When the Beam Tracker Trusted Stale Pilots: Tracing a Sub-THz Outage Back to Its Hidden Invariant
The beams were correct when they were computed. By the time they were used, they no longer were. In a sub-THz cell running AI-assisted beam tracking on massive MIMO CSI, outages bunched up around fast-moving users while the predictor kept reporting high confidence. Finding the trigger was easy. Finding the assumption underneath it took much longer, because nobody had ever written it down.
The trigger, and why it is not the cause
The incident started after a load-balancing policy raised the CSI-RS periodicity to 20 ms on the busiest cells, so that resource elements could go to data during peak hours. The change stayed inside specification. The 3GPP reference signal framework, with CSI-RS for downlink channel acquisition and SRS for uplink sounding, is designed to let operators trade pilot density against overhead. Within minutes, radio link failures rose among users crossing the cell at vehicular speed, while stationary and walking users saw almost nothing. Rolling back the policy made the symptom go away, and that is exactly why the investigation could not end there. A legal configuration change that other layers absorb without complaint only breaks a system that was quietly relying on something else.
The assumption that stopped holding
The tracker had been trained and validated on the premise that every CSI snapshot it received was younger than the channel's coherence time. At sub-THz carriers, Doppler spread grows with carrier frequency, so coherence time for a vehicle at urban speed comes out at roughly a millisecond by a rough estimate, which is more than an order of magnitude shorter than the new pilot interval. CSI age was never an input to the model. It extrapolated angle and gain from evidence that had already decorrelated, and its confidence score measured how close the input was to the training distribution, not how fresh it was. Pedestrian traffic dominated the training set, and the premise held for pedestrians even at the relaxed periodicity, so validation never reached the boundary. No single team got it wrong: the scheduler owned periodicity, the ML pipeline owned the predictor, and the dependency between them existed only as an invariant nobody had stated.
What kept the rest of the cell alive
The users who stayed connected did so because the wide-beam control channel anchor and the lower-band fallback link never consulted the predictor, and that independence is the only reason the event was a degradation instead of a cell-wide blackout.
Turning the invariant into a monitored quantity
The fix that survives review is neither a shorter pilot interval nor a better model. It is making the hidden invariant explicit and observable. Each CSI report now carries its age, together with a coherence time estimate derived from the Doppler measured on the same user's SRS, and the predictor's output is gated on the ratio between the two. The decision rule is this: once CSI age exceeds half the estimated coherence time, the beam manager stops trusting the prediction, widens the beam by one codebook level, and requests aperiodic sounding for that user, returning to narrow tracking only when fresh reports agree again. This spends pilot overhead exactly where the load-balancing policy wanted to save it, and that trade-off is still open: two controllers now compete for the same resource elements, and an explicit priority has to say which one wins at peak load before the next incident makes that call for you.
Technical FAQ: Would retraining the predictor on vehicular traces have been enough?
Retraining widens the region where the freshness assumption holds, but it does not remove the assumption. A later change in carrier, pilot periodicity or mobility mix would move the boundary again without anyone noticing. Age gating keeps the failure observable and bounded whatever the model has learned, so retraining belongs next to the gate, not in place of it.