Entanglement and Classical Simulability in Quantum Extreme Learning Machines

TL;DR

This work analyzes Quantum Extreme Learning Machines (QELMs) where only the output layer is trained and a fixed quantum reservoir evolves data encoded with dense-angle states under the $H_{XX}$ Hamiltonian. It reveals a sharp accuracy transition at a characteristic time $t_*$ followed by a saturation $A_*$, with $A_*$ matching the performance of random unitaries despite the integrable, local nature of $H_{XX}$. The onset of entanglement coincides with improved data embedding in the Hilbert space, enabling better class separation in global measurement probabilities $p(s)$, though local observables saturate differently. Crucially, the transition time is largely system-size independent, implying QELMs operate with shallow quantum circuits and remain classically simulable for many learning tasks, while still illustrating how moderate quantum correlations can bridge quantum dynamics and classical feature learning. The analysis uses a probability polytope framework and a digital surrogate to demonstrate that limited entanglement and local propagation can achieve near-random-unitary performance, offering a controlled platform to study quantum-classical boundaries in learning.

Abstract

Quantum Machine Learning (QML) has emerged as a promising framework to exploit quantum mechanics for computational advantage. Here we investigate Quantum Extreme Learning Machines (QELMs), a quantum analogue of classical Extreme Learning Machines in which training is restricted to the output layer. Our architecture combines dimensionality reduction (via PCA or Autoencoders), quantum state encoding, evolution under an XX Hamiltonian, and projective measurement to produce features for a classical single-layer classifier. By analyzing the classification accuracy as a function of evolution time, we identify a sharp transition between low- and high-accuracy regimes, followed by saturation. Remarkably, the saturation value coincides with that obtained using random unitaries that generate maximally complex dynamics, even though the XX model is integrable and local. We show that this performance enhancement correlates with the onset of entanglement, which improves the embedding of classical data in Hilbert space and leads to more separable clusters in measurement probability space. Thus, entanglement contributes positively to the structure of the data embedding, improving learnability without necessarily implying computational advantage. For the image classification tasks studied in this work (namely MNIST, Fashion-MNIST, and CIFAR-10) the required evolution time corresponds to information exchange among nearest neighbors and is independent of the system size. This implies that QELMs rely on limited entanglement and remain classically simulable for a broad class of learning problems. Our results clarify how moderate quantum correlations bridge the gap between quantum dynamics and classical feature learning.

Figures (12)

• Figure 1: Schematic representation of the QELM architecture. The workflow consists of the following steps: dimensionality reduction using either PCA or an Autoencoder (AE); encoding of the classical data into an initial quantum state; time evolution through the quantum layer; measurement of the evolved quantum state; and final classification via a classical single-layer neural network. Image for illustrative purposes only.
• Figure 2: Sample images from three widely used benchmark datasets in this study. The top row displays handwritten digits from MNIST (digits 0–9), the middle row shows clothing items from Fashion-MNIST (10 clothing categories), and the bottom row presents natural object categories from CIFAR-10. Each column corresponds to one class, labeled above each image.
• Figure 3: Training (left panel) and testing (right panel) accuracy as a function of evolution time using the MNIST dataset. The time evolution and the encoding have been performed with the Hamiltonian XX model and the dense angle, respectively. The horizontal dashed lines indicate the performance obtained using a random unitary matrix, specifically by performing 10 measurements and computing the mean and standard deviation.
• Figure 4: Training (left panel) and testing (right panel) accuracy as a function of evolution time using the Fashion-MNIST dataset. The time evolution and the encoding have been performed with the Hamiltonian XX model and the dense angle, respectively. The horizontal dashed lines indicate the performance obtained using a random unitary matrix, specifically by performing 10 measurements and computing the mean and standard deviation.
• Figure 5: Training (left panel) and testing (right panel) accuracy as a function of evolution time using the CIFAR-10 dataset. The time evolution and the encoding have been performed with the Hamiltonian XX model and the dense angle, respectively. The horizontal dashed lines indicate the performance obtained using a random unitary matrix, specifically by performing 10 measurements and computing the mean and standard deviation.