Solving a 1D ODE

In this tutorial we will show how to use Qadence to solve a basic 1D Ordinary Differential Equation (ODE) with a QNN using Differentiable Quantum Circuits (DQC) ¹.

Consider the following non-linear ODE and boundary condition:

\[ \frac{df}{dx}= 5\times(4x^3+x^2-2x-\frac12), \qquad f(0)=0 \]

It admits an exact solution:

\[ f(x)=5\times(x^4+\frac13x^3-x^2-\frac12x) \]

Our goal will be to find this solution for \(x\in[-1, 1]\).

For the purpose of this tutorial, we will compute the derivative of the circuit using torch.autograd. The point of the DQC algorithm is to use differentiable circuits with parameter shift rules. In Qadence, PSR is implemented directly as custom overrides of the derivative function in the autograd engine, and thus we can later change the derivative method for the model itself if we wish.

Defining the loss function

The essential part of solving this problem is to define the right loss function to represent our goal. In this case, we want to define a model that has the capacity to learn the target solution, and we want to minimize: - The derivative of this model in comparison with the exact derivative in the equation; - The output of the model at the boundary in comparison with the value for the boundary condition;

We can write it like so:

Different loss criterions could be considered, and we could also play with the balance between the sum of the two loss terms. For now, let's proceed with the definition above.

Note that so far we have not used any quantum specific assumption, and we could in principle use the same loss function with a classical neural network.

Defining a QNN with Qadence

Now, we can finally use Qadence to write a QNN. We will use a feature map to encode the input values, a trainable ansatz circuit, and an observable to measure as the output.

We used a Chebyshev feature map with a tower-like scaling of the input reupload, and a standard hardware-efficient ansatz. You can check the qml constructors tutorial to see how you can customize these components. In the observable, for now we consider the simple case of measuring the magnetization of the first qubit.

from qadence.draw import display

# display(circuit)

Training the model

Now that the model is defined we can proceed with the training. the QNN class can be used like any other torch.nn.Module. Here we write a simple training loop, but you can also look at the ml tools tutorial to use the convenience training functions that Qadence provides.

To train the model, we will select a random set of collocation points uniformly distributed within \(-1.0< x <1.0\) and compute the loss function for those points.

Note the values of \(x\) are only picked from \(x\in[-0.99, 0.99]\) since we are using a Chebyshev feature map, and derivative of \(\text{acos}(x)\) diverges at \(-1\) and \(1\).

Plotting the results

import matplotlib.pyplot as plt

def f_exact(x: torch.Tensor) -> torch.Tensor:
    return 5*(x**4 + (1/3)*x**3 - x**2 - 0.5*x)

x_test = torch.arange(xmin, xmax, step = 0.01).unsqueeze(1)

result_exact = f_exact(x_test).flatten()

result_model = model(x_test).flatten().detach()

plt.plot(x_test, result_exact, label = "Exact solution")
plt.plot(x_test, result_model, label = " Trained model")

Clearly, the result is not optimal.

Improving the solution

One point to consider when defining the QNN is the possible output range, which is bounded by the spectrum of the chosen observable. For the magnetization of a single qubit, this means that the output is bounded between -1 and 1, which we can clearly see in the plot.

One option would be to define the observable as the total magnetization over all qubits, which would allow a range of -3 to 3.

And we again plot the result:

x_test = torch.arange(xmin, xmax, step = 0.01).unsqueeze(1)

result_exact = f_exact(x_test).flatten()

result_model = model(x_test).flatten().detach()

plt.plot(x_test, result_exact, label = "Exact solution")
plt.plot(x_test, result_model, label = "Trained model")

References

Kyriienko et al., Solving nonlinear differential equations with differentiable quantum circuits. ↩