Tensor network inner product

194 Views Asked by At

Tensor networks can be thought of as linear maps between Hilbert spaces. They are used in physics to scale down huge calculations of many body systems, and have a strong mathematical basis. Tensor networks are basically multiple tensors connected together to form a network.

Take the tensor $ R_1 $ as a linear map between two Hilbert spaces $ H_t $ and $ H_s $, where $t_1, t_2, s_1, $ and $ s_2 $ are the labels of the indices of $ R_1 $:

$ R_1: H_{t_1} \otimes H_{t_2} \to H_{s_1} \otimes H_{s_2} $.

Another tensor $ R_2 $ is defined similarly as:

$ R_2: H_{t_3} \otimes H_{t_4} \to H_{s_3} \otimes H_{s_4} $.

$ R_1 $ and $ R_2 $ are rank four tensors because they have four indices.

$ R_1 $ and $ R_2 $ are "joined" to form a lattice of rank four tensors.

Say there are nine rank four tensors. So we can describe this in a set,

$ G_1 = (R_1,R_2,...,R_9), $

These nine rank four tensors form an inner cycle and each of them except the one at the center is joined to a tensor of rank five. The tensors of rank five are defined similarly to the tensors of rank four. Here is how $ E_1 $ is defined:

$ E_1: H_{t_5} \otimes H_{t_6} \to H_{s_5} \otimes H_{s_6} $.

And the set of these tensors is:

$ G_2 = (E_1,E_2,...,E_4), $.

The tensor $ E_1 $ is the rank five tensor a the top left corner of the network. The tensor $ E_2 $ is the rank five tensor in the lower right corner of the network.

Removing all interior tensors (the inner cycle of nine rank four tensors) we can view $ E_1 $ and $ E_2 $ as an inner product, correct? If we can, then I guess we can also view $ E_3 $ and $ E_4 $ alone as an inner product as well, where $ E_3 $ and $ E_4 $ are the upper right and lower left corners respectively.

Let the two inner products overlap orthogonally and treat each intersection as a rank four tensor. Is this network I described a valid tensor network?

What might be some strengths and weaknesses of this tensor network?

enter image description here