Most radio resource algorithms for cellular networks are designed as if each base station were alone. Interference from neighbouring cells is folded into a fixed noise floor, and under that simplification power control, scheduling and network slicing become clean optimization problems. In dense deployments this assumption fails, but not everywhere and not all the time. Some cells are strongly affected by what their neighbours do. Others are nearly isolated. Operators have no cheap way to tell the two apart without exchanging instantaneous channel measurements over the links between base stations, which is expensive. We propose to estimate, from routine per-cell performance logs alone, how strongly each neighbour's activity degrades each cell. We then use that estimate to decide where a single-cell optimizer can be trusted and where coordination is worth paying for.
The mechanism adapts an idea from cloud monitoring. There, a directed form of dynamic time warping aligns one service's metrics against another's to measure how much the first depends on the second. Dynamic time warping is an algorithm that matches two time series while letting one run faster or slower than the other. In the directed form, matches are allowed only in the causal direction and only within a bounded delay. In a cellular network, the "cause" series is a neighbour's resource usage and transmit activity. The "effect" series is an interference-sensitive quality metric at the target cell. The delay between them drifts during the day with queue backlogs, handover timers and crowd movement. The claim is specific and falsifiable. When that delay is non-stationary, a latency-bounded directed warping score recovers true interference coupling better than lagged correlation, and a network that coordinates only the flagged cell pairs matches full coordination at a fraction of the signalling cost. When the delay is roughly constant, we expect no advantage, and the experiments are built to show which regime real networks sit in.
What the literature gives and leaves open
The motivation comes from wireless work showing that single-cell designs break under coupling. Uplink Pilot and Data Power Control for Single Cell Massive MIMO Systems with MRC derives joint pilot and data power control for one cell. It guarantees "at least 2.2 bit/s/Hz for every user", but its authors note that ignoring out-of-cell pilot contamination invalidates the policy in multi-cell settings. Massive Connectivity with Massive MIMO-Part II shows why: reused pilots from adjacent cells corrupt channel estimates. A Heuristic Scheduling Scheme in Multiuser OFDMA Networks shows that one cell's scheduling sets its neighbours' interference. Two learning-based controllers also stop at the cell boundary and name multi-cell interference as their open problem: the single-cell policy-gradient allocator of Resource Allocation in Information-Centric Wireless Networking With D2D-Enabled MEC and the multicast power controller of Scheduling and Power Control for Wireless Multicast Systems via Deep Reinforcement Learning. The slicing work does the same. Energy Efficient Communication and Computation Resource Slicing for eMBB and URLLC Coexistence solves slicing for one base station and leaves inter-cell interference as an open problem. None of these papers offers a way to measure where its own assumption is violated. That measurement is what we add.
The method comes from AID. AID defines Dynamic Status Warping, a directed, round-trip-time-bounded variant of dynamic time warping over one-minute microservice metrics. On a Huawei Cloud trace it reduced dependency-intensity error by 61.1% in MAE and 45.8% in cross entropy relative to its best baseline, which it describes as including Pearson, Spearman and Kendall correlation. Our proposal differs from AID in three ways.
- Physics of the link. Cellular interference is a broadcast side effect, not an explicit caller-callee call. There is no call graph to follow, so every nearby cell is a candidate cause.
- Source of the delay bound. The bound comes from scheduling, queueing and handover timing, not network round-trip time.
- Confounding. Neighbouring cells share commuters, so their loads move together even without any interference between them. A lead-lag score can therefore detect correlated demand rather than coupling. AID did not face this problem.
The third difference is the main scientific risk, and the evaluation is built around it.
Two further papers play smaller roles. DTWNet supplies an optional learnable extension: gradients flow along the fixed optimal alignment path, so a learned local cost is cheap to train. It is evaluated only on UCR-style benchmarks, so its wireless value is untested. Existing cellular traffic models align time rigidly. MPGAT uses temporal convolutions over graph attention. MetaSTNet uses fixed windows with a GCN. Understanding Mobile Traffic Patterns of Large Scale Cellular Towers uses Fourier analysis and Euclidean clustering. These papers forecast load; none of them estimates interference coupling.
Core Claim
For a target cell $i$ and neighbour $j$, let $x^j_s$ be $j$'s activity at time bin $s$. We use PRB occupancy per subband, weighted by transmit power; a PRB (physical resource block) is the basic time-frequency unit a base station allocates. Let $y^i_t$ be an interference-sensitive quantity at $i$ with $i$'s own load regressed out. We use the residual of per-PRB spectral efficiency or channel-quality reports given $i$'s own occupancy. Directed warping restricts alignment paths to pairs $(s,t)$ with $0 \le t - s \le L$, where $L$ is a physical upper bound on the effect delay. We compute $L$ from scheduler queue length and handover time-to-trigger. The path is found by the usual dynamic-programming recursion under this band constraint. The coupling score is
$$ C_{j \to i} = \frac{\bar{d}_{\text{null}} - \min_{W \in \mathcal{W}_L} \frac{1}{|W|}\sum_{(s,t)\in W} c(x^j_s, y^i_t)}{\bar{d}_{\text{null}}}, $$
where $\mathcal{W}_L$ is the set of causal paths within delay $L$, $c$ is a local cost (negative standardized product, or a learned cost in the DTWNet variant), and $\bar{d}_{\text{null}}$ is the same minimum computed on block-shuffled copies of $x^j$. The shuffling keeps daily periodicity but breaks alignment, so the null already absorbs the periodic co-movement that Understanding Mobile Traffic Patterns of Large Scale Cellular Towers found carries over 94% of traffic energy. A cell is declared "single-cell safe" if $\max_j C_{j\to i}$ falls below a threshold calibrated on the null.
Research Program
We first build a multi-cell simulator by extending the 3GPP TR 36.885 Manhattan setup, which Resource Allocation in Information-Centric Wireless Networking With D2D-Enabled MEC uses with a 500 m radius and 10 RBs, from one cell to a grid of cells. The simulator has mobility-driven load waves whose propagation speed varies by time of day. The feature that makes ground truth possible is an interventional switch. We can remove $j$'s interference from $i$'s SINR while leaving the mobility and demand processes untouched. True coupling $C^*_{j\to i}$ is the drop in $i$'s throughput per PRB when the switch is on. This separates real interference coupling from shared-demand correlation, which observational scores cannot do by themselves.
Second, we compute directed warping scores and every baseline from logs aggregated to operator-realistic granularities of one second, one minute and fifteen minutes. We then measure how well each score recovers the interventional coupling.
Third, we close the loop with selective coordination. Flagged cell pairs run a multi-cell-aware power control that exchanges coarse load summaries. Unflagged cells keep the single-cell Max-Min policy from the MRC paper.
Fourth, we check stability on real traffic from the Milano and Trento datasets used by MetaSTNet. These datasets contain load but no interference. They can only tell us whether lead-lag delays between neighbouring cells are actually non-stationary, which is the precondition for warping to matter. They cannot validate coupling.
Baselines
The naive reference point has two parts. The first is to trust the single-cell assumption everywhere: every cell runs the single-cell Max-Min policy with no coordination. The second is its opposite, full coordination of all neighbour pairs. Together they bracket both throughput and signalling cost. Selective coordination beats the naive point if it recovers at least most of the gap in worst-user throughput between these two while exchanging messages on a minority of pairs.
The strongest competitor for coupling inference is maximum lagged cross-correlation. For each pair we search integer lags over the same window $[0, L]$ and use the same load residualization and the same shuffled null. This isolates the only thing warping adds: letting the lag change within a trace. Tuning is matched by giving both methods the same grid over $L$, the same aggregation levels and the same validation split. We also run Granger-style lagged regression and Pearson, Spearman and Kendall correlation, which are the correlation baselines AID compared against. We found no reported result on Lacuna for any coupling-inference method on wireless networks. The only relevant figure is AID's 61.1% MAE reduction over correlation baselines, and it comes from cloud microservice traces. It motivates the question but does not transfer as a target. Beating lagged correlation means a clearly higher area under the ROC curve (AUROC) and rank correlation against $C^*$, with confidence intervals over simulation seeds, specifically in the non-stationary-delay regime.
The ablations isolate the key ingredients. Undirected warping removes the causal constraint. Unbounded warping removes the latency band. Warping without load residualization tests the confounding control. The DTWNet learned-cost variant is included only if the exact-path version saturates.
For the secondary forecasting check, we use the warping-derived neighbour graph as the adjacency for a spatiotemporal forecaster. The reference number is MPGAT's reported 5-minute MAPE of 0.1511, against 0.1576 for GWNET, on its Taiwan intersection data (MPGAT). That dataset is from a telecom provider and may not be obtainable. If so, the comparison runs on Milano with a re-trained MPGAT-style model, and we will not claim to beat the 0.1511 figure.
Why not just…?
Why not just use lagged cross-correlation? This is the real threat. If coupling delays are close to constant, a single best lag captures them and warping adds only variance and the risk of overfitting. We do not argue around this. The simulator varies delay non-stationarity as an explicit knob, and the report will show the crossover point where warping starts to help. The real-data analysis measures where actual cells fall relative to that point. If real networks sit on the stationary side, the conclusion is that lagged correlation with the same null is sufficient, and that is still a useful answer.
Why not just read neighbour signal strength from UE measurement reports or propagation maps? Handsets already report neighbour received power, so the geometry of interference is partly known. Received power, however, says nothing about whether the neighbour is transmitting on the same subbands at the moments that matter, or whether the target cell's quality is limited by interference rather than by its own load. We include a measurement-report coupling score in the simulator as a baseline and also combine it with warping. If geometry alone matches the interventional ground truth, the temporal machinery is unnecessary for flagging cells, and we will report that.
Why not just let multi-agent reinforcement learning absorb the coupling? The single-cell learning controllers in Scheduling and Power Control for Wireless Multicast Systems and Resource Allocation in ICWN With D2D-Enabled MEC name multi-cell extension as unsolved because the state space grows too large. A coupling map is complementary to these methods: it tells a learner which neighbours to include in its state. We test this by comparing learners given all neighbours, no neighbours, and only the flagged neighbours.
Why not Fourier decomposition? Removing daily and weekly periodic components handles the dominant cycles but assumes they are stationary. The shuffled null already removes periodic co-movement, so the Fourier baseline is subsumed. We report it anyway.
Decisive Checks
The first check uses the interventional ground truth. Directed warping must beat lagged correlation in AUROC against $C^*$ when delay drift is high, and must tie it when drift is low. A win everywhere would suggest leakage in the evaluation, not insight.
The second check is the confounding test. We build scenarios with strong shared mobility and zero interference, meaning the switch is on for that pair. In these scenarios, scores must stay at the null level. If they rise, the method is detecting correlated demand and fails as a coupling estimator.
The third check is operational. Selective coordination must recover most of the gap between all-single-cell and full coordination in worst-user throughput while coordinating few pairs, and must do so at log granularities an operator actually has.
The fourth check uses the real Milano and Trento traces. It determines whether the non-stationary regime occurs in practice, and therefore whether the method matters outside simulation.
References
- Uplink pilot and data power control for single cell massive MIMO systems with MRC
- Massive Connectivity With Massive MIMO - Part II: Achievable Rate Characterization
- A Heuristic Scheduling Scheme in Multiuser OFDMA Networks
- Resource Allocation in Information-Centric Wireless Networking With D2D-Enabled MEC: A Deep Reinforcement Learning Approach
- Scheduling and Power Control for Wireless Multicast Systems via Deep Reinforcement Learning
- Energy Efficient Communication and Computation Resource Slicing for eMBB and URLLC Coexistence in 5G and Beyond
- AID: Efficient Prediction of Aggregated Intensity of Dependency in Large-scale Cloud Systems
- DTWNet: a Dynamic Time Warping Network
- Multivariate and Propagation Graph Attention Network for Spatial-Temporal Prediction with Outdoor Cellular Traffic
- MetaSTNet: Multimodal Meta-Learning for Cellular Traffic Conformal Prediction
- Understanding Mobile Traffic Patterns of Large Scale Cellular Towers in Urban Environment