Quantum Circuits Enhance Machine Learning Data

A new quantum machine learning framework for estimating classical covariance matrices using parameterised quantum circuits has been developed by Vicente P. Soloviev and Bibhas Adhikari at the Fujitsu Research. They present two quantum covariance estimators, the C-Estimator and the E-Estimator, analysing their trade-offs regarding qubit requirements and learning complexity. The framework offers a key approach to learning covariance matrices and potentially extends the application of quantum machine learning to complex, high-dimensional statistical estimation problems.

The team also detail a method to mitigate the barren plateau phenomenon during training, improving the efficiency of their E-Estimator through careful selection of regularization parameters. Quantum machine learning achieves sub-one percent error in high-dimensional covariance estimation Error rates in estimating covariance matrices dropped to below 0.01 using the newly developed quantum machine learning framework, a threshold previously unattainable with classical methods for high-dimensional datasets. This advance stems from a novel application of the Pauli-Correlation-Encoding (PCE) model, translating the complex task of covariance estimation into a quantum circuit optimisation problem. The PCE paradigm represents classical data as quantum states through the encoding of correlations, allowing quantum algorithms to operate directly on the data’s covariance structure.

The team created and analysed two estimators, the C-Estimator and E-Estimator; the C-Estimator utilises Cholesky factorization to guarantee positive definiteness, a crucial property for valid covariance matrices, while the E-Estimator offers computational efficiency through direct estimation of covariance entries from observable expectation values. Cholesky factorization decomposes the covariance matrix into the product of a lower triangular matrix and its transpose, ensuring that the resulting matrix is positive semi-definite, a mathematical requirement for representing valid probability distributions. The E-Estimator, in contrast, bypasses this explicit factorization, directly estimating the covariance elements, potentially reducing computational overhead. Numerical simulations demonstrate strong performance in low-rank approximation, covariance completion with incomplete data, and estimation of precision matrices. Low-rank approximation is particularly useful when dealing with datasets where many variables are correlated, allowing for a simplified representation of the covariance structure. Covariance completion addresses the common problem of missing data, reconstructing the full covariance matrix from partial observations. Precision matrices, which are the inverse of covariance matrices, are important in various statistical applications, including graphical modelling and Bayesian network analysis. Regularization parameters within the loss function used for the HEA ansatz, a specific type of parameterised quantum circuit, were carefully selected to mitigate the ‘barren plateau’ phenomenon, a specific challenge in training quantum circuits where gradients vanish exponentially with increasing circuit depth. Analysis revealed sufficient conditions for these parameters to ensure the resulting covariance estimators remained mathematically valid, specifically positive semi-definite, which is important for representing realistic data distributions. The selection of appropriate regularization parameters is critical to prevent the optimisation process from getting stuck in local minima or producing invalid covariance matrices. The E-Estimator required fewer qubits than the C-Estimator for equivalent performance, highlighting a trade-off between computational resources and algorithmic complexity, and potentially key for scaling the approach to larger datasets. Qubit count is a significant constraint in current quantum hardware, making the E-Estimator a potentially more practical option for near-term quantum devices. Simulations also explored the framework’s ability to reconstruct covariance matrices from limited information and complete partially observed matrices, mimicking real-world scenarios with missing data. This capability is particularly relevant in fields like finance and healthcare, where data is often incomplete or noisy. Currently, these results rely on numerical simulations with randomly generated data; demonstrating a clear advantage over existing classical methods with real-world financial or biological datasets remains a vital hurdle, and future work will focus on this validation. Rigorous benchmarking against established classical algorithms using real-world datasets is essential to establish the practical utility of this quantum framework. Quantum computation offers potential for high-dimensional covariance matrix estimation Covariance matrices underpin countless statistical analyses, from financial risk modelling to signal processing, yet accurately estimating them becomes exponentially harder as data dimensions increase. This phenomenon, known as the ‘curse of dimensionality’, limits the applicability of classical methods to high-dimensional datasets. A quantum approach was developed to tackle this challenge, utilising a method of representing data for quantum computers. The PCE model allows for a potentially exponential speedup in certain computational tasks by leveraging the principles of quantum superposition and entanglement. Efficient estimation of these matrices could unlock improvements across diverse fields, as they are vital for understanding relationships within data and building accurate predictive models. This framework shows potential for tackling high-dimensional statistical challenges and could unlock improvements across finance and signal processing. In finance, accurate covariance matrices are crucial for portfolio optimisation and risk management. In signal processing, they are used for noise reduction and feature extraction. Establishing a quantum framework for estimating covariance matrices represents a crucial task in multivariate statistics. By introducing both the C-Estimator and E-Estimator, a pathway to construct these matrices using quantum circuits was demonstrated, balancing the need for mathematical validity with computational efficiency. The C-Estimator prioritizes mathematical correctness through Cholesky decomposition, while the E-Estimator prioritizes computational speed. Careful parameter selection further improves the viability of the E-Estimator by mitigating the ‘barren plateau’, a common obstacle in quantum algorithm training, and allowing for stronger performance. The barren plateau arises from the vanishing of gradients during the optimisation of quantum circuits, hindering the learning process. By carefully tuning the regularization parameters, the researchers were able to alleviate this issue and improve the convergence of the E-Estimator. The framework’s ability to handle incomplete data is also a significant advantage, as real-world datasets often contain missing values. Further research will focus on exploring the scalability of this framework and demonstrating its performance on larger, more complex datasets. The researchers successfully developed a quantum machine learning framework for estimating classical covariance matrices using parameterized quantum circuits. This is important because covariance matrices are fundamental to understanding relationships within data and are used in fields like finance and signal processing. Two estimators, the C-Estimator and E-Estimator, were created, offering a trade-off between mathematical accuracy and computational efficiency.

The team also demonstrated a method to improve the performance of the E-Estimator by addressing the ‘barren plateau’ phenomenon, a common challenge in quantum algorithm training. 👉 More information 🗞 Quantum Learning of Classical Correlations with continuous-domain Pauli Correlation Encoding 🧠 ArXiv: https://arxiv.org/abs/2604.05637 Tags: