Separating Known and Unknown Interferers in Open RAN with Signature-Style Matrix Factorization

When a phone at the edge of a cell gets poor service, the base station sees only the total interference on each slice of spectrum. It cannot tell which transmitter caused it. Part of that interference comes from neighbouring cells that the network itself scheduled, so the network knows exactly when and where they transmitted. Part comes from sources it does not control: another operator's cell, a cell from a different vendor outside the controller's reach, a misbehaving device, or a jammer. Cancer genomics faces the same structure. A tumour's mutation counts are a non-negative sum of known processes, such as the aging-related signature present in almost every tumour, plus unknown ones that must be discovered. Genomics handles this with non-negative matrix factorization (NMF), which writes a table of non-negative counts as a sum of a few non-negative "signatures" times non-negative "exposures." The best tools in that field hold known components fixed and push the new ones to be sparse, so that ubiquitous processes do not leak into the discovered ones.

We propose to bring this split between fixed known components and discovered sparse ones into Open RAN interference management. Here the radio access network is controlled by software apps running on a near-real-time controller, the RAN Intelligent Controller or RIC, on a 10 ms to 1 s loop. Each cell's per-resource-block uplink interference-plus-noise power becomes the data matrix. The network's own scheduling logs fix the activity of controlled transmitters. Factorization recovers the spatial footprint and time-frequency activity of the uncontrolled ones. The output is an interpretable and cheap answer to the question "what is hurting this cell, and when," computed inside the controller's time budget.

A second point follows from the reproducibility literature. The method depends on the scheduling log matching what actually went over the air, and host-level timing glitches break that match. The evaluation is therefore built as a hermetic replay benchmark, with a deliberate test of whether execution-environment artifacts show up as phantom interferers.

Core Claim

Let $V$ be a $C \times T$ matrix. Each row is a receiving port across the cells, and each column is one (resource block, slot) pair. Entries are measured uplink interference-plus-noise power on a linear scale. Let $S$ be the $M \times T$ known activity of the $M$ transmitters under RIC control, meaning their scheduled power per resource block and slot taken from scheduler logs. Let $n$ be the per-port thermal noise floor, estimated from idle blocks and held fixed. We fit

$$ \min_{A \ge 0,\; W \ge 0,\; H \ge 0} \; \left\| V - A S - W H - n \mathbf{1}^\top \right\|_F^2 + \lambda \|H\|_1 , $$

where $A$ ($C \times M$) holds the unknown coupling gains from controlled transmitters to each port, $W$ ($C \times K$) holds the spatial footprints of $K$ uncontrolled sources, $H$ ($K \times T$) holds their time-frequency activity, and $\lambda$ sets how sparse that activity is.

The model has three parts, each borrowed from genomics. The fixed noise floor plays the role of the clamped background signature in SparseSignatures. That method holds a known background such as COSMIC SBS5 constant, places a LASSO penalty on the signatures, and selects both the number of signatures and the penalty by repeated bi-cross-validation. The known-activity term $AS$ corresponds to refitting known catalogue signatures. The $WH$ term is de novo discovery.

We depart from SparseSignatures in two ways. The sparsity penalty sits on activity rather than on footprints, because uncontrolled interferers are intermittent. Bi-cross-validation chooses $K$, which here means how many uncontrolled sources exist. We borrow bootstrap-based stability scoring from CANCERSIGN, which runs 600 bootstrap resamplings so that only signatures that recur are reported.

The claim is specific. With $S$ clamped, the residual is identified much better than either of two simpler options. Unconstrained NMF on all of $V$ lets controlled scheduling bleed into discovered components, which is the radio analogue of SBS5 bleeding. Plain regression on $S$ lumps all uncontrolled sources into one undifferentiated residual. The fit should also run well inside the RIC window.

Research Program

The first component is the estimator: multiplicative-update NMF with the clamped blocks, warm-started across controller cycles, plus a bootstrap stability score per discovered source. The second is a recording and replay harness. Live over-the-air experiments cannot be repeated exactly, so every comparison runs on recorded traces. Each trace records the interference matrix, scheduler logs, and ground-truth injection logs. Traces are replayed inside version-pinned containers with append-only provenance logs, following the design of the audit-grade SciAgent harness. SciAgent runs scientific agents in pinned containers and audits them with a verifier that sees only logs and artifacts. We reuse its logging and isolation discipline but not its language-model verifier, which would add nothing here. Hermetic isolation also prevents leakage between consecutive runs, which Haben et al. identify as a main source of flaky results.

The third component is a small downstream policy. It converts discovered sources into soft resource-block preferences, de-weighting blocks where a stable uncontrolled source is active. This mirrors the "soft preference" interface that xDiff uses to bridge the RIC's 10 ms to 1 s loop and the distributed unit's 1 ms scheduling.

Data come from two places. One is an OpenAirInterface testbed with the FlexRIC controller, configured like xDiff's deployment of 3 cells and 10 commercial phones. We add software-defined radios that inject uncontrolled interferers whose patterns are logged but hidden from the method: narrowband, periodic, a wideband edge leak, and a "foreign cell" carrying its own scheduled traffic. The other is a ray-traced scene built with SionnaRT, in the manner of Campus5G. Campus5G calibrated SionnaRT against a 15 km walk test of about 11,000 measurements over 20 radio units and names AI-driven interference management for low-SINR areas as future work. The scene lets us scale $C$ to 20 or more receiving units, where footprints become far more identifiable than with 3 cells.

Baselines

The naive reference is a static per-block interference map: the time-averaged $V$, with blocks above a threshold flagged as bad. It costs nothing, and any attribution method must beat it on downstream throughput to be worth running.

The most important baseline is non-negative least squares on the known activity alone. It fits $A$, fixes $W = 0$, and treats the residual above the noise floor as one "external" component. This is what a network engineer with scheduler logs would do first. It is the honest competitor because it shares our clamped term exactly. It also serves as the ablation that isolates our key ingredient, the de novo sparse factorization of the residual. We beat it if we resolve distinct injected sources it cannot. Concretely, we want mean cosine similarity of at least 0.9 between recovered and true footprints and activities for each injected source, and a downstream throughput gain, both on the same replayed traces with matched compute.

The second baseline is unconstrained NMF on all of $V$, with $K$ chosen by the same bi-cross-validation. It isolates the value of clamping. We expect it to split controlled cells into spurious components or absorb them into external sources. We count it beaten if our method shows lower bleed, measured as correlation between discovered activity and the controlled $S$, at equal reconstruction error.

The strongest existing interference-management method in the file is xDiff, a conditional diffusion model with double Q-learning that outputs soft block preferences. Its authors report it outperforms DDQN and DDPG and prevents phone disconnections. They report an inference latency of about 21.8 ms. On Lacuna we found no reported throughput number for xDiff that we could reuse, and no reported result for interference-source attribution on any O-RAN benchmark. We will retrain xDiff on our traces with the same training-data budget and an equal hyperparameter search budget. Our policy beats it if it matches xDiff's cell-edge throughput within noise at a fraction of its latency, or exceeds xDiff when an uncontrolled interferer appears that was absent from training.

Why not just…?

Why not just regress on the scheduler logs? For controlled cells, that is the right answer, and our model already does it. The problem is everything else. The residual from regression is one blob, so a policy cannot tell a periodic narrowband source, which you avoid on specific blocks and times, from a foreign cell, which you avoid by spatial coordination. This is a real threat. If the regression baseline's downstream throughput matches ours when interferers are injected, the de novo term is not earning its keep, and we would report that.

Why not just run plain NMF, PCA, or independent component analysis on the whole matrix? PCA and independent component analysis allow negative footprints, which are meaningless for received power. As the SparseSignatures authors argue for mutation counts, cancellation between components destroys parts-based interpretation. Plain NMF respects non-negativity but has no reason to keep known scheduling separate. The unconstrained-NMF baseline measures directly how much it bleeds.

Why not just train a deep RL or diffusion controller end to end? xDiff shows this works within the RIC window. It also documents that DDQN and DDPG were unstable under heavy interference. A learned controller does not say what it is reacting to, and it has no mechanism for a new external source except retraining. The decisive test is the held-out-interferer condition. If xDiff adapts as well as our method there, our advantage reduces to interpretability and roughly an order of magnitude less compute. We would state it that way.

Why not just use zero-forcing equalization? Zero-forcing nulls interference between users it has channel estimates for. Chu et al. needed it in their cell-free testbed and note that its matrix inversions scale poorly at wider bandwidths. It also cannot null an uncontrolled source with no pilot. Our method operates at the controller layer and is complementary to it, not a replacement.

Proposed Experiments

The identification experiment sweeps three factors on replayed testbed and SionnaRT traces: the number of injected uncontrolled sources (one to five), their duty cycle, and the number of receiving ports (3 to 20). We report three things for each method. The first is source-recovery cosine similarity after Hungarian matching. The second is the accuracy of the bi-cross-validated choice of $K$. The third is bootstrap stability.

The prediction is that clamping and sparsity together keep recovery high where unconstrained NMF degrades. We also expect identifiability to fail predictably once $K$ approaches $C$, because $W$ has only $C$ rows. Finding that boundary is itself a result.

The latency check times warm-started fits per controller cycle on the testbed's edge server. It must stay well under the 10 ms to 1 s window, and we compare it against xDiff's reported 21.8 ms.

The downstream check runs the soft-preference policy over the air against the static map, regression on logs, and retrained xDiff. Metrics are cell-edge (5th-percentile) throughput and disconnection counts. Each comparison is repeated across days to measure run-to-run spread.

The final check tests whether the execution environment fakes interferers. Campus5G found that real-time baseband processing needed core isolation, pinned threads, and disabled CPU power-saving states to avoid latency spikes. We rerun identical injection scenarios with those protections on and then off. If late or missed slots make the logged $S$ diverge from what was transmitted, the clamped model should push the mismatch into the residual. A stable spurious "source" correlated with controlled traffic would then appear only in the unprotected runs.

This check settles whether host-level reproducibility is a correctness requirement for log-clamped attribution rather than mere hygiene. It also produces a diagnostic: a discovered source whose activity correlates with $S$ flags corrupted logs rather than a real interferer.