In this talk I will discuss recent advances in Neural-Network parametrisations of pure and mixed quantum states that can be efficiently sampled.  In the first part I will generalise the Neural Density Operator ansatz originally proposed by Torlai and Melko by combining two general ingredients that can be used to construct deep, autoregressive ansatze that automatically enforce positive definitness.  In the second part, instead, I will show that any matrix product state can be exactly represented by a recurrent neural network with a linear memory update. I will then discuss how to generalise this linear RNN architecture to 2D lattices, comparing both approaches to standard DMRG calculations.

Zoom link:  https://pitp.zoom.us/j/98833163484?pwd=bEZTeTB0c1l1QmkrQXdEc3dBQ0dJZz09


Talk Number 22070028
Speaker Profile Filippo Vincentini
Collection Quantum Matter
Source Repository PIRSA