Quantum kernel methods and their limits

Short answer: Quantum kernel methods were proposed as a candidate for quantum advantage in data analysis, but a Nature Communications paper shows that, under certain conditions, quantum kernel values can concentrate exponentially in the number of qubits toward a fixed value. With a polynomial number of measurements, the trained model is then trivial: its predictions on unseen inputs do not depend on the input. A 2026 preprint studies local and multi-scale kernels as mitigations in simulation.

What kernel methods promise

The Nature Communications paper "Exponential concentration in quantum kernel methods", also on arXiv, begins by saying kernel methods in quantum machine learning (QML) have gained significant attention as a potential candidate for achieving a quantum advantage in data analysis. Among their attractive properties, it says, training a kernel-based model is guaranteed to find the optimal parameters because the training landscape is convex.

The catch: estimating the kernel

The authors say that guarantee is based on the assumption that the quantum kernel can be efficiently obtained from quantum hardware. They study kernel models from the perspective of the resources needed to accurately estimate kernel values. They show that, under certain conditions, the values of quantum kernels over different inputs can be exponentially concentrated, in the number of qubits, toward some fixed value. As a result, training with a polynomial number of measurements ends up with a trivial model in which predictions on unseen inputs are independent of the input data.

The paper argues that quantum kernel methods experience a barrier similar to barren plateaus, which we cover in a separate article on variational algorithms.

Four sources of concentration

The paper identifies four sources that can lead to concentration: the expressivity of the data embedding, global measurements, entanglement and noise. For each, the authors derive an associated concentration bound for quantum kernels analytically. They also show that, with classical data, training a parametrized data embedding with a kernel alignment method is also susceptible to exponential concentration. They say their results are verified through numerical simulations for several QML tasks, and offer guidelines saying that certain features should be avoided to ensure the efficient evaluation of quantum kernels.

What happens to fidelity kernels

A 2026 arXiv preprint on mitigation strategies describes fidelity-based quantum kernels as providing a direct interface between quantum feature maps and classical kernel methods. It says they can exhibit exponential concentration: as system size or circuit expressivity grows, the Gram matrix approaches the identity and suppresses informative similarity structure. When feature-map states resemble random states, off-diagonal entries become small, which can suppress label-relevant structure and can make reliable estimation of the kernel expensive in the number of measurement shots. The preprint calls concentration both a statistical issue and an operational issue, the second being measurement cost.

Proposed mitigations

The preprint presents what it calls an empirical study of two mitigation strategies implemented in Qiskit: local, or patch-wise, kernels that aggregate subsystem similarities, and multi-scale kernels that mix local and global similarity across patch granularities. It is a preprint, and this article does not report its numerical results.

What this does not tell you

This article summarises abstracts and introductions of two papers. It does not say that quantum kernels cannot work in any setting. The Nature Communications paper states conditions under which concentration occurs, and the preprint describes simulation studies, not hardware advantage.

FAQ

Why were quantum kernel methods attractive?

The paper says training a kernel-based model is guaranteed to find the optimal parameters because the training landscape is convex.

What is exponential concentration?

Under certain conditions, quantum kernel values over different inputs concentrate exponentially, in the number of qubits, toward a fixed value, giving a trivial model.

What causes it?

The paper names expressivity of the data embedding, global measurements, entanglement and noise.

Does it apply to classical data?

The paper says training a parametrized data embedding with a kernel alignment method on classical data is also susceptible.

Can it be mitigated?

A 2026 preprint studies local and multi-scale kernels in Qiskit simulations. It is an empirical study, not a hardware result.