EIG() does NOT return sorted values; it returns them to the user in the order in which the underlying LAPACK routine calculates them. That they often happen to appear to have been numerically sorted is simply a result of that routine applied to the input array.
In general, there is no such thing as "sorted" or "unsorted" order of the eigenvalues that eig() can know of -- they are merely the set of polynomial roots in the complex plane.
That there is a physical problem associated with the particular array is unknown to eig() and that there is a sequence to come even further from its knowledge base. However, there is a File Exchange submission that may help resolve the problem you're trying to solve -- it isn't unique but there isn't necessarily a hard and fast manner in which it can be solved.
Try John d'Errico's eigenshuffle()
