They are stack-allocated, dynamically sized vectors with a maximum size. This is useful for slicing small matrices, without allocation.
Sébastien Crozet