Detecting underdetermination in parameterized quantum circuits

Marie C. Kempkes, Jakob Spiegelberg, Evert P. L. van Nieuwenburg, Vedran Dunjko

TL;DR

This paper introduces a method for detecting "underdetermination" in parameterized quantum circuits (PQCs), a phenomenon where multiple circuit configurations can explain the training data but disagree on new predictions. By adapting a classical machine learning technique that analyzes the Hessian matrix of the loss function, the authors demonstrate that this approach can effectively identify unreliable predictions in PQCs, even in the presence of noise.

What Problem Does This Paper Address?

As quantum machine learning (QML) moves towards practical applications, ensuring the reliability of its predictions is paramount, especially for safety-critical tasks. A key challenge is the "underdetermination problem," where a model family can have multiple valid configurations (hypotheses) that perform equally well on training data but yield vastly different predictions for new, unseen data points. This ambiguity can lead to unreliable outputs. While this problem is well-studied in classical machine learning, its implications and detection methods in the context of quantum machine learning, particularly with parameterized quantum circuits (PQCs), are less understood. This paper aims to bridge this gap by investigating whether classical methods for detecting underdetermination can be effectively applied to PQCs.

Methods

The core of the paper's methodology is the adaptation of a classical machine learning technique for detecting underdetermination, as proposed by Madras et al. [4]. This method relies on analyzing the Hessian matrix of the training loss function at the optimized parameters. The Hessian matrix captures the curvature of the loss landscape. By examining the eigenvalue spectrum of the Hessian, the authors identify regions of "low curvature."

The key idea is to project the gradient of a PQC's prediction with respect to its parameters onto this low-curvature subspace. This projection, termed the extrapolation score ($E_m$), quantifies how much the prediction would change if the parameters were slightly perturbed within these flat regions. A higher extrapolation score suggests a greater potential for disagreement among different hypotheses, thus indicating underdetermination.

Specifically, the method involves:

  1. Computing the Hessian matrix ($H$) of the loss function at the trained parameters.
  2. Performing spectral decomposition of the Hessian to obtain its eigenvectors and eigenvalues.
  3. Identifying a subspace of low curvature by selecting eigenvectors corresponding to the smallest eigenvalues. This subspace is represented by a matrix $U_m$.
  4. Calculating the gradient of the prediction with respect to the parameters, $g_{\theta^*} (x')$.
  5. Computing the extrapolation score as $E_m(x') = ||U_m^T g_{\theta^*} (x')||^2$.

The authors also investigate the correlation between the extrapolation score and the ensemble standard deviation, a more direct measure of underdetermination obtained by training multiple models and observing the variance in their predictions.

Datasets

The paper utilizes the following datasets:

  • Iris dataset: Specifically, the first two classes (setosa and versicolor) are used for binary classification. The data is normalized to the interval [0, π].
  • Synthetic data: Generated according to the function $y = \sin(4x) + \mathcal{N}(0, 1)$ and $y = x + \mathcal{N}(0, 1)$. For these datasets, training data is restricted to specific intervals, while test data spans a wider range to simulate out-of-distribution scenarios.
  • Fashion-MNIST dataset: A larger, real-world dataset used to test the method's robustness and applicability to more complex scenarios. For this dataset, PCA is applied to reduce dimensionality to match the number of qubits.

Key Insights and Ideas

  • PQCs exhibit underdetermination: The study confirms that PQCs, like classical neural networks, are susceptible to the underdetermination problem. This means that even with good training performance, their predictions might not be universally reliable.
  • Hessian-based method is effective for PQCs: The classical method of using local second-order information (Hessian) to detect underdetermination is shown to be effective for PQCs. This suggests that the loss landscape of PQCs around trained minima has structures that can be exploited for reliability assessment.
  • Extrapolation score as a reliable indicator: The extrapolation score ($E_m$) serves as a good proxy for underdetermination, correlating well with the ensemble standard deviation, especially in controlled synthetic scenarios.
  • Robustness to shot noise: A crucial finding is that the extrapolation score remains a reliable indicator of underdetermination even when the PQC outputs are estimated using a limited number of shots (i.e., in the presence of shot noise). This is vital for practical quantum computing where perfect expectation values are not directly accessible.
  • Potential for safety-critical QML: By providing a quantifiable measure of prediction reliability, this work lays the groundwork for deploying QML models in applications where safety is a concern.

Key Results

  • Correlation: A Pearson correlation coefficient of $R = 0.86$ was observed between the extrapolation score and the ensemble standard deviation on the Iris dataset, indicating a strong positive relationship.
  • Synthetic Data Visualization: On synthetic datasets ($y = \sin(4x) + \mathcal{N}(0, 1)$ and $y = x + \mathcal{N}(0, 1)$), the extrapolation score clearly distinguished between data points within the training domain and those outside, demonstrating its ability to detect distributional shifts.
  • Noise Robustness: The extrapolation score showed robustness against moderate levels of shot noise. In a binary classification scenario simulating underdetermination, a linear model trained on the extrapolation score achieved higher accuracy than one trained on the ensemble standard deviation when shot noise was present.
  • Real-World Data: On the Fashion-MNIST dataset, the extrapolation score increased with larger rotation angles applied to test images, simulating a distributional shift. This indicates the score's ability to reflect growing degrees of unreliability.

Limitations and Future Work

  • Hyperparameter Tuning: The effectiveness of the extrapolation score depends on the choice of the hyperparameter $m$, which determines the size of the low-curvature subspace. Selecting an appropriate $m$ requires analyzing the eigenvalue spectrum, and a clear distinction between small and large eigenvalues is beneficial.
  • "Local" Ensemble Assumption: The theoretical underpinnings of the extrapolation score rely on the concept of a "local" ensemble. While the correlation with ensemble standard deviation is good, it's not perfect, suggesting that the ensemble might not always be strictly local in practice.
  • PQC Architecture: The study primarily uses commonly explored PQC architectures and does not delve into designing architectures that are inherently more trainable or resistant to dequantization.
  • Future Work: The authors suggest further research into designing PQC architectures that balance trainability and robustness. They also highlight the need for more work on identifying and constructing such architectures. The paper's contribution is seen as a step towards reliable QML, paving the way for its application in safety-critical domains.

Summary updated

Abstract

A central question in machine learning is how reliable the predictions of a trained model are. Reliability includes the identification of instances for which a model is likely not to be trusted based on an analysis of the learning system itself. Such unreliability for an input may arise from the model family providing a variety of hypotheses consistent with the training data, which can vastly disagree in their predictions on that particular input point. This is called the underdetermination problem, and it is important to develop methods to detect it. With the emergence of quantum machine learning (QML) as a prospective alternative to classical methods for certain learning problems, the question arises to what extent they are subject to underdetermination and whether similar techniques as those developed for classical models can be employed for its detection. In this work, we first provide an overview of concepts from Safe AI and reliability, which in particular received little attention in QML. We then explore the use of a method based on local second-order information for the detection of underdetermination in parameterized quantum circuits through numerical experiments. We further demonstrate that the approach is robust to certain levels of shot noise. Our work contributes to the body of literature on Safe Quantum AI, which is an emerging field of growing importance.
BibTeX
@article{kempkes2026detecting,
  author = {Marie C. Kempkes and Jakob Spiegelberg and Evert P. L. van Nieuwenburg and Vedran Dunjko},
  title = {{Detecting underdetermination in parameterized quantum circuits}},
  journal = {Mach. Learn. Sci. Technol},
  year = {2026},
  doi = {10.1088/2632-2153/ae304e},
}
Figure