Paper review · October 10, 2026
Toward Joint Optimization of Circuit Depth and Training Data Size in Adaptively Grown Quantum Classifiers
Running the Q-FLAIR circuit-growing method on MNIST while varying the training data size, the authors found no predictable relationship between data size and final circuit size.
Problem
Quantum models require deciding together how complex to make the circuit and how much training data to use. Caro et al. gave a bound showing that fewer trainable gates allow generalization from less data, and Q-FLAIR grows the circuit gate by gate and stops when loss improvement stalls. Whether combining the two results yields a scaling law for circuit size as a function of data size had not been checked.
Method
The authors reimplemented Q-FLAIR's gate pool, analytic reconstruction and stopping rule as-is, using bounded scalar optimization instead of grid search for selecting among 784-dimensional features. They classified MNIST 3 vs. 5 on the raw 784 pixels with a 5-qubit ring structure, running 15 times: 3 seeds for each of five sizes from N=2000 to 10000. After growth, all weights were fine-tuned with Adam, and the number of gates K whose weights changed by more than 0.05 was measured and compared with the Caro et al. bound.
Results
- Final circuit size ranged from 16 to 23 gates and test accuracy from 84% to 89%, varying non-monotonically with N.
- The standard deviation across seeds was up to about 9 gates, comparable to the differences across the full range of N.
- In all 14 runs, excluding 1 run where fine-tuning brought no improvement, the actual generalization gap did not exceed the Caro et al. bound.
- The correlation between the generalization gap and the bound value was weak at r=0.12, and not meaningfully better than using only the total gate count T (r=0.05).
Limitations
It uses a single dataset and single task with only 3 seeds per N, so trends smaller than the seed variance cannot be detected. The fine-tuning stage, its learning rate and the threshold δ were added in this study and only one value was tested; K and the correlation results may change with other values. Generalization to QSVM variants or other high-dimensional inputs was not checked. All circuits were run only in simulation.
Why it matters
It shows that even when a theoretical generalization bound holds, it may do little to predict actual performance. The authors argue that a large search space, such as 784 candidate features, makes the stopping rule independent of data size, and propose that a stopping rule reflecting data size and the number of candidates is needed.
Key terms
- Q-FLAIR
- A method that grows a quantum feature map circuit from an empty state by adding one gate at a time, stopping when the improvement in training loss falls below a threshold.
- Active gate count K
- In the Caro et al. generalization bound, the number of gates that changed significantly during training. If it is smaller than the total gate count T, it gives a tighter bound.
- Generalization Gap
- The difference between performance on training data and performance on unseen test data.
- Data Re-uploading
- A design that feeds input data repeatedly into multiple layers of a circuit to increase the expressive power of a quantum model.
Original
Authors: Saeefa Rubaiyet Nowmi, Md Mahmuduzzaman Kamol, Mohammad Saidur Rahman · Posted 2026-10-08 · License CC BY-NC-ND 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.