rigetticomputing/grove

Switch branches/tags
Nothing to show
Fetching contributors…
Cannot retrieve contributors at this time
419 lines (287 sloc) 12.4 KB

Variational-Quantum-Eigensolver (VQE)

Overview

The Variational-Quantum-Eigensolver (VQE) [1, 2] is a quantum/classical hybrid algorithm that can be used to find eigenvalues of a (often large) matrix H. When this algorithm is used in quantum simulations, H is typically the Hamiltonian of some system [3, 4, 5]. In this hybrid algorithm a quantum subroutine is run inside of a classical optimization loop.

The quantum subroutine has two fundamental steps:

1. Prepare the quantum state |\Psi(\vec{\theta})\rangle, often called the ansatz.
2. Measure the expectation value \langle\,\Psi(\vec{\theta})\,|\,H\,|\,\Psi(\vec{\theta})\,\rangle.

The variational principle ensures that this expectation value is always greater than the smallest eigenvalue of H.

This bound allows us to use classical computation to run an optimization loop to find this eigenvalue:

1. Use a classical non-linear optimizer to minimize the expectation value by varying ansatz parameters \vec{\theta}.
2. Iterate until convergence.

Practically, the quantum subroutine of VQE amounts to preparing a state based off of a set of parameters \vec{\theta} and performing a series of measurements in the appropriate basis. The paramaterized state (or ansatz) preparation can be tricky in these algorithms and can dramatically affect performance. Our VQE module allows any Python function that returns a pyQuil program to be used as an ansatz generator. This function is passed into vqe_run as the variational_state_evolve argument. More details are in the source documentation.

Measurements are then performed on these states based on a Pauli operator decomposition of H. Using Quil, these measurements will end up in classical memory. Doing this iteratively followed by a small amount of postprocessing, one may compute a real expectation value for the classical optimizer to use.

Below there is a very small first example of VQE and Grove's implementation of the Quantum Approximate Optimization Algorithm QAOA also makes use of the VQE module.

Basic Usage

Here we will take you through an example of a very small variational quantum eigensolver problem. In this example we will use a quantum circuit that consists of a single parametrized gate to calculate an eigenvalue of the Pauli Z matrix.

First we import the necessary pyQuil modules to construct our ansatz pyQuil program.

from pyquil.quil import Program
import pyquil.api as api
from pyquil.gates import *
qvm = api.QVMConnection()

Any Python function that takes a list of numeric parameters and outputs a pyQuil program can be used as an ansatz function. We will see some more examples of this later. For now, we just take a parameter list with a single parameter.

def small_ansatz(params):
return Program(RX(params[0], 0))

print(small_ansatz([1.0]))
RX(1.0) 0


This small_ansatz function is our \Psi(\vec{\theta}). To construct the Hamiltonian that we wish to simulate, we use the pyquil.paulis module.

from pyquil.paulis import sZ
initial_angle = [0.0]
# Our Hamiltonian is just \sigma_z on the zeroth qubit
hamiltonian = sZ(0)

We now use the vqe module in Grove to construct a VQE object to perform our algorithm. In this example, we use scipy.optimize.minimize() with Nelder-Mead as our classical minimizer, but you can choose other parameters or write your own minimizer.

from grove.pyvqe.vqe import VQE
from scipy.optimize import minimize
import numpy as np

vqe_inst = VQE(minimizer=minimize,
minimizer_kwargs={'method': 'nelder-mead'})

Before we run the minimizer, let us look manually at what expectation values \langle\,\Psi(\vec{\theta})\,|\,H\,|\,\Psi(\vec{\theta})\,\rangle we calculate for fixed parameters of \vec{\theta}.

angle = 2.0
vqe_inst.expectation(small_ansatz([angle]), hamiltonian, None, qvm)
-0.4161468365471423


The expectation value was calculated by running the pyQuil program output from small_ansatz, saving the wavefunction, and using that vector to calculate the expectation value. We can sample the wavefunction as you would on a quantum computer by passing an integer, instead of None, as the samples argument of the expectation() method.

angle = 2.0
vqe_inst.expectation(small_ansatz([angle]), hamiltonian, 10000, qvm)
-0.42900000000000005


We can loop over a range of these angles and plot the expectation value.

angle_range = np.linspace(0.0, 2 * np.pi, 20)
data = [vqe_inst.expectation(small_ansatz([angle]), hamiltonian, None, qvm)
for angle in angle_range]

import matplotlib.pyplot as plt
plt.ylabel('Expectation value')
plt.plot(angle_range, data)
plt.show()

Now with sampling...

angle_range = np.linspace(0.0, 2 * np.pi, 20)
data = [vqe_inst.expectation(small_ansatz([angle]), hamiltonian, 1000, qvm)
for angle in angle_range]

import matplotlib.pyplot as plt
plt.ylabel('Expectation value')
plt.plot(angle_range, data)
plt.show()

We can compare this plot against the value we obtain when we run the our variational quantum eigensolver.

result = vqe_inst.vqe_run(small_ansatz, hamiltonian, initial_angle, None, qvm=qvm)
print(result)
{'fun': -0.99999999954538055, 'x': array([ 3.1415625])}


Running Noisy VQE

A great thing about VQE is that it is somewhat insensitive to noise. We can test this out by running the previous algorithm on a noisy qvm.

Remember that Pauli channels are defined as a list of three probabilities that correspond to the probability of a random X, Y, or Z gate respectively. First we'll study the impact of a channel that has the same probability of each random Pauli.

pauli_channel = [0.1, 0.1, 0.1] #10% chance of each gate at each timestep
noisy_qvm = api.QVMConnection(gate_noise=pauli_channel)

Let us check that this QVM has noise:

p = Program(X(0), X(1)).measure(0, [0]).measure(1, [1])
noisy_qvm.run(p, [0, 1], 10)
[[1, 1],
[0, 1],
[1, 0],
[0, 1],
[0, 0],
[1, 1],
[0, 1],
[1, 0],
[1, 0],
[0, 1]]


We can run the VQE under noise. Let's modify the classical optimizer to start with a larger simplex so we don't get stuck at an initial minimum.

vqe_inst.minimizer_kwargs = {'method': 'Nelder-mead', 'options': {'initial_simplex': np.array([[0.0], [0.05]]), 'xatol': 1.0e-2}}
result = vqe_inst.vqe_run(small_ansatz, hamiltonian, initial_angle, samples=10000, qvm=noisy_qvm)
print(result)
{'fun': 0.5875999999999999, 'x': array([ 0.01874886])}


10% error is a huge amount of error! We can plot the effect of increasing noise on the result of this algorithm:

data = []
noises = np.linspace(0.0, 0.01, 4)
for noise in noises:
pauli_channel = [noise] * 3
noisy_qvm = api.QVMConnection(gate_noise=pauli_channel)
# We can pass the noise params directly into the vqe_run instead of passing the noisy connection
result = vqe_inst.vqe_run(small_ansatz, hamiltonian, initial_angle,
gate_noise=pauli_channel)
data.append(result['fun'])
plt.xlabel('Noise level %')
plt.ylabel('Eigenvalue')
plt.plot(noises, data)
plt.show()

It looks like this algorithm is pretty robust to noise up until 0.6% error. However measurement noise might be a different story.

meas_channel = [0.1, 0.1, 0.1] #10% chance of each gate at each measurement
noisy_meas_qvm = api.QVMConnection(measurement_noise=meas_channel)

Measurement noise has a different effect:

p = Program(X(0), X(1)).measure(0, [0]).measure(1, [1])
noisy_meas_qvm.run(p, [0, 1], 10)
[[1, 1],
[1, 1],
[1, 1],
[1, 1],
[1, 1],
[1, 1],
[0, 1],
[1, 0],
[1, 1],
[1, 0]]

data = []
noises = np.linspace(0.0, 0.01, 4)
for noise in noises:
meas_channel = [noise] * 3
noisy_qvm = api.QVMConnection(measurement_noise=meas_channel)
result = vqe_inst.vqe_run(small_ansatz, hamiltonian, initial_angle, samples=10000, qvm=noisy_qvm)
data.append(result['fun'])
plt.xlabel('Noise level %')
plt.ylabel('Eigenvalue')
plt.plot(noises, data)
plt.show()

We see this particular VQE algorithm is generally more sensitive to measurement noise than gate noise.

More Sophisticated Ansatzes

Because we are working with Python, we can leverage the full language to make much more sophisticated ansatzes for VQE. As an example we can easily change the number of gates.

def smallish_ansatz(params):
return Program(RX(params[0], 0), RX(params[1], 0))

print(smallish_ansatz([1.0, 2.0]))
RX(1.0) 0
RX(2.0) 0

vqe_inst = VQE(minimizer=minimize,
initial_angles = [1.0, 1.0]
result = vqe_inst.vqe_run(smallish_ansatz, hamiltonian, initial_angles, None, qvm=qvm)
print(result)
{'fun': -1.0000000000000004, 'x': array([ 1.61767133,  1.52392133])}


We can even dynamically change the gates in the circuit based on a parameterization:

def variable_gate_ansatz(params):
gate_num = int(np.round(params[1])) # for scipy.minimize params must be floats
p = Program(RX(params[0], 0))
for gate in range(gate_num):
p.inst(X(0))
return p

print(variable_gate_ansatz([0.5, 3]))
RX(0.5) 0
X 0
X 0
X 0

initial_params = [1.0, 3]
result = vqe_inst.vqe_run(variable_gate_ansatz, hamiltonian, initial_params, None, qvm=qvm)
print(result)
{'fun': -1.0, 'x': array([  2.65393312e-09,   3.42891875e+00])}


Note that the restriction that the ansatz function take a single list of floats as parameters only comes from our choice of minimizer (this is what scipy.optimize.minimize takes). One could easily imagine writing a custom minimizer that takes more sophisticated forms of arguments.

This concludes our brief tour of VQE. There is a lot of fascinating literature about this algorithm out there and we encourage you to both explore those topics as well as come up with new ideas using this library. Let us know if you have ideas about anything that you would like to see added!

Here are some links to get you started:

Source Code Docs

Here you can find documentation for the different submodules in pyVQE.

grove.pyvqe.vqe

.. automodule:: grove.pyvqe.vqe
:members:
:undoc-members:
:show-inheritance: