Given an instance of ScanVectorMachine V' (V S), we can produce a type V'' and instance ScanVectorMachine V'' (V' (V S)). In other words, given an implementation of vectors with some nonzero nesting depth, this will produce an implementation with nesting depth one level deeper.

This is different from SegmentedVectors, which uses flat vectors (0-deep nesting) to emulate segmented vectors (1-deep nesting) by cutting the size of the scalars in half. Here, there is no need to assume that the flat-vector scalars are twice as wide (in terms of bits) as the segmented scalars, so arbitrarily deep nesting may be achieved without sacrificing any additional bit-width. In addition, NestedVectors introduces less overhead than SegmentedVectors. For this reason, many hardware/platform providers choose to implement ScanVectorMachine V' (V S) instead of ScanVectorMachine (V S); this requires more work (more methods to implement), but eliminates the overhead of SegmentedVectors.