control//hierarchical control//time-scale separation

Time-scale separation is the condition that the parts of a system evolve at rates far enough apart that each can be analysed with the faster parts treated as already settled and the slower parts treated as constant, and it is the assumption that lets engineers design a large system as a stack of small loops (hierarchical control). A drone's position loop asks for a tilt and assumes it gets it at once; the attitude loop holding that tilt assumes the position setpoint does not move during its own transient. Both assumptions are false and both are good enough, as long as the loops are far enough apart in speed.

Time-scale separation is the condition that the parts of a system evolve at rates far enough apart that each can be analysed with the faster parts treated as already settled and the slower parts treated as constant, and it is the assumption that lets engineers design a large system as a stack of small loops (hierarchical control). A drone's position loop asks for a tilt and assumes it gets it at once; the attitude loop holding that tilt assumes the position setpoint does not move during its own transient. Both assumptions are false and both are good enough, as long as the loops are far enough apart in speed.

The working figure is a factor of five to ten between neighbours. An attitude loop with a bandwidth of 20 Hz serves a position loop of 2 to 4 Hz well: by the time the position loop has changed its request noticeably, the attitude loop has long since delivered the previous one. The mathematics behind the habit is singular perturbation theory, which shows that the fast dynamics can be replaced by their equilibrium when the ratio of time scales is small, with an error that shrinks with the ratio.

It is what makes the inner loop look like an actuator. With enough separation an outer loop can be designed as if the inner loop were a gain of one, so each layer is tuned and tested on its own, inner loops first (cascade control).

When the separation breaks, the loops fight. Detuning an inner loop (a softer gain after a vibration problem, a heavier filter) or pushing an outer loop faster brings their bandwidths together, and the two start to exchange energy: the outer loop corrects an error the inner loop is still correcting, and the result is a slow oscillation that neither loop shows when tested alone. A request the inner loop cannot follow (a tilt beyond its authority) breaks it the same way, which is why the setpoints passed down are limited.

It holds far beyond control. A machine-learning system runs on the same ladder: microseconds of control, milliseconds of inference, days of training, weeks between retrainings on fleet data. A sensor fusion runs it in estimation, a fast inertial prediction between slow absolute fixes (multi-rate fusion), and a plant runs it from the drive to the ERP (automation pyramid).

It is also an argument for where computation goes. Each layer can live on hardware matched to its rate (a microcontroller at kilohertz, an onboard computer at tens of hertz, a server at minutes), and the interfaces between layers carry setpoints, which are few and slow, instead of raw measurements.