Toward Joint Optimization of Circuit Depth and Training Data Size in Adaptively Grown Quantum Classifiers

What it claims

On raw 784-pixel MNIST 3-vs-5, Q-FLAIR's converged circuit size T\ does not track training-set size N. Across N = 2000–10000, mean T\ ranges from 16 to 23 gates and mean test accuracy from 84% to 89%, neither moving monotonically. The authors' explanation is that with 784 candidate features, some gate clears the fixed loss-improvement threshold at almost any N. This decouples the stopping decision from how much data is available.

One clarification on the summary's "14 of 15": the 15th run (N=10000) did not violate the Caro et al. bound. Fine-tuning moved no weights, so K = 0 and the bound is trivially zero. The authors therefore excluded that run as "not applicable."

How it works

Phase 1, growth (Q-FLAIR reimplementation):

• Uses Q-FLAIR's exact gate pool: Rx, Ry, Rxx, Ryy (fixed and data-encoding variants) plus H.
• Runs on a 5-qubit ring, which is Q-FLAIR's own QNN qubit ceiling.
• Each candidate gate's output is reconstructed as a closed-form sinusoid of its rotation angle, using 2 extra quantum evaluations.
• The loss-minimizing weight and feature are found classically. At d = 784, brute-force grid search over features is too slow, so they use SciPy's bounded scalar minimizer (tolerance 10⁻³). This cuts classical cost by about 10×.
• Growth stops when the best candidate improves training loss by less than ΔL = 10⁻³. There is no cap on circuit size.

Phase 2, joint fine-tune (new in this paper):

• Q-FLAIR freezes each gate's weight once found, so K is undefined on its own.
• The authors warm-start from the grown circuit and run Adam over all weights, with mini-batches and early stopping.
• K is the number of gates whose weight moves by more than δ = 0.05.
• Whichever weights (original or fine-tuned) give lower validation loss are kept, so Phase 2 can never report a worse model than Phase 1.

The bound reported is the full Caro et al. Theorem 3 form: min_K[√(K log(MT)/N) + Σ_{k>K} MΔ_k].

Evidence

Results are means over 3 seeds per N, each evaluated on the same 1000-example held-out test set:

• The seed standard deviation in T\* (up to ±9 gates) is comparable to the whole spread across N.
• Across the 14 applicable runs, the gap correlates with the bound at r = 0.12. Raw gate count T alone gives r = 0.05, so K adds almost nothing as a predictor.
• Setup:
• Simulation is exact statevector in PennyLane.
• Each run takes minutes to tens of minutes.
• Pixels are rescaled to [−π, π] using training-set statistics.
• The loss is negative log-likelihood.
• The decision threshold comes from an ROC-style search that maximizes balanced accuracy on the training set.

Limits

• Single task: only one dataset and task. Q-FLAIR's QSVM variant is untested.
• Low statistical power: 3 seeds per N means a trend smaller than the seed variance would go undetected.
• Untuned Phase 2: the fine-tune learning rate and δ were each chosen once and never varied. Other values could change K and the r = 0.12 result.
• No code release: the page does not mention a repository.
• Proposed fix is untested: the authors suggest a stopping threshold that scales with both N and candidate-pool size, such as a multiple-comparisons correction. They do not run it. The high-dimensionality explanation for the missing trend is a hypothesis, not something they test, for example by reducing the number of features.