Hierarchical Tucker tensor optimization - applications to 4D seismic data interpolation

There has been a swell of research in the scientific computing community in the last couple of years which tries to extend notions of linear algebra (rank, the SVD, linear systems, etc.) to higher dimensional arrays, or tensors. Much work has been proposed to try to overcome the so called "curse of dimensionality", the O(N^d) storage required for a d-dimensional array, where N is the size of each dimension. The hierarchical Tucker format is one such tensor representation which manages to decompose a hierarchy of dimensions into parameter matrices of very manageable size, requiring at most dNK + (d - 2)K^3 + K^2 parameters, where K is an internal rank parameter. In this work, we extend ideas of matrix completion to the tensor case, where we only know a small number of randomly distributed entries from various 4D frequency slices, and try to recover the fully sampled tensor based on the knowledge that it has low hierarchical tucker rank in a particular arrangement of dimensions. Using this approach, we exploit the multi-dimensional dependencies within the full data in order to achieve very promising interpolation results even from heavily subsampled data.