A mathematical study of a quantum search problem finds that a square-root measurement can approach the best possible collective strategy for pinpointing a single temporary change, even as the sequence grows. The result holds when the interval length is known and when it is unknown, although the two settings settle on different large-sequence benchmarks. The task is exact labeling: the measurement must identify where the changed stretch begins and ends. The document is a preprint, arXiv:2608.24543v2, in the quantum-information category and dated 3 Sep 2026.
The problem is exact location
The setup is deliberately specific. A known anomalous pure state occupies one nonempty contiguous interval in an otherwise known sequence of reference states. An unrestricted collective measurement can use the full sequence together and must return the exact interval label. The comparison is between the minimum-error strategy that is optimal under the stated prior and the square-root measurement, or SRM. The analysis combines Gram-matrix Toeplitz comparison, Følner transfer, and exact decompositions by excitation number and interval hull.
The limit depends on what is known
When the interval length is fixed in advance, the optimal measurement and SRM converge to the same success probability as the number of possible placements grows. The theorem covers every fixed length of at least one site and every squared local overlap from zero to one. For a fixed length and an overlap strictly between zero and one, the nonnegative gap between optimal and SRM success shrinks at inverse-square-root order in the number of placements. The theorem gives an order of decay, not a finite-size constant.
If the known interval itself grows, the same agreement survives. When both the interval length and the number of allowed placements grow without bound, the two probabilities converge to a one-dimensional benchmark evaluated at the squared local overlap, with no condition on which grows faster. The unknown-length model changes the prior as well as the label set: every nonempty interval is assigned equal weight, rather than every length. Under that rule, the average interval length is one-third of the sequence length plus two-thirds.
Unknown length makes the geometry harder. The physical state-overlap matrix is not globally two-dimensional Toeplitz, because for disjoint intervals the separating gap changes the relationship between endpoint displacement and the sites on which the intervals differ. The authors handle that mismatch with comparison and transfer arguments. Even so, under the uniform interval prior, both optimal and SRM success converge to the square of the one-dimensional benchmark evaluated at the local overlap.
A harder boundary
The analysis also follows a boundary case in which the anomalous and reference states become increasingly alike as the sequence grows. In a compact critical window, the normalized success probabilities are approximated by a limiting function whose input tracks sequence length, the loss of overlap, and the square of the logarithm of sequence length. The paper gives a uniform error of order the logarithm of the logarithm of sequence length divided by the logarithm of sequence length in that window. If the boundary scale settles to a fixed value, the normalized trace benchmark, SRM result and optimum converge to the same limiting function at that value.
A related hull-dominance condition gives another route to the unknown-length result. When the overlap approaches one from below while the scaled one-dimensional benchmark diverges, the trace benchmark, SRM and optimal probabilities are each asymptotic to the squared benchmark. The condition is sufficient in the paper's analysis; whether it is also necessary remains open.
Adding a no-change option
Allowing the possibility of no change produces a weighted version of the same picture. With a fixed no-change prior strictly between zero and one, the best joint success in detecting a change and naming its interval approaches the no-change prior plus the anomalous prior multiplied by the relevant conditional localization limit. The limit is the fixed-length benchmark for a known interval, the one-dimensional benchmark at squared overlap for a growing known length, and the squared benchmark for unknown length. In the two growing-length models, a prior-weighted SRM reaches that joint limit when the overlap is fixed below one.
At unit overlap, where the physical states are identical, the edge case is different. The exact weighted SRM success with K competing labels is the squared no-change prior plus the squared anomalous prior divided by K, while the Bayes-optimal rule selects whichever label has the larger prior. As the number of labels grows, the weighted SRM tends to the squared no-change prior, whereas the joint optimum tends to the no-change prior. A final most-probable-label step, called MAP postprocessing, is asymptotically Bayes optimal in this case.
Checks, and the limits of the model
Finite calculations were used as checks on the analysis. Across 30 unknown-length cases, the largest reported raw primal-dual gap was 6.782 times 10 to the minus ninth, the largest postprocessed bracket width was 1.025 times 10 to the minus eighth, and the largest rechecked completeness residual was 1.41 times 10 to the minus fifteenth. The authors describe these as IEEE-double floating-point diagnostics, not rigorous enclosures.
The limits come with a narrow scope. The model assumes conditionally independent pure-state outputs, known states, one nonempty returning interval, uniform priors and unrestricted collective measurements. Mixed or correlated outputs, unknown anomaly states, imperfect returns, restricted measurements and multiple intervals are outside the stated analysis and would require separate work. The results are therefore benchmarks for a calibrated theoretical ensemble, not evidence that the same performance has been established for other measurement settings.
Paper data and sources
Original title: Quantum Change Intervals: Exact Asymptotic Localization with Collective Measurements
Authors: Xu Chen, Xue Ma
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text