Update pce_cross_validation.rst.

python/doc/theory/meta_modeling/pce_cross_validation.rst

@@ -57,7 +57,13 @@ then the fast methods presented in :ref:`cross_validation` can be applied:
 Fast methods are implemented in :class:`~openturns.FunctionalChaosValidation`.
 
 If model selection is used (such as :class:`LARS`), then only the active
-functions in the basis must be used.
+functions in the basis from the training step must be used.
+Indeed, with naive cross-validation, the selection method may consider different
+active bases.
+The fast method, on the other hand, only consider the active basis which is the
+result of the training step.
+If the selection method is not stable in the cross-validation method, then
+the results with the naive and fast cross-validation methods may be different.
 
 .. topic:: API: