Rethinking Quantum Noise in Quantum Machine Learning: When Noise Improves Learning

TL;DR

This work challenges the notion that quantum noise is always detrimental for near-term quantum machine learning by showing initialization-dependent, heterogeneous noise effects in quantum graph neural networks trained to predict the HOMO-LUMO gap from QM9 data. Using a single-layer EDU circuit with 12 qubits and a depolarizing-like noise model across four per-gate error rates, the authors analyze 55 independently initialized models, revealing that approximately one-third benefit from moderate noise while others degrade or remain unaffected. A strong negative correlation ($r = -0.62$) between baseline noiseless performance and noise benefit indicates that noise serves as an implicit regularizer for under-optimized models, with the optimal observed noise level ($ u=0.005$) lower than simple theoretical predictions due to possible error cancellation in structured circuits. These results motivate structure- and noise-aware optimization strategies and suggest adaptive, initialization-conditioned approaches to mitigating or leveraging noise in NISQ-era quantum machine learning.

Abstract

Quantum noise is conventionally viewed as a fundamental obstacle in near-term quantum computing, motivating extensive error correction and mitigation strategies. We present numerical evidence that challenges this consensus. Through experiments on quantum graph neural networks for molecular property prediction, we discover that quantum noise induces heterogeneous, initialization-dependent responses. Among randomly initialized models with identical architecture, approximately one-third show performance improvement under moderate noise, while a smaller fraction deteriorate and the remainder are marginally affected. We identify a strong negative correlation ($r = -0.62$) between baseline model performance and noise benefit, suggesting that noise acts as an implicit regularizer for under-optimized models while disrupting well-converged ones. The observed optimal noise level falls below theoretical predictions, indicating error cancellation in structured quantum circuits. These findings demonstrate that quantum noise effects depend critically on initialization quality and need not be uniformly detrimental, suggesting a shift from universal noise mitigation toward structure- and noise-aware optimization strategies.

Figures (3)

• Figure 1: EDU bond encoding mechanism. (a) Molecular graph representation showing nodes (atoms) and edges with different bond types (single and double bonds). (b) EDU circuit structure. The quantum circuit implements the decomposition $\mathrm{EDU} = (V \otimes V) \cdot D \cdot (V^\dagger \otimes V^\dagger)$, where node-local unitaries $V$ use bond-type-specific parameters ($\theta_{\mathrm{type}}, \phi_{\mathrm{type}}$) and the diagonal unitary $D$ uses layer-shared parameters ($\psi_{\mathrm{layer}}$).
• Figure 2: Heterogeneous response to quantum noise across 55 random initializations. (a) Waterfall plot showing individual model responses sorted by performance change. (b) Distribution of performance changes reveals 36.4% beneficial (>2%), 14.5% detrimental (<-2%), and 49.1% marginal responses. (c) Optimal noise level distribution with 49.1% of models performing best at $\varepsilon = 0.005$. (d) Strong negative correlation ($r = -0.620$, $p < 0.001$) between baseline performance and noise benefit.
• Figure 3: Dose-response analysis and mechanistic basis. (a) Average performance change as a function of noise level for beneficial, detrimental, and marginal groups. (b) Baseline performance distribution across response categories reveals systematic differences.