De Jolliffe (2002), page 339 : "It is true that variances, covariances and correlations have especial relevance for multivariate normal x, and that linear functions of binary variables are less readily interpretable than linear functions of continuous variaibles. However, the basic objective of PCA --  to summarize most of the 'variation' that is present in the original set of p variables using a smaller number of derived variables -- can be achieved regardless of the nature of the original variables.

For data in which all variables are binary, Gower (1966) points out that using PCA *does* provide a plausible low-dimensional representation. This follows because PCA is equivalent to a principal coordinate analysis based on the commonly used definition of similarity between two individuals (observations) as the proportion of the p variables for which the two individals take the same value."

Donc oui, on peut faire de l'ACP sur données binaires et ça a du sens. Je stocke ça ici.