Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2412.03182 · Annales Henri Poincaré

Quantitative Convergence of Trained Quantum Neural Networks to a Gaussian Process

Anderson Melchor Hernandez, Filippo Girardi, Davide Pastorello, Giacomo De Palma

Published Oct 8, 2025• 4 citations
arXivHTMLPDF

Abstract

Abstract We study quantum neural networks where the generated function is the expectation value of the sum of single-qubit observables across all qubits. In (Girardi et al., CMP (2025), it is proven that the probability distributions of such generated functions converge in distribution to a Gaussian process in the limit of infinite width for both untrained networks with randomly initialized parameters and trained networks. In this paper, we provide a quantitative proof of this convergence in terms of the Wasserstein distance of order 1. First, we establish an upper bound on the distance between the probability distribution of the function generated by any untrained network with finite width and the Gaussian process with the same covariance. This proof utilizes Stein’s method to estimate the Wasserstein distance of order 1. Next, we analyze the training dynamics of the network via gradient flow, proving an upper bound on the distance between the probability distribution of the function generated by the trained network and the corresponding Gaussian process. This proof is based on a quantitative upper bound on the maximum variation of a parameter during training. This bound implies that for sufficiently large widths, training occurs in the lazy regime, i.e., each parameter changes only by a small amount. While the convergence result of Girardi et al., CMP (2025) holds at a fixed training time, our upper bounds are uniform in time and hold even as $$t \rightarrow \infty $$ t → ∞ .

DOI
10.1007/s00023-025-01631-6

Related papers

Quantum neural networks form Gaussian processes

Diego García-Martín, Martín Larocca, Marco Cerezo

Classical artificial neural networks initialized from independent and identically distributed priors converge to Gaussian processes in the limit of a large number of neurons per hidden layer. This correspondence plays an important role in the current understanding of the capabilities of neural networks. Here we prove an analogous result for quantum neural networks. We show that the outputs of certain models based on Haar-random unitary or orthogonal quantum neural networks converge to Gaussian processes in the limit of large Hilbert space dimension d. The derivation of this result is more nuanced than in the classical case due to the role played by the input states, the measurement observable and because the entries of unitary matrices are not independent. We show that the efficiency of predicting measurements at the output of a quantum neural network using Gaussian process regression depends on the number of measured qubits. Furthermore, our theorems imply that the concentration of measure phenomenon in Haar-random quantum neural networks is worse than previously thought, because expectation values and gradients concentrate as $${\mathcal{O}}\left({1}/{\operatorname{e}^{d}\sqrt{d}}\right)$$ . The connection between classical neural networks and Gaussian processes is a fundamental result in machine learning. It has now been shown that many quantum neural networks converge to Gaussian processes, enabling their use for regression tasks.

arXiv:2305.09957•Nature Physics•May 21, 2025•44 citations

Quantum Lazy Training

Erfan Abedi, Salman Beigi, Leila Taghavi

In the training of over-parameterized model functions via gradient descent, sometimes the parameters do not change significantly and remain close to their initial values. This phenomenon is called lazy training and motivates consideration of the linear approximation of the model function around the initial parameters. In the lazy regime, this linear approximation imitates the behavior of the parameterized function whose associated kernel, called the tangent kernel , specifies the training performance of the model. Lazy training is known to occur in the case of (classical) neural networks with large widths. In this paper, we show that the training of geometrically local parameterized quantum circuits enters the lazy regime for large numbers of qubits. More precisely, we prove bounds on the rate of changes of the parameters of such a geometrically local parameterized quantum circuit in the training process, and on the precision of the linear approximation of the associated quantum model function; both of these bounds tend to zero as the number of qubits grows. We support our analytic results with numerical simulations.

arXiv:2202.08232•Quantum•Apr 27, 2023•17 citations

Classical Computation of the Quantum NTK

BRUNO DANTI

Girardi and De Palma [Girardi and De Palma, Commun. Math. Phys. 406, 92 (2025)] proved that wide quantum neural networks (QNNs) satisfying certain hypotheses, mainly having light cone sizes (i.e. number of qubits that influence the measurement output of a given qubit) that scale slowly enough, trained under gradient flow, converge to a Gaussian Process (GP), whose expectation value depends only on the QNN’s neural tangent kernel (NTK), i.e. the inner product between the gradients of the model function computed at two different inputs. Hernandez et al. [Hernandez et al., Quantum 10, 2118 (2026)] proposed a classical Monte Carlo algorithm estimating the NTK of a class of QNNs with sample complexity that is polynomial in all the relevant parameters of the QNN, but whose upper bound is astronomically large, leaving open whether the algorithm is feasible in practice. In this thesis, we implement this algorithm, and use it to simulate the infinite-width limit of the QNN proposed by Tognini et al. [Tognini et al., arXiv:2505.14789] on the binary MNIST classification task. The predictions stabilize at approximately 104 samples, which our code generates in less than a day on a desktop computer, while the upper bound guarantees the same precision only at 1054 samples. We also extend the sample complexity upper bound to kernel ridge regression with the NTK, including hyperparameter selection, whose model reaches a test accuracy of 98.4%, similar to Tognini’s 97.5% accuracy, and find the sample complexity bound to be similarly pessimistic in this regime. Since we have no reason to believe the MNIST task or Tognini’s QNN to be special in some regard, we believe Hernandez’s sample complexity bound to be astronomically pessimistic in general, due to our results, which would rule out a quantum advantage for all wide QNNs converging to a GP and in Hernandez’s class. As such, we believe the most likely candidate subclass of wide QNNs that could exhibit a quantum advantage to be that of wide QNNs with light cones too large to converge to a GP, but which are not too deep, in order not to incur barren plateaus.

•Oct 14, 2026•0 citations

A Unified Theory of Quantum Neural Network Loss Landscapes

Eric R. Anschuetz

Classical neural networks with random initialization famously behave as Gaussian processes in the limit of many neurons, which allows one to completely characterize their training and generalization behavior. No such general understanding exists for quantum neural networks (QNNs), which -- outside of certain special cases -- are known to not behave as Gaussian processes when randomly initialized. We here prove that QNNs and their first two derivatives instead generally form what we call "Wishart processes," where certain algebraic properties of the network determine the hyperparameters of the process. This Wishart process description allows us to, for the first time: give necessary and sufficient conditions for a QNN architecture to have a Gaussian process limit; calculate the full gradient distribution, generalizing previously known barren plateau results; and calculate the local minima distribution of algebraically constrained QNNs. Our unified framework suggests a certain simple operational definition for the "trainability" of a given QNN model using a newly introduced, experimentally accessible quantity we call the "degrees of freedom" of the network architecture.

arXiv:2408.11901•Aug 21, 2024•25 citations

Lazy training of quantum physics informed neural networks

Anderson Melchor Hernandez, Giacomo De Palma

We study the gradient-flow training dynamics of quantum physics-informed neural networks (QPINNs) for the solution of second-order elliptic partial differential equations with Dirichlet boundary conditions. We consider parameterized quantum circuits as function approximators and analyze their overparameterized regime through the lens of the neural tangent kernel (NTK). Our contribution is a nonasymptotic lazy-training theory for QPINNs and their variational formulation: we prove that, for sufficiently large circuit width, the nonlinear gradient flow is quantitatively approximated by a linearized NTK model, with explicit bounds depending on the number of qubits, circuit depth, circuit light-cone geometry, and the dimension of the domain of the solution to the PDE.

quant-pharXiv:2609.19239•Sep 17, 2026•0 citations

Analytic Theory for the Dynamics of Wide Quantum Neural Networks

Junyu Liu, Khadijeh Najafi, Kunal Sharma +3 more

Parametrized quantum circuits can be used as quantum neural networks and have the potential to outperform their classical counterparts when trained for addressing learning problems. To date, much of the results on their performance on practical problems are heuristic in nature. In particular, the convergence rate for the training of quantum neural networks is not fully understood. Here, we analyze the dynamics of gradient descent for the training error of a class of variational quantum machine learning models. We define wide quantum neural networks as parametrized quantum circuits in the limit of a large number of qubits and variational parameters. Then, we find a simple analytic formula that captures the average behavior of their loss function and discuss the consequences of our findings. For example, for random quantum circuits, we predict and characterize an exponential decay of the residual training error as a function of the parameters of the system. Finally, we validate our analytic results with numerical experiments.

arXiv:2203.16711•Physical Review Letters•Apr 10, 2023•70 citations
Abstract We study quantum neural networks where the generated function is the expectation value of the sum of single-qubit observables across all qubits. In (Girardi et al., CMP (2025), it is proven that the probability distributions of such generated functions converge in distribution to a Gaussian process in the limit of infinite width for both untrained networks with randomly initialized parameters and trained networks. In this paper, we provide a quantitative proof of this convergence in terms of the Wasserstein distance of order 1. First, we establish an upper bound on the distance between the probability distribution of the function generated by any untrained network with finite width and the Gaussian process with the same covariance. This proof utilizes Stein’s method to estimate the Wasserstein distance of order 1. Next, we analyze the training dynamics of the network via gradient flow, proving an upper bound on the distance between the probability distribution of the function generated by the trained network and the corresponding Gaussian process. This proof is based on a quantitative upper bound on the maximum variation of a parameter during training. This bound implies that for sufficiently large widths, training occurs in the lazy regime, i.e., each parameter changes only by a small amount. While the convergence result of Girardi et al., CMP (2025) holds at a fixed training time, our upper bounds are uniform in time and hold even as t→∞t \rightarrow \infty t→∞ t → ∞ .
Classical artificial neural networks initialized from independent and identically distributed priors converge to Gaussian processes in the limit of a large number of neurons per hidden layer. This correspondence plays an important role in the current understanding of the capabilities of neural networks. Here we prove an analogous result for quantum neural networks. We show that the outputs of certain models based on Haar-random unitary or orthogonal quantum neural networks converge to Gaussian processes in the limit of large Hilbert space dimension d. The derivation of this result is more nuanced than in the classical case due to the role played by the input states, the measurement observable and because the entries of unitary matrices are not independent. We show that the efficiency of predicting measurements at the output of a quantum neural network using Gaussian process regression depends on the number of measured qubits. Furthermore, our theorems imply that the concentration of measure phenomenon in Haar-random quantum neural networks is worse than previously thought, because expectation values and gradients concentrate as O(1/e⁡dd){\mathcal{O}}\left({1}/{\operatorname{e}^{d}\sqrt{d}}\right)O(1/edd​) . The connection between classical neural networks and Gaussian processes is a fundamental result in machine learning. It has now been shown that many quantum neural networks converge to Gaussian processes, enabling their use for regression tasks.
In the training of over-parameterized model functions via gradient descent, sometimes the parameters do not change significantly and remain close to their initial values. This phenomenon is called lazy training and motivates consideration of the linear approximation of the model function around the initial parameters. In the lazy regime, this linear approximation imitates the behavior of the parameterized function whose associated kernel, called the tangent kernel , specifies the training performance of the model. Lazy training is known to occur in the case of (classical) neural networks with large widths. In this paper, we show that the training of geometrically local parameterized quantum circuits enters the lazy regime for large numbers of qubits. More precisely, we prove bounds on the rate of changes of the parameters of such a geometrically local parameterized quantum circuit in the training process, and on the precision of the linear approximation of the associated quantum model function; both of these bounds tend to zero as the number of qubits grows. We support our analytic results with numerical simulations.
Girardi and De Palma [Girardi and De Palma, Commun. Math. Phys. 406, 92 (2025)] proved that wide quantum neural networks (QNNs) satisfying certain hypotheses, mainly having light cone sizes (i.e. number of qubits that influence the measurement output of a given qubit) that scale slowly enough, trained under gradient flow, converge to a Gaussian Process (GP), whose expectation value depends only on the QNN’s neural tangent kernel (NTK), i.e. the inner product between the gradients of the model function computed at two different inputs. Hernandez et al. [Hernandez et al., Quantum 10, 2118 (2026)] proposed a classical Monte Carlo algorithm estimating the NTK of a class of QNNs with sample complexity that is polynomial in all the relevant parameters of the QNN, but whose upper bound is astronomically large, leaving open whether the algorithm is feasible in practice. In this thesis, we implement this algorithm, and use it to simulate the infinite-width limit of the QNN proposed by Tognini et al. [Tognini et al., arXiv:2505.14789] on the binary MNIST classification task. The predictions stabilize at approximately 104 samples, which our code generates in less than a day on a desktop computer, while the upper bound guarantees the same precision only at 1054 samples. We also extend the sample complexity upper bound to kernel ridge regression with the NTK, including hyperparameter selection, whose model reaches a test accuracy of 98.4%, similar to Tognini’s 97.5% accuracy, and find the sample complexity bound to be similarly pessimistic in this regime. Since we have no reason to believe the MNIST task or Tognini’s QNN to be special in some regard, we believe Hernandez’s sample complexity bound to be astronomically pessimistic in general, due to our results, which would rule out a quantum advantage for all wide QNNs converging to a GP and in Hernandez’s class. As such, we believe the most likely candidate subclass of wide QNNs that could exhibit a quantum advantage to be that of wide QNNs with light cones too large to converge to a GP, but which are not too deep, in order not to incur barren plateaus.
Classical neural networks with random initialization famously behave as Gaussian processes in the limit of many neurons, which allows one to completely characterize their training and generalization behavior. No such general understanding exists for quantum neural networks (QNNs), which -- outside of certain special cases -- are known to not behave as Gaussian processes when randomly initialized. We here prove that QNNs and their first two derivatives instead generally form what we call "Wishart processes," where certain algebraic properties of the network determine the hyperparameters of the process. This Wishart process description allows us to, for the first time: give necessary and sufficient conditions for a QNN architecture to have a Gaussian process limit; calculate the full gradient distribution, generalizing previously known barren plateau results; and calculate the local minima distribution of algebraically constrained QNNs. Our unified framework suggests a certain simple operational definition for the "trainability" of a given QNN model using a newly introduced, experimentally accessible quantity we call the "degrees of freedom" of the network architecture.
We study the gradient-flow training dynamics of quantum physics-informed neural networks (QPINNs) for the solution of second-order elliptic partial differential equations with Dirichlet boundary conditions. We consider parameterized quantum circuits as function approximators and analyze their overparameterized regime through the lens of the neural tangent kernel (NTK). Our contribution is a nonasymptotic lazy-training theory for QPINNs and their variational formulation: we prove that, for sufficiently large circuit width, the nonlinear gradient flow is quantitatively approximated by a linearized NTK model, with explicit bounds depending on the number of qubits, circuit depth, circuit light-cone geometry, and the dimension of the domain of the solution to the PDE.
Parametrized quantum circuits can be used as quantum neural networks and have the potential to outperform their classical counterparts when trained for addressing learning problems. To date, much of the results on their performance on practical problems are heuristic in nature. In particular, the convergence rate for the training of quantum neural networks is not fully understood. Here, we analyze the dynamics of gradient descent for the training error of a class of variational quantum machine learning models. We define wide quantum neural networks as parametrized quantum circuits in the limit of a large number of qubits and variational parameters. Then, we find a simple analytic formula that captures the average behavior of their loss function and discuss the consequences of our findings. For example, for random quantum circuits, we predict and characterize an exponential decay of the residual training error as a function of the parameters of the system. Finally, we validate our analytic results with numerical experiments.