Fig. 4.

Impact of data quality and dimensionality on optimal stopping time. In (A) we plot the impact of SNR on the optimal stopping time with low measurement density $\alpha =.05$, and compare this to the predictions of (16) with $\lambda =1$ because the non-zero values of the MP distribution are highly peaked around this value at low measurement density. In (B) we plot the impact of measurement density on the optimal stopping time in the low noise (SNR $=$ 100) limit, where longer training is required near $\alpha =1$ because the high quality of the data makes it beneficial to learn weights even in the small eigenvalue directions. In both plots green curves are optimal stopping time numerical predictions computed via gradient descent on $t$ using (14). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)