Channel-Independent Step Sizes for Unfolded Energy-Efficiency Power Control

Deciding how much power each transmitter in a wireless network should use is an optimization problem that must be re-solved every few milliseconds, because the radio channel keeps changing. When the goal is energy efficiency, meaning bits delivered per Joule spent, the objective is a ratio, and ratios are not concave. The standard fix is Dinkelbach's method. It turns the ratio into a sequence of "rate minus price times power" problems, each solved by gradient steps. Those steps have to be tiny, because the safe step size under ordinary Euclidean geometry is set by the largest channel-gain-to-noise ratio in the network, and that ratio varies over many orders of magnitude with distance and fading. Recent work sidesteps this by unfolding the solver. Deep unfolding runs a fixed, small number of iterations as network layers and lets a neural network predict aggressive step sizes. This buys large speedups, but it gives up any guarantee that each layer improves the objective. It also forces the network to extrapolate step sizes whenever the channel statistics shift.

This proposal rests on a simple observation. If one measures distance between power vectors with the Burg entropy Bregman divergence, the weighted sum-rate objective of an interference network has a curvature bound that does not depend on the channel at all. Burg entropy is the divergence generated by $-\sum_i \log p_i$, which weights changes to small powers more heavily than changes to large ones. Under this geometry, a certified safe step exists in closed form for every channel draw. The mirror-descent update is elementwise and costs no more than a Euclidean gradient step, and the power budget is enforced exactly by clipping. We propose to unfold Dinkelbach's method in this geometry. A permutation-equivariant graph network predicts per-user multipliers on the certified step. It does not predict raw step sizes. The claim is that this yields unfolded energy-efficiency solvers that improve monotonically layer by layer when the multiplier is capped at one. They should also transfer across path-loss models and network sizes better than Euclidean unfolded solvers, because the network only has to learn a bounded, channel-normalized correction.

Core Claim

Consider $n$ transmitter–receiver pairs with powers $p \in (0, P_{\max}]^n$, channel gains $g_{kj} \ge 0$ from transmitter $j$ to receiver $k$, noise $\sigma_k > 0$, and per-user rates $R_k(p) = \log(\sigma_k + \sum_j g_{kj}p_j) - \log(\sigma_k + \sum_{j\ne k} g_{kj}p_j)$. Dinkelbach's method maximizes the global energy efficiency $\sum_k w_k R_k(p) / (\mathbf{1}^\top p + P_c)$, where $w_k$ are user weights and $P_c$ is circuit power. It does so by repeatedly minimizing

$$F_q(p) = -\sum_k w_k R_k(p) + q\,(\mathbf{1}^\top p + P_c),$$

and then resetting the price $q$ to the current ratio. Each rate term is a difference of two functions of the form $\log(c + a^\top p)$ with $a \ge 0$ and $c > 0$. For any direction $d$, a weighted Cauchy–Schwarz inequality gives $(a^\top d)^2/(c + a^\top p)^2 \le \sum_j d_j^2/p_j^2$. The right-hand side is exactly the curvature of the Burg kernel $\psi(p) = -\sum_j \log p_j$. The price term is linear and adds no curvature. Hence $F_q$ is smooth relative to $\psi$ with constant

$$L = 2\sum_k w_k,$$

which holds for every channel realization, every noise level, and every price $q \ge 0$. The Euclidean smoothness constant, by contrast, grows like $\max_{k,j}(g_{kj}/\sigma_k)^2$. This is the Poisson-type relative-smoothness argument applied to interference networks. To our knowledge, nothing in the literature we reviewed applies it to fractional power control.

The resulting mirror step with step $\eta$ has a closed form, coordinate by coordinate. The new power is $p_i^+ = 1/(1/p_i + \eta\,\partial_i F_q(p))$, set to $P_{\max}$ if the denominator is non-positive or the value exceeds the budget. Whenever $\eta \le 1/L$, the relative-smoothness descent lemma guarantees $F_q(p^+) \le F_q(p)$. By the standard Dinkelbach argument, every layer that takes such a step and then updates $q$ cannot decrease energy efficiency. The feasibility guarantee needs no learned projection, and the monotonicity guarantee needs no eigenvalue computation.

Relation to Prior Work

Three unfolded solvers motivate the work and supply the architecture. Graph-based Algorithm Unfolding for Energy-aware Power Allocation in Wireless Networks (USCA) unfolds successive concave approximation for weighted sum energy efficiency. It uses one graph network for the update direction and another for the step size, and adds a monotonicity-in-budget regularizer to the loss. Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation unfolds the weighted minimum mean-square-error algorithm with graph networks and shows zero-shot transfer across network sizes. DeepFP: Deep-Unfolded Fractional Programming for MIMO Beamforming replaces FastFP's eigenvalue-based step size with learned, user-specific steps, and reports that the objective fluctuates across layers as a result. All three learn step sizes in Euclidean geometry and get no per-layer guarantee. We keep their equivariant graph parameterization. We change the geometry so that the learned quantity is a dimensionless multiplier with a known safe range.

Two optimization papers supply the method and constrain the design. Bregman Finito/MISO for Nonconvex Regularized Finite Sum Minimization without Lipschitz Gradient Continuity shows that relative smoothness supports convergence analysis for nonconvex objectives without global Lipschitz gradients, using a Bregman–Moreau envelope. Fast Frank–Wolfe Algorithms with Adaptive Bregman Step-Size for Weakly Convex Functions shows empirically that Bregman steps beat Euclidean ones by orders of magnitude in primal gap on non-Lipschitz losses. Its rates are only local, however, which warns us not to over-claim global optimality. Neither paper treats ratios or unfolding.

ADMM for Structured Fractional Minimization handles single-ratio problems with Euclidean majorization-minimization and names sum-of-ratios as open. We do not solve that general problem. We give a guarantee for the single-ratio global energy efficiency and treat the sum-of-ratios case empirically. Learning to be green: robust energy efficiency maximization in dynamic MIMO-OFDM systems tracks energy efficiency online through a Charnes–Cooper reformulation with Euclidean semidefinite projections. It is the closest online reference point, but it is limited to settings without multi-cell interference ratios.

Research Program

The first deliverable is the analysis. It covers the relative-smoothness constant above, including a tighter weighted version that counts only the interferers a receiver actually sees. It also covers the monotonicity of the unfolded Dinkelbach–Burg iteration when multipliers are capped at one, and a check of whether the descent lemma survives replacing the global price with per-user prices. The per-user-price version is the standard parametric heuristic for weighted sum energy efficiency. There, the guarantee applies only to each layer's parametric objective, and we will measure how often true weighted sum energy efficiency still increases.

The second deliverable is the unfolded model. It has $K$ layers, with $K$ between 4 and 10 to match the prior unfolded systems. In each layer, a shared graph network reads channel gains, current powers, gradients, and the price. It outputs per-user multipliers $m_{k,i}$ and sets the step to $\eta_{k,i} = m_{k,i}/L$. In the certified variant, $m$ passes through a sigmoid and lies in $(0,1]$. In the free variant, $m$ may exceed one, and a cheap accept/reject test on $F_q$ falls back to $m = 1$ for users whose step failed. The test adds one extra objective evaluation per layer. Training is unsupervised on the final energy efficiency, with progressive layer-wise training as in USCA. Per-layer cost is dominated by the $O(n^2)$ interference gradient, the same as USCA's.

A follow-on question, which we flag honestly, is the MIMO covariance case. There, $\log\det$ rate terms are relatively smooth with respect to $-\log\det Q$, where $Q$ is a transmit covariance matrix. The mirror step for this kernel requires a matrix inverse, however, which conflicts with DeepFP's inversion-free motivation. We will treat that case as a separate study rather than assume it transfers.

Baselines

The naive reference points are full power and uniform power allocation. Learning to be green: robust energy efficiency maximization in dynamic MIMO-OFDM systems reports energy-efficiency gains of up to 600% over uniform allocation, which indicates how large the headroom is. Alongside them we run the classical solver, full successive concave approximation. Graph-based Algorithm Unfolding for Energy-aware Power Allocation in Wireless Networks reports it at about 5.23 seconds per allocation. We also run classical Dinkelbach with Euclidean projected gradient, both truncated to $K$ iterations and run to convergence.

The strongest existing competitor is USCA. It targets the same weighted sum energy efficiency problem on the same channel models, uses the same kind of equivariant graph network, and already learns step sizes. It reports a 65% reduction in the suboptimality gap of classical successive concave approximation on small instances where the global optimum is known, and a roughly 58x speedup (about 0.09 s versus 5.23 s). We will reimplement it on identical training channels and user counts, with identical layer count, parameter count, training epochs, and hyperparameter search budget. Our method beats it if it closes more of the gap to the global optimum at equal layers and wall time, and if it degrades less under the distribution shifts described below.

For pure sum-rate (price $q = 0$), the relevant baseline is UWMMSE on the 20-pair random-geometric network. It reports an average sum-rate of 83.21 in 4 layers, against 82.94 for 100-iteration WMMSE, at about 2 ms against 16 ms per sample. We must at least match that sum-rate at 4 layers for the geometry to be credible beyond energy efficiency.

Three ablations isolate the key ingredient. The first is the Euclidean twin: the same graph network and training, but steps taken in Euclidean geometry with projection onto the box. The second is a diagonal-preconditioned Euclidean twin with preconditioner $\mathrm{diag}(p_i^2)$, which matches the Burg curvature locally but carries no global certificate. The third is an untrained Burg iteration at fixed $\eta = 1/L$, which tests whether learning adds anything once the certified step is available.

Why not just…?

Why not just keep learned Euclidean step sizes, as USCA and DeepFP do? They work well on the training distribution. The problem is that the ideal Euclidean step scales with the inverse squared channel-to-noise ratio, so a network trained under one path-loss model must extrapolate its outputs by orders of magnitude under another. DeepFP's own reports of layerwise fluctuation DeepFP: Deep-Unfolded Fractional Programming for MIMO Beamforming show the cost even without a distribution shift. The Euclidean-twin ablation under cross-channel-model testing settles whether this matters in practice. If the Euclidean twin transfers as well as ours, the core claim fails.

Why not just precondition the Euclidean gradient by $p_i^2$? This is the most serious threat, because locally it is the same metric. The differences are that a preconditioned Euclidean step can leave the positive orthant and has no descent certificate, while the mirror step stays positive and satisfies the descent lemma at a known step. If the diagonal-preconditioned ablation matches our method in energy efficiency, monotonicity rate, and transfer, then the contribution shrinks to a clean guarantee rather than a performance gain. We will report it as such.

Why not just truncate classical Dinkelbach, or estimate the Lipschitz constant cheaply by power iteration? The Euclidean constant is cheap to estimate, so the cost of estimating it is not the issue. Its size is, because it forces tiny steps. The UWMMSE and DeepFP results show that truncated classical solvers fall far short: FastFP needed 56 iterations to match DeepFP. Our untrained fixed-step Burg baseline tests directly whether the geometry alone, without learning, already closes most of that gap. If it does, the simpler answer wins and we would say so.

Proposed Experiments

The main benchmark is the 8-to-100-user weighted sum energy efficiency setup of Graph-based Algorithm Unfolding for Energy-aware Power Allocation in Wireless Networks, using both its wideband spatial and Hata-COST231 channel models. We report both global energy efficiency (the certified single-ratio case) and weighted sum energy efficiency. On small instances we compute global optima as that paper did, report the gap to optimum, and compare the gap reduction against USCA's reported 65%. We also evaluate sum-rate on the 20-pair network of Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation, testing on 6,400 channels and transferring to 10 and 30 pairs.

Beyond energy efficiency, we measure the fraction of test instances in which the objective increases at every layer, constraint violation (zero by construction for our method, measured for the baselines), and wall time per allocation on the same CPU.

The decisive check is a distribution-shift matrix. We train on one channel model and a narrow range of transmitter–receiver distances. We then test on the other channel model, on distances shifted to change the gain-to-noise range by 20–40 dB, and on user counts up to 100. The prediction is specific: the Euclidean unfolded baselines lose a growing share of energy efficiency as the gain range widens, while the certified Burg model degrades gracefully and never drops below its untrained $\eta = 1/L$ floor. A second check counts how often the free variant's accept/reject fallback fires, which tells us how far beyond the certificate the learned steps actually want to go. If the fallback rarely fires, the certificate is nearly tight. If it fires constantly, the channel-independent constant is too loose, and the weighted-interferer refinement becomes the priority.