Task-Aligned Spectral Inductive Biases in Parameterized Quantum Circuits

Forum topic · 小凯 · 2026-10-20

Summary

This arXiv paper (2610.23834) by Felix Frohnert, Adrián Pérez-Salinas, Sebastian Lange, Marcus Schwaebisch, Evert van Nieuwenburg, and Vedran Dunjko investigates how data reuploading parameterized quantum circuits can provide task-aligned inductive biases for quantum machine learning. The circuits take two distinct inputs: a signal varying between samples and a fixed, task-defined spectral query specifying how the signal should be probed. The authors show that encoding the query modifies the circuit's hypothesis class over the signal by modulating how existing signal frequencies are combined, without adding new sample-dependent information. This mechanism naturally matches tasks based on nonuniform Fourier transforms, where evaluation frequencies define the query and sampling positions guide its encoding. The connection is demonstrated on a synthetic aperture radar (SAR)-inspired classification task, where acquisition geometry determines the query for classifying unfocused spectral segments before imaging. Experiments on small-scale spectral and radar-like tasks show that aligned queries yield better learning with limited data compared to parameter-matched circuits with mismatched or missing queries. The results offer a path for designing and validating inductive biases in quantum circuits from known spectral task structure.

Paper Overview

Research area: Machine Learning (Quantum Machine Learning) Authors: Felix Frohnert, Adrián Pérez-Salinas, Sebastian Lange, Marcus Schwaebisch, Evert van Nieuwenburg, Vedran Dunjko Published: 2026-10-16 arXiv: 2610.23834

Abstract

Parameterized quantum circuits are widely investigated in quantum machine learning, yet how their architecture can provide task-aligned inductive biases remains poorly understood. We study data reuploading quantum circuits with two distinct inputs: a signal that varies between samples and a fixed, task-defined spectral query that specifies how the signal should be probed. We show that encoding the query can change the circuit's hypothesis class over the signal by modulating how its existing signal frequencies are combined, without introducing new sample-dependent information. This mechanism matches tasks built on nonuniform Fourier transforms, whose evaluation frequencies specify the query and whose sampling positions guide its encoding.

We examine this connection in a classification task inspired by synthetic aperture radar, where the acquisition geometry determines the query used to classify unfocused spectral segments before imaging. On small-scale spectral and radar-like tasks, circuits with aligned queries learn better with limited data than parameter-matched circuits with mismatched or missing queries.

Key Contributions

Identifies a mechanism by which encoding a fixed spectral query reshapes the hypothesis class of a data reuploading quantum circuit over the signal input, without adding sample-dependent information.
Establishes a natural correspondence between such circuits and tasks based on nonuniform Fourier transforms: evaluation frequencies define the query, sampling positions guide its encoding.
Demonstrates the advantage empirically on a SAR-inspired classification task where acquisition geometry dictates the spectral query.
Provides a pathway for designing and verifying inductive biases in parameterized quantum circuits from known spectral task structure.

Takeaway

When the spectral query encoded into a quantum circuit matches the frequency structure of the underlying task (e.g., nonuniform Fourier-based imaging), the circuit exhibits a beneficial inductive bias, improving learning performance especially in limited-data regimes.

---

*Auto-collected on 2026-10-20.*