Statistical consequences of applying a PCA noise filter on EELS spectrum images.

Principal component analysis (PCA) noise filtering is a popular method to remove noise from experimental electron energy loss (EELS) spectrum images. Here, we investigate the statistical behaviour of this method by applying it on a simulated data set with realistic noise levels. This phantom data set provides access to the true values contained in the data set as well as to many different realizations of the noise. Using least squares fitting and parameter estimation theory, we demonstrate that even though the precision on the estimated parameters can be better as the Cramér-Rao lower bound, a significant bias is introduced which can alter the conclusions drawn from experimental data sets. The origin of this bias is in the incorrect retrieval of the principal loadings for noisy data. Using an expression for the bias and precision of the singular values from literature, we present an evaluation criterion for these singular values based on the noise level and the amount of information present in the data set. This criterion can help to judge when to avoid PCA noise filtering in practical situations. Further we show that constructing elemental maps of PCA noise filtered data using the background subtraction method, does not guarantee an increase in the signal to noise ratio due to correlation of the spectral data as a result of the filtering process.