No Reference-Free Generalization in Quantum Machine Learning

The Challenge of Reference-Free Generalization

Quantum machine learning (QML) is often promoted for its ability to leverage the exponentially large Hilbert space of quantum systems. However, this paper argues that the mere existence of a large state space does not guarantee generalization. The author identifies a fundamental "identifiability problem": if a learner is provided with training data but no external reference frame—such as a fixed measurement basis, a specific feature map, or a known Hamiltonian—it lacks the orienting structure necessary to assign different meanings to unseen quantum directions.

The Symmetry Constraint

The paper defines a "reference-free" learner as one whose predictions are invariant under arbitrary unitary rotations of the training data. Because the learner has no external reference, it must respect any unitary symmetry that the training data itself leaves unbroken. The author proves that if the training states do not span the entire Hilbert space, the learner is forced to treat all pure states in the orthogonal complement as equivalent. Consequently, these states must receive the same prediction, even if they are perfectly distinguishable to an observer who possesses a calibrated measurement device.

Implications for QML Practice

This finding demonstrates that the limitation is not a result of insufficient computational power, poor optimization, or statistical noise, but rather a structural consequence of missing reference information. The paper concludes that successful generalization in QML is essentially a form of symmetry breaking. To achieve meaningful predictions on unseen data, a model must utilize "operational resources"—such as feature maps, locality priors, or specific measurement bases—that provide the necessary physical structure to distinguish between quantum directions. The author emphasizes that Hilbert-space dimension alone is not a learnable feature space; rather, it is the physical structure imposed by the model that gives unseen directions semantic meaning.