qutip-tensorflow working example

In this blog post I will show an example of use for qutip-tensorflow. We will optimize an arbitrary single qubit gate to perform a full rotation from the state $| 0 \rangle$ to the state $| 1 \rangle$. Even though this is a very simple example with analytical solution, it will show how qutip-tensorflow allows seamless use of QuTiP’s Qobj with TensorFlow’s auto-differentiation features through GradientTape.

Hence we start by importing the modules.

import qutip as qt
import tensorflow as tf
import qutip_tensorflow

Parametrized single qubit gate

We begin defining the parametrized unitary operation (single qubit gate) that we aim to optimize. Besides, we will do this using QuTiP’s native functions. The first challenge we encounter is that these functions return a Qobj represented with the Dense data backend. This is one of the native data backends of QuTiP 5.0, but we want the data to be backed by a tf.Tensor in such a way that we can benefit from TensorFlow’s GradientTape. For this we employ qutip-tensorflow’s TfTensor data backend, which is registered when importing qutip-tensorflow. We then change the underlying data with the to method. As an example, consider Pauli’s X operator:

sx = qt.sigmax()  # Qobj
sx.data # This is an instance of `Dense` class
sx = sx.to('tftensor')  # Transform data to TfTensor
sx.data  # This is an instance of the TfTensor class

When the data is represented as TfTensor, it is possible to access the underlying tf.Tensor with the attribute _tf. However, this should not be needed in most cases, which is way it is accessed as a private attribute.

sx.data._tf  # This is an instance of tf.Tensor

An arbitrary single qubit gate can be composed with a consecutive set of three rotations in the Z-Y-Z directions. The parameters representing the rotation angle of each unitary-rotation are $\alpha$, $\beta$ and $\gamma$. Hence, the single qubit gate we need can be written as:

def single_qubit_gate(alpha, beta, gamma):
    # Pauli matrices backed with TfTensor
    sz = qt.sigmax().to('tftensor')
    sy = qt.sigmay().to('tftensor')
    # Tensorflow does not support automatic casting of real 
    # tf.Variable to complex se we do it manuall.
    alpha = tf.cast(alpha, dtype=tf.complex128)
    beta = tf.cast(beta, dtype=tf.complex128)
    gamma = tf.cast(gamma, dtype=tf.complex128)

    # Note that all this operations are backed with tf.Tensors
    # so they will support backpropagation if alpha, beta and/or gamma
    # are instaces of tf.Variable.
    rot = (1j*alpha/2*sz).expm(dtype='tftensor')
    rot *= (1j*beta/2*sy).expm(dtype='tftensor')
    rot *= (1j*gamma/2*sz).expm(dtype='tftensor')

    return rot

Loss function and minimization

We now define the loss function that we will minimize. This will consist on the evolution of an initial $|0\rangle$ state with the unitary, $U(\alpha, \beta, \gamma)$, defined in single_qubit_gate. We then compute the overlap of the evolved state with the desired $|1\rangle$ state. This defines an output subjected to maximization. We can minimize the output instead if we take the absolute negative value. Hence, the loss function subject to minimization is:

$$ \text{loss}(\alpha, \beta, \gamma) = -| \langle 1 | U(\alpha, \beta, \gamma)| 0 \rangle |$$

Once more, with qutip-tensorflow imported, we can code this function using native QuTiP methods:

# Initial values
# Note that the variables are defined out of the scope of the funtion.
alpha = tf.Variable(0.01, dtype=tf.float64, name='alpha')
beta = tf.Variable(0.01, dtype=tf.float64, name='beta')
gamma = tf.Variable(0.01, dtype=tf.float64, name='gamma')

var_list = [alpha, beta, gamma]

# loss function to be optimized
def loss():
    # initial (|0>) and final (|1>) states   
    initial = qt.basis(2,0).to('tftensor')
    final = qt.basis(2,1).to('tftensor')
    # Unitary gate. Note that variables are defined out of the
    # function scope.
    U = single_qubit_gate(*var_list)

    # Overlap between the inital state and the desired state. 
    overlap = final.dag()*U*initial
    return - tf.abs(overlap)

Now we carry the minimization of the loss function with TensorFlow’s native optimizers:

# Create an optimizer. We choose the SGD method.
opt = tf.keras.optimizers.SGD(learning_rate=1)

for epoch in range(10):
    opt.minimize(loss, var_list=var_list)  # One optimization step
    print(f"Epoch: {epoch+1:2f} | Cost: {loss():3f}")

print(f"Cost: {loss():3f}")  # -0.999992 
print(f"alpha: {alpha.numpy():3f}")  # alpha: 1.566813
print(f"beta: {beta.numpy():3f}")  # beta: 0.397178
print(f"gamma: {gamma.numpy():3f}")  # gamma: 1.566813

As we can see, we obtain a set of parameters, $\alpha$, $\beta$ and $\gamma$ that define, with a good approximation, a unitary gate that transforms the state $|0\rangle$ to the state $| 1 \rangle$.


To finalise this blog post let me show another selling point for qutip-tensorflow which once more relies in the seamless integration we aim to provide between QuTiP and Tensorflow. QuTiP implements a few useful visualization tools. Among these, the Bloch class facilitates plotting the Bloch sphere for a state. For example, we can store the rotations obtained after each epoch in the optimization process and visualize them using the Bloch class1:

# Save initial rotation. Note that we store a Qobj backed with tf.Tensor
initial = qt.basis(2,0).to('tftensor')
state_hist = [single_qubit_gate(*var_list)*initial]

# Create an optimizer.
opt = tf.keras.optimizers.SGD(learning_rate=1)

for epoch in range(10):
    opt.minimize(loss, var_list=var_list)  # One optimization step
    print(f"Epoch: {epoch+1:2f} | Cost: {loss():3f}")
    # Keep track of the evolution after each epoch
# Code adapted from qgrad: 
from qutip import Bloch
import numpy as np
from mpl_toolkits.mplot3d import Axes3D
import matplotlib.animation as animation
import matplotlib.pyplot as plt

plt.rcParams["animation.html"] = "jshtml"
fig = plt.figure()
ax = Axes3D(fig, azim=-40, elev=30)
sphere = Bloch(axes=ax)

def animate(i):
    # Note that the input is a Qobj backed with TfTensor
    return ax

def init():
    sphere.vector_color = ["b"]
    return ax

ani = animation.FuncAnimation(

The code above will output the following animation:

In this animation we see that after a few epochs, the Unitary operation is optimized to perform the desired rotation: $|0\rangle \rightarrow |1\rangle$ .


In this blog post we have seen how QuTiP tensorflow allows us to use QuTiP’s native functions to optimize an operation on Quantum objects. In the future, we aim to extend qutip-tensorflow to seamlessly work with keras.layers and keras.models. This will facilitate writing Machine learning models using QuTiP. Another future goal is to make qutip-tensorflow work with qutip-qip. This will facilitate writing quantum models that rely on quantum circuits.

Note: at the moment of writing this post some of the code snipets may not fully work yet as there are a few PR waiting to be merged.

  1. The code for the animation was obtained from here. ↩︎

Asier Galicia Martínez
Asier Galicia Martínez
Master student at TU Delft