Using a custom Hamiltonian function with ParaQeet

In this example we demonstrate how a custom Hamiltonian function H(t, params) can be used to simulate and optimize quantum systems.

import jax.numpy as jnp
import matplotlib.pyplot as plt

import paraqeet as pq

1. Define the parameters, Hamiltonian function and gradient functions

Define a Hamiltonian as a function of time and optimizable parameters. The optimizable parameter need to be of the type pq.Quantity.

Here we define a two level system (TLS) Hamiltonian, with a cosine drive (with optimizable parameters Amplitude and Frequency).

amplitude = pq.Quantity(
    value=jnp.array(1.55e8),
    min_value=jnp.array(0.0),
    max_value=jnp.array(5 * 1e8),
    unit="Hz",
    name="Amplitude",
    two_pi=True,
)

frequency = pq.Quantity(
    value=jnp.array(4.8e9 * 2 * jnp.pi),
    min_value=jnp.array(0.8 * 4.8e9 * 2 * jnp.pi),
    max_value=jnp.array(1.2 * 4.8e9 * 2 * jnp.pi),
    unit="Hz",
    name="lo_freq",
    two_pi=True,
)


def cos_envelope(t: pq.Array, amp, freq):
    """Define a cosine envelope."""
    return amp * jnp.cos(freq * t)


def tls_hamiltonian(t: pq.Array, amp, freq):
    """Define a Two level system Hamiltonian."""
    return 0.5 * freq * sigma_z + sigma_x * cos_envelope(t, amp, freq)


sigma_x = jnp.array([[0j, 1], [1, 0]])
sigma_z = jnp.diag(jnp.array([1.0, -1.0]))
freq = 4.8e9 * 2 * jnp.pi

from paraqeet.model.closed_system import ClosedSystem
from paraqeet.model.custom_hamiltonian import CustomHamiltonian

tls = CustomHamiltonian(
    hamiltonian_function=tls_hamiltonian,
    parameters=[amplitude, frequency],
)
model = ClosedSystem(tls)

2. Define propagation method and measurement function

Here we pick the standard ScipyExpmGOAT method for propagation and StateTransferFidelity as our measurement function

from paraqeet.measurement.state_transfer_fidelity import StateTransferFidelity
from paraqeet.propagation.scipy_expm_goat import ScipyExpmGOAT

t_final = 12e-9

prop = ScipyExpmGOAT(model, resolution=100e9)
times = jnp.array([0.0, t_final])

init = jnp.array([[1.0], [0]])  # |0>
target = jnp.array([[0.0], [1]])  # |1>
zeroone = StateTransferFidelity(
    propagation=prop,
    initial_state=init,
    target_state=target,
)

def make_plot():
    """Plot the signal."""
    ts = jnp.linspace(0, t_final, 1001)
    states = jnp.reshape(prop.propagate(ts), (-1, 2))
    sig = cos_envelope(ts, amp=amplitude.get_value()[0], freq=frequency.get_value()[0])

    fig, ax = plt.subplots(2, figsize=(4, 4), sharex=True)
    ax[0].plot(ts / 1e-9, sig)
    ax[0].set_ylabel("Field [MHz]")
    ax[0].grid(True, linestyle=(1, (1, 5)), linewidth=1)

    ax[1].plot(ts / 1e-9, jnp.abs(states) ** 2)
    ax[1].set_ylabel("Population")
    ax[1].grid(True, linestyle=(1, (1, 5)), linewidth=1)

    ax[-1].set_xlabel("Time [ns]")
    return fig, ax


make_plot()

(<Figure size 400x400 with 2 Axes>,
 array([<Axes: ylabel='Field [MHz]'>,
        <Axes: xlabel='Time [ns]', ylabel='Population'>], dtype=object))

3. Gradient based optimization

While using the above setup one can perform gradient free optimization. To do a gradient based optimization, we need to provide the gradient of the Hamiltonian wrt each parameter in the Hamiltonian function.

These gradient functions can be written as analytical functions or constructed using automatic differentiation using jax.grad.

Here we demonstrate both the cases.

1. Analytical functions for gradient of the Hamiltonian.

def grad_amp(t: pq.Array, amp, freq):
    """Gradient of Hamiltonian wrt amplitude."""
    return sigma_x * jnp.cos(freq * t)


def grad_frequency(t: pq.Array, amp, freq):
    """Gradient of Hamiltonian wrt frequency."""
    return -1 * sigma_x * amp * t * jnp.sin(freq * t)


analytical_grad_funcs = [grad_amp, grad_frequency]
tls.gradient_functions = analytical_grad_funcs

2. Gradient functions using Automatic differentiation

from jax import grad

# env_grads(t, amp, freq) returns a tuple of gradients wrt amp and freq
env_grads = grad(cos_envelope, argnums=(1, 2))


def grad_amp(t, amp, freq):
    """Gradient of Hamiltonian wrt amplitude."""
    return sigma_x * env_grads(t, amp, freq)[0]


def grad_freq(t, amp, freq):
    """Gradient of Hamiltonian wrt frequency."""
    return sigma_x * env_grads(t, amp, freq)[1]


tls.gradient_functions = [grad_amp, grad_freq]

zeroone.measure(times)

0.6404027521371757

4. Create optmap and optimize

from paraqeet.optimization_map import OptimizationMap
from paraqeet.optimizers.scipy_optimizer_gradient import ScipyOptimizerGradient

optmap = OptimizationMap()
optmap.add(tls)
opt = ScipyOptimizerGradient(zeroone, optimization_map=optmap)

opt.optimize(times)

{'status': 1, 'value': 1.4876988529977098e-14, 'iterations': 9, 'message': 'CONVERGENCE: NORM OF PROJECTED GRADIENT <= PGTOL'}

make_plot()

(<Figure size 400x400 with 2 Axes>,
 array([<Axes: ylabel='Field [MHz]'>,
        <Axes: xlabel='Time [ns]', ylabel='Population'>], dtype=object))

zeroone.measure(times)

0.9999999999999851