Technical Summary: Overcoming Fourier Locking in Quantum Data Re-uploading Classifiers via Spectral Homotopy

1. Problem Statement

Data Re-uploading Parameterized Quantum Circuits (DRU-PQCs) are theoretically universal function approximators capable of modeling complex, non-linear decision boundaries by mimicking deep neural networks through sequential data injection. However, their high expressivity creates highly oscillatory, non-convex loss landscapes that resist gradient-based optimization.

While recent literature attributes training failures in these architectures to "barren plateaus" (where gradients vanish due to chaotic expressivity) or insufficient capacity, this paper argues that the primary bottleneck is a distinct structural failure mode termed Fourier Locking (FL). In DRU-PQCs, encoding weights (which select input frequencies) and entangling layers (which route quantum information) are non-linearly coupled. When initialized randomly on high-frequency targets, the optimizer often collapses the encoding parameters into isolated, spurious local minima. Once "locked," the encoding weights select a frequency spectrum misaligned with the target decision boundary. Consequently, the entangling layers cannot route a signal that effectively does not exist, causing the model to fail despite retaining full geometric sensitivity to its parameters.

2. Methodology and Theoretical Framework

Diagnostic Metrics

The authors propose that standard gradient magnitudes and parameter-space Quantum Fisher Information Matrix (QFIM) are insufficient for diagnosing FL. Instead, they utilize two specific Fisher diagnostics:

1. Input-Space Quantum Fisher Information ($F_x$): Measures the effective frequency content of the encoded state (the Fubini–Study susceptibility to input data). In a locked circuit, $F_x$ remains frozen at a misaligned value from initialization.
2. Fisher Discriminant Ratio (FDR): Measures the alignment of measured features with class labels. In a locked circuit, the FDR collapses toward zero, indicating a "class-degenerate readout" where the state oscillates with the input but fails to map to the labels.

Crucially, the authors demonstrate that trapped circuits retain a non-degenerate parameter-space QFIM (indicating the state remains responsive to parameters), proving the failure is one of spectral misalignment, not geometric insensitivity.

Experimental Design

The study employs a 4-qubit DRU-PQC with $L=2$ re-uploading layers and 50 trainable parameters (40 quantum, 10 classical). Experiments were conducted across 50 independent random seeds to map the bimodal probability space of the optimizer (trapped vs. escaped).

The core intervention is Spectral Homotopy, a curriculum learning protocol that stages the target frequency ($f$) from low to high ($f: 1.0 \to 2.0 \to 3.0$). This approach aims to convexify the early loss landscape, allowing the joint parameters to be transported into the target frequency band before high-frequency attractors materialize.

Control conditions included:

• Standard Training (STD): Direct training on the high-frequency target ($f=3.0$).
• Routing Ablations: Testing if pre-aligning entangling layers (Identity initialization or Coupled Warmup) could rescue a randomly initialized encoding state.
• Reverse Curriculum: Training from high to low frequency ($3.0 \to 1.0$) to test for irreversibility.
• Random Curriculum: Randomly ordering frequency phases to control for data diversity effects.

3. Key Results

Diagnosis of Fourier Locking

• Readout Collapse: Trapped circuits exhibited a Fisher Discriminant Ratio (FDR) collapsing toward zero ($r_{pb} = 0.912, p < 0.001$), confirming a failure of label alignment at the readout.
• Spectral Immobility: The defining signature of a trapped circuit was the lack of movement in its effective frequency content ($F_x$). Trapped circuits held a static, misaligned $F_x$ throughout training, whereas escaping circuits migrated their frequency content significantly ($r_{pb} = -0.48, p < 0.001$).
• Parameter Sensitivity: Contrary to the barren plateau hypothesis, trapped circuits did not show a collapse in the parameter-space QFIM trace, confirming the optimizer was paralyzed by spectral misalignment rather than a loss of geometric sensitivity.

Efficacy of Spectral Homotopy

• Escape Rate Improvement: Standard training on high-frequency targets yielded a 6% escape rate (final accuracy > 0.65). The forward spectral homotopy curriculum ($1.0 \to 2.0 \to 3.0$) tripled this rate to 18%.
• Mechanism of Action: Successful "escapees" under the forward curriculum showed a monotonic increase in $F_x$ in step with the curriculum phases, while trapped circuits remained spectrally static. The low-frequency phases did not improve accuracy immediately (accuracy remained at chance levels until Phase 3) but acted as a "latent scaffold," transporting the encoding weights into the correct spectral funnel.
• Routing Failure: Ablation studies showed that fixing the entangling layers (via Identity initialization or warmup) did not rescue performance. This confirms that the encoding initialization is the dominant variable; if the encoding weights commit to a wrong frequency, the entangling layers cannot route the signal.
• Irreversibility: The reverse curriculum ($3.0 \to 1.0$) resulted in a 0% escape rate. Even seeds that successfully learned the high-frequency boundary in Phase 1 suffered "catastrophic spectral forgetting" when forced into lower-frequency geometries, indicating the architecture cannot retain competing frequency representations once the high-frequency attractor is lost.

4. Significance and Claims

The paper reframes the optimization bottleneck in expressive DRU-PQCs. It posits that failure is not due to a lack of representational capacity or the chaotic flattening of the landscape (barren plateaus), but rather a Fourier Locking phenomenon where the encoding weights become structurally misaligned with the target frequency band.

The primary contribution is the identification of spectral mobility (the migration of $F_x$) as the critical diagnostic for successful training, distinguishing it from static parameter sensitivity. The study demonstrates that frequency pacing via spectral homotopy can bypass these locks by convexifying the early landscape, effectively acting as a scaffold to transport parameters into the target basin.

However, the authors maintain a modest stance regarding the solution's completeness. Despite the success of spectral homotopy, 82% of random initializations remained permanently trapped, indicating that while frequency pacing mitigates the issue, the "initialization lottery" remains a dominant factor. The paper concludes that random initialization is insufficient for reliably training expressive DRU-PQCs on high-frequency targets and suggests future work must focus on deterministic encoding initialization (e.g., using classical spectral analysis to pre-set weights) rather than relying solely on optimization curricula.