Adds Double Excitation operations #1123

ixfoduap · 2021-03-04T20:58:32Z

Context:
Current tools for constructing quantum circuits for quantum chemistry are inefficient. By relying on fermionic-to-qubit mappings to define operations and circuit decompositions to compute gradients, the resulting circuits for creating single and double excitation operations lead to lengthy computations. Through research work, we've identified that Givens rotations perform the same role as fermionic excitation operations, while allowing for simple analytical gradients and faster implementation in hardware and simulators.

Description of the Change:
This PR adds the 4-qubit DoubleExcitation operation with support across all native devices. Its analytical gradient requires a decomposition into two gates, DoubleExcitationPlus and DoubleExcitationMinus, which are also added as part of this PR.

Benefits:
Much faster implementation of quantum circuits for quantum chemistry.

Note:
The DoubleExcitationPlus and DoubleExcitationMinus have generators G such that G^2=1 (self-inverse). This means they satisfy the standard parameter-shift rule and no custom gradient formula needs to be passed

josh146

Just did a quick run through, looks great so far!

Small comment re: autograd, jax, and tensorflow: it will be important to have some backprop tests, just to ensure everything remains differentiable!

josh146 · 2021-03-05T13:22:28Z

.github/CHANGELOG.md

+- Added the `DoubleExcitation` four-qubit operation, which is particularly useful for quantum
+  chemistry applications. [(#1121)](https://github.com/PennyLaneAI/pennylane/pull/1121)


Minor: the second half of the sentence "which is particularly useful for quantum chemistry applications", sounds a bit weird to me in a changelog 🤔

Suggest making it more specific (saying in 2-3 extra words why it is useful), or removing 'particularly'?

josh146 · 2021-03-05T13:27:48Z

pennylane/devices/tf_ops.py

+    U = [[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+     [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+     [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+     [0, 0, 0, c, 0, 0, 0, 0, 0, 0, 0, 0, -s, 0, 0, 0],
+     [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+     [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+     [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+     [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0],
+     [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0],
+     [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0],
+     [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0],
+     [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0],
+     [0, 0, 0, s, 0, 0, 0, 0, 0, 0, 0, 0, c, 0, 0, 0],
+     [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0],
+     [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0],
+     [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1]]


is this because tensor indexing doesn't work in tensorflow? You could try using tf.scatter_nd to avoid writing out the full matrix.

Thanks Josh! I took a look and it's still pretty cumbersome to write down the matrix using tf.scatter_nd . As far as I can tell the only way to use tf.scatter is to populate an entire tensor index at a time for a tensor of all zeros. In this case it means populating entire rows. This would still be helpful if some rows where all zeros, but instead it seems we have to specify each full row directly, which is as difficult as writing down the entire matrix. For example, to build the 2x2 identity we'd have to do something like:

indices = tf.constant([[0], [1]]) updates = tf.constant([[1, 0], [0, 1]]) shape = tf.constant([2, 2]) scatter = tf.scatter_nd(indices, updates, shape) print(scatter) >>> tf.Tensor( [[1 0] [0 1]], shape=(2, 2), dtype=int32)

Ah. To be honest I don't 100% understand scatter_nd 😆

josh146 · 2021-03-05T13:28:26Z

pennylane/devices/autograd_ops.py

+    U = np.zeros((16, 16))
+    U[3, 3] = c  # 3 (dec) = 0011 (bin)
+    U[3, 12] = -s  # 12 (dec) = 1100 (bin)
+    U[12, 3] = s
+    U[12, 12] = c


Might be important to check that this remains differentiable --- I think autograd (and maybe JAX) do not like setting array elements via indexing

Thanks for the suggestion! How would that check look like exactly? 🤔

Is this not checked in the device tests?

Ah, so what you will need to do is just add a new integration test, which

Creates a QNode parametrized on the diff method (both parameter-shift and backprop)

Uses the new gate

Differentiates the QNode and compares it against the expected result 🙂

If everything is fine, the tests will pass for both diff methods.

.github/CHANGELOG.md

pennylane/devices/tf_ops.py

ixfoduap · 2021-03-10T22:25:24Z

pennylane/devices/tf_ops.py

+
+    return tf.convert_to_tensor(U)
+
+    return U


Suggested change

return U

codecov · 2021-03-11T21:49:41Z

Codecov Report

Merging #1123 (f23ba82) into master (6aa1228) will increase coverage by 0.02%.
The diff coverage is 97.42%.

@@            Coverage Diff             @@
##           master    #1123      +/-   ##
==========================================
+ Coverage   98.02%   98.05%   +0.02%     
==========================================
  Files         142      142              
  Lines       10398    10522     +124     
==========================================
+ Hits        10193    10317     +124     
  Misses        205      205

Impacted Files	Coverage Δ
pennylane/devices/default_mixed.py	`100.00% <ø> (ø)`
pennylane/devices/default_qubit.py	`100.00% <ø> (ø)`
pennylane/devices/default_qubit_tf.py	`90.66% <90.66%> (ø)`
pennylane/devices/default_qubit_jax.py	`94.02% <94.02%> (ø)`
pennylane/devices/jax_ops.py	`97.22% <97.22%> (+0.85%)`	⬆️
pennylane/devices/autograd_ops.py	`100.00% <100.00%> (ø)`
pennylane/devices/default_qubit_autograd.py	`100.00% <100.00%> (ø)`
pennylane/devices/tf_ops.py	`100.00% <100.00%> (ø)`
pennylane/ops/qubit.py	`98.19% <100.00%> (+0.16%)`	⬆️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6aa1228...f23ba82. Read the comment docs.

tests/ops/test_qubit_ops.py

pennylane/devices/autograd_ops.py

pennylane/devices/jax_ops.py

pennylane/devices/tf_ops.py

pennylane/ops/qubit.py

antalszava

Left some suggestions, nearing approval from my side. 👍 I'd double-check that changes required for the docs are made in a quick other iteration. Checking the docs is great to see that rendering is good for docstrings.

tests/ops/test_qubit_ops.py

.github/CHANGELOG.md

antalszava · 2021-03-19T18:08:02Z

pennylane/devices/autograd_ops.py

+
+
+def DoubleExcitationPlus(phi):
+    r"""


The previous suggestion for the docstrings are applicable here and for other additions too.

pennylane/devices/tf_ops.py

pennylane/ops/qubit.py

tests/ops/test_qubit_ops.py

antalszava · 2021-03-19T18:31:11Z

pennylane/devices/tf_ops.py

+
+
+def DoubleExcitation(phi):
+    r"""


Side-note: formatting of the docstring is required for the docs too, not just for cosmetics:
https://pennylane--1123.org.readthedocs.build/en/1123/code/api/pennylane.devices.tf_ops.DoubleExcitation.html?highlight=doubleexcitation#pennylane.devices.tf_ops.DoubleExcitation

antalszava · 2021-03-19T18:33:41Z

pennylane/ops/qubit.py

+    r"""DoubleExcitationPlus(phi, wires)
+    Double excitation rotation with positive phase-shift outside the rotation subspace.
+
+    .. math:: U_+(\phi) = \begin{bmatrix}


From my browser, the matrix seems to be too big when rendered in the docs:
https://pennylane--1123.org.readthedocs.build/en/1123/code/api/pennylane.DoubleExcitationPlus.html?highlight=doubleexcitationplus#pennylane.DoubleExcitationPlus

agran2018

Thanks @ixfoduap. Left few minor suggestions. Happy to approve after checking them.

pennylane/devices/tf_ops.py

pennylane/ops/qubit.py

agran2018 · 2021-03-19T21:16:38Z

pennylane/devices/default_qubit_jax.py

@@ -145,6 +145,9 @@ def circuit():
        "CRY": jax_ops.CRY,
        "CRZ": jax_ops.CRZ,
        "MultiRZ": jax_ops.MultiRZ,
+        "DoubleExcitation": jax_ops.DoubleExcitation,
+        "DoubleExcitationPlus": jax_ops.DoubleExcitationPlus,
+        "DoubleExcitationMinus": jax_ops.DoubleExcitationMinus,
    }

    C_DTYPE = jnp.complex64


just out of curiosity, why is the precision lowered here as compare with autograd and tensorflow?

@Thenerdstation

Co-authored-by: antalszava <antalszava@gmail.com> Co-authored-by: agran2018 <45397799+agran2018@users.noreply.github.com>

…to double_excitations

antalszava

Looks good to me! 🥇 🥳

antalszava · 2021-03-22T20:32:25Z

pennylane/ops/qubit.py

+    |1100\rangle,|0011\rangle\}`. More precisely, it performs the transformation
+
+    .. math::
+
+        &|0011\rangle \rightarrow \cos(\phi) |0011\rangle - \sin(\phi) |1100\rangle\\
+        &|1100\rangle \rightarrow \cos(\phi) |1100\rangle + \sin(\phi) |0011\rangle,


antalszava · 2021-03-22T20:36:21Z

.github/CHANGELOG.md

+  expectation value calculations, following results from
+  [Kottmann et al.](https://arxiv.org/abs/2011.05938).
+
+


Suggested change

antalszava · 2021-03-22T20:36:44Z

pennylane/devices/autograd_ops.py

+        phi (float): rotation angle
+    Returns:
+        array[float]: Double excitation rotation matrix
+


Suggested change

antalszava · 2021-03-22T20:37:18Z

pennylane/devices/jax_ops.py

+        phi (float): rotation angle
+    Returns:
+        jnp.Tensor[float]: Double excitation rotation matrix
+


Suggested change

antalszava · 2021-03-22T20:37:37Z

pennylane/devices/tf_ops.py

+        phi (float): rotation angle
+    Returns:
+        tf.Tensor[complex]: Double excitation rotation matrix
+


Suggested change

antalszava · 2021-03-22T20:41:19Z

pennylane/ops/qubit.py


+    while leaving all other basis states unchanged.


(Minor): is it worth further specifying that it's the computational basis?

Leaves them all unchanged, so the basis doesn't matter!

agran2018

Thanks @ixfoduap, looks great!.

Adds first verison of double excitation operations

81cc3a5

ixfoduap self-assigned this Mar 4, 2021

ixfoduap requested review from josh146, trbromley and agran2018 March 4, 2021 20:58

excludes ops from reversible test

cd2a788

josh146 reviewed Mar 5, 2021

View reviewed changes

ixfoduap commented Mar 10, 2021

View reviewed changes

.github/CHANGELOG.md Outdated Show resolved Hide resolved

ixfoduap and others added 4 commits March 10, 2021 14:12

Update .github/CHANGELOG.md

892ce8b

Adds gate_data for testing

542654e

Adds first patch of tests

99b91dc

Fixes indent error

2063319

ixfoduap changed the title ~~Adds double excitation operations [WIP, no tests for now]~~ Adds double excitation operations Mar 10, 2021

ixfoduap commented Mar 10, 2021

View reviewed changes

pennylane/devices/tf_ops.py Outdated Show resolved Hide resolved

Update pennylane/devices/tf_ops.py

9ea144b

ixfoduap commented Mar 10, 2021

View reviewed changes

pennylane/devices/tf_ops.py Outdated

return tf.convert_to_tensor(U)

return U

Copy link

Contributor Author

ixfoduap Mar 10, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change

return U

ixfoduap and others added 2 commits March 11, 2021 16:33

Merge branch 'master' into double_excitations

0379d65

black

a2e531f

ixfoduap added 3 commits March 11, 2021 16:54

removes interface tests

499ffe8

fixes docstring warning

267134e

fixes docstring warning

bc3f68d

ixfoduap changed the title ~~Adds double excitation operations~~ Adds Double Excitation operations Mar 12, 2021

ixfoduap and others added 2 commits March 17, 2021 11:49

adds interface tests for plus/minus

5f3eeba

Merge branch 'master' into double_excitations

edf55e1

ixfoduap commented Mar 17, 2021

View reviewed changes

tests/ops/test_qubit_ops.py Outdated Show resolved Hide resolved

tests/ops/test_qubit_ops.py Outdated Show resolved Hide resolved

tests/ops/test_qubit_ops.py Outdated Show resolved Hide resolved

ixfoduap and others added 4 commits March 17, 2021 14:19

Apply suggestions from code review

eb47c9b

adds grad interface tests

55b8bc3

Merge branch 'master' into double_excitations

dde727e

Merge branch 'master' into double_excitations

88a39ea

this should fix it

7809fea

ixfoduap added core Core functionality (Operations, QNodes, CircuitGraphs) qchem ⚛️ Related to the QChem package labels Mar 19, 2021

ixfoduap commented Mar 19, 2021

View reviewed changes

pennylane/devices/autograd_ops.py Outdated Show resolved Hide resolved

pennylane/devices/jax_ops.py Outdated Show resolved Hide resolved

pennylane/devices/tf_ops.py Outdated Show resolved Hide resolved

pennylane/ops/qubit.py Outdated Show resolved Hide resolved

Apply suggestions from code review

b058620

ixfoduap requested review from josh146 and antalszava March 19, 2021 14:04

josh146 removed their request for review March 19, 2021 15:12

antalszava reviewed Mar 19, 2021

View reviewed changes

agran2018 requested changes Mar 19, 2021

View reviewed changes

josh146 and others added 5 commits March 22, 2021 16:28

Merge branch 'master' into double_excitations

47c261e

Merge branch 'master' into double_excitations

6762d30

Apply suggestions from code review

bdbdd76

Co-authored-by: antalszava <antalszava@gmail.com> Co-authored-by: agran2018 <45397799+agran2018@users.noreply.github.com>

changes from code review

89cb97b

Merge branch 'double_excitations' of github.com:XanaduAI/pennylane in…

fb5f082

…to double_excitations

ixfoduap requested review from agran2018 and antalszava March 22, 2021 19:57

ixfoduap added 2 commits March 22, 2021 16:01

improves docstrings

96d44dd

removes useless import

8a8c6aa

antalszava approved these changes Mar 22, 2021

View reviewed changes

agran2018 approved these changes Mar 22, 2021

View reviewed changes

ixfoduap and others added 4 commits March 22, 2021 17:47

Merge branch 'master' into double_excitations

505cb21

black

61ff30c

fix indent

35b1d30

Merge branch 'master' into double_excitations

f23ba82

ixfoduap merged commit c05978e into master Mar 23, 2021

ixfoduap deleted the double_excitations branch March 23, 2021 12:59

antalszava mentioned this pull request May 5, 2021

Define decomposition for single and double excitation operations #1275

Closed

This was referenced May 28, 2021

Refactor UCCSD template using the excitation operations #1370

Closed

Add template to prepare quantum states of molecules using the excitation operations #1383

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adds Double Excitation operations #1123

Adds Double Excitation operations #1123

ixfoduap commented Mar 4, 2021

josh146 left a comment

josh146 Mar 5, 2021

josh146 Mar 5, 2021

ixfoduap Mar 10, 2021

josh146 Mar 11, 2021

josh146 Mar 5, 2021

ixfoduap Mar 10, 2021

josh146 Mar 11, 2021

ixfoduap Mar 10, 2021

codecov bot commented Mar 11, 2021 •

edited

Loading

antalszava left a comment

antalszava Mar 19, 2021

antalszava Mar 19, 2021

antalszava Mar 19, 2021

agran2018 left a comment

agran2018 Mar 19, 2021

josh146 Mar 20, 2021

antalszava left a comment

antalszava Mar 22, 2021

antalszava Mar 22, 2021

antalszava Mar 22, 2021

antalszava Mar 22, 2021

antalszava Mar 22, 2021

antalszava Mar 22, 2021

ixfoduap Mar 22, 2021

agran2018 left a comment

		- Added the `DoubleExcitation` four-qubit operation, which is particularly useful for quantum
		chemistry applications. [(#1121)](https://github.com/PennyLaneAI/pennylane/pull/1121)

		expectation value calculations, following results from
		[Kottmann et al.](https://arxiv.org/abs/2011.05938).

Adds Double Excitation operations #1123

Adds Double Excitation operations #1123

Conversation

ixfoduap commented Mar 4, 2021

josh146 left a 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

codecov bot commented Mar 11, 2021 • edited Loading

Codecov Report

antalszava left a 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

agran2018 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

antalszava left a 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

agran2018 left a comment

Choose a reason for hiding this comment

codecov bot commented Mar 11, 2021 •

edited

Loading