• We implement a variant of Tensor Cross Interpolation designed for learning a huge number of user-provided pivots.
• We can sample millions of points separately, and later combine the learned tensor trains without losing information.
• The cost scales linear with the number of pivots.

We provide:

• A matrix interpolative decomposition (ID) based on rank revealing QR (from lapack)
• A basic TensorTrain class
• A class doing something like Tensor Cross Interpolation (TCI) , but using ID instead of CUR
• A new implementation of linear-cost interpolation of global pivots
• The quantics grid and qtt learning

• libtorch c++ for tensor manipulation
• Catch2 for testing
• lapack for linear algebra