Point Group Mapping

Here we discuss how to transform an MPS with a point group (PG) (such as D2h) to an MPS with another PG (such as C2v or Cs).

Limitations:

  • Can transform from high-order PG (ket) to low-order PG (bra) or low-order PG (ket) to high-order PG (bra);

  • The mapping between the two PG must be a homomorphism. As long as it is a homomorphism, any mapping can be used.

  • For normal MPS, the transformation only have a tiny fitting error. For MPS with big site, the matrix elements in the big-site tensor in the MPS will be artificially reordered. (Because the “fusing order” inside the big site can have an influence on the order of states within each symmetry block. Since “fusing order” in the big site can be arbitrary, the mapping code itself cannot figure out the correct mapping for the order of states within each symmetry block. For normal site, there is only one state for each symmetry block so there is no problem.) So the transformed big-site tensor will not be accurate (but the normal site tensors in a big-site MPS will still be accurate).

  • The integral (FCIDUMP) with the two PG must be exactly the same. Consider the following case: if you generate the integral from pyscf, you calculate one integral with D2h symmetry in the molecule, and another integral with C2v symmetry in the molecule. Then there can be small (or big) changes in the integral. Then you cannot use this feature, since point group symmetry is not the only thing that changed.

Example

The example integral file C2.CAS.PVDZ.FCIDUMP can be found in the data folder.

First we do a ground state calculation using D2h point group:

from block2 import *
from block2.su2 import *
import numpy as np
SX = SU2

Global.frame = DataFrame(10 * 1024 ** 2, 10 * 1024 ** 3, "nodex")
n_threads = Global.threading.n_threads_global
Global.threading = Threading(
    ThreadingTypes.OperatorBatchedGEMM | ThreadingTypes.Global,
    n_threads, n_threads, 1)
Global.threading.seq_type = SeqTypes.Tasked
Global.frame.fp_codec = DoubleFPCodec(1E-16, 1024)
Global.frame.minimal_disk_usage = True
Global.frame.use_main_stack = False
print(Global.frame)
print(Global.threading)

fcidump = FCIDUMP()
fcidump.read('C2.CAS.PVDZ.FCIDUMP')

# D2H Hamiltonian
pg = "d2h"
swap_pg = getattr(PointGroup, "swap_" + pg)
vacuum = SX(0)
target = SX(fcidump.n_elec, fcidump.twos, swap_pg(fcidump.isym))
n_sites = fcidump.n_sites
orb_sym = VectorUInt8(map(swap_pg, fcidump.orb_sym))
hamil = HamiltonianQC(vacuum, n_sites, orb_sym, fcidump)
print("D2H ORB SYM = ", hamil.orb_sym)

# MPS
mps_info = MPSInfo(n_sites, vacuum, target, hamil.basis)
mps_info.tag = 'KET'
mps_info.set_bond_dimension(250)
mps_info.save_data('./mps_info.bin')
mps = MPS(n_sites, 0, 2)
mps.initialize(mps_info)
mps.random_canonicalize()
mps.save_mutable()
mps_info.save_mutable()

# MPO
mpo = MPOQC(hamil, QCTypes.Conventional)
mpo = SimplifiedMPO(mpo, RuleQC(), True, True, OpNamesSet((OpNames.R, OpNames.RD)))

# DMRG
me = MovingEnvironment(mpo, mps, mps, "DMRG")
me.delayed_contraction = OpNamesSet.normal_ops()
me.cached_contraction = True
me.init_environments(True)
dmrg = DMRG(me, VectorUBond([250, 500]), VectorDouble([1E-5] * 5 + [1E-6] * 5 + [0]))
dmrg.noise_type = NoiseTypes.ReducedPerturbativeCollected
dmrg.davidson_conv_thrds = VectorDouble([1E-6] * 5 + [1E-7] * 5)
ener = dmrg.solve(20, mps.center == 0, 1E-8)
print('DMRG Energy = %20.15f' % ener)

Then we define a Hamiltonian with a different PG (but the same integral). The mapping should be based on the XOR notation. Here we create hamil_c2v by mapping D2h irreps to C2v irreps. The mapping is not unique, you may need to figure out the actual mapping based on how you will need to mix orbitals with different irreps. Note that something like pg_map = lambda x: x & 6 or pg_map = lambda x: x & 3 should also work.

# C2V Hamiltonian
pg_map = lambda x: (x & 6) >> 1
orb_sym_c2v = VectorUInt8([pg_map(x) for x in orb_sym]) # the mapping is not unique
hamil_c2v = HamiltonianQC(vacuum, n_sites, orb_sym_c2v, fcidump)
target_c2v = SX(target.n, target.twos, pg_map(target.pg))
print("C2V ORB SYM = ", hamil_c2v.orb_sym)

To transform MPS, we need a special identity MPO. This identity will not have bond dimension 1 since it has to mix different PG irreps. If the MPS does not have any big-site, the last two parameters orb_sym_c2v, orb_sym can be omitted.

# Identity MPO for PG mapping
delta_target = (target_c2v - target)[0]
impo = IdentityMPO(hamil_c2v.basis, hamil.basis, vacuum,
    delta_target, hamil.opf, orb_sym_c2v, orb_sym)
impo = SimplifiedMPO(impo, NoTransposeRule(RuleQC()))

Next, we can perform the transformation of MPS using fitting.

# C2V MPS
mps_info_c2v = MPSInfo(n_sites, vacuum, target_c2v, hamil_c2v.basis)
mps_info_c2v.tag = 'KET-C2V'
mps_info_c2v.set_bond_dimension(500)
mps_info_c2v.save_data('./mps_info_c2v.bin')
mps_c2v = MPS(n_sites, mps.center, 2)
mps_c2v.initialize(mps_info_c2v)
mps_c2v.random_canonicalize()
mps_c2v.save_mutable()
mps_info_c2v.save_mutable()

# Linear
me = MovingEnvironment(impo, mps_c2v, mps, "LIN")
me.delayed_contraction = OpNamesSet.normal_ops()
me.cached_contraction = True
me.init_environments(True)
cps = Linear(me, VectorUBond([500]), VectorUBond([500]))
norm = cps.solve(20, mps.center == 0, 1E-8)
print('Norm = %20.15f' % norm)

Finally, we can check whether the MPS gives the correct energy in the new C2v basis:

# C2V MPO
mpo_c2v = MPOQC(hamil_c2v, QCTypes.Conventional)
mpo_c2v = SimplifiedMPO(mpo_c2v, RuleQC(), True, True, OpNamesSet((OpNames.R, OpNames.RD)))

# Expectation
me = MovingEnvironment(mpo_c2v, mps_c2v, mps_c2v, "DMRG")
me.delayed_contraction = OpNamesSet.normal_ops()
me.cached_contraction = True
me.init_environments(True)
ex = Expect(me, 500, 500)
ener_c2v = ex.solve(False)
print('C2V Energy = %20.15f' % ener_c2v)

The printed energy should be very close to the D2h sweep energy at the last site of the last sweep. Note that this may not be the same as the DMRG energy, which is the lowest energy in the last sweep, because here the MPS is transformed from the previous D2h MPS with the center at the last site.

If the MPS contains big-site, there can be a much larger error in the energy due to the reordering of states in the big-site MPS tensor. Re-optimizing the big-site tensor may solve this problem. In addition, me.delayed_contraction = OpNamesSet.normal_ops() must not be set. Otherwise, the following assertion occurs:

Assertion`a->get_type() == SparseMatrixTypes::Normal && b->get_type() == SparseMatrixTypes::Normal && c->get_type() == SparseMatrixTypes::Normal && v->get_type() == SparseMatrixTypes::Normal && da->get_type() == SparseMatrixTypes::Normal && db->get_type() == SparseMatrixTypes::Normal' failed.

Some reference output for this example:

D2H ORB SYM =  VectorUInt8[ 5 0 6 5 3 5 0 0 5 0 3 6 5 0 3 6 7 2 7 2 7 2 1 4 0 5 ]
<-- Site =    0-   1 .. Mmps =    3 Ndav =   1 E =    -75.7284493902 Error = 1.14e-16 FLOPS = 8.66e+05 Tdav = 0.00 T = 0.01
DMRG Energy =  -75.728475543752168
C2V ORB SYM =  VectorUInt8[ 2 0 3 2 1 2 0 0 2 0 1 3 2 0 1 3 3 1 3 1 3 1 0 2 0 2 ]
Norm =    1.000000000000001
C2V Energy =  -75.728449390238850

Inverse Mapping

The inverse mapping from C2v to D2h is also supported. The script is basically the same (except the exchange between C2v and D2h):

from block2 import *
from block2.su2 import *
import numpy as np
SX = SU2

Global.frame = DataFrame(10 * 1024 ** 2, 10 * 1024 ** 3, "nodex")
n_threads = Global.threading.n_threads_global
Global.threading = Threading(
    ThreadingTypes.OperatorBatchedGEMM | ThreadingTypes.Global,
    n_threads, n_threads, 1)
Global.threading.seq_type = SeqTypes.Tasked
Global.frame.fp_codec = DoubleFPCodec(1E-16, 1024)
Global.frame.minimal_disk_usage = True
Global.frame.use_main_stack = False
print(Global.frame)
print(Global.threading)

fcidump = FCIDUMP()
fcidump.read('C2.CAS.PVDZ.FCIDUMP')

# C2V Hamiltonian
pg = "d2h"
pg_map = lambda x: (x & 6) >> 1
swap_pg = getattr(PointGroup, "swap_" + pg)
vacuum = SX(0)
target_d2h = SX(fcidump.n_elec, fcidump.twos, swap_pg(fcidump.isym))
target_c2v = SX(target_d2h.n, target_d2h.twos, pg_map(target_d2h.pg))
n_sites = fcidump.n_sites
orb_sym_d2h = VectorUInt8(map(swap_pg, fcidump.orb_sym))
orb_sym_c2v = VectorUInt8([pg_map(x) for x in orb_sym_d2h]) # the mapping is not unique
hamil = HamiltonianQC(vacuum, n_sites, orb_sym_c2v, fcidump)
print("C2V ORB SYM = ", hamil.orb_sym)

# C2V MPS
mps_info = MPSInfo(n_sites, vacuum, target_c2v, hamil.basis)
mps_info.tag = 'KET'
mps_info.set_bond_dimension(250)
mps_info.save_data('./mps_info.bin')
mps = MPS(n_sites, 0, 2)
mps.initialize(mps_info)
mps.random_canonicalize()
mps.save_mutable()
mps_info.save_mutable()

# C2V MPO
mpo = MPOQC(hamil, QCTypes.Conventional)
mpo = SimplifiedMPO(mpo, RuleQC(), True, True, OpNamesSet((OpNames.R, OpNames.RD)))

# C2V DMRG
me = MovingEnvironment(mpo, mps, mps, "DMRG")
me.delayed_contraction = OpNamesSet.normal_ops()
me.cached_contraction = True
me.init_environments(True)
dmrg = DMRG(me, VectorUBond([250, 500]), VectorDouble([1E-5] * 5 + [1E-6] * 5 + [0]))
dmrg.noise_type = NoiseTypes.ReducedPerturbativeCollected
dmrg.davidson_conv_thrds = VectorDouble([1E-6] * 5 + [1E-7] * 5)
ener = dmrg.solve(20, mps.center == 0, 1E-8)
print('DMRG Energy = %20.15f' % ener)

# D2H Hamiltonian
hamil_d2h = HamiltonianQC(vacuum, n_sites, orb_sym_d2h, fcidump)
print("D2H ORB SYM = ", hamil_d2h.orb_sym)

# Identity MPO for PG mapping
delta_target = (target_d2h - target_c2v)[0]
impo = IdentityMPO(hamil_d2h.basis, hamil.basis, vacuum,
    delta_target, hamil.opf, orb_sym_d2h, orb_sym_c2v)
impo = SimplifiedMPO(impo, NoTransposeRule(RuleQC()))

# D2H MPS
mps_info_d2h = MPSInfo(n_sites, vacuum, target_d2h, hamil_d2h.basis)
mps_info_d2h.tag = 'KET-D2H'
mps_info_d2h.set_bond_dimension(500)
mps_info_d2h.save_data('./mps_info_d2h.bin')
mps_d2h = MPS(n_sites, mps.center, 2)
mps_d2h.initialize(mps_info_d2h)
mps_d2h.random_canonicalize()
mps_d2h.save_mutable()
mps_info_d2h.save_mutable()

# Linear
me = MovingEnvironment(impo, mps_d2h, mps, "LIN")
me.delayed_contraction = OpNamesSet.normal_ops()
me.cached_contraction = True
me.init_environments(True)
cps = Linear(me, VectorUBond([500]), VectorUBond([500]))
norm = cps.solve(20, mps.center == 0, 1E-8)
print('Norm = %20.15f' % norm)

# D2H MPO
mpo_d2h = MPOQC(hamil_d2h, QCTypes.Conventional)
mpo_d2h = SimplifiedMPO(mpo_d2h, RuleQC(), True, True, OpNamesSet((OpNames.R, OpNames.RD)))

# D2H Expectation
me = MovingEnvironment(mpo_d2h, mps_d2h, mps_d2h, "DMRG")
me.delayed_contraction = OpNamesSet.normal_ops()
me.cached_contraction = True
me.init_environments(True)
ex = Expect(me, 500, 500)
ener_d2h = ex.solve(False) / norm ** 2
print('D2H Energy = %20.15f' % ener_d2h)

Some reference outputs for this example:

C2V ORB SYM =  VectorUInt8[ 2 0 3 2 1 2 0 0 2 0 1 3 2 0 1 3 3 1 3 1 3 1 0 2 0 2 ]
--> Site =   24-  25 .. Mmps =    3 Ndav =   1 E =    -75.7284490538 Error = 1.62e-19 FLOPS = 3.87e+05 Tdav = 0.00 T = 0.01
DMRG Energy =  -75.728475021520978
D2H ORB SYM =  VectorUInt8[ 5 0 6 5 3 5 0 0 5 0 3 6 5 0 3 6 7 2 7 2 7 2 1 4 0 5 ]
Norm =    0.999999999998821
D2H Energy =  -75.728449053829152

Initial Guess for Compression

For large systems, the initial guess for Linear (mps_c2v or mps_d2h in the above examples) may be too bad, and very small overlap (F value) with mps can be observed. The MPS bond dimension will be kept as 1 or a very small number (it should be at least one, since by default the random FCI initial guess is used, where at least one state is kept for each quantum number in the initial guess).

To solve this problem, one can add cps.cutoff = 0 before the line norm = cps.solve(...). Alternatively, one can add cps.trunc_type = TruncationTypes.KeepOne * n before the line norm = cps.solve(...), where n is a small positive integer.

Generating initial guess using occupation numbers may also alleviate this problem, but using the above settings, better initial guess with occupation numbers is not mandatory.