arXiv:2610.12015
Who can sample forever? Shot noise effects in quantum linear regression
Gabriele Lo Monaco, Salvatore Lorenzo, Alessandro Ferraro, Mauro Paternostro, G. Massimo Palma, Luca Innocenti
Abstract
Quantum fidelity kernels and quantum extreme learning machines share the same basic architecture: a fixed quantum device generates features that are combined through a classically trained linear model. We refer to this broad non-variational setting as quantum linear regression. A defining aspect of these models is that the quantum features are not directly available, but must be estimated from a finite number of shots. We show that finite statistics makes quantum linear regression unavoidably ill-conditioned. Ideal feature vectors are confined to a subspace determined by the dimension of the Hilbert space, whereas finite-shot estimates generically acquire components outside this subspace. Training then attempts to fit directions that disappear in the infinite-statistics limit, making the learned model increasingly sensitive to sampling noise. We analyze this mechanism for quantum extreme learning machines trained on $n_{\rm tr}$ states with $N$ shots per state. Despite the ill-conditioning, predictions remain controlled when the training targets are exact. Finite sampling produces a systematic prediction error that decreases quadratically with $N$, while fluctuations induced by the random training measurements decrease inversely with the total training budget. Thus, unlike in standard linear regression with exactly known features, increasing the dataset size alone cannot remove the error caused by sampling noise. If the training targets are also noisy, due for instance to state-preparation errors, the otherwise hidden directions become active. The test error can then increase with $N$, a behavior that we ascribe to a form of quantum overfitting. Our results expose a fundamental tradeoff between dataset size and per-state statistics in non-variational quantum learning and provide a baseline for resource-aware training and regularization strategies.
AI Summary
The paper asks how finite-shot feature estimation changes non-variational quantum learning with classically trained linear readouts. It analyzes quantum extreme learning machines using multinomial probability features and unregularized minimum-norm least squares, with $n_{\rm tr}$ training states, $N$ shots per training state, and $M$ test shots. For informationally complete measurements and spanning training states in the overcomplete regime $n_{\rm tr},n_{\rm out}>d^2$, the ideal feature matrix has rank $d^2$, whereas shot noise generically increases its rank. Noise-only singular values scale as $N^{-1/2}$, producing pseudoinverse contributions growing as $\sqrt N$. Exact achievable labels and exact test probabilities project out the divergent block. Under conditioning and asymptotic-growth assumptions, the infinite-test-statistics MSE has leading terms $b^2/N^2+V/(Nn_{\rm tr})$, with explicitly derived coefficients. Thus increasing dataset size at fixed per-state statistics generally leaves a bias floor. Independent label noise of variance $\gamma^2$, combined with finite test statistics, activates the divergent directions. A homogeneous Gaussian reference model gives an additional leading test-error contribution $\gamma^2NC/M$, where $C=(n_{\rm out}-d^2)/(n_{\rm tr}-n_{\rm out}-1)$ and $n_{\rm tr}>n_{\rm out}+1$: more training shots can then worsen predictions. The analysis combines blockwise pseudoinverse expansions, Gaussian/Wishart calculations, and direct multinomial concentration and remainder bounds, supported by simulations. Rigorous multinomial control is principally large-dataset, fixed-dimension and requires uniform conditioning; Gaussian coefficient averages alone do not control expansion remainders. Detailed extensions to fidelity kernels and comparisons with regularized estimators are left open.
Central claim may not hold: Finite-shot quantum linear regression is not unavoidably subject to the paper's rank-changing instability: an informationally complete measurement with n_out=d^2 and spanning training data has a full-row-rank ideal feature matrix with a locally continuous, bounded pseudoinverse. The abstract states the N^{-2} squared-bias and 1/(N n_tr) fluctuation laws without the asymptotic, conditioning, and nonzero-coefficient qualifications required by the analysis. The abstract's explanation omits that the growing target-noise contribution requires finite test-shot noise as well as training-shot noise and target noise; exact test probabilities eliminate the divergent block's contribution.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- Main: 15 pages, 6 figures. Supplementary: 37 pages, 6 figures