Abstract

We show that any matrix product state (MPS) can be exactly represented by arecurrent neural network (RNN) with a linear memory update. We generalize thisRNN architecture to 2D lattices using a multilinear memory update. It supportsperfect sampling and wave function evaluation in polynomial time, and canrepresent an area law of entanglement entropy. Numerical evidence shows that itcan encode the wave function using a bond dimension lower by orders ofmagnitude when compared to MPS, with an accuracy that can be systematicallyimproved by increasing the bond dimension.

 

Quick Read (beta)