The control cost of resetting a qubit without revealing its past

Sinan Bugu’s study, submitted to arXiv on November 8, develops mathematical limits for this task and constructs feedback strategies that approach them. It combines analytical proofs with computer simulations, not laboratory measurements. The supplied publication record does not establish peer review.

A reset judged by its whole history

The model compares two situations. In the reference case, a qubit starts in its ground state and remains undriven. In the other, it starts excited, and a controller uses its ongoing measurement signal to guide a drive toward the ground state. Monitoring begins immediately, so the record includes evidence produced before and during the reset.

An observer sees that record, but not the drive or control hardware. The controller cannot remove earlier evidence; it can only influence what gets recorded next.

The study measures the statistical difference between the two possible record distributions using Kullback–Leibler divergence, called D. Smaller D means less distinguishability, rather than identical individual traces. Control effort, J, is the expected time integral of the squared Rabi frequency, which describes how strongly the drive rotates the qubit. This is not a calculation of heat or thermodynamic work, and converting it into physical energy requires device-specific calibration.

Why a fixed pulse is not enough

The analysis assumes continuous measurement at a fixed rate, a known initially excited state and no spontaneous relaxation. Measurement efficiency describes how much of the measurement is captured in the available signal.

Within this model, Bugu proves that every predetermined drive leaves infinite record divergence over the complete, infinite observation period. Driving creates quantum coherence needed for rotation, but measurement dephases that coherence. Without feedback or relaxation, residual excitation leaves continuing evidence in the record.

That conclusion is specific to the full-history criterion. It does not mean every fixed pulse is useless over a finite observation window.

Feedback changes the situation because the controller can respond to the conditional state inferred from the signal. The paper constructs a two-phase strategy: first rotate the qubit close to the ground state, then watch the record and apply corrective pulses when it indicates a return toward excitation.

Evidence can arrive late

The proofs use quantum filtering, stochastic calculus and inequalities relating state motion to control effort. A central mathematical identity separates the record’s information cost into two contributions.

One is evidence written while the qubit still has excited-state population. The other comes from measurement backaction that the controller does not observe. This unrecorded component destroys coherence and leaves the conditional state mixed, or impure. Evidence associated with that impurity appears later, as continued measurement drives the conditional state back toward purity.

Counting only the rapid rotation would therefore miss part of the cost. The construction must also account for later monitoring and rare corrective restarts.

The main result gives an exact limiting constant, K, for the product of divergence and effort, divided by the measurement rate, as divergence approaches zero. K depends on measurement efficiency. Explicit feedback policies approach it while their divergence tends to zero and their effort grows without bound. Making the record increasingly hard to distinguish thus requires increasingly large effort in this idealized setting.

This does not establish the best product at every finite effort. For intermediate efficiencies, the paper brackets the unrestricted optimum within a factor of two but leaves open whether a finite-effort strategy can beat the high-effort constant.

Simulations, not a hardware demonstration

Monte Carlo simulations illustrate the proposed strategy. At the largest simulated feedback gain, the normalized products were approximately 3–6% above the limiting constant across three efficiencies. The simulations do not supply the proof. They also involve rare, costly restarts; numerical refinement checks support stability but do not rigorously bound discretization error.

Hardware limitations are substantial. The theory assumes instantaneous feedback and unlimited bandwidth. Under the paper’s illustrative superconducting-qubit parameters, strong drives can populate higher levels, invalidating the two-level model, while calculated rotations can be faster than the cited feedback latency. The paper describes the asymptotic regime as out of reach for present transmon parameters.

Finally, any nonzero relaxation makes the unrestricted minimum product vanish: an undriven qubit can relax with zero drive effort and finite divergence. That does not guarantee a nearly indistinguishable record. The low-divergence limit with relaxation remains unanswered. The result is therefore a precise theoretical benchmark for a restricted model, not a demonstrated way to erase a qubit’s recorded past.