tensorly.contrib.decomposition
.matrix_product_state_cross

matrix_product_state_cross
(input_tensor, rank, tol=1e05, n_iter_max=100)[source] MPS (tensortrain) decomposition via crossapproximation (TTcross) [1]
Decomposes input_tensor into a sequence of order3 tensors of given rank. (factors/cores) Rather than directly decompose the whole tensor, we sample fibers based on skeleton decomposition. We initialize a random tensortrain and sweep from left to right and right to left. On each core, we shape the core as a matrix and choose the fibers indices by finding maximumvolume submatrix and update the core.
 Advantage: faster
The main advantage of TTcross is that it doesn’t need to evaluate all the entries of the tensor. For a tensor_shape^tensor_order tensor, SVD needs O(tensor_shape^tensor_order) runtime, but TTcross’ runtime is linear in tensor_shape and tensor_order, which makes it feasible in high dimension.
 Disadvantage: less accurate
TTcross may underestimate the error, since it only evaluates partial entries of the tensor. Besides, in contrast to its practical fast performance, there is no theoretical guarantee of it convergence.
 Parameters
input_tensor : tensorly.tensor
The tensor to decompose.
rank : {int, int list}
maximum allowable MPS rank of the factors if int, then this is the same for all the factors if int list, then rank[k] is the rank of the kth factor
tol : float
accuracy threshold for outer whileloop
n_iter_max : int
maximum iterations of outer whileloop (the ‘crosses’ or ‘sweeps’ sampled)
 Returns
factors : MPS factors
order3 tensors of the MPS decomposition
Notes
Pseudocode [2]: 1. Initialization tensor_order cores and column indices 2. while (error > tol) 3. update the tensortrain from left to right:
 for Core 1 to Core tensor_order
approximate the skeletondecomposition by QR and maxvol
 update the tensortrain from right to left:
 for Core tensor_order to Core 1
approximate the skeletondecomposition by QR and maxvol
end while
Acknowledgement: the main body of the code is modified based on TensorToolbox by Daniele Bigoni
References
 R12
Ivan Oseledets and Eugene Tyrtyshnikov. Ttcross approximation for multidimensional arrays. LinearAlgebra and its Applications, 432(1):70–88, 2010.
 R13
Sergey Dolgov and Robert Scheichl. A hybrid alternating least squares–tt cross algorithm for parametricpdes. arXiv preprint arXiv:1707.04562, 2017.
Examples
Generate a 5^3 tensor, and decompose it into tensortrain of 3 factors, with rank = [1,3,3,1] >>> tensor = tl.tensor(np.arange(5**3).reshape(5,5,5)) >>> rank = [1, 3, 3, 1] >>> factors = matrix_product_state_cross(tensor, rank) print the first core: >>> print(factors[0]) .[[[ 24. 0. 4.]
[ 49. 25. 29.] [ 74. 50. 54.] [ 99. 75. 79.] [124. 100. 104.]]]