University of Melbourne Team Develops Fourier Analysis for Non-Linear Variational Quantum Circuits

Fourier analysis is being applied to variational quantum circuits with non-linear data embedding to understand their capabilities and limitations. Haiyue Kang and colleagues at University of Melbourne present a new Fourier analysis of quantum neural networks utilising amplitude embedding, a compact encoding scheme. The analysis extends beyond noiseless environments and angle embedding, revealing how input feature domains impact expressivity and showing that the mean of Fourier coefficients concentrates around zero with variance decaying exponentially with frequency magnitude. The team demonstrate, through both analytical derivation using Weingarten calculus and supporting simulations, that this framework provides theoretical guarantees for trainability scaling and practical insights for deployment on noisy quantum devices, even when target functions exhibit noninteger frequencies.

Noise-induced variance suppression enables strong variational quantum computation

The variance of Fourier coefficients in Variational Quantum Circuits (VQC) utilising amplitude embedding is suppressed by a factor of (∑k p2 k) Q, representing a significant improvement over previous analyses. This factor, where Q denotes the number of noise channel instances, arises from the interaction between the amplitude embedding and the noise model considered. Previous investigations into VQC expressivity and trainability were largely limited to noiseless environments and angle embedding, failing to account for the crucial impact of noise and alternative data encoding strategies. This limitation meant that the true potential of VQCs in realistic quantum hardware scenarios remained largely unexplored. The suppression effect discovered by Kang et al. permits reliable quantum computation even with substantial noise, exceeding previously established thresholds for error tolerance. Understanding the precise mechanisms of noise mitigation is paramount for developing fault-tolerant quantum algorithms and assessing the feasibility of near-term quantum devices.

Kang and colleagues at University of Melbourne demonstrate that the mean of these coefficients concentrates around zero, with variance decaying exponentially with multidimensional frequency magnitude, thus providing theoretical guarantees for both expressivity and trainability. This exponential decay is a key finding, suggesting that higher-frequency components of the target function require increasingly complex circuits to represent accurately. Detailed investigations at University of Melbourne have revealed how noise impacts the performance of Variational Quantum Circuits (VQC) employing amplitude embedding, specifically showing a suppression of variance in Fourier coefficients. This suppression isn’t merely a reduction in signal strength; it fundamentally alters the distribution of Fourier coefficients, shifting the balance towards lower frequencies and enhancing the circuit’s ability to generalise. Confirming this behaviour, the team tested target functions utilising noninteger frequencies, which are known to pose challenges for traditional Fourier analysis and highlight the robustness of their new framework. The use of noninteger frequencies is particularly relevant in scenarios where the underlying data is continuous or sampled at irregular intervals.

Noiseless and noisy simulations corroborated these analytical findings, establishing a framework for understanding expressivity and trainability within the frequency domain. These simulations employed a variety of noise models, including depolarising noise and amplitude damping, to assess the robustness of the framework under different conditions. Amplitude embedding requires a number of qubits that scales logarithmically with the number of features, a contrast to the linear scaling of angle embedding, which is particularly relevant for large datasets common in modern machine learning. This logarithmic scaling offers a significant advantage in terms of qubit efficiency, potentially enabling the processing of larger and more complex datasets with limited quantum resources. A detailed framework for analysing variational quantum circuits utilising amplitude embedding, a data encoding technique representing information via the strength of quantum states, has been developed by researchers at University of Melbourne and CSIRO. The choice of data encoding strategy is critical for the performance of VQCs, and amplitude embedding offers a compelling alternative to angle embedding due to its compact representation and potential for noise resilience.

Fourier analysis clarifies learning and error susceptibility in amplitude-encoded quantum circuits

A new analytical perspective on variational quantum circuits, particularly those employing amplitude embedding, has been provided by researchers at University of Melbourne. This technique efficiently encodes data, and the new framework offers theoretical guarantees regarding trainability and expressivity, while also accounting for the detrimental effects of noise. The theoretical foundation of this work relies on the principles of Fourier analysis, which decomposes complex functions into a sum of simpler sinusoidal components. By analysing the Fourier spectrum of the quantum circuit, researchers can gain insights into its ability to represent different functions and its susceptibility to errors. The team’s work rests on the assumption that the generated unitaries must form at least a 2-design, a condition that may not always be met. A 2-design ensures that the circuit can approximate any unitary transformation up to second order, which is sufficient for many quantum machine learning tasks. However, verifying this condition can be computationally challenging, and deviations from a 2-design can impact the accuracy of the analysis.

Establishing theoretical guarantees, even with caveats, aids the development of more durable quantum machine learning algorithms. The ability to mathematically prove the trainability of a quantum circuit is crucial for ensuring its practical viability. Without such guarantees, it is difficult to predict whether the circuit will converge to a meaningful solution during training. The analysis details a link between the expressivity and trainability of quantum circuits and the frequencies present within the data. High-frequency components of the target function require more complex circuits to represent accurately, while low-frequency components can be captured with simpler circuits. This relationship highlights the importance of feature selection and data preprocessing in quantum machine learning. Predictably, the variance of important Fourier coefficients diminishes as the circuit’s complexity increases. As the number of parameters in the quantum circuit grows, its ability to represent different functions also increases, leading to a reduction in the variance of the Fourier coefficients. Furthermore, noise suppresses this variance, contributing to the overall robustness of the computation. This noise-induced suppression is a key finding, suggesting that noise can sometimes play a beneficial role in quantum machine learning by regularising the circuit and preventing overfitting. These findings offer valuable insights into the behaviour of amplitude-encoded quantum circuits and their potential for practical applications, including classification, regression, and generative modelling.

The research demonstrated that the variance of Fourier coefficients in variational quantum circuits scales exponentially with frequency magnitude and is further reduced by noise. This matters because understanding how information is encoded within these circuits is vital for developing more reliable quantum machine learning algorithms. The authors used Weingarten calculus to show that, under certain conditions, these coefficients are concentrated around zero, indicating a predictable relationship between circuit design and data representation. They validated these analytical results through simulations, including those with noisy data, and suggest this approach has practical utility for analysing circuit expressivity.