Use Mollist2 as internal system information data structure #103

liwt31 · 2020-07-15T16:07:02Z

This the first commit for #100 . The majority of this commit is to use MolList2 as the only way to construct MPO. Now Mpo.__init__ is by default constructing MPOs using the general algorithm. The biggest modification apart from deleting old code is to correct the zero-point energies due to different definitions between MolList and MolList2. The quasiboson algorithm #25 , chain representation #90 , and hybrid TDH algorithm #61 are removed.

Apart from issues mentioned in #100, there are other things that are in the todo list:

The removal of hybrid TDH is not complete. Basically only the tests are removed. (Done in 4dd41ff)
The Bogoliubov formulation should be updated to work with MolList2. Although the algorithm is not necessary I'd like to keep the example. (The example is removed in 5d6b3a9)
The finite temperature SBM won't work because thermal propagation calculates electron occupations which are not defined for SBM. This is considered to be a side-effect of using "e" to represent spins. (Fixed in 5d6b3a9)

These things will be addressed in future commits.

Lastly, I observed that the svd compression test in test_mp.py won't pass for Mpo. I have to increase the bond dimension a little bit to let it pass. I tried to figure out why but I best speculation is that svd compression will stick into some local minimum when compressing the new Mpo structure by MolList2. After all, it is not easy to understand why Mpdm will work while Mpo will not, and the distance metric is actually rather strict for large tensor. I'll keep an eye on this though.

liwt31 · 2020-07-16T14:38:49Z

The second commit 9293d9d makes MolList a subclass of MolList2 and renames it to HolsteinModel. Now every mol_list in Reno is essentially MolList2.

Every if-else clause that distinguishes MolList and MolList2 is simplified and only the branch with MolList2 is retained.
The ModelTranslator mechanism is deleted because in principle every model should be converted to the general model first.
- Currently, the SBM model is implemented in the HolsteinModel and it'll be moved to another class class SpinBosonModel(MolList2) in the future. (update: done in 778525c)
- class VibronicModel(MolList2) should also be implemented. (update: done in e112ab9)
Tests based on cross-validation between different schemes and between MolList and MolList2 are deleted. The MPO constructed by MolList2 should be tested against qutip in small systems in the future, and this will be a part of this PR. (Done in 38ab90b)
I found scipy.linalg.norm sometimes produces a nan with perfectly normal input, so I changed all of them to np.linalg.norm.

I've thought of retaining MolList by making it just a wrapper for HolsteinModel for maximum backward compatibility but finally decided not to do so. But I'm happy to make one if you think it's necessary.

liwt31 · 2020-07-16T15:08:18Z

In e112ab9 the VibronicModel is implemented. The condition for MPO svd compression test in test_mp.py is further relaxed to make the test pass. I haven't found a clue why the test cannot pass yet. But the fact that in my machine the atol=1e-4 condition will work while in Travis it will not may indicate it's a numerical issue.

Implement `SpinBosonModel` as a subclass of `MolList2`. Original interface in `HolsteinModel` is removed. Original `SpinBosonModel` in `sbm/sbm.py` is renamed to `SpinBosonDynamics`.

liwt31 · 2020-07-19T07:27:49Z

In the build for 4dd41ff I found a performance regression on finite temperature CV DMRG:

============================ slowest test durations ============================
89.19s call     renormalizer/cv/tests/test_abs.py::test_ft_abs

After some closer examination, It is concluded the regression is caused by the zero-point energy included in MolList2, and it is fixed by d6d6332. I don't understand what's going on @jiangtong1000 if you are interested you can have a look.

liwt31 · 2020-07-28T01:37:10Z

Contents of the last three commits:

MolList2 is renamed to Model. Almost all variables similar to mol_list or mlist is renamed to model.
A new interface is designed for Model.
- The model dict in the original MolList2 is replaced by ham_terms list. The list elements are Op class (an augmented version of the original Op class). The representation is essentially sum-of-product.
- The order argument is removed because basis now carries information on the DoF it contains.
The new interface relies on the enhancement of Op and BasisSet. Both the class is moved from utils directory to model directory.
- In Op the DoF it represents must be specified via the dof argument. The arguments accept any hashable object, which means we can use simple int or tuple to represent the DoF instead of "e_1" etc.
- In BasisSet the DoF also has to be specified.
VibronicModel introduced in e112ab9 is removed because drastic changes have been made to BasisSet and Op.

The new interface generally reduces boilerplate when constructing a new model, and most importantly the new interface is natural and consistent for different model and basis. An example in 1d-Heisenberg.ipynb:

ham_terms = []
for ispin in range(nspin-1):
    op1 = Op("sigma_z sigma_z", [ispin, ispin+1], 1.0/4)
    op2 = Op("sigma_+ sigma_-", [ispin, ispin+1], 1.0/2)
    op3 = Op("sigma_- sigma_+", [ispin, ispin+1], 1.0/2)
    ham_terms.extend([op1, op2, op3])

# set the spin order and local basis
basis = [BasisHalfSpin(i) for i in range(nspin)]

# construct Hamiltonian MPO
model = Model(basis, ham_terms)
mpo = Mpo(model)

By now the PR is almost complete. I understand the content could be overwhelming. I could introduce the changes in a coding group meeting if necessary.

jjren · 2020-07-28T08:43:34Z

It's a rather big update. Thanks, Weitang!

liwt31 · 2020-07-28T08:44:41Z

Now the only problem is the SVD compress test for MPO in test_mp.py. My opinion is still that it is not a big issue and I'm happy to be corrected.

jjren · 2020-08-10T06:16:52Z

renormalizer/model/phonon.py

-        self.qbtrunc = qbtrunc
-        self.base = int(round(n_phys_dim ** (1.0 / nqboson)))
-        self.hartree = hartree
+
        if hartree:


Should we remove the hartree part?
The Hatree (SCF, TDH) is actually limited to Holstein model.
I think the same calculations could be carried out with DMRG/TD-DMRG with M=1.

Indeed. I think it's OK to remove them.

jjren · 2020-08-10T06:18:01Z

renormalizer/model/phonon.py

@@ -161,7 +154,7 @@ def e0(self):

    @property
    def is_simple(self):
-        return self.force3rd == (0.0, 0.0) and self.nqboson == 1 and self.omega[0] == self.omega[1]
+        return self.force3rd == (0.0, 0.0)  and self.omega[0] == self.omega[1]


I think the force3d is not needed anymore.

It is now only used in TDH module. This feature will be removed along with the module.

jjren · 2020-08-10T06:25:04Z

renormalizer/mps/mp.py

-            mt.sigmaqn = np.zeros(mt.pdim_prod, dtype=np.int)
-        else:
-            mt.sigmaqn = self._get_sigmaqn(idx)
+            mt.sigmaqn = np.zeros_like(mt.sigmaqn)


Is use_dummy_qn necessary or not now? we could set sigmaqn=0 to achieve the same purpose, right?

I'm not sure about this. I'll try to remove this to see if it is necessary or not.

I think it would be more consistent to use sigmaqn=0 if not considering the quantum number.

jjren · 2020-08-10T06:43:13Z

renormalizer/mps/tests/test_mpo.py

+
+
+@pytest.mark.parametrize("nsites", [5, 10])
+# More states make MPO representation not efficient


More sites?

jjren · 2020-08-10T08:25:55Z

renormalizer/model/op.py

+        self.symbol: str = symbol
+        # replace " + " so that " " can be used to split the symbol ("b^\dagger + b")
+        # the logic of Op is based on multiplication of symbols. So special treatment on addition is inevitable
+        self.split_symbol : List[str] = [s.replace("plus", " + ") for s in symbol.replace(" + ", "plus").split(" ")]


Do you think it is better to replace " + " in defining "b^\dagger + b" with "b^\dagger+b"?

Well, this is a possible improvement but I usually prefer b^\dagger + b in writing LaTeX. In the near future, I don't think the change is necessary

jjren · 2020-08-10T09:28:26Z

renormalizer/mps/mpo.py

+    # Construct mapping from easy-to-manipulate integer to actual Op
+    # There has to be an Identity for the algorithm to work
+    # even if Identity is not in the table
+    primary_ops = {0: Op.identity()}
+    # unique operators with DoF names taken into consideration
+    # The inclusion of DoF names is necessary fo multi-dof basis.
+    unique_op: Set[Op] = set(np.array(table).ravel())
+    if primary_ops[0] in unique_op:
+        unique_op.remove(primary_ops[0])


I don't understand why Op.identity is necessary here (restricting the key to be 0).
Though the left-most in_ops is an identity operator physically, it is technically just a placeholder with qn=0 and corresponds to an additional column np.zeros attached to op_table to make the logic more consistent when looping from 0 to N-1.
For example primary_ops = {0:xxx, 1: op.indentity()} will also obtain the correct answer.

The key could be anything. The problem is that if the operator "identity" is not present in the model, there will be actually no primary_ops key for the operator "identity". This is not a serious issue because the algorithm still works as you've pointed out, but I feel somewhat unhappy with the situation. I agree that the resulting code is ugly and hard to grasp. I'll clean it up and add a few comments on the code.

ok. If you think it is necessary to include operator "identity" in primary_ops, I think the key should be 0, to be consistent with the first and the last artificial column.

liwt31 · 2020-08-16T03:38:47Z

In this commit, the TDH module is removed. The removal of use_dummy_qn is not as easy as the removal of the TDH module. Apart from cases when qn is not preserved, the attribute is also used in the conj_trans method for MpDm as a result of the simple qn design (only one set of qn available). To remove conj_trans I also choose to remove the MpDmFull and SuperMPO class. MpDmFull relies on conj_trans and the dissipation dynamics are rarely used.

After this PR I think it'll be appropriate to list the obsolete features in the project wiki for future reference.

jjren · 2020-08-20T03:35:14Z

I‘m ok with this PR. The qn problem in mpdm could be fixed in the future.
I plan to update the qn system in the near future to support more quantum numbers by replacing the single int with a tuple.
After that, mpdm could be characterized by the qn of <bra| and qn of |ket>. The transpose operation could work then.

liwt31 · 2020-08-20T05:15:41Z

The outdated CI information sees to be a persisted issue. I'm merging this PR because CI results have been manually checked.

liwt31 added 2 commits July 15, 2020 23:46

Use MolList2 as when construcing MPO

ab211fd

make MolList a subclass of MolList2 and rename it to HolsteinModel

9293d9d

implement VibronicModel

e112ab9

liwt31 added 3 commits July 17, 2020 10:07

implement SpinBosonModel

778525c

Implement `SpinBosonModel` as a subclass of `MolList2`. Original interface in `HolsteinModel` is removed. Original `SpinBosonModel` in `sbm/sbm.py` is renamed to `SpinBosonDynamics`.

remove hybrid TDH algorithm

4dd41ff

fix ft_cv regression and set logging level in example

d6d6332

liwt31 added 3 commits July 23, 2020 15:28

rename MolList2 to Model

8ff4fbc

New interface for general Model

5d6b3a9

fix import numpy bug in h2o_qc.py

59b5052

add test for mpo construction

38ab90b

liwt31 changed the title ~~[Work in Progress] Use Mollist2 as internal system information data structure~~ Use Mollist2 as internal system information data structure Jul 28, 2020

add print symbolic mpo

1ea7177

jjren reviewed Aug 10, 2020

View reviewed changes

remove TDH module and use_dummy_qn

a85ca3f

jjren approved these changes Aug 20, 2020

View reviewed changes

Merge branch 'master' into clean_up_mpo

14b2b10

liwt31 merged commit c6ea6d8 into master Aug 20, 2020

jjren deleted the clean_up_mpo branch October 8, 2020 04:43

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use Mollist2 as internal system information data structure #103

Use Mollist2 as internal system information data structure #103

liwt31 commented Jul 15, 2020 •

edited by jjren

Loading

liwt31 commented Jul 16, 2020 •

edited

Loading

liwt31 commented Jul 16, 2020

liwt31 commented Jul 19, 2020

liwt31 commented Jul 28, 2020 •

edited

Loading

jjren commented Jul 28, 2020

liwt31 commented Jul 28, 2020

jjren Aug 10, 2020 •

edited

Loading

liwt31 Aug 13, 2020

jjren Aug 10, 2020

liwt31 Aug 13, 2020

jjren Aug 10, 2020

liwt31 Aug 13, 2020

jjren Aug 14, 2020

jjren Aug 10, 2020

liwt31 Aug 13, 2020

jjren Aug 10, 2020

liwt31 Aug 16, 2020

jjren Aug 17, 2020

jjren Aug 10, 2020

liwt31 Aug 13, 2020

jjren Aug 14, 2020 •

edited

Loading

liwt31 commented Aug 16, 2020

jjren commented Aug 20, 2020

liwt31 commented Aug 20, 2020



		@pytest.mark.parametrize("nsites", [5, 10])
		# More states make MPO representation not efficient

Use Mollist2 as internal system information data structure #103

Use Mollist2 as internal system information data structure #103

Conversation

liwt31 commented Jul 15, 2020 • edited by jjren Loading

liwt31 commented Jul 16, 2020 • edited Loading

liwt31 commented Jul 16, 2020

liwt31 commented Jul 19, 2020

liwt31 commented Jul 28, 2020 • edited Loading

jjren commented Jul 28, 2020

liwt31 commented Jul 28, 2020

jjren Aug 10, 2020 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jjren Aug 14, 2020 • edited Loading

Choose a reason for hiding this comment

liwt31 commented Aug 16, 2020

jjren commented Aug 20, 2020

liwt31 commented Aug 20, 2020

liwt31 commented Jul 15, 2020 •

edited by jjren

Loading

liwt31 commented Jul 16, 2020 •

edited

Loading

liwt31 commented Jul 28, 2020 •

edited

Loading

jjren Aug 10, 2020 •

edited

Loading

jjren Aug 14, 2020 •

edited

Loading