Work

Erasure Check Timing from an Inferred Structured Bath: Dual Rail Qubits with Near Field Memory

Year: 2026
Type: preprint
Abstract: A schedule of erasure checks built from error rates cannot know when the near field of a qubit exchanges the excitation with it, because a rate carries no timing and no phase. The erasure check of a dual rail qubit returns one bit: the excitation is on the rails, or it is not. Here the near field of a dual rail superconducting qubit is kept as a finite quantum bath and inferred from the qubit's own calibration signal before any check is made, by Fourier features, principal components and gradient boosted regression, followed by a fit of the bath model to the signal. The inferred bath is a map of the exchange: it predicts when the excitation is in the near field and when it has returned to the rails, and the erasure check, a correcting rotation and the decoder weights are set from that prediction. With three calibration inputs all six parameters of the layered bath are recovered exactly (R^2 = 1.0000) on 49 unseen baths, for both ways of attaching the rails, and the check times and error channels built from the inferred bath equal those of the exact bath. In a rotated surface code at distance 7, with every data qubit in its own unseen bath, the gain over a clocked check given a perfectly calibrated frame grows with the wait. The logical error rate is lower by factors of 2.3 and 3.9 in the bit flip memory and 1.6 and 1.8 in the phase flip memory at clock periods 1 and 2, by factors of 108 and 33 at period 4, and from period 6 to period 50 the check placed from the inferred bath gives 0 to 9 logical errors in 20000 shots where the clocked check gives 1500 to 9900. With the two rails attached to different nodes of the near field the factors are 8.8 and 8.9 at period 4 and 36 and 33 at period 6, and at periods 25 and 50 there are 3 to 8 logical errors in 20000 shots against 7200 to 9000. Under shot noise, readout error, preparation error and drift of the bath at the level of a hardware calibration (1000 to 10000 shots per point), with the contrast and the preparation error left free in the fit, the six parameters are recovered with R^2 of 0.906 or higher and the logical error rate equals that of the exactly known bath within the sampling scatter on every rung. In a second part every link of the near field is drawn independently: all eleven bath parameters, the two horizontal bonds of the second layer included, are recovered exactly in 48 of 49 unseen baths, and a check placed from this map at clock periods 12 to 50 gives 0 to 12 logical errors in 20000 shots at distance 7, against 5400 to 9400 with the clocked check; with the vertical links alone the logical error rate at periods 25 and 50 is 0.03 to 0.42. The defects of the near field do not move at the operating temperature, so the map is taken once per cooldown. No hardware and no checks are added.
Language: English
Open Access status: green
OpenAlex ID: W7235389745
Cites: 2
Cited by:
Related to: 10