Uncertainty-Aware Power Control: Guaranteeing Every User's Rate When Channel Estimates Are Wrong

Wireless base stations must decide, every few milliseconds, how much power each transmitter uses so that users interfering with one another all get usable data rates. Neural networks now make these decisions in under a millisecond, but they are trained on channel measurements as if those measurements were exact. Real measurements are stale or noisy, and some links are far less reliable than others. A fast-moving user's estimate goes out of date quickly, and a link with weak pilot signals is estimated poorly. When an estimate is wrong, a network trained only for average throughput quietly leaves some users below their minimum rate.

We propose a graph-neural power allocator that receives, alongside each channel estimate, a per-link number saying how wrong that estimate could be. It is trained against an adversary that pushes every link's channel to its worst value inside that error radius, with a per-user price enforcing each rate floor. Because the radius is an input rather than a fixed training setting, one network is cautious on unreliable links and aggressive on reliable ones. It still runs in a single forward pass, and when the radii shrink to zero it should behave like an ordinary allocator.

Core Claim

Robust optimization methods choose an "uncertainty set," a bounded region of possible true channels around the estimate, and optimize for the worst case inside it. In nearly all of them the size of that region is one static hyperparameter. Set it too small and the policy is brittle; set it too large and it gives up throughput everywhere. In wireless systems the right size is neither unknown nor uniform. It follows from quantities the base station already observes, chiefly pilot signal-to-noise ratio and the age of the estimate relative to Doppler spread. Our claim is that conditioning a permutation-equivariant graph allocator on these per-link radii, and training it as a min-max game with dual prices on rate floors, reduces worst-case rate-floor violations sharply relative to standard training. It should do this at a small cost in average sum-rate and with no added inference latency. It should also beat both random-noise augmentation and fixed-radius adversarial training on worst-case and structured-shift tests.

Concretely, let $H$ be the estimated channel matrix, $\rho_{ij}$ the error radius on the link from transmitter $j$ to receiver $i$, and $p_\theta(H,\rho)$ the power vector output by a graph neural network with weights $\theta$. Let $R_i(p; G)$ be user $i$'s rate under power $p$ and true channel $G$, and let $r_i$ be that user's rate floor. Training solves

$$\max_\theta \; \mathbb{E}_{(H,\rho)} \Big[ \min_{|\Delta_{ij}|\le \rho_{ij}} \Big( \sum_i R_i(p_\theta; H+\Delta) + \sum_i \lambda_i \big(R_i(p_\theta; H+\Delta) - r_i\big) \Big) \Big],$$

where $\Delta$ is the adversary's perturbation and $\lambda_i \ge 0$ is a per-user dual multiplier, a price on violating user $i$'s floor, updated by dual ascent. The uncertainty set is a product of per-edge intervals, and relabeling users relabels the set in the same way, so the optimal min-max policy is itself permutation-equivariant. A graph network can therefore represent it without breaking the symmetry that makes wireless graph networks scale.

Relation to Prior Work

The allocator side builds on graph networks for interference management. IGCNet (A Graph Neural Network Approach for Scalable Wireless Power Control) represents transmitter-receiver pairs as nodes and interference links as edges. It reached 102.4% of the sum-rate of the classical iterative solver WMMSE (weighted minimum mean-square error) on 30 users in 0.48 ms versus 31.4 ms. It also kept over 92% of peak performance with 70% of channel entries missing. Graph neural networks for scalable radio resource management ties message passing to distributed optimization and scales to 1000 users. Both supply our backbone and our nominal reference. Neither treats estimation error as anything but random input damage, and neither enforces per-user rate floors.

Unfolded WMMSE (Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks) turns WMMSE iterations into graph-network layers. It tolerated up to 20% Gaussian coefficient distortion, but it was never trained or tested against worst-case perturbations, and its authors list quality-of-service constraints as future work. That gap is exactly the one we target. HMGAT (GNN-Enabled Robust Hybrid Beamforming with Score-Based CSI Generation and Denoising) takes the opposite route to robustness: it cleans noisy estimates with an iterative score-based denoiser before beamforming. That approach improves the average case, offers no worst-case guarantee, and adds sampling latency. We test it as a competitor rather than reuse it.

The method side borrows from adversarial and distributionally robust learning. Adversarial Deep Learning for Online Resource Allocation trains an algorithm network against an adversary network and uses Lagrangian dual multipliers as per-step marginal prices to avoid high-variance credit assignment. We reuse the dual-price device. That work, however, operates on abstract budget sequences with dense networks and no channels, topology, or interference. Learning to Solve Routing Problems via Distributionally Robust Optimization motivates the proposal by showing that graph solvers trained by ordinary risk minimization collapse on structured distributions. Its POMO baseline had a 17.57% average optimality gap on TSPLib/CVRPLib, which its robust variant cut to 8.08%. Its robustness, though, needs hand-specified discrete groups, while ours is a continuous, observation-driven set.

Two theoretical results constrain the design. Adversarial generalization of unfolding (model-based) networks bounds robustness of unfolded networks only under single-step FGSM (fast gradient sign method) perturbations and notes that this under-approximates multi-step attacks. We therefore train and evaluate with multi-step projected gradient attacks, never one-step ones. The Curious Price of Distributional Robustness in Reinforcement Learning with a Generative Model shows that the geometry of the uncertainty set changes how hard robust learning is. We choose bounded per-edge intervals, which are physically interpretable and closer in spirit to the favorable total-variation case than to divergence-based sets. We do not claim that this theory transfers. Finally, Distributionally Robust Power Policies for Wireless Systems Under Power Fluctuation Risk and Distributionally Robust Resource Allocation with Trust-aided Parametric Information Fusion formulate robust allocation without real-time neural execution or graph structure. Deep Reinforcement Learning for Robust Beamforming in IRS-assisted Wireless Communications handles uncertainty with heuristic reward penalties. None conditions on per-link reliability.

Research Program

The first stage is single-antenna interference-channel power control in the IGCNet setting with Rayleigh fading. Each sample draws user positions, true channels, a pilot SNR and a CSI age per link. The estimate is the true channel corrupted by estimation noise plus temporal decorrelation under a Jakes model, in which correlation decays as a Bessel function of Doppler times age. The radius $\rho_{ij}$ is set to a fixed multiple of the resulting error standard deviation, a multiple chosen so the interval covers a target fraction of errors. The allocator is an IGCNet-style network that takes $\rho_{ij}$ as an additional edge feature. During training, the inner minimum is approximated by multi-step projected gradient attacks on edge values, with random restarts, while $\lambda_i$ follows dual ascent on the observed floor violations. All of this cost is paid offline; inference is one forward pass.

The second stage moves to multi-antenna beamforming by applying the same objective to the unfolded WMMSE architecture on Rayleigh and Rician channels. This tests whether radius conditioning survives inside an unfolded solver without reintroducing per-layer worst-case subproblems, which the research file identifies as the main threat to unfolding's speed. The third stage uses DeepMIMO ray-tracing channels to construct structured shifts, namely line-of-sight blockages on a subset of links. These test whether the per-edge set, trained on statistical errors, transfers to physically plausible correlated failures it never saw.

Analysis includes an empirical check of permutation equivariance under the adversary and a measurement of how the learned policy's power margin scales with $\rho$. The latter tells us whether the network actually reads the radius or ignores it.

Baselines

The naive reference is IGCNet trained by ordinary sum-rate maximization on estimated channels with no knowledge of errors. This is what current practice would deploy. Its reported nominal result, 102.4% of WMMSE sum-rate at 0.48 ms on 30 users, is the bar for clean-channel performance; we must stay close to it as $\rho \to 0$.

The strongest simple competitor is the same network trained by ordinary risk minimization on realistically noisy estimates, given the radius as an input, with rate floors enforced by the same dual prices. This model sees exactly the same information as ours and differs only in optimizing the expected rather than the worst-case rate. It is therefore the cleanest test of whether the adversary matters.

A second strong competitor is adversarial training with a single global radius, the standard practice in the robust literature. A third is the Group DRO procedure of Jiang et al., adapted to groups defined by mobility class.

For latency-insensitive quality reference we solve a classical worst-case robust power control problem per sample with an iterative solver. We report its runtime against the 31.4 ms WMMSE and 1,305 ms MIMO WMMSE figures in the file. On DeepMIMO we compare with HMGAT plus its denoising front end, for which the file reports 37.61 bps/Hz sum-rate on the Los Angeles 16-user, 32-antenna scenario. We found no reported result on Lacuna for worst-case rate-floor satisfaction by any wireless graph allocator, so the robust metric baselines are all ones we run ourselves.

Fairness rules apply to all learned methods: identical channel generators, identical architecture and parameter count, the same number of training gradient steps with adversary cost counted, and equal hyperparameter search budgets. Two ablations isolate the key ingredients. The first removes the radius input while keeping per-edge adversarial training. The second replaces dual prices with a fixed penalty. Beating the baselines means at least halving the worst-case floor-violation rate relative to the noise-augmented competitor at matched average sum-rate, with inference within a few percent of IGCNet's latency.

Why not just…?

Why not just train on noisy channels with the radius as input? This is the real threat. If true errors are Gaussian and the floors are soft, an expected-rate model trained this way may approach the right behavior. The difference shows up in the tail and in structured shifts. Expected-rate training tolerates rare large errors on a few users, while the min-max objective does not. We will report violation rates both under realistic random errors and under worst-case attacks. If the noise-augmented model matches ours on realistic errors, we will say so, and the claim narrows to worst-case and blockage robustness.

Why not just add a safety margin to each rate floor? Inflating $r_i$ by a constant, or by a multiple of $\rho$, is cheap and needs no adversary. It is applied per user, though, while interference errors couple users: one unreliable cross-link threatens another user's floor. We include margin-scaled-by-radius as an extra baseline, sweeping the margin to trace its full trade-off curve against ours.

Why not just use one global uncertainty radius or Group DRO? These are the standard robust choices. The research file identifies the static radius as the central weakness of distributionally robust methods, since it is either brittle or pessimistic. The global-radius and Group DRO baselines, together with the no-radius ablation, measure directly how much throughput the per-link conditioning recovers on reliable links.

Why not just run a classical robust solver, or denoise first? Iterative solvers are already too slow without robustness: 1,305 ms versus 54 ms for unfolded WMMSE. Worst-case formulations only add cost. Denoising, as in HMGAT, reduces average error but cannot bound the remaining error, and its iterative sampling costs time. We measure both on accuracy and latency rather than assume.

Decisive Checks

The proposal succeeds if four things hold on the interference-channel and MIMO benchmarks. Worst-case rate-floor violation falls substantially below that of the noise-augmented and global-radius models at matched sum-rate. Sum-rate on clean channels stays close to IGCNet's level as radii shrink. Learned power margins grow monotonically with $\rho$. Latency stays sub-millisecond. On DeepMIMO blockage shifts, the radius-conditioned model should keep more users above their floors than both HMGAT and the expected-rate model.

The proposal fails informatively in two cases. If the noise-augmented competitor matches it everywhere, the adversary contributes nothing beyond the radius input. If performance on blockage shifts collapses, independent per-edge intervals are the wrong shape for physical failures, and correlated uncertainty sets become the next problem to solve.