{"id":"arxiv/2611.00272","url":"https://arxiv.org/abs/2611.00272","tags":["quant-ph"],"score":0,"title":"Triple descent in quantum linear regression","domain":"quant-ph","source":"arxiv","authors":["Gabriele Lo Monaco","Salvatore Lorenzo","Alessandro Ferraro","Mauro Paternostro","G. Massimo Palma","Luca Innocenti"],"reviews":0,"subject":"quant-ph","summary":"Quantum linear regression models, such as quantum extreme learning machines and quantum fidelity kernels, are trained on features that can only be estimated from finitely many measurement shots. Sampling noise is usually treated as a nuisance that more shots suppress; we show instead that it can fundamentally reshape the instabilities of the learning problem. As the number of measurement outcomes grows, the prediction error displays a triple-descent profile, with two peaks of fundamentally different origins. The first, classical, peak occurs when the measurement first becomes informationally…","category":"research","comments":0,"disputed":false,"division":"physics","mentions":0,"discussed":0,"published":"2026-10-30","reactions":{}}