Reproduce “Representation Learning with Quantum Signal Processing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junqi Wang, Junyu Liu
This research utilizes Quantum Signal Processing (QSP) as an analytically solvable model to rigorously prove representation learning in quantum systems. It demonstrates that QSP's quantum neural tangent kernel (QNTK) evolves dynamically during training, moving beyond the frozen-kernel approximation, and reveals distinct learning behaviors based on data density.
Key insights
- The diagonal QNTK entries remain non-self-averaging even at large depths (coefficient of variation converges to 1/3), showing that relative kernel fluctuations persist despite the underlying QSP unitary approaching Haar randomness.
- The ensemble-mean QNTK exhibits a spectral bias in the Chebyshev basis, with its eigenvalues revealing an effective spectral bandwidth of $O(\sqrt{d})$, motivating the definition of a sampling density parameter $w = n/\sqrt{d}$.
- Representation learning dynamics transition from independent scalar flows in sparse data regimes to coupled evolution in denser regimes, where the evolving QNTK reshapes shared features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.