Canonical quantization of neurons

Abstract: Canonical quantization provides a systematic procedure for constructing quantum models from classical Hamiltonians. Here, we apply this principle to a fundamental computational primitive of machine learning: the neuron. Specifically, by viewing a neuron as a composition of an energy function and an activation function, we quantize this model by replacing the energy function with a quantum Hamiltonian and applying the activation function to it through matrix functional calculus. This results in an activation observable that can be measured on an input quantum state. We investigate the use of these quantized neurons for function approximation, where the objective is to learn an unknown observable from labeled quantum data. For this purpose, we develop hybrid quantum-classical algorithms for training and evaluation, including procedures for measuring the activation observable and estimating gradients of the squared loss error. Our algorithms for gradient estimation rely on basic primitives like classical random sampling, the Hadamard test, and Hamiltonian simulation, and those for measuring an activation observable rely on quantum algorithms known as the power of one qumode and Schroedingerization. Numerical experiments demonstrate that our quantized neurons exhibit enhanced expressive capabilities relative to corresponding classical neurons on representative learning tasks. Our work establishes canonical quantization as a principled framework for constructing quantum machine learning primitives and provides a foundation for developing neural architectures tailored to quantum data.

• The paper demonstrates that quantizing the neuron’s energy function into a quantum Hamiltonian enables measurable activation observables through matrix functional calculus.
• The methodology employs hybrid quantum–classical algorithms that efficiently estimate gradients via techniques like the Hadamard test and Hamiltonian simulation.
• Numerical experiments on a seven-qubit system reveal that noncommuting Hamiltonians significantly enhance expressive capacity, lowering squared loss in quantum tasks.

Canonical Quantization of Neurons: A Quantum Framework for Machine Learning Primitives

Overview

The paper "Canonical quantization of neurons" (2607.05000) introduces a formal procedure for constructing quantum versions of classical neurons via canonical quantization. Instead of encoding neural information in quantum states or amplitudes as in prior quantum machine learning models, this approach quantizes the underlying energy function of a classical neuron into a quantum Hamiltonian and then applies the activation function through matrix functional calculus. The resultant activation observable is measurable on quantum input states. The authors present hybrid quantum–classical (HQC) algorithms for training and evaluation, including practical protocols for gradient estimation and observable measurement. Numerical experiments demonstrate that quantized neurons, particularly those employing noncommuting Hamiltonians, exhibit heightened expressive capacity relative to classical neurons on quantum learning tasks.

Canonical Quantization Methodology

Classical neurons, specifically perceptrons, process inputs $z$ via an energy function $w^T z + b$, transformed by an activation function $\varphi$. The canonical quantization approach replaces each classical input variable with a Hermitian operator ($\hat{z}_i = \sigma_Z^{(i)}$ for spin variables), constructing a Hamiltonian $$H_C(\theta) = \sum_i w_i \sigma_Z^{(i)} + b I^{\otimes n}$$. The activation function is then promoted to a matrix function, yielding the activation observable $\varphi(H_C(\theta))$. The expected value of this observable on a quantum state $\rho$ is given by $\Tr[\varphi(H_C(\theta)) \rho]$.

The generalization allows for quantum Hamiltonians with interaction and noncommuting terms across all Pauli operators, expanding the operator space beyond classical models. This enhances the representational power and enables direct application to quantum data. The quantum neuron reduces to the classical model when input states are computational basis eigenstates, ensuring conceptual consistency.

Hybrid Quantum–Classical Algorithms

Training quantized neurons to approximate unknown observables is framed as a function approximation problem. Quantum training data comprises pairs $(\rho_m, y_m)$, with $y_m = \Tr[O \rho_m]$ and $w^T z + b$0 the target observable. The parameterized Hamiltonian $w^T z + b$1 is optimized within the loss function $w^T z + b$2, where $w^T z + b$3 is the temperature-scaled activation.

The paper details HQC algorithms utilizing classical random sampling, the Hadamard test, and Hamiltonian simulation for gradient estimation. Specifically, the expected value and gradients of the activation observable are efficiently estimated by expressing $w^T z + b$4 through an integral representation and exploiting quantum phase estimation-like primitives. For observable measurement, procedures based on the "power of one qumode" and Schrödingerization leverage continuous-variable quantum control and momentum quadrature measurements, producing unbiased estimates of the activation observable.

Numerical Results

Empirical investigations compare quantized neurons (with transverse-field Ising Hamiltonians, TFIM) to classical neurons (Ising Hamiltonians, IM) for function approximation in a seven-qubit system. The target observable $w^T z + b$5 is a nonlinear function of a TFIM Hamiltonian. Quantized neurons utilizing noncommuting Hamiltonians consistently achieve lower squared loss than classical neurons, even with identical parameter counts. This result underscores the critical role of operator noncommutativity and richer operator algebra structures in improving the learnability and expressivity for quantum-induced data distributions.

Notably, activation observables generated by noncommuting Hamiltonians contain higher-order Pauli terms inaccessible to commuting models, further bolstering expressive power. The Taylor expansion of the activation function introduces products of Pauli operators ($w^T z + b$6), enabling representation of directions unavailable in classical frameworks.

Implications and Future Prospects

The canonical quantization of neurons establishes a principled foundation for quantum machine learning primitives directly tailored to quantum information processing. The approach complements quantum Boltzmann machine models by quantizing the nonlinear processing element rather than the underlying probability distribution. Practically, this provides new paradigms for designing quantum neural networks capable of processing quantum data in ways unattainable by classical models.

Theoretically, the enhanced expressive power implies that quantized neurons may unlock robust architectures for learning complex quantum observables. The paper raises open questions regarding trainability and potential barren-plateau phenomena, as current analyses for variational quantum circuits do not directly extend to this framework. Further, the modular composition of quantized neurons into hybrid networks (quantum layers followed by classical ones) offers promising avenues for integrating quantum and classical learning strategies.

Scaling simulations to larger systems, benchmarking on experimental quantum datasets, and formal investigations into barren-plateau behavior and compositionality remain essential research directions. The identification of a BQP-complete decision problem and extensions to continuous-variable quantum data in the companion paper suggest foundational computational implications yet to be fully explored.

Conclusion

The canonical quantization of neurons, as formalized herein, provides a rigorous mechanism for transposing neural computation into the quantum domain. Through the activation of quantum Hamiltonians via matrix functional calculus, the framework enables direct measurement and training of quantum neural primitives with demonstrable advantages in expressivity and performance. The HQC algorithms introduced are practically efficient and conceptually aligned with quantum information processing. This work forms a foundational step toward architecting neural networks operative on quantum data and paves the way for new explorations in quantum learning theory and practice.