Skip to content

Latest commit

 

History

History
423 lines (307 loc) · 15.6 KB

analog-basics.md

File metadata and controls

423 lines (307 loc) · 15.6 KB

!!! warning The digital-analog emulation framework is under construction and more changes to the interface may still occur.

Qadence includes primitives for the construction of programs implemented on a set of interacting qubits. The goal is to build digital-analog programs that better represent the reality of interacting qubit platforms, such as neutral-atoms, while maintaining a simplified interface for users coming from a digital quantum computing background that may not be as familiar with pulse-level programming.

To build the intuition for the interface in Qadence, it is important to go over some of the underlying physics. We can write a general Hamiltonian for a set of $n$ interacting qubits as

$$ \mathcal{H} = \sum_{i=0}^{n-1}\left(\mathcal{H}^\text{d}{i}(t) + \sum{j<i}\mathcal{H}^\text{int}_{ij}\right), $$

where the driving Hamiltonian $\mathcal{H}^\text{d}{i}$ describes the pulses used to control single-qubit rotations, and the interaction Hamiltonian $\mathcal{H}^\text{int}{ij}$ describes the natural interaction between qubits.

Rydberg atoms

For the purpose of digital-analog emulation of neutral-atom systems in Qadence, we now consider a simplified time-independent global driving Hamiltonian, written as

$$ \mathcal{H}^\text{d}_{i} = \frac{\Omega}{2}\left(\cos(\phi) X_i - \sin(\phi) Y_i \right) - \delta N_i $$

where $\Omega$ is the Rabi frequency, $\delta$ is the detuning, $\phi$ is the phase, $X_i$ and $Y_i$ are the standard Pauli operators, and $N_i=\frac{1}{2}(I_i-Z_i)$ is the number operator. This Hamiltonian allows arbitrary global single-qubit rotations to be written, meaning that the values set for $(\Omega,\phi,\delta)$ are the same accross the qubit support.

For the interaction term, Rydberg atoms typically allow both an Ising and an XY mode of operation. For now, we focus on the Ising interaction, where the Hamiltonian is written as

$$ \mathcal{H}^\text{int}{ij} = \frac{C_6}{r{ij}^6}N_iN_j $$

where $r_{ij}$ is the distance between atoms $i$ and $j$, and $C_6$ is a coefficient depending on the specific Rydberg level of the excited state used in the computational logic states. A typical value for rydberg level of 60 is $C_6\approx 866~[\text{rad} . \mu \text{m}^6 / \text{ns}]$.

For a given register of atoms prepared in some spatial coordinates, the Hamiltonians described will generate the dynamics of some unitary operation as

$$ U(t, \Omega, \delta, \phi) = \exp(-i\mathcal{H}t) $$

where we specify the final parameter $t$, the duration of the operation.

Qadence uses the following units for user-specified parameters:

  • Rabi frequency and detuning $\Omega$, $\delta$: $[\text{rad}/\mu \text{s}]$
  • Phase $\phi$: $[\text{rad}]$
  • Duration $t$: $[\text{ns}]$
  • Atom coordinates: $[\mu \text{m}]$

In practice

Given the Hamiltonian description in the previous section, we will now go over a few examples of the standard operations available in Qadence.

Arbitrary rotation

To start, we will exemplify the a general rotation on a set of atoms. To create an arbitrary register of atoms, we refer the user to the register creation tutorial. Below, we create a line register of three qubits with a separation of $8~\mu\text{m}$. This is a typical value used in combination with a standard experimental setup of neutral atoms such that the interaction term in the Hamiltonian can effectively be used for computations.

from qadence import Register

reg = Register.line(3, spacing=8.0)  # Atom spacing in μm

Currently, the most general rotation operation uses the AnalogRot operation, which essentially implements $U(t, \Omega, \delta, \phi)$ defined above.

from qadence import AnalogRot, PI

rot_op = AnalogRot(
    duration = 500., # [ns]
    omega = PI, # [rad/μs]
    delta = PI, # [rad/μs]
    phase = PI, # [rad]
)

Note that in the code above a specific qubit support is not defined. By default this operation applies a global rotation on all qubits. We can define a circuit using the 3-qubit register and run it in the pyqtorch backend:

from qadence import BackendName, run

wf = run(reg, rot_op, backend = BackendName.PYQTORCH)

print(wf)
Under the hood of AnalogRot
To be fully explicit about what goes on under the hood of `AnalogRot`, we can look at the example
code below.

```python exec="on" source="material-block" result="json"
from qadence import BackendName, HamEvo, X, Y, N, add, run, PI
from qadence.analog.constants import C6_DICT
from math import cos, sin

# Following the 3-qubit register above
n_qubits = 3
dx = 8.0

# Parameters used in the AnalogRot
duration = 500.
omega = PI
delta = PI
phase = PI

# Building the terms in the driving Hamiltonian
h_x = (omega / 2) * cos(phase) * add(X(i) for i in range(n_qubits))
h_y = (-1.0 * omega / 2) * sin(phase) * add(Y(i) for i in range(n_qubits))
h_n = -1.0 * delta * add(N(i) for i in range(n_qubits))

# Building the interaction Hamiltonian

# Dictionary of coefficient values for each Rydberg level, which is 60 by default
c_6 = C6_DICT[60]

h_int = c_6 * (
    1/(dx**6) * (N(0)@N(1)) +
    1/(dx**6) * (N(1)@N(2)) +
    1/((2*dx)**6) * (N(0)@N(2))
)

hamiltonian = h_x + h_y + h_n + h_int

# Convert duration to µs due to the units of the Hamiltonian
explicit_rot = HamEvo(hamiltonian, duration / 1000)

wf = run(n_qubits, explicit_rot, backend = BackendName.PYQTORCH)

# We get the same final wavefunction
print(wf)
```

When sending the AnalogRot operation to the pyqtorch backend, Qadence automatically builds the correct Hamiltonian and the corresponding HamEvo operation with the added qubit interactions, as shown explicitly in the minimized section above. However, this operation is also supported in the Pulser backend, where the correct pulses are automatically created.

wf = run(
    reg,
    rot_op,
    backend = BackendName.PULSER,
)

print(wf)

RX / RY / RZ rotations

The AnalogRot provides full control over the parameters of $\mathcal{H}^\text{d}$, but users coming from a digital quantum computing background may be more familiar with the standard RX, RY and RZ rotations, also available in Qadence. For the emulated analog interface, Qadence provides alternative AnalogRX, AnalogRY and AnalogRZ operations which call AnalogRot under the hood to represent the rotations accross the respective axis.

For a given angle of rotation $\theta$ provided to each of these operations, currently a set of hardcoded assumptions are made on the tunable Hamiltonian parameters:

$$ \begin{aligned} \text{RX}:& \quad \Omega = \pi, \quad \delta = 0, \quad \phi = 0, \quad t = (\theta/\Omega)\times 10^3 \\ \text{RY}:& \quad \Omega = \pi, \quad \delta = 0, \quad \phi = -\pi/2, \quad t = (\theta/\Omega)\times 10^3 \\ \text{RZ}:& \quad \Omega = 0, \quad \delta = \pi, \quad \phi = 0, \quad t = (\theta/\delta)\times 10^3 \\ \end{aligned} $$

Note that the $\text{RZ}$ operation as defined above includes a global phase compared to the standard $\text{RZ}$ rotation since it evolves $\exp\left(-i\frac{\theta}{2}\frac{I-Z}{2}\right)$ instead of $\exp\left(-i\frac{\theta}{2}Z\right)$ given the detuning operator in $\mathcal{H}^\text{d}$.

!!! warning As shown above, the values of $\Omega$ and $\delta$ are currently hardcoded in these operators, and the effective angle of rotation is controlled by varying the duration of the evolution. Currently, the best way to overcome this is to use AnalogRot directly, but more general and convenient options will be provided soon in an improved interface.

Below we exemplify the usage of AnalogRX:

from qadence import Register, BackendName
from qadence import RX, AnalogRX, random_state, equivalent_state, kron, run, PI

n_qubits = 3
reg = Register.line(n_qubits, spacing=8.0)

# Rotation angle
theta = PI

# Analog rotation using the Rydberg Hamiltonian
rot_analog = AnalogRX(angle = theta)

# Equivalent full-digital global rotation
rot_digital = kron(RX(i, theta) for i in range(n_qubits))

# Some random initial state
init_state = random_state(n_qubits)

# Compare the final state using the full digital and the AnalogRX
wf_analog_pyq = run(
    reg,
    rot_analog,
    state = init_state,
    backend = BackendName.PYQTORCH
)


wf_digital_pyq = run(
    reg,
    rot_digital,
    state = init_state,
    backend = BackendName.PYQTORCH
)

bool_equiv = equivalent_state(wf_analog_pyq, wf_digital_pyq, atol = 1e-03)

print("States equivalent: ", bool_equiv)

As we can see, running a global RX or the AnalogRX does not result in equivalent states at the end, given that the digital RX operation does not include the interaction between the qubits. By setting dx very high in the code above the interaction will be less significant and the results will match.

However, if we compare with the Pulser backend, we see that the results for AnalogRX are consistent with the expected results from a real device:

wf_analog_pulser = run(
    reg,
    rot_analog,
    state = init_state,
    backend = BackendName.PULSER,
)

bool_equiv = equivalent_state(wf_analog_pyq, wf_analog_pulser, atol = 1e-03)

print("States equivalent: ", bool_equiv)

Evolving the interaction term

Finally, besides applying specific qubit rotations, we can also choose to evolve only the interaction term $\mathcal{H}^\text{int}$, equivalent to setting $\Omega = \delta = \phi = 0$. To do so, Qadence provides the function AnalogInteraction which does exactly this.

from qadence import Register, BackendName, random_state, equivalent_state, AnalogInteraction, run

n_qubits = 3
reg = Register.line(n_qubits, spacing=8.0)

duration = 1000.
op = AnalogInteraction(duration = duration)

init_state = random_state(n_qubits)

wf_pyq = run(reg, op, state = init_state, backend = BackendName.PYQTORCH)
wf_pulser = run(reg, op, state = init_state, backend = BackendName.PULSER)

bool_equiv = equivalent_state(wf_pyq, wf_pulser, atol = 1e-03)

print("States equivalent: ", bool_equiv)

Device specifications in Qadence

As a way to control other specifications of the interacting Rydberg atoms, Qadence provides a RydbergDevice class, which is currently used for both the pyqtorch and the pulser backends. Below we initialize a Rydberg device showcasing all the possible options.

from qadence import RydbergDevice, DeviceType, Interaction, PI

device_specs = RydbergDevice(
    interaction=Interaction.NN, # Or Interaction.XY, supported only for pyqtorch
    rydberg_level=60, # Integer value affecting the C_6 coefficient
    coeff_xy=3700.00, # C_3 coefficient for the XY interaction
    max_detuning=2 * PI * 4, # Max value for delta, currently only used in pulser
    max_amp=2 * PI * 3, # Max value for omega, currently only used in pulser
    pattern=None, # Semi-local addressing pattern, see the relevant tutorial
    type=DeviceType.IDEALIZED, # Pulser device to which the qadence device is converted in that backend
)

The values above are the defaults when simply running device_specs = RydbergDevice(). The convenience wrappers IdealDevice() or RealisticDevice() can also be used which simply change the type for the Pulser backend, but also allow an AddressingPattern passed in the pattern argument (see the relevant tutorial here).

!!! warning Currently, the options above are not fully integrated in both backends and this class should mostly be used if a user wishes to experiment with a different rydberg_level, or to change the device type for the pulser backend.

Planned features to add to the RydbergDevice include the definition of custom interaction functions,
the control of other drive Hamiltonian parameters so that $\Omega$, $\delta$ and $\phi$ are
not hardcoded when doing analog rotations, and the usage of the `max_detuning` and `max_amp` to control those
respective parameters when training models in the pyqtorch backend.

Finally, to change a given simulation, the device specifications are integrated in the Qadence Register. By default, all registers initialize an IdealDevice() under the hood. Below we run a quick test for a different rydberg level.

from qadence import Register, BackendName, random_state, equivalent_state, run
from qadence import AnalogRX, RydbergDevice, PI

device_specs = RydbergDevice(rydberg_level = 70)

n_qubits_side = 2
reg = Register.square(
    n_qubits_side,
    spacing = 8.0,
    device_specs = device_specs
)

rot_analog = AnalogRX(angle = PI)

init_state = random_state(n_qubits = 4)

wf_analog_pyq = run(
    reg,
    rot_analog,
    state = init_state,
    backend = BackendName.PYQTORCH
)

wf_analog_pulser = run(
    reg,
    rot_analog,
    state = init_state,
    backend = BackendName.PULSER
)

bool_equiv = equivalent_state(wf_analog_pyq, wf_analog_pulser, atol = 1e-03)

print("States equivalent: ", bool_equiv)

Technical details

!!! warning The details described here are relevant in the current version but will be lifted soon for the next version of the emulated analog interface.

In the previous section we have exemplified the main ingredients of the current user-facing functionalities of the emulated analog interface, and in the next tutorial on Quantum Circuit Learning we will exmplify its usage in a simple QML example. Here we specify some extra details of this interface.

In the block system, all analog rotation operators initialize a [ConstantAnalogRotation][qadence.blocks.analog.ConstantAnalogRotation] block, while the AnalogInteraction operation initializes an [InteractionBlock][qadence.blocks.analog.InteractionBlock]. As we have shown, by default, these blocks use a global qubit support, which can be passed explicitly by setting qubit_support = QubitSupportType.GLOBAL. However, composing blocks using kron with local qubit supports and different durations is not allowed.

from qadence import AnalogRX, AnalogRY, Register, kron

dx = 8.0
reg = Register.from_coordinates([(0, 0), (dx, 0)])

# Does not work (the angle affects the duration, as seen above):
rot_0 = AnalogRX(angle = 1.0, qubit_support = (0,))
rot_1 = AnalogRY(angle = 2.0, qubit_support = (1,))

try:
    block = kron(rot_0, rot_1)
except ValueError as error:
    print("Error:", error)

# Works:
rot_0 = AnalogRX(angle = 1.0, qubit_support = (0,))
rot_1 = AnalogRY(angle = 1.0, qubit_support = (1,))

block = kron(rot_0, rot_1)

Using chain is only supported between analog blocks with global qubit support:

from qadence import chain

rot_0 = AnalogRX(angle = 1.0, qubit_support = "global")
rot_1 = AnalogRY(angle = 2.0, qubit_support = "global")

block = chain(rot_0, rot_1)

The restrictions above only apply to the analog blocks, and analog and digital blocks can currently be composed.

from qadence import RX

rot_0 = AnalogRX(angle = 1.0, qubit_support = "global")
rot_1 = AnalogRY(angle = 2.0, qubit_support = (0,))
rot_digital = RX(1, 1.0)

block_0 = chain(rot_0, rot_digital)
block_1 = kron(rot_1, rot_digital)