Although quantum machine learning is rapidly advancing, challenges remain in optimizing performance and understanding the mechanisms underlying these complex systems. Andrei Kardashin from Skolkovo University of Science and Technology, together with Konstantin Antipin from MV Lomonosov Moscow State University and Skolkovo University of Science and Technology, are investigating critical issues affecting the reliability of variational quantum circuits. Their work focuses on the act of measurement within these circuits, specifically how measuring only a portion of the quantum state increases the variance of label predictions during regression tasks. This study demonstrates that the number of distinct eigenvalues obtained from these limited measurements is directly correlated with the instability of the predictions, providing important insights into the design of more robust quantum neural networks.
Optimizing variational quantum circuits with limited measurements
Variational quantum circuits are widely used tools to perform quantum machine learning (QML) tasks on labeled quantum states. For some specific tasks, or for certain variational analyses, measurements may be performed on a restricted portion of the entire input state. This is the case, for example, with quantum convolutional neural networks (QCNN). After each layer of the circuit, a subset of the qubits in the processed state is measured or tracked. Reducing the number of qubits processed in subsequent layers significantly reduces computational costs and memory requirements.
In this work, we investigate how to efficiently implement such limited measurements within larger variational quantum circuits, focusing on performance and scalability optimization. The research approach focuses on developing new techniques to implement partial measurements using a combination of controlled swapping and selective qubit readout. This allows unwanted qubits to be effectively “tracked” without the need for full state tomography or complex post-processing. The method is designed to be compatible with a variety of variational circuit architectures and quantum hardware platforms. The efficiency and accuracy of the proposed method is demonstrated through numerical simulations using circuits with up to 10 qubits.
The specific contributions of this study include a detailed theoretical analysis of the proposed measurement scheme, demonstrating its potential to reduce circuit complexity. Furthermore, the researchers present a practical implementation of this technique within a commonly used quantum computing framework. The results show a significant reduction in the number of quantum gates required to perform equivalent measurements compared to standard approaches, especially for circuits with large numbers of qubits. This advancement paves the way for more efficient and scalable QML algorithms.
Observable limitations and predictive differences in QML regression
This study investigates the impact of observable constraints on the variance of predictions within a quantum machine learning (QML) regression task. The researchers designed a variational quantum computing framework to investigate how the choice of measured observables affects the accuracy of label prediction. The core of their methodology focuses on parameterizing the observable as a sum of orthogonal projectors Λi, each with a corresponding real coefficient λi, allowing for flexible observable design. This approach enables a systematic investigation of the relationship between observable properties and estimator variance, which is important for designing sample-efficient QML architectures.
In the experiment, a variational quantum circuit Uθ was used to transform the input quantum state ρα before measurement. The circuit parameters θ were optimized to accurately predict the label α by minimizing the prediction bias bα and variance Δ2ραM. Specifically, the team constructed an observable Mλ,θ by applying a unitary transformation Uθ to the input state and projecting it onto a subset of m qubits. Here, m is less than or equal to the total number of qubits, n. The expected value of this observable serves as the predicted label a for the input state. For even more methodological control, the scientists took advantage of an extension to Naimark, a technique that introduces auxiliary qubits to effectively simulate measurements on projectors of arbitrary rank.
This allowed us to investigate a wider range of observable structures without changing the underlying measurement process. This process involves connecting the ma auxiliary qubit to |0⟩⟨0|. Design a circuit Uθ that calculates the states and acts on the combined states ρα ⊗|0⟩⟨0|⊗ma to reproduce the desired measurement result. In this work, we closely analyzed the dependence of the dispersion on observable properties, such as the number of qubits over which it is distributed and the degeneracy of the spectrum, across several regression tasks. These include predicting weights in convex combinations of states and estimating parameters of local Hamiltonian models. By combining analytical formulas for variance and numerical experiments, this study shows that limited support measurements can indeed lead to an increase in label prediction variance, and this finding has important implications for the development of practical and efficient QML algorithms.
Observable structure causes differences in QML predictions
Scientists have demonstrated an important connection between the structure of observables used in quantum machine learning (QML) and the resulting predictive variance of regression tasks. This study reveals that measurements performed on a restricted portion of a quantum state lead to an increase in label prediction variance, a phenomenon that is directly correlated to the number of distinct eigenvalues present in the measured observable. The experiments focused on regression problems where the goal is to predict labels associated with quantum states, and established a framework in which label predictions are derived from the expected values of observables. The team measured the variance of the observables to understand how it affects the accuracy of label estimation and found that the variance is essentially tied to the support size of the first observable.
The results show that observables acting on more qubits can significantly reduce the variance for a given number of measurement repetitions, which is important for achieving higher accuracy. Conversely, using observables with limited support, such as single-qubit Pauli operators, tends to increase the variance of the estimates, an important consideration for architectures like quantum convolutional neural networks (QCNNs). Further analysis quantified this trade-off between measurement constraints and prediction variance in quantum data regression, revealing that while QCNN offers shallow circuitry and advantageous scaling, it inherently limits the structural richness of readout observations. In this study, we analytically derived variance equations for two regression tasks: finding the weights of convex combinations of states and predicting the parameters of a local Hamiltonian model, and verified them through numerical experiments.
Measurements confirm that the degeneracy of the observed spectrum also plays an important role in determining the prediction variance. This study establishes that the number of distinct eigenvalues of an observable directly affects the accuracy of label prediction, and that a higher number of eigenvalues generally results in lower variance. This work highlights the importance of carefully designing readout observables in QML architectures to balance computational and sample efficiency, paving the way for more robust and accurate quantum machine learning algorithms. Understanding this relationship is essential for developing QML architectures that are experimentally feasible and capable of providing accurate results with limited measurement resources.
Observable dispersion and quantum Fisher information
Researchers demonstrated the relationship between the variance of label predictions in a regression quantum machine learning task and the observables used to measure them. Their work revealed that measurements performed on a restricted portion of a quantum state lead to an increase in variance, a phenomenon related to the number of distinct eigenvalues present in the measured observations. Specifically, this work proves that for a parametrized family of pure states, observables with a real basis can achieve a dispersion that saturates the inverse quantum Fisher information. Further investigations focused on scenarios where these pure states exist within a two-dimensional real subspace.
The researchers demonstrated that in such cases, it is always possible to find an optimal observable that saturates the inverse classical Fisher information, and thus the inverse quantum Fisher information. This suggests that there are fundamental limits to the accuracy that can be achieved when predicting parameters based on these limited measurements. Numerical experiments performed using a transverse field Ising Hamiltonian confirmed these theoretical observations. The authors acknowledge that their analysis primarily concerns pure states in real subspaces and that further research is needed to extend these findings to mixed states or complex subspaces. They suggest that future work could investigate the effect of varying the dimensions of the real subspace and the number of measured qubits on the efficiency of parameter prediction. These ongoing efforts aim to improve our understanding of the interaction between measurement strategies and the ultimate performance of quantum machine learning algorithms.
