The size of tt_reshape

[Yukwah1214]

When I was trying to use the function tt_reshape, I encountered the following problem: It works when the code is :

tt_reshape(tt_eye(8),[2,2,2;2,2,2]')

but if I change the size: tt_reshape(tt_eye(8),[4,4,4;1,1,1]') it went wrong.

and if I change the size to: tt_reshape(tt_eye(8),[4,1,4;1,4,1]') it worked again!

I wonder what is the reason and what should I pay attention to when reshaping a TT tensor/matrix? Maybe this is a simple question, but I really can't find the answer. Sincerely hope someone can help me answer

[dolgov]

Actually, I don't know how to reshape a tt_matrix. tt_tensor is easy: split a core of required size if you can, otherwise merge original cores until you can. For tt_matrix, you need to permute dimensions in different ways depending on what you want (you can't even tell this from sizes alone).

In your example in particular, splitting out a 4 x 1 core from a 8 x 8 matrix can be understood in at least the following ways:

• consider the matrix as tt_eye(2,3), and reshape each core into a 4 x 1 matrix (perhaps what you want);
• split the original full matrix into 4 x 2 x 1 x 8;
• split the original full matrix into 4 x 1 x ... (the total number of ways is combinatorial I think).

The current code works if both row and column sizes can be reorganized by the same sequence of splits and merges. For anything more general I don't see how to write it consistently.