A bug in "MPS.apply_local_op"?

How do I use this algorithm? What does that parameter do?
Post Reply
hainingpan
Posts: 1
Joined: 12 Jan 2024, 21:34

A bug in "MPS.apply_local_op"?

Post by hainingpan »

I encountered a very strange problem.

The minimal example is a chain of qubits of 4, and then followed by a 2-site gate. The strange thing is that: the answer is incorrect when the 2 site gate is applied to qubits (1,2) using MPS.apply_local_op", but it is correct when applying to qubits (0,1). What is even stranger is that, the correctness depends on the specific applied "2-site gate". Below I will provide the example to reproduce it.

I am using the version of
tenpy 0.10.0 (compiled without HAVE_MKL),
git revision 0ada42d89dcf186a22ff820159dec378f2cde077 using
python 3.11.4 (main, Jul 5 2023, 14:15:25) [GCC 11.2.0]
numpy 1.24.3, scipy 1.11.4

Python: Select all

import tenpy
import numpy as np
from tenpy.linalg import np_conserved as npc

# construct a simple |0000> state
qubit_site = tenpy.networks.site.SpinHalfSite(conserve=None,)
legs=[f'p{i}' for i in range(4)]
vec=np.zeros((2**4,),dtype=complex)
vec[0]=1
wavefunction_tenpy = npc.Array.from_ndarray(vec.reshape((2,2,2,2)), [qubit_site.leg] * 4,labels=legs,)
# convert mps back to state vector 
def toarray(psi_mps):
    wf=psi_mps.get_B(0)
    for mps in psi_mps._B[1:]:
        wf=npc.tensordot(wf,mps,axes=([-1],[0]))
    wf_contracted = npc.trace(wf, leg1=0, leg2=-1)
    return wf_contracted.to_ndarray().flatten()

wf_mps=tenpy.networks.mps.MPS.from_full( [qubit_site]*4, wavefunction_tenpy,form='B',normalize=True)

toarray(wf_mps)

# The output is correct.
array([1.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j,
       0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j])
Then I define a simple "2-qubit gate", and apply to (1,2) qubits,

Python: Select all

U1=(1+1j)*np.arange(1,17).reshape((2,2,2,2))
U_npc1=npc.Array.from_ndarray(U1, [qubit_site.leg] * 4,labels=['p0','p1','p0*','p1*'],)
wf_mps=tenpy.networks.mps.MPS.from_full( [qubit_site]*4, wavefunction_tenpy,form='B',normalize=True)
wf_mps.apply_local_op(1,U_npc1)
toarray(wf_mps)

# Output:
array([0.04256283+0.04256283j, 0.        +0.j        ,
       0.21281413+0.21281413j, 0.        +0.j        ,
       0.38306544+0.38306544j, 0.        +0.j        ,
       0.55331674+0.55331674j, 0.        +0.j        ,
       0.        +0.j        , 0.        +0.j        ,
       0.        +0.j        , 0.        +0.j        ,
       0.        +0.j        , 0.        +0.j        ,
       0.        +0.j        , 0.        +0.j        ])

wf_mps=tenpy.networks.mps.MPS.from_full( [qubit_site]*4, wavefunction_tenpy,form='B',normalize=True)
wf_mps.apply_local_op(0,U_npc1)

toarray(wf_mps)

# Output: 
array([0.04256283+0.04256283j, 0.        +0.j        ,
       0.        +0.j        , 0.        +0.j        ,
       0.21281413+0.21281413j, 0.        +0.j        ,
       0.        +0.j        , 0.        +0.j        ,
       0.38306544+0.38306544j, 0.        +0.j        ,
       0.        +0.j        , 0.        +0.j        ,
       0.55331674+0.55331674j, 0.        +0.j        ,
       0.        +0.j        , 0.        +0.j        ])
So far eveything is correct, the output state vector is normalized internally (I guess).

Now comes to the strange part: If the applied gate is the following (which I attach it in the pickle), then the result is wrong.

Python: Select all

import pickle
with open('UU.pkl','rb') as f:
    U2=pickle.load(f)
U2
# Output:
array([[[[ 0.04578135-0.19817758j, -0.0670271 -0.11887706j],
         [ 0.21649947+0.26349682j, -0.17076864+0.89137018j]],

        [[-0.1950499 -0.04680251j,  0.236489  +0.71568491j],
         [ 0.56152975-0.10577503j, -0.25196233-0.04054729j]]],


       [[[-0.25624666+0.32897489j, -0.62812977-0.04047919j],
         [ 0.02579847-0.56901467j, -0.29334038+0.13941218j]],

        [[-0.84659876-0.16666881j,  0.08502642+0.09913931j],
         [-0.44095103+0.19573532j, -0.02838949+0.06984177j]]]])
Then applying this 2-site gate to the chain on qubit (1,2):

Python: Select all

U_npc2=npc.Array.from_ndarray(U2, [qubit_site.leg] * 4,labels=['p0','p1','p0*','p1*'],)
wf_mps=tenpy.networks.mps.MPS.from_full( [qubit_site]*4, wavefunction_tenpy,form='B',normalize=True)
wf_mps.apply_local_op(1,U_npc2)
toarray(wf_mps)
# Output:
array([ 0.29762525-0.81477149j,  0.        +0.j        ,
       -0.49749284-0.00821637j,  0.        +0.j        ,
       -0.26306051+0.42233375j,  0.        +0.j        ,
       -0.84588209-0.19213763j,  0.        +0.j        ,
        0.        +0.j        ,  0.        +0.j        ,
        0.        +0.j        ,  0.        +0.j        ,
        0.        +0.j        ,  0.        +0.j        ,
        0.        +0.j        ,  0.        +0.j        ])
toarray(wf_mps)


This is already wrong. The vector components do not even exist in the original U2 matrix.

However, if I just apply the gate to (0,1), then it is correct again:

Python: Select all

wf_mps=tenpy.networks.mps.MPS.from_full( [qubit_site]*4, wavefunction_tenpy,form='B',normalize=True)
wf_mps.apply_local_op(0,U_npc2)
toarray(wf_mps)
# Output:
array([ 0.04578135-0.19817758j,  0.        +0.j        ,
        0.        +0.j        ,  0.        +0.j        ,
       -0.1950499 -0.04680251j,  0.        +0.j        ,
        0.        +0.j        ,  0.        +0.j        ,
       -0.25624666+0.32897489j,  0.        +0.j        ,
        0.        +0.j        ,  0.        +0.j        ,
       -0.84659876-0.16666881j,  0.        +0.j        ,
        0.        +0.j        ,  0.        +0.j        ])
So I wonder the reason why the third output (array([ 0.29762525-0.81477149j, 0. +0.j ...])) is wrong? Is it because of some caveat when using "MPS.apply_local_op" that I didn't notice?

I have also provide the ipynb which can directly run to reproduce it.
Attachments
MPS_bug.zip
(2.09 KiB) Downloaded 655 times
Post Reply