Glossary

Abbreviations used in the documentation, the examples and the API reference.

AD

Automatic differentiation. Derivatives of a computation are obtained by applying the chain rule to its elementary operations. torchtt builds on the PyTorch autograd engine, so gradients can be taken with respect to the TT cores (see torchtt.grad).

AMEn

Alternating minimal energy method [Dolgov and Savostyanov, 2014]. It updates one TT core at a time and enriches the approximation with a low-rank approximation of the residual, so the TT ranks adapt during the iteration. torchtt uses it to solve linear systems (torchtt.solvers.amen_solve()), for matrix products (method='amen' in torchtt.matvec() and torchtt.matmat()), for elementwise division (torchtt.elementwise_divide()) and for cross interpolation (method='amen' in torchtt.interpolate.function_interpolate()).

BiCGSTAB

Biconjugate gradient stabilized method, an iterative solver for non-symmetric linear systems. torchtt.solvers.amen_solve() can use it for its local systems (local_solver=2).

DMRG

Density matrix renormalization group. The name comes from quantum physics [White, 1992]; for the TT format it denotes schemes that optimize two neighbouring TT cores at a time, which lets the TT ranks adapt [Oseledets, 2011, Savostyanov and Oseledets, 2011]. torchtt uses it for matrix-vector products (method='dmrg' in torchtt.matvec()), elementwise products (method='dmrg' in torchtt.hadamard()) and cross interpolation (method='dmrg' in torchtt.interpolate.function_interpolate()).

GMRES

Generalized minimal residual method, an iterative solver for non-symmetric linear systems. torchtt.solvers.amen_solve() uses it for its local systems by default (local_solver=1).

QR

QR decomposition. It factors a matrix into a matrix with orthonormal columns (\(Q\)) and an upper triangular matrix (\(R\)). It is used to orthogonalize TT cores.

QTT

Quantized tensor train [Khoromskij, 2011, Oseledets, 2010]: the TT format applied to a tensor that has been reshaped so that every mode has size 2. See Quantized TT (QTT).

SVD

Singular value decomposition. Truncating it is how TT ranks are reduced, both when a full tensor is decomposed and when a TT tensor is rounded.

TT

Tensor train [Oseledets, 2011]. A format that represents a \(d\)-dimensional tensor by \(d\) three-dimensional tensors, the TT cores. See Overview.

TTM
TT matrix

A linear operator (tensor operator) in the TT format. Its cores are four-dimensional, with one row and one column index each. A torchtt.TT is a TT matrix when its shape is given as a list of (rows, columns) tuples.