How Quantum Circuits Avoid Spurious Local Minima via Homotopy

Researchers have discovered that quantum circuits can become frozen during training, not due to a lack of sensitivity, but because of a misalignment of their frequency content. In two independent experiments with fifty different initializations each, circuits initialized on high-frequency targets collapsed into spurious local minima, remaining locked throughout the run, while circuits that escaped migrated their frequency content. The work demonstrates that standard methods for diagnosing optimization failures miss this, as trapped circuits retain a fully non-degenerate parameter-space QFIM. A frequency-staged homotopy protocol that paces the target frequency tripled the escape rate from these frozen states, increasing it from 6% to 18%, and revealed that Fourier locking is a frequency-alignment problem, and its remedy is frequency pacing.

Data Re-uploading Circuits and Universal Expressivity

Initializing quantum circuits on high-frequency data targets can lead to complete optimization failure, a phenomenon researchers are calling Fourier locking. Unlike typical barren plateaus where optimization simply stalls, these circuits become entirely frozen during training, unable to learn despite possessing theoretical expressivity. This isn’t a matter of insufficient capacity, but a fundamental structural issue within the circuit itself. Researchers measured the parameter-space Quantum Fisher Information Matrix (QFIM) throughout training and found it did not collapse in trapped circuits; the locked model remained fully sensitive to its parameters. Instead, a misalignment of the circuit’s frequency content, as measured by the input-space quantum Fisher information, is the core of the problem. The team identified two quantities that indicate this issue: the Fisher discriminant ratio, which measures label alignment, and the input-space quantum Fisher information, which reveals the effective frequency content of the encoded state.

Trapped circuits exhibit a collapsed Fisher discriminant ratio, indicating a failure of label alignment at the readout. A frequency-staged homotopy protocol that paces the target frequency convexifies the early loss landscape, allowing circuits to escape the frozen state. The escape rate tripled, increasing from 6% to 18%. Escaping circuits migrate their frequency content over training, while those trapped remain at a static, misaligned frequency from initialization onward.

Fourier Locking as an Optimization Bottleneck

This spectral mobility is the replicated signature, not any endpoint value, and trapped circuits retain a fully non-degenerate parameter-space QFIM; 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 convexifies the early loss landscape; escaping circuits progressed in step with the curriculum, and the escape rate tripled (18% vs. 6%). Fourier locking is a frequency-alignment problem, and its remedy is frequency pacing. The Fisher discriminant ratio (FDR) of the measured features collapses when a circuit locks, indicating a failure of label alignment at the readout.

The input-space quantum Fisher information, the Fubini, Study susceptibility of the state to the encoded data, identifies its dynamical signature: the effective frequency content of a trapped circuit is frozen at a misaligned value from initialization onward, while escaping circuits migrate theirs over training. This paper demonstrates that standard gradient magnitudes fail to diagnose Fourier locking traps, and that natural quantum-geometric measures also fall short. Initial investigations revealed that circuits, in repeated experiments with fifty different initializations each, could remain entirely frozen for the entire training run, a complete stagnation rather than gradual progress.

Fisher Diagnostics of Fourier Locking

Spencer Topel at Moth, a Brooklyn-based organization, is developing new methods for diagnosing optimization failures in data re-uploading parameterized quantum circuits (DRU-PQCs). Researchers there have identified a critical bottleneck not stemming from insufficient circuit capacity, but from a phenomenon termed Fourier locking (FL). This occurs when encoding weights and entangling layers become misaligned, causing the circuit to collapse into spurious local minima during training. Two key diagnostics carry the diagnostic signal, moving beyond simple gradient analysis. Researchers measured the parameter-space Quantum Fisher Information Matrix (QFIM) throughout training and found it did not collapse in trapped circuits; the locked model remained fully sensitive to its parameters. Instead, the input-space quantum Fisher information, the Fubini, Study susceptibility of the state to the encoded data, revealed a crucial signature.

Trapped circuits remained frozen for the entire run, while escaping circuits migrated their frequency content. The Fisher discriminant ratio (FDR) of the measured features collapses when a circuit locks, indicating a failure of label alignment at the readout. In two independent experiments with fifty different initializations each, circuits could remain entirely frozen for the entire training run, a complete stagnation rather than gradual progress. Importantly, trapped circuits retained a fully non-degenerate parameter-space QFIM, confirming the issue is spectral misalignment of a responsive state, not a loss of geometric sensitivity. This detailed analysis reframes the optimization challenge, shifting focus from parameter tuning to spectral mobility.

Initial assumptions about quantum circuit optimization often center on a lack of capacity, yet recent work reveals a more nuanced failure mode. The core of this misalignment lies in the circuit’s frequency content. This understanding prompted the development of a surprisingly simple solution: a frequency-staged homotopy protocol that paces the target frequency. This approach convexifies the early loss landscape, allowing circuits to escape the frozen state.

Data re-uploading parameterized quantum circuits, or DRU-PQCs, possess a theoretical capacity for complex pattern recognition, yet consistently encounter optimization challenges during training. Recent work clarifies that these failures aren’t simply due to insufficient circuit capacity, but a structural failure mode termed Fourier locking (FL). This phenomenon arises from the nonlinear coupling between encoding weights, which select input frequencies, and entangling layers responsible for routing quantum information. Initializing circuits on high-frequency targets, researchers found, causes encoding parameters to collapse into spurious local minima, effectively halting learning. In two independent experiments with fifty different initializations each, the locking is literal: trapped circuits remained frozen for the entire run, while escaping circuits migrated their frequency content. The team identified two quantities that carry the diagnostic signal. Conversely, escaping circuits migrate their frequency content throughout training.

Researchers at Moth demonstrate that the primary obstacle isn’t insufficient circuit capacity, but a structural failure they call Fourier locking (FL). In two independent experiments with fifty different initializations each, the locking is literal: trapped circuits remained frozen for the entire run, while escaping circuits migrated their frequency content. A frequency-staged homotopy protocol that paces the target frequency convexifies the early loss landscape; escaping circuits progressed 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. The replicated signature is spectral mobility, not any endpoint value, and trapped circuits retain a fully non-degenerate parameter-space QFIM; the failure is spectral misalignment of a responsive state, not a loss of geometric sensitivity. The Fisher discriminant ratio (FDR) of the measured features collapses when a circuit locks, indicating a failure of label alignment at the readout. The input-space quantum Fisher information measures the effective frequency content of the encoded state. Conversely, escaping circuits migrate theirs over training.