# Release notes¶

This page contains the release notes for PennyLane.

- orphan

## Release 0.33.0 (current release)¶

### New features since last release

#### Postselection and statistics in mid-circuit measurements 📌

It is now possible to request postselection on a mid-circuit measurement. (#4604)

This can be achieved by specifying the

`postselect`

keyword argument in`qml.measure`

as either`0`

or`1`

, corresponding to the basis states.import pennylane as qml dev = qml.device("default.qubit") @qml.qnode(dev, interface=None) def circuit(): qml.Hadamard(wires=0) qml.CNOT(wires=[0, 1]) qml.measure(0, postselect=1) return qml.expval(qml.PauliZ(1)), qml.sample(wires=1)

This circuit prepares the \(| \Phi^{+} \rangle\) Bell state and postselects on measuring \(|1\rangle\) in wire

`0`

. The output of wire`1`

is then also \(|1\rangle\) at all times:>>> circuit(shots=10) (-1.0, array([1, 1, 1, 1, 1, 1]))

Note that the number of shots is less than the requested amount because we have thrown away the samples where \(|0\rangle\) was measured in wire

`0`

.Measurement statistics can now be collected for mid-circuit measurements. (#4544)

dev = qml.device("default.qubit") @qml.qnode(dev) def circ(x, y): qml.RX(x, wires=0) qml.RY(y, wires=1) m0 = qml.measure(1) return qml.expval(qml.PauliZ(0)), qml.expval(m0), qml.sample(m0)

>>> circ(1.0, 2.0, shots=10000) (0.5606, 0.7089, array([0, 1, 1, ..., 1, 1, 1]))

Support is provided for both finite-shot and analytic modes and devices default to using the deferred measurement principle to enact the mid-circuit measurements.

#### Exponentiate Hamiltonians with flexible Trotter products 🐖

Higher-order Trotter-Suzuki methods are now easily accessible through a new operation called

`TrotterProduct`

. (#4661)Trotterization techniques are an affective route towards accurate and efficient Hamiltonian simulation. The Suzuki-Trotter product formula allows for the ability to express higher-order approximations to the matrix exponential of a Hamiltonian, and it is now available to use in PennyLane via the

`TrotterProduct`

operation. Simply specify the`order`

of the approximation and the evolution`time`

.coeffs = [0.25, 0.75] ops = [qml.PauliX(0), qml.PauliZ(0)] H = qml.dot(coeffs, ops) dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(): qml.Hadamard(0) qml.TrotterProduct(H, time=2.4, order=2) return qml.state()

>>> circuit() [-0.13259524+0.59790098j 0. +0.j -0.13259524-0.77932754j 0. +0.j ]

Approximating matrix exponentiation with random product formulas, qDrift, is now available with the new

`QDrift`

operation. (#4671)As shown in 1811.08017, qDrift is a Markovian process that can provide a speedup in Hamiltonian simulation. At a high level, qDrift works by randomly sampling from the Hamiltonian terms with a probability that depends on the Hamiltonian coefficients. This method for Hamiltonian simulation is now ready to use in PennyLane with the

`QDrift`

operator. Simply specify the evolution`time`

and the number of samples drawn from the Hamiltonian,`n`

:coeffs = [0.25, 0.75] ops = [qml.PauliX(0), qml.PauliZ(0)] H = qml.dot(coeffs, ops) dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(): qml.Hadamard(0) qml.QDrift(H, time=1.2, n = 10) return qml.probs()

>>> circuit() array([0.61814334, 0. , 0.38185666, 0. ])

#### Building blocks for quantum phase estimation 🧱

A new operator called

`CosineWindow`

has been added to prepare an initial state based on a cosine wave function. (#4683)As outlined in 2110.09590, the cosine tapering window is part of a modification to quantum phase estimation that can provide a cubic improvement to the algorithm’s error rate. Using

`CosineWindow`

will prepare a state whose amplitudes follow a cosinusoidal distribution over the computational basis.import matplotlib.pyplot as plt dev = qml.device('default.qubit', wires=4) @qml.qnode(dev) def example_circuit(): qml.CosineWindow(wires=range(4)) return qml.state() output = example_circuit() plt.style.use("pennylane.drawer.plot") plt.bar(range(len(output)), output) plt.show()

Controlled gate sequences raised to decreasing powers, a sub-block in quantum phase estimation, can now be created with the new

`ControlledSequence`

operator. (#4707)To use

`ControlledSequence`

, specify the controlled unitary operator and the control wires,`control`

:dev = qml.device("default.qubit", wires = 4) @qml.qnode(dev) def circuit(): for i in range(3): qml.Hadamard(wires = i) qml.ControlledSequence(qml.RX(0.25, wires = 3), control = [0, 1, 2]) qml.adjoint(qml.QFT)(wires = range(3)) return qml.probs(wires = range(3))

>>> print(circuit()) [0.92059345 0.02637178 0.00729619 0.00423258 0.00360545 0.00423258 0.00729619 0.02637178]

#### New device capabilities, integration with Catalyst, and more! ⚗️

`default.qubit`

now uses the new`qml.devices.Device`

API and functionality in`qml.devices.qubit`

. If you experience any issues with the updated`default.qubit`

, please let us know by posting an issue. The old version of the device is still accessible by the short name`default.qubit.legacy`

, or directly via`qml.devices.DefaultQubitLegacy`

. (#4594) (#4436) (#4620) (#4632)This changeover has a number of benefits for

`default.qubit`

, including:The number of wires is now optional — simply having

`qml.device("default.qubit")`

is valid! If wires are not provided at instantiation, the device automatically infers the required number of wires for each circuit provided for execution.dev = qml.device("default.qubit") @qml.qnode(dev) def circuit(): qml.PauliZ(0) qml.RZ(0.1, wires=1) qml.Hadamard(2) return qml.state()

>>> print(qml.draw(circuit)()) 0: ──Z────────┤ State 1: ──RZ(0.10)─┤ State 2: ──H────────┤ State

`default.qubit`

is no longer silently swapped out with an interface-appropriate device when the backpropagation differentiation method is used. For example, consider:import jax dev = qml.device("default.qubit", wires=1) @qml.qnode(dev, diff_method="backprop") def f(x): qml.RX(x, wires=0) return qml.expval(qml.PauliZ(0)) f(jax.numpy.array(0.2))

In previous versions of PennyLane, the device will be swapped for the JAX equivalent:

>>> f.device <DefaultQubitJax device (wires=1, shots=None) at 0x7f8c8bff50a0> >>> f.device == dev False

Now,

`default.qubit`

can itself dispatch to all the interfaces in a backprop-compatible way and hence does not need to be swapped:>>> f.device <default.qubit device (wires=1) at 0x7f20d043b040> >>> f.device == dev True

A QNode that has been decorated with

`qjit`

from PennyLane’s Catalyst library for just-in-time hybrid compilation is now compatible with`qml.draw`

. (#4609)import catalyst @catalyst.qjit @qml.qnode(qml.device("lightning.qubit", wires=3)) def circuit(x, y, z, c): """A quantum circuit on three wires.""" @catalyst.for_loop(0, c, 1) def loop(i): qml.Hadamard(wires=i) qml.RX(x, wires=0) loop() qml.RY(y, wires=1) qml.RZ(z, wires=2) return qml.expval(qml.PauliZ(0)) draw = qml.draw(circuit, decimals=None)(1.234, 2.345, 3.456, 1)

>>> print(draw) 0: ──RX──H──┤ <Z> 1: ──H───RY─┤ 2: ──RZ─────┤

### Improvements 🛠

#### More PyTrees!

`MeasurementProcess`

and`QuantumScript`

objects are now registered as JAX PyTrees. (#4607) (#4608)It is now possible to JIT-compile functions with arguments that are a

`MeasurementProcess`

or a`QuantumScript`

:import jax tape0 = qml.tape.QuantumTape([qml.RX(1.0, 0), qml.RY(0.5, 0)], [qml.expval(qml.PauliZ(0))]) dev = qml.device('lightning.qubit', wires=5) execute_kwargs = {"device": dev, "gradient_fn": qml.gradients.param_shift, "interface":"jax"} jitted_execute = jax.jit(qml.execute, static_argnames=execute_kwargs.keys()) jitted_execute((tape0, ), **execute_kwargs)

#### Improving QChem and existing algorithms

Computationally expensive functions in

`integrals.py`

,`electron_repulsion`

and`_hermite_coulomb`

, have been modified to replace indexing with slicing for better compatibility with JAX. (#4685)`qml.qchem.import_state`

has been extended to import more quantum chemistry wavefunctions, from MPS, DMRG and SHCI classical calculations performed with the Block2 and Dice libraries. #4523 #4524 #4626 #4634Check out our how-to guide to learn more about how PennyLane integrates with your favourite quantum chemistry libraries.

The qchem

`fermionic_dipole`

and`particle_number`

functions have been updated to use a`FermiSentence`

. The deprecated features for using tuples to represent fermionic operations are removed. (#4546) (#4556)The tensor-network template

`qml.MPS`

now supports changing the`offset`

between subsequent blocks for more flexibility. (#4531)Builtin types support with

`qml.pauli_decompose`

have been improved. (#4577)`AmplitudeEmbedding`

now inherits from`StatePrep`

, allowing for it to not be decomposed when at the beginning of a circuit, thus behaving like`StatePrep`

. (#4583)`qml.cut_circuit`

is now compatible with circuits that compute the expectation values of Hamiltonians with two or more terms. (#4642)

#### Next-generation device API

`default.qubit`

now tracks the number of equivalent qpu executions and total shots when the device is sampling. Note that`"simulations"`

denotes the number of simulation passes, whereas`"executions"`

denotes how many different computational bases need to be sampled in. Additionally, the new`default.qubit`

tracks the results of`device.execute`

. (#4628) (#4649)`DefaultQubit`

can now accept a`jax.random.PRNGKey`

as a`seed`

to set the key for the JAX pseudo random number generator when using the JAX interface. This corresponds to the`prng_key`

on`default.qubit.jax`

in the old API. (#4596)The

`JacobianProductCalculator`

abstract base class and implementations`TransformJacobianProducts`

`DeviceDerivatives`

, and`DeviceJacobianProducts`

have been added to`pennylane.interfaces.jacobian_products`

. (#4435) (#4527) (#4637)`DefaultQubit`

dispatches to a faster implementation for applying`ParametrizedEvolution`

to a state when it is more efficient to evolve the state than the operation matrix. (#4598) (#4620)Wires can be provided to the new device API. (#4538) (#4562)

`qml.sample()`

in the new device API now returns a`np.int64`

array instead of`np.bool8`

. (#4539)The new device API now has a

`repr()`

method. (#4562)`DefaultQubit`

now works as expected with measurement processes that don’t specify wires. (#4580)Various improvements to measurements have been made for feature parity between

`default.qubit.legacy`

and the new`DefaultQubit`

. This includes not trying to squeeze batched`CountsMP`

results and implementing`MutualInfoMP.map_wires`

. (#4574)`devices.qubit.simulate`

now accepts an interface keyword argument. If a QNode with`DefaultQubit`

specifies an interface, the result will be computed with that interface. (#4582)`ShotAdaptiveOptimizer`

has been updated to pass shots to QNode executions instead of overriding device shots before execution. This makes it compatible with the new device API. (#4599)`pennylane.devices.preprocess`

now offers the transforms`decompose`

,`validate_observables`

,`validate_measurements`

,`validate_device_wires`

,`validate_multiprocessing_workers`

,`warn_about_trainable_observables`

, and`no_sampling`

to assist in constructing devices under the new device API. (#4659)Updated

`qml.device`

,`devices.preprocessing`

and the`tape_expand.set_decomposition`

context manager to bring`DefaultQubit`

to feature parity with`default.qubit.legacy`

with regards to using custom decompositions. The`DefaultQubit`

device can now be included in a`set_decomposition`

context or initialized with a`custom_decomps`

dictionary, as well as a custom`max_depth`

for decomposition. (#4675)

#### Other improvements

The

`StateMP`

measurement now accepts a wire order (e.g., a device wire order). The`process_state`

method will re-order the given state to go from the inputted wire-order to the process’s wire-order. If the process’s wire-order contains extra wires, it will assume those are in the zero-state. (#4570) (#4602)Methods called

`add_transform`

and`insert_front_transform`

have been added to`TransformProgram`

. (#4559)Instances of the

`TransformProgram`

class can now be added together. (#4549)Transforms can now be applied to devices following the new device API. (#4667)

All gradient transforms have been updated to the new transform program system. (#4595)

Multi-controlled operations with a single-qubit special unitary target can now automatically decompose. (#4697)

`pennylane.defer_measurements`

will now exit early if the input does not contain mid circuit measurements. (#4659)The density matrix aspects of

`StateMP`

have been split into their own measurement process called`DensityMatrixMP`

. (#4558)`StateMeasurement.process_state`

now assumes that the input is flat.`ProbabilityMP.process_state`

has been updated to reflect this assumption and avoid redundant reshaping. (#4602)`qml.exp`

returns a more informative error message when decomposition is unavailable for non-unitary operators. (#4571)Added

`qml.math.get_deep_interface`

to get the interface of a scalar hidden deep in lists or tuples. (#4603)Updated

`qml.math.ndim`

and`qml.math.shape`

to work with built-in lists or tuples that contain interface-specific scalar dat (e.g.,`[(tf.Variable(1.1), tf.Variable(2.2))]`

). (#4603)When decomposing a unitary matrix with

`one_qubit_decomposition`

and opting to include the`GlobalPhase`

in the decomposition, the phase is no longer cast to`dtype=complex`

. (#4653)`_qfunc_output`

has been removed from`QuantumScript`

, as it is no longer necessary. There is still a`_qfunc_output`

property on`QNode`

instances. (#4651)`qml.data.load`

properly handles parameters that come after`'full'`

(#4663)The

`qml.jordan_wigner`

function has been modified to optionally remove the imaginary components of the computed qubit operator, if imaginary components are smaller than a threshold. (#4639)`qml.data.load`

correctly performs a full download of the dataset after a partial download of the same dataset has already been performed. (#4681)The performance of

`qml.data.load()`

has been improved when partially loading a dataset (#4674)Plots generated with the

`pennylane.drawer.plot`

style of`matplotlib.pyplot`

now have black axis labels and are generated at a default DPI of 300. (#4690)Shallow copies of the

`QNode`

now also copy the`execute_kwargs`

and transform program. When applying a transform to a`QNode`

, the new qnode is only a shallow copy of the original and thus keeps the same device. (#4736)`QubitDevice`

and`CountsMP`

are updated to disregard samples containing failed hardware measurements (record as`np.NaN`

) when tallying samples, rather than counting failed measurements as ground-state measurements, and to display`qml.counts`

coming from these hardware devices correctly. (#4739)

### Breaking changes 💔

`qml.defer_measurements`

now raises an error if a transformed circuit measures`qml.probs`

,`qml.sample`

, or`qml.counts`

without any wires or observable, or if it measures`qml.state`

. (#4701)The device test suite now converts device keyword arguments to integers or floats if possible. (#4640)

`MeasurementProcess.eigvals()`

now raises an`EigvalsUndefinedError`

if the measurement observable does not have eigenvalues. (#4544)The

`__eq__`

and`__hash__`

methods of`Operator`

and`MeasurementProcess`

no longer rely on the object’s address in memory. Using`==`

with operators and measurement processes will now behave the same as`qml.equal`

, and objects of the same type with the same data and hyperparameters will have the same hash. (#4536)In the following scenario, the second and third code blocks show the previous and current behaviour of operator and measurement process equality, determined by

`==`

:op1 = qml.PauliX(0) op2 = qml.PauliX(0) op3 = op1

Old behaviour:

>>> op1 == op2 False >>> op1 == op3 True

New behaviour:

>>> op1 == op2 True >>> op1 == op3 True

The

`__hash__`

dunder method defines the hash of an object. The default hash of an object is determined by the objects memory address. However, the new hash is determined by the properties and attributes of operators and measurement processes. Consider the scenario below. The second and third code blocks show the previous and current behaviour.op1 = qml.PauliX(0) op2 = qml.PauliX(0)

Old behaviour:

>>> print({op1, op2}) {PauliX(wires=[0]), PauliX(wires=[0])}

New behaviour:

>>> print({op1, op2}) {PauliX(wires=[0])}

The old return type and associated functions

`qml.enable_return`

and`qml.disable_return`

have been removed. (#4503)The

`mode`

keyword argument in`QNode`

has been removed. Please use`grad_on_execution`

instead. (#4503)The CV observables

`qml.X`

and`qml.P`

have been removed. Please use`qml.QuadX`

and`qml.QuadP`

instead. (#4533)The

`sampler_seed`

argument of`qml.gradients.spsa_grad`

has been removed. Instead, the`sampler_rng`

argument should be set, either to an integer value, which will be used to create a PRNG internally, or to a NumPy pseudo-random number generator (PRNG) created via`np.random.default_rng(seed)`

. (#4550)The

`QuantumScript.set_parameters`

method and the`QuantumScript.data`

setter have been removed. Please use`QuantumScript.bind_new_parameters`

instead. (#4548)The method

`tape.unwrap()`

and corresponding`UnwrapTape`

and`Unwrap`

classes have been removed. Instead of`tape.unwrap()`

, use`qml.transforms.convert_to_numpy_parameters`

. (#4535)The

`RandomLayers.compute_decomposition`

keyword argument`ratio_imprivitive`

has been changed to`ratio_imprim`

to match the call signature of the operation. (#4552)The private

`TmpPauliRot`

operator used for`SpecialUnitary`

no longer decomposes to nothing when the theta value is trainable. (#4585)`ProbabilityMP.marginal_prob`

has been removed. Its contents have been moved into`process_state`

, which effectively just called`marginal_prob`

with`np.abs(state) ** 2`

. (#4602)

### Deprecations 👋

The following decorator syntax for transforms has been deprecated and will raise a warning: (#4457)

@transform_fn(**transform_kwargs) @qml.qnode(dev) def circuit(): ...

If you are using a transform that has supporting

`transform_kwargs`

, please call the transform directly using`circuit = transform_fn(circuit, **transform_kwargs)`

, or use`functools.partial`

:@functools.partial(transform_fn, **transform_kwargs) @qml.qnode(dev) def circuit(): ...

The

`prep`

keyword argument in`QuantumScript`

has been deprecated and will be removed from`QuantumScript`

.`StatePrepBase`

operations should be placed at the beginning of the`ops`

list instead. (#4554)`qml.gradients.pulse_generator`

has been renamed to`qml.gradients.pulse_odegen`

to adhere to paper naming conventions. During v0.33,`pulse_generator`

is still available but raises a warning. (#4633)

### Documentation 📝

A warning section in the docstring for

`DefaultQubit`

regarding the start method used in multiprocessing has been added. This may help users circumvent issues arising in Jupyter notebooks on macOS for example. (#4622)Documentation improvements to the new device API have been made. The documentation now correctly states that interface-specific parameters are only passed to the device for backpropagation derivatives. (#4542)

Functions for qubit-simulation to the

`qml.devices`

sub-page of the “Internal” section have been added. Note that these functions are unstable while device upgrades are underway. (#4555)A documentation improvement to the usage example in the

`qml.QuantumMonteCarlo`

page has been made. An integral was missing the differential \(dx\). (#4593)A documentation improvement for the use of the

`pennylane`

style of`qml.drawer`

and the`pennylane.drawer.plot`

style of`matplotlib.pyplot`

has been made by clarifying the use of the default font. (#4690)

### Bug fixes 🐛

Jax jit now works when a probability measurement is broadcasted onto all wires. (#4742)

Fixed

`LocalHilbertSchmidt.compute_decomposition`

so that the template can be used in a QNode. (#4719)Fixes

`transforms.transpile`

with arbitrary measurement processes. (#4732)Providing

`work_wires=None`

to`qml.GroverOperator`

no longer interprets`None`

as a wire. (#4668)Fixed an issue where the

`__copy__`

method of the`qml.Select()`

operator attempted to access un-initialized data. (#4551)Fixed the

`skip_first`

option in`expand_tape_state_prep`

. (#4564)`convert_to_numpy_parameters`

now uses`qml.ops.functions.bind_new_parameters`

. This reinitializes the operation and makes sure everything references the new NumPy parameters. (#4540)`tf.function`

no longer breaks`ProbabilityMP.process_state`

, which is needed by new devices. (#4470)Fixed unit tests for

`qml.qchem.mol_data`

. (#4591)Fixed

`ProbabilityMP.process_state`

so that it allows for proper Autograph compilation. Without this, decorating a QNode that returns an`expval`

with`tf.function`

would fail when computing the expectation. (#4590)The

`torch.nn.Module`

properties are now accessible on a`pennylane.qnn.TorchLayer`

. (#4611)`qml.math.take`

with Pytorch now returns`tensor[..., indices]`

when the user requests the last axis (`axis=-1`

). Without the fix, it would wrongly return`tensor[indices]`

. (#4605)Ensured the logging

`TRACE`

level works with gradient-free execution. (#4669)

### Contributors ✍️

This release contains contributions from (in alphabetical order):

Guillermo Alonso, Utkarsh Azad, Thomas Bromley, Isaac De Vlugt, Jack Brown, Stepan Fomichev, Joana Fraxanet, Diego Guala, Soran Jahangiri, Edward Jiang, Korbinian Kottmann, Ivana Kurečić Christina Lee, Lillian M. A. Frederiksen, Vincent Michaud-Rioux, Romain Moyard, Daniel F. Nino, Lee James O’Riordan, Mudit Pandey, Matthew Silverman, Jay Soni.

- orphan

## Release 0.32.0¶

### New features since last release

#### Encode matrices using a linear combination of unitaries ⛓️️

It is now possible to encode an operator

`A`

into a quantum circuit by decomposing it into a linear combination of unitaries using PREP (qml.StatePrep) and SELECT (qml.Select) routines. (#4431) (#4437) (#4444) (#4450) (#4506) (#4526)Consider an operator

`A`

composed of a linear combination of Pauli terms:>>> A = qml.PauliX(2) + 2 * qml.PauliY(2) + 3 * qml.PauliZ(2)

A decomposable block-encoding circuit can be created:

def block_encode(A, control_wires): probs = A.coeffs / np.sum(A.coeffs) state = np.pad(np.sqrt(probs, dtype=complex), (0, 1)) unitaries = A.ops qml.StatePrep(state, wires=control_wires) qml.Select(unitaries, control=control_wires) qml.adjoint(qml.StatePrep)(state, wires=control_wires)

>>> print(qml.draw(block_encode, show_matrices=False)(A, control_wires=[0, 1])) 0: ─╭|Ψ⟩─╭Select─╭|Ψ⟩†─┤ 1: ─╰|Ψ⟩─├Select─╰|Ψ⟩†─┤ 2: ──────╰Select───────┤

This circuit can be used as a building block within a larger QNode to perform algorithms such as QSVT and Hamiltonian simulation.

Decomposing a Hermitian matrix into a linear combination of Pauli words via

`qml.pauli_decompose`

is now faster and differentiable. (#4395) (#4479) (#4493)def find_coeffs(p): mat = np.array([[3, p], [p, 3]]) A = qml.pauli_decompose(mat) return A.coeffs

>>> import jax >>> from jax import numpy as np >>> jax.jacobian(find_coeffs)(np.array(2.)) Array([0., 1.], dtype=float32, weak_type=True)

#### Monitor PennyLane's inner workings with logging 📃

Python-native logging can now be enabled with

`qml.logging.enable_logging()`

. (#4377) (#4383)Consider the following code that is contained in

`my_code.py`

:import pennylane as qml qml.logging.enable_logging() # enables logging dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def f(x): qml.RX(x, wires=0) return qml.state() f(0.5)

Executing

`my_code.py`

with logging enabled will detail every step in PennyLane’s pipeline that gets used to run your code.$ python my_code.py [1967-02-13 15:18:38,591][DEBUG][<PID 8881:MainProcess>] - pennylane.qnode.__init__()::"Creating QNode(func=<function f at 0x7faf2a6fbaf0>, device=<DefaultQubit device (wires=2, shots=None) at 0x7faf2a689b50>, interface=auto, diff_method=best, expansion_strategy=gradient, max_expansion=10, grad_on_execution=best, mode=None, cache=True, cachesize=10000, max_diff=1, gradient_kwargs={}" ...

Additional logging configuration settings can be specified by modifying the contents of the logging configuration file, which can be located by running

`qml.logging.config_path()`

. Follow our logging docs page for more details!

#### More input states for quantum chemistry calculations ⚛️

Input states obtained from advanced quantum chemistry calculations can be used in a circuit. (#4427) (#4433) (#4461) (#4476) (#4505)

Quantum chemistry calculations rely on an initial state that is typically selected to be the trivial Hartree-Fock state. For molecules with a complicated electronic structure, using initial states obtained from affordable post-Hartree-Fock calculations helps to improve the efficiency of the quantum simulations. These calculations can be done with external quantum chemistry libraries such as PySCF.

It is now possible to import a PySCF solver object in PennyLane and extract the corresponding wave function in the form of a state vector that can be directly used in a circuit. First, perform your classical quantum chemistry calculations and then use the qml.qchem.import_state function to import the solver object and return a state vector.

```
>>> from pyscf import gto, scf, ci
>>> mol = gto.M(atom=[['H', (0, 0, 0)], ['H', (0,0,0.71)]], basis='sto6g')
>>> myhf = scf.UHF(mol).run()
>>> myci = ci.UCISD(myhf).run()
>>> wf_cisd = qml.qchem.import_state(myci, tol=1e-1)
>>> print(wf_cisd)
[ 0. +0.j 0. +0.j 0. +0.j 0.1066467 +0.j
1. +0.j 0. +0.j 0. +0.j 0. +0.j
2. +0.j 0. +0.j 0. +0.j 0. +0.j
-0.99429698+0.j 0. +0.j 0. +0.j 0. +0.j]
The state vector can be implemented in a circuit using ``qml.StatePrep``.
```

```
>>> dev = qml.device('default.qubit', wires=4)
>>> @qml.qnode(dev)
... def circuit():
... qml.StatePrep(wf_cisd, wires=range(4))
... return qml.state()
>>> print(circuit())
[ 0. +0.j 0. +0.j 0. +0.j 0.1066467 +0.j
1. +0.j 0. +0.j 0. +0.j 0. +0.j
2. +0.j 0. +0.j 0. +0.j 0. +0.j
-0.99429698+0.j 0. +0.j 0. +0.j 0. +0.j]
The currently supported post-Hartree-Fock methods are RCISD, UCISD, RCCSD, and UCCSD which
denote restricted (R) and unrestricted (U) configuration interaction (CI) and coupled cluster (CC)
calculations with single and double (SD) excitations.
```

#### Reuse and reset qubits after mid-circuit measurements ♻️

PennyLane now allows you to define circuits that reuse a qubit after a mid-circuit measurement has taken place. Optionally, the wire can also be reset to the \(|0\rangle\) state. (#4402) (#4432)

Post-measurement reset can be activated by setting

`reset=True`

when calling qml.measure. In this version of PennyLane, executing circuits with qubit reuse will result in the defer_measurements transform being applied. This transform replaces each reused wire with an*additional*qubit. However, future releases of PennyLane will explore device-level support for qubit reuse without consuming additional qubits.Qubit reuse and reset is also fully differentiable:

dev = qml.device("default.qubit", wires=4) @qml.qnode(dev) def circuit(p): qml.RX(p, wires=0) m = qml.measure(0, reset=True) qml.cond(m, qml.Hadamard)(1) qml.RX(p, wires=0) m = qml.measure(0) qml.cond(m, qml.Hadamard)(1) return qml.expval(qml.PauliZ(1))

>>> jax.grad(circuit)(0.4) Array(-0.35867804, dtype=float32, weak_type=True)

You can read more about mid-circuit measurements in the documentation, and stay tuned for more mid-circuit measurement features in the next few releases!

### Improvements 🛠

#### A new PennyLane drawing style

Circuit drawings and plots can now be created following a PennyLane style. (#3950)

The

`qml.draw_mpl`

function accepts a`style='pennylane'`

argument to create PennyLane themed circuit diagrams:def circuit(x, z): qml.QFT(wires=(0,1,2,3)) qml.Toffoli(wires=(0,1,2)) qml.CSWAP(wires=(0,2,3)) qml.RX(x, wires=0) qml.CRZ(z, wires=(3,0)) return qml.expval(qml.PauliZ(0)) qml.draw_mpl(circuit, style="pennylane")(1, 1)

PennyLane-styled plots can also be drawn by passing

`"pennylane.drawer.plot"`

to Matplotlib’s`plt.style.use`

function:import matplotlib.pyplot as plt plt.style.use("pennylane.drawer.plot") for i in range(3): plt.plot(np.random.rand(10))

If the font Quicksand Bold isn’t available, an available default font is used instead.

#### Making operators immutable and PyTrees

Any class inheriting from

`Operator`

is now automatically registered as a pytree with JAX. This unlocks the ability to jit functions of`Operator`

. (#4458)>>> op = qml.adjoint(qml.RX(1.0, wires=0)) >>> jax.jit(qml.matrix)(op) Array([[0.87758255-0.j , 0. +0.47942555j], [0. +0.47942555j, 0.87758255-0.j ]], dtype=complex64, weak_type=True) >>> jax.tree_util.tree_map(lambda x: x+1, op) Adjoint(RX(2.0, wires=[0]))

All

`Operator`

objects now define`Operator._flatten`

and`Operator._unflatten`

methods that separate trainable from untrainable components. These methods will be used in serialization and pytree registration. Custom operations may need an update to ensure compatibility with new PennyLane features. (#4483) (#4314)The

`QuantumScript`

class now has a`bind_new_parameters`

method that allows creation of new`QuantumScript`

objects with the provided parameters. (#4345)The

`qml.gradients`

module no longer mutates operators in-place for any gradient transforms. Instead, operators that need to be mutated are copied with new parameters. (#4220)PennyLane no longer directly relies on

`Operator.__eq__`

. (#4398)`qml.equal`

no longer raises errors when operators or measurements of different types are compared. Instead, it returns`False`

. (#4315)

#### Transforms

Transform programs are now integrated with the QNode. (#4404)

def null_postprocessing(results: qml.typing.ResultBatch) -> qml.typing.Result: return results[0] @qml.transforms.core.transform def scale_shots(tape: qml.tape.QuantumTape, shot_scaling) -> (Tuple[qml.tape.QuantumTape], Callable): new_shots = tape.shots.total_shots * shot_scaling new_tape = qml.tape.QuantumScript(tape.operations, tape.measurements, shots=new_shots) return (new_tape, ), null_postprocessing dev = qml.devices.experimental.DefaultQubit2() @partial(scale_shots, shot_scaling=2) @qml.qnode(dev, interface=None) def circuit(): return qml.sample(wires=0)

>>> circuit(shots=1) array([False, False])

Transform Programs,

`qml.transforms.core.TransformProgram`

, can now be called on a batch of circuits and return a new batch of circuits and a single post processing function. (#4364)`TransformDispatcher`

now allows registration of custom QNode transforms. (#4466)QNode transforms in

`qml.qinfo`

now support custom wire labels. #4331`qml.transforms.adjoint_metric_tensor`

now uses the simulation tools in`qml.devices.qubit`

instead of private methods of`qml.devices.DefaultQubit`

. (#4456)Auxiliary wires and device wires are now treated the same way in

`qml.transforms.metric_tensor`

as in`qml.gradients.hadamard_grad`

. All valid wire input formats for`aux_wire`

are supported. (#4328)

#### Next-generation device API

The experimental device interface has been integrated with the QNode for JAX, JAX-JIT, TensorFlow and PyTorch. (#4323) (#4352) (#4392) (#4393)

The experimental

`DefaultQubit2`

device now supports computing VJPs and JVPs using the adjoint method. (#4374)New functions called

`adjoint_jvp`

and`adjoint_vjp`

that compute the JVP and VJP of a tape using the adjoint method have been added to`qml.devices.qubit.adjoint_jacobian`

(#4358)`DefaultQubit2`

now accepts a`max_workers`

argument which controls multiprocessing. A`ProcessPoolExecutor`

executes tapes asynchronously using a pool of at most`max_workers`

processes. If`max_workers`

is`None`

or not given, only the current process executes tapes. If you experience any issue, say using JAX, TensorFlow, Torch, try setting`max_workers`

to`None`

. (#4319) (#4425)`qml.devices.experimental.Device`

now accepts a shots keyword argument and has a`shots`

property. This property is only used to set defaults for a workflow, and does not directly influence the number of shots used in executions or derivatives. (#4388)`expand_fn()`

for`DefaultQubit2`

has been updated to decompose`StatePrep`

operations present in the middle of a circuit. (#4444)If no seed is specified on initialization with

`DefaultQubit2`

, the local random number generator will be seeded from NumPy’s global random number generator. (#4394)

#### Improvements to machine learning library interfaces

`pennylane/interfaces`

has been refactored. The`execute_fn`

passed to the machine learning framework boundaries is now responsible for converting parameters to NumPy. The gradients module can now handle TensorFlow parameters, but gradient tapes now retain the original`dtype`

instead of converting to`float64`

. This may cause instability with finite-difference differentiation and`float32`

parameters. The machine learning boundary functions are now uncoupled from their legacy counterparts. (#4415)`qml.interfaces.set_shots`

now accepts a`Shots`

object as well as`int`

‘s and tuples of`int`

‘s. (#4388)Readability improvements and stylistic changes have been made to

`pennylane/interfaces/jax_jit_tuple.py`

(#4379)

#### Pulses

A

`HardwareHamiltonian`

can now be summed with`int`

or`float`

objects. A sequence of`HardwareHamiltonian`

s can now be summed via the builtin`sum`

. (#4343)`qml.pulse.transmon_drive`

has been updated in accordance with 1904.06560. In particular, the functional form has been changed from \(\Omega(t)(\cos(\omega_d t + \phi) X - \sin(\omega_d t + \phi) Y)$ to $\Omega(t) \sin(\omega_d t + \phi) Y\). (#4418) (#4465) (#4478) (#4418)

#### Other improvements

The

`qchem`

module has been upgraded to use the fermionic operators of the`fermi`

module. #4336 #4521The calculation of

`Sum`

,`Prod`

,`SProd`

,`PauliWord`

, and`PauliSentence`

sparse matrices are orders of magnitude faster. (#4475) (#4272) (#4411)A function called

`qml.math.fidelity_statevector`

that computes the fidelity between two state vectors has been added. (#4322)`qml.ctrl(qml.PauliX)`

returns a`CNOT`

,`Toffoli`

, or`MultiControlledX`

operation instead of`Controlled(PauliX)`

. (#4339)When given a callable,

`qml.ctrl`

now does its custom pre-processing on all queued operators from the callable. (#4370)The

`qchem`

functions`primitive_norm`

and`contracted_norm`

have been modified to be compatible with higher versions of SciPy. The private function`_fac2`

for computing double factorials has also been added. #4321`tape_expand`

now uses`Operator.decomposition`

instead of`Operator.expand`

in order to make more performant choices. (#4355)CI now runs tests with TensorFlow 2.13.0 (#4472)

All tests in CI and pre-commit hooks now enable linting. (#4335)

The default label for a

`StatePrepBase`

operator is now`|Ψ⟩`

. (#4340)`Device.default_expand_fn()`

has been updated to decompose`qml.StatePrep`

operations present in the middle of a provided circuit. (#4437)`QNode.construct`

has been updated to only apply the`qml.defer_measurements`

transform if the device does not natively support mid-circuit measurements. (#4516)The application of the

`qml.defer_measurements`

transform has been moved from`QNode.construct`

to`qml.Device.batch_transform`

to allow more fine-grain control over when`defer_measurements`

should be used. (#4432)The label for

`ParametrizedEvolution`

can display parameters with the requested format as set by the kwarg`decimals`

. Array-like parameters are displayed in the same format as matrices and stored in the cache. (#4151)

### Breaking changes 💔

Applying gradient transforms to broadcasted/batched tapes has been deactivated until it is consistently supported for QNodes as well. (#4480)

Gradient transforms no longer implicitly cast

`float32`

parameters to`float64`

. Finite difference differentiation with`float32`

parameters may no longer give accurate results. (#4415)The

`do_queue`

keyword argument in`qml.operation.Operator`

has been removed. Instead of setting`do_queue=False`

, use the`qml.QueuingManager.stop_recording()`

context. (#4317)`Operator.expand`

now uses the output of`Operator.decomposition`

instead of what it queues. (#4355)The gradients module no longer needs shot information passed to it explicitly, as the shots are on the tapes. (#4448)

`qml.StatePrep`

has been renamed to`qml.StatePrepBase`

and`qml.QubitStateVector`

has been renamed to`qml.StatePrep`

.`qml.operation.StatePrep`

and`qml.QubitStateVector`

are still accessible. (#4450)Support for Python 3.8 has been dropped. (#4453)

`MeasurementValue`

‘s signature has been updated to accept a list of`MidMeasureMP`

‘s rather than a list of their IDs. (#4446)The

`grouping_type`

and`grouping_method`

keyword arguments have been removed from`qchem.molecular_hamiltonian`

. (#4301)`zyz_decomposition`

and`xyx_decomposition`

have been removed. Use`one_qubit_decomposition`

instead. (#4301)`LieAlgebraOptimizer`

has been removed. Use`RiemannianGradientOptimizer`

instead. (#4301)`Operation.base_name`

has been removed. (#4301)`QuantumScript.name`

has been removed. (#4301)`qml.math.reduced_dm`

has been removed. Use`qml.math.reduce_dm`

or`qml.math.reduce_statevector`

instead. (#4301)The

`qml.specs`

dictionary no longer supports direct key access to certain keys. (#4301)Instead, these quantities can be accessed as fields of the new

`Resources`

object saved under`specs_dict["resources"]`

:`num_operations`

is no longer supported, use`specs_dict["resources"].num_gates`

`num_used_wires`

is no longer supported, use`specs_dict["resources"].num_wires`

`gate_types`

is no longer supported, use`specs_dict["resources"].gate_types`

`gate_sizes`

is no longer supported, use`specs_dict["resources"].gate_sizes`

`depth`

is no longer supported, use`specs_dict["resources"].depth`

`qml.math.purity`

,`qml.math.vn_entropy`

,`qml.math.mutual_info`

,`qml.math.fidelity`

,`qml.math.relative_entropy`

, and`qml.math.max_entropy`

no longer support state vectors as input. (#4322)The private

`QuantumScript._prep`

list has been removed, and prep operations now go into the`_ops`

list. (#4485)

### Deprecations 👋

`qml.enable_return`

and`qml.disable_return`

have been deprecated. Please avoid calling`disable_return`

, as the old return system has been deprecated along with these switch functions. (#4316)`qml.qchem.jordan_wigner`

has been deprecated. Use`qml.jordan_wigner`

instead. List input to define the fermionic operator has also been deprecated; the fermionic operators in the`qml.fermi`

module should be used instead. (#4332)The

`qml.RandomLayers.compute_decomposition`

keyword argument`ratio_imprimitive`

will be changed to`ratio_imprim`

to match the call signature of the operation. (#4314)The CV observables

`qml.X`

and`qml.P`

have been deprecated. Use`qml.QuadX`

and`qml.QuadP`

instead. (#4330)The method

`tape.unwrap()`

and corresponding`UnwrapTape`

and`Unwrap`

classes have been deprecated. Use`convert_to_numpy_parameters`

instead. (#4344)The

`mode`

keyword argument in QNode has been deprecated, as it was only used in the old return system (which has also been deprecated). Please use`grad_on_execution`

instead. (#4316)The

`QuantumScript.set_parameters`

method and the`QuantumScript.data`

setter have been deprecated. Please use`QuantumScript.bind_new_parameters`

instead. (#4346)The

`__eq__`

and`__hash__`

dunder methods of`Operator`

and`MeasurementProcess`

will now raise warnings to reflect upcoming changes to operator and measurement process equality and hashing. (#4144) (#4454) (#4489) (#4498)The

`sampler_seed`

argument of`qml.gradients.spsa_grad`

has been deprecated, along with a bug fix of the seed-setting behaviour. Instead, the`sampler_rng`

argument should be set, either to an integer value, which will be used to create a PRNG internally or to a NumPy pseudo-random number generator created via`np.random.default_rng(seed)`

. (4165)

### Documentation 📝

The

`qml.pulse.transmon_interaction`

and`qml.pulse.transmon_drive`

documentation has been updated. #4327`qml.ApproxTimeEvolution.compute_decomposition()`

now has a code example. (#4354)The documentation for

`qml.devices.experimental.Device`

has been improved to clarify some aspects of its use. (#4391)Input types and sources for operators in

`qml.import_operator`

are specified. (#4476)

### Bug fixes 🐛

`qml.Projector`

is pickle-able again. (#4452)`_copy_and_shift_params`

does not cast or convert integral types, just relying on`+`

and`*`

‘s casting rules in this case. (#4477)Sparse matrix calculations of

`SProd`

s containing a`Tensor`

are now allowed. When using`Tensor.sparse_matrix()`

, it is recommended to use the`wire_order`

keyword argument over`wires`

. (#4424)`op.adjoint`

has been replaced with`qml.adjoint`

in`QNSPSAOptimizer`

. (#4421)`jax.ad`

(deprecated) has been replaced by`jax.interpreters.ad`

. (#4403)`metric_tensor`

stops accidentally catching errors that stem from flawed wires assignments in the original circuit, leading to recursion errors. (#4328)A warning is now raised if control indicators are hidden when calling

`qml.draw_mpl`

(#4295)`qml.qinfo.purity`

now produces correct results with custom wire labels. (#4331)`default.qutrit`

now supports all qutrit operations used with`qml.adjoint`

. (#4348)The observable data of

`qml.GellMann`

now includes its index, allowing correct comparison between instances of`qml.GellMann`

, as well as Hamiltonians and Tensors containing`qml.GellMann`

. (#4366)`qml.transforms.merge_amplitude_embedding`

now works correctly when the`AmplitudeEmbedding`

s have a batch dimension. (#4353)The

`jordan_wigner`

function has been modified to work with Hamiltonians built with an active space. (#4372)When a

`style`

option is not provided,`qml.draw_mpl`

uses the current style set from`qml.drawer.use_style`

instead of`black_white`

. (#4357)`qml.devices.qubit.preprocess.validate_and_expand_adjoint`

no longer sets the trainable parameters of the expanded tape. (#4365)`qml.default_expand_fn`

now selectively expands operations or measurements allowing more operations to be executed in circuits when measuring non-qwc Hamiltonians. (#4401)`qml.ControlledQubitUnitary`

no longer reports`has_decomposition`

as`True`

when it does not really have a decomposition. (#4407)`qml.transforms.split_non_commuting`

now correctly works on tapes containing both`expval`

and`var`

measurements. (#4426)Subtracting a

`Prod`

from another operator now works as expected. (#4441)The

`sampler_seed`

argument of`qml.gradients.spsa_grad`

has been changed to`sampler_rng`

. One can either provide an integer, which will be used to create a PRNG internally. Previously, this lead to the same direction being sampled, when`num_directions`

is greater than 1. Alternatively, one can provide a NumPy PRNG, which allows reproducibly calling`spsa_grad`

without getting the same results every time. (4165) (4482)`qml.math.get_dtype_name`

now works with autograd array boxes. (#4494)The backprop gradient of

`qml.math.fidelity`

is now correct. (#4380)

### Contributors ✍️

This release contains contributions from (in alphabetical order):

Utkarsh Azad, Thomas Bromley, Isaac De Vlugt, Amintor Dusko, Stepan Fomichev, Lillian M. A. Frederiksen, Soran Jahangiri, Edward Jiang, Korbinian Kottmann, Ivana Kurečić, Christina Lee, Vincent Michaud-Rioux, Romain Moyard, Lee James O’Riordan, Mudit Pandey, Borja Requena, Matthew Silverman, Jay Soni, David Wierichs, Frederik Wilde.

- orphan

## Release 0.31.0¶

### New features since last release

#### Seamlessly create and combine fermionic operators 🔬

Fermionic operators and arithmetic are now available. (#4191) (#4195) (#4200) (#4201) (#4209) (#4229) (#4253) (#4255) (#4262) (#4278)

There are a couple of ways to create fermionic operators with this new feature:

`qml.FermiC`

and`qml.FermiA`

: the fermionic creation and annihilation operators, respectively. These operators are defined by passing the index of the orbital that the fermionic operator acts on. For instance, the operators`a⁺(0)`

and`a(3)`

are respectively constructed as>>> qml.FermiC(0) a⁺(0) >>> qml.FermiA(3) a(3)

These operators can be composed with (

`*`

) and linearly combined with (`+`

and`-`

) other Fermi operators to create arbitrary fermionic Hamiltonians. Multiplying several Fermi operators together creates an operator that we call a Fermi word:>>> word = qml.FermiC(0) * qml.FermiA(0) * qml.FermiC(3) * qml.FermiA(3) >>> word a⁺(0) a(0) a⁺(3) a(3)

Fermi words can be linearly combined to create a fermionic operator that we call a Fermi sentence:

>>> sentence = 1.2 * word - 0.345 * qml.FermiC(3) * qml.FermiA(3) >>> sentence 1.2 * a⁺(0) a(0) a⁺(3) a(3) - 0.345 * a⁺(3) a(3)

via qml.fermi.from_string: create a fermionic operator that represents multiple creation and annihilation operators being multiplied by each other (a Fermi word).

>>> qml.fermi.from_string('0+ 1- 0+ 1-') a⁺(0) a(1) a⁺(0) a(1) >>> qml.fermi.from_string('0^ 1 0^ 1') a⁺(0) a(1) a⁺(0) a(1)

Fermi words created with

`from_string`

can also be linearly combined to create a Fermi sentence:>>> word1 = qml.fermi.from_string('0+ 0- 3+ 3-') >>> word2 = qml.fermi.from_string('3+ 3-') >>> sentence = 1.2 * word1 + 0.345 * word2 >>> sentence 1.2 * a⁺(0) a(0) a⁺(3) a(3) + 0.345 * a⁺(3) a(3)

Additionally, any fermionic operator, be it a single fermionic creation/annihilation operator, a Fermi word, or a Fermi sentence, can be mapped to the qubit basis by using qml.jordan_wigner:

>>> qml.jordan_wigner(sentence) ((0.4725+0j)*(Identity(wires=[0]))) + ((-0.4725+0j)*(PauliZ(wires=[3]))) + ((-0.3+0j)*(PauliZ(wires=[0]))) + ((0.3+0j)*(PauliZ(wires=[0]) @ PauliZ(wires=[3])))

Learn how to create fermionic Hamiltonians describing some simple chemical systems by checking out our fermionic operators demo!

#### Workflow-level resource estimation 🧮

PennyLane’s Tracker now monitors the resource requirements of circuits being executed by the device. (#4045) (#4110)

Suppose we have a workflow that involves executing circuits with different qubit numbers. We can obtain the resource requirements as a function of the number of qubits by executing the workflow with the

`Tracker`

context:dev = qml.device("default.qubit", wires=4) @qml.qnode(dev) def circuit(n_wires): for i in range(n_wires): qml.Hadamard(i) return qml.probs(range(n_wires)) with qml.Tracker(dev) as tracker: for i in range(1, 5): circuit(i)

The resource requirements of individual circuits can then be inspected as follows:

>>> resources = tracker.history["resources"] >>> resources[0] wires: 1 gates: 1 depth: 1 shots: Shots(total=None) gate_types: {'Hadamard': 1} gate_sizes: {1: 1} >>> [r.num_wires for r in resources] [1, 2, 3, 4]

Moreover, it is possible to predict the resource requirements without evaluating circuits using the

`null.qubit`

device, which follows the standard execution pipeline but returns numeric zeros. Consider the following workflow that takes the gradient of a`50`

-qubit circuit:n_wires = 50 dev = qml.device("null.qubit", wires=n_wires) weight_shape = qml.StronglyEntanglingLayers.shape(2, n_wires) weights = np.random.random(weight_shape, requires_grad=True) @qml.qnode(dev, diff_method="parameter-shift") def circuit(weights): qml.StronglyEntanglingLayers(weights, wires=range(n_wires)) return qml.expval(qml.PauliZ(0)) with qml.Tracker(dev) as tracker: qml.grad(circuit)(weights)

The tracker can be inspected to extract resource requirements without requiring a 50-qubit circuit run:

>>> tracker.totals {'executions': 451, 'batches': 2, 'batch_len': 451} >>> tracker.history["resources"][0] wires: 50 gates: 200 depth: 77 shots: Shots(total=None) gate_types: {'Rot': 100, 'CNOT': 100} gate_sizes: {1: 100, 2: 100}

Custom operations can now be constructed that solely define resource requirements — an explicit decomposition or matrix representation is not needed. (#4033)

PennyLane is now able to estimate the total resource requirements of circuits that include one or more of these operations, allowing you to estimate requirements for high-level algorithms composed of abstract subroutines.

These operations can be defined by inheriting from ResourcesOperation and overriding the

`resources()`

method to return an appropriate Resources object:class CustomOp(qml.resource.ResourcesOperation): def resources(self): n = len(self.wires) r = qml.resource.Resources( num_wires=n, num_gates=n ** 2, depth=5, ) return r

>>> wires = [0, 1, 2] >>> c = CustomOp(wires) >>> c.resources() wires: 3 gates: 9 depth: 5 shots: Shots(total=None) gate_types: {} gate_sizes: {}

A quantum circuit that contains

`CustomOp`

can be created and inspected using qml.specs:dev = qml.device("default.qubit", wires=wires) @qml.qnode(dev) def circ(): qml.PauliZ(wires=0) CustomOp(wires) return qml.state()

>>> specs = qml.specs(circ)() >>> specs["resources"].depth 6

#### Community contributions from UnitaryHack 🤝

ParametrizedHamiltonian now has an improved string representation. (#4176)

>>> def f1(p, t): return p[0] * jnp.sin(p[1] * t) >>> def f2(p, t): return p * t >>> coeffs = [2., f1, f2] >>> observables = [qml.PauliX(0), qml.PauliY(0), qml.PauliZ(0)] >>> qml.dot(coeffs, observables) (2.0*(PauliX(wires=[0]))) + (f1(params_0, t)*(PauliY(wires=[0]))) + (f2(params_1, t)*(PauliZ(wires=[0])))

The quantum information module now supports trace distance. (#4181)

Two cases are enabled for calculating the trace distance:

A QNode transform via qml.qinfo.trace_distance:

dev = qml.device('default.qubit', wires=2) @qml.qnode(dev) def circuit(param): qml.RY(param, wires=0) qml.CNOT(wires=[0, 1]) return qml.state()

>>> trace_distance_circuit = qml.qinfo.trace_distance(circuit, circuit, wires0=[0], wires1=[0]) >>> x, y = np.array(0.4), np.array(0.6) >>> trace_distance_circuit((x,), (y,)) 0.047862689546603415

Flexible post-processing via qml.math.trace_distance:

>>> rho = np.array([[0.3, 0], [0, 0.7]]) >>> sigma = np.array([[0.5, 0], [0, 0.5]]) >>> qml.math.trace_distance(rho, sigma) 0.19999999999999998

It is now possible to prepare qutrit basis states with qml.QutritBasisState. (#4185)

wires = range(2) dev = qml.device("default.qutrit", wires=wires) @qml.qnode(dev) def qutrit_circuit(): qml.QutritBasisState([1, 1], wires=wires) qml.TAdd(wires=wires) return qml.probs(wires=1)

>>> qutrit_circuit() array([0., 0., 1.])

A new transform called one_qubit_decomposition has been added to provide a unified interface for decompositions of a single-qubit unitary matrix into sequences of X, Y, and Z rotations. All decompositions simplify the rotations angles to be between

`0`

and`4`

pi. (#4210) (#4246)>>> from pennylane.transforms import one_qubit_decomposition >>> U = np.array([[-0.28829348-0.78829734j, 0.30364367+0.45085995j], ... [ 0.53396245-0.10177564j, 0.76279558-0.35024096j]]) >>> one_qubit_decomposition(U, 0, "ZYZ") [RZ(tensor(12.32427531, requires_grad=True), wires=[0]), RY(tensor(1.14938178, requires_grad=True), wires=[0]), RZ(tensor(1.73305815, requires_grad=True), wires=[0])] >>> one_qubit_decomposition(U, 0, "XYX", return_global_phase=True) [RX(tensor(10.84535137, requires_grad=True), wires=[0]), RY(tensor(1.39749741, requires_grad=True), wires=[0]), RX(tensor(0.45246584, requires_grad=True), wires=[0]), (0.38469215914523336-0.9230449299422961j)*(Identity(wires=[0]))]

The

`has_unitary_generator`

attribute in`qml.ops.qubit.attributes`

no longer contains operators with non-unitary generators. (#4183)PennyLane Docker builds have been updated to include the latest plugins and interface versions. (#4178)

#### Extended support for differentiating pulses ⚛️

The stochastic parameter-shift gradient method can now be used with hardware-compatible Hamiltonians. (#4132) (#4215)

This new feature generalizes the stochastic parameter-shift gradient transform for pulses (

`stoch_pulse_grad`

) to support Hermitian generating terms beyond just Pauli words in pulse Hamiltonians, which makes it hardware-compatible.A new differentiation method called qml.gradients.pulse_generator is available, which combines classical processing with the parameter-shift rule for multivariate gates to differentiate pulse programs. Access it in your pulse programs by setting

`diff_method=qml.gradients.pulse_generator`

. (#4160)`qml.pulse.ParametrizedEvolution`

now uses*batched*compressed sparse row (`BCSR`

) format. This allows for computing Jacobians of the unitary directly even when`dense=False`

. (#4126)def U(params): H = jnp.polyval * qml.PauliZ(0) # time dependent Hamiltonian Um = qml.evolve(H, dense=False)(params, t=10.) return qml.matrix(Um) params = jnp.array([[0.5]], dtype=complex) jac = jax.jacobian(U, holomorphic=True)(params)

#### Broadcasting and other tweaks to Torch and Keras layers 🦾

The

`TorchLayer`

and`KerasLayer`

integrations with`torch.nn`

and`Keras`

have been upgraded. Consider the following`TorchLayer`

:n_qubits = 2 dev = qml.device("default.qubit", wires=n_qubits) @qml.qnode(dev) def qnode(inputs, weights): qml.AngleEmbedding(inputs, wires=range(n_qubits)) qml.BasicEntanglerLayers(weights, wires=range(n_qubits)) return [qml.expval(qml.PauliZ(wires=i)) for i in range(n_qubits)] n_layers = 6 weight_shapes = {"weights": (n_layers, n_qubits)} qlayer = qml.qnn.TorchLayer(qnode, weight_shapes)

The following features are now available:

Native support for parameter broadcasting. (#4131)

>>> batch_size = 10 >>> inputs = torch.rand((batch_size, n_qubits)) >>> qlayer(inputs) >>> dev.num_executions == 1 True

The ability to draw a

`TorchLayer`

and`KerasLayer`

using`qml.draw()`

and`qml.draw_mpl()`

. (#4197)>>> print(qml.draw(qlayer, show_matrices=False)(inputs)) 0: ─╭AngleEmbedding(M0)─╭BasicEntanglerLayers(M1)─┤ <Z> 1: ─╰AngleEmbedding(M0)─╰BasicEntanglerLayers(M1)─┤ <Z>

Support for

`KerasLayer`

model saving and clearer instructions on`TorchLayer`

model saving. (#4149) (#4158)>>> torch.save(qlayer.state_dict(), "weights.pt") # Saving >>> qlayer.load_state_dict(torch.load("weights.pt")) # Loading >>> qlayer.eval()

Hybrid models containing

`KerasLayer`

or`TorchLayer`

objects can also be saved and loaded.

### Improvements 🛠

#### A more flexible projector

`qml.Projector`

now accepts a state vector representation, which enables the creation of projectors in any basis. (#4192)dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(state): return qml.expval(qml.Projector(state, wires=[0, 1])) zero_state = [0, 0] plusplus_state = np.array([1, 1, 1, 1]) / 2

>>> circuit(zero_state) tensor(1., requires_grad=True) >>> circuit(plusplus_state) tensor(0.25, requires_grad=True)

#### Do more with qutrits

Three qutrit rotation operators have been added that are analogous to

`RX`

,`RY`

, and`RZ`

:`qml.TRX`

: an X rotation`qml.TRY`

: a Y rotation`qml.TRZ`

: a Z rotation

Qutrit devices now support parameter-shift differentiation. (#2845)

#### The qchem module

`qchem.molecular_hamiltonian()`

,`qchem.qubit_observable()`

,`qchem.import_operator()`

, and`qchem.dipole_moment()`

now return an arithmetic operator if`enable_new_opmath()`

is active. (#4138) (#4159) (#4189) (#4204)Non-cubic lattice support for all electron resource estimation has been added. (3956)

The

`qchem.molecular_hamiltonian()`

function has been upgraded to support custom wires for constructing differentiable Hamiltonians. The zero imaginary component of the Hamiltonian coefficients have been removed. (#4050) (#4094)Jordan-Wigner transforms that cache Pauli gate objects have been accelerated. (#4046)

An error is now raised by

`qchem.molecular_hamiltonian`

when the`dhf`

method is used for an open-shell system. This duplicates a similar error in`qchem.Molecule`

but makes it clear that the`pyscf`

backend can be used for open-shell calculations. (#4058)Updated various qubit tapering methods to support operator arithmetic. (#4252)

#### Next-generation device API

The new device interface has been integrated with

`qml.execute`

for autograd, backpropagation, and no differentiation. (#3903)Support for adjoint differentiation has been added to the

`DefaultQubit2`

device. (#4037)A new function called

`measure_with_samples`

that returns a sample-based measurement result given a state has been added. (#4083) (#4093) (#4162) (#4254)`DefaultQubit2.preprocess`

now returns a new`ExecutionConfig`

object with decisions for`gradient_method`

,`use_device_gradient`

, and`grad_on_execution`

. (#4102)Support for sample-based measurements has been added to the

`DefaultQubit2`

device. (#4105) (#4114) (#4133) (#4172)The

`DefaultQubit2`

device now has a`seed`

keyword argument. (#4120)Added a

`dense`

keyword to`ParametrizedEvolution`

that allows forcing dense or sparse matrices. (#4079) (#4095) (#4285)Adds the Type variables

`pennylane.typing.Result`

and`pennylane.typing.ResultBatch`

for type hinting the result of an execution. (#4018)`qml.devices.ExecutionConfig`

no longer has a`shots`

property, as it is now on the`QuantumScript`

.

It now has a`use_device_gradient`

property.`ExecutionConfig.grad_on_execution = None`

indicates a request for`"best"`

, instead of a string. (#4102)The new device interface for Jax has been integrated with

`qml.execute`

. (#4137)The new device interface is now integrated with

`qml.execute`

for Tensorflow. (#4169)The experimental device

`DefaultQubit2`

now supports`qml.Snapshot`

. (#4193)The experimental device interface is integrated with the

`QNode`

. (#4196)The new device interface in integrated with

`qml.execute`

for Torch. (#4257)

#### Handling shots

`QuantumScript`

now has a`shots`

property, allowing shots to be tied to executions instead of devices. (#4067) (#4103) (#4106) (#4112)Several Python built-in functions are now properly defined for instances of the

`Shots`

class.`print`

: printing`Shots`

instances is now human-readable`str`

: converting`Shots`

instances to human-readable strings`==`

: equating two different`Shots`

instances`hash`

: obtaining the hash values of`Shots`

instances

`qml.devices.ExecutionConfig`

no longer has a`shots`

property, as it is now on the`QuantumScript`

. It now has a`use_device_gradient`

property.`ExecutionConfig.grad_on_execution = None`

indicates a request for`"best"`

instead of a string. (#4102)`QuantumScript.shots`

has been integrated with QNodes so that shots are placed on the`QuantumScript`

during`QNode`

construction. (#4110)The

`gradients`

module has been updated to use the new`Shots`

object internally (#4152)

#### Operators

`qml.prod`

now accepts a single quantum function input for creating new`Prod`

operators. (#4011)`DiagonalQubitUnitary`

now decomposes into`RZ`

,`IsingZZ`

and`MultiRZ`

gates instead of a`QubitUnitary`

operation with a dense matrix. (#4035)All objects being queued in an

`AnnotatedQueue`

are now wrapped so that`AnnotatedQueue`

is not dependent on the has of any operators or measurement processes. (#4087)A

`dense`

keyword to`ParametrizedEvolution`

that allows forcing dense or sparse matrices has been added. (#4079) (#4095)Added a new function

`qml.ops.functions.bind_new_parameters`

that creates a copy of an operator with new parameters without mutating the original operator. (#4113) (#4256)`qml.CY`

has been moved from`qml.ops.qubit.non_parametric_ops`

to`qml.ops.op_math.controlled_ops`

and now inherits from`qml.ops.op_math.ControlledOp`

. (#4116)`qml.CZ`

now inherits from the`ControlledOp`

class and supports exponentiation to arbitrary powers with`pow`

, which is no longer limited to integers. It also supports`sparse_matrix`

and`decomposition`

representations. (#4117)The construction of the Pauli representation for the

`Sum`

class is now faster. (#4142)`qml.drawer.drawable_layers.drawable_layers`

and`qml.CircuitGraph`

have been updated to not rely on`Operator`

equality or hash to work correctly. (#4143)

#### Other improvements

A transform dispatcher and program have been added. (#4109) (#4187)

Reduced density matrix functionality has been added via

`qml.math.reduce_dm`

and`qml.math.reduce_statevector`

. Both functions have broadcasting support. (#4173)The following functions in

`qml.qinfo`

now support parameter broadcasting:`reduced_dm`

`purity`

`vn_entropy`

`mutual_info`

`fidelity`

`relative_entropy`

`trace_distance`

The following functions in

`qml.math`

now support parameter broadcasting:`purity`

`vn_entropy`

`mutual_info`

`fidelity`

`relative_entropy`

`max_entropy`

`sqrt_matrix`

`pulse.ParametrizedEvolution`

now raises an error if the number of input parameters does not match the number of parametrized coefficients in the`ParametrizedHamiltonian`

that generates it. An exception is made for`HardwareHamiltonian`

s which are not checked. (#4216)The default value for the

`show_matrices`

keyword argument in all drawing methods is now`True`

. This allows for quick insights into broadcasted tapes, for example. (#3920)Type variables for

`qml.typing.Result`

and`qml.typing.ResultBatch`

have been added for type hinting the result of an execution. (#4108)The Jax-JIT interface now uses symbolic zeros to determine trainable parameters. (4075)

A new function called

`pauli.pauli_word_prefactor()`

that extracts the prefactor for a given Pauli word has been added. (#4164)Variable-length argument lists of functions and methods in some docstrings is now more clear. (#4242)

`qml.drawer.drawable_layers.drawable_layers`

and`qml.CircuitGraph`

have been updated to not rely on`Operator`

equality or hash to work correctly. (#4143)Drawing mid-circuit measurements connected by classical control signals to conditional operations is now possible. (#4228)

The autograd interface now submits all required tapes in a single batch on the backward pass. (#4245)

### Breaking changes 💔

The default value for the

`show_matrices`

keyword argument in all drawing methods is now`True`

. This allows for quick insights into broadcasted tapes, for example. (#3920)`DiagonalQubitUnitary`

now decomposes into`RZ`

,`IsingZZ`

, and`MultiRZ`

gates rather than a`QubitUnitary`

. (#4035)Jax trainable parameters are now

`Tracer`

instead of`JVPTracer`

. It is not always the right definition for the JIT interface, but we update them in the custom JVP using symbolic zeros. (4075)The experimental Device interface

`qml.devices.experimental.Device`

now requires that the`preprocess`

method also returns an`ExecutionConfig`

object. This allows the device to choose what`"best"`

means for various hyperparameters like`gradient_method`

and`grad_on_execution`

. (#4007) (#4102)Gradient transforms with Jax no longer support

`argnum`

. Use`argnums`

instead. (#4076)`qml.collections`

,`qml.op_sum`

, and`qml.utils.sparse_hamiltonian`

have been removed. (#4071)The

`pennylane.transforms.qcut`

module now uses`(op, id(op))`

as nodes in directed multigraphs that are used within the circuit cutting workflow instead of`op`

. This change removes the dependency of the module on the hash of operators. (#4227)`Operator.data`

now returns a`tuple`

instead of a`list`

. (#4222)The pulse differentiation methods,

`pulse_generator`

and`stoch_pulse_grad`

, now raise an error when they are applied to a QNode directly. Instead, use differentiation via a JAX entry point (`jax.grad`

,`jax.jacobian`

, …). (#4241)

### Deprecations 👋

`LieAlgebraOptimizer`

has been renamed to`RiemannianGradientOptimizer`

. [(#4153)(https://github.com/PennyLaneAI/pennylane/pull/4153)]`Operation.base_name`

has been deprecated. Please use`Operation.name`

or`type(op).__name__`

instead.`QuantumScript`

‘s`name`

keyword argument and property have been deprecated. This also affects`QuantumTape`

and`OperationRecorder`

. (#4141)The

`qml.grouping`

module has been removed. Its functionality has been reorganized in the`qml.pauli`

module.The public methods of

`DefaultQubit`

are pending changes to follow the new device API, as used in`DefaultQubit2`

. Warnings have been added to the docstrings to reflect this. (#4145)`qml.math.reduced_dm`

has been deprecated. Please use`qml.math.reduce_dm`

or`qml.math.reduce_statevector`

instead. (#4173)`qml.math.purity`

,`qml.math.vn_entropy`

,`qml.math.mutual_info`

,`qml.math.fidelity`

,`qml.math.relative_entropy`

, and`qml.math.max_entropy`

no longer support state vectors as input. Please call`qml.math.dm_from_state_vector`

on the input before passing to any of these functions. (#4186)The

`do_queue`

keyword argument in`qml.operation.Operator`

has been deprecated. Instead of setting`do_queue=False`

, use the`qml.QueuingManager.stop_recording()`

context. (#4148)`zyz_decomposition`

and`xyx_decomposition`

are now deprecated in favour of`one_qubit_decomposition`

. (#4230)

### Documentation 📝

The documentation is updated to construct

`QuantumTape`

upon initialization instead of with queuing. (#4243)The docstring for

`qml.ops.op_math.Pow.__new__`

is now complete and it has been updated along with`qml.ops.op_math.Adjoint.__new__`

. (#4231)The docstring for

`qml.grad`

now states that it should be used with the Autograd interface only. (#4202)The description of

`mult`

in the`qchem.Molecule`

docstring now correctly states the value of`mult`

that is supported. (#4058)

### Bug Fixes 🐛

Fixed adjoint jacobian results with

`grad_on_execution=False`

in the JAX-JIT interface. (4217)Fixed the matrix of

`SProd`

when the coefficient is tensorflow and the target matrix is not`complex128`

. (#4249)Fixed a bug where

`stoch_pulse_grad`

would ignore prefactors of rescaled Pauli words in the generating terms of a pulse Hamiltonian. (4156)Fixed a bug where the wire ordering of the

`wires`

argument to`qml.density_matrix`

was not taken into account. (#4072)A patch in

`interfaces/autograd.py`

that checks for the`strawberryfields.gbs`

device has been removed. That device is pinned to PennyLane <= v0.29.0, so that patch is no longer necessary. (#4089)`qml.pauli.are_identical_pauli_words`

now treats all identities as equal. Identity terms on Hamiltonians with non-standard wire orders are no longer eliminated. (#4161)`qml.pauli_sentence()`

is now compatible with empty Hamiltonians`qml.Hamiltonian([], [])`

. (#4171)Fixed a bug with Jax where executing multiple tapes with

`gradient_fn="device"`

would fail. (#4190)A more meaningful error message is raised when broadcasting with adjoint differentiation on

`DefaultQubit`

. (#4203)The

`has_unitary_generator`

attribute in`qml.ops.qubit.attributes`

no longer contains operators with non-unitary generators. (#4183)Fixed a bug where

`op = qml.qsvt()`

was incorrect up to a global phase when using`convention="Wx""`

and`qml.matrix(op)`

. (#4214)Fixed a buggy calculation of the angle in

`xyx_decomposition`

that causes it to give an incorrect decomposition. An`if`

conditional was intended to prevent divide by zero errors, but the division was by the sine of the argument. So, any multiple of $pi$ should trigger the conditional, but it was only checking if the argument was 0. Example:`qml.Rot(2.3, 2.3, 2.3)`

(#4210)Fixed bug that caused

`ShotAdaptiveOptimizer`

to truncate dimensions of parameter-distributed shots during optimization. (#4240)`Sum`

observables can now have trainable parameters. (#4251) (#4275)

### Contributors ✍️

This release contains contributions from (in alphabetical order):

Venkatakrishnan AnushKrishna, Utkarsh Azad, Thomas Bromley, Isaac De Vlugt, Lillian M. A. Frederiksen, Emiliano Godinez Ramirez Nikhil Harle Soran Jahangiri, Edward Jiang, Korbinian Kottmann, Christina Lee, Vincent Michaud-Rioux, Romain Moyard, Tristan Nemoz, Mudit Pandey, Manul Patel, Borja Requena, Modjtaba Shokrian-Zini, Mainak Roy, Matthew Silverman, Jay Soni, Edward Thomas, David Wierichs, Frederik Wilde.

- orphan

## Release 0.30.0¶

### New features since last release

#### Pulse programming on hardware ⚛️🔬

Support for loading time-dependent Hamiltonians that are compatible with quantum hardware has been added, making it possible to load a Hamiltonian that describes an ensemble of Rydberg atoms or a collection of transmon qubits. (#3749) (#3911) (#3930) (#3936) (#3966) (#3987) (#4021) (#4040)

Rydberg atoms are the foundational unit for neutral atom quantum computing. A Rydberg-system Hamiltonian can be constructed from a drive term —

`qml.pulse.rydberg_drive`

— and an interaction term —`qml.pulse.rydberg_interaction`

:from jax import numpy as jnp atom_coordinates = [[0, 0], [0, 4], [4, 0], [4, 4]] wires = [0, 1, 2, 3] amplitude = lambda p, t: p * jnp.sin(jnp.pi * t) phase = jnp.pi / 2 detuning = 3 * jnp.pi / 4 H_d = qml.pulse.rydberg_drive(amplitude, phase, detuning, wires) H_i = qml.pulse.rydberg_interaction(atom_coordinates, wires) H = H_d + H_i

The time-dependent Hamiltonian

`H`

can be used in a PennyLane pulse-level differentiable circuit:dev = qml.device("default.qubit.jax", wires=wires) @qml.qnode(dev, interface="jax") def circuit(params): qml.evolve(H)(params, t=[0, 10]) return qml.expval(qml.PauliZ(0))

>>> params = jnp.array([2.4]) >>> circuit(params) Array(0.6316659, dtype=float32) >>> import jax >>> jax.grad(circuit)(params) Array([1.3116529], dtype=float32)

The qml.pulse page contains additional details. Check out our release blog post for a demonstration of how to perform the execution on actual hardware!

A pulse-level circuit can now be differentiated using a stochastic parameter-shift method. (#3780) (#3900) (#4000) (#4004)

The new qml.gradient.stoch_pulse_grad differentiation method unlocks stochastic-parameter-shift differentiation for pulse-level circuits. The current version of this new method is restricted to Hamiltonians composed of parametrized Pauli words, but future updates to extend to parametrized Pauli sentences can allow this method to be compatible with hardware-based systems such as an ensemble of Rydberg atoms.

This method can be activated by setting

`diff_method`

to qml.gradient.stoch_pulse_grad:>>> dev = qml.device("default.qubit.jax", wires=2) >>> sin = lambda p, t: jax.numpy.sin(p * t) >>> ZZ = qml.PauliZ(0) @ qml.PauliZ(1) >>> H = 0.5 * qml.PauliX(0) + qml.pulse.constant * ZZ + sin * qml.PauliX(1) >>> @qml.qnode(dev, interface="jax", diff_method=qml.gradients.stoch_pulse_grad) >>> def ansatz(params): ... qml.evolve(H)(params, (0.2, 1.)) ... return qml.expval(qml.PauliY(1)) >>> params = [jax.numpy.array(0.4), jax.numpy.array(1.3)] >>> jax.grad(ansatz)(params) [Array(0.16921353, dtype=float32, weak_type=True), Array(-0.2537478, dtype=float32, weak_type=True)]

#### Quantum singular value transformation 🐛➡️🦋

PennyLane now supports the quantum singular value transformation (QSVT), which describes how a quantum circuit can be constructed to apply a polynomial transformation to the singular values of an input matrix. (#3756) (#3757) (#3758) (#3905) (#3909) (#3926) (#4023)

Consider a matrix

`A`

along with a vector`angles`

that describes the target polynomial transformation. The`qml.qsvt`

function creates a corresponding circuit:dev = qml.device("default.qubit", wires=2) A = np.array([[0.1, 0.2], [0.3, 0.4]]) angles = np.array([0.1, 0.2, 0.3]) @qml.qnode(dev) def example_circuit(A): qml.qsvt(A, angles, wires=[0, 1]) return qml.expval(qml.PauliZ(wires=0))

This circuit is composed of

`qml.BlockEncode`

and`qml.PCPhase`

operations.>>> example_circuit(A) tensor(0.97777078, requires_grad=True) >>> print(example_circuit.qtape.expand(depth=1).draw(decimals=2)) 0: ─╭∏_ϕ(0.30)─╭BlockEncode(M0)─╭∏_ϕ(0.20)─╭BlockEncode(M0)†─╭∏_ϕ(0.10)─┤ <Z> 1: ─╰∏_ϕ(0.30)─╰BlockEncode(M0)─╰∏_ϕ(0.20)─╰BlockEncode(M0)†─╰∏_ϕ(0.10)─┤

The qml.qsvt function creates a circuit that is targeted at simulators due to the use of matrix-based operations. For advanced users, you can use the operation-based

`qml.QSVT`

template to perform the transformation with a custom choice of unitary and projector operations, which may be hardware compatible if a decomposition is provided.The QSVT is a complex but powerful transformation capable of generalizing important algorithms like amplitude amplification. Stay tuned for a demo in the coming few weeks to learn more!

#### Intuitive QNode returns ↩️

An updated QNode return system has been introduced. PennyLane QNodes now return exactly what you tell them to! 🎉 (#3957) (#3969) (#3946) (#3913) (#3914) (#3934)

This was an experimental feature introduced in version 0.25 of PennyLane that was enabled via

`qml.enable_return()`

. Now, it’s the default return system. Let’s see how it works.Consider the following circuit:

import pennylane as qml dev = qml.device("default.qubit", wires=1) @qml.qnode(dev) def circuit(x): qml.RX(x, wires=0) return qml.expval(qml.PauliZ(0)), qml.probs(0)

In version 0.29 and earlier of PennyLane,

`circuit()`

would return a single length-3 array:>>> circuit(0.5) tensor([0.87758256, 0.93879128, 0.06120872], requires_grad=True)

In versions 0.30 and above,

`circuit()`

returns a length-2 tuple containing the expectation value and probabilities separately:>>> circuit(0.5) (tensor(0.87758256, requires_grad=True), tensor([0.93879128, 0.06120872], requires_grad=True))

You can find more details about this change, along with help and troubleshooting tips to solve any issues. If you still have questions, comments, or concerns, we encourage you to post on the PennyLane discussion forum.

#### A bunch of performance tweaks 🏃💨

Single-qubit operations that have multi-qubit control can now be decomposed more efficiently using fewer CNOT gates. (#3851)

Three decompositions from arXiv:2302.06377 are provided and compare favourably to the already-available

`qml.ops.ctrl_decomp_zyz`

:wires = [0, 1, 2, 3, 4, 5] control_wires = wires[1:] @qml.qnode(qml.device('default.qubit', wires=6)) def circuit(): with qml.QueuingManager.stop_recording(): # the decomposition does not un-queue the target target = qml.RX(np.pi/2, wires=0) qml.ops.ctrl_decomp_bisect(target, (1, 2, 3, 4, 5)) return qml.state() print(qml.draw(circuit, expansion_strategy="device")())

0: ──H─╭X──U(M0)─╭X──U(M0)†─╭X──U(M0)─╭X──U(M0)†──H─┤ State 1: ────├●────────│──────────├●────────│─────────────┤ State 2: ────├●────────│──────────├●────────│─────────────┤ State 3: ────╰●────────│──────────╰●────────│─────────────┤ State 4: ──────────────├●───────────────────├●────────────┤ State 5: ──────────────╰●───────────────────╰●────────────┤ State

A new decomposition to

`qml.SingleExcitation`

has been added that halves the number of CNOTs required. (3976)>>> qml.SingleExcitation.compute_decomposition(1.23, wires=(0,1)) [Adjoint(T(wires=[0])), Hadamard(wires=[0]), S(wires=[0]), Adjoint(T(wires=[1])), Adjoint(S(wires=[1])), Hadamard(wires=[1]), CNOT(wires=[1, 0]), RZ(-0.615, wires=[0]), RY(0.615, wires=[1]), CNOT(wires=[1, 0]), Adjoint(S(wires=[0])), Hadamard(wires=[0]), T(wires=[0]), Hadamard(wires=[1]), S(wires=[1]), T(wires=[1])]

The adjoint differentiation method can now be more efficient, avoiding the decomposition of operations that can be differentiated directly. Any operation that defines a

`generator()`

can be differentiated with the adjoint method. (#3874)For example, in version 0.29 the

`qml.CRY`

operation would be decomposed when calculating the adjoint-method gradient. Executing the code below shows that this decomposition no longer takes place in version 0.30 and`qml.CRY`

is differentiated directly:import jax from jax import numpy as jnp def compute_decomposition(self, phi, wires): print("A decomposition has been performed!") decomp_ops = [ qml.RY(phi / 2, wires=wires[1]), qml.CNOT(wires=wires), qml.RY(-phi / 2, wires=wires[1]), qml.CNOT(wires=wires), ] return decomp_ops qml.CRY.compute_decomposition = compute_decomposition dev = qml.device("default.qubit", wires=2) @qml.qnode(dev, diff_method="adjoint") def circuit(phi): qml.Hadamard(wires=0) qml.CRY(phi, wires=[0, 1]) return qml.expval(qml.PauliZ(1)) phi = jnp.array(0.5) jax.grad(circuit)(phi)

Derivatives are computed more efficiently when using

`jax.jit`

with gradient transforms; the trainable parameters are now set correctly instead of every parameter having to be set as trainable. (#3697)In the circuit below, only the derivative with respect to parameter

`b`

is now calculated:dev = qml.device("default.qubit", wires=2) @qml.qnode(dev, interface="jax-jit") def circuit(a, b): qml.RX(a, wires=0) qml.RY(b, wires=0) qml.CNOT(wires=[0, 1]) return qml.expval(qml.PauliZ(0)) a = jnp.array(0.4) b = jnp.array(0.5) jac = jax.jacobian(circuit, argnums=[1]) jac_jit = jax.jit(jac) jac_jit(a, b) assert len(circuit.tape.trainable_params) == 1

### Improvements 🛠

#### Next-generation device API

In this release and future releases, we will be making changes to our device API with the goal in mind to make developing plugins much easier for developers and unlock new device capabilities. Users shouldn’t yet feel any of these changes when using PennyLane, but here is what has changed this release:

Several functions in

`devices/qubit`

have been added or improved:`sample_state`

: returns a series of samples based on a given state vector and a number of shots. (#3720)`simulate`

: supports measuring expectation values of large observables such as`qml.Hamiltonian`

,`qml.SparseHamiltonian`

, and`qml.Sum`

. (#3759)`apply_operation`

: supports broadcasting. (#3852)`adjoint_jacobian`

: supports adjoint differentiation in the new qubit state-vector device. (#3790)

`qml.devices.qubit.preprocess`

now allows circuits with non-commuting observables. (#3857)`qml.devices.qubit.measure`

now computes the expectation values of`Hamiltonian`

and`Sum`

in a backpropagation-compatible way. (#3862)

#### Pulse programming

Here are the functions, classes, and more that were added or improved to facilitate simulating ensembles of Rydberg atoms: (#3749) (#3911) (#3930) (#3936) (#3966) (#3987) (#3889) (#4021)

`HardwareHamiltonian`

: an internal class that contains additional information about pulses and settings.`rydberg_interaction`

: a user-facing function that returns a`HardwareHamiltonian`

containing the Hamiltonian of the interaction of all the Rydberg atoms.`transmon_interaction`

: a user-facing function for constructing the Hamiltonian that describes the circuit QED interaction Hamiltonian of superconducting transmon systems.`drive`

: a user-facing function function that returns a`ParametrizedHamiltonian`

(`HardwareHamiltonian`

) containing the Hamiltonian of the interaction between a driving electro-magnetic field and a group of qubits.`rydberg_drive`

: a user-facing function that returns a`ParametrizedHamiltonian`

(`HardwareHamiltonian`

) containing the Hamiltonian of the interaction between a driving laser field and a group of Rydberg atoms.`max_distance`

: a keyword argument added to`qml.pulse.rydberg_interaction`

to allow for the removal of negligible contributions from atoms beyond`max_distance`

from each other.

`ParametrizedEvolution`

now takes two new Boolean keyword arguments:`return_intermediate`

and`complementary`

. They allow computing intermediate time evolution matrices. (#3900)Activating

`return_intermediate`

will return intermediate time evolution steps, for example for the matrix of the Operation, or of a quantum circuit when used in a QNode. Activating`complementary`

will make these intermediate steps be the*remaining*time evolution complementary to the output for`complementary=False`

. See the docstring for details.Hardware-compatible pulse sequence gradients with

`qml.gradient.stoch_pulse_grad`

can now be calculated faster using the new keyword argument`use_broadcasting`

. Executing a`ParametrizedEvolution`

that returns intermediate evolutions has increased performance using the state vector ODE solver, as well. (#4000) (#4004)

#### Intuitive QNode returns

The QNode keyword argument

`mode`

has been replaced by the boolean`grad_on_execution`

. (#3969)The

`"default.gaussian"`

device and parameter-shift CV both support the new return system, but only for single measurements. (#3946)Keras and Torch NN modules are now compatible with the new return type system. (#3913) (#3914)

`DefaultQutrit`

now supports the new return system. (#3934)

#### Performance improvements

The efficiency of

`tapering()`

,`tapering_hf()`

and`clifford()`

have been improved. (3942)The peak memory requirements of

`tapering()`

and`tapering_hf()`

have been improved when used for larger observables. (3977)Pauli arithmetic has been updated to convert to a Hamiltonian more efficiently. (#3939)

`Operator`

has a new Boolean attribute`has_generator`

. It returns whether or not the`Operator`

has a`generator`

defined.`has_generator`

is used in`qml.operation.has_gen`

, which improves its performance and extends differentiation support. (#3875)The performance of

`CompositeOp`

has been significantly improved now that it overrides determining whether it is being used with a batch of parameters (see`Operator._check_batching`

).`Hamiltonian`

also now overrides this, but it does nothing since it does not support batching. (#3915)The performance of a

`Sum`

operator has been significantly improved now that`is_hermitian`

checks that all coefficients are real if the operator has a pre-computed Pauli representation. (#3915)The

`coefficients`

function and the`visualize`

submodule of the`qml.fourier`

module now allow assigning different degrees for different parameters of the input function. (#3005)Previously, the arguments

`degree`

and`filter_threshold`

to`qml.fourier.coefficients`

were expected to be integers. Now, they can be a sequences of integers with one integer per function parameter (i.e.`len(degree)==n_inputs`

), resulting in a returned array with shape`(2*degrees[0]+1,..., 2*degrees[-1]+1)`

. The functions in`qml.fourier.visualize`

accordingly accept such arrays of coefficients.

#### Other improvements

A

`Shots`

class has been added to the`measurements`

module to hold shot-related data. (#3682)The custom JVP rules in PennyLane also now support non-scalar and mixed-shape tape parameters as well as multi-dimensional tape return types, like broadcasted

`qml.probs`

, for example. (#3766)The

`qchem.jordan_wigner`

function has been extended to support more fermionic operator orders. (#3754) (#3751)The

`AdaptiveOptimizer`

has been updated to use non-default user-defined QNode arguments. (#3765)Operators now use

`TensorLike`

types dunder methods. (#3749)`qml.QubitStateVector.state_vector`

now supports broadcasting. (#3852)`qml.SparseHamiltonian`

can now be applied to any wires in a circuit rather than being restricted to all wires in the circuit. (#3888)Operators can now be divided by scalars with

`/`

with the addition of the`Operation.__truediv__`

dunder method. (#3749)Printing an instance of

`MutualInfoMP`

now displays the distribution of the wires between the two subsystems. (#3898)`Operator.num_wires`

has been changed from an abstract value to`AnyWires`

. (#3919)`qml.transforms.sum_expand`

is not run in`Device.batch_transform`

if the device supports`Sum`

observables. (#3915)The type of

`n_electrons`

in`qml.qchem.Molecule`

has been set to`int`

. (#3885)Explicit errors have been added to

`QutritDevice`

if`classical_shadow`

or`shadow_expval`

is measured. (#3934)`QubitDevice`

now defines the private`_get_diagonalizing_gates(circuit)`

method and uses it when executing circuits. This allows devices that inherit from`QubitDevice`

to override and customize their definition of diagonalizing gates. (#3938)`retworkx`

has been renamed to`rustworkx`

to accommodate the change in the package name. (#3975)`Exp`

,`Sum`

,`Prod`

, and`SProd`

operator data is now a flat list instead of nested. (#3958) (#3983)`qml.transforms.convert_to_numpy_parameters`

has been added to convert a circuit with interface-specific parameters to one with only numpy parameters. This transform is designed to replace`qml.tape.Unwrap`

. (#3899)`qml.operation.WiresEnum.AllWires`

is now -2 instead of 0 to avoid the ambiguity between`op.num_wires = 0`

and`op.num_wires = AllWires`

. (#3978)Execution code has been updated to use the new

`qml.transforms.convert_to_numpy_parameters`

instead of`qml.tape.Unwrap`

. (#3989)A sub-routine of

`expand_tape`

has been converted into`qml.tape.tape.rotations_and_diagonal_measurements`

, a helper function that computes rotations and diagonal measurements for a tape with measurements with overlapping wires. (#3912)Various operators and templates have been updated to ensure that their decompositions only return lists of operators. (#3243)

The

`qml.operation.enable_new_opmath`

toggle has been introduced to cause dunder methods to return arithmetic operators instead of a`Hamiltonian`

or`Tensor`

. (#4008)>>> type(qml.PauliX(0) @ qml.PauliZ(1)) <class 'pennylane.operation.Tensor'> >>> qml.operation.enable_new_opmath() >>> type(qml.PauliX(0) @ qml.PauliZ(1)) <class 'pennylane.ops.op_math.prod.Prod'> >>> qml.operation.disable_new_opmath() >>> type(qml.PauliX(0) @ qml.PauliZ(1)) <class 'pennylane.operation.Tensor'>

A new data class called

`Resources`

has been added to store resources like the number of gates and circuit depth throughout a quantum circuit. (#3981)A new function called

`_count_resources()`

has been added to count the resources required when executing a`QuantumTape`

for a given number of shots. (#3996)`QuantumScript.specs`

has been modified to make use of the new`Resources`

class. This also modifies the output of`qml.specs()`

. (#4015)A new class called

`ResourcesOperation`

has been added to allow users to define operations with custom resource information. (#4026)For example, users can define a custom operation by inheriting from this new class:

>>> class CustomOp(qml.resource.ResourcesOperation): ... def resources(self): ... return qml.resource.Resources(num_wires=1, num_gates=2, ... gate_types={"PauliX": 2}) ... >>> CustomOp(wires=1) CustomOp(wires=[1])

Then, we can track and display the resources of the workflow using

`qml.specs()`

:>>> dev = qml.device("default.qubit", wires=[0,1]) >>> @qml.qnode(dev) ... def circ(): ... qml.PauliZ(wires=0) ... CustomOp(wires=1) ... return qml.state() ... >>> print(qml.specs(circ)()['resources']) wires: 2 gates: 3 depth: 1 shots: 0 gate_types: {'PauliZ': 1, 'PauliX': 2}

`MeasurementProcess.shape`

now accepts a`Shots`

object as one of its arguments to reduce exposure to unnecessary execution details. (#4012)

### Breaking changes 💔

The

`seed_recipes`

argument has been removed from`qml.classical_shadow`

and`qml.shadow_expval`

. (#4020)The tape method

`get_operation`

has an updated signature. (#3998)Both JIT interfaces are no longer compatible with JAX

`>0.4.3`

(we raise an error for those versions). (#3877)An operation that implements a custom

`generator`

method, but does not always return a valid generator, also has to implement a`has_generator`

property that reflects in which scenarios a generator will be returned. (#3875)Trainable parameters for the Jax interface are the parameters that are

`JVPTracer`

, defined by setting`argnums`

. Previously, all JAX tracers, including those used for JIT compilation, were interpreted to be trainable. (#3697)The keyword argument

`argnums`

is now used for gradient transforms using Jax instead of`argnum`

.`argnum`

is automatically converted to`argnums`

when using Jax and will no longer be supported in v0.31 of PennyLane. (#3697) (#3847)`qml.OrbitalRotation`

and, consequently,`qml.GateFabric`

are now more consistent with the interleaved Jordan-Wigner ordering. Previously, they were consistent with the sequential Jordan-Wigner ordering. (#3861)Some

`MeasurementProcess`

classes can now only be instantiated with arguments that they will actually use. For example, you can no longer create`StateMP(qml.PauliX(0))`

or`PurityMP(eigvals=(-1,1), wires=Wires(0))`

. (#3898)`Exp`

,`Sum`

,`Prod`

, and`SProd`

operator data is now a flat list, instead of nested. (#3958) (#3983)`qml.tape.tape.expand_tape`

and, consequentially,`QuantumScript.expand`

no longer update the input tape with rotations and diagonal measurements. Note that the newly expanded tape that is returned will still have the rotations and diagonal measurements. (#3912)`qml.Evolution`

now initializes the coefficient with a factor of`-1j`

instead of`1j`

. (#4024)

### Deprecations 👋

Nothing for this release!

### Documentation 📝

The documentation of

`QubitUnitary`

and`DiagonalQubitUnitary`

was clarified regarding the parameters of the operations. (#4031)A typo has been corrected in the documentation for the introduction to

`inspecting_circuits`

and`chemistry`

. (#3844)`Usage Details`

and`Theory`

sections have been separated in the documentation for`qml.qchem.taper_operation`

. (3977)

### Bug fixes 🐛

`ctrl_decomp_bisect`

and`ctrl_decomp_zyz`

are no longer used by default when decomposing controlled operations due to the presence of a global phase difference in the zyz decomposition of some target operators. (#4065)Fixed a bug where

`qml.math.dot`

returned a numpy array instead of an autograd array, breaking autograd derivatives in certain circumstances. (#4019)Operators now cast a

`tuple`

to an`np.ndarray`

as well as`list`

. (#4022)Fixed a bug where

`qml.ctrl`

with parametric gates was incompatible with PyTorch tensors on GPUs. (#4002)Fixed a bug where the broadcast expand results were stacked along the wrong axis for the new return system. (#3984)

A more informative error message is raised in

`qml.jacobian`

to explain potential problems with the new return types specification. (#3997)Fixed a bug where calling

`Evolution.generator`

with`coeff`

being a complex ArrayBox raised an error. (#3796)`MeasurementProcess.hash`

now uses the hash property of the observable. The property now depends on all properties that affect the behaviour of the object, such as`VnEntropyMP.log_base`

or the distribution of wires between the two subsystems in`MutualInfoMP`

. (#3898)The enum

`measurements.Purity`

has been added so that`PurityMP.return_type`

is defined.`str`

and`repr`

for`PurityMP`

are also now defined. (#3898)`Sum.hash`

and`Prod.hash`

have been changed slightly to work with non-numeric wire labels.`sum_expand`

should now return correct results and not treat some products as the same operation. (#3898)Fixed bug where the coefficients where not ordered correctly when summing a

`ParametrizedHamiltonian`

with other operators. (#3749) (#3902)The metric tensor transform is now fully compatible with Jax and therefore users can provide multiple parameters. (#3847)

`qml.math.ndim`

and`qml.math.shape`

are now registered for built-ins and autograd to accomodate Autoray 0.6.1. #3864Ensured that

`qml.data.load`

returns datasets in a stable and expected order. (#3856)The

`qml.equal`

function now handles comparisons of`ParametrizedEvolution`

operators. (#3870)`qml.devices.qubit.apply_operation`

catches the`tf.errors.UnimplementedError`

that occurs when`PauliZ`

or`CNOT`

gates are applied to a large (>8 wires) tensorflow state. When that occurs, the logic falls back to the tensordot logic instead. (#3884)Fixed parameter broadcasting support with

`qml.counts`

in most cases and introduced explicit errors otherwise. (#3876)An error is now raised if a QNode with Jax-jit in use returns

`counts`

while having trainable parameters (#3892)A correction has been added to the reference values in

`test_dipole_of`

to account for small changes (~`2e-8`

) in the computed dipole moment values resulting from the new PySCF 2.2.0 release. (#3908)`SampleMP.shape`

is now correct when sampling only occurs on a subset of the device wires. (#3921)An issue has been fixed in

`qchem.Molecule`

to allow basis sets other than the hard-coded ones to be used in the`Molecule`

class. (#3955)Fixed bug where all devices that inherit from

`DefaultQubit`

claimed to support`ParametrizedEvolution`

. Now, only`DefaultQubitJax`

supports the operator, as expected. (#3964)Ensured that parallel

`AnnotatedQueues`

do not queue each other’s contents. (#3924)Added a

`map_wires`

method to`PauliWord`

and`PauliSentence`

, and ensured that operators call it in their respective`map_wires`

methods if they have a Pauli rep. (#3985)Fixed a bug when a

`Tensor`

is multiplied by a`Hamiltonian`

or vice versa. (#4036)

### Contributors ✍️

This release contains contributions from (in alphabetical order):

Komi Amiko, Utkarsh Azad, Thomas Bromley, Isaac De Vlugt, Olivia Di Matteo, Lillian M. A. Frederiksen, Diego Guala, Soran Jahangiri, Korbinian Kottmann, Christina Lee, Vincent Michaud-Rioux, Albert Mitjans Coma, Romain Moyard, Lee J. O’Riordan, Mudit Pandey, Matthew Silverman, Jay Soni, David Wierichs.

- orphan

## Release 0.29.0¶

### New features since last release

#### Pulse programming 🔊

Support for creating pulse-based circuits that describe evolution under a time-dependent Hamiltonian has now been added, as well as the ability to execute and differentiate these pulse-based circuits on simulator. (#3586) (#3617) (#3645) (#3652) (#3665) (#3673) (#3706) (#3730)

A time-dependent Hamiltonian can be created using

`qml.pulse.ParametrizedHamiltonian`

, which holds information representing a linear combination of operators with parametrized coefficents and can be constructed as follows:from jax import numpy as jnp f1 = lambda p, t: p * jnp.sin(t) * (t - 1) f2 = lambda p, t: p[0] * jnp.cos(p[1]* t ** 2) XX = qml.PauliX(0) @ qml.PauliX(1) YY = qml.PauliY(0) @ qml.PauliY(1) ZZ = qml.PauliZ(0) @ qml.PauliZ(1) H = 2 * ZZ + f1 * XX + f2 * YY

>>> H ParametrizedHamiltonian: terms=3 >>> p1 = jnp.array(1.2) >>> p2 = jnp.array([2.3, 3.4]) >>> H((p1, p2), t=0.5) (2*(PauliZ(wires=[0]) @ PauliZ(wires=[1]))) + ((-0.2876553231625218*(PauliX(wires=[0]) @ PauliX(wires=[1]))) + (1.517961235535459*(PauliY(wires=[0]) @ PauliY(wires=[1]))))

The time-dependent Hamiltonian can be used within a circuit with

`qml.evolve`

:def pulse_circuit(params, time): qml.evolve(H)(params, time) return qml.expval(qml.PauliX(0) @ qml.PauliY(1))

Pulse-based circuits can be executed and differentiated on the

`default.qubit.jax`

simulator using JAX as an interface:>>> dev = qml.device("default.qubit.jax", wires=2) >>> qnode = qml.QNode(pulse_circuit, dev, interface="jax") >>> params = (p1, p2) >>> qnode(params, time=0.5) Array(0.72153819, dtype=float64) >>> jax.grad(qnode)(params, time=0.5) (Array(-0.11324919, dtype=float64), Array([-0.64399616, 0.06326374], dtype=float64))

Check out the qml.pulse documentation page for more details!

#### Special unitary operation 🌞

A new operation

`qml.SpecialUnitary`

has been added, providing access to an arbitrary unitary gate via a parametrization in the Pauli basis. (#3650) (#3651) (#3674)`qml.SpecialUnitary`

creates a unitary that exponentiates a linear combination of all possible Pauli words in lexicographical order — except for the identity operator — for`num_wires`

wires, of which there are`4**num_wires - 1`

. As its first argument,`qml.SpecialUnitary`

takes a list of the`4**num_wires - 1`

parameters that are the coefficients of the linear combination.To see all possible Pauli words for

`num_wires`

wires, you can use the`qml.ops.qubit.special_unitary.pauli_basis_strings`

function:>>> qml.ops.qubit.special_unitary.pauli_basis_strings(1) # 4**1-1 = 3 Pauli words ['X', 'Y', 'Z'] >>> qml.ops.qubit.special_unitary.pauli_basis_strings(2) # 4**2-1 = 15 Pauli words ['IX', 'IY', 'IZ', 'XI', 'XX', 'XY', 'XZ', 'YI', 'YX', 'YY', 'YZ', 'ZI', 'ZX', 'ZY', 'ZZ']

To use

`qml.SpecialUnitary`

, for example, on a single qubit, we may define>>> thetas = np.array([0.2, 0.1, -0.5]) >>> U = qml.SpecialUnitary(thetas, 0) >>> qml.matrix(U) array([[ 0.8537127 -0.47537233j, 0.09507447+0.19014893j], [-0.09507447+0.19014893j, 0.8537127 +0.47537233j]])

A single non-zero entry in the parameters will create a Pauli rotation:

>>> x = 0.412 >>> theta = x * np.array([1, 0, 0]) # The first entry belongs to the Pauli word "X" >>> su = qml.SpecialUnitary(theta, wires=0) >>> rx = qml.RX(-2 * x, 0) # RX introduces a prefactor -0.5 that has to be compensated >>> qml.math.allclose(qml.matrix(su), qml.matrix(rx)) True

This operation can be differentiated with hardware-compatible methods like parameter shifts and it supports parameter broadcasting/batching, but not both at the same time. Learn more by visiting the qml.SpecialUnitary documentation.

#### Always differentiable 📈

The Hadamard test gradient transform is now available via

`qml.gradients.hadamard_grad`

. This transform is also available as a differentiation method within`QNode`

s. (#3625) (#3736)`qml.gradients.hadamard_grad`

is a hardware-compatible transform that calculates the gradient of a quantum circuit using the Hadamard test. Note that the device requires an auxiliary wire to calculate the gradient.>>> dev = qml.device("default.qubit", wires=2) >>> @qml.qnode(dev) ... def circuit(params): ... qml.RX(params[0], wires=0) ... qml.RY(params[1], wires=0) ... qml.RX(params[2], wires=0) ... return qml.expval(qml.PauliZ(0)) >>> params = np.array([0.1, 0.2, 0.3], requires_grad=True) >>> qml.gradients.hadamard_grad(circuit)(params) (tensor(-0.3875172, requires_grad=True), tensor(-0.18884787, requires_grad=True), tensor(-0.38355704, requires_grad=True))

This transform can be registered directly as the quantum gradient transform to use during autodifferentiation:

>>> dev = qml.device("default.qubit", wires=2) >>> @qml.qnode(dev, interface="jax", diff_method="hadamard") ... def circuit(params): ... qml.RX(params[0], wires=0) ... qml.RY(params[1], wires=0) ... qml.RX(params[2], wires=0) ... return qml.expval(qml.PauliZ(0)) >>> params = jax.numpy.array([0.1, 0.2, 0.3]) >>> jax.jacobian(circuit)(params) Array([-0.3875172 , -0.18884787, -0.38355705], dtype=float32)

The gradient transform

`qml.gradients.spsa_grad`

is now registered as a differentiation method for QNodes. (#3440)The SPSA gradient transform can now be used implicitly by marking a QNode as differentiable with SPSA. It can be selected via

>>> dev = qml.device("default.qubit", wires=1) >>> @qml.qnode(dev, interface="jax", diff_method="spsa", h=0.05, num_directions=20) ... def circuit(x): ... qml.RX(x, 0) ... return qml.expval(qml.PauliZ(0)) >>> jax.jacobian(circuit)(jax.numpy.array(0.5)) Array(-0.4792258, dtype=float32, weak_type=True)

The argument

`num_directions`

determines how many directions of simultaneous perturbation are used and therefore the number of circuit evaluations, up to a prefactor. See the SPSA gradient transform documentation for details. Note: The full SPSA optimization method is already available as`qml.SPSAOptimizer`

.The default interface is now

`auto`

. There is no need to specify the interface anymore; it is automatically determined by checking your QNode parameters. (#3677) (#3752) (#3829)import jax import jax.numpy as jnp qml.enable_return() a = jnp.array(0.1) b = jnp.array(0.2) dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(a, b): qml.RY(a, wires=0) qml.RX(b, wires=1) qml.CNOT(wires=[0, 1]) return qml.expval(qml.PauliZ(0)), qml.expval(qml.PauliY(1))

>>> circuit(a, b) (Array(0.9950042, dtype=float32), Array(-0.19767681, dtype=float32)) >>> jac = jax.jacobian(circuit)(a, b) >>> jac (Array(-0.09983341, dtype=float32, weak_type=True), Array(0.01983384, dtype=float32, weak_type=True))

The JAX-JIT interface now supports higher-order gradient computation with the new return types system. (#3498)

Here is an example of using JAX-JIT to compute the Hessian of a circuit:

import pennylane as qml import jax from jax import numpy as jnp jax.config.update("jax_enable_x64", True) qml.enable_return() dev = qml.device("default.qubit", wires=2) @jax.jit @qml.qnode(dev, interface="jax-jit", diff_method="parameter-shift", max_diff=2) def circuit(a, b): qml.RY(a, wires=0) qml.RX(b, wires=1) return qml.expval(qml.PauliZ(0)), qml.expval(qml.PauliZ(1)) a, b = jnp.array(1.0), jnp.array(2.0)

>>> jax.hessian(circuit, argnums=[0, 1])(a, b) (((Array(-0.54030231, dtype=float64, weak_type=True), Array(0., dtype=float64, weak_type=True)), (Array(-1.76002563e-17, dtype=float64, weak_type=True), Array(0., dtype=float64, weak_type=True))), ((Array(0., dtype=float64, weak_type=True), Array(-1.00700085e-17, dtype=float64, weak_type=True)), (Array(0., dtype=float64, weak_type=True), Array(0.41614684, dtype=float64, weak_type=True))))

The

`qchem`

workflow has been modified to support both Autograd and JAX frameworks. (#3458) (#3462) (#3495)The JAX interface is automatically used when the differentiable parameters are JAX objects. Here is an example for computing the Hartree-Fock energy gradients with respect to the atomic coordinates.

import pennylane as qml from pennylane import numpy as np import jax symbols = ["H", "H"] geometry = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 1.0]]) mol = qml.qchem.Molecule(symbols, geometry) args = [jax.numpy.array(mol.coordinates)]

>>> jax.grad(qml.qchem.hf_energy(mol))(*args) Array([[ 0. , 0. , 0.3650435], [ 0. , 0. , -0.3650435]], dtype=float64)

The kernel matrix utility functions in

`qml.kernels`

are now autodifferentiation-compatible. In addition, they support batching, for example for quantum kernel execution with shot vectors. (#3742)This allows for the following:

dev = qml.device('default.qubit', wires=2, shots=(100, 100)) @qml.qnode(dev) def circuit(x1, x2): qml.templates.AngleEmbedding(x1, wires=dev.wires) qml.adjoint(qml.templates.AngleEmbedding)(x2, wires=dev.wires) return qml.probs(wires=dev.wires) kernel = lambda x1, x2: circuit(x1, x2)

We can then compute the kernel matrix on a set of 4 (random) feature vectors

`X`

but using two sets of 100 shots each via>>> X = np.random.random((4, 2)) >>> qml.kernels.square_kernel_matrix(X, kernel)[:, 0] tensor([[[1. , 0.86, 0.88, 0.92], [0.86, 1. , 0.75, 0.97], [0.88, 0.75, 1. , 0.91], [0.92, 0.97, 0.91, 1. ]], [[1. , 0.93, 0.91, 0.92], [0.93, 1. , 0.8 , 1. ], [0.91, 0.8 , 1. , 0.91], [0.92, 1. , 0.91, 1. ]]], requires_grad=True)

Note that we have extracted the first probability vector entry for each 100-shot evaluation.

#### Smartly decompose Hamiltonian evolution 💯

Hamiltonian evolution using

`qml.evolve`

or`qml.exp`

can now be decomposed into operations. (#3691) (#3777)If the time-evolved Hamiltonian is equivalent to another PennyLane operation, then that operation is returned as the decomposition:

>>> exp_op = qml.evolve(qml.PauliX(0) @ qml.PauliX(1)) >>> exp_op.decomposition() [IsingXX((2+0j), wires=[0, 1])]

If the Hamiltonian is a Pauli word, then the decomposition is provided as a

`qml.PauliRot`

operation:>>> qml.evolve(qml.PauliZ(0) @ qml.PauliX(1)).decomposition() [PauliRot((2+0j), ZX, wires=[0, 1])]

Otherwise, the Hamiltonian is a linear combination of operators and the Suzuki-Trotter decomposition is used:

>>> qml.evolve(qml.sum(qml.PauliX(0), qml.PauliY(0), qml.PauliZ(0)), num_steps=2).decomposition() [RX((1+0j), wires=[0]), RY((1+0j), wires=[0]), RZ((1+0j), wires=[0]), RX((1+0j), wires=[0]), RY((1+0j), wires=[0]), RZ((1+0j), wires=[0])]

#### Tools for quantum chemistry and other applications 🛠️

A new method called

`qml.qchem.givens_decomposition`

has been added, which decomposes a unitary into a sequence of Givens rotation gates with phase shifts and a diagonal phase matrix. (#3573)unitary = np.array([[ 0.73678+0.27511j, -0.5095 +0.10704j, -0.06847+0.32515j], [-0.21271+0.34938j, -0.38853+0.36497j, 0.61467-0.41317j], [ 0.41356-0.20765j, -0.00651-0.66689j, 0.32839-0.48293j]]) phase_mat, ordered_rotations = qml.qchem.givens_decomposition(unitary)

>>> phase_mat tensor([-0.20604358+0.9785369j , -0.82993272+0.55786114j, 0.56230612-0.82692833j], requires_grad=True) >>> ordered_rotations [(tensor([[-0.65087861-0.63937521j, -0.40933651-0.j ], [-0.29201359-0.28685265j, 0.91238348-0.j ]], requires_grad=True), (0, 1)), (tensor([[ 0.47970366-0.33308926j, -0.8117487 -0.j ], [ 0.66677093-0.46298215j, 0.5840069 -0.j ]], requires_grad=True), (1, 2)), (tensor([[ 0.36147547+0.73779454j, -0.57008306-0.j ], [ 0.2508207 +0.51194108j, 0.82158706-0.j ]], requires_grad=True), (0, 1))]

A new template called

`qml.BasisRotation`

has been added, which performs a basis transformation defined by a set of fermionic ladder operators. (#3573)import pennylane as qml from pennylane import numpy as np V = np.array([[ 0.53672126+0.j , -0.1126064 -2.41479668j], [-0.1126064 +2.41479668j, 1.48694623+0.j ]]) eigen_vals, eigen_vecs = np.linalg.eigh(V) umat = eigen_vecs.T wires = range(len(umat)) def circuit(): qml.adjoint(qml.BasisRotation(wires=wires, unitary_matrix=umat)) for idx, eigenval in enumerate(eigen_vals): qml.RZ(eigenval, wires=[idx]) qml.BasisRotation(wires=wires, unitary_matrix=umat)

>>> circ_unitary = qml.matrix(circuit)() >>> np.round(circ_unitary/circ_unitary[0][0], 3) tensor([[ 1. -0.j , -0. +0.j , -0. +0.j , -0. +0.j ], [-0. +0.j , -0.516-0.596j, -0.302-0.536j, -0. +0.j ], [-0. +0.j , 0.35 +0.506j, -0.311-0.724j, -0. +0.j ], [-0. +0.j , -0. +0.j , -0. +0.j , -0.438+0.899j]], requires_grad=True)

A new function called

`qml.qchem.load_basisset`

has been added to extract`qml.qchem`

basis set data from the Basis Set Exchange library. (#3363)A new function called

`qml.math.max_entropy`

has been added to compute the maximum entropy of a quantum state. (#3594)A new template called

`qml.TwoLocalSwapNetwork`

has been added that implements a canonical 2-complete linear (2-CCL) swap network described in arXiv:1905.05118. (#3447)dev = qml.device('default.qubit', wires=5) weights = np.random.random(size=qml.templates.TwoLocalSwapNetwork.shape(len(dev.wires))) acquaintances = lambda index, wires, param: (qml.CRY(param, wires=index) if np.abs(wires[0]-wires[1]) else qml.CRZ(param, wires=index)) @qml.qnode(dev) def swap_network_circuit(): qml.templates.TwoLocalSwapNetwork(dev.wires, acquaintances, weights, fermionic=False) return qml.state()

>>> print(weights) tensor([0.20308242, 0.91906199, 0.67988804, 0.81290256, 0.08708985, 0.81860084, 0.34448344, 0.05655892, 0.61781612, 0.51829044], requires_grad=True) >>> print(qml.draw(swap_network_circuit, expansion_strategy = 'device')()) 0: ─╭●────────╭SWAP─────────────────╭●────────╭SWAP─────────────────╭●────────╭SWAP─┤ State 1: ─╰RY(0.20)─╰SWAP─╭●────────╭SWAP─╰RY(0.09)─╰SWAP─╭●────────╭SWAP─╰RY(0.62)─╰SWAP─┤ State 2: ─╭●────────╭SWAP─╰RY(0.68)─╰SWAP─╭●────────╭SWAP─╰RY(0.34)─╰SWAP─╭●────────╭SWAP─┤ State 3: ─╰RY(0.92)─╰SWAP─╭●────────╭SWAP─╰RY(0.82)─╰SWAP─╭●────────╭SWAP─╰RY(0.52)─╰SWAP─┤ State 4: ─────────────────╰RY(0.81)─╰SWAP─────────────────╰RY(0.06)─╰SWAP─────────────────┤ State

### Improvements 🛠

#### Pulse programming

A new function called

`qml.pulse.pwc`

has been added as a convenience function for defining a`qml.pulse.ParametrizedHamiltonian`

. This function can be used to create a callable coefficient by setting the timespan over which the function should be non-zero. The resulting callable can be passed an array of parameters and a time. (#3645)>>> timespan = (2, 4) >>> f = qml.pulse.pwc(timespan) >>> f * qml.PauliX(0) ParametrizedHamiltonian: terms=1

The

`params`

array will be used as bin values evenly distributed over the timespan, and the parameter`t`

will determine which of the bins is returned.>>> f(params=[1.2, 2.3, 3.4, 4.5], t=3.9) DeviceArray(4.5, dtype=float32) >>> f(params=[1.2, 2.3, 3.4, 4.5], t=6) # zero outside the range (2, 4) DeviceArray(0., dtype=float32)

A new function called

`qml.pulse.pwc_from_function`

has been added as a decorator for defining a`qml.pulse.ParametrizedHamiltonian`

. This function can be used to decorate a function and create a piecewise constant approximation of it. (#3645)>>> @qml.pulse.pwc_from_function((2, 4), num_bins=10) ... def f1(p, t): ... return p * t

The resulting function approximates the same of

`p**2 * t`

on the interval`t=(2, 4)`

in 10 bins, and returns zero outside the interval.# t=2 and t=2.1 are within the same bin >>> f1(3, 2), f1(3, 2.1) (DeviceArray(6., dtype=float32), DeviceArray(6., dtype=float32)) # next bin >>> f1(3, 2.2) DeviceArray(6.6666665, dtype=float32) # outside the interval t=(2, 4) >>> f1(3, 5) DeviceArray(0., dtype=float32)

Add

`ParametrizedHamiltonianPytree`

class, which is a pytree jax object representing a parametrized Hamiltonian, where the matrix computation is delayed to improve performance. (#3779)

#### Operations and batching

The function

`qml.dot`

has been updated to compute the dot product between a vector and a list of operators. (#3586)>>> coeffs = np.array([1.1, 2.2]) >>> ops = [qml.PauliX(0), qml.PauliY(0)] >>> qml.dot(coeffs, ops) (1.1*(PauliX(wires=[0]))) + (2.2*(PauliY(wires=[0]))) >>> qml.dot(coeffs, ops, pauli=True) 1.1 * X(0) + 2.2 * Y(0)

`qml.evolve`

returns the evolution of an`Operator`

or a`ParametrizedHamiltonian`

. (#3617) (#3706)`qml.ControlledQubitUnitary`

now inherits from`qml.ops.op_math.ControlledOp`

, which defines`decomposition`

,`expand`

, and`sparse_matrix`

rather than raising an error. (#3450)Parameter broadcasting support has been added for the

`qml.ops.op_math.Controlled`

class if the base operator supports broadcasting. (#3450)The

`qml.generator`

function now checks if the generator is Hermitian, rather than whether it is a subclass of`Observable`

. This allows it to return valid generators from`SymbolicOp`

and`CompositeOp`

classes. (#3485)The

`qml.equal`

function has been extended to compare`Prod`

and`Sum`

operators. (#3516)`qml.purity`

has been added as a measurement process for purity (#3551)In-place inversion has been removed for qutrit operations in preparation for the removal of in-place inversion. (#3566)

The

`qml.utils.sparse_hamiltonian`

function has been moved to thee`qml.Hamiltonian.sparse_matrix`

method. (#3585)The

`qml.pauli.PauliSentence.operation()`

method has been improved to avoid instantiating an`SProd`

operator when the coefficient is equal to 1. (#3595)Batching is now allowed in all

`SymbolicOp`

operators, which include`Exp`

,`Pow`

and`SProd`

. (#3597)The

`Sum`

and`Prod`

operations now have broadcasted operands. (#3611)The XYX single-qubit unitary decomposition has been implemented. (#3628)

All dunder methods now return

`NotImplemented`

, allowing the right dunder method (e.g.`__radd__`

) of the other class to be called. (#3631)The

`qml.GellMann`

operators now include their index when displayed. (#3641)`qml.ops.ctrl_decomp_zyz`

has been added to compute the decomposition of a controlled single-qubit operation given a single-qubit operation and the control wires. (#3681)`qml.pauli.is_pauli_word`

now supports`Prod`

and`SProd`

operators, and it returns`False`

when a`Hamiltonian`

contains more than one term. (#3692)`qml.pauli.pauli_word_to_string`

now supports`Prod`

,`SProd`

and`Hamiltonian`

operators. (#3692)`qml.ops.op_math.Controlled`

can now decompose single qubit target operations more effectively using the ZYZ decomposition. (#3726)The

`qml.qchem.Molecule`

class raises an error when the molecule has an odd number of electrons or when the spin multiplicity is not 1. (#3748)

`qml.qchem.basis_rotation`

now accounts for spin, allowing it to perform Basis Rotation Groupings for molecular hamiltonians. (#3714) (#3774)The gradient transforms work for the new return type system with non-trivial classical jacobians. (#3776)

The

`default.mixed`

device has received a performance improvement for multi-qubit operations. This also allows to apply channels that act on more than seven qubits, which was not possible before. (#3584)`qml.dot`

now groups coefficients together. (#3691)>>> qml.dot(coeffs=[2, 2, 2], ops=[qml.PauliX(0), qml.PauliY(1), qml.PauliZ(2)]) 2*(PauliX(wires=[0]) + PauliY(wires=[1]) + PauliZ(wires=[2]))

`qml.generator`

now supports operators with`Sum`

and`Prod`

generators. (#3691)The

`Sum._sort`

method now takes into account the name of the operator when sorting. (#3691)A new tape transform called

`qml.transforms.sign_expand`

has been added. It implements the optimal decomposition of a fast-forwardable Hamiltonian that minimizes the variance of its estimator in the Single-Qubit-Measurement from arXiv:2207.09479. (#2852)

#### Differentiability and interfaces

The

`qml.math`

module now also contains a submodule for fast Fourier transforms,`qml.math.fft`

. (#1440)The submodule in particular provides differentiable versions of the following functions, available in all common interfaces for PennyLane

Note that the output of the derivative of these functions may differ when used with complex-valued inputs, due to different conventions on complex-valued derivatives.

Validation has been added on gradient keyword arguments when initializing a QNode — if unexpected keyword arguments are passed, a

`UserWarning`

is raised. A list of the current expected gradient function keyword arguments can be accessed via`qml.gradients.SUPPORTED_GRADIENT_KWARGS`

. (#3526)The

`numpy`

version has been constrained to`<1.24`

. (#3563)Support for two-qubit unitary decomposition with JAX-JIT has been added. (#3569)

`qml.math.size`

now supports PyTorch tensors. (#3606)Most quantum channels are now fully differentiable on all interfaces. (#3612)

`qml.math.matmul`

now supports PyTorch and Autograd tensors. (#3613)Add

`qml.math.detach`

, which detaches a tensor from its trace. This stops automatic gradient computations. (#3674)Add

`typing.TensorLike`

type. (#3675)`qml.QuantumMonteCarlo`

template is now JAX-JIT compatible when passing`jax.numpy`

arrays to the template. (#3734)`DefaultQubitJax`

now supports evolving the state vector when executing`qml.pulse.ParametrizedEvolution`

gates. (#3743)`SProd.sparse_matrix`

now supports interface-specific variables with a single element as the`scalar`

. (#3770)Added

`argnum`

argument to`metric_tensor`

. By passing a sequence of indices referring to trainable tape parameters, the metric tensor is only computed with respect to these parameters. This reduces the number of tapes that have to be run. (#3587)The parameter-shift derivative of variances saves a redundant evaluation of the corresponding unshifted expectation value tape, if possible (#3744)

#### Next generation device API

The

`apply_operation`

single-dispatch function is added to`devices/qubit`

that applies an operation to a state and returns a new state. (#3637)The

`preprocess`

function is added to`devices/qubit`

that validates, expands, and transforms a batch of`QuantumTape`

objects to abstract preprocessing details away from the device. (#3708)The

`create_initial_state`

function is added to`devices/qubit`

that returns an initial state for an execution. (#3683)The

`simulate`

function is added to`devices/qubit`

that turns a single quantum tape into a measurement result. The function only supports state based measurements with either no observables or observables with diagonalizing gates. It supports simultaneous measurement of non-commuting observables. (#3700)The

`ExecutionConfig`

data class has been added. (#3649)The

`StatePrep`

class has been added as an interface that state-prep operators must implement. (#3654)`qml.QubitStateVector`

now implements the`StatePrep`

interface. (#3685)`qml.BasisState`

now implements the`StatePrep`

interface. (#3693)New Abstract Base Class for devices

`Device`

is added to the`devices.experimental`

submodule. This interface is still in experimental mode and not integrated with the rest of pennylane. (#3602)

#### Other improvements

Writing Hamiltonians to a file using the

`qml.data`

module has been improved by employing a condensed writing format. (#3592)Lazy-loading in the

`qml.data.Dataset.read()`

method is more universally supported. (#3605)The

`qchem.Molecule`

class raises an error when the molecule has an odd number of electrons or when the spin multiplicity is not 1. (#3748)`qml.draw`

and`qml.draw_mpl`

have been updated to draw any quantum function, which allows for visualizing only part of a complete circuit/QNode. (#3760)The string representation of a Measurement Process now includes the

`_eigvals`

property if it is set. (#3820)

### Breaking changes 💔

The argument

`mode`

in execution has been replaced by the boolean`grad_on_execution`

in the new execution pipeline. (#3723)`qml.VQECost`

has been removed. (#3735)The default interface is now

`auto`

. (#3677) (#3752) (#3829)The interface is determined during the QNode call instead of the initialization. It means that the

`gradient_fn`

and`gradient_kwargs`

are only defined on the QNode at the beginning of the call. Moreover, without specifying the interface it is not possible to guarantee that the device will not be changed during the call if you are using backprop (such as`default.qubit`

changing to`default.qubit.jax`

) whereas before it was happening at initialization.The tape method

`get_operation`

can also now return the operation index in the tape, and it can be activated by setting the`return_op_index`

to`True`

:`get_operation(idx, return_op_index=True)`

. It will become the default in version`0.30`

. (#3667)`Operation.inv()`

and the`Operation.inverse`

setter have been removed. Please use`qml.adjoint`

or`qml.pow`

instead. (#3618)For example, instead of

>>> qml.PauliX(0).inv()

use

>>> qml.adjoint(qml.PauliX(0))

The

`Operation.inverse`

property has been removed completely. (#3725)The target wires of

`qml.ControlledQubitUnitary`

are no longer available via`op.hyperparameters["u_wires"]`

. Instead, they can be accesses via`op.base.wires`

or`op.target_wires`

. (#3450)The tape constructed by a

`QNode`

is no longer queued to surrounding contexts. (#3509)Nested operators like

`Tensor`

,`Hamiltonian`

, and`Adjoint`

now remove their owned operators from the queue instead of updating their metadata to have an`"owner"`

. (#3282)`qml.qchem.scf`

,`qml.RandomLayers.compute_decomposition`

, and`qml.Wires.select_random`

now use local random number generators instead of global random number generators. This may lead to slightly different random numbers and an independence of the results from the global random number generation state. Please provide a seed to each individual function instead if you want controllable results. (#3624)`qml.transforms.measurement_grouping`

has been removed. Users should use`qml.transforms.hamiltonian_expand`

instead. (#3701)`op.simplify()`

for operators which are linear combinations of Pauli words will use a builtin Pauli representation to more efficiently compute the simplification of the operator. (#3481)All

`Operator`

‘s input parameters that are lists are cast into vanilla numpy arrays. (#3659)`QubitDevice.expval`

no longer permutes an observable’s wire order before passing it to`QubitDevice.probability`

. The associated downstream changes for`default.qubit`

have been made, but this may still affect expectations for other devices that inherit from`QubitDevice`

and override`probability`

(or any other helper functions that take a wire order such as`marginal_prob`

,`estimate_probability`

or`analytic_probability`

). (#3753)

### Deprecations 👋

`qml.utils.sparse_hamiltonian`

function has been deprecated, and usage will now raise a warning. Instead, one should use the`qml.Hamiltonian.sparse_matrix`

method. (#3585)`qml.op_sum`

has been deprecated. Users should use`qml.sum`

instead. (#3686)The use of

`Evolution`

directly has been deprecated. Users should use`qml.evolve`

instead. This new function changes the sign of the given parameter. (#3706)Use of

`qml.dot`

with a`QNodeCollection`

has been deprecated. (#3586)

### Documentation 📝

Revise note on GPU support in the circuit introduction. (#3836)

Make warning about vanilla version of NumPy for differentiation more prominent. (#3838)

The documentation for

`qml.operation`

has been improved. (#3664)The code example in

`qml.SparseHamiltonian`

has been updated with the correct wire range. (#3643)A hyperlink has been added in the text for a URL in the

`qml.qchem.mol_data`

docstring. (#3644)A typo was corrected in the documentation for

`qml.math.vn_entropy`

. (#3740)

### Bug fixes 🐛

Fixed a bug where measuring

`qml.probs`

in the computational basis with non-commuting measurements returned incorrect results. Now an error is raised. (#3811)Fixed a bug where measuring

`qml.probs`

in the computational basis with non-commuting measurements returned incorrect results. Now an error is raised. (#3811)Fixed a bug in the drawer where nested controlled operations would output the label of the operation being controlled, rather than the control values. (#3745)

Fixed a bug in

`qml.transforms.metric_tensor`

where prefactors of operation generators were taken into account multiple times, leading to wrong outputs for non-standard operations. (#3579)Local random number generators are now used where possible to avoid mutating the global random state. (#3624)

The

`networkx`

version change being broken has been fixed by selectively skipping a`qcut`

TensorFlow-JIT test. (#3609) (#3619)Fixed the wires for the

`Y`

decomposition in the ZX calculus transform. (#3598)`qml.pauli.PauliWord`

is now pickle-able. (#3588)Child classes of

`QuantumScript`

now return their own type when using`SomeChildClass.from_queue`

. (#3501)A typo has been fixed in the calculation and error messages in

`operation.py`

(#3536)`qml.data.Dataset.write()`

now ensures that any lazy-loaded values are loaded before they are written to a file. (#3605)`Tensor._batch_size`

is now set to`None`

during initialization, copying and`map_wires`

. (#3642) (#3661)`Tensor.has_matrix`

is now set to`True`

. (#3647)Fixed typo in the example of

`qml.IsingZZ`

gate decomposition. (#3676)Fixed a bug that made tapes/qnodes using

`qml.Snapshot`

incompatible with`qml.drawer.tape_mpl`

. (#3704)`Tensor._pauli_rep`

is set to`None`

during initialization and`Tensor.data`

has been added to its setter. (#3722)`qml.math.ndim`

has been redirected to`jnp.ndim`

when using it on a`jax`

tensor. (#3730)Implementations of

`marginal_prob`

(and subsequently,`qml.probs`

) now return probabilities with the expected wire order. (#3753)This bug affected most probabilistic measurement processes on devices that inherit from

`QubitDevice`

when the measured wires are out of order with respect to the device wires and 3 or more wires are measured. The assumption was that marginal probabilities would be computed with the device’s state and wire order, then re-ordered according to the measurement process wire order. Instead, the re-ordering went in the inverse direction (that is, from measurement process wire order to device wire order). This is now fixed. Note that this only occurred for 3 or more measured wires because this mapping is identical otherwise. More details and discussion of this bug can be found in the original bug report.Empty iterables can no longer be returned from QNodes. (#3769)

The keyword arguments for

`qml.equal`

now are used when comparing the observables of a Measurement Process. The eigvals of measurements are only requested if both observables are`None`

, saving computational effort. (#3820)Only converts input to

`qml.Hermitian`

to a numpy array if the input is a list. (#3820)

### Contributors ✍

This release contains contributions from (in alphabetical order):

Gian-Luca Anselmetti, Guillermo Alonso-Linaje, Juan Miguel Arrazola, Ikko Ashimine, Utkarsh Azad, Miriam Beddig, Cristian Boghiu, Thomas Bromley, Astral Cai, Isaac De Vlugt, Olivia Di Matteo, Lillian M. A. Frederiksen, Soran Jahangiri, Korbinian Kottmann, Christina Lee, Albert Mitjans Coma, Romain Moyard, Mudit Pandey, Borja Requena, Matthew Silverman, Jay Soni, Antal Száva, Frederik Wilde, David Wierichs, Moritz Willmann.

- orphan

## Release 0.28.0¶

### New features since last release

#### Custom measurement processes 📐

Custom measurements can now be facilitated with the addition of the

`qml.measurements`

module. (#3286) (#3343) (#3288) (#3312) (#3287) (#3292) (#3287) (#3326) (#3327) (#3388) (#3439) (#3466)Within

`qml.measurements`

are new subclasses that allow for the possibility to create*custom*measurements:`SampleMeasurement`

: represents a sample-based measurement`StateMeasurement`

: represents a state-based measurement`MeasurementTransform`

: represents a measurement process that requires the application of a batch transform

Creating a custom measurement involves making a class that inherits from one of the classes above. An example is given below. Here, the measurement computes the number of samples obtained of a given state:

from pennylane.measurements import SampleMeasurement class CountState(SampleMeasurement): def __init__(self, state: str): self.state = state # string identifying the state, e.g. "0101" wires = list(range(len(state))) super().__init__(wires=wires) def process_samples(self, samples, wire_order, shot_range, bin_size): counts_mp = qml.counts(wires=self._wires) counts = counts_mp.process_samples(samples, wire_order, shot_range, bin_size) return counts.get(self.state, 0) def __copy__(self): return CountState(state=self.state)

We can now execute the new measurement in a QNode as follows.

dev = qml.device("default.qubit", wires=1, shots=10000) @qml.qnode(dev) def circuit(x): qml.RX(x, wires=0) return CountState(state="1")

>>> circuit(1.23) tensor(3303., requires_grad=True)

Differentiability is also supported for this new measurement process:

>>> x = qml.numpy.array(1.23, requires_grad=True) >>> qml.grad(circuit)(x) 4715.000000000001

For more information about these new features, see the documentation for ``qml.measurements` <https://docs.pennylane.ai/en/stable/code/qml_measurements.html>`_.

#### ZX Calculus 🧮

QNodes can now be converted into ZX diagrams via the PyZX framework. (#3446)

ZX diagrams are the medium for which we can envision a quantum circuit as a graph in the ZX-calculus language, showing properties of quantum protocols in a visually compact and logically complete fashion.

QNodes decorated with

`@qml.transforms.to_zx`

will return a PyZX graph that represents the computation in the ZX-calculus language.dev = qml.device("default.qubit", wires=2) @qml.transforms.to_zx @qml.qnode(device=dev) def circuit(p): qml.RZ(p[0], wires=1), qml.RZ(p[1], wires=1), qml.RX(p[2], wires=0), qml.PauliZ(wires=0), qml.RZ(p[3], wires=1), qml.PauliX(wires=1), qml.CNOT(wires=[0, 1]), qml.CNOT(wires=[1, 0]), qml.SWAP(wires=[0, 1]), return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))

>>> params = [5 / 4 * np.pi, 3 / 4 * np.pi, 0.1, 0.3] >>> circuit(params) Graph(20 vertices, 23 edges)

Information about PyZX graphs can be found in the PyZX Graphs API.

#### QChem databases and basis sets ⚛️

The symbols and geometry of a compound from the PubChem database can now be accessed via

`qchem.mol_data()`

. (#3289) (#3378)>>> import pennylane as qml >>> from pennylane.qchem import mol_data >>> mol_data("BeH2") (['Be', 'H', 'H'], tensor([[ 4.79404621, 0.29290755, 0. ], [ 3.77945225, -0.29290755, 0. ], [ 5.80882913, -0.29290755, 0. ]], requires_grad=True)) >>> mol_data(223, "CID") (['N', 'H', 'H', 'H', 'H'], tensor([[ 0. , 0. , 0. ], [ 1.82264085, 0.52836742, 0.40402345], [ 0.01417295, -1.67429735, -0.98038991], [-0.98927163, -0.22714508, 1.65369933], [-0.84773114, 1.373075 , -1.07733286]], requires_grad=True))

Perform quantum chemistry calculations with two new basis sets:

`6-311g`

and`CC-PVDZ`

. (#3279)>>> symbols = ["H", "He"] >>> geometry = np.array([[1.0, 0.0, 0.0], [0.0, 0.0, 0.0]], requires_grad=False) >>> charge = 1 >>> basis_names = ["6-311G", "CC-PVDZ"] >>> for basis_name in basis_names: ... mol = qml.qchem.Molecule(symbols, geometry, charge=charge, basis_name=basis_name) ... print(qml.qchem.hf_energy(mol)()) [-2.84429531] [-2.84061284]

#### A bunch of new operators 👀

The controlled CZ gate and controlled Hadamard gate are now available via

`qml.CCZ`

and`qml.CH`

, respectively. (#3408)>>> ccz = qml.CCZ(wires=[0, 1, 2]) >>> qml.matrix(ccz) [[ 1 0 0 0 0 0 0 0] [ 0 1 0 0 0 0 0 0] [ 0 0 1 0 0 0 0 0] [ 0 0 0 1 0 0 0 0] [ 0 0 0 0 1 0 0 0] [ 0 0 0 0 0 1 0 0] [ 0 0 0 0 0 0 1 0] [ 0 0 0 0 0 0 0 -1]] >>> ch = qml.CH(wires=[0, 1]) >>> qml.matrix(ch) [[ 1. 0. 0. 0. ] [ 0. 1. 0. 0. ] [ 0. 0. 0.70710678 0.70710678] [ 0. 0. 0.70710678 -0.70710678]]

Three new parametric operators,

`qml.CPhaseShift00`

,`qml.CPhaseShift01`

, and`qml.CPhaseShift10`

, are now available. Each of these operators performs a phase shift akin to`qml.ControlledPhaseShift`

but on different positions of the state vector. (#2715)>>> dev = qml.device("default.qubit", wires=2) >>> @qml.qnode(dev) >>> def circuit(): ... qml.PauliX(wires=1) ... qml.CPhaseShift01(phi=1.23, wires=[0,1]) ... return qml.state() ... >>> circuit() tensor([0. +0.j , 0.33423773+0.9424888j, 1. +0.j , 0. +0.j ], requires_grad=True)

A new gate operation called

`qml.FermionicSWAP`

has been added. This implements the exchange of spin orbitals representing fermionic-modes while maintaining proper anti-symmetrization. (#3380)dev = qml.device('default.qubit', wires=2) @qml.qnode(dev) def circuit(phi): qml.BasisState(np.array([0, 1]), wires=[0, 1]) qml.FermionicSWAP(phi, wires=[0, 1]) return qml.state()

>>> circuit(0.1) tensor([0. +0.j , 0.99750208+0.04991671j, 0.00249792-0.04991671j, 0. +0.j ], requires_grad=True)

Create operators defined from a generator via

`qml.ops.op_math.Evolution`

. (#3375)`qml.ops.op_math.Evolution`

defines the exponential of an operator $hat{O}$ of the form $e^{ixhat{O}}$, with a single trainable parameter, $x$. Limiting to a single trainable parameter allows the use of`qml.gradients.param_shift`

to find the gradient with respect to the parameter $x$.dev = qml.device('default.qubit', wires=2) @qml.qnode(dev, diff_method=qml.gradients.param_shift) def circuit(phi): qml.ops.op_math.Evolution(qml.PauliX(0), -.5 * phi) return qml.expval(qml.PauliZ(0))

>>> phi = np.array(1.2) >>> circuit(phi) tensor(0.36235775, requires_grad=True) >>> qml.grad(circuit)(phi) -0.9320390495504149

The qutrit Hadamard gate,

`qml.THadamard`

, is now available. (#3340)The operation accepts a

`subspace`

keyword argument which determines which variant of the qutrit Hadamard to use.>>> th = qml.THadamard(wires=0, subspace=[0, 1]) >>> qml.matrix(th) array([[ 0.70710678+0.j, 0.70710678+0.j, 0. +0.j], [ 0.70710678+0.j, -0.70710678+0.j, 0. +0.j], [ 0. +0.j, 0. +0.j, 1. +0.j]])

#### New transforms, functions, and more 😯

Calculating the purity of arbitrary quantum states is now supported. (#3290)

The purity can be calculated in an analogous fashion to, say, the Von Neumann entropy:

`qml.math.purity`

can be used as an in-line function:>>> x = [1, 0, 0, 1] / np.sqrt(2) >>> qml.math.purity(x, [0, 1]) 1.0 >>> qml.math.purity(x, [0]) 0.5 >>> x = [[1 / 2, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 1 / 2]] >>> qml.math.purity(x, [0, 1]) 0.5

`qml.qinfo.transforms.purity`

can transform a QNode returning a state to a function that returns the purity:dev = qml.device("default.mixed", wires=2) @qml.qnode(dev) def circuit(x): qml.IsingXX(x, wires=[0, 1]) return qml.state()

>>> qml.qinfo.transforms.purity(circuit, wires=[0])(np.pi / 2) 0.5 >>> qml.qinfo.transforms.purity(circuit, wires=[0, 1])(np.pi / 2) 1.0

As with the other methods in

`qml.qinfo`

, the purity is fully differentiable:>>> param = np.array(np.pi / 4, requires_grad=True) >>> qml.grad(qml.qinfo.transforms.purity(circuit, wires=[0]))(param) -0.5

A new gradient transform,

`qml.gradients.spsa_grad`

, that is based on the idea of SPSA is now available. (#3366)This new transform allows users to compute a single estimate of a quantum gradient using simultaneous perturbation of parameters and a stochastic approximation. A QNode that takes, say, an argument

`x`

, the approximate gradient can be computed as follows.>>> dev = qml.device("default.qubit", wires=2) >>> x = np.array(0.4, requires_grad=True) >>> @qml.qnode(dev) ... def circuit(x): ... qml.RX(x, 0) ... qml.RX(x, 1) ... return qml.expval(qml.PauliZ(0)) >>> grad_fn = qml.gradients.spsa_grad(circuit, h=0.1, num_directions=1) >>> grad_fn(x) array(-0.38876964)

The argument

`num_directions`

determines how many directions of simultaneous perturbation are used, which is proportional to the number of circuit evaluations. See the SPSA gradient transform documentation for details. Note that the full SPSA optimizer is already available as`qml.SPSAOptimizer`

.Multiple mid-circuit measurements can now be combined arithmetically to create new conditionals. (#3159)

dev = qml.device("default.qubit", wires=3) @qml.qnode(dev) def circuit(): qml.Hadamard(wires=0) qml.Hadamard(wires=1) m0 = qml.measure(wires=0) m1 = qml.measure(wires=1) combined = 2 * m1 + m0 qml.cond(combined == 2, qml.RX)(1.3, wires=2) return qml.probs(wires=2)

>>> circuit() [0.90843735 0.09156265]

A new method called

`pauli_decompose()`

has been added to the`qml.pauli`

module, which takes a hermitian matrix, decomposes it in the Pauli basis, and returns it either as a`qml.Hamiltonian`

or`qml.PauliSentence`

instance. (#3384)`Operation`

or`Hamiltonian`

instances can now be generated from a`qml.PauliSentence`

or`qml.PauliWord`

via the new`operation()`

and`hamiltonian()`

methods. (#3391)A

`sum_expand`

function has been added for tapes, which splits a tape measuring a`Sum`

expectation into mutliple tapes of summand expectations, and provides a function to recombine the results. (#3230)

#### (Experimental) More interface support for multi-measurement and gradient output types 🧪

The autograd and Tensorflow interfaces now support devices with shot vectors when

`qml.enable_return()`

has been called. (#3374) (#3400)Here is an example using Tensorflow:

import tensorflow as tf qml.enable_return() dev = qml.device("default.qubit", wires=2, shots=[1000, 2000, 3000]) @qml.qnode(dev, diff_method="parameter-shift", interface="tf") def circuit(a): qml.RY(a, wires=0) qml.RX(0.2, wires=0) qml.CNOT(wires=[0, 1]) return qml.expval(qml.PauliZ(0)), qml.probs([0, 1])

>>> a = tf.Variable(0.4) >>> with tf.GradientTape() as tape: ... res = circuit(a) ... res = tf.stack([tf.experimental.numpy.hstack(r) for r in res]) ... >>> res <tf.Tensor: shape=(3, 5), dtype=float64, numpy= array([[0.902, 0.951, 0. , 0. , 0.049], [0.898, 0.949, 0. , 0. , 0.051], [0.892, 0.946, 0. , 0. , 0.054]])> >>> tape.jacobian(res, a) <tf.Tensor: shape=(3, 5), dtype=float64, numpy= array([[-0.345 , -0.1725 , 0. , 0. , 0.1725 ], [-0.383 , -0.1915 , 0. , 0. , 0.1915 ], [-0.38466667, -0.19233333, 0. , 0. , 0.19233333]])>

The PyTorch interface is now fully supported when

`qml.enable_return()`

has been called, allowing the calculation of the Jacobian and the Hessian using custom differentiation methods (e.g., parameter-shift, finite difference, or adjoint). (#3416)import torch qml.enable_return() dev = qml.device("default.qubit", wires=2) @qml.qnode(dev, diff_method="parameter-shift", interface="torch") def circuit(a, b): qml.RY(a, wires=0) qml.RX(b, wires=1) qml.CNOT(wires=[0, 1]) return qml.expval(qml.PauliZ(0)), qml.probs([0, 1])

>>> a = torch.tensor(0.1, requires_grad=True) >>> b = torch.tensor(0.2, requires_grad=True) >>> torch.autograd.functional.jacobian(circuit, (a, b)) ((tensor(-0.0998), tensor(0.)), (tensor([-0.0494, -0.0005, 0.0005, 0.0494]), tensor([-0.0991, 0.0991, 0.0002, -0.0002])))

The JAX-JIT interface now supports first-order gradient computation when

`qml.enable_return()`

has been called. (#3235) (#3445)import jax from jax import numpy as jnp jax.config.update("jax_enable_x64", True) qml.enable_return() dev = qml.device("lightning.qubit", wires=2) @jax.jit @qml.qnode(dev, interface="jax-jit", diff_method="parameter-shift") def circuit(a, b): qml.RY(a, wires=0) qml.RX(b, wires=0) return qml.expval(qml.PauliZ(0)), qml.expval(qml.PauliZ(1)) a, b = jnp.array(1.0), jnp.array(2.0)

>>> jax.jacobian(circuit, argnums=[0, 1])(a, b) ((Array(0.35017549, dtype=float64, weak_type=True), Array(-0.4912955, dtype=float64, weak_type=True)), (Array(5.55111512e-17, dtype=float64, weak_type=True), Array(0., dtype=float64, weak_type=True)))

### Improvements 🛠

`qml.pauli.is_pauli_word`

now supports instances of`qml.Hamiltonian`

. (#3389)When

`qml.probs`

,`qml.counts`

, and`qml.sample`

are called with no arguments, they measure all wires. Calling any of the aforementioned measurements with an empty wire list (e.g.,`qml.sample(wires=[])`

) will raise an error. (#3299)Made

`qml.gradients.finite_diff`

more convenient to use with custom data type observables/devices by reducing the number of magic methods that need to be defined in the custom data type to support`finite_diff`

. (#3426)The

`qml.ISWAP`

gate is now natively supported on`default.mixed`

, improving on its efficiency. (#3284)Added more input validation to

`qml.transforms.hamiltonian_expand`

such that Hamiltonian objects with no terms raise an error. (#3339)Continuous integration checks are now performed for Python 3.11 and Torch v1.13. Python 3.7 is dropped. (#3276)

`qml.Tracker`

now also logs results in`tracker.history`

when tracking the execution of a circuit.

The execution time of

`Wires.all_wires`

has been improved by avoiding data type changes and making use of`itertools.chain`

. (#3302)Printing an instance of

`qml.qchem.Molecule`

is now more concise and informational. (#3364)The error message for

`qml.transforms.insert`

when it fails to diagonalize non-qubit-wise-commuting observables is now more detailed. (#3381)Extended the

`qml.equal`

function to`qml.Hamiltonian`

and`Tensor`

objects. (#3390)`QuantumTape._process_queue`

has been moved to`qml.queuing.process_queue`

to disentangle its functionality from the`QuantumTape`

class. (#3401)QPE can now accept a target operator instead of a matrix and target wires pair. (#3373)

The

`qml.ops.op_math.Controlled.map_wires`

method now uses`base.map_wires`

internally instead of the private`_wires`

property setter. (#3405)A new function called

`qml.tape.make_qscript`

has been created for converting a quantum function into a quantum script. This replaces`qml.transforms.make_tape`

. (#3429)Add a

`_pauli_rep`

attribute to operators to integrate the new Pauli arithmetic classes with native PennyLane objects. (#3443)Extended the functionality of

`qml.matrix`

to qutrits. (#3508)The

`qcut.py`

file in`pennylane/transforms/`

has been reorganized into multiple files that are now in`pennylane/transforms/qcut/`

. (#3413)A warning now appears when creating a

`Tensor`

object with overlapping wires, informing that this can lead to undefined behaviour. (#3459)Extended the

`qml.equal`

function to`qml.ops.op_math.Controlled`

and`qml.ops.op_math.ControlledOp`

objects. (#3463)Nearly every instance of

`with QuantumTape()`

has been replaced with`QuantumScript`

construction. (#3454)Added

`validate_subspace`

static method to`qml.Operator`

to check the validity of the subspace of certain qutrit operations. (#3340)`qml.equal`

now supports operators created via`qml.s_prod`

,`qml.pow`

,`qml.exp`

, and`qml.adjoint`

. (#3471)Devices can now disregard observable grouping indices in Hamiltonians through the optional

`use_grouping`

attribute. (#3456)Add the optional argument

`lazy=True`

to functions`qml.s_prod`

,`qml.prod`

and`qml.op_sum`

to allow simplification. (#3483)Updated the

`qml.transforms.zyz_decomposition`

function such that it now supports broadcast operators. This means that single-qubit`qml.QubitUnitary`

operators, instantiated from a batch of unitaries, can now be decomposed. (#3477)The performance of executing circuits under the

`jax.vmap`

transformation has been improved by being able to leverage the batch-execution capabilities of some devices. (#3452)The tolerance for converting openfermion Hamiltonian complex coefficients to real ones has been modified to prevent conversion errors. (#3367)

`OperationRecorder`

now inherits from`AnnotatedQueue`

and`QuantumScript`

instead of`QuantumTape`

. (#3496)Updated

`qml.transforms.split_non_commuting`

to support the new return types. (#3414)Updated

`qml.transforms.mitigate_with_zne`

to support the new return types. (#3415)Updated

`qml.transforms.metric_tensor`

,`qml.transforms.adjoint_metric_tensor`

,`qml.qinfo.classical_fisher`

, and`qml.qinfo.quantum_fisher`

to support the new return types. (#3449)Updated

`qml.transforms.batch_params`

and`qml.transforms.batch_input`

to support the new return types. (#3431)Updated

`qml.transforms.cut_circuit`

and`qml.transforms.cut_circuit_mc`

to support the new return types. (#3346)Limit NumPy version to

`<1.24`

. (#3346)

### Breaking changes 💔

Python 3.7 support is no longer maintained. PennyLane will be maintained for versions 3.8 and up. (#3276)

The

`log_base`

attribute has been moved from`MeasurementProcess`

to the new`VnEntropyMP`

and`MutualInfoMP`

classes, which inherit from`MeasurementProcess`

. (#3326)`qml.utils.decompose_hamiltonian()`

has been removed. Please use`qml.pauli.pauli_decompose()`

instead. (#3384)The

`return_type`

attribute of`MeasurementProcess`

has been removed where possible. Use`isinstance`

checks instead. (#3399)Instead of having an

`OrderedDict`

attribute called`_queue`

,`AnnotatedQueue`

now inherits from`OrderedDict`

and encapsulates the queue. Consequentially, this also applies to the`QuantumTape`

class which inherits from`AnnotatedQueue`

. (#3401)The

`ShadowMeasurementProcess`

class has been renamed to`ClassicalShadowMP`

. (#3388)The

`qml.Operation.get_parameter_shift`

method has been removed. The`gradients`

module should be used for general parameter-shift rules instead. (#3419)The signature of the

`QubitDevice.statistics`

method has been changed fromdef statistics(self, observables, shot_range=None, bin_size=None, circuit=None):

to

def statistics(self, circuit: QuantumTape, shot_range=None, bin_size=None):

The

`MeasurementProcess`

class is now an abstract class and`return_type`

is now a property of the class. (#3434)

### Deprecations 👋

Deprecations cycles are tracked at doc/developement/deprecations.rst.

The following methods are deprecated: (#3281)

`qml.tape.get_active_tape`

: Use`qml.QueuingManager.active_context()`

instead`qml.transforms.qcut.remap_tape_wires`

: Use`qml.map_wires`

instead`qml.tape.QuantumTape.inv()`

: Use`qml.tape.QuantumTape.adjoint()`

instead`qml.tape.stop_recording()`

: Use`qml.QueuingManager.stop_recording()`

instead`qml.tape.QuantumTape.stop_recording()`

: Use`qml.QueuingManager.stop_recording()`

instead`qml.QueuingContext`

is now`qml.QueuingManager`

`QueuingManager.safe_update_info`

and`AnnotatedQueue.safe_update_info`

: Use`update_info`

instead.

`qml.transforms.measurement_grouping`

has been deprecated. Use`qml.transforms.hamiltonian_expand`

instead. (#3417)The

`observables`

argument in`QubitDevice.statistics`

is deprecated. Please use`circuit`

instead. (#3433)The

`seed_recipes`

argument in`qml.classical_shadow`

and`qml.shadow_expval`

is deprecated. A new argument`seed`

has been added, which defaults to None and can contain an integer with the wanted seed. (#3388)`qml.transforms.make_tape`

has been deprecated. Please use`qml.tape.make_qscript`

instead. (#3478)

### Documentation 📝

Added documentation on parameter broadcasting regarding both its usage and technical aspects. (#3356)

The quickstart guide on circuits as well as the the documentation of QNodes and Operators now contain introductions and details on parameter broadcasting. The QNode documentation mostly contains usage details, the Operator documentation is concerned with implementation details and a guide to support broadcasting in custom operators.

The return type statements of gradient and Hessian transforms and a series of other functions that are a

`batch_transform`

have been corrected. (#3476)Developer documentation for the queuing module has been added. (#3268)

More mentions of diagonalizing gates for all relevant operations have been corrected. (#3409)

The docstrings for

`compute_eigvals`

used to say that the diagonalizing gates implemented $U$, the unitary such that $O = U Sigma U^{dagger}$, where $O$ is the original observable and $Sigma$ a diagonal matrix. However, the diagonalizing gates actually implement $U^{dagger}$, since $langle psi | O | psi rangle = langle psi | U Sigma U^{dagger} | psi rangle$, making $U^{dagger} | psi rangle$ the actual state being measured in the $Z$-basis.A warning about using

`dill`

to pickle and unpickle datasets has been added. (#3505)

### Bug fixes 🐛

Fixed a bug that prevented

`qml.gradients.param_shift`

from being used for broadcasted tapes. (#3528)Fixed a bug where

`qml.transforms.hamiltonian_expand`

didn’t preserve the type of the input results in its output. (#3339)Fixed a bug that made

`qml.gradients.param_shift`

raise an error when used with unshifted terms only in a custom recipe, and when using any unshifted terms at all under the new return type system. (#3177)The original tape

`_obs_sharing_wires`

attribute is updated during its expansion. (#3293)An issue with

`drain=False`

in the adaptive optimizer has been fixed. Before the fix, the operator pool needed to be reconstructed inside the optimization pool when`drain=False`

. With this fix, this reconstruction is no longer needed. (#3361)If the device originally has no shots but finite shots are dynamically specified, Hamiltonian expansion now occurs. (#3369)

`qml.matrix(op)`

now fails if the operator truly has no matrix (e.g.,`qml.Barrier`

) to match`op.matrix()`

. (#3386)The

`pad_with`

argument in the`qml.AmplitudeEmbedding`

template is now compatible with all interfaces. (#3392)`Operator.pow`

now queues its constituents by default. (#3373)Fixed a bug where a QNode returning

`qml.sample`

would produce incorrect results when run on a device defined with a shot vector. (#3422)The

`qml.data`

module now works as expected on Windows. (#3504)

### Contributors ✍️

This release contains contributions from (in alphabetical order):

Guillermo Alonso, Juan Miguel Arrazola, Utkarsh Azad, Samuel Banning, Thomas Bromley, Astral Cai, Albert Mitjans Coma, Ahmed Darwish, Isaac De Vlugt, Olivia Di Matteo, Amintor Dusko, Pieter Eendebak, Lillian M. A. Frederiksen, Diego Guala, Katharine Hyatt, Josh Izaac, Soran Jahangiri, Edward Jiang, Korbinian Kottmann, Christina Lee, Romain Moyard, Lee James O’Riordan, Mudit Pandey, Kevin Shen, Matthew Silverman, Jay Soni, Antal Száva, David Wierichs, Moritz Willmann, and Filippo Vicentini.

- orphan

## Release 0.27.0¶

### New features since last release

#### An all-new data module 💾

The

`qml.data`

module is now available, allowing users to download, load, and create quantum datasets. (#3156)Datasets are hosted on Xanadu Cloud and can be downloaded by using

`qml.data.load()`

:>>> H2_datasets = qml.data.load( ... data_name="qchem", molname="H2", basis="STO-3G", bondlength=1.1 ... ) >>> H2data = H2_datasets[0] >>> H2data <Dataset = description: qchem/H2/STO-3G/1.1, attributes: ['molecule', 'hamiltonian', ...]>

Datasets available to be downloaded can be listed with

`qml.data.list_datasets()`

.To download or load only specific properties of a dataset, we can specify the desired properties in

`qml.data.load`

with the`attributes`

keyword argument:>>> H2_hamiltonian = qml.data.load( ... data_name="qchem", molname="H2", basis="STO-3G", bondlength=1.1, ... attributes=["molecule", "hamiltonian"] ... )[0] >>> H2_hamiltonian.hamiltonian <Hamiltonian: terms=15, wires=[0, 1, 2, 3]>

The available attributes can be found using

`qml.data.list_attributes()`

:To select data interactively, we can use

`qml.data.load_interactive()`

:>>> qml.data.load_interactive() Please select a data name: 1) qspin 2) qchem Choice [1-2]: 1 Please select a sysname: ... Please select a periodicity: ... Please select a lattice: ... Please select a layout: ... Please select attributes: ... Force download files? (Default is no) [y/N]: N Folder to download to? (Default is pwd, will download to /datasets subdirectory): Please confirm your choices: dataset: qspin/Ising/open/rectangular/4x4 attributes: ['parameters', 'ground_states'] force: False dest folder: datasets Would you like to continue? (Default is yes) [Y/n]: <Dataset = description: qspin/Ising/open/rectangular/4x4, attributes: ['parameters', 'ground_states']>

Once a dataset is loaded, its properties can be accessed as follows:

>>> dev = qml.device("default.qubit",wires=4) >>> @qml.qnode(dev) ... def circuit(): ... qml.BasisState(H2data.hf_state, wires = [0, 1, 2, 3]) ... for op in H2data.vqe_gates: ... qml.apply(op) ... return qml.expval(H2data.hamiltonian) >>> print(circuit()) -1.0791430411076344

It’s also possible to create custom datasets with

`qml.data.Dataset`

:>>> example_hamiltonian = qml.Hamiltonian(coeffs=[1,0.5], observables=[qml.PauliZ(wires=0),qml.PauliX(wires=1)]) >>> example_energies, _ = np.linalg.eigh(qml.matrix(example_hamiltonian)) >>> example_dataset = qml.data.Dataset( ... data_name = 'Example', hamiltonian=example_hamiltonian, energies=example_energies ... ) >>> example_dataset.data_name 'Example' >>> example_dataset.hamiltonian (0.5) [X1] + (1) [Z0] >>> example_dataset.energies array([-1.5, -0.5, 0.5, 1.5])

Custom datasets can be saved and read with the

`qml.data.Dataset.write()`

and`qml.data.Dataset.read()`

methods, respectively.>>> example_dataset.write('./path/to/dataset.dat') >>> read_dataset = qml.data.Dataset() >>> read_dataset.read('./path/to/dataset.dat') >>> read_dataset.data_name 'Example' >>> read_dataset.hamiltonian (0.5) [X1] + (1) [Z0] >>> read_dataset.energies array([-1.5, -0.5, 0.5, 1.5])

We will continue to work on adding more datasets and features for

`qml.data`

in future releases.

#### Adaptive optimization 🏃🏋️🏊

Optimizing quantum circuits can now be done

*adaptively*with`qml.AdaptiveOptimizer`

. (#3192)The

`qml.AdaptiveOptimizer`

takes an initial circuit and a collection of operators as input and adds a selected gate to the circuit at each optimization step. The process of growing the circuit can be repeated until the circuit gradients converge to zero within a given threshold. The adaptive optimizer can be used to implement algorithms such as ADAPT-VQE as shown in the following example.Firstly, we define some preliminary variables needed for VQE:

symbols = ["H", "H", "H"] geometry = np.array([[0.01076341, 0.04449877, 0.0], [0.98729513, 1.63059094, 0.0], [1.87262415, -0.00815842, 0.0]], requires_grad=False) H, qubits = qml.qchem.molecular_hamiltonian(symbols, geometry, charge = 1)

The collection of gates to grow the circuit is built to contain all single and double excitations:

n_electrons = 2 singles, doubles = qml.qchem.excitations(n_electrons, qubits) singles_excitations = [qml.SingleExcitation(0.0, x) for x in singles] doubles_excitations = [qml.DoubleExcitation(0.0, x) for x in doubles] operator_pool = doubles_excitations + singles_excitations

Next, an initial circuit that prepares a Hartree-Fock state and returns the expectation value of the Hamiltonian is defined:

hf_state = qml.qchem.hf_state(n_electrons, qubits) dev = qml.device("default.qubit", wires=qubits) @qml.qnode(dev) def circuit(): qml.BasisState(hf_state, wires=range(qubits)) return qml.expval(H)

Finally, the optimizer is instantiated and then the circuit is created and optimized adaptively:

opt = qml.optimize.AdaptiveOptimizer() for i in range(len(operator_pool)): circuit, energy, gradient = opt.step_and_cost(circuit, operator_pool, drain_pool=True) print('Energy:', energy) print(qml.draw(circuit)()) print('Largest Gradient:', gradient) print() if gradient < 1e-3: break

Energy: -1.246549938420637 0: ─╭BasisState(M0)─╭G²(0.20)─┤ ╭<𝓗> 1: ─├BasisState(M0)─├G²(0.20)─┤ ├<𝓗> 2: ─├BasisState(M0)─│─────────┤ ├<𝓗> 3: ─├BasisState(M0)─│─────────┤ ├<𝓗> 4: ─├BasisState(M0)─├G²(0.20)─┤ ├<𝓗> 5: ─╰BasisState(M0)─╰G²(0.20)─┤ ╰<𝓗> Largest Gradient: 0.14399872776755085 Energy: -1.2613740231529604 0: ─╭BasisState(M0)─╭G²(0.20)─╭G²(0.19)─┤ ╭<𝓗> 1: ─├BasisState(M0)─├G²(0.20)─├G²(0.19)─┤ ├<𝓗> 2: ─├BasisState(M0)─│─────────├G²(0.19)─┤ ├<𝓗> 3: ─├BasisState(M0)─│─────────╰G²(0.19)─┤ ├<𝓗> 4: ─├BasisState(M0)─├G²(0.20)───────────┤ ├<𝓗> 5: ─╰BasisState(M0)─╰G²(0.20)───────────┤ ╰<𝓗> Largest Gradient: 0.1349349562423238 Energy: -1.2743971719780331 0: ─╭BasisState(M0)─╭G²(0.20)─╭G²(0.19)──────────┤ ╭<𝓗> 1: ─├BasisState(M0)─├G²(0.20)─├G²(0.19)─╭G(0.00)─┤ ├<𝓗> 2: ─├BasisState(M0)─│─────────├G²(0.19)─│────────┤ ├<𝓗> 3: ─├BasisState(M0)─│─────────╰G²(0.19)─╰G(0.00)─┤ ├<𝓗> 4: ─├BasisState(M0)─├G²(0.20)────────────────────┤ ├<𝓗> 5: ─╰BasisState(M0)─╰G²(0.20)────────────────────┤ ╰<𝓗> Largest Gradient: 0.00040841755397108586

For a detailed breakdown of its implementation, check out the Adaptive circuits for quantum chemistry demo.

#### Automatic interface detection 🧩

QNodes now accept an

`auto`

interface argument which automatically detects the machine learning library to use. (#3132)from pennylane import numpy as np import torch import tensorflow as tf from jax import numpy as jnp dev = qml.device("default.qubit", wires=2) @qml.qnode(dev, interface="auto") def circuit(weight): qml.RX(weight[0], wires=0) qml.RY(weight[1], wires=1) return qml.expval(qml.PauliZ(0)) interface_tensors = [[0, 1], np.array([0, 1]), torch.Tensor([0, 1]), tf.Variable([0, 1], dtype=float), jnp.array([0, 1])] for tensor in interface_tensors: res = circuit(weight=tensor) print(f"Result value: {res:.2f}; Result type: {type(res)}")

Result value: 1.00; Result type: <class 'pennylane.numpy.tensor.tensor'> Result value: 1.00; Result type: <class 'pennylane.numpy.tensor.tensor'> Result value: 1.00; Result type: <class 'torch.Tensor'> Result value: 1.00; Result type: <class 'tensorflow.python.framework.ops.EagerTensor'> Result value: 1.00; Result type: <class 'jaxlib.xla_extension.Array'>

#### Upgraded JAX-JIT gradient support 🏎

JAX-JIT support for computing the gradient of QNodes that return a single vector of probabilities or multiple expectation values is now available. (#3244) (#3261)

import jax from jax import numpy as jnp from jax.config import config config.update("jax_enable_x64", True) dev = qml.device("lightning.qubit", wires=2) @jax.jit @qml.qnode(dev, diff_method="parameter-shift", interface="jax") def circuit(x, y): qml.RY(x, wires=0) qml.RY(y, wires=1) qml.CNOT(wires=[0, 1]) return qml.expval(qml.PauliZ(0)), qml.expval(qml.PauliZ(1)) x = jnp.array(1.0) y = jnp.array(2.0)

>>> jax.jacobian(circuit, argnums=[0, 1])(x, y) (Array([-0.84147098, 0.35017549], dtype=float64, weak_type=True), Array([ 4.47445479e-18, -4.91295496e-01], dtype=float64, weak_type=True))

Note that this change depends on

`jax.pure_callback`

, which requires`jax>=0.3.17`

.

#### Construct Pauli words and sentences 🔤

We’ve reorganized and grouped everything in PennyLane responsible for manipulating Pauli operators into a

`pauli`

module. The`grouping`

module has been deprecated as a result, and logic was moved from`pennylane/grouping`

to`pennylane/pauli/grouping`

. (#3179)`qml.pauli.PauliWord`

and`qml.pauli.PauliSentence`

can be used to represent tensor products and linear combinations of Pauli operators, respectively. These provide a more performant method to compute sums and products of Pauli operators. (#3195)`qml.pauli.PauliWord`

represents tensor products of Pauli operators. We can efficiently multiply and extract the matrix of these operators using this representation.>>> pw1 = qml.pauli.PauliWord({0:"X", 1:"Z"}) >>> pw2 = qml.pauli.PauliWord({0:"Y", 1:"Z"}) >>> pw1, pw2 (X(0) @ Z(1), Y(0) @ Z(1)) >>> pw1 * pw2 (Z(0), 1j) >>> pw1.to_mat(wire_order=[0,1]) array([[ 0, 0, 1, 0], [ 0, 0, 0, -1], [ 1, 0, 0, 0], [ 0, -1, 0, 0]])

`qml.pauli.PauliSentence`

represents linear combinations of Pauli words. We can efficiently add, multiply and extract the matrix of these operators in this representation.>>> ps1 = qml.pauli.PauliSentence({pw1: 1.2, pw2: 0.5j}) >>> ps2 = qml.pauli.PauliSentence({pw1: -1.2}) >>> ps1 1.2 * X(0) @ Z(1) + 0.5j * Y(0) @ Z(1) >>> ps1 + ps2 0.0 * X(0) @ Z(1) + 0.5j * Y(0) @ Z(1) >>> ps1 * ps2 -1.44 * I + (-0.6+0j) * Z(0) >>> (ps1 + ps2).to_mat(wire_order=[0,1]) array([[ 0. +0.j, 0. +0.j, 0.5+0.j, 0. +0.j], [ 0. +0.j, 0. +0.j, 0. +0.j, -0.5+0.j], [-0.5+0.j, 0. +0.j, 0. +0.j, 0. +0.j], [ 0. +0.j, 0.5+0.j, 0. +0.j, 0. +0.j]])

#### (Experimental) More support for multi-measurement and gradient output types 🧪

`qml.enable_return()`

now supports QNodes returning multiple measurements, including shots vectors, and gradient output types. (#2886) (#3052) (#3041) (#3090) (#3069) (#3137) (#3127) (#3099) (#3098) (#3095) (#3091) (#3176) (#3170) (#3194) (#3267) (#3234) (#3232) (#3223) (#3222) (#3315)In v0.25, we introduced

`qml.enable_return()`

, which separates measurements into their own tensors. The motivation of this change is the deprecation of ragged`ndarray`

creation in NumPy.With this release, we’re continuing to elevate this feature by adding support for:

Execution (

`qml.execute`

)Jacobian vector product (JVP) computation

Gradient transforms (

`qml.gradients.param_shift`

,`qml.gradients.finite_diff`

,`qml.gradients.hessian_transform`

,`qml.gradients.param_shift_hessian`

).Interfaces (Autograd, TensorFlow, and JAX, although without JIT)

With this added support, the JAX interface can handle multiple shots (shots vectors), measurements, and gradient output types with

`qml.enable_return()`

:import jax qml.enable_return() dev = qml.device("default.qubit", wires=2, shots=(1, 10000)) params = jax.numpy.array([0.1, 0.2]) @qml.qnode(dev, interface="jax", diff_method="parameter-shift", max_diff=2) def circuit(x): qml.RX(x[0], wires=[0]) qml.RY(x[1], wires=[1]) qml.CNOT(wires=[0, 1]) return qml.var(qml.PauliZ(0) @ qml.PauliX(1)), qml.probs(wires=[0])

>>> jax.hessian(circuit)(params) ((Array([[ 0., 0.], [ 2., -3.]], dtype=float32), Array([[[-0.5, 0. ], [ 0. , 0. ]], [[ 0.5, 0. ], [ 0. , 0. ]]], dtype=float32)), (Array([[ 0.07677898, 0.0563341 ], [ 0.07238522, -1.830669 ]], dtype=float32), Array([[[-4.9707499e-01, 2.9999996e-04], [-6.2500127e-04, 1.2500001e-04]], [[ 4.9707499e-01, -2.9999996e-04], [ 6.2500127e-04, -1.2500001e-04]]], dtype=float32)))

For more details, please refer to the documentation.

#### New basis rotation and tapering features in qml.qchem 🤓

Grouped coefficients, observables, and basis rotation transformation matrices needed to construct a qubit Hamiltonian in the rotated basis of molecular orbitals are now calculable via

`qml.qchem.basis_rotation()`

. (#3011)>>> symbols = ['H', 'H'] >>> geometry = np.array([[0.0, 0.0, 0.0], [1.398397361, 0.0, 0.0]], requires_grad = False) >>> mol = qml.qchem.Molecule(symbols, geometry) >>> core, one, two = qml.qchem.electron_integrals(mol)() >>> coeffs, ops, unitaries = qml.qchem.basis_rotation(one, two, tol_factor=1.0e-5) >>> unitaries [tensor([[-1.00000000e+00, -5.46483514e-13], [ 5.46483514e-13, -1.00000000e+00]], requires_grad=True), tensor([[-1.00000000e+00, 3.17585063e-14], [-3.17585063e-14, -1.00000000e+00]], requires_grad=True), tensor([[-0.70710678, -0.70710678], [-0.70710678, 0.70710678]], requires_grad=True), tensor([[ 2.58789009e-11, 1.00000000e+00], [-1.00000000e+00, 2.58789009e-11]], requires_grad=True)]

Any gate operation can now be tapered according to \(\mathbb{Z}_2\) symmetries of the Hamiltonian via

`qml.qchem.taper_operation`

. (#3002) (#3121)>>> symbols = ['He', 'H'] >>> geometry = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 1.4589]]) >>> mol = qml.qchem.Molecule(symbols, geometry, charge=1) >>> H, n_qubits = qml.qchem.molecular_hamiltonian(symbols, geometry) >>> generators = qml.qchem.symmetry_generators(H) >>> paulixops = qml.qchem.paulix_ops(generators, n_qubits) >>> paulix_sector = qml.qchem.optimal_sector(H, generators, mol.n_electrons) >>> tap_op = qml.qchem.taper_operation(qml.SingleExcitation, generators, paulixops, ... paulix_sector, wire_order=H.wires, op_wires=[0, 2]) >>> tap_op(3.14159) [Exp(1.5707949999999993j PauliY)]

Moreover, the obtained tapered operation can be used directly within a QNode.

>>> dev = qml.device('default.qubit', wires=[0, 1]) >>> @qml.qnode(dev) ... def circuit(params): ... tap_op(params[0]) ... return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1)) >>> drawer = qml.draw(circuit, show_all_wires=True) >>> print(drawer(params=[3.14159])) 0: ──Exp(0.00+1.57j Y)─┤ ╭<Z@Z> 1: ────────────────────┤ ╰<Z@Z>

Functionality has been added to estimate the number of measurements required to compute an expectation value with a target error and estimate the error in computing an expectation value with a given number of measurements. (#3000)

#### New functions, operations, and observables 🤩

Wires of operators or entire QNodes can now be mapped to other wires via

`qml.map_wires()`

. (#3143) (#3145)The

`qml.map_wires()`

function requires a dictionary representing a wire map. Use it witharbitrary operators:

>>> op = qml.RX(0.54, wires=0) + qml.PauliX(1) + (qml.PauliZ(2) @ qml.RY(1.23, wires=3)) >>> op (RX(0.54, wires=[0]) + PauliX(wires=[1])) + (PauliZ(wires=[2]) @ RY(1.23, wires=[3])) >>> wire_map = {0: 10, 1: 11, 2: 12, 3: 13} >>> qml.map_wires(op, wire_map) (RX(0.54, wires=[10]) + PauliX(wires=[11])) + (PauliZ(wires=[12]) @ RY(1.23, wires=[13]))

A

`map_wires`

method has also been added to operators, which returns a copy of the operator with its wires changed according to the given wire map.entire QNodes:

dev = qml.device("default.qubit", wires=["A", "B", "C", "D"]) wire_map = {0: "A", 1: "B", 2: "C", 3: "D"} @qml.qnode(dev) def circuit(): qml.RX(0.54, wires=0) qml.PauliX(1) qml.PauliZ(2) qml.RY(1.23, wires=3) return qml.probs(wires=0)

>>> mapped_circuit = qml.map_wires(circuit, wire_map) >>> mapped_circuit() tensor([0.92885434, 0.07114566], requires_grad=True) >>> print(qml.draw(mapped_circuit)()) A: ──RX(0.54)─┤ Probs B: ──X────────┤ C: ──Z────────┤ D: ──RY(1.23)─┤

The

`qml.IntegerComparator`

arithmetic operation is now available. (#3113)Given a basis state \(\vert n \rangle\), where \(n\) is a positive integer, and a fixed positive integer \(L\),

`qml.IntegerComparator`

flips a target qubit if \(n \geq L\). Alternatively, the flipping condition can be \(n < L\) as demonstrated below:dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(): qml.BasisState(np.array([0, 1]), wires=range(2)) qml.broadcast(qml.Hadamard, wires=range(2), pattern='single') qml.IntegerComparator(2, geq=False, wires=[0, 1]) return qml.state()

>>> circuit() [-0.5+0.j 0.5+0.j -0.5+0.j 0.5+0.j]

The

`qml.GellMann`

qutrit observable, the ternary generalization of the Pauli observables, is now available. (#3035)When using

`qml.GellMann`

, the`index`

keyword argument determines which of the 8 Gell-Mann matrices is used.dev = qml.device("default.qutrit", wires=2) @qml.qnode(dev) def circuit(): qml.TClock(wires=0) qml.TShift(wires=1) qml.TAdd(wires=[0, 1]) return qml.expval(qml.GellMann(wires=0, index=8) + qml.GellMann(wires=1, index=3))

>>> circuit() -0.42264973081037416

Controlled qutrit operations can now be performed with

`qml.ControlledQutritUnitary`

. (#2844)The control wires and values that define the operation are defined analogously to the qubit operation.

dev = qml.device("default.qutrit", wires=3) @qml.qnode(dev) def circuit(U): qml.TShift(wires=0) qml.TAdd(wires=[0, 1]) qml.ControlledQutritUnitary(U, control_wires=[0, 1], control_values='12', wires=2) return qml.state()

>>> U = np.array([[1, 1, 0], [1, -1, 0], [0, 0, np.sqrt(2)]]) / np.sqrt(2) >>> circuit(U) tensor([0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 1.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j], requires_grad=True)

### Improvements

PennyLane now supports Python 3.11! (#3297)

`qml.sample`

and`qml.counts`

work more efficiently and track if computational basis samples are being generated when they are called without specifying an observable. (#3207)The parameters of a basis set containing a different number of Gaussian functions are now easier to differentiate. (#3213)

Printing a

`qml.MultiControlledX`

operator now shows the`control_values`

keyword argument. (#3113)`qml.simplify`

and transforms like`qml.matrix`

,`batch_transform`

,`hamiltonian_expand`

, and`split_non_commuting`

now work with`QuantumScript`

as well as`QuantumTape`

. (#3209)A redundant flipping of the initial state in the UCCSD and kUpCCGSD templates has been removed. (#3148)

`qml.adjoint`

now supports batching if the base operation supports batching. (#3168)`qml.OrbitalRotation`

is now decomposed into two`qml.SingleExcitation`

operations for faster execution and more efficient parameter-shift gradient calculations on devices that natively support`qml.SingleExcitation`

. (#3171)The

`Exp`

class decomposes into a`PauliRot`

class if the coefficient is imaginary and the base operator is a Pauli Word. (#3249)Added the operator attributes

`has_decomposition`

and`has_adjoint`

that indicate whether a corresponding`decomposition`

or`adjoint`

method is available. (#2986)Structural improvements are made to

`QueuingManager`

, formerly`QueuingContext`

, and`AnnotatedQueue`

. (#2794) (#3061) (#3085)`QueuingContext`

is renamed to`QueuingManager`

.`QueuingManager`

should now be the global communication point for putting queuable objects into the active queue.`QueuingManager`

is no longer an abstract base class.`AnnotatedQueue`

and its children no longer inherit from`QueuingManager`

.`QueuingManager`

is no longer a context manager.Recording queues should start and stop recording via the

`QueuingManager.add_active_queue`

and`QueuingContext.remove_active_queue`

class methods instead of directly manipulating the`_active_contexts`

property.`AnnotatedQueue`

and its children no longer provide global information about actively recording queues. This information is now only available through`QueuingManager`

.`AnnotatedQueue`

and its children no longer have the private`_append`

,`_remove`

,`_update_info`

,`_safe_update_info`

, and`_get_info`

methods. The public analogues should be used instead.`QueuingManager.safe_update_info`

and`AnnotatedQueue.safe_update_info`

are deprecated. Their functionality is moved to`update_info`

.

`qml.Identity`

now accepts multiple wires.

>>> id_op = qml.Identity([0, 1]) >>> id_op.matrix() array([[1., 0., 0., 0.], [0., 1., 0., 0.], [0., 0., 1., 0.], [0., 0., 0., 1.]]) >>> id_op.sparse_matrix() <4x4 sparse matrix of type '<class 'numpy.float64'>' with 4 stored elements in Compressed Sparse Row format> >>> id_op.eigvals() array([1., 1., 1., 1.])

Added

`unitary_check`

keyword argument to the constructor of the`QubitUnitary`

class which indicates whether the user wants to check for unitarity of the input matrix or not. Its default value is`false`

. (#3063)Modified the representation of

`WireCut`

by using`qml.draw_mpl`

. (#3067)Improved the performance of

`qml.math.expand_matrix`

function for dense and sparse matrices. (#3060) (#3064)Added support for sums and products of operator classes with scalar tensors of any interface (NumPy, JAX, Tensorflow, PyTorch…). (#3149)

>>> s_prod = torch.tensor(4) * qml.RX(1.23, 0) >>> s_prod 4*(RX(1.23, wires=[0])) >>> s_prod.scalar tensor(4)

Added

`overlapping_ops`

property to the`Composite`

class to improve the performance of the`eigvals`

,`diagonalizing_gates`

and`Prod.matrix`

methods. (#3084)Added the

`map_wires`

method to the operators, which returns a copy of the operator with its wires changed according to the given wire map. (#3143)>>> op = qml.Toffoli([0, 1, 2]) >>> wire_map = {0: 2, 2: 0} >>> op.map_wires(wire_map=wire_map) Toffoli(wires=[2, 1, 0])

Calling

`compute_matrix`

and`compute_sparse_matrix`

of simple non-parametric operations is now faster and more memory-efficient with the addition of caching. (#3134)Added details to the output of

`Exp.label()`

. (#3126)`qml.math.unwrap`

no longer creates ragged arrays. Lists remain lists. (#3163)New

`null.qubit`

device. The`null.qubit`

performs no operations or memory allocations. (#2589)`default.qubit`

favours decomposition and avoids matrix construction for`QFT`

and`GroverOperator`

at larger qubit numbers. (#3193)`qml.ControlledQubitUnitary`

now has a`control_values`

property. (#3206)Added a new

`qml.tape.QuantumScript`

class that contains all the non-queuing behavior of`QuantumTape`

. Now,`QuantumTape`

inherits from`QuantumScript`

as well as`AnnotatedQueue`

. (#3097)Extended the

`qml.equal`

function to MeasurementProcesses (#3189)`qml.drawer.draw.draw_mpl`

now accepts a`style`

kwarg to select a style for plotting, rather than calling`qml.drawer.use_style(style)`

before plotting. Setting a style for`draw_mpl`

does not change the global configuration for matplotlib plotting. If no`style`

is passed, the function defaults to plotting with the`black_white`

style. (#3247)

### Breaking changes

`QuantumTape._par_info`

is now a list of dictionaries, instead of a dictionary whose keys are integers starting from zero. (#3185)`QueuingContext`

has been renamed to`QueuingManager`

. (#3061)Deprecation patches for the return types enum’s location and

`qml.utils.expand`

are removed. (#3092)`_multi_dispatch`

functionality has been moved inside the`get_interface`

function. This function can now be called with one or multiple tensors as arguments. (#3136)>>> torch_scalar = torch.tensor(1) >>> torch_tensor = torch.Tensor([2, 3, 4]) >>> numpy_tensor = np.array([5, 6, 7]) >>> qml.math.get_interface(torch_scalar) 'torch' >>> qml.math.get_interface(numpy_tensor) 'numpy'

`_multi_dispatch`

previously had only one argument which contained a list of the tensors to be dispatched:>>> qml.math._multi_dispatch([torch_scalar, torch_tensor, numpy_tensor]) 'torch'

To differentiate whether the user wants to get the interface of a single tensor or multiple tensors,

`get_interface`

now accepts a different argument per tensor to be dispatched:>>> qml.math.get_interface(*[torch_scalar, torch_tensor, numpy_tensor]) 'torch' >>> qml.math.get_interface(torch_scalar, torch_tensor, numpy_tensor) 'torch'

`Operator.compute_terms`

is removed. On a specific instance of an operator,`op.terms()`

can be used instead. There is no longer a static method for this. (#3215)

### Deprecations

`QueuingManager.safe_update_info`

and`AnnotatedQueue.safe_update_info`

are deprecated. Instead,`update_info`

no longer raises errorsif the object isn’t in the queue. (#3085)

`qml.tape.stop_recording`

and`QuantumTape.stop_recording`

have been moved to`qml.QueuingManager.stop_recording`

. The old functions will still be available until v0.29. (#3068)`qml.tape.get_active_tape`

has been deprecated. Use`qml.QueuingManager.active_context()`

instead. (#3068)`Operator.compute_terms`

has been removed. On a specific instance of an operator, use`op.terms()`

instead. There is no longer a static method for this. (#3215)`qml.tape.QuantumTape.inv()`

has been deprecated. Use`qml.tape.QuantumTape.adjoint`

instead. (#3237)`qml.transforms.qcut.remap_tape_wires`

has been deprecated. Use`qml.map_wires`

instead. (#3186)The grouping module

`qml.grouping`

has been deprecated. Use`qml.pauli`

or`qml.pauli.grouping`

instead. The module will still be available until v0.28. (#3262)

### Documentation

The code block in the usage details of the UCCSD template has been updated. (#3140)

Added a “Deprecations” page to the developer documentation. (#3093)

The example of the

`qml.FlipSign`

template has been updated. (#3219)

### Bug fixes

`qml.SparseHamiltonian`

now validates the size of the input matrix. (#3278)Users no longer see unintuitive errors when inputing sequences to

`qml.Hermitian`

. (#3181)The evaluation of QNodes that return either

`vn_entropy`

or`mutual_info`

raises an informative error message when using devices that define a vector of shots. (#3180)Fixed a bug that made

`qml.AmplitudeEmbedding`

incompatible with JITting. (#3166)Fixed the

`qml.transforms.transpile`

transform to work correctly for all two-qubit operations. (#3104)Fixed a bug with the control values of a controlled version of a

`ControlledQubitUnitary`

. (#3119)Fixed a bug where

`qml.math.fidelity(non_trainable_state, trainable_state)`

failed unexpectedly. (#3160)Fixed a bug where

`qml.QueuingManager.stop_recording`

did not clean up if yielded code raises an exception. (#3182)Returning

`qml.sample()`

or`qml.counts()`

with other measurements of non-commuting observables now raises a QuantumFunctionError (e.g.,`return qml.expval(PauliX(wires=0)), qml.sample()`

now raises an error). (#2924)Fixed a bug where

`op.eigvals()`

would return an incorrect result if the operator was a non-hermitian composite operator. (#3204)Fixed a bug where

`qml.BasisStatePreparation`

and`qml.BasisEmbedding`

were not jit-compilable with JAX. (#3239)Fixed a bug where

`qml.MottonenStatePreparation`

was not jit-compilable with JAX. (#3260)Fixed a bug where

`qml.MottonenStatePreparation`

was not jit-compilable with JAX. (#3260)Fixed a bug where

`qml.expval(qml.Hamiltonian())`

would not raise an error if the Hamiltonian involved some wires that are not present on the device. (#3266)Fixed a bug where

`qml.tape.QuantumTape.shape()`

did not account for the batch dimension of the tape (#3269)

### Contributors

This release contains contributions from (in alphabetical order):

Kamal Mohamed Ali, Guillermo Alonso-Linaje, Juan Miguel Arrazola, Utkarsh Azad, Thomas Bromley, Albert Mitjans Coma, Isaac De Vlugt, Olivia Di Matteo, Amintor Dusko, Lillian M. A. Frederiksen, Diego Guala, Josh Izaac, Soran Jahangiri, Edward Jiang, Korbinian Kottmann, Christina Lee, Romain Moyard, Lee J. O’Riordan, Mudit Pandey, Matthew Silverman, Jay Soni, Antal Száva, David Wierichs,

- orphan

## Release 0.26.0¶

### New features since last release

#### Classical shadows 👤

PennyLane now provides built-in support for implementing the classical-shadows measurement protocol. (#2820) (#2821) (#2871) (#2968) (#2959) (#2968)

The classical-shadow measurement protocol is described in detail in the paper Predicting Many Properties of a Quantum System from Very Few Measurements. As part of the support for classical shadows in this release, two new finite-shot and fully-differentiable measurements are available:

QNodes returning the new measurement

`qml.classical_shadow()`

will return two entities;`bits`

(0 or 1 if the 1 or -1 eigenvalue is sampled, respectively) and`recipes`

(the randomized Pauli measurements that are performed for each qubit, labelled by integer):dev = qml.device("default.qubit", wires=2, shots=3) @qml.qnode(dev) def circuit(): qml.Hadamard(wires=0) qml.CNOT(wires=[0, 1]) return qml.classical_shadow(wires=[0, 1])

>>> bits, recipes = circuit() >>> bits tensor([[0, 0], [1, 0], [0, 1]], dtype=uint8, requires_grad=True) >>> recipes tensor([[2, 2], [0, 2], [0, 2]], dtype=uint8, requires_grad=True)

QNodes returning

`qml.shadow_expval()`

yield the expectation value estimation using classical shadows:dev = qml.device("default.qubit", wires=range(2), shots=10000) @qml.qnode(dev) def circuit(x, H): qml.Hadamard(0) qml.CNOT((0,1)) qml.RX(x, wires=0) return qml.shadow_expval(H) x = np.array(0.5, requires_grad=True) H = qml.Hamiltonian( [1., 1.], [qml.PauliZ(0) @ qml.PauliZ(1), qml.PauliX(0) @ qml.PauliX(1)] )

>>> circuit(x, H) tensor(1.8486, requires_grad=True) >>> qml.grad(circuit)(x, H) -0.4797000000000001

Fully-differentiable QNode transforms for both new classical-shadows measurements are also available via

`qml.shadows.shadow_state`

and`qml.shadows.shadow_expval`

, respectively.For convenient post-processing, we’ve also added the ability to calculate general Renyi entropies by way of the

`ClassicalShadow`

class’`entropy`

method, which requires the wires of the subsystem of interest and the Renyi entropy order:>>> shadow = qml.ClassicalShadow(bits, recipes) >>> vN_entropy = shadow.entropy(wires=[0, 1], alpha=1)

#### Qutrits: quantum circuits for tertiary degrees of freedom ☘️

An entirely new framework for quantum computing is now simulatable with the addition of qutrit functionalities. (#2699) (#2781) (#2782) (#2783) (#2784) (#2841) (#2843)

Qutrits are like qubits, but instead live in a

*three*-dimensional Hilbert space; they are not binary degrees of freedom, they are*tertiary*. The advent of qutrits allows for all sorts of interesting theoretical, practical, and algorithmic capabilities that have yet to be discovered.To facilitate qutrit circuits requires a new device:

`default.qutrit`

. The`default.qutrit`

device is a Python-based simulator, akin to`default.qubit`

, and is defined as per usual:>>> dev = qml.device("default.qutrit", wires=1)

The following operations are supported on

`default.qutrit`

devices:The qutrit shift operator,

`qml.TShift`

, and the ternary clock operator,`qml.TClock`

, as defined in this paper by Yeh et al. (2022), which are the qutrit analogs of the Pauli X and Pauli Z operations, respectively.The

`qml.TAdd`

and`qml.TSWAP`

operations which are the qutrit analogs of the CNOT and SWAP operations, respectively.Custom unitary operations via

`qml.QutritUnitary`

.`qml.state`

and`qml.probs`

measurements.Measuring user-specified Hermitian matrix observables via

`qml.THermitian`

.

A comprehensive example of these features is given below:

dev = qml.device("default.qutrit", wires=1) U = np.array([ [1, 1, 1], [1, 1, 1], [1, 1, 1] ] ) / np.sqrt(3) obs = np.array([ [1, 1, 0], [1, -1, 0], [0, 0, np.sqrt(2)] ] ) / np.sqrt(2) @qml.qnode(dev) def qutrit_state(U, obs): qml.TShift(0) qml.TClock(0) qml.QutritUnitary(U, wires=0) return qml.state() @qml.qnode(dev) def qutrit_expval(U, obs): qml.TShift(0) qml.TClock(0) qml.QutritUnitary(U, wires=0) return qml.expval(qml.THermitian(obs, wires=0))

>>> qutrit_state(U, obs) tensor([-0.28867513+0.5j, -0.28867513+0.5j, -0.28867513+0.5j], requires_grad=True) >>> qutrit_expval(U, obs) tensor(0.80473785, requires_grad=True)

We will continue to add more and more support for qutrits in future releases.

#### Simplifying just got... simpler 😌

The

`qml.simplify()`

function has several intuitive improvements with this release. (#2978) (#2982) (#2922) (#3012)`qml.simplify`

can now perform the following:simplify parametrized operations

simplify the adjoint and power of specific operators

group like terms in a sum

resolve products of Pauli operators

combine rotation angles of identical rotation gates

Here is an example of

`qml.simplify`

in action with parameterized rotation gates. In this case, the angles of rotation are simplified to be modulo \(4\pi\).>>> op1 = qml.RX(30.0, wires=0) >>> qml.simplify(op1) RX(4.867258771281655, wires=[0]) >>> op2 = qml.RX(4 * np.pi, wires=0) >>> qml.simplify(op2) Identity(wires=[0])

All of these simplification features can be applied directly to quantum functions, QNodes, and tapes via decorating with

`@qml.simplify`

, as well:dev = qml.device("default.qubit", wires=2) @qml.simplify @qml.qnode(dev) def circuit(): qml.adjoint(qml.prod(qml.RX(1, 0) ** 1, qml.RY(1, 0), qml.RZ(1, 0))) return qml.probs(wires=0)

>>> circuit() >>> list(circuit.tape) [RZ(11.566370614359172, wires=[0]) @ RY(11.566370614359172, wires=[0]) @ RX(11.566370614359172, wires=[0]), probs(wires=[0])]

#### QNSPSA optimizer 💪

A new optimizer called

`qml.QNSPSAOptimizer`

is available that implements the quantum natural simultaneous perturbation stochastic approximation (QNSPSA) method based on Simultaneous Perturbation Stochastic Approximation of the Quantum Fisher Information. (#2818)`qml.QNSPSAOptimizer`

is a second-order SPSA algorithm, which combines the convergence power of the quantum-aware Quantum Natural Gradient (QNG) optimization method with the reduced quantum evaluations of SPSA methods.While the QNSPSA optimizer requires additional circuit executions (10 executions per step) compared to standard SPSA optimization (3 executions per step), these additional evaluations are used to provide a stochastic estimation of a second-order metric tensor, which often helps the optimizer to achieve faster convergence.

Use

`qml.QNSPSAOptimizer`

like you would any other optimizer:max_iterations = 50 opt = qml.QNSPSAOptimizer() for _ in range(max_iterations): params, cost = opt.step_and_cost(cost, params)

Check out our demo on the QNSPSA optimizer for more information.

#### Operator and parameter broadcasting supplements 📈

Operator methods for exponentiation and raising to a power have been added. (#2799) (#3029)

The

`qml.exp`

function can be used to create observables or generic rotation gates:>>> x = 1.234 >>> t = qml.PauliX(0) @ qml.PauliX(1) + qml.PauliY(0) @ qml.PauliY(1) >>> isingxy = qml.exp(t, 0.25j * x) >>> isingxy.matrix() array([[1. +0.j , 0. +0.j , 1. +0.j , 0. +0.j ], [0. +0.j , 0.8156179+0.j , 1. +0.57859091j, 0. +0.j ], [0. +0.j , 0. +0.57859091j, 0.8156179+0.j , 0. +0.j ], [0. +0.j , 0. +0.j , 1. +0.j , 1. +0.j ]])

The

`qml.pow`

function raises a given operator to a power:>>> op = qml.pow(qml.PauliX(0), 2) >>> op.matrix() array([[1, 0], [0, 1]])

An operator called

`qml.PSWAP`

is now available. (#2667)The

`qml.PSWAP`

gate – or phase-SWAP gate – was previously available within the PennyLane-Braket plugin only. Enjoy it natively in PennyLane with v0.26.Check whether or not an operator is hermitian or unitary with

`qml.is_hermitian`

and`qml.is_unitary`

. (#2960)>>> op1 = qml.PauliX(wires=0) >>> qml.is_hermitian(op1) True >>> op2 = qml.PauliX(0) + qml.RX(np.pi/3, 0) >>> qml.is_unitary(op2) False

Embedding templates now support parameter broadcasting. (#2810)

Embedding templates like

`AmplitudeEmbedding`

or`IQPEmbedding`

now support parameter broadcasting with a leading broadcasting dimension in their variational parameters.`AmplitudeEmbedding`

, for example, would usually use a one-dimensional input vector of features. With broadcasting, we can now compute>>> features = np.array([ ... [0.5, 0.5, 0., 0., 0.5, 0., 0.5, 0.], ... [1., 0., 0., 0., 0., 0., 0., 0.], ... [0.5, 0.5, 0., 0., 0., 0., 0.5, 0.5], ... ]) >>> op = qml.AmplitudeEmbedding(features, wires=[1, 5, 2]) >>> op.batch_size 3

An exception is

`BasisEmbedding`

, which is not broadcastable.

### Improvements

The

`qml.math.expand_matrix()`

method now allows the sparse matrix representation of an operator to be extended to a larger hilbert space. (#2998)>>> from scipy import sparse >>> mat = sparse.csr_matrix([[0, 1], [1, 0]]) >>> qml.math.expand_matrix(mat, wires=[1], wire_order=[0,1]).toarray() array([[0., 1., 0., 0.], [1., 0., 0., 0.], [0., 0., 0., 1.], [0., 0., 1., 0.]])

`qml.ctrl`

now uses`Controlled`

instead of`ControlledOperation`

. The new`Controlled`

class wraps individual`Operator`

‘s instead of a tape. It provides improved representations and integration. (#2990)`qml.matrix`

can now compute the matrix of tapes and QNodes that contain multiple broadcasted operations or non-broadcasted operations after broadcasted ones. (#3025)A common scenario in which this becomes relevant is the decomposition of broadcasted operations: the decomposition in general will contain one or multiple broadcasted operations as well as operations with no or fixed parameters that are not broadcasted.

Lists of operators are now internally sorted by their respective wires while also taking into account their commutativity property. (#2995)

Some methods of the

`QuantumTape`

class have been simplified and reordered to improve both readability and performance. (#2963)The

`qml.qchem.molecular_hamiltonian`

function is modified to support observable grouping. (#2997)`qml.ops.op_math.Controlled`

now has basic decomposition functionality. (#2938)Automatic circuit cutting has been improved by making better partition imbalance derivations. Now it is more likely to generate optimal cuts for larger circuits. (#2517)

By default,

`qml.counts`

only returns the outcomes observed in sampling. Optionally, specifying`qml.counts(all_outcomes=True)`

will return a dictionary containing all possible outcomes. (#2889)>>> dev = qml.device("default.qubit", wires=2, shots=1000) >>> >>> @qml.qnode(dev) >>> def circuit(): ... qml.Hadamard(wires=0) ... qml.CNOT(wires=[0, 1]) ... return qml.counts(all_outcomes=True) >>> result = circuit() >>> result {'00': 495, '01': 0, '10': 0, '11': 505}

Internal use of in-place inversion is eliminated in preparation for its deprecation. (#2965)

`Controlled`

operators now work with`qml.is_commuting`

. (#2994)`qml.prod`

and`qml.op_sum`

now support the`sparse_matrix()`

method. (#3006)>>> xy = qml.prod(qml.PauliX(1), qml.PauliY(1)) >>> op = qml.op_sum(xy, qml.Identity(0)) >>> >>> sparse_mat = op.sparse_matrix(wire_order=[0,1]) >>> type(sparse_mat) <class 'scipy.sparse.csr.csr_matrix'> >>> sparse_mat.toarray() [[1.+1.j 0.+0.j 0.+0.j 0.+0.j] [0.+0.j 1.-1.j 0.+0.j 0.+0.j] [0.+0.j 0.+0.j 1.+1.j 0.+0.j] [0.+0.j 0.+0.j 0.+0.j 1.-1.j]]

Provided

`sparse_matrix()`

support for single qubit observables. (#2964)`qml.Barrier`

with`only_visual=True`

now simplifies via`op.simplify()`

to the identity operator or a product of identity operators. (#3016)More accurate and intuitive outputs for printing some operators have been added. (#3013)

Results for the matrix of the sum or product of operators are stored in a more efficient manner. (#3022)

The computation of the (sparse) matrix for the sum or product of operators is now more efficient. (#3030)

When the factors of

`qml.prod`

don’t share any wires, the matrix and sparse matrix are computed using a kronecker product for improved efficiency. (#3040)`qml.grouping.is_pauli_word`

now returns`False`

for operators that don’t inherit from`qml.Observable`

instead of raising an error. (#3039)Added functionality to iterate over operators created from

`qml.op_sum`

and`qml.prod`

. (#3028)>>> op = qml.op_sum(qml.PauliX(0), qml.PauliY(1), qml.PauliZ(2)) >>> len(op) 3 >>> op[1] PauliY(wires=[1]) >>> [o.name for o in op] ['PauliX', 'PauliY', 'PauliZ']

### Deprecations

In-place inversion is now deprecated. This includes

`op.inv()`

and`op.inverse=value`

. Please use`qml.adjoint`

or`qml.pow`

instead. Support for these methods will remain till v0.28. (#2988)Don’t use:

>>> v1 = qml.PauliX(0).inv() >>> v2 = qml.PauliX(0) >>> v2.inverse = True

Instead use:

>>> qml.adjoint(qml.PauliX(0)) Adjoint(PauliX(wires=[0])) >>> qml.pow(qml.PauliX(0), -1) PauliX(wires=[0])**-1 >>> qml.pow(qml.PauliX(0), -1, lazy=False) PauliX(wires=[0]) >>> qml.PauliX(0) ** -1 PauliX(wires=[0])**-1

`qml.adjoint`

takes the conjugate transpose of an operator, while`qml.pow(op, -1)`

indicates matrix inversion. For unitary operators,`adjoint`

will be more efficient than`qml.pow(op, -1)`

, even though they represent the same thing.The

`supports_reversible_diff`

device capability is unused and has been removed. (#2993)

### Breaking changes

Measuring an operator that might not be hermitian now raises a warning instead of an error. To definitively determine whether or not an operator is hermitian, use

`qml.is_hermitian`

. (#2960)The

`ControlledOperation`

class has been removed. This was a developer-only class, so the change should not be evident to any users. It is replaced by`Controlled`

. (#2990)The default

`execute`

method for the`QubitDevice`

base class now calls`self.statistics`

with an additional keyword argument`circuit`

, which represents the quantum tape being executed. Any device that overrides`statistics`

should edit the signature of the method to include the new`circuit`

keyword argument. (#2820)The

`expand_matrix()`

has been moved from`pennylane.operation`

to`pennylane.math.matrix_manipulation`

(#3008)`qml.grouping.utils.is_commuting`

has been removed, and its Pauli word logic is now part of`qml.is_commuting`

. (#3033)`qml.is_commuting`

has been moved from`pennylane.transforms.commutation_dag`

to`pennylane.ops.functions`

. (#2991)

### Documentation

Updated the Fourier transform docs to use

`circuit_spectrum`

instead of`spectrum`

, which has been deprecated. (#3018)Corrected the docstrings for diagonalizing gates for all relevant operations. The docstrings used to say that the diagonalizing gates implemented \(U\), the unitary such that \(O = U \Sigma U^{\dagger}\), where \(O\) is the original observable and \(\Sigma\) a diagonal matrix. However, the diagonalizing gates actually implement \(U^{\dagger}\), since \(\langle \psi | O | \psi \rangle = \langle \psi | U \Sigma U^{\dagger} | \psi \rangle\), making \(U^{\dagger} | \psi \rangle\) the actual state being measured in the Z-basis. (#2981)

### Bug fixes

Fixed a bug with

`qml.ops.Exp`

operators when the coefficient is autograd but the diagonalizing gates don’t act on all wires. (#3057)Fixed a bug where the tape transform

`single_qubit_fusion`

computed wrong rotation angles for specific combinations of rotations. (#3024)Jax gradients now work with a QNode when the quantum function was transformed by

`qml.simplify`

. (#3017)Operators that have

`num_wires = AnyWires`

or`num_wires = AnyWires`

now raise an error, with certain exceptions, when instantiated with`wires=[]`

. (#2979)Fixed a bug where printing

`qml.Hamiltonian`

with complex coefficients raises`TypeError`

in some cases. (#3004)Added a more descriptive error message when measuring non-commuting observables at the end of a circuit with

`probs`

,`samples`

,`counts`

and`allcounts`

. (#3065)

### Contributors

This release contains contributions from (in alphabetical order):

Juan Miguel Arrazola, Utkarsh Azad, Tom Bromley, Olivia Di Matteo, Isaac De Vlugt, Yiheng Duan, Lillian Marie Austin Frederiksen, Josh Izaac, Soran Jahangiri, Edward Jiang, Ankit Khandelwal, Korbinian Kottmann, Meenu Kumari, Christina Lee, Albert Mitjans Coma, Romain Moyard, Rashid N H M, Zeyue Niu, Mudit Pandey, Matthew Silverman, Jay Soni, Antal Száva, Cody Wang, David Wierichs.

- orphan

## Release 0.25.1¶

### Bug fixes

Fixed Torch device discrepencies for certain parametrized operations by updating

`qml.math.array`

and`qml.math.eye`

to preserve the Torch device used. (#2967)

### Contributors

This release contains contributions from (in alphabetical order):

Romain Moyard, Rashid N H M, Lee James O’Riordan, Antal Száva

- orphan

## Release 0.25.0¶

### New features since last release

#### Estimate computational resource requirements 🧠

Functionality for estimating molecular simulation computations has been added with

`qml.resource`

. (#2646) (#2653) (#2665) (#2694) (#2720) (#2723) (#2746) (#2796) (#2797) (#2874) (#2944) (#2644)The new resource module allows you to estimate the number of non-Clifford gates and logical qubits needed to implement quantum phase estimation algorithms for simulating materials and molecules. This includes support for quantum algorithms using first and second quantization with specific bases:

First quantization using a plane-wave basis via the

`FirstQuantization`

class:>>> n = 100000 # number of plane waves >>> eta = 156 # number of electrons >>> omega = 1145.166 # unit cell volume in atomic units >>> algo = FirstQuantization(n, eta, omega) >>> print(algo.gates, algo.qubits) 1.10e+13, 4416

Second quantization with a double-factorized Hamiltonian via the

`DoubleFactorization`

class:symbols = ["O", "H", "H"] geometry = np.array( [ [0.00000000, 0.00000000, 0.28377432], [0.00000000, 1.45278171, -1.00662237], [0.00000000, -1.45278171, -1.00662237], ], requires_grad=False, ) mol = qml.qchem.Molecule(symbols, geometry, basis_name="sto-3g") core, one, two = qml.qchem.electron_integrals(mol)() algo = DoubleFactorization(one, two)

>>> print(algo.gates, algo.qubits) 103969925, 290

The methods of the

`FirstQuantization`

and the`DoubleFactorization`

classes, such as`qubit_cost`

(number of logical qubits) and`gate_cost`

(number of non-Clifford gates), can be also accessed as static methods:>>> qml.resource.FirstQuantization.qubit_cost(100000, 156, 169.69608, 0.01) 4377 >>> qml.resource.FirstQuantization.gate_cost(100000, 156, 169.69608, 0.01) 3676557345574

#### Differentiable error mitigation ⚙️

Differentiable zero-noise-extrapolation (ZNE) error mitigation is now available. (#2757)

Elevate any variational quantum algorithm to a

*mitigated*algorithm with improved results on noisy hardware while maintaining differentiability throughout.In order to do so, use the

`qml.transforms.mitigate_with_zne`

transform on your QNode and provide the PennyLane proprietary`qml.transforms.fold_global`

folding function and`qml.transforms.poly_extrapolate`

extrapolation function. Here is an example for a noisy simulation device where we mitigate a QNode and are still able to compute the gradient:# Describe noise noise_gate = qml.DepolarizingChannel noise_strength = 0.1 # Load devices dev_ideal = qml.device("default.mixed", wires=1) dev_noisy = qml.transforms.insert(noise_gate, noise_strength)(dev_ideal) scale_factors = [1, 2, 3] @mitigate_with_zne( scale_factors, qml.transforms.fold_global, qml.transforms.poly_extrapolate, extrapolate_kwargs={'order': 2} ) @qml.qnode(dev_noisy) def qnode_mitigated(theta): qml.RY(theta, wires=0) return qml.expval(qml.PauliX(0))

>>> theta = np.array(0.5, requires_grad=True) >>> qml.grad(qnode_mitigated)(theta) 0.5712737447327619

#### More native support for parameter broadcasting 📡

`default.qubit`

now natively supports parameter broadcasting, providing increased performance when executing the same circuit at various parameter positions compared to manually looping over parameters, or directly using the`qml.transforms.broadcast_expand`

transform. (#2627)dev = qml.device("default.qubit", wires=1) @qml.qnode(dev) def circuit(x): qml.RX(x, wires=0) return qml.expval(qml.PauliZ(0))

>>> circuit(np.array([0.1, 0.3, 0.2])) tensor([0.99500417, 0.95533649, 0.98006658], requires_grad=True)

Currently, not all templates have been updated to support broadcasting.

Parameter-shift gradients now allow for parameter broadcasting internally, which can result in a significant speedup when computing gradients of circuits with many parameters. (#2749)

The gradient transform

`qml.gradients.param_shift`

now accepts the keyword argument`broadcast`

. If set to`True`

, broadcasting is used to compute the derivative:dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(x, y): qml.RX(x, wires=0) qml.RY(y, wires=1) return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))

>>> x = np.array([np.pi/3, np.pi/2], requires_grad=True) >>> y = np.array([np.pi/6, np.pi/5], requires_grad=True) >>> qml.gradients.param_shift(circuit, broadcast=True)(x, y) (tensor([[-0.7795085, 0. ], [ 0. , -0.7795085]], requires_grad=True), tensor([[-0.125, 0. ], [0. , -0.125]], requires_grad=True))

The following example highlights how to make use of broadcasting gradients at the QNode level. Internally, broadcasting is used to compute the parameter-shift rule when required, which may result in performance improvements.

@qml.qnode(dev, diff_method="parameter-shift", broadcast=True) def circuit(x, y): qml.RX(x, wires=0) qml.RY(y, wires=1) return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))

>>> x = np.array(0.1, requires_grad=True) >>> y = np.array(0.4, requires_grad=True) >>> qml.grad(circuit)(x, y) (array(-0.09195267), array(-0.38747287))

Here, only 2 circuits are created internally, rather than 4 with

`broadcast=False`

.To illustrate the speedup, for a constant-depth circuit with Pauli rotations and controlled Pauli rotations, the time required to compute

`qml.gradients.param_shift(circuit, broadcast=False)(params)`

(“No broadcasting”) and`qml.gradients.param_shift(circuit, broadcast=True)(params)`

(“Broadcasting”) as a function of the number of qubits is given here.Operations for quantum chemistry now support parameter broadcasting. (#2726)

>>> op = qml.SingleExcitation(np.array([0.3, 1.2, -0.7]), wires=[0, 1]) >>> op.matrix().shape (3, 4, 4)

#### Intuitive operator arithmetic 🧮

New functionality for representing the sum, product, and scalar-product of operators is available. (#2475) (#2625) (#2622) (#2721)

The following functionalities have been added to facilitate creating new operators whose matrix, terms, and eigenvalues can be accessed as per usual, while maintaining differentiability. Operators created from these new features can be used within QNodes as operations or as observables (where physically applicable).

Summing any number of operators via

`qml.op_sum`

results in a “summed” operator:>>> ops_to_sum = [qml.PauliX(0), qml.PauliY(1), qml.PauliZ(0)] >>> summed_ops = qml.op_sum(*ops_to_sum) >>> summed_ops PauliX(wires=[0]) + PauliY(wires=[1]) + PauliZ(wires=[0]) >>> qml.matrix(summed_ops) array([[ 1.+0.j, 0.-1.j, 1.+0.j, 0.+0.j], [ 0.+1.j, 1.+0.j, 0.+0.j, 1.+0.j], [ 1.+0.j, 0.+0.j, -1.+0.j, 0.-1.j], [ 0.+0.j, 1.+0.j, 0.+1.j, -1.+0.j]]) >>> summed_ops.terms() ([1.0, 1.0, 1.0], (PauliX(wires=[0]), PauliY(wires=[1]), PauliZ(wires=[0])))

Multiplying any number of operators via

`qml.prod`

results in a “product” operator, where the matrix product or tensor product is used correspondingly:>>> theta = 1.23 >>> prod_op = qml.prod(qml.PauliZ(0), qml.RX(theta, 1)) >>> prod_op PauliZ(wires=[0]) @ RX(1.23, wires=[1]) >>> qml.eigvals(prod_op) [-1.39373197 -0.23981492 0.23981492 1.39373197]

Taking the product of a coefficient and an operator via

`qml.s_prod`

produces a “scalar-product” operator:>>> sprod_op = qml.s_prod(2.0, qml.PauliX(0)) >>> sprod_op 2.0*(PauliX(wires=[0])) >>> sprod_op.matrix() array([[ 0., 2.], [ 2., 0.]]) >>> sprod_op.terms() ([2.0], [PauliX(wires=[0])])

Each of these new functionalities can be used within QNodes as operators or observables, where applicable, while also maintaining differentiability. For example:

dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(angles): qml.prod(qml.PauliZ(0), qml.RY(angles[0], 1)) qml.op_sum(qml.PauliX(1), qml.RY(angles[1], 0)) return qml.expval(qml.op_sum(qml.PauliX(0), qml.PauliZ(1)))

>>> angles = np.array([1.23, 4.56], requires_grad=True) >>> circuit(angles) tensor(0.33423773, requires_grad=True) >>> qml.grad(circuit)(angles) array([-0.9424888, 0. ])

All PennyLane operators can now be added, subtracted, multiplied, scaled, and raised to powers using

`+`

,`-`

,`@`

,`*`

,`**`

, respectively. (#2849) (#2825) (#2891)You can now add scalars to operators, where the interpretation is that the scalar is a properly-sized identity matrix;

>>> sum_op = 5 + qml.PauliX(0) >>> sum_op.matrix() array([[5., 1.], [1., 5.]])

The

`+`

and`-`

operators can be used to combine all Pennylane operators:>>> sum_op = qml.RX(phi=1.23, wires=0) + qml.RZ(phi=3.14, wires=0) - qml.RY(phi=0.12, wires=0) >>> sum_op RX(1.23, wires=[0]) + RZ(3.14, wires=[0]) + -1*(RY(0.12, wires=[0])) >>> qml.matrix(sum_op) array([[-0.18063077-0.99999968j, 0.05996401-0.57695852j], [-0.05996401-0.57695852j, -0.18063077+0.99999968j]])

Note that the behavior of

`+`

and`-`

with*observables*is different; it still creates a Hamiltonian.The

`*`

and`@`

operators can be used to scale and compose all PennyLane operators.>>> prod_op = 2*qml.RX(1, wires=0) @ qml.RY(2, wires=0) >>> prod_op 2*(RX(1, wires=[0])) @ RY(2, wires=[0]) >>> qml.matrix(prod_op) array([[ 0.94831976-0.80684536j, -1.47692053-0.51806945j], [ 1.47692053-0.51806945j, 0.94831976+0.80684536j]])

The

`**`

operator can be used to raise PennyLane operators to a power.>>> exp_op = qml.RZ(1.0, wires=0) ** 2 >>> exp_op RZ**2(1.0, wires=[0]) >>> qml.matrix(exp_op) array([[0.54030231-0.84147098j, 0. +0.j ], [0. +0.j , 0.54030231+0.84147098j]])

A new class called

`Controlled`

is available in`qml.ops.op_math`

to represent a controlled version of any operator. This will eventually be integrated into`qml.ctrl`

to provide a performance increase and more feature coverage. (#2634)Arithmetic operations can now be simplified using

`qml.simplify`

. (#2835) (#2854)>>> op = qml.adjoint(qml.adjoint(qml.RX(x, wires=0))) >>> op Adjoint(Adjoint(RX))(tensor([1.04719755, 1.57079633], requires_grad=True), wires=[0]) >>> qml.simplify(op) RX(tensor([1.04719755, 1.57079633], requires_grad=True), wires=[0])

A new function called

`qml.equal`

can be used to compare the equality of parametric operators. (#2651)>>> qml.equal(qml.RX(1.23, 0), qml.RX(1.23, 0)) True >>> qml.equal(qml.RY(4.56, 0), qml.RY(7.89, 0)) False

#### Marvelous mixed state features 🙌

The

`default.mixed`

device now supports backpropagation with the`"jax"`

interface, which can result in significant speedups. (#2754) (#2776)dev = qml.device("default.mixed", wires=2) @qml.qnode(dev, diff_method="backprop", interface="jax") def circuit(angles): qml.RX(angles[0], wires=0) qml.RY(angles[1], wires=1) return qml.expval(qml.PauliZ(0) + qml.PauliZ(1))

>>> angles = np.array([np.pi/6, np.pi/5], requires_grad=True) >>> qml.grad(circuit)(angles) array([-0.8660254 , -0.25881905])

Additionally, quantum channels now support Jax and TensorFlow tensors. This allows quantum channels to be used inside QNodes decorated by

`tf.function`

,`jax.jit`

, or`jax.vmap`

.The

`default.mixed`

device now supports readout error. (#2786)A new keyword argument called

`readout_prob`

can be specified when creating a`default.mixed`

device. Any circuits running on a`default.mixed`

device with a finite`readout_prob`

(upper-bounded by 1) will alter the measurements performed at the end of the circuit similarly to how a`qml.BitFlip`

channel would affect circuit measurements:>>> dev = qml.device("default.mixed", wires=2, readout_prob=0.1) >>> @qml.qnode(dev) ... def circuit(): ... return qml.expval(qml.PauliZ(0)) >>> circuit() array(0.8)

#### Relative entropy is now available in qml.qinfo 💥

The quantum information module now supports computation of relative entropy. (#2772)

We’ve enabled two cases for calculating the relative entropy:

A QNode transform via

`qml.qinfo.relative_entropy`

:dev = qml.device('default.qubit', wires=2) @qml.qnode(dev) def circuit(param): qml.RY(param, wires=0) qml.CNOT(wires=[0, 1]) return qml.state()

>>> relative_entropy_circuit = qml.qinfo.relative_entropy(circuit, circuit, wires0=[0], wires1=[0]) >>> x, y = np.array(0.4), np.array(0.6) >>> relative_entropy_circuit((x,), (y,)) 0.017750012490703237

Support in

`qml.math`

for flexible post-processing:>>> rho = np.array([[0.3, 0], [0, 0.7]]) >>> sigma = np.array([[0.5, 0], [0, 0.5]]) >>> qml.math.relative_entropy(rho, sigma) tensor(0.08228288, requires_grad=True)

#### New measurements, operators, and more! ✨

A new measurement called

`qml.counts`

is available. (#2686) (#2839) (#2876)QNodes with

`shots != None`

that return`qml.counts`

will yield a dictionary whose keys are bitstrings representing computational basis states that were measured, and whose values are the corresponding counts (i.e., how many times that computational basis state was measured):dev = qml.device("default.qubit", wires=2, shots=1000) @qml.qnode(dev) def circuit(): qml.Hadamard(wires=0) qml.CNOT(wires=[0, 1]) return qml.counts()

>>> circuit() {'00': 495, '11': 505}

`qml.counts`

can also accept observables, where the resulting dictionary is ordered by the eigenvalues of the observable.dev = qml.device("default.qubit", wires=2, shots=1000) @qml.qnode(dev) def circuit(): qml.Hadamard(wires=0) qml.CNOT(wires=[0, 1]) return qml.counts(qml.PauliZ(0)), qml.counts(qml.PauliZ(1))

>>> circuit() ({-1: 470, 1: 530}, {-1: 470, 1: 530})

A new experimental return type for QNodes with multiple measurements has been added. (#2814) (#2815) (#2860)

QNodes returning a list or tuple of different measurements return an intuitive data structure via

`qml.enable_return()`

, where the individual measurements are separated into their own tensors:qml.enable_return() dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(x): qml.Hadamard(wires=[0]) qml.CRX(x, wires=[0, 1]) return (qml.probs(wires=[0]), qml.vn_entropy(wires=[0]), qml.probs(wires=0), qml.expval(wires=1))

>>> circuit(0.5) (tensor([0.5, 0.5], requires_grad=True), tensor(0.08014815, requires_grad=True), tensor([0.5, 0.5], requires_grad=True), tensor(0.93879128, requires_grad=True))

In addition, QNodes that utilize this new return type support backpropagation. This new return type can be disabled thereafter via

`qml.disable_return()`

.An operator called

`qml.FlipSign`

is now available. (#2780)Mathematically,

`qml.FlipSign`

functions as follows: \(\text{FlipSign}(n) \vert m \rangle = (-1)^\delta_{n,m} \vert m \rangle\), where \(\vert m \rangle\) is an arbitrary qubit state and $n$ is a qubit configuration:basis_state = [0, 1] dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(): for wire in list(range(2)): qml.Hadamard(wires = wire) qml.FlipSign(basis_state, wires = list(range(2))) return qml.state()

>>> circuit() tensor([ 0.5+0.j -0.5+0.j 0.5+0.j 0.5+0.j], requires_grad=True)

The simultaneous perturbation stochastic approximation (SPSA) optimizer is available via

`qml.SPSAOptimizer`

. (#2661)The SPSA optimizer is suitable for cost functions whose evaluation may involve noise. Use the SPSA optimizer like you would any other optimizer:

max_iterations = 50 opt = qml.SPSAOptimizer(maxiter=max_iterations) for _ in range(max_iterations): params, cost = opt.step_and_cost(cost, params)

#### More drawing styles 🎨

New PennyLane-inspired

`sketch`

and`sketch_dark`

styles are now available for drawing circuit diagram graphics. (#2709)

### Improvements 📈

`default.qubit`

now natively executes any operation that defines a matrix except for trainable`Pow`

operations. (#2836)Added

`expm`

to the`qml.math`

module for matrix exponentiation. (#2890)When adjoint differentiation is requested, circuits are now decomposed so that all trainable operations have a generator. (#2836)

A warning is now emitted for

`qml.state`

,`qml.density_matrix`

,`qml.vn_entropy`

, and`qml.mutual_info`

when using a device with finite shots or a shot list since these measurements are always analytic. (#2918)The efficiency of the Hartree-Fock workflow has been improved by removing repetitive steps. (#2850)

The coefficients of the non-differentiable molecular Hamiltonians generated with openfermion now have

`requires_grad = False`

by default. (#2865)Upgraded performance of the

`compute_matrix`

method of broadcastable parametric operations. (#2759)Jacobians are now cached with the Autograd interface when using the parameter-shift rule. (#2645)

The

`qml.state`

and`qml.density_matrix`

measurements now support custom wire labels. (#2779)Add trivial behaviour logic to

`qml.operation.expand_matrix`

. (#2785)Added an

`are_pauli_words_qwc`

function which checks if certain Pauli words are pairwise qubit-wise commuting. This new function improves performance when measuring hamiltonians with many commuting terms. (#2789)Adjoint differentiation now uses the adjoint symbolic wrapper instead of in-place inversion. (#2855)

### Breaking changes 💔

The deprecated

`qml.hf`

module is removed. Users with code that calls`qml.hf`

can simply replace`qml.hf`

with`qml.qchem`

in most cases, or refer to the qchem documentation and demos for more information. (#2795)`default.qubit`

now uses`stopping_condition`

to specify support for anything with a matrix. To override this behavior in inheriting devices and to support only a specific subset of operations, developers need to override`stopping_condition`

. (#2836)Custom devices inheriting from

`DefaultQubit`

or`QubitDevice`

can break due to the introduction of parameter broadcasting. (#2627)A custom device should only break if all three following statements hold simultaneously:

The custom device inherits from

`DefaultQubit`

, not`QubitDevice`

.The device implements custom methods in the simulation pipeline that are incompatible with broadcasting (for example

`expval`

,`apply_operation`

or`analytic_probability`

).The custom device maintains the flag

`"supports_broadcasting": True`

in its`capabilities`

dictionary*or*it overwrites`Device.batch_transform`

without applying`broadcast_expand`

(or both).

The

`capabilities["supports_broadcasting"]`

is set to`True`

for`DefaultQubit`

. Typically, the easiest fix will be to change the`capabilities["supports_broadcasting"]`

flag to`False`

for the child device and/or to include a call to`broadcast_expand`

in`CustomDevice.batch_transform`

, similar to how`Device.batch_transform`

calls it.Separately from the above, custom devices that inherit from

`QubitDevice`

and implement a custom`_gather`

method need to allow for the kwarg`axis`

to be passed to this`_gather`

method.The argument

`argnum`

of the function`qml.batch_input`

has been redefined: now it indicates the indices of the batched parameters, which need to be non-trainable, in the quantum tape. Consequently, its default value (set to 0) has been removed. (#2873)Before this breaking change, one could call

`qml.batch_input`

without any arguments when using batched inputs as the first argument of the quantum circuit.dev = qml.device("default.qubit", wires=2, shots=None) @qml.batch_input() # argnum = 0 @qml.qnode(dev, diff_method="parameter-shift", interface="tf") def circuit(inputs, weights): # argument `inputs` is batched qml.RY(weights[0], wires=0) qml.AngleEmbedding(inputs, wires=range(2), rotation="Y") qml.RY(weights[1], wires=1) return qml.expval(qml.PauliZ(1))

With this breaking change, users must set a value to

`argnum`

specifying the index of the batched inputs with respect to all quantum tape parameters. In this example the quantum tape parameters are`[ weights[0], inputs, weights[1] ]`

, thus`argnum`

should be set to 1, specifying that`inputs`

is batched:dev = qml.device("default.qubit", wires=2, shots=None) @qml.batch_input(argnum=1) @qml.qnode(dev, diff_method="parameter-shift", interface="tf") def circuit(inputs, weights): qml.RY(weights[0], wires=0) qml.AngleEmbedding(inputs, wires=range(2), rotation="Y") qml.RY(weights[1], wires=1) return qml.expval(qml.PauliZ(1))

PennyLane now depends on newer versions (>=2.7) of the

`semantic_version`

package, which provides an updated API that is incompatible which versions of the package prior to 2.7. If you run into issues relating to this package, please reinstall PennyLane. (#2744) (#2767)

### Documentation 📕

Added a dedicated docstring for the

`QubitDevice.sample`

method. (#2812)Optimization examples of using JAXopt and Optax with the JAX interface have been added. (#2769)

Updated IsingXY gate docstring. (#2858)

### Bug fixes 🐞

Fixes

`qml.equal`

so that operators with different inverse properties are not equal. (#2947)Cleans up interactions between operator arithmetic and batching by testing supported cases and adding errors when batching is not supported. (#2900)

Fixed a bug where the parameter-shift rule wasn’t defined for

`qml.kUpCCGSD`

. (#2913)Reworked the Hermiticity check in

`qml.Hermitian`

by using`qml.math`

calls because calling`.conj()`

on an`EagerTensor`

from TensorFlow raised an error. (#2895)Fixed a bug where the parameter-shift gradient breaks when using both custom

`grad_recipe`

s that contain unshifted terms and recipes that do not contain any unshifted terms. (#2834)Fixed mixed CPU-GPU data-locality issues for the Torch interface. (#2830)

Fixed a bug where the parameter-shift Hessian of circuits with untrainable parameters might be computed with respect to the wrong parameters or might raise an error. (#2822)

Fixed a bug where the custom implementation of the

`states_to_binary`

device method was not used. (#2809)`qml.grouping.group_observables`

now works when individual wire labels are iterable. (#2752)The adjoint of an adjoint now has a correct

`expand`

result. (#2766)Fixed the ability to return custom objects as the expectation value of a QNode with the Autograd interface. (#2808)

The WireCut operator now raises an error when instantiating it with an empty list. (#2826)

Hamiltonians with grouped observables are now allowed to be measured on devices which were transformed using

`qml.transform.insert()`

. (#2857)Fixed a bug where

`qml.batch_input`

raised an error when using a batched operator that was not located at the beginning of the circuit. In addition, now`qml.batch_input`

raises an error when using trainable batched inputs, which avoids an unwanted behaviour with duplicated parameters. (#2873)Calling

`qml.equal`

with nested operators now raises a`NotImplementedError`

. (#2877)Fixed a bug where a non-sensible error message was raised when using

`qml.counts`

with`shots=False`

. (#2928)Fixed a bug where no error was raised and a wrong value was returned when using

`qml.counts`

with another non-commuting observable. (#2928)Operator Arithmetic now allows

`Hamiltonian`

objects to be used and produces correct matrices. (#2957)

### Contributors

This release contains contributions from (in alphabetical order):

Juan Miguel Arrazola, Utkarsh Azad, Samuel Banning, Prajwal Borkar, Isaac De Vlugt, Olivia Di Matteo, Kristiyan Dilov, David Ittah, Josh Izaac, Soran Jahangiri, Edward Jiang, Ankit Khandelwal, Korbinian Kottmann, Meenu Kumari, Christina Lee, Sergio Martínez-Losa, Albert Mitjans Coma, Ixchel Meza Chavez, Romain Moyard, Lee James O’Riordan, Mudit Pandey, Bogdan Reznychenko, Shuli Shu, Jay Soni, Modjtaba Shokrian-Zini, Antal Száva, David Wierichs, Moritz Willmann

- orphan

## Release 0.24.0¶

### New features since last release

#### All new quantum information quantities 📏

Functionality for computing quantum information quantities for QNodes has been added. (#2554) (#2569) (#2598) (#2617) (#2631) (#2640) (#2663) (#2684) (#2688) (#2695) (#2710) (#2712)

This includes two new QNode measurements:

The Von Neumann entropy via

`qml.vn_entropy`

:>>> dev = qml.device("default.qubit", wires=2) >>> @qml.qnode(dev) ... def circuit_entropy(x): ... qml.IsingXX(x, wires=[0,1]) ... return qml.vn_entropy(wires=[0], log_base=2) >>> circuit_entropy(np.pi/2) 1.0

The mutual information via

`qml.mutual_info`

:>>> dev = qml.device("default.qubit", wires=2) >>> @qml.qnode(dev) ... def circuit(x): ... qml.IsingXX(x, wires=[0,1]) ... return qml.mutual_info(wires0=[0], wires1=[1], log_base=2) >>> circuit(np.pi/2) 2.0

New differentiable transforms are also available in the

`qml.qinfo`

module:The classical and quantum Fisher information via

`qml.qinfo.classical_fisher`

,`qml.qinfo.quantum_fisher`

, respectively:dev = qml.device("default.qubit", wires=3) @qml.qnode(dev) def circ(params): qml.RY(params[0], wires=1) qml.CNOT(wires=(1,0)) qml.RY(params[1], wires=1) qml.RZ(params[2], wires=1) return qml.expval(qml.PauliX(0) @ qml.PauliX(1) - 0.5 * qml.PauliZ(1)) params = np.array([0.5, 1., 0.2], requires_grad=True) cfim = qml.qinfo.classical_fisher(circ)(params) qfim = qml.qinfo.quantum_fisher(circ)(params)

These quantities are typically employed in variational optimization schemes to tilt the gradient in a more favourable direction — producing what is known as the natural gradient. For example:

>>> grad = qml.grad(circ)(params) >>> cfim @ grad # natural gradient [ 5.94225615e-01 -2.61509542e-02 -1.18674655e-18] >>> qfim @ grad # quantum natural gradient [ 0.59422561 -0.02615095 -0.03989212]

The fidelity between two arbitrary states via

`qml.qinfo.fidelity`

:dev = qml.device('default.qubit', wires=1) @qml.qnode(dev) def circuit_rx(x): qml.RX(x[0], wires=0) qml.RZ(x[1], wires=0) return qml.state() @qml.qnode(dev) def circuit_ry(y): qml.RY(y, wires=0) return qml.state()

>>> x = np.array([0.1, 0.3], requires_grad=True) >>> y = np.array(0.2, requires_grad=True) >>> fid_func = qml.qinfo.fidelity(circuit_rx, circuit_ry, wires0=[0], wires1=[0]) >>> fid_func(x, y) 0.9905158135644924 >>> df = qml.grad(fid_func) >>> df(x, y) (array([-0.04768725, -0.29183666]), array(-0.09489803))

Reduced density matrices of arbitrary states via

`qml.qinfo.reduced_dm`

:dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) def circuit(x): qml.IsingXX(x, wires=[0,1]) return qml.state()

>>> qml.qinfo.reduced_dm(circuit, wires=[0])(np.pi/2) [[0.5+0.j 0.+0.j] [0.+0.j 0.5+0.j]]

Similar transforms,

`qml.qinfo.vn_entropy`

and`qml.qinfo.mutual_info`

exist for transforming QNodes.

Currently, all quantum information measurements and transforms are differentiable, but only support statevector devices, with hardware support to come in a future release (with the exception of

`qml.qinfo.classical_fisher`

and`qml.qinfo.quantum_fisher`

, which are both hardware compatible).For more information, check out the new qinfo module and measurements page.

In addition to the QNode transforms and measurements above, functions for computing and differentiating quantum information metrics with numerical statevectors and density matrices have been added to the

`qml.math`

module. This enables flexible custom post-processing.Added functions include:

`qml.math.reduced_dm`

`qml.math.vn_entropy`

`qml.math.mutual_info`

`qml.math.fidelity`

For example:

>>> x = torch.tensor([1.0, 0.0, 0.0, 1.0], requires_grad=True) >>> en = qml.math.vn_entropy(x / np.sqrt(2.), indices=[0]) >>> en tensor(0.6931, dtype=torch.float64, grad_fn=<DivBackward0>) >>> en.backward() >>> x.grad tensor([-0.3069, 0.0000, 0.0000, -0.3069])

#### Faster mixed-state training with backpropagation 📉

The

`default.mixed`

device now supports differentiation via backpropagation with the Autograd, TensorFlow, and PyTorch (CPU) interfaces, leading to significantly more performant optimization and training. (#2615) (#2670) (#2680)As a result, the default differentiation method for the device is now

`"backprop"`

. To continue using the old default`"parameter-shift"`

, explicitly specify this differentiation method in the QNode:dev = qml.device("default.mixed", wires=2) @qml.qnode(dev, interface="autograd", diff_method="backprop") def circuit(x): qml.RY(x, wires=0) qml.CNOT(wires=[0, 1]) return qml.expval(qml.PauliZ(wires=1))

>>> x = np.array(0.5, requires_grad=True) >>> circuit(x) array(0.87758256) >>> qml.grad(circuit)(x) -0.479425538604203

#### Support for quantum parameter broadcasting 📡

Quantum operators, functions, and tapes now support broadcasting across parameter dimensions, making it more convenient for developers to execute their PennyLane programs with multiple sets of parameters. (#2575) (#2609)

Parameter broadcasting refers to passing tensor parameters with additional leading dimensions to quantum operators; additional dimensions will flow through the computation, and produce additional dimensions at the output.

For example, instantiating a rotation gate with a one-dimensional array leads to a broadcasted

`Operation`

:>>> x = np.array([0.1, 0.2, 0.3], requires_grad=True) >>> op = qml.RX(x, 0) >>> op.batch_size 3

Its matrix correspondingly is augmented by a leading dimension of size

`batch_size`

:>>> np.round(qml.matrix(op), 4) tensor([[[0.9988+0.j , 0. -0.05j ], [0. -0.05j , 0.9988+0.j ]], [[0.995 +0.j , 0. -0.0998j], [0. -0.0998j, 0.995 +0.j ]], [[0.9888+0.j , 0. -0.1494j], [0. -0.1494j, 0.9888+0.j ]]], requires_grad=True) >>> qml.matrix(op).shape (3, 2, 2)

This can be extended to quantum functions, where we may mix-and-match operations with batched parameters and those without. However, the

`batch_size`

of each batched`Operator`

within the quantum function must be the same:>>> dev = qml.device('default.qubit', wires=1) >>> @qml.qnode(dev) ... def circuit_rx(x, z): ... qml.RX(x, wires=0) ... qml.RZ(z, wires=0) ... qml.RY(0.3, wires=0) ... return qml.probs(wires=0) >>> circuit_rx([0.1, 0.2], [0.3, 0.4]) tensor([[0.97092256, 0.02907744], [0.95671515, 0.04328485]], requires_grad=True)

Parameter broadcasting is supported on all devices, hardware and simulator. Note that if not natively supported by the underlying device, parameter broadcasting may result in additional quantum device evaluations.

A new transform,

`qml.transforms.broadcast_expand`

, has been added, which automates the process of transforming quantum functions (and tapes) to multiple quantum evaluations with no parameter broadcasting. (#2590)>>> dev = qml.device('default.qubit', wires=1) >>> @qml.transforms.broadcast_expand() >>> @qml.qnode(dev) ... def circuit_rx(x, z): ... qml.RX(x, wires=0) ... qml.RZ(z, wires=0) ... qml.RY(0.3, wires=0) ... return qml.probs(wires=0) >>> print(qml.draw(circuit_rx)([0.1, 0.2], [0.3, 0.4])) 0: ──RX(0.10)──RZ(0.30)──RY(0.30)─┤ Probs \ 0: ──RX(0.20)──RZ(0.40)──RY(0.30)─┤ Probs

Under-the-hood, this transform is used for devices that don’t natively support parameter broadcasting.

To specify that a device natively supports broadcasted tapes, the new flag

`Device.capabilities()["supports_broadcasting"]`

should be set to`True`

.To support parameter broadcasting for new or custom operations, the following new

`Operator`

class attributes must be specified:`Operator.ndim_params`

specifies expected number of dimensions for each parameter

Once set,

`Operator.batch_size`

and`QuantumTape.batch_size`

will dynamically compute the parameter broadcasting axis dimension, if present.

#### Improved JAX JIT support 🏎

JAX just-in-time (JIT) compilation now supports vector-valued QNodes, enabling new types of workflows and significant performance boosts. (#2034)

Vector-valued QNodes include those with:

`qml.probs`

;`qml.state`

;`qml.sample`

ormultiple

`qml.expval`

/`qml.var`

measurements.

Consider a QNode that returns basis-state probabilities:

dev = qml.device('default.qubit', wires=2) x = jnp.array(0.543) y = jnp.array(-0.654) @jax.jit @qml.qnode(dev, diff_method="parameter-shift", interface="jax") def circuit(x, y): qml.RX(x, wires=[0]) qml.RY(y, wires=[1]) qml.CNOT(wires=[0, 1]) return qml.probs(wires=[1])

>>> circuit(x, y) Array([0.8397495 , 0.16025047], dtype=float32)

Note that computing the jacobian of vector-valued QNode is not supported with JAX JIT. The output of vector-valued QNodes can, however, be used in the definition of scalar-valued cost functions whose gradients can be computed.

For example, one can define a cost function that outputs the first element of the probability vector:

def cost(x, y): return circuit(x, y)[0]

>>> jax.grad(cost, argnums=[0])(x, y) (Array(-0.2050439, dtype=float32),)

#### More drawing styles 🎨

New

`solarized_light`

and`solarized_dark`

styles are available for drawing circuit diagram graphics. (#2662)

#### New operations & transforms 🤖

The

`qml.IsingXY`

gate is now available (see 1912.04424). (#2649)The

`qml.ECR`

(echoed cross-resonance) operation is now available (see 2105.01063). This gate is a maximally-entangling gate and is equivalent to a CNOT gate up to single-qubit pre-rotations. (#2613)The adjoint transform

`adjoint`

can now accept either a single instantiated operator or a quantum function. It returns an entity of the same type / call signature as what it was given: (#2222) (#2672)>>> qml.adjoint(qml.PauliX(0)) Adjoint(PauliX)(wires=[0]) >>> qml.adjoint(qml.RX)(1.23, wires=0) Adjoint(RX)(1.23, wires=[0])

Now,

`adjoint`

wraps operators in a symbolic operator class`qml.ops.op_math.Adjoint`

. This class should not be constructed directly; the`adjoint`

constructor should always be used instead. The class behaves just like any other`Operator`

:>>> op = qml.adjoint(qml.S(0)) >>> qml.matrix(op) array([[1.-0.j, 0.-0.j], [0.-0.j, 0.-1.j]]) >>> qml.eigvals(op) array([1.-0.j, 0.-1.j])

A new symbolic operator class

`qml.ops.op_math.Pow`

represents an operator raised to a power. When`decomposition()`

is called, a list of new operators equal to this one raised to the given power is given: (#2621)>>> op = qml.ops.op_math.Pow(qml.PauliX(0), 0.5) >>> op.decomposition() [SX(wires=[0])] >>> qml.matrix(op) array([[0.5+0.5j, 0.5-0.5j], [0.5-0.5j, 0.5+0.5j]])

A new transform

`qml.batch_partial`

is available which behaves similarly to`functools.partial`

, but supports batching in the unevaluated parameters. (#2585)This is useful for executing a circuit with a batch dimension in some of its parameters:

dev = qml.device("default.qubit", wires=1) @qml.qnode(dev) def circuit(x, y): qml.RX(x, wires=0) qml.RY(y, wires=0) return qml.expval(qml.PauliZ(wires=0))

>>> batched_partial_circuit = qml.batch_partial(circuit, x=np.array(np.pi / 4)) >>> y = np.array([0.2, 0.3, 0.4]) >>> batched_partial_circuit(y=y) tensor([0.69301172, 0.67552491, 0.65128847], requires_grad=True)

A new transform

`qml.split_non_commuting`

is available, which splits a quantum function or tape into multiple functions/tapes determined by groups of commuting observables: (#2587)dev = qml.device("default.qubit", wires=1) @qml.transforms.split_non_commuting @qml.qnode(dev) def circuit(x): qml.RX(x,wires=0) return [qml.expval(qml.PauliX(0)), qml.expval(qml.PauliZ(0))]

>>> print(qml.draw(circuit)(0.5)) 0: ──RX(0.50)─┤ <X> \ 0: ──RX(0.50)─┤ <Z>

### Improvements

Expectation values of multiple non-commuting observables from within a single QNode are now supported: (#2587)

>>> dev = qml.device('default.qubit', wires=1) >>> @qml.qnode(dev) ... def circuit_rx(x, z): ... qml.RX(x, wires=0) ... qml.RZ(z, wires=0) ... return qml.expval(qml.PauliX(0)), qml.expval(qml.PauliY(0)) >>> circuit_rx(0.1, 0.3) tensor([ 0.02950279, -0.09537451], requires_grad=True)

Selecting which parts of parameter-shift Hessians are computed is now possible. (#2538)

The

`argnum`

keyword argument for`qml.gradients.param_shift_hessian`

is now allowed to be a two-dimensional Boolean`array_like`

. Only the indicated entries of the Hessian will then be computed.A particularly useful example is the computation of the diagonal of the Hessian:

dev = qml.device("default.qubit", wires=1) @qml.qnode(dev) def circuit(x): qml.RX(x[0], wires=0) qml.RY(x[1], wires=0) qml.RX(x[2], wires=0) return qml.expval(qml.PauliZ(0)) argnum = qml.math.eye(3, dtype=bool) x = np.array([0.2, -0.9, 1.1], requires_grad=True)

>>> qml.gradients.param_shift_hessian(circuit, argnum=argnum)(x) tensor([[-0.09928388, 0. , 0. ], [ 0. , -0.27633945, 0. ], [ 0. , 0. , -0.09928388]], requires_grad=True)

Commuting Pauli operators are now measured faster. (#2425)

The logic that checks for qubit-wise commuting (QWC) observables has been improved, resulting in a performance boost that is noticable when many commuting Pauli operators of the same type are measured.

It is now possible to add

`Observable`

objects to the integer`0`

, for example`qml.PauliX(wires=[0]) + 0`

. (#2603)Wires can now be passed as the final argument to an

`Operator`

, instead of requiring the wires to be explicitly specified with keyword`wires`

. This functionality already existed for`Observable`

s, but now extends to all`Operator`

s: (#2432)>>> qml.S(0) S(wires=[0]) >>> qml.CNOT((0,1)) CNOT(wires=[0, 1])

The

`qml.taper`

function can now be used to consistently taper any additional observables such as dipole moment, particle number, and spin operators using the symmetries obtained from the Hamiltonian. (#2510)Sparse Hamiltonians’ representation has changed from Coordinate (COO) to Compressed Sparse Row (CSR) format. (#2561)

The CSR representation is more performant for arithmetic operations and matrix-vector products. This change decreases the

`expval()`

calculation time for`qml.SparseHamiltonian`

, specially for large workflows. In addition, the CSR format consumes less memory for`qml.SparseHamiltonian`

storage.IPython now displays the

`str`

representation of a`Hamiltonian`

, rather than the`repr`

. This displays more information about the object. (#2648)The

`qml.qchem`

tests have been restructured. (#2593) (#2545)OpenFermion-dependent tests are now localized and collected in

`tests.qchem.of_tests`

. The new module`test_structure`

is created to collect the tests of the`qchem.structure`

module in one place and remove their dependency to OpenFermion.Test classes have been created to group the integrals and matrices unit tests.

An

`operations_only`

argument is introduced to the`tape.get_parameters`

method. (#2543)The

`gradients`

module now uses faster subroutines and uniform formats of gradient rules. (#2452)Instead of checking types, objects are now processed in the

`QuantumTape`

based on a new`_queue_category`

property. This is a temporary fix that will disappear in the future. (#2408)The

`QNode`

class now contains a new method`best_method_str`

that returns the best differentiation method for a provided device and interface, in human-readable format. (#2533)Using

`Operation.inv()`

in a queuing environment no longer updates the queue’s metadata, but merely updates the operation in place. (#2596)A new method

`safe_update_info`

is added to`qml.QueuingContext`

. This method is substituted for`qml.QueuingContext.update_info`

in a variety of places. (#2612) (#2675)`BasisEmbedding`

can accept an int as argument instead of a list of bits. (#2601)For example,

`qml.BasisEmbedding(4, wires = range(4))`

is now equivalent to`qml.BasisEmbedding([0,1,0,0], wires = range(4))`

(as`4==0b100`

).Introduced a new

`is_hermitian`

property to`Operator`

s to determine if an operator can be used in a measurement process. (#2629)Added separate

`requirements_dev.txt`

for separation of concerns for code development and just using PennyLane. (#2635)The performance of building sparse Hamiltonians has been improved by accumulating the sparse representation of coefficient-operator pairs in a temporary storage and by eliminating unnecessary

`kron`

operations on identity matrices. (#2630)Control values are now displayed distinctly in text and matplotlib drawings of circuits. (#2668)

The

`TorchLayer`

`init_method`

argument now accepts either a`torch.nn.init`

function or a dictionary which should specify a`torch.nn.init`

/`torch.Tensor`

for each different weight. (#2678)The unused keyword argument

`do_queue`

for`Operation.adjoint`

is now fully removed. (#2583)Several non-decomposable

`Adjoint`

operators are added to the device test suite. (#2658)The developer-facing

`pow`

method has been added to`Operator`

with concrete implementations for many classes. (#2225)The

`ctrl`

transform and`ControlledOperation`

have been moved to the new`qml.ops.op_math`

submodule. The developer-facing`ControlledOperation`

class is no longer imported top-level. (#2656)

### Deprecations

`qml.ExpvalCost`

has been deprecated, and usage will now raise a warning. (#2571)Instead, it is recommended to simply pass Hamiltonians to the

`qml.expval`

function inside QNodes:@qml.qnode(dev) def ansatz(params): some_qfunc(params) return qml.expval(Hamiltonian)

### Breaking changes

When using

`qml.TorchLayer`

, weights with negative shapes will now raise an error, while weights with`size = 0`

will result in creating empty Tensor objects. (#2678)PennyLane no longer supports TensorFlow

`<=2.3`

. (#2683)The

`qml.queuing.Queue`

class has been removed. (#2599)The

`qml.utils.expand`

function is now removed;`qml.operation.expand_matrix`

should be used instead. (#2654)The module

`qml.gradients.param_shift_hessian`

has been renamed to`qml.gradients.parameter_shift_hessian`

in order to distinguish it from the identically named function. Note that the`param_shift_hessian`

function is unaffected by this change and can be invoked in the same manner as before via the`qml.gradients`

module. (#2528)The properties

`eigval`

and`matrix`

from the`Operator`

class were replaced with the methods`eigval()`

and`matrix(wire_order=None)`

. (#2498)`Operator.decomposition()`

is now an instance method, and no longer accepts parameters. (#2498)Adds tests, adds no-coverage directives, and removes inaccessible logic to improve code coverage. (#2537)

The base classes

`QubitDevice`

and`DefaultQubit`

now accept data-types for a statevector. This enables a derived class (device) in a plugin to choose correct data-types: (#2448)>>> dev = qml.device("default.qubit", wires=4, r_dtype=np.float32, c_dtype=np.complex64) >>> dev.R_DTYPE <class 'numpy.float32'> >>> dev.C_DTYPE <class 'numpy.complex64'>

### Bug fixes

Fixed a bug where returning

`qml.density_matrix`

using the PyTorch interface would return a density matrix with wrong shape. (#2643)Fixed a bug to make

`param_shift_hessian`

work with QNodes in which gates marked as trainable do not have any impact on the QNode output. (#2584)QNodes can now interpret variations on the interface name, like

`"tensorflow"`

or`"jax-jit"`

, when requesting backpropagation. (#2591)Fixed a bug for

`diff_method="adjoint"`

where incorrect gradients were computed for QNodes with parametrized observables (e.g.,`qml.Hermitian`

). (#2543)Fixed a bug where

`QNGOptimizer`

did not work with operators whose generator was a Hamiltonian. (#2524)Fixed a bug with the decomposition of

`qml.CommutingEvolution`

. (#2542)Fixed a bug enabling PennyLane to work with the latest version of Autoray. (#2549)

Fixed a bug which caused different behaviour for

`Hamiltonian @ Observable`

and`Observable @ Hamiltonian`

. (#2570)Fixed a bug in

`DiagonalQubitUnitary._controlled`

where an invalid operation was queued instead of the controlled version of the diagonal unitary. (#2525)Updated the gradients fix to only apply to the

`strawberryfields.gbs`

device, since the original logic was breaking some devices. (#2485) (#2595)Fixed a bug in

`qml.transforms.insert`

where operations were not inserted after gates within a template. (#2704)`Hamiltonian.wires`

is now properly updated after in place operations. (#2738)

### Documentation

The centralized Xanadu Sphinx Theme is now used to style the Sphinx documentation. (#2450)

Added a reference to

`qml.utils.sparse_hamiltonian`

in`qml.SparseHamiltonian`

to clarify how to construct sparse Hamiltonians in PennyLane. (2572)Added a new section in the Gradients and Training page that summarizes the supported device configurations and provides justification. In addition, code examples were added for some selected configurations. (#2540)

Added a note for the Depolarization Channel that specifies how the channel behaves for the different values of depolarization probability

`p`

. (#2669)The quickstart documentation has been improved. (#2530) (#2534) (#2564 (#2565 (#2566) (#2607) (#2608)

The quantum chemistry quickstart documentation has been improved. (#2500)

Testing documentation has been improved. (#2536)

Documentation for the

`pre-commit`

package has been added. (#2567)Documentation for draw control wires change has been updated. (#2682)

### Contributors

This release contains contributions from (in alphabetical order):

Guillermo Alonso-Linaje, Mikhail Andrenkov, Juan Miguel Arrazola, Ali Asadi, Utkarsh Azad, Samuel Banning, Avani Bhardwaj, Thomas Bromley, Albert Mitjans Coma, Isaac De Vlugt, Amintor Dusko, Trent Fridey, Christian Gogolin, Qi Hu, Katharine Hyatt, David Ittah, Josh Izaac, Soran Jahangiri, Edward Jiang, Nathan Killoran, Korbinian Kottmann, Ankit Khandelwal, Christina Lee, Chae-Yeun Park, Mason Moreland, Romain Moyard, Maria Schuld, Jay Soni, Antal Száva, tal66, David Wierichs, Roeland Wiersema, WingCode.

- orphan

## Release 0.23.1¶

### Bug fixes

Fixed a bug enabling PennyLane to work with the latest version of Autoray. (#2548)

### Contributors

This release contains contributions from (in alphabetical order):

Josh Izaac

- orphan

## Release 0.23.0¶

### New features since last release

#### More powerful circuit cutting ✂️

Quantum circuit cutting (running

`N`

-wire circuits on devices with fewer than`N`

wires) is now supported for QNodes of finite-shots using the new`@qml.cut_circuit_mc`

transform. (#2313) (#2321) (#2332) (#2358) (#2382) (#2399) (#2407) (#2444)With these new additions, samples from the original circuit can be simulated using a Monte Carlo method, using fewer qubits at the expense of more device executions. Additionally, this transform can take an optional classical processing function as an argument and return an expectation value.

The following

`3`

-qubit circuit contains a`WireCut`

operation and a`sample`

measurement. When decorated with`@qml.cut_circuit_mc`

, we can cut the circuit into two`2`

-qubit fragments:dev = qml.device("default.qubit", wires=2, shots=1000) @qml.cut_circuit_mc @qml.qnode(dev) def circuit(x): qml.RX(0.89, wires=0) qml.RY(0.5, wires=1) qml.RX(1.3, wires=2) qml.CNOT(wires=[0, 1]) qml.WireCut(wires=1) qml.CNOT(wires=[1, 2]) qml.RX(x, wires=0) qml.RY(0.7, wires=1) qml.RX(2.3, wires=2) return qml.sample(wires=[0, 2])

we can then execute the circuit as usual by calling the QNode:

>>> x = 0.3 >>> circuit(x) tensor([[1, 1], [0, 1], [0, 1], ..., [0, 1], [0, 1], [0, 1]], requires_grad=True)

Furthermore, the number of shots can be temporarily altered when calling the QNode:

>>> results = circuit(x, shots=123) >>> results.shape (123, 2)

The

`cut_circuit_mc`

transform also supports returning sample-based expectation values of observables using the`classical_processing_fn`

argument. Refer to the`UsageDetails`

section of the transform documentation for an example.The

`cut_circuit`

transform now supports automatic graph partitioning by specifying`auto_cutter=True`

to cut arbitrary tape-converted graphs using the general purpose graph partitioning framework KaHyPar. (#2330) (#2428)Note that

`KaHyPar`

needs to be installed separately with the`auto_cutter=True`

option.For integration with the existing low-level manual cut pipeline, refer to the documentation of the function .

@qml.cut_circuit(auto_cutter=True) @qml.qnode(dev) def circuit(x): qml.RX(x, wires=0) qml.RY(0.9, wires=1) qml.RX(0.3, wires=2) qml.CZ(wires=[0, 1]) qml.RY(-0.4, wires=0) qml.CZ(wires=[1, 2]) return qml.expval(qml.grouping.string_to_pauli_word("ZZZ"))

>>> x = np.array(0.531, requires_grad=True) >>> circuit(x) 0.47165198882111165 >>> qml.grad(circuit)(x) -0.276982865449393

#### Grand QChem unification ⚛️ 🏰

Quantum chemistry functionality — previously split between an external

`pennylane-qchem`

package and internal`qml.hf`

differentiable Hartree-Fock solver — is now unified into a single, included,`qml.qchem`

module. (#2164) (#2385) (#2352) (#2420) (#2454)

(#2199) (#2371) (#2272) (#2230) (#2415) (#2426) (#2465)The

`qml.qchem`

module provides a differentiable Hartree-Fock solver and the functionality to construct a fully-differentiable molecular Hamiltonian.For example, one can continue to generate molecular Hamiltonians using

`qml.qchem.molecular_hamiltonian`

:symbols = ["H", "H"] geometry = np.array([[0., 0., -0.66140414], [0., 0., 0.66140414]]) hamiltonian, qubits = qml.qchem.molecular_hamiltonian(symbols, geometry, method="dhf")

By default, this will use the differentiable Hartree-Fock solver; however, simply set

`method="pyscf"`

to continue to use PySCF for Hartree-Fock calculations.Functions are added for building a differentiable dipole moment observable. Functions for computing multipole moment molecular integrals, needed for building the dipole moment observable, are also added. (#2173) (#2166)

The dipole moment observable can be constructed using

`qml.qchem.dipole_moment`

:symbols = ['H', 'H'] geometry = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 1.0]]) mol = qml.qchem.Molecule(symbols, geometry) args = [geometry] D = qml.qchem.dipole_moment(mol)(*args)

The efficiency of computing molecular integrals and Hamiltonian is improved. This has been done by adding optimized functions for building fermionic and qubit observables and optimizing the functions used for computing the electron repulsion integrals. (#2316)

The

`6-31G`

basis set is added to the qchem basis set repo. This addition allows performing differentiable Hartree-Fock calculations with basis sets beyond the minimal`sto-3g`

basis set for atoms with atomic number 1-10. (#2372)The

`6-31G`

basis set can be used to construct a Hamiltonian assymbols = ["H", "H"] geometry = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 1.0]]) H, qubits = qml.qchem.molecular_hamiltonian(symbols, geometry, basis="6-31g")

External dependencies are replaced with local functions for spin and particle number observables. (#2197) (#2362)

#### Pattern matching optimization 🔎 💎

Added an optimization transform that matches pieces of user-provided identity templates in a circuit and replaces them with an equivalent component. (#2032)

For example, consider the following circuit where we want to replace sequence of two

`pennylane.S`

gates with a`pennylane.PauliZ`

gate.def circuit(): qml.S(wires=0) qml.PauliZ(wires=0) qml.S(wires=1) qml.CZ(wires=[0, 1]) qml.S(wires=1) qml.S(wires=2) qml.CZ(wires=[1, 2]) qml.S(wires=2) return qml.expval(qml.PauliX(wires=0))

We specify use the following pattern that implements the identity:

with qml.tape.QuantumTape() as pattern: qml.S(wires=0) qml.S(wires=0) qml.PauliZ(wires=0)

To optimize the circuit with this identity pattern, we apply the

`qml.transforms.pattern_matching`

transform.>>> dev = qml.device('default.qubit', wires=5) >>> qnode = qml.QNode(circuit, dev) >>> optimized_qfunc = qml.transforms.pattern_matching_optimization(pattern_tapes=[pattern])(circuit) >>> optimized_qnode = qml.QNode(optimized_qfunc, dev) >>> print(qml.draw(qnode)()) 0: ──S──Z─╭C──────────┤ <X> 1: ──S────╰Z──S─╭C────┤ 2: ──S──────────╰Z──S─┤ >>> print(qml.draw(optimized_qnode)()) 0: ──S⁻¹─╭C────┤ <X> 1: ──Z───╰Z─╭C─┤ 2: ──Z──────╰Z─┤

For more details on using pattern matching optimization you can check the corresponding documentation and also the following paper.

#### Measure the distance between two unitaries📏

Added the

`HilbertSchmidt`

and the`LocalHilbertSchmidt`

templates to be used for computing distance measures between unitaries. (#2364)Given a unitary

`U`

,`qml.HilberSchmidt`

can be used to measure the distance between unitaries and to define a cost function (`cost_hst`

) used for learning a unitary`V`

that is equivalent to`U`

up to a global phase:# Represents unitary U with qml.tape.QuantumTape(do_queue=False) as u_tape: qml.Hadamard(wires=0) # Represents unitary V def v_function(params): qml.RZ(params[0], wires=1) @qml.qnode(dev) def hilbert_test(v_params, v_function, v_wires, u_tape): qml.HilbertSchmidt(v_params, v_function=v_function, v_wires=v_wires, u_tape=u_tape) return qml.probs(u_tape.wires + v_wires) def cost_hst(parameters, v_function, v_wires, u_tape): return (1 - hilbert_test(v_params=parameters, v_function=v_function, v_wires=v_wires, u_tape=u_tape)[0])

>>> cost_hst(parameters=[0.1], v_function=v_function, v_wires=[1], u_tape=u_tape) tensor(0.999, requires_grad=True)

For more information refer to the documentation of qml.HilbertSchmidt.

#### More tensor network support 🕸️

Adds the

`qml.MERA`

template for implementing quantum circuits with the shape of a multi-scale entanglement renormalization ansatz (MERA). (#2418)MERA follows the style of previous tensor network templates and is similar to quantum convolutional neural networks.

def block(weights, wires): qml.CNOT(wires=[wires[0],wires[1]]) qml.RY(weights[0], wires=wires[0]) qml.RY(weights[1], wires=wires[1]) n_wires = 4 n_block_wires = 2 n_params_block = 2 n_blocks = qml.MERA.get_n_blocks(range(n_wires),n_block_wires) template_weights = [[0.1,-0.3]]*n_blocks dev= qml.device('default.qubit',wires=range(n_wires)) @qml.qnode(dev) def circuit(template_weights): qml.MERA(range(n_wires),n_block_wires,block, n_params_block, template_weights) return qml.expval(qml.PauliZ(wires=1))

It may be necessary to reorder the wires to see the MERA architecture clearly:

>>> print(qml.draw(circuit,expansion_strategy='device',wire_order=[2,0,1,3])(template_weights)) 2: ───────────────╭C──RY(0.10)──╭X──RY(-0.30)───────────────┤ 0: ─╭X──RY(-0.30)─│─────────────╰C──RY(0.10)──╭C──RY(0.10)──┤ 1: ─╰C──RY(0.10)──│─────────────╭X──RY(-0.30)─╰X──RY(-0.30)─┤ <Z> 3: ───────────────╰X──RY(-0.30)─╰C──RY(0.10)────────────────┤

#### New transform for transpilation ⚙️

Added a swap based transpiler transform. (#2118)

The transpile function takes a quantum function and a coupling map as inputs and compiles the circuit to ensure that it can be executed on corresponding hardware. The transform can be used as a decorator in the following way:

dev = qml.device('default.qubit', wires=4) @qml.qnode(dev) @qml.transforms.transpile(coupling_map=[(0, 1), (1, 2), (2, 3)]) def circuit(param): qml.CNOT(wires=[0, 1]) qml.CNOT(wires=[0, 2]) qml.CNOT(wires=[0, 3]) qml.PhaseShift(param, wires=0) return qml.probs(wires=[0, 1, 2, 3])

>>> print(qml.draw(circuit)(0.3)) 0: ─╭C───────╭C──────────╭C──Rϕ(0.30)─┤ ╭Probs 1: ─╰X─╭SWAP─╰X────╭SWAP─╰X───────────┤ ├Probs 2: ────╰SWAP─╭SWAP─╰SWAP──────────────┤ ├Probs 3: ──────────╰SWAP────────────────────┤ ╰Probs

### Improvements

`QuantumTape`

objects are now iterable, allowing iteration over the contained operations and measurements. (#2342)with qml.tape.QuantumTape() as tape: qml.RX(0.432, wires=0) qml.RY(0.543, wires=0) qml.CNOT(wires=[0, 'a']) qml.RX(0.133, wires='a') qml.expval(qml.PauliZ(wires=[0]))

Given a

`QuantumTape`

object the underlying quantum circuit can be iterated over using a`for`

loop:>>> for op in tape: ... print(op) RX(0.432, wires=[0]) RY(0.543, wires=[0]) CNOT(wires=[0, 'a']) RX(0.133, wires=['a']) expval(PauliZ(wires=[0]))

Indexing into the circuit is also allowed via

`tape[i]`

:>>> tape[0] RX(0.432, wires=[0])

A tape object can also be converted to a sequence (e.g., to a

`list`

) of operations and measurements:>>> list(tape) [RX(0.432, wires=[0]), RY(0.543, wires=[0]), CNOT(wires=[0, 'a']), RX(0.133, wires=['a']), expval(PauliZ(wires=[0]))]

Added the

`QuantumTape.shape`

method and`QuantumTape.numeric_type`

attribute to allow extracting information about the shape and numeric type of the output returned by a quantum tape after execution. (#2044)dev = qml.device("default.qubit", wires=2) a = np.array([0.1, 0.2, 0.3]) def func(a): qml.RY(a[0], wires=0) qml.RX(a[1], wires=0) qml.RY(a[2], wires=0) with qml.tape.QuantumTape() as tape: func(a) qml.state()

>>> tape.shape(dev) (1, 4) >>> tape.numeric_type complex

Defined a

`MeasurementProcess.shape`

method and a`MeasurementProcess.numeric_type`

attribute to allow extracting information about the shape and numeric type of results obtained when evaluating QNodes using the specific measurement process. (#2044)The parameter-shift Hessian can now be computed for arbitrary operations that support the general parameter-shift rule for gradients, using

`qml.gradients.param_shift_hessian`

(#2319)Multiple ways to obtain the gradient recipe are supported, in the following order of preference:

A custom

`grad_recipe`

. It is iterated to obtain the shift rule for the second-order derivatives in the diagonal entries of the Hessian.Custom

`parameter_frequencies`

. The second-order shift rule can directly be computed using them.An operation’s

`generator`

. Its eigenvalues will be used to obtain`parameter_frequencies`

, if they are not given explicitly for an operation.

The strategy for expanding a circuit can now be specified with the

`qml.specs`

transform, for example to calculate the specifications of the circuit that will actually be executed by the device (`expansion_strategy="device"`

). (#2395)The

`default.qubit`

and`default.mixed`

devices now skip over identity operators instead of performing matrix multiplication with the identity. (#2356) (#2365)The function

`qml.eigvals`

is modified to use the efficient`scipy.sparse.linalg.eigsh`

method for obtaining the eigenvalues of a`SparseHamiltonian`

. This`scipy`

method is called to compute \(k\) eigenvalues of a sparse \(N \times N\) matrix if`k`

is smaller than \(N-1\). If a larger \(k\) is requested, the dense matrix representation of the Hamiltonian is constructed and the regular`qml.math.linalg.eigvalsh`

is applied. (#2333)The function

`qml.ctrl`

was given the optional argument`control_values=None`

. If overridden,`control_values`

takes an integer or a list of integers corresponding to the binary value that each control value should take. The same change is reflected in`ControlledOperation`

. Control values of`0`

are implemented by`qml.PauliX`

applied before and after the controlled operation (#2288)Operators now have a

`has_matrix`

property denoting whether or not the operator defines a matrix. (#2331) (#2476)Circuit cutting now performs expansion to search for wire cuts in contained operations or tapes. (#2340)

The

`qml.draw`

and`qml.draw_mpl`

transforms are now located in the`drawer`

module. They can still be accessed via the top-level`qml`

namespace. (#2396)Raise a warning where caching produces identical shot noise on execution results with finite shots. (#2478)

### Deprecations

The

`ObservableReturnTypes`

`Sample`

,`Variance`

,`Expectation`

,`Probability`

,`State`

, and`MidMeasure`

have been moved to`measurements`

from`operation`

. (#2329) (#2481)

### Breaking changes

The caching ability of devices has been removed. Using the caching on the QNode level is the recommended alternative going forward. (#2443)

One way for replicating the removed

`QubitDevice`

caching behaviour is by creating a`cache`

object (e.g., a dictionary) and passing it to the`QNode`

:n_wires = 4 wires = range(n_wires) dev = qml.device('default.qubit', wires=n_wires) cache = {} @qml.qnode(dev, diff_method='parameter-shift', cache=cache) def expval_circuit(params): qml.templates.BasicEntanglerLayers(params, wires=wires, rotation=qml.RX) return qml.expval(qml.PauliZ(0) @ qml.PauliY(1) @ qml.PauliX(2) @ qml.PauliZ(3)) shape = qml.templates.BasicEntanglerLayers.shape(5, n_wires) params = np.random.random(shape)

>>> expval_circuit(params) tensor(0.20598436, requires_grad=True) >>> dev.num_executions 1 >>> expval_circuit(params) tensor(0.20598436, requires_grad=True) >>> dev.num_executions 1

The

`qml.finite_diff`

function has been removed. Please use`qml.gradients.finite_diff`

to compute the gradient of tapes of QNodes. Otherwise, manual implementation is required. (#2464)The

`get_unitary_matrix`

transform has been removed, please use`qml.matrix`

instead. (#2457)The

`update_stepsize`

method has been removed from`GradientDescentOptimizer`

and its child optimizers. The`stepsize`

property can be interacted with directly instead. (#2370)Most optimizers no longer flatten and unflatten arguments during computation. Due to this change, user provided gradient functions

*must*return the same shape as`qml.grad`

. (#2381)The old circuit text drawing infrastructure has been removed. (#2310)

`RepresentationResolver`

was replaced by the`Operator.label`

method.`qml.drawer.CircuitDrawer`

was replaced by`qml.drawer.tape_text`

.`qml.drawer.CHARSETS`

was removed because unicode is assumed to be accessible.`Grid`

and`qml.drawer.drawable_grid`

were removed because the custom data class was replaced by list of sets of operators or measurements.`qml.transforms.draw_old`

was replaced by`qml.draw`

.`qml.CircuitGraph.greedy_layers`

was deleted, as it was no longer needed by the circuit drawer and did not seem to have uses outside of that situation.`qml.CircuitGraph.draw`

was deleted, as we draw tapes instead.The tape method

`qml.tape.QuantumTape.draw`

now simply calls`qml.drawer.tape_text`

.In the new pathway, the

`charset`

keyword was deleted, the`max_length`

keyword defaults to`100`

, and the`decimals`

and`show_matrices`

keywords were added.

The deprecated QNode, available via

`qml.qnode_old.QNode`

, has been removed. Please transition to using the standard`qml.QNode`

. (#2336) (#2337) (#2338)In addition, several other components which powered the deprecated QNode have been removed:

The deprecated, non-batch compatible interfaces, have been removed.

The deprecated tape subclasses

`QubitParamShiftTape`

,`JacobianTape`

,`CVParamShiftTape`

, and`ReversibleTape`

have been removed.

The deprecated tape execution method

`tape.execute(device)`

has been removed. Please use`qml.execute([tape], device)`

instead. (#2339)

### Bug fixes

Fixed a bug in the

`qml.PauliRot`

operation, where computing the generator was not taking into account the operation wires. (#2466)Fixed a bug where non-trainable arguments were shifted in the

`NesterovMomentumOptimizer`

if a trainable argument was after it in the argument list. (#2466)Fixed a bug with

`@jax.jit`

for grad when`diff_method="adjoint"`

and`mode="backward"`

. (#2460)Fixed a bug where

`qml.DiagonalQubitUnitary`

did not support`@jax.jit`

and`@tf.function`

. (#2445)Fixed a bug in the

`qml.PauliRot`

operation, where computing the generator was not taking into account the operation wires. (#2442)Fixed a bug with the padding capability of

`AmplitudeEmbedding`

where the inputs are on the GPU. (#2431)Fixed a bug by adding a comprehensible error message for calling

`qml.probs`

without passing wires or an observable. (#2438)The behaviour of

`qml.about()`

was modified to avoid warnings being emitted due to legacy behaviour of`pip`

. (#2422)Fixed a bug where observables were not considered when determining the use of the

`jax-jit`

interface. (#2427) (#2474)Fixed a bug where computing statistics for a relatively few number of shots (e.g.,

`shots=10`

), an error arose due to indexing into an array using a`Array`

. (#2427)PennyLane Lightning version in Docker container is pulled from latest wheel-builds. (#2416)

Optimizers only consider a variable trainable if they have

`requires_grad = True`

. (#2381)Fixed a bug with

`qml.expval`

,`qml.var`

,`qml.state`

and`qml.probs`

(when`qml.probs`

is the only measurement) where the`dtype`

specified on the device did not match the`dtype`

of the QNode output. (#2367)Fixed a bug where the output shapes from batch transforms are inconsistent with the QNode output shape. (#2215)

Fixed a bug caused by the squeezing in

`qml.gradients.param_shift_hessian`

. (#2215)Fixed a bug in which the

`expval`

/`var`

of a`Tensor(Observable)`

would depend on the order in which the observable is defined: (#2276)>>> @qml.qnode(dev) ... def circ(op): ... qml.RX(0.12, wires=0) ... qml.RX(1.34, wires=1) ... qml.RX(3.67, wires=2) ... return qml.expval(op) >>> op1 = qml.Identity(wires=0) @ qml.Identity(wires=1) @ qml.PauliZ(wires=2) >>> op2 = qml.PauliZ(wires=2) @ qml.Identity(wires=0) @ qml.Identity(wires=1) >>> print(circ(op1), circ(op2)) -0.8636111153905662 -0.8636111153905662

Fixed a bug where

`qml.hf.transform_hf()`

would fail due to missing wires in the qubit operator that is prepared for tapering the HF state. (#2441)Fixed a bug with custom device defined jacobians not being returned properly. (#2485)

### Documentation

The sections on adding operator and observable support in the “How to add a plugin” section of the plugins page have been updated. (#2389)

The missing arXiv reference in the

`LieAlgebra`

optimizer has been fixed. (#2325)

### Contributors

This release contains contributions from (in alphabetical order):

Karim Alaa El-Din, Guillermo Alonso-Linaje, Juan Miguel Arrazola, Ali Asadi, Utkarsh Azad, Sam Banning, Thomas Bromley, Alain Delgado, Isaac De Vlugt, Olivia Di Matteo, Amintor Dusko, Anthony Hayes, David Ittah, Josh Izaac, Soran Jahangiri, Nathan Killoran, Christina Lee, Angus Lowe, Romain Moyard, Zeyue Niu, Matthew Silverman, Lee James O’Riordan, Maria Schuld, Jay Soni, Antal Száva, Maurice Weber, David Wierichs.

- orphan

## Release 0.22.2¶

### Bug fixes

Most compilation transforms, and relevant subroutines, have been updated to support just-in-time compilation with jax.jit. This fix was intended to be included in

`v0.22.0`

, but due to a bug was incomplete. (#2397)

### Documentation

The documentation run has been updated to require

`jinja2==3.0.3`

due to an issue that arises with`jinja2`

`v3.1.0`

and`sphinx`

`v3.5.3`

. (#2378)

### Contributors

This release contains contributions from (in alphabetical order):

Olivia Di Matteo, Christina Lee, Romain Moyard, Antal Száva.

- orphan

## Release 0.22.1¶

### Bug fixes

Fixes cases with

`qml.measure`

where unexpected operations were added to the circuit. (#2328)

### Contributors

This release contains contributions from (in alphabetical order):

Guillermo Alonso-Linaje, Antal Száva.

- orphan

## Release 0.22.0¶

### New features since last release

#### Quantum circuit cutting ✂️

You can now run

`N`

-wire circuits on devices with fewer than`N`

wires, by strategically placing`WireCut`

operations that allow their circuit to be partitioned into smaller fragments, at a cost of needing to perform a greater number of device executions. Circuit cutting is enabled by decorating a QNode with the`@qml.cut_circuit`

transform. (#2107) (#2124) (#2153) (#2165) (#2158) (#2169) (#2192) (#2216) (#2168) (#2223) (#2231) (#2234) (#2244) (#2251) (#2265) (#2254) (#2260) (#2257) (#2279)The example below shows how a three-wire circuit can be run on a two-wire device:

dev = qml.device("default.qubit", wires=2) @qml.cut_circuit @qml.qnode(dev) def circuit(x): qml.RX(x, wires=0) qml.RY(0.9, wires=1) qml.RX(0.3, wires=2) qml.CZ(wires=[0, 1]) qml.RY(-0.4, wires=0) qml.WireCut(wires=1) qml.CZ(wires=[1, 2]) return qml.expval(qml.grouping.string_to_pauli_word("ZZZ"))

Instead of executing the circuit directly, it will be partitioned into smaller fragments according to the

`WireCut`

locations, and each fragment executed multiple times. Combining the results of the fragment executions will recover the expected output of the original uncut circuit.>>> x = np.array(0.531, requires_grad=True) >>> circuit(0.531) 0.47165198882111165

Circuit cutting support is also differentiable:

>>> qml.grad(circuit)(x) -0.276982865449393

For more details on circuit cutting, check out the qml.cut_circuit documentation page or Peng et. al.

#### Conditional operations: quantum teleportation unlocked 🔓🌀

Support for mid-circuit measurements and conditional operations has been added, to enable use cases like quantum teleportation, quantum error correction and quantum error mitigation. (#2211) (#2236) (#2275)

Two new functions have been added to support this capability:

`qml.measure()`

places mid-circuit measurements in the middle of a quantum function.`qml.cond()`

allows operations and quantum functions to be conditioned on the result of a previous measurement.

For example, the code below shows how to teleport a qubit from wire 0 to wire 2:

dev = qml.device("default.qubit", wires=3) input_state = np.array([1, -1], requires_grad=False) / np.sqrt(2) @qml.qnode(dev) def teleport(state): # Prepare input state qml.QubitStateVector(state, wires=0) # Prepare Bell state qml.Hadamard(wires=1) qml.CNOT(wires=[1, 2]) # Apply gates qml.CNOT(wires=[0, 1]) qml.Hadamard(wires=0) # Measure first two wires m1 = qml.measure(0) m2 = qml.measure(1) # Condition final wire on results qml.cond(m2 == 1, qml.PauliX)(wires=2) qml.cond(m1 == 1, qml.PauliZ)(wires=2) # Return state on final wire return qml.density_matrix(wires=2)

We can double-check that the qubit has been teleported by computing the overlap between the input state and the resulting state on wire 2:

>>> output_state = teleport(input_state) >>> output_state tensor([[ 0.5+0.j, -0.5+0.j], [-0.5+0.j, 0.5+0.j]], requires_grad=True) >>> input_state.conj() @ output_state @ input_state tensor(1.+0.j, requires_grad=True)

For a full description of new capabilities, refer to the Mid-circuit measurements and conditional operations section in the documentation.

Train mid-circuit measurements by deferring them, via the new

`@qml.defer_measurements`

transform. (#2211) (#2236) (#2275)If a device doesn’t natively support mid-circuit measurements, the

`@qml.defer_measurements`

transform can be applied to the QNode to transform the QNode into one with*terminal*measurements and*controlled*operations:dev = qml.device("default.qubit", wires=2) @qml.qnode(dev) @qml.defer_measurements def circuit(x): qml.Hadamard(wires=0) m = qml.measure(0) def op_if_true(): return qml.RX(x**2, wires=1) def op_if_false(): return qml.RY(x, wires=1) qml.cond(m==1, op_if_true, op_if_false)() return qml.expval(qml.PauliZ(1))

>>> x = np.array(0.7, requires_grad=True) >>> print(qml.draw(circuit, expansion_strategy="device")(x)) 0: ──H─╭C─────────X─╭C─────────X─┤ 1: ────╰RX(0.49)────╰RY(0.70)────┤ <Z> >>> circuit(x) tensor(0.82358752, requires_grad=True)

Deferring mid-circuit measurements also enables differentiation:

>>> qml.grad(circuit)(x) -0.651546965338656

#### Debug with mid-circuit quantum snapshots 📷

A new operation

`qml.Snapshot`

has been added to assist in debugging quantum functions. (#2233) (#2289) (#2291) (#2315)`qml.Snapshot`

saves the internal state of devices at arbitrary points of execution.Currently supported devices include:

`default.qubit`

: each snapshot saves the quantum state vector`default.mixed`

: each snapshot saves the density matrix`default.gaussian`

: each snapshot saves the covariance matrix and vector of means

During normal execution, the snapshots are ignored:

dev = qml.device("default.qubit", wires=2) @qml.qnode(dev, interface=None) def circuit(): qml.Snapshot() qml.Hadamard(wires=0) qml.Snapshot("very_important_state") qml.CNOT(wires=[0, 1]) qml.Snapshot() return qml.expval(qml.PauliX(0))

However, when using the

`qml.snapshots`

transform, intermediate device states will be stored and returned alongside the results.>>> qml.snapshots(circuit)() {0: array([1.+0.j, 0.+0.j, 0.+0.j, 0.+0.j]), 'very_important_state': array([0.70710678+0.j, 0. +0.j, 0.70710678+0.j, 0. +0.j]), 2: array([0.70710678+0.j, 0. +0.j, 0. +0.j, 0.70710678+0.j]), 'execution_results': array(0.)}

#### Batch embedding and state preparation data 📦

Added the

`@qml.batch_input`

transform to enable batching non-trainable gate parameters. In addition, the`qml.qnn.KerasLayer`

class has been updated to natively support batched training data. (#2069)As with other transforms,

`@qml.batch_input`

can be used to decorate QNodes:dev = qml.device("default.qubit", wires=2, shots=None) @qml.batch_input(argnum=0) @qml.qnode(dev, diff_method="parameter-shift", interface="tf") def circuit(inputs, weights): # add a batch dimension to the embedding data qml.AngleEmbedding(inputs, wires=range(2), rotation="Y") qml.RY(weights[0], wires=0) qml.RY(weights[1], wires=1) return qml.expval(qml.PauliZ(1))

Batched input parameters can then be passed during QNode evaluation:

>>> x = tf.random.uniform((10, 2), 0, 1) >>> w = tf.random.uniform((2,), 0, 1) >>> circuit(x, w) <tf.Tensor: shape=(10,), dtype=float64, numpy= array([0.46230079, 0.73971315, 0.95666004, 0.5355225 , 0.66180948, 0.44519553, 0.93874261, 0.9483197 , 0.78737918, 0.90866411])>

#### Even more mighty quantum transforms 🐛➡🦋

New functions and transforms of operators have been added:

`qml.matrix()`

for computing the matrix representation of one or more unitary operators. (#2241)`qml.eigvals()`

for computing the eigenvalues of one or more operators. (#2248)`qml.generator()`

for computing the generator of a single-parameter unitary operation. (#2256)

All operator transforms can be used on instantiated operators,

>>> op = qml.RX(0.54, wires=0) >>> qml.matrix(op) [[0.9637709+0.j 0. -0.26673144j] [0. -0.26673144j 0.9637709+0.j ]]

Operator transforms can also be used in a functional form:

>>> x = torch.tensor(0.6, requires_grad=True) >>> matrix_fn = qml.matrix(qml.RX) >>> matrix_fn(x, wires=[0]) tensor([[0.9553+0.0000j, 0.0000-0.2955j], [0.0000-0.2955j, 0.9553+0.0000j]], grad_fn=<AddBackward0>)

In its functional form, it is fully differentiable with respect to gate arguments:

>>> loss = torch.real(torch.trace(matrix_fn(x, wires=0))) >>> loss.backward() >>> x.grad tensor(-0.2955)

Some operator transform can also act on multiple operations, by passing quantum functions or tapes:

>>> def circuit(theta): ... qml.RX(theta, wires=1) ... qml.PauliZ(wires=0) >>> qml.matrix(circuit)(np.pi / 4) array([[ 0.92387953+0.j, 0.+0.j , 0.-0.38268343j, 0.+0.j], [ 0.+0.j, -0.92387953+0.j, 0.+0.j, 0. +0.38268343j], [ 0. -0.38268343j, 0.+0.j, 0.92387953+0.j, 0.+0.j], [ 0.+0.j, 0.+0.38268343j, 0.+0.j, -0.92387953+0.j]])

A new transform has been added to construct the pairwise-commutation directed acyclic graph (DAG) representation of a quantum circuit. (#1712)

In the DAG, each node represents a quantum operation, and edges represent non-commutation between two operations.

This transform takes into account that not all operations can be moved next to each other by pairwise commutation:

>>> def circuit(x, y, z): ... qml.RX(x, wires=0) ... qml.RX(y, wires=0) ... qml.CNOT(wires=[1, 2]) ... qml.RY(y, wires=1) ... qml.Hadamard(wires=2) ... qml.CRZ(z, wires=[2, 0]) ... qml.RY(-y, wires=1) ... return qml.expval(qml.PauliZ(0)) >>> dag_fn = qml.commutation_dag(circuit) >>> dag = dag_fn(np.pi / 4, np.pi / 3, np.pi / 2)

Nodes in the commutation DAG can be accessed via the

`get_nodes()`

method, returning a list of the form`(ID, CommutationDAGNode)`

:>>> nodes = dag.get_nodes() >>> nodes NodeDataView({0: <pennylane.transforms.commutation_dag.CommutationDAGNode object at 0x7f461c4bb580>, ...}, data='node')

Specific nodes in the commutation DAG can be accessed via the

`get_node()`

method:>>> second_node = dag.get_node(2) >>> second_node <pennylane.transforms.commutation_dag.CommutationDAGNode object at 0x136f8c4c0> >>> second_node.op CNOT(wires=[1, 2]) >>> second_node.successors [3, 4, 5, 6] >>> second_node.predecessors []

### Improvements

The text-based drawer accessed via

`qml.draw()`

has been optimized and improved. (#2128) (#2198)The new drawer has:

a

`decimals`

keyword for controlling parameter roundinga

`show_matrices`

keyword for controlling display of matricesa different algorithm for determining positions

deprecation of the

`charset`

keywordadditional minor cosmetic changes

@qml.qnode(qml.device('lightning.qubit', wires=2)) def circuit(a, w): qml.Hadamard(0) qml.CRX(a, wires=[0, 1]) qml.Rot(*w, wires=[1]) qml.CRX(-a, wires=[0, 1]) return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))

>>> print(qml.draw(circuit, decimals=2)(a=2.3, w=[1.2, 3.2, 0.7])) 0: ──H─╭C─────────────────────────────╭C─────────┤ ╭<Z@Z> 1: ────╰RX(2.30)──Rot(1.20,3.20,0.70)─╰RX(-2.30)─┤ ╰<Z@Z>

The frequencies of gate parameters are now accessible as an operation property and can be used for circuit analysis, optimization via the

`RotosolveOptimizer`

and differentiation with the parameter-shift rule (including the general shift rule). (#2180) (#2182) (#2227)>>> op = qml.CRot(0.4, 0.1, 0.3, wires=[0, 1]) >>> op.parameter_frequencies [(0.5, 1.0), (0.5, 1.0), (0.5, 1.0)]

When using

`qml.gradients.param_shift`

, either a custom`grad_recipe`

or the parameter frequencies are used to obtain the shift rule for the operation, in that order of preference.See Vidal and Theis (2018) and Wierichs et al. (2021) for theoretical background information on the general parameter-shift rule.

No two-term parameter-shift rule is assumed anymore by default. (#2227)

Previously, operations marked for analytic differentiation that did not provide a

`generator`

,`parameter_frequencies`

or a custom`grad_recipe`

were assumed to satisfy the two-term shift rule. This now has to be made explicit for custom operations by adding any of the above attributes.Most compilation transforms, and relevant subroutines, have been updated to support just-in-time compilation with

`jax.jit`

. (#1894)The

`qml.draw_mpl`

transform supports a`expansion_strategy`

keyword argument. (#2271)The

`qml.gradients`

module has been streamlined and special-purpose functions moved closer to their use cases, while preserving existing behaviour. (#2200)Added a new

`partition_pauli_group`

function to the`grouping`

module for efficiently measuring the`N`

-qubit Pauli group with`3 ** N`

qubit-wise commuting terms. (#2185)The Operator class has undergone a major refactor with the following changes:

**Matrices**: the static method`Operator.compute_matrices()`

defines the matrix representation of the operator, and the function`qml.matrix(op)`

computes this for a given instance. (#1996)**Eigvals**: the static method`Operator.compute_eigvals()`

defines the matrix representation of the operator, and the function`qml.eigvals(op)`

computes this for a given instance. (#2048)**Decompositions**: the static method`Operator.compute_decomposition()`

defines the matrix representation of the operator, and the method`op.decomposition()`

computes this for a given instance. (#2024) (#2053)**Sparse matrices**: the static method`Operator.compute_sparse_matrix()`

defines the sparse matrix representation of the operator, and the method`op.sparse_matrix()`

computes this for a given instance. (#2050)**Linear combinations of operators**: The static method`compute_terms()`

, used for representing the linear combination of coefficients and operators representing the operator, has been added. The method`op.terms()`

computes this for a given instance. Currently, only the`Hamiltonian`

class overwrites`compute_terms()`

to store coefficients and operators. The`Hamiltonian.terms`

property hence becomes a proper method called by`Hamiltonian.terms()`

. (#2036)**Diagonalization**: The`diagonalizing_gates()`

representation has been moved to the highest-level`Operator`

class and is therefore available to all subclasses. A condition`qml.operation.defines_diagonalizing_gates`

has been added, which can be used in tape contexts without queueing. In addition, a static`compute_diagonalizing_gates`

method has been added, which is called by default in`diagonalizing_gates()`

. (#1985) (#1993)Error handling has been improved for Operator representations. Custom errors subclassing

`OperatorPropertyUndefined`

are raised if a representation has not been defined. This replaces the`NotImplementedError`

and allows finer control for developers. (#2064) (#2287)A

`Operator.hyperparameters`

attribute, used for storing operation parameters that are*never*trainable, has been added to the operator class. (#2017)The

`string_for_inverse`

attribute is removed. (#2021)The

`expand()`

method was moved from the`Operation`

class to the main`Operator`

class. (#2053) (#2239)

### Deprecations

There are several important changes when creating custom operations: (#2214) (#2227) (#2030) (#2061)

The

`Operator.matrix`

method has been deprecated and`Operator.compute_matrix`

should be defined instead. Operator matrices should be accessed using`qml.matrix(op)`

. If you were previously defining the class method`Operator._matrix()`

, this is a a**breaking change**— please update your operation to instead overwrite`Operator.compute_matrix`

.The

`Operator.decomposition`

method has been deprecated and`Operator.compute_decomposition`

should be defined instead. Operator decompositions should be accessed using`Operator.decomposition()`

.The

`Operator.eigvals`

method has been deprecated and`Operator.compute_eigvals`

should be defined instead. Operator eigenvalues should be accessed using`qml.eigvals(op)`

.The

`Operator.generator`

property is now a method, and should return an*operator instance*representing the generator. Note that unlike the other representations above, this is a**breaking change**. Operator generators should be accessed using`qml.generator(op)`

.The

`Operation.get_parameter_shift`

method has been deprecated and will be removed in a future release.Instead, the functionalities for general parameter-shift rules in the

`qml.gradients`

module should be used, together with the operation attributes`parameter_frequencies`

or`grad_recipe`

.

Executing tapes using

`tape.execute(dev)`

is deprecated. Please use the`qml.execute([tape], dev)`

function instead. (#2306)The subclasses of the quantum tape, including

`JacobianTape`

,`QubitParamShiftTape`

,`CVParamShiftTape`

, and`ReversibleTape`

are deprecated. Instead of calling`JacobianTape.jacobian()`

and`JacobianTape.hessian()`

, please use a standard`QuantumTape`

, and apply gradient transforms using the`qml.gradients`

module. (#2306)`qml.transforms.get_unitary_matrix()`

has been deprecated and will be removed in a future release. For extracting matrices of operations and quantum functions, please use`qml.matrix()`

. (#2248)The

`qml.finite_diff()`

function has been deprecated and will be removed in an upcoming release. Instead,`qml.gradients.finite_diff()`

can be used to compute purely quantum gradients (that is, gradients of tapes or QNode). (#2212)The

`MultiControlledX`

operation now accepts a single`wires`

keyword argument for both`control_wires`

and`wires`

. The single`wires`

keyword should be all the control wires followed by a single target wire. (#2121) (#2278)

### Breaking changes

The representation of an operator as a matrix has been overhauled. (#1996)

The “canonical matrix”, which is independent of wires, is now defined in the static method

`compute_matrix()`

instead of`_matrix`

. By default, this method is assumed to take all parameters and non-trainable hyperparameters that define the operation.>>> qml.RX.compute_matrix(0.5) [[0.96891242+0.j 0. -0.24740396j] [0. -0.24740396j 0.96891242+0.j ]]

If no canonical matrix is specified for a gate,

`compute_matrix()`

raises a`MatrixUndefinedError`

.The generator property has been updated to an instance method,

`Operator.generator()`

. It now returns an instantiated operation, representing the generator of the instantiated operator. (#2030) (#2061)Various operators have been updated to specify the generator as either an

`Observable`

,`Tensor`

,`Hamiltonian`

,`SparseHamiltonian`

, or`Hermitian`

operator.In addition,

`qml.generator(operation)`

has been added to aid in retrieving generator representations of operators.The argument

`wires`

in`heisenberg_obs`

,`heisenberg_expand`

and`heisenberg_tr`

was renamed to`wire_order`

to be consistent with other matrix representations. (#2051)The property

`kraus_matrices`

has been changed to a method, and`_kraus_matrices`

renamed to`compute_kraus_matrices`

, which is now a static method. (#2055)The

`pennylane.measure`

module has been renamed to`pennylane.measurements`

. (#2236)

### Bug fixes

The

`basis`

property of`qml.SWAP`

was set to`"X"`

, which is incorrect; it is now set to`None`

. (#2287)The

`qml.RandomLayers`

template now decomposes when the weights are a list of lists. (#2266)The

`qml.QubitUnitary`

operation now supports just-in-time compilation using JAX. (#2249)Fixes a bug in the JAX interface where

`Array`

objects were not being converted to NumPy arrays before executing an external device. (#2255)The

`qml.ctrl`

transform now works correctly with gradient transforms such as the parameter-shift rule. (#2238)Fixes a bug in which passing required arguments into operations as keyword arguments would throw an error because the documented call signature didn’t match the function definition. (#1976)

The operation

`OrbitalRotation`

previously was wrongfully registered to satisfy the four-term parameter shift rule. The correct eight-term rule will now be used when using the parameter-shift rule. (#2180)Fixes a bug where

`qml.gradients.param_shift_hessian`

would produce an error whenever all elements of the Hessian are known in advance to be 0. (#2299)

### Documentation

The developer guide on adding templates and the architecture