On fundamental aspects of quantum extreme learning machines

TL;DR

This work analyzes the expressivity and scalability of Quantum Extreme Learning Machines (QELMs) when learning from classical data by mapping predictions to a Fourier series whose frequencies are set by the encoding generator and whose Fourier coefficients depend on the reservoir dynamics and chosen observables.The authors show that the Fourier spectrum $\\Omega$ is determined by eigenvalue differences of the encoding Hamiltonian, while coefficients $a_\omega$ and the final weights control which Fourier modes contribute to the prediction, linking expressivity to encoding and measurement design.They identify four sources of exponential concentration—Haar-expressivity of encoders and reservoirs, entanglement, global measurements, and noise—that can drive observables to input-independent values and undermine scalability, cautioning against highly random reservoirs (2-designs) for large systems.The work contrasts Pauli-based encodings (polynomial-frequency spectra) with exponential encodings (exponentially many frequencies) and discusses classical surrogates (Fourier surrogates and Random Fourier Features), clarifying when a quantum advantage is plausible and highlighting the need to balance expressivity with generalization for QELMs.

Abstract

Quantum Extreme Learning Machines (QELMs) have emerged as a promising framework for quantum machine learning. Their appeal lies in the rich feature map induced by the dynamics of a quantum substrate - the quantum reservoir - and the efficient post-measurement training via linear regression. Here we study the expressivity of QELMs by decomposing the prediction of QELMs into a Fourier series. We show that the achievable Fourier frequencies are determined by the data encoding scheme, while Fourier coefficients depend on both the reservoir and the measurement. Notably, the expressivity of QELMs is fundamentally limited by the number of Fourier frequencies and the number of observables, while the complexity of the prediction hinges on the reservoir. As a cautionary note on scalability, we identify four sources that can lead to the exponential concentration of the observables as the system size grows (randomness, hardware noise, entanglement, and global measurements) and show how this can turn QELMs into useless input-agnostic oracles. In particular, our result on the reservoir-induced concentration strongly indicates that quantum reservoirs drawn from a highly random ensemble make QELM models unscalable. Our analysis elucidates the potential and fundamental limitations of QELMs, and lays the groundwork for systematically exploring quantum reservoir systems for other machine learning tasks.

Figures (11)

• Figure 1: Framework of a QELM. A QELM encodes classical data onto accessible qubits (red arrows in panel (a) or the encoding unitary $U(\boldsymbol{x})$ in panel (b)). The accessible and hidden qubits then undergo a unitary (reservoir) evolution $U_R$, and a set of Hermitian observables $\{O_k\}_{k=1}^{M}$ is measured. The estimates of the observables are classically post-processed to predict the output $f(\boldsymbol{x})$ via linear regression. The reservoir unitary is fixed and only the linear regression weights $\bm{\eta}$ are classically optimized.
• Figure 2: Encoding strategies. The encoding strategy determines the achievable Fourier frequencies of the prediction. Pauli re-uploading and the exponential encoding (as defined in Sec. \ref{['subsec:encoding']}) lead to polynomially and exponentially many frequencies in panels (a) and (b) respectively. Hence, as shown in the plots of the Fourier transform $\hat{f}(\omega$) of the output $f(\boldsymbol{x})$ against the Fourier frequency $\omega$, the prediction of a QELM using Pauli encoding has a more concentrated Fourier spectrum, which is more efficiently simulated classically. The exponential encoding, which corresponds to the partial control regime $M<|\Omega|$, allows for a wider range of target functions compared to the Pauli re-uploading, where $M>|\Omega|$ and might offer a quantum advantage.
• Figure 3: Comparison between two encoding schemes. The mean performance of two QELMs with exponential and Pauli encoding respectively as a function of the number of observables, averaged over 20 randomly selected target functions. The Mean Squared Error (MSE) evaluated on test data is then normalized to the range $[0,1]$. The dashed lines show the standard deviation of the normalized MSE. The vertical green line shows the number of frequencies expressible by exponential encoding ($3^6 +1= 730$). This line separates two controllability regimes, which are defined and discussed in Sec. \ref{['Sec:Control']}.
• Figure 4: Exemplary Fourier spectra of a QELM with $n_{a}=4$ and $n_{h}=4$. The Fourier coefficients of three observables, $O_1=X^{\otimes 4}\otimes \mathcal{I}$, $O_2=X\otimes Y \otimes Y\otimes Z \otimes \mathcal{I}$ and $O_3=O_{rand}\otimes \mathcal{I}$ (where $O_{rand}$ is a random Hermitian in the accessible space), resulting from six $reservoirs$: no reservoir, integrable Ising model ($J=-1,\ B_x=0,\ B_z=1$), chaotic Ising model ($J=-1,\ B_x=0.7,\ B_z=1.5$) and three different Haar random reservoirs. Each row (from top to bottom) corresponds to an observable (from $O_1$ to $O_3$ respectively), while each column corresponds to a reservoir unitary indicated on the top. We remark that the motivation of showing plots for three different Haar random reservoirs is to avoid special cases.
• Figure 5: Richness of Fourier modes. The richness of Fourier modes, i.e. the proportion of non-zero Fourier coefficients averaged over all the Pauli observables in the accessible space $\{I,\ X,\ Y,\ Z\}^{\otimes n_{a}}\otimes \mathcal{I}$, are plotted on a logarithmic scale against the number of accessible qubits $n_{a}$ (the number of hidden qubits is fixed to $n_{h}=4$).

Theorems & Definitions (36)

• Definition 1: Fourier-expressivity of QELM
• Theorem 1: Upper bound of QELM's Fourier-expressivity
• Definition 2: Probabilistic exponential concentration
• Definition 3: Deterministic exponential concentration
• Theorem 2: Encoding Haar-expressivity-induced concentration
• Theorem 3: Reservoir Haar-expressivity-induced concentration
• Theorem 4: Entanglement-induced concentration
• Theorem 5: Global measurement-induced concentration
• Theorem 6: Noise-induced concentration
• Theorem 1: Upper bound of QELM's Fourier-expressivity