Persistency of Excitation
Persistent excitation is the richness condition on input signals that guarantees adaptive and identification algorithms converge to true parameters.
Why richness matters
An adaptive law drives tracking error to zero, but that does not mean the estimated parameters equal the true ones; many parameter values can produce the same output for a given input. To pin down the parameters uniquely, the input must exercise all the modes of the system. Persistency of excitation formalizes exactly how rich the signal must be.
The condition
A regressor vector phi(t) is persistently exciting if there exist positive constants and a window length such that the integral of phi phi-transpose over every window of that length is bounded below by a positive-definite matrix. Equivalently, the signal spans all directions in parameter space repeatedly over time. A sinusoid at n distinct frequencies persistently excites a regressor of dimension up to 2n.
Consequences
When PE holds, parameter estimates in recursive least squares and Lyapunov-based adaptive laws converge exponentially to the true values, and the closed loop inherits robustness. When PE fails, estimates may converge to wrong values or drift, and small disturbances can cause slow instability. This is why deliberately injecting probing signals, dither, is sometimes necessary, connecting to the exploration-exploitation tension of dual control.
- Tracking to zero does not imply correct parameters
- PE requires the regressor to span parameter space repeatedly
- n frequencies excite up to 2n parameters
- Failure causes estimate drift and fragile adaptation
Because PE cannot always be assumed, robust adaptive designs add safeguards, sigma-modification, projection, dead zones, that keep estimates bounded even without excitation, trading exact convergence for guaranteed stability.
PE is the shared foundation of MRAC, self-tuning regulators, and system identification generally. Recent concurrent-learning methods relax it by reusing stored data, but the underlying requirement, that the data be informative, is unchanged.