WorksPhysics and Astronomy
A feedback-driven central limit theorem with -Gaussian limits
Abstract
We establish a central limit theorem, the feedback-driven central limit theorem, for sums whose summands are independent noise variables multiplied by an amplitude that depends on the preceding normalized sum . The summands are martingale differences whose conditional variance, , depends quadratically on , with feedback strength ; unlike in the martingale central limit theorem, this conditional variance does not settle to a constant. The noise has finite variance; for positive feedback we assume, as in Lyapunov's theorem, a finite third moment, and for negative feedback bounded noise, so that the conditional variance stays nonnegative. For positive feedback, the normalized sum converges to a Student law, equivalently a -Gaussian with , independently of the noise distribution. For negative feedback, it converges to a compactly supported density, which is a -Gaussian with when . The Gaussian law, , is recovered at zero feedback, while the Cauchy law, , is obtained only as a limiting case, in a triangular array in which the feedback grows slowly with the number of terms. Thus every is obtained for a fixed feedback strength, and as a limit. We further show that the balance identity behind the theorem singles out quadratic feedback: a -Gaussian with satisfies it only for a quadratic conditional variance, and no -Gaussian with satisfies it for any conditional variance. The proofs use the method of moments for negative feedback and a Wasserstein coupling argument for positive feedback, including cases in which the limiting variance is infinite.