Training a quantum circuit with PyTorch

Author: Juan Miguel Arrazola — Posted: 11 October 2019. Last updated: 25 January 2021.

In this notebook, we build and optimize a circuit to prepare arbitrary single-qubit states, including mixed states. Along the way, we also show how to:

  1. Construct compact expressions for circuits composed of many layers.

  2. Succinctly evaluate expectation values of many observables.

  3. Estimate expectation values from repeated measurements, as in real hardware.

The most general state of a qubit is represented in terms of a positive semi-definite density matrix \(\rho\) with unit trace. The density matrix can be uniquely described in terms of its three-dimensional Bloch vector \(\vec{a}=(a_x, a_y, a_z)\) as:


where \(\sigma_x, \sigma_y, \sigma_z\) are the Pauli matrices. Any Bloch vector corresponds to a valid density matrix as long as \(\|\vec{a}\|\leq 1\).

The purity of a state is defined as \(p=\text{Tr}(\rho^2)\), which for a qubit is bounded as \(1/2\leq p\leq 1\). The state is pure if \(p=1\) and maximally mixed if \(p=1/2\). In this example, we select the target state by choosing a random Bloch vector and renormalizing it to have a specified purity.

To start, we import PennyLane, NumPy, and PyTorch for the optimization:

import pennylane as qml
import numpy as np
import torch
from torch.autograd import Variable

# we generate a three-dimensional random vector by sampling
# each entry from a standard normal distribution
v = np.random.normal(0, 1, 3)

# purity of the target state
purity = 0.66

# create a random Bloch vector with the specified purity
bloch_v = Variable(
    torch.tensor(np.sqrt(2 * purity - 1) * v / np.sqrt(np.sum(v ** 2))),

# array of Pauli matrices (will be useful later)
Paulis = Variable(torch.zeros([3, 2, 2], dtype=torch.complex128), requires_grad=False)
Paulis[0] = torch.tensor([[0, 1], [1, 0]])
Paulis[1] = torch.tensor([[0, -1j], [1j, 0]])
Paulis[2] = torch.tensor([[1, 0], [0, -1]])

Unitary operations map pure states to pure states. So how can we prepare mixed states using unitary circuits? The trick is to introduce additional qubits and perform a unitary transformation on this larger system. By “tracing out” the ancilla qubits, we can prepare mixed states in the target register. In this example, we introduce two additional qubits, which suffices to prepare arbitrary states.

The ansatz circuit is composed of repeated layers, each of which consists of single-qubit rotations along the \(x, y,\) and \(z\) axes, followed by three CNOT gates entangling all qubits. Initial gate parameters are chosen at random from a normal distribution. Importantly, when declaring the layer function, we introduce an input parameter \(j\), which allows us to later call each layer individually.

# number of qubits in the circuit
nr_qubits = 3
# number of layers in the circuit
nr_layers = 2

# randomly initialize parameters from a normal distribution
params = np.random.normal(0, np.pi, (nr_qubits, nr_layers, 3))
params = Variable(torch.tensor(params), requires_grad=True)

# a layer of the circuit ansatz
def layer(params, j):
    for i in range(nr_qubits):
        qml.RX(params[i, j, 0], wires=i)
        qml.RY(params[i, j, 1], wires=i)
        qml.RZ(params[i, j, 2], wires=i)

    qml.CNOT(wires=[0, 1])
    qml.CNOT(wires=[0, 2])
    qml.CNOT(wires=[1, 2])

Here, we use the default.qubit device to perform the optimization, but this can be changed to any other supported device.

dev = qml.device("default.qubit", wires=3)

When defining the QNode, we introduce as input a Hermitian operator \(A\) that specifies the expectation value being evaluated. This choice later allows us to easily evaluate several expectation values without having to define a new QNode each time.

Since we will be optimizing using PyTorch, we configure the QNode to use the PyTorch interface:

@qml.qnode(dev, interface="torch")
def circuit(params, A):

    # repeatedly apply each layer in the circuit
    for j in range(nr_layers):
        layer(params, j)

    # returns the expectation of the input matrix A on the first qubit
    return qml.expval(qml.Hermitian(A, wires=0))

Our goal is to prepare a state with the same Bloch vector as the target state. Therefore, we define a simple cost function

\[C = \sum_{i=1}^3 \left|a_i-a'_i\right|,\]

where \(\vec{a}=(a_1, a_2, a_3)\) is the target vector and \(\vec{a}'=(a'_1, a'_2, a'_3)\) is the vector of the state prepared by the circuit. Optimization is carried out using the Adam optimizer. Finally, we compare the Bloch vectors of the target and output state.

# cost function
def cost_fn(params):
    cost = 0
    for k in range(3):
        cost += torch.abs(circuit(params, Paulis[k]) - bloch_v[k])

    return cost

# set up the optimizer
opt = torch.optim.Adam([params], lr=0.1)

# number of steps in the optimization routine
steps = 200

# the final stage of optimization isn't always the best, so we keep track of
# the best parameters along the way
best_cost = cost_fn(params)
best_params = np.zeros((nr_qubits, nr_layers, 3))

print("Cost after 0 steps is {:.4f}".format(cost_fn(params)))

# optimization begins
for n in range(steps):
    loss = cost_fn(params)

    # keeps track of best parameters
    if loss < best_cost:
        best_cost = loss
        best_params = params

    # Keep track of progress every 10 steps
    if n % 10 == 9 or n == steps - 1:
        print("Cost after {} steps is {:.4f}".format(n + 1, loss))

# calculate the Bloch vector of the output state
output_bloch_v = np.zeros(3)
for l in range(3):
    output_bloch_v[l] = circuit(best_params, Paulis[l])

# print results
print("Target Bloch vector = ", bloch_v.numpy())
print("Output Bloch vector = ", output_bloch_v)


Cost after 0 steps is 1.0179
Cost after 10 steps is 0.1467
Cost after 20 steps is 0.0768
Cost after 30 steps is 0.0813
Cost after 40 steps is 0.0807
Cost after 50 steps is 0.0940
Cost after 60 steps is 0.0614
Cost after 70 steps is 0.0932
Cost after 80 steps is 0.0455
Cost after 90 steps is 0.0752
Cost after 100 steps is 0.0301
Cost after 110 steps is 0.0363
Cost after 120 steps is 0.1332
Cost after 130 steps is 0.0687
Cost after 140 steps is 0.0505
Cost after 150 steps is 0.0800
Cost after 160 steps is 0.0644
Cost after 170 steps is 0.0813
Cost after 180 steps is 0.0592
Cost after 190 steps is 0.0502
Cost after 200 steps is 0.0573
Target Bloch vector =  [ 0.33941241 -0.09447812  0.44257553]
Output Bloch vector =  [ 0.3070776  -0.07421944  0.47393046]

About the author

Juan Miguel Arrazola

Juan Miguel Arrazola

Juan Miguel is the Algorithms Senior Team Lead at Xanadu. His work focuses on developing quantum algorithms and related software for simulating molecules and materials.

Total running time of the script: ( 0 minutes 9.352 seconds)

Gallery generated by Sphinx-Gallery