Cycle-Checked Phase Calibration for Cell-Free Joint Transmission

In cell-free massive MIMO, many distributed radio units send to the same user at once. This only helps if their oscillators are phase-aligned, so their signals add constructively at the receiver. In practice the oscillators drift. A hardware study on commercial radio units measured phase drift of up to ±30°, and found that a calibration only 5 ms out of date cut joint-transmission throughput by up to 46% (Experimental Performance Evaluation of Cell-Free Massive MIMO Systems Using COTS RRU With OTA Reciprocity Calibration and Phase Synchronization).

The network must therefore keep spending airtime measuring relative phases between pairs of radio units. It must also decide which units it trusts enough to group for coherent transmission, and how to set their beamformers. These are a discrete choice and a continuous choice coupled together, the kind of mixed-integer problem that wireless resource allocation already struggles to solve on millisecond deadlines.

This note proposes borrowing one idea from 3D vision and robotics: loop closure. When you measure relative offsets around a closed triangle, they must sum to zero modulo $2\pi$. How far a measured triangle misses zero is a free, label-free signal of how wrong those measurements are. We propose to use that miss in two ways. The first is to clean up the pairwise phase measurements. The second, which is the new part, is to turn the miss into a per-radio-unit uncertainty estimate that the resource allocator consumes directly. The allocator uses it to decide which units join a coherent cluster, how to beamform, and when to spend airtime recalibrating. In short, the triangle check tells the scheduler which radios it can afford to trust, and the scheduler then decides how to spend power, cluster membership, and calibration time accordingly.

Core Claim

Phase calibration across radio units is an angular synchronization problem. The goal is to recover absolute phases $\theta_i$ from noisy, sometimes corrupted relative measurements $\theta_{ij}$. Corruption comes from interference or multipath during the calibration slot. The vision community has strong tools for this problem. Message passing over cycles reweights edges that repeatedly break closure, as in MPLS and CEMP (Message Passing Least Squares Framework and its Application to Rotation Synchronization, Robust Group Synchronization via Cycle-Edge Message Passing). A quadratic program can estimate per-edge corruption levels directly, as in DESC (Robust Group Synchronization via Quadratic Programming). A directed graph neural network can be trained with a triangle-inconsistency loss and no ground truth, as in GNNSync (Robust Angular Synchronization via Directed Graph Neural Networks). GNNSync's stated limitation to the 2D rotation group SO(2) is not a limitation here, because oscillator phases live in exactly that group.

Our claim goes beyond "apply angular synchronization to radios." Cycle residuals yield calibrated per-unit phase-error variances $\hat\sigma_i^2$. Feeding these variances into the joint allocation problem should beat two alternatives: designs that use only point phase estimates, and designs that assume a uniform worst-case phase error. The mechanism is simple. Suppose unit $i$'s residual phase error is roughly Gaussian with variance $\sigma_i^2$, and drifts as a random walk with rate $\kappa_i$ since the last calibration $\tau$ seconds ago. Then the expected coherent amplitude it contributes is scaled by

$$\rho_i(\tau) = \exp\!\left(-\tfrac{1}{2}\left(\hat\sigma_i^2 + \kappa_i \tau\right)\right).$$

The scale follows from the characteristic function of a Gaussian phase. The linear growth in $\tau$ follows from the random-walk drift model, and the measured 5 ms sensitivity suggests that this term dominates quickly.

Replacing each channel $h_{k,i}$ in the SINR expression with $\rho_i h_{k,i}$ gives a coherence-aware signal-to-interference-plus-noise ratio. The allocator then chooses three things: which unit pairs to measure in the next calibration slot, which units to activate and cluster for each user, and the transmit beamformers. The discrete part includes the calibration graph itself. Each triangle added to that graph buys redundancy, and so a better $\hat\sigma_i^2$, at the cost of data airtime.

Research Program

The first component is the uncertainty estimator. We will take DESC's estimated edge corruption levels and MPLS-style cycle reweighting, and convert them into per-node variance estimates. We will then test whether these estimates are calibrated, meaning that predicted variance matches empirical squared phase error. Calibration is the property the downstream allocator depends on. Prior synchronization work reports accuracy, not calibration, so this is a genuine gap. We will also train a GNNSync variant with its cycle-inconsistency loss, add a variance output head, and compare the two estimators.

The second component is the calibration graph design. Long chains of radio units with few triangles are precisely where cycle methods fail, a limitation Cycle-Sync notes for sparse topologies (Cycle-Sync: Robust Global Camera Pose Estimation through Enhanced Cycle-Consistent Synchronization). Unlike a robot's trajectory, a radio network can choose its measurement graph. We will study greedy and learned rules for adding triangles. These rules target units whose $\rho_i$ most limits the users they serve. In this sense, loop closure becomes something the system allocates rather than something it finds.

The third component is the joint solver. The problem with fixed $\rho_i$ is the Cloud-RAN problem of selecting radio heads and beamforming under SINR targets. We will solve it with exact branch-and-bound and with the imitation-learned pruning policy of Scalable Network Adaptation for Cloud-Rans: An Imitation Learning Approach. To that policy's size-independent tree features we add the $\rho_i$ values. The calibration-slot decision sits in an outer loop that chooses the recalibration interval and edge set by comparing expected throughput across a small set of candidates.

Baselines

The naive reference point is a spanning-tree calibration refreshed on a fixed schedule. It measures the minimum number of pairs, uses no redundancy, and all calibrated units are clustered as if they were perfectly coherent. It shows what happens when no structure is exploited at all.

The strongest existing calibration method is the averaged Argos procedure with genetic grouping from Experimental Performance Evaluation of Cell-Free Massive MIMO Systems Using COTS RRU With OTA Reciprocity Calibration and Phase Synchronization. It is the only method in the file that has been run on real radios. That paper reports the ±30° drift and the 46% throughput loss at 5 ms delay. No reported calibration-error figure for the method itself was found on Lacuna. We will reimplement it on the same simulated drift and measurement model, give it the same number of calibration slots, and tune its grouping hyperparameters with the same search budget as our method.

On the allocation side, the strongest baseline is imitation-learned branch-and-bound, which reports over 20× speedup with under 0.6% objective loss at 10 radio heads and 15 users (Scalable Network Adaptation for Cloud-Rans: An Imitation Learning Approach). We run it at that same scale, given perfect-coherence assumptions and the best point phase estimates.

Beating these baselines means higher delivered throughput, or lower power at equal SINR-target satisfaction, at equal total airtime including calibration. Wins must appear across at least the 10-unit, 15-user setting and a sparser larger network.

The key ablation keeps everything identical but passes only point estimates to the allocator, with $\rho_i = 1$ for every unit. If this ablation matches the full system, the uncertainty transfer is not doing the work. For the synchronization stage alone, the reference results are DESC outperforming linear programming and CEMP above 40% corruption, and Cycle-Sync's exact recovery up to 80% corrupted edges on synthetic uniform-corruption graphs. Both are reported in the cited papers.

Why not just…?

Why not just calibrate more often with pairwise Argos? Calibration slots are airtime taken from users. The 46% loss at 5 ms shows that drift races calibration. More frequent full calibration directly trades against data throughput, and our airtime-matched comparison measures exactly this trade. If frequent Argos wins at every overhead level, the triangle redundancy is not worth its cost, and we will report that.

Why not just run robust angular synchronization and then allocate with point estimates? This sequential approach is the real threat. Good synchronization may make the residual phase error small enough that uncertainty barely changes cluster decisions. The point-estimate ablation settles the question. We expect the gap to open mainly with stale calibrations and with weakly connected units, where $\rho_i$ varies most across the network. If the gap is negligible everywhere, the contribution reduces to a better calibration step, and we would narrow the claim to that.

Why not just use worst-case robust beamforming with a uniform ±30° bound? A uniform bound is safe but forces every unit to be treated as the least reliable one. The worst-case robust beamforming variant will be run as an extra baseline. Per-unit estimates should let the allocator keep well-measured units coherent while excluding drifting ones. If the uniform design loses little, phase heterogeneity across units is too small to matter in practice.

Proposed Experiments

No wireless phase-synchronization benchmark appears in the file, so evaluation has two parts.

Synchronization quality is checked on the GNNSync synthetic topologies (Erdős–Rényi, Barabási–Albert, and random geometric graphs) and on uniform-corruption graphs. In addition to mean angular error, we report the calibration of $\hat\sigma_i^2$, using expected calibration error on squared phase error.

The system test is a cell-free simulator built on the Cloud-RAN instances of the imitation-learning paper (10 radio heads and 15 users, plus a 6-head, 8-user training scale). We add a random-walk oscillator drift model fitted to the reported ±30° range and 5 ms sensitivity. Calibration measurements suffer SNR-dependent noise plus a controlled fraction of corrupted slots. Metrics are delivered downlink throughput, total power at fixed SINR targets, fraction of users missing their target, calibration overhead fraction, and solve time against the millisecond scheduling budget.

The decisive checks are three. First, the full system must beat the point-estimate ablation by a clear margin when calibrations are stale, for example by recovering a meaningful share of the 46% loss at 5 ms. Second, adding chosen triangles must outperform spanning-tree calibration at matched overhead. Third, the imitation-learned solver with $\rho_i$ features must stay within the reported sub-1% optimality gap of exact branch-and-bound on the coherence-aware objective. Failing the first check means the core claim is wrong, whatever the other results show.