Exponential concentration in quantum kernel methods.

Area of Science:

• Quantum Machine Learning (QML)
• Computational Quantum Physics

Background:

• Kernel methods are a promising area in Quantum Machine Learning (QML) for achieving quantum advantage.
• The convexity of the training landscape in kernel-based models guarantees optimal parameter convergence, assuming efficient quantum kernel estimation.

Purpose of the Study:

• To investigate the resource requirements for accurate quantum kernel value estimation.
• To identify conditions under which quantum kernel values concentrate, impacting QML model performance.
• To provide guidelines for efficient quantum kernel evaluation in QML.

Main Methods:

• Analytical derivation of concentration bounds for quantum kernels.
• Identification of four key sources contributing to kernel value concentration: data embedding expressivity, global measurements, entanglement, and noise.
• Numerical simulations to verify findings on various QML tasks.

Main Results:

• Quantum kernel values can concentrate exponentially with the number of qubits under certain conditions.
• Concentration leads to trivial QML models where predictions are input-independent, even with polynomial measurements.
• Parametrized data embedding with kernel alignment is also susceptible to exponential concentration for classical data.

Conclusions:

• Efficient quantum kernel evaluation is critical for the success of QML kernel methods.
• Specific features like data embedding expressivity, global measurements, entanglement, and noise can lead to performance degradation.
• Guidelines are provided to avoid quantum kernel concentration, ensuring the practical utility of QML algorithms.