Traffic routing is the canonical argument for real-time data: conditions change minute to minute, accidents reshape the network, rush hours move through the city like weather. You need live data. Everyone knows this.
Oriel-Singer et al. decomposed the Seoul traffic matrix into spatial and temporal components and found that the temporal variation concentrates on a small number of modes. A low-rank predictor using just those modes — no live feeds, no sensors, no streaming infrastructure — incurs less than 1.5 minutes of excess travel time compared to full real-time routing. Over a full year of validation.
The tail matched too. Not just the average case — the worst cases. The predictor's distribution of excess time tracks the real-time system's distribution. The disasters you'd expect from flying blind aren't there, because the system isn't actually chaotic. It's periodic with noise.
The through-claim: systems that appear to require continuous monitoring are often secretly low-dimensional in time. The complexity lives in the spatial structure — which roads connect to which, how the network geometry shapes flow — not in the temporal dynamics. The hours of the day, the days of the week, the seasons of the year account for most of the variation. What looks like unpredictable flux is mostly a clock.
This matters beyond traffic. Markets have this property (most variation is intraday seasonality, not signal). Hospital census has it. Power grid load has it. The question “do we need real-time data?” often decomposes into “is the temporal dimension actually high-rank?” and the answer, surprisingly often, is no. A good model of the period beats a noisy measurement of the instant.