai, H. Q. Ngo, and L.-N. Tran, “Energy Efficiency Maximiza- tion in Large-Scale Cell-Free Massive MIMO: A Projected Gradient Approach,” IEEE Transactions on Wireless Communications, Aug. 2022

work page 2022
[1]

Foundations of User- Centric Cell-Free Massive MIMO,

¨O. T. Demir, E. Bj ¨ornson, and L. Sanguinetti, “Foundations of User- Centric Cell-Free Massive MIMO,” F oundations and Trends® in Signal Processing, vol. 14, no. 3–4, p. 162–472, 2021

work page 2021
[3]

How much energy is needed to run a wireless network?,

G. Auer, V . Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson, M. Imran, D. Sabella, M. Gonzalez, O. Blume, and A. Fehske, “How much energy is needed to run a wireless network?,” IEEE Wireless Communications, vol. 18, p. 40–49, Oct. 2011

work page 2011
[4]

Machine Learning and Analytical Power Consumption Models for 5G Base Stations,

N. Piovesan, D. Lopez-Perez, A. De Domenico, X. Geng, H. Bao, and M. Debbah, “Machine Learning and Analytical Power Consumption Models for 5G Base Stations,” IEEE Communications Magazine, vol. 60, p. 56–62, Oct. 2022

work page 2022
[5]

Cell-Free Massive MIMO in O-RAN: Energy-Aware Joint Orchestration of Cloud, Fronthaul, and Radio Resources,

¨O. T. Demir, M. Masoudi, E. Bj ¨ornson, and C. Cavdar, “Cell-Free Massive MIMO in O-RAN: Energy-Aware Joint Orchestration of Cloud, Fronthaul, and Radio Resources,” IEEE Journal on Selected Areas in Communications, vol. 42, p. 356–372, Feb. 2024

work page 2024
[6]

Location-Based Load Balancing for Energy-Efficient Cell-Free Net- works,

R. Beerten, V . Ranjbar, A. P. Guevara, H. Sallouha, and S. Pollin, “Location-Based Load Balancing for Energy-Efficient Cell-Free Net- works,” in 2024 Joint European Conference on Networks and Commu- nications & 6G Summit (EuCNC/6G Summit) , pp. 558–563, 2024

work page 2024

Show all 16 references

[8]

Information-Theoretic Study of Time-Domain Energy- Saving Techniques in Radio Access,

F. Rottenberg, “Information-Theoretic Study of Time-Domain Energy- Saving Techniques in Radio Access,” IEEE Transactions on Green Communications and Networking , 2024

2024
[10]

Utility Maxi- mization for Large-Scale Cell-Free Massive MIMO Downlink,

M. Farooq, H. Q. Ngo, E.-K. Hong, and L.-N. Tran, “Utility Maxi- mization for Large-Scale Cell-Free Massive MIMO Downlink,” IEEE Transactions on Communications , vol. 69, p. 7050–7062, Oct. 2021

2021
[11]

Massive MIMO Networks: Spectral, Energy, and Hardware Efficiency,

E. Bj ¨ornson, J. Hoydis, and L. Sanguinetti, “Massive MIMO Networks: Spectral, Energy, and Hardware Efficiency,” F oundations and Trends® in Signal Processing , vol. 11, no. 3-4, pp. 154–655, 2017

2017
[12]

CVX: Matlab software for disciplined convex programming, version 2.1

M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex programming, version 2.1.” https://cvxr.com/cvx, Mar. 2014

2014
[13]

Lecture Notes Optimization III: Con- vex Analysis, Nonlinear Programming Theory, Nonlinear Programming Algorithms,

A. Ben-Tal and A. Nemirovski, “Lecture Notes Optimization III: Con- vex Analysis, Nonlinear Programming Theory, Nonlinear Programming Algorithms,” 2023

2023
[14]

Smooth minimization of non-smooth functions,

Y . Nesterov, “Smooth minimization of non-smooth functions,” Mathe- matical Programming, vol. 103, p. 127–152, May 2005

2005
[15]

Fast Gradient-Based Algorithms for Con- strained Total Variation Image Denoising and Deblurring Problems,

A. Beck and M. Teboulle, “Fast Gradient-Based Algorithms for Con- strained Total Variation Image Denoising and Deblurring Problems,” IEEE Transactions on Image Processing , vol. 18, Nov. 2009

2009
[16]

Projecting onto the Intersection of a Cone and a Sphere,

H. H. Bauschke, M. N. Bui, and X. Wang, “Projecting onto the Intersection of a Cone and a Sphere,” SIAM Journal on Optimization , vol. 28, no. 3, pp. 2158–2188, 2018

2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [7]

Amplifier-Aware Multiple- Input Multiple-Output Power Allocation,

D. Persson, T. Eriksson, and E. G. Larsson, “Amplifier-Aware Multiple- Input Multiple-Output Power Allocation,” IEEE Communications Let- ters, vol. 17, p. 1112–1115, June 2013

work page 2013

[2] [9]

Energy-Saving Precoder Design for Narrowband and Wideband Massive MIMO,

E. Peschiera and F. Rottenberg, “Energy-Saving Precoder Design for Narrowband and Wideband Massive MIMO,” IEEE Transactions on Green Communications and Networking , vol. 7, Dec. 2023

work page 2023

[3] [2]

Energy Efficiency Maximiza- tion in Large-Scale Cell-Free Massive MIMO: A Projected Gradient Approach,

T. C. Mai, H. Q. Ngo, and L.-N. Tran, “Energy Efficiency Maximiza- tion in Large-Scale Cell-Free Massive MIMO: A Projected Gradient Approach,” IEEE Transactions on Wireless Communications, Aug. 2022

work page 2022

[4] [1]

Foundations of User- Centric Cell-Free Massive MIMO,

¨O. T. Demir, E. Bj ¨ornson, and L. Sanguinetti, “Foundations of User- Centric Cell-Free Massive MIMO,” F oundations and Trends® in Signal Processing, vol. 14, no. 3–4, p. 162–472, 2021

work page 2021

[5] [3]

How much energy is needed to run a wireless network?,

G. Auer, V . Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson, M. Imran, D. Sabella, M. Gonzalez, O. Blume, and A. Fehske, “How much energy is needed to run a wireless network?,” IEEE Wireless Communications, vol. 18, p. 40–49, Oct. 2011

work page 2011

[6] [4]

Machine Learning and Analytical Power Consumption Models for 5G Base Stations,

N. Piovesan, D. Lopez-Perez, A. De Domenico, X. Geng, H. Bao, and M. Debbah, “Machine Learning and Analytical Power Consumption Models for 5G Base Stations,” IEEE Communications Magazine, vol. 60, p. 56–62, Oct. 2022

work page 2022

[7] [5]

Cell-Free Massive MIMO in O-RAN: Energy-Aware Joint Orchestration of Cloud, Fronthaul, and Radio Resources,

¨O. T. Demir, M. Masoudi, E. Bj ¨ornson, and C. Cavdar, “Cell-Free Massive MIMO in O-RAN: Energy-Aware Joint Orchestration of Cloud, Fronthaul, and Radio Resources,” IEEE Journal on Selected Areas in Communications, vol. 42, p. 356–372, Feb. 2024

work page 2024

[8] [6]

Location-Based Load Balancing for Energy-Efficient Cell-Free Net- works,

R. Beerten, V . Ranjbar, A. P. Guevara, H. Sallouha, and S. Pollin, “Location-Based Load Balancing for Energy-Efficient Cell-Free Net- works,” in 2024 Joint European Conference on Networks and Commu- nications & 6G Summit (EuCNC/6G Summit) , pp. 558–563, 2024

work page 2024