Deep-Unfolded Accelerated Projected Gradient for Energy-Efficient Cell-Free Massive MIMO

arXiv:2608.03237 · eess.SP · Submitted 2026-08-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Deep-Unfolded Accelerated Projected Gradient for Energy-Efficient Cell-Free Massive MIMO".

Dev: Here's the gist of the paper. This paper investigates energy efficiency (EE) maximization for the downlink of cell-free massive multipleinput multiple-output systems under quality-of-service and peraccess point power constraints.

Rosa: Problem Formulation The energy efficiency maximization problem is formulated as maximizing EE subject to minimum SE requirements and transmit power limits.

Dev: The original problem (P1) is non-convex due to the fractional structure of the objective function and the coupling of the optimization variables in the SINR expressions.

Rosa: To enable gradient-based optimization, a reformulation introduces variable θl,k = √ρl,k to avoid square-root terms in the SINR and QoS expressions.

Paper summary: Rosa: So looking at the whole paper, "Deep-Unfolded Accelerated Projected Gradient for Energy-Efficient Cell-Free Massive MIMO", this work proposes a specific way to handle energy efficiency maximization in cell-free massive MIMO systems with both quality of service and per access point power constraints.

Dev: The authors are Phuong Nam Tran, Nhan Thanh Nguyen, Hien Quoc Ngo, and Markku Juntti from the Centre for Wireless Communications at the University of Oulu in Finland and Queen's University Belfast in the UK.

Taro: What this means for us is that if you're running massive MIMO systems cell-free, you can use this framework to get good energy efficiency results without having to worry as much about tuning your step sizes or penalty parameters manually during the optimization process.

Rosa: It suggests that deep unfolding can be effectively applied to solve complex optimization problems in wireless communications because it gives you a fixed computational cost determined by T, which is really important when computational cost is a critical concern.

Dev: Exactly. The ability of the method to learn step sizes and penalty parameters makes it highly adaptable for various system settings, and that fixed complexity offers a significant advantage in scenarios where running the optimization needs to be fast without introducing extra per-iteration complexity.

Conclusion: Rosa: So, we're wrapping up on this paper about Deep-Unfolded Accelerated Projected Gradient for Energy-Efficient Cell-Free Massive MIMO by Phuong Nam Tran and his team.

Dev: Basically, it takes a complex optimization problem with quality of service and power limits and turns it into a deep unfolded network.

Rosa: It maps the iterative updates of this APG method into several trainable layers, so you're learning how to solve the problem instead of tuning everything by hand.

Taro: The big thing is that they learn the step sizes and those penalty parameters directly from data, which cuts out all that tedious line search stuff we usually have to do.

Dev: That fixed computational cost determined by the number of layers, T, is a huge win for real-time systems where you need predictable performance.

Rosa: The results show it gets energy efficiency numbers comparable to standard APG but with way less computation required.

Taro: For the MRT precoding setup, it's almost as good as standard APG after just ten iterations, which suggests the algorithm finds a good spot fast.

Dev: But for PZF precoding, it shows a huge efficiency gain, like nearly two hundred percent improvement at twenty-five dBm power levels.

Rosa: It keeps the performance stable even when the transmit power gets really high, unlike some other methods that might wobble.

Taro: It seems they've really managed to make this deep unfolding work for these kinds of massive cell-free systems under those tough constraints.

Dev: So, what does this mean for us? It suggests we can use these deep unfolding techniques to tackle similar heavy optimization problems in other wireless setups.

Centre for Wireless Communications, University of Oulu, Finland · School of Electronics, Electrical Engineering and Computer Science, Queen's University Belfast, UK

eess.SP

Submitted: 2026-08-04

Updated: 2026-11-09

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 73/100

The gist: This paper investigates energy efficiency (EE) maximization for the downlink of cell-free massive multipleinput multiple-output systems under quality-of-service and peraccess point power constraints.

Key concepts

Deep Unfolding
This technique transforms an iterative optimization algorithm (like APG) into a fixed, multi-layer neural network structure. Instead of running the full iterative process every time, the network learns how to perform each step in a single forward pass, making it faster and more efficient.
APG Optimization Framework
This is an optimization technique that accelerates convergence by using information from previous iterations to determine the current search direction. It helps find the optimal solution much quicker than standard gradient methods by intelligently combining past steps into the next update rule.
Energy Efficiency (EE) Maximization
The goal is to maximize EE, which measures how effectively energy is used to transmit data. The system tries to find the best way for cell-free massive MIMO access points to allocate transmit power while meeting user quality requirements and staying within power limits.
Penalty Formulation
Because the quality-of-service constraints make the optimization problem difficult (non-convex), these constraints are added to the main objective function using a penalty term. This allows the optimization algorithm to handle complex rules by adding a cost for violating those rules.

Terminology

Summary

This paper investigates energy efficiency (EE) maximization for the downlink of cell-free massive multipleinput multiple-output systems under quality-of-service and peraccess point power constraints. p. 0

Problem Formulation

The energy efficiency maximization problem is formulated as maximizing EE subject to minimum SE requirements and transmit power limits. The original problem (P1) is non-convex due to the fractional structure of the objective function and the coupling of the optimization variables in the SINR expressions. p. 1 To enable gradient-based optimization, a reformulation introduces variable θl,k = √ρl,k to avoid square-root terms in the SINR and QoS expressions. p. 1 The problem is then transformed into (P2) by incorporating QoS constraints into the objective using a penalty formulation. p. 2

How the training signal is built

The deep-unfolded approach maps the iterative APG updates into a finite number of trainable layers, where step sizes and penalty coefficients are learned from data. Each layer corresponds to an iteration of the APG method in (8)–(11). p. 3 The network structure replaces the non-differentiable selection step in (11) with a learnable weighted combination θ(t+1) = wt z(t+1) + (1 − wt) v(t+1), where wt ∈ (0, 1) is a learnable parameter. p. 3 Learnable parameters include step sizes αy,n and αθ,n which are constructed by applying learnable scaling factors to the Barzilai–Borwein (BB) estimates. p. 3

Where the gains come from

The primary advantage of this method is that it eliminates the need for line search or iterative tuning by learning step sizes and penalty parameters from data. This results in a fixed computational cost determined by T layers, unlike conventional APG which has variable computational cost depending on channel realizations. p. 4 Numerical results show that the proposed method achieves EE performance comparable to iterative APG while requiring significantly lower computational cost. p. 0

Computational Complexity Analysis

The computational complexity of the unfolded network is O(T LK 2), which is fixed regardless of channel realizations. In contrast, the iterative APG algorithm has complexity O(IP IAPGLK 2), where IP and IAPG denote the number of outer and inner iterations, respectively. p. 4 The paper demonstrates that for (L, K) = (20, 6), iterative APG requires approximately 30 times more FLOPs than the unfolded method. p. 5 This confirms the scalability of the proposed approach. p. 5

Performance Comparison

The unfolded APG consistently achieves higher EE than APG and HCD across all transmit power levels for both PZF and MRT precoding. For MRT, its performance is close to that of APG under the same number of APG iterations, indicating that APG already reaches a near-stationary performance level within 10 iterations. p. 4 PZF achieves significantly higher EE under all settings; for example, at ρmax = 25 dBm, PZF provides an EE gain of approximately 198%. p. 5 The proposed method maintains more stable performance at higher transmit power levels compared to APG and HCD. p. 5

Conclusion

This paper proposes a deep-unfolded APG method for EE maximization in downlink CF-mMIMO systems under QoS and per-AP power constraints. The proposed method achieves EE performance comparable to iterative APG while requiring significantly lower computational cost, indicating its scalability for CF-mMIMO systems. p. 5 The method successfully addresses the limitation of iterative APG methods by mapping updates into a finite number of trainable layers. p. 0 The resulting framework provides a fixed computational complexity determined by T, ensuring fast convergence without introducing extra per-iteration complexity. p. 4 The method's effectiveness is validated through numerical results comparing it against iterative APG and HCD schemes. The work suggests that deep unfolding can be effectively applied to solve complex optimization problems in wireless communications. p. 0 The method's ability to learn step sizes and penalty parameters makes it highly adaptable for various system settings. The fixed complexity offers a significant advantage in scenarios where computational cost is a critical concern. p. 4 The paper provides a robust framework for energy-efficient power control in cell-free massive MIMO architectures. p. 5 The proposed deep-unfolded APG method is highly scalable for CF-mMIMO systems. p. 5

--- Page 1 ---

Deep-Unfolded Accelerated Projected Gradient for Energy-Efficient Cell-Free Massive MIMO Phuong Nam Tran∗, Nhan Thanh Nguyen∗, Hien Quoc Ngo†, Markku Juntti∗ Centre for Wireless Communications (CWC), University of Oulu, Finland, †School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast, UK

--- Page 2 ---

A. System Model A. p. 0 System Model We consider the downlink of a CF-mMIMO system operating in time-division duplex (TDD) mode. p. 1 p. 1 The network consists of L APs, each equipped with M antennas, that jointly serve K single-antenna user equipments (UEs) over the same time-frequency resource, with LM ≫ K [9]. p. 1 p. 1 All APs are connected to a central processing unit (CPU) via fronthaul links. p. 1 p. 1 The channel between AP l and UE k is denoted by hl,k ∈ CM×1 and follows hl,k ∼ CN (0, βl,kIM), where βl,k represents the large-scale fading coefficient. p. 1 p. 1 During uplink training, each UE transmits a pilot sequence ϕik ∈ C τp with∥ϕik∥2 = τp, where ik is the pilot index and τp is the pilot length. p. 1 p. 1 We assume τp ≤ K, and thus pilot reuse occurs. p. 1 p. 1 The set of UEs sharing the same pilot as UE k is denoted by Pk = 1t: it = ik, which leads to pilot contamination. p. 1 p. 1 Let pk denote the uplink transmit power of UE k. p. 1 p. 1 The MMSE estimate of hl,k, denoted by hˆ l,k, has zero mean and covariance γl,kIM, where γl,k = τppkβ2 l, k τp P t∈Pk ptβl + 1. p. 1 (1). We assume that each AP l employs PZF precoding, as it provides very good performance [9]. p. 1 p. 1 The set of UEs is partitioned into a strong set Sl and a weak set Wl = 1,. p. 1 p. 1.., K) Sl based on the large-scale fading coefficients βl,k. p. 1 p. 1 Zero-forcing (ZF) is applied to UEs in Sl to suppress interference within pilot subspaces, while MRT is used for UEs in Wl to reduce computational complexity. p. 1 p. 1 This structure allows interference suppression for UEs with strong channels while maintaining low complexity for the remaining UEs. p. 1 p. 1 Since ZF is applied over pilot subspaces, if k ∈ Sl, then all UEs sharing the same pilot as UE k belong to Sl. p. 1 p. 1 Let τSl denote the number of distinct pilots assigned to Sl, and the condition M > τSl is required for ZF. p. 1 p. 1 The downlink transmit signal from AP l is xl = K k=1 √ρl,k wl,kqk, where ρl,k ≥ 0 denotes the transmit power allocated by AP l to UE k, wl,k is the precoding vector. p. 1 An achievable downlink SE for UE k under local PZF is given by SEk = τc−τp τc log2 1 + SINRk, where τc is the coherence block length. p. 1 p. 1 The corresponding signal-to-interference-plus-noise ratio (SINR) is given by [9] SINRk = P L l=1 l=1 alk√ρl,k2 P t∈Pk k 1 P L l=1 bl,k t√ρl t2 + P K t=1 P L l=1 dl,k t ρl t + 1, where δl,k equals one if k ∈ Sl and zero otherwise. p. 1

--- Page 3 ---

II. System Model and Problem Formulation II. p. 1 System Model and Problem Formulation A. p. 1 System Model We consider the downlink of a CF-mMIMO system operating in time-division duplex (TDD) mode. The network consists of L APs, each equipped with M antennas, that jointly serve K single-antenna user equipments (UEs) over the same time-frequency resource, with LM ≫ K [9]. All APs are connected to a central processing unit (CPU) via fronthaul links. The channel between AP l and UE k is denoted by hl,k ∈ CM×1 and follows hl,k ∼ CN (0, βl,kIM), where βl,k represents the large-scale fading coefficient. During uplink training, each UE transmits a pilot sequence ϕik ∈ C τp with∥ϕik∥2 = τp, where ik is the pilot index and τp is the pilot length. We assume τp ≤ K, and thus pilot reuse occurs. The set of UEs sharing the same pilot as UE k is denoted by Pk = 1t: it = ik, which leads to pilot contamination. Let pk denote the uplink transmit power of UE k. The MMSE estimate of hl,k, denoted by hˆ l,k, has zero mean and covariance γl,kIM, where γl k = τppkβ2 l, k τp P t∈Pk ptβl + 1. p. 1 (1). We assume that each AP l employs PZF precoding, as it provides very good performance [9]. The set of UEs is partitioned into a strong set Sl and a weak set Wl = 1,..., K) Sl based on the large-scale fading coefficients βl,k. Zero-forcing (ZF) is applied to UEs in Sl to suppress interference within pilot subspaces, while MRT is used for UEs in Wl to reduce computational complexity. This structure allows interference suppression for UEs with strong channels while maintaining low complexity for the remaining UEs. Since ZF is applied over pilot subspaces, if k ∈ Sl, then all UEs sharing the same pilot as UE k belong to Sl. Let τSl denote the number of distinct pilots assigned to Sl, and the condition M > τSl is required for ZF. The downlink transmit signal from AP l is xl = K k=1 √ρl,k wl,kqk, where ρl k ≥ 0 denotes the transmit power allocated by AP l to UE k, wl k is the precoding vector. p. 1 An achievable downlink SE for UE k under local PZF is given by SEk = τc−τp τc log2 1 + SINRk, where τc is the coherence block length. The corresponding signal-to-interference-plus-noise ratio (SINR) is given by [9] SINRk = P L l=1 l=1 alk√ρl k2 P t∈Pk k 1 P L l=1 bl,k t√ρl t2 + P K t=1 P L l=1 dl,k t ρl t + 1, where δl,k equals one if k ∈ Sl and zero otherwise. p. 1

--- Page 4 ---

III. Proposed Deep-Unfolded APG Method III. p. 0 Proposed Deep-Unfolded APG Method To address (P1), we first develop an APG-based optimization framework and then construct a deep-unfolded network by mapping APG iterations into a finite number of trainable layers. p. 1 The problem is reformulated by introducing variable θl,k = √ρl,k ≥ 0 to avoid square-root terms in the SINR and QoS expressions and facilitate more efficient projection and gradient updates. p. 1 The feasible sets for the power allocation variables θl and θ are defined as Cl = n θl ∈ R K: θl,k ≥ 0, ∀k, PK k=1 θ squared l,k ≤ ρ max l o and C = Θ ∈ R L×K: θl ∈ Cl, ∀l, respectively. p. 2 The QoS constraints introduce non-convex coupling between the variables and cannot be handled directly within the APG updates. p. 2 Therefore, they are incorporated into the objective using a penalty formulation. p. 2 The reformulated problem is given by (P2): max θ∈C fξ(θ) = EE(θ)−ξ X K k=1 max(0, gk(θ)) squared, where ξ > 0 is a penalty parameter that controls the strength of constraint enforcement. p. 2

APG Optimization Framework

The APG method exploits the difference between successive iterates to capture the recent search direction, accelerating convergence. The update rule for y(n) incorporates information from previous updates, defined as y(n) = θ(n)+ sn-1 sn z(n−θ(n) + sn-1 - 1/sn θ(n−1), where sn = (1 + q 4s2 n−1)/2 controls the contribution of previous updates. p. 2 The update z(n+1) incorporates information from previous iterations and promotes faster progress. p. 2 The projection PC(·) enforces the per-AP power constraints independently to each θl. p. 2

Gradient of the Objective Function

The gradient of fξ(θ) with respect to θ ∈ R L×K is given by ∇fξ = ∇Ee − ξ ∇Ψ. The gradient of the energy efficiency (∇Ee) is derived from the total power consumption model. p. 1 The gradient of the penalty term (∇Ψ) is defined as ∇Ψ = X k: gk>0 2 gk ∇gk. p. 1

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Deep Unfolded APG Framework for Power Allocation

The system can solve non-convex energy efficiency maximization problems under quality-of-service and per access point power constraints by mapping iterative APG updates onto a finite number of trainable neural network layers. This eliminates the need for line search or manual parameter tuning, producing power allocation solutions through a fixed number of gradient based updates with fixed computational complexity.

  1. Learning Step Sizes and Penalty Parameters

The proposed deep unfolded approach allows step sizes, penalty coefficients, and the update selection mechanism to be learned from data training rather than being manually tuned. This eliminates the need for line search and iterative tuning, leading to a fixed inference complexity of O(T LK2).

  1. Enhanced Energy Efficiency Performance

The method achieves EE performance comparable to iterative APG while requiring significantly lower computational cost, with up to a 30-fold reduction in floating point operations under considered system settings. It maintains more stable performance at higher transmit power levels compared to iterative APG and HCD schemes.

  1. Scalable Computational Complexity

The proposed method exhibits fixed complexity determined by T, resulting in O(T LK2) overall complexity for the T layer network. This is superior to iterative APG which has variable computational cost depending on channel realizations and convergence criteria, demonstrating scalability for larger system sizes (L, K).

Sources

Related papers