arXiv:2610.24426
Margin-Based Generalization for Quantum Machine Learning with Classical Data Encoding
Ju-Young Ryu, Tak Hur, Daniel K. Park
Abstract
Understanding generalization is essential for assessing the practical viability of quantum machine learning. In many applications, quantum models operate on classical data that must first be encoded into quantum states, making the encoding part of the effective hypothesis class. We develop a margin-based generalization framework for this setting, covering both fixed and trainable quantum data encodings. For fixed encodings, the existing margin bound transfers unchanged. For trainable encodings, we derive a bound using a two-scale covering argument that combines the original margin-dependent term with an additive encoding-capacity term of order $\tilde O(\sqrt{L/m})$, where $L$ is the number of trainable encoding parameters and $m$ is the training-sample size. Across 19,440 simulated training runs spanning various image datasets, circuit architectures, and optimization settings, the inverse margin performs best among five baseline and two bound-derived complexity measures under conditional-independence and 90th-percentile sign-error criteria, within bootstrap uncertainty. These results provide a unified margin-based view of generalization in quantum machine learning with classical data.
AI Summary
The paper studies margin-based generalization bounds for quantum neural networks classifying classical data, focusing on the capacity introduced by trainable quantum encodings. Fixed encodings reduce directly to the existing quantum-input setting. For an encoder with $L$ trainable parameters, it derives, uniformly over hypotheses, $R(h)\le\widehat R_\gamma(h)+\widetilde O((bE/\gamma)\sqrt{n/m}+\sqrt{L/m}+\sqrt{\ln(2/\delta)/m})$, where $b$ bounds the post-encoding unitary class, $E$ characterizes the measurement, and $m$ is sample size. The proof combines separate covers of encoder parameters and unitary images with a Lipschitz estimate and Dudley’s entropy integral. Absorbing all trainable gates into the encoder gives a parameter-count bound with only logarithmic margin dependence, $O(\sqrt{(L/m)\ln(L/\gamma)})$ for Pauli rotations and fixed-size projective measurements. An empirical study trains 19,440 simulated models on three binary image-classification tasks, reduced to two PCA features, using 4–8 qubits. Inverse median training margin outperforms parameter-count and theorem-derived measures under conditional-independence and 90th-percentile sign-error criteria. Its mean sign-error is 0.480 and 90th percentile 0.712, versus 0.504 and 0.772 for parameter count. A fitted additive capacity term changes the latter score by only 0.0007, within bootstrap uncertainty. Empirical margin loss at fixed thresholds predicts better than inverse median margin. The stated scope excludes general classical neural encoders; broader data–parameter coupling requires additional assumptions. Hardware noise remains untested. Small generalization gaps can be comparable to finite-test-set fluctuations, and fitted composite weights risk in-sample optimism. Bootstrap uncertainty is also limited by five training seeds.
Central claim may not hold: The sign-error weighting incorrectly squares a potentially negative concentration-bound expression, giving maximal reliability weight at zero estimated gap difference and positive weight to differences below approximately 0.01780, so its claimed reliability filtering and the empirical comparisons relying on it are not established as written.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 22 pages, 6 figures, 7 tables