Designing a quantum sensor by shaping its answer

A sensor often has to do more than measure. If its task is to sort an unknown phase into one of two classes, estimating the phase first and classifying it afterward may be an unnecessary detour. Could the quantum dynamics be designed to produce the decision directly—and could that design be found reliably rather than by a lucky choice of starting controls?

The paper studies quantum computational sensing for binary phase classification. The unknown phase lies within one 2π period, and each class has a probability distribution over possible phases. A single-qubit generalized quantum signal processing (GQSP) sequence makes a fixed number, Q, of queries to the phase-dependent signal, with adjustable rotations between queries. Measuring the qubit gives a probability p(θ) of reporting class 1. The goal is to choose the sequence that minimizes the average classification error.

The key change is what gets optimized. Directly adjusting all the rotations creates a nonconvex problem: a numerical optimizer can depend on its initial settings and may settle on a poor solution. Instead, the authors optimize p(θ), the sensor’s response as a function of phase. For Q queries, GQSP restricts that response to a trigonometric polynomial with degree at most Q, while quantum probabilities require it to stay between zero and one. The classification error is linear in this response.

Those ingredients lead to an exact finite-dimensional semidefinite program (SDP). Positive-semidefinite matrices encode both p(θ) and its complement, enforcing the probability bounds across the entire phase range—not merely at sampled points. The paper proves that every response satisfying these constraints can be implemented by a Q-query GQSP sequence. After solving the SDP, the researchers use Fejér–Riesz factorization and GQSP synthesis to recover physical control rotations. Thus the convex solution is not just a useful approximation: it is globally optimal within the stated fixed-query model.

The framework also accommodates labeled training samples when the phase distributions are not known, by replacing the expected error with empirical error. For additive phase uncertainty shared across signal interactions, the method modifies the optimization weights according to the noise distribution’s Fourier components. In the Gaussian case, higher-frequency components are attenuated more strongly, so the redesign can account for the loss of fine response features.

The evidence is numerical, using two constructed classification tasks and query budgets from 2 to 20. In one reported task at Q = 16, the SDP’s error was 3.33% compared with 3.99% for the best of 64 direct angle-optimization starts; in the interleaved task, the respective errors were 12.52% and 13.92%. The authors report that the direct optimizer’s results varied more across random initializations as Q increased, even though good angle settings remained attainable. Their SDP took 3.51 seconds in the listed Q = 16 design-time comparison, versus 88.37 seconds for the multi-start angle search; these timings describe their particular implementation and setup, not a universal speed guarantee.

When they add Gaussian phase uncertainty, noise-aware redesign increasingly improves on a design that ignores the uncertainty as noise grows. These comparisons are simulations, not a laboratory demonstration. They depend on the assumed GQSP architecture and task distributions; sample-based design also depends on the available labeled data. The result establishes an exact optimization route for the paper’s model, not a claim that real quantum sensors already classify better or that hardware noise and implementation costs have been resolved. The work is a 2026 arXiv preprint, not evidence of peer review.