When a transmitter turns up its power to protect a confidential message, the eavesdropper hears more too. With perfect channel knowledge, precoding can steer energy away from eavesdroppers and keep the secure rate growing. With stale channel knowledge it cannot. The part of the signal that the stale estimate fails to null leaks to the eavesdropper in proportion to transmit power, so the secrecy rate (the rate Bob can decode minus the rate Eve can) first flattens and then collapses. In fast-moving networks the channel keeps changing after it was last measured, an effect called channel aging. This means the safe power level is not a fixed number: it depends on how old and how uncertain the channel estimate is. Distributed power control makes this worse. Each transmitter adjusts its own power in rounds of a game, and the convergence guarantees for such games assume the channels hold still while the rounds run.
This note proposes one object that addresses both problems: a pessimistic potential, meaning a lower confidence bound on the network's secrecy sum-rate given the channel predictor's uncertainty. Every transmitter's utility is built as its marginal contribution to that bound. The claim is that this choice does three things at once. It keeps the game an exact potential game, so distributed best-response updates still terminate. It produces a power back-off that tracks channel age automatically, generalizing the fixed power caps used against high-SNR secrecy collapse. And it pulls the dynamics toward power levels where the secrecy objective is least sensitive to channel drift, which is where game iterations stop oscillating. The quantity that causes secrecy collapse, namely how strongly the secrecy rate reacts to channel error, is the same quantity that destabilizes the game. A single penalty on it should fix both.
Background and the gap
Three bodies of work set up the problem without solving it. The first is the large-system analysis of linear precoding with confidential messages (Geraci et al.). It shows that when users outnumber antennas and act as cooperating eavesdroppers, secrecy collapses at high SNR. It fixes this with Regularized Channel Inversion with Power Reduction (RCI-PR), a centralized precoder with a fixed finite power cap. It also shows that quantized channel feedback must grow linearly with SNR in dB to avoid collapse. The RIS study (Zhang et al.) shows that regularized zero-forcing (RZF) escapes the secrecy floor that maximum-ratio transmission hits, and that the optimal regularization scales as 1/SNR. Separately, Hyadi et al. show that imperfect channel knowledge imposes a secrecy ceiling that added power cannot break. These works motivate the proposal. All of them assume block fading, where the channel is constant within a block. None makes the power cap depend on how old the estimate is, and none is distributed.
The second body of work is potential games: games in which every player's utility change equals the change in one shared scalar function, so better-response updates can only climb that function and must stop at a pure Nash equilibrium. These supply the convergence machinery. Della Penda et al. reach a stable equilibrium in 4 to 5 rounds for D2D channel allocation, within 15% of branch-and-bound. They explicitly leave high-speed mobility as future work. Chen and Huang report convergence in under 20 iterations with a 7 to 8% price of anarchy, assuming static channel availability. He et al. embed a potential game in per-slot Lyapunov control but treat channels as quasi-static within each 1-second slot. These works constrain the design: whatever is done must keep the exact-potential identity, or the termination proof is lost.
The third body of work is pessimistic value estimation from offline RL, which supplies the uncertainty penalty. It subtracts a confidence width from a nominal estimate so that a maximizer cannot exploit poorly supported regions. Yin and Wang show that variance-aware penalties need coverage only along the optimal policy. Yin et al. use a gradient-times-inverse-Fisher penalty for nonlinear models, and we borrow that form directly. The closest multi-agent work is SOLAR (Zhang et al.). It throttles potential-based reward shaping when a world-model ensemble disagrees, and it is evaluated on StarCraft. In SOLAR the "potential" is a shaping signal inside actor-critic training, not a game potential with a termination guarantee, and there is no channel or secrecy model.
The obvious combination is to make each agent pessimistic about its own utility. That combination is wrong. A penalty $\Gamma_i$ that depends on the other agents' actions breaks the identity $\Delta U_i = \Delta\Phi$, and the finite improvement property goes with it. The contribution here is to put the pessimism in the potential and derive the utilities from it.
Core Claim
Consider $N$ transmitters, such as small-cell base stations or D2D pairs. Each chooses a power level $p_i$ from a finite set and serves a legitimate user in the presence of internal or external eavesdroppers. Each transmitter uses a channel predictor that outputs a mean channel $\hat h$ and a covariance $\Sigma_\tau$ for the channel at the moment of transmission, where $\tau$ is the age of the last pilot. Let $\hat C_s(p)$ be the secrecy sum-rate evaluated at $\hat h$, and let $g(p)$ be its gradient with respect to the channel. The pessimistic potential is
$$\Phi_{\text{pess}}(p) = \hat C_s(p) - \beta \sqrt{g(p)^\top \Sigma_\tau\, g(p)},$$
with $\beta \ge 0$ the pessimism level. This is a delta-method lower confidence bound, the same gradient-covariance form as PFQL's penalty, with channel uncertainty in place of parameter uncertainty. Each agent's utility is its marginal contribution:
$$U_i(p) = \Phi_{\text{pess}}(p_i, p_{-i}) - \Phi_{\text{pess}}(p_i^{0}, p_{-i}),$$
where $p_i^0$ is a fixed reference level such as zero power. This makes the game an exact potential game with potential $\Phi_{\text{pess}}$ by construction.
The physics predicts what the penalty does. When the channel error, rather than receiver noise, dominates the leakage at the eavesdropper, the eavesdropper's received leakage scales like $p\,\mathrm{tr}(\Sigma_\tau)$. The penalty then grows with power while the nominal secrecy rate saturates near $\log_2(\alpha/\beta_{\text{eve}})$, where $\alpha$ and $\beta_{\text{eve}}$ are Bob's and Eve's channel gains. The maximizer of $\Phi_{\text{pess}}$ is therefore a finite power that falls as the channel ages, roughly holding $p\,\mathrm{tr}(\Sigma_\tau)$ constant. This is RCI-PR's cap made age-dependent and decentralized.
The second prediction concerns drift. The potential's sensitivity to channel change is largest at high power, where leakage scales with $p$. Pessimism therefore steers equilibria toward a region where aging moves $\Phi$ least between rounds. That is the condition under which the dynamics can keep up.
Research Program
The analytical part has two pieces. The first is a tracking theorem for $\eta$-better-response dynamics on a drifting potential, where an agent switches only if its gain exceeds a threshold $\eta$. Suppose aging changes $\Phi_{\text{pess}}$ by at most $\delta$ per round, with $\delta$ bounded by the potential's Lipschitz constant in $h$ times the per-round channel change (itself a function of Doppler). If $\eta > \delta$, the number of accepted updates between equilibria is bounded, and the dynamics remain within an $\eta$-equilibrium set. This gives the maximum Doppler spread that preserves convergence, which the file identifies as unformalized. We will show that the Lipschitz constant restricted to the equilibria of $\Phi_{\text{pess}}$ is smaller than for the certainty-equivalent potential, so the admissible Doppler grows with $\beta$. The second piece is a closed-form characterization of the age-dependent optimal power in the single-cell, error-dominated regime, to be checked against simulation.
The systems part builds the uncertainty source. The predictor is a small ensemble of sequence models of the hybrid state-space type in MambaCSP, trained offline on logged channel traces. Ensemble disagreement plus residual variance at each horizon $\tau$ gives $\Sigma_\tau$. The offline-RL lens matters here. CeBed reports that deep estimators "saturate and can even degrade below LS at high SNRs when tested on unseen SNR ranges." High SNR is exactly where a secrecy optimizer pushes power, so an overconfident predictor would trigger collapse there. A pessimistic potential is only safe if $\Sigma_\tau$ widens out of distribution, and we test that directly.
Baselines
The naive reference is full power with certainty-equivalent best response: the potential game runs on $\hat C_s$ alone, with predicted CSI treated as truth. This is what Della Penda et al.-style formulations do once a secrecy objective is plugged in. It isolates what pessimism adds. Under static channels it should match our method, and under aging it should collapse at high power.
The strongest existing method on the secrecy side is centralized RCI-PR (Geraci et al.), with its power cap tuned on the same training traces and given the same predicted CSI. It is the strongest competitor because it already prevents high-SNR collapse. The file reports this only qualitatively: at load $\beta = 1.2$ and $\rho = 25$ dB, RCI-PR "prevents the total secrecy rate from falling to zero". No reported result was found on Lacuna for RCI-PR, or for any method, under channel aging or Doppler. Our method beats it if it achieves higher realized ergodic secrecy sum-rate at velocities of 60 km/h and above, with equal or lower secrecy-outage probability, while being distributed. We also report a centralized oracle with perfect current CSI as an upper bound.
The ablation that isolates the key ingredient is per-agent pessimism, where each agent penalizes its own utility independently. This breaks the potential identity, and we measure whether it produces strategy cycles. A second ablation replaces the learned $\Sigma_\tau$ with the analytic Gauss–Markov/Jakes aging covariance, which scales with $1-J_0^2(2\pi f_D\tau)$, where $J_0$ is the zeroth-order Bessel function and $f_D$ is the Doppler frequency. A third is a worst-case robust potential with an ellipsoidal error bound. All methods share the same channel traces, the same power grid, the same signaling budget per round, and the same validation budget for tuning $\beta$, the power cap, or the robust radius.
Why not just…?
Why not just send pilots more often? Lu et al. show that under peak-power limits, longer training directly takes symbols from payload. At high mobility, training would eat the coherence block. We include a pilot-density sweep in the evaluation: the certainty-equivalent game run with doubled and quadrupled pilot rates, scored on net secure throughput after training overhead. If extra pilots close the gap at 100 km/h, the proposal's value is confined to pilot-starved regimes, and we will say so.
Why not just apply a fixed, analytically derived age-dependent power back-off? This is the real threat. Under Rayleigh fading with Jakes aging, the error covariance is known in closed form, so a lookup-table cap $p^*(\tau)$ might match everything learned. The Jakes ablation settles it. On stationary synthetic channels we expect it to tie, and that would be fine, because the potential-game construction still contributes convergence under drift. Learned uncertainty only earns its place if it wins on non-stationary QuaDRiGa traces with cluster birth-death, where the Jakes model is misspecified. If it does not win there, we drop the learned predictor from the claim.
Why not just run offline multi-agent RL with uncertainty-throttled shaping, as SOLAR does? SOLAR (Zhang et al.) offers no termination guarantee for the per-slot power game. It adds roughly 1.2× training cost over model-free baselines on StarCraft, and its policy explores configurations whose leakage is irreversible. We will nonetheless train a SOLAR-style offline learner on the same logs as a comparison. If it matches our method's secrecy rate and oscillation rate, the structural argument is weaker than claimed.
Why not just centralize? Centralized secure precoding needs all users' CSI at one point, which ages further in transit. The centralized RCI-PR baseline, with backhaul delay added to its CSI age, measures this cost directly.
Proposed Experiments
Channels come from QuaDRiGa in the 3GPP UMa NLOS configuration, at 10 to 100 km/h, under both TDD and FDD, following the MambaCSP protocol of 8,000/2,000/10,000 train/validation/test traces. The network is a 7-cell layout borrowed from the D2D benchmark of Della Penda et al., with $M=24$ base-station antennas as in Zhang et al.. Loads span underloaded, $\beta=1.2$, and near-$2$ cases following Geraci et al.. Predictor calibration across SNR shift is checked on CeBed's SNR-split scenarios.
Metrics are computed on the true channel at transmission time: realized ergodic secrecy sum-rate, secrecy-outage probability, the fraction of slots with zero secrecy, rounds to an $\eta$-equilibrium, strategy flips per round, and mean transmit power.
Four checks decide the outcome. The first is the secrecy-versus-power curve at each velocity: the certainty-equivalent game should collapse at high power while the pessimistic one should not. The fitted relation between the chosen power and $\mathrm{tr}(\Sigma_\tau)$ should be approximately inverse. The second compares measured oscillation onset against the predicted Doppler threshold from the tracking theorem, for several values of $\beta$; the admissible velocity should rise with $\beta$. The third is per-agent pessimism, which should show cycles that the potential-level version does not. The fourth is the learned-versus-Jakes comparison on non-stationary traces. The project counts as failed if pessimism lowers secrecy relative to tuned RCI-PR at every velocity, or if the predicted Doppler threshold misses the measured onset by more than a factor of two.
References
- Large System Analysis of Linear Precoding in MISO Broadcast Channels with Confidential Messages
- RIS-Aided Secure Communications With Regularized Zero-Forcing Precoding
- Securing Multi-User Broadcast Wiretap Channels with Finite CSI Feedback
- Distributed Channel Allocation for D2D-Enabled 5G Networks Using Potential Games
- Database-Assisted Distributed Spectrum Sharing
- An Online Joint Optimization Approach for QoE Maximization in UAV-Enabled Mobile Edge Computing
- Towards Instance-Optimal Offline Reinforcement Learning with Pessimism
- Offline Reinforcement Learning with Differentiable Function Approximation is Provably Efficient
- SOLAR for Offline MARL: Plateau-Triggered Potential Shaping under World-Model Uncertainty
- MambaCSP: Hybrid-Attention State Space Models for Hardware-Efficient Channel State Prediction
- CeBed: A Benchmark for Deep Data-Driven OFDM Channel Estimation
- Joint optimization of power allocation and training duration for uplink multiuser MIMO communications