Adds a convenience `prob()` measurement function to return the probabilities of all computational basis states in a single QNode evaluation #432

josh146 · 2019-11-21T21:56:46Z

Context:

Currently, due to the 'expectation of observable' abstraction we use, the only way to return the probabilities from a QNode is to use a single QNode per computational basis state, and use return qml.expval(qml.Hermitian(|x><x|, wires=All)). This can be computationally expensive, especially if the size of each projector is huge.

However, probability is already computed via a single QNode evaluation (analytically on simulators and estimated on hardware), so there is no reason not to provide a way to return it from QNodes. Furthermore, since it is simply an array of expectation values of projectors, it supports analytic differentiation. That said, it does necessitate a slightly different UI, since it doesn't correspond to a single expectation of a single observable (as we currently require).

Example:

import numpy as np
import pennylane as qml

dev = qml.device("default.qubit", wires=2)

@qml.qnode(dev)
def circuit(x):
    qml.Hadamard(wires=0)
    qml.RY(x, wires=0)
    qml.RX(x, wires=1)
    qml.CNOT(wires=[0, 1])
    return qml.probs(wires=[0])

print("Marginal probability of wire 1:", circuit(0.2))
print("Analytic jacobian:", circuit.jacobian([0.2], method='A').flatten())
print("Numeric jacobian:", circuit.jacobian([0.2], method='F').flatten())

Output:

Marginal probability of wire 1: [0.40066533 0.59933467]
Analytic jacobian: [-0.49003329  0.49003329]
Numeric jacobian: [-0.49003329  0.49003328]

Description of the Change:

Adds the qml.probs() measurement function (from a UI perspective, roughly equivalent to the expval, sample, var measurement functions).
Unlike the other measurement functions, it does not accept an observable. Instead, probs() accepts a required argument wires, determining the marginal probabilities to return. prob(wires=range(num_wires)) will return the full joint probability distribution.
Unlike the other measurement functions, which return a scalar, probs(wires=range(num_wires)) will return an array of size (2**num_wires,)
Adds a new observable return type enum to qml.operation, Probability.
If an observable with obs.return_type=Probability is present, the QNode has a new logic flow:
- self.output_conversion = np.squeeze (to avoid superfluous nesting)
- self.output_dim = 2**len(obs.wires)
- Device.execute() returns the result of self.probability(wires=wires).
The ability to return marginal probabilities has been added to DefaultQubit.probability().

Benefits:

Now possible to return and differentiate the probabilities of a quantum state, with very little modification.
UI should apply equally well to the CV case, with the addition of a cutoff argument.

Possible Drawbacks:

Ordering of the computational basis states is implicit (attempting to label the states would break autodiff). This is instead explained in the documentation.

Related GitHub Issues: n/a

codecov · 2019-11-21T22:01:18Z

Codecov Report

❗ No coverage uploaded for pull request base (master@6642817). Click here to learn what that means.
The diff coverage is 100%.

@@            Coverage Diff            @@
##             master     #432   +/-   ##
=========================================
  Coverage          ?   99.29%           
=========================================
  Files             ?       45           
  Lines             ?     3426           
  Branches          ?        0           
=========================================
  Hits              ?     3402           
  Misses            ?       24           
  Partials          ?        0

Impacted Files	Coverage Δ
pennylane/plugins/default_qubit.py	`99.45% <100%> (ø)`
pennylane/_device.py	`100% <100%> (ø)`
pennylane/qnodes/base.py	`100% <100%> (ø)`
pennylane/measure.py	`100% <100%> (ø)`
pennylane/operation.py	`99.62% <100%> (ø)`

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 6642817...df134b4. Read the comment docs.

antalszava · 2019-12-20T19:59:32Z

pennylane/plugins/default_qubit.py


-        return OrderedDict(zip(states, probs))
+        wires = wires or range(self.num_wires)
+        wires = np.hstack(wires)


How come this line would be needed? One of the reasons could be flattening wires here, however, would it not be required to have wires already as a flat list or an array? (Also this performs the implicit conversion on range(self.num_wires))

Good question. I can't remember why I needed to add this line 🤔

antalszava · 2019-12-20T20:20:20Z

pennylane/measure.py

+
+    This measurement function accepts no observables, and instead
+    instructs the QNode to return a flat array containing the
+    probabilities of each quantum state.


As an extra sentence, it might be worth adding something like the explanation in the example such as
The dimension of the computational basis states is equal to the number of wires specified.

Updated to say

Marginal probabilities may also be requested by restricting the wires to a subset of the full system; the size of the returned array will be ``[2**len(wires)]``.

antalszava · 2019-12-20T20:36:37Z

pennylane/measure.py

+        wires (Sequence[int] or int): the wire the operation acts on
+    """
+    # pylint: disable=protected-access
+    op = qml.Identity(wires=wires, do_queue=False)


Currently qml.Identity inherits from CVObservable, how come this works in the qubit case? Would have expected that there is a check somewhere that would throw an error

It looks the qnodes/base.py explicitly makes an exception for Identity:

# True if op is a CV, False if it is a discrete variable (Identity could be either) are_cvs = [ isinstance(op, CV) for op in self.queue + list(res) if not isinstance(op, qml.Identity) ]

antalszava · 2019-12-20T20:39:11Z

tests/test_prob.py

+@pytest.fixture
+def init_state(scope="session"):
+    """Fixture that creates an initial state"""
+    def _init_state(n):
+        """An initial state over n wires"""
+        state = np.random.random([2 ** n]) + np.random.random([2 ** n]) * 1j
+        state /= np.linalg.norm(state)
+        return state
+
+    return _init_state


Maybe it would be nice to have this in the conftest.py.

True, that's a good point. I'm a bit hesitant to put it in the conftest.py file right away, as the conftest file is a bit... messy, and needs cleaning up (there are two many fixtures which aren't used or aren't needed).

antalszava

Looks really nice @josh146 ! Had a couple of minor comments, most we have already discussed

initial commit

9a0bdc5

josh146 added enhancement ✨ New feature or request WIP 🚧 Work-in-progress labels Nov 21, 2019

linting

7598bcd

add tests

f5c5953

josh146 added review-ready 👌 PRs which are ready for review by someone from the core team. and removed WIP 🚧 Work-in-progress labels Nov 21, 2019

josh146 added 6 commits November 27, 2019 13:41

Merge branch 'master' into prob_fn

773da6a

Merge branch 'master' into prob_fn

7f10806

merge master

4cc2e53

renamed prob to probs

357a7d2

update docs

cabb8f0

fix bugs

e49577d

josh146 requested review from antalszava and mariaschuld December 20, 2019 18:29

josh146 and others added 2 commits December 20, 2019 14:27

update changelog

20401a7

Merge branch 'master' into prob_fn

2507599

antalszava reviewed Dec 20, 2019

View reviewed changes

antalszava approved these changes Dec 20, 2019

View reviewed changes

josh146 added 2 commits December 20, 2019 17:18

suggested changes

6dff8c9

Merge branch 'master' into prob_fn

df134b4

josh146 merged commit 5dbbdcc into master Dec 24, 2019

josh146 deleted the prob_fn branch December 24, 2019 00:24

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adds a convenience `prob()` measurement function to return the probabilities of all computational basis states in a single QNode evaluation #432

Adds a convenience `prob()` measurement function to return the probabilities of all computational basis states in a single QNode evaluation #432

josh146 commented Nov 21, 2019 •

edited

Loading

codecov bot commented Nov 21, 2019 •

edited

Loading

antalszava Dec 20, 2019

josh146 Dec 20, 2019

antalszava Dec 20, 2019

josh146 Dec 20, 2019

antalszava Dec 20, 2019

josh146 Dec 20, 2019

antalszava Dec 20, 2019

josh146 Dec 20, 2019

antalszava left a comment

Adds a convenience prob() measurement function to return the probabilities of all computational basis states in a single QNode evaluation #432

Adds a convenience prob() measurement function to return the probabilities of all computational basis states in a single QNode evaluation #432

Conversation

josh146 commented Nov 21, 2019 • edited Loading

codecov bot commented Nov 21, 2019 • edited Loading

Codecov Report

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

antalszava left a comment

Choose a reason for hiding this comment

Adds a convenience `prob()` measurement function to return the probabilities of all computational basis states in a single QNode evaluation #432

Adds a convenience `prob()` measurement function to return the probabilities of all computational basis states in a single QNode evaluation #432

josh146 commented Nov 21, 2019 •

edited

Loading

codecov bot commented Nov 21, 2019 •

edited

Loading