When Does Multi-Operator Cellular Redundancy Help V2X? A Cross-Layer Measurement Study

Abstract

Multi-operator cellular redundancy is widely proposed for the ultra-reliable, low-latency requirements of vehicle-to-everything (V2X) safety services, increasingly through predictive or learned schedulers. Such schedulers rest on an assumption: that outages on different operators correlate in ways a scheduler could exploit. Using three open vehicular cellular datasets spanning the radio layer and a V2X application-layer trace, we test it directly. The correlation is real but small: dominated by shared geography, it barely dents the benefit simple redundancy provides and is too weak for a scheduler to exploit. Instantaneous radio quality predicts throughput only weakly and inconsistently under operational mobility — a nonlinear model evaluated out of sample still leaves most of the variance unexplained — so, rather than rank operators by predicted rate, we pose operator selection at the outage layer. There, joint outages exceed location-conditional independence, but only modestly, dominated by physical co-location not operator coupling (Mantel–Haenszel rate ratio R = 1.47 off-mast, 95% CI [1.15, 1.98]); pairs that do not share low-band spectrum behave as negative controls. This weak dependence reduces the realized dual-operator availability gain by about 2.5% (retaining ${\sim }97.5\%$; at most ${\sim }5\%$ for shared-mast pairs), so simple redundancy keeps essentially all its independence-predicted benefit, and at the V2X application layer it leaves even a perfect correlated-outage scheduler less than a single 3GPP reliability step. The practical answer is simple: field two operators and use whichever link is up or, under duplication-cost constraints, prefer the more reliable one with reactive failover. Under a binary per-second availability metric, a learned scheduler whose value rests on exploiting correlated-outage structure is unwarranted by this evidence; predictive objectives independent of that structure (mobility-, load-, or history-driven prediction, latency-tail control) lie outside our scope. A reproduction package emits every number from one fixed-seed pipeline.