Harnessing Data Augmentation to Quantify Uncertainty in the Early Estimation of Single-Photon Source Quality

Novel methods for rapidly estimating single-photon source (SPS) quality havebeen promoted in recent literature to address the expensive and time-consumingnature of experimental validation via intensity interferometry. However, thefrequent lack of uncertainty discussions and reproducible details raisesconcerns about their reliability. This study investigates the use of dataaugmentation, a machine learning technique, to supplement experimental datawith bootstrapped samples and quantify the uncertainty of such estimates. Eightdatasets obtained from measurements involving a single InGaAs/GaAs epitaxialquantum dot serve as a proof-of-principle example. Analysis of one of the SPSquality metrics derived from efficient histogram fitting of the syntheticsamples, i.e. the probability of multi-photon emission events, revealssignificant uncertainty contributed by stochastic variability in the Poissonprocesses that describe detection rates. Ignoring this source of error riskssevere overconfidence in both early quality estimates and claims forstate-of-the-art SPS devices. Additionally, this study finds that standardleast-squares fitting is comparable to using a Poisson likelihood, andexpanding averages show some promise for early estimation. Also, reducingbackground counts improves fitting accuracy but does not address thePoisson-process variability. Ultimately, data augmentation demonstrates itsvalue in supplementing physical experiments; its benefit here is to emphasisethe need for a cautious assessment of SPS quality.