Probabilistic PCA for missing data: learning curves shows a phase transition and missing rate acts as an effective reduction in the signal-to-noise ratio, not the sample size.
