Wireless networks trade spectral efficiency, meaning bits per second per hertz of spectrum, against energy efficiency, meaning bits delivered per joule. Once you count the fixed power drawn by the hardware, this tradeoff has a peak: spend too little power and you deliver almost nothing, spend too much and the extra bits cost far more than they are worth. Learned and iterative controllers increasingly decide how much power each base station and user spends. Those controllers are themselves hardware. They burn joules on computation and on the signaling messages that carry channel measurements to them, and they take time, so by the time a solution is applied the channel may have changed. Almost none of the energy-efficiency literature puts that cost in the denominator.
This note proposes to organize radio resource control the way a journal's editorial board handles manuscripts, and then to measure honestly what that organization costs. Local "associate editor" controllers, one per user cluster, settle most allocation problems themselves. Only the cases they cannot settle are escalated to a central "editor-in-chief" solver. How readily a controller escalates becomes a tunable quantity. The claim is that end-to-end energy efficiency, counting radio, circuit, compute and signaling energy together, is bell-shaped in how often the system escalates. The peak shifts in a predictable way with how fast the channel changes and how expensive computation is.
Core Claim
Define end-to-end energy efficiency for one scheduling interval of length $T$ as
$$\mathrm{EE}_{\text{e2e}}(\tau) = \frac{R(\tau)}{P_{\text{tx}}(\tau) + P_{\text{circ}} + \big(E_{\text{comp}}(\tau) + E_{\text{sig}}(\tau)\big)/T}.$$
Here $\tau$ is the escalation threshold, and $R(\tau)$ is the sum rate actually achieved once the allocation has been applied to the channel as it exists after the controller's decision delay. $P_{\text{tx}}$ is radiated power and $P_{\text{circ}}$ is the hardware circuit power, modeled as in Designing Multi-User MIMO for Energy Efficiency: When is Massive MIMO the Answer?. $E_{\text{comp}}$ and $E_{\text{sig}}$ are the measured joules spent on controller computation and on carrying channel information and decisions between tiers.
A low threshold sends many clusters to the central solver. That gives better interference coordination but costs more compute, more signaling and more delay. A high threshold leaves clusters isolated, which is cheap and fast but blind to leakage between clusters. We predict an interior optimum $\tau^*$. The prediction rests on the same structural argument that produces the bell-shaped spectral-energy curve: one term in the denominator grows with escalation, while the rate in the numerator saturates or degrades through staleness. We further predict that $\tau^*$ moves toward less escalation as user speed rises (shorter coherence time) and as compute energy per operation rises.
The proposal adds two things to existing work. The first is a selective escalation rule, as opposed to a fixed decomposition. The second is an energy ledger that charges the controller for itself.
Relation to Prior Work
Two papers supply the motivation. Björnson et al. showed that digital processing power, which grows with antennas and users, can erase physical-layer gains. That result produces their "transmit power paradox", in which transmit power must rise to justify a large array's static draw. Their model counts the baseband processing that applies a precoder, not the optimizer that chooses one. We extend the same accounting to the controller. AI-Ready Energy Modelling for Next Generation RAN reports a 14.8% energy-efficiency gain at a 3.9% spectral-efficiency cost from tuning a subset of base stations in the AIMM simulator. It also observes that cutting power by 6 dBm can raise throughput by removing interference. As described in the file, it does not charge the optimizer's own energy.
Four papers supply the closest methods, and all of them use fixed delegation. Deep Reinforcement Learning-Based Block Coordinate Descent for Downlink Weighted Sum-Rate Maximization on AI-Native Wireless Networks has a reinforcement-learning agent predict initializations for a block coordinate descent solver, an iterative method that updates one group of variables at a time. It reports beating both pure learning and pure solver baselines from 4 to 128 users, with no circuit or controller energy modeled. Meta Federated Reinforcement Learning for Distributed Resource Allocation pushes decisions to user devices so that raw channel state never has to travel to the base station. Collaborative Multi-BS Power Management for Dense Radio Access Network Using Deep Reinforcement Learning centralizes control in a cloud Q-network. Mechanism design for energy efficiency in wireless networks runs a flat price-broadcast loop between one base station and its users that converges in about 20 iterations. Each of these picks one point on the centralization axis. None makes the degree of centralization a per-interval decision, and none charges for it.
The organizational side supplies the structure and also constrains it. The publication information of IEEE Transactions on Information Theory and IEEE Transactions on Cybernetics documents the three-tier editor-in-chief, associate editor and referee arrangement. In that arrangement, filtering lower in the hierarchy keeps raw volume away from the top. Embodied LLM Agents Learn to Cooperate in Organized Teams found that designated leadership cut task time by 9.76% for three-agent teams and kept communication cost linear in team size up to nine agents. It also found that dynamic leader election caused large transient spikes in communication. That last result is a negative finding, and we take it seriously. It is the reason our tiers and cluster heads are fixed within each run, and only the escalation decision adapts.
Research Program
The system has three tiers. At the bottom, users report quantized channel measurements to their cluster head. In the middle, each cluster head runs a cheap local energy-efficiency solver over its own users. This is block coordinate descent wrapped in Dinkelbach's method, the standard fractional-programming procedure used by Energy-Efficient Power Allocation in Uplink mmWave Massive MIMO with NOMA, and it produces a locally optimal allocation. Clusters are formed by spatial channel correlation, following the same paper.
Each cluster head also computes a leakage score, an estimate of how much interference its allocation would push into neighboring clusters and absorb from them. If the score exceeds $\tau$, the cluster escalates. The central tier then jointly re-solves only the escalated clusters and their interfering neighbors, warm-started from the local solutions. This is the "desk reject versus send to the editor" decision, and it is the one ingredient we treat as essential.
Every computation is instrumented. We count operations, and we measure joules on a fixed reference processor for each solver call. Signaling bits per tier are multiplied by a fronthaul energy-per-bit figure, and the sensitivity to that figure is reported as a sweep. Decision delay is converted into staleness by applying each allocation to a channel that has evolved under a temporally correlated fading model for the measured latency.
The analysis has three parts. First, we sweep $\tau$ to trace $\mathrm{EE}_{\text{e2e}}$ against escalation rate. Second, we derive a first-order condition for $\tau^*$: escalation is worth it when the marginal rate recovered per escalated cluster, discounted for staleness, exceeds the marginal joules spent. Third, we test whether this condition predicts the empirical optimum across speeds and compute-cost settings.
Baselines
The naive reference is fixed-power operation with no optimization: every base station transmits at nominal power, combined with the base-station sleep heuristic. This is the floor that Chang et al. beat. They report 10% higher energy efficiency than sleep schemes and 5% higher than tabular Q-learning. Those figures come from their simulator, not ours, so we re-implement the heuristic rather than borrow the number.
The strongest existing method on this question is the learned-initialization block coordinate descent of Chen, Tan and Poor. We run it as a single central controller over all users, re-targeted from weighted sum rate to energy efficiency with the same Dinkelbach outer loop we use. It is the strongest competitor because it already combines learning with structured optimization and is reported to beat pure learning, graph network and pure solver baselines up to 128 users. Its reported numbers are qualitative in the file: no reported sum-rate value was found on Lacuna for this method. We therefore reproduce it on the same channel draws.
The second strong competitor is a fully centralized Dinkelbach and block coordinate descent solve with no learning, run to convergence. It is the quality ceiling when delay is ignored. The third is the flat pricing mechanism of Chorppath and Alpcan, which represents cheap one-tier coordination.
Fairness rules are fixed in advance. All methods see identical channel realizations, user placements and circuit model, and all are charged under the same energy ledger. Learned components get the same training budget in environment steps and the same hyperparameter-search trials. For centralized methods, delay is set by their measured computation plus collection of channel state from every user.
There are two ablations that isolate the proposal's key ingredient. The first is the hierarchy with escalation disabled, meaning clusters only, so that $\tau=\infty$. The second is the hierarchy with every cluster always escalated, so that $\tau=0$.
The proposal beats the baselines only if, at its tuned $\tau^*$, it achieves higher $\mathrm{EE}_{\text{e2e}}$ than the learned-initialization method and the centralized solve. It must also stay within a pre-registered spectral-efficiency loss of 4%, chosen to bracket the 3.9% tradeoff Sthankiya et al. accepted, and it must do this across at least two of the three mobility settings. Beating the clusters-only ablation is also required. If it is not, escalation adds nothing and the editorial structure is ornamental.
Why not just…?
Why not just add a controller-power constant to Björnson's circuit model and keep a centralized solver? This is the most serious threat. If controller cost is roughly constant per interval, a flat solver that knows about that constant would choose the same radio operating point, and hierarchy would add nothing. Our answer is that controller cost is not constant. It scales with how many users are solved jointly and how much channel state is collected, and the delay cost depends on the mobility regime. The experiment settles it directly. We give the centralized baseline the exact measured controller energy as a known penalty in its objective. If it then matches the hierarchy across all speeds, we report that the energy ledger, not the escalation structure, is the contribution.
Why not just fix the clusters and never escalate, as in the NOMA clustering of Zeng et al.? Static clustering is cheap and may be enough when clusters are well separated. The $\tau=\infty$ ablation measures this directly. Our prediction is that it loses in dense deployments where inter-cluster leakage dominates, which is the regime where Sthankiya et al. saw interference reversals from a 6 dBm change. If it does not lose there, we drop the escalation claim.
Why not just let every node optimize itself with no hierarchy, as in flat game-theoretic or federated schemes? Meshkati et al., in A Game-Theoretic Approach to Energy-Efficient Modulation in CDMA Networks with Delay QoS Constraints, analyze selfish users under a utility in bits per joule and show that tightening delay constraints pushes them toward higher-order modulation and worse efficiency. Chorppath and Alpcan show that selfish price-takers need a coordinator's prices to reach the welfare optimum. The federated approach of Ji et al. avoids moving channel state to the center, but it is evaluated only on orthogonal channel access. We include Chorppath-Alpcan pricing as a baseline rather than assume it loses. Under very high mobility, cheap flat coordination may well win, and if it does the bell-curve claim predicts it.
Proposed Experiments
The primary testbed is the AIMM simulator extension of Sthankiya et al. It models a 19-site hexagonal macrocell layout with users at 1256 per km² under 3D urban-macro non-line-of-sight propagation, and its sub-3-second runs allow wide sweeps. Circuit power follows the 3GPP and EARTH-derived model of Björnson et al. As a sanity check, our centralized single-cell solve should recover their reported optimum of about 165 antennas serving 85 users under their parameters. Scaling experiments use the synthetic downlink generator of Chen, Tan and Poor from 4 to 128 users, and then extend to 512 users to probe where central solving breaks. Mobility transfer uses the 3GPP urban micro, indoor, urban macro and rural macro scenarios as in Ji et al. Results are reported in bits per joule under both the radio-only ledger and the end-to-end ledger, alongside spectral efficiency, escalation rate and decision latency.
Three checks decide the outcome. The first is whether $\mathrm{EE}_{\text{e2e}}$ against $\tau$ has an interior maximum that the radio-only ledger does not show. If both ledgers peak in the same place, controller accounting is irrelevant. The second is whether $\tau^*$ shifts toward less escalation, as predicted, when user speed or compute energy per operation is increased. The third is the head-to-head against the cost-aware centralized solver described above. Rankings that change between the radio-only and end-to-end ledgers would by themselves be a reportable finding. They would mean current energy-efficiency comparisons of learned controllers are measuring the wrong thing.
References
- Designing multi-user MIMO for energy efficiency: When is massive MIMO the answer?
- AI-Ready Energy Modelling for Next Generation RAN
- Deep Reinforcement Learning-Based Block Coordinate Descent for Downlink Weighted Sum-Rate Maximization on AI-Native Wireless Networks
- Meta Federated Reinforcement Learning for Distributed Resource Allocation
- Collaborative Multi-BS Power Management for Dense Radio Access Network Using Deep Reinforcement Learning
- Mechanism design for energy efficiency in wireless networks
- IEEE Transactions on Information Theory publication information
- IEEE Transactions on Cybernetics publication information
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Energy-Efficient Power Allocation in Uplink mmWave Massive MIMO With NOMA
- A Game-Theoretic Approach to Energy-Efficient Modulation in CDMA Networks with Delay QoS Constraints