Approximation and Generalization Capacities of Parametrized Quantum Circuits for Functions in Sobolev Spaces

Alberto Manzano, David Dechant, Jordi Tura, Vedran DunjkoQuantum 2025 · DOI

TL;DR

This paper extends the theoretical understanding of how well parametrized quantum circuits (PQCs) can approximate functions, moving beyond the standard L2 distance to cover broader function spaces like Sobolev spaces and using more general distances. The authors also establish novel generalization bounds that connect different distance metrics and loss functions, providing a more robust theoretical foundation for using PQCs in machine learning tasks.

What Problem Does This Paper Address?

Parametrized quantum circuits (PQCs) are increasingly used as powerful models in quantum machine learning (QML), analogous to neural networks in classical ML. A key question is their "expressivity" – their ability to approximate various functions. While previous work has shown PQCs can approximate functions in the L2 space (square-integrable functions) with respect to the L2 distance, this has limitations. Specifically, L2 approximation doesn't guarantee good performance at all points, which is crucial for applications like risk assessment or solving differential equations. Furthermore, the theoretical understanding of PQC approximation capabilities in other function spaces (like continuous functions or Sobolev spaces) and under different distance metrics (like the supremum or L_p norms) has been less explored. This paper aims to fill these gaps by providing theoretical guarantees for PQC approximation in these more general settings and by developing generalization bounds that connect different distance and loss functions.

Methods

The paper builds upon existing work showing that PQCs can represent generalized trigonometric series. The authors extend this by proving that PQCs can approximate functions in:

  • L_p spaces: For 1 < p < ∞, demonstrating that PQCs can approximate functions with respect to the L_p norm.
  • C⁰ space: They show PQCs can uniformly approximate continuous functions on compact domains.
  • Hᵏ Sobolev spaces: This is a significant extension, proving PQCs can approximate functions and their derivatives up to order k in the Hᵏ norm.

A key technique for achieving these broader approximation capabilities is min-max feature scaling of the input data. This normalization strategy, particularly mapping inputs to the [-1, 1] range, is shown to significantly enhance PQC expressivity.

For generalization bounds, the authors introduce novel connections between different distance metrics and loss functions. They propose a new loss function, l<h1>, which incorporates derivatives, and show it is consistent with the supremum distance (C⁰ distance) in one-dimensional problems. They also establish generalization bounds that relate empirical risks (computed from data) to true risks (in terms of target distances) for various function spaces and distances, including L_p and Hᵏ spaces.

Datasets

Not discussed in the paper. The paper focuses on theoretical analysis and uses synthetic functions for numerical experiments.

Key Insights and Ideas

  • Beyond L²: The paper demonstrates that PQCs are not limited to L² approximation. By carefully choosing the PQC architecture and, crucially, by normalizing the input data, PQCs can approximate functions in Lₚ, C⁰, and Hᵏ spaces.
  • Data Normalization is Key: The min-max feature scaling of input data (mapping to [-1, 1]) is highlighted as a critical factor in unlocking the broader approximation capabilities of PQCs. This suggests that how data is preprocessed has a profound impact on what a PQC can learn.
  • Connecting Distances and Loss Functions: The development of generalization bounds that bridge different distance metrics (e.g., Lₚ, C⁰, Hᵏ) and loss functions is a major contribution. This is crucial for practical machine learning, as it helps us understand how minimizing a computable loss function leads to good performance in terms of a desired, often intractable, target distance.
  • Derivative Information Matters: The introduction of loss functions that incorporate derivative information (like l<h1>) is shown to be beneficial for approximating functions in higher-order spaces (Hᵏ) and for achieving better generalization, especially when such derivative data is available.

Key Results

  • Theoretical Approximation Guarantees: PQCs are proven to approximate functions in Lₚ (1 < p < ∞), C⁰, and Hᵏ Sobolev spaces under corresponding distances.
  • Improved Expressivity via Normalization: A min-max feature scaling strategy (mapping inputs to [-1, 1]) is shown to enable PQCs to approximate a wider range of function spaces.
  • Generalization Bounds: Novel bounds are derived that connect empirical risks to target distances across different function spaces and loss functions, including those that utilize derivative information.
  • Consistency of Loss Functions: The paper shows that certain loss functions (e.g., l<h1>) are consistent with more stringent distances (e.g., C⁰ distance) than standard L² loss functions, leading to better generalization.

Limitations and Future Work

  • Curse of Dimensionality: The authors acknowledge that as the input dimension increases, the complexity of the loss functions (especially those involving derivatives) grows exponentially, leading to a "curse of dimensionality." This implies that for high-dimensional problems, a significant amount of data for partial derivatives is required.
  • Data Requirements for Derivatives: The proposed loss functions that utilize derivative information require access to training data containing these derivatives, which may not always be available.
  • Practical Thresholds: While theoretical bounds are provided, the paper suggests further research into the practical thresholds at which these techniques might fail, particularly concerning the input dimension and the number of training data points.
  • Alternative Normalization Strategies: The paper mentions that other normalization strategies (like the classical [-1, 1] normalization for input data) are also valid and suggests exploring combinations of normalization techniques with mappings that resemble Chebyshev polynomials.
  • Application to Differential Equations: The authors highlight the relevance of their work for physics-informed neural networks (PINNs) and differential machine learning (DML) and suggest further investigation into how their findings can improve solvers for differential equations.

Summary updated

BibTeX
@article{manzano2025approximation,
  author = {Alberto Manzano and David Dechant and Jordi Tura and Vedran Dunjko},
  title = {{Approximation and Generalization Capacities of Parametrized Quantum Circuits for Functions in Sobolev Spaces}},
  journal = {Quantum},
  year = {2025},
  doi = {10.22331/q-2025-03-10-1658},
}
Figure