Work

Spectral Decay and Fading-Memory Bounds for Reset-Gated Quantum Recurrent Maps

Year: 2026
Type: preprint
Abstract: This paper studies a discrete-time recurrent map on density matrices in which the memory state is updated by a data-dependent re-uploading unitary and gated by an input-dependent replacement channel R_f(ρ) = fρ + (1 − f)σ, a quantum analogue of the forget gate of a long short-term memory cell. The starting point is the last-renewal representation familiar from stochastic resetting, extended from a fixed unitary and constant reset rate to input-dependent unitaries and input-dependent gates: the state after T steps is a convex combination of T reset-free horizon states with weights in closed form. On this basis three quantitative results are proved. First, as a function of any single input the output is a trigonometric polynomial whose frequencies lie in an explicit finite set determined by the encoding alone, and, for a constant gate, whose coefficients at nonzero frequencies are bounded by f^m ‖O‖, where m is the age of that input. Second, the sensitivity of the output to each input admits a bound that does not depend on the trainable unitaries, includes the contribution of the gate derivative, and is summable uniformly in T; no bound of this type holds for the classical cell. Third, the representation yields an exactly unbiased estimator of the output, using only reset-free unitary circuits, whose expected depth equals Σ_m P_m ≤ (1 − f_max)^(−1) uniformly in T, together with an explicit truncation bound. All identities are verified numerically to machine precision, and all bounds on random instances, in two independent implementations, and the reconstruction is executed on two superconducting quantum processors.
Language: English
Open Access status: green
OpenAlex ID: W7225886087
Cites: 16
Cited by:
Related to: 10