Tensor Train Approximation of Multivariate Functions

The tensor train is a popular model for approximating high-dimensional rectangular data structures that cannot fit in any computer memory due to their size. The tensor train can approximate complex functions with many variables in the continuous domain. The traditional method for obtaining the tensor train model is based on a skeleton decomposition, which is better known for matrices. The skeleton (cross) decomposition has the property that the tensor approximation is accurate on certain tensor fibers but may be poor on other fibers. In this paper, we propose a technique for fitting a tensor train to an arbitrary number of tensor fibers, allowing flexible modeling of multivariate functions that contain noise. Two examples are studied: a noisy Rosenbrock function and a noisy quadratic function, both of order 20.