Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper identifies a critical optimization challenge in data re-uploading parameterized quantum circuits (DRU-PQCs) known as Fourier locking (FL), which arises from the nonlinear coupling of encoding weights and entangling layers, leading to spurious local minima during gradient-based optimization. Through two independent experiments, the authors demonstrate that circuits affected by FL exhibit frozen input-space quantum Fisher information, while those that escape show significant spectral mobility, indicating that the optimization failure is due to spectral misalignment rather than insufficient capacity. By implementing a frequency-staged homotopy protocol that adjusts the target frequency, the authors successfully convexify the loss landscape, resulting in a tripling of the escape rate from FL, thus providing a novel solution to this optimization bottleneck.
Fourier locking in quantum circuits can freeze optimization, but a clever frequency pacing strategy can triple escape rates from local minima.
Data re-uploading parameterized quantum circuits (DRU-PQCs) are universal function approximators, yet their expressivity produces oscillatory, non-convex loss landscapes that resist gradient-based optimization. We show that the primary optimization bottleneck in DRU-PQCs is not insufficient capacity but a structural failure mode we term Fourier locking (FL): because encoding weights and entangling layers are nonlinearly coupled, random initialization on high-frequency targets collapses the encoding parameters into spurious local minima. Two Fisher diagnostics characterize FL. The input-space quantum Fisher information $F_x$ measures the effective frequency content of the encoded state; the Fisher discriminant ratio of the measured features measures their alignment with the class labels. In two independent 50-seed experiments, the locking is literal: trapped circuits hold $F_x$ frozen for the entire run, while escaping circuits migrate their frequency content (direct training: $r_{pb} = -0.48$; curriculum: $d = 1.34$; both $p < 0.001$). The replicated signature is this spectral mobility, not any endpoint value of $F_x$, and trapped circuits retain a fully non-degenerate parameter-space QFIM ($r_{pb} \approx 0$): the failure is spectral misalignment of a responsive state, not a loss of geometric sensitivity. A frequency-staged homotopy protocol that paces the target frequency ($f: 1.0 \to 3.0$) convexifies the early loss landscape; escaping circuits raise $F_x$ in step with the curriculum, and the escape rate triples (18% vs. 6%). Fourier locking is a frequency-alignment problem, and its remedy is frequency pacing.