Shared Inducing Points for Decentralized Radio Maps and Cell-Free Power Control

In a cell-free network, many small access points cooperate to serve users. Deciding which access points serve which user, and at what transmit power, depends on how strongly each access point's signal reaches each location. Today those decisions are either computed as if the channels were perfectly known, or computed at a central processor that first gathers every measurement over a capacity-limited link. We propose a different approach. The access points jointly learn a probabilistic map of signal strength over the service area. They do this by agreeing on a small shared summary: the values of a Gaussian process at a few hundred "inducing points," which are virtual locations that stand in for the full dataset. They then feed both the map's prediction and its uncertainty into a decentralized power-control solver.

The central idea is that the inducing-point summary can serve as the thing the access points agree on, replacing raw channel matrices. The summary has a fixed size set by the number of inducing points, whatever the number of measurements or antennas. We use two tools for agreement. The first is exact averaging of statistics, wherever the model is simple enough that averaging gives the right answer. The second is the alternating direction method of multipliers (ADMM, a method that splits one optimization problem across agents and adds penalty prices that pull their local copies into agreement). We use ADMM only where averaging is not enough: when measurements are cut off by receiver sensitivity, when fitting kernel settings, and when solving the coupled power-control decision.

Core Claim

Consider a Gaussian process with a Gaussian likelihood and fixed kernel settings. The optimal distribution over the inducing values is determined by sums of per-agent statistics. So one round of averaging among neighbors (gossip), or one upload to a central processor, already gives the exact centralized answer. No iterative optimizer is needed for that part.

Real radio measurements break this in a specific way. A user device reports nothing useful when the signal falls below its sensitivity threshold. Reports are also quantized. The measurement model therefore becomes censored: for a missing report we only learn that the true signal was below the threshold. Under a censored model, the local evidence no longer reduces to additive statistics. Treating censored reports as exact values at the threshold biases the map upward exactly in the weak-coverage regions that decide outage.

Our claim has two parts. First, a consensus-ADMM formulation over the inducing distribution's natural parameters, the kernel hyperparameters, and the inducing locations should recover the centralized censored-likelihood posterior. It should do so while exchanging only messages whose size depends on the number of inducing points $M$, never on the number of samples. Second, a power-control and access-point-selection solver that uses the posterior variance should cut true user outage relative to a solver that plugs in the posterior mean. It should do so at the same total communication volume.

Concretely, access point $p$ holds reports $y_{pj}$ at locations $x_{pj}$ and solves its share of

$$ \min_{\{\eta_p,\theta_p\},\,\eta,\theta}\ \sum_{p=1}^{P}\Big(-\sum_j \mathbb{E}_{q_{\eta_p}(f)}\big[\log p(y_{pj}\mid f(x_{pj}))\big] + \tfrac{1}{P}\,\mathrm{KL}\big(q_{\eta_p}(u)\,\|\,p_{\theta_p}(u)\big)\Big)\quad \text{s.t.}\ \eta_p=\eta,\ \theta_p=\theta . $$

Here $\eta_p$ is agent $p$'s copy of the natural parameters of $q(u)$, the Gaussian over inducing values $u$. The symbol $\theta_p$ is its copy of the kernel hyperparameters and inducing locations, and $\eta,\theta$ are the agreed values. Each likelihood term $p(y\mid f)$ is Gaussian for an observed report and a Gaussian-CDF probit term $\Phi((s-f)/\sigma)$ for a report censored below sensitivity $s$. This is the per-point variational bound of Gaussian Processes for Big Data, with the non-conjugate extension of Scalable Variational Gaussian Process Classification. The global KL penalty is split evenly across the $P$ agents.

How This Sits Against Prior Work

Two papers motivate the gap. Decentralized Beamforming Design for Intelligent Reflecting Surface-enhanced Cell-free Networks solves joint base-station beamforming and surface phase design with incremental ADMM that passes a token around a ring of base stations. It reaches centralized sum-rate in about 30 iterations, but it assumes perfect channel knowledge. Its ring topology also makes latency grow with network size. We keep its lesson that ADMM can coordinate coupled physical decisions without a central processor. We change two things: we replace known channels with a learned posterior, and we replace the ring with parallel neighbor exchanges. SIM-Enabled Hybrid Digital-Wave Beamforming for Fronthaul-Constrained Cell-Free Massive MIMO Systems treats the link to the central processor (the fronthaul) as a binding resource. That is why we count every transmitted bit.

Three papers supply methods or constrain the design.

  • Kernel consensus without data sharing. COKE: Communication-Censored Decentralized Kernel Learning observes that kernel models tied to local samples have no shared parameterization. It fixes this with random Fourier features (a fixed random basis of sinusoids) and adds a rule that skips a transmission when the local update is small, saving about 50% of messages. COKE holds hyperparameters fixed and outputs no posterior variance. We borrow its skipping rule and replace its random basis with inducing points, which have learnable locations and carry uncertainty.
  • Parallel ADMM schedules. GADMM splits workers into two alternating groups and reports large communication savings for convex losses. It explicitly leaves non-convex objectives open. Our hyperparameter and inducing-location updates are non-convex, so its guarantees do not carry over. We treat that as a risk to measure, not a solved problem.
  • Tuning the penalty weight. Adaptive Consensus ADMM for Distributed Optimization supplies the adaptive rule we use for ADMM's penalty weight.

Distributed inducing-point GPs exist, but none use peer-to-peer consensus.

Both rely on a server, and neither couples the learned model to a physical decision.

One paper constrains the design. A General O(n2) Hyper-Parameter Optimization for Gaussian Process Regression with Cross-Validation and Non-linearly Constrained ADMM uses ADMM to split an exact GP solve on one machine and deliberately rejects inducing points as too lossy. We take that concern seriously. Our first decisive check is whether the decentralized sparse model matches a centralized one, not whether it matches an exact GP.

The closest wireless competitor is Physics-Inspired Distributed Radio Map Estimation. It learns city radio maps across clients by combining a shared path-loss autoencoder with local shadowing autoencoders and a soft consensus penalty. It gives no predictive variance, which is exactly the quantity our power controller consumes.

Research Program

The model follows the physics-inspired decomposition of the radio-map work above. Each access point's log path gain is a deterministic log-distance path-loss term plus a Gaussian-process shadowing field. The kernel hyperparameters are shared across access points: correlation distance, variance, and a noise level. So are the inducing locations spread over the service area, together with a common shadowing component modeled as a shared latent function. Per-access-point residual fields stay local.

When all reports are uncensored, we compute the shared inducing posterior in closed form by averaging statistics over the neighbor graph. The ADMM machinery is not used. When reports are censored, each access point runs a few natural-gradient steps on its local objective in the primal update. It then exchanges its $O(M^2)$ natural parameters, and the $O(M)$ hyperparameters and inducing locations, with its neighbors. Transmissions whose change falls below a threshold are skipped, following COKE. The whole scheme is built so the conjugate special case falls back to averaging automatically. That lets us attribute any gain specifically to the censored and hyperparameter parts.

The decision layer chooses user-to-access-point associations and uplink powers. The goal is to maximize the rate of the weakest 5% of users, subject to a chance constraint that each served user's SINR (signal-to-interference-plus-noise ratio) exceeds a target with probability at least $1-\epsilon$. Because the posterior over log gains is Gaussian, we approximate the chance constraint deterministically. Each gain is replaced by its posterior mean minus $\kappa$ posterior standard deviations, with $\kappa$ set by $\epsilon$ under the Gaussian posterior. This is conservative, since interference terms are themselves uncertain; we will measure how conservative it is.

The coupling between access points runs through the interference sums. We split those sums with consensus ADMM, using the parallel two-group schedule of GADMM rather than a token ring. The local updates are non-convex because association variables are discrete. We relax them and round, and we report the number of iterations to a fixed tolerance rather than claiming a convergence guarantee.

A further analysis asks when the censored model matters. We predict that the gap between averaging-with-threshold-imputation and proper censored inference grows with the fraction of censored reports. We will sweep that fraction.

Baselines

The naive reference points are two. The first is purely local Gaussian processes, one per access point, with no exchange at all. The second is a pure log-distance path-loss model with no learned shadowing. Both run on the same measurements. Beating them by a clear margin in map error and in 5th-percentile rate is the minimum bar.

The accuracy ceiling for learning is a centralized sparse variational GP (Gaussian Processes for Big Data, with the censored likelihood from Scalable Variational Gaussian Process Classification). It is trained at a central processor on all raw reports, with the same $M$ and the same optimizer budget. It is the strongest competitor because it is the thing we approximate. The decentralized method "beats" it only in communication. On accuracy, success means matching it within noise across seeds.

As an external anchor, Gaussian Processes for Big Data reports an MSE of 0.426 on 75,000 UK property sales, against about 0.50 for a subset GP. We will partition that dataset geographically across agents and require the decentralized method to land near 0.426. For radio maps, no reported result found on Lacuna for RMSE on RadioMapSeer. The PI-DRME paper reports that it achieves the lowest RMSE at 1% measurement coverage but gives no number in the file. We will therefore re-run PI-DRME's decomposition ourselves.

The strongest server-free learning competitor is COKE-style consensus ADMM on random Fourier features (COKE). We give it a feature count chosen so its per-round message size equals ours, and the same number of communication rounds. We give it a squared-loss fit to censored-at-threshold values, because it has no likelihood model. Beating it means lower map error at matched bytes and, for the decision layer, lower true outage. COKE supplies no variance, so its power controller must use the plug-in mean.

The strongest server-based competitor is exact statistic aggregation. Each access point uploads its $O(M^2)$ statistics once to a central processor, which is exact under the Gaussian likelihood. A centralized optimizer then solves the same chance-constrained power control. We take this as the stronger version of the "why not a parameter server" objection. We run it at matched fronthaul bits.

For the decision layer, the upper bound is the same power-control solver given the true maps. That is our analogue of the perfect-CSI setting of Decentralized Beamforming Design for Intelligent Reflecting Surface-enhanced Cell-free Networks. We also run that paper's ring-scheduled incremental ADMM on our learned maps, to isolate the latency effect of the parallel schedule.

Two ablations isolate the proposal's ingredients. The first removes the variance by setting $\kappa=0$. It tests whether uncertainty-aware design is what reduces outage. The second replaces censored inference with threshold imputation plus exact averaging. It tests whether ADMM earns its place in the learning layer at all.

Why not just…?

Why not just average sufficient statistics by gossip and skip ADMM entirely? For Gaussian likelihoods with fixed kernel settings this is exact, and we adopt it. The question is only whether censoring and hyperparameter learning make it insufficient. This is a real threat. If sensitivity cut-offs rarely bind, or if threshold imputation barely hurts calibration, the ADMM layer is unnecessary for learning. The censoring-fraction sweep and the imputation ablation settle this directly. We will report the crossover point honestly, including the case where it is never reached.

Why not just ship statistics or raw reports to the central processor that cell-free networks already have? Where a well-provisioned central processor exists and noise is Gaussian, one-shot upload of $O(M^2)$ statistics is as good as our method. We include it as a baseline rather than argue against it. Our case rests on three settings. The first is fronthaul-constrained deployments (SIM-Enabled Hybrid Digital-Wave Beamforming for Fronthaul-Constrained Cell-Free Massive MIMO Systems). The second is censored measurements, where one-shot upload is not exact. The third is decision coupling, which needs repeated rounds anyway. Parameter-server communication totals in GADMM (495,792 for centralized LAG-PS versus 1,030 for GADMM on Derm logistic regression) suggest the gap can be large for iterative problems. That comparison is convex and not radio, so we measure our own.

Why not just use random Fourier features with COKE, which already has consensus and message skipping? It would give point maps of comparable accuracy at modest input dimension (two-dimensional location). Accuracy alone may therefore not separate the methods. The separation should come from the decision layer, since COKE yields no variance for the chance constraint. The matched-bytes comparison of true outage is what decides this.

Why not just estimate the map and then optimize power with the plug-in mean? This decoupled pipeline is the default in practice. Joint-design papers such as Delay-Constrained Joint Power Control, User Detection and Passive Beamforming in Intelligent Reflecting Surface-Assisted Uplink mmWave System show that ignoring coupling between design stages costs performance. However, they concern coupling among decision variables, not between learning and deciding. The $\kappa=0$ ablation is the direct test. If it matches the variance-aware version on true outage, the second half of our claim fails.

Proposed Experiments and Decisive Checks

The main testbed is RadioMapSeer: 700 simulated urban regions of 256 by 256 m at 1 m resolution. In each region we place several transmitters as access points, from 8 to 16. We sample user reports at 1% coverage, add quantization, and censor reports below a sensitivity threshold. We sweep the threshold so the censored fraction ranges from near zero to roughly half. Maps are scored against held-out pixels by RMSE, negative log predictive density, and empirical coverage of 90% predictive intervals. The interval coverage is the calibration check that the sparse approximation's known tendency to underestimate variance away from inducing points does not undermine the chance constraint.

The decision layer runs a large-scale-fading cell-free uplink simulation driven by the true maps. It uses designs computed from each method's predicted maps. We report true 5th-percentile user rate, the fraction of users missing the SINR target, total bits exchanged per access point, and wall-clock rounds. We vary the neighbor graph (ring, random geometric, complete) and the number of inducing points.

The generic sanity check is the geographically partitioned UK housing dataset. There the decentralized learner should approach the 0.426 MSE reported for the centralized method in Gaussian Processes for Big Data.

Four outcomes are decisive. First, decentralized learning must match the centralized sparse GP within seed variance on both UK housing and RadioMapSeer; otherwise the objection from the cross-validation ADMM paper holds. Second, the gap between censored ADMM and averaging-with-imputation must grow with the censored fraction, and must be visible in interval coverage near cell edges. Third, variance-aware power control must reduce true outage relative to $\kappa=0$ and relative to the COKE pipeline at matched bytes. Fourth, the parallel schedule must reach the same design as the ring schedule in fewer wall-clock rounds as the number of access points grows.

If the first holds but the second and third do not, the honest conclusion is narrower: averaging statistics is all a cell-free radio-map system needs, and ADMM belongs only in the decision layer.