Although researchers have long known that distance metrics are central to machine learning, quantifying distances between ensembles of quantum states is a major challenge due to inherent measurement limitations. Jian Yao, Pengtao Li, and Xiaohui Chen from the University of Southern California, along with Quntao Zhuang and colleagues, introduced a hierarchy of integral probability metrics, called MMD-, that extends maximum-mean-discrepancy to quantum ensembles. This study is important because it reveals a clear trade-off between the ability of a metric to distinguish between states and the statistical efficiency with which it can be measured, and shows that estimation of MMD requires fewer samples than alternative methods such as Wasserstein distance to achieve comparable performance. These findings provide important guidance for designing effective loss functions in quantum machine learning, and the authors illustrate their application through training a denoised diffusion stochastic model.
MMD-k metrics and quantum state identification
Scientists have established a hierarchy of integer probability metrics, called MMD-k, to address the challenge of measuring distances between ensembles of quantum states, previously hampered by fundamental quantum measurement constraints. This work introduces a new approach that generalizes maximum mean mismatch to quantum ensembles and reveals a strict trade-off between the ability to discriminate between ensembles and the statistical efficiency achieved as the moment order increases. For a pure state ensemble of size N, estimating MMD-k using an experimentally viable SWAP test-based estimator requires Θ(N 2−2/k) samples for constant k, and Θ(N 3) samples to achieve full discriminative power when k equals N. This is in contrast to the quantum Wasserstein distance, which achieves perfect discriminatory power with Θ(N 2 log N) samples.
The research team achieved these results by focusing on integral probability metrics and investigating how the ability to distinguish between ensembles is limited by the number of samples required for estimation. They developed a family of metrics, MMD-k, that extends the classical MMD distance. This shows that the discriminative power of MMD-k increases with moment order k and saturates approximately where k equals N. However, this enhanced discriminatory power requires a corresponding statistical cost. This is because estimating MMD-k using a constant k for a fixed additive error using the SWAP test-based protocol requires a sample number on the scale of ~N. 2−2/k. This study revealed that approximately ~N 3 samples are required to achieve full discriminatory power with k = N.
These findings provide principled guidance for the design of loss functions in quantum machine learning, and the researchers illustrate their contribution through training quantum denoising diffusion stochastic models. Numerical simulations confirm the theoretical predictions regarding the quantum measurement protocol and validate the proposed methodology. This study establishes a fundamental trade-off between discriminative power and sample complexity due to the measurement uncertainties inherent in quantum physics. This study suggests that loss function hierarchy may be unavoidable in learning scenarios with limited or noisy data access, providing a new perspective on the challenges of quantum data analysis. This work paves the way to optimizing the balance between the ability to detect ensemble properties and the efficiency with which they can be learned from measurements, potentially leading to more effective quantum machine learning algorithms and a deeper understanding of quantum state characterization.
MMD-k estimation using SWAP test is powerful
Scientists have developed a new hierarchy of integral probability metrics called MMD-k to address the challenge of quantifying distances between ensembles of quantum states. The research team developed a method to generalize maximum mean discrepancy. This allows for ensemble comparison while exhibiting a quantifiable trade-off between discriminative power and statistical efficiency as moment order k increases. This study pioneered the use of an experimentally viable SWAP test-based estimator that requires only N = 200 samples for a constant k to estimate MMD-k for a pure state ensemble of size N. However, to achieve full discriminative power with k = 2, N = 2000 samples are required.
The researchers used quantum circuits to implement the SWAP test, a key element of the estimation procedure. The team built these circuits using basic single-qubit gates such as Hadamard gates, Pauli X-gates, and Pauli Z-gates in parallel with two-qubit gates such as controlled NOT gates and SWAP gates. Controlled SWAP gates (Fredkin gates) are also employed for multi-qubit operations to selectively swap quantum states based on the state of the control qubit. The SWAP gate itself, denoted by SWAP(|φ⟩⊗|ψ⟩) = (|ψ⟩⊗|φ⟩), was decomposed into three CNOT gates to facilitate efficient implementation in quantum circuits.
This innovative circuit design allows for accurate measurements of state distances. In this study, we closely compared the performance of MMD-k with Wasserstein distance, a well-established metric. Experiments revealed that Wasserstein distance can achieve full discriminative power in as few as 200 samples, highlighting its potential efficiency advantage over MMD-k in certain scenarios. This comparison provides principled guidance for designing loss functions in machine learning, especially within the context of training denoised diffuse stochastic models. The research team demonstrated that this method achieves rigorous evaluation of ensemble distances and provides a valuable tool for quantum information processing and machine learning applications.
Additionally, this work details the mathematical basis of MMD-k and proves important properties regarding its relationship with other distance metrics and its behavior with partial tracing operations. The researchers demonstrated that if two ensembles produce the same kth moment, they are indistinguishable using MMD-k for any k. This theoretical framework is supported by rigorous proofs and supports the practical implementation and interpretation of the developed methodology. This approach allows for a nuanced understanding of the tradeoffs inherent in quantifying distances between quantum ensembles, paving the way for more informed design choices in quantum machine learning algorithms.
Consider the trade-off between MMD-k hierarchy and sample complexity
The researchers established a new hierarchy of integral probability metrics, called MMD-k, that extends maximum-mean discrepancy to ensembles of quantum states. This hierarchy presents a clear trade-off between the ability to discriminate ensembles and the statistical efficiency required for accurate estimation, with higher moment orders of MMD-k increasing discriminatory power at the expense of more samples. Specifically, estimating MMD-k for a pure-state ensemble of a given size requires a large number of samples that scale with the ensemble size with a constant moment order and scale linearly with the ensemble size to achieve full discriminatory power. Compared to the Wasserstein distance, which achieves perfect discriminative power with different numbers of samples, these findings provide guidance for designing loss functions in quantum machine learning.
The authors illustrate this by applying their results to training a denoising diffusion stochastic model. Although they acknowledge that their results are specific to quantum data, they highlight the broader principle that data access limitations, whether physical or algorithmic, may require trade-offs between discriminative power and statistical cost. Future studies may investigate the behavior of MMD-k with non-integer values of k and investigate the relationship between MMD-k and Wasserstein distance. This study reveals the fundamental relationship between discriminatory power and sample complexity in quantum data analysis. The established hierarchy of loss functions suggests that scenarios with limited data access may require inherently higher statistical costs to achieve greater discrimination. The authors also point out that further research is needed to fully understand the properties of MMD-k and its relationship with other distance metrics, which could lead to more efficient learning algorithms in various situations where data access is limited or noisy.
